PMCID
string | Title
string | FullText
string |
|---|---|---|
PMC4606567
|
An automated shotgun lipidomics platform for high throughput, comprehensive, and quantitative analysis of blood plasma intact lipids
|
Blood plasma has gained protagonism in lipidomics studies due to its availability, uncomplicated collection and preparation, and informative readout of physiological status. At the same time, it is also technically challenging to analyze due to its complex lipid composition affected by many factors, which can hamper the throughput and/or lipidomics coverage. To tackle these issues, we developed a comprehensive, high throughput, and quantitative mass spectrometry-based shotgun lipidomics platform for blood plasma lipid analyses. The main hallmarks of this technology are (i) it is comprehensive, covering 22 quantifiable different lipid classes encompassing more than 200 lipid species; (ii) it is amenable to high-throughput, with less than 5 min acquisition time allowing the complete analysis of 200 plasma samples per day; (iii) it achieves absolute quantification, by inclusion of internal standards for every lipid class measured; (iv) it is highly reproducible, achieving an average coefficient of variation of <10% (intra-day), approx. 10% (inter-day), and approx. 15% (inter-site) for most lipid species; (v) it is easily transferable allowing the direct comparison of data acquired in different sites. Moreover, we thoroughly assessed the influence of blood stabilization with different anticoagulants and freeze-thaw cycles to exclude artifacts generated by sample preparation. Practical applications: This shotgun lipidomics platform can be implemented in different laboratories without compromising reproducibility, allowing multi-site studies and inter-laboratory comparisons. This possibility combined with the high-throughput, broad lipidomic coverage and absolute quantification are important aspects for clinical applications and biomarker research.Lipidomics is the systematic study of pathways and networks of cellular lipids by profiling and understanding the role of lipids in biological systems 1. Blood plasma is widely used in lipidomics studies mainly due to its availability. Furthermore, it carries the lipids incorporated into lipoproteins over the circulatory system from the intestine and liver to all the tissues in the body 2. Because this fluid has access to peripheral tissues, and its lipoproteins exchange lipids with them, its lipid composition should contain a detailed picture of the individual metabolic state 3,4. This makes plasma a prime target for diagnostic research but at the same time very challenging due to its compositional complexity 5. In order to establish reliable lipid diagnostic biomarkers, one needs to dissect the many factors affecting the plasma lipidome which are not necessarily disease-related, e.g., different genetic backgrounds, diet, gender, age, and life style 6–11. This creates a severe problem and often plasma lipidomics studies are either performed in a high-throughput manner but targeting only a subset of the lipidome 12 or they cover the whole lipidome for a small sample set 5. In lipidomics methodologies, like in other omics techniques, there is an inverse correlation between a method's throughput versus lipidomic structural elucidation and coverage 13. The most comprehensive plasma lipidomic coverage available, offered by the LipidMaps consortium 5, is achieved by the combination of multiple analytical set-ups running in different laboratories 14 which greatly reduces the throughput and feasibility. On the other hand, one-step chromatography-based approaches offer improved throughput at the expense of lower lipidomic coverage 15–19. Nowadays, the available shotgun lipidomics methodologies offer the highest throughput. This is achieved by direct infusion of the extract without previous chromatographic separation with the drawbacks of not providing details about lipid molecular species 20 or limited lipidomic coverage 21. In this paper, we developed a high throughput mass spectrometry (MS)-based shotgun lipidomics platform that offers quantitative and extended lipidomic coverage down to the molecular lipid species level for systematic lipidomic profiling of large populations. Our platform achieves unprecedented acquisition speed (5 min per acquisition, allowing the analysis of 200 samples per day per MS instrument including sample preparation) combined with high lipidomic coverage (22 quantifiable lipid classes encompassing more than 200 lipid species) in a single acquisition. In order to achieve this performance, we optimized the sample preparation, MS acquisition, and lipid identification approaches. All sample preparation steps were automatized for increased throughput and precision. On the instrumentation side, we took advantage of the configuration of the quadrupole-Orbitrap hardware that allows fast polarity switching, safe precursor selection for MSMS fragmentation analysis, and high mass accuracy and resolution. The combination of these features allows for obtaining multiple complementary MS scans in the same acquisition that enabled us to cover the full spectrum of intact lipids. The increase in sample number and spectral complexity (inclusion of MSMS spectra) greatly inflates the information to be processed. To handle this, we developed a new approach for individual molecular species identification and quantification. Lipid species from high resolution MS spectrum are identified and quantified followed by deconvolution of these lipid species into the different lipid molecular species using the corresponding tandem MS fragments. In addition to the high coverage and throughput, this technology is also very robust. We observed very small technical variation within the same day, between different days of acquisition and between different acquisition sites. We present, for the first time, a technology so robust that it can be implemented in other laboratories without any compromise in precision. These are important features required in clinical diagnostics screens. Moreover, we assess the impact of sample collection and storage on the plasma lipidome stability. Water, propan-2-ol, and methanol were purchased from Fischer Scientific. Methyl tert-butyl ether, chloroform, ammonium bicarbonate, and ammonium acetate were purchased from Sigma–Aldrich. All chemicals were analytical grade. Synthetic lipid standards were purchased from Avanti Polar Lipids, Larodan Fine Chemicals, and Sigma–Aldrich. For plasma isolation, blood was collected and centrifuged (2000×g, 10 min) and the supernatant was collected 22. Prior to extraction, the plasma was diluted 1:50 v/v by mixing 15 μL of it with 735 μL of 150 mM ammonium bicarbonate aqueous solution, aliquoted, and stored at −80°C. Since we use 1 μL of blood plasma per extraction, this dilution is required to minimize the error arising from handling low volumes of liquid. The whole procedure was finished in 2 h after blood collection. It was used to prepare the reference samples, where plasma from three healthy, unfasted donors was combined in equivolumetric ratios separated in batches according to the anticoagulant used. For all analysis, except the comparison of anticoagulants, the plasma derived from EDTA-stabilized blood samples was used and all samples measured were coming from the same batch of reference samples. For the anticoagulation comparison, blood was collected into S-Monovette EDTA K3, sodium citrate, and lithium–heparin (all Sarstedt) anticoagulant vacutainers (each donor donated blood to all three containers) according to the producer's manual. The data were corrected for the initial dilution of blood by citrate (blood to citrate 10:1 ratio; v/v). In the freeze and thaw assessment, the samples were prepared as described above and were thawed at 4°C for 2 h, mixed by vortexing, and frozen by placing back at −80°C for 24 h. One freeze and thaw cycle was performed on a daily basis and immediately after all cycles were completed samples were extracted and analyzed. The lipid extraction (adapted from Matyash et al. 23) was carried out in high grade polypropylene deep well plates. Fifty microliters of diluted plasma (50×) (equivalent of 1 μL of undiluted plasma) was mixed with 130 μL of ammonium bicarbonate solution and 810 μL of methyl tert-butyl ether/methanol (7:2, v/v) solution was added. Twenty-one microliters of internal standard mixture was pre-mixed with the organic solvents mixture. The internal standard mixture contained: 50 pmol of lysophasphatidylglycerol (LPG) 17:1, 50 pmol of lysophosphatic acid (LPA) 17:0, 500 pmol of phosphatidylcholine (PC) 17:0/17:0, 30 pmol of hexosylceramide (HexCer) 18:1;2/12:0, 50 pmol of phosphatidylserine (PS) 17:0/17:0, 50 pmol of phosphatidylglycerol (PG) 17:0/17:0, 50 pmol of phosphatic acid (PA) 17:0/17:0, 50 pmol of lysophposphatidylinositol (LPI 17:1), 50 pmol of lysophosphatidylserine (LPS) 17:1, 1 nmol cholesterol (Chol) D6, 100 pmol of diacylglycerol (DAG) 17:0/17:0, 50 pmol of triacylglycerol (TAG) 17:0/17:0/17:0, 50 pmol of ceramide (Cer) 18:1;2/17:0, 200 pmol of sphingomyelin (SM) 18:1;2/12:0, 50 pmol of lysophosphatidylcholine (LPC) 12:0, 30 pmol of lysophosphatidylethanolamine (LPE) 17:1, 50 pmol of phosphatidylethanolamine (PE) 17:0/17:0, 100 pmol of cholesterol ester (CE) 20:0, 50 pmol of phosphatidylinositol (PI) 16:0/16:0. The plate was then sealed with a teflon-coated lid, shaken at 4°C for 15 min, and spun down (3000 g, 5 min) to facilitate separation of the liquid phases and clean-up of the upper organic phase. Hundred microliters of the organic phase was transferred to an infusion plate and dried in a speed vacuum concentrator. Dried lipids were re-suspended in 40 μL of 7.5 mM ammonium acetate in chloroform/methanol/propanol (1:2:4, v/v/v) and the wells were sealed with an aluminum foil to avoid evaporation and contamination during infusion. All liquid handling steps were performed using Hamilton STARlet robotic platform with the Anti Droplet Control feature for organic solvents pipetting. Samples were analyzed by direct infusion in a QExactive mass spectrometer (Thermo Fisher Scientific) equipped with a TriVersa NanoMate ion source (Advion Biosciences). Five microliters were infused with gas pressure and voltage set to 1.25 psi and 0.95 kV, respectively. The delivery time was set to 4 min and 55 s with contact closure delay of 20 s to avoid initial spray instability. Polarity switch from positive to negative mode was set at 135 s after contact closure. Samples were analyzed in both polarities in a single acquisition. The MS acquisition method starts with positive ion mode by acquiring the m/z 402–412 in MS + mode at Rm/z = 200 = 140 000 to monitor the [Chol + NH4] ion for 12 s. All individual scans in every segment are the average of 2 microscans. Automatic gain control (AGC) was set to 5 × 10 and maximum ion injection time (IT) was set to 200 ms. Then we scan the m/z 550–1000 in MS + (Rm/z = 200 = 140 000) with lock mass activated at a common background (m/z = 680.48022) for 18 s. AGC was set to 10 and IT was set to 50 ms. This is followed by a MSMS + (Rm/z = 200 = 17 500) data independent analysis triggered by an inclusion list for 105 s. The inclusion list contains all the masses from 500.5 to 999.75 with 1 Da intervals. AGC was set to 10 and IT was set to 64 ms. The isolation width was set to 1.0 Da, first mass of MSMS acquisition was 250 Da and normalized collision energy was set to 20%. Both MS+ and MSMS+ data are combined to monitor SE, DAG, and TAG ions as ammonium adducts. After polarity switch to negative ion mode, a lag of 15 s before acquisition was inserted to allow spray stabilization. Then, we scan for the m/z 400–650 in FTMS − (Rm/z = 200 = 140 000) for 15 s with lock mass activated at a common background (m/z = 529.46262) to monitor LPG, LPA, LPI, LPS, and LPE as deprotonated anions and LPC and LPC O– as acetate adducts. AGC was set to 10 and IT was set to 50 ms. We then scan the m/z 520–940 in FTMS − (Rm/z = 200 = 140 000) for 15 s with lock mass activated at a common background (m/z = 529.46262). AGC was set to 10 and IT was set to 50 ms. Finally, we scan MSMS- (Rm/z= 200 = 17 500) by data independent analysis triggered by an inclusion list for 90 s. This inclusion list contains all the masses from 590.5 to 939.5 with 1 Da intervals. AGC was set to 10 and IT was set to 64 ms. Isolation width was set to 1.0 Da, first mass of MSMS acquisition was 150 Da, and normalized collision energy was set to 35%. Both MS and MSMS data were combined in order to monitor PC, PC O–, HexCer, Cer, SM as acetate adducts and PS, PG, PA, PE, PE O–, and PI as deprotonated anions. All data were analyzed with an in-house developed lipid identification software based on LipidXplorer 24,25. All Molecular Fragmentation Query Language queries used in this work are available and can be found in Supplementary Material 1. Tolerance for MS and MSMS identification was set to 2 ppm in scans where we have lock mass activated and 8 ppm when lock mass was not available. Data post-processing and normalization were performed using an in-house developed data management system. Data visualization, linear regression (linear least squares method), and correlation (two-tailed Pearson correlation) calculations were performed on Prism 6.0e software (GraphPad Software, Inc.). Automation is essential in order to handle large number of samples because it increases the speed and throughput of the method, makes it more cost-efficient, and most importantly it improves reproducibility 26 (see section 3.5). To this end, we chose the Hamilton Robotics STARlet system. It allows pipetting of wide volume ranges (1–1000 μL) at high precision with the Anti Droplet Control function limiting organic phase dripping from pipette tips. Remarkably, most lipidomics studies still use one of almost 60 years old classical extraction protocols at times with minor modifications 27,28. More recently, Matyash et al. 23 introduced a new high throughput oriented extraction procedure where chloroform is replaced by methyltert-butyl ether (MTBE). This extraction achieves high recovery of multiple lipid classes with several important advantages 23. MTBE, compared to chloroform, is non-toxic and non-carcinogenic. The organic lipid-containing phase remains on top of the aqueous phase, reducing the water-soluble contaminants during extract collection. All these factors put together greatly facilitate automated sample handling procedures. For this study, we defined the optimal time of MTBE-based lipid extraction at which maximum lipid recovery is achieved. To this end, a mixture containing standards representing the most abundant lipid classes was added post-extraction to blood plasma lipids extracted for different times. Surprisingly, we observed that in less than 2 min, equilibrium is reached for all lipid classes tracked (Fig. S1). However, the recovery was more reproducible after 5 min. For this reason, we decided that extracting for 15 min is the best compromise between extraction reproducibility and speed. The quadrupole-Orbitrap configuration allows multiple possibilities for lipid identification. Besides the consecutive acquisition in both positive and negative ion mode 20, some lipids can be detected by a top-down approach where the unambiguous identification is possible due to the high resolution of the detector, relying solely on the precursor mass accuracy 29. In order to increase the identification specificity, it can also be combined to a bottom-up approach by introducing fragmentation experiments made possible by the presence of the quadrupole. Moreover, this bottom-up approach allows additional structural elucidation which enables molecular lipid species identification 30,31. To date, molecular lipid species identification and quantification required multiple rounds of long tandem MS acquisitions in order to obtain the statistics for an accurate measurement. To bypass this and decrease the acquisition time, we propose a new approach that takes advantage of the Q-Orbitrap features. Firstly, from the FTMS high-resolution spectrum, we normalize the de-isotoped (type I and II) intensity of the monoisotopic peak of an endogenous species to the de-isotoped (type I and II) intensity of the monoisotopic peak of the standard of the same class to obtain absolute quantification at the species level: where the concentration of a given lipid species [lip spec] is given by the ratio of its de-isotoped monoisotopic intensity (I(lip spec)MS) to the added lipid standard de-isotoped monoisotopic intensity (I(lip std)MS) multiplied by the concentration of the lipid standard [lip std]. After quantification at the species level, the amounts of all x molecular phospholipid species with overlapping sum compositions, can then be deconvoluted from the n acyl anions fragmentation information contained in the tandem MS data, as follows: where the concentration of a given molecular species x, [mol spec]x, is given by the sum of the intensities of its complementary acyl anions FAx1 and FAx2, divided by the sum of the intensities of n acyl anions from all other isobaric lipid molecular species present in the same MSMS scan, which corresponds to the molar fraction of each lipid molecular species. With this approach, we circumvent the need for multiple cycles of MSMS in order to achieve accurate quantification greatly reducing the time of acquisition. Phospholipids display variable but in general efficient fragmentation that allows the elucidation of the fatty acid composition 31, but other lipid classes can be more difficult to assess. For example, because triacylglycerides contain combinations of three fatty acids, this makes the assignment of molecular lipid species without chromatographic separation and MS type of experiments virtually impossible 32. Also, the fragmentation of sphingolipids is very inefficient and not suitable for structure elucidation. The most notable case is sphingomyelin. It presents only one abundant phosphocholine fragment in positive mode that does not give any additional structural information. Conversely, in negative ion mode there is a fragment corresponding to the loss of the amide-linked fatty acid allowing the molecular lipid species elucidation but it cannot be used for identification due to its low intensity. In Table 1, we present a strategy to identify each lipid class and the level of structural detail that can be achieved. Mode of acquisition and identification (see section 2.4 for details), structural detail, optimal sample amounts and their r, dynamic range, its quantification slopes and their r, LOQ for every lipid class (see Figs. S1 and S3 for additional information) Although MSMS was used to increase the confidence of identification, we could not assign molecular species for the TAG (see text for details). Not measurable in this reference sample. One of the biggest challenges of comprehensive lipid analysis is the huge dynamic range of lipid abundances observed in biological samples like blood plasma. Chromatographic techniques minimize the ion suppression effect of highly abundant lipid species by resolving the different lipids in a temporal dimension 33. This allows for extracting a wide range of sample amounts that to some extent can be tuned for the detection of low abundant species. In a shotgun lipidomics approach, sample infusion is continuous and the acquisition is done simultaneously for all the different lipid classes, which makes it more susceptible to ion suppression 34. To minimize this, the fine tuning of the sample amount to be extracted is of paramount importance for shotgun-based experiments 21. We extracted different sample amounts together with fixed amounts of internal standards in order to determine the minimum and maximum amounts that give reliable lipid compositions. We observed that most lipid classes give a linear response over two orders of magnitude of sample amount (0.2–20 μL) (Fig. S2). Importantly, we observe a consensual sample range (1–2 μL), which provides a satisfying compromise between lipidomic coverage and sensitivity, and where reproducibility is the highest (Table1 and Fig. S1). It is important to note that this amount is optimized for healthy individuals and needs to be carefully assessed in metabolically challenged conditions where the concentration of specific lipid classes such as triglycerides in blood can increase significantly. This sample amount corresponds to a final sample dilution in the infusate of 1/320 to 1/160 which is in agreement with a previous study for the optimal sample amount for shotgun lipidomics experiments 21. For quantification, we used one internal standard per lipid class. These internal standards were chosen based on two criteria: first, its absence in plasma samples and second, its similarity to the structure and properties (e.g., extraction recovery and ionization efficiency) of the analytes. This approach is valid due to the fact that the recovery and ionization efficiency of different lipid species within a lipid class is predominantly dependent on their charged head group while their differential acyl chains only minimally alter their ionization at low lipid concentrations 35,36 (Fig. S3). Two notable exceptions are present in the SM profile. SM 42:1;2 apparent relative decrease with sample amount increase is explained by a small background peak of a similar overlapping mass observed already in the blank samples. Importantly, this effect is only significant at very low sample amounts and its impact at the optimal sample amount is minimal. For SM 34:1;2, we observed a relative increase up to 15% in its proportion across the two orders of magnitude of sample amount. Although this is a significant difference, it is not observed in other SM species. This effect is most likely related to differential matrix effects introduced by variable amounts of plasma sample extracted and its impact is irrelevant when we fix the sample amount. In conclusion, we obtain identical lipid species profiles at different sample concentrations, confirming the assumption that ionization within a lipid class is not lipid species specific for the sample amounts tested allowing to express the quantities in molar amounts. This observation allows us to conclude that, for the optimal sample amount, all lipid species measured within a lipid class have similar response factors enabling absolute quantification with a single standard addition per lipid class. Different lipid classes exhibit different extraction, ionization, and fragmentation efficiencies, so the limit of quantification (LOQ) and dynamic ranges are lipid class-specific. In order to determine all these parameters, we added different amounts of each lipid class standard to the fixed optimal plasma sample amount (1 μL) and we recorded how the signal acquired correlates with the standard amount added. The LOQ was defined by the minimum concentration at which the slope and the correlation coefficient (r) are not significantly compromised and the variation of the standard intensity is below 20%. The summary of the results obtained can be found in Table1 and Fig. S3. We observed that the lipid classes showing a higher LOQ (>1 µM) correspond to abundant lipid classes in plasma (e.g., Chol, SE, and TAG); therefore, a higher detection limit should not have a significant impact on their species. Interestingly, they are all acquired in positive ion mode as an ammonium adduct, so an increase in the ammonium acetate concentration in the MS-infusion mixture can potentially improve their detection sensitivity, if required. On the other hand, anionic lipids that are usually acquired in negative ion mode, exhibit LOQs down to 50 nM. We conclude that the absence or low number of lipid species identified in some classes such as HexCer, LPA, LPS, and PA is due to their low abundance in this particular sample rather than low sensitivity for these lipid classes. Reproducibility is the ability to obtain similar results by the same researcher on different days and/or by someone else working independently. In this work, we aimed at maximizing method reproducibility by automation of most of the sample preparation, extraction, data acquisition and, to some extent, lipid identification, and post-processing of the data. In order to make a thorough assessment of the reproducibility of our method, we performed a systematic series of experiments where we aliquoted the same reference sample 540 times across six independent 96-well plates. Three of these plates (270 aliquots in total) were subjected on three different days (90 aliquots per day) to extraction and analysis in one laboratory and the other three plates were processed similarly in another laboratory. Both sites used identical platforms for the sample processing and analysis. Since every sample is an aliquot of the same batch of plasma (see section 2.2), the variation observed across different acquisitions, accurately reflects the intra-plate variation (variation within the same day of acquisition), inter-day variation (variation on different days of acquisition), and inter-site variation (variation between different sites). It is important to note that in analytical methods, there is usually an inverse correlation between lipid amount present in the sample and the coefficient of variation 21. Because different lipid species can range in amounts by several orders of magnitude, one has to assess how the amount affects the reproducibility of a given molecular lipid species. To assess method variation, we sorted the lipids by abundance and divided the data in quartiles and averaged the coefficient of variation within each quartile (Fig. 1A and Table2). (A) Correlation between coefficients of variation and average lipid concentrations from 270 individual measurements performed on three different days. The dashed lines separate the quartiles according to lipid concentration (see Table2 for additional information). (B) Pearson correlation plot of averaged concentrations determined at the two sites. Every point represents the average concentration of lipid species measured from 270 independent acquisitions. Correlation coefficient (r) is given. Averaged coefficients of variation within the different lipid concentration quartiles obtained on the same day (intra-plate), on different days (inter-day), and in different sites (inter-site) (see Fig. 1 for additional information) We observed that the coefficient of variation indeed correlates inversely with the lipid amount but importantly we still obtain low CVs when the plates are processed on different days and in different laboratories (Table2 and Fig. 1B). It is important to note that, although the first and second quartiles correspond to half the number of lipid species analyzed, when taking in consideration their amounts, they correspond to 98% of the lipidome and they can be measured with CVs lower than 15% on average even when acquired at different sites. For the third and fourth quartiles, the least abundant species, we can also conclude that the higher CV can be attributed to their intrinsic low abundance and not to any irreproducibility bias. From the sample collection to spectral acquisition, there are several steps that might affect the state of the sample and as a consequence alter the final result. During the sample collection, it is necessary to add anticoagulants to the blood sample when plasma preparation is considered. However, throughout the literature, different anticoagulants have been used indiscriminately without any verification of the impact that they might have on the lipid analysis, though it is known that they can influence multiple parameters of blood 37. To assess if different anticoagulants used for blood stabilization affected the plasma lipidome, we compared plasma samples derived from blood stabilized with the three most commonly used substances: EDTA, citrate, and heparin. All samples, regardless of the anticoagulant used, yielded remarkably comparable lipidomes, with Pearson correlation coefficients exceeding 0.999 and had almost identical lipid class profiles (Fig. 2). Interestingly, we observed that citrate containing samples display systematically lower lipid amounts when compared to samples containing other anticoagulants. This is due to the fact that citrate in vacutainers is present as an aqueous solution and is mixed with the blood during drawing, therefore, diluting it. Although we corrected for the dilution to the best of our ability, we did not manage to retrieve results identical to the other anticoagulants. This has been observed previously 38, which makes us conclude that citrate should be avoided if possible due to the introduction of uncertainty in the sample volume determination. Effect of anticoagulants on the blood plasma lipidome analysis. (A) Lipid class profile comparison of plasma collected with EDTA, citrate, and heparin as anticoagulants. Averaged values are shown with error bars depicting standard deviation of 27 independent experiments for each anticoagulant. (B) Pearson correlation of the lipids species quantified. Every point represents the average concentration of lipid species calculated from 27 independent experiments for each condition. Correlation coefficients (r) versus EDTA are given. In a given study, it might be useful to re-analyze the same sample in order to obtain additional information. As it is often assumed, freeze-thawing cycles might influence sample and analyte properties: in this particular case, the lipids. We evaluated how freezing and thawing affected the plasma lipidome by performing 10 freeze-thaw cycles prior to extraction. We did not detect any significant systematic alterations in lipid levels, nor in molecular species composition of lipid classes that could be correlated to the number of freeze and thaw cycles (Fig. 3A). More importantly, we did not observe any decrease in the unsaturation level of fatty acids present in lipids, which would be a sign of oxidation of polyunsaturated fatty acids (Fig. 3B). These results corroborate previous observations 39,40. It is worth to mention that the definition of the maximum number of freeze and thaw cycles without sample damage is still controversial 41. We believe that the controversy might be related more to the specific conditions of freeze-thaw cycles in different studies than to the number of cycles performed. With this in mind, instead of claiming that one can perform 10 freeze and thaw cycles without sample degradation, we would rather emphasize that thawing the samples at 4°C as quickly as possible, should allow for at least 10 cycles without sample degradation as has been shown in this study. Taken together, these results provide guidelines for consistent sample collection, storage, and re-acquisition. However, although the effects of sample collection and freeze-thawing on the plasma lipidome were ruled out, other factors leading to degradation of lipids cannot be excluded. For instance, it is known that physical stress imposed on leukocytes during blood drawing may activate them (especially when hemolysis occurs that might take place in blood from anemic patients) and induce the arachidonic acid cascade, which in turn can influence the plasma lipidome 5. However, as these effects potentially appear during sample collection, their influence on the measured lipidome is beyond the reach of the methodology described here. Effect of the number of freeze and thaw cycles on the plasma lipidome. (A) Lipid class profile of the same sample frozen and thawed up to 10 times. (B) Double bond profile of the same sample frozen and thawed up to 10 times. Averaged values are shown with error bars depicting standard deviation of five experiments per freeze and thaw cycle. We have established a fully automated, high throughput shotgun-based lipidomics method for systematic screening of plasma samples. Currently, this is the most comprehensive MS-based lipidomics method from a single acquisition, and most importantly, it is quantitative and highly reproducible. More important than speed and reproducibility is accuracy. We sampled the literature in order to compare our data with the results obtained in studies providing quantitative data and similar lipidomic coverage to ours 5,20,21,42–44. We observed that our data fit the range of values observed in literature, with the exception of TAG and SM (Figs. 4 and S5). These discrepancies are most likely related to the increase in chylomicron content in the plasma that we used from unfasted donors. Chylomicrons are known to transport lipids from the intestine and an increase in TAG content should be expected in unfasted subjects, since they are the most TAG enriched lipoproteins particles present in plasma 45. The main lipid classes’ amount distribution described in literature 5,20,21,42–44 compared with this study. Medians of average for control samples reported in these papers are presented. Error bars denote minimal and maximal values (range). Note that not all classes were analyzed in every study. To the best of our knowledge, for the first time for an omics-type analytical technology, we showed high method precision not only on different days but also in different laboratories. The implications of these are twofold. First, it means that this technology can be easily implemented in different sites without compromising data reproducibility and second, it facilitates direct integration of data between different laboratories, allowing multi-site studies for higher throughput, if required. We believe this method paves the way for making lipidomics an accessible and indispensable tool not only in biological basic research, but also for clinical diagnostics and nutrition by offering unprecedented throughput and accuracy. The authors would like to thank to Andrej Shevchenko for feedback on the paper. This study was supported by Nestlé Institute of Health Sciences and the Klaus Tschira Foundation. Conflict of interest statement: MAS, RH, AV, CK, KS, JLS have paid employment at Lipotype GmbH and NC, DM-R, MM have paid employment at Nestlé Institute of Health Sciences S.A. This does not alter the authors’ adherence to all policies on sharing data and materials.
|
PMC10513930
|
Dynamic lipidome alterations associated with human health, disease and ageing
|
Lipids can be of endogenous or exogenous origin and affect diverse biological functions, including cell membrane maintenance, energy management and cellular signalling. Here, we report >800 lipid species, many of which are associated with health-to-disease transitions in diabetes, ageing and inflammation, as well as cytokine–lipidome networks. We performed comprehensive longitudinal lipidomic profiling and analysed >1,500 plasma samples from 112 participants followed for up to 9 years (average 3.2 years) to define the distinct physiological roles of complex lipid subclasses, including large and small triacylglycerols, ester- and ether-linked phosphatidylethanolamines, lysophosphatidylcholines, lysophosphatidylethanolamines, cholesterol esters and ceramides. Our findings reveal dynamic changes in the plasma lipidome during respiratory viral infection, insulin resistance and ageing, suggesting that lipids may have roles in immune homoeostasis and inflammation regulation. Individuals with insulin resistance exhibit disturbed immune homoeostasis, altered associations between lipids and clinical markers, and accelerated changes in specific lipid subclasses during ageing. Our dataset based on longitudinal deep lipidome profiling offers insights into personalized ageing, metabolic health and inflammation, potentially guiding future monitoring and intervention strategies.Lipids are an important and highly diverse class of molecules that have critical roles in cell structure, cell signalling and bioenergetics. Despite their critical roles in many biological processes, there is much to be learned about the diversity of lipids in humans, how their composition differs across people, and how they change over time at an individual level and during disease. Such information is expected to provide insights into biological processes such as ageing as well as the possible roles of lipids in health and disease. High-throughput omics technologies provide new avenues to understand the molecular landscape of human physiology and its dynamic changes during health and disease. To date, many studies have used next-generation sequencing owing to its accessibility and cost-effectiveness. Recently, mass spectrometry (MS) strategies have provided quantitative insights into the proteome at scale and depth. Metabolites, which can also be investigated using MS, have been studied to a lesser extent given their complex chemical diversity. Lipids comprise a major, heterogeneous family of biomolecules within the metabolome and remain challenging to characterize owing to their wide range of physicochemical properties and the relatively small number of lipidomics studies. Complex lipids can be divided into several classes and subclasses that are distinguished by lipid head groups and linkages to different aliphatic chains. Lipids such as triacylglycerols (TAGs), diacylglycerols (DAGs), phosphatidylcholines (PCs), phosphatidylethanolamines (PEs), ceramides (CERs), sphingomyelins (SMs) and cholesterol esters (CEs) each consist of a specific backbone architecture conjugated to various fatty acids (FAs). The attached FAs can vary in the number of unsaturated bonds and their positions within acyl chains; together with the backbone, FAs confer distinct physicochemical properties and physiological roles. Lipids carry out and regulate many key functions, including redox homoeostasis, energy storage, intracellular and extracellular signalling, induction and resolution of acute and chronic inflammation, and maintenance of electrochemical gradients across subcellular compartments. Abnormal lipid profiles (dyslipidaemia) have been associated with a range of diseases, including metabolic syndrome, type 2 diabetes (T2D), cancer, nephropathy and cardiovascular and neurodegenerative diseases, and may result from a combination of factors such as genetic heterogeneity, lifestyle and, as recently shown, inflammation related to coronavirus disease 2019 infection. One of the key roles of lipids in maintaining metabolic homoeostasis is to mediate the induction and attenuation of inflammatory processes (for example, leukotriene, prostanoid and endocannabinoid signalling). Because of the various roles lipids have in maintaining homoeostasis in humans, different lipid species or classes may influence perturbations that induce acute inflammation (for example, respiratory viral infections (RVIs)), as well as the resolution of inflammation, metabolic diseases (for example, T2D) and physiological processes (for example, ageing) that have been associated with changes in the regulation of chronic inflammation. In light of the diverse roles of lipids, it is important to understand their quantitative differences among individuals and their dynamics across phenotypes to characterize their potential roles in health and disease. Here, we characterize the lipidome dynamics in >100 human participants followed for up to 9 years, covering periods of health and disease, using an MS-based approach that allows a broad array of lipid types to be measured rapidly, quantitatively and rigorously. We identified distinct longitudinal lipid signatures that link lipid profiles to the microbiome, ageing and different clinical pathophysiologies, including insulin resistance (IR) and chronic and acute inflammation. Our results provide valuable insights into the associations of key lipids and lipid subclasses with distinct metabolic health states in humans, and serve as a unique resource to the scientific community. From a cohort of >100 participants with IR or insulin sensitivity (IS), we previously collected longitudinal molecular data comprising genome, transcriptome, proteome, metabolome and 16S microbiome data across different timepoints (~1,000 in total). Within this cohort, we explored various molecular signatures in health and disease and identified hundreds of molecular pathways associated with metabolic, cardiovascular and oncologic pathophysiologies. Here, we investigate the dynamics of a largely unexplored molecular layer—the ‘plasma lipidome’—and extend the longitudinal duration by 2 years to obtain a total of 1,539 samples. To investigate lipidome alterations associated with health, disease and lifestyle changes, plasma samples from 112 participants were profiled at a median of ten timepoints across 2–9 years (average 3.2 years; one participant was sampled 163 times across 9 years; Fig. 1a, Supplementary Data 1 and 2 and Supplementary Figs. 1 and 2). Samples were collected every 3 months when the participants were healthy and with an increased frequency of three to seven collections over 3 weeks during periods of illness (for example, RVI) or notable stress, as previously reported. In addition to lipid profiling, we collected 50 clinical laboratory measurements at each sampling timepoint along with medical records (Supplementary Data 2). Finally, because samples were collected during periods of stress and illness, we also profiled 62 cytokines, chemokines and growth factors in plasma at the same timepoints.Fig. 1Longitudinal lipidomics profiling.a, Profiling, using >1,500 biosamples, across 112 participants followed for up to 9 years. Dynamic changes in the lipidome were characterized in the context of health status and medication history and in comparison with the participants’ cytokine, chemokine and metabolic profiles, as well as microbiome. b, Lipid subclasses investigated in this study. Lipid species, defined by a specific combination of backbone architecture and FAs, can be grouped based on their physicochemical properties. c, We analysed 846 lipids (y axis) across multiple subclasses. d, Across all 112 participants (median estimated concentration across all participant-specific samples), lipid species (846) spanned a dynamic range of more than four orders of magnitude, with distinct estimated concentration ranges for each lipid species and subclass. e, Comparison of the CVs of QC (n = 104), intraparticipant and interparticipant samples. All boxplots report the 25% (lower hinge), 50% (centre line) and 75% (upper hinge) quantiles. Whiskers indicate observations equal to or outside the hinge ± 1.5× the interquartile range (IQR). Outliers (beyond 1.5× the IQR) are not plotted.Source data a, Profiling, using >1,500 biosamples, across 112 participants followed for up to 9 years. Dynamic changes in the lipidome were characterized in the context of health status and medication history and in comparison with the participants’ cytokine, chemokine and metabolic profiles, as well as microbiome. b, Lipid subclasses investigated in this study. Lipid species, defined by a specific combination of backbone architecture and FAs, can be grouped based on their physicochemical properties. c, We analysed 846 lipids (y axis) across multiple subclasses. d, Across all 112 participants (median estimated concentration across all participant-specific samples), lipid species (846) spanned a dynamic range of more than four orders of magnitude, with distinct estimated concentration ranges for each lipid species and subclass. e, Comparison of the CVs of QC (n = 104), intraparticipant and interparticipant samples. All boxplots report the 25% (lower hinge), 50% (centre line) and 75% (upper hinge) quantiles. Whiskers indicate observations equal to or outside the hinge ± 1.5× the interquartile range (IQR). Outliers (beyond 1.5× the IQR) are not plotted. Source data The human lipidome was characterized using a high-throughput quantitative lipidomics pipeline (Lipidyzer) consisting of a triple-quadrupole mass spectrometer (Sciex QTRAP 5500) in conjunction with a differential mobility separation (DMS) device. This setup allows the identification and robust quantification (estimated concentrations) of >1,000 lipid species across 16 subclasses (free FA (FFA), TAG, DAG, CE, PC, lysophosphatidylcholine (LPC), PE, alkyl ether substituent containing PE (PE-O), alkenyl ether (Plasmalogen) substituent containing PE (PE-P), lysophosphatidylethanolamine (LPE), SM, PI, CER, hexosylceramide (HCER), lactosylceramide (LCER) and dihydroceramide (DCER); Fig. 1b). In addition, we observed the differential behaviour of smaller and larger TAGs, which comprise ≤48 and ≥49 carbons across all FAs, respectively, and evaluated these separately in most analyses. For accurate quantification and to control for variance introduced during lipid extraction, we included a mix of 54 deuterated spike-in standards for nine lipid subclasses at known concentrations. Lipid species that were not present as labelled spike-in standards were normalized against the spike-in standards based on structural similarity and signal correlation (described in Methods). We randomized the samples separately for lipid extraction and MS data acquisition. After filtering (described in Methods), we quantified, on average, 778 lipids in each sample and 846 lipid species across >1,600 samples (including quality control (QC) samples). We found the highest number (373) of lipid species in the large TAG subclass and the smallest number (4) in the DCER subclass (Fig. 1c). Lipids comprise chemically heterogeneous molecules that exert a broad spectrum of biological functions ranging from bioenergetics to cellular signalling. This is partially visible in lipid subclass-specific abundance distributions. Figure 1d shows the abundance distributions across more than four orders of magnitude and for each lipid subclass, and depicts two distinct properties: (1) the median abundance of that subclass and (2) the abundance range across all interrogated plasma samples (including healthy and disease timepoints). SMs and FFAs were observed, on average, as the most abundant subclasses, but they spanned a relatively small dynamic range. Other lipid subclasses, including LPCs, CEs and TAGs, had a lower median abundance but a much wider dynamic range. Our study demonstrated high technical reproducibility. As anticipated, the 104 QC samples clustered distinctly (Extended Data Fig. 1); the median coefficient of variation (CV) for the QC samples was low, with values between 6.5% (small TAGs) and 20.7% (DAGs). In contrast, CVs calculated across participants and sampling timepoints ranged from 19.9% (SMs) to 91.4% (small TAGs), indicating sufficient assay reproducibility to discern biological differences. To ensure the highest robustness in our analysis, we focused on 736 lipid species for which (1) QC CVs were <20% and (2) CVs in biosamples were larger than CVs in QC samples. Except for FFAs, intraparticipant variance was consistently lower than interparticipant variance, suggesting that individual lipid signatures are distinct and stable over time (Fig. 1e). Interestingly, both small and large TAGs and ester- and ether-linked PEs (PE versus PE-O and PE-P) exhibited significant differences within their respective subclasses in terms of variance (Fig. 1e) and abundance distribution (Fig. 1d). This implies the existence of unique physiological and participant-specific differences, which may provide new insights into biological processes. We first sought to investigate lipid abundance differences across individuals by characterizing the lipidome in ‘healthy’ baseline samples, defined as samples from participants in the absence of any self-reported acute disease. This does not preclude latent, asymptomatic chronic conditions such as prediabetes or potential undiagnosed conditions. Overall, we analysed 802 healthy baseline samples derived from 96 participants from whom we collected samples at two or more timepoints. The number of baseline samples per participant is shown in Supplementary Fig. 3; most participants had approximately ten healthy visits, except one outlier with 52 healthy baseline samples. In comparison with the transcriptome, proteome and general metabolome, lipid signatures can be highly personalized when assessed longitudinally. To investigate the participant specificity of lipid profiles for healthy sampling timepoints at timescales of months to years, we examined which lipid subclasses show the largest interindividual differences and quantified how much of the variance observed for each lipid species can be attributed to interparticipant differences (Fig. 2a). Many lipids, in particular among TAGs, SMs, HCERs and CEs, showed a high degree of participant-specific variance, in some cases >50%. In contrast, FFAs were found to have relatively low participant-specific variation. To further illustrate participant specificity, we performed t-distributed stochastic neighbour embedding (t-SNE) on data from participants with >12 healthy visits, based on 100 lipids that we determined to be the most personalized (Fig. 2b). Most, but not all, samples clustered by individual participants (Fig. 2b,c), showing that some lipids can comprise personalized signatures even across years.Fig. 2Interindividual differences in healthy baseline.a, Top, bar plot showing the number of lipid species per class ordered by the variance explained by the participant factor; bottom, boxplot showing the variance explained by participants in each lipid class (left y axis) and line graph showing the mean log10(estimated concentration) (red line, right y axis) of each lipid class. Variance decomposition analysis was conducted using n = 802 healthy samples. b, t-SNE clustering of 11 participants who contributed ≥12 healthy samples (n = 191), based on the 100 most personalized lipids. c, Intraparticipant distance, which refers to the Euclidean distance between each pair of samples belonging to the same participant, and interparticipant distance, which refers to the distance between the centroids from each pair of participants, for the t-SNE results. Boxplots report the 25% (lower hinge), 50% (centre line) and 75% (upper hinge) quantiles. Whiskers indicate observations equal to or outside the hinge ± 1.5× the IQR. Outliers (beyond 1.5× the IQR) are not plotted. The intraparticipant and interparticipant distances were compared using a two-sided t test. d, WGCNA modules and their correlation (BH-adjusted FDR cut-off of 5%) with clinical measures. Dot size depicts the BH-adjusted −log10(FDR). The colour scale indicates the degree and direction of the correlation. TGL, total triglyceride; CHOL, total cholesterol; NHDL, non-HDL; CHOLHDL, cholesterol to HDL ratio; LDLHDL, LDL to HDL ratio; GLU, glucose; INSF, fasting insulin; HSCRP, high-sensitivity CRP; WBC, white blood cell count; NEUT, neutrophil percent; NEUTAB, neutrophil absolute count; LYM, lymphocyte percent; LYMAB, lymphocyte absolute count; MONO, monocyte percent; MONOAB, monocyte absolute count; EOS, eosinophil percent; EOSAB, eosinophil absolute count; BASO, basophil percent; BASOAB, basophil absolute count; IGM, immunoglobulin M; RBC, red blood cell count; HGB, haemoglobin; HCT, haematocrit; MCV, mean corpuscular volume; MCH, mean corpuscular haemoglobin; MCHC, mean corpuscular haemoglobin concentration; RDW, red cell distribution width; PLT, platelet; AG, albumin to globulin ratio; CR, creatinine; BUN, blood urea nitrogen; EGFR, estimated glomerular filtration rate; UALB, urine albumin; ALCRU, aluminium to creatinine ratio, urine; UALBCR, urine albumin to creatinine ratio; TP, total protein; ALB, albumin; TBIL, total bilirubin; ALKP, alkaline phosphatase; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GLOB, globulin. e, Module composition for the WGCNA analysis shown in c, coloured by lipid subclass. f, Enrichment analysis results based on Fisher’s exact test, depicting the BH-adjusted −log10(FDR) for enriched subclasses for each WGCNA module.Source data a, Top, bar plot showing the number of lipid species per class ordered by the variance explained by the participant factor; bottom, boxplot showing the variance explained by participants in each lipid class (left y axis) and line graph showing the mean log10(estimated concentration) (red line, right y axis) of each lipid class. Variance decomposition analysis was conducted using n = 802 healthy samples. b, t-SNE clustering of 11 participants who contributed ≥12 healthy samples (n = 191), based on the 100 most personalized lipids. c, Intraparticipant distance, which refers to the Euclidean distance between each pair of samples belonging to the same participant, and interparticipant distance, which refers to the distance between the centroids from each pair of participants, for the t-SNE results. Boxplots report the 25% (lower hinge), 50% (centre line) and 75% (upper hinge) quantiles. Whiskers indicate observations equal to or outside the hinge ± 1.5× the IQR. Outliers (beyond 1.5× the IQR) are not plotted. The intraparticipant and interparticipant distances were compared using a two-sided t test. d, WGCNA modules and their correlation (BH-adjusted FDR cut-off of 5%) with clinical measures. Dot size depicts the BH-adjusted −log10(FDR). The colour scale indicates the degree and direction of the correlation. TGL, total triglyceride; CHOL, total cholesterol; NHDL, non-HDL; CHOLHDL, cholesterol to HDL ratio; LDLHDL, LDL to HDL ratio; GLU, glucose; INSF, fasting insulin; HSCRP, high-sensitivity CRP; WBC, white blood cell count; NEUT, neutrophil percent; NEUTAB, neutrophil absolute count; LYM, lymphocyte percent; LYMAB, lymphocyte absolute count; MONO, monocyte percent; MONOAB, monocyte absolute count; EOS, eosinophil percent; EOSAB, eosinophil absolute count; BASO, basophil percent; BASOAB, basophil absolute count; IGM, immunoglobulin M; RBC, red blood cell count; HGB, haemoglobin; HCT, haematocrit; MCV, mean corpuscular volume; MCH, mean corpuscular haemoglobin; MCHC, mean corpuscular haemoglobin concentration; RDW, red cell distribution width; PLT, platelet; AG, albumin to globulin ratio; CR, creatinine; BUN, blood urea nitrogen; EGFR, estimated glomerular filtration rate; UALB, urine albumin; ALCRU, aluminium to creatinine ratio, urine; UALBCR, urine albumin to creatinine ratio; TP, total protein; ALB, albumin; TBIL, total bilirubin; ALKP, alkaline phosphatase; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GLOB, globulin. e, Module composition for the WGCNA analysis shown in c, coloured by lipid subclass. f, Enrichment analysis results based on Fisher’s exact test, depicting the BH-adjusted −log10(FDR) for enriched subclasses for each WGCNA module. Source data Given the high variation in specific lipid classes among individuals, we next investigated the degree to which global lipidome profiles from healthy baseline samples are associated with clinical measures. We first grouped lipids into seven modules by applying weighted gene correlation network analysis (WGCNA) based on the similarity of lipid profiles and then associated these seven modules with 50 clinical measures while controlling for the covariates sex, age, ethnicity and body mass index (BMI; Fig. 2d–f and Supplementary Fig. 4). Controlling for these covariates allows investigation of the direct associations between the lipid modules and clinical measures by ruling out potentially confounding effects from sex, age and BMI. Modules M1 and M5, which were enriched for CER and PE, as well as small TAG (mainly M1) and large TAG (mainly M5), showed the strongest positive association with T2D measures, including glycated haemoglobin (A1C), fasting blood glucose and fasting insulin. Moreover, they showed a positive association with inflammatory markers, including high-sensitivity C-reactive protein (CRP) level and white blood cell count, and a negative association with high-density lipoprotein (HDL; ‘good cholesterol’) levels. Hence, lipids in M1 and M2 have negative health associations based on conventional clinical measures. In contrast, M7, which contained some FFAs and LPCs, correlated with lower CRP and A1C levels. M3, which was enriched for PE-P and PE-O, showed an association with higher levels of HDL and lower levels of fasting insulin, and, compared with the dominant T2D patterns in M1 and M5, demonstrated a signature that is generally considered healthier. In addition, we investigated lipid–microbiome associations and observed mostly negatively correlated lipids, including several TAG species for the bacterial family Oscillospiraceae and (L)PE, PC and CE for Clostridiaceae (Supplementary Fig. 5 and Supplementary Note 1). These microorganisms are known to be abundant in the gut of participants with IS in this cohort, suggesting a potentially beneficial role of Clostridia in host lipid metabolism. Finally, an outlier analysis identified participants with abnormally high or low lipid signatures, some of which we could correlate with underlying medical conditions such as hepatic steatosis (Supplementary Fig. 6 and Supplementary Note 2). Overall, this global analysis suggests that many lipid subclasses are associated with and potentially have a role (for example, proinflammatory, anti-inflammatory or metabolic role) in clinical conditions, or may serve as biomarkers to stratify health states. As many clinical measures were associated with specific lipid subclasses, we next determined how the lipidome is influenced by the chronic metabolic disorder IR. IR commonly occurs in T2D and is a condition in which cells, mainly muscle cells and adipocytes, are unresponsive to insulin, leading to high glucose levels in the blood. IR is often associated with chronic inflammation as well as metabolic syndrome, including dyslipidaemia, and can lead to non-alcoholic fatty liver disease. Elucidating how the lipid network is perturbed in individuals with IR is important to better understand the molecular mechanisms and prognosis of metabolic disorders. IR can be diagnosed by measuring the steady-state plasma glucose (SSPG) level after endogenous insulin secretion is suppressed and insulin and glucose are infused at fixed concentrations. IR or IS (IR/IS) status was measured using SSPG assays in 69 participants, of whom 36 and 33 were classified as having IR (SSPG >150 mg dl) and IS (SSPG ≤150 mg dl), respectively. At the global level, we observed some capacity of lipid signatures to distinguish IR and IS (Fig. 3a). Using regression analyses that controlled for age, sex, ethnicity and baseline BMI, we resolved comprehensive differences between IR and IS across most lipid subclasses (Fig. 3b–d), such that more than half of the lipids (424) were significantly associated with SSPG levels. Lipids and lipid subclasses that had a significant positive correlation with SSPG included TAGs and DAGs, which is consistent with our observations (Fig. 2d) and previous reports of higher levels of these lipids in individuals with dyslipidaemia and metabolic syndrome. We also observed subsets of CERs to have increased abundance, contributing to the development of obesity-induced IR in mice and humans (Fig. 3b,c), and making possible the lipid-based differentiation of IR and IS (Fig. 3a and Supplementary Fig. 7).Fig. 3IR- and IS-associated lipid signatures.a, Principal component analysis comparing IR and IS. The density plot on the right indicates the distribution of eigenvectors for each data point along the second principal component (PC2). Eigenvector comparison between IR and IS was conducted using a two-sided t test. b, Regression analysis (n = 69): 424 of 736 lipids had significant correlations with SSPG (BH FDR < 5%; corrected for age, sex, ethnicity and baseline BMI). c, Boxplot depicting regression coefficients for the respective lipid classes by using 69 samples for which the SSPG level was measured at the visit. Larger coefficients indicate stronger associations with higher SSPG levels. Colour indicates distributions for which the 25th or 75th percentile is positive or negative. Boxplots report the 25% (lower hinge), 50% (centre line) and 75% (upper hinge) quantiles. Whiskers indicate observations equal to or outside the hinge ± 1.5× the IQR. Outliers (beyond 1.5× the IQR) are not plotted. d, Proportional differences between IR and IS detected in participants. Centre numbers indicate the total number of lipids in each class. Enzyme names are shown in red. CDP-Cho, cytidine diphosphocholine; CDP-Eth, cytidine diphosphoethanolamine; CPT, choline phosphotransferase; EPT, ethanolamine phosphotransferase; GPAT, glycerol-3-phosphate acyltransferase; LPAAT, lysophosphatidic acid acyltransferases; PAP, phosphatidate phosphatase; DGAT, DAG acyltransferase; G-3-P, glyceraldehyde-3-phosphate; CDS, CDP-diacylglycerol synthase; PSD, PS decarboxylase; PSS, PS synthase; PGS, PG synthase; PIS, PI synthase; SPT, serine palmitoyl transferase; CerS, ceramide synthase; SMase, sphingomyelinase; DES, dihydroceramide desaturase; Acetyl-CoA, acetyl coenzyme A; TCA, tricarboxylic acid. e, Enrichment analysis (Fisher’s exact test) performed on the coefficients from SSPG regression. Enriched annotations were calculated for positive coefficients with BH FDR < 10% (positive log2(odds)) and negative coefficients with BH FDR < 10% (negative log2(odds)). For enriched annotations, a BH FDR cut-off of 5% was applied. f, Correlations between clinical measures and lipid profiles for IR and IS. Correlations are shown when the correlations in IR and IS were significantly different and the absolute Δ correlations in IR and IS were >0.2. In addition, the overall correlations between lipids and clinical measures across IR and IS are depicted when the aforementioned two criteria were met.Source data a, Principal component analysis comparing IR and IS. The density plot on the right indicates the distribution of eigenvectors for each data point along the second principal component (PC2). Eigenvector comparison between IR and IS was conducted using a two-sided t test. b, Regression analysis (n = 69): 424 of 736 lipids had significant correlations with SSPG (BH FDR < 5%; corrected for age, sex, ethnicity and baseline BMI). c, Boxplot depicting regression coefficients for the respective lipid classes by using 69 samples for which the SSPG level was measured at the visit. Larger coefficients indicate stronger associations with higher SSPG levels. Colour indicates distributions for which the 25th or 75th percentile is positive or negative. Boxplots report the 25% (lower hinge), 50% (centre line) and 75% (upper hinge) quantiles. Whiskers indicate observations equal to or outside the hinge ± 1.5× the IQR. Outliers (beyond 1.5× the IQR) are not plotted. d, Proportional differences between IR and IS detected in participants. Centre numbers indicate the total number of lipids in each class. Enzyme names are shown in red. CDP-Cho, cytidine diphosphocholine; CDP-Eth, cytidine diphosphoethanolamine; CPT, choline phosphotransferase; EPT, ethanolamine phosphotransferase; GPAT, glycerol-3-phosphate acyltransferase; LPAAT, lysophosphatidic acid acyltransferases; PAP, phosphatidate phosphatase; DGAT, DAG acyltransferase; G-3-P, glyceraldehyde-3-phosphate; CDS, CDP-diacylglycerol synthase; PSD, PS decarboxylase; PSS, PS synthase; PGS, PG synthase; PIS, PI synthase; SPT, serine palmitoyl transferase; CerS, ceramide synthase; SMase, sphingomyelinase; DES, dihydroceramide desaturase; Acetyl-CoA, acetyl coenzyme A; TCA, tricarboxylic acid. e, Enrichment analysis (Fisher’s exact test) performed on the coefficients from SSPG regression. Enriched annotations were calculated for positive coefficients with BH FDR < 10% (positive log2(odds)) and negative coefficients with BH FDR < 10% (negative log2(odds)). For enriched annotations, a BH FDR cut-off of 5% was applied. f, Correlations between clinical measures and lipid profiles for IR and IS. Correlations are shown when the correlations in IR and IS were significantly different and the absolute Δ correlations in IR and IS were >0.2. In addition, the overall correlations between lipids and clinical measures across IR and IS are depicted when the aforementioned two criteria were met. Source data To investigate over- and under-representations of specific subgroups of lipids, we performed an enrichment analysis on positive and negative model coefficients. As TAGs comprise the largest subclass of lipids in our data and could dominate the results, we performed enrichment analyses separately for each lipid subclass and across all lipids (Fig. 3e). Enrichments were evaluated at the subclass level (Fig. 1b) and for FA composition (global saturation level and specific FAs). Importantly, to our knowledge, new associations were found, including an association of ether-linked PE (PE-P)—in contrast to PE in general—with lower SSPG levels. Ether-linked PEs are involved in cell signalling and can act as antioxidants. Together with increased levels of TAGs with higher SSPG levels, reduced PE-P levels suggest IR-associated inflammation and may indicate a PE-mediated link between oxidative stress, inflammation and IR. In Fig. 2d, we demonstrated lipid modules that correlate with a variety of clinical measures. As it is well documented that IR affects both lipid regulation (dyslipidaemia) and clinical phenotypes, we investigated whether associations between lipids and clinical measures are affected by the IR/IS status, which would have important health implications for these participants (Fig. 3f and Supplementary Fig. 8). Intriguingly, we found many significant differences in both the effect sizes and the direction of the correlation of lipid signatures and clinical measures between participants with IR and those with IS. For instance, in IR unlike in IS, the low-density lipoprotein (LDL) to HDL ratio was positively associated with the ether-linked PE-P and PE-O, and negatively associated with LPE (Fig. 3f). Moreover, in participants with IR and IS, we observed opposite correlations of immune and blood cell measurements with lipid subclasses, including A1C–SM, SSPG–CER and SSPG–PI, as well as immunoglobulin M–PE-P/PE-O, monocyte–PE-P/PE-O, eosinophil–TAG and white blood cell–PI (Fig. 3f). Overall, these data indicate that, depending on the IR/IS status, lipid–clinical measure associations can vary significantly and the key lipids involved in energy regulation, cell signalling and immune homoeostasis exhibit broad dysregulation in IR. In addition to their role in chronic inflammatory and metabolic conditions such as IR, complex lipids are key mediators of acute inflammatory responses, for example, by releasing arachidonic acid (FA(20:4)). Hence, complex lipids may be modified, released and activated during RVIs and possibly vaccinations while also having important roles in these processes in an IR-dependent manner. Participants in this cohort were densely sampled during periods of RVI (72 distinct RVI episodes in 36 participants for a total of 390 samples) and vaccination (44 episodes in 24 participants for a total of 275 visits; Supplementary Fig. 9). For both RVI and vaccination, we classified longitudinally collected samples as early-phase (days 1–6), later-phase (days 7–14) and recovery-phase (weeks 3–5) samples (Fig. 4a). Using linear models, we identified 210 lipids that were significantly changed during RVI (false discovery rate (FDR) < 10%) across most subclasses (Fig. 4b), some of which have previously been implicated in acute inflammation. For instance, PEs have been reported to have a critical role in apoptotic cell clearance and the aetiology of various viruses. Another example is PIs, which bind to the respiratory syncytial virus with high affinity, preventing virus attachment to epithelial cells. LPCs, which we observed in increased abundance during inflammation, have been demonstrated to have therapeutic effects (after intraperitoneal administration in mice) in severe infections through immune cell recruitment and modulation.Fig. 4RVIs and vaccination.a, Longitudinal sampling at five timepoints during RVIs: before infection (healthy), early event, late event, recovery and after infection (healthy). b, Lipid class breakdown for all detected lipids. Dark green depicts 210 significantly changed lipids throughout RVI. Enriched subclass. Fisher’s exact test was used for the lipid class enrichment analysis of the significant lipids (BH FDR for each lipid subclass: CE, 3.35 × 10; CER, 0.95; DCER, 0.49; HCER, 0.87; LCER, 1; DAG, 1; FFA, 0.56; LPC, 6.32 × 10; LPE, 8.40 × 10; PC, 3.01 × 10; PE, 0.27; PE-O, 1.01 × 10; PE-P, 1.00 × 10; PI, 7.65 × 10; SM, 1; large TAG, 1; small TAG, 3.66 × 10). c, Lipid enrichment analysis for significantly changed lipids during RVI, across (left column) and within classes. d, Trajectory analysis of the 210 significantly changed lipids following RVI and their corresponding profiles in each cluster. e, Associations of lipid profiles in RVI and clinical measures. Depicted are correlations between the identified lipid clusters (d) and 50 clinical laboratory measures (BH FDR cut-off of 5%). Dot size depicts −log10(FDR); colour scale represents the correlation direction and degree. f, Differential profile of lipids that were significantly changed during RVI, comparing IR and IS. For each lipid feature, the shaded blocks demonstrate the time intervals during which the corresponding lipid was significantly different between IR and IS. The orange shaded blocks representing the lipid profiles at this time interval are dominant (with higher lipid levels) in IR, and the blue shaded blocks representing the lipid profiles at this time interval are dominant in IS.Source data a, Longitudinal sampling at five timepoints during RVIs: before infection (healthy), early event, late event, recovery and after infection (healthy). b, Lipid class breakdown for all detected lipids. Dark green depicts 210 significantly changed lipids throughout RVI. Enriched subclass. Fisher’s exact test was used for the lipid class enrichment analysis of the significant lipids (BH FDR for each lipid subclass: CE, 3.35 × 10; CER, 0.95; DCER, 0.49; HCER, 0.87; LCER, 1; DAG, 1; FFA, 0.56; LPC, 6.32 × 10; LPE, 8.40 × 10; PC, 3.01 × 10; PE, 0.27; PE-O, 1.01 × 10; PE-P, 1.00 × 10; PI, 7.65 × 10; SM, 1; large TAG, 1; small TAG, 3.66 × 10). c, Lipid enrichment analysis for significantly changed lipids during RVI, across (left column) and within classes. d, Trajectory analysis of the 210 significantly changed lipids following RVI and their corresponding profiles in each cluster. e, Associations of lipid profiles in RVI and clinical measures. Depicted are correlations between the identified lipid clusters (d) and 50 clinical laboratory measures (BH FDR cut-off of 5%). Dot size depicts −log10(FDR); colour scale represents the correlation direction and degree. f, Differential profile of lipids that were significantly changed during RVI, comparing IR and IS. For each lipid feature, the shaded blocks demonstrate the time intervals during which the corresponding lipid was significantly different between IR and IS. The orange shaded blocks representing the lipid profiles at this time interval are dominant (with higher lipid levels) in IR, and the blue shaded blocks representing the lipid profiles at this time interval are dominant in IS. Source data To further investigate the lipid-associated processes that are involved in acute infection, we examined enriched lipids during infection (Fig. 4c). We observed significant changes in specific lipid subclasses, including ether-linked PEs and TAGs containing saturated FAs (SFAs) such as dodecanoic acid (FA(12:0)), following RVI. Dodecanoic acid and palmitic acid (FA(16:0)) are proinflammatory compounds that upregulate cyclooxygenase 2 (ref. ) and have key roles in the activation of inflammatory responses. Overall, this suggests that different key lipid subclasses may be important for various aspects of viral biology as well as the immune response, and undergo significant changes during RVI. To explore the choreography of lipid dynamics over time, we examined their trajectory during the different phases of RVI. The 210 significantly changed lipids were mapped to four major clusters, using a hierarchical clustering approach based on the Euclidean distance between lipid species as the similarity measure (Fig. 4d), and these main clusters were linked with clinical measures (Fig. 4e). Except for the green cluster, which was significantly enriched for PC, all profiles showed decreased levels during infection. The blue cluster was significantly enriched for small TAG and showed sharply decreased lipid levels in early RVI, with a rapid recovery that correlated with clinical measures of total lipids, including cholesterol and LDL. This indicates a metabolic shift in early infection, potentially to support increased energy metabolism. The orange cluster, enriched for LPC, large TAG and ether-linked PE, showed a similar profile to the blue cluster but a delayed recovery to baseline levels. Lipids in this cluster were positively correlated with the clinical lipid panel and blood glucose levels but negatively correlated with CRP level and neutrophils. This suggests that early changes in energy metabolism (reduction in lipid and blood glucose levels) are coupled with increased inflammation (reduction in LPC and ether-linked PE levels, as well as an increase in the CRP level and neutrophils) followed by a slow attenuation of inflammation at later stages of RVI. The purple cluster, which was enriched for FFA, represents slowly decreasing lipid levels and reached the lowest point during the RVI recovery phase before reverting to the baseline levels. In particular, the late-stage correlation with immune-related parameters (CRP level, lymphocyte count) suggests that reduction in the levels of some lipids in this cluster may relate to a temporary strong immunosuppression to attenuate early- to mid-phase inflammation and promote a return to homoeostasis. Overall, our data suggest links between differential responses of lipids and specific biological roles, with rapid shifts in energy metabolism to support inflammation early in infection and possible attenuation in later stages. Reflecting important global shifts in cell signalling, metabolism and inflammation during RVI, these lipids may allow the assessment of disease severity and prognosis or offer an opportunity for therapeutic intervention. We next investigated whether individuals with IR and IS respond differently to infections and vaccination (Fig. 4f and Supplementary Fig. 10). Through a longitudinal differential analysis, we found distinct longitudinal profiles for IR and IS. We observed a higher abundance of several FFAs during the early stage of RVI and greater elevation of PC levels in the middle to late stage in participants with IR than in participants with IS. In contrast, TAGs and some PEs were differentially elevated in IS compared with IR throughout the middle to later stages of infection. The IR/IS-specific FFA and TAG responses may reflect the altered energy metabolism in IR, whereas differences in PCs and other lipid classes may indicate changes in immune-associated signalling pathways. Importantly, we found that the patterns after vaccination were distinct from those during infection (Supplementary Fig. 10). For example, fewer TAG species showed elevated levels in IS, whereas a distinct population of LCERs were upregulated in IR after vaccination. As individuals with T2D associated with IR often exert a more compromised immune response to RVI, such changes may be biologically significant. Ageing increases the risk of cardiovascular diseases and is accompanied by a variety of diseases including T2D and chronic inflammation. In our study, the participants spanned an age range of 20–79 years (healthy timepoints, median 57 years) and were longitudinally sampled on average over 3 years (Fig. 5a). Across the cohort, we observed an increase in BMI with higher age (Fig. 5b). We previously identified age-associated molecular signatures in a subset of this cohort, including inflammation (acute-phase proteins), blood glucose and lipid metabolism (A1C, apolipoprotein A-IV protein), but had not investigated age-associated lipidome changes. To identify lipids and pathways that change with ageing and may be associated with the development of age-related pathologies, such as chronic low-grade inflammation, we investigated longitudinal changes in the lipidome. In cross-sectional studies, lipid content can differ across participants with different ages, owing to biological ageing or the period during which the cohort aged, or other cohort effects. Periods and cohorts are social contexts affecting individuals and are inherently and mathematically confounded by the individuals’ age. These comprise environmental factors differently affecting young and old participants, due to them being born in different generations, and include generation-dependent exposures that may also affect lipidome compositions (for example, diet, lifestyle and/or diseases) rather than actual age. However, we note that the longitudinal nature of our data better enabled us to eliminate some biases and focus on the same individual across time. Furthermore, we previously did not observe major dietary changes in the cohort. To identify lipid changes that occur with ageing in our longitudinal cohort, we used a linear model that estimates relative lipid changes as a function of the change in age (Δage model) while also controlling for sample storage length and BMI. With this model, we determined the ‘ageing’ effect (β coefficient) for each lipid subclass (Fig. 5c) and across lipid species (Fig. 5d).Fig. 5Age-associated changes in the lipidome.a, Median ages, age range (horizontal lines) and number of visits (y axis) of 90 healthy participants. Violin plot shows the distribution of age within the cohort. Inner boxplot reports the 25% (left hinge), 50% (centre line) and 75% (right hinge) quantiles. Whiskers indicate observations equal to or outside the hinge ± 1.5× the IQR. Outliers (beyond 1.5× the IQR) are not plotted. b, Correlation of median BMI and median age across healthy participants. Vertical lines depict the BMI range for each participant across all collected healthy samples. Regression line (red) from a linear model is shown with the 95% confidence band (grey). c, Significantly (BH FDR < 5%) changed lipid subclasses (percentage change for the summed untransformed concentration of respective lipid species) with ageing across 5 years based on the Δage model controlling for BMI and sample storage length. d, Fisher’s exact test enrichment analysis comparing physicochemical properties associated with higher age (positive log2(odds), red, determined for all positive Δage model coefficients at the lipid species level with a BH FDR of <10%) and those associated with lower age (negative log2(odds), blue, determined for all negative Δage model coefficients at the lipid species level with a BH FDR of <10%). Enrichments were calculated independently within lipid subclasses, as well as across all lipid species (‘all’). log2(odds) values are depicted for significant associations with lower or higher age (BH FDR < 5%). Infinite log2(odds) values are imputed with 0.5× the mean value of positive/negative log2(odds) determined across all data. MUFA, monounsaturated FA. e, Δage coefficients (ageing–sex) of individual lipid subclasses for male and female participants, controlling for sample storage length and BMI. f, Δage coefficients (ageing–IR/IS) of individual lipid subclasses for IR and IS, controlling for storage length, BMI and sex. For e and f, data are presented as the mean of estimated coefficients ± s.d., determined using an ordinary least-squares regression test.Source data a, Median ages, age range (horizontal lines) and number of visits (y axis) of 90 healthy participants. Violin plot shows the distribution of age within the cohort. Inner boxplot reports the 25% (left hinge), 50% (centre line) and 75% (right hinge) quantiles. Whiskers indicate observations equal to or outside the hinge ± 1.5× the IQR. Outliers (beyond 1.5× the IQR) are not plotted. b, Correlation of median BMI and median age across healthy participants. Vertical lines depict the BMI range for each participant across all collected healthy samples. Regression line (red) from a linear model is shown with the 95% confidence band (grey). c, Significantly (BH FDR < 5%) changed lipid subclasses (percentage change for the summed untransformed concentration of respective lipid species) with ageing across 5 years based on the Δage model controlling for BMI and sample storage length. d, Fisher’s exact test enrichment analysis comparing physicochemical properties associated with higher age (positive log2(odds), red, determined for all positive Δage model coefficients at the lipid species level with a BH FDR of <10%) and those associated with lower age (negative log2(odds), blue, determined for all negative Δage model coefficients at the lipid species level with a BH FDR of <10%). Enrichments were calculated independently within lipid subclasses, as well as across all lipid species (‘all’). log2(odds) values are depicted for significant associations with lower or higher age (BH FDR < 5%). Infinite log2(odds) values are imputed with 0.5× the mean value of positive/negative log2(odds) determined across all data. MUFA, monounsaturated FA. e, Δage coefficients (ageing–sex) of individual lipid subclasses for male and female participants, controlling for sample storage length and BMI. f, Δage coefficients (ageing–IR/IS) of individual lipid subclasses for IR and IS, controlling for storage length, BMI and sex. For e and f, data are presented as the mean of estimated coefficients ± s.d., determined using an ordinary least-squares regression test. Source data We found that the levels of most lipid subclasses increased with ageing, most prominently CERs (LCER, HCER, DCER), SMs, LPCs and CEs, with some of the observed variance suggesting more complex lipid–ageing dependencies (Fig. 5c). A general increase in the levels of multiple lipid species and subclasses is consistent with previous observations. Intriguingly, the levels of TAGs generally increased over time (Supplementary Fig. 11), but this trend disappeared when controlling for BMI. We performed an enrichment analysis on the Δage model coefficients at the species level and observed a shift in the physicochemical properties of lipids associated with ageing, including increased levels of SFAs and monounsaturated FAs, whereas the levels of polyunsaturated FAs (PUFAs) were reduced (Fig. 5d). This pattern has been previously associated with dyslipidaemia and inflammation, underlining progressive deterioration of metabolic health during ageing. We also observed depleted levels of (beneficial) omega-3 FAs. In particular, the levels of docosahexaenoic acid (FA(22:6), TAG) and eicosapentaenoic acid (FA(20:5), PE) decreased with ageing. These omega-3 FAs have been indicated to have beneficial health effects by lowering plasma cholesterol levels and serving as precursors for mediators that resolve inflammation, such as resolvins, protectins and maresins. In addition, decreased levels of linoleic acid (FA(18:2)) have been reported in aged skin; our data show that this is also a significant ageing biomarker in blood plasma, suggesting a more systemic decrease. Through desaturation and elongation, linoleic acid is metabolized to arachidonic acid (FA(20:4)), which we found to increase in abundance with increasing age when we applied less stringent filtering (Supplementary Fig. 12), further substantiating a general shift towards inflammation with ageing. Furthermore, large and small TAGs showed distinct patterns, underlining the different functional roles along the TAG spectrum. Interestingly, the levels of LPCs, which have been implicated in cardiovascular diseases and neurodegeneration and some of which are anti-correlated with CRP (Fig. 2), increased with ageing, further underlining their pleiotropic role in human health. We also observed a strong sex dimorphism for multiple subclasses (Fig. 5e). Beyene et al. reported sex-associated differences in lyso- and ether-phospholipid metabolism, which we confirmed in our study. In addition, we observed sex-associated differences for small TAGs as a prominent signature in ageing, with higher levels in men and lower levels in women. We next investigated the extent to which IR alters molecular ageing signatures and observed that participants with IR had larger coefficients for multiple subclasses, including HCER, LCER, SM and CE, than participants with IS. Larger coefficients indicate that ageing-related changes may be accelerated in IR versus IS (controlling for storage length, sex and BMI; Fig. 5f). In contrast to previous reports that did not distinguish IR status, our study identified a negative association between DAGs and ageing in participants with IR. Intriguingly, higher DAG levels are commonly linked to dyslipidaemia and IR; however, similar to TAGs (see above), DAGs may have a stronger association with BMI, which was controlled for in the model. Moreover, PI and PE showed opposite ageing effects in participants with IR and IS, which suggests IR-specific changes in phospholipid metabolism with ageing. In sum, the composition of many lipid subclasses (that is, degree of unsaturation, omega-3 FAs, large TAGs, ether-linked PEs) changes significantly with ageing, a process that—for some lipid subclasses—differs between the sexes and is distinctly accelerated in the presence of IR. Given the importance of cytokines, chemokines and growth factors in diverse biological processes, we characterized their relationship to lipids across homoeostasis and various pathophysiological disease processes in our longitudinal cohort. We investigated the degree to which the abundance of a particular lipid predicts the level of cytokines, chemokines or growth factors, controlling for BMI, sex, ethnicity and multiple measurements per participant as random effects across all samples and timepoints for which both measures were available (1,180 samples). Overall, we found 1,245 significant (FDR < 5%) positive and negative associations between a majority of lipids (580) and 40 cytokines, chemokines and growth factors (Fig. 6a).Fig. 6Lipid–cytokine associations.a–e, Network of 1,245 significant (BH FDR < 5%) lipid–cytokine associations, indicating positive (red) and negative (blue) associations calculated across 1,180 samples, across all lipids (a) and for PCs (b), PEs (c), LPCs (d) and LPEs (e). Networks were pruned based on a BH FDR of 5% for coefficients determined in linear mixed-effects models. Colour indicates lipid class; edge width represents coefficients; and node size represents node connectivity (popularity). The network was assembled using the ‘graphopt’ layout algorithm. f, Fisher’s exact test enrichment analysis comparing the physicochemical properties of lipids (y axis), at the subclass, global FA and individual FA level, that are associated with a particular cytokine (x axis). The analysis was performed for TAGs only (i), for all non-TAG lipids (ii) and across all lipids (iii). Enrichments (log2(odds)) among lipids with positive β coefficients (BH FDR < 10%) are indicated in red; enrichments (log2(odds)) among lipids with negative β coefficients (BH FDR < 10%) are indicated in blue; black denotes cases for which a certain property was enriched in both populations (positive and negative associations). log2(odds) values are depicted when the respective annotation was significantly associated with a BH FDR of <5%. Infinite log2(odds) values are imputed with 0.5× the positive/negative log2(odds) values determined across all data. IL-1Ra, IL-1 receptor antagonist; ICAM1, intercellular adhesion molecule 1; SDF1⍺, stromal cell-derived factor 1⍺; RANTES, regulated on activation, normal T cell expressed and secreted; PDGF-BB, platelet-derived growth factor-BB; GRO⍺, growth-regulated ⍺ protein; FasL, Fas ligand; TRAIL, tumour necrosis factor-related apoptosis-inducing ligand.Source data a–e, Network of 1,245 significant (BH FDR < 5%) lipid–cytokine associations, indicating positive (red) and negative (blue) associations calculated across 1,180 samples, across all lipids (a) and for PCs (b), PEs (c), LPCs (d) and LPEs (e). Networks were pruned based on a BH FDR of 5% for coefficients determined in linear mixed-effects models. Colour indicates lipid class; edge width represents coefficients; and node size represents node connectivity (popularity). The network was assembled using the ‘graphopt’ layout algorithm. f, Fisher’s exact test enrichment analysis comparing the physicochemical properties of lipids (y axis), at the subclass, global FA and individual FA level, that are associated with a particular cytokine (x axis). The analysis was performed for TAGs only (i), for all non-TAG lipids (ii) and across all lipids (iii). Enrichments (log2(odds)) among lipids with positive β coefficients (BH FDR < 10%) are indicated in red; enrichments (log2(odds)) among lipids with negative β coefficients (BH FDR < 10%) are indicated in blue; black denotes cases for which a certain property was enriched in both populations (positive and negative associations). log2(odds) values are depicted when the respective annotation was significantly associated with a BH FDR of <5%. Infinite log2(odds) values are imputed with 0.5× the positive/negative log2(odds) values determined across all data. IL-1Ra, IL-1 receptor antagonist; ICAM1, intercellular adhesion molecule 1; SDF1⍺, stromal cell-derived factor 1⍺; RANTES, regulated on activation, normal T cell expressed and secreted; PDGF-BB, platelet-derived growth factor-BB; GRO⍺, growth-regulated ⍺ protein; FasL, Fas ligand; TRAIL, tumour necrosis factor-related apoptosis-inducing ligand. Source data The largest numbers of positive associations were between granulocyte–macrophage colony-stimulating factor (GM-CSF) and TAGs and between leptin and TAGs (Fig. 6a and Supplementary Fig. 13). The adipokine leptin regulates caloric intake and is commonly present in elevated levels in obesity, contributing to the associated inflammatory state. Its amount in the blood correlates with the amount of adipose tissue. Its receptor is expressed in the hypothalamus, hippocampus and many immune cells; thus, it also acts as a neuroregulator and an immunoregulator. The cytokine GM-CSF, originally defined as a haemopoietic growth factor, has other biological roles, including exerting proinflammatory effects. These signatures are consistent with the inflammatory effect of the high TAG levels that we observed and are also found as a consequence of a high-fat diet, obesity and hepatic adiposity. The pleiotropic cytokine interleukin-6 (IL-6), whose inflammatory and anti-inflammatory effects are context dependent, together with the anti-inflammatory cytokine IL-10 (ref. ), showed negative associations with some TAGs and clustered distinctly from the positive TAG–leptin and TAG–GM-CSF associations, suggesting functional differences among different TAG species in immunoregulatory networks (Fig. 6a). TAGs showed the overall highest number of associations with leptin and GM-CSF, whereas lipids from other subclasses, such as PE, PC and DAG, were also positively associated (Fig. 6b,c and Supplementary Fig. 13). In contrast, lyso species of PE and PC (Fig. 6d,e) showed fewer associations with and less central roles for GM-CSF and leptin. Overall, these results suggest regulatory commonalities across lipid classes (for example, positive associations of TAGs, DAGs, PCs and PEs with leptin) and differences within subclasses for proinflammatory and immunoregulatory pathways. To elucidate the extent to which specific subsets of lipid species are associated with cytokines and chemokines, we performed an enrichment analysis (Fig. 6f). Overall, we observed strong associations of FAs with distinct cytokines. For instance, positive leptin–TAG associations were significantly enriched for SFA, the polyunsaturated FA(18:3) and small TAGs. In contrast, large TAGs were negatively associated with IL-6 and IL-10. Moreover, we observed a hub of negative associations between TAGs containing FA(22:5) and multiple cytokines, including the anti-inflammatory IL-10 and the proinflammatory IL-23, as well as IL-6. Enrichment of TAG subclasses for positive and negative associations within both proinflammatory and immunoregulatory cytokines suggests that TAG subclasses (in terms of both the length and saturation of the acyl chain) have distinct roles in immunoregulation and signalling. In Fig. 2, we found that some LPCs were associated with anti-inflammatory, hence healthier, signatures. Here, LPCs were positively associated with several growth factors, such as epidermal growth factor (EGF), vascular endothelial growth factor (VEGF) and brain-derived neurotrophic factor (BDNF), and resistin. VEGF is involved in promoting angiogenesis, whereas BDNF and EGF promote cell proliferation, with BDNF having a cardinal role in neurogenesis and plasticity. In addition, LPCs were positively associated with the soluble CD40 ligand (sCD40L), which is secreted by activated T cells and platelets during inflammation, as well as with the inflammatory cytokine IL-1⍺ and the adipose tissue-specific secretory factor resistin, which induces other cytokines and has been suggested to contribute to a chronic proinflammatory cascade in T2D. Together, LPCs demonstrated contrasting associations, including some anti-inflammatory and tissue repair as well as proinflammatory signatures. These associations may highlight a difference between chronic inflammation (for example, mediated by factors such as resistin during T2D) and acute inflammation (for example, during an infection), which is strongly associated with high CRP levels. It may also reflect that both inflammatory and anti-inflammatory mediators are present in amounts that regulate a response so that it is effective but not excessive. Moreover, PCs containing linoleic acid (FA(18:2)) were negatively associated with the chemokines CXC motif ligand 9 (CXCL9; also known as MIG (monokine induced by interferon-γ (IFNγ))) and CXCL10 (also known as IP-10 (IFNγ-induced protein 10 kDa); Fig. 6b,f). CXCL9 and CXCL10 are induced by IFNγ to recruit cells to sites of inflammation; they bind to the same chemokine receptor, CXCR3. This association suggests that these lipids may affect immune cell migration during inflammation, in addition to their immune modulation role that we observed during RVI (Fig. 4). Overall, our multiomics data outline complex associations between cytokines and lipid subclasses as well as differential associations of lipids with specific FA compositions, suggesting distinct roles ranging from immune activation to immunosuppression. Until recently, most omics studies have focused on transcriptomics, proteomics and, more recently, metabolomics, which is more closely associated with many phenotypes. However, lipidomics remains largely underexplored despite lipids’ important roles in cell signalling, cell structure and energy management. The lipidome has been difficult to study owing to its complexity and the fact that it is derived from both endogenous and exogenous factors, such as the microbiome and lifestyle (diet and physical activity). Investigating the dynamic range of lipid changes, including which lipids change with acute or chronic conditions over what period, may reveal markers of early disease onset and progression as well as mechanistic insights that can be used to develop better and personalized treatments. By following participants for up to 9 years, we identified highly participant-specific lipids and lipid subclasses (Figs. 1e and 2a–c), functional modules of lipids that map to clinical measures at baseline and throughout diseases (Figs. 2d,f and 3f), and lipid outlier signatures that may be predictive of diseases such as hepatic steatosis (Supplementary Fig. 6). Across perturbations such as IR (Fig. 3), viral infections (Fig. 4), ageing (Fig. 5) and cytokine–lipid associations (Fig. 6), we observed distinct behaviours among many lipid subclasses, such as ether- and ester-linked PEs, small and large TAGs, and lipids with specific FA configurations (for example, omega-3/omega-6 FAs, PUFAs and SFAs). Overall, our results point to the distinct biological roles of lipid subclasses and demonstrate that conventional clinical lipid profiles (that is, overall TAG levels) do not resolve many changes relevant to metabolic health. Throughout our analyses, we consistently observed distinct behaviours between ester-linked and ether-linked PEs (PE versus PE-O/PE-P), as well as between two functionally distinct subgroups of TAGs (small TAGs (≤48 carbons across all FAs) and large TAGs (≥49 carbons across all FAs)). Ether-linked PEs have been implicated in cell signalling and as antioxidants, and we found them to be significantly associated with healthy phenotypes including low SSPG levels and high HDL levels. In addition, ether-linked PEs are depleted early during infection, putatively increasing the inflammatory state, and/or depleted by scavenging radical oxygen species resulting from inflammation. In addition, PE, PE-P and PE-O show sex- and IR/IS-specific signatures during ageing (see below). Together, these observations suggest that ether-linked PEs are associated with health and chronically low levels of these PEs may have detrimental effects in humans. In this context, it will also be interesting to investigate other ether-linked lipid subclasses known to be detectable in blood plasma, such as PCs with alkyl ether substituent (PC-O) and PCs with alkenyl ether (Plasmalogen) substituent (PC-P), in future studies. Recently, we reported the differential regulation of small and large TAGs within a 60-min recovery phase after exercise. Here, we confirm and expand our previous observations suggesting new clinically relevant physiological roles of these TAGs within a much larger longitudinal cohort. Our data demonstrate that (1) many biological variations are captured in TAGs’ abundance profiles (Figs. 1e and 2a), which (2) are distinct for TAG subgroups including large and small TAGs, and those containing specific FAs (Figs. 5d and 6f). For instance, small TAGs show distinct associations with certain cytokines and chemokines and are rapidly depleted during early RVI, followed by a rapid recovery to baseline levels. Depletion of small TAGs during infection suggests an important role in energy metabolism and signalling to support early inflammation. In addition, during ageing, large and small TAGs differ markedly, suggesting distinct roles in ageing-related energy metabolism and lipid-mediated signalling. TAGs also showed the highest technical reproducibility in this study, making them an ideal target for new biomarkers at the subclass and species levels. We therefore propose that small and large TAGs as well as ether-linked PEs could be further explored as health biomarkers. Moreover, dietary supplements affecting plasma levels may provide a therapeutic avenue to reduce chronic inflammation and the detrimental effects of ageing and other conditions. In addition to identifying lipids that decrease in abundance with ageing, such as PUFAs, omega-3 FAs, FA(18:3) and FA(18:2), we identified many lipid subclasses and properties that are enhanced with ageing, including the proinflammatory CEs, CERs, SMs, arachidonic acid and SFAs, as well as LPCs whose role could be more ambiguous. Intriguingly, some of these effects were stronger (for example, HCERs, CEs, SMs) or directionally different (for example, PEs, DAGs) in participants with IR, which can be interpreted as accelerated or differential ageing. IR is associated with chronic inflammation and can lead to dyslipidaemia and metabolic syndrome, which, in turn, increases the risk of age-related morbidities such as diabetes or cardiovascular diseases. Together with a distinct regulation between IR and IS, these observations indicate a large-scale realignment of chronic inflammatory processes during ageing, which may be accelerated in individuals with IR. Moreover, we observed strong sex-specific ageing signatures for lipid subclasses including CE. As cholesterol is a precursor of steroid hormones, many of these sex differences in CE levels may relate to or even cause differences in sex-specific hormone levels. In addition to ageing-related differences, we found significant differences in both the effect sizes and the direction of the correlation of lipid subclasses and clinical measures between participants with IR and those with IS. Importantly, our models controlled for—among other variables—BMI, which is commonly associated with IR. Our analysis therefore highlights significant IR/IS differences separate from BMI effects. Key IR/IS-related differences include clinical markers related to T2D and the immune response; haematological, hepatic and renal measures; as well as electrolytes. For instance, PE-P/PE-O, LPE and LPC showed distinct directions of associations with multiple clinical measures, such as the LDL to HDL ratio and various cell populations, for IR and IS, expanding the previously suggested interplay between clinical measures and complex lipids. As many lipids have both bioenergetic and signalling functions, our observation indicates important differences in cellular signalling in participants with IR. Overall, these results underscore that the holistic assessment of metabolic health state is highly useful, and in some cases necessary, to improve the interpretation of conventional clinical measures such as the LDL to HDL ratio. We identified distinct lipid subclasses and species that change in various disease states such as acute (RVI) and chronic (diabetes, ageing) inflammatory conditions. The role of lipids in immunity is only partially understood given the complexity of both the immune system and lipid metabolism. Our analysis covered myriad positive and negative cytokine–lipid associations, providing a valuable reference and resource for future studies. An important finding in this study is the strong positive associations of GM-CSF, leptin and the chemokine eotaxin (CCL11) with small TAG species (Fig. 6). As GM-CSF and leptin are involved in regulating and promoting inflammation, their strong associations with small TAGs underscore the central and distinct role these lipids may have in immunoregulation in comparison with groups of large TAGs. Moreover, our data provide strong evidence for the pleiotropic role of LPCs, which are categorized as proinflammatory. We found a strong positive association with proinflammatory signalling molecules, including IL-1⍺ and sCD40L, in addition to positive associations with growth factors such as BDNF and negative correlations with the proinflammatory markers CRP and SSPG. There may be feedback loops such that a response is balanced, leading to the production of both inflammatory and anti-inflammatory cytokines with different kinetics. Inflammatory responses must be proportionate to the stimulus so that they are effective but not excessive (leading to damage, for example, a cytokine storm). After the initial acute inflammation, the response needs to diminish and anti-inflammatory signals need to increase. This process needs to be tuned in both magnitude and kinetics, and associated lipids may have a role in these responses. Although this study was not designed to mechanistically resolve the role of LPCs and other lipid subclasses that we observed to be longitudinally associated with, for instance, viral infections, our data provide a resource for future studies and highlight putatively competing roles that may depend on physiological and pathological contexts, including comorbidities, age, BMI or IR/IS status. Although the study cohort is ethnically diverse and sex balanced similar to the US population, there are some limitations. (1) There is a bias towards middle-aged and highly educated participants with a higher proportion of individuals living in northern California. (2) Some of the insights were generated based on small sample numbers (for example, outlier analysis; Supplementary Fig. 6). Therefore, not all observations and findings may be generalizable to a wider population that is subject to different lifestyles. Although we identified many signatures that have been observed in previous studies, supporting the validity of our findings, lifestyle differences can have an effect and future studies should be designed to confirm and extend our observations. (3) Our lipidomics pipeline targets >1,100 lipid species but does not always resolve the exact molecule identity (for example, position of a double bond in FAs). To expand the set of lipids investigated in this study, we included phosphatidylinositol (PI) as a new lipid class in multiple-reaction-monitoring transitions (MRMs) and observed multiple PI species to be significantly altered, for example, during RVIs. For this lipid class, however, no internal spike-in standards were included in the sample preparation, which limits the accuracy of the abundance information for all PIs. Moreover, not all putatively detectable blood plasma lipids, including PC-O and PC-P, were monitored. (4) The biphasic extraction method used here is an established and efficient procedure that is compatible with high-quality non-glass laboratory equipment. However, we have observed that FFAs have higher baseline levels when extracted with non-glass material, which can lead to an overestimation of the levels of certain FFAs. This ratio compression can reduce the sensitivity for detecting subtle changes between participants. (5) We used descriptive models (that is, we did not evaluate the predictive power by using cross-validation and hold-out data). Although we confirmed many previous findings and observed new lipid–phenotype associations, correlations are not proof of causation. In addition, our models controlled for multiple covariates (for example, sex and BMI), allowing for separation of effects, such as lipid–ageing versus lipid–BMI associations. Ageing is a complex process, and increasing BMI may be a default process for industrialized societies and lifestyles. Moreover, high BMI has been linked to inflammation. Thus, controlling for BMI may not always be desirable for understanding age- and inflammation-related changes in these cohorts. (6) The roles of complex lipids are diverse, and the physiological impacts of lipids with specific properties (for example, the fraction of complex lipids containing arachidonic acid) may not be due to direct effects. Our study provides a longitudinal, in-depth analysis of intricate lipid–health relationships during acute and chronic inflammation, metabolic diseases and ageing. The multitude of lipid–phenotype connections we have revealed here serve as a valuable resource for biomarker discovery, a starting point to investigate disease mechanisms, and a basis to conceive therapeutic and preventive strategies. For instance, although many lipids undergo intestinal lipolysis driven by endogenous enzymes and microbial processes, some dietary lipids are absorbed directly. Our research suggests a variety of potential dietary interventions that could improve human health. Supplements such as ether-linked PEs may alleviate inflammation, providing effective treatment for chronic inflammation and facilitating recovery after RVI by minimizing damage related to reactive oxidative agents. Alongside adjusting the small TAG to large TAG ratio, these lipids could also be instrumental in addressing ageing- and IR-associated dyslipidaemia. Moreover, some lipid species and subclasses, including FFAs, ether-linked PEs, CEs or CERs, demonstrate sex-specific and ageing-related patterns. This prominent sex dimorphism suggests that sex- and age-specific interventions should be considered and might enhance therapeutic effects. It would also be worthwhile to assess the impact of genetic polymorphisms, particularly in relation to lipid species and subclasses exhibiting high variance, such as certain TAG, DAG and PE species, which may be relevant for personalized interventions. Together, future studies should explore how altering the exogenous lipid intake (for example, through diet) or targeting lipid conversion enzymes can affect both plasma lipid signatures and clinical phenotypes such as IR, acute and chronic inflammation, and molecular ageing. Participants were enrolled as ‘healthy volunteers’ in the framework of the National Institutes of Health integrated Human Microbiome Project 2 (ref. ). Inclusion and exclusion criteria were previously described in detail. Participants provided informed written consent for the study under research study protocol 23602 approved by the Stanford University Institutional Review Board. Plasma samples were prepared and analysed in a randomized order. Plasma was thawed on ice, and lipids were extracted using a biphasic separation technique (ice-cold methanol, methyl tert-butyl ether and water). A 260-μl volume of methanol and 40 μl of a spike-in standard (cat. no. 5040156, Sciex) were added to 40 μl of plasma, and the mixture was vortexed for 20 s. Lipids were extracted by adding 1,000 μl of methyl tert-butyl ether and incubating the samples under agitation for 30 min at 4 °C. Phase separation was induced by adding 250 μl of ice-cold water, followed by vortexing for 1 min and centrifugation at 14,000g for 15 min at 4 °C. The upper phase containing the lipids was collected, dried down under nitrogen and stored at −20 °C in 200 μl of methanol. On the day of MS acquisition, lipids were dried down under nitrogen and reconstituted with 300 μl of 10 mM ammonium acetate in a 9:1 mixture of methanol and toluene. The QTRAP 5500 system (Sciex) equipped with a DMS device (Lipidyzer) was operated with a Shimadzu SIL30AC autosampler for targeted lipidomics, with a modified strategy to include additional lipid species in the acquisition. To ensure robustness of results, the Lipidyzer was cleaned and tuned after each batch (every 48 h; Supplementary Fig. 14). The tuning solution contained 40 μl of the SPLASH internal standard mix, 100 μl of the Sciex tuning mix, 100 μl of Lyso-tune mix (1 mg ml 17:1 lysophosphatidylglycerol (LPG), 1 mg ml 17:1 lysophosphatidylserine (LPS), 0.1 mg ml 17:1 lysophosphatidylinositol (LPI), 10 μg ml lysophosphatidic acid (LPA)) and 760 μl toluene–methanol (1:9) with 10 mM ammonium acetate. For lipid extracts from 40 μl of plasma, three acquisition methods were used. The injection volumes were 42, 50 and 39 μl for methods 1, 2 and 3, respectively. The source temperature was set to 150 °C for all methods. Methods 1 and 3 were operated with DMS enabled and at a separation voltage of 3,700 V. Lipid classes were monitored as follows: method 1—PC (140), PE (119), PE-O (36), PE-P (61), LPC (26), LPE (26); method 2—CE (26), CER (12), DCER (12), HCER (12), LCER (12), FFA (26), TAG (519), DAG (59); method 3—SM (12), phosphatidic acid (PA; 77), LPA (12), phosphatidylglycerol (PG; 78), LPG (16), PI (77), LPI (16), phosphatidylserine (PS; 78), LPS (16). Each transition was acquired 20 times (see Supplementary Data 2 for the compensation voltage, Q1 and Q3 masses, and dwell times). Method 1, method 2 and the positive mode of method 3 (SM) contained transitions of the Lipidyzer original setup. Method 3 negative mode targets additional lipids. Data acquisition was performed similarly to the processing of the Lipidyzer Workflow Manager. First, *.wiff files were converted to *.mzML files with MSConvert (v.3.0), setting ‘write index’ and ‘TPP compatibility’ to true. For each raw file, data extraction was performed in R. In brief, *.mzML files were imported with openMSfile(FileAndPath, backend = “pwiz”)chromatograms(‘openMSfile_output) using mzR (v.2.6.2). Next, all transitions with more than two zero intensities throughout the 20 repeated measurements were excluded (reported as ‘not available’). For all remaining transitions, the mean intensity was calculated (excluding zero-intensity recordings). Lipid species identities were matched based on the Q1 and Q3 masses and the corresponding scan index (the order in which MRMs were scheduled) in methods 1 (negative mode: PC, PE, LPC, LPE), 2 (negative mode: FFA; positive mode: TAG, CE, DAG, CER, DCER, HCER, LCER) and 3 (negative mode: LPG, PG, LPI, PI, LPS, PS, LPA, PA; positive mode: SM). Note that the transitions PG, PS, PA and their respective lyso forms were not considered for analysis. For lipids monitored in methods 1 and 2, as well as SM (method 3), internal standards from the Sciex Lipidyzer internal spike (LPISTDKIT-102b) were matched according to the Sciex Lipidyzer protocol. Individual concentrations were estimated based on the known abundance of the corresponding spike-ins. In brief, concentrations (‘actual concentration’) of all spike-in standards were retrieved from the ‘certificate of analysis’ of the internal spikes and converted to ‘nmol ml’. Lipidyzer assumes that, in plasma, nmol g = nmol ml. Internal spike stocks of individual lipid classes were mixed, dried down and resuspended in a volume to adjust their respective stock concentration to the expected plasma levels (using the Lipidyzer Workflow Manager as a reference). The internal spike area measured by MS was compared with that of the respective endogenous lipids to approximate the absolute concentrations of the endogenous lipids. For complex lipids with two identical FAs, it was assumed that the measured signal from the fragment ions was at 2× intensity. Lipids belonging to the additional classes in method 3 had no corresponding standards and were normalized based on one of the other spiked-in lipids, as detailed in the next section. The Lipidyzer resolves an individual FA as part of a TAG within each transition, a layer of information that we use to evaluate changes in FA compositions. For analyses that do not rely on the specific FA composition depicted in Fig. 3c, we aggregated TAGs to groups defined by summed FA carbons and the number of unsaturations, summing the untransformed concentrations of the corresponding TAGs. Spike-in standards (internal spikes) were not available at the time of lipid extraction for the lipids monitored here. Although this allows a relative comparison across samples given a reproducible workflow, we desired to leverage the information of the other internal spikes to further normalize for variances introduced across samples. To that end, we performed a correlation analysis by using QC samples derived from the same stocks that were measured across all batches. The lipid intensities in these samples are expected to be the same, and the ratios of internal lipids compared with internal spikes will be the same if the variation introduced by lipid extraction or MS analysis affects them in a similar manner. This setup allowed us to identify spike-in standards to normalize the additional lipids monitored in method 3 (PA, LPA, PG, LPG, PS, LPS, PI, LPI). For this normalization, we only considered internal spikes of the classes PC, PE, LPC and LPE, as those were also acquired in positive mode with DMS enabled. We calculated the Pearson correlation coefficients for log10(intensities) comparing internal spikes and the new lipid species, and selected pairs according to the following hierarchy: (1) the highest correlating internal spike with at least 50% complete observations across QC samples was selected; (2) if a match could not be determined for a lipid species–internal spike pair, we selected the internal spike that showed the highest correlation with any other lipid of the same class; and (3) if both (1) and (2) did not select an internal spike, the highest correlating spike-in standard across all additional lipid classes was selected. For all original lipid classes, internal spikes of known concentrations allow the approximation of absolute abundances. As described above, the missing internal spikes for new lipids in method 3 do not allow direct inference of absolute abundances. Using a linear regression model based on all the known concentrations of lipids in the samples, we predicted the concentrations of the new lipids. As the normalized abundances for these lipids (method 3) are not based on labelled spike-in of the same molecular class and thus do not account for ionization efficiency differences, they provide an estimate of the absolute abundance range of the new classes. Importantly, this normalization does not affect the comparison of the relative abundance of the same lipid species across samples. To ensure the accuracy and reliability of our analysis, we implemented several data filtering criteria. First, we excluded from the analysis biosamples with >25% missing data. Additionally, lipids with <10% valid values, as determined by the Lipidyzer reporting requirements, were also excluded. To further ensure high-quality quantitative results for the results presented in Figs. 2–6, we removed any lipid with a CV of >20% in QC samples and the few lipids for which the CV in QC samples was higher than the CV across the remaining biosamples. Furthermore, owing to limitations in separation by DMS, we did not include PAs in our analysis. We also excluded PS/LPS and PG/LPG from the analysis as they showed a significant number of missing values. PI(16:0/18:3) was removed from the dataset owing to its association with incorrect masses. Finally, QC_73 from batch 21 was removed owing to separate clustering compared with all other QC samples. Four internal spikes were not consistently quantified across the samples (missingness rate >5%) and were substituted with similar deuterated (d) standards belonging to the same class: dDAG(16:0/18:3) missing in 21% of the samples was substituted with dDAG(16:0/18:2); dDAG(16:0/20:5) missing in 12% was substituted with dDAG(16:0/20:4); dPE(18:0/22:5) missing in 44% was substituted with dPE(18:0/20:4); and dPE(18:0/20:5) missing in 8% was substituted with dPE(18:0/20:4). Lipids were normalized based on the internal spike-in standards (see above), similar to the standard Lipidyzer workflow that has been validated previously. Within an expanded method published by Su et al., additional MRMs can be used for isotope correction. Here, we did not acquire all of the MRMs needed for isotope correction. Although the extent of correction depends on the abundance of interfering species and can significantly mask the signal of a targeted lipid species, Su et al. reported that no species was corrected by >6% and, outside of TAGs, no species was corrected by >3% (ref. ). Cytokines were obtained in three separate batches. Data were log2 transformed and corrected for the effect of the batches using the ‘dbnorm’ (v.0.2.2) package. The ComBat model (sva (v.3.38.0)) showed the best performance and thus was considered in this study. Missing values were imputed using a K-nearest-neighbour strategy that accounts for a truncated distribution (Extended Data Fig. 2). This approach involves drawing from intensities at the detection limit defined for each lipid class separately. This was a reasonable yet conservative assumption that allowed for the imputation of missing values without inflating fold changes by considering the sensitivity of MS. Missing weight measures in the ageing analysis were imputed by taking the mean between the two closest adjacent timepoints (Supplementary Data 2). Missing levels of several cytokines in batch 1 (that is, hepatocyte growth factor (HGF), basic fibroblast growth factor (FGFb), IL-8, IL-9, MIP-1⍺, stem cell factor (SCF) and tumour necrosis factor-β (TNFβ)) and batch 2 (that is, IFN⍺2 and FGFb) were imputed using the K-nearest-neighbour method with the number of neighbours being 10. CV was calculated for non-imputed, untransformed data. If, for a participant, multiple samples were collected on the same day, those were excluded. Only participants with at least three sampling timepoints were considered. Lipids with fewer than three quantifications were excluded. Note that, although the average intraparticipant CV was larger than the average QC CV, a small subset of low-abundance lipid species for a subset of participants showed lower CVs. These CVs emerged from the low signal that results in discrete quantifications from the mass analyser. t-SNE scatterplots were generated after log2 transformation and z-score scaling of the data using the R package ‘Rtsne’ (v.0.15) with the following parameters: perplexity = 5, θ = 0.5. Network analysis using self-reported healthy samples (Fig. 2d) was performed using the WGCNA R package (v.1.70-3). The soft-threshold power was optimized to achieve approximate scale-free topology (R > 0.8). Networks were constructed using the ‘blockwiseModules’ function. The network dendrogram was created using average linkage hierarchical clustering of the topological overlap dissimilarity matrix (1 − TOM). Modules were defined as branches of the dendrogram by using the hybrid dynamic tree-cutting method, selecting a minimum module size of 5. Modules were presented by their first principal component (module eigengene) of the standardized expression profiles. Modules with eigengene correlations of >0.8 were merged together, generating seven lipid modules. Next, the Pearson correlation coefficients between the module eigengene and clinical measures were calculated using the ‘cor.test’ function in R (stats (v.3.6.2)), and all the obtained P values for the correlations were corrected for multiple hypotheses through the Benjamini–Hochberg (BH) procedure (stats (v.3.6.2)). To determine over- and under-representation of functional subgroups of lipids, we classified all lipid species based on their physicochemical properties (Supplementary Data 2) as reported in the LION database. Under- or over-representation was evaluated using a hypergeometric test (Fisher’s exact test) or one-dimensional annotation enrichment. Our dataset is dominated by TAG species, which could bias the enrichment analysis results. For that reason, we performed the enrichment analysis across all lipids and within subclasses. For each figure, the following statistics were used to calculate enrichments for categories: ‘OddEven_All’, ‘Omega_All’, ‘Saturation_All’, ‘Lipid_Class_Detailed_Special’ and ‘FA_All’, applying a BH FDR set to 0.1 by using the ‘p.adjust()’ function. Figure 3e shows the Fisher’s exact test comparing positive SSPG coefficients with negative SSPG coefficients. Figure 4c shows the Fisher’s exact test determining whether the significantly changed lipids during infection (RVI) were enriched. Figure 5f shows the Fisher’s exact test comparing positive Δage coefficients with negative Δage coefficients. Figure 6f shows the Fisher’s exact test calculating enrichments in positive coefficients as well as negative coefficients. Negative infinite log2(odds) and positive infinite log2(odds) were imputed with 0.5× the minimum log2(odds) and 0.5× the maximum log2(odds), respectively. If a category was enriched among negative and positive coefficients for a lipid (that is, an over-representation was observed in the FDR-significant positive and negative coefficients), the enrichment was set to 0 and highlighted in black in the heat map. Correlations between lipid levels and clinical measures were calculated using the ‘cor.test’ function in R. Lipidomics data and laboratory measures were both standardized before correlation calculation. To investigate differences between the correlations in IR and IS, we calculated the correlations by using only the healthy samples from participants with IR and IS, and we highlight only the correlation contrasts that were significantly different between IR and IS. The correlations between lipid levels and clinical measures from using all healthy samples from participants with IR and IS were also calculated and are presented as reference values. K-means clustering was performed to investigate lipid similarities following infection by using the lipidomics data of infection events after log2 transformation and z-score scaling. We calculated the minimum centroid distance for a range of cluster numbers, and the optimal number was chosen using the ‘elbow’ method. The median values of the lipid profiles belonging to each cluster were correlated with the clinical measures to indicate medical implications. The correlations were calculated using the ‘cor.test’ function in R, and all the obtained P values for the correlations were corrected for multiple hypotheses using the BH procedure. To detect the time intervals of differentially abundant lipids between IR and IS during infection events, we used a longitudinal analysis method, OmicsLonDA. For each lipid in each group (IR or IS), we used a generalized additive mixed model for modelling nonlinear time-series abundance during the inflammation episodes. OmicsLonDA is an extension of MetaLonDA to account for correlated data, repeated measurements and multiple covariates. We accounted for sex, age, ethnicity and BMI as covariates, whereas participant identifiers were used as random effects. The P values for each lipid at each time interval (the time interval unit was set to 1 day) were obtained and then adjusted for multiple testing by using the BH procedure. We implemented this process on both infection and immunization events, identified the significantly different time intervals between IR and IS in both kinds of events, and compared these significant time intervals. To identify lipids that were associated with SSPG, linear mixed models were applied using log-transformed lipid measurements, controlling for participants, sex, ethnicity, age and BMI (Fig. 3). The R package ‘lme4’ (v.1.1-27.1) was used to construct the linear mixed models, as well as output estimates and nominal P values. The obtained raw P values were corrected for multiple hypotheses through the BH procedure by using the ‘p.adjust’ function in R. To identify lipids that were significantly changed during infection episodes, linear mixed models were applied using log-transformed lipid measurements, controlling for participants, sex, ethnicity, age and BMI (Fig. 4). The R package ‘lme4’ (v.1.1-27.1) was used to construct the linear mixed models, as well as output estimates and nominal P values. The obtained raw P values were corrected for multiple hypotheses through the BH procedure by using the ‘p.adjust’ function in R. For each individual, ageing-associated lipid changes were calculated by subtracting measurements obtained at each visit from the baseline values (Fig. 5). Accordingly, the number of years since onset was calculated as the number of years from the first recorded measurement. To estimate the fractional changes in lipid measurements, we used a linear regression model with log-transformed lipid measurements, controlling for BMI and storage length (and sex if indicated in the figure). To control for potential biases related to the number of samples per individual, we excluded measurements from one participant with a uniquely large number of samples. To control for potential biases related to a few participants with measurements spread across a longer enrolment time, we excluded a few samples collected >5 years since onset. All coefficients and s.d. values were estimated using the ordinary least-squares implementation of the linear regression method in the ‘statsmodels’ package with the default parameters in Python (v.3.7). Linear models were run either at the lipid species level (Fig. 5d) or the lipid class level (sum of raw concentrations) for all participants (Fig. 5c), as well as for sex and IR/IS (Fig. 5e,f). We used linear mixed-effects models (lmer ) controlling for BMI, sex, ethnicity and participants (random effects) to estimate cytokine levels as a function of estimated lipid concentrations (Fig. 6). Both cytokine levels and lipid signals were scaled and centred (scale()). Restricted maximum likelihood was set to false, and P values were estimated using summ and corrected for multiple-hypothesis testing with p.adjust() applying a BH FDR of <5% for network generation. The cytokine–lipid network was constructed based on model coefficients by using the ‘graphopt’ layout algorithm in graphlayout (v.2.1.0), igraph (v.1.5.0) and tidygraph (v.1.2). The network was pruned to exclude all coefficients with a BH FDR of >0.05. Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
|
PMC6691112
|
Alterations in lipid metabolism of spinal cord linked to amyotrophic lateral sclerosis
|
Amyotrophic lateral sclerosis (ALS) is characterized by progressive loss of upper and lower motor neurons leading to muscle paralysis and death. While a link between dysregulated lipid metabolism and ALS has been proposed, lipidome alterations involved in disease progression are still understudied. Using a rodent model of ALS overexpressing mutant human Cu/Zn-superoxide dismutase gene (SOD1-G93A), we performed a comparative lipidomic analysis in motor cortex and spinal cord tissues of SOD1-G93A and WT rats at asymptomatic (~70 days) and symptomatic stages (~120 days). Interestingly, lipidome alterations in motor cortex were mostly related to age than ALS. In contrast, drastic changes were observed in spinal cord of SOD1-G93A 120d group, including decreased levels of cardiolipin and a 6-fold increase in several cholesteryl esters linked to polyunsaturated fatty acids. Consistent with previous studies, our findings suggest abnormal mitochondria in motor neurons and lipid droplets accumulation in aberrant astrocytes. Although the mechanism leading to cholesteryl esters accumulation remains to be established, we postulate a hypothetical model based on neuroprotection of polyunsaturated fatty acids into lipid droplets in response to increased oxidative stress. Implicated in the pathology of other neurodegenerative diseases, cholesteryl esters appear as attractive targets for further investigations.The central nervous system (CNS) is characterized by the presence of high amounts and a wide variety of lipids. According to their molecular characteristics and cellular localization, lipids play a critical role in the CNS controlling membrane fluidity, improving transmissions of electrical signals and serving as precursors for various second messengers. Although molecular alterations in CNS at different stages of life are part of the physiological development, aging can modulate alterations in the lipidome, especially in response to increased reactive oxygen species (ROS). As a consequence, alterations in lipid metabolism could contribute to the onset of neurodegenerative disorders. This paper is focused on ALS, a neurodegenerative disorder characterized by loss of motor neurons in CNS. Symptomatic stages of ALS lead to muscular atrophy, paralysis and death within 5 years after the onset of symptoms. Among familial cases of ALS, G93A mutation in the gene that encodes for the antioxidant enzyme Cu/Zn-superoxide dismutase (SOD1-G93A) is one of the most studied models. The putative mechanisms involved in ALS progression consist in protein aggregation of mutant SOD1, abnormal production of ROS and alterations in mitochondrial functions. Lipid alterations in ALS have been mainly investigated through targeted analysis of some specific lipid species. Historically, a pioneer study by Cutler et al. (2002) revealed increased levels of sphingolipids and cholesteryl esters based on targeted lipid analysis of the spinal cord from an ALS mouse model. In addition, these authors suggested a link between sphingolipid metabolism and synthesis of cholesteryl esters in the CNS. Alterations in sphingolipid metabolism were also emphasized in skeletal muscle and spinal cord from SOD1 mice. Increased sphingolipid content, specially ceramides and glucosylceramides, as well as accumulation of phosphatidylcholine (PC 36:4) were reported as the main lipid alterations in cerebrospinal fluid of ALS patients. Collectively, these studies and others have found evidence for significant lipid metabolism alterations in advanced stages of ALS. Here, we sought to capture global lipidome alterations linked to ALS progression by performing an untargeted mass spectrometry-based lipidomic analysis of motor cortex and spinal cord tissues from ALS asymptomatic (SOD1-G93A 70 days old) and symptomatic (SOD1-G93A 120 days old) rats in comparison to their wild types as controls. Our analysis revealed drastic changes in lipid metabolism in the CNS both according to aging and disease progression. To investigate how lipids are affected in ALS, we performed a global lipidome analysis of motor cortex and spinal cord from asymptomatic (SOD1-G93A 70 days) and symptomatic (SOD1-G93A 120 days) ALS rats, and their respective age-matched wild type (WT 70 days and WT 120 days) as control groups. Results from these analyses are presented separately below. Using untargeted lipidomics, we identified and quantified 285 lipid molecular species in the motor cortex, which were sorted into 26 lipid subclasses (Fig. 1A). Glycerophospholipids (n = 120) followed by sphingolipids (n = 78), glycerolipids (n = 69) and free fatty acids (n = 14) showed the highest diversity of individual lipid molecular species. In terms of relative abundance, plasmenyl and diacyl phosphatidylethanolamine (pPE and PE) encompassed together about 70% in mass of the total identified lipids in motor cortex of both SOD1-G93A and WT rats (Fig. 1B). The remaining 30% of lipids were mainly composed of phosphatidylserine (PS, ~7%), diacylglycerol (DAG, ~5%), free fatty acids (FFA, ~2%), among others.Figure 1Lipidome profile in motor cortex of SOD1-G93A and WT groups at 70 and 120 days old. (A) Number of the identified lipid molecular species per lipid subclass. (B) Relative abundance of summed concentrations of each lipid subclass. Abbreviations: FFA, free fatty acids; CL, cardiolipin; PC, phosphatidylcholine; PE, phosphatidylethanolamine; PG, phosphatidylglycerol; PI, phosphatidylinositol; PS, phosphatidylserine; Cer, ceramide; GalC, galactosylceramide; 1G-aC, monoglycosylated acylceramide; SM, sphingomyelin; 1G-AEG, monoglucosyl-acyl-ether-glycerol; DAG, diacylglycerol; TAG, triacylglycerol; Ch, cholesterol; CE, cholesteryl ester; UbQ, ubiquinone; p- or o- before phospholipids (PC, PE, PG and PS) indicate plasmenyl or plasmanyl, respectively. Lipidome profile in motor cortex of SOD1-G93A and WT groups at 70 and 120 days old. (A) Number of the identified lipid molecular species per lipid subclass. (B) Relative abundance of summed concentrations of each lipid subclass. Abbreviations: FFA, free fatty acids; CL, cardiolipin; PC, phosphatidylcholine; PE, phosphatidylethanolamine; PG, phosphatidylglycerol; PI, phosphatidylinositol; PS, phosphatidylserine; Cer, ceramide; GalC, galactosylceramide; 1G-aC, monoglycosylated acylceramide; SM, sphingomyelin; 1G-AEG, monoglucosyl-acyl-ether-glycerol; DAG, diacylglycerol; TAG, triacylglycerol; Ch, cholesterol; CE, cholesteryl ester; UbQ, ubiquinone; p- or o- before phospholipids (PC, PE, PG and PS) indicate plasmenyl or plasmanyl, respectively. Multivariate analysis was performed by sparse partial least squares analysis (sPLS-DA) in the motor cortex lipidome. The score plot of the sPLS-DA showed a relatively clear segregation of experimental groups according to age (component 1 = 18.1%, Fig. 2A). This trend was related to high levels of several sphingolipids observed in older compared to younger animals as revealed by the top 20 loadings 1 values (Fig. 2B). Univariate analysis performed by one-way ANOVA yielded 61 altered lipid molecular species when comparing all groups. These altered lipids are displayed as clusters in the heatmap plot (Fig. 2C). Of note, most samples from both SOD1-G93A 120d and WT 120d clustered together displaying increased levels of galactosyl ceramides (GalC; 14 species), sulfatides (16 species) and monoglycosylated acylceramides (1G-aC; 8 species). The clustering of samples in the heatmap plot and sPLS-DA suggest that age, rather than the disease per se, modulates major lipidome alterations in motor cortex.Figure 2Multivariate analysis and heatmap plot of motor cortex lipidomes from SOD1-G93A and WT groups. Log transformation of lipid concentrations was applied to lipid concentrations prior to statistical analysis in Metaboanalyst. (A) Score plot of the sparse partial least squares analysis (sPLS-DA); (B) Top 20 lipid species according to loadings 1 values of the sPLS-DA (C) Heatmap plot displaying clusters of samples and the 61 significantly altered lipid molecular species according to one-way ANOVA followed by Tukey’s post-test (p < 0.05; FDR-adjusted). Individual lipids are shown in rows and samples displayed in columns, according to cluster analysis (clustering distance was calculated by Pearson and clustering algorithm estimated by Ward). Each colored cell on the heatmap plot corresponds to values above (red) or below (blue) the mean normalized concentrations for a given lipid. Abbreviations of lipid subclasses are described in Fig. 1. Nomenclature: “p” before a given fatty acyl chain indicates sn-1 plasmenyl followed by a sn-2 linked fatty acid (as in pPE (p18:0/18:1); “d” refers to the sphingoid base followed by a n-acyl chain, for example, Sulfatide (d18:1/24:1). “OH” refers to the hydroxyl group in the n-acyl chain, for example, Cer (18:0/24:1-OH). Please note that fatty acid composition of phospholipids, TAG and DAG does not reflect their specific position in the glycerol backbone. GL, glycerolipids; GP, glycerophospholipids; SP, sphingolipids and TG, triglycerides. Multivariate analysis and heatmap plot of motor cortex lipidomes from SOD1-G93A and WT groups. Log transformation of lipid concentrations was applied to lipid concentrations prior to statistical analysis in Metaboanalyst. (A) Score plot of the sparse partial least squares analysis (sPLS-DA); (B) Top 20 lipid species according to loadings 1 values of the sPLS-DA (C) Heatmap plot displaying clusters of samples and the 61 significantly altered lipid molecular species according to one-way ANOVA followed by Tukey’s post-test (p < 0.05; FDR-adjusted). Individual lipids are shown in rows and samples displayed in columns, according to cluster analysis (clustering distance was calculated by Pearson and clustering algorithm estimated by Ward). Each colored cell on the heatmap plot corresponds to values above (red) or below (blue) the mean normalized concentrations for a given lipid. Abbreviations of lipid subclasses are described in Fig. 1. Nomenclature: “p” before a given fatty acyl chain indicates sn-1 plasmenyl followed by a sn-2 linked fatty acid (as in pPE (p18:0/18:1); “d” refers to the sphingoid base followed by a n-acyl chain, for example, Sulfatide (d18:1/24:1). “OH” refers to the hydroxyl group in the n-acyl chain, for example, Cer (18:0/24:1-OH). Please note that fatty acid composition of phospholipids, TAG and DAG does not reflect their specific position in the glycerol backbone. GL, glycerolipids; GP, glycerophospholipids; SP, sphingolipids and TG, triglycerides. Alterations in sphingolipids concentration evidenced by both multivariate and univariate analyses occurred in several individual molecular species and included variations in the composition of sphingosine base, hydroxylation and length of n-acyl chains (Fig. S1). Taken together, our findings suggest that age plays a critical role modulating sphingolipids metabolism in motor cortex. In line with previous studies, our data indeed revealed that some sphingolipids, particularly the summed concentrations of GalC and 1G-aC (Fig. S2), were significantly increased in the SOD1-G93A 120d group relative to the other groups. This finding suggests a link between ALS progression and lipidome alterations in motor cortex, and we thus decided to examine specific differences between SOD1-G93A 120d and WT 120d groups. For this purpose, we performed a comparison between these groups through an OPLS-DA and volcano plot (Fig. 3). As shown in Fig. 3A,B, the discrimination between these groups in multivariate analysis is mainly linked to alterations in the levels of TAG, DAG and one specimen of pPE (top 20 lipid species ranked according to loading 1 values of the OPLS-DA). Differences in lipidome alterations evidenced by multivariate analysis need to be interpreted with caution since univariate analysis by volcano plot yielded a single altered specimen (i.e., pPE (p18:0/16:0)). These results reinforce the idea that lipid alterations in motor cortex are clearly more linked to age than disease progression per se.Figure 3Pairwise comparison by orthogonal partial least squares–discriminate analysis (OPLS-DA) and volcano plot of motor cortex lipidomes of SOD1-G93A 120d and WT 120d groups. Data used for this comparison have been Log-transformed prior to statistical analysis in Metaboanalyst. (A) Score plot of the OPLS-DA revealing a clear segregation of groups; (B) Loadings plot of the OPLS-DA. Colored features representing lipid subclasses displaying the top 20 lipid species according to highest p values; (C) Volcano plot of the pairwise comparison SOD1-G93A 120d versus WT 120d groups represented by the log2 (fold change) plotted against the –log10 (p value). This analysis revealed only one molecular specimen altered between groups. Statistical significance was evaluated by t-test (p < 0.05; FDR-adjusted) and the fold change was set to >1.5. Pairwise comparison by orthogonal partial least squares–discriminate analysis (OPLS-DA) and volcano plot of motor cortex lipidomes of SOD1-G93A 120d and WT 120d groups. Data used for this comparison have been Log-transformed prior to statistical analysis in Metaboanalyst. (A) Score plot of the OPLS-DA revealing a clear segregation of groups; (B) Loadings plot of the OPLS-DA. Colored features representing lipid subclasses displaying the top 20 lipid species according to highest p values; (C) Volcano plot of the pairwise comparison SOD1-G93A 120d versus WT 120d groups represented by the log2 (fold change) plotted against the –log10 (p value). This analysis revealed only one molecular specimen altered between groups. Statistical significance was evaluated by t-test (p < 0.05; FDR-adjusted) and the fold change was set to >1.5. The lipidomic analysis of spinal cord revealed 406 lipid molecular species sorted into 33 lipid subclasses. Glycerophospholipids (n = 177) followed by sphingolipids (n = 101) and glycerolipids (n = 103) showed the highest diversity of individual lipid molecular species in spinal cord tissues (Fig. 4A). The most abundant lipid subclasses were PE and PS which represented together approximately 50% in mass of total lipid molecular species (Fig. 4B). Sphingolipids representing ~13% were the third most abundant lipid subclass, which showed sphingomyelins (SM) and sulfatides as the most abundant subclasses. In comparison to motor cortex lipidomes (Fig. 1), higher percentages of storage lipids (represented by triacylglycerols and cholesteryl esters) and cardiolipin (CL) were found in spinal cord tissues.Figure 4Lipidome profile identified in spinal cord of SOD1-G93A and WT groups at 70 and 120 days old. (A) Number of lipid species identified in each lipid subclass. (B) Relative abundance of summed concentrations of lipid subclasses. Abbreviations: FFA, free fatty acids; CL, cardiolipin; PA, phosphatidic acid; PC, phosphatidylcholine; PE, phosphatidylethanolamine; PG, phosphatidylglycerol; PI, phosphatidylinositol; PS, phosphatidylserine; Cer, ceramide; GalC, galactosylceramide; 1G-aC, monoglycosylated acylceramide; SM, sphingomyelin; 1G-AEG, monoglucosyl-acyl-ether-glycerol; DAG, diacylglycerol; TAG, triacylglycerol; Ch, cholesterol; CE, cholesteryl ester; UbQ, ubiquinone; p- or o- before lipid subclass mean plasmenyl or plasmanyl, respectively. Lipidome profile identified in spinal cord of SOD1-G93A and WT groups at 70 and 120 days old. (A) Number of lipid species identified in each lipid subclass. (B) Relative abundance of summed concentrations of lipid subclasses. Abbreviations: FFA, free fatty acids; CL, cardiolipin; PA, phosphatidic acid; PC, phosphatidylcholine; PE, phosphatidylethanolamine; PG, phosphatidylglycerol; PI, phosphatidylinositol; PS, phosphatidylserine; Cer, ceramide; GalC, galactosylceramide; 1G-aC, monoglycosylated acylceramide; SM, sphingomyelin; 1G-AEG, monoglucosyl-acyl-ether-glycerol; DAG, diacylglycerol; TAG, triacylglycerol; Ch, cholesterol; CE, cholesteryl ester; UbQ, ubiquinone; p- or o- before lipid subclass mean plasmenyl or plasmanyl, respectively. As shown in Fig. 5A, the spinal cord lipidome of SOD1-G93A 120d group is spatially segregated from the others as explained by component 1 of the sPLS-DA (10.3%). Analysis of the top 20 lipid species ranked according to values of loadings 1 of the sPLS-DA reveals enrichment of ceramides and cholesteryl esters in SOD1-G93A 120d group and reduced amounts of cardiolipins relative to the other groups (Fig. 5B). A clear separation of the SOD1-G93A 120d group from others is also observed in the heatmap plot (Fig. 5C), displaying the top 12 significantly altered lipid species from the spinal cord (one-way ANOVA). Univariate analysis revealed modulation of 4 sphingolipids (ceramides and 1G-aC) and all cholesteryl esters species identified by lipidomic analysis. Importantly, 2 ceramides and all cholesteryl esters were increased exclusively in the symptomatic group (Fig. 5C), linking these lipids to ALS progression.Figure 5Multivariate analysis and heatmap plot of spinal cord lipidomes of SOD1-G93A and WT groups. (A) Score plot of the sPLS-DA; (B) Top 20 lipid species according to loadings 1 values of the sPLS-DA; (C) Heatmap plot displaying clusters of 12 significantly altered lipid molecular species (according to one-way ANOVA) and samples. For statistical parameters and graphic details see Fig. 2 caption. Abbreviations of lipid subclasses are described in Fig. 4. Nomenclature: “d” refers to the sphingoid base followed by a n-acyl chain, for example, Cer (d18:0/22:0). “OH” refers to the hydroxyl group in the n-acyl chain, for example, Cer (18:0/24:1-OH). Multivariate analysis and heatmap plot of spinal cord lipidomes of SOD1-G93A and WT groups. (A) Score plot of the sPLS-DA; (B) Top 20 lipid species according to loadings 1 values of the sPLS-DA; (C) Heatmap plot displaying clusters of 12 significantly altered lipid molecular species (according to one-way ANOVA) and samples. For statistical parameters and graphic details see Fig. 2 caption. Abbreviations of lipid subclasses are described in Fig. 4. Nomenclature: “d” refers to the sphingoid base followed by a n-acyl chain, for example, Cer (d18:0/22:0). “OH” refers to the hydroxyl group in the n-acyl chain, for example, Cer (18:0/24:1-OH). To examine detailed differences linked to symptomatic stage of ALS, we performed a pairwise comparison between SOD1-G93A 120d and WT 120d groups through OPLS-DA and volcano plot (Fig. 6). Together, multivariate and univariate analyses confirmed significant alterations in ceramides and cholesteryl esters, which were both found markedly elevated in SOD1-G93A 120d group. In addition, several cardiolipins (CL) species were found reduced in the SOD1-G93A 120d group.Figure 6Pairwise comparison by OPLS-DA and volcano plot of spinal cord lipidomes of SOD1-G93A 120d and WT 120d groups. (A) Multivariate statistical analysis using orthogonal partial least squares–discriminate analysis (OPLS-DA); (B) Feature importance of the top 20 lipid species altered according to the OPLS-DA analysis; (C) Volcano plot is represented by the log2 (fold change) of SOD1-G93A 120d/WT 120d groups plotted against the –log10 (p). Statistical significance was evaluated by t-test (p < 0.05; FDR-adjusted) and the fold change was set to >1.5. The significantly altered lipid molecular species are colored according to lipid subclasses. List of altered cardiolipins is sorted by log2 (fold change). Pairwise comparison by OPLS-DA and volcano plot of spinal cord lipidomes of SOD1-G93A 120d and WT 120d groups. (A) Multivariate statistical analysis using orthogonal partial least squares–discriminate analysis (OPLS-DA); (B) Feature importance of the top 20 lipid species altered according to the OPLS-DA analysis; (C) Volcano plot is represented by the log2 (fold change) of SOD1-G93A 120d/WT 120d groups plotted against the –log10 (p). Statistical significance was evaluated by t-test (p < 0.05; FDR-adjusted) and the fold change was set to >1.5. The significantly altered lipid molecular species are colored according to lipid subclasses. List of altered cardiolipins is sorted by log2 (fold change). Among ceramide species, remarkable changes were observed in Cer (d18:0/24:0-OH) and Cer (d18:0/24:1-OH). These hydroxylated ceramides were drastically elevated in the SOD1-G93A 120d group compared to the age-matched WT group (Fig. S3). In addition, changes in ceramides levels were examined by calculating the ratios of hydroxylated/non-hydroxylated (OH/n-OH) and very long chain/long chain (VLC/LC) species. Both ratios were found significantly increased in the SOD1-G93A 120d group, whereas the ratio ceramide/dihydroceramide (Cer/dh-Cer) was reduced (Fig. S4B). Therefore, these data confirm previous findings indicating an association between ceramide metabolism/remodeling and ALS disease development. The most conspicuous changes in the lipidome of spinal cord in the SOD1-G93A 120d group were related to alterations in cholesteryl esters and cardiolipin levels (Fig. 7). Of note, the SOD1-G93A 120d group showed a 6-fold increase in total concentration of cholesteryl esters (Fig. 7A). Elevated cholesteryl esters levels were mainly linked to those molecular species esterified to polyunsaturated fatty acids (PUFAs), with arachidonic (20:4), eicosapentaenoic (20:5) and adrenic (22:4) acids (Fig. 7B). Contrasting with the increase in cholesteryl esters, 10 out of 32 cardiolipin species were found reduced in the SOD1-G93A 120d group relative to the age-matched WT (Fig. S5). This change was reflected not only in total concentration of cardiolipins, which was lower in SOD1-G93A 120d group (Fig. 7C), but also in the major fatty acids esterified to cardiolipins (Fig. 7D).Figure 7Altered concentrations of cholesteryl esters (CE) and cardiolipins (CL) identified in the spinal cord of SOD1-G93A 120d and WT 120d groups. (A) Concentration of total CE. (B) Fatty acid composition of CE. (C) Concentration of total CL. (D) Fatty acid composition of CL. Data are shown as mean ± standard error of mean (SEM). Statistical significance was evaluated by t-test (p < 0.05; FDR-adjusted) using Metaboanalyst. (*) Statistically different when compared to the WT 120d group. Altered concentrations of cholesteryl esters (CE) and cardiolipins (CL) identified in the spinal cord of SOD1-G93A 120d and WT 120d groups. (A) Concentration of total CE. (B) Fatty acid composition of CE. (C) Concentration of total CL. (D) Fatty acid composition of CL. Data are shown as mean ± standard error of mean (SEM). Statistical significance was evaluated by t-test (p < 0.05; FDR-adjusted) using Metaboanalyst. (*) Statistically different when compared to the WT 120d group. Guided by findings that astrocytes isolated from the spinal cord of an identical ALS rat model has aberrant features and accumulates lipid droplets, we performed immunohistochemistry of the astrocyte-specific glial fibrillary acidic protein (GFAP). This analysis was conducted to confirm astrogliosis, a hallmark of ALS, and it was restricted to the spinal cords of SOD1-G93A 120d group and its age-matched control (WT 120d). As depicted in the Fig. S6, the spinal cord of the SOD1-G93A 120d group exhibited severe astrogliosis (stellate appearance) in the grey matter of the spinal cord ventral horn, as reported in previous studies. Lipids play a critical role in structuring the CNS through membrane fluidity control, transmission of electrical signals and stabilization of synapses. Alterations of lipid metabolism in neurons and glial cells (astrocytes, oligodendrocytes and microglia) modulate processes linked to aging and neurodegenerative diseases. To address lipidome alterations in the CNS, we investigated the motor cortex and spinal cord of SOD1-G93A transgenic rats as model for ALS and compared them to wild type as control. In this study, an untargeted lipidomic analysis using high resolution UHPLC-Q-TOF-MS was performed. Manual identification of the most abundant molecular ions based exclusively on their MS/MS profile allowed us to unambiguously annotate a wide range of lipid molecular species occurring in motor cortex (285 species, Fig. 1) and spinal cord (406 species, Fig. 4) tissues. Major lipidome changes in the motor cortex were linked to altered sphingolipid metabolism and to animal age (Figs 2 and 3). In contrast, lipidome alterations in the spinal cord were strongly associated with ALS at symptomatic stage (SOD1-G93A 120d group) (Figs 5 and 6). Alterations in the lipidome of motor cortex were mostly related to age, except for a single lipid (pPE (p18:0/16:0)) that was significantly decreased in SOD1-G93A 120d relative to age-matched WT (Fig. 3C). Among sphingolipids, several GalC, sulfatide and 1G-aC were increased in both WT and SOD1-G93A groups at 120d compared to 70d (Figs 2 and S1). Our findings are in line with active production of myelin by adults, either from newly formed oligodendrocytes or synthesis of new membranes by existing cells, reflecting the constant increase in sphingolipids content with age. Interestingly, increased content of 1G-aC was observed in motor cortex of older animals (Fig. S2). To our knowledge, 1G-aC have never been reported in the CNS and their modulation with age together with increased GalC and sulfatide concentrations suggest a possible link of these sphingolipids to myelination and brain development. In contrast to motor cortex, drastic differences in lipidome profiles were observed in the spinal cord of SOD1-G93A 120d rats, particularly when compared to its age-matched control (WT 120d) (Fig. 6). Major alterations were linked to cholesteryl esters and cardiolipin levels and to a minor extent in the abundance of ceramides (Figs 7 and S3). Differences in ceramide metabolism have been repeatedly reported in spinal cord tissues of ALS patients and rodent models. Here, we detected increased concentrations of two hydroxylated ceramides, the Cer (d18:0/24:0-OH) and Cer (d18:0/24:1-OH). Hydroxylated ceramides are essential for myelin stability and maturation, as demonstrated by knockdown and mutations in fatty acid 2-hydroxylase (FA2H, the enzyme responsible for hydroxylation of n-acyl chains) that results in neural impairments in mice and humans. Ceramides have also been implicated in cellular signaling, and in ALS they are thought to activate the oxysterol-binding protein (OSBP). According to these authors, OSBP binds and delivers oxysterols to the endoplasmic reticulum thereby activating acylcoenzyme A:cholesterol acyltransferase (ACAT) yielding cholesteryl esters as products. Accumulation of either hydroxylated ceramides or dihydroceramides could also trigger cell survival events such as autophagy. Since one of the hallmarks of ALS is the presence of cytoplasmic inclusions or protein aggregates in affected motor neurons, the stimulation of autophagic process by these ceramides could be a potential pathway for elimination of protein aggregates and damaged organelles. As highlighted by our data, ALS symptomatic rats displayed conspicuously elevated levels of cholesteryl esters and decreased concentrations of cardiolipin species in the spinal cord as major lipid signatures of the disease (Fig. 6). Changes in cholesteryl ester metabolism were not reflected in alterations in the pools of free cholesterol (Fig. S7) or 24-hydroxycholesterol, the latter only observed in trace amounts in spinal cord tissues (data not shown). That is, it appears that cholesterol synthesis is upregulated but does not result in accumulation of free cholesterol nor 24-hydroxycholesterol, which is generally more soluble and can cross the brain-blood-barrier at a much faster rate than cholesterol. A growing body of evidence implicates impaired energy metabolism in ALS patients and models. Of note, ROS generation, protein aggregation and mitochondrial dysfunction in motor neurons represent a clinical hallmark of ALS. Whether elevated ROS is caused by SOD1 aggregation and glutamate accumulation leading to imbalanced calcium homeostasis in neurons remains unanswered. Nonetheless, elevated ROS is associated with mitochondrial dysfunction, leading to changes that range from decreased energy metabolism to major morphological modifications. Abnormal mitochondrial morphology has been repeatedly observed in motor neurons of diverse familial cases of ALS models. These alterations consist of mitochondria clusters, swollen mitochondria, loss of cristae and vacuoles derived from degenerating mitochondria. Reflecting dysfunctional mitochondria, our data reveal a significant decrease in cardiolipin levels in spinal cord of SOD1-G93A 120d group when compared to the WT 120d group (Fig. 6). Cardiolipin is a major phospholipid of mitochondria specifically located at mitochondrial inner membranes. The functions of cardiolipin have been primarily related to ATP production, curvature stress control and mitochondria cristae morphology. Given its function in structuring mitochondrial function and morphology, we suggest that a decrease in cardiolipin levels may partially reflect the loss of mitochondrial cristae and thus dysfunctional mitochondria in the spinal cord of ALS symptomatic rats. Neurons are known to use lactate from glia to fuel the glycolytic pathway. By inducing ROS formation in fruit flies neurons, Liu et al. (2017) described lactate being converted to pyruvate and acetyl-CoA, with the latter driving synthesis of fatty acids that are stored as triacylglycerols in lipid droplets. Importantly, these lipid droplets are transported to and accumulate in glial cells for neuroprotection. A similar coordination of neurons and astrocytes in the metabolism of fatty acids was recently demonstrated in primary hippocampal neurons and astrocytes from rats. To our knowledge, a detailed lipidomic analysis was not performed by these studies. According to our findings, neutral or storage lipid accumulation was evidenced almost exclusively as a significant increase in cholesteryl esters concentration in spinal cords of ALS symptomatic rats, with no major alterations in the pool of triacylglycerols (Fig. S4C,D). A pertinent question is therefore why cholesteryl esters and not triacylglycerols are stored as neutral lipids in droplets? We hypothesize that the answer relies in part in the efficiency by which lipid droplets can be shuttled from neurons to astrocytes by apolipoproteins (ApoE/D) for neuroprotection as demonstrated by Liu et al.. Accumulation of storage lipids in spinal cords of ALS SOD1-G93A rats was already evidenced by isolation of aberrant astrocytes bearing abundant lipid droplets. Not surprisingly and similar to previous studies with ALS models, immunohistochemistry of the astrocyte-specific glial fibrillary acidic protein (GFAP) revealed extensive reactive astrogliosis in the spinal cord ventral horn of the SOD1-G93A 120d group (Fig. S6). Nevertheless, the studies evidencing either astrogliosis or aberrant astrocytes in ALS models have not conducted lipidomic analysis. It is thus tempting to suggest that lipids accumulating in aberrant astrocytes are in fact lipid droplets composed mainly of cholesteryl esters. Of note, the 6-fold increase of cholesteryl esters in spinal cords of SOD1-G93A 120d group relative to WT 120d group was mostly related to accumulation of molecular species esterified to PUFAs such as arachidonic (20:4), eicosapentaenoic (20:5) and adrenic (22:4) acids. Since mammals cannot synthesize these fatty acids, we speculate that this accumulation likely reflects a protective mechanism against oxidative stress. That is, under elevated ROS, neuronal and glial cells store their highly susceptible membrane-bound PUFAs as cholesteryl esters or triacylglycerols, thereby avoiding membrane lipid peroxidation. It is becoming apparent that sequestering toxic fatty acids into lipid droplets is a widespread cellular strategy to avoid lipotoxicity. Once protected in lipid droplets, these neutral lipids may be shuttled to glial cells and undergo fatty acid beta-oxidation. Whether fatty acids can be quantitatively used as fuel for energy metabolism in glial cells remains debatable. The controversy is that fatty acid beta-oxidation requires more oxygen than glucose, and may deplete the oxygen demand by neurons. Furthermore, beta-oxidation also generates superoxide anion that may enhance oxidative stress, and for this reason neurons likely rely on glial cells for storing excess lipids. We suggest that a delicate balance between neuron-glia lipid transport and energy metabolism (i.e. significant contribution of beta-oxidation) is likely relevant to neuroprotection. In retrospect, any perturbation of this feedback loop may result in augmented neurodegeneration. Here, we attempted to formulate a model shown in Fig. 8 based on: 1) elevated ROS and dysfunctional neuronal mitochondria in ALS (supported by the data showing decreased cardiolipin content in spinal cords of SOD1-G93A 120d group); 2) upregulation of cholesterol synthesis due to altered lactate metabolism in neuronal mitochondria; 3) accumulation of cholesteryl esters composed of PUFAs as a protective mechanism against lipid peroxidation. This hypothetical model is adapted from previously published models, with the exception that in our study we performed lipidomics analysis, which revealed significant changes regarding cardiolipin and cholesteryl esters levels related to ALS progression in spinal cord.Figure 8Proposed model for the accumulation of lipid droplets in astrocytes. Under normal conditions, circulating glucose is taken up by astrocytes, converted to lactate and shuttled to neurons for ATP production. Increasing evidences indicate that elevated ROS and mitochondrial defects in neurons trigger lipid droplet formation in astrocytes. In this scenario, our lipidomic data demonstrate that lipid accumulating in the spinal cord are mainly consisted of cholesteryl ester species. Scheme adapted from Liu et al., 2017 and Ioannou et al.. Proposed model for the accumulation of lipid droplets in astrocytes. Under normal conditions, circulating glucose is taken up by astrocytes, converted to lactate and shuttled to neurons for ATP production. Increasing evidences indicate that elevated ROS and mitochondrial defects in neurons trigger lipid droplet formation in astrocytes. In this scenario, our lipidomic data demonstrate that lipid accumulating in the spinal cord are mainly consisted of cholesteryl ester species. Scheme adapted from Liu et al., 2017 and Ioannou et al.. Our findings reiterate that alterations in cholesterol metabolism of CNS are associated with ALS. In other neurodegenerative diseases such as Alzheimer’s disease (AD), inhibition of cholesteryl ester synthesis by deletion of ACAT1 prevented Tau phosphorylation and restored cognitive deficits. More recently, the mechanism of cholesteryl esters accumulation avoiding phosphorylated Tau degradation was attributed to a negative regulation of ubiquitin-proteasome system (UPS). The UPS activity is also down-regulated in ALS, leading to accumulation of cytotoxic protein aggregates and ER stress, a hallmark of ALS. In addition, emerging evidence suggests that lipid droplets play a fundamental role in protein aggregation and clearance. Thus, metabolic pathways linked to synthesis and degradation of cholesteryl esters represent potential targets for reduction of ALS disease progression, similar to those reported for AD. Overall, our study shows the spinal cord is more susceptible to lipid alterations occurring during ALS disease progression than the motor cortex. In addition, our results provide evidence for lipids accumulating as cholesteryl esters in the spinal cords of SOD1-G93A 120d rats, which are under apparent astrogliosis. We proposed a model based on the neuroprotective mechanism against ROS that supports lipid droplets accumulation as cholesteryl esters in astrocytes from the spinal cords of ALS rats. Since fatty acid beta-oxidation is likely limited in the CNS, cholesteryl esters might represent a suitable vehicle for lipid transport within and perhaps across the CNS. Thus, cellular processes involving cholesteryl esters metabolism and transport may represent potential targets for treatment of neurodegenerative diseases. The precise analysis of molecular species of lipids as performed in our study may help to further explore the role of oxidative stress regulating lipid metabolism in neurodegenerative diseases and aging. The lipids used as internal standards were 1,2-diheptadecanoyl-sn-glycero-3-phosphocholine (PC 17:0/17:0), 1,2-diheptadecanoyl-sn-glycero-3-phosphoethanolamine (PE 17:0/17:0), 1-heptadecanoyl-2-hydroxy-sn-glycero-3-phosphocoline (LPC 17:0), 1,2-diheptadecanoyl-sn-glycero-3-phosphoserine (PS 17:0/17:0), 1,2-diheptadecanoyl-sn-glycero-3-phospho-(1′-rac-glycerol) (PG 17:0/17:0), 1′, 3′-bis[1,2-dimyristoyl-sn-glycero-3-phospho]-glycerol (CL 14:0 × 4), N-decanoyl-D-erythro-sphingosine (Ceramide d18:1/10:0), 3-O-sulfo-D-galactosyl-ß1-1′-N-heptadecanoyl-D-erythro-sphingosine (Sulfatide d18:1/17:0) and N-heptadecanoyl-D-erytro-sphingosylphosphorylcholine (SM d18:1/17:0), all purchased from Avanti Polar Lipids Inc. (Alabaster, AL. USA). In our analysis, we observed co-elution of internal standards PC (17:0/17:0) and PE (17:0/17:0) with isomers contained in the samples that were composed of 16:0/18:0 fatty acid chain combination. In addition, several other lipid subclasses lacked sufficient amounts or were unavailable (e.g. ubiquinone) in our laboratory to be added as internal standards or were not available for purchase (e.g. 1G-aC, Semino, 1G-AEG). In cases of co-elution and lack of internal standards mentioned above, we used external calibration curves to calculate lipid concentrations. The lipids used for construction of external calibration curves were 1,2-diheptadecanoyl-sn-glycero-3-phosphoinositol (PI 17:0/14:1), free fatty acid (FFA 17:0), (PC 17:0/17:0), (PE 17:0/17:0), d5-1, 3-diheptadecanoyl-glycerol (d5-DAG 17:0 × 2), triacylglycerol (TAG 17:0 × 3), 27-hydroxy-cholesterol and cholesteryl ester (CE 15:0), all obtained from Avanti Polar Lipids Inc (Alabaster, AL. USA). These lipids were all corrected to LPC (17:0) for normalization (see details below). Ammonium formate and ammonium acetate were obtained from Sigma-Aldrich (St Louis, MO, USA) as well as all HPLC grade organic solvents used in this study. Animal study was conducted in accordance with the ethical principles for animal experimentation and conducted according to the guidelines of the National Council for Animal Experimentation Control (Conselho Nacional de Controle de Experimentação Animal – CONCEA, Ministry of Science, Technology, Innovation and Communications, Brazil). Male Sprague Dawley rats overexpressing human SOD1-G93A were obtained from Taconic and bred with wild-type Sprague-Dawley females to establish a colony. Genotyping by PCR to detect exogenous hSOD1 transgene was performed by amplification of ear DNA at 20 days of age. Rats were housed under controlled laboratory conditions, including room temperature, a 12 hours light/12 hours dark cycle with food and water ad libitum. Asymptomatic ALS rats (SOD1-G93A 70d; n = 7) and their wild type controls (WT 70d; n = 7) were sacrificed at 73 ± 4 days of age, while symptomatic ALS rats (SOD1-G93A 120d; n = 13) and their control (WT 120d; n = 15) were sacrificed at 122 ± 6 days of age. The criterion for euthanasia of symptomatic SOD1-G93A rats was loss of 15% of their maximum body weight. Rats were fasted for 6 h and anesthetized by isoflurane inhalation at dose of 4% for induction and 2% for maintenance. Motor cortex and spinal cord were collected and stored at −80 °C until further processing. All animal procedures were approved by University of Sao Paulo - Chemistry Institute’s Animal Care and Use Committee under the protocol number 14/2013. Lipid extraction was performed according to the method established by Yoshida et al.. Motor cortex or spinal cord tissues (200 mg) were homogenized in ice by a tissue grinder in 1 mL of 10 mM phosphate buffer (pH 7.4) containing deferoxamine mesylate 100 μM. Briefly, 100 μL of motor cortex or spinal cord homogenate were mixed with 400 μL of phosphate buffer, 400 μL of ice-cold methanol and 100 μL of internal standards (Supplementary Table 1). Next, 1.5 mL of chloroform/ethyl acetate (4:1) was added to the mixture, which was thoroughly vortexed for 30 s. After centrifugation at 1500 g for 2 min at 4 °C, the lower phase containing the total lipid extract (TLE) was transferred to a new tube and dried under N2 gas. Simultaneously, yeast samples (10 mg) were extracted and used as quality controls for reproducibility analysis. Dried TLE were redissolved in 100 μL of isopropanol and the injection volume was set at 1 µL. Blanks and quality controls were injected every 5 and 10 samples, respectively. TLEs were analyzed by ESI-Q-TOFMS (Triple TOF 6600, Sciex, Concord, US) interfaced with an ultra-high performance liquid chromatography (UHPLC Nexera, Shimadzu, Kyoto, Japan). The samples were loaded into a CORTECS (UPLC C18 column, 1.6 µm, 2.1 mm i.d. × 100 mm) with a flow rate of 0.2 mL min and the oven temperature maintained at 35 °C. For reverse-phase LC, mobile phase A consisted of water/acetonitrile (60:40), while mobile phase B composed of isopropanol/acetonitrile/water (88:10:2). Mobile phases A and B contained ammonium acetate or ammonium formate (at a final concentration of 10 mM) for experiments performed in negative or positive ionization mode, respectively. Lipids were separated by a 20 min linear gradient as follows: from 40 to 100% B over the first 10 min., hold at 100% B from 10–12 min., decreased from 100 to 40% B during 12–13 min., and hold at 40% B from 13–20 min. The MS was operated in both positive and negative ionization modes, and the scan range set at a mass-to-charge ratio of 200–2000 Da. Data for lipid molecular species identification and quantification was obtained by Information Dependent Acquisition (IDA). Data acquisition using Analyst® 1.7.1 was performed with a cycle time period of 1.05 s with 100 ms acquisition time for MS1 scan and 25 ms acquisition time to obtain the top 36 precursor ions. An ion spray voltage of −4.5 kV and 5.5 kV (for negative and positive modes, respectively) and the cone voltage at +/−80 V were set to analysis. Additional parameters included curtain gas set at 25 psi, nebulizer and heater gases at 45 psi and interface heater of 450 °C. The LC-MS/MS data were analyzed with PeakView. Lipid molecular species were manually identified based on their exact masses, specific fragments and/or neutral losses and with the help of an in-house manufactured Excel-based macro. Also, a maximum error of 5 mDa was defined for the attribution of the precursor ion. After identification, the area of lipid species was obtained by MS data using MultiQuant. Each peak integration was carefully inspected for correct peak detection and accurate area determination. For quantification, the area ratio of each lipid was calculated by dividing the peak area of the lipid by the corresponding internal standard or using external calibration (Supplementary Table 1). The concentration of lipid species was calculated by either multiplying the area ratio by the concentration of the corresponding internal standard or by external calibration curves relative to LPC (17:0). The external calibration curves used for FFA, PC, PI, PE, cholesterol, CE, DAG and TAG to determine subclass-specific response factors are presented in Supplementary Tables 2 and 3. The total amount of lipids was expressed in µg/g of tissue. Data are presented as mean ± standard error of mean (SEM) (calculated by summing up individual lipid species within each subclass). Note that standards of some lipid subclasses (GalC, 1G-aC, Semino, 1G-AEG and UbQ) were unavailable in the laboratory or commercially and thus lack absolute concentrations. Thus, their concentrations can be compared among samples, but not with other compounds. Data reproducibility analysis was performed by quality controls in both negative and positive ion mode (Supplementary Tables 4 and 5). The peak area and retention time of selected lipids in quality controls (extracted from a yeast sample) were measured at the beginning, after every 10 samples and at the end of the LC-MS/MS experiments. Tissue preparation for histological analysis was performed as previously described with minor modifications. Animals were deeply anesthetized via an intraperitoneal injection of a mixture of ketamine hydrochloride and xylazine and then rapidly perfused transcardially with 0.9% saline, followed by 4% formaldehyde in phosphate-saline buffer, at 4 °C. Spinal cords were removed and postfixed for several days, and then placed in a solution containing 20% sucrose diluted in buffered 4% formaldehyde. The frozen spinal cords (lombar) were mounted on a freezing microtome and cut into 30 μm coronal sections. The slices were collected in cold cryoprotectant solution (0.05 M sodium phosphate buffer, pH 7.3, 30% ethylene glycol, and 20% glycerol) and stored at −20 °C. The sections were mounted onto adhesive microscope slides and subjected to antigen retrieval by incubation with proteinase K 10 µg/ml at 37 °C for 30 minutes. Following three washes with KPBS, conventional immunofluorescence in fixed tissue was performed as described elsewhere by incubating washed sections in blocking solution (KPBS containing 1% BSA and 0.4% Triton™ X-100) for 30 min. Using the same buffer solution composition, the sections were incubated overnight at 4 °C with primary antibody anti-glial fibrillary acidic protein (GFAP; Millipore #MAB360). After incubation with the primary antibody, the sections were rinsed in KPBS and incubated with Alexa Fluor® -488 anti-mouse secondary antibodies for 120 min. Sections were then rinsed with KPBS, counterstained with DAPI for nuclear labeling, and coverslip with PVA/DABCO mounting media. Epifluorescence photomicrographs (Nikon microscope) were processed to enhance contrast and image quality using GIMP (http://www.gimp.org/) and were assembled using Inkscape (http://inkscape.org). Pictures representing different groups received equivalent image treatment. All statistical analyses were performed with Metaboanalyst (website: www.metaboanalyst.ca). In brief, data were log transformed prior to statistical analyses. All groups were compared by multivariate analysis (sPLS-DA) and one-way ANOVA followed by Tukey’s post-hoc (p < 0.05; FDR-adjusted). We also performed heatmap plots using clusters of lipids that were statistically altered in the different groups (one-way ANOVA p < 0.05; FDR-adjusted) and clusters of samples. For pairwise comparisons of SOD1-G93A 120d group with WT 120d group, we performed multivariate analysis (OPLS-DA) and volcano plots consisting of p-values of t-test (assuming unequal variances; Welch’s test; p < 0.05; FDR-adjusted) and fold change set to >1.5. Graphics were generated using R 3.5.0 base packages (website: www.r-project.org), ggplot2, Metaboanalyst, GraphPad Prism 6 and Excel. Throughout our analysis using Metaboanalyst, we followed the protocols by Xia and others.
|
PMC9352691
|
Comprehensive lipid and lipid-related gene investigations of host immune responses to characterize metabolism-centric biomarkers for pulmonary tuberculosis
|
Despite remarkable success in the prevention and treatment of tuberculosis (TB), it remains one of the most devastating infectious diseases worldwide. Management of TB requires an efficient and timely diagnostic strategy. In this study, we comprehensively characterized the plasma lipidome of TB patients, then selected candidate lipid and lipid-related gene biomarkers using a data-driven, knowledge-based framework. Among 93 lipids that were identified as potential biomarker candidates, ether-linked phosphatidylcholine (PC O–) and phosphatidylcholine (PC) were generally upregulated, while free fatty acids and triglycerides with longer fatty acyl chains were downregulated in the TB group. Lipid-related gene enrichment analysis revealed significantly altered metabolic pathways (e.g., ether lipid, linolenic acid, and cholesterol) and immune response signaling pathways. Based on these potential biomarkers, TB patients could be differentiated from controls in the internal validation (random forest model, area under the curve [AUC] 0.936, 95% confidence interval [CI] 0.865–0.992). PC(O-40:4), PC(O-42:5), PC(36:0), and PC(34:4) were robust biomarkers able to distinguish TB patients from individuals with latent infection and healthy controls, as shown in the external validation. Small changes in expression were identified for 162 significant lipid-related genes in the comparison of TB patients vs. controls; in the random forest model, their utilities were demonstrated by AUCs that ranged from 0.829 to 0.956 in three cohorts. In conclusion, this study introduced a potential framework that can be used to identify and validate metabolism-centric biomarkers.Tuberculosis (TB), a communicable disease caused by Mycobacterium tuberculosis (Mtb), remains a global health crisis. The World Health Organization (WHO) estimated that TB was responsible for 1.5 million deaths and approximately 10 million new patients in 2020. The persistently high incidence and prevalence of TB in part reflect inadequate diagnostic approaches; it has been estimated that only 60% of new cases are detected, especially in countries with a high disease burden and low treatment coverage. A sensitive and easy-to-implement test would provide an important initial improvement in diagnostic accuracy. However, the current standard in the diagnosis of TB is smear microscopy or culture tests, both of which have a low sensitivity, are laborious, and require a specialist laboratory. Molecular tests, such as the Xpert MTB/RIF assay, have been introduced; however, their use is not economically feasible in primary care settings. Moreover, TB tests often rely on sputum and thus have sub-optimal sensitivity, especially for patients with early active TB—such patients cannot consistently provide sputum. To improve TB healthcare quality, the WHO has urged the development of a rapid and sensitive biomarker-based non-sputum test that can be implemented at the clinical site and utilizes accessible samples, such as blood, urine, or breath condensate. Omics-based discovery studies in TB patients—which involved comprehensive profiling of the host transcriptome, metabolome, and proteome—identified several biomarkers for the diagnosis of TB. Moreover, implementing network analysis with multi-omics data could potentially verify and choosing the best among those biomarkers. Transcriptomics is the most matured technology that identifies promising transcript biosignatures for TB diagnosis, treatment monitoring, and outcome prediction. Among the proposed signatures, Sweeney3, a host-response three-gene signature, has met the WHO’s target product profiles for a triage test, whereas lipidomics research applications in TB management remain limited. Consequently, there is a need for further research, particularly into the altered lipidome of TB patients; lipids and lipid-related genes also have potential for use as diagnostic and prognostic biomarkers. Moreover, large-scale lipid profiling using plasma can provide insights into the disease because host lipids constitute a significant nutrition source for Mtb growth and reproduction. Mutual metabolic alterations constitute important aspects of host–pathogen interactions; together with regulatory factors, such alterations are responsible for drug tolerance but can be exploited to design effective host-tailored therapies. Mtb lipid metabolism in host macrophages has a vital role in TB pathogenesis. Lipid droplet (LD) formation, an important event in Mtb lipid metabolism, is a multifaceted process related to Mtb intracellular growth and drug tolerance; it also acts as a host defense mechanism to combat the pathogen. Accordingly, studies that examine biomarkers related to lipid metabolism and immunology are expected to be fruitful. There have been several investigations of the biological fluid lipidome in TB patients, with the goal of identifying biomarkers for TB diagnosis. Chen et al. described changes in lipid levels during TB treatment, and the unbiased lipidomics approach of Shivakoti et al. revealed an association between the host lipidome and treatment failure. These pioneering studies indicate significant differences in lipid profiles of patients with active TB and their counterparts; thus, they highlight potential applications of lipid and lipid-gene biomarkers in diverse clinical scenarios. In the current study, a robust workflow was developed that facilitates the identification and validation of multi-omics metabolism-centric lipid and lipid-gene biomarkers for the diagnosis of active pulmonary TB. The Institutional Review Board of Korea University Guro Hospital reviewed and approved the study (No. 2017GR0012). All procedures were carried out following the Declaration of Helsinki. Written informed consent was obtained from all participants that allowed the blood and clinical data analysis to be used. As mentioned elsewhere, plasma and clinical data were obtained from the Biobank of Korea University Guro Hospital. Patients with malignant diseases, diabetes mellitus, hyperlipidemia, human immunodeficiency virus infection, and chronic liver or renal diseases were excluded. Thus, 35 patients with confirmed pulmonary TB and 37 controls were included in this study. The demographic information of included populations was described in Supplementary Table S1. There were no statistically significant differences between the two groups in terms of age (Wilcoxon rank sum test) or sex (Fisher’s exact test). Three data sets with baseline gene expression profiles of TB patients and the counterparts were selected for the differentially expressed analysis and machine learning (ML)-based classification studies. The data sets are: GSE107991 (21 TB, 21 latent tuberculosis infection (LTBI), and 12 Control), E-MTAB-8290 (54 TB and 127 non-TB, including presumptive symptomatic adults with negative TB diagnosis controls and with or without human immunodeficiency virus infection), and GSE101705 (28 TB and 16 LTBI). The LC–MS grade ammonium formate, formic acid, methyl tertbutyl ether (MTBE), and toluene were purchased Sigma Aldrich (St. Louis, Missouri, USA). LiChroSolv® LC–MS grade solvents including water, methanol, acetonitrile, and isopropanol were purchased from Merck KGaA (Darmstadt, Germany). The SPLASH Lipidomix® Mass Spec Standard was purchased from Avanti Polar Lipids (Alabama, USA). Acquity charged surface hybrid technology (CSH) C18 2.1 × 100 mm, 1.7 μm column and Acquity VanGuard CSH C18 2.1 × 5 mm, 1.7 μm pre-column were purchased from Waters (Milford, MA, USA). Sample preparation and lipid extraction were performed in accordance with previously established methods, with a few modifications. In brief, 55-μL plasma samples were thawed on ice for approximately 30 min; subsequently, 5 μL were removed from each sample and pooled to obtain a quality control (QC) sample. Five microliters of the lipid internal standard mixture were injected into each sample (1:10, v/v) and the sample was briefly vortexed. After the sample had been incubated on ice for 20 min with intermittent vortexing, 300 μL of methanol (− 20 °C) and 1000 μL of MTBE (− 20 °C) were added. The mixture was vortexed vigorously for 10 s, then incubated at 4 °C for 1 h with occasional vortexing. After the addition of 250 μL of water, vigorous vortexing for 20 s, and a 10-min incubation at 4 °C, the sample was centrifuged for 2 min at 4 °C and 14,000 rcf. The two supernatants (lipid fraction), each comprising 500 μL, were collected. One half was used for assessments in positive ion mode, and the other half was used for assessments in negative ion mode. The lipid fraction was completely dried in a vacuum at room temperature and stored at − 80 °C until needed. An Acquity charged surface hybrid technology C18 column (2.1 × 100 mm, 1.7 μm) and Acquity VanGuard charged surface hybrid technology C18 pre-column (2.1 × 5 mm, 1.7 μm) were used for lipid separation with a binary gradient elution as described in Supplementary Table S2. A Shimadzu Nexera LC system (Kyoto, Japan) was utilized for the experiment. Lipid extracts were resuspended in methanol/toluene (9:1, v/v) and kept at 4 °C in an autosampler. The injected volume was ion-mode- and data-acquisition-dependent. From 200 μL of resuspended volume, 1 μL (scan profiling) and 2 μL (information-dependent acquisition and SWATH-based data-independent acquisition) were injected in positive ion mode; 3 μL (scan profiling) and 6 μL (information-dependent acquisition and SWATH) were injected in negative ion mode. The separated lipid ions were analyzed using an X500R QTOF with a Turbo V™ ion source with a TwinSpray probe (SCIEX, MA, USA). For the tandem MS analyses, either 45 eV (spread of 15 eV) or 25 eV (spread of 15 eV) were used. The MS parameters are shown in Supplementary Table S3. Mass calibration was automatically performed after every fifth injection through the instrument’s CDS system, using X500R positive or negative calibration solutions. Raw data (wiff files) were directly input to MS-DIAL (version 4.8) for data processing, alignment, and lipid identification. The parameters were ion mode-dependent, as described in Supplementary Table S4. The aligned data were exported for subsequent use. Post-alignment data processing was performed using MetaboAnalyst 5.0 and features with missing rates ≥ 50% were removed; otherwise, the k-nearest neighbors algorithm was used to impute the missing features. Features with relative standard deviation of ≥ 25% in the pooled QC were also removed. The MS-DIAL inbuilt library and Fiehn’s lab lipidomics library were used for lipid identification. Raw counts of transcripts mapped into genes were summarized using the sum level. The annotated gene-level raw counts were normalized using Trimmed Mean of M-values. The pipeline was implemented using NetworkAnalyst 3.0. Differentially expressed analysis was applied for lipid-related genes in the three transcriptome profiles (i.e., E-MTAB-8290, GSE107991, GSE101705) using two-sided unpaired t-test (rstatix package version 0.7.0, implemented in R 4.1.2). Genes with a false discovery rate (FDR) less than 0.05 were considered as significant. An unsupervised method, principal component analysis (PCA), was employed to explore and visualize the lipidome data. Prior to the analysis, the data were normalized (using the median method), log-transformed, and Pareto scaled. PCs that explained the most sample variance were plotted in a two-dimensional space (MetaboAnalyst 5.0) or three-dimensional space (R package, Plotly version 4.10.0). Heatmap and volcano plots (MetaboAnalyst 5.0) were also used for data visualization. Prior to univariate analysis using an unpaired t-test, the data were normalized using the median method and log-transformed. An FDR of 0.05 was set as the threshold for significant features. Fold-change (FC) thresholds of 1.2, 1.5, and 2 were also tested for biomarker candidate selection. Class discrimination between the lipid profiles of the two groups was achieved using partial least squares-discriminant analysis (PLS-DA). Because the discriminant model has tuning parameters (e.g., the number of components), the optimal model was selected in a tenfold cross-validation process. The variable importance in projection (VIP) score of the PC1 of the optimal model was set at ≥ 1.2 as the threshold of important features used to detect potential biomarker candidates. Statistical analyses were conducted using MetaboAnalyst 5.0 unless stated otherwise. Univariate receiver operating characteristic (ROC) curve analysis was conducted to examine the potential biomarker applications of individual lipids. Random forest and linear support vector machine (SVM) were carried out to investigate the discriminatory capacity of the lipid biomarker candidates. Random forest is an ensemble method that generates many decision trees, then aggregates their outcomes to obtain greater prediction accuracy. This powerful tool uses bagging and random feature selection to build multiple base learners. In SVM, a hyperplane is identified that maximizes the margin from data points. A larger margin leads to greater separation by the hyperplane, thus reducing generalization error. The performances of random forest and SVM are stable, regardless of the domain and data types. For internal validation, the ROC curve-based exploratory analysis was utilized because it can automate important feature identification and performance evaluation. In the external validation, the biomarkers that were overlapped with the quantified lipids in the data of Cho et al. (at the fatty acyl/alkyl sum composition) were used to validate their performance in classifying TB patients from latent infection and controls. All matched biomarkers were used to establish the biomarker models. The analyses were carried out in three different scenarios: TB vs. LTBI + control, TB vs. LTBI, and TB vs. control. The biomarker models using lipid-related genes in the datasets E-MTAB-8290 (54 TB, 127 control/non-TB), GSE107991 (21 TB, 12 controls, 21 LTBI), and GSE101705 (28 TB, 16 LTBI) were trained and validated using the same approach. In particular, the gene expression of matched lipid-related genes in four different data sets were utilized for the ROC-curve-based exploratory analysis. In all analyses, the area under the curve (AUC) and 95% confident interval (CI) of the best models are reported. The analyses were performed in the “Biomarker Analysis” module of MetaboAnalyst 5.0. Normalized expression levels of lipid biomarker candidates were visualized by correlation network analysis in the R package corrr (version 0.4.3). The network shows variables as nodes and their association as edges. The proximity of two nodes is determined by their correlation strength; their locations (or Euclidean coordinates) are found by multidimensional scaling. This method reduces the number of data dimensions to facilitate variable visualization. Biomarker candidate data were submitted to Lipid Ontology (LION) for lipid ontology enrichment analysis via the “LION-PCA heatmap” module. In addition, lipid-gene association networks were analyzed using Lipidsig and lipid-genes were extracted. For visualization, the R package ggplot2 (version 3.3.5) was used. PCA was performed in positive ion mode to explore sample tendencies independent of sample source. The analysis was conducted using a total of 3791 detected lipid features of TB and control samples, with and without QC samples. In the PCA scores plot with QC samples, all QC samples clustered tightly together (Fig. S1A), and the relative standard deviation of the raw total ion chromatogram among QC samples was only 6.5%. These data indicated satisfactory repeatability of the untargeted lipid profiling analysis, which allowed subsequent data analyses and interpretation. In the PCA scores plot with TB and control samples, the three first PCs explained 52.1% of the variance: 23.2%, 21.1%, and 7.8% for PCs 1, 2, and 3, respectively. The relative separation of samples into two separate groups is evident in the three-dimensional PCA plot (Fig. 1A). Heatmap analysis captured relative differences between the two groups at the feature level; differences in lipid features were relatively clear (Fig. 1B). In addition, PCA analysis, which included 762 detected features, were also conducted in negative ion mode. Similar to positive ion mode, the QC samples clustered together (relative standard deviation of total ion chromatogram: 5.6%, Fig. S1B). The three-dimensional PCA plot indicated relative separation of the samples into two groups (Fig. 1C). At the feature level in the heatmap, we could also notice a proportionately contrast between the two groups (Fig. 1D). Taken together, the data exploration analyses in positive and negative ion modes indicated considerable differences between the lipid profiles of TB patients and of controls.Figure 1Plasma lipidome data visualization of Tuberculosis patients (N = 35) and Control (N = 37) group. (a) Principal components analysis 3D score plot of the two group in the positive ion mode. (b) Heatmap of all lipidome features between two group in the positive ion mode. (c) Principal components analysis 3D score plot of the two group in the negative ion mode. (d) Heatmap of all lipidome features between the two group in the negative ion mode. C control group, T Tuberculosis group. Plasma lipidome data visualization of Tuberculosis patients (N = 35) and Control (N = 37) group. (a) Principal components analysis 3D score plot of the two group in the positive ion mode. (b) Heatmap of all lipidome features between two group in the positive ion mode. (c) Principal components analysis 3D score plot of the two group in the negative ion mode. (d) Heatmap of all lipidome features between the two group in the negative ion mode. C control group, T Tuberculosis group. Data exploration indicated considerable differences in lipid metabolic profiles between TB patients and controls, but a more sophisticated statistical approach was needed to identify lipids that could be regarded as biomarker candidates. Supervised investigation was conducted using PLS-DA and the lipid profiles of TB patients and controls. In positive ion mode, a PLS-DA model with five components classified the two groups with appropriate performance metrics (Fig. 2A, accuracy = 0.90, R = 0.92, and Q = 0.58). Similarly, in negative ion mode, a PLS-DA model with five components provided satisfactory classification (Fig. 2B, accuracy = 0.90, R = 0.95, and Q = 0.62) (Fig. S2A,B). The VIP score of the first PC, which explained the most sample variance, was extracted as an additional metric of biomarker candidate potential. A VIP score ≥ 1.2 was determined for 821 (21.66%) and 139 (15.71%) features in positive and negative ion modes, respectively. Finally, the random forest model demonstrated satisfactory performance in distinguishing the two groups. The cross-validated out-of-bag errors were 19.7% and 11.3% for positive and negative ion modes, respectively.Figure 2Partial least squares-discriminant analysis (PLS-DA) score plots of Tuberculosis patients and controls plasma lipidome. (a) PLS-DA 3D score plot of the two group in the positive ion mode. (b) PLS-DA 3D score plot of the two group in the negative ion mode. C control group, T Tuberculosis group. Partial least squares-discriminant analysis (PLS-DA) score plots of Tuberculosis patients and controls plasma lipidome. (a) PLS-DA 3D score plot of the two group in the positive ion mode. (b) PLS-DA 3D score plot of the two group in the negative ion mode. C control group, T Tuberculosis group. Univariate analysis using a t-test was employed to further explore potential biomarker candidates. In positive ion mode, 752 significant features (351 up- and 401 downregulated in TB patients) were found based on an FDR threshold of 0.05. Among these features, 743 (343 up- and 400 downregulated in TB patients), 404 (115 up- and 289 downregulated in TB patients), and 195 (51 up- and 144 downregulated in TB patients) exceeded the FC thresholds of 1.2, 1.5, and 2.0, respectively. In negative ion mode, 175 significant features (94 up- and 81 downregulated in TB patients) were identified with an FDR threshold of 0.05. With the FC thresholds of 1.2, 1.5, and 2.0, 156 (78 up- and 78 downregulated in TB patients), 74 (23 up- and 51 downregulated in TB patients), and 33 (8 up- and 25 downregulated in TB patients) features were selected, respectively. The volcano plots in Supplementary Fig. S3A (positive ion mode) and S3B (negative ion mode) show significant features based on the FC threshold of 1.5 and FDR threshold of 0.05. The intersection of two criteria, a VIP score (PC1, PLS-DA model) of ≥ 1.2 and an FC (t-test) of ≥ 1.5, revealed 89 and 28 potential biomarker candidates in positive and negative ion modes, respectively. Among the selected features, 73 (positive ion mode) and 26 (negative ion mode) were successfully annotated as lipids, thus yielding 93 non-overlapping lipid biomarkers (Table 1).Table 1Statistics information of the potential biomarkers for TB versus control distinguish.IDAnalyteIon modeRegulation in TBVIP scoreFold changeFDRConcentration* (µM)In TBIn control1CAR(20:4)PositiveUp1.4881.6423.16E−02NANA2Cer(d34:1)PositiveUp1.6401.5316.74E−03NANA3DG(40:7)PositiveDown1.6660.6448.85E−030.871.524DG(40:8)PositiveDown2.5910.4829.50E−040.461.185Hex2Cer(d42:2)PositiveUp1.6231.5546.99E−03NANA6LPC(20:3)PositiveUp1.5821.8462.02E−020.090.057LPC(22:4)PositiveUp1.8211.7273.30E−030.090.068LPC(O-18:0)PositiveUp1.6561.5122.75E−030.230.169LPC(O-18:1)PositiveUp1.8221.6121.28E−030.530.3510LPE(O-16:1)PositiveUp2.6021.9618.08E−070.420.2311NAE(16:1)PositiveDown2.4130.5242.92E−03NANA12PC(34:4)PositiveDown1.6660.6038.85E−030.470.8813PC(34:5)PositiveDown1.9060.4968.45E−030.561.1114PC(35:5)PositiveDown1.7780.5171.09E−020.501.0115PC(36:0)PositiveUp2.3191.7271.82E−061.961.1516PC(36:6)PositiveDown1.7270.6602.55E−032.513.8017PC(38:3)PositiveUp1.4271.8153.76E−029.145.3018PC(38:7)PositiveDown1.8650.5951.51E−033.966.2019PC(41:7)PositiveDown1.6000.6512.20E−020.270.4220PC(42:8)PositiveUp1.7061.5903.65E−030.800.4721PC(45:11)PositiveDown1.7840.5242.36E−020.521.0022PC(O-32:1)PositiveDown1.7130.5801.62E−020.350.6823PC(O-34:0)PositiveUp1.9461.6114.04E−041.500.9724PC(O-36:0)PositiveUp2.6602.0645.56E−060.350.1725PC(O-37:5)PositiveDown1.2840.6493.28E−0220.6834.0626PC(O-38:4)PositiveUp1.6791.5555.93E−035.743.9027PC(O-39:5)PositiveUp1.5511.5439.08E−030.250.1528PC(O-40:4)PositiveUp1.5761.5151.14E−020.940.6529PC(O-42:5)PositiveUp1.4001.5242.81E−025.073.5630PC(O-44:5)PositiveUp1.4371.5242.70E−027.955.6331PE(34:1)PositiveUp1.7471.5235.01E−031.771.1732PE(36:1)PositiveUp2.4921.8764.19E−061.250.7633PE(38:4)PositiveUp1.9131.5311.12E−0410.707.5334PE(O-40:5)PositiveUp2.2281.7671.77E−040.960.5735PE(O-40:5)PositiveUp3.3862.7396.02E−090.480.1936PI(38:5)PositiveUp1.8791.8625.35E−03NANA37TG(36:0)PositiveUp2.3827.8666.74E−030.260.0338TG(38:0)PositiveUp2.5977.8772.88E−030.330.0339TG(40:0)PositiveUp2.6769.2773.30E−030.570.0840TG(42:0)PositiveUp2.2755.6801.37E−020.810.1741TG(42:1)PositiveUp2.2896.9392.75E−020.530.1042TG(42:2)PositiveUp2.17013.3374.47E−020.390.0443TG(51:6)PositiveDown1.4830.6202.97E−020.040.0944TG(52:5)PositiveDown1.4690.6553.67E−0226.5141.2645TG(52:6)PositiveDown2.2990.4942.81E−030.902.4846TG(54:7)PositiveDown1.5080.6332.59E−023.746.3147TG(54:7)PositiveDown1.5840.5993.16E−023.856.5648TG(54:7)PositiveDown1.7050.6101.14E−020.090.1749TG(54:8)PositiveDown2.0360.5574.96E−030.861.5950TG(54:8)PositiveDown1.8700.4411.27E−020.240.5851TG(55:7)PositiveDown1.7150.6071.02E−020.290.4852TG(56:8)PositiveDown2.0800.5502.92E−0312.2523.7653TG(56:9)PositiveDown2.4660.4521.46E−030.792.0954TG(56:9)PositiveDown2.3030.5051.51E−031.042.0055TG(56:9)PositiveDown1.5720.6213.64E−020.550.8956TG(57:8)PositiveDown1.6850.5681.81E−020.210.4257TG(57:9)PositiveDown1.9760.4488.85E−030.050.1558TG(58:10)PositiveDown1.5070.6114.77E−020.851.2259TG(58:10)PositiveDown2.1990.5212.92E−031.382.7360TG(58:11)PositiveDown2.2310.3381.26E−020.190.5761TG(58:11)PositiveDown2.2950.4385.01E−030.310.5762TG(58:12)PositiveDown2.3010.3765.85E−030.020.0663TG(58:9)PositiveDown1.6230.6611.60E−024.277.0264TG(58:9)PositiveDown1.3960.6624.93E−023.436.0265TG(60:12)PositiveDown2.7380.2932.24E−030.401.3766TG(60:12)PositiveDown1.7980.3793.12E−020.320.8467TG(60:13)PositiveDown2.4910.3583.99E−030.030.1168TG(60:13)PositiveDown2.2420.2861.52E−020.120.4269TG(62:12)PositiveDown1.9210.4821.37E−020.080.1870TG(62:13)PositiveDown2.1020.3501.86E−020.190.6071TG(62:14)PositiveDown2.3560.3279.57E−030.120.4172TG(62:14)PositiveDown1.8550.3783.31E−020.000.0173TG(64:17)PositiveDown2.5080.1322.05E−020.010.0674FA(14:0)NegativeDown1.5640.6665.14E−03NANA75FA(16:1)NegativeDown2.4080.5312.27E−03NANA76FA(18:1)NegativeDown2.5420.5243.88E−04NANA77FA(18:2)NegativeDown2.8780.4268.41E−05NANA78FA(18:3)NegativeDown2.6160.3841.31E−04NANA79FA(20:1)NegativeDown2.3530.5151.31E−04NANA80FA(20:3)NegativeDown2.8900.3762.30E−07NANA81FA(20:4)NegativeDown1.9880.5571.53E−04NANA82FA(20:5)NegativeDown2.9380.3026.57E−07NANA83FA(22:4)NegativeDown1.9740.5429.61E−04NANA84FA(22:5)NegativeDown3.2590.2641.27E−06NANA85FA(22:6)NegativeDown2.6990.3761.09E−05NANA86LPC(O-18:1)NegativeUp1.7101.5677.46E−050.390.2387LPE(18:1)NegativeUp1.4771.6472.96E−031.100.6688LPE(O-16:1)NegativeUp2.4251.9496.76E−080.440.2189LPE(O-18:1)NegativeUp2.1241.7303.14E−060.470.2690PC(34:4)NegativeDown1.7310.6031.28E−030.100.1591PC(34:5)NegativeDown1.7910.5017.23E−030.030.0692PC(36:6)NegativeDown1.9730.4779.61E−040.190.3593PC(36:6)NegativeDown1.6950.6465.88E−040.580.7994PC(38:7)NegativeDown1.8640.5968.41E−050.901.2895PE(34:1)NegativeUp1.7401.6127.85E−041.080.6396PE(36:1)NegativeUp1.9851.7462.94E−041.951.0797PE(36:3)NegativeUp1.5631.5967.09E−031.550.8798PE(O-38:5)NegativeUp2.6552.1148.57E−100.910.3299PE(O-40:5)NegativeUp3.2452.7502.65E−100.710.24TB tuberculosis, VIP variable importance in projection, FDR false discovery rate, NA no information, CAR acylcarnitine, Cer ceramide, Hex2Cer hexosylceramide, LPC lysophosphatidylcholines, LPC (O-) Ether-linked lysophosphatidylcholines, PC phosphatidylcholine, PC (O-) Ether-linked phosphatidylcholine, LPE lysophosphatidylethanolamines, LPE (O-) Ether-linked lysophosphatidylethanolamines, PE phosphatidylethanolamine, PE (O-) Ether-linked phosphatidylethanolamine, PI phosphatidylinositol, NAE N-acetyl ethanolamine, DG diacylglycerol, TG triacylglycerol, FA free fatty acid.Analyte detected in both positive and negative mode.*Single point quantification by using the peak area ratios with matched lipid class of available internal standards. Statistics information of the potential biomarkers for TB versus control distinguish. TB tuberculosis, VIP variable importance in projection, FDR false discovery rate, NA no information, CAR acylcarnitine, Cer ceramide, Hex2Cer hexosylceramide, LPC lysophosphatidylcholines, LPC (O-) Ether-linked lysophosphatidylcholines, PC phosphatidylcholine, PC (O-) Ether-linked phosphatidylcholine, LPE lysophosphatidylethanolamines, LPE (O-) Ether-linked lysophosphatidylethanolamines, PE phosphatidylethanolamine, PE (O-) Ether-linked phosphatidylethanolamine, PI phosphatidylinositol, NAE N-acetyl ethanolamine, DG diacylglycerol, TG triacylglycerol, FA free fatty acid. Analyte detected in both positive and negative mode. *Single point quantification by using the peak area ratios with matched lipid class of available internal standards. Annotated lipid biomarker candidates were first subjected to univariate biomarker analysis. The ROC curves for those candidates were significantly associated with the TB status (Supplementary Table S5). Among 93 candidate lipid biomarkers, 21 had AUC values < 0.7, whereas 72 were considered promising (AUC ≥ 0.7); of the 72, 13 were considered good (AUC ≥ 0.8) and 2 were considered excellent (AUC > 0.9). The “excellent” lipid biomarkers were two ether-linked phosphatidylethanolamines: PE(O-38:5) and PE(O-40:5). The “good” biomarker candidates were from six lipid sub-classes: two phosphatidylcholine (PC), PC(36:0) and PC(38:7); two ether-linked phosphatidylcholines (PC(O-)), PC(O-36:0) and PC(O-34:0); two ether-linked lysophosphatidylethanolamines (LPE(O-)), LPE(O-16:1) and LPE(O-18:1); two phosphatidylethanolamines (PE), (PE(36:1) and PE(38:4); one PE(O-), PE(O-40:5); and four free fatty acids (FAs), FA(20:3), FA(20:5), FA(22:5), and FA(22:6). Multivariate biomarker analysis using the random forest method revealed that models with 93 variables had the best performance (AUC = 0.921, 95% confidence interval [95% CI] 0.834–0.987) (Fig. 3A). The result of the linear SVM method was approximately similar to the result of the random forest method (Fig. 3B). Correlation analysis showed a significant linear correlation among biomarkers for both TB patients (Fig. 3C) and controls (Fig. 3D), suggesting that a small number of lipids could be used as biomarkers to differentiate TB patients from controls.Figure 3Lipid biomarkers multivariate and correlation analysis. (a) Random Forest predictive model of the lipid biomarkers. (b) Linear Support Vector Machine predictive model of the lipid biomarkers. (c) Correlation of the lipid biomarkers in Tuberculosis group (d) Correlation of the lipid biomarkers Control group. Var variable, AUC area under the curve, CI confidence interval, CAR acylcarnitine, Cer ceramide, Hex2Cer hexosylceramide, LPC lysophosphatidylcholines, LPC (O-) Ether-linked lysophosphatidylcholines, PC phosphatidylcholine, PC (O-) Ether-linked phosphatidylcholine, LPE lysophosphatidylethanolamines, LPE (O-) Ether-linked lysophosphatidylethanolamines, PE phosphatidylethanolamine, PE (O-) Ether-linked phosphatidylethanolamine, PI phosphatidylinositol, NAE N-acetyl ethanolamine, DG diacylglycerol, TG triacylglycerol, FA free fatty acid. Lipid biomarkers multivariate and correlation analysis. (a) Random Forest predictive model of the lipid biomarkers. (b) Linear Support Vector Machine predictive model of the lipid biomarkers. (c) Correlation of the lipid biomarkers in Tuberculosis group (d) Correlation of the lipid biomarkers Control group. Var variable, AUC area under the curve, CI confidence interval, CAR acylcarnitine, Cer ceramide, Hex2Cer hexosylceramide, LPC lysophosphatidylcholines, LPC (O-) Ether-linked lysophosphatidylcholines, PC phosphatidylcholine, PC (O-) Ether-linked phosphatidylcholine, LPE lysophosphatidylethanolamines, LPE (O-) Ether-linked lysophosphatidylethanolamines, PE phosphatidylethanolamine, PE (O-) Ether-linked phosphatidylethanolamine, PI phosphatidylinositol, NAE N-acetyl ethanolamine, DG diacylglycerol, TG triacylglycerol, FA free fatty acid. To rule out the possibility that the internal validation overestimated candidate biomarker performance, external validation was conducted using the dataset from Cho et al. (21 active TB patients, 20 patients with latent infections, 28 controls) but restricted to overlapping lipids (i.e., one LPC, 4 PCs, and 7 PC(O-)s). The validation was first conducted by dividing the samples into active TB vs. non-TB (patients with LTBI and controls) groups. In the univariate ROC analysis, six lipids exhibited an AUC ≥ 0.7. While LPC(20:3) and PC(O-34:0) were unable to differentiate between groups, the AUC of PC(O-40:4) and PC(O-42:5) was 1. Satisfactory performance (AUC 1, 95% CI 1–1) was obtained using the random forest model. Similar results were achieved in the comparison of active TB vs. LTBI or TB vs. control (Table 2). The results of the external validation partially supported the results of correlation network analysis, i.e., only a few lipids could differentiate TB from LTBI or controls.Table 2External validation performance of 12 overlapping lipid biomarkers.AnalyteRegulation in TBUnivariate ROC analysis, AUC (CI)Cho et al.OursCho et al.Our analysisTB vs. non-TBTB vs. LTBITB vs. controllysoPC a C20:3LPC(20:3)UpUp0.581(0.421–0.735)0.671 (0.498–0.823)0.483 (0.327–0.638)PC aa C34:4PC(34:4)DownDown0.899 (0.806–0.961)0.976 (0.914–1.000)0.854 (0.736–0.940)PC aa C36:0PC(36:0)UpUp0.964 (0.919–0.992)0.924 (0.819–0.986)0.997 (0.980–1.000)PC aa C36:6PC(36:6)DownDown0.704 (0.584–0.822)0.857 (0.733–0.962)0.600 (0.436–0.752)PC aa C38:3PC(38:3)DownUp0.734 (0.584–0.864)0.729 (0.562–0.861)0.734 (0.566–0.879)PC ae C32:1PC(O-32:1)DownDown0.630 (0.478–0.759)0.656 (0.481–0.811)0.607 (0.436–0.754)PC ae C34:0PC(O-34:0)DownUp0.497 (0.354–0.647)0.614 (0.431–0.783)0.581 (0.406–0.725)PC ae C36:0PC(O-36:0)DownUp0.660 (0.490–0.807)0.748 (0.562–0.894)0.607 (0.429–0.781)PC ae C38:4PC(O-38:4)UpUp0.699 (0.535–0.846)0.764 (0.581–0.893)0.663 (0.479–0.809)PC ae C40:4PC(O-40:4)UpUp1.000 (1.000–1.000)1.000 (1.000–1.000)1.000 (1.000–1.000)PC ae C42:5PC(O-42:5)UpUp1.000 (1.000–1.000)1.000 (1.000–1.000)1.000 (1.000–1.000)PC ae C44:5PC(O-44:5)UpUp0.479 (0.327–0.651)0.545 (0.364–0.710)0.447 (0.295–0.631)TB tuberculosis, LTBI latent tuberculosis infection, AUC area under the curve, CI confidence interval, lysoPC lysophosphatidylcholines, aa diacyl, ae acyl-alkyl, LPC lysophosphatidylcholines, PC phosphatidylcholine, PC (O-) Ether-linked phosphatidylcholine, LPE lysophosphatidylethanolamines, PE phosphatidylethanolamine, PI phosphatidylinositol, NAE N-acetyl ethanolamine, DG diacylglycerol, TG triacylglycerol, FA free fatty acid. External validation performance of 12 overlapping lipid biomarkers. TB tuberculosis, LTBI latent tuberculosis infection, AUC area under the curve, CI confidence interval, lysoPC lysophosphatidylcholines, aa diacyl, ae acyl-alkyl, LPC lysophosphatidylcholines, PC phosphatidylcholine, PC (O-) Ether-linked phosphatidylcholine, LPE lysophosphatidylethanolamines, PE phosphatidylethanolamine, PI phosphatidylinositol, NAE N-acetyl ethanolamine, DG diacylglycerol, TG triacylglycerol, FA free fatty acid. The LION ontology results indicated that PC(O-), PC, and PE were generally enriched in the TB group, whereas FAs and triacylglycerols (TAGs) with longer acyl chains were downregulated. The TB group also exhibited enrichment of lipids associated with mitochondrion, endoplasmic reticulum, and membrane components; it showed decreased levels of LD-related lipid species (Fig. 4A).Figure 4Lipid ontology enrichment and lipid-gene association network analysis. (a) Lipid ontology (LION) PCA-heatmap of Tuberculosis and Control group. (b) Bubble plot of lipid-gene association pathways. C control group, T Tuberculosis group, PC phosphatidylcholine, TG triacylglycerol, LION Lipid ontology. Lipid ontology enrichment and lipid-gene association network analysis. (a) Lipid ontology (LION) PCA-heatmap of Tuberculosis and Control group. (b) Bubble plot of lipid-gene association pathways. C control group, T Tuberculosis group, PC phosphatidylcholine, TG triacylglycerol, LION Lipid ontology. The reported biomarker candidates belonged to 15 (sub)-classes, of which six contained ≥ 3 lipid species. Those six sub-classes were subjected to lipid-gene analysis to identify TB-associated functional dysregulation and potential gene biomarkers. As shown in Fig. 4B, numerous pathways were altered. Significantly enriched biological processes included the PI3K-Akt, Rap1, calcium, and chemokine signaling pathways. Ether lipid metabolism, fat digestion and absorption, linolenic acid metabolism, and cholesterol metabolism were also altered. The full list of altered pathways and associated genes is provided in Supplementary Table S6. Among dysregulated pathways detected in the lipid-gene analysis (p < 0.01), 162 unique genes were identified. These genes were tested for their ability to differentiate active TB from LTBI or controls in three different data sets. The random forest classifier established from the expression of lipid-related genes was able to distinguish TB from its counterparts in three different TB cohorts, with an AUC ranging from 0.829 (95% CI 0.707–0.931, E-MTAB-8290) to 0.958 (95% CI 0.909–1, GSE101705) (Fig. 5A–D). The linear SVM model showed similar results (Fig. S4A–D). However, most genes exhibited a small FC.Figure 5Tuberculosis (TB) and non-TB classification in three cohort by lipid-genes biomarkers using Random Forest predictive model. (a) Model performance (AUC = 0.919) of TB versus Control classification in GSE107991 dataset. (b) Model performance (AUC = 0.884) of TB versus latent tuberculosis infection (LTBI) classification in GSE107991 dataset. (c) Model performance (AUC = 0.829) of TB versus non-TB classification in E-MTAB-8290 dataset. (d) Model performance (AUC = 0.958) of TB versus Control classification in GSE101705 dataset. Var variable, AUC area under the curve, CI confidence interval, TB Tuberculosis, LTBI Latent tuberculosis infection. Tuberculosis (TB) and non-TB classification in three cohort by lipid-genes biomarkers using Random Forest predictive model. (a) Model performance (AUC = 0.919) of TB versus Control classification in GSE107991 dataset. (b) Model performance (AUC = 0.884) of TB versus latent tuberculosis infection (LTBI) classification in GSE107991 dataset. (c) Model performance (AUC = 0.829) of TB versus non-TB classification in E-MTAB-8290 dataset. (d) Model performance (AUC = 0.958) of TB versus Control classification in GSE101705 dataset. Var variable, AUC area under the curve, CI confidence interval, TB Tuberculosis, LTBI Latent tuberculosis infection. This study demonstrated differences in the lipidomes of TB patients and non-TB controls; it revealed 73 and 26 potential biomarker candidates, identified in positive and negative mode, respectively (six biomarkers were detected in both). In TB patients, the biomarkers had at least a 1.5-fold difference (in either direction). Among the significantly altered lipid sub-classes, ceramide (Cer), LPC, PC(O-), and PE were generally upregulated; certain PCs, diacylglycerols, and FAs were downregulated in TB patients. TAGs with shorter acyl chains were strongly increased in TB patients, while TAGs with longer acyl chains were decreased. The biomarkers mostly belonged to the lipid classes PC, PC(O-), PE, PE(O-), FA, and TAG, suggesting that these lipid classes are important in TB pathophysiology. Among the putative lipid biomarkers, 12 were matched with previously reported lipid profiles that utilized a targeted approach. PC(O-40:4), PC(O-42:5), PC(36:0), and PC(34:4) were prominent biomarker candidates identified by internal and external validation. Besides, our biomarkers showed concordance partially with the top biomarkers reported by Chen et al. and Han et al. in terms of lipid species. PC and PC (O-) were dominant in the top list, which suggests the importance of these lipids in differentiating TB from the controls. However, different biological and technical factors can be attributed to the heterogeneity of the findings among studies, such as cohort characteristics, genetic background, sample treatment and instrumental data acquisition, and utilized statistical methods. Lipid-related genes were associated with various TB pathologically comparable pathways; they also formed a distinctive signature that differentiated active TB from LTBI and other non-TB controls. Although gene expression was only subtly altered, it provided a consistent signature. Among the 162 lipid-related genes, 96 genes were found to be differentially expressed in TB versus non-TB in at least one comparison (Supplementary Table S7); these genes are presumably involved in crucial biological processes that underlie TB pathophysiology. For example, differentially expressed genes related to the biochemical regulations of PC, which is profoundly changed in TB, included CHPT1, LPCAT2, LPCAT4, PLA2G4A, PLA2G4C, PLD2, PLD4, MBOAT2, ADORA2A and ADORA2B. A significant increase in Cer(d34:1) was identified in TB patients. This biomarker has been previously reported to exhibit consistently higher levels in TB patients than in healthy individuals or patients with other respiratory diseases. Shivakoti et al. showed that Cer(d34:1) is also related to TB treatment outcome; patients with the highest Cer levels had an increased risk of treatment failure. The involvement of Cer in host immune responses against Mtb—via immune cell activation, phagocytosis, and other mechanisms—might explain its higher level in TB patients than in controls. Although Mtb relies on different carbon sources at different stages of pathogenesis, host lipids are generally the primary carbon source for Mtb in vivo. The genome of Mtb laboratory strain H37Rv has > 250 genes related to lipid metabolism. The infection of macrophages by Mtb triggers the formation of foamy macrophages through the accumulation of lipid bodies of TAGs and cholesterol esters. Consistent with previous findings, our study showed a decrease in TAGs in host plasma; this may be related to the uptake of host TAGs into foamy macrophages to form LDs, which can serve as nutrient sources. While LD formation may be a host-driven immune response rather than an Mtb-mediated process, the resulting physiological changes in Mtb lead to TAG accumulation, LD formation, growth reduction, decreased metabolic activity, and development of phenotypic drug resistance; these processes are associated with the persistent and non-dividing stages of Mtb. We also found the downregulation of FAs in TB patients; this hinders the formation of longer-chain FAs that form the main components of Mtb cell wall lipids. Fas I/II-induced elongation could partially explain the decrease in plasma FAs. Lysophosphatidylcholine acyltransferase 2 (LPCAT2) induces LD accumulation in cancer patients during the onset of chemoresistance. Our study is presumably the first to report an association of the upregulation of LPCAT2 with metabolic alterations in TB patients vs. non-TB controls, suggesting that LPCAT2 can serve as a biomarker in the diagnosis of TB. There is also emerging evidence concerning the crucial role of metabolism in host–pathogen dynamics, with the transcription factor PPAR (peroxisome proliferator-activated receptor) implicated in LD buildup during inflammation and infectious diseases. Our analysis demonstrated the enrichment of several lipid-related genes associated with PPAR signaling pathways; these genes include PPARA, CD36, FABP4, and ACSL1. Lipid mediators, cytokines, and chemokines may act in a paracrine manner to induce LD formation. We also identified the involvement of lipids and lipid-related genes in chemokine signaling (CDC42, FGR, IKBKB, RAC1), ether lipid metabolism, glycerophospholipid metabolism, sphingolipid metabolism, and phospholipase D signaling pathways. In a mouse model of TB and a study of T cells from TB patients, Mtb was found to inhibit host proinflammatory cytokine production through the PI3K-Akt signaling pathway. In the ontology analysis of lipid genes, the PI3K-Akt signaling pathway exhibited the most significant functional dysregulations in TB patients. Together, these observations support a significant role of lipid metabolism and lipid-related genes in the host immune response. Our study had several limitations. First, its focus was on the discovery and validation of lipid biomarkers; however, there is evidence to support the use of hydrophilic metabolites (e.g., glutamic acid and glutamine) as biomarkers. The combined use of these metabolites and lipids would significantly improve the detection of active TB in clinical settings. Second, other infectious respiratory diseases (e.g., community-acquired pneumonia) were not included in the lipidomics analysis. Nevertheless, some markers were able to reliably distinguish TB and LTBI. Subsequent studies should examine the differential diagnostic performance of those biomarker candidates in other infectious respiratory diseases. Third, quantitative information is available for some biomarker candidates, based on isotopically labeled internal standards at ratios relative to human plasma. However, an accurate quantification strategy (e.g., AdipoAtlas) is needed to facilitate clinical application. This can be readily achieved through targeted analysis of a subset of the most promising biomarkers. Fourth, through our exploratory analysis, we enable the identification of several altered lipids and lipid genes, as well as lipid-related metabolism and immune response pathway in TB patients. Experimental studies on in vitro or animal models are required to substantiate our findings. Finally, a prospective validation cohort study with actual concentrations of lipid biomarkers is required to examine the relevance of the identified biomarkers in TB manifestations. In summary, our study identified and validated lipid-focused biomarkers. Multiple data mining methods with lipidome and lipid-related transcript signatures were used to obtain robust biomarkers and gain new mechanistic insights into TB. Lipid species that belonged to the PC(O-), PCs, TAGs, FAs, and Cer were identified as excellent candidate biomarkers. PC(O-40:4), PC(O-42:5), PC(36:0), and PC(34:4) were externally validated and had a good performance. Additionally, our study revealed systemic and multi-omics levels of biologically relevant processes involved in host responses to Mtb infection. Overall, comprehensive omics analyses employing a data-driven, knowledge-based approach can support metabolism-centric biomarker discovery and validation.
|
PMC11127996
|
Lipidome atlas of the adult human brain
|
Lipids are the most abundant but poorly explored components of the human brain. Here, we present a lipidome map of the human brain comprising 75 regions, including 52 neocortical ones. The lipidome composition varies greatly among the brain regions, affecting 93% of the 419 analyzed lipids. These differences reflect the brain’s structural characteristics, such as myelin content (345 lipids) and cell type composition (353 lipids), but also functional traits: functional connectivity (76 lipids) and information processing hierarchy (60 lipids). Combining lipid composition and mRNA expression data further enhances functional connectivity association. Biochemically, lipids linked with structural and functional brain features display distinct lipid class distribution, unsaturation extent, and prevalence of omega-3 and omega-6 fatty acid residues. We verified our conclusions by parallel analysis of three adult macaque brains, targeted analysis of 216 lipids, mass spectrometry imaging, and lipidome assessment of sorted murine neurons.The human brain indisputably represents one of the most complex biological structures. Current efforts to characterize its detailed organization tend to follow two directions: (i) molecular studies focusing on the brain region macroscale architecture mainly represented by gene expression analysis of individual cells, brain structures, and isolated cell populations; and (ii) macroscale studies of the structural and functional organization of the brain mainly based on brain imaging techniques. Several layers of the human brain organization potentially bridging the two research streams, however, largely escape the current scientific focus. Among these layers, we believe, is the human brain lipidome. Lipids are the main components of brain tissue, comprising 78% of the dry weight of axon myelin sheath and 35-40% of the neuron-rich gray matter. Over the past 150 years, research on the human brain lipidome has provided substantial information on the biochemical composition of selected gray and white matter regions, axon-wrapping myelin, and major brain cell types. These pioneering studies identified a multitude of lipid classes comprising brain tissue, listing phosphatidylcholines and phosphatidylethanolamines representing phospholipids, ceramides representing sphingolipids, and cholesterols representing sterols among its main components. These studies further revealed substantial variation of lipid class content and fatty acid composition among isolated brain structures including, among others, higher levels of sphingolipids, particularly ceramides, cholesterol, and oleic acid in myelin membranes of oligodendrocyte cells compared to other brain components, neutral lipid deposits in astrocytes, and particular lipid class and fatty acid profiles of synaptic membranes. More recently, mass spectrometry-based methods have provided compound-level resolution of lipids in gray and white matter, selected brain regions, and cultured brain cells from humans, macaques, and mice. Building on previous work, these studies have identified additional differences in lipid composition among brain regions, cell types, and cellular projections. As our understanding of the brain’s lipidome composition improves, we are also gaining insight into the functional roles of lipids. The lipid class ratio, the size and the charge of lipid head groups and the number, length and unsaturation extent of their fatty acid residues define geometry, fluidity, and compartmentalization of membrane bilayers, including formation of structured lipid rafts. In addition to these well-established roles, lipids are now recognized to critically contribute to brain energy metabolism, cell type differentiation, neuronal and glial signaling, control of inflammatory response and modulation of protein complexes. Furthermore, alterations in the brain lipidome have been linked to cognitive disorders: autism, schizophrenia, Alzheimer’s disease, alcohol-related brain damage, and others. Rodent studies similarly showed significant aging-related brain lipidome alterations and lipid composition differences among brain regions, brain cell types, and individual brain cell types and their projections. Despite the increasing understanding of the brain lipidome, there are still limitations in our knowledge regarding the connection between lipidome variation and the structural and functional organization of the brain. Previous lipidome studies have focused on specific brain structures comprising less than ten regions in total, limiting our ability to associate lipidome variation with data on brain structural and functional networks. To address this limitation, we conducted both untargeted and targeted lipidome characterization of 75 anatomically and functionally distinct human brain regions. The data we acquired allowed us to examine the distribution of lipid features across the brain and integrate it with the known molecular, anatomical, and functional brain features, providing a foundation for future systematic studies of the human brain’s lipidome. To construct a lipidome map of a neurotypical adult human brain, we investigated the lipidome composition in 75 anatomically and functionally distinct regions dissected from four adult cognitively healthy humans (Fig. 1a, b; Supplementary Table 1; Supplementary Data 1). In addition to humans, we assessed the brain lipidome in 38 of the 75 brain regions of three adult macaques (Fig. 1a; Supplementary Table 1; Supplementary Data 1). Unlike humans, macaques were raised in a standardized environment, followed by rapid and controlled tissue collection, thus providing a robust reference for the human brain lipidome quality evaluation.Fig. 1Lipidome analysis of human and rhesus macaque brain regions.a Experimental scheme displaying numbers of subjects and lipids analyzed by mass spectrometry-based techniques (HRMS and MRM). b Numbers of lipids detected in the human brain by HRMS and MRM in lipid classes defined as corresponding LIPID MAPS subclasses. c, d List of assessed brain regions (c) and their anatomical localization in the human brain (d). Brain images were adapted from ref. . e Visualization of the total lipidome variation in the human and macaque brains using t-SNE based on HRMS and MRM measurements. Each circle represents a brain region, colors according to (d). f, g Correlation of lipid intensity profiles across brain regions between HRMS and MRM (f) and between humans (n = 4 individuals) and macaques (n = 3 animals) (g). Random pairs distributions represent the correlation between lipid intensity profiles of two datasets with permuted region labels. Source data are provided as a Source Data file. a Experimental scheme displaying numbers of subjects and lipids analyzed by mass spectrometry-based techniques (HRMS and MRM). b Numbers of lipids detected in the human brain by HRMS and MRM in lipid classes defined as corresponding LIPID MAPS subclasses. c, d List of assessed brain regions (c) and their anatomical localization in the human brain (d). Brain images were adapted from ref. . e Visualization of the total lipidome variation in the human and macaque brains using t-SNE based on HRMS and MRM measurements. Each circle represents a brain region, colors according to (d). f, g Correlation of lipid intensity profiles across brain regions between HRMS and MRM (f) and between humans (n = 4 individuals) and macaques (n = 3 animals) (g). Random pairs distributions represent the correlation between lipid intensity profiles of two datasets with permuted region labels. Source data are provided as a Source Data file. The 75 evaluated human brain regions comprised 11 gross anatomical structures. The 56 regions represented four groups of neocortical areas: primary sensory or motor cortices (six regions), secondary cortical areas (16 regions), associative cortical areas (21 regions), and limbic cortex (13 regions). The remaining structures included basal ganglia (five regions), thalamus (four regions), hypothalamus, two midbrain regions, four dispersed white matter regions, and three cerebellar gray and white matter regions (Fig. 1a, c; Supplementary Data 1). The 38 analyzed macaque areas covered all 11 anatomical structures (Fig. 1a, c, d; Supplementary Data 1). For all samples, we collected lipid abundance data using two complementary methods: (i) untargeted high-resolution mass-spectrometry (HRMS) and (ii) targeted tandem mass spectrometry using multiple reaction monitoring (MRM). Untargeted analysis of the human samples yielded intensities of 419 lipids annotated using ion fragmentation comprising 21 lipid classes corresponding to LIPID MAPS subclasses (exceptions described in Materials and Methods) (Fig. 1b; Supplementary Fig. 1; Supplementary Data 2). MRM measurements involved 216 lipids representing the same lipids classes, with 169 shared between the techniques (Supplementary Data 2). Macaque brain tissue measurements covered 394 of the 419 HRMS lipids and all 216 MRM lipids (Fig. 1a, b; Supplementary Data 2). To minimize the biological variation that is inevitably present among humans and, to a lesser extent, among macaque samples, we standardized average lipid abundance levels among individuals. This was achieved by dividing the abundance of each lipid in each of the 75 human or 38 macaque regions by its per-individual mean. Consequently, our subsequent analysis is based on these normalized lipid abundance levels. This approach allows us to estimate the relative representation of each lipid among the investigated brain regions. However, it reduces our ability to compare different lipids with each other in terms of their absolute abundance in the brain tissue. Visualization of the human brain lipidome variation based on the normalized abundance of 419 HRMS-assessed or 216 MRM-assessed lipids revealed a reproducible gradient of brain structures—from associative and limbic cortical regions to the central white matter tracts (Fig. 1e, top row). The macaque lipidome measurements yielded the same gradient (Fig. 1e, bottom row). Accordingly, normalized lipid intensity varied significantly among brain regions for 391 of 419 (93%) compounds in humans and 279 of 394 (71%)—in macaques (ANOVA, BH-adjusted p < 0.01). Further, profiles of normalized lipid intensity differences among brain regions correlated positively and significantly between the techniques (HRMS and MRM) and the species (humans and macaques), indicating the robustness and reproducibility of our lipidome measurements (comparison to random pair correlations, Mann–Whitney U test p < 0.0001; Fig. 1f, g; Supplementary Fig. 2,3). Visualization of the relative representation of 21 biochemical lipid classes sorted by their geometry and lipids sorted by the extent of unsaturation in their fatty acid residues revealed distinct differences separating myelin-rich subcortical and white matter regions, as well as more subtle lipid gradients across neocortex and subcortical structures (Fig. 2a, b). For example, consistent with previous findings, cholesterol levels were elevated in the subcortical white matter, while lipids containing polyunsaturated fatty acids (PUFAs), such as docosahexaenoic acid (22:6), were decreased (Fig. 2a, b). In addition to this difference, both cholesterol and PUFA-containing lipids exhibited considerable variation across neocortical regions, with lower relative abundance in the prefrontal regions and elevated levels in the motor, visual, and parietal cortices.Fig. 2Normalized abundance profiles of lipids sorted by class and unsaturation across brain regions.a Heatmaps illustrating the relative abundance of each lipid class in every brain region. The normalized abundance refers to the average signal intensity of each lipid class in a given brain region normalized to its global average intensity, calculated across all 75 regions, and then averaged across individuals (n = 4 individuals). Each row represents a lipid class, each column corresponds to a brain region. The color bars underneath the plots here and in panel b indicate the anatomical assignment of brain regions as shown in Fig. 1d. Lipid classes are grouped based on their geometry, which is illustrated in an insert on the right side of the figure. The gray bars on the right side represent the number of detected lipids in each lipid class (Supplementary Data 2). The brain regions on the right side are colored according to the relative abundance levels of cholesterol. Brain images were adapted from ref. . b Heatmaps representing the normalized abundance levels of lipids containing a defined total number of double bonds in their fatty acid residues (rows) across brain regions (columns), averaged across individuals (n = 4 individuals). The gray bars on the right side represent the number of detected lipids in each fatty acid unsaturation group (Supplementary Data 2). The brain regions on the right side are colored according to the relative abundance levels of lipids with a total of six double bonds in their fatty acid residues. Brain images were adapted from ref. . Source data are provided as a Source Data file. a Heatmaps illustrating the relative abundance of each lipid class in every brain region. The normalized abundance refers to the average signal intensity of each lipid class in a given brain region normalized to its global average intensity, calculated across all 75 regions, and then averaged across individuals (n = 4 individuals). Each row represents a lipid class, each column corresponds to a brain region. The color bars underneath the plots here and in panel b indicate the anatomical assignment of brain regions as shown in Fig. 1d. Lipid classes are grouped based on their geometry, which is illustrated in an insert on the right side of the figure. The gray bars on the right side represent the number of detected lipids in each lipid class (Supplementary Data 2). The brain regions on the right side are colored according to the relative abundance levels of cholesterol. Brain images were adapted from ref. . b Heatmaps representing the normalized abundance levels of lipids containing a defined total number of double bonds in their fatty acid residues (rows) across brain regions (columns), averaged across individuals (n = 4 individuals). The gray bars on the right side represent the number of detected lipids in each fatty acid unsaturation group (Supplementary Data 2). The brain regions on the right side are colored according to the relative abundance levels of lipids with a total of six double bonds in their fatty acid residues. Brain images were adapted from ref. . Source data are provided as a Source Data file. Our analysis of lipid class variation among brain regions indicated that myelin content plays a substantial role in defining regional lipidome composition. This observation aligns with previous work showing a substantial contribution of myelin to the lipidome variation among cell types and brain regions in mice. To identify lipids abundance patterns explainable by myelin representation and search for the ones displaying myelin-independent patterns, we first estimated the relative myelin content of the brain regions using human and macaque structural MRI (sMRI) T1w/T2w image data (Fig. 3a, b; Supplementary Data 1), recognized to provide a reasonable approximation of the myelin content. For both species, myelin content distribution correlated strongly with the first principal component of the lipidome intensity distribution, even though the lipid and sMRI data were acquired from entirely distinct sets of individuals (Pearson’s R = 0.78 and 0.77 for humans and macaques, respectively, p < 0.0001; Fig. 3c, d). Consistently, the intensity profiles of individual lipids correlated positively and significantly with the myelin content for 65% of detected compounds (N = 274; myelin lipids) and negatively for another 17% (N = 71; myelin lipids) (ANOVA, BH-corrected p < 0.01; Fig. 3e, Supplementary Data 2).Fig. 3Lipidome patterns within the human brain.a, b Schematic representation of the human (a) and macaque (b) brains colored according to the relative myelin content determined using sMRI T1w/T2w image data. Brain images were adapted from. c, d Correlation between lipidome variation based on PC1 of HRMS measurements and myelin content derived from sMRI image data in human (n = 210 individuals) (c) and macaque (n = 19 animals) (d) brains. e Schematic drawing of the lipid profile classification protocol and visualization of the resulting average category profiles (red line) and profiles of individual lipids within the category (gray lines). For each category plot: the y-axis shows normalized lipid intensity and the x-axis corresponds to 75 brain regions arranged according to hierarchical clustering outcome (see Fig. 2a). f Volcano plot showing significant lipid intensity differences between human white and gray matter calculated using public data (n = 5 individuals) and colored according to lipid category classification from this study. P-values were calculated using two-sided t test, BH-corrected. g, h Visualization of spatial intensity distributions of 19 lipids detected by HRMS using MALDI imaging of human prefrontal cortical sections (n = 4 individuals) (h) and their intensities’ ratio between white and gray matter areas of the sections (g). Here and in Figs. 4b, d, 5d, h, 6c, j, box plots represent the median, 25th and 75th percentiles; the whiskers extend to the largest and smallest value no further than 1.5 interquartile range. MALDI image panel headings indicate category placement of the lipids determined by HRMS data analysis. i Correspondence percentage of the lipid placement into the five categories between human (n = 4 individuals) and macaque (n = 3 animals) HRMS data. Green color indicates a perfect match, red—a mismatch, and gray—a non-contradictory alternative assignment. Color intensity reflects the number of overlapping lipids. j Distributions of the correlation coefficients for lipid intensity profile comparisons between humans (n = 4 individuals) and macaques (n = 3 animals) for the three main lipid categories. Source data are provided as a Source Data file. a, b Schematic representation of the human (a) and macaque (b) brains colored according to the relative myelin content determined using sMRI T1w/T2w image data. Brain images were adapted from. c, d Correlation between lipidome variation based on PC1 of HRMS measurements and myelin content derived from sMRI image data in human (n = 210 individuals) (c) and macaque (n = 19 animals) (d) brains. e Schematic drawing of the lipid profile classification protocol and visualization of the resulting average category profiles (red line) and profiles of individual lipids within the category (gray lines). For each category plot: the y-axis shows normalized lipid intensity and the x-axis corresponds to 75 brain regions arranged according to hierarchical clustering outcome (see Fig. 2a). f Volcano plot showing significant lipid intensity differences between human white and gray matter calculated using public data (n = 5 individuals) and colored according to lipid category classification from this study. P-values were calculated using two-sided t test, BH-corrected. g, h Visualization of spatial intensity distributions of 19 lipids detected by HRMS using MALDI imaging of human prefrontal cortical sections (n = 4 individuals) (h) and their intensities’ ratio between white and gray matter areas of the sections (g). Here and in Figs. 4b, d, 5d, h, 6c, j, box plots represent the median, 25th and 75th percentiles; the whiskers extend to the largest and smallest value no further than 1.5 interquartile range. MALDI image panel headings indicate category placement of the lipids determined by HRMS data analysis. i Correspondence percentage of the lipid placement into the five categories between human (n = 4 individuals) and macaque (n = 3 animals) HRMS data. Green color indicates a perfect match, red—a mismatch, and gray—a non-contradictory alternative assignment. Color intensity reflects the number of overlapping lipids. j Distributions of the correlation coefficients for lipid intensity profile comparisons between humans (n = 4 individuals) and macaques (n = 3 animals) for the three main lipid categories. Source data are provided as a Source Data file. Comparison of myelin and myelin normalized lipid intensities to published human lipidome data derived from cortical white and gray matter enriched and depleted in myelin verified our classification (Fisher test, p < 0.05; Fig. 3f). Further, direct visualization of the lipids’ intensity distributions in the human prefrontal cortical sections using MALDI imaging similarly confirmed the sMRI-based assignment of myelin lipids in the white matter and myelin lipids—in the gray matter (Fisher test, p < 0.05; Fig. 3g, h; Supplementary Fig. 4). Additionally, gene expression patterns correlating with lipid intensity profiles across 35 brain regions (Pearson correlation, R > 0.5, p < 0.05; Supplementary Fig. 5) displayed enrichment in relevant Gene Ontology (GO) terms: axons and myelin—for genes correlated with myelin lipids, and synaptic functions—for genes correlated with myelin lipids (hypergeometric test, BH-corrected p < 0.05; Supplementary Data 3). Among the remaining 18% of the analyzed lipidome, 11% (N = 46) differed among brain regions in a reproducible manner not explained by the myelin content (unexplained lipid category). Further, 5% of the lipids (N = 20) did not show any intensity differences among brain regions (housekeeping lipids), and 2% (N = 8) varied substantially among individuals (variable lipids) (Fig. 3e; Supplementary Data 2). Parallel analysis of the macaque brain lipidome and sMRI data yielded a consistent lipid assignment into the five pattern-based categories for 84% of the 394 overlapping HRMS-based lipids and 88% for MRM-based ones (Fig. 3i; Supplementary Fig. 6). Further, lipid intensity profiles within the categories correlated positively and significantly between the two species, with myelin lipids showing the highest concordance, followed by myelin and unexplained lipid categories (Fig. 3j; Supplementary Fig. 7). Analysis of lipid chemical properties revealed an overrepresentation of five lipid classes among myelin lipids, including three previously assigned to the central nervous system white matter (hypergeometric test, p < 0.1; Fig. 4a; Supplementary Data 2). The myelin lipids showed an overrepresentation of two phospholipid classes, lysophosphatidylcholines and phosphatidylethanolamines (hypergeometric test, p < 0.01; Fig. 4a; Supplementary Data 2), and housekeeping lipids—excess of free fatty acids (hypergeometric test, p < 0.001; Fig. 4a; Supplementary Data 2). In addition to lipid class content, lipids comprising the pattern-based categories displayed differences in fatty acid chain unsaturation and length. Particularly, myelin lipids tended to have polyunsaturated fatty acid residues in PE and PE-P classes (Mann–Whitney U test, BH-corrected p < 0.05; Fig. 4a), longer residues’ chain length in PE-P class and shorter—in PC class (Mann–Whitney U test, BH-corrected p < 0.1; Fig. 4a). By contrast, housekeeping lipids as a category included unsaturated (Mann–Whitney U test, BH-corrected p < 0.05; Fig. 4a) and short-chain free fatty acids (FA) (Mann–Whitney U test, BH-corrected p < 0.1; Fig. 4a). These differences resulted in variation in the predicted membrane fluidity, with myelin lipids yielding the highest fluidity levels and myelin lipids—the lowest, consistent with reported white matter properties (Mann–Whitney U test, BH-corrected p < 0.001; Fig. 4b). Experimentally, we further verified the predicted differences between categories by direct visualization of molecular ions corresponding to lipid head groups representing 12 lipid classes in human cerebellar sections using time-of-flight secondary ion mass spectrometry (ToF-SIMS), (Fig. 4b–e; Supplementary Fig. 8; Supplementary Table 2).Fig. 4Characterization of lipids within five brain-profile categories.a Distribution of detected lipid compounds among lipid classes within each category (left panel, total number of lipids within each category, from the top: 274, 71, 46, 20, 8) in humans (n = 4 individuals). Сolored bars indicate significantly enriched classes (two-sided hypergeometric test, BH-corrected, p < 0.001; p < 0.01; p < 0.05;p < 0.1). Hatched bars mark lipid classes previously assigned to brain white matter. Distribution of total double bond count (central panel) and carbon chain lengths (right panel) of fatty acid residues within each lipid class in each category. Circles represent lipid compounds. Lipid compounds are represented by circles, with circle colors indicating higher (orange) or lower (blue) number of double bonds or fatty acid residues of particular length in a given class compared to the other clusters (p < 0.05 for number of double bonds and p < 0.1 for carbon chain length). Background color represents geometry groups as indicated in Fig. 2a. b Distribution of predicted membrane fluidity effects in humans (n = 4 individuals) for each lipid in a category (two-sided Mann–Whitney U test, BH-corrected, ***p < 0.001). c, d Visualization of the spatial intensity distributions of five lipid head group ions using SIMS imaging of human cerebellar sections (n = 3 individuals) (c) and their intensities’ ratio between white and gray matter areas of the sections (d). e Difference in prevalence of lipids in myelin and myelin clusters shown as a proportion difference calculated for lipid class sets containing the same head group, as measured by ToF-SIMS in HRMS data (two-sided hypergeometric test, BH-corrected, **p < 0.01). Source data are provided as a Source Data file. a Distribution of detected lipid compounds among lipid classes within each category (left panel, total number of lipids within each category, from the top: 274, 71, 46, 20, 8) in humans (n = 4 individuals). Сolored bars indicate significantly enriched classes (two-sided hypergeometric test, BH-corrected, p < 0.001; p < 0.01; p < 0.05;p < 0.1). Hatched bars mark lipid classes previously assigned to brain white matter. Distribution of total double bond count (central panel) and carbon chain lengths (right panel) of fatty acid residues within each lipid class in each category. Circles represent lipid compounds. Lipid compounds are represented by circles, with circle colors indicating higher (orange) or lower (blue) number of double bonds or fatty acid residues of particular length in a given class compared to the other clusters (p < 0.05 for number of double bonds and p < 0.1 for carbon chain length). Background color represents geometry groups as indicated in Fig. 2a. b Distribution of predicted membrane fluidity effects in humans (n = 4 individuals) for each lipid in a category (two-sided Mann–Whitney U test, BH-corrected, ***p < 0.001). c, d Visualization of the spatial intensity distributions of five lipid head group ions using SIMS imaging of human cerebellar sections (n = 3 individuals) (c) and their intensities’ ratio between white and gray matter areas of the sections (d). e Difference in prevalence of lipids in myelin and myelin clusters shown as a proportion difference calculated for lipid class sets containing the same head group, as measured by ToF-SIMS in HRMS data (two-sided hypergeometric test, BH-corrected, **p < 0.01). Source data are provided as a Source Data file. Analysis of primary cell cultures derived from the mouse brain reported substantial lipidome differences among major neural cell types. To assess the contribution of cell type composition to the lipidome variation in the human brain, we first estimated the relative representation of main neural cell types (OD—oligodendrocytes, MG—microglia, OPC—oligodendrocyte progenitor cells, In—inhibitory neurons, Ex—excitatory neurons, Ast—astrocytes) using the mRNA expression of 15 established marker genes available for 35 of the 75 regions (Fig. 5a). For 353 of the 419 HRMS-assessed lipids (84%), the intensity profile spanning the brain regions correlated positively and significantly with at least one marker gene expression profile (cell-type-associated lipids, Pearson’s R > 0.5, p < 0.01; Supplementary Data 2). The majority (82%) of cell-type-associated myelin lipids correlated best with the oligodendrocyte marker profiles, reflecting the fact that myelin sheaths are made by oligodendrocyte membranes (Fig. 5b, c). By contrast, the intensities of myelin lipids correlated best with the expression profiles of inhibitory and excitatory neuron markers (Fig. 5b, c). Housekeeping and variable lipids were distributed across all cell types, while unexplained lipids were associated significantly with markers of astrocytes, oligodendrocyte progenitor cells, and inhibitory neurons (Fig. 5b, c) (hypergeometric test, BH-corrected p < 0.05).Fig. 5Associations between lipid profiles and particular cell types.a Correlation of 35 brain region mRNA expression profiles of marker genes used in cell type deconvolution analysis in humans (n = 4 individuals here and in b–e). b Clustering of 419 HRMS lipids based on their intensity profiles across brain regions using tSNE colored by category (left) and assigned cell type (right). c Proportion of lipids assigned to cell types within each category and their relative enrichment (one-sided hypergeometric test, BH-corrected, p < 0.0001; p < 0.001; p < 0.05). d Distributions of correlation coefficients for human-mouse (n = 4 humans and n = 3 mice) comparisons based on lipid intensity profiles matched and mismatched between the species. e Lipid class association with main brain cell types. Two lipid classes (PE and HexCer) show significant and specific associations. P-values were calculated using one-sided hypergeometric test, BH-corrected. f Comparison of the relative lipid species proportion of PE and HexCer among four cell types in the human brain (n = 4 individuals) and relative intensities of these two lipid classes in the corresponding cell type lines in the mouse brain (n = 3 animals). Error bars represent the standard deviation. g Scheme of mouse cell sorting experiment. h The intensity ratio between lipids associated with neuronal and non-neuronal cell types in our data in the lipidomes of sorted pyramidal neurons and the rest of sorted cells (n = 2 animals). P-value was calculated using one-sided Wilcoxon test. Source data are provided as a Source Data file. a Correlation of 35 brain region mRNA expression profiles of marker genes used in cell type deconvolution analysis in humans (n = 4 individuals here and in b–e). b Clustering of 419 HRMS lipids based on their intensity profiles across brain regions using tSNE colored by category (left) and assigned cell type (right). c Proportion of lipids assigned to cell types within each category and their relative enrichment (one-sided hypergeometric test, BH-corrected, p < 0.0001; p < 0.001; p < 0.05). d Distributions of correlation coefficients for human-mouse (n = 4 humans and n = 3 mice) comparisons based on lipid intensity profiles matched and mismatched between the species. e Lipid class association with main brain cell types. Two lipid classes (PE and HexCer) show significant and specific associations. P-values were calculated using one-sided hypergeometric test, BH-corrected. f Comparison of the relative lipid species proportion of PE and HexCer among four cell types in the human brain (n = 4 individuals) and relative intensities of these two lipid classes in the corresponding cell type lines in the mouse brain (n = 3 animals). Error bars represent the standard deviation. g Scheme of mouse cell sorting experiment. h The intensity ratio between lipids associated with neuronal and non-neuronal cell types in our data in the lipidomes of sorted pyramidal neurons and the rest of sorted cells (n = 2 animals). P-value was calculated using one-sided Wilcoxon test. Source data are provided as a Source Data file. To verify the assignment of lipids to the main neural cell types, we first compared the human data to the two published mouse datasets, each covering eight of the 75 brain regions. The comparison revealed evident and significant similarities between the lipidome variation profiles of the two species (Fig. 5d; Supplementary Fig. 9; Pearson’s R = 0.66, p = 0.00012). Further, lipid classes showing significant cell-type specificity in the human brain, HexCer and PE, displayed the same cell type assignment in mice (hypergeometric test, BH-corrected p < 0.001; Fig. 5e, f). We next carried out direct experimental verification of the neuronal cell type assignment by the lipidome assessment of fluorescently labeled pyramidal neurons and unlabeled cells isolated using cell sorting from the brain of Thy1-ChR2-YFP transgenic mice (Fig. 5g; Supplementary Fig. 10). The experiment confirmed the enrichment of lipids assigned to neurons in the human brain among lipids comprising sorted mouse neurons compared to sorted unlabeled cells (Wilcoxon test, p = 0.0077; Fig. 5h; Supplementary Fig. 10). Our analysis indicates that the human brain lipidome composition could be associated with the brain’s structural organization features, such as myelin content and cell type composition. We next examined whether the lipidome composition of the brain could be linked to the well-described functional brain architecture features: hierarchical information processing and functional connectivity. The hierarchy of signal processing among brain regions (HR) reflects the information flow within the brain: from the primary to the higher-level associative areas for sensory inputs and in reverse for motor commands. Among the 75 brain regions, 56 could be assigned to one of four levels of the signal processing hierarchy defined according to (Fig. 6a). The lipidome composition varied significantly among the four HR levels: the second principal component of the total lipid intensity variation correlated significantly with HR assignment (Pearson’s R = 0.58, p = 0.0000034; Fig. 6b). The single region deviating substantially from this relationship was the piriform cortex—an entry point of the odor sensation pathway (Fig. 6b). Further, individual lipid intensities correlated with HR well above the chance expectation (HR lipids, N = 60, lm(lipid intensity ~ hierarchy level), BH-corrected p < 0.05; Supplementary Data 2), 26 of them positively (HR lipids) and 34—negatively (HR lipids).Fig. 6Association between brain lipidome and the functional architecture.a Location of 56 brain regions assigned and colored according to their processing hierarchy (HR) within the brain. Brain images were adapted from ref. . b Relationship between PC2 of the total variation of 419 HRMS lipids and HR in the human brain (n = 4 individuals here and in c–k). Circles represent brain regions. Circle marks the piriform cortex. c Significance of the relationship between lipid intensity and HR levels shown as correlation test p-value distribution based on all lipids within a category: negative correlations—left, positive—right. Asterisks indicate the significance of the difference among categories (one-sided Mann–Whitney U test, p < 0.01; p < 0.05). d–g Volcano (top) and distribution (bottom) plots showing properties of lipids significantly correlated with HR: lipid class allocation (d), fatty acid residue unsaturation (e), omega-3/omega-6 prevalence based on docosahexaenoic (DHA) or adrenic (AdA) acid occurrence (f), and cell type assignment (g). P-values were calculated via slope for a linear model, BH-corrected. h Location of 59 brain regions assigned and colored according to the inverted first principal component (–PC1) of their functional connectivity within the brain. Brain images were adapted from ref. . i Relationship between PC2 of the total variation of 419 HRMS lipids and –PC1 of FC. Circles represent brain regions. j Distributions of correlation coefficients between lipid intensity and FC in each lipid category. Asterisks indicate the significance of the difference among categories (one-sided Mann–Whitney U test, p < 0.01; p < 0.05). k Significantly higher correlation with FC for individual lipid classes (one-sided Mann–Whitney U test). Source data are provided as a Source Data file. a Location of 56 brain regions assigned and colored according to their processing hierarchy (HR) within the brain. Brain images were adapted from ref. . b Relationship between PC2 of the total variation of 419 HRMS lipids and HR in the human brain (n = 4 individuals here and in c–k). Circles represent brain regions. Circle marks the piriform cortex. c Significance of the relationship between lipid intensity and HR levels shown as correlation test p-value distribution based on all lipids within a category: negative correlations—left, positive—right. Asterisks indicate the significance of the difference among categories (one-sided Mann–Whitney U test, p < 0.01; p < 0.05). d–g Volcano (top) and distribution (bottom) plots showing properties of lipids significantly correlated with HR: lipid class allocation (d), fatty acid residue unsaturation (e), omega-3/omega-6 prevalence based on docosahexaenoic (DHA) or adrenic (AdA) acid occurrence (f), and cell type assignment (g). P-values were calculated via slope for a linear model, BH-corrected. h Location of 59 brain regions assigned and colored according to the inverted first principal component (–PC1) of their functional connectivity within the brain. Brain images were adapted from ref. . i Relationship between PC2 of the total variation of 419 HRMS lipids and –PC1 of FC. Circles represent brain regions. j Distributions of correlation coefficients between lipid intensity and FC in each lipid category. Asterisks indicate the significance of the difference among categories (one-sided Mann–Whitney U test, p < 0.01; p < 0.05). k Significantly higher correlation with FC for individual lipid classes (one-sided Mann–Whitney U test). Source data are provided as a Source Data file. Lipids positively and negatively correlating with HR (HR and HR lipids) displayed distinct categorical, chemical, and biological properties. HR lipids were overrepresented in the myelin category, enriched in phosphatidylcholines, and contained polyunsaturated fatty acid residues: particularly omega-3 DHA (docosahexaenoic acid) containing six double bonds (Binomial test, BH-corrected p < 0.05; Fig. 6c–g). By contrast, HR lipids were overrepresented in the unexplained category, enriched in sphingomyelins, and preferentially contained saturated or oligo-unsaturated fatty acid residues with up to four double bonds (Binomial test, BH-corrected p < 0.05; Fig. 6c–e; Supplementary Fig. 11). Further, in contrast to the HR group, fatty acid components of HR lipids displayed significant depletion of omega-3 and overrepresentation of omega-6 residues (Binomial test, BH-corrected p < 0.05; Fig. 6f). Similarly, cell type association analysis preferentially assigned HR lipids to astrocytes, OPCs, and microglia, while HR lipids – to oligodendrocytes and inhibitory neurons (Fig. 6g). Functional brain connectivity (FC) data, reflecting the correlated activity of dispersed brain regions at rest measured using functional magnetic resonance imaging (rs-fMRI) and potentially reflecting the topology of the brain’s functional networks, was available for 59 of the 75 regions (Fig. 6h). Тhe first principal component of the FC distance matrix correlated significantly with the second principal component of the normalized lipid intensities matrix, indicating the relationship between the lipidome composition and FC (Pearson’s R = 0.35, p = 0.006; Fig. 6i). Accordingly, for 18% (N = 76) of analyzed lipids, the intensity matrices tended to correlate positively with FC ones (Mantel test, nominal p < 0.005). Markedly, myelin lipids constituting the axonal tracts’ sheath showed a significantly better correlation with FC than lipids in the other four categories, potentially reflecting better physical connectivity of the regions within activity-synchronized networks (Mann–Whitney U test, p = 1.5 × 10; Fig. 6j). Further, three of the 21 lipid classes, sulfatides, hexosylceramides, and diacylglycerols, all three enriched in myelin lipids, demonstrated significantly higher correlation with FC than the bulk of detected lipids (Mann–Whitney U test, BH-corrected p < 0.05; Fig. 6k). In this study, we aimed to bridge the gap between brain structural and functional architecture studies and microscale molecular organization analyses by constructing a systematic map of the human brain lipidome and connecting it to the brain’s structural and functional organization. Previous studies have investigated the composition of the human, macaque, and rodent brain lipidome, revealing general brain lipid class and fatty acid composition, as well as identifying key biochemical features of myelin, synaptosomes, gray and white matter neocortical sections, isolated cells, and cultured brain cell types. These studies have emphasized variations in lipid composition among different brain regions and cell types, with myelin displaying the most distinct lipidome composition. However, despite these advancements, there still exists a gap in our understanding of the specific roles played by lipids in brain structural and functional features, including cellular composition diversity and functional network architecture. In our study, we sought to evaluate the connection between lipid abundance variation in 75 human brain regions and their structural and functional attributes. To this end, we first sorted lipids according to their structural properties using a well-established lipid class nomenclature, revealing distinct differences between myelin-rich and myelin-poor brain regions such as subcortical white matter areas and neocortical gray matter (Fig. 2). Consistent with prior studies, our analysis of lipid class and fatty acid residue unsaturation confirmed an overrepresentation of ceramides, hexosylceramides, sulfohexosylceramides, cholesterol, and lipids containing saturated fatty acid residues or ones with up to two double bonds in myelin-rich regions. Diacylglycerides (DG) and plasmalogens of phosphatidylcholine and phosphatidylethanolamine were also found to be enriched in myelin-rich regions, corroborating previous findings. Conversely, PUFA-containing lipids were concentrated in the neocortex and other myelin-poor brain regions, aligning with previous observations (Fig. 2). Given the substantial influence of myelin content on lipid abundance variations among brain regions, we developed our own lipid classification based on their distribution across the brain and the relationship between this abundance variation and myelin levels determined using structural MRI data. Our results from the human brain, confirmed by a parallel assessment using macaque brain data, indicated that the majority of detected compounds (82%) belonged to the myelin-associated profile categories, which include myelin and myelin lipids. Consistent with our lipid class distribution analysis, the biochemical composition of these categories aligned with published work, with sphingolipids associating positively and phospholipids containing long polyunsaturated fatty acid residues associating negatively with myelin levels (Figs. 3f and 4a). Nonetheless, despite the significant role of myelin in determining brain lipid profiles, 18% of the detected compounds were not associated with myelin levels. Of these lipids, over a quarter did not exhibit measurable variation across the entire brain. This category, termed housekeeping lipids, was significantly enriched in short-chain unesterified fatty acids, aligning with their role as universal building blocks in brain lipid biosynthesis. Additionally, two-thirds of the myelin-independent lipids consistently varied across both neocortical and subcortical regions, a variation that could not be explained by myelin levels and, therefore, termed as unexplained category. To assess the functional significance of lipid level variation among brain regions, we utilized a gene expression-based tissue composition deconvolution procedure to assign lipids to the main brain cell types. This assignment process, which relied on the correlation between lipid intensity and mRNA expression profiles of cell type marker genes across 35 brain regions, proved particularly effective for myelin-dependent categories. In the other three categories, particularly among housekeeping and variable lipids that did not display consistent variation profiles across brain regions, a substantial proportion of lipids remained unassigned to specific cell types (Figs. 5c and 7). Nonetheless, our analysis suggests that the characteristic profiles of the unexplained lipids could be linked to the abundance of astrocytes and oligodendrocyte precursor cells in the brain regions, suggesting an association with these cell types (Fig. 5b, c and Fig. 7). The link between the unexplained lipid category and astrocytes was further suggested by the low conservation of their profiles between the human and macaque brains (Fig. 3i and Fig. 7). This finding aligns with the reported rapid evolution of human astrocytes, detected using single-cell transcriptome assessment. By contrast, the profiles of myelin and myelin lipids were highly conserved between humans and macaques and correlated with oligodendrocyte and microglia markers for the myelin and neuronal markers for the myelin categories, respectively (Fig. 3f, g, Fig. 5d–h).Fig. 7Summary of lipid properties in the five profile-defined categories.Columns, from left to right, show the following information: (1) The number and percentage of all detected lipids occupied by the category. (2) Significantly enriched lipid classes. (3) Bar plots depicting the distribution of the total number of double bonds within lipid fatty acid residues, color-coded based on their number (light gray—one, gray—two, deep gray—three). (4) Density plots showing the total lipid fatty acid chain length distribution, combining information for lipids with different numbers of fatty acid residues (left peak—one, middle peak—two, right peak—three). The counts for each group were normalized to maintain the relative heights of the peaks. (5) Numbers of lipids associated with the main brain cell types within each category. Lipid-cell type assignments were based on a Pearson correlation threshold of 0.5 between lipid intensity and mRNA expression profiles of cell-type marker genes across 35 brain regions. If correlations did not meet this threshold, the lipid was not assigned to a particular cell type (cell type = “none” including also ubiquitous lipids). (6) Conservation of lipid profiles between human (n = 4 individuals) and macaque (n = 3 animals) brains as the percentage of human lipids falling into the same category in macaques. (7) Correlation of lipid profiles with brain information processing hierarchy, with a darker shade of gray indicating lipids with correlation p < 0.05. (8) The strength and significance of the correlation between the lipid intensity matrices and the functional connectivity matrix. Asterisks, when present, indicate the enrichment of a specific lipid group within a given lipid category (hypergeometric test, BH-corrected, *p < 0.001; **p < 0.01; *p < 0.05). Source data are provided as a Source Data file. Columns, from left to right, show the following information: (1) The number and percentage of all detected lipids occupied by the category. (2) Significantly enriched lipid classes. (3) Bar plots depicting the distribution of the total number of double bonds within lipid fatty acid residues, color-coded based on their number (light gray—one, gray—two, deep gray—three). (4) Density plots showing the total lipid fatty acid chain length distribution, combining information for lipids with different numbers of fatty acid residues (left peak—one, middle peak—two, right peak—three). The counts for each group were normalized to maintain the relative heights of the peaks. (5) Numbers of lipids associated with the main brain cell types within each category. Lipid-cell type assignments were based on a Pearson correlation threshold of 0.5 between lipid intensity and mRNA expression profiles of cell-type marker genes across 35 brain regions. If correlations did not meet this threshold, the lipid was not assigned to a particular cell type (cell type = “none” including also ubiquitous lipids). (6) Conservation of lipid profiles between human (n = 4 individuals) and macaque (n = 3 animals) brains as the percentage of human lipids falling into the same category in macaques. (7) Correlation of lipid profiles with brain information processing hierarchy, with a darker shade of gray indicating lipids with correlation p < 0.05. (8) The strength and significance of the correlation between the lipid intensity matrices and the functional connectivity matrix. Asterisks, when present, indicate the enrichment of a specific lipid group within a given lipid category (hypergeometric test, BH-corrected, *p < 0.001; **p < 0.01; *p < 0.05). Source data are provided as a Source Data file. To further evaluate possible connection between lipidome variation among regions and the human brain functionality, we assessed the relationship between the lipid abundance profiles and the functional architecture of information processing networks, specifically the signal processing hierarchy and resting-state synchronicity. Furthermore, lipids that are linked to these functional brain features display distinct biochemical and cell-type-specific properties. For instance, lipids inversely correlated with the signal processing hierarchy tend to contain more DHA residues and are associated with inhibitory neurons. On the other hand, lipids that are positively correlated with the signal-processing hierarchy contain an excess of omega-3 polyunsaturated residues and are associated with astrocytes (Fig. 6f, g). These findings suggest the existence of gradients in neuronal and astrocyte density or subtype composition aligning with processing hierarchy. The existence of such anatomical patterns is supported by studies reporting anteroposterior neuronal density gradient in primates and gradual changes in astrocytic representation between dorsoventral and rostrocaudal cortices in mice, both partially aligning with the signal processing architecture. Moreover, recent spatial transcriptomic analysis of the macaque brain has directly shown the relationship between variations in gene expression and the traditional anatomy and neurophysiology-based information flow hierarchy of the visual signal network. It should be noted that the myelination signal does not determine the relationship between the composition of the lipidome and the signal processing hierarchy. This is evident from the fact that the unexplained lipid category, which is independent of myelin content, shows the strongest association with the brain’s signal-processing architecture (Fig. 6c and Fig. 7). In contrast to the myelin-independent nature of the lipid association with the signal processing hierarchy, our findings indicate that myelin lipids have the strongest association with the brain’s resting-state network architecture. This suggests that there are differential properties of the lipidome related to axon myelination or differential representation of myelinated axon subtypes within the tracks connecting regions within these networks. The relationship between the microstructure of cortical regions, as represented by intracortical myelin content, and functional network connectivity has indeed been demonstrated in humans using ultrahigh-resolution MRI, and aligns with earlier studies conducted in macaques and mice. Our study has several limitations that should be considered. Firstly, our brain lipidome assessment was based on a small sample size of only four individuals. While this number aligns with previous transcriptome brain map studies, it may limit the generalizability of our findings. However, we were able to construct representative average lipid intensity estimates and the stability of these estimates was supported by assessing interindividual variation in humans and reproducibility in macaques, which served as a better-controlled sample group. Additionally, our study design may have been underpowered to detect intensity alterations specific to a single brain region. This could potentially lead to the misassignment of certain lipids to either a housekeeping or variable category. Furthermore, our lipidome analysis did not include potentially informative biochemical lipid classes, such as cardiolipins – mitochondrial lipids known to have brain-specific variants, or gangliosides – glycosylated lipids abundant in the nervous system and implicated in neurodegenerative diseases. In the association analysis of lipids and brain cell types, we were constrained by the indirect assignment of lipid profiles to cell types, which was based on the deconvolution of tissue sample composition using known cell type mRNA markers. Although this method is well established, it does have its limitations. In particular, while this deconvolution procedure is effective at separating major cell types, it struggles with differentiation of cellular subtypes, which often display strongly correlated expression profiles. Moreover, our brain region selection was biased towards the neocortex, neglecting the detailed exploration of subcortical structures, connective axonal tracks, and infracortical white matter. This limits our understanding of the overall lipid composition of the brain. Additionally, our LC-MS analysis lacked information about the spatial distribution of lipids within the examined regions. To partially address this limitation, we utilized mass spectrometry imaging for selected regions. Despite certain limitations, our study underscores the potential of lipid measurements to yield valuable insights into the structural and functional properties of the brain that may not be fully captured by other techniques. More broadly, we illustrate that the brain’s lipidome content can serve as a bridge between the local molecular and cellular composition of individual regions and the large-scale structural and functional characteristics of the organ. This type of analysis could be particularly beneficial for studies examining molecular alterations in psychiatric and cognitive disorders that affect distributed brain structures. Several studies have identified lipids as promising molecular markers of brain disorders. Consequently, our exploration of lipid variation across diverse regions in a healthy brain could provide a crucial baseline for future lipidome studies of brains affected by such disorders. Post-mortem human and macaque brain samples were obtained from the National BioService, St. Petersburg, Russia. Informed consent for the use of human brain tissues for research was obtained from all donors or their next of kin by the tissue provider bank. The protocol was approved by the Skoltech Institutional Review Board. All human subjects were defined as healthy with respect to the sampled brain tissue by medical pathologists (Supplementary Table 1). Sex was determined based on mRNA expression of Y chromosome genes. All human subjects suffered sudden death with no prolonged agony state from causes not related to brain function. The macaques were not subjected to any pharmaceutical or immunologically related treatment anytime during their lifetime and were sacrificed for reasons others than participation in this study. They were part of a placebo group in a non-commercial study. The use and care of the non-human primates were carried out in accordance with Directive 2010/63/EU of the European Parliament and of the Council of 22 September 2010 on the protection of animals used for scientific purposes. All post-mortem brain slices were stored at −80 °C. For sample dissection, The Atlas of the Human Brain and The Rhesus Monkey Brain were used to locate the areas of interest in the human and macaque brains, respectively. Frozen brain slices were atemperated to −20 °C prior to dissection. Tissue pieces of approximately 10–15 mg in weight were cut out from selected brain areas using a metal scalpel. Dissected samples were then collected with tweezers, put into tubes, and immediately stored at −80 °C. All materials used during dissection (scalpels, tweezers, tubes) were sterile and chilled using dry ice before use. For each of the four human individuals, we dissected 75 samples representing anatomically defined brain structures (Supplementary Table 1). Similarly, for each of the three macaques, we dissected 38 samples from brain regions homologous to a subset of 75 human brain areas (Supplementary Table 1). Prior to extraction, 10–15 mg tissue pieces were dissected from the frozen tissue samples, weighed, and transferred to cooled 2 ml round bottom reinforced Precellys tubes (Bertin Technologies). The extraction buffer (methanol:methyl-tert-butyl-ether (1:3 (v/v)) was spiked with 0.5 µg/ml triacylglycerol (TAG 15:0/18:1-d7/15:0, Avanti Polar Lipids, 791648C), diacylglycerol (DAG 15:0/18:1-d7, Avanti Polar Lipids, 791647C), ceramide (Cer d18:1-d7/15:0, Avanti Polar Lipids, 860681), lysophosphatidylcholine (LPC 18:1-d7, Avanti Polar Lipids, 791643C), phosphatidylglycerol (PG 15:0/18:1-d7, Avanti Polar Lipids, 791640C), phosphatidylcholine (PC 15:0/18:1-d7, Avanti Polar Lipids, 791637C), and phosphatidylethanolamine (PE 15:0/18:1-d7, Avanti Lipids, 791638C) as internal standards. The buffer was prepared once for all samples and saturated with nitrogen prior to cooling to −20 °C. After stratified randomization, lipid extraction was performed in batches of 48 samples. After every 23rd sample, an extraction blank sample consisting of an empty tube without a tissue sample was inserted. Following the addition of 0.5 ml of −20 °C cold extraction buffer to each tube, including the extraction blanks, we performed homogenization of the tissue pieces using a Precellys Evolution Tissue Homogenizer (Bertin Technologies). After the addition of 0.5 ml extraction buffer, the homogenates were vortexed and then incubated for 30 min at 4 °C on an orbital shaker, followed by 10 min ultra-sonication in an ice-cooled sonication bath. The homogenate was transferred to 2 ml Eppendorf tubes followed by the addition of 700 μl of water:methanol mix (3:1 (v/v)). After brief vortexing, the homogenates were centrifuged for 5 min at 14.000 g to achieve separation of the hydrophobic and aqueous phases. Next, 540 μl of the upper phase containing hydrophobic compounds (predominantly lipids) was transferred to a 1.5 ml Eppendorf tube. The solvent was then evaporated using a Speed Vac concentrator at room temperature. For quality control (QC) samples, 5 µl of the upper phase containing lipid compounds of each analyzed sample was additionally collected and pooled. Lipid samples were then stored at −80 °C till the mass spectrometry analysis stage. Prior to mass spectrometry analysis, samples derived from one brain were grouped and randomized within the group for the processing order. All samples were run in one sequence without interruption. Batches of approximately 100 samples were prepared simultaneously. To resuspend the dried lipid pellets, 200 µl ice-cold acetonitrile:isopropanol (3:1 (v:v)) was added to each sample. After brief rigorous vortexing, the samples were shaken for 10 min, sonicated in an ice-cooled sonication bath for 10 min, and centrifuged for 5 min at 14.000g. The supernatant was transferred to a fresh 1.7 ml Eppendorf tube. For mass spectrometry analysis, 25 µl of lipid sample was transferred to a 2 ml autosampler vial and diluted 1:15 with 350 µl acetonitrile:isopropanol (3:1 (v:v)). Then 3 µl per diluted sample were separated on a Reversed Phase Bridged Ethyl Hybrid (BEH) C8 column reverse (100 mm × 2.1 mm, containing 1.7 µm diameter particles, Waters) coupled to a Vanguard precolumn with the same dimensions, using a Waters Acquity UPLC system (Waters, Manchester, UK). The mobile phases used for the chromatographic separation included: water containing 10 mM ammonium acetate, 0.1% formic acid (Buffer A), and acetonitrile:isopropanol (7:3 (v:v)) containing 10 mM ammonium acetate, 0.1% formic acid (Buffer B). The gradient separation protocol included the following steps: 1 min 55% Buffer B, 3 min linear gradient from 55% to 80% Buffer B, 8 min linear gradient from 80% Buffer B to 85% Buffer B, and 3 min linear gradient from 85% Buffer B to 100% Buffer B. After 4.5 min washing with 100% Buffer B the column was re-equilibrated with 55 % Buffer B for 4.5 min. The flow rate was set to 400 µl/min. The mass spectra were acquired in positive and negative modes using a heated electrospray ionization source in combination with a Bruker Impact II QTOF (quadrupole-Time-of-Flight) mass spectrometer (Bruker Daltonics, Bremen, Germany). Negative ion mode samples were run after completion of the positive ion mode run using a 2 µl injection of non-diluted lipid samples. The spectra were recorded using full scan mode, covering a mass range from 50−1200 m/z. MS settings in positive polarity were as follows: capillary voltage 4000 V, nebulizer 2 Bar, dry gas 6.0 l/min, dry temperature 180 °C. MS settings for negative polarity were as follows: capillary voltage 4000 V, nebulizer 2 Bar, dry gas 6.0 l/min, dry temperature 200 °C. For quality control (QC), a pooled sample of all the lipid extract was prepared in and injected six times before initiating the runs to condition the column, at least five times after each sub-cohort, and after the completion of the runs. In addition, the QC sample was injected after every 10 regular brain lipid samples to assess instrument stability and analyte reproducibility. Four blank samples consisting of acetonitrile:isopropanol (3:1 (v:v)) were measured at the start of the run. The extraction blank samples were queued at the end of the run preceded by eight injections of pure acetonitrile. After the acquisition, Bruker raw data output (.d files) were automatically calibrated by internal calibration followed by lock mass calibration and converted into mzXML format using a custom DataAnalysis script (Bruker, Version 4.3). Obtained mzXML files were then transferred to XCMS software using the xcms package for R version 3.8.2. Mass spectrometric peaks falsely duplicated during the XCMS peak merging procedure were identified using a 10 ppm mass threshold within 1 s retention time difference. Peaks detected during the first minute of the run (retention time <1 min) and after retention time = 18.3 min were excluded from further analysis. The ‘fillpeaks’ procedure implemented in the xcms package was used for missing value imputation. The lipid intensity values missing after the ‘fillpeaks’ procedure were filled by random sampling from a normal distribution with the mean equaling the median of minimal intensity values of detected lipid peaks and the standard deviation equaling the 16th percentile of this distribution. To annotate mass spectrometry peaks, we considered 23 lipid classes referenced by the literature with respect to the primate brains (PE, PC, PE O, PC O, PG, PS, PA, PI, LPC, LPE, LPC O, LPE O, LPS, LPA, Cer, SM, HexCer, GM3, CE, SulfoHexCer, TAG, DAG, MAG and cholesterol) and searched for the corresponding lipid species in our data. Lipid classes were defined according to corresponding LIPID MAPS subclasses, except for several special cases: 1) all free fatty acids were merged in one class; 2) Cer class also included dihydroceramides and phytoceramides subclasses from LIPID MAPS; 3) HexCer class contained both galactosylceramides and glucosylceramides. Mass spectrometry peak annotation included the following steps. First, we constructed a custom database of mass-to-charge values, varying the chain lengths and the number of double bonds, based on selected 23 lipid classes, to match with the mass spectrometry peaks of our dataset. Next, we matched the mass-to-charge values of all the lipid species of each lipid class with a fixed adduct using a 10 ppm threshold. We kept only those lipid classes for which a visually distinguishable ‘grid’ on the mass-to-charge versus retention time scatterplot could be found, filtering out the mass spectrometry peaks with retention times that did not match the grid-like pattern, additionally using internal standard retention time as an anchor point, when available. This procedure was first performed in each ionization mode with a class-specific adduct of choice. Putative annotations were then verified by cross-validation between two ionization modes (when available for the lipid class) by matching the retention times of the peaks between positive and negative ionization modes. We further validated putative annotations based on molecular ion masses and retention times using fragment ions information obtained in a tandem mass-spectrometry experiment. The following statistical analysis included lipids with confirmed lipid class annotation. Fragment ions data also yielded more detailed information about the fatty acid composition of the lipids. Fragmentation was performed on pooled samples, representing mixed aliquots of brain regions including both white and gray matter components. The brain regions with the highest average intensity of a given target lipid class were chosen for this analysis. For the fragment ion analysis, conditions of chromatographic separation were maintained the same as described in the section “MS sample preparation and measurements” and mass spectra acquisition was performed on a hybrid QExactive instrument in data-dependent (DDA) mode. Spectra were recorded separately in positive and negative ionization modes. In each mode, resulting spectra were obtained by averaging three fragmentation spectra at different collision energies. Mass spectrometer was equipped with heated electrospray ionization (H-ESI) and tune parameters were set as follows: capillary temperature: 320 °C; aux gas heater temperature: 350 ° С; capillary voltage: 4.5 kV (−3.5 kV); S-lens RF level: 60; sheath gas flow rate (N2): 45 arbitrary units (a.u.); auxiliary gas flow rate (N2): 20 a.u., sweep gas flow rate (N2): 4 a.u. Probe parameters were identical for both polarities. The operating parameters of the mass spectrometer for full scan mode were set as listed: resolution: 70000 at m/z 200, automatic gain control (AGC target): 5e5, maximum injection time (IT): 50 ms, scan range: 200 to 2000 Da. For DDA mode: resolution 17500 at m/z 200, AGC: 2e4, IT: 100 ms, mass isolation window: 1.2 Da, retention time window width: expected time ± 1 minute, stepped normalized collision energy: 15, 25, 30%, dynamic exclusion 12 s, inclusion: on, customize tolerances: 10 ppm. The spectra were recorded in the profile mode. The lipid formulas proposed in this work are based on the match of the precursor and fragment ion masses with the theoretical ones with an accuracy of less than 10 and 15 ppm, respectively, and expected chromatographic behavior (increasing retention time for series of homologs and decreasing with the addition of double bonds). Acquired MS2 spectra were manually curated, and lipid class assignments were performed based on their distinctive fragmentation behavior. Lipids were attributed to phospholipids by the presence of fragment (or neutral loss) signals related to the polar head group, to glycerolipids – by neutral loss of ammonia for the precursor ion, and to sphingolipids – by characteristic ions of the sphingoid base. The fatty acid composition for phospholipids was interpreted by the presence of deprotonated fatty acid ions in MS2 spectra, for glycerolipids by the neutral loss of fatty acids as ammonium adducts, and for sphingolipids by the difference in the masses of precursor ions and ions of sphingoid bases. Our lipid structure elucidation refers to “FA acyl/alkyl level” according to the proposed lipid nomenclature or to “putatively characterized compound classes (level 2)” according to the metabolomics standard initiative. It implies that sn-attachments of FA, positions of double bounds, and stereochemistry are not declared. The retention behavior of seven lipid classes was confirmed by class-specific isotopically labeled standard (see section “Lipid extraction”). Since ether-linked lipids (plasmanyl- and plasmenyl-phospholipids) cannot be distinguished by class-specific fragments, we utilized their chromatographic behavior for annotation. As we observed sufficient separation between closely eluted isomeric species differing in double bond position (e.g., PC-P 36:2 and PC-O 36:3), both species were reported. The elution order was validated using the authentic standard of PC-P 18:0/18:1(9Z) (Avanti Polar Lipids, 852467 C). Lipids were attributed to phospholipids by the presence of fragment or neutral loss (NL) signals related to the polar head group*, to glycerolipids by neutral loss of ammonia for the precursor ion**, and to sphingolipids by characteristic ions of the sphingoid base***. The fatty acid composition for phospholipids was interpreted by the presence of deprotonated fatty acid ions in MS2 spectra, for glycerolipids by the neutral loss of fatty acids as ammonium adducts, and for sphingolipids by the difference in the masses of precursor ions and ions of sphingoid bases. Free fatty acids were annotated based on the exact mass of deprotonated molecules in MS1 scanning mode. * positive polarity: 184.0739 (C5H15NO4P + ) – LPC, PC, PC-O, PC-P; 18.0106 (NL H2O) – LPC; 141.0191 (NL C2H8NO4P) – LPE, PE, PE-P, PE-O; loss of the glycerol moiety as a neutral olefin (unique for each plasmalogen sn-1 group) - PE-P; 185.0089 (NL C3H8NO6P) – PS; 172.0137 (NL C3H9O6P), 190.0475 (C3H13NO6P + ) – PG; negative polarity: 168.0458 (C4H11O4NP-), 224.0688 (C7H15O5NP-), 60.0211(NL C2H4O2), 74.03678 (NL C3H6O2) - LPC, PC, PC-O, PC-P; 140.0113 (C2H7O4NP-), 196.0375 (C5H11O5NP-) - LPE, PE, PE-P, PE-O; 241.0113 (C6H10O9P-), 259.0226 (C6H12O10P-), 223.0008 (C6H8O8P-)– PI; 87.0320 (NL C3H5NO2) – PS; ** positive polarity: 17.026549 (NL NH3) - TAG, 35.0371 (NL NH3 + H2O), 18.0106 (NL H2O) – DAG. Fatty acid composition was established by NL of a neutral carboxylic acid and neutral ammonia together. *** positive polarity: 184.0739 (C5H15NO4P+) – SM; 180.0634 (NL C6H12O6) – HexCer; the characteristic product ions of the sphingoid bases for Cer, HexCer and sulfotides: 250.2534 (C17H32N+- d17:1), 264.2691 (C18H34N+- d18:1), 262.2534 (C18H32N+- d18:2), 292.3004 (C20H38N+- d20:1), 268.3004 (C18H36N+- m18:0), 266.2847 (C18H34N+- m18:1), 296.3317 (C20H38N+- m20:0), 294.3160 (C20H36N+- m20:1), 268.2640 (C17H34NO+), 250.2534 (C17H32N+- t17:0), 282.2796 (C18H36NO+), 264.2691 (C18H34N+- t18:0). Due to in-source water loss from protonated Cer and HexCer molecules, monitoring of [M+nxH2O]+ ions and [M-nxH2O]+ ms1-ions along with [M + H]+ ions was performed to distinguish between sphingolipid bases with different number for hydroxyl groups. The following fragment ions were used to establish sphingoid base composition of Cer, HexCer, sulfotides: [base+H-H2O]+ for mCer, [base+H-2xH2O]+ for dCer, [base+H-2xH2O]+ and [base+H-3xH2O]+ for tCer. negative polarity: 96.960 (HSO4-) – sulfatides. Other lipids: positive polarity: 85.028 (C4H5O2+) - CAR, 369.352 (C27H45+) – Cholesterol, CE. For the targeted analysis, we selected 341 transitions from the total list of signals obtained in an untargeted experiment. Depending on the lipid class identity, we prepared two sample dilutions. For the measurement of SM, PE, PE-O, PE-P, PC, PC-O, PC-P, LPC, and LPC transitions, samples were diluted 1:100 with acetonitrile:isopropanol (3:1 (v:v)). For the measurement of PG, PI, PS, DAG, Cer, HexCer, SHexCer, and CE transitions, samples were diluted 1:10 with acetonitrile:isopropanol (3:1 (v:v)). 2 µl per diluted sample were separated on an EclipsePlus C18 RRHD column (2.1 × 50 mm, 1.8μm, Agilent Technologies), using an Agilent 1290 Infinity Binary Pump. The mobile phases used for the chromatographic separation were acetonitrile:water (4:6 (v:v)) containing 10 mM ammonium formate (Buffer A) and acetonitrile:isopropanol (1:9 (v:v)) containing 10 mM ammonium formate (Buffer B). The gradient separation was: 2 min 60 % B, 7 min 100 % B, 9 min 100 % B, 9.01 min 20 % B, 10.80 min 20 % B. The flow rate was set to 400 µl/min. MRM were acquired in positive and negative modes using a heated electrospray ionization source with an Agilent jet stream in combination with an Agilent 6495 Triple Quadrupole mass spectrometer (Agilent Technologies, USA). All samples were measured in one uninterrupted run including positive ion mode for all transitions and negative ion mode for PI transitions. MS settings: gas temperature 120 °C, gas flow 11 l/min, nebulizer 40 psi, SheathGasHeater 300, SheathGasFlow 10, capillary voltage 3500 V for positive and 3000 V for negative polarity, VCharging 500 for positive and 1500 for negative polarity. QC dilution series (DiQC) were run at the beginning (0.0625, 0.125, 0.25, 0.5, 1) and at the end of the queue (1, 0.5, 0.25, 0.125, 0.0625). Pooled QC sample was run after every 10th brain sample, followed by a blank sample containing only solvent; every second blank was followed by an extraction blank; every third QC was preceded by the long-term reference sample (LTR, extract of the plasma lipids). All transitions were integrated with MassHunter Quantitative Analysis software version B.08. Peaks with intensity levels in blanks exceeding 0.1 of the intensity levels in real samples were excluded; peaks with the coefficient of variation in QC samples above 25% were excluded; peaks with the correlation coefficient between the DiQC series below 0.9 were excluded. Out of 341 transitions, 169 passed all filters and were considered for subsequent analysis. To account for extraction and technical noise, we applied filtering procedures to the resulting lipid compound target list. First, a blank samples filter was applied: only the features with mean intensity in samples being at least two times higher than in blanks were selected for further analysis. Second, a variance filter was applied: for each peak coefficient of variation was calculated across QC samples in each batch, and only peaks with median CoV lower than 0.25 were selected. Lipid intensities, defined as areas under mass spectrometric peaks, were log10-transformed, and all resulting lipid intensities were normalized on the median value of standards in a sample and the wet weight of the sample. For normalization wet weights and standard intensities were log10-transformed, then the difference of sample values from the mean of a parameter was subtracted. Resulting lipid intensity values were further normalized using the mean intensity of this lipid calculated within each individual’s brain to adjust for inter-individual variability. RNA sequencing was performed for 35 human brain regions representing a subset of 75 brain regions used in lipid analysis in the same four human individuals. Data from 33 brain regions were previously published, while data for an additional two brain regions generated in the same experiment was not previously released. The experiment was performed as described in ref. . Briefly, total RNA was isolated using Direct-zol-96 RNA (Zymo Research) from 10 mg of the frozen tissue per sample. Sequencing libraries were prepared with NEBNext Ultra II RNA Library Prep Kit (New England Biolabs) according to the manufacturer’s instructions. Libraries were then sequenced on the Illumina HiSeq 4000 system using the 150-bp paired-end sequencing protocol. Quality assessment of raw data with the fastQC revealed that many reads have Illumina universal adapters at the end. To trim low-quality bases and remove adapter sequences we used the trimmomatic tool with the following parameters: “PE -phred33 ILLUMINACLIP:all.fa:2:30:10:2:true SLIDINGWINDOW:4:15 LEADING:3 TRAILING:3 MINLEN:20”, we used the union of all adapter sequences provided with trimmomatic and, additionally, we included the sequence of Illumina universal adapter found by the fastQC (AGATCGGAAGAG) for palindrome clipping (see trimmomatic manual). Then reads were mapped to corresponding genomes using hisat2 with the following parameters: “--no-softclip --max-intronlen 1000000 -k 20”. Gene expression values (Transcripts Per Kilobase Million, TPM) were estimated using the stringtie with the following parameters: “-e -G reference.gtf -B out -A out.tab”. Genome sequences (GRCh38, Mmul_8.0.1, panpan1.1 and Pan_tro_3.0), gene annotations used in the analysis, and a table of orthologous genes for all four species were obtained from Ensembl v91. Acquired sequence read counts were next transformed to TPMs, and only genes with TPM > 1 were selected for further analysis. All TPM values were log10-transformed and normalized using the median value of each sample. Only protein-coding genes were used in the analysis. As in the case of lipids, the expression values of each gene were normalized by the mean expression level of this gene within each brain. Prior to hierarchical clustering, we calculated the mean lipid intensity profiles across all individuals spanning all analyzed regions. We used Euclidean distance as a dissimilarity measure and a “complete linkage” method for hierarchical clustering implemented in R hclust() function. We used the tSNE method for lipidome-based sample similarity visualization. We used ANOVA method to select lipids with significantly different intensity levels in different regions with the model: lipid ~ region. To compare spatial distances between regions to lipid intensity and gene expression distances, Euclidean distance matrices were calculated for each dataset. Next, Pearson’s R was calculated for each region’s vector of distances and the corresponding vectors of distances in two other datasets: lipidome and transcriptome. For this analysis, we focused on 19 neocortical regions contained in both lipidome and transcriptome datasets. The anatomical positions of dissected samples were aligned to the standardized MNI coordinate space of human and macaque brains based on sample preparation notes and images. We used preprocessed sMRI data recorded from alive individuals from Human Connectome Project 1200 release from. The myelin content of each region was evaluated as the mean intensity of T1w/T2w signal in a cube 9 × 9 mm centered around the coordinate of the center for each sample. Similar data from macaque was taken from and preprocessed in the same way. For the functional connectivity analysis, data extraction regions of interest were matched to regions from ref. , covering 59 of the 75 analyzed regions. In the case of multiple matching between the regions, the mean value of multiple matched regions was taken. The categorical classification of lipids based on their profiles across 75 regions was performed stepwise. First, the ANOVA method was applied to select lipids with significantly different intensity levels among brain regions (BH-corrected p-value < 0.01). Second, a linear model was built for each ANOVA-selected lipid comparing their intensity with the calculated myelin content of the regions: lipids with significant positive correlation (BH-corrected p-value for linear model <0.05) were classified as myelin+ lipids; lipids with significant negative correlation (BH-corrected p-value for linear model <0.05) were classified as myelin- lipids; the remaining lipids were marked as unexplained. Finally, lipids showing no significant intensity differences among brain regions were divided into two groups based on their within-region variability. For this purpose, the Gaussian component of the within-regions variability was extracted, and the threshold was set at μ + σ. Lipids with within-region intensity variability falling below the threshold (inside the Gaussian bell curve) were classified as housekeeping, and the remaining lipids were classified as variable. Lipid class, number of double bonds, and chain length enrichment of lipid categories were evaluated using a hypergeometric test, implemented in R as phyper() function. The hypergeometric test is employed to assess the statistical significance of drawing a specific number of successes from a population with a known quantity of successes. For instance, in the enrichment analysis of lipid classes, all measured lipids serve as the population. This population contains a specific number of lipids in a particular class, representing the possible number of successes. The lipids within a specific profile category represent the lipids we draw. Gene Ontology (GO) enrichment analysis was conducted for gene sets correlated with the mean lipid cluster profile drawn across 69 regions (based on Allen Brain data) with a coefficient greater than 0.5. For each category of GO terms (cellular compartment, biological processes, and molecular function) the enrichment analysis was conducted using the function enrichGO() from ClusterProfiler package for BF (biological process), CC (cellular compartment), and MF (molecular function) subontologies. All genes detected in the experiment were used as background. Simplify method was applied to selected categories to remove highly intersecting terms (threshold = 0.5), and a significance threshold for BH-adjusted enrichment p-values was set to 0.001. To analyze the macaque brain lipidome dataset, we applied the lipid annotation procedure used for the human samples, including LC-MS/MS validation, without modifications. All lipids annotated in macaque data and matching the list of annotated human lipid compounds were selected for further analysis. Lipid clustering and classification procedures were identical to the ones used for the human data but relied on macaque sMRI data calculated for the 38 analyzed regions. Differential stability scores were calculated, as described in ref. and cross-correlation coefficients between individuals were averaged for each lipid. To evaluate the significance of the correlation between corresponding lipids in the human and macaque datasets, for each lipid we permuted sample labels and compared Pearson’s R between two datasets for true and permuted labels by paired Wilcoxon test. To test the stability of the lipid classification procedure, all cluster assignments were classified either as “good” (the same cluster assignment in human and macaque datasets), “neutral” (assignment to a non-matching cluster with no significant difference between regions compared to the matching one in at least one dataset) and “bad” (assignment in different clusters with significant difference between regions in human and macaque datasets). To find an association between lipid intensities distribution across regions and the cellular composition of the brain we used mRNA data derived from 35 regions, representing a subset of 75 regions analyzed for their lipid content. The same tissue samples were used both for mRNA and lipid measurements. We used a manually selected from the literature well-characterized set of cell type markers corresponding to the six main brain cell types: oligodendrocytes (OD), microglia (MG), astrocytes (Ast), oligodendrocyte progenitor cells (OPCs), inhibitory (In) and excitatory (Ex) neurons (Fig. 3f). Lipid assignment to cell types was based on the correlation between lipid intensity profiles constructed across 35 brain regions and mRNA expression profiles of cell type marker genes constructed across the same 35 regions. A lipid was assigned to a cell type based on the highest Pearson correlation with expression marker profiles and exceeding the correlation coefficient threshold = 0.5. If no correlations to the expression profiles of any cell type markers passed the threshold, the corresponding lipid was marked as non-assigned to a cell type (cell type = “none”). Cell type markers enrichment analysis of lipid categories was carried out using the hypergeometric test. The test was conducted for each cell type and lipid category using all lipids as a background comparison population. Comparison of lipid intensities before and after the cell sorting experiment was carried out using a list of top-20 lipid cell type markers showing the highest correlation with expression markers of a given cell type (or less, if a total number of lipid markers was included less than 20 lipids for a given cell type). In (inhibitory neuron) and Ex (excitatory neuron) markers were merged together in a group of “neuronal” markers, while OD, OPC, Ast, and MG lipid markers were merged as a “non-neuronal” group. Two measures of a membrane fluidity computation were used for each lipid: mean number of double bonds per carbon chain and mean chain length. For this purpose, the total number of double bonds and the total number of carbon atoms were divided by the number of carbon chain residues typical for the class. For this type of analysis, cholesterol was not considered. Each value was normalized to range between 0 and 1 by subtracting the minimal value (among all lipids) and dividing by the difference between the maximal and minimal value (among all lipids). Since the membrane fluidity gets higher with an increasing number of double bonds and gets lower with increasing carbon chain length, the normalized chain length was subtracted from 1 and averaged with the normalized number of double bonds. The resulting value reflected the relative normalized influence on membrane fluidity provided by the given lipid. For the lipidome analysis of sorted cell populations, we used Tg(Thy1-COP4/EYFP)9Gfng (Thy1-ChR2-YFP) transgenic mice (n = 2 animals). Expression of YFP in this line was reported in layer 5 cortical neurons, CA1 and CA3 pyramidal neurons of the hippocampus, cerebellar mossy fibers, neurons in the thalamus, midbrain, and brainstem, and the olfactory bulb mitral cells. Adult female mice were anesthetized and decapitated following the standard, ethically acceptable procedure. The protocol was approved by the Skoltech Institutional Review Board in accordance with the guidelines on the ethical use of animals. All possible efforts were made to minimize animal suffering, and to reduce the number of animals used per condition by calculating the necessary sample size before performing the experiments. Mice were housed in standard breeding cages at constant temperature (22 ± 1 °C) and relative humidity (50%), with a 12:12 h light:dark cycle. Intact brains were isolated from the sacrificed animals within 3 min postmortem, minced by two cold razor blades on ice cold glass plate and placed in an ice-cold solution of zinc fixative (0.1 M Tris-HCl, pH = 6.5, 0.5% ZnCl2, 0.5% zinc acetate, 0.05% CaCl2, final pH = 6.3) in at least 10 × volume at 4 °C for 2 h. Tissues were washed 3 times in PBS (20 min/wash). Fixed and washed tissue was dissociated by Medimax machine with Medicons-P disposable disaggregator with about 50-100 µm separator screen for 10 s in 1 ml of PBS. The suspension was filtered through SmartStrainers 70 µm filters. Filtered cells were spun at 250 g for 3 min in a swinging bucket centrifuge at 4 °C. Supernatant was carefully removed, the cells were resuspended in 200 μL of PBS, and stored at 4 °C until sorting. FACS experiments were performed using a FACSAria SORP instrument (BD Biosciences). The samples were additionally stained with Hoechst33342 to distinguish the nuclei-containing cell bodies. The excitation and emission wavelengths for YFP and Hoechst33342 detection were as follows: Ex. 488 nm, Em. 505LP + 525/20 nm BP and Ex 407 nm, Em. 450/50 nm BP. Gates were set on two populations of interest: non-neuronal nuclei-containing parts (Hoechst33342-positive, YFP-negative) and neuronal YFP-containing parts without nuclei (Hoechst33342-negative, YFP-positive). Sorting was conducted in Purity mode, using an 85 μm nozzle with the corresponding pressure settings. Sorted specimens were frozen at dry ice immediately after the experiment and kept at −80 °C for further lipid analysis. To compare our human HRMS lipid intensity dataset to published mouse brain lipidome atlas data and murine cell type-specific primary culture lipid data, we downloaded the original published raw intensities and preprocessed them using our human HRMS data processing pipeline with no modifications. We then calculated Pearson’s correlation coefficients based on the intensity measurement vectors of lipid compounds matched by annotation between the datasets in the eight regions present both in human and mouse data (both murine datasets had eight regions in common with our human data). Median diagonal and off-diagonal values in matrices of Pearson’s R between two datasets were calculated in 100 region label permutations to test the non-random level of correlation between the two datasets. The resulting distributions were compared using paired Wilcoxon test. For each cell type, raw intensities equal to zero were replaced by random noise values, as described in the section Data extraction and preprocessing of the Methods, and all values were normalized using total signal intensity within each sample. For the cell type comparison between human and mouse datasets, the mean intensity for each cell type was calculated for each lipid class. The standard deviation was calculated across all mouse individuals. For the human dataset, the proportion of lipids from a given lipid class serving as cell-type markers was taken for comparison. To compare human white matter lipidome data to HRMS measurements, raw data from ref. was upper-quartile normalized for each sample and log10-transformed. The difference in the median intensity of white matter and gray matter samples was calculated for each peak. The statistical significance of this difference was assessed using t test and p-values were BH-corrected for multiple testing. Lipid peaks were matched between the two datasets based on their retention time value and mass-to-charge ratio dataset within the ±30 seconds and ±10 ppm windows. Prior to this procedure retention times were aligned between the two datasets based on a subset of lipid features matching within a more relaxed time window of ±120 seconds. Retention time correction was performed using the svm() function from the e1071 R package. For lipids passing the adjusted p-value threshold of 0.05 and fold-change threshold of two, an association between the sign of the difference and the myelin+/myelin- category was tested using Fisher exact test. MALDI imaging was conducted on the fresh frozen sections from the posterior part of superior temporal gyrus dissected from postmortem brain samples of 3 healthy donors. Fresh frozen brain tissue was sectioned into 10 μm slices using a cryostat Leica CM1950 at −18 °C. Slices of tissue sections were thaw-mounted on ITO slides (15Ω, HST, US). Sections were dried in the vacuum chamber for 2–3 h, and matrix was applied by spraying. α-Cyano-4-hydroxycinnamic acid (CHCA) was used as the MALDI matrix. A saturated solution of CHCA at 5 mg/mL−1 in 50/50 water/acetonitrile with 0.1% TFA was diluted two times. This solution was applied using an airbrush (Iwata Micron CM-B2) for 2 s and left to dry for 2.5 min, the cycle was performed 20 times. Neighboring sections were mounted on SuperFrost slides (Thermo Scientific, US) for subsequent histological staining, and stored at −80°C until processing. Histology of white and gray matter on the brain sections was revealed by luxol fast blue staining (blue color for lipid-rich compartments) and by eosin (pink color for protein-rich cytoplasm). Brain sections were briefly de-fated to increase dye penetration: placed gradually in ethyl alcohol (50%, 75%, 95% and 100%) and back to 95% for 1 min in each solution. Then sections were left in 0,1% luxol fast blue solution in ethyl alcohol with 0,5% glacial acetic acid in 56 C oven for 12–14 h. Excess stain was rinsed off with 95% ethyl alcohol and then in distilled water. Staining was differentiated in the 0,05% lithium carbonate solution in water for 30 seconds and rinsed in distilled water several times with constant control by microscopic examination if gray matter is clear and white matter sharply defined. Then sections were counterstained with 1% eosin solution in water for 30-40 seconds, rinsed in distilled water, dehydrated briefly in IsoPrep (BioVitrum), cleared in Bio Clear (Bio-Optica) and coversliped with Bio Mount HM (Bio-Optica). Histology images were acquired with Zeiss Axio.Observer.Z1 transmitted light microscope system. MALDI images were obtained using a modified MALDI-Orbitrap mass spectrometer (Thermo Scientific Q-Exactive orbitrap with MALDI/ESI Injector from Spectroglyph, LLC) equipped with a 355 nm Nd:YAG Laser Garnet (Laser-export. Co. Ltd, Moscow, Russia). For positive ion induction, the laser power was set to a 20 J repetition rate of 1.7 kHz. The distance between the sample on a coordinate table and ion funnel was 5 cm. Produced ions were captured by the ion funnel and transferred to QExactive Orbitrap mass spectrometer (Thermo). Mass spectra were obtained in the mass range of m/z 500–1000 and a mass resolution was 140,000. The tissue region to be imaged and the raster step size were controlled using the Spectroglyph MALDI Injector Software. To generate images, the spectra were collected at 40-μm intervals in both the x and y dimensions across the surface of sample. Ion images were generated from raw files (obtained from Orbitrap tune software) and coordinate files (obtained from MALDI Injector Software) by Image Insight software from Spectroglyph LLC (www.spectroglyph.com). MALDI raw mass spectra were converted to *.ibd and *.imzML formats using Spectroglyph software (https://spectroglyph.com/), with the background noise threshold set to zero. All further processing was done using Cardinal 2.8.0, an R package designed for mass spectrometry imaging data analysis. For image analysis, duplicated coordinates were removed from converted files, then spectra were normalized by the total ion current. We performed peak picking based on a signal-to-noise ratio threshold equal to three to select peak centers for the downstream analysis. Signal to noise ratio was calculated based on the difference between the mean peak height in a window of a predefined size and the mean height in the window of the manually selected flat part of the spectrum. Peaks present in less than 7% of the sample spectra were removed from further analysis. Peaks originating from the glass slide surface not covered by tissue and uninformative parts of the spectra containing no biologically relevant peaks were removed from each sample spectral data. To do so, the sample image area was divided into two parts using the spatial k-means algorithm; one of the clusters corresponding to the tissue sample and the other – to the sample-free matrix-covered surface of the glass slide. The mapping of the clusters to the sample and the surrounding area was manually curated by visual inspection of the slides. Mean feature intensities were calculated for the two clusters and only peaks with 1.5 times greater mean intensities within the sample cluster compared to the surrounding sample-free area were kept. After this filtration step, all sample spectra were aligned to a spectrum with the largest number of detected peaks. For the clustering of pixels within the sample area, the first principal component of the ion intensity matrix was used. The resulting clusters represented white and gray matter and were annotated based on histological staining of the adjacent sample sections. Log10-transformed difference between median intensities of non-zero values in the white matter and gray matter clusters was calculated for each peak in each sample. To match HRMS annotated peaks to MALDI data, only peaks with 99% quantile intensities greater than 1000 were selected. Peaks were considered matched between the datasets if the mass difference between HRMS and MALDI was below 10 ppm. Non-uniquely matched peaks were removed from further analysis. Brain samples were sectioned and processed as described above in the MALDI imaging section. ToF-SIMS measurements were performed using a ToF-SIMS 5 instrument (ION-TOF Gmbh, Germany) equipped with 30 keV Bi3+ primary ions. Mass spectra were recorded in spectroscopy mode with pixel size ~ 5 μm. First mass spectra and chemical maps were acquired for positive secondary ions and then for negative. The region of interest and analysis area was 1.8 × 1.8 mm (384 × 384 pixels) while primary ion dose density (PIDD) did not exceed 3.2x1010ions/cm2. Stage scan was utilized for all experiments. A low-energy electron flood gun was activated to avoid charging effects. Ion yields were calculated as the intensity of the corresponding peak of interest normalized to the total ion count amount. In the standards validation experiment, droplets of lipids standard were applied and dried on ITO coated slide type II 1.1 mm thick (P/N PL-IC-000010-P100, Hudson Surface Technology inc.). Eleven lipid standards manufactured by Avanti Polar lipids used in the experiment included: PE(17:0/17:0) (P/N 830756X), PG(17:0/17:0) (P/N 830456X), PI(8:0/8:0) (P/N 850181P), 17:0 SM (P/N 860585P), Lyso PC (17:0/0:0) (P/N 855676C), Lyso PE(14:0/0:0) (P/N 856735X), C17 Glucosyl(ß) Ceramide (d18:1/17:0) (P/N 860569P), Cer(d18:1-d7/15:0) (P/N 860681), d7-cholesterol (P/N 70004), C17 Ceramide (d18:1/17:0) (P/N 860517P) and mix of Brain Sulfatides (P/N 131305P-5mg). The concentration of all standards was 1 mole/ml, and droplets of 1–2 μl were dried at room temperature. ToF-SIMS measurement was performed as described above. Five ions, validated by standards and previously reported in literature were selected for comparison to HRMS data. Ion images of four samples were divided into white matter and gray matter spatial clusters based on histological staining of adjacent sections. For each sample mean intensity through all pixels was calculated for each cluster and the log10-transformed difference between white matter and gray matter clusters was calculated for each ion. For groups of lipid classes, corresponding to selected characteristic ions, the difference between proportions of lipids from a tested group contained within myelin+ and myelin- clusters based on HRMS data were normalized by the mean value of two fractions. For comparison of lipid intensity profiles to the hierarchical organization of the human brain, we build a linear model based on a subset of neocortical regions corresponding to primary, secondary, associative, and limbic cortices. Corresponding numerical values of cortical hierarchy levels were set to be from 1 to 4, respectively. To compare lipid intensity data to rs-fMRI data, a functional connectivity (FC) matrix was constructed for a subset of 59 brain regions, as described in the corresponding section of Methods. Outer product matrices were built for each lipid, and the corresponding correlation coefficient with the FC matrix was calculated. Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
|
PMC6097988
|
Phospholipidome of endothelial cells shows a different adaptation response upon oxidative, glycative and lipoxidative stress
|
Endothelial dysfunction has been widely associated with oxidative stress, glucotoxicity and lipotoxicity and underlies the development of cardiovascular diseases (CVDs), atherosclerosis and diabetes. In such pathological conditions, lipids are emerging as mediators of signalling pathways evoking key cellular responses as expression of proinflammatory genes, proliferation and apoptosis. Hence, the assessment of lipid profiles in endothelial cells (EC) can provide valuable information on the molecular alterations underlying CVDs, atherosclerosis and diabetes. We performed a lipidomic approach based on hydrophilic interaction liquid chromatography-tandem mass spectrometry (HILIC-MS/MS) for the analysis of the phospholipidome of bovine aortic EC (BAEC) exposed to oxidative (H2O2), glycative (glucose), or lipoxidative (4-hydroxynonenal, HNE) stress. The phospholipid (PL) profile was evaluated for the classes PC, PE, PS, PG, PI, SM, LPC and CL. H2O2 induced a more acute adaptation of the PL profile than glucose or HNE. Unsaturated PL molecular species were up-regulated after 24 h incubation with H2O2, while an opposite trend was observed in glucose- and HNE-treated cells. This study compared, for the first time, the adaptation of the phospholipidome of BAEC upon different induced biochemical stresses. Although further biological studies will be necessary, our results unveil specific lipid signatures in response to characteristic types of stress.The endothelium is the tissue responsible for the regulation of the hemodynamics of the whole circulatory system. Endothelial cells (EC) under oxidative stress play a pathogenic role in the onset and the development of cardiovascular diseases (CVDs) and atherosclerosis. The scenario in which the endothelium is involved in the initiation and progression of CVDs and other oxidative stress-related disorders is referred to as endothelial dysfunction, an array of maladaptive changes in the functional phenotype of EC that is known to occur upon exposure to minimally oxidized low-density lipoproteins (LDL) and overproduction of reactive oxygen species (ROS). Moreover, diabetes and hyperglycemia can also trigger oxidative stress, leading to endothelial dysfunction, which contributes to diabetic retinopathy, cardiovascular complications and atherosclerosis. Also, hyperglycemia has been associated with endothelial dysfunction through the decrease of cell viability and the induction of EC apoptosis. Oxidative stress can also lead to the formation of aldehydic lipid peroxidation products (ALPP), as 4-hydroxy-2-nonenal (HNE), which can further induce oxidative stress in EC. HNE is able to exert prominent cytotoxic effects in human umbilical EC (HUVEC) that result in morphological changes, diminished cellular viability, impaired endothelial barrier function, and cell apoptosis. Therefore, the endothelial barrier dysfunction promoted by HNE may contribute to the vascular changes that lead to the development of atherosclerosis. In the CVDs that are related to endothelial dysfunction and, more broadly, in chronic inflammatory diseases related to oxidative stress, lipids have progressively been considered as key molecules mediating the outbreak and the progression of such pathologies. During the last decade, we have assisted to the rapid development of lipidomics, a group of analytical platforms and protocols aimed at the assessment of lipid metabolic profiles and networks in biological systems. Lipidomics can provide information about the molecular basis of CVDs, highlight the links between lipid functions and pharmacological treatments, and allow a more in-depth monitoring of the response to therapies. However, the evaluation of the lipidome of EC is still limited. Murphy and co-authors reported the phospholipid (PL) compositions of cultured EC from human artery, saphenous vein, and umbilical vein, and observed a similar profile for the three cell types. Héliè-Toussaint and co-authors further studied the lipidomic pathways of HUVEC, observing a preferential homeostasis leading to the synthesis of PL rather than triacylglycerols, and a fast incorporation of palmitic acid and arachidonic acid in the membrane PL pool. More recent insights in cardiovascular lipidomics have allowed the characterization of the lipidome of human atherosclerotic plaques, pinpointing an enrichment in phosphatidylcholines (PC), oxidized phosphatidylcholines (ox-PC) and lyso-PL within the cells. Nevertheless, the understanding of the pathogenic mechanisms underlying CVDs requires the study of the phospholipidome of EC upon stressing conditions such as hyperglycemia and overproduction of ROS. However, up to date, only Yang and co-authors have investigated the variations in the lipidome of human EC upon oxidative stress. A phospholipidomic fingerprinting of EC subjected to biochemical stress would represent a very informative model of cardiovascular pathobiology, aimed to understand the molecular mechanisms of adaptation that may occur during endothelial dysfunction and contribute to the onset of CVDs. In the present study, we wanted to assess whether specific stress conditions would induce distinctive changes in the lipidome of EC. For this, we employed hydrophilic interaction liquid chromatography coupled to mass spectrometry (HILIC-MS/MS) for the phospholipidomic profiling of cultured bovine aortic EC (BAEC), which constitute a widely used model for vascular biology studies, upon oxidative (H2O2), glycative (glucose) or lipoxidative (HNE) stress conditions. The workflow used to carry out the entire experiment is shown in Fig. 1. Our results show for the first time that the lipidome of EC is exquisitely responsive to diverse stress conditions, and thus, may mediate specific adaptive changes.Figure 1Schematic diagram showing the experimental workflow including cell treatment, cell lysis, lipid extraction, chromatographic separation and MS analysis. Schematic diagram showing the experimental workflow including cell treatment, cell lysis, lipid extraction, chromatographic separation and MS analysis. Lipids have recently emerged as key mediators in the onset of chronic inflammatory diseases characterized by endothelial dysfunction and oxidative stress. Hence, we employed a HILIC-LC-MS/MS platform to analyse the phospholipid profile of BAEC treated in control conditions and in response to different stressing agents (H2O2, glucose, or HNE). The data sets resulting from the HILIC-LC-MS/MS analysis of four sample groups (control, H2O2, glucose, and HNE) were later subjected to both univariate and multivariate statistical analysis, aiming to identify significant changes occurring in the BAEC phospholipid profile upon induced biochemical stress. We performed the identification and the relative quantification of PL species belonging to 8 different classes: phosphatidylcholine (PC), phosphatidylethanolamine (PE), phosphatidylserine (PS), phosphatidylglycerol (PG), phosphatidylinositol (PI), lyso-PC (LPC), cardiolipin (CL) and sphingomyelin (SM). The whole list of the 109 PL species (correspondent to the most abundant species in all the identified classes) that were identified and quantified after MS and MS/MS analysis of each sample can be found in Supplementary Table S1. The total chain length (C) and degree of unsaturation (N) are included. Also, the different isomers of the same classes that bear different esterified fatty acids and correspond to each C:N composition were included. These isomers cannot be resolved by the LC-HILIC method, but the fatty acyl composition was determined by MS/MS analysis. Negative ion mode MS/MS data were used to analyse fatty acid carboxylate anions fragments, which allowed to assign the fatty acyl chains esterified to the PL molecular species. For the relative quantification of all the PL listed in Supplementary Table S1, the peak areas of the extracted ion chromatograms (XICs) of each PL species (C:N) within each class were normalized using the peak area of the internal standard (IS) selected for the class. Data were subsequently autoscaled and then subjected to a principal component analysis (PCA) to display the clustering trends of the four experimental groups of BAEC: control, H2O2-, glucose-, and HNE-treated. The PCA showed that all the groups were separated from each other in a two-dimensional score plot which represented the analyses describing 66.8% of the total variance, including principal component 1 (51.1%) and principal component 2 (15.7%), where principal component 1 was the major discriminating component (Fig. 2). From the loading values, PE (34:3), PE (36:2), PE (38:2), PE (36:1) and PE (36:3) were the major contributors from component 1, whereas PI (36:2), PE (34:1), PC (O-32:0), PI (36:1) and PG (34:2) were the main contributors for component 2. Control samples were scattered on the central region of the plot. Glucose or HNE-treated samples were scattered on the left region of the plot according to the order: glucose, HNE. Interestingly, H2O2-treated samples formed the only group that was scattered on the right region of the plot.Figure 2Principal component analysis score plot of the phospholipid profiles obtained from BAEC treated with glucose, H2O2 or HNE. Control, vehicle medium; Glucose, 25 mM glucose; H2O2, 1 mM hydrogen peroxide; HNE, 10 µM 4-hydroxy-2-nonenal. All samples were analysed after 24 hours of treatment. Principal component analysis score plot of the phospholipid profiles obtained from BAEC treated with glucose, H2O2 or HNE. Control, vehicle medium; Glucose, 25 mM glucose; H2O2, 1 mM hydrogen peroxide; HNE, 10 µM 4-hydroxy-2-nonenal. All samples were analysed after 24 hours of treatment. Additionally, we carried out a hierarchical clustering analysis (HCA) on the phospholipid data sets from the four conditions (Fig. 3). The resulting HCA dendrogram depicted a noticeable separation of the four data sets of control, H2O2, glucose, and HNE. The first level of separation was evidenced between H2O2-treated samples and the remaining conditions. The second level of separation distinguished control samples from the two remaining groups (glucose and HNE-treated). However, glucose- and HNE-treated cells differentiated in two clusters in the third level of separation.Figure 3Hierarchical cluster analysis of the phospholipid profiles obtained from BAEC treated with glucose, H2O2 or HNE. Control, vehicle medium; Glucose, 25 mM glucose; H2O2, 1 mM hydrogen peroxide; HNE, 10 µM 4-hydroxy-2-nonenal. All samples were analysed after 24 hours of treatment. Hierarchical cluster analysis of the phospholipid profiles obtained from BAEC treated with glucose, H2O2 or HNE. Control, vehicle medium; Glucose, 25 mM glucose; H2O2, 1 mM hydrogen peroxide; HNE, 10 µM 4-hydroxy-2-nonenal. All samples were analysed after 24 hours of treatment. Then, we performed a projection to latent structures discriminant analysis (PLS-DA) in order to maximize the phenotypic classification of samples, which showed the performance statistics of R2X = 0.94241, R2Y = 0.96805 and a high prediction parameter Q2 of 0.80026 (X) and 0.90986 (Y). The four groups were well separated in the resulting two-dimensional score plot (Fig. 4). The PLS-DA score plot described 65.8% of the total variance, including component 1 (16.8%) and component 2 (49%). Along with component 2, control samples were scattered at the central region of the plot. Glucose- and HNE-treated samples were scattered on the top region of the plot. H2O2-treated samples formed the only group that was scattered at the bottom region of the plot. Along with component 1, control- and glucose-treated samples were scattered at the left side of the plot, while HNE- and H2O2-treated samples were scattered on the right side of the plot.Figure 4Projection to latent structures discriminant analysis score plot of the phospholipid profiles obtained from BAEC treated with glucose, H2O2 or HNE. Control, vehicle medium; Glucose, 25 mM glucose; H2O2, 1 mM hydrogen peroxide; HNE, 10 µM 4-hydroxy-2-nonenal. All samples were analysed after 24 hours of treatment. Projection to latent structures discriminant analysis score plot of the phospholipid profiles obtained from BAEC treated with glucose, H2O2 or HNE. Control, vehicle medium; Glucose, 25 mM glucose; H2O2, 1 mM hydrogen peroxide; HNE, 10 µM 4-hydroxy-2-nonenal. All samples were analysed after 24 hours of treatment. Besides multivariate statistical analyses, we used another approach to facilitate the interpretation of the extensive dataset produced in the present study, namely the analysis of semi-quantitative phospholipidomic features across distinct depths of detail. For a clearer overview of the phospholipidome, we included the analysis of PL by features, assessing the adaptation of all the PL species bearing the same number of unsaturation, or bearing the same total carbon chain length, as already performed by other authors. Therefore, the cumulative levels of all the PL species comprised of the same number of unsaturation, ranging from zero, one, two, three and four double bonds, were summarized in the singular lipidomic features PL-DB0, PL-DB1, PL-DB2, PL-DB3, and PL-DB4, respectively. When comparing globally the degree of unsaturation observed for BAEC PL upon the four tested conditions (control, H2O2, glucose and HNE), the highest cumulative levels of (poly-)unsaturated species (PL-DB1, PL-DB2, PL-DB3, and PL-DB4) were always observed after the treatment with H2O2 (Fig. 5). Conversely, treatment of BAEC with glucose and HNE downregulated the levels of PL-DB1, PL-DB2, PL-DB3 and PL-DB4 when compared with control (Fig. 5). Only when comparing HNE with control, changes in PL-DB4 were not statistically significant.Figure 5(A) Phospholipid species comprised of the same number of unsaturation on the fatty acyl chains, given as normalized XIC area for each category (the contribution of CL is not included for clarity). (B) Phospholipid species comprised of the same number of carbon atoms on the hydrocarbon chains, given as normalized XIC area for each category (the contributions of CL and LPC are not included for clarity). ****** Statistically significant variation between selected conditions (p < 0.05, p < 0.01 and p < 0.001, respectively). (A) Phospholipid species comprised of the same number of unsaturation on the fatty acyl chains, given as normalized XIC area for each category (the contribution of CL is not included for clarity). (B) Phospholipid species comprised of the same number of carbon atoms on the hydrocarbon chains, given as normalized XIC area for each category (the contributions of CL and LPC are not included for clarity). ****** Statistically significant variation between selected conditions (p < 0.05, p < 0.01 and p < 0.001, respectively). Analogously, the cumulative levels of all the PL species comprised of the same number of carbons in their hydrocarbon chain were summarized in the singular lipidomic features PL-C32, PL-C34, PL-C36, PL-C38, and PL-C40, respectively. Regardless of the treatment, PL-C36 species were always the most abundant in BAEC. Treatment of BAEC with H2O2 lead to an increase of the features PL-C34, PL-C36, and PL-C38 when compared with control (Fig. 5). Conversely, we observed a downregulation of PL-C32, PL-C34, PL-C36, PL-C38, and PL-C40 for BAEC treated with glucose, and a downregulation PL-C32, PL-C34, PL-C36, and PL-C40 for BAEC treated with HNE, in comparison with control (Fig. 5). We did not observe any statistically significant variation for PL-C32 and PL-C40 after treatment with H2O2, nor for PL-C38 after treatment with HNE. For a more detailed interpretation of the data, we further addressed the adaptation of single PL species induced by the different stressing treatments. The most abundant PC molecular species was PC (34:1) followed by PC (36:2), in all the conditions. We observed a statistically significant increase of the levels of PC (34:1), PC (34:2) and PC (36:4) in cells treated with H2O2, when compared to control cells (p < 0.05). On the other hand, we observed a statistically significant decrease of the levels of PC (30:0) and PC (32:0) in H2O2-treated cells in comparison with control. Interestingly, all these five PC molecular species, PC (34:1), PC (34:2), PC (36:4), PC (30:0) and PC (32:0), were decreased in high glucose-treated cells when compared to control cells (p < 0.05). PC (30:0), PC (32:0) and PC (34:1) were also decreased in cells treated with HNE in comparison with controls (p < 0.05). No significant alterations between HNE-treated cells and control cells were observed for PC (34:2) and PC (36:4) (Fig. 6 and Table 1).Figure 6Box plots of the 24 most discriminant PL molecular species from BAEC treated with glucose, H2O2 or HNE. CTL, vehicle medium; GLU, 25 mM glucose; H2O2, 1 mM hydrogen peroxide; HNE, 10 µM 4-hydroxy-2-nonenal. All samples were analysed after 24 hours of treatment.Table 1Summary of the alterations observed in the molecular species of PC, PE, PG, PS, PI, and SM from BAEC comparing control with H2O2-treated, control with glucose-treated and control with HNE-treated cells, along with their respective fold changes.ClassSpecies (C:N)H2O2 vs. controlGlucose vs. controlHNE vs. controlAdaptationFold changeAdaptationFold changeAdaptationFold changePC30:0↓1.60↓1.13↓1.3032:0↓1.64↓1.47↓1.3834:1↑0.87↓1.09↓1.1336:4↑0.70↓1.4434:2↑0.75↓1.21PE34:3↑0.73↓1.21↓1.1338:6↑0.90↓1.79↓1.2236:3↑0.79↓1.37↓1.1536.2↑0.82↓1.29↓1.1638:3↑0.84↓1.48↓1.26PG38:5↓1.49↓1.2532:1↓1.40↓1.3634:1↓1.53↓1.27↓1.5940:6↓1.87↓2.14↓1.3236:0↓1.34↓1.44PS36:1↓0.91↓2.6136:2↓1.30↓1.06↓1.5640:5↓3.57↓1.85↓4.4038:4↑0.2640:7↑0.29PI38:5↑0.95↓1.38↓1.2638:6↑0.79↓1.36↓1.3036:1↓1.20↑0.75↓1.5536:2↑0.91↑0.82↓1.3938:3↑0.85↓1.29↓1.23SM34:2↑0.81↓1.21↓1.1036:2↑0.78↓1.32↓1.1438:2↑0.79↓1.70↓1.1740:2↑0.82↓1.36All the alterations are significant at the p < 0.05 level. CTL, vehicle medium; GLU, 25 mM glucose; H2O2, 1 mM hydrogen peroxide; HNE, 10 µM 4-hydroxy-2-nonenal. All samples were analysed after 24 hours of treatment. Box plots of the 24 most discriminant PL molecular species from BAEC treated with glucose, H2O2 or HNE. CTL, vehicle medium; GLU, 25 mM glucose; H2O2, 1 mM hydrogen peroxide; HNE, 10 µM 4-hydroxy-2-nonenal. All samples were analysed after 24 hours of treatment. Summary of the alterations observed in the molecular species of PC, PE, PG, PS, PI, and SM from BAEC comparing control with H2O2-treated, control with glucose-treated and control with HNE-treated cells, along with their respective fold changes. All the alterations are significant at the p < 0.05 level. CTL, vehicle medium; GLU, 25 mM glucose; H2O2, 1 mM hydrogen peroxide; HNE, 10 µM 4-hydroxy-2-nonenal. All samples were analysed after 24 hours of treatment. In all the conditions, PE (36:2) was the most abundant PE molecular species followed by PE (36:3). The levels of PE (36:3), along with PE (34:3), PE (38:6), PE (36:2) and PE (38:3), were increased in H2O2-treated cells when compared to controls (p < 0.05). Conversely, we observed a statistically significant down-regulation for all these five PE species in both glucose- and HNE-treated cells compared to controls (p < 0.05) (Fig. 6 and Table 1). When comparing cells treated with H2O2 with controls we found that PS (38:5), PS (38:4) and PS (40:7) levels were increased, while PS (36:2) and PS (40:5) were decreased (p < 0.05). We also observed a significant decrease of PS (36:2), PS (40:5) and PS (36:1) in both glucose and HNE-treated cells compared with controls (p < 0.05) (Fig. 6 and Table 1). A significant increase of the levels of PI (38:5), PI (38:6), PI (36:2) and PI (38:3), and a significant decrease of PI (36:1), were observed in cells treated with H2O2 compared to control cells (p < 0.05). When comparing glucose-treated cells with controls, the levels of PI (38:5), PI (38:6), PI (38:3) were decreased while PI (36:1) and PI (36:2) were increased (p < 0.05). We also observed a statistically significant decrease for all these five PI species in cells treated with HNE when compared with controls (p < 0.05) (Fig. 6 and Table 1). The treatment of cells with H2O2 induced a decrease on the levels of PG (32:1), PG (34:1) and PG (40:6), compared to controls (p < 0.05). We also observed decrease of the levels of PG (34:1) and PG (40:6), along with PG (38:4) and PG (36:0), in glucose-treated cells when compared with controls. The levels of five PG molecular species - PG (32:1), PG (34:1), PG (40:6), PG (38:5) and PG (36:0) – were found to be decreased in cells treated with HNE compared to controls (p < 0.05) (Fig. 6 and Table 1). The most abundant SM molecular species in all conditions was SM (34:1). We observed increased levels of the species SM (34:2), SM (36:2), SM (38:2), SM (40:2) in cells treated with H2O2, compared to controls (p < 0.05). Conversely, all these five species of SM were decreased in glucose-treated cells, when compared to controls (p < 0.05). Treatment of cells with HNE also decreased the levels of the species SM (34:2), SM (36:2), SM (38:2), in comparison with controls (Fig. 6 and Table 1). We did not observe any statistically significant difference for the levels of the molecular species belonging to the CL and LPC classes. Lipids have been highlighted as biomolecules involved in the onset of CVDs, a family of chronic inflammatory diseases in which vascular pathobiology is associated with endothelial dysfunction. Cellular lipid profiling can provide evidence at the molecular level for the role of lipids in these diseases. In our study, we performed a phospholipidomic profiling on BAEC subjected to H2O2, glucose, or HNE treatment, aiming to elucidate the adaptations in the PL of EC upon three models of biochemical stress associated with the onset and/or progression of CVDs. In a first instance, the multivariate analyses that we performed (PCA, PLS-DA, and HCA) indicated that the phospholipid profiles of BAEC were significantly altered in response to each one of the induced biochemical stresses (H2O2, glucose or HNE). The two-dimensional plots of PCA (Fig. 2) and PLS-DA (Fig. 4), along with the dendrogram of HCA (Fig. 3), showed that the most evident differentiation of the BAEC-treated phospholipidome, in comparison with controls, occurred in the model of oxidative stress, 24 h after exposure to 1 mM H2O2. This suggests that the H2O2-mediated toxicity induced more particular and specific changes in the PL profile of BAEC when compared with glucotoxicity and lipotoxicity, presumably due to the adaptation mechanisms established by the cells to survive the exposure to oxidative stress. Yang and co-authors studied a hybrid cell line (EA.hy296) subjected to oxidative stress, and using phospholipidomic data found that samples were clearly clustered depending on the different exposure times to 0.2 mM H2O2 (0, 1, 2, 3 and 6 h). Besides the present study, no other published works have reported the adaptation of the phospholipid profile of EC upon hyperglycemia. In our study, BAEC treated with 25 mM glucose for 24 h showed a significant alteration of the phospholipidome. PCA, HCA and PLS-DA (Figs 2, 3 and 4, respectively) all exhibited good clustering of the phospholipid profile of glucose-treated samples when compared to controls, differing from the trend observed for H2O2-treated samples. The clustering evidenced by HCA, in which high glucose samples are relatively closer to controls in comparison to H2O2 samples, suggests that the cellular effects mediated by glucose affected the phospholipid profile of BAEC via less harsh changes when compared to a directly induced acute oxidative stress. In this study, we also assessed the phospholipidome of BAEC upon HNE exposure. Although the biological effects of HNE on several EC types have already been widely studied (as reviewed by Chapple and co-authors), this is also the first study that aimed to evaluate the variations in the phospholipid profile of EC treated with an ALPP. In this study, we used 10 µM HNE, a concentration in the range of those measured in disease states. BAEC treated for 24 h with HNE showed a significant phenotypical differentiation of their phospholipid profile, as shown in Figs 2, 3 and 4. Treatment of BAEC with H2O2 increased the cumulative levels of (poly-)unsaturated PL species PL-DB1, PL-DB2, PL-DB3, and PL-DB4, in comparison with control (Fig. 5). Remarkably, it had been previously described that direct oxidative stress (H2O2 1 mM) induced PL peroxidation within the first hours of treatment, while along 24 hours of treatment, BAEC may have established an adaptive mechanism of re-acylation of unsaturated fatty acyl chains in the PL pool, aimed to counteract the ROS-mediated peroxidation. An increase in (poly)-unsaturated PL species was also reported by Peterson and co-authors, in primary neurocortical cells incubated with H2O2 for 24 h. Hence, this is in agreement with our results that show augmented levels of (poly)-unsaturated species in H2O2-exposed BAEC. The amount of (poly)-unsaturated PL species in the plasma membrane is known to be an effective modulator of its physical properties, as phase transition temperature. In our study, the up-regulation observed for the (poly)-unsaturated species of BAEC subjected to oxidative stress might have resulted in increased membrane fluidity. We found this adaptation of interest since an augmented membrane fluidity was also observed by Sergent in rat primary hepatocytes subjected to ethanol-induced oxidative stress, which would be by our conjectures. However, it is important to remark that in the present study the phospholipidome composition was not strictly studied for the plasma membrane. Hence PL (poly)-unsaturated species might also belong to other cellular organelles. Additionally, other lipids that were not analysed in the present study (e.g., cholesterol) are also regulators of membrane fluidity. The attempt to evaluate significant changes in both high and low abundant PL molecular species led us to perform a univariate statistical analysis using autoscaled data. Importantly, whereas highly abundant PL species are central for maintaining structural and biophysical properties, lower abundant PL species might more likely act as mediators of signalling functions. Treatment with H2O2 resulted in a decrease of two PC molecular species esterified to saturated fatty acids, namely PC (30:0) and PC (32:0), along with an increase of three (poly)-unsaturated PC molecular species, namely PC (34:1), PC (34:2) and PC (36:4) (Fig. 6 and Table 1). First, these results reflect the general increase of PL-DB1, PL-DB2, PL-DB3, and PL-DB4. More deeply, Cai and Harrison reviewed the regulation of EC by H2O2 and highlighted the fundamental role of this ROS in promoting the inflammatory state of the endothelium. In our conditions, treatment of BAEC with H2O2 led to the upregulation of PC (34:1), a ligand of the nuclear receptor PPARα, that upon activation regulates several anti-inflammatory genes. Thus, the up-regulation of PC (34:1) could be part of a protective strategy adopted by the cells in response to the inflammatory state promoted by oxidative stress. We observed a significant up-regulation in the molecular species PE (34:3), PE (38:6), PE (36:3), PE (36:2) and PE (38:3) in BAEC treated with H2O2. PE is the second most abundant PL class in mammalian cells, hence the increase of singular unsaturated molecular species correlates with the increase that we observed for PL-DB1, PL-DB2, PL-DB3, and PL-DB4. Due to the relatively small head group and the high unsaturation degree characterizing PE molecules, significant changes in the composition of PE species may affect the curvature of the cell membrane and contribute to an increase in its fluidity Of interest, Hailey and co-authors found mitochondrial PE to be the recruiters of autophagy markers in starved mammalian cells, while Rockenfeller and co-authors highlighted that increased levels of PE enhanced the lifespan of cultured mammalian cells via promotion of autophagy. Since high levels of H2O2 are known to induce apoptosis in cultured EC, we postulate that the up-regulation of PE species herein observed is an adaptive mechanism of survival adopted by the cells in response to the early apoptotic signals triggered by the H2O2 treatment. Curiously, H2O2 treatment led to a complex regulation of (poly)-unsaturated PS species in BAEC, since PS (38:4) and PS (40:7) were up-regulated, while PS (36:2) and PS (40:5) were down-regulated. The anti-inflammatory properties of PS in mammalian cells have recently been highlighted, hence the increase of PS (38:4) and PS (40:7) could represent a mechanism promoted by BAEC to compensate the inflammatory state triggered by H2O2. Since the formation of ox-PS in mammalian cell membranes is a hallmark of apoptosis, we interpret that the decrease in (poly)-unsaturated PS is a consequence of the early apoptotic signals that might have been triggered by incubation with H2O2. Five SM species were up-regulated after treatment with H2O2, namely SM (34:2), SM (36:2), SM (38:2) and SM (40:2). In eukaryotic cells, SM is known to co-localize and associate with sterol lipids and display structural roles related to cholesterol homeostasis and membrane distribution. The interaction between SM and cholesterol forms ordered domains that are primarily involved in the regulation of cell membrane fluidity. In this sense, the oxidative stress-induced up-regulation of SM species observed in this study might reflect a mechanism of perturbation in the fluidity of the plasma membrane of treated BAEC. Nevertheless, there are other SM cellular functions that deserve discussion. Yang et al. treated yeast cells with increasing H2O2 concentrations and found that cell death effects were strongly abrogated in cells expressing sphingomyelin synthase 1 (SMS1), the principal enzyme catalysing the synthesis of cellular SM species. Since BAEC always maintained cell viability along the 24 h treatment with H2O2 (no cell death was observed), the significant increase in SM molecular species might represent an augmented SMS1 activity, adopted by the cells to abolish the cell death signals induced by H2O2. Altogether, the significant variations herein observed for several PL and SM molecular species might arise from a wide array of BAEC adaptive mechanisms occurring upon H2O2-induced stress, which include response to inflammation, impaired membrane fluidity, and cell death. However, it is important to remark that none of the above-listed responses was measured in the present study, hence future works will be necessary to validate what has been herein conjectured. The global observation of PL features highlighted that the treatment with glucose reduced the levels of PL-DB1, PL-DB2, PL-DB3, and PL-DB4 when compared with control (Fig. 5). A similar trend was observed for PL-C32, PL-C34, PL-C36, and PL-C38 (Fig. 5). To the best of our knowledge, the effects of hyperglycemia on the phospholipidome of EC have never been reported before. In this regard, it is interesting to pinpoint the experiments of Mīinea et al., which found a diminished Δ5 desaturase activity and a markedly decreased biosynthesis of arachidonic acid-containing PL in Schwann cells grown in 30 mM glucose. The majority of the PL species that showed significant variation, reported in Fig. 6 and Table 1, were down-regulated after the treatment with glucose. The alterations induced by high glucose may reflect several mechanisms including biological signalling, redox imbalance and formation of advanced glycation end products (AGEs). Hempel and co-authors observed an increased EC permeability mediated by the activation of PKC, upon hyperglycemia. Duffy and co-authors reported that 72 h hyperglycemia induced apoptosis in human aortic EC (HAEC). Noteworthy, an excess of glucose can lead to NADPH consumption and formation of sorbitol, which in turn can induce oxidative and osmotic stress, respectively. Several authors have reported the promotion of ROS generation and the redox imbalance mediated by high glucose in EC. The ability of high glucose to induce 12/15-lipoxygenase (12/15-LOX) in mesangial cells and in EC was also highlighted. Glucose is also known to modify proteins and aminophospholipids (PE and PS). Altogether, these mechanisms are referred as glucotoxicity and contribute to the endothelial dysfunction that occurs in diabetes. However, the cellular mechanisms linking glucotoxicity and PL turnover in mammalian cells are still unclear. On one hand, the formation of aminophospholipid-AGEs could have contributed to the decreased levels observed for PE and PS species; additionally, PE AGEs were found to trigger oxidative stress, which can be a further explanation of the down-regulation of (poly)-unsaturated PL species observed in glucose-treated BAEC. On the other hand, the decrease of (poly)-unsaturated PL species may have been induced by the radical-based oxidative stress promoted by the 24 h exposition to high glucose. However, the ability of 12/15-LOX to oxidize (poly)-unsaturated fatty acid esterified to PL has recently been suggested, hence a glucose-induced up-regulation of 12/15-LOX could be another explanation for the decrease in (poly)-unsaturated PL. We analysed the global changes occurring in the PL features of BAEC after the treatment with HNE. The levels of PL-DB1, PL-DB2, and PL-DB3, along with the levels of PL-C32, PL-C34, PL-C36, and PL-C40 were downregulated when compared with controls (Fig. 5). Our work constitutes the first report on the effects of HNE on the phospholipidome of BAEC. The treatment with HNE also led to a down-regulation of all the PL molecular species reported in Fig. 6 and Table 1. HNE induces several intracellular cascades in EC, as reviewed by Chapple and co-authors. Increased levels of this ALPP (1–100 µM) were found in disease states, mediating cellular-damaging pathways, including increased ROS generation and endoplasmic reticulum (ER) stress, altogether leading to EC dysfunction. Although the interplay between HNE-mediated toxicity (lipotoxicity) and the alteration of EC lipidome is mostly unknown at present, the ER stress induced by HNE in EC should be carefully considered, since this organelle is directly involved in the biosynthesis of several PL classes. The down-regulation that we observed for all the PL species can be due to a HNE-induced perturbation of the PL biosynthetic pathways in the ER. However, BAEC treated with HNE were also found to suffer increased intracellular ROS levels, increased apoptosis and down-regulation of antioxidant defences. This oxidative stress response triggered by HNE in BAEC can finally lead to membrane PL peroxidation, which would explain the decreased levels of (poly)-unsaturated PL species that were herein observed. Moreover, similarly to glucose, HNE is able to form covalent adducts with nucleophilic sites in aminophospholipids, which could further explain the decrease of PS and PE molecular species in HNE-treated BAEC. In summary, the results from the present study point out that the phospholipidome of BAEC suffers statistically significant changes upon different biochemical stresses (H2O2, glucose, and HNE). For the first time, the phopsholipidomic profiling of BAEC was compared between homeostasis, oxidative stress, glucotoxicity and lipotoxicity, and specific lipidome alterations were reported for all the tested conditions. The molecular adaptation observed for the treatment with H2O2 was distinctive when compared with the alterations induced by glucose or HNE. More deeply, we found H2O2 to increase the cellular levels of (poly)-unsaturated PL molecular species, while the same species were down-regulated after the treatment with glucose or HNE. These evidences highlight that PL are fundamental players in the response of vascular cells to such external stresses. A specific adaptation of the whole PL profile of EC can represent a cellular hallmark for the onset and the development of CVDs, atherosclerosis, and diabetes. Nevertheless, the biochemical mechanisms adopted by BAEC that lead to the alteration of the phospholipidome are still unclear, particularly in the case of glucolipotoxicity, and a considerable gap of knowledge exists regarding the lipid profiling of EC in non-homeostatic or pathological conditions. These results open new avenues of research towards further studies necessary to unveil the interplay between biochemical stress, PL turnover, and onset of cardiovascular pathologies and diabetic complications. Phospholipid internal standards 1,2-dimyristoyl-sn-glycero-3-phosphocholine (dMPC), 1,2-dimyristoyl-sn-glycero-3-phosphoethanolamine (dMPE), 1,2-dimyristoyl-sn-glycero-3-phospho-(10-rac-)glycerol (dMPG), 1,2-dimyristoyl-sn-glycero-3-phospho-L-serine (dMPS), tetramyristoylcardiolipin (TMCL), 1,2-dipalmitoyl-sn-glycero-3-phosphatidylinositol (dPPI), N-palmitoyl-D-erythro-sphingosylphosphorylcholine (NPSM), 1-nonadecanoyl-2-hydroxy-sn-glycero-3-phosphocholine (LPC) were purchased from Avanti Polar Lipids, Inc. (Alabaster, AL). Chloroform, methanol and acetonitrile were purchased from Fisher scientific (Leicestershire, UK); all the solvents were of high performance liquid chromatography (HPLC) grade and were used without further purification. All the other reagents and chemicals used were of the highest grade of purity commercially available. The water was of Milli-Q purity (Synergy1, Millipore Corporation, Billerica, MA). Bovine aortic endothelial cells (BAEC) were obtained from Lonza, Inc., (Walkersville, MD) and cultured in RPMI1640 medium supplemented with antibiotics (100 U/mL penicillin and 100 µg/mL streptomycin) and 10% calf serum from Gibco (Life Technologies). For each experiment, BAEC were used between passages 8 and 16 and were grown to near confluence for experiments (80–90% density). Milli-Q water was used as a solvent throughout the experiments (H2O2, glucose, HNE). Treatment of BAEC with the different stressing agents (H2O2, glucose, HNE) was carried out in serum-free medium. This induces a near-quiescent state and does not reduce cell viability. Control cells received an equivalent amount of water as required. Cells were subjected to biochemical stress with 1 mM H2O2, 25 mM glucose, or 10 µM HNE during 24 h. After treatment, cells were collected by scraping in PBS on ice and centrifuged at 1000 rpm for 5 min. Cell pellets were stored at −80 °C. The whole experimental procedure, including cell culture and treatments, lipid extraction and analysis, was repeated four times. Thereafter, total lipids were extracted using the Bligh and Dyer method, and the phospholipid (PL) amount in each lipid extract was quantified with the phosphorus assay, performed according to Bartlett and Lewis. For the detailed experimental procedures of lipid extraction and PL quantification, the reader is referred to a previously published work in which the same methodologies were applied. Phospholipids were separated by hydrophilic interaction liquid chromatography (HILIC-LC-MS) using a high performance-LC (HPLC) system (Thermo scientific Accela) with an autosampler coupled online to the Q-Exactive® hybrid quadrupole Orbitrap® mass spectrometer (Thermo Fisher Scientific, Bremen, Germany). The solvent system consisted of two mobile phases as follows: mobile phase A [acetonitrile:methanol:water 50:25:25 (v/v/v) with 1 mM ammonium acetate] and mobile phase B [acetonitrile:methanol 60:40 (v/v) with 1 mM ammonium acetate]. Initially, 0% of mobile phase A was held isocratically for 8 min, followed by a linear increase to 60% of A within 7 min and a maintenance period of 15 min, returning to the initial conditions in 10 min. A volume of 5 µL of each sample containing 5 µg of phospholipid extract, 4 µL of phospholipid standards mix (dMPC - 0.02 µg, dMPE - 0.02 µg, NPSM - 0.02 µg, LPC - 0.02 µg, TMCL - 0.08 µg, dPPI - 0.08 µg, dMPG - 0.012 µg, dMPS - 0.04 µg) and 91 µL of eluent B was introduced into the Ascentis® Si column (15 cm × 1 mm, 3 µm, Sigma-Aldrich) with a flow rate of 40 µL min and at 30 °C. The mass spectrometer with Orbitrap® technology operated simultaneously in positive (electrospray voltage 3.0 kV) and negative (electrospray voltage −2.7 kV) modes with a high resolution of 70,000 and AGC target 1e6. The capillary temperature was 250 °C and the sheath gas flow was 15 U. No intensity threshold was used during full MS acquisitions. PC and LPC species were analysed in the LC-MS spectra in the negative ion mode as acetate anions adducts [M + CH3COO]. PE, PS, PI, PG and CL species were also analysed in the LC-MS spectra in the negative ion mode, as [M − H] ions. SM molecular species were analysed in the LC-MS spectra in positive ion mode as [M + H] ions. In MS/MS experiments, a resolution of 17,500 and AGC target of 1e5 were used and the cycles consisted of one full scan mass spectrum and ten data-dependent MS/MS scans that were repeated continuously throughout the experiments, with the dynamic exclusion of 60 seconds and intensity threshold of 1e4. Normalized collision energy™ (CE) ranged between 25, 30 and 35 eV. At least one blank run was performed between different treatment samples to prevent cross-contamination. Data acquisition was carried out using the Xcalibur data system (V3.3, Thermo Fisher Scientific, USA). Phospholipid peak integration and assignments were performed using MZmine 2.30. The software allows for filtering and smoothing, peak detection, peak processing, and assignment against an in-house database. During the processing of the raw data acquired in full MS mode, all the peaks with raw intensity lower than 1e4 were excluded. For all assignments, peaks within 6 ppm of the lipid exact mass were considered. Consequently, assigned PL were further validated by manual analysis of the MS/MS data (Supplementary Table S1). Analysis of the MS/MS spectra acquired in positive ion mode was performed to confirm the identity of the molecular species belonging to the PC, LPC, SM, and PE classes. The fragment ion at m/z 184.07, when observed in the MS/MS acquired in the positive mode, indicates the presence of phosphocholine head group, allowing to identify the precursor ions as belonging to the PC, LPC and SM classes. The species belonging to these three classes were further differentiated by the characteristic retention times. The neutral loss of 141 Da, when observed in the MS/MS acquired in the positive mode, allows identifying precursor ions of the PE class. The MS/MS spectra acquired in negative ion mode were used to confirm the identity of PG, PI and PS phospholipids. In particular, the fragment ion at m/z 171.01 was used to confirm [M-H] PG class precursor ions and the fragment ion at m/z 241.01 was used to confirm [M − H] PI class precursor ions. The neutral loss of 87 Da, when observed in the MS/MS spectra acquired in the negative mode, allows identifying precursor ions belonging to the PS class. Negative ion mode MS/MS data were used to identify the fatty acid carboxylate anions fragments RCOO, which allowed the assignment of the fatty acyl chains esterified to the PL precursor. All the MS/MS fragmentation patterns characteristic for the lipid classes analysed in the present study, acquired in positive and negative ion modes, are available online as supplementary information. Relative quantitation was performed by exporting integrated peak areas values into a computer spreadsheet (Excel, Microsoft, Redmond, WA). For normalization of the data, the peak areas of the XICs of the PL precursors of each class (listed in Supplementary Table S1), (C:N composition), were divided for the peak area of the IS selected for the class. Initial data assessment was performed using Metaboanalyst. Shapiro-Wilk test was used for analysing the normality of the BAEC treatment data. Univariate statistical analysis was performed using ANOVA test following post hoc least significant difference (LSD) test, with Benjamin–Hochberg correction for the false discovery rate (FDR). Univariate data processing and analyses were performed using the SPSS software package (IBM SPSS Statistics Version 24). Principal component analysis (PCA), projection to latent structures discriminant analysis (PLS-DA) and hierarchical clustering were performed on auto-scaled data using the R software (R version 3.4.2 downloaded from https://www.R-project.org) with the packages RFmarkerDetector, FactoMineR, Factoextra and Ropls. PCA was performed using the FactoMineR and ellipses were drawn assuming a multivariate normal distribution. Hierarchical clustering was performed using Ward’s method using the method hclust. Boxplots were created using the R package ggplot. Bar graphs were created using the software GraphPad Prism 7. The datasets generated during and/or analysed during the current study are available online as supplementary material.
|
PMC5050431
|
LC-MS Based Sphingolipidomic Study on A2780 Human Ovarian Cancer Cell Line and its Taxol-resistant Strain
|
Drug resistance elicited by cancer cells continue to cause huge problems world-wide, for example, tens of thousands of patients are suffering from taxol-resistant human ovarian cancer. However, its biochemical mechanisms remain unclear. Sphingolipid metabolic dysregulation has been increasingly regarded as one of the drug-resistant mechanisms for various cancers, which in turn provides potential targets for overcoming the resistance. In the current study, a well-established LC-MS based sphingolipidomic approach was applied to investigate the sphingolipid metabolism of A2780 and taxol-resistant A2780 (A2780T) human ovarian cancer cell lines. 102 sphingolipids (SPLs) were identified based on accurate mass and characteristic fragment ions, among which 12 species have not been reported previously. 89 were further quantitatively analyzed by using multiple reaction monitoring technique. Multivariate analysis revealed that the levels of 52 sphingolipids significantly altered in A2780T cells comparing to those of A2780 cells. These alterations revealed an overall increase of sphingomyelin levels and significant decrease of ceramides, hexosylceramides and lactosylceramides, which concomitantly indicated a deviated SPL metabolism in A2780T. This is the most comprehensive sphingolipidomic analysis of A2780 and A2780T, which investigated significantly changed sphingolipid profile in taxol-resistant cancer cells. The aberrant sphingolipid metabolism in A2780T could be one of the mechanisms of taxol-resistance.Ovarian cancer is the most aggressive gynecologic cancer and thus a leading cause of cancer-related mortality in women worldwide1. At present, the most effective strategy for ovarian cancer is combination therapy based on cytoreductive surgery and chemotherapy with taxanes (e.g. taxol), but intrinsic or acquired tumor chemoresistance remains the most important clinical problem and a major obstacle to a successful therapy2. According to a systematic literature review, 69 of the total 137 acquired drug-resistant cell lines were resistant to taxol3. Seventy-five percent of ovarian cancer patients initially respond to platinum or taxane based chemotherapy; however, most of them eventually develop chemotherapy resistance4. Many factors can lead to drug resistance, including increased drug efflux, drug inactivation, alterations in drug target, processing of drug-induced damage, and evasion of apoptosis5. Mechanisms including overexpression of drug resistant associated proteins6 and activation of some signaling pathways7 have been implicated in resistance to taxol, but the overall molecular mechanisms of taxol resistance still need further elucidation. Sphingolipids (SPLs) are a kind of membrane and intracellular lipids that typically play structural roles and act as signaling molecules and/or modulators of signaling pathways associated with cell survival8. Besides the most widely studied bioactive SPL - ceramide, the relationship between cancer and other SPL has been extensively studied, including sphingosine 1-phosphate (S1P)9, glucosylceramide (GluCer)10, sphingosine and C1P11. Growing evidence showed that sphingolipids are deeply involved in the regulation of apoptosis as well as the apoptosis resistance that is displayed by cancer cells12. Qualitative and quantitative assessment of SPLs could reveal novel biomarkers for early diagnosis of cancer13. There are several studies focused on the sphingolipidomics of A2780 Human Ovarian Cancer cell line1415, as well as its fenretinide-resistant16 and multidrug-resistant strains1718. Valsecchi M et al. have characterized the sphingolipidomes in N-(4-hydroxyphenyl)retinamide (4-HPR) and 4-oxo-N-(4-hydroxyphenyl)retinamide (4-oxo-4-HPR) treated A2780 cells by ESI-MS, revealed that the two drugs differentially affect the early steps of SPL synthesis19. In 4-HPR resistant A2780 cells (A2780/HPR), a remarkable alteration of sphingolipid metabolism with respect to both of the parental sensitive A2780 cells and 2780AD cells has been revealed20. Increasing evidence suggests the change of SPL metabolism can be (one of) the crucial mechanism of drug resistance in A2780 cells. However, till now, there is no sphingolipidomic study on taxol resistant A2780 cells (named as A2780T, TA2780, A2780/Taxol, or A2780/PTX in literature). Therefore, a comprehensive sphingolipidomic study is required for elucidating the mechanisms underlying the resistance of A2780T cells to taxol treatment. In the current study, SPLs in A2780 and A2780T were comprehensively profiled and quantitatively determined by using a well-established LC-MS approach developed in our lab21. It appears to be a promising tool for viewing overall sphingolipidomic difference between taxol-sensitive and -resistant strain of A2780. Duplicate analyses of pooled samples of A2780 and A2780T cells (QC samples) were carried out to achieve comprehensive profiling of SPLs in these two cell lines. Ultra-high performance liquid chromatography coupled with Q-TOF mass spectrometry (UHPLC-Q-TOF MS) is an effective and sensitive analytical tool to separate and identify SPLs in a complex mixture. By integrating the high efficient separation offered by UHPLC, high-resolution mass spectrum obtained by MS and MS/MS on Q-TOF, as well as comparing the data with those of reference standards and searching against our personal database, totally 102 SPLs have been identified in the pooled samples, among which six ceramides (d18:1/17:3; d18:1/15:3(OH); d18:1/14:3(OH); d18:2/23:1; d18:0/18:3 and d17:0/13:0(OH)), two ceramide-1-phosphates (d18:1/19:0(OH) and d18:1/12:2), one hexosylceramide (d18:1/20:1), and three sphingosines (d16:3; d15:3 and t19:2) are new SPLs. Sixty-seven out of the 102 SPLs were reported for the first time in A2780 cells. MS signals might be masked by isomeric, isotopic or isobaric ions. For sphingolipidomic profiling of A2780, our improved sphingolipidomic approach showed great potential in differentiating isomeric and isotopic species as that have been observed in PC12 cells21. A major interference in the identification of SPLs is the isomeric species that have exact identical molecular elemental compositions, thus MS/MS data together with optimized separation are essential for discrimination. For instance, the extracted ion chromatogram of m/z 620.5903 at 5 ppm mass accuracy yielded two peaks at 15.894 and 16.061 min. Targeted MS/MS of m/z 620.6 at respective time points gave distinct product ions corresponding to backbone of Cer (d18:1/22:1) (m/z 264.3) and Cer (d18:2/22:0) (m/z 262.3), providing evidence for the identification of these two species (Fig. 1). The targeted ion pairs together with complete chromatographic separation also enabled subsequent quantification of such isomers by using multiple reaction monitoring (MRM) technique. Notably, 4 pairs of isomeric species (A1–A4) were clearly distinguished in our study (Table 1). The comprehensive profiling of SPLs provided an overall “picture” of the sphingolipidome of A2780 cells. Generally, sphingomyelin (SM) is the most abundant subclass of SPLs in this cell line. Totally 43 SMs, including 31 dehydrosphingomyelins and 12 dihydrosphingomyelins (DHSMs), were identified based on exact mass and characteristic product ions obtained in targeted MS/MS experiments, 31 of which are reported for the first time in A2780 cell line. All the SMs were found to possess a C18 sphingoid base chain, with d18:1 account for the majority, comparing to the d18:0 and d18:2 backbones. The length of N-acyl chain varies from 14 to 26, and the unsaturation degree ranges from 0 to 5. Notably, the N-acyl chains of all the 12 DHSMs are fully saturated. Two highly unsaturated (total unsaturation degree no less than 4) SMs, SM (d18:1/24:3) and SM (d18:2/24:3), have been detected in A2780 cells for the first time. In A2780 cells, 26 Cers, including 19 dehydroceramides and 7 dihydroceramides (DHCers), were identified based on the MS information and, in some cases, by comparing the retention time with that of SPLs in PC12 cells in our previous study21. Most Cers detected in the sample were with a d18:1 sphingoid backbone, with carbon number of N-acyl chain ranged from 14 to 24. Three dehydroceramides and 4 DHCers with a hydroxyl group on N-acyl chain have been characterized, among which 2 dehydroceramides and 1 DHCer with short N-acyl chain (carbon number less than 16) were reported for the first time to the best of our knowledge. The other 3 new Cers were species with high degree of unsaturation, for instance, Cer (d18:1/17:3), Cer (d18:2/23:1) and a new DHCer (d18:0/18:3). A notable ceramide was DHCer (d17:0/13:0(OH)), which was a very uncommon DHCer with odd carbon number sphingoid backbone. Due to the limitation of chromatographic separation, galactosylceramide and glucosylceramide cannot be distinguished, thus these two hexose-linked ceramides were represented as HexCer. All C1P, HexCer, and lactosylceramide (LacCer) species exclusively bared a d18:1 sphingoid base backbone. The dominant HexCers and LacCers are d18:1/24:1, d18:1/24:0 and d18:1/16:0. Two novel C1Ps, i.e. C1P (d18:1/19:0(OH)) and C1P (d18:1/12:2), were identified in A2780. The former one has an N-acyl chain with odd carbon number and a hydroxyl group, while the latter one has two degrees of unsaturation on the N-acyl fatty chain. Eighteen sphingoid bases with carbon number ranging from 14 to 20 were successfully identified. Short chain sphingosines with high unsaturation degree (d16:3 & d15:3) and a sphingosine with 3 hydroxyl groups (t19:2) have been discovered as uncommon species. Comparing to routine LC-MS based approaches, UHPLC coupled with QQQ mass spectrometer in MRM mode provides more sensitive and accurate quantification with wider dynamic range of SPLs. However, the quantification of SPLs cannot be accomplished accurately in LC-MS/MS analysis with a QQQ analyzer solely, as triple-quadruple cannot distinguish isotopic/isobaric ions within 0.1 Da when selecting the precursor ions. For instance, each unsaturated SPL could generate an isotopic interference on SPLs with less degree of unsaturation as exemplified by SM (d18:1/14:0) and SM (d18:0/14:0) (Fig. 2). In our approach, based on foregoing comprehensive profiling by Q-TOF and the optimized chromatographic separation, all the structurally similar SPLs were accurately quantified with elimination of such isotopic/isobaric interference. With the optimized MRM conditions, 89 species from 9 subclasses out of 102 identified SPLs were quantified by using the UHPLC-QQQ MS method. It was found that A2780 and A2780T share most common sphingolipid molecules, except for HexCer (d18:1/20:1) which is only present in A2780. The amounts of SPLs were measured by using the internal standards previously mentioned, duplicate measurements for each sample yielded consistent results in all cases. The quantitative results showed that SMs take the highest proportion of all the SPLs, among which SMs with C18 sphingoid base backbone are the dominant species (Fig. 3). In A2780 cells the most dominant species are SM (d18:1/16:0) which corresponding to [M + H] at m/z 703, followed by DHSM (d18:0/16:0) (m/z 705), SM (d18:1/16:1) (m/z 701) with less relative abundance. The d18:1 SMs with C16/C18/C22/C24 N-acyl chain showed relative high levels in both A2780 and A2780T. Forty-two out of the 43 SMs were quantified except for DHSM (d18:0/25:0), whose content is lower than the limit of quantitation (LOQ). A total of 20 Cers have been quantified, but most of them are d18:1 species due to the weak intensity of d18:0 backbone fragment ions (m/z 266.4). According to the finding of Koyanagi et al. in tumors only the content of C16 N-acyl chain ceramide (C16-Cer) are significantly high22, that can explain why other DHCers cannot be quantified exactly. As shown in Fig. 4, the content of individual Cers differ dramatically (at most 500 times), for some common species like d18:1/18:1, d18:1/24:0 and d18:1/24:1, the contents are significantly higher than that of highly unsaturated species d18:1/24:2. In general, the amounts of Cers are significantly higher in A2780 than those in A2780T. Among all the HexCers and LacCers, only d18:1 sphingoid base backbone type was found. All the 12 HexCers and LacCers showed higher intensity in A2780 cells than that in A2780T. Sphinganine, as the precursor of DHCer, showed decrease in A2780T. The overall content of sphingosines was similar in both cell types, but the expression of individual sphingosine was quite different. Higher level of Cer1P (d18:0/20:0) was detected in A2780T (data not shown). Figure 5 showed the trends of all 7 marker HexCers and LacCers between A2780 and A2780T. Multivariate analyses were further carried out to view the overall differences between A2780 and A2780Ts, and to identify SPL markers that were significantly changed in A2780T. PLS-DA was used to visualize general clustering among A2780, A2780T, and QC groups firstly (Fig. 6A). After auto scaling of data sets, discrimination feature between the profiles were identified for each model by displaying loadings plots. Loading plots and VIP value in PLS-DA model are commonly used for biomarker selection and identification. According to the results, potential SPL markers that were differentially expressed between A2780 and A2780T groups were identified (Fig. 6B and Table 2), suggesting a SPL alternation was involved in A2780T. A total of 52 potential biomarkers were identified according to the VIP value and scattering-plot, among which most of them are sphingomyelins, several highly unsaturated SPLs [SM (d18:1/24:3), SM (d18:1/24:2) and SM (d18:1/22:2)] were also included. LacCer (d18:1/24:1) showed the largest decline in A2780T, whose content decreased by approximately 70 folds, which contributes most significantly to the classification. In order to drive study on the metabolism of sphingolipids, a reliable and informative analytical method for the comprehensive profiling of SPLs is essential. By using a combined analytical strategy, which enables the reliable identification and sensitive quantification, the dynamic distribution and interconversion of SPLs have been comprehensively monitored. Our improved sphingolipidomic analyses on A2780 and A2780T encompassed most of the important SPLs including sphinganine, sphingosine, ceramide-1-phosphase, hexosylceramide, lactosylceramide, dehydroceramide and dihydroceramide as well dehydrosphingomyelin and dihydrosphingomyelin subclasses. It is the most comprehensive sphingolipidomic study on A2780/A2780T cells to date, as evidenced by the identification of up to 102 SPLs including 67 species that are reported in the cell line for the first time. Distinguished from previous studies, this research of SPL took advantage of a well-established LC-MS method, and looked into the content variation of individual SPL species instead of the overall content of each subclass, thus provided much more detailed and useful information for revealing the mechanism of taxol-resistance. Most of the identified SPLs can be the metabolic pathway related biomarkers, especially the low abundance species of which the subtle changes may result in altered biological function like drug resistance23. It’s noted that all the rare SPLs (odd number of carbons/high level of unsaturation) in A2780/A2780T are with the low abundance. Similarly, SPLs with odd number of carbons (C15 and C17) have been reported24, and highly unsaturated SPLs were isolated from halotolerant fungus with poor natural abundance25. Even in A2780 cell line, Cer with C23 and C25 N-acetyl chain have already been found19. Discovery of these rare SPL species is one of the research highlights of this study. Comparing to taxol sensitive A2780 cells, the most notable alteration in A2780T cells was the overall decrease of Cers. Eleven Cers are recognized as biomarkers of A2780T, along with a 1.5 to 13-fold decrease has been observed. Cer is known as an intracellular messenger that is able to regulate many intracellular effectors mediating activation of the apoptotic process. It has been recognized as a kind of tumor suppressor and has been found to act as a major player in the action of many chemotherapeutic drugs26. Dysregulated metabolism of Cers has been identified as a feature of many drug-resistant cancers27, as well as in taxol resistant human ovarian cancer cell line CABA-PTX28. In A2780T cells the depletion of Cers could potentially help the cells to evade Cer-induced apoptosis, and thereby can be a crucial mechanism responsible for the drug resistance of A2780T. Totally 34 SMs (including 26 dehydrosphingomyelins and 8 DHSMs) were identified as biomarkers, which took most proportion of the biomarkers. Among all the marker SMs, the content of 6 dehydrosphingomyelins increased in taxol-resistant cells compared to sensitive cells by 0.4–1.2 times. Especially, C16-SMs, a kind of high abundance SPL in both A2780 and A2780T, were found to be significantly higher in the taxol resistant cells than those in the sensitive cells. This leads to the increased total SM level in A2780T, same as previously reported in 2780AD cells29. However, the other 20 dehydrosphingomyelin biomarkers decreased by up to 90%. Individually, the content of all d18:2 SMs, d18:1 SMs with unsaturated double bond(s) at the N-acyl chain, as well as d18:1 SMs with saturated N-Acyl chain of C18 to 23 decreased significantly in A2780T comparing to that of A2780. Whilst the content of d18:1 SMs with saturated N-Acyl chain of C16–C17 and C24–C26, as well as all DHSMs, were found to increase significantly in A2780T comparing to A2780. Of note, the increase of DHSMs in A2780T is high up to 8-fold for most species. Dihydrosphingolipids have received increasing attention. Wang et al. have determined that 4-HPR treated MDR cancer cells displayed elevations in DHCer but not dehydroceramides, together with elevated DHSM species rather than dehydrosphingomyelins30. It indicates that dihydrosphingolipids may fulfil a distinctive role in the metabolic pathway comparing to unsaturated sphingolipids. In A2780T, significant increase of dehydrosphingomyelins and DHSMs concomitant with decrease of corresponding DHCers (which was not identified as markers) have been observed. These variations are consistent with the hypothetical “DHCer - DHSM - dehydrosphingomyelin” pathway, and the activity of related enzymes (dihydroceramide desaturase and dihydroceramide synthase) may be altered31. Cer plays a central role in the sphingolipid metabolism. All the Cers showed consistent trend of decrease in A2780T, except for some extremely low species (DHCers) whose content cannot meet the limit of quantitation. The overall decrease of Cers and accompanying increase of most SMs in A2780T cells, especially, the decrease of two most abundant Cers [Cer (d18:1/24:0) and Cer (d18:1/16:0)] and concomitant increase of corresponding species of SMs [SM (d18:1/24:0) and SM (d18:1/16:0)], clearly indicated SM-related depletion of Cers in A2780T cells. The roles of SMase and SMS in cancer treatment have been well recognized for decades. Their actions have been defined as one of the main routes for the alteration of Cer8. Sphingomyelinases are key enzymes of sphingolipid metabolism that regulate the formation and degradation of ceramide32. Drugs (including taxol) enhanced ceramide-governed cytotoxic response by activating sphingomyelinase27. While SM is the end product in the SM-Cer related pathway, and inhibition of SMS will result in Cer accumulation with effect solely on SM33. Thus, it can be proposed that in A2780T cells, the decreased level of Cers might be resulted from the down-regulated expression/activity of SMase or up-regulated SMS expression/activity. Similar mechanism has been reported that a decrease of the ceramide level via activation of glucosylceramide synthase (GCS) and SMS was detected in chemoresistant HL-60/ADR human promyelocytic leukemia cells34. Besides Cer and SM, other SPLs and SPL metabolites also have biological activities that could be responsible for the acquisition of a drug resistance phenotype35. Ceramide glycosylation by the enzyme glucosylceramide synthase, which forms glucosylceramide and has been noted in some drug-resistant cell lines, is an important pathway for bypassing apoptosis3637. Because SPLs comprised of d18:1 sphingosine backbone are the major species found in mammals38, in A2780T only HexCer and LacCer with d18:1 backbone can be detected and further quantified as markers. Additionally, glucosylceramide is known as an intermediate metabolite in the synthesis and degradation of the more complex gangliosides, and a number of drug-resistant cancer cell lines accumulate this noncytotoxic metabolite27. In our case of A2780T, the decrease of glucosylceramide and LacCer can be explained as “activation of ganglioside pathway”8, which enable cancer cells convert Cer into gangliosides to evade the pro-apoptotic function of Cer. The enzymes related to the “glucosylceramide-lactosylceramide-ganglioside pathway”, including glucosylceramidase, glucosylceramide synthase, and lactosylceramide synthase, could have participated in the biological progress. In A2780T, reduced syntheses of Cer, HexCer, and LacCer were observed, with the concomitant increase of DHSM and total SM levels, in which C16-SMs contributes the vast majority. These represent the main sphingolipid metabolism pattern in A2780T, which is significantly different from the SPL profiles in similar ovarian cancer cell lines. On one hand, comparing with the sphingolipidome in another taxol resistant human ovarian cell line CABA-PTX28, the level of sphingomyelin in A2780T changed significantly. On the other hand, A2780T also showed different sphingolipidomic profile from A2780 cell lines resistant to other drugs. In sharp contrast to the well-studied MDR A2780 cells29, the rise of cellular HexCer (including galactosylceramide and glucosylcermide) levels was not observed in A2780T. And in A2780/HPR cells the glycosphigolipid-dominated alteration20 is also different from the SPL pattern in A2780T, which possesses a distinctive feature of “SM-related depletion of Cers”. It indicated that the resistant mechanism of A2780T could be different from that of either other taxol-resistant cancer cells (CABA-PTX) or A2780 cells resistant to other drugs (MDR A2780 & A2780/HPR). Such interdisciplinary basic scientific research has close relevance to the medical community and it facilitates the applications in rapid detection and classification of disease type (taxol-resistant or not) and medication direction. Since the role of sphingolipids in cancer cell has been widely recognized, comprehensive sphingolipidomic study is essential for exploring its drug resistance mechanism. The most comprehensive and accurate method described in this paper fully identified SPLs in A2780 human ovarian cancer cell line and the taxol-resistant cell line A2780T. Most individual species, including some low abundance but biologically important SPLs, have been accurately quantified. It provides more detailed information than general overview of a whole subclass, which is significant for studying the alterations39. The sphingolipid metabolism in A2780T is oriented toward down-regulation of ceramides. We propose A2780T cells may escape from the ceramide-caused apoptosis mainly via sphingomyelin/ceramide pathway, while SMS was expressed more in A2780T than in the sensitive cell line, or the activity of SMase was inhibited. These enzymes related to the marker SPLs and altered pathways, are the potential targets. Based on the sphingolipidomic study, adjusting the sphingolipid metabolism purposively may represent a winning strategy to overcome taxol resistance and improve cancer therapy. This study facilitates not only development of new drugs against taxol resistance, but also clinical diagnosis of taxol-resistant ovarian cancer. Human ovarian cancer cell line (A2780) and its taxol-resistant strain (A2780T) were purchased from KeyGen Biotech Co., Ltd. (Nanjing, China). The LIPID MAPS internal standard (IS) cocktail in ethanol, composed of 25 μM each of nine sphingolipid standards including SM (d18:1/12:0), Cer (d18:1/12:0), C1P (d18:1/12:0), HexCer (d18:1/12:0), LacCer (d18:1/12:0), Sphinganine (d17:0), Sphingosine (d17:1), Sphinganine-1-Phosphate (d17:0) and Sphingosine-1-Phosphae (d17:1), was purchased from Avanti Polar Lipids (Alabaster, AL, USA). HPLC-grade methanol (MeOH), chloroform (CHCl3) and isopropanol (IPA) were purchased from Merck (Darmstadt, Germany). Ammonium acetate (NH4OAc), potassium hydroxide (KOH), acetic acid (CH3COOH) and formic acid (HCOOH) were purchased from Sigma-Aldrich (St. Louis, MO, USA). Dulbecco’s Modified Eagle’s Medium (DMEM), Roswell Park Memorial Institute (RPMI) 1640 medium, Fetal Bovine Serum (FBS), Penicillin-Streptomycin (PS) were purchased from Gibco, New Zealand. Sodium dodecyl sulfate (SDS) and 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT) were purchased from Acros, USA. A pooled sample equally aliquoted from all samples can provide the most comprehensive information within a specific study. Hence, equivalent amount of A2780T was spiked into A2780 to prepare a pooled quality control (QC) sample. A2780 human ovarian cancer cell line were cultured in DMEM supplemented with 1% PS and 10% FBS in a humidified 5% CO2 atmosphere at 37 °C. For lipid analysis, cells were seeded into dishes and grown to confluence. Cells were rinsed twice with ice-cold PBS and scraped into a borosilicate glass tube with polytetrafluoroethylene coated top. After adding 0.5 mL of MeOH, 0.25 mL of CHCl3 and 10 μL of internal standards cocktail (2.5 μM) successively, the mixture was sonicated at room temperature for 30 s then incubated at 48 °C for 12 h to extract SPLs. After 75 μL of KOH in MeOH (1M) was added in, the mixture was incubated in a shaking water bath for 2 h at 37 °C to cleave potentially interfering glycerolipids. After cooling and neutralization with acetic acid, a four-step extraction procedure was performed as reported to prepare the SPLs for LC-MS analysis. In order to verify the taxol resistance in the commercial A2780T cell line, MTT assay was employed. A2780T cells were cultured in RPMI 1640 medium containing 10% FBS, 1% PS and 800 ng/mL taxol solution at 37 °C in a 5% CO2/95% air atmosphere. For assessment of cell viability, A2780 and A2780T were respectively seeded in a 96-well plate at a density of 5 × 10 cells/well and were allowed to adhere for 16 h before treatment. Following incubation for 48 h, MTT solution (10 μL per well, 5 mg/mL solution) was added to each well and incubated for 4 h at 37 °C. Thereafter 100 μL 10% SDS was added for lysing and the plate was maintained overnight at 37 °C in a 5% CO2/95% air atmosphere. The optical densities were determined at 570 nm using a microplate reader. The procedure was repeated three times. The sensitivity of A2780 and A2780T cell lines to taxol was assayed by using MTT assay. The IC50 for taxol in A2780 and A2780T was 73.16 nM and 149.6 μM, respectively. The result indicates that resistance to taxol of A2780T is at least 1000 fold greater than that of A2780. Sphingolipid analysis was performed by using our well-established LC-MS method21 with minor optimization, an Agilent 1290 UHPLC tandem 6550 quadrupole time-of-flight (Q-TOF) MS system and an Agilent 1290 UHPLC tandem 6460 triple quadrupole (QQQ) MS system were employed for qualitative profiling and quantitative analysis respectively. Chromatographic separation was performed as described previously, while the injection volume was 10 μL for Q-TOF and 5 μL for QQQ, respectively. An Agilent Eclipse Plus C18 column (100 × 2.1 mm, 1.8 μm) was used to separate the endogenous SPLs. The mobile phase consisted of (A) MeOH/H2O/HCOOH (60:40:0.2, v/v/v) and (B) MeOH/IPA/HCOOH (60:40:0.2, v/v/v), both containing 10 mM NH4OAc. The flow rate was 0.4 mL/min, and the column temperature was maintained at 40 °C for each run. A linear gradient was optimized as follows: 0–3 min, 0% to 10% B; 3–5 min, 10% to 40% B; 5–5.3 min, 40% to 55% B; 5.3–8 min, 55% to 60% B; 8–8.5 min, 60% to 80% B; 8.5–10.5 min, 80% to 80% B; 10.5–16 min, 80% to 90% B; 16–19 min, 90% to 90% B; 19–22 min, 90% to 100% B, followed by washing with 100% B and equilibration with 0% B. A typical data acquisition time was 20 min. The above UHPLC system was interfaced with an Agilent ultrahigh definition (UHD) 6550 Q-TOF mass spectrometer equipped with an ESI source (Santa Clara, CA, USA). The source parameters were: drying gas (N2) temperature 150 °C, flow rate 15 L/min, nebulizer pressure 25 psi, sheath gas (N2) temperature was set at 200 °C with a flow-rate at 12 L/min. The scan parameters were: positive ion mode over m/z 110–1200, capillary voltage 4000 V, nozzle voltage 300 V, fragmentor voltage 175 V, skimmer voltage 65 V, octopole RF peak 500 V, drying gas 6 L/min at 300 °C. A reference solution was nebulized for continuous calibration in positive ion mode using the reference mass m/z 922.00979800. The acquisition and data analysis were controlled using Agilent Mass Hunter Workstation Software (Agilent, USA). The UHPLC conditions for quantitative analysis were the same as those mentioned above. The LC system was coupled to an Agilent 6460 triple-quadrupole mass spectrometer (Santa Clara, CA, USA). The ESI parameters were optimized as follows: positive ion mode, drying gas (N2) temperature 325 °C, flow rate 11 L/min, nebulizer pressure 45 psi, capillary voltage 4000 V, nozzle voltage 300 V, sheath gas (N2) temperature was set at 200 °C with a flow-rate at 12 L/min. Data were processed with Agilent Mass Hunter Workstation Software. Further detail of the parameters, such as characteristic transitions, fragmentor and CE voltages optimized for each compound, and the methodology validations are similar as described before. The screening and identification of SPLs were performed by searching in our personal database, which was built and updated based on the Agilent Mass Hunter Personal Compound Database and Library (PCDL) software and LIPID MAPS information (31643 SPLs until July 08 2015). The sphingolipidomic approach was applied in qualitative research of SPLs by analyzing a pooled sample equally mixed by A2780 and A2780T. In quantitative research, A2780 cells (models, n = 10) and A2780T cells (models, n = 10) were analyzed in parallel. Supervised multivariate statistical analysis, partial least squares to latent structure-discriminant analysis (PLS-DA) method, was used to differentiate the amounts of SPLs between the two strains. Potential biomarkers were selected according to Variable Importance in Projection (VIP) value and the loading scattering-plot, using SIMCA-P+ software version 13.0 (Umetrics, Umea, Sweden). VIP values higher than 1.00 were considered significant. How to cite this article: Huang, H. et al. LC-MS Based Sphingolipidomic Study on A2780 Human Ovarian Cancer Cell Line and Its Taxol-resistant Strain. Sci. Rep. 6, 34684; doi: 10.1038/srep34684 (2016).
|
PMC9011034
|
Quantitative Lipidomics and Spatial MS-Imaging Uncovered Neurological and Systemic Lipid Metabolic Pathways Underlying Troglomorphic Adaptations in Cave-Dwelling Fish
|
Sinocyclocheilus represents a rare, freshwater teleost genus endemic to China that comprises the river-dwelling surface fish and the cave-dwelling cavefish. Using a combinatorial approach of quantitative lipidomics and mass-spectrometry imaging (MSI), we demonstrated that neural compartmentalization of lipid distribution and lipid metabolism is associated with the evolution of troglomorphic traits in Sinocyclocheilus. Attenuated docosahexaenoic acid (DHA) biosynthesis via the Δ4 desaturase pathway led to reductions in DHA-phospholipids in cavefish cerebellum. Instead, cavefish accumulates arachidonic acid-phospholipids that may disfavor retinotectal arbor growth. Importantly, MSI of sulfatides coupled with immunostaining of myelin basic protein and transmission electron microscopy images of hindbrain axons revealed demyelination in cavefish raphe serotonergic neurons. Demyelination in cavefish parallels the loss of neuroplasticity governing social behavior such as aggressive dominance. Outside the brain, quantitative lipidomics and qRT-PCR revealed systemic reductions in membrane esterified DHAs in the liver, attributed to suppression of genes along the Sprecher pathway (elovl2, elovl5, and acox1). Development of fatty livers was observed in cavefish; likely mediated by an impeded mobilization of storage lipids, as evident in the diminished expressions of pnpla2, lipea, lipeb, dagla, and mgll; and suppressed β-oxidation of fatty acyls via both mitochondria and peroxisomes as reflected in the reduced expressions of cpt1ab, hadhaa, cpt2, decr1, and acox1. These neurological and systemic metabolic adaptations serve to reduce energy expenditure, forming the basis of recessive evolution that eliminates nonessential morphological and behavioral traits and giving cavefish a selective advantage to thrive in caves where proper resource allocation becomes a major determinant of survival.Sinocyclocheilus (Cypriniformes, Cyprinidae) is a rare, freshwater teleost genus endemic to Southwestern China, one of world’s largest cave-rich karst geomorphologic regions (Meng et al. 2013). Sinocyclocheilus can exist in different forms, predominantly the surface-dwelling species and the cave-dwelling species (Meng et al. 2013; Zhao et al. 2021). Sinocyclocheilus cave dwellers first colonized cave habitats to seek refuge in deeper waters in response to widespread drying associated with the aridification of China during the late Miocene and Pliocene (Mao et al. 2021). Cave habitat is considered as an extreme environment due to perpetual darkness and food scarcity (Simon et al. 2017). Cave-dwelling fish across the world independently evolved a series of troglodyte characteristics and behavioral adaptations to enhance survival, such as enhanced sensation, eyesight degeneration, loss of pigmentation, and dominance aggressiveness, as well as a disrupted circadian rhythm (Elipot et al. 2013, 2014; Meng et al. 2013; Soares and Niemiller 2013). As the brain of vertebrates displays a high degree of conservation both in terms of anatomical structures and neuromodulatory signaling (Katz and Lillvis 2014), rapid evolution of behavior may instead rely on distinct compartmentalization of neuromodulatory signaling networks (Chattopadhyay et al. 2012). Lipids, as biophysical constituents of biological membranes, partake in the morphological and functional compartmentalization in the brain via membrane remodeling (Aureli et al. 2015). For example, the formation and remodeling of myelin sheaths around specific axons can be harnessed to fine-tune neural plasticity in achieving bidirectional regulation of functional behavior (Chang et al. 2016). The surface fish and cavefish thus provide a natural setting for understanding how brain lipid metabolism may regulate neural plasticity leading to distinct behavioral traits during evolution. Serotonin (5-HT) signaling mediated by different subtypes of 5-HT receptors (5-HTR) is phylogenetically conserved. Aberrant 5-HT signaling has been implicated in various neurological diseases including depression and schizophrenia (Bjork et al. 2010). Differential regulation of serotonergic signaling in specific neuronal populations was found to constitute the basis of differential behavioral patterns in Mexican surface fish and cavefish (Elipot et al. 2013). Enhanced hypothalamic serotoninergic signaling in cavefish mediates passive foraging behavior, although downregulation of raphe serotoninergic signaling modulates social dominance and aggressive behavior in river-dwelling surface fish (Elipot et al. 2013). In a later study, it was found that a mutation in monoamine oxidase (MAO) leads to impeded 5-HT turnover and thus higher 5-HT content throughout the brain of cavefish, partly explaining why cavefish are unable to downregulate raphe serotonergic signaling to establish dominance aggressiveness like their surface counterparts (Elipot et al. 2014).This does not explain, however, why surface fish can elicit experience-dependent downregulation in raphe 5-HT signaling. Since membrane lipid micro milieu can possibly alter receptor ligand-binding efficiency and its subsequent activity (Sjögren and Svenningsson 2007), dynamic changes in neural lipid membrane remodeling may offer a quick avenue for modulating 5-HT signal transduction. In this study, we investigated region-specific differences in the spatial distribution and metabolism of lipids in the brain, in association with distinct behaviors of the Sinocyclocheilus cavefish (Sinocyclocheilu anophthalmus) and surface fish (Sinocyclocheilu augustiporus). We leveraged on a combination of high-sensitivity targeted mass spectrometry based on multiple reaction monitoring and matrix-assisted laser desorption ionization-Fourier-transform ion cyclotron resonance mass spectrometry (MALDI-FTICR MS) on adjacent brain tissue sections to obtain a spatial atlas of lipid distribution in four major brain regions of Sinocyclocheilus species. Investigating these differences can help one to understand how lipids modulate neuroplasticity and behavioral traits and confers new evidence that neural compartmentalization of lipid distribution and lipid metabolism may function to regulate behavior in Sinocyclocheilus species. In addition to the brain, we also examined quantitative lipidome changes in the eyes and livers of the two fish species, uncovering alterations in systemic metabolism that facilitate troglomorphic adaptations to life in cave environments. In all, the neurological and systemic adaptations in cavefish follow a common theme to eliminate nonessential morphological and behavioral traits, such as the loss of advanced eye functions and social behavior, which cumulatively serve to reduce energy expenditure and confer a selective advantage to surviving in cave environments with irregular food supply. Sinocyclocheilus fishes (declared as national second-class protected animals in China in 2020) were wild-caught near Kunming, Yunnan province of China between the years of 2016 and 2018. The cave species, S. anophthalmus, were found in a cave along the Nanpanjiang River (supplementary fig. 1, Supplementary Material online), whereas the surface fish, S. angustiporus, were obtained at sites along the Huangnihe River (fig. 1A). Akin to Astyanax mexicanus, we found that the surface species of Sinocyclocheilus displays obvious aggregation behavior in the wild and in the laboratory tank, whereas the cavefish S. anophthalmus behave as a single individual in the cave and in a dispersed state in the fish tank (supplementary fig. 1, Supplementary Material online). Using whole-brain transcriptome data published previously (Meng et al. 2013, 2018) (supplementary table S1, Supplementary Material online), we first performed Gene Set Enrichment Analysis (GSEA) to investigate the differences in KEGG pathways between these two species. GSEA revealed enhanced arachidonic acid (ARA) metabolism and oxidative phosphorylation in the brain of cavefish relative to surface fish (fig. 1B). We then analyzed the whole-brain lipidomes (comprising >750 individual lipids) (supplementary table S2, Supplementary Material online) of the two species using targeted LC-MS/MS approaches (Song, et al. 2020; Lam, et al. 2021). Principal component analysis showed that the brain lipidomes of cavefish and surface fish were clearly segregated (fig. 1C), and the clustering heatmap of major lipid classes illustrated appreciable elevations of complex glycosphingolipids in the brains of surface fish (fig. 1D). To examine global alterations in lipid coregulations, we analyzed lipid correlations in the whole-brain lipidome of cavefish and surface fish, respectively (fig. 1E). Chord diagrams revealed a strong negative correlation (red shades) between storage triacylglycerols (TAGs) and mitochondria-resident cardiolipins (CLs) in the brain of cavefish, which was absent in surface fish. CLs denote the signature phospholipid class of the mitochondria, which are localized at the foldings of the inner mitochondrial membrane, known as crista—the primary site of oxidative phosphorylation where protein components of the electron transport chains reside (Ikon and Ryan 2017). A negative correlation, therefore, indicates high CLs at the expense of storage TAGs, implying an enhanced mobilization of storage TAGs into free fatty acyls, which are consumed via mitochondrial oxidative phosphorylation in cavefish brain (Ikon and Ryan 2017). Lipid correlation analyses thus corroborated pathway analysis based on GSEA that cavefish brains elicit enhanced oxidative phosphorylation. Indeed, inhibition of brain oxidative phosphorylation was previously shown to increase aggressive behavior in honeybees and fruit flies (Li-Byarlay et al. 2014). Results based on the Sinocyclocheilus cavefish and surface fish thus support the preceding finding that enhanced neural oxidative phosphorylation is negatively associated with aggressive behavior. Changes in whole-brain transcriptome and lipidome between cavefish and surface fish. (A) Schematic illustration on river drainages near the collection sites of cavefish ( Sinocyclocheilus anopthalmus) (orange mark) and surface fish (S. angustiporus) (green marks) near the city of Kunming, Yunnan province. Scale bar: 1 cm. (B) Enrichment plot illustrates top dysregulated pathways based on whole-brain transcriptomics analysis from Meng et al. (2013, 2018) on the brains of cavefish relative to surface fish. Gene ratio denotes the number of differentially expressed genes relative to the total number of genes under the specific term. (C) Principal component analysis based on whole-brain lipidome of surface fish and cavefish. n = 5 biological replicates for each fish species. (D) Hierarchical clustering heatmap illustrates patterns of changes in major lipid classes between the brains of cavefish and surface fish. Box indicates complex glycosphingolipids including sulfatides (SL), glucosylceramide (GluCer), and galactosylceramide (GalCer) were appreciably upregulated in the brains of surface fish. n = 5 biological replicates for each fish species. (E) Chord diagrams show the changes in lipid correlations in the brain of cavefish and surface fish, respectively. Lipid correlations were calculated by Spearman correlation analyses with cutoffs in correlation coefficients at ≥0.7 and P < 0.05. Band width indicates the number of correlations and color indicates direction of correlation (red: negative correlation; blue: positive correlation). n = 5 biological replicates for each fish species. To elucidate region-specific differences in lipid metabolism and spatial lipid distribution, we then systematically sectioned the whole-brain of each species transversely. The brain sections were classified according to their positions along the longitudinal axis into four major brain regions, namely, the telencephalon (Tel), tectum opticum (TeO), corpus cerebelli (CC), and medulla oblongata (MO). In earlier studies conducted in the Mexican Astyanax morphs, it was found that region-specific differences in serotonergic signaling determine the disparate behavior between surface fish and cavefish (Elipot et al. 2013; Rétaux and Elipot 2013). In teleost, serotonergic neurons are mainly found in the hindbrain raphe nucleus and in three hypothalamic nuclei located in the anterior brain, and the latter is absent in mammals (Elipot et al. 2013) (fig. 2A). A larger anterior paraventricular nucleus in cavefish results in enhanced hypothalamic serotonergic signaling, which was shown to drive foraging behavior characteristic of cave dwellers (Elipot et al. 2013). On the other hand, experience-dependent downregulation of raphe serotonergic signaling was observed in the dominant fish among a school of surface fish (Rétaux and Elipot 2013). In our brain-sectioning scheme, sections from the TeO region cut across both superior raphe and hypothalamic serotonergic projections, whereas sections from CC and MO regions comprise mainly inferior raphe serotonergic projections. Region-specific changes in lipidomes across four brain regions of cavefish and surface fish. Tel, telencephalon; TeO, tectum opticum; CC, corpus cerebelli; MO, medulla oblongata. (A) Schematic illustration of the lateral and ventral views of serotonergic neuron populations (regions colored yellow and red) in the brain of teleosts adapted from Prasad et al. (2015) and Elipot et al. (2013). Four distinct regions laterally span the teleost’s brain, which includes the telencephalon (Tel), the midbrain region largely covered by the tectum opticum (TeO), the hindbrain/brain stem region rostrally covered by the cerebellum (CC), and the medulla oblongata (MO) that grades into the spinal cord. The major serotonergic populations in the brain of teleost include the hypothalamic populations (yellow) and raphe populations (red). Most teleosts preserve two l-tryptophan hydroxylases (TPHs) for biosynthesis of serotonin from l-tryptophan. The raphe serotonergic populations (red) utilize TPH2 for serotonin production, whereas the hypothalamic populations rely on TPH1 (yellow). The superior raphe nuclei extend into the forebrain and midbrain, whereas the inferior raphe populations project into the hindbrain-spinal cord region (MO). (B) Volcano plots display top lipids that were most significantly different in each pairwise comparison between surface fish and cavefish brain sections from each brain region, based on magnitudes of P-value and fold-changes. Two-sided Welch’s t-test was used for pairwise comparisons. n = 3–4 brain sections from three biological replicates for Tel, TeO, and CC; n = 2 brain sections from three biological replicates for MO. (C) Hierarchical clustering was performed to aggregate lipid classes exhibiting comparable patterns of change across four regions (Tel, TeO, CC, and MO) between the brains of cavefish and surface fish. Patterns were visually examined and three major clusters were selected and expanded on the right, which included arachidonyl (ARA)-phospholipids, acylcarnitines, and sulfatides (SL). n = 3–4 brain sections from three biological replicates for Tel, TeO, and CC, n = 2 brain sections from three biological replicates for MO. Please refer to supplementary table S4, Supplementary Material online for the comprehensive list of lipids ordered by hierarchical clustering (arranged from top to bottom of the heatmap). (D) Barplots on changes in total acylcarnitine, sulfatides (SL), and docosahexaenoic acid (FFA 22:6) across Tel, TeO, CC, and MO in cavefish versus surface fish. P-values from two-sided Welch’s t-test for each pairwise comparison were illustrated. n = 3–4 brain sections from three biological replicates for Tel, TeO, and CC, n = 2 brain sections from three biological replicates for MO. Quantitative lipidomics revealed reductions in membrane phospholipids containing ARAs in surface fish relative to cavefish consistently across all four brain regions (fig. 2B and C and supplementary table S3, Supplementary Material online). Enriched ARA-phospholipids in cavefish brain was in agreement with GSEA analysis based on whole-brain transcriptome, indicating an enhanced ARA metabolism (fig. 1B). In contrast, other lipid classes exhibited region-specific changes that became evident only from region-specific lipidomics. For example, acylcarnitine levels in TeO and CC of cavefish were markedly elevated compared with surface fish (fig. 2C and D). Since acylcarnitines supply fatty acids to the mitochondria for oxidative phosphorylation (Jones et al. 2010), their higher levels in TeO and CC of cavefish suggest that these denote the primary brain regions adapted to enhanced oxidative phosphorylation compared with surface fish. Of interest, we noticed that the levels of free docosahexaenoic acids (DHAs), contrary to ARA, were significantly reduced in TeO and CC of cavefish compared with surface fish (Fig. 2D). Major DHA-phospholipids including PC 38:6, PC 40:6, and PE 40:6 were also reduced in the TeO and CC regions of cavefish relative to surface fish (supplementary fig. 2, Supplementary Material online). In accordance with results from quantitative lipidomics, spatial MS-imaging (MSI) of brain sections illustrated that ARA-phospholipids, such as PC 36:4 and PC 38:4 (phosphatidylcholines, PCs; see table 1 for a comprehensive list of lipid name abbreviations used in-text), were consistently increased in cavefish sections from TeO, CC, and MO regions (Fig. 3). On the other hand, DHA-phospholipids, such as PC 40:6, PC 38:6, and PE 40:6, were markedly reduced in cavefish compared with surface fish. The enrichment in DHA-phospholipids were more pronounced in surface fish TeO sections, and localized increases in DHA-phospholipids were particularly evident in the medial division of the valvula cerebelli that comprises Purkinje cells (Wullimann et al. 1996). Indeed, knockout of a member of the major facilitator superfamily, Mfsd2a, demonstrated to be a transporter of DHAs across the blood–brain barrier into the brain of mice, led to a significant loss of Purkinje cells in the cerebellum (Nguyen et al. 2014) (fig. 3). DHAs, therefore, are likely critical to cerebellar functions in surface fish that might have diminished importance in the evolution of cavefish. The enrichment in DHA-phospholipids extended to the eye and liver of surface fish relative to cavefish, with prominent clusters of DHA-PCs being present at higher levels in surface fish (fig. 4C). As for ARA-phospholipids, we noticed that their enrichment in cavefish was localized at the periventricular gray zone of the degenerated optic lobes (Wullimann et al. 1996) (fig. 3). The preferential accumulations of ARA- over DHA-phospholipids were validated also in the whole-eyes and whole-liver samples of cavefish relative to their surface-dwelling counterparts (fig. 4 and supplementary tables S5 and S6, Supplementary Material online). ARA-containing phospholipids, including phosphatidylserine PS 38:4(18:0_20:4) and phosphatidylinositol PI 36:4(16:0_20:4), were among the top significantly different lipids found in higher levels in cavefish eyes (CE) relative to surface fish eyes (SE) (fig. 4A). Distinct enrichment in clusters of ARA-PCs was noted in both the eye and liver of cavefish relative to surface fish (fig. 4C). Previous investigation conducted in zebrafish identified free ARAs released by cytoplasmic phospholipase A2 (cPLA2) cleavage of ARA-phospholipids as modulators of retinotectal arbor growth dynamics. In particular, inhibition of tectal cPLA2 (i.e., increased esterified ARAs) impedes the growth of retinal axonal arbor leading to retinotectal unsharpening (Leu and Schmidt 2008). Thus, the selective sequestration of ARAs in membrane phospholipids in cavefish may drive neuroplasticity underlying the loss of vision in the evolution of cavefish, since a neural map of an external environment of perpetual darkness is no longer essential to survival. Furthermore, the degeneration of the eyes and loss of vision also conserve energy to maximize survival in cave environment. In addition, drugs targeting the ARA-cascade, which inhibit the release of free ARAs from membrane phospholipids and lower the subsequent production of eicosanoids, have long been used as mood stabilizers in the treatment of human bipolar disorder (Rao et al. 2008). Sequestration of ARAs in neural membrane phospholipids of cavefish may, therefore, partly explain its attenuated aggressive behavior. Mass spectrometric imaging of spatial lipid distribution across four brain regions in cavefish and surface fish. Frozen fish brain tissues were sectioned at 10 µm thickness, and images were acquired using a Bruker solariX mass spectrometer equipped with a 9.4 T superconducting magnet operating in the positive ion mode. The intensity of colors corresponds to lipid abundances as illustrated by the intensity bars. Scale bar: 5 mm. The m/z of protonated parent ions illustrated were used for data acquisition, unless otherwise indicated. M + K refer to potassium adducts. Tel, telencephalon; TeO, tectum opticum; CC, corpus cerebelli; MO, medulla oblongata. Lipidome changes in the whole-eye and whole-liver of cavefish and surface fish. (A) and (B) Differential metabolite plots illustrates top significantly different lipids between the eyes and livers of surface fish relative to cavefish. Lipids were ranked by P-value and log2(fold-changes) were plotted. In the eye, the levels of arachidonyl (20:4)-phospholipids such as PC40:6(20:4_20:2) and PS 38:4(18:0_20:4) were significantly higher in cavefish than surface fish. In the liver, total TAG and plasmalogen PC were significantly higher in cavefish than surface fish. SE, surface fish eye; CE, cavefish eye; SL, surface fish liver; CL, cavefish liver; PC, phosphatidylcholines; PS, phosphatidylserines; TAG, triacylglycerols. P-values from Welch’s t-tests were illustrated. n = 2 technical replicates from three cavefish and surface fish, respectively. (C) Heatmaps with hierarchical clustering of lipid species were plotted for the classes of PC and plasmalogen PC (PCp) in the eye and liver of cavefish and surface fish. Notable clusters of PCs containing arachidonic acids (C20:4; ARAs) and docosahexaenoic acids (C22:6, DHAs) were boxed up in red and blue, respectively. DHA-clusters were enriched in the surface fish eye and liver, whereas ARA-clusters were enriched in the cavefish eye and liver. In addition, most PCps were increased in the cavefish liver relative to the surface fish liver. n = 2 technical replicates from three cavefish and surface fish, respectively. SE, surface fish eye; CE, cavefish eye; SL, surface fish liver; CL, cavefish liver; PC, phosphatidylcholines; PCp, plasmalogen phosphatidylcholines. List of Lipid Name Abbreviations. Besides a differential accumulation of ARA- relative to DHA-phospholipids, we also observed a notable accumulation of neutral lipids triacylglycerols (TAGs) in the liver of cavefish relative to surface fish. In addition to an enhanced level of storage TAGs, cavefish liver also displayed elevated levels of plasmalogen phosphatidylcholines (PCp). The preferential accumulation of energy-dense lipids particularly in the adipose tissues of cavefish was previously reported by others (Xiong et al. 2018; Minchin 2020), which was postulated to serve as an “energy insurance” to buffer against episodes of food deprivation in caves. Cavefish were found to possess elevated body fat that enables them to sustain extended period of nutrient scarcity. Because of a lack of photosynthesis-driven primary producers in caves, cave dwellers experience prolonged nutrient limitation and are largely dependent on energy inputs from external sources including seasonal floods and bat droppings to thrive (Wilkens et al. 2000). In agreement with our observation on Sinocyclocheilus cavefish, fatty livers were also previously observed in the Tinaja cave form of Astyana mexicanus (Riddle et al. 2018). Based on region-specific lipidomics, we noted that sulfatides (SLs) were significantly elevated in the TeO, CC, and MO regions of surface fish relative to cavefish (fig. 2B–D). SL, together with its metabolic precursor galactosylceramide (GalCer), constitutes the predominant lipid components that ensure the normal structural and functional attributes of myelin sheath (Coetzee et al. 1996). MSI revealed that SL-enrichment (SL t42:1, SL t44:1) in surface fish TeO (fig. 3) was localized at the posterior tuberculum abundant in neuronal projections from the superior raphe serotonergic populations (Lillesaar 2011). In contrast, the dorsal zone of the periventricular hypothalamus of TeO sections that transverse the hypothalamic serotonergic neurons carried no observable signals of SLs but were, instead, enriched in DHA-phospholipids (PE 40:6, PC 40:6, and PC 38:6) (fig. 3). Similarly, SL-enrichment was observed in the area of the intermediate reticular formation in CC and MO sections from surface fish relative to cavefish (fig. 3), which are rich in projections from the inferior raphe serotonergic neurons (Lillesaar et al. 2009). Immunostaining of myelin distribution, that is, red fluorescent signals from myelin basic protein (supplementary fig. 3, Supplementary Material online) was in agreement with MSI data. Myelin distribution spatially overlapped with SL signals in regions corresponding to the inferior raphe serotonergic neurons in CC and the superior raphe serotonergic neurons in the TeO but not in the dorsal periventribular hypothalamus containing hypothalamic serotonergic neurons (supplementary fig. 3, Supplementary Material online). We then validated our observations based on lipidomics and immunostaining by imaging axon myelination using transmission electron microscopy (TEM) (fig. 5). We selected the hindbrain CC region for TEM because this region contains predominantly inferior raphe serotonergic innervations but not the hypothalamic serotonergic innervations. Representative TEM images from CC region revealed a greater number of myelin sheaths around individual axons captured in the visual field (fig. 5A), as well as a greater thickness of myelin sheaths around larger axons (fig. 5B and C and supplementary table S7, Supplementary Material online), in surface fish relative to cavefish. The g-ratio, given by the ratio of axon diameter to the summed diameter of axon and its surrounding myelin sheath, is an indicator of the degree of myelination in myelinated axons. A smaller g-ratio indicates a higher degree of myelination (Buckley, et al. 2010; Niu, et al. 2021). A plot of g-ratios against axon diameters revealed that axons with larger diameters possessed significantly thicker myelination (i.e., smaller g-ratios) in the CC region of surface fish compared with cavefish (fig. 5C). Based on a combination of MSI, immunostaining and TEM images examining the extent of myelination, therefore, it appears that the raphe serotonergic neuronal populations, predominantly in the hindbrain region, were more heavily myelinated in Sinocyclocheilus surface fish relative to cavefish. The raphe serotonergic neurons in cavefish possibly underwent appreciable demyelination in the course of their evolution. Immunostaining also revealed that cavefish possessed higher levels of 5-HT receptor 4 (5-HTR4) (green fluorescence) in both TeO and CC regions compared with surface fish (fig. 5D), indicating enhanced 5-HT signaling in cavefish was in agreement with previous report (Elipot et al. 2013). The dorsal periventricular hypothalamus exhibited appreciable 5-HTR4 signals indicating the presence of hypothalamic serotonergic populations in both fish species (fig. 5D), unlike the posterior tuberculum; however, myelin signal was absent in this region (supplementary fig. 3A, Supplementary Material online). Our results cumulatively suggest that relative to surface fish, cavefish undergo demyelination specifically in their raphe serotonergic neurons but not hypothalamic serotonergic neurons. Demyelination in raphe serotonergic neurons in cavefish. (A) TEM images taken from the CC region of the brain of cavefish and surface fish. Scale bar = 1 µm for the top panel and 0.5 µm for the bottom panel. A greater number of myelin sheaths around individual axons and thicker myelin sheath wrapping around larger axons were observed in the CC region of surface fish relative to cavefish. Representative images from two to three independent experiments were shown. n = 20 sections from two cavefish and three surface fish, respectively. CC, corpus cerebelli. (B) Barplot illustrates the number of myelin sheaths captured per section taken from the CC region of cavefish (CAV-CC) and surface fish (SUR-CC), respectively. n = 20 sections from two cavefish and three surface fish, respectively. P-value from Welch’s t-test was indicated. CC, corpus cerebelli. (C) A plot of g-ratios against axon diameters captured in TEM images taken from the CC region of cavefish and surface fish. n = 88 axons from two cavefish and n = 132 axons from three surface fish, respectively. P-value from Chow’s test was used to test the statistical significance between the true coefficients in the two linear regressions constructed based on data sets from the CC regions of cavefish (CAV-CC) and surface fish (SUR-CC). CC, corpus cerebelli. (D) Confocal images on the distribution of serotonin receptor 4 (5-HTR4) in the TeO and CC regions of cavefish and surface fish. Representative images were from two independent experiments. Primary scale bar: 2 mm; inset scale bar: 200 µm. TeO, tectum opticum; CC, corpus cerebelli. In our lipidomic investigation of the brain, the eye and liver of cavefish versus surface fish, we discovered four essential differences in lipid metabolism between the two fish species, which might relate to the differential selection pressures of their distinct habitats. Relative to surface fish, cavefish exhibited 1) enhanced oxidative phosphorylation in the brain; 2) preferential accumulation of DHA-phospholipids over ARA-phospholipids in the brain, eye, and liver; 3) accumulation of fat (in the form of storage TAGs) and plasmalogens PC in the liver; and (4) selective demyelination of raphe serotonergic neurons in its hindbrain. To elucidate the candidate genes underlying the differential lipid metabolism between cavefish and surface fish, we examined the relative expressions of several genes along the pathways of DHA biosynthesis, uptake and phospholipid remodeling, fat mobilization, mitochondrial and peroxisomal β-oxidation, as well as plamalogen biosynthesis in the brain and liver (fig. 6, supplementary table S8, Supplementary Material online). DHA biosynthesis in teleosts can proceed either via the Δ4 desaturase pathway or the Sprecher pathway (Oboh et al. 2017; Matsushita et al. 2020). The biosynthesis of DHAs involves fatty acid desaturases (fads) and elongation of very-long-chain fatty acid (Elovl) proteins. Unlike mammals that carry both fads1 and fads2 genes, virtually all teleosts possess only the fads2 gene but with varying copy numbers. fads2 has acquired diverse functions during the process of teleost evolution and is able to introduce double bonds at different positions along the fatty acyl chains (Xie et al. 2021). Elovls catalyze the two-carbon elongation of preexisting fatty acyl moieties. Elovl2 displays substrate preference for C20 and C22 polyunsaturated fatty acids (PUFAs), and plays a significant role in the biosynthesis of DHAs via the Sprecher pathway (Monroig et al. 2018). Our qRT-PCR showed that in the brain, fads2 expression was significantly higher in surface fish relative to cavefish but not for other genes along the Sprecher pathway (elovl2, elovl5, and acyl-coenzyme A oxidase 1: acox1). In contrast, while the levels of fads2 in the liver were not significantly different between cavefish and surface fish, other genes of the Sprecher pathway (elovl2, elovl5, and acox1) were significantly elevated in the liver of surface fish compared with cavefish. Based on our results, it appears that surface fish increases DHA biosynthesis in the brain via the Δ4 desaturase pathway but instead relies on the Sprecher pathway to enhance DHA biosynthesis in the liver. In addition, the level of mfsd2a in the brain, which mediates DHA uptake from the systemic circulation into the brain (Nguyen et al. 2014), was not significantly different between cavefish and surface fish. Hence, surface fish likely retains a local, neurological supply of DHAs to sustain critical brain functions, instead of relying on DHA uptake from the systemic circulation. Indeed, it was reported that among different tissues, the teleost brain possesses the highest expression of fads2, and this enriched pattern of expression in marine fish brain serves to retain a local, functional Δ6 desaturase to ensure a sufficient supply of DHAs to neural tissues (Monroig et al. 2018). The enhanced biosynthesis of DHAs was particularly evident for the TeO region of the surface fish’s brain, which may be related to the essential roles of DHAs in facilitating advanced functions of photoreceptors in the eye (Crawford et al. 1999). In addition to augmented local biosynthesis of DHAs, increased incorporation of DHAs into membrane phospholipids via remodeling processes were also evident in the brain. The relative expressions of lysophospholipid acyltransferases with substrate specificity for PUFAs, including membrane bound-O-acyltransferase domain containing 7 (mboat7) and lysophosphatidylcholine acyltransferase 3 (lpcat3) (Hishikawa et al. 2008; Zarini et al. 2014), were significantly elevated in surface fish particularly in the TeO region (fig. 6). mRNA expression of genes in the brain and liver involved in the metabolic adaptations of cavefish and surface fish to their distinct habitats. Relative expression values of genes were normalized to those of β-actin. Barplots display changes in the relative expression of genes in the whole-brain or whole-liver of cavefish and surface fish. Boxplots illustrate changes in relative expression of genes in cavefish and surface fish across the four regions of the brains. For barplots, n = 3 technical replicates from one cavefish and two surface fish, respectively. For boxplots, n = 3–4 brain sections from three biological replicates for Tel, TeO, and CC, n = 2 brain sections from three biological replicates for MO. In barplots, means ± SEM were plotted. In boxplots, the median is indicated by the horizontal line and the first and third quartiles were represented by the box edges. The lower and upper whiskers extend from the hinges to the smallest and largest values, respectively, with individual samples indicated as dots. In all plots, P-values from Welch’s t-test were indicated. fads2: fatty acid desaturase 2; elovl2: elongation of very-long-chain fatty acids 2; elovl5: elongation of very-long-chain fatty acids 5; acox1: acyl-coenzyme A oxidase 1; mfsd2a: member of the major facilitator superfamily 2a; cpt1ab: carnitine-palmitoyltransferase 1a or 1b; hadhaa: hydroxyl-coenzyme A dehydrogenase alpha subunit; mboat7: membrane bound-O-acyltransferase domain containing 7; lpcat3: lysophosphatidylcholine acyltransferase 3; pnpla2: patatin-like phospholipase domain containing 2; lipea: lipase, hormone-sensitive a; lipeb: lipase, hormone-sensitive b; dagla: diacylglycerol lipase, alpha; mgll: monoglyceride lipase; gnpat2: glyceronephosphate O-acyltransferase 2; cpt2: carnitine-palmitoyltransferase II; decr1: 2,4-dienoyl CoA reductase 1, mitochondrial. Tel, telencephalon; TeO, tectum opticum; CC, corpus cerebelli; MO, medulla oblongata. Corroborating our lipidomic observations, cavefish brain exhibited marginally elevated expressions of genes mediating mitochondria β-oxidation, such as the carnitine-palmitoyltransferase 1a or 1b (cpt1ab) and hydroxyl-coenzyme A dehydrogenase alpha subunit (hadhaa) compared with surface fish. Contrary to the brain, the liver of surface fish displayed markedly elevated levels of genes involved in fat mobilization (pnpla2: patatin-like phospholipase domain containing 2; lipea: lipase, hormone-sensitive a; lipeb: lipase, hormone-sensitive b; dagla: diacylglycerol lipase, alpha; mgll: monoglyceride lipase that mediate the breakdown of TAGs to diacylglycerols DAGs to monoacylglycerols MAGs and, finally, to free fatty acyls), as well as those involved in mitochondrial β-oxidation (cpt1ab, hadhaa, cpt2: carnitine-palmitoyltransferase II, and decr1: 2,4-dienoyl CoA reductase 1) and peroxisomal β-oxidation (acox1) (fig. 6). Thus, fatty livers in cavefish might have resulted from attenuated mobilization and oxidation of fatty acyls compared with surface fish. In addition, cavefish liver also exhibited enhanced expression of glyceronephosphate O-acyltransferase 2 (gnpat2) (fig. 6), the rate-limiting enzyme in peroxisomal matrix governing plasmalogens biosynthesis (Brites et al. 2004), in agreement with the observed increases in PCps in cavefish liver. We also attempted to investigate the expressions of genes mediating SL biosynthesis and breakdown in the hindbrain regions of Sinocyclocheilus, including galactose 3-O-sulfotransferase 4 (gal3st4) and arylsulfatase A (arsa), but the levels of gene expression were too low to render reliable quantitative comparisons between cavefish and surface fish. Other yet unknown genes candidates might possibly regulate the changes in myelination observed in the hindbrains of Sinocyclocheilus that warrant further investigation in future studies. As cavefish thrive in nutrient-poor environments under perpetual darkness, obtaining food becomes a top priority that determines its survival (Xiong et al. 2018). In the process of adapting to their cave habitats, for example, the Mexican Astyanax has evolved various behavioral changes including the loss of aggression and schooling and swim randomly to facilitate foraging and maximize their chances of coming across food (Elipot et al. 2013). As a result, multiple independent cave species in physically separated caves evolved similar behavioral and morphological traits. Molecular mechanisms underlying such parallelism or convergence in evolution have remained largely unclear (Gross 2012). Our lipid-centric investigation has identified 1) enhanced oxidative phosphorylation in cavefish brain, 2) reduced biosynthesis of DHAs in the brain, eye, and liver of cavefish, 3) accumulation of fat and plasmalogens in cavefish liver, as well as 4) the selective demyelination of raphe serotonergic neurons in cavefish as the prominent lipid pathways possibly underlying behavioral adaptations to caves over freshwater environments. Region-specific quantitative lipidomics and spatial MSI revealed reduced DHA biosynthesis and incorporation into membrane phospholipids in cavefish brains relative to surface fish, which led us to postulate that DHAs likely facilitate cerebellar functions in surface fish that are no longer pivotal to survival in cavefish. In freshwater ecosystems, schooling behavior reduces the risk of predation (Magurran 1990). DHA-enriched diet was previously shown in larval yellowtail Seriola to affect the volumetric growth of TeO and CC regions important to visual acuity and swimming performance. In particular, the onset of schooling behavior was closely matched with the time-point at which significant volumetric expansion in TeO and CC was observed (Ishizaki et al. 2000). Similarly in human cohorts, increased maternal DHA intake from seafood consumption was positively associated with improvements in prosocial behavior and fine motor skills of their children (Hibbeln et al. 2007). The loss of DHA-enriched neural domains might also be partly explained by vision loss in cavefish driven by an environment of total darkness, since DHA represents the PUFA uniquely utilized by photoreceptors (Crawford et al. 1999). We then validated via qRT-PCR that the specific reductions in membrane DHAs in the brain of cavefish were attributed to reduction in DHA biosynthetic capacity via fads2 along the Δ4 desaturase pathway (fig. 7) coupled with a diminished remodeling of polyunsaturated DHAs into membrane phospholipids mediated by mboat7 and lpcat3. Indeed, fads2 was identified as a key metabolic gene that increases survival on DHA-deficient diets within freshwater ecosystems (Twining et al. 2016)—a pivotal factor to the colonization of freshwater habitats in fishes (Ishikawa et al. 2019). Kitano and colleagues demonstrated that additional copy number of fads2 in Pacific Ocean sticklebacks contributes to survivorship on DHA-deficient freshwater habitats and that freshwater stickleback populations with longer evolution history in freshwater are associated with higher fads2 copy numbers (Ishikawa et al. 2019). Sinocyclocheilus ancestors may have acquired enhanced biosynthetic capacity of DHAs during the course of evolution. The elevated incorporation of DHAs into neural and optic membrane lipids enhance vision and facilitate neural execution of complex behavioral traits, such as schooling, important to the survival of surface fish in freshwater habitats. In comparison, food scarcity replaces predation as the major determinant of survivorship in cave habitats; coupled with the total darkness in the cave environments that S. anophthalmus reside, the maintenance of vision and complex social behavior such as schooling diminished in importance relative to minimizing energy expenditure. Schematic diagram summarizes major pathways in the brain and liver of cavefish and surface fish underlying metabolic adaptations to their distinct environments as revealed by lipidomics and qRT-PCR. The cave environment entails a solitary life marked by food scarcity and darkness under which resource utilization and minimizing energy expenditure become the major determinants of extended survival. Surface fish, in comparison, is subjected to an environment of light and constant predation and developed social behaviors including schooling and aggressive dominance to reduce risk of predation. Our lipidomics and qRT-PCR supported reduced DHA biosynthesis along the Δ4 desaturase pathway and enhanced oxidative phosphorylation in the brains of cavefish as the prominent neurological pathways contributing to troglomorphic adaptations. Outside the brain, a global reduction in DHA biosynthesis along the Sprecher pathway coupled with attenuated mobilization of fat storage and reduced β-oxidation in the mitochondria and peroxisomes denotes key metabolic changes underlying troglomorphic adaptions that primarily serve to reduce energy expenditure. fads2: fatty acid desaturase 2; elovl2: elongation of very-long-chain fatty acids 2; elovl5: elongation of very-long-chain fatty acids 5; acox1: acyl-coenzyme A oxidase 1; mfsd2a: member of the major facilitator superfamily 2a; cpt1ab: carnitine-palmitoyltransferase 1a or 1b; hadhaa: hydroxyl-coenzyme A dehydrogenase alpha subunit; mboat7: membrane bound-O-acyltransferase domain containing 7; lpcat3: lysophosphatidylcholine acyltransferase 3; pnpla2: patatin-like phospholipase domain containing 2; lipea: lipase, hormone-sensitive a; lipeb: lipase, hormone-sensitive b; dagla: diacylglycerol lipase, alpha; mgll: monoglyceride lipase; gnpat2: glyceronephosphate O-acyltransferase 2; cpt2: carnitine-palmitoyltransferase II; decr1: 2,4-dienoyl CoA reductase 1, mitochondrial. In addition to the brain, cavefish also experienced similar reductions in DHA lipids in its liver—a key organ governing systemic lipid metabolism. In contrast to the brain, the reduction of DHA biosynthesis in cavefish liver results from an attenuated flow along the Sprecher pathway (fig. 7). This observation suggests that the loss of DHA biosynthetic capacity probably has evolutionary significance outside of the brain. The membrane pacemaker theory of metabolism proposes the central role of DHA as a major contributor to the degree of cellular membrane polyunsaturation, which, in turn, governs the molecular activities of membrane ion pumps that determine the basal metabolic rate of an organism (Hulbert 2007). For example, a strong positive correlation was observed between the DHA content of phospholipids and the heart rate of a diverse group of mammals ranging from mice to whales (de Duve and Hayalshl 1978). DHA content of membranes was found to increase the molecular activity of sodium/potassium ATPase pumps that denote a major contributor to basal metabolic rate (Turner et al. 2003). Cavefish are known to possess lower energy expenditure and requirement relative to surface fish, in order to withstand prolonged period of nutrient deprivation in an environment with infrequent food supply. For example, the degeneration of the eyes, which have high energy requirement, allows cavefish to conserve up to 15% of energy (Krishnan and Rohner 2017), whereas the loss of a functional circadian rhythm saves another 27% of energy (Moran et al. 2014). Thus, based on the membrane pacemaker theory of cellular metabolism, a global reduction in membrane DHAs in Sinocyclocheilus cavefish may serve to lower basal metabolic rate and conserve energy to maximize survival in a nutrient-limiting environment. On a similar note, we also observed that cavefish greatly elevate fat storage in the liver, possibly via suppressed breakdown of storage TAGs into fatty acyl constituents and reduced oxidation of fatty acyls through the mitochondria and peroxisomes (fig. 7). The reductions in fatty acyls oxidation are in agreement with a lower energy requirement and reduced basal metabolic rate. Cavefish was postulated to have evolved compensatory mechanisms that enable them to remain physiologically healthy, with comparable lifespan and without appreciable accumulation of advanced glycation end-products (AGEs) compared with surface fish; in spite of deleterious diabetes-related pathologies characterized by excess fat accumulation (i.e., fatty livers), insulin resistance and hyperglycemia that serve to cope with irregular food supply (Riddle et al. 2018). In this aspect, it is worthy to note that plasma levels of plasmalogen PCs were negatively correlated with the abundances of AGEs in human patients (Zhang et al. 2017). Plasmalogens, by virtue of their vinyl ether linkages, can counter the generation of reactive oxygen species induced by AGEs (Braverman and Moser 2012). It remains an interesting question whether the upregulation of plasmalogen biosynthesis denotes a metabolic adaptation to cope with excess fat accumulation in cavefish. Finally, a selective demyelination of raphe serotonergic neurons might be attributed to social isolation of cavefish ancestors and a loss in requirement for neural plasticity that shapes complex social behavior, since it can be potentially energetically costly to maintain oligodendrocytes needed for myelination (Harris and Attwell 2012). Indeed, demyelination had been identified as a contributing factor of cognitive impairment resulting from social isolation in humans and rodent models (Baraban et al. 2016). Our observation of enhanced myelination specifically at raphe serotonergic neurons in surface fish provides a plausible explanation on why only surface fish, but not cavefish, are able to execute experience-dependent downregulation of raphe serotonergic signaling in establishing social dominance. Indeed, preliminary results from our resident-intruder assay indicate that surface fish exhibit a markedly higher level of aggressive dominance compared with cavefish (supplementary fig. S4, Supplementary Material online). Regulation of myelination based on social experience can alter the intensity of raphe serotonergic signaling. Experience-dependent shaping of myelination to achieve neural plasticity had also been demonstrated in adult human brains (Tomassy et al. 2016). Our current work, however, did not identify the gene candidates central to the regulation of raphe serotonergic neuron myelination in Sinocyclocheilus. Through a combination of quantitative lipidomics with spatial MSI, we revealed that neural lipid metabolic plasticity, in particular enhanced oxidative phosphorylation, reduced DHA biosynthesis and membrane incorporation, and demyelination of raphe serotonergic neurons might contribute to neuroplasticity underlying troglomorphic behavioral adaptations, that is, loss of aggressive dominance and schooling in the evolution of cavefish. Given that the gross morphological anatomy of the brains between the two fish species are largely conserved apart from degeneration of the optic lobes in cave-dwelling fish, our study illustrates how regional compartmentalization in metabolism and distribution of lipids can alter neural plasticity to modulate behavior during the course of evolution. On top of neurological adaptations, we uncovered that systemic metabolic adaptations, such as the development of fatty livers, global reductions in membrane DHAs, and the accretion of liver plasmalogens, also contribute to troglomorphic adaptations to a cave habitat. The central theme of troglomorphic adaptations entails the loss of nonessential morphological and behavioral traits (regressive evolution) in order conserve energy in an environment of limiting, irregular food supply, as proper resource allocation becomes the major determinant of survival in a cave habitat. The selection pressure for these troglomorphic adaptations in turn drives parallel or convergent evolution in cave dwellers. Notwithstanding these findings, our study has numerous limitations. First, our study results do not allow the interpretation of causality. Although the investigated genes may underlie changes in lipid metabolic pathways that are differentially regulated between cavefish and surface fish, functional studies and the generation of genetic knockouts are imperative to determine whether these genes cause the metabolic changes that contribute to the evolution of troglomorphic traits in Sinocyclocheilus. Such experimental validations are currently infeasible in Sinocyclocheilus, largely circumscribed by the rarity of this genus, long growth term, and the recent classification of Sinocyclocheilus as second-class protected animals in China. We also failed to elucidate the key genes regulating the myelination of raphe serotonergic neurons in Sinocyclocheilus, owing to the low expression levels of our candidate genes investigated (gal3st4, arsa). A functional characterization of fads2 in different tissues of Sinocyclocheilus, for example, using heterologous expression system in yeast (Matsushita et al. 2020), which can help determine whether qualitative changes in fatty acid metabolic enzymes in addition to gene expressions may modulate the biosynthetic capacity of DHAs, was not conducted in this study. In addition, it will be meaningful to investigate in future studies the upstream evolutionary genetic basis for the altered expressions of genes underlying the differences in behavioral and morphological adaptations between the two fish species reported in this study. Collection sites for cavefish (S. anopthalmus) and surface fish (S. angustiporus) were near the city of Kunming, Yunnan province (fig. 1A). The orange mark indicates collection location of cavefish at Jiuxiang cave N 25.05478°, E 103.37975°. Surface fish were collected from Huangnihe River in Agang Town (green marks), Luoping, at N 25.00905°, E 103.59256°. After collection, fish were allowed to recover overnight in laboratory tanks to minimize metabolic distress due to transport, then euthanized with 0.05% tricaine methanesulfonate, and dissected within 2 days with no feeding in between. The dorsal surface of the head was dissected away to expose the brain, the fresh brains were carefully dissected from the head and frozen on the dry ice for MSI. For immunohistochemistry, the brains were fixed in 4% paraformaldehyde overnight at 4 °C and equilibrated in 30% sucrose prior to cross-cryosectioning. All experimental procedures involving animals were conducted and approved by the Animal Care and Use Committee of the Institute of Zoology, Chinese Academy of Sciences (approved protocol: IOZ18002). Frozen fish brain tissues were fixed in position on the cutting stage. All tissues were sectioned at 10 µm thickness using a Leica CM1950 cryostat (Leica Microsystems GmbH, Wetzlar, Germany) at −18 °C and mounted onto indium tin oxide-coated glass slides (Type I 0.7 mm/100ea, HST Inc., Newark, NJ, USA). The slide was put into a vacuum desiccator and dried for approximately 1 h. For analysis of lipids, a mixture (250 µl) of AgNPs and DHB (1 mg/ml AgNPs and 10 mg/ml DHB in ACN/H2O 8:2) solution was sprayed on the tissue section using an in-house electrospray-based matrix deposition device and at a solvent flow rate of 800 µl/h (Guan et al. 2018). The glass slides were dried in the vacuum desiccator for approximately 1 h prior to analysis on MALDI MS-Imaging (MSI). Lipids were extracted from whole-brains/brain sections using a modified Bligh and Dyer’s protocol as previously described (Lam et al. 2021). Tissues were homogenized in extraction solvent, that is, chloroform: methanol 1:2 (v/v) containing 10% MilliQ water on a bead ruptor (OMNI, USA). Following incubation, 350 µl of MilliQ water and 250 µl of ice-cold chloroform were added to induce phase separation. Samples were then centrifuged at 16, 260 × g at 4 °C for 5 min. The lower organic phase containing lipids were transferred to a new tube. The extraction was repeated once via addition of another 500 µl of ice-cold chloroform to the remaining aqueous phase, and the extractions were pooled and dried in a SpeedVac under the organic mode. Dried lipid extracts were resuspended in chloroform: methanol (1:1) prior to LC-MS/MS analysis (Song et al. 2020). Analyses of polar lipids in whole-brain and brain sections were conducted on Exion-UPLC coupled with Sciex 6500 Plus QTRAP that runs on Analyst 1.6.3, whereas neutral lipids were analyzed on an Agilent 1260 HPLC connected to Sciex 5500 QTRAP. Analyses were conducted in the electrospray ionization mode, using the following source parameters, 5500: CUR 10, CAD High, TEM 350 °C, GS1 35, GS2 35; 6500: CUR 10, CAD High, TEM 400, GS1 20, GS2 20. Internal standard cocktail used for normalization of lipidome data in whole-brain and brain sections included DMPC, DMPE, PA-C17:0, d31-PS, DMPG, C14:0-BMP, SL d18:1/17:0, CL22:1(3)-14:1, LPC 17:0, LPS 17:1, LPA 17:0, LPE 17:1, SM d18:1/12:0, Cer d18:1/17:0, GluCer d18:1/8:0, d3-16:0-carnitine from Avanti Polar Lipids, PI-8:0/8:0 was purchased from Echelon Biosciences, whereas d31-FFA 16:0 and d8-FFA-20:4 were obtained from Sigma-Aldrich and Cayman Chemicals, respectively. Analyses of lipidomes from whole-eye and whole-liver samples of cavefish and surface fish were conducted on a Shimadzu Nexera X2 LC-30AD UPLC coupled with Sciex Triple Quad 7500. Analysis was performed in the electrospray ionization mode with the following source parameters: Ion Source Gas 1 35, Ion Source Gas 2 70, Curtain Gas 32, Temperature 450 °C, CAD Gas 9. Internal standard cocktail for normalization of the eye and liver lipidome data included d9-PC32:0(16:0/16:0), d9-PC36:1p(18:0p/18:1), d7-PE33:1(15:0/18:1), d9-PE36:1p(18:0p/18:1), d31-PS(16:0/18:1), d7-PG33:1(15:0/18:1), d7-PI33:1(15:0/18:1), d7PA33:1(15:0/18:1), C14-BMP, d8-SM d18:1/18:1, SL-d18:1/17:0, d7-LPC 18:1, d7-LPE 18:1, LPA-C17:0, LPI-C17:1, LPS-C17:1, LPG-C17:1, DAG(16:0/16:0)-d5, and DAG(18:1/18:1)-d5 obtained from Avanti Polar Lipids; TAG(14:0)3-d5, TAG(16:0)3-d5, TAG(18:0)3-d5, d6-CE18:0, and d6-Cho purchased from CDN isotopes; and d3-16:0-carnitine from Cambridge Isotope Laboratories. d31-FFA-16:0 from Sigma-Aldrich and d8-FFA-20:4 from Cayman Chemicals were used for quantitation of saturated/monounsaturated fatty acids and PUFAs, respectively. MALDI-FTICR mass spectrometric analysis of the tissue sections were performed using a Bruker solariX mass spectrometer equipped with a 9.4 T superconducting magnet (Li et al. 2016). Data were collected in the positive ion mode, in broadband over a mass range of 100–1,200 m/z with a resolution of 200,000 at m/z 200; mass calibrations were performed externally using sodium trifluoroacetate (NaTFA), using 150 shots per scan by a Smart Beam II laser operating at 150 Hz, a laser focus of 50 µm. For MALDI MSI analysis, the entire tissue section was analyzed averaging 1 scan per spectrum (per pixel) with fixed raster step size (Tel region: 70 µm; TeO region: 85 µm; CC region: 80 µm; MO region: 50 µm). All the data were processed using DataAnalysis 4.0 (Bruker Daltonics) and FlexImaging 3.0 sofware (Bruker Daltonics). Fish brain sections were incubated in PBS with 0.5% Triton-100 for permeabilizing. Antigens were unmasked by microwaving sections in 10 mM citrate buffer, pH 6.0, 5 min. Then blocked in 5% donkey serum in PBS for 1 h at room temperature and incubated overnight at 4 °C with the rabbit antiserotonin antibody (1:2000, Sigma No. S5545), the rabbit anti-5HT4 (1:100, Abcam No. ab60359), and rat antimyelin basic protein (1:200, Abcam No. 7349). Brain sections were washed three times with PBS for 5 min and incubated with the secondary antibodies (Alexa 488 conjugated or Alexa-568 conjugated, 1:1000, Thermo Fisher Scientific) at room temperature for 1 h. Images were required by Leica Aperio Versa 200. The corpus cerebelli (CC) region of the brain from cavefish and surface fish were dissected and immediately fixed with 2.5% glutaraldehyde overnight and postfixed with 1% osmium tetroxide for 1 h at 4°C. Samples were then stained with 3% uranyl acetate for 30 min at room temperature, washed with deionized water five times for 10 min each round, and dehydrated in a series of acetone treatments and infiltrated in embed-812 resin. The embedded tissues were cut into 70 nm slices and observed using a transmission electron microscope (TEM) (JEM 1400) at 80 kV. The number of myelin sheaths in individual TEM sections was counted, and g-ratios were calculated as the diameter of the axon divided by the diameter of the axon and its surrounding myelin sheath using ImageJ (Buckley et al. 2010; Lam et al. 2021). Total RNAs were extracted from brain sections using 1 ml of TRIzol reagent (Invitrogen, Cat No: 15596-026). For each sample preparation, 0.2 ml of chloroform was added and the aqueous phase was collected in fresh tubes and then mixed with 0.5 ml of isopropyl alcohol for pelleting RNA. To remove DNA contamination, RNA pellets were washed with 75% ethanol, digested with DNase for 30 min, and then pelleted via addition of sodium acetate and lithium chloride. RNA pellets were then washed in 75% ethanol and reconstituted in nuclease-free water. cDNA was synthesized using the iScript cDNA Synthesis Kit (Bio-Rad, Cat No 1708891). qRT-PCR was performed using the SYBR Green PCR kit (Bio-Rad, Cat No 1725120) on a Bio-Rad CFX Connect 384-Real-Time PCR Detection machine. Relative expression values of genes were normalized to those of β-actin. Gene-specific primers used for amplification were listed in table 2. Gene-specific Primers for qRT-PCR Analyses. To examine changes in lipidome composition, lipids in whole-brains were expressed in µmol/g dry mass, whereas lipids in individual brain sections were expressed in µmoles/section. To examine changes in membrane lipidomes, lipid levels in whole-livers and whole-eyes were expressed in molar fractions of total polar lipids, where total polar lipids denote the sum total of all phospholipids and sphingolipids detected. Hierarchical clustering using ward method (ward.D2 in R hclust) was performed on log2-transformed lipid levels in µmol/g dry mass. Heatmaps were drawn using ComplexHeatmap v2.7.10.9001. Clustering was visually evaluated and patterns of interests were manually selected. For analysis of lipid correlations in the brain of cavefish and surface fish, data were log-transformed and correlations between lipid pairs were calculated based on the Spearman correlation analysis using Hmisc v4.4-2. Correlation coefficient cutoff was set at ≥0.7 and P-value cutoff was set at P < 0.05. Correlations with corresponding P < 0.05 were visualized using chord diagrams drawn with chorddiag v0.1.2 and circlize v0.4.12, where bandwidth indicates number of correlations and color indicates direction of correlation. The blue shade indicates positive correlations, whereas the red shade indicates negative correlation between lipids. For brain sections, only two-group comparisons were made between brain sections of cavefish and surface fish obtained from the same brain region (i.e., Tel, TeO, CC, or MO) using Welch’s t-test. Fold change and −log10P-values were presented in volcano plots. For GSEA, a ranked gene list comparing whole-brain transcriptomes between cavefish and surface fish was obtained from Meng et al. (2013, 2018) (supplementary table S1, Supplementary Material online). GSEA was performed using gseGO function from R package clusterProfiler 3.14.3 with P-value cutoff set at <0.05. Statistical analyses were performed using R 4.0.2.
|
PMC8342504
|
Evaluation of two highly effective lipid-lowering therapies in subjects with acute myocardial infarction
|
For cardiovascular disease prevention, statins alone or combined with ezetimibe have been recommended to achieve low-density lipoprotein cholesterol targets, but their effects on other lipids are less reported. This study was designed to examine lipid changes in subjects with ST-segment elevation myocardial infarction (STEMI) after two highly effective lipid-lowering therapies. Twenty patients with STEMI were randomized to be treated with rosuvastatin 20 mg QD or simvastatin 40 mg combined with ezetimibe 10 mg QD for 30 days. Fasting blood samples were collected on the first day (D1) and after 30 days (D30). Lipidomic analysis was performed using the Lipidyzer platform. Similar classic lipid profile was obtained in both groups of lipid-lowering therapies. However, differences with the lipidomic analysis were observed between D30 and D1 for most of the analyzed classes. Differences were noted with lipid-lowering therapies for lipids such as FA, LPC, PC, PE, CE, Cer, and SM, notably in patients treated with rosuvastatin. Correlation studies between classic lipid profiles and lipidomic results showed different information. These findings seem relevant, due to the involvement of these lipid classes in crucial mechanisms of atherosclerosis, and may account for residual cardiovascular risk. Randomized clinical trial: ClinicalTrials.gov, NCT02428374, registered on 28/09/2014.Therapies aiming to reduce LDL-C changed substantially the natural history of cardiovascular disease (CVD), especially coronary heart disease (CHD). However, despite the achievement of very low levels of LDL-C, recurrent events after acute coronary syndromes are still observed, suggesting that other lipid components including fatty acids (FA) may contribute to the atherothrombotic disease. In addition, decreased plasma levels of plasmalogens were reported in subjects with CHD. In fact, their role in atherosclerosis is still poorly understood, involving possibly antioxidant properties that impact the movement of molecules in and out of the cells, as plasmalogens are components of cell membranes. High free FA concentrations can induce activation of NLRP3 inflammasome, triggering a pro-inflammatory response related to atherosclerosis, and lipid-lowering therapies, such as statin combined with the inhibitor of intestinal cholesterol absorption (simvastatin + ezetimibe), can partially revert many inflammatory biomarkers. Following rosuvastatin treatment, a lipidomic study revealed significant decrease in sphingomyelin (SM), triglycerides (TG), phosphatidylinositol (PI) and phosphatidylethanolamines (PE) levels, but for lysophosphatidylcholines (LPC) and phosphatidylcholines (PC), no significant changes were reported. Despite effective achievement of LDL-C and non HDL-C targets by the use of less potent statin combined with ezetimibe, their effects in other lipids are less reported. Therefore, this study aimed to compare lipid composition in the plasma of subjects with very high cardiovascular risk with STEMI at baseline (first day) and after 30 days of exposure to two highly effective lipid-lowering therapies (rosuvastatin alone or simvastatin combined with ezetimibe). Prior to the lipidomic study, a classic lipid profile was obtained. Figure 1 shows the box plots of measurements of total cholesterol (TC), high-density lipoprotein-cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C) and triglycerides (TG*) in each group of samples. As expected, after one month of both statin therapies, main changes were observed for TC (q < 0.001) and LDL-C (q < 0.001) with comparable magnitude for both treatments as observed by similar % changes and q-values from Post Hoc test with Bonferroni correction for multiple comparisons of the repeated measures ANOVA (Table S8). It was observed that rosuvastatin therapy decreased on a small scale HDL-C at D30 (− 17%, q-value = 0.04).Figure 1Evaluation of clinical results for the infarcted patients under investigation. G1: Patients randomized to the rosuvastatin group at the first day of myocardial infarction (D1); G2: Patients treated by rosuvastatin after 30 days (D30); G3: Patients randomized to the simvastatin + ezetimibe group at the first day of myocardial infarction (D1); G4: Patients treated by simvastatin plus ezetimibe after 30 days (D30). Total cholesterol (TC), low density lipoprotein—cholesterol (LDL-C), high density lipoprotein—cholesterol (HDL-C) and triglycerides (TG*). Figure was created in Minitab 17.0 (Minitab Statistical Software; URL: https://www.minitab.com/pt-br/)) and Microsoft PowerPoint 2013 (URL: https://www.microsoft.com/pt-br/microsoft-powerpoint-2013). Evaluation of clinical results for the infarcted patients under investigation. G1: Patients randomized to the rosuvastatin group at the first day of myocardial infarction (D1); G2: Patients treated by rosuvastatin after 30 days (D30); G3: Patients randomized to the simvastatin + ezetimibe group at the first day of myocardial infarction (D1); G4: Patients treated by simvastatin plus ezetimibe after 30 days (D30). Total cholesterol (TC), low density lipoprotein—cholesterol (LDL-C), high density lipoprotein—cholesterol (HDL-C) and triglycerides (TG*). Figure was created in Minitab 17.0 (Minitab Statistical Software; URL: https://www.minitab.com/pt-br/)) and Microsoft PowerPoint 2013 (URL: https://www.microsoft.com/pt-br/microsoft-powerpoint-2013). Figure 2 shows the plasma concentrations of the analyzed lipids, categorized by classes (CE, Cer, FA, LPC, LPE, PC, PE, SM, and TG) (Fig. 2a) and by groups (G1, G2, G3 and G4) (Fig. 2b). Comparing the mean values in each group, lipid classes were decreased in both groups after the exposure to the treatments at D30, except for LPC, LPE and TG compound class. However, significant changes were observed for FA, LPC, PC, PE, SM, CE, and Cer only for rosuvastatin group, while SM changed in both groups, as evaluated by q-values from post hoc test with Bonferroni correction for multiple comparisons of the repeated measures ANOVA (Table S8).Figure 2Evaluation of lipid classes for the infarcted patients under evaluation itemized by class (a) and by group (b). CE cholesterol ester, Cer ceramides, FA free fat acids, LPC lysophosphatidylcholine, LPE lysophosphatidylethanolamine, PC phosphatidyl choline, PE phosphatidylethanolamine, SM sphingomyelin, TG triacylglycerides; group labels as in Fig. 1. Figure was created in Minitab 17.0 (Minitab Statistical Software; URL: https://www.minitab.com/pt-br/) and Microsoft PowerPoint 2013 (URL: https://www.microsoft.com/pt-br/microsoft-powerpoint-2013). ±Mean Value. Evaluation of lipid classes for the infarcted patients under evaluation itemized by class (a) and by group (b). CE cholesterol ester, Cer ceramides, FA free fat acids, LPC lysophosphatidylcholine, LPE lysophosphatidylethanolamine, PC phosphatidyl choline, PE phosphatidylethanolamine, SM sphingomyelin, TG triacylglycerides; group labels as in Fig. 1. Figure was created in Minitab 17.0 (Minitab Statistical Software; URL: https://www.minitab.com/pt-br/) and Microsoft PowerPoint 2013 (URL: https://www.microsoft.com/pt-br/microsoft-powerpoint-2013). ±Mean Value. Considering rosuvastatin administration at D30, there was a decrease for CE (− 40%, q-value = 0.006), Cer (− 37%, q-value = 0.02), FA (− 19%, q-value = 0.006), PC (− 65%, q-value < 0.001), PE (− 63%, q-value < 0.001) and SM (− 36%, q < 0.001). LPC was the only lipid class that increased (27%, q-value = 0.002) in G2. Following simvastatin plus ezetimibe therapy at D30, there was a decrease in SM (− 27%, q-value = 0.002). Although not significant (q-value > 0.05), simvastatin treatment after D30 also showed a trend for decrease in PE (− 36%, q-value = 0.08) and PC (− 23%, q-value = 0.1). FA, CE, Cer, LPC, TG and LPE were almost unaltered in G4 (q-value > 0.05, % change: between − 11% and 11%). Comparing the two arms of treatment at D1 (G1 and G3) lower levels of PC (− 54%, q-value < 0.001) and PE (− 51%, q-value = 0.001) were found for simvastatin plus ezetimibe group. Regarding the effects of treatments at D30 (G2 and G4), only the sum of LPC were 14% decreased (q-value = 0.02) in G4 compared to rosuvastatin arm (G2). Levels of LDL-C, HDL-C, TC, and TG* were correlated with levels of lipid classes (CE, Cer, FA, LPC, LPE, PC, PE, SM, and TG) for all studied patient groups (Fig. 3). More correlations between clinical parameters and lipids were found considering rosuvastatin administration at D1 in comparison to D30 and to the other lipid-lowering therapy. Figure 3a,b present the correlation plots related to the two therapies at D1 of STEMI, respectively, to G1 and G3. Considering rosuvastatin administration at D1 (Fig. 3a) a series of correlations between clinical parameters and lipid classes were observed. CE, PC, SM and TG were highly correlated with LDL-C, TC and TG*. Otherwise, TG* was negatively correlated to FA. Interestingly, considering simvastatin plus ezetimibe at D1 (Fig. 3b), no negative correlation was found and fewer positive correlations are present. PC was positively correlated to LDL-C and TC, and SM was positively correlated only with TC. Figure 3c,d depicts correlation plots related to the two therapies at D30 of STEMI, respectively, to G2 and G4. Considering rosuvastatin at D30 (G2), only one negative correlation between TG and HDL-C was found. Positive correlations between PC vs. LDL-C and TG*; SM vs. LDL-C and TC; LPC vs. LDL-C were found considering simvastatin plus ezetimibe at D30 and, a negative correlation was found between FA and HDL-C.Figure 3Pearson’s correlation of lipid classes concentrations and clinical parameters. (a) for rosuvastatin treatment at D1 (G1), (b) for simvastatin plus ezetimibe treatment at D1 (G3), (c) for rosuvastatin treatment (G2) at D30, and (d) for simvastatin plus ezetimibe treatment at D30 (G4). Figure was created in R 3.6.3 (The R Project for Statistical Computing; https://www.r-project.org/packages: corrplot, Hmisc, RColorBrewer, Cairo). Pearson’s correlation of lipid classes concentrations and clinical parameters. (a) for rosuvastatin treatment at D1 (G1), (b) for simvastatin plus ezetimibe treatment at D1 (G3), (c) for rosuvastatin treatment (G2) at D30, and (d) for simvastatin plus ezetimibe treatment at D30 (G4). Figure was created in R 3.6.3 (The R Project for Statistical Computing; https://www.r-project.org/packages: corrplot, Hmisc, RColorBrewer, Cairo). The OPLS-DA scores plot (Fig. 4) shows a separation among the four studied groups (G1, G2, G3 and G4). It is possible to observe a higher separation between time of treatment, D1 (G1 and G3) and D30 (G2 and G4), in comparison to the separation between the two lipid-lowering therapies, rosuvastatin (G1 and G2) and simvastatin plus ezetimibe (G3 and G4). However, a worst separation was observed between the two lipid-lowering therapies at D30 (G2 × G4). The OPLS-DA model, with R = 0.637 and Q = 0.246, was validated by CV-ANOVA (p-value = 0.032) and permutation test (100 permutations) of selected Y variables. The poor predictability of the model, indicated by Q < 0.5, is provided by the low numbers of observations (n = 10) and the properties of the dataset. Thus, this model must be validated by permutation test and by cross-validation (pcv-ANOVA). The QVY values for each y variables, from the permutation test, had shown that the Y variable G4 was responsible for the poor predictability of this model (Q = 0.246) as shown in the Supplementary Table S1. Triba et al., recommend to permute the lines of their dataset to control that the Q value calculated is stable regarding this permutation. In addition, the stable Q with low error (0.251 ± 0.06; Q2 mean ± 2*Error; Error = ta2*sd/√n) obtained towards a permutation test with five randomized dataset, indicating that the model is trustful (See Supplementary Table S2).Figure 4OPLS-DA scores plot (PC1 vs PC2) of all infarcted patients in the four groups considered. G1: Patients randomized to the rosuvastatin group at first day after myocardial infarction (D1); G2: Patients treated by rosuvastatin after 30 days of treatment (D30); G3: Patients randomized to the simvastatin plus ezetimibe group at the first day after myocardial infarction (D1); G4: Patients treated by simvastatin plus ezetimibe after 30 days (D30). Figure was created in SIMCA 16 (Statistical Software Package, Umetrics, Sweden; URL: http://umetrics.com/product/simca). OPLS-DA scores plot (PC1 vs PC2) of all infarcted patients in the four groups considered. G1: Patients randomized to the rosuvastatin group at first day after myocardial infarction (D1); G2: Patients treated by rosuvastatin after 30 days of treatment (D30); G3: Patients randomized to the simvastatin plus ezetimibe group at the first day after myocardial infarction (D1); G4: Patients treated by simvastatin plus ezetimibe after 30 days (D30). Figure was created in SIMCA 16 (Statistical Software Package, Umetrics, Sweden; URL: http://umetrics.com/product/simca). Discriminant variables were selected by the combination of multivariate analysis (VIP > 1), and univariate evaluation by post hoc test with Bonferroni correction for multiple comparisons of the repeated measures ANOVA (q-value < 0.05 for one of the comparisons) as shown in Supplementary Table S3 and displayed in Fig. 5. VIP value for all variables studied and, post-hoc (q-value) for variables considered statistically significant are shown in Supplementary Table S4 and S5, respectively.Figure 5Discriminant lipids from multivariate (VIP > 1) and univariate analysis (repeated measures ANOVA). (a) % change of metabolites considering the temporal treatment effect (*metabolite statistically significant are highlighted with black borders); (b) % of change of metabolites considering the lipid-lowering therapy effect (*metabolite statistically significant are highlighted with black border); (c) Venn Diagram considering lipid-lowering therapy; (d) Venn Diagram considering temporal treatment effect. Figure was created in Microsoft Excel 2013 (URL: https://www.microsoft.com/pt-br/microsoft-excel-2013) and Microsoft PowerPoint 2013 (URL: https://www.microsoft.com/pt-br/microsoft-powerpoint-2013). Discriminant lipids from multivariate (VIP > 1) and univariate analysis (repeated measures ANOVA). (a) % change of metabolites considering the temporal treatment effect (*metabolite statistically significant are highlighted with black borders); (b) % of change of metabolites considering the lipid-lowering therapy effect (*metabolite statistically significant are highlighted with black border); (c) Venn Diagram considering lipid-lowering therapy; (d) Venn Diagram considering temporal treatment effect. Figure was created in Microsoft Excel 2013 (URL: https://www.microsoft.com/pt-br/microsoft-excel-2013) and Microsoft PowerPoint 2013 (URL: https://www.microsoft.com/pt-br/microsoft-powerpoint-2013). Figure 5a,b show the percentage change related to the two lipid-lowering therapy effect (G1 vs G3 and G2 vs G4) and temporal treatment effect (G1 vs G2 and G3 vs G4). As expected, fewer metabolites are altered related to the two lipid-lowering therapy comparison than temporal treatment. Figure 5c shows that 2 FA, 4 PC and 2 SM were significantly altered between the two lipid-lowering therapies at D1 (G1 vs G3), only 2 LPC at D30 (G2 vs G4) and, FA 16:1 was similar for both comparison (D1 and D30). Considering the temporal treatment evaluation (G1 vs G2 and G3 vs G4), Fig. 5d shows that 1 FA, 4 CE, 1 Cer, 1 TG, 1 PE, 2 SM, 2 LPC and 5 PC levels were affected by rosuvastatin therapy (G1 vs G2), 2 PC and only 1 PE levels were decreased only by simvastatin + ezetimibe treatment (G3 vs G4). Both treatments promoted changes in the concentration levels of 5 FA, 3 LPC, 5 PE, 7PC and 6 SM. Figure 5a shows that, LPC 20:4, PE 16:0_18:2, LPC 17:0, LPC 16:0, LPC 18:1, SM 18:1;O2/20:0, presented the largest alterations by the rosuvastatin treatment. The largest alterations promoted by the simvastatin plus ezetimibe treatment were for PE 16:0_18:2, LPC 16:1 and LPC 20:4. This study revealed differences in many lipid concentrations in subjects with STEMI despite comparable classic lipid profiles. After highly effective lipid-lowering therapies (rosuvastatin or simvastatin + ezetimibe), similar trend in the lipid composition were observed for both therapies. However, differences were noted between the lipid-lowering therapies for many lipids such as FA, LPC, PC, PE, CE, Cer and SM and these findings seem relevant, due to the involvement of these lipid classes in crucial mechanisms of atherosclerosis . These changes were more pronounced by rosuvastatin therapy, and the significant decrease in Cer observed only with this lipid-lowering drug seems important, as Cer has been considered a robust predictor for cardiovascular events, independently of LDL-C as well as for cardiovascular mortality. It should be pointed out that, although in the lipid class evaluation, the sum of individual lipids from the same class presented significant changes only for SM class with simvastatin plus ezetimibe therapy, the statistical analysis of individual lipid species has not shown that the treatment altered lipids from many other classes (FA, PE, LPC, PC). These findings highlight the importance of evaluating lipid species individually. Lipids with different fatty acid chains, from the same class, may present opposite effects in CVD risk. In fact, lipidomic studies have shown that PC containing polyunsaturated fatty acyl chains were negatively associated with CVD risk, while saturated and monounsaturated fatty acyl chains were deleterious for CVD risk. Inflammation and oxidative stress have been considered as a part of the pathophysiology of acute coronary syndromes and their recurrences. Although FA and their esters constitute the major sources of energy for the heart muscle, their excess has profound effects on the heart causing an enhanced susceptibility to inflammation, oxidative stress and ischemic damage (fibrosis and hypertrophy). Different FA can be metabolized into pro- and anti-inflammatory signalling molecules. In particular, some n-6 FA (first and foremost arachidonic acid) are precursors to pro-inflammatory molecules (primarily prostaglandins), while saturated fatty acids (SAFAs) act as major inducers of inflammation through several mechanisms, such as the activation of toll-like receptor-4 (TLR4), which promotes the production of inflammatory cytokines (IL-1beta; IL-6). Endothelial dysfunction due to the NF-kappa B activation is also induced by SAFA, resulting in increased superoxide production, while NLRP3 inflammation activation increases endothelial permeability. In our study, in the two arms of therapy, lower concentrations of FA (SAFAs, monounsaturated fatty acids—MUFAs and polyunsaturated fatty acids—PUFAs) were observed after 1 month of treatment. These results suggest a potential beneficial effect of therapies on CVD. Interesting to note, both therapies promoted similar reduction in FA, but for FA 16:0, it was significantly reduced at D30 only in the rosuvastatin arm. Our results are in line with a meta-analysis of clinical studies involving atorvastatin and simvastatin published by Sahebkar et al. showing reduction in free fatty acids levels independently of treatment duration, dose and magnitude of reduction in LDL-C levels. LPC is increasingly recognized as a key factor positively associated with atherosclerosis development and cardiovascular diseases. However, findings from recent clinical studies have suggested potential biomarkers of positive prognostic making those studies controversial. A key issue is the complexity of the enzyme cascade involved in LPC metabolism, which shows that a long way must be traveled until the understanding of the results reported in literature. In our study, LPCs were significantly increased with rosuvastatin treatment at D30, of which five were responsible for this increase (LPC 15:0, LPC 16:1, LPC 17:0, LPC 18:1 and LPC 20:4), while only three of them had their levels altered after simvastatin + ezetimibe at D30 (Table S3). Interestingly, in agreement with our findings, LPC 18:1 and LPC 20:4 has shown a negative association with cardiovascular events in three different lipidomic studies, as reported by Ding et al.. Fernandez, et al.studying plasma lipid composition and risk of developing cardiovascular disease also observed a decrease in LPCs associated with CVD. Other interesting study on lipid profile of rosuvastatin in humans has also highlighted changes in LPC and PC compound classes, suggesting an effect of rosuvastatin in LPC and PC metabolism. Although no significant differences between therapies were observed in the D1 (G1 vs. G3) and D30 (G2 vs. G4) for LPC compounds class, temporal % changes in the rosuvastatin arm were higher than simvastatin/ezetimibe. Some lipid composed of odd chain fatty acids were significantly altered by both treatments. Aforetime, odd chain fatty acids were associated solely with diet, however nowadays they are recognized as products of catabolism of branched-chain amino acids. Branched-chain amino acids were positively associated with CVD risk and type 2 diabetes. Furthermore, Khaw et al. have associated odd chain phospholipids with lower cardiovascular risk. Also, in the study of Ward-Caviness et al. a standard deviation increase in log-transformed concentration of LPC 17:0 negatively correlated with the risk of incident myocardial infarction in three cohorts. Interestingly, our study shows that rosuvastatin therapy increased LPC 17:0 in more than 50% at D30, which could contribute to a cardiovascular protective effect. PE and PC were reduced after treatment with both lipid-lowering therapies. High levels were found in rosuvastatin group for most of the compounds. In addition, our study showed that both lipid-lowering therapies decreased SM concentrations at D30. Sphingomyelin was measured in the large Multi-Ethnic Study of Atherosclerosis (MESA), and the authors reported a modest negative association with incident CVDs, after adjustment for lipoproteins and full adjustment for other risk factors. Higher decrease with rosuvastatin compared with atorvastatin treatments was found in the SM/SM + PC ratio. The same was found here, for rosuvastatin and simvastatin + ezetimibe comparison. These results corroborate Choi et al. findings, that reported a significant decrease in SM, TG and PE, especially in PE 18:0_18:2 levels, and an increase in LPC 20:4 and LPC 18:1. In 2018, Lee et al. hypothesized that these increases may be due to pleiotropic effects of statins, since phospholipase A2, which is the main enzyme involved in LPC metabolism, is inhibited by rosuvastatin. Besides the superiority in changing lipid concentration for rosuvastatin in the altered classes compared to simvastatin + ezetimibe, only rosuvastatin promoted significant reduction in CE and Cer compounds. Ng et al. studying the dose-dependent effects on plasma sphingolipidome and phospholipidome in the metabolic syndrome found that rosuvastatin at both 10 and 40 mg/d significantly reduced the concentrations of total and individual plasma sphingolipids and phospholipids with evidence of dose-dependent effects. For CE, as the results found for LPC, controversial findings have been reported, but Stegemann et al. applying mass spectrometry-based lipidomic profiling, reported an association of CE with cardiovascular disease over a 10-year observation period. Considering the correlations between lipids and classic clinical parameters, it is interesting to note the number of correlations found at G1 compared to other groups (G2, G3, and G4). The lack of correlations for those groups indicates that lipidomic measurements and classic lipid profile addresses different information. In this work, PC, SM and TG were positively correlated to LDL-C for both lipid-lowering therapies at D1, but a negative or no correlation between HDL-C and TG was found, mainly when rosuvastatin was administrated at D30. There is an expected inverse association between HDL-C and TG in diabetic and overweight patients. Patients with insulin resistance have delayed hydrolysis of TG-rich lipoproteins due to reduced lipoprotein lipase activity. As a result, fewer phospholipids and surface components are transferred to HDL, decreasing its effectiveness for the reverse cholesterol transport. Therefore, lower concentrations of cholesterol from HDL are found. A previous study reported that LPC are predominantly found in HDL, Cer in LDL, and PC are present in both lipoproteins (HDL and LDL). Studies examining CVDs with lipidomics found that after adjusting for HDL-C and LDL-C levels, only PC remained associated with cardiovascular events. In this work, LDL-C was reduced with both lipid-lowering therapies at similar % changes. Otherwise, Cer was reduced significantly, only with rosuvastatin at D30. As a reduction of LDL-C and Cer is expected to contribute to lower residual cardiovascular risk, rosuvastatin presented better results than simvastatin + ezetimibe. Considering classic lipid parameters, rosuvastatin treatment decreased in 17% HDL-C level after one month. Another study showed that patients on rosuvastatin therapy may have either increased or decreased the HDL-c after 4 weeks of treatment. However, these changes in HDL-C did not affect the incidence of major cardiovascular events in one year follow-up. In summary, regarding to the changes in the classic lipid profile, our results are in agreement with previous studies. Interestingly, the sum of our findings corroborates with a review article recently published by Rai & Bhatnagar showing decreased LPC 16:0 in hyperlipidemia causative disorders, such as high-fat diet, obesity and diabetes. LPC decrease was accompanied by increase in FA and Cer. The authors also highlighted the association of hyperlipidemia with an increase in small-chain fatty acids, SAFA content of DG, TG and PC lipid classes, factors associated with CVD risk. No direct comparisons between rosuvastatin and simvastatin + ezetimibe in cardiovascular outcomes have been reported. This study compared two highly effective lipid-lowering therapies in the acute phase of myocardial infarction. However, we are unable to estimate the effects of the acute myocardial infarction per se on lipid composition. It is expected some decrease in lipids due to the healing process involving the necrotic and ischemic myocardium, but these changes have been reported as insignificant in the following days after the acute coronary event. In addition, the myocardial injury estimated by troponins was similar in both arms of lipid-lowering treatment. Our sample size is relatively small, but the patients included in the study were all submitted to same treatment strategy (pharmacoinvasive) with similar characteristics at baseline. In spite of comparable classic lipid profile at baseline and after the exposure to treatments, significant differences in lipid composition were found between the two highly effective lipid-lowering therapies. To note, higher % changes in the rosuvastatin arm of therapy compared to simvastatin + ezetimibe were identified, and significant changes for CE and Cer were observed only in the rosuvastatin group. Regarding the correlation studies, we found that lipidomic analysis and classic clinical exams account for different information in both lipid-lowering therapies. In summary, our results indicate important differences in lipid composition that cannot be identified by the classic lipid profile between the studied lipid-lowering therapies. These differences may account for residual cardiovascular risk. Methanol, 1-propanol and dichloromethane in HPLC grade were purchased from JTBAKER (Avantor Performance Materials, Mexico, Mexico). Water was purified by the Milli Q system (Millipore Waters, Darmstadt, Germany). Ammonium acetate was obtained from Sigma-Aldrich (Saint Louis, MO, USA). The Lipidyzer isotope labeled internal standards mixture kit consisting of 54 isotopes from 13 lipid classes (LPC, lysophosphatidylethanolamines (LPE), PC, PE, Sphingomyelin (SM), diacylglycerols (DG), TG, FA, Cholesterol ester (CE), Ceramide (Cer); dihydroceramides (DhCer), hexosylceramides (HexCer), lactosylceramides (LacCer) was purchased from Sciex (Framingham, MA, USA). This prospective, randomized, open label study was delineated to evaluate differences in the composition of lipids in patients with STEMI at D1 and at D30 after implementing the two lipid-lowering therapies (rosuvastatin 20 mg [Crestor, AstraZeneca] or simvastatin 40 mg combined with ezetimibe10 mg [Vytorin, MSD]). Plasma of patients were categorized in four groups: patients in the rosuvastatin group at D1 (G1) and at D30 (G2); patients treated with simvastatin + ezetimibe at D1 (G3) and at D30 (G4). These two lipid-lowering therapies were chosen to promote similar changes in the classic lipid profile, allowing the comparison of the more effective inhibition of cholesterol synthesis (rosuvastatin) with the combined mechanisms of LDL-C lowering (inhibition of cholesterol synthesis and inhibition of intestinal cholesterol absorption by simvastatin/ezetimibe). Patients were randomized 1:1 for the lipid-lowering therapy using a central computerized system (battle-ami.huhsp.org.br). All patients followed similar protocol, receiving dual antiplatelet therapies, betablockers, and renin-angiotensin system blockers and they were referred to coronary angiogram, and percutaneous coronary intervention, when needed, in the first 24 h of STEMI. The study included mainly middle aged males, approximately half of them with type 2 diabetes. From the 25 patients consecutively screened for the trial, three were not eligible due to inclusion/exclusion criteria and two did not complete trial (one patient died in the first month e other was hospitalized due to heart failure). Table 1 shows the main characteristics of study population. The cohort is part of the B And T Types of Lymphocytes Evaluation in Acute Myocardial Infarction (BATTLE-AMI) study (NCT02428374). They had no prior MI and were naive for lipid-lowering treatment. All patients were submitted to pharmacological thrombolysis with tenecteplase in the first 6 h of STEMI, followed by coronary angiogram and percutaneous intervention when needed in the first 24 h of STEMI (pharmacoinvasive strategy). Key exclusion criteria included hemodynamic instability, autoimmune disease, known malignancy, pregnancy and signs of active infections. All patients received the study medications from the hospitalization as well as at hospital discharge. These patients were monitored by phone and followed up in our outpatient clinic (Hospital Sao Paulo—UNIFESP). At each visit, the patients brought the boxes of their medications to check their adherence.Table 1Characteristics of the study population and classic lipid profile at baseline and after treatment.ParametersRosuvastatin (n = 10)Simvastatin/ezetimibe (n = 10)p-valueAge, years, median (IQR)62 (59–64)53 (48–62)0.06Male gender, n (%)7 (70)8 (80)0.61Weight, kg76.2 ± 11.774.5 ± 10.90.78BMI, kg m, median (IQR)25.3 (24.4–29.7)29.1 (28.5–32.30.32Diabetes, n (%)4 (40)6 (60)0.37Hypertensives, n (%)7 (70)7 (70)1.00BaselineHbA1c, %6.2 ± 1.66.9 ± 1.70.83Glucose, mg dL155 ± 61171 ± 780.41Creatinine, mg dL0.96 ± 0.210.95 ± 0.200.85GFR, mL min m77 ± 1586 ± 210.32Troponin T, ng L5713 ± 33667770 ± 70530.42Cholesterol, mg dL202 ± 44204 ± 470.91LDL-cholesterol, mg dL131 ± 29134 ± 390.85HDL-cholesterol, mg dL41 ± 939 ± 130.71Triglycerides, mg dL171 ± 106189 ± 1260.75After 30 daysCholesterol, mg dL106 ± 13118 ± 240.18LDL-cholesterol, mg dL51 ± 1159 ± 210.30HDL-cholesterol, mg dL34 ± 335 ± 60.60Triglycerides, mg dL123 ± 33152 ± 660.23Data are mean ± SD unless otherwise stated. Comparisons were examined by unpaired Student´s t test or by the non-parametric Mann–Whitney U test. Categorical variables were tested by Pearson’s Chi-square test. BMI body mass index, HbA1c glycated hemoglobin, GFR glomerular filtration rate. Characteristics of the study population and classic lipid profile at baseline and after treatment. Data are mean ± SD unless otherwise stated. Comparisons were examined by unpaired Student´s t test or by the non-parametric Mann–Whitney U test. Categorical variables were tested by Pearson’s Chi-square test. BMI body mass index, HbA1c glycated hemoglobin, GFR glomerular filtration rate. Fasting blood samples were collected, in the morning at D1 and at D30 after lipid-lowering therapy, in tubes containing EDTA, followed by centrifugation at 1300g for 15 min, at room temperature and storage at − 80 °C before analysis. All samples for general biochemical tests, including the classic lipid profile were performed in the Central Laboratory of the University Hospital and the LDL-C was estimated by the Friedewald equation. Biochemical determination of TC, HDL-c and TG* was performed by enzymatic colorimetric assays with commercial kits from Roche in Cobas C 501 module. Lipid extraction was carried out by a modified Blight-Dyer protocol as described elsewhere. Briefly, 100 μL of plasma were transferred to a borosilicate glass culture tube (16 × 100 mm). Next, 900 μL water, 2 mL methanol, and 900 μL dichloromethane were added to all samples and the mixture was vortexed for 5 s. Samples were left to incubate at room temperature for 30 min. Next, another 1 mL water and 900 μL dichloromethane were added to the tube, followed by gentle vortexing for 5 s, and centrifugation at 2500g at 15 °C for 10 min. The bottom organic layer was transferred to a new tube and 1.8 mL dichloromethane were added to the original tube for a second extraction. The combined extracts were concentrated under nitrogen. Exactly 100 μL of the isotope labeled internal standards mixture were added to the dried extract and another 30 min incubation was allowed until equilibrium is reached. Finally, 250 μL mobile phase solution (10 mmol L ammonium acetate in 50:50 methanol:dichloromethane) were added. IS mixture is composed by 54 labeled lipid species that covers 10 main lipid classes found in human plasma with different final concentrations reflecting their physiological concentrations. Quantitative lipidomics was performed with the Sciex Lipidyzer platform configured by an ExionLC AD instrument (Sciex) coupled to a QTRAP 5500 mass spectrometer (Sciex) equipped with SelexION for differential mobility spectrometry (DMS) and electrospray ionization (ESI) source. The solvent 1-propanol was used as the chemical modifier for the DMS. Samples were introduced to the mass spectrometer by flow injection at 8 µL min. Each sample was injected twice, with the DMS on (PC/PE/LPC/LPE/SM) and off (CE/Cer/DG/FA/TG). Over 1100 lipid species and 54 labeled internal standards were monitored by selected reaction monitoring (SRM) in positive/negative polarity switching. Positive ion mode was used to detect the lipid classes SM/DG/CE/Cer/TG and the negative ion mode to detect the lipid classes LPE/LPC/PC/PE/FA. Lipid annotation is achieved by measuring specific SRM transitions, where the monitored fragment is related to the fatty acid composition. Also, lipid class is determined by ramping the compensation voltages in the differential mobility unit of the Lipidyzer platform. DMS parameters used were: temperature = low; separation voltage = 3.5 kV and differential mobility spectrometric resolution = low. Electrospray ion source parameters were as follow: voltage (ESI +): 4.1 kV, voltage (ESI -) = − 2.5 kV, curtain gas = 17, CAD gas = Medium, Temperature = 200 °C, Nebulizing gas = 17 and heater gas − 25. All data obtained from the Lipidyzer Platform were automatically processed in the Lipidomics Workflow Manager (LWM). Signals of all lipids obtained for each sample were quantified using the intensity of internal standard applying the Lipidyzer platform. The software calculates concentration as average intensity of the analyte MRM/average intensity of the most structurally similar IS MRM multiplied by its concentration in nmol/mL. Lipidyzer platform allowed for automated data acquisition, data processing, and reporting. A detailed description of the method can be found in previous studies. Quality control (QC) consisted of a standard plasma sample obtained from the Lipydizer kit. The reconstituted lyophilized plasma was extracted following the procedure described previously. The QC sample was injected five times at the beginning of the randomized sample batch, every 10 injections and, at the end of the sample batch. Box-plot was performed using Minitab 17.0 (Minitab Statistical Software; https://www.minitab.com/pt-br/) and Microsoft Excel 2013. Univariate analysis (Repeated Measures ANOVA) was performed in MATLAB (The MathWorks, Inc., Natick, Massachusetts, United States) using an in-house script. Multivariate analysis was performed in SIMCA 16 (Statistical Software Package, Umetrics, Sweden; http://umetrics.com/product/simca). Correlation graphics were performed in R 3.6.3 (The R Project for Statistical Computing; https://www.r-project.org/packages: corrplot, Hmisc, RColorBrewer, Cairo). From the original generated table compiling the lipids identified (Supplementary Table S7), only the ones that had presented concentrations (µmol L-1) in at least 80% of the samples of one group were used for data treatment. Discriminant variables were obtained not only by multivariate analysis (VIP > 1), but also by univariate evaluation by Post Hoc test with Bonferroni correction for multiple comparisons of the Repeated Measures ANOVA, as shown in Supplementary Table S3. Data quality was carried out by inspecting the repeatability of lipids in the QC plasma sample, analyzed throughout data acquisition. More than 80% of the quantified metabolites in the QCs have acceptable coefficient of variation percentages (% CV) for peak areas; in this work, < 20% CV was used as a criterion to retain that particular component in the dataset for further evaluation, which was in agreement with the recomendation. The study protocol was approved by the local ethics committee (UNIFESP IRB 0297/2014; CAAE: 71652417.3.0000.5505), which follows the Declaration of Helsinki, and written informed consent was provided by all subjects before their inclusion in the study.
|
PMC9170690
|
Comprehensive genetic analysis of the human lipidome identifies loci associated with lipid homeostasis with links to coronary artery disease
|
We integrated lipidomics and genomics to unravel the genetic architecture of lipid metabolism and identify genetic variants associated with lipid species putatively in the mechanistic pathway for coronary artery disease (CAD). We quantified 596 lipid species in serum from 4,492 individuals from the Busselton Health Study. The discovery GWAS identified 3,361 independent lipid-loci associations, involving 667 genomic regions (479 previously unreported), with validation in two independent cohorts. A meta-analysis revealed an additional 70 independent genomic regions associated with lipid species. We identified 134 lipid endophenotypes for CAD associated with 186 genomic loci. Associations between independent lipid-loci with coronary atherosclerosis were assessed in ∼456,000 individuals from the UK Biobank. Of the 53 lipid-loci that showed evidence of association (P < 1 × 10), 43 loci were associated with at least one lipid endophenotype. These findings illustrate the value of integrative biology to investigate the aetiology of atherosclerosis and CAD, with implications for other complex diseases.Lipids comprise thousands of individual species, spanning many classes and subclasses. Genome-wide association studies (GWAS) of lipid species can provide novel insights into human physiology, inborn errors of metabolism and mechanisms for complex traits and diseases. Dyslipidaemia, a broad term for disordered lipid and lipoprotein, is a major risk factor for atherosclerotic cardiovascular disease and a therapeutic target for the primary and secondary prevention of coronary artery disease (CAD). Defined by elevated low-density lipoprotein (LDL) cholesterol and triglycerides with decreased high-density lipoprotein (HDL) cholesterol —these ‘clinical lipid’ measures provide only a partial view of the complex lipoprotein structures and their metabolism. Lipidomic technologies can now measure hundreds of individual molecular lipid species that make up the human lipidome, providing a more complete snapshot of the underlying lipid metabolism occurring within an individual. Genome-wide association studies have uncovered thousands of genetic variants linked to traditional clinical lipids (LDL-cholesterol, HDL-cholesterol, triglycerides). Genes implicated at these loci show functional links between lipid levels and CAD. The human lipidome is heritable and predictive of CAD, furthering our understanding of the biology of CAD. The individual lipid species that make up the lipidome are biologically simpler measures that may reside closer to the causal action of genes, making them valuable endophenotypes for gene identification. Genetic interrogation of the human lipidome may therefore reveal further genetic variants that play a role in lipid metabolism and CAD. Compared with other complex traits, relatively few genomic loci have been associated with lipid species in GWAS of the human serum/plasma lipidome, although these studies have generally interrogated a restricted subset of lipid species. The serum lipidome is complex and consists of many isobaric and isomeric species that share elemental composition but are structurally distinct. Existing lipidomic studies often employ techniques that provide poor resolution of these species, limiting their biological interpretation. We have recently expanded our lipidomic platform to better characterise isomeric lipid species, now measuring 596 lipids from 33 classes. Our methodology focuses on the precise measurement of a broad number of lipid and lipid-like compounds, utilising extensive chromatographic separation. Here, we report a GWAS of 596 targeted lipid species (across 33 lipid classes) in an Australian population-based cohort of 4492 individuals, validation of significant loci in two independent cohorts and a meta-analysis of all results. Using robust procedures, we disentangle the genetic effects of lipid species from lipoproteins. Integration of multiple datasets, including expression quantitative trait loci (eQTL), methylation QTL (meQTL), and protein QTL (pQTL), and in-depth analysis of significant loci highlights putative susceptibility genes for CAD. We demonstrate robust associations between lipid species and CAD using genetic correlations, polygenic risk scores and phenotypic associations. Many lipid-associated loci show pleiotropy with CAD in co-localisation analysis. Assessment of loci with coronary atherosclerosis in 456,486 UK Biobank participants reveals genetic associations, independent of clinical lipid measures. We measured 596 individual lipid species within 33 lipid classes, covering the major glycerophospholipid, sphingolipid, glycerolipid, sterol, and fatty acyl classes in serum and plasma samples from three independent cohorts (Supplementary Table 1, Supplementary Data 1, 2). Assay performance was monitored using pooled plasma quality control samples, enabling the determination of coefficient of variation (%CV) values for each lipid class and species. In the Busselton Health Study (BHS) discovery cohort, the median %CV was 8.6% with 570 (95.6%) lipid species showing a %CV less than 20%. All lipids were measured in every individual, with the exception of three values which were below the limit of detection. The lipidomic analysis of the Australian Imaging, Biomarker, and Lifestyle (AIBL) and Alzheimer’s Disease Neuroimaging Initiative (ADNI) validation cohorts showed similar assay performance. We performed a GWAS of the human serum lipidome (Fig. 1) in the BHS discovery cohort (4492 individuals of European ancestry) followed by validation against a meta-analysis of the two validation cohorts (ADNI and AIBL; 670 and 895 individuals of European ancestry, respectively). We further performed a discovery meta-analysis of all three studies. All summary-level statistics are available at our PheWeb data portal (https://metabolomics.baker.edu.au/).Fig. 1Study design for the genetic analysis of the human lipidome.Representation of genome-wide association studies (GWAS) of the lipidome in the BHS discovery cohort (blue boxes), ADNI and AIBL validation cohorts (green boxes), discovery meta-analysis (orange box), and downstream analyses (grey boxes). ADNI Alzheimer’s Disease Neuroimaging Initiative, AIBL Australian Imaging, Biomarker & Lifestyle Flagship Study of Ageing, BHS Busselton Health Study, CAD coronary artery disease, Chol cholesterol, eQTL expression quantitative trait loci, GRM genetic relatedness matrix, GWAS genome-wide association study, IVW inverse-variance weighted, LD linkage disequilibrium, MAC minor allele count, meQTL methylation quantitative trait loci, mQTL metabolite quantitative trait loci, PC principal component, PRS polygenic risk score, pQTL protein quantitative trait loci, SD standard deviation, SNP single nucleotide polymorphism, Trig triglycerides. Representation of genome-wide association studies (GWAS) of the lipidome in the BHS discovery cohort (blue boxes), ADNI and AIBL validation cohorts (green boxes), discovery meta-analysis (orange box), and downstream analyses (grey boxes). ADNI Alzheimer’s Disease Neuroimaging Initiative, AIBL Australian Imaging, Biomarker & Lifestyle Flagship Study of Ageing, BHS Busselton Health Study, CAD coronary artery disease, Chol cholesterol, eQTL expression quantitative trait loci, GRM genetic relatedness matrix, GWAS genome-wide association study, IVW inverse-variance weighted, LD linkage disequilibrium, MAC minor allele count, meQTL methylation quantitative trait loci, mQTL metabolite quantitative trait loci, PC principal component, PRS polygenic risk score, pQTL protein quantitative trait loci, SD standard deviation, SNP single nucleotide polymorphism, Trig triglycerides. Within the discovery GWAS, 70,831 genome-wide significant SNP-lipid species and 3474 SNP-lipid class associations were identified (P < 5.0 × 10; Fig. 2). All lipid classes and 543 (of 596; 91.1%) lipid species had at least one significant association (Supplementary Data 3, 4). All significantly associated SNPs were in Hardy-Weinberg Equilibrium (HWE; all P ≥ 1.53 × 10) and were relatively common (minor allele frequency; MAF < 0.01: 4%; MAF > 0.05: 91%, Supplementary Data 5). LD-clumping identified 2279 independent SNP-lipid species associations, and 132 independent SNP-lipid class associations at a genome-wide significance (P < 5.0 × 10; r < 0.1; Fig. 2; Supplementary Data 6).Fig. 2Circular presentation of loci associated with circulating lipid species identified in our Discovery GWAS.The −log10(P) for genetic association with lipid species are arranged by chromosomal position, indicated by alternating blue and green points. Association P-values are truncated at P < 1 × 10. Genome-wide significance (P < 5 × 10) is indicated by the red line. For details about significant associations, see Supplementary Data 2, 3. Genes identified in our candidate gene analysis are highlighted in blue, otherwise the closest gene is indicated in black. The purple band indicates lipid-loci that co-localise with coronary artery disease (CAD) or show association with CAD after adjusting for clinical lipids. The inner circle shows a Fuji plot of SNP-lipid associations, coloured by broad lipid category. Colour keys representing broad lipid categories are indicated in the plot centre. Chromosomes are indicated by numbered panels 1–22. The −log10(P) for genetic association with lipid species are arranged by chromosomal position, indicated by alternating blue and green points. Association P-values are truncated at P < 1 × 10. Genome-wide significance (P < 5 × 10) is indicated by the red line. For details about significant associations, see Supplementary Data 2, 3. Genes identified in our candidate gene analysis are highlighted in blue, otherwise the closest gene is indicated in black. The purple band indicates lipid-loci that co-localise with coronary artery disease (CAD) or show association with CAD after adjusting for clinical lipids. The inner circle shows a Fuji plot of SNP-lipid associations, coloured by broad lipid category. Colour keys representing broad lipid categories are indicated in the plot centre. Chromosomes are indicated by numbered panels 1–22. Each SNP was associated with between 1 and 222 lipids (Supplementary Fig. 1). SNPs associated with a large number of lipids were in regions known to be involved in lipid regulation, including FADS1/FADS2/FADS3, APOE, and LIPC. The most significant associations were observed between PC(18:0_20:4) and rs174564 (FADS2; P = 4.63 × 10) and between Cer(d19:1/22:0) and the intergenic SNP rs364585 (flanking SPTLC3; P = 7.81 × 10). In fact, the most significant 26 SNP-lipid species associations were with SNPs in these two regions. The median genomic inflation factors were 1.01 (range: 0.99–1.03), and 1.02 (range: 1.00–1.03) for lipid species and class analyses, respectively. SNP-based heritability estimates were moderately correlated (r = 0.45) with lambda estimates, for each of the lipid species and classes (Supplementary Fig. 2a), as expected. We performed an additional GWAS, adjusting for clinical lipids (total cholesterol, HDL-cholesterol, triglycerides), to identify SNP-lipid species associations independent of clinical lipid traits (Adjusted Discovery GWAS). The median genomic inflation factors were 1.01 (range: 0.99–1.03), and 1.01 (range: 1.00–1.03) for lipid species and classes, respectively; with heritability estimates moderately correlated (r = 0.51) with lambda estimates, for each of the lipid species and classes (Supplementary Fig. 2b). Adjustment for clinical lipids identified 2424 independent SNP-lipid species associations, and 124 independent SNP-lipid class associations (Supplementary Data 6). There were 1545 SNP-lipid species and 72 SNP-lipid class associations that were significant in both the unadjusted and the adjusted analyses, with an r between beta coefficients of 0.93 (Fig. 3). Adjustment for clinical lipids identified an additional 879 significant SNP-lipid species associations, for 387 lipid species. However, 726 SNP-lipid species associations previously associated in the unadjusted analysis, fell below our significance threshold. Approximately 24% of these lipid species are members of the cholesteryl ester (n = 93), and phosphatidylcholine (n = 81) classes (Supplementary Data 6). We also identified an additional 52 significant SNP-lipid class associations, particularly for trihexosylceramide (6 associations) and hexosylceramide (6 associations) classes. However, 60 SNP-lipid class associations fell below our significance threshold, with the classes diacylglycerol, GM3 ganglioside, lysophosphatidylcholine, lysoalkenylphosphatidylethanolamine, phosphatidylcholine, alkylphosphatidylethanolamine, alkenylphosphatidylethanolamine, phosphatidylserine, sphingomyelin, and triacylglycerol no longer associated (P < 5.0 × 10) with any genetic variants.Fig. 3Comparison of estimated lipidomic effect sizes between clinical lipid adjusted and unadjusted models.a Beta coefficients for independent unadjusted SNP-lipid associations (x-axis) are plotted against clinical lipid-adjusted SNP-lipid associations (y-axis). b Z-scores for unadjusted SNP-lipid associations (x-axis) are plotted against clinical lipid-adjusted SNP-lipid associations (y-axis). Z-scores for SNP associations reaching genome-wide significance (P < 5 × 10) in either the clinical lipid adjusted or unadjusted models. Variant effect signs are fixed so adjusted associations are positive. Variants showing greater (positive) associations in clinical lipid-adjusted analysis are shown in red, and variants showing reduced associations are shown in blue. Circle diameter is proportional of −log10(P) t-test of effect differences. a Beta coefficients for independent unadjusted SNP-lipid associations (x-axis) are plotted against clinical lipid-adjusted SNP-lipid associations (y-axis). b Z-scores for unadjusted SNP-lipid associations (x-axis) are plotted against clinical lipid-adjusted SNP-lipid associations (y-axis). Z-scores for SNP associations reaching genome-wide significance (P < 5 × 10) in either the clinical lipid adjusted or unadjusted models. Variant effect signs are fixed so adjusted associations are positive. Variants showing greater (positive) associations in clinical lipid-adjusted analysis are shown in red, and variants showing reduced associations are shown in blue. Circle diameter is proportional of −log10(P) t-test of effect differences. Results from multi-trait conditional and joint (mtCOJO; Supplementary Data 3, 4) analyses using clinical lipid traits (total cholesterol, HDL-cholesterol, triglycerides) GWAS results from the UK Biobank, to minimise the risk of pleiotropy/collider bias introduced by heritable covariates, were largely consistent with those of the clinical lipid-adjusted analysis (r of beta coefficients = 0.91, Supplementary Fig. 3). A comparison of the clinical lipid-adjusted Z-scores and mtCOJO Z-scores identified three gene regions (APOE, FADS1/FADS2/FADS3, TMEM229B/PLEKHH1) with substantial differences (P < 1.0 × 10) indicating the possibility of biased effect measures for the adjusted analyses in these regions. Overall, results were overwhelmingly consistent between mtCOJO and clinical lipid-adjusted analyses. Conditional analysis (sequentially conditioning on the lead SNP) identified 386 secondary signals (across both unadjusted and clinical lipid-adjusted analyses), associated with 163 lipid species/classes (Supplementary Data 7). Two gene regions, LIPC and ATP10D, each contained five independent signals (PCONDITIONAL < 5.0 × 10). The LIPC genomic region was strongly associated with phosphatidylethanolamine species and class, while ATP10D was associated with hexosylceramide species and class. The SPTLC3 region harboured four independent signals, strongly associated with sphingolipids containing a d19:1 sphingoid base. For each lipid, significantly associated SNPs were linkage disequilibrium (LD)-clumped to remove variants in LD (r > 0.1). We assessed whether the 2411 independent lipid species/class associations identified in the BHS discovery cohort (unadjusted) were validated within a combined ADNI and AIBL validation cohort meta-analysis (Validation meta-analysis). There were 273 SNP-lipid associations not available for validation in the meta-analysis, either due to lipids not available in the ADNI and AIBL cohorts; missing SNPs (and proxies) on the imputation panel; or monomorphic/very-low-frequency MAF in ADNI/AIBL. Therefore, we attempted to validate the remaining 2137 significant SNP-lipid associations. We considered a SNP-lipid association to be validated if (i) the SNP was significantly associated (P < 5 × 10) in the unadjusted BHS discovery GWAS; (ii) the direction of effect was concordant between the validation meta-analysis and the BHS discovery analysis; and (iii) the association was nominally significant (P < 0.05; less conservative) or reached the Bonferroni significance threshold (P < 2.34 × 10) in the validation meta-analysis. We identified 1474 (69.2%) SNP-lipid associations that reached nominal significance (P < 0.05), and 644 (30.1%) reaching Bonferroni-corrected significance (Supplementary Data 8). Almost all associations (>99%) had the same direction of effect, with a very strong correlation between validation meta-analysis and significant (P < 5 × 10) discovery effect sizes (r = 0.53 overall, and r = 0.80 for SNPs with MAF > 0.05 in the BHS; Supplementary Fig. 4). At a stringent significance threshold of P < 3.47 × 10 (5 × 10/144 effective lipid dimensions), the meta-analysis of all three studies identified 65,563 significant SNP-lipid associations (Supplementary Data 9), involving 499 lipid species/classes and 7600 SNPs. We identified 5658 new associations not observed in the BHS discovery GWAS alone, involving 352 lipids and 2914 SNPs. The majority of these (n = 5543; 98%) showed some evidence of association in the BHS discovery GWAS (5 × 10 < P < 5 × 10). However, 89 associations were not nominally significant (P > 0.05) in the BHS discovery GWAS, indicating that the effects observed in the meta-analysis were largely due to the AIBL and ADNI samples. For each lipid, significantly associated SNPs were LD-clumped to remove variants in LD (r > 0.1). Lead variants from the BHS discovery GWAS (adjusted and unadjusted) and conditional analyses, were clumped if the index SNPs were in linkage disequilibrium (r > 0.1). We identified 3361 independent loci-lipid associations, involving 610 lipid species/classes, each associated with between 1 and 30 independent SNPs. To identify genomic regions associated with lipid metabolism, a single dataset was produced by identifying the smallest P-value for each SNP across all lipids and analyses. LD-clumping of this dataset resulted in 667 independent genomic regions (Supplementary Data 10; filtered by column ‘Lead SNP in BHS GWAS’). This procedure was repeated, including SNP-lipid associations passing our discovery meta-analysis significance threshold (P < 3.47 × 10), resulting in 682 independent genomic regions (Supplementary Data 10; filtered by column ‘Lead SNP in Discovery-Meta analysis’), 612 of which overlap with those identified in BHS alone (737 in total). The variants within a genomic region and the lipids associated with those variants are collectively termed a genetically influenced lipotype. Using the prioritisation of candidate causal Genes at Molecular QTLs (ProGeM) framework to prioritise candidate causal genes, biologically plausible genes were identified in 573 of the 737 genomic regions (Supplementary Data 10-12), with an overlap of 498 genomic regions between genetic-based (bottom-up) and biological knowledge (top-down) based approaches. A total of 2321 SNP-gene pairs were identified, where the gene has previously been implicated in the regulation of metabolism or a molecular phenotype (Fig. 4a). Of these genes, 970 (41.8%) are present in lipid-metabolism-specific databases.Fig. 4Identification of putative causal genes using genetic prioritisation and knowledge-based approaches.Assignment of putative causal genes was performed using the ProGeM framework, incorporating genetic-based prioritisation (bottom-up), and biological knowledge-based approaches (top-down). a Venn diagram showing the number of loci with annotations for candidate genes using the distinct approaches and the overlap. Top-down annotations were divided into lipid-specific databases and generic databases. b Venn diagram of distinct genes identified in genetic-based prioritisation analysis. c Summary of putative causal genes with overlapping annotations for closest gene, protein consequences, eQTL and meQTL (left). Summary of putative causal SNP-gene pairs for which pQTL evidence was identified (right). eQTL expression quantitative trait loci, meQTL methylation quantitative trait loci, pQTL protein quantitative trait loci. Assignment of putative causal genes was performed using the ProGeM framework, incorporating genetic-based prioritisation (bottom-up), and biological knowledge-based approaches (top-down). a Venn diagram showing the number of loci with annotations for candidate genes using the distinct approaches and the overlap. Top-down annotations were divided into lipid-specific databases and generic databases. b Venn diagram of distinct genes identified in genetic-based prioritisation analysis. c Summary of putative causal genes with overlapping annotations for closest gene, protein consequences, eQTL and meQTL (left). Summary of putative causal SNP-gene pairs for which pQTL evidence was identified (right). eQTL expression quantitative trait loci, meQTL methylation quantitative trait loci, pQTL protein quantitative trait loci. A total of 62 SNPs were annotated as either missense (n = 59), stop gain (n = 2), structural interaction (n = 1), start loss (n = 1), or splice donor (n = 1) mutations. Of these, three were annotated as having a putative ‘high’ impact, and the remaining as ‘moderate’ impact. These SNPs are linked to 55 protein products (Fig. 4b). Comparing our lead SNPs and proxies against previously published eQTL associations, 2058 SNP-gene pairs were identified (Fig. 4b). Published meQTL associations revealed 879 SNP-gene pairs, 587 (66.8%) of which replicated eQTL associations. In contrast to eQTL and meQTL, the overlap of published pQTL associations was much less evident, with only 16 SNP-gene pairs identified (Fig. 4c). In total, 18 SNP-gene pairs were identified with evidence from the closest gene, protein consequences, eQTL and meQTL. The overlap of top-down and bottom-up candidates supported the annotation of 1031 SNP-gene pairs. For each of the 737 lead variants, we assessed whether they (or their proxies) had been previously reported as being associated with any lipid or metabolite. From 35 previous metabolomic/lipidomic studies (Supplementary Table 2), 228 lead variants (31%) had been reported as associating with a lipid or metabolite, resulting in 509 unreported genetically influenced lipotypes (Supplementary Data 13). We looked at the overlap between 10 hard cardiovascular disease (CVD) endpoints from the GWAS Catalog and the lead SNP (or proxy) from each of the 737 regions, identifying a total of 23 lead SNPs, or their proxies, associated (P < 5 × 10) with 10 hard CVD endpoints (Supplementary Data 14). The most frequently overlapping GWAS Catalog hard CVD endpoints were CAD (n = 14 SNPs), CVD (n = 10 SNPs), coronary artery calcification (n = 8 SNPs), and myocardial infarction (n = 8 SNPs). Three additional lead SNPs were associated with CAD in the CARDIoGRAMplusC4D and UK Biobank meta-analysis. Eighty-four lead SNPs were associated with 101 CVD-related traits, including chronic kidney disease (n = 18,) C-reactive protein (n = 14), metabolic syndrome (n = 12), body mass index (n = 8), and systolic blood pressure (n = 4). As expected, lead SNPs frequently overlapped with 186 lipid-related traits, with 99 lead SNPs or proxies observed in the GWAS Catalog. Using nominal significance (P < 0.05), we identified 243 lipid species/classes phenotypically associated with incident CAD in the BHS (Fig. 5a; Supplementary Data 15), with 88% in the positive direction. The strongest association was between TG(50:2) [NL-18:2] and incident CAD (0.311 ± 0.046, P = 1.74 × 10, FDR q = 1.09 × 10). Overall, the most strongly associated lipid species were those in the triacylglycerol, diacylglycerol, phosphatidylethanolamine, and cholesteryl ester classes.Fig. 5Genetic and phenotypic associations of the lipidome with coronary artery disease.Forest plots of lipid-coronary artery disease; circles represent effect sizes and horizontal bars represent ±standard errors. a Phenotypic associations (logistic regression; two-sided) between lipid species and incident coronary artery disease in the BHS cohort (551 cases and 3703 controls), adjusted for age, sex, and the first 10 genomic principal components. b Association of lipid species with polygenic risk for coronary artery disease. Individuals in the discovery cohort (n = 4492) were assessed for risk using the metaGRS polygenic score, consisting of ∼1.7 million genetic variants. Linear regressions (two-sided) were performed to test the association between an individual’s polygenic score and lipid species concentrations, adjusting for age, sex, and the 10 first principal components. c Genetic correlations of lipid species (n = 4492) against coronary artery disease (meta-analysis of CARDIoGRAMplusC4D and UK Biobank; 122,733 cases and 424,528 controls), performed with Linkage Disequilibrium Score Regression (LDSC; v1.0.1). Nominally significant and Benjamini–Hochberg corrected significance is indicated by light- and dark-grey circles, respectively. The 10 most significant lipid species are highlighted in blue, red, or green. Forest plots of lipid-coronary artery disease; circles represent effect sizes and horizontal bars represent ±standard errors. a Phenotypic associations (logistic regression; two-sided) between lipid species and incident coronary artery disease in the BHS cohort (551 cases and 3703 controls), adjusted for age, sex, and the first 10 genomic principal components. b Association of lipid species with polygenic risk for coronary artery disease. Individuals in the discovery cohort (n = 4492) were assessed for risk using the metaGRS polygenic score, consisting of ∼1.7 million genetic variants. Linear regressions (two-sided) were performed to test the association between an individual’s polygenic score and lipid species concentrations, adjusting for age, sex, and the 10 first principal components. c Genetic correlations of lipid species (n = 4492) against coronary artery disease (meta-analysis of CARDIoGRAMplusC4D and UK Biobank; 122,733 cases and 424,528 controls), performed with Linkage Disequilibrium Score Regression (LDSC; v1.0.1). Nominally significant and Benjamini–Hochberg corrected significance is indicated by light- and dark-grey circles, respectively. The 10 most significant lipid species are highlighted in blue, red, or green. We identified 265 lipid species/classes that showed a nominally significant (P < 0.05) association with the CAD polygenic risk score in the BHS (Fig. 5b; Supplementary Data 15). These were positive associations except for lipids in the alkenyl-phosphatidylcholine and alkenyl-phosphatidylethanolamine classes. The strongest association was observed for LPE(18:0) [sn2] (0.075 ± 0.014, P = 8.9 × 10, FDR q = 5.59 × 10). Next, we estimated the genetic correlation between lipid species/classes and CAD. Using linkage disequilibrium score regression, we identified nominally significant genetic correlations (P < 0.05) between 199 lipid species/classes and CAD, with 50 of these negatively correlated (Fig. 5c; Supplementary Data 15). The strongest genetic correlations were between TG(51:2) [NL-16:0] (0.275 ± 0.058, P = 2.22 × 10, FDR q = 8.94 × 10) and CAD. Overall, using a significance threshold of P < 0.05, we identified 134 lipid species/classes that were significantly associated in each of the three analyses—association with incident CVD (phenotypic), CAD polygenic risk (PRS), and genetic correlation. Importantly, these lipid species/classes showed concordant directions of effects in all three analyses, defining these lipid species/classes as lipid endophenotypes for CAD. We performed pairwise co-localisation analysis, within each QTL, between lipid species and CAD to assess whether they share common variants (Supplementary Data 16). We identified evidence of 43 shared variants for CAD and any lipid species (Table 1; Supplementary Note 1; Fig. 6). The strongest evidence was between CE(18:1) and CAD at the APOE rs7412 loci (H3 + H4 = 1.00; H4/H3 = 1.17 × 10). There was strong evidence for the sharing of this variant between CAD and 184 lipid species from 23 lipid classes (with and without clinical lipid adjustment). There was also strong evidence for rs603424, near a likely candidate SCD (Stearoyl-CoA desaturase), and 24 lipid species/classes (0.936 < H3 + H4 < 0.998; 16 < H4/H3 < 1.8 × 10).Table 1Genomic regions showing co-localisation with lipid species and coronary artery disease.#rsIDPositionEA/ OACo-localised lipid classesNumber of lipids co-localisedStrongest co-localisationMinimum CAD P-value in regionNearby genes1rs115911471:55505647G/TCE, DE, Hex2Cer, Hex3Cer, PC(P), SHexCer, SM, TG(O)32CE(18:1)1.86 × 10PCSK9, USP24, BSND2rs6026331:109821511G/THexCer2HexCer(d18:1/24:1)3.63 × 10PSRC1, CELSR2, MYBPHL3rs22817191:230297659C/TDG, PI, TG [NL]5DG(18:0_18:1)6.41 × 10GALNT2, PGBD5, COG24rs107798351:230299949C/TDG, TG [NL]4TG(54:2) [NL-18:0]6.41 × 10GALNT2, PGBD5, COG25rs5151352:21286057C/TCE, PC4PC(16:0_18:0)5.74 × 10APOB, TDRD15, LDAH6rs67138652:23899807A/GAC2AC(16:0)2.86 × 10KLHL29, ATAD2B, UBXN2A7rs65447132:44073881C/TCE6CE(20:1)1.84 × 10ABCG8, ABCG5, DYNC2LI18rs27361776:31586094C/TTG [NL]2TG(50:2) [NL-18:2]4.86 × 10AIF1, PRRC2A, BAG69rs412796337:44580876G/TCE1CE(18:0)1.72 × 10NPC1L1, DDX56, TMED410rs69825028:126479362C/TSM1SM(d18:0/22:0)7.67 × 10TRIB1, NSMCE2, WASHC511rs29808698:126488250C/TPC1PC(36:0)7.67 × 10TRIB1, NSMCE2, WASHC512rs350934639:107586238A/CHex3Cer2Hex3Cer(d18:1/22:0)4.00 × 10ABCA1, NIPSNAP3B, NIPSNAP3A13rs18009789:107665978C/GHex3Cer1Hex3Cer(d18:1/24:1)4.00 × 10ABCA1, NIPSNAP3B, NIPSNAP3A149:1361418709:136141870C/TCE1CE(18:0)2.03 × 10ABO, SURF6, OBP2B15rs60342410:102075479A/GAC, CE, DG, Hex2Cer, LPC, PC, PC(P), TG [NL]24LPC(16:1) [sn2]7.41 × 10PKD2L1, BLOC1S2, SCD16rs735048111:116586283C/TCE, DG2DG(18:1_18:2)5.64 × 10BUD13, ZPR1, APOA517rs658956311:116590787A/GCE, DG, TG [NL]4DG(18:0_18:1)5.64 × 10BUD13, ZPR1, APOA518rs155886111:116607437C/TCE, DG, PI, TG [NL]25TG(54:4) [NL-18:2]5.64 × 10BUD13, ZPR1, APOA519rs96418411:116648917C/GCE, DE, DG, LPI, PC, PE, PG, PI, TG [NL]64TG(54:2) [NL-18:0]7.03 × 10ZPR1, BUD13, APOA520rs65182111:116662579C/TCE, PE3CE(22:0)7.03 × 10APOA5, ZPR1, BUD1321rs116928812:121416650A/CCer(d), PC, SM6PC(36:0)1.26 × 10HNF1A, C12orf43, OASL22rs224460812:121416988A/GSM1SM(d18:0/22:0)1.26 × 10HNF1A, C12orf43, OASL23rs204308515:58680954C/TPE1PE(18:0_18:1)7.24 × 10ALDH1A2, LIPC, AQP924rs153208515:58683366A/GPE, PG16PE(18:1_18:2)7.24 × 10ALDH1A2, LIPC, ADAM1025rs107783515:58723426A/GPE7PE(15-MHDA_22:6)7.24 × 10ALDH1A2, LIPC, ADAM1026rs180058815:58723675C/TDG, LPE, PE, PE(O), PG, TG(O)19LPE(20:4) [sn1]7.24 × 10ALDH1A2, LIPC, ADAM1027rs207089515:58723939A/GCE, PE, PG, PS16PG(34:2)7.24 × 10ALDH1A2, LIPC, ADAM1028rs58813615:58730498C/TDG, PC, PC(P), PS, TG(O)10Total PC7.24 × 10ALDH1A2, LIPC, ADAM1029rs26134215:58731153C/GLPE, TG [NL]3LPE(20:4) [sn1]7.24 × 10ALDH1A2, LIPC, ADAM1030rs1244651516:56987015C/TPC, PC(O)3PC(16:0_16:0)1.19 × 10CETP, HERPUD1, NLRC531rs5615692216:56987369C/THex3Cer, PC, PC(O), PC(P), PE(P)22PC(P-16:0/16:1)1.19 × 10CETP, HERPUD1, NLRC532rs5622860916:56987765C/TCE, PC(O), PE(O), PI, TG(O)6CE(18:0)1.19 × 10CETP, HERPUD1, NLRC533rs24761616:56989590C/TPC1PC(16:0_18:3) (a)1.19 × 10CETP, HERPUD1, NLRC534rs1214954516:56993161A/GPC(O), PC(P), PE(O), PI, TG(O)11TG(O-50:1) [NL-16:0]1.19 × 10CETP, HERPUD1, NLRC535rs376426116:56993324A/CPC1PC(18:2_18:2)1.19 × 10CETP, HERPUD1, NLRC536rs1723150616:56994528C/THex2Cer, Hex3Cer, PC, PC(O), PC(P), PE(P), TG(O)40TG(O-50:1) [NL-16:0]1.19 × 10CETP, HERPUD1, NLRC537rs5628982119:11188247A/GCE, Cer(d), COH, GM3, Hex2Cer, Hex3Cer, HexCer, PC, PC(O), PC(P), SHexCer, SM60SM(35:2) (b)1.93 × 10LDLR, SMARCA4, SPC2438rs7299903319:19366632C/TCer(d)1Cer(d16:1/24:1)3.18 × 10HAPLN4, NCAN, TM6SF239rs5854292619:19379549C/TLPC, PC2LPC(20:3) [sn1]3.18 × 10TM6SF2, HAPLN4, SUGP140rs1040196919:19407718C/TCer(d), DG, LPC, PC, PE, TG [NL]38DG(18:1_20:4)3.18 × 10SUGP1, TM6SF2, MAU241rs7300106519:19460541C/GCer(d), TG [NL]3Cer(d18:1/24:0)3.18 × 10MAU2, SUGP1, GATAD2A42rs15026854819:19494483A/GCer(d)3Total Cer3.18 × 10GATAD2A, MAU2, SUGP143rs741219:45412079C/TCE, Cer(d), COH, DE, DG, GM1, GM3, Hex2Cer, Hex3Cer, HexCer, LPC, LPC(O), LPC(P), LPE(P), PC, PC(O), PC(P), PE(P), SHexCer, SM, TG [NL], TG(O)184CE(16:0)2.14 × 10APOE, TOMM40, APOC1Co-localisation analyses performed using coronary artery disease in UK Biobank and CARDIoGRAMplusC4D. Minimum CAD P-values were obtained from the meta-analysis performed in van der Harst & Verweij 2018.CAD coronary artery disease, EA effect allele, OA other allele.Genomic position based on Genome Reference Consortium Human Build 37 (GRCh37).Closest three protein coding genes to causal variant.Fig. 6Co-localisation of lipid-loci with coronary artery disease.Summary of lipid classes which contain at least one lipid specie that co-localises with coronary artery disease. Colours indicate broad lipid categories—green, sphingolipids; orange, phospholipids; blue, neutral lipids/others. Indicated variants were identified as the most likely causal variant for each of the identified co-localisation analysis. Genetic variants are ordered according to the number of co-localisations across lipid classes. Evidence of co-localisation included H3 + H4 > 0.8 and H4/H3 > 10. Genomic regions showing co-localisation with lipid species and coronary artery disease. Co-localisation analyses performed using coronary artery disease in UK Biobank and CARDIoGRAMplusC4D. Minimum CAD P-values were obtained from the meta-analysis performed in van der Harst & Verweij 2018. CAD coronary artery disease, EA effect allele, OA other allele. Genomic position based on Genome Reference Consortium Human Build 37 (GRCh37). Closest three protein coding genes to causal variant. Summary of lipid classes which contain at least one lipid specie that co-localises with coronary artery disease. Colours indicate broad lipid categories—green, sphingolipids; orange, phospholipids; blue, neutral lipids/others. Indicated variants were identified as the most likely causal variant for each of the identified co-localisation analysis. Genetic variants are ordered according to the number of co-localisations across lipid classes. Evidence of co-localisation included H3 + H4 > 0.8 and H4/H3 > 10. To further define pleiotropic effects between lipid species and CAD, we performed association analysis of 737 lead SNPs and coronary atherosclerosis in 456,486 participants of the UK Biobank (Supplementary Data 17). Eleven of the lipid-associated SNPs had genome-wide significant (P < 5 × 10) associations with coronary atherosclerosis. Adjustment for clinical lipids (total cholesterol, HDL-cholesterol, triglycerides) increased this number to 17; however, adjustment for clinical lipids using mtCOJO, which is free of the bias introduced by heritable covariates, resulted in only 14 associations with coronary atherosclerosis. Importantly, 11 of these associations were sub-genome-wide significant in the initial analysis, suggesting the presence of strong pleiotropy in these regions. After comparing effect estimates between the standard GWAS and mtCOJO clinical lipid-adjusted analysis, eight lead SNPs (with P < 5 × 10 in the standard GWAS) showed the opposite directions of associations. These regions contain prototypical lipid/lipoprotein regulating genes, such as APOE, CETP, LDLR, and PCSK9. Interestingly, for all lead SNPs with marginal association with coronary atherosclerosis (P < 1.0 × 10; with and without conditioning on clinical lipids), 43 (81%) were associated with lipid endophenotypes for CAD. By integrative analysis of the human lipidome and CAD phenotypes, we have identified candidate risk genes for CAD, providing evidence for the role of these lipid species in the development of CAD. Our high-resolution genome-wide association analyses of the human lipidome have identified 737 independent genomic regions associated with lipid metabolism, of which 509 represent genetic loci not previously associated with lipid metabolism. This is a substantial increase over previous studies with similar or larger sample sizes. Our expanded lipidomic platform utilises extensive chromatographic separation to increase the diversity of measured lipid species and distinguish lipid isomers and isotopes over those measured in previous studies. Combined with the extended pedigree study design of the BHS, we identify many rare/low-frequency variants with large effect sizes. The majority (69.2%) of the 2137 SNP-lipid associations identified in our discovery GWAS were validated in a meta-analysis of two independent cohorts. Adjustment for clinical lipids (both as standard covariates and mtCOJO analysis), confirmed that the majority of SNP-lipid associations observed were not acting directly through clinical lipids (i.e. associations were not the result of mediated pleiotropy). Discovery meta-analysis of all three studies identified an additional 5658 SNP-lipid associations (from 122 loci)—involving 352 lipid species—that were not identified in the BHS discovery GWAS alone. Overall, nearly all lipid species (95%) had at least one genome-wide significant SNP association, highlighting the genetic contribution to lipid metabolism and homeostasis. We identified 134 lipid species/classes showing consistent and significant associations with CAD when assessed with genetic correlation, phenotypic association, and PRS association. These lipids are potential endophenotypes for CAD, which can facilitate the identification of susceptibility genes. Of those loci associated with this subset of lipids, we identified 32 regions with evidence of shared genetic effects (co-localisation) with lipids and CAD. We assessed the association of lipid-loci with coronary atherosclerosis in ~456,000 individuals of the UK Biobank, considering the independence of clinical lipid traits. A total of 53 loci showed evidence of association (P < 1 × 10) in at least one analysis. Of these, 43 loci were associated with at least one of the 134 lipid species identified above. Our lipidomic profiling provided improved resolution and precision in the measurement of lipid species. Prior studies examined lipid phenotypes that were mixtures of similar, but distinct species; lacked structural characterisation of lipid species, or were contaminated through isotopic overlap. Many of the associations between lipid species and prototypical lipid regulating genes observed in our study—such as FADS1/FADS2, APOE, and LDLR—have been reported in earlier GWAS. With our expanded lipidomic profile, we have built on these earlier studies, identifying many new loci associated with lipid species and classes. Previous studies, containing mis-annotation of lipid species, report associations between SNPs in the FADS region and sphingomyelin species as containing a mono-unsaturated (16:1, 18:1, or 20:1) n-acyl chain. Here, we show the associations of sphingomyelins with SNPs in the FADS gene region are disproportional with species containing the d18:2 sphingoid base. This is supported by recent experimental evidence, suggesting FADS3 is a ceramide-specific desaturase, targeting the sphingoid bases. Early dogma suggested the dominant isoform of sphingomyelins was d18:1 leading to the aforementioned annotations (i.e. SM(d18:1/16:1)). However, chromatographic separation and characterisation identify the predominant species as SM(d18:2/16:0). While these associations are not novel per se, the additional specificity of our lipidomics methodology extends across all lipid species and classes, leading to greater confidence in defining true relationships. We also observed strong associations between specific sphingolipid isoforms and variants in the SPTLC3 gene region. Serine palmitoyltransferase long chain base subunits (SPTLC) are a series of enzymes responsible for the de novo synthesis of sphingolipids through condensation of serine with palmitoyl-CoA. Three mammalian isoforms have been identified (SPTLC1-3), which form a heterodimer in situ, of which SPTLC1 is requisite for function. The subunit SPTLC3 was discovered more recently and was thought to facilitate the synthesis of shorter-chain sphingolipids. However, we identify strong associations of SNPs in the SPTLC3 gene region with atypical sphingolipids, containing a d19:1 sphingoid base (Supplementary Data 4). This supports the recent report that SPTLC3 has broader substrate specificity, with capacity to metabolise branched isomers of palmitate (anteiso-branched-C16) leading to the synthesis of d19:1 sphingoid bases. The atypical structure of these sphingolipids has previously led to mis-annotation resulting in reported associations of the SPTLC3 gene with hydroxylated sphingomyelins, when hydroxylated sphingomyelins in the n-acyl chain are unlikely to exist in human plasma. Many genes associated with CAD risk were identified as also associated with lipid species and classes, including HMGCR, PCSK9, and LDLR (Table 1), thereby providing new avenues for investigation into mechanistic pathways. We also provide new evidence to support potential roles for genes not reaching genome-wide significance and identify possible mechanisms linking these genes to CAD; we identified strong associations between ten independent signals in the LIPC/ALDH1A2/AQP9 gene region with phosphatidylethanolamine, lysophosphatidylethanolamine, and phosphatidylglycerol lipid species independent of clinical lipids. Two lead variants were associated with functional consequences, including a start loss for gene ALDH1A2 and a missense variant for gene LIPC. The LIPC gene on chromosome 15 encodes hepatic lipase, which is functionally described as a triglyceride lipase and as possessing phospholipase A1 activity (hydrolyses sn-1 fatty acid from phospholipids). The role of hepatic lipase in lipoprotein remodelling is complex, being intimately involved in HDL-, IDL-, and chylomicron remnant-metabolism. Consequently, the role of hepatic lipase in cardiovascular disease risk has been controversial, with both pro- and anti-atherogenic mechanisms identified. These mechanisms are often viewed through the lens of lipoprotein kinetics. However, the associations of variants in the LIPC gene region with phosphatidylethanolamine species are independent of lipoprotein metabolism (Supplementary Data 3, 4)—notionally as these lipids are direct substrates for hepatic lipase. Interestingly, the strength of association of LIPC variants with coronary atherosclerosis is considerably increased when conditioned on clinical lipids (both standard adjustment and mtCOJO analyses; Fig. 7c, Supplementary Data 17) further supporting a direct mechanistic link. Phenotypically, phosphatidylethanolamine species are associated with incident CAD (Supplementary Data 15), with a direction of effect concordant with the SNP associations (Fig. 7a). Visual comparison of regional association plots and SNP effect scatter plot supports consistent effects (Figs. 7b, d). We selected independent SNPs (r < 0.05) in the LIPC gene region associated with the phosphatidylethanolamine class and assessed the similarity of effects with CAD (Fig. 7d). Inverse-variance weighted meta-analysis of SNP effects using Generalised Summary-data-based Mendelian Randomisation (GSMR) support strong pleiotropy consistent with a causal relationship (Fig. 7e).Fig. 7Genetic analysis of the LIPC gene region and circulating levels of phosphatidylethanolamine.a Lipid-wide association with the genetic variant, rs2043085, in the BHS cohort (n = 4492). Symbol colour is used to distinguish lipid classes. The symbol orientation indicates the effect sign, inverted triangles indicate negative associations, while regular triangles indicate positive associations. The dashed line indicates genome-wide significance (P < 5 × 10). b Regional association plots for Total PE and coronary artery disease (van der Harst & Verweij 2018), focusing on the LIPC region. Variants are coloured based on LD with the lead variant, rs2043085. Linkage disequilibrium plot showing correlation between variants following clumping (r > 0.8; P < 5 × 10). Variant correlations were obtained from 10,000 unrelated individuals from the UK Biobank. c Plot of genetic instrument effect sizes against Total PE (n = 4492) and coronary artery disease (n = 547,261). Variants were selected based on association with Total PE from within the LIPC region. Eight approximately independent variants were left following clumping (r > 0.05; P < 5 × 10). Generalised Summary-data based Mendelian Randomisation (GSMR) was used to estimate effect of Total PE on coronary artery disease, accounting for the variant correlations and uncertainty in both bzx and bzy. Data are presented as mean ± SE. d Forest plot of single variant tests and GSMR estimates from panel c. Data presented as mean ± 95% confidence interval. e Diagram of mediated pleiotropy, showing effect sizes estimated across multiple datasets. Exposure modifying variant effect sizes were estimated in the BHS cohort, as well as odds ratio of phosphatidylethanolamine lipid species against incident cardiovascular disease. Total effect represents the sum of genetics effects on coronary artery disease, whether mediated through phosphatidylethanolamine or not. Coronary artery disease effect size was obtained from van der Harst & Verweij 2018. Source data are provided as a Source Data file. MAF minor allele frequency, MR Mendelian randomisation, OR odds ratio, PE phosphatidylethanolamine, SNP single nucleotide polymorphism. a Lipid-wide association with the genetic variant, rs2043085, in the BHS cohort (n = 4492). Symbol colour is used to distinguish lipid classes. The symbol orientation indicates the effect sign, inverted triangles indicate negative associations, while regular triangles indicate positive associations. The dashed line indicates genome-wide significance (P < 5 × 10). b Regional association plots for Total PE and coronary artery disease (van der Harst & Verweij 2018), focusing on the LIPC region. Variants are coloured based on LD with the lead variant, rs2043085. Linkage disequilibrium plot showing correlation between variants following clumping (r > 0.8; P < 5 × 10). Variant correlations were obtained from 10,000 unrelated individuals from the UK Biobank. c Plot of genetic instrument effect sizes against Total PE (n = 4492) and coronary artery disease (n = 547,261). Variants were selected based on association with Total PE from within the LIPC region. Eight approximately independent variants were left following clumping (r > 0.05; P < 5 × 10). Generalised Summary-data based Mendelian Randomisation (GSMR) was used to estimate effect of Total PE on coronary artery disease, accounting for the variant correlations and uncertainty in both bzx and bzy. Data are presented as mean ± SE. d Forest plot of single variant tests and GSMR estimates from panel c. Data presented as mean ± 95% confidence interval. e Diagram of mediated pleiotropy, showing effect sizes estimated across multiple datasets. Exposure modifying variant effect sizes were estimated in the BHS cohort, as well as odds ratio of phosphatidylethanolamine lipid species against incident cardiovascular disease. Total effect represents the sum of genetics effects on coronary artery disease, whether mediated through phosphatidylethanolamine or not. Coronary artery disease effect size was obtained from van der Harst & Verweij 2018. Source data are provided as a Source Data file. MAF minor allele frequency, MR Mendelian randomisation, OR odds ratio, PE phosphatidylethanolamine, SNP single nucleotide polymorphism. Angiopoietin-like 3 (ANGPTL3) has been implicated in CAD risk, with a deficiency being associated with cardioprotective effects. ANGPTL3 acts as an inhibitor to two other lipases, lipoprotein lipase (LPL)—a rate-limiting enzyme in the clearance of triglyceride-rich lipoproteins—and the phospholipase endothelial lipase (LIPG). Indeed, loss of function mutations in the ANGPTL3 gene has been linked to hypolipidemia. Most previous research has focused on the lipoprotein modulating effects of ANGPTL3 through LPL. However, a recent Mendelian Randomization analysis, using NMR lipoprotein profiling, revealed a divergence in the metabolic effects of genetic variants in ANGPTL3 and LPL. We recently identified a rare frameshift deletion (rs398122988) associated with decreased ANGPTL3 protein levels in extended Mexican American families; the variant was also associated with a ~1.3 standard deviation decrease in phosphatidylinositol species. In this study, we validate this observation, with SNPs in the ANGPTL3 gene region associated with a decrease in phosphatidylinositol species, again these associations persisted even after adjustment for clinical lipids (total cholesterol, HDL-cholesterol, triglycerides). Interestingly, we also observe associations of phosphatidylinositol species with SNPs in the LIPG region, suggesting a larger metabolic effect of the ANGPTL3-LIPG pathway, at least in fasting subjects. Commonly, phosphatidylinositol species have been studied for their intracellular messaging roles following phosphorylation of the inositol ring by kinases, including PI-3-kinase, which lead to downstream cardio-metabolic effects. However, the role of phosphatidylinositol species in CVD risk is still largely unknown. We have previously observed the change in the ratio of phosphatidylinositol to phosphatidylcholine species as a predictor of CVD risk reduction from statin treatment. Further work is now required to unravel the role of phosphatidylinositol in mediating the effect of these genes on CVD risk. Limitations to the study warrant mention. First, our samples were restricted to individuals with European ancestry, complicating generalisability to individuals of non-European ancestry. Previous studies have shown conservation of lipid-metabolism genetics across different ancestries; however, future studies in non-European ancestry individuals are required. Second, adjustment for many combinations of lipid-lowering medications and doses is not practical. As a majority of lipid-lowering medications were statins and the assumption that medication dose was titrated, a single lipid species/class correction was applied to all individuals taking these medications. However, as only 2% of the BHS discovery cohort were taking lipid-lowering medications, the putative impact is unlikely to be large. A larger proportion of the two validation samples were taking lipid-lowering medications (ADNI: 49%; AIBL: 22%). Nonetheless, a substantial number of our associations were validated; therefore, the single adjustment was also unlikely to have greatly affected our results. Third, we did not have access to an independent validation sample for our discovery meta-analysis. We consider the discovery meta-analysis to be exploratory, with the potential to provide evidence of associations that can be followed up in future studies. Finally, lipidomic profiling was performed on serum in the discovery BHS and validation ADNI cohorts, whereas the validation study AIBL was plasma. While the absolute concentration of some blood metabolites may differ between plasma and serum, measurements are generally highly correlated between matrices. We have previously shown lipid associations are consistent between serum and plasma. In summary, using our expanded lipidomic profiling platform, we have investigated the largest number of targeted lipid species in a GWAS, and have reported significant genetic associations with lipid species that have not previously been reported in any genetic association studies to date. Our strategy to use lipid species as endophenotypes in the search for CVD genes is the tip of the iceberg. We have previously reported phenotypic associations of lipid species with other complex traits, including diabetes, Alzheimer’s disease, and atrial fibrillation; we believe the same integrative genomics approach may now be used to elucidate the mechanistic underpinnings of lipid metabolism in these and other complex diseases. These data now represent a valuable resource for the future exploration of the genetic analysis of the lipidome to identify lipid metabolic pathways and regulatory genes associated with complex disease and identify new therapeutic targets. To this end we provide all summary statistics and an online searchable resource of association plots of lipid species and classes with genetic variants and regional association plots with individual lipid species and classes (https://metabolomics.baker.edu.au/). Participants in the discovery cohort (n = 4492) were all participants of the 1994/95 survey of the long-running epidemiological study, the BHS, for whom genome-wide SNP data, extensive longitudinal phenotype data, and blood serum were available. The BHS is a community-based study in Western Australia that includes both related and unrelated individuals (predominantly of European ancestry) and has been described in more detail elsewhere. Informed consent was obtained from all participants and the 1994/95 health survey was approved by the University of Western Australia Human Research Ethics Committee (UWA HREC). The current study was also approved by UWA HREC (RA/4/1/7894) and the Western Australian Department of Health HREC (RGS03656). The two validation cohorts used in this study were the AIBL study and the ADNI study; both of which were established to discover biomarkers, health and lifestyle factors for the development, early detection, and tracking of Alzheimer’s disease. The AIBL study is a longitudinal study which recruited 1112 individuals aged over 60 years within Australia. Time points for blood/data collection were every 18 months from baseline. For each individual, lipidomic data obtained from the earliest blood collection was used. At baseline, 768 individuals were characterised as cognitively normal, 133 with mild cognitive impairment and 211 with Alzheimer’s disease. The ADNI study is a longitudinal study, starting in 2004 and recruited 800 individuals at baseline, from sites across the United States of America and Canada. Serum samples obtained at baseline were analysed. Study data analysed here were obtained from the ADNI database, which is available online (http://adni.loni.usc.edu/). For the lipidomics analysis, the AIBL study was deemed low risk (The Alfred Ethics Committee; Project 183/19), and the ADNI study was deemed ‘research not involving human subjects’ (Duke Institute review board; ID:Pro00053208). Targeted lipidomic profiling was performed using liquid chromatography coupled electrospray ionisation-tandem mass spectrometry from fasting blood serum (BHS discovery), fasting blood plasma (AIBL validation), and a combination of fasting and non-fasting blood serum (ADNI validation; 90% fasting, 10% non-fasting). We quantified 596 lipid species (from 33 lipid classes) in the BHS discovery cohort, 573 lipid species (from 32 lipid classes) in the validation AIBL cohort, and 581 lipid species (from 32 lipid classes) in the validation ADNI cohort. Due to strict quality control, lipid species may be removed from a dataset and typically represent very low abundant species and/or those requiring near-optimal chromatographic separation. All lipid classes were consistent across the studies, except for the Oxidised sterol ester which was only available in the discovery BHS cohort. Overall, 596 lipid species were quantified; 570 of which were quantified within all three cohorts; five lipid species were present only within BHS and ADNI; and 21 lipid species were present only in the BHS cohort (Supplementary Data 1, 2). Lipidomic profiling of each cohort was performed using the standardised methodology described by Huynh et al.. Lipidomic profiling has been described previously for BHS and ADNI/AIBL. Briefly, 10 μL of serum/plasma was spiked with an internal standard mix (Supplementary Data 1) and lipid species were isolated using a single-phase butanol:methanol (1:1; BuOH:MeOH) extraction. Analysis of serum/plasma extracts was performed on an Agilent 6490 QqQ mass spectrometer with an Agilent 1290 series HPLC, as previously described. Mass spectrometry settings and transitions for each lipid class are shown in Supplementary Data 1. A total of 497 transitions, representing 596 lipid species (BHS discovery), 573 lipid species (AIBL validation), and 581 lipid species (ADNI validation), were measured using dynamic multiple reaction monitoring (dMRM), where data was collected during a retention time window specific to each lipid species. Raw mass spectrometry data were analysed using MassHunter Quant B08 (Agilent Technologies). Lipid concentrations were calculated by relating the area under the chromatographic peak, for each lipid species, to the corresponding internal standard. Correction factors were applied to adjust for differences in response factors, where these were known. In-house pipelines were used for quality control and filtering of lipid concentrations. Across the entire BHS dataset, only three missing values were evident. Lipids below the limit of detection (missing values) were imputed to half the minimum observed value. To remove technical batch variation, the lipid data in each analytical batch (approximately 486 samples per batch) was aligned to the median value in pooled plasma quality control samples included in each analytical run. Unwanted variation in the discovery cohort was identified using a modified remove unwanted variation-2 (RUV-2) approach. In brief, lipid data were residualised in a linear mixed model, against age, sex, body mass index (BMI), clinical lipids and the genetic relatedness matrix (described below) as the random effects. Principal component analysis was performed on the residualised data. The first two components showed clear trends along with samples in collection order. Therefore, variation associated with these first two principal components was removed from the original dataset. Lipid class totals were generated by summing the concentration of the individual species within each class. Validation cohorts were processed in a similar manner. Details of the BHS data collection have been published previously. Serum cholesterol and triglycerides were calculated by standard enzymatic methods on a Hitachi 747 (Roche Diagnostics, Sydney, Australia) from fasting blood collected in 1994/95. HDL-cholesterol was determined on a serum supernatant after polyethylene glycol precipitation using an enzymatic cholesterol assay and LDL-cholesterol was estimated using the Friedewald formula. Height and weight (used to calculate BMI) were collected from participants at the time of the interview (1994/95). The use of lipid-lowering medication was recorded at the time of the interview (1994/95). Diagnosis of incident CAD was defined as either hospitalisation or death due to CAD (ICD9: 410-414; ICD10: I20-I25) after the blood collection date (and until June 2015). Hospitalisations and deaths were identified from the Western Australian Department of Health Hospital Morbidity Data Collection and Death Registrations. For individuals taking lipid-lowering medication (BHS, n = 108; AIBL, n = 198; ADNI, n = 328), lipid species and clinical lipid concentrations were adjusted using previously identified effects of lipid-lowering medication. Changes in lipid species and clinical lipids following one year of statin use were calculated from a placebo randomised controlled trial (LIPID study; n = 4991). To calculate correction factors, lipid measures were centred and scaled by the mean and standard deviation of baseline measures (prior to statin usage), and the change in lipid abundance was calculated and regressed on age, sex, BMI, and statin usage. Statin usage beta coefficients (effect of the lipid-lowering medication) were added to standardised lipid species concentrations of the individuals taking lipid-lowering medication in the current study. For lipid species present in both this study and the LIPID study (overlap of 314 lipid species), species-specific correction factors were calculated. For those lipid species not measured in the LIPID study (n = 282), class-specific correction factors were used in place of species-specific correction factors i.e. a ceramide-specific correction factor (average beta coefficient of overlapping ceramide species) was used for ceramide species not measured in the LIPID study. Due to the large proportion of ADNI participants taking lipid-lowering medication, we performed a sensitivity analysis, comparing the above correction against residualising lipid concentrations adjusting for medication usage as a covariate (Supplementary Note 2). For the BHS discovery cohort, genotyping was performed on the Illumina Human 610 K Quad-Bead Chip (Illumina Inc., San Diego, CA, USA) at the Centre National de Genotypage in Paris, France (n = 1468), and on the Illumina 660 W Quad Array Bead Chip (Illumina Inc., San Diego, CA, USA) at the PathWest Laboratory Medicine WA (Nedlands, WA, Australia (n = 3428). Complete linkage clustering based on pairwise identity by state distance in PLINK showed no batch effects, therefore the batches were merged. Standard genotype data quality control was performed as described previously. Briefly, individuals were excluded if: >3% of SNP data were missing (n = 11), reported sex did not match genotyped sex (n = 48), duplicates (n = 123), missing phenotype data (n = 11), or >5 standard deviations above/below mean heterozygosity (n = 28). Individuals with non-European ancestry (n = 4) were also excluded. To prepare genotype data for imputation, SNPs were excluded if: call rates <95%, minor allele count <10, deviations from HWE (P < 5.0 × 10), no matching Haplotype Reference Consortium (HRC) reference panel SNP, palindromic (A/T, G/C) SNPs with MAF greater than 0.4 from the HRC (n = 5), and SNPs with >0.2 MAF difference compared to HRC (n = 150). After quality control, SNP data was available for 513,634 SNPs. Imputation was performed to the HRC reference panel using the Michigan Imputation Server. Following imputation, 39,117,105 SNPs were available for analysis. We excluded variants if the number of copies of the minor allele <5 or if imputation quality (r) <0.3. This resulted in 13,887,524 variants available for analysis. Genotyping in ADNI was performed on the Human 610-Quad BeadChip (Illumina, Inc., San Diego, CA). Following standard quality control procedures performed in Plink (minimum SNP and individual call rate >95%, MAF > 0.05, HWE test P > 1 × 10), the sample was imputed to the 1000 Genomes Phase 3 reference panel using Impute2, with pre-phasing using ShapeIT. Genotyping in AIBL was performed on the Infinium OmniExpressExome array (Illumina, Inc., San Diego, CA). Quality control procedures were performed in Plink. After removing individuals with ambiguous sex, Plink was used to remove individuals with call rate <0.90; SNPs were removed if call rate <0.95, HWE test P < 1.0 × 10, or MAF < 0.05. SNPs were flipped to the positive strand before imputation to the 1000 Genomes Phase 3 reference panel using the Michigan Imputation Server (using Minimac 4). Both the AIBL and ADNI validation cohorts were restricted to individuals of non-Hispanic European ancestry, based on projection onto the 1000 Genomes reference panel. The discovery sample, BHS, used in this study consisted of related and unrelated individuals; therefore, all analyses included a genetic relatedness matrix. Twenty-two genetic relatedness matrices were calculated. First, a hard-call set of imputed SNPs was created in Plink (i.e. SNP genotypes were called if SNP imputation quality r > 0.8 and if genotype probability >0.9). The HLA region on chromosome 6 was also excluded. SNPs were then pruned in Plink using ‘indep-pairwise 500 50 0.3’ [window of size 500, moving 50 SNPs along each time, removing variants with r > 0.3] to create a set of 486,553 independent SNPs. Twenty-two genetic relatedness matrices were created (using the option ‘gk 1’ which specifies a centred relatedness matrix), with each omitting one chromosome, in GEMMA. Genome-wide association analyses for the 596 lipid species and 33 lipid classes in the discovery cohort were performed using imputed genotype dosages in linear mixed models, as implemented in GEMMA. To avoid proximal contamination, analyses were performed using genetic relatedness matrices implementing a leave-one-chromosome out scheme. Analyses were performed using rank-based inverse normal transformed residuals, after adjustment by age, sex, age, age*sex, age*sex, and the first 10 principal components (generated from Eigenstrat). Validation cohorts, ADNI and AIBL, were analysed using an additive linear model, as implemented in Plink. Analyses were performed using rank-based inverse normal transformed residuals, after adjustment by age, sex, age, age*sex, age*sex, study-specific covariates (including fasting status for ADNI) and a number of principal components deemed sufficient to capture population structure. Meta-analysis between all three studies was performed using an inverse-variance weighted fixed-effects model, as implemented in METAL. Due to the correlation between lipid species, the effective number of tests was calculated as the number of principal components required to explain at least 95% variance of the lipidome (144 components). Statistical significance was defined using the standard genome-wide significance (P < 5 × 10) in the BHS discovery analysis, P < 0.05 in AIBL/ADNI validation, and P < 3.47 × 10 in the three-study meta-analysis (5 × 10/144 lipid dimensions; Bonferroni correction using the effective number of tests). A more stringent threshold was used for the meta-analysis due to the lack of validation samples available. For each lipid, significantly associated SNPs were LD-clumped (r > 0.1) using correlation measures obtained from 10,000 unrelated individuals from the UK Biobank, the 1000 Genomes, or the BHS. A singular dataset was created by retrieving the smallest P-value across all analyses. This dataset was LD-clumped (r > 0.1) to determine the number of independent genomic regions. For each locus, a regional association plot was produced using LocusZoom. Conditional analysis was performed to detect independent association signals at each genome-wide significant loci using GEMMA. For each lipid, we iteratively clumped regions within a 2 Mb window centred on the lead SNP until no more genome-wide significant associations were left. Regions with overlapping windows were merged. Conditional analysis was iteratively performed, including the lead variant as a covariate until no more conditionally independent signals (P < 5 × 10) remained. Within the discovery cohort, to determine whether SNP-lipid associations were independent of clinical lipid traits (total cholesterol, HDL-cholesterol, triglycerides), all SNPs were tested with and without adjustment for clinical lipid traits. We compared loci effect sizes between analyses run with and without clinical lipid adjustment using a pooled standard deviation t-test (Supplementary Note 3). Bonferroni adjustment (0.05/number of loci) was used to identify loci which differed substantially following adjustment. As adjusting for heritable covariates can introduce collider bias, we further validated these using multi-trait conditional and joint analysis (mtCOJO), conditioning on GWAS summary-level data for clinical lipids obtained from the UK Biobank. Proxies for lead SNPs were found by identifying those in high LD (r > 0.8) within the BHS dataset; in an unrelated subset of white, British individuals from the UK Biobank; or in the 1000 Genomes. Lead SNPs and their proxies were annotated using SNPEff. SNiPA database v3.3 was used to retrieve the combined annotation dependent depletion (CADD) score. Expression QTL associations (cis-eQTL) were obtained from GTEx (release v8) and eQTLGen (release 2019-12-20). SNiPA metabolite QTL (mQTL) associations were supplemented with mQTL associations reported in PhenoScanner and recently published lipidomic GWAS. SNiPA protein QTL (pQTL) associations were supplemented with cis-pQTL associations from ref. . Methylation QTL (meQTL) associations were obtained from ref. . A locus was defined as previously unreported if the lead SNP or its proxies have not been identified as an mQTL or lipid-related trait loci. Putative causal genes, for each loci, were identified using a slightly modified approach to that previously described (ProGeM). For the bottom-up approach, the three closest protein coding genes (within a 1 Mb window) were identified, for each lead SNP. Genes were noted if a lead SNP or its proxies were annotated by SNPEff as missense, start loss, stop gain, or with an annotation impact as High. As performed by ProGeM, the top-down analysis reports genes within 500 kb of the lead SNP that are present in a curated database of known metabolic-related genes. A list of primary candidates was generated based on the overlap of top-down and bottom-up genes. To assess whether our lead SNPs were previously associated with CVD-related traits, we performed a look-up within the GWAS Catalog v1.02 (release 2020-08-26) of 10 hard CVD endpoints, 72 CVD-related traits, and 141 lipid-related traits. We also performed a look-up against a meta-analysis of CAD between CARDIoGRAMplusC4D and UK Biobank. Within the discovery cohort, the association of lipid species with incident CAD was assessed using logistic regression, adjusting for age, sex, and the first 10 genomic principal components. Prevalent CAD cases were removed prior to analysis; defined as individuals hospitalised with CAD between the start of the Hospital Morbidity Data Collection (1970), and an individual’s serum collection date. Incident CAD events (CAD hospitalisations or death) were included up to the end of follow-up (July 2015). Results are displayed as log-odds ratios. Polygenic risk for CAD was calculated for each individual in the discovery cohort using the metaGRS polygenic score, consisting of ~1.7 million genetic variants. Linear regression in R was performed to test the association between an individual’s polygenic score and lipid species concentrations, adjusting for age, sex, and the 10 first principal components. Genetic correlations of lipid species against CAD were assessed using Linkage Disequilibrium Score Regression (v1.0.1). Regression weights and scores were obtained from 1000 Genomes European data, as previously described. Summary statistics from all datasets were restricted to SNPs from the HapMap 3 panel, with 1000 Genomes European MAF greater than 5%. Where available, SNPs were filtered to an imputation quality r > 0.9. Similarly, SNPs were removed if the reported MAF deviated from 1000 Genomes European MAF by greater than 0.1. Summary statistics for CAD were obtained from the meta-analysis of CARDIoGRAMplusC4D and UK Biobank by van der Harst and Verweij. Due to no overlapping samples between BHS and other summary results, the genetic covariance intercept was constrained to 0. Co-localisation between lipid species genome-wide significant loci and CAD was performed using the R package COLOC. For each loci, all variants within a 400 kb window centred on the lead SNP were selected. Priors were kept at default settings. Evidence for shared variants was determined as the posterior probability of both traits containing causal variants in the region (H3 + H4 > 0.8) and a larger probability of a shared variant (H4/H3 > 10). Sensitivity analysis for regions with shared variants is shown in Supplementary Note 1. Lead SNPs (or proxies) were tested for association with coronary atherosclerosis in the UK Biobank. In a subset of white, British individuals (n = 456,486), electronic health records (updated 14 December 2020) were converted into PheCodes using the R package PheWAS. Coronary atherosclerosis (phecode 411.4) was exported for genome-wide association analysis. FastGWA was used to assess the association of lipid-loci with these phenotypes, adjusting for age, sex, age, age*sex, age*sex, the first 20 principal components as provided by the UK Biobank, and the genetic relatedness matrix as the random effect. The analysis was repeated, additionally adjusting for clinical lipids (total cholesterol, HDL-cholesterol, triglycerides; measurements obtained from the first available blood collection). Individuals with missing values were excluded from the analysis. As clinical lipids are heritable, mtCOJO analysis was also performed using GWAS summary statistics obtained above. Further information on research design is available in the Nature Research Reporting Summary linked to this article.
|
PMC7417482
|
A Quantitative Analysis of Cellular Lipid Compositions During Acute Proteotoxic ER Stress Reveals Specificity in the Production of Asymmetric Lipids
|
The unfolded protein response (UPR) is central to endoplasmic reticulum (ER) homeostasis by controlling its size and protein folding capacity. When activated by unfolded proteins in the ER-lumen or aberrant lipid compositions, the UPR adjusts the expression of hundreds of target genes to counteract ER stress. The proteotoxic drugs dithiothreitol (DTT) and tunicamycin (TM) are commonly used to induce misfolding of proteins in the ER and to study the UPR. However, their potential impact on the cellular lipid composition has never been systematically addressed. Here, we report the quantitative, cellular lipid composition of Saccharomyces cerevisiae during acute, proteotoxic stress in both rich and synthetic media. We show that DTT causes rapid remodeling of the lipidome when used in rich medium at growth-inhibitory concentrations, while TM has only a marginal impact on the lipidome under our conditions of cultivation. We formulate recommendations on how to study UPR activation by proteotoxic stress without interferences from a perturbed lipid metabolism. Furthermore, our data suggest an intricate connection between the cellular growth rate, the abundance of the ER, and the metabolism of fatty acids. We show that Saccharomyces cerevisiae can produce asymmetric lipids with two saturated fatty acyl chains differing substantially in length. These observations indicate that the pairing of saturated fatty acyl chains is tightly controlled and suggest an evolutionary conservation of asymmetric lipids and their biosynthetic machineries.Biological membranes are complex assemblies of proteins and lipids forming the boundary of cellular life and compartmentalizing biochemical processes in different organelles (van Meer et al., 2008; Bigay and Antonny, 2012). A major fraction of cellular bioactivity is localized to biological membranes: one third of all proteins and the majority of therapeutic drug targets are either membrane embedded or membrane associated (Uhlén et al., 2015). The interactions, activities, and subcellular localizations of membrane proteins are modulated by their complex and dynamically regulated environment (Lee, 2004; Phillips et al., 2009; Lorent et al., 2017). The lipidome of a eukaryotic cell comprises hundreds, if not thousands, of lipid species and it can be remodeled upon dietary perturbation, by the growth phase, and in response to external cues such as temperature or nutrient availability (Shevchenko and Simons, 2010; Klose et al., 2012; Casanovas et al., 2015; Han, 2016; Levental et al., 2020). This membrane responsiveness down to the level of individual lipid species is essential to sustain cellular fitness, by maintaining physicochemical membrane properties such as fluidity, permeability, phase behavior, and surface charge density in a regime acceptable for membrane function (Bigay and Antonny, 2012; Sezgin et al., 2017; Ernst et al., 2018; Harayama and Riezman, 2018). Our understanding of these complex remodeling processes, their purposes and the underlying principles, remains rather limited. The endoplasmic reticulum (ER) is the central hub for membrane biogenesis in eukaryotic cells (van Meer et al., 2008). The vast majority of membrane proteins is targeted to ER-localized machineries for membrane insertion (Aviram and Schuldiner, 2017). Likewise, a major fraction of membrane lipids including sterols, glycerophospholipids, and ceramides is produced in the ER (van Meer et al., 2008). In the past years it became increasingly clear that protein quality control and lipid metabolism are intimately connected both on the cellular and the molecular level (Jonikas et al., 2009; De Kroon et al., 2013; Stordeur et al., 2014; Volmer and Ron, 2015; Fun and Thibault, 2020; Goder et al., 2020). A prominent example is the unfolded protein response (UPR) (Walter and Ron, 2011). Both an accumulation of unfolded protein in the lumen of the ER and stiffening of the ER membrane due to lipid imbalances serve as activating signals for the UPR (Halbleib et al., 2017; Karagöz et al., 2017; Adams et al., 2019; Preissler and Ron, 2019). How precisely these activating signals from the lumen of the ER and the ER membrane are integrated by the transducers of the UPR is matter of active debate (Volmer and Ron, 2015; Covino et al., 2018; Fun and Thibault, 2020). Once activated, the UPR increases the size of the ER and its folding capacity in order to reestablish ER homeostasis even under adverse conditions (Bernales et al., 2006; Schuck et al., 2009). This is accomplished by a global attenuation of protein production (Walter and Ron, 2011), by upregulating the ER-associated degradation machinery, and by increasing the number of ER chaperones (Cox et al., 1993; Jonikas et al., 2009). At the same time, the UPR induces the expression of a large number of genes involved in membrane-related processes such as lipid biosynthesis, membrane protein sorting, and vesicular traffic (Travers et al., 2000). Unbiased genetic screens and targeted perturbations of lipid metabolism have clearly established the mutual dependency of the UPR and lipid metabolism (Jonikas et al., 2009; Pineau et al., 2009; Schuck et al., 2009; Promlek et al., 2011; Thibault et al., 2012; Surma et al., 2013). Given its central importance for ER homeostasis and cell physiology, it is not surprising that the UPR plays also a crucial role in pathologic processes such as viral infections, neurodegeneration, and cancer (Wang and Kaufman, 2012,Wang and Kaufman, 2014; Hetz et al., 2019). Metabolic diseases associated with chronic UPR signaling such as type II diabetes and non-alcoholic steatohepatitis (Kaufman, 2002; Fonseca et al., 2009) are historically studied with a focus on the role of unfolded, soluble proteins in the ER lumen while the contribution of signals from the ER membrane remains understudied. The eukaryotic model organism Saccharomyces cerevisiae (S. cerevisiae) has facilitated the identification of numerous key components of the secretory pathway, lipid metabolism, and the proteostasis network (Novick et al., 1980; Wolf and Schäfer, 2006; Henry et al., 2012; De Kroon et al., 2013). In contrast to metazoans, the UPR in S. cerevisiae relies on a single, ER-localized UPR transducer (Kimata and Kohno, 2011): the Inositol-requiring enzyme 1 (Ire1p). It is conserved from yeast to humans and comprises an N-terminal sensor domain facing the ER-lumen, a single transmembrane helix, and cytosolic effector domains with kinase and RNase functions. The formation of dimers and higher oligomers of Ire1p during stress from unfolded proteins or from the ER membrane (Kimata et al., 2007; Korennykh et al., 2009; Halbleib et al., 2017) triggers the trans-autophosphorylation of the cytosolic kinase domain and the activation of the adjacent RNase domain. The RNase activity of Ire1p contributes to an unconventional splicing of the HAC1 mRNA in the cytosol (Cox and Walter, 1996; Mori et al., 1996) thereby facilitating the production of the active transcription factor Hac1p regulating several hundred UPR-target genes. For studying and assaying the UPR, it is common practice to stress the cells acutely either with dithiothreitol (DTT), a reducing agent interfering with disulfide bridge formation in the ER-lumen, or tunicamycin (TM), a natural inhibitor of the N-linked glycosylation of proteins in the ER (Azim and Surani, 1979). It is generally assumed that DTT and TM exclusively cause proteotoxic ER stress. However, the impact of DTT and TM on the cellular lipid composition has never been systematically tested. Here, we have studied the impact of DTT or TM on the lipidome of S. cerevisiae in both rich and synthetic medium. Serendipitously, we find evidence for a remarkable selectivity of S. cerevisiae in the generation and metabolism of highly asymmetric glycerophospholipids with one saturated, medium fatty acyl chain (C10 or C12) and a long, saturated one (C16 or C18). Despite a high overall abundance of unsaturated fatty acyl chains (C16:1 or C18:1), we find an almost exclusive paring of two saturated fatty acyl chains in these asymmetric glycerophospholipids, thereby implying a strong, inherent selectivity of acyl chain pairing. With respect to the UPR, we find that (1) DTT and TM impair cellular growth, (2) the medium has a significant impact on the cellular lipidome thereby potentially tuning the sensitivity of the UPR, (3) DTT at growth-inhibitory concentrations causes a substantial and rapid remodeling of the lipidome in rich medium, and (4) TM under our experimental conditions has only a marginal impact on the cellular lipidome in both synthetic and rich medium. Based on these findings, we provide a guideline to predictably and unambiguously activate the UPR by proteotoxic stress, whilst minimizing potential artifacts from lipid bilayer stress. Yeasts used were the standard laboratory wild-type S. cerevisiae strain BY4741 MATa his3Δ1 leu2Δ0 met15Δ0 ura3Δ0 and the ire1Δ-derivative BY4741 MATa; ura3Δ0; leu2Δ0; his3Δ1; met15Δ0; YHR079c::kanMX4. Chemicals and solvents of HPLC/LC-MS analytical grade were used. TM (#T7765), DTT (#D0632), and ammonium bicarbonate (#9830) were purchased from Sigma-Aldrich. Ammonium sulfate (#9218) was purchased from Carl Roth. All media were prepared according to standard protocols (Dymond, 2013). D(+)-Glucose (#HN06, tryptone/peptone (#8952), and yeast extract (#2363) were purchased from Carl Roth. Yeast nitrogen base (YNB) (#CYN0602), agar–agar (#AGA03), and complete supplement mixture (CSM complete) (#DCS0019) were purchased from FORMEDIUM. Cells were cultivated under constant agitation at 30°C at 220 rpm, if not stated otherwise. Every lipidomic sample started from an individual, single colony on either yeast peptone dextrose (YPD) or synthetic complete dextrose (SCD) agar plates. A single colony was used to inoculate a pre-culture, which was then cultivated overnight for 21 h in either YPD or SCD liquid medium. The resulting stationary culture was used to inoculate a main culture in fresh medium to a final OD600 of 0.1. When the culture reached an OD600 of 0.8 ± 0.05, the cells were either stressed with DTT, TM or left untreated. DTT was used at a final concentration 8 mM and 2 mM in YPD and SCD, respectively. TM was used at a final concentration of 1.0 μg/ml and 1.5 μg/ml in YPD and SCD, respectively. After an additional hour of cultivation, 20 OD units of cells were harvested by centrifugation (3,500 × g, 5 min, 4°C), and washed three-times with ice-cold 155 mM ammonium bicarbonate supplemented with 10 mM sodium azide in 1.5 ml reaction tubes using quick centrifugation (10.000 × g, 20 s, 4°C). The resulting cell pellets were snap-frozen with liquid nitrogen and stored for up to 4 weeks at −80°C. Prior to cell lysis, pellets were thawed on ice and then resuspended in 1 ml 155 mM ammonium bicarbonate. 200 μl zirconia beads were added to the suspension and cells were disrupted by vigorous shaking using a DisruptorGenie for 10 min at 4°C. 500 μl of the resulting lysate was snap-frozen and used for further analysis via shotgun mass spectrometry. Cultures in YPD and SCD were inoculated precisely as described for lipidomic experiments. The density of the culture was monitored over a prolonged period of time by determining the OD600 for up to 5 h after they had reached an OD600 = 0.8. For determining the doubling time of an exponentially growing culture, all data points with an OD600 between 0.2 and 2.5 were considered. The data were fitted to the exponential (Malthusian) growth function using Prism 8 for macOS Version 8.4.1. Stationary overnight cultures in YPD were used to inoculate a pre-culture in either YPD or SCD to an OD600 of 0.2. The cells were then cultivated for 6 h to reach the exponential growth phase. These cultures were used to inoculate a main culture in a 96-well plate to an OD600 of 0.01 using fresh medium (either YPD or SCD) containing different concentrations of DTT. After cultivation for 16 h at 30°C with no agitation, the final OD600 was determined after intense mixing of the culture using a microplate reader (Tecan Microplate Reader Spark). Mass spectrometry-based lipid analysis was performed by Lipotype GmbH (Dresden, Germany) as described (Ejsing et al., 2009; Klose et al., 2012). Lipids were extracted using a two-step chloroform/methanol procedure (Ejsing et al., 2009). Samples were spiked with internal lipid standard mixture containing: CDP-DAG 17:0/18:1, ceramide 18:1;2/17:0 (Cer), diacylglycerol 17:0/17:0 (DAG), lyso-phosphatidate 17:0 (LPA), lyso-phosphatidylcholine 12:0 (LPC), lyso-phosphatidylethanolamine 17:1 (LPE), lyso-phosphatidylinositol 17:1 (LPI), lyso-phosphatidylserine 17:1 (LPS), phosphatidate 17:0/14:1 (PA), phosphatidylcholine 17:0/14:1 (PC), phosphatidylethanolamine 17:0/14:1 (PE), phosphatidylglycerol 17:0/14:1 (PG), phosphatidylinositol 17:0/14:1 (PI), phosphatidylserine 17:0/14:1 (PS), ergosterol ester 13:0 (EE), triacylglycerol 17:0/17:0/17:0 (TAG), stigmastatrienol, inositolphosphorylceramide 44:0;2 (IPC), mannosyl-inositolphosphorylceramide 44:0;2 (MIPC), mannosyl-di-(inositolphosphoryl)ceramide 44:0;2 (M(IP)2C). After extraction, the organic phase was transferred to an infusion plate and dried in a speed vacuum concentrator. 1st step dry extract was resuspended in 7.5 mM ammonium acetate in chloroform/methanol/propanol (1:2:4, V:V:V) and 2nd step dry extract in 33% ethanol solution of methylamine in chloroform/methanol (0.003:5:1; V:V:V). All liquid handling steps were performed using Hamilton Robotics STARlet robotic platform with the Anti Droplet Control feature for organic solvents pipetting. Samples were analyzed by direct infusion on a QExactive mass spectrometer (Thermo Scientific) equipped with a TriVersa NanoMate ion source (Advion Biosciences). Samples were analyzed in both positive and negative ion modes with a resolution of Rm/z = 200 = 280000 for MS and Rm/z = 200 = 17500 for MSMS experiments, in a single acquisition. MSMS was triggered by an inclusion list encompassing corresponding MS mass ranges scanned in 1 Da increments (Surma et al., 2015). Both MS and MSMS data were combined to monitor EE, DAG, and TAG ions as ammonium adducts; PC as an acetate adduct; and CL, PA, PE, PG, PI, and PS as deprotonated anions. MS only was used to monitor LPA, LPE, LPI, LPS, IPC, MIPC, M(IP)2C as deprotonated anions; Cer and LPC as acetate adducts and ergosterol as protonated ion of an acetylated derivative (Liebisch et al., 2006). Data were analyzed by Lipotype GmbH using an in-house developed lipid identification software based on LipidXplorer (Herzog et al., 2011, 2012). Data post-processing and normalization were performed using an in-house developed data management system. Only lipid identifications with a signal-to-noise ratio > 5, and a signal intensity five-fold higher than in corresponding blank samples were considered for further data analysis. We investigated the impact of two potent, proteotoxic inducers of ER stress, namely DTT and TM, on cellular growth and the cellular lipid composition. Our ultimate goal was to faithfully induce proteotoxic stress in the lumen of the ER whilst minimizing potential artifacts from aberrant ER lipid compositions. We wanted to know the impact of DTT and TM on cellular growth. To this end, we cultivated the S. cerevisiae wildtype (WT) strain BY4741 and isogenic ire1Δ cells in either rich medium (YPD) or synthetic medium (SCD) to the exponential growth phase. Using these cells, we inoculated fresh cultures in a 96-well plate to an OD600 of 0.01 in the respective medium supplemented with various concentrations of DTT and TM. After overnight cultivation, cellular growth was assayed via the OD600 (Figures 1A–D). Not surprisingly, WT cells are more resistant to DTT- or TM-induced ER stress than ire1Δ in both rich and synthetic medium (Figures 1A–D). This suggests that UPR-activation via Ire1p contributes to cellular fitness under conditions of prolonged proteotoxic stress. Notably, the choice of the medium affects the growth-inhibitory concentrations of DTT and TM such that higher initial concentrations of DTT are required to inhibit overnight growth in rich medium, while lower concentrations are sufficient in minimal medium (Figures 1A,B). In contrast, lower concentrations of TM are required to block overnight growth in rich versus synthetic medium (Figures 1C,D). The underlying reasons remain unclear. Among other possibilities, the medium might affect the drug per se (e.g., DTT oxidation), the uptake and extrusion of the compound, or -via diverse mechanisms- the cellular resistance to proteotoxic stress. Nevertheless, our data help choosing appropriate concentrations to effectively inhibit overnight growth for each medium and both drugs. Acute and prolonged ER stress inhibits cellular growth. (A–D) Determination of the minimal inhibitory concentration (MIC) of DTT and TM for the indicated strains. An overnight culture of the indicated strains in rich medium (YPD) was used to inoculate a main culture in either YPD (E) or SCD (F) to an OD600 of 0.2. After 6 h of cultivation to reach the exponential growth phase, a fresh culture in a 96-well plate was inoculated to an OD600 = 0.01 and adjusted to the indicated, final concentrations of either DTT or TM. After 16 h of cultivation, the OD600 as a measure for the overnight growth was plotted against the concentration of the proteotoxic drug. The data in panels (A–D) are plotted as the average ± standard deviation (SD) from three independent experiments (n = 3) with technical duplicates. (E–H) BY4741 WT and the isogenic ire1Δ strain were cultivated in either synthetic (SCD) or rich (YPD) medium at 30°C. The main cultures were inoculated to an OD600 of 0.1 from a stationary overnight culture in the respective medium. Cell proliferation is monitored by plotting the OD600 against the time of cultivation. At an OD600 of 0.80 ± 0.05 the cells were either left untreated (black) or stressed with either TM (green) or DTT (orange). An arrow indicates the time point of drug addition. DTT was used at a final concentration 8 mM and 2 mM in YPD and SCD, respectively. TM was used at a final concentration of 1.0 μg/ml and 1.5 μg/ml in YPD and SCD, respectively. The data in panels (A–D) are from a single, representative experiment. Raw data can be found in the Supplementary Data Sheet 2. Next, we wanted to study the impact of acute ER stress on cellular growth. We cultivated WT cells in rich (YPD) and synthetic (SCD) medium (Figures 1E,F) under conditions most commonly used for studying the biology of S. cerevisiae (Sherman, 2002). We inoculated liquid cultures to an OD600 of 0.1 using stationary, overnight cultures in the respective medium and then followed the cellular growth over time. When the cultures reached an OD600 of 0.75 to 0.8, the cells were either left untreated or stressed with DTT or TM at concentrations causing a near-complete inhibition of overnight growth to account for the different dose–response curves in different media (Figures 1A–D). Specifically, DTT was used at a concentration of 8 mM and 2 mM, while TM was used at a concentration of 1.0 μg/ml and 1.5 μg/ml in rich and synthetic medium, respectively. Notably, 8 mM of DTT has previously been used to study ER membrane expansion in stressed cells, while 1–2 μg/ml of TM are known to reorganize Golgi traffic and mitochondrial enlargement by activating the UPR (Bernales et al., 2007; Schuck et al., 2009; Hsu et al., 2016; Tran et al., 2019). Expectedly, we find that unstressed cells grow faster in rich medium (doubling time 86 min) than in synthetic medium (doubling time 107 min) (Figures 1E,F), which underscores the previous finding that BY4741 strains requires an additional supplementation of the SCD medium for optimal growth (Hanscho et al., 2012). Furthermore, DTT- and TM-stressed cells grow markedly slower compared to the unstressed cells in both media (Figures 1E,F). Consistent with previous observations (Pincus et al., 2010), the reduced rate of growth becomes apparent as early as 1 h after the addition of the stress-inducing agents (Supplementary Figures S1A,B). Notably, the impact of DTT is more pronounced than the impact of TM at the given concentrations (Figures 1E,F). Next, we wanted to test if the reduced growth of the stressed cells is due to an activation of the UPR, which is known to peak within 1 h after the addition of DTT or TM to the medium and which largely remodels the cellular transcriptome (Kawahara et al., 1997; Travers et al., 2000; Promlek et al., 2011). Surprisingly, the growth of both stressed and unstressed ire1Δ cells was indistinguishable from WT cells in both rich and synthetic medium (Figures 1E–H and Supplementary Figures S1C–H). This suggests that DTT and TM impair cellular growth during this early phase of stress predominantly via their impact on protein folding and not by processes downstream of UPR activation. The slightly higher potency of DTT to impede cellular growth compared to TM at the given concentrations may reflect the fact that these compounds affect the folding of different sets of proteins: proteins with disulfide bonds in the case of DTT and N-linked glycosylated proteins in the case of TM. Furthermore, DTT can reduce already formed disulfide bonds and is known to directly affect also other cellular processes outside the ER such as the protein import into mitochondria (Mesecke et al., 2005) and protein palmitoylation (Levental et al., 2010). In contrast, TM affects only the glycosylation of freshly synthesized proteins. Our data suggest that the growth inhibition observed in acutely stressed cells is independent of UPR activation. We used shotgun mass spectrometry-based lipidomics to comprehensively and quantitatively dissect the impact of acute proteotoxic stress on the cellular lipid composition. In light of the pronounced impact of DTT and TM on cellular growth, we focused on their immediate effects within 1 h of treatment. We analyzed the lipid composition for six conditions and two different strains each as biological triplicates (Figure 2A). A principal component analysis (PCA) of the entire dataset at the level of individual lipid species revealed a close clustering of all samples from WT and ire1Δ cells cultivated in SCD, thereby suggesting UPR activity itself has little impact on their lipidomes (Figure 2B). In contrast, we observed two clusters for cells cultivated in rich medium. One cluster contained samples from WT and ire1Δ cells that were either left untreated or stressed with TM, while the other cluster contained samples from cells that were stressed with DTT at concentrations commonly used for UPR activation. This suggests that DTT causes a substantial remodeling of the lipidome, while TM treatments have a lesser impact on the cellular lipid composition. Not surprisingly, the loadings plot suggests a correlation of specific groups of lipids (Supplementary Figure S2A). The total amount of lipids quantified from 1 OD unit of cells (Supplementary Figure S2B) and the amount of storage lipids (Supplementary Figure S2C) highlight a low variability between replicates and show that storage lipids are more abundant in synthetic (SCD) medium. Storages lipids comprise all triacylglycerol (TAG) species and ergosterol esters. Experimental conditions affecting lipidome variability. (A) Overview of the cultivation and stress conditions use for lipidomic analysis. (B) A two-dimensional principal component analysis (PCA) of lipid species abundances reveals the degree of variations between different cultivation and stress conditions. Data from BY4741 WT and ire1Δ cells are indicated by circles and triangles, respectively. The color of the data points indicates unstressed cells in gray and cells stressed either with DTT (orange) or TM (green). Data from stressed and unstressed cells are indicated red and gray, respectively. PCA reveals that the lipidomes from both stressed and unstressed cells cultivated in synthetic medium (SCD) are very similar. In contrast, DTT induces a characteristic remodeling of the lipidome of both WT and ire1Δ cells cultivated in rich medium (YPD). For representing this complex dataset, we assorted the identified lipid species to one of four major lipid categories: sterols, sphingolipids (SLs), membrane glycerolipids (MGLs), and storage lipids. MGLs comprise all glycerophospholipids and diacylglycerol (DAG) species (Figure 3 and Supplementary Figure S3). Overall, we find a remarkable impact of the medium on the cellular lipid composition (Figure 3A). The lipid composition of S. cerevisiae WT indifferent media. A single colony of the BY4741 WT strain was used to inoculate a preculture in either synthetic (SCD) or rich (YPD) medium. After overnight cultivation for 21 h, a fresh culture was inoculated to an OD600 of 0.1. When the cells reached an OD600 of 0.80 ± 0.05, they were cultivated for one more hour. 20 OD equivalents of these cells were harvested and analyzed by lipid mass spectrometry. The data represented by black and white bars relate to cells cultivated in rich and synthetic medium, respectively [except for panels (F and G)]. (A) Lipid class composition in mol% of all quantified lipids in the sample organized by the categories: sterol (orange), sphingolipids (magenta), membrane glycerolipids (green, classes with two acyl chains highlighted by green box), and storage lipids (blue). Erg, ergosterol; Cer, ceramide; IPC, inositolphosphorylceramide; MIPC, mannosy-IPC; M(IP)2C, mannosyl-di-IPC; CL, cardiolipin; PA, phosphatidic acid; PC, phosphatidylcholine; PE, phosphatidylethanolamine; PI, phosphatidylinositol; PS, phosphatidylserine; DAG, diacylglycerol; LPC, lyso-PC; LPE, lyso-PE; LPI, lyso-PI; LPS, lyso-PS; TAG, triacylglycerol; EE, ergosteryl ester. (B) Profile of PA lipids in mol% of the class. (C) Total length of lipids in a sub-category of membrane glycerolipids (MGLs) including PA, PC, PE, PI, PS, DAG. The total length is given as the sum of carbon atoms in both fatty acyl chains in mol% of this sub-category. (D) Total number double bonds in a sub-category of MGLs (PA, PC, PE, PI, PS, DAG) is given as the sum of double bonds in both acyl chains, in mol% of this sub-category. (E) The acyl chain composition of PA, PC, PI and of a s sub-category of MGLs (PA, PC, PE, PI, PS, DAG) is normalized either to the individual lipid class or the sub-category and given in mol%. (F,G) The pairing of fatty acyl chains in MGLs (PA, PC, PE, PI, PS, DAG) is plotted for cells cultivated in (F) rich (YPD) and (G) synthetic (SCD) medium. The left panel indicates the pairing of fatty acyl chains in a sub-category of MGLs (PA, PC, PE, PI, PS, DAG) normalized to each particular fatty acyl chain and is given as mol%. The right panel indicates the abundance of acyl chain pairs in the sub-category of MGLs and is given in mol%. (H) Profile of sphingolipids in mol% of the category. The least abundant species in each panel are omitted for clarity. Each bar represents the average ± SD from n = 3 independent experiments. Statistical significance was tested by an unpaired two-tailed t-test using GraphPad Prism, *p < 0.05, **p < 0.01, ***p < 0.001. Yeast SLs comprise inositolphosphorylceramide (IPC), mannosyl-inositol phosphorylceramide (MIPC), mannosyl-di-(inositolphosphoryl) ceramide (M(IP)2C), and – less abundantly – ceramides (Cer). With the exception of Cer, all SLs have a significantly lower level in cells cultivated in rich medium compared to cells cultivated synthetic medium (Figure 3A). The same trend was observed for all SLs in ire1Δ cells (Supplementary Figure S3A). Thus, the lower level of sphingolipids in YPD-cultured cells is not due to possible differences in basal, constitutive UPR signaling. However, because SLs are concentrated along the secretory pathway and highly abundant in the plasma membrane (van Meer et al., 2008; Klemm et al., 2009; Hannich et al., 2011; Surma et al., 2011) it is tempting to speculate that cells cultivated in rich medium feature a higher abundance of inner membranes relative to the plasma membrane. In fact, the rapidly growing cells in rich medium have a particularly high demand for membrane biogenesis and both ER size and the rate of membrane lipid production can be controlled independently of UPR signaling (Loewen, 2004; Schuck et al., 2009; Henry et al., 2012). Membrane glycerolipids constitute the most abundant lipid category comprising the lipid classes phosphatidic acid (PA), phosphatidylcholine (PC), phosphatidyl-ethanolamine (PE), phosphatidylinositol (PI), phosphatidylserine (PS), and diacylglycerol (DAG) (Henry et al., 2012; Klose et al., 2012). Less abundant are cardiolipin (CL) and the lyso-lipid derivatives of the major glycerophospholipid classes with one fatty acid (FA) chain. We find that phosphatidylinositol (PI) is by far the most abundant lipid class in both WT and ire1Δ cells when cultivated in rich medium (Figure 3A and Supplementary Figure S3A). However, when cells are cultivated in synthetic medium, PI is much less abundant and found at levels comparable with phosphatidylcholine (PC) and phosphatidylethanolamine (PE) (Figure 3A and Supplementary Figure S3A). This is particularly relevant when studying the UPR: Ire1p has been identified as inositol-requiring enzyme (Nikawa and Yamashita, 1992), it is activated by inositol-depletion (Promlek et al., 2011), and sensitive to the stiffness of the ER membrane (Halbleib et al., 2017) thereby implying an important role of PI lipids. But not only PI, also the abundance of PE is affected by the choice the medium. PE is significantly less abundant in cells cultivated in rich versus synthetic medium (Figure 3A). Again, this is likely to affect the sensitivity and activity of the UPR as aberrant PE-to-PC ratios have been firmly implicated in chronic activation of the UPR in yeast, worms, and mammals (Fu et al., 2011; Thibault et al., 2012; Hou et al., 2014). Although qualitatively consistent with previous reports (Ejsing et al., 2009; Klose et al., 2012; Surma et al., 2013; Casanovas et al., 2015) our observations also highlight an important caveat for the use of rich medium. Because YPD is not a defined medium (in contrast to the synthetic-defined SCD), its use will inevitably lead to inconsistencies with respect to the lipid composition. As a consequence, it is almost impossible to compare data coming from different laboratories using media from different suppliers or even media batches. It is also impossible to exclude a technical bias as contributing factor for seemingly divergent observations: different procedures for lipid extraction into an organic phase may affect the spectrum of lipids that can be detected, different modes of sample separation and/or detection [such as thin-layer chromatography (TLC), liquid chromatography coupled to mass spectrometry (LC-MS) or the shotgun lipidomics platform used for this study] might bias the detection of some lipids/lipid classes over others. Ideally, the use of internal standards should correct the bias from extraction and detection. Furthermore, only fully quantitative data expressed in absolute units (such as pmol or derived molar fractions) can be compared to each other, as any relative data depend on the experimental setup and reference points applicable only within a given experiment. The choice of the medium has a striking impact on the abundancies of storage lipids. Exponentially growing cells cultivated in rich medium store much less TAGs and EEs than those cultivated in synthetic medium (Figure 3A). The same trend is observed in ire1Δ cells (Supplementary Figure S3A). It is tempting to speculate that rapidly growing cells depend more heavily on the production of membrane lipids than relatively slow growing cells, which can ‘afford’ to store some lipids for future use. Clearly, the different growth rates in the two media (Figure 1 and Supplementary Figure S1) must be considered in light of the intricate connections between membrane biogenesis, lipid droplet formation, and the UPR (Gaspar et al., 2011; Stordeur et al., 2014; Casanovas et al., 2015). Our data underscore the importance to study the UPR under tightly controlled conditions. We wanted to know the impact of rich versus synthetic medium on lipid acyl chains and initially focused our attention on the profile of PA lipids (Figure 3B). Normally, in S. cerevisiae, the fatty acid composition of PA lipids is mostly limited to palmitic (C16:0), palmitoleic (C16:1) and oleic acid (C18:1) with low amounts of shorter fatty acids (Klose et al., 2012). However, the cells cultivated in rich medium exhibited a significantly higher abundance of PA lipids with shorter, saturated acyl chains, as evidenced by the level of PA lipids with a cumulative acyl chain length of C26, C28, and C30 (Figure 3B). A combined analysis of all MGLs with two fatty acyl chains (PA, PC, PE, PI, PS, DAG) further highlighted this general and significant shift toward shorter (Figure 3C) and more saturated lipid species (Figure 3D). Notably, identical trends can be observed for ire1Δ cells indicating that a low, basal activation of the UPR does not contribute to this global trend in all MGLs caused by the cultivation in different media (Supplementary Figures S3B–D). We then studied the usage of different fatty acyl chains in PA, PC, and PI in cells cultivated in both rich and synthetic medium (Figure 3E) and also calculated the abundance of the different fatty acyl chains in all MGLs (with two fatty acyl chains) (Figure 3E). Consistent with previous reports, we find that the distribution of the fatty acyl chains differs between the individual glycerophospholipids classes (Figure 3E) (Ejsing et al., 2009; Klose et al., 2012; Casanovas et al., 2015). Strikingly, our data reveal that saturated, medium fatty acyl chains (C10, C12, and C14) are much more prominently found in cells cultivated in rich medium throughout all glycerophospholipid classes (Figure 3E). Notably, similar asymmetric lipids with two acyl chains differing substantially in length have only recently been found at high abundance in the lipidome of Schizosaccharomyces japonicus (Makarova et al., 2020). In order to gain more insight in the molecular rules that govern the production of asymmetric lipids, we studied the pairing of fatty acyl chains in MGLs. We found an almost exclusive pairing of saturated, medium fatty acyl chains (C10:0 or C12:0) with longer saturated fatty acyl chains (C16:0 or C18:0) in S. cerevisiae (Figure 3F; left panel). In fact, we find virtually no pairing of unsaturated fatty acyl chain (C16:1 or C18:1) with medium fatty acyl chains (C10:0 or C12:0). Also when considering the relative abundance of the different fatty acyl chains (C16:0, C16:1 and C18:1 are most abundant), the same, remarkable selectivity for certain pairs of fatty acyl chains becomes clear: medium fatty acyl chains preferentially pair with saturated, but not with unsaturated fatty acyl chains (Figure 3F; right panel). YPD is not a fully defined medium and it may contain minor concentrations of short and medium chain fatty acids. In order to test if S. cerevisiae can synthesize asymmetric lipids from scratch, we analyzed the acyl chain pairing in cells cultivated in synthetic medium (Figure 3G). Again, we found a strong preferential pairing of saturated, medium fatty acyl chains with saturated, long fatty acid chains, but not with unsaturated ones (Figure 3G). Our data suggest that tight and evolutionary conserved rules underly the pairing of fatty acyl chains in these highly asymmetric lipids. Another remarkably specific impact of the medium can be observed in the profile of SLs for both WT (Figure 3H) and ire1Δ cells (Supplementary Figure S3E). The most abundant sphingolipid species of cells cultivated in rich medium is IPC 44:0,3 (with three hydroxylations) contributing to more than 50 mol% of all SLs, while IPC 44:0,4 (with four hydroxylations) is much less abundant. In contrast, IPC 44:0;4 is contributing to more than 40 mol% to the pool of SLs and is the most abundant SL species in SCD-cultured cells. The molecular underpinnings of this remarkable shift in the species composition and its physiological relevance remains to be studied in greater detail. We wanted to study the impact of DTT and TM on the lipid composition of cells cultivated in both rich medium (Figure 4) and synthetic medium (Figure 5) and focus our attention on an early time point after 1 h of treatment. As important controls, we also determined the lipid compositions of stressed and unstressed ire1Δ cells in both media to test a possible contribution of UPR to changes of the lipidome (Supplementary Figures S4, S5). The impact of proteotoxic stress on the lipidome of S. cerevisiae WT in rich medium. A single colony of the BY4741 WT strain was used to inoculate a preculture in rich (YPD) medium. After overnight cultivation for 21 h, a fresh culture was inoculated to an OD600 of 0.1 and then cultivated to an OD600 of 0.80 ± 0.05. The cells were then either left untreated (white bars), stressed by the addition of either with 8 mM DTT (black bars) or 1.0 μg/ml TM (gray bars). After one additional hour of cultivation, 20 OD equivalents of these cells were harvested and analyzed by lipid mass spectrometry. (A) Lipid class composition in mol% of all quantified lipids in the sample organized by lipid categories. (B) The difference in lipid class abundance in stressed minus unstressed cells highlights the impact of DTT (black) and TM (gray) on the cellular lipid composition in rich medium. The difference in abundance in mol% was calculated by subtracting the abundance in unstressed cells from the abundance in either DTT- or TM-stressed cells under consideration of error propagation. (C) Profile of PA lipids in mol% of the class. (D) Total length of lipids in a sub-category of MGLs (PA, PC, PE, PI, PS, DAG). The total length is given as the sum of carbon atoms in both fatty acyl chains in mol% of this sub-category. (E) Total number double bonds in a sub-category of MGLs (PA, PC, PE, PI, PS, DAG) is given as the sum of double bonds in both acyl chains and represented in mol% of this sub-category. (F) Profile of sphingolipids in mol% of this category. The least abundant species are omitted for clarity. Each bar represents the average ± SD from n = 3 independent experiments. Statistical significance was tested by an unpaired two-tailed t-test using GraphPad Prism, *p < 0.05, **p < 0.01, ***p < 0.001. The impact of proteotoxic stress on the lipidome ofS. cerevisiae WT cultivated in synthetic medium. A single colony of the BY4741 WT strain was used to inoculate a preculture in synthetic medium (SCD). After overnight cultivation for 21 h, a fresh culture was inoculated to an OD600 of 0.1 and then cultivated to an OD600 of 0.80 ± 0.05. The cells were then either left untreated (white bars), stressed by the addition of either with 8 mM DTT (black bars) or 1.0 μg/ml TM (gray bars). After one additional hour of cultivation, 20 OD equivalents of these cells were harvested and analyzed by lipid mass spectrometry. (A) Lipid class composition in mol% of all quantified lipids in the sample organized by lipid categories. (B) The difference in lipid class abundance in stressed minus unstressed cells highlights the impact of DTT (black) and TM (gray) on the cellular lipid composition in synthetic medium. The difference in abundance in mol% was calculated by subtracting the abundance in unstressed cells from the abundance in either DTT- or TM-stressed cells under consideration of error propagation. (C) Profile of PA lipids in mol% of the class. (D) Total length of lipids in a sub-category of MGLs (PA, PC, PE, PI, PS, DAG). The total length is given as the sum of carbon atoms in both fatty acyl chains in mol% of this sub-category. (E) Total number double bonds in a sub-category of MGLs (PA, PC, PE, PI, PS, DAG) is given as the sum of double bonds in both acyl chains and represented in mol% of this sub-category. (F) Profile of sphingolipids in mol% of this category. The least abundant species are omitted for clarity. Each bar represents the average ± SD from n = 3 independent experiments. Statistical significance was tested by an unpaired two-tailed t-test using GraphPad Prism, *p < 0.05, **p < 0.01, ***p < 0.001. Consistent with our PCA analysis (Figure 2B), we find that the TM-induced stress has barely any impact on the lipid composition of WT cells (Figure 4A). This lack-of-impact is best evident when representing difference of abundance between stressed and unstressed cells for each lipid class (Figure 4B). There is also barely any impact of TM on the profile of PA lipids (Figure 4C), the length distribution of the fatty acyl chains (Figure 4D), the degree of saturation (Figure 4E), and the species composition of SLs (Figure 4F). Notably, this is true for both for WT and ire1Δ cells (Supplementary Figures S4A–F). These data suggest TM can be used as proteotoxic drug in YPD-cultivated cells without severely affecting the cellular lipid composition. In contrast to that, DTT at the given concentration has a significant impact on the lipidome of YPD-cultured cells (Figure 4A), which is most apparent when plotting the difference of lipid class abundance between stressed and unstressed cells (Figure 4B). Clearly, treating the cells for only 1 h with DTT is sufficient to cause a significant and substantial increase of PA in the stressed cells. Given that PA lipids are important signaling lipids involved in regulating membrane biogenesis (Loewen, 2004; Schuck et al., 2009; Henry et al., 2012; Hofbauer et al., 2018), a two-fold increase of the cellular PA level is likely to have broad impact on lipid metabolism and the cellular transcriptome. Intriguingly, treating the cells with DTT induces also a shift in the profile of PA lipids toward a higher average acyl chain length and more unsaturation for both WT (Figure 4C) and ire1Δ cells (Supplementary Figure S4C). While these observations suggest that DTT affects PA lipids either directly (by affecting fatty acid metabolism) or indirectly (by its impact on cellular growth or other means; Supplementary Figure S1A), these trends are not pronounced in the wider group of MGLs with two fatty acyl chains – neither for WT nor ire1Δ cells (Figures 4D,E and Supplementary Figures S4D,E). Based on its position in the lipid metabolic network (Henry et al., 2012; Ernst et al., 2016), we speculate that PA is a class of ‘early-responding’ lipids, which change most readily upon an environmental perturbation, while other MGLs are affected only after prolonged periods of stress. We also find some impact of ER stress on the profile of SLs of WT and ire1Δ cells (Figure 4F and Supplementary Figure S4F), but these changes are relatively small. The overall similarity of the lipidomic changes observed for WT and ire1Δ cells upon DTT treatments, suggest that most DTT impacts on lipid metabolism are independently of UPR-signaling, at least under the given conditions (Figure 4 and Supplementary Figure S4). In summary, we find that even a short cultivation in the presence of DTT induces substantial changes of the cellular lipidome. We wanted to know the impact of DTT and TM on the lipid composition of S. cerevisiae cultivated in synthetic medium. We therefore determined the lipidomes of both stressed and unstressed WT and ire1Δ cells (Figure 5 and Supplementary Figure S5). Overall, we find an almost identical lipid class composition for cells stressed either with DTT or TM compared to untreated control cells with only minor changes in the abundance of several lipid classes (Figures 5A,B and Supplementary Figures S5A,B). While the changes between stressed and unstressed cells are significant, the difference in abundance seems rather moderate (Figure 5B). For example, the ∼30% lower level of PA in both DTT- and TM-stressed cells (Figure 5A), has most probably a lower impact on cellular physiology than the two-fold increase of PA in DTT-treated cells in rich medium (Figure 4A). Other significant, moderate changes are observed for several SLs, Erg, and EEs (Figure 5A). We doubt that these lipidome changes are substantial enough to impact on the activity of the UPR via a membrane-based mechanism. This speculation is supported by the observation that proteotoxic stress clearly dominates over membrane-based stress when cells are treated with either DTT or TM for 1 h or less (Promlek et al., 2011). Beyond the overall very similar lipid class distribution of stressed and unstressed cells cultivated in synthetic medium, we find significant, but minor stress-induced changes in the PA species composition (Figure 5C), a similar total fatty acyl chain length in MGLs (Figure 5D), a mildly affected degree of lipid saturation (Figure 5E), and a largely similar profile of SLs (Figure 5F). Again, the lipidomes observed for WT and ire1Δ cells are extremely similar. In summary, we find, somewhat surprisingly, that the cellular lipidome is quite ‘resistant’ to perturbations from DTT- and TM-treatments when cells are treated for 1 h in synthetic medium. Our lipidomic survey of stressed and unstressed S. cerevisiae reveals that TM treatments for 1 h are suitable in both rich and synthetic medium to activate the UPR via proteotoxic stress with only little interference from a perturbed lipid composition. However, more precautions should be taken when using DTT. We have performed a systematic, quantitative analysis of the lipidomic changes associated with acute ER stress caused by the proteotoxic agents DTT and TM. Given the central relevance of the UPR for cellular homeostasis and the adverse effects associated with chronic UPR activation, it is crucial to better understand the signals that underlie prolonged and chronic activation of the UPR in the future. Using S. cerevisiae as a model, we highlight the importance of tightly controlled cultivation conditions and quantitative lipid analyses for a more holistic approach toward understanding the interplay of UPR-activating signals. We show that (1) the proteotoxic drugs DTT and TM impair cellular growth, thereby confirming previous observations (Pincus et al., 2010), (2) DTT causes within 1 h of treatment substantial changes of the lipidome of YPD-cultured cells, (3) DTT and TM have only a minor impact on the lipidome of SCD-cultured cells, (4) unstressed WT and ire1Δ cells feature virtually identical lipidomes under the tested conditions, thereby suggesting that basal, low-level UPR signaling (or lack of thereof) does not substantially affect the cellular lipid composition. Overall, our data are consistent with the general assumption that TM predominantly causes proteotoxic stress, at least within the first hour of treatment (Promlek et al., 2011). The rapid, DTT-dependent remodeling of the lipidome observed in YPD-cultured cells and the strongly impaired growth of stressed cultures, however, may serve as a warning to carefully interpret data derived from acutely stressed cells. We like to stress that we investigated only the impact of a single concentration for each drug in both media on the cellular lipid composition. It is possible that TM, when used at higher concentrations, might have a more severe impact on the cellular lipid composition. Nevertheless, for studying the UPR and its response to proteotoxic signals with little to no interference from a perturbed lipid metabolism, we suggest (1) the use of defined SCD over ill-defined YPD, (2) the use of TM over DTT, because of its more specific impact on protein folding in the ER, and (3) applying TM or DTT stress for a maximum of 1 h. Our systematic lipidomic analyses of WT and ire1Δ cells in two different media provides insights into the orchestration of membrane biogenesis by rapidly growing, eukaryotic cells. Even though rich medium provides a rich supply of nutrients, cells cultivated in this medium do not accumulate substantial amounts of storage lipids during the exponential growth phase (Figure 3A). Cells cultivated in synthetic medium, however, accumulate a five times higher level of neutral lipids (TAGs + EEs) although the medium is not as rich (Figure 3A). We speculate that these marked differences are at least partly due to the growth rate, which is substantially higher for YPD-cultured cells. This interpretation is consistent with the previous observation that storage lipids are dynamically regulated in a growth-stage dependent fashion (Klose et al., 2012; Kohlwein et al., 2013; Casanovas et al., 2015) and suggests an increased flux of fatty acids into membrane lipids in rapidly growing cells. In fact, the lipid phenotypes associated with a genetically disrupted fatty acid desaturation are more pronounced in YPD-cultured cells (Surma et al., 2013). The global role of the cellular growth rate on lipid metabolism remains to be established in the future. Along these lines, we also find that the ratio of SLs to MGLs is low in rapidly growing, YPD-cultured cells, but higher in slow-growing SCD-cultured cells (Figure 3A). While it is clear that many distinct regulatory circuits of the cell are affected by the composition of the medium, we speculate that the low SL-to-MGL ratio reflects an increased global rate of the lipid biosynthesis in the ER and -as a consequence- a low ratio of plasma membrane-to-ER abundance. Notably, our finding that WT and ire1Δ cells feature almost identical lipidomes (Supplementary Figure S3) would suggest that ER abundance is controlled independently of the UPR in this case. This is consistent with previous findings that ER proliferation can be uncoupled from the UPR via the Opi1 and Ino2/Ino4 regulatory circuit (Loewen, 2004; Schuck et al., 2009; Henry et al., 2012). Opi1 binding to PA lipids at the ER membrane (Loewen, 2004; Hofbauer et al., 2018) de-represses Ino2/Ino4-regulated genes involved in MGL biosynthesis and tip the balance from storing neutral lipids toward membrane proliferation. In fact, we find an increased level of PA lipids in YPD-cultured cells (Figure 3A and Supplementary Figure S3A). Our findings fuel that idea that the flux of fatty acids into either membrane or storage lipids is affected by the cellular growth rate, which shall investigated by dedicated experiments in the future. We were surprised by the high content of saturated lipids in YPD-grown cells (Figures 3D,E), which differs from previous studies (Klose et al., 2012; Surma et al., 2013; Casanovas et al., 2015) and which resembles lipid phenotypes observed only in genetically modified S. cerevisiae with a disrupted fatty acid desaturation (Pineau et al., 2009; Surma et al., 2013; Ballweg and Ernst, 2017; Budin et al., 2018). The majority of these saturated lipids contain a medium fatty acyl chain (C10:0 or C12:0) paired with a second, long fatty acyl chain (C16:0 or C18:0). Such asymmetric, saturated lipids at even higher abundancies have been recently identified in Schizosaccharomyces japonicus (Makarova et al., 2020). The length difference of the two acyl chains may allow for an interdigitation of the acyl chains, which increases lipid packing, whilst maintaining a fluid bilayer (Xu and Huang, 1987; Schram and Thompson, 1995; Makarova et al., 2020). This way, the asymmetric lipids can provide an alternative handle to balance competing demands in the homeostasis of physicochemical membrane properties, e.g., by maintaining membrane barrier function whilst increasing membrane fluidity (Lande et al., 1995; Schram and Thompson, 1995; Radanović et al., 2018). This may become very relevant under conditions when medium-chain fatty acids accumulate in the cell and/or when the desaturation of long-chain fatty acids via the fatty acid desaturase Ole1 becomes limiting (Stukey et al., 1989; Ballweg and Ernst, 2017). How precisely saturated, asymmetric lipids provide a feedback to the production of unsaturated fatty acid via the lipid saturation sensor Mga2 (Covino et al., 2016; Ballweg et al., 2020) is an interesting question for the future. Our finding that asymmetric, saturated lipids can be observed at substantial levels in S. cerevisiae suggests that such lipids may play a much wider role than previously anticipated. The machineries and mechanisms mediating a finely tuned production of asymmetric lipids in S. cerevisiae and other fungi are still unknown, but tracking the origin and fate of medium chain fatty acids might help identifying them. Our finding that saturated, asymmetric lipids are produced even upon cultivation in fatty acid-free SCD medium (Figure 3G) suggests that at least a significant portion of the esterified medium chain fatty acids originate from de novo biosynthesis. The fungal fatty acid synthase produces fatty acids of different lengths determined by the cellular ratio of acetyl-CoA to malonyl-CoA, which are used for priming and fatty acid elongation (Sumper et al., 1969; Okuyama et al., 1979; Heil et al., 2019). It is possible that a different product spectrum of the fatty acid synthase contributes to the production and abundance of asymmetric lipids in S. cerevisiae. Intriguingly, a hyper-active mutation in the rate-limiting enzyme for fatty acid biosynthesis leads to an increased production of malony-CoA and increased average fatty acyl chain length in glycerophospholipids (Hofbauer et al., 2014). In line with previous reports (Makarova et al., 2020), we therefore propose that the profile of de novo synthesized fatty acids is a major determinant for the abundance of saturated, asymmetric lipids. Our finding that saturated, medium fatty acyl chain pair almost exclusive with saturated, but not with the more abundant unsaturated fatty acyl chains (Figures 3E–G), reveals a remarkable, inherent selectivity in the biosynthesis and metabolism of asymmetric lipids. It will be intriguing to learn about the processes that contribute to this selectivity and to dissect the contribution of fatty acid biosynthesis and activation, acyl transferases, phospholipases, and substrate channeling (Henry et al., 2012; Ernst et al., 2016; Patton-Vogt and de Kroon, 2020). Because the key enzymes of lipid metabolism and the principle mechanisms of membrane adaptivity are conserved throughout evolution (Henry et al., 2012; Ernst et al., 2016), it will be intriguing to test if similarly specific mechanisms of acyl chain pairing are at work in organisms from bacteria to humans. The UPR has been implicated in a wide array of pathologies and is gaining increasing attention as a potential drug target (Hetz et al., 2019). While the fatal consequences of prolonged ER stress have been intensively studied (Tabas and Ron, 2011), the molecular events that cause chronic UPR activation remain poorly characterized. In the future, it will be crucial to develop new tools and assays to better understand the signals that perpetuate ER stress. Only quantitative information on the ER load with unfolded proteins and on the composition of the ER membrane during acute and prolonged ER stress can unambiguously establish the relative importance of signals from the ER lumen and the ER membrane. As a first step in this direction, we have studied the impact of different media and two UPR-activating, proteotoxic drugs on the cellular lipid composition. Our data will provide an important reference point for future endeavors, and has already proven useful for highlighting possible connections between the cellular growth rate, proteotoxic ER stress, and lipid metabolism. Our finding that S. cerevisiae produces asymmetric lipids with two saturated acyl chains of different lengths provides evidence for a remarkable specificity in paring of saturated fatty acyl chains. All datasets presented in this study are included in the article/Supplementary Material.
|
PMC5286405
|
Lipidomics reveals dramatic lipid compositional changes in the maturing postnatal lung
|
Lung immaturity is a major cause of morbidity and mortality in premature infants. Understanding the molecular mechanisms driving normal lung development could provide insights on how to ameliorate disrupted development. While transcriptomic and proteomic analyses of normal lung development have been previously reported, characterization of changes in the lipidome is lacking. Lipids play significant roles in the lung, such as dipalmitoylphosphatidylcholine in pulmonary surfactant; however, many of the roles of specific lipid species in normal lung development, as well as in disease states, are not well defined. In this study, we used liquid chromatography-mass spectrometry (LC-MS/MS) to investigate the murine lipidome during normal postnatal lung development. Lipidomics analysis of lungs from post-natal day 7, day 14 and 6–8 week mice (adult) identified 924 unique lipids across 21 lipid subclasses, with dramatic alterations in the lipidome across developmental stages. Our data confirmed previously recognized aspects of post-natal lung development and revealed several insights, including in sphingolipid-mediated apoptosis, inflammation and energy storage/usage. Complementary proteomics, metabolomics and chemical imaging corroborated these observations. This multi-omic view provides a unique resource and deeper insight into normal pulmonary development.The majority of perinatal deaths are of infants born prematurely (~10% of all births in the U.S. annually1), with the incidence of infant death increasing with decreasing gestational age at delivery2. Lung immaturity is a major cause of morbidity and mortality in premature infants with respiratory distress syndrome (RDS) most commonly the cause of death within the first 14 days of life (accounting for 49.5% of deaths) as well as within the first month of life (42.8%)3. Bronchopulmonary dysplasia, another pulmonary disorder, represents the leading cause of death after 60 days of life3. Strategies to promote maturation of under-developed lungs in premature infants remain an unsolved clinical challenge. A thorough understanding of the molecular mechanisms driving normal lung development is necessary to identify potential mechanisms for promoting alveologenesis and maturation of the under-developed lungs in premature infants. Towards this goal, comprehensive, untargeted omics studies of normal lung development have been conducted, including several transcriptomics analyses and a single proteomics analysis (see review in ref. 4). While there have been lipidomics studies on lung and lung surfactant in the context of disease and/or injury56, to our knowledge no untargeted omics investigation of lipid profiles during normal lung development has been reported. Lipids are a large and structurally diverse group of biomolecules and are increasingly appreciated to play crucial roles in maintaining energy balance, guiding intercellular communication, and regulating membrane dynamics. In the lung, lipids are highly significant molecules influencing many processes, and for example, are the primary component of pulmonary surfactant, a lipoprotein complex that functions to reduce the surface tension at the air/liquid interface in alveoli and prevent lung collapse7. Deficiency of pulmonary surfactant in the immature lung is the primary cause of RDS. Furthermore, variations in species and distributions of lipids in the lungs are implicated in the outcomes of cystic fibrosis8, lung cancer9, and asthma10. Thus, a thorough characterization of the lung lipidome (the totality of lipids in a biological system) is critical to understand lung function, development and disease states. In this study, we used liquid chromatography coupled with a tandem mass spectrometer (LC-MS/MS) to profile the temporal ontogeny of lipid changes during post-natal lung development employing a label-free relative quantification approach51112. Specifically we examined whole lung homogenates from mice at post-natal day 7 (PND7), 14 (PND14) and 6–8 weeks after birth (adult), spanning early alveolar to adult developmental stages1314. Our LC-MS/MS-based lipidomics analysis resulted in identification of 924 lipid species across 21 lipid subclasses, providing a deep and comprehensive view of the lung lipidome during normal development. Furthermore, we complemented these untargeted lipidomics measurements with additional untargeted proteomics (8289 total proteins, Table S2) and metabolomics (178 total metabolites, Table S3) analyses from the same samples to provide a more comprehensive picture of lipid metabolism during normal lung development. These types of multi-omics analyses have been shown to enhance our understanding of biological systems1215. Our results confirmed several previously recognized aspects of post-natal lung development and revealed several new insights into sphingolipid-mediated apoptosis, inflammation and energy storage/usage, providing detailed molecular signatures underlying these processes. Importantly, the lipid-based observations were corroborated by corresponding changes in relevant proteins and metabolites as judged from the complementary proteomics and metabolomics measurements of identical samples. Our multi-omic measurement approach offers unique insights into mammalian respiratory system development. We used a lipidomics platform consisting of a commercial UPLC system coupled with a Velos Orbitrap mass spectrometer operating in data-dependent MS/MS mode to profile the murine lung lipidome during post-natal development. Specifically we examined whole lung homogenates from mice at post-natal day 7 (PND7), 14 (PND14), and 6–8 weeks after birth (adult) (n = 3, each). The PND7 and PND14 time-points span peak stages of alveologenesis14, during which the alveoli that are critical for the air/gas exchange function of the lung form. Alveologenesis is largely completed by post-natal day 28, therefore, the 6–8 week time point used in our study represents the mature lung. Lipid tandem mass spectra were analyzed using LIQUID, an in-house software for lipid identification and quantification. Confident lipid identifications were made by examining the tandem mass spectra for diagnostic fragment ions along with associated acyl chain fragment information. In addition, the mass measurement error, isotopic profile, extracted ion chromatogram, and retention time of each lipid precursor ion was examined. Across all samples, 924 unique lipid species covering 3 lipid categories and 21 lipid subclasses (Fig. 1) were confidently identified based on their MS/MS fragmentation patterns, representing one of the largest lipidome datasets reported to date. The higher lipidome coverage obtained in this study compared to prior studies (see review in ref. 16) likely stems from the use of an UPLC system with a long gradient (90 min) that afforded efficient lipid separation and minimized under-sampling1718. The most commonly identified lipid species in the developing lung lipidome belonged to the triacylglycerol (TG) subclass with 253 identifications, followed by diacylglycerophosphocholine (PC) and the diacylglycerophosphoglycerol (PG) subclasses with 128 and 105 identifications, respectively. Further examination of the data revealed the most abundant species in PC and PG subclasses to be PC(14:0_16:0), PC(16:0_16:0), PG(16:0_18:1), and PG(16:0_16:0), in agreement with a previous report on the lipids known to be most abundant in pulmonary surfactant19. Relative quantification of the identified lipids51112 revealed that 161 species varied in a statistically significant manner (t-test; p < 0.05) between the PND7 and PND14 lungs, while 452 lipids and 438 lipids varied in a statistically significant manner (t-test; p < 0.05) between PND7 and adult samples and between PND14 and adult samples, respectively (Table S1). The high degree of dissimilarity between the PND7/PND14 samples versus the adult samples is highlighted by the dendogram in Figure S1, where PND7 and PND14 clustered most closely to each other relative to adult. This trend of a large degree of difference between murine lungs undergoing active alveolarization (PND7/PND14) and murine lungs that had completed alveolarization (adult) was also documented using nanospray desorption electrospray ionization mass spectrometry imaging (nano-DESI MSI) on independent (i.e. obtained from an outside laboratory) murine lung samples (PND7 vs. PND28) (Figure. S2). Focusing on lipids that showed at least a 1.3-fold change between time points from the nano-DESI analysis and also changed in a statistically significant manner between PND7 and adult time points in the LC-MS/MS lipidomics analysis, we defined a subset of 11 lipids for cross comparison between the nano-DESI lipid chemical imaging analysis and the LC-MS/MS lipidomics analysis. Each of the 11 lipids selected had a direction of change across time points as judged from nano-DESI in agreement with observations from the LC-MS/MS lipidomics analysis (Figure S2). Taken together, our data show broad changes in the murine lung lipidome spanning alveologenesis to adulthood and provide an unprecedented detailed compendium of temporal changes occurring in lipid species during post-natal lung development. Within the sphingolipid category, 75 unique lipids were identified covering five sphingolipid subclasses. Sphingomyelin (SM) species (26 identifications), underwent the greatest amount of change, with 67% significantly increased (t-test; p < 0.05) in the PND7 sample compared to the adult sample and 25% significantly decreased (t-test; p < 0.05) in the same comparison (Tables S1–9, Figure S3). Of the gangliosides, only GM3 species were detected (11 identifications); all except two decreased significantly (p < 0.05) from the PND7 and PND14 time-points to the adult time-point (Tables S1–10, Figure S4). Sphingosine (d18:1) and sphinganine (d18:0) remained relatively unchanged from PND7 to PND14 and then decreased in the adult time-point (Figure S5). Approximately 75% and 55% of identified ceramides (Cer) changed in a statistically significant manner (t-test; p < 0.05) from PND7 to adult and from PND14 to adult, respectively (37% changed between PND7 and PND14), with a relatively even distribution between lipids increasing or decreasing over this time-frame (Tables S1–11, Figure S6). A total of 332 lipids in triacylglycerol (TG) and diacylglycerol (DG) subclasses were identified. The largest change in glycerolipids occurred from PND14 to adult with 78% of the quantified glycerolipids changing in a statistically significant manner (t-test; p < 0.05) (Table S1). Both TGs and DGs with medium chain saturated fatty acids (MCSFA; 8:0, 10:0, 12:0, 14:0) (Fig. 2, Tables S1–7) and long chain polyunsaturated fatty acids (LCPUFA; greater than 20 carbons and with 4 or more double bonds in at least one fatty acid chain) were more abundant in lungs from PND7 and PND14 mice relative to adult mice (Fig. 3, Tables S1–8). Further, PND7 lungs contained more TGs with longer fatty acid chains, relative to adult lungs – the median total number of carbons in TG acyl chains for PND7 samples is 58 per intact lipid as compared to 52 for adult samples (Tables S1–12, Figure S7). Additionally the TG species identified in PND7 mice had on average almost 3 times as many double bonds as compared to adult mice – the median number of double bonds in TGs from PND7 samples is 8, whereas the median number of double bonds in adult samples is 3 (Tables S1–12, Figure S7). 517 unique glycerophospholipids were identified, spanning 14 subclasses and which exhibited a wide variety of trends supportive of diverse roles for glycerophospholipids during lung development. A major trend was MCSFA being observed in highest abundance in PND7 and PND14 samples relative to adult samples (Tables S1–7, Fig. 2). Additionally, the majority of glycerophosphoethanolamine plasmalogen (PE(P-)) species (17 of 28) increased in abundance over time, whereas only 1 species decreased in abundance in a statistically significant manner (Tables S1–13, Figure S8). Lastly, 46% monoacylglycerophosphocholines (LPCs) statistically increased (t-test; p < 0.05) from the PND7 to the adult time-point, whereas only 20% of LPCs decreased over the same time span (Tables S1–14, Figure S9). The LPC species that were most abundant in PND7 and PND14 samples tended to have either a MCSFA or a LCPUFA. Sphingolipids are a class of lipids that are intricately involved in numerous cell processes, including proliferation, senescence, differentiation, and signaling20. In the lung, sphingolipids have been used both as a marker for pulmonary development as well as for disease21. Ceramides, sphingosine, and sphingosine-1-phosphate (S1P) determine an apoptotic balance within cells, in which relative increases in ceramides and sphingosine result in cell death2223. Further, ceramides with different chain lengths have different effects on the cell; some species promote proliferation, some inhibit proliferation, and others can induce autophagy or apoptosis20. In this study, we observed an increase in pro-apoptotic and pro-autophagic ceramide species (Cer 14:0, Cer 16:0, and Cer 18:0) and pro-apoptotic sphingosine in lungs of PND7 and PND14 mice2425 (Fig. 4, Tables S1–3). Although S1P was not detected in our study, a corresponding proteomics analysis of the same samples (see Supplementary Methods) revealed the sphingosine-1-phosphate receptors (S1PR1, S1PR2 and S1PR3 detected), which mediate the proliferative/protective effects of S1P26, all exhibited lower abundance in lungs from PND7 and PND14 mice relative to those from adult mice (Fig. 4, Table S2). In the lung specifically, S1P and the S1PRs are known to have strong effects on maintenance of the endothelial barrier27, with the ability to rearrange the cytoskeleton to alter vascular permeability28. Corresponding metabolomics analysis of the same samples (see Supplementary Methods) showed the ceramide precursor L-serine increased in PND7 relative to adult lungs (Fig. 4, Table S3), with the corresponding proteomics analysis (described above) showing serine palmitoyltransferase, the rate-limiting enzyme in de-novo ceramide synthesis from serine, increased in PND7 relative to adult samples (Fig. 4, Table S2). Taken together, the above multi-omic observations support the notion of a shift in balance towards a more apoptotic state in the PND7 samples (Fig. 4). Additionally, almost all GM3 lipids were detected in higher abundance in PND7 and PND14 samples, relative to adult. In some tissues, these lipids are known to have pro-apoptotic effects; however, more research is necessary to elucidate the precise role of GM3 lipids in the lung2930. Recent studies have shown that BCL-2 genes are downregulated relative to BAX genes in association with sphingolipid-mediated cell death, which promotes apoptosis3132. The corresponding proteomics data in this study revealed that more pro-apoptotic BCL-2 and BAX family proteins were elevated in PND7 and PND14, mice while more proliferative proteins were elevated in the adult mice (Fig. 4, Table S2). The increase in pro-apoptotic molecules in samples from PND7 and PND14 mice is likely related to tissue remodeling and thinning associated with alveolarization and gas exchange optimization33. Previous studies have shown that during alveolarization as many as 20% of fibroblasts eventually undergo apoptosis in rat lungs32, and up to 12% of mesenchymal and epithelial cells in rats at postnatal day 1 are undergoing apoptosis34. Our multi-omics data support these earlier observations of increased apoptosis during alveolarization and for the first time provide a detailed view of the molecular signatures driving this process. Complex lipids containing LCPUFA, in particular glycerophospholipids within cell membranes, have been linked to inflammatory processes3536. These LCPUFA can impact the inflammatory process multiple ways including influencing membrane fluidity, protein function, and also acting as substrates for the generation of lipid signaling molecules (i.e. lipid mediators); the latter of which have potent effects on the regulation of inflammation in tissues and are themselves tightly regulated within a system37. These bioactive lipid mediators are generated by the actions of phospholipases on intact lipids containing 20:4, 20:5, and 22:6 fatty acids, which then release arachidonic acid (AA), eicosapentaenoic acid (EPA), or docosahexaenoic acid (DHA), respectively38. These fatty acids are then enzymatically modified by cytochrome P450 proteins, cyclooxygenases, and lipoxygenases, among others, to create pro-inflammatory, anti-inflammatory, and pro-resolution bioactive lipid mediators37. Although many bioactive lipid mediators can have both pro-inflammatory and pro-resolution effects (i.e. lipoxins), it is generally accepted that derivatives of arachidonic acid (AA, 20:4) tend to be pro-inflammatory, whereas derivatives of eicosapentaenoic acid (EPA, 20:5) and particularly docosahexaenoic acid (DHA, 22:6) tend to be pro-resolution39. We found increased glycerophospholipids with LCPUFA in adult mice as compared to PND7 and PND14 mice, with the PND7 and PND14 mice having most LCPUFAs localized in glycerolipid species (Fig. 3A, Tables S1–8). The corresponding proteomics data revealed lipases (diacylglycerol lipase alpha, phospholipase A2, adipose triglyceride lipase) that cleave LCPUFA from the intact lipids were increased in adult mice as compared to the PND7 and PND14 mice. Additionally, the proteins (cytochrome P450, cyclooxygenases, and lipoxygenases) that then generate the bioactive lipid mediators from the cleaved fatty acids were also found in higher abundance in the lungs from adult mice (Table S2). In the arachidonic acid (AA) metabolic pathway alone, 70% of proteins detected were significantly more abundant (p < 0.05) in adult (Fig. 3B, Figure S11). Taken together with the increased glycerophospholipids with LCPUFA in adult mice, this suggests that the source of arachidonic acids likely derives from glycerophospholipids in this study. Indeed, it is well known that PCs and PIs act as sources for arachidonic acids. For example, the protein PA24A selectively hydrolyzes arachidonyl phospholipids in the sn-2 position releasing arachidonic acid40. These data collectively highlight the ability of multi-omics analyses to provide insight on the complex interplay of molecules that regulate inflammation in the lungs during development. In addition to LCPUFA, lysoPCs (LPC) were more abundant in adult animals (Figure S9, Tables S1–14). Several studies have shown that LPC species are involved in pro-inflammatory processes in situations of acute injury as well as chronic stress41. LPCs are produced by the actions of phospholipase A2 (PLA2) and aid in the recruitment of T-lymphocytes42, promote production of pro-inflammatory cytokines by immune cells4344, increase reactive oxygen species production45, and upregulate cell adhesion molecules46. In the lungs, LPCs increase the permeability of the alveolar epithelium47 and inactivate surfactant function48. In our study, the majority of LPC species were more abundant in lungs from adult mice (Figure S9, Tables S1–14), correlating with the increase in phospholipase A2 (PLA2) in the adult lungs. Plasmalogen species are lipids that modulate membrane structure49. In addition, due to the instability of the vinyl ether bond connecting the glycerol backbone to the fatty acid in plasmalogen species, these lipids are able to act as lipophilic antioxidants in cells5051. Plasmalogens can have a protective effect against iron-induced lipid peroxidation, and it has been suggested that they are some of the most significant antioxidants in the lung and pulmonary surfactant5051. In our study, glycerophosphoethanolamine plasmalogen (PE(P-)) species were found in highest abundance in lungs from adult mice (Figure S8, Tables S1–13). Together with the above observations related to LCPUFAs and LPCs, the increased (PE(P-)) levels in adult animals supports the notion that the adult lungs are in a greater state of inflammation, with a carefully coordinated interaction between pro-inflammatory and pro-resolution lipids and associated protein mediators to maintain a finely balanced homeostasis to prevent overt injury to the lung. For each lipid subclass in which MCSFA were detected (Cer, DG, LPC, LPE, PC, PE, PE(P-, PG, PI, PS, SM, TG), the highest abundances were observed in lungs from PND7 and PND14 mice, relative to those from adults. Specifically, of the 92 lipid species that were detected with at least one MCSFA and no co-eluting species lacking MCSFA, 88% were more abundant in young relative to adult mice (Fig. 2, Tables S1–7). The higher abundance of MCSFA in younger mice may suggest a role for MCSFA as a rapid energy source in the earlier alveolar stage. Free MCFAs do not need to be transported by carnitine palmitoyl transferase to the mitochondria in order to undergo beta-oxidation; because of this they are preferentially oxidized over long chain fatty acids for energy52. Oxidation of free MCFAs may also contribute to a ketotic state, because an increase in beta-oxidation of fatty acids results in an increase in ketone bodies5354. In our study, an increase in beta-hydroxybutyrate, a ketone body, was observed in metabolomics data from PND7/PND14 lung tissue as compared to adult (Table S3 and Figure S10), suggesting a high rate of fatty acid oxidation in the younger animals (acetone and acetoacetate, the other two ketone bodies, were not detected; Table S3)). Recent work has shown that mild ketosis is frequently observed in suckling mammalian infants and that ketone bodies are vital to organ development, particularly the brain55. Ketone bodies are preferentially utilized over glucose for de novo surfactant production in lungs of newborn rats56. These observations suggest that MCSFAs may be important for newborn lung development, primarily by providing a source of rapidly available energy or intermediates for new and evolving postnatal functions. In this study, we used an untargeted lipidomics approach to provide a comprehensive picture of the murine lipidome throughout alveolarization, resulting in the confident identification (i.e. based primarily on MS/MS fragmentation patterns) of over 900 lipids in 21 lipid subclasses and representing one of the largest lipidome datasets reported to date. The data showed large global shifts in the lipidome over time and provided insight into the molecular level events driving cellular processes and function. Specifically, differential lipid abundances were observed in molecular species that are known to have significant roles in regulating apoptosis, inflammation, and energy storage. These observations were supported by data from complementary proteomics and metabolomics analysis on the same samples, as well as lipid chemical imaging analysis. Our work highlights the utility of leveraging deep lipidome measurements and complementary multi-omic data to provide unique insights into pulmonary development and potentially discover therapeutic targets to promote lung maturation. All animal care procedures were approved by PNNL Institutional Animal Care and Use Committee, and carried out in accordance with PNNL Institutional Animal Care and Use Committee guidelines and regulations. A total of 6, 6–7 week old mice and 6 time-mated C57BL/6 J mice were ordered from The Jackson Laboratory and kept in standard shoebox caging (72 °F ± 3°, humidity 50% rH) with ALPHA-dri bedding (Shepherd Specialty Papers) at Pacific Northwest National Laboratory. The light-dark cycle was 12 hours each and the mice were fed ad libitum food (Lab Diet 5002 Certified Rodent Diet) and water. Mice were euthanized by cervical dislocation (>PND14) or decapitation (PND7) and exsanguinated via an incision in the left ventricle. The heart and lungs were washed in situ with an injection of 10 ml of 4 °C PBS (with Ca2+ and Mg2+) through the right ventricle. The lungs were then removed en bloc and washed twice in cold PBS. Any contaminating tissue was carefully removed at this point before flash freezing in liquid nitrogen and storage at −70 °C until further sample prep. Frozen tissue samples were transferred into a tarred, pre-chilled Eppendorf Safe-Lock tube and the total mass was determined. The tissue samples were then homogenized using a Qiagen TissueLyser II with a 2 × 24 adapter (chilled to −20 °C) following the vendor suggested protocol for tissue samples, with modifications. In short, a 3 mm tungsten bead was added to each tube along with 0.5 mL of chilled methanol prior to processing twice in TissueLyser II for 3 minutes at 30 Hz. After homogenization, the samples were again chilled to −20 °C prior to transferring to a chilled Sorenson MμlTI™ SafeSeal™ microcentrifuge. The aqueous polar metabolites, lipids, and proteins were extracted from each homogenate using a modified Folch extraction57 as previously described58. Keeping each sample on ice, a volume of chilled chloroform and water were added to a final ratio of 3:8:4 water-chloroform-methanol, mixing gently after each addition. The samples were chilled on ice for 5 minutes before mixing well and separating the layers by centrifugation (10 k × g, 10 minutes, 4 °C). The aqueous polar metabolites, lipids and proteins were isolated and concentrated to dryness in a vacuum concentrator and stored at −70 °C until ready for further processing. The lower organic layer of the Folch extraction was reconstituted in MeOH and subjected to LC-ESI-MS/MS analyses, using a Waters NanoAquity UPLC system (Waters column, HSS T3 1.0 mm × 150 mm × 1.8 μm particle size) interfaced with a Velos-ETD Orbitrap mass spectrometer (Thermo Scientific, San Jose, CA). Data dependent scan events occurred both in the ion trap (collision-induced dissociation, CID) and Orbitrap (HCD) using normalized collision energy (NCE) of 30 and 35 arbitrary units, respectively, in the same run as outlined previously59. Scan events were between the mass ranges of m/z 200–2000 for both positive and negative ionization. A full scan event occurring in the Orbitrap with a resolving power of 60,000. The full scan event is followed by MS/MS of the top 6 ions, alternating between HCD (resolving power of 7500) and CID, respectively, with an isolation width of 2 and dynamic exclusion of 60 sec. The ion time for the full scan MSn was 500 ms, for MS/MS in the Orbitrap (HCD) 1000 ms, and MS/MS in the ion trap (CID) 50 ms. The automatic gain control for the full scan was 1,000,000, for HCD MS/MS 50,000, and for CID MS/MS 10,000. Unique lipid species were separated chromatographically over a 90 min gradient elution (mobile phase A: ACN/H2O (40:60) containing 10 mM ammonium acetate; mobile phase B: ACN/IPA (10:90) containing 10 mM ammonium acetate) at a flow rate of 30 μl/min. The LC gradient details can be found in the supplementary Tables S1–6. Higher-energy collision dissociation (HCD) and collision-induced dissociation (CID) were used to analyze samples in both positive and negative ionization modes in order to gain high coverage of lipid species. LC-MS/MS raw data files were imported into the in-house developed software LIQUID (Lipid Informed Quantitation and Identification) for semi-automated identification of lipid molecular species. Lipid identifications were confirmed by examination of the isotopic profiles, precursor masses, XICs, and tandem mass spectra. In the tandem mass spectra, the diagnostic ion in addition to fatty acid fragment ions were used to confirm the identification. Each ionization mode of the datasets was then separately aligned and gap-filled based on m/z and retention time using MZmine 260, with manual verification of every feature. Peak apex intensities were exported for statistical analysis. All lipid nomenclature follows the “Comprehensive Classification System for Lipids” developed by the International Lipid Classification and Nomenclature Committee6162. Extracted proteins were denatured, alkylated, digested with trypsin and desalted on a C18 SPE cartridge (Discovery C18, 1 mL, 50 mg, Sulpelco). The peptide concentration was measured by BCA assay (Thermo Scientific). Peptides were labeled with 10-plex TMT reagents (Life technology) according to the manufacturer’s instructions prior to be pooled together. They were then separated using an off-line reversed-phase chromatography column as previously described63. 24 fractions were collected. 5 μL of 0.1 μg/μL of peptides from each fraction were analyzed by reverse phase LC-MS/MS using a Waters nanoEquityTM UPLC system (Millford) coupled with a Q-Exactive mass spectrometer (Thermo Scientific). The LC was configured to load the sample first on a solid phase extraction (SPE) column followed by separation on an analytical column. Analytical columns were made in-house by slurry packing 3-μm Jupiter C18 stationary phase (Phenomenex, Torrence, CA) into a 70-cm long, 360 μm OD × 75 μm ID fused silica capillary tubing (Polymicro Technologies Inc., Phoenix, AZ). Samples were loaded on the SPE column via a 5 μL sample loop for 30 minutes at a flow rate of 3 μL per minute and then separated by the analytical column using a 110 minute gradient from 99% mobile phase A (MP-A) to 5% MP-A at a flow rate of 0.3 μL per minute. Mass spectrometry analysis was started 15 minutes after the sample was moved to the analytical column. After the gradient was completed, column was washed with 100% mobilie phase B (MP- B) first and then reconditioned with 99% MP- A for 30 minutes. The effluents from the LC column were ionized by electrospray ionization and mass analyzed with a QExactive hybrid quadrupole/Orbitrap mass spectrometer operated in the data-dependent analysis mode. Top 10 ions from the survey scan were selected by a quadrupole mass filter for high-energy collision dissociation (HCD) in collision with nitrogen and mass analyzed by the Orbitrap. An isolation window of 2 Daltons was used for the isolation of ions and a collision energy of 28% was used for HCD with AGC setting of 1e5 ions. Mass spectra were recorded for 100 minutes by repeating this process with a dynamic exclusion of previously selected ions for 60 seconds. Raw mass spectrometry data were converted to peak lists (DTA files) using the DeconMSn (version 2.3.1.2) and searched with MS-GF+64 against Uniprot/SwissProt mus musculus database (downloaded 2013-09-18), bovine trypsin and human keratin sequences. The identified spectra were filtered based on their MSGF+ scores and only the proteins with two proteospecific peptides were conserved resulting in a protein and peptide false discovery rate <1%. For the quantitative analysis, the TMT reporter ion intensities were extracted with MASIC65. TMT reporter intensities were summed from the different peptides belonging to the same proteins. Proteins with missing data were excluded for the quantification analysis. Analysis of the dried polar metabolites was performed as previously described66. Prior to analysis, the carbonyl groups were protected by treating with 20 uL of methoxyamine solution (30 mg/mL in pyridine) and incubated for 90 minutes at 37 °C with shaking. The hydroxyl and amine groups were then derivatized with 80 uL of N-methyl-N-(trimethylsilyl) trifluoroacetamide (MSTFA) with 1% trimethylchlorosilane (TMCS), incubating for 30 minutes at 37 °C with shaking. The samples were then allowed to cool to room temperature and were analyzed by an Agilent GC 7890 A coupled with MSD 5975 C mass spectrometer (Agilent Technologies, Santa Clara, CA). Separations were performed on a HP-5MS column (30 m × 0.25 mm × 0.25 μm; Agilent Technologies). The injection mode was splitless, and the injection port temperature was held at 250 °C. The column oven was initially maintained at 60 °C for 1 min and then ramped to 325 °C by 10 °C/min, followed by a 10 min hold at 325 °C. GC-MS raw data files from each Experiment were processed using the Metabolite Detector software, version 2.5 beta. Briefly, Agilent. D files were converted to netCDF format using Agilent Chemstation, followed by conversion to binary files using Metabolite Detector (PMID: 19358599). Retention indices (RI) of detected metabolites were calculated based on the analysis of the FAMEs mixture, followed by their chromatographic alignment across all analyses after deconvolution. Metabolites were initially identified by matching experimental spectra to an PNNL augmented version of FiehnLib67, containing spectra and validated retention indices for over 850 metabolites, using a Metabolite Detector match probability threshold of 0.6 (combined retention index and spectral probability). All metabolite identifications were manually validated to reduce deconvolution errors during automated data-processing and to eliminate false identifications. For lipidomics, metabolomics and proteomics, the data was log transformed and median normalized within each sample and statistically significant changes were determined using two-tailed, homoscedastic t-test and/or ANOVA using R stats package. Pearson’s correlation and hierarchical clustering were also performed using this stats package. The principal component analysis were realized using the ‘mixOmics’ package68. All mice used for imaging data were housed in the Cincinnati Children’s Hospital Medical Center Animal Care Facility according to National Institutes of Health and institutional guidelines for the use of laboratory animals. All protocols of the present study were reviewed and approved by Cincinnati Children’s Hospital Research Foundation Institutional Animal Care and Use Committee. C57BL/6 mice from JAX Mice (Jackson Laboratory, Bar Harbor, ME, USA) were sacrificed at day 7 (PND7 - early alveolar) and day 28 (PND28 - late alveolar) by CO2 overdose. Lungs were collected, cleaned, embedded in carboxymethyl cellulose (CMC) and stored in −80 °C freezer. Samples were sectioned into 10 μm thick slices with a Thermo CryoStar NX70 (Thermo Scientific, Waltham, MA) microtome to generate coronal sections of the lungs. Sections were thaw-mounted onto regular glass slides and stored at −80 °C until analysis. The section was allowed to thaw at room temperature right before mounting onto the nano-DESI sample holder for analysis. A custom made nano-DESI source69 comprised of a sample holder attached to a high-resolution motorized XYZ sample stage (Zaber Technologies, Vancouver, BC) controlled via a custom-designed Labview software69 was mounted onto an XL LTQ/Orbitrap mass spectrometer (Thermo Scientific, Waltham, MA). The OD 150 μm × ID 50 μm fused silica capillaries were used to make both nano-DESI primary and secondary capillaries. A solvent consisting of 90% methanol (Fisher Scientific) and 10% water (HPLC grade, Fisher Scientific) was used with the addition of 3 μM of LPC 19:0 and PC 23:0 as internal standards. Solvent was delivered at a flow rate of 500 nl/min with an applied voltage of 3.5 kV through the primary capillary. The heated capillary inlet was held at 30 V and 250 °C. The primary and secondary capillaries were positioned using micromanipulators (XYZ 500MIM, Quater Research and Development, Bend, OR) under monitoring of two Dino-Lite digital microscopes (AnMo Electronics Corporation, Sanchong, New Taipei, Taiwan). For each time point, three biological and three technical replicates were acquired in both positive and negative mode. Imaging experiments were performed by scanning the sample line by line under the nano-DESI probe at a constant velocity while acquiring mass spectra. The imaging scan rate was 70 μm/s and spacing between lines was set at 100 μm. The nano-DESI stage and mass spectrometer was synchronized by triggering at the beginning of each line scan. The mass spectrometer was operated with a mass resolution of 60 000 (m/Δm) at m/z 400. Mass spectral data acquired by the Xcalibur software were subsequently processed by MSI QuickView70, a visualization software developed at PNNL. Accession codes: Data deposited and freely available at ProteomeXchange data repository, ProteomeXchangeID: PXD004651 and MassIVE data repository, MassIVE ID: MSV000080000. How to cite this article: Dautel, S. E. et al. Lipidomics reveals dramatic lipid compositional changes in the maturing postnatal lung. Sci. Rep. 7, 40555; doi: 10.1038/srep40555 (2017). Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
|
PMC8748654
|
Lipidomic profiling of human serum enables detection of pancreatic cancer
|
Pancreatic cancer has the worst prognosis among all cancers. Cancer screening of body fluids may improve the survival time prognosis of patients, who are often diagnosed too late at an incurable stage. Several studies report the dysregulation of lipid metabolism in tumor cells, suggesting that changes in the blood lipidome may accompany tumor growth. Here we show that the comprehensive mass spectrometric determination of a wide range of serum lipids reveals statistically significant differences between pancreatic cancer patients and healthy controls, as visualized by multivariate data analysis. Three phases of biomarker discovery research (discovery, qualification, and verification) are applied for 830 samples in total, which shows the dysregulation of some very long chain sphingomyelins, ceramides, and (lyso)phosphatidylcholines. The sensitivity and specificity to diagnose pancreatic cancer are over 90%, which outperforms CA 19-9, especially at an early stage, and is comparable to established diagnostic imaging methods. Furthermore, selected lipid species indicate a potential as prognostic biomarkers.Non-invasive cancer screening methods based on blood analysis have been intensively investigated in medical research over the last decades, with special focus on the detection of early cancer stages. Some cancer types, such as pancreatic cancer, do not show specific symptoms, which makes the diagnosis of early stages difficult with established screening methods. Pancreatic ductal adenocarcinoma (PDAC), accounting for 90% of pancreatic cancers, is mostly diagnosed at the late stage resulting in the worst 5-year survival rate (7%) among all cancers. Imaging modalities used to diagnose PDAC in clinical practice include magnetic resonance imaging, computed tomography, endoscopic ultrasound, and positron emission tomography, with accuracy reported in the meta-analysis of 5,399 patients from 52 studies of 90, 89, 89, and 84%, respectively. Invasive procedures, i.e., biopsies, are performed only for the final confirmation of PDAC. Several types of blood tests were considered for PDAC screening, such as carbohydrate antigen (CA) 19-9 measured alone or with other blood proteins, e.g., carcinoembryonic antigen. The sensitivity and specificity values of CA 19-9 drop for early cancer stages, which prevents the applicability for early screening. However, the sensitivity increases for late stage, and therefore CA 19-9 is used for monitoring of cancer treatment. The analysis of circulating tumor DNA, extracellular vesicles, and circulating tumor cells shows a potential for the diagnosis of PDAC and is under investigation. Kirsten-ras (KRAS) mutation testing is currently used in clinical practice for the epithelial cancer screening (e.g., lung or colorectal cancers) and was evaluated as well for the diagnosis of PDAC using liquid biopsies. However, the sensitivity for KRAS mutation testing is low, even though this mutation is encountered in more than 90% of PDAC. KRAS may be involved in the metabolic reprogramming of fast proliferating tumor cell populations towards elevated glucose and glutamine flows, defined as one of the hallmarks of cancer. Furthermore, the uptake of nutrients in KRAS mutated cells can include blood lipids for cell proliferation and survival. KRAS mutation has been reported to be associated with lipid metabolism in pancreatic cancer cells. Lipids serve numerous functions in human metabolism, such as cell membrane constituents, signaling molecules, energy supply, and storage. Changes in lipid concentrations were already reported in other cancer types, mostly for cell lines, tissues, and less frequently for body fluids, too. Here we show differences in serum lipidome concentrations between samples obtained from PDAC patients and healthy controls using mass spectrometry (MS) based approaches followed by statistical analysis. Preliminary results showed that the monitoring of single lipid species did not perform well for the differentiation between cases and controls, unlike the multi-analyte approach. Furthermore, lipid species and classes are interrelated, thus it was assumed that the analysis of the lipidome may provide not only molecular biological insights of PDAC but also a more reliable experimental design for clinical diagnostics. The overall methodology is summarized in Fig. 1. Lipid species were quantified by using exogenous lipid class internal standards (IS) added to the serum before the sample preparation (Supplementary Tables 1–4). This allows intra- and inter laboratory comparison because the results are expressed independently from the instrumental signal response. Prepared extracts were analyzed using MS-based approaches, and lipidomic MS data were processed with an in-house script allowing automated lipid identification and quantitation. Finally, the data were statistically evaluated using descriptive and explorative approaches. All lipid species analyzed with the various MS approaches and within different study phases fulfilled the defined inclusion criteria, i.e., concentrations have to be reported for more than 25% of the samples, otherwise, the lipid species is excluded from the statistical evaluation. This exclusion criterium results in different lipid coverages for individual methods and phases due to natural differences in the sensitivity.Fig. 1Overview of study design for the differentiation of PDAC patients (T, red) from normal healthy controls (N, blue) and pancreatitis patients (Pan, green) based on the lipidomic profiling of human serum using various mass spectrometry-based approaches.a Phase I (discovery) for 364 samples (262 T + 102 N) divided into training (213 T + 79 N) and validation (49 T + 23 N) sets measured by UHPSFC/MS, shotgun MS (LR), and MALDI-MS. b Phase II (qualification) for 554 samples (444 T + 98 N + 12 Pan) divided into training (328 T + 82 N + 12 Pan) and validation (116 T + 16 N) sets measured by UHPSFC/MS, shotgun MS (LR and HR), and RP-UHPLC/MS at 3 different laboratories. c Phase III (verification) for 830 samples (546 T + 262 N + 22 Pan) divided into training (430 T + 246 N + 22 Pan) and validation (116 T + 16 N) sets measured by UHPSFC/MS for samples obtained from four collection sites. a Phase I (discovery) for 364 samples (262 T + 102 N) divided into training (213 T + 79 N) and validation (49 T + 23 N) sets measured by UHPSFC/MS, shotgun MS (LR), and MALDI-MS. b Phase II (qualification) for 554 samples (444 T + 98 N + 12 Pan) divided into training (328 T + 82 N + 12 Pan) and validation (116 T + 16 N) sets measured by UHPSFC/MS, shotgun MS (LR and HR), and RP-UHPLC/MS at 3 different laboratories. c Phase III (verification) for 830 samples (546 T + 262 N + 22 Pan) divided into training (430 T + 246 N + 22 Pan) and validation (116 T + 16 N) sets measured by UHPSFC/MS for samples obtained from four collection sites. The study was divided into individual phases (Fig. 1) called discovery, qualification, and verification phases, whereby each phase had their own purpose. The sample sets in individual phases were classified into the training and validation sets before applying multivariate data analysis (MDA) to ensure unbiased statistical evaluation. The training set was used to build statistical models, and the validation set for the independent evaluation of the model performance to differentiate samples of cancer patients from healthy controls. The influence of the blood collection tube on the lipidomic analysis was evaluated as a part of the preliminary testing, method optimization, and validation using ultrahigh-performance supercritical fluid chromatography (UHPSFC)/MS. Results showed slightly higher lipid concentrations in serum in comparison to plasma, which yields an enhanced sensitivity. Therefore, serum was used as the sample matrix of choice for the presented PDAC screening study. The discovery phase was a proof-of-concept study with the goal to find differences between serum lipidomic profiles of cases and controls. In total, samples of 262 PDAC patients and 102 healthy controls were analyzed by UHPSFC/MS and shotgun MS, and a limited subset of 64 samples also by matrix-assisted laser desorption/ionization (MALDI)-MS. All methods differ in the detection coverage of lipids, whereby shotgun MS has the highest number of 270 detected lipid species (Supplementary Data 1), followed by UHPSFC/MS with 168 lipid species (Supplementary Data 2), where both methods are based on the positive ion mode. Lipid species belonged to glycerolipids, phospholipids, sphingolipids, and cholesteryl esters for both methods. 42 lipid species from sphingomyelins and sulfatide classes were detected by MALDI-MS in the negative ion mode (Supplementary Data 3). Differences between case and control samples based on the lipidomic profile were visualized by MDA. A partial discrimination between cases and controls was already observed for principal component analysis (PCA) score plots, and the distinct group differentiation was achieved by supervised orthogonal projections to latent structures discriminant analysis (OPLS-DA) models. The influence of gender on the differentiation of case and control samples was investigated (Fig. 2) by performing OPLS-DA for both genders and gender-separated. The accuracy to assign the sample type correctly was slightly better for gender-separated models, therefore gender-separated models were further used in this work, which is in accordance with previously published results.Fig. 2Effect of gender separation on the quality of OPLS-DA models used for the differentiation of human serum samples obtained from PDAC patients (T) and healthy controls (N) for the training set using UHPSFC/MS in the Phase I.a Both genders. b Males. c Females. d Specificity, sensitivity, and accuracy for individual models. Source data are provided as a Source Data file. a Both genders. b Males. c Females. d Specificity, sensitivity, and accuracy for individual models. Source data are provided as a Source Data file. The lipidomic profiling approach for cancer and control samples seems to be independent of the cancer stage because a random distribution of cancer stages is observed in OPLS-DA plots without any clustering (Fig. 3a–d). This finding suggests that the lipidomic profiles differ even for early stage cancer from control samples, which is further verified in the subsequent study phases. The ROC curves and accuracies for training and validation sets were comparable for all methods in the discovery phase (Fig. 3e, f). The most dysregulated lipids are shown in Fig. 3g–j. UHPSFC/MS was used for subsequent studies due to the highest robustness and throughput among the compared methods and also supported by extensive experiences in our group including the full method validation and the stability test for samples collected during one year. The whole sample preparation protocol was optimized including the development of quality control (QC) system.Fig. 3Results for the Phase I obtained in lab 1. Individual samples are colored according to tumor (T) stage: T1 - yellow, T2 - orange, T3 - red, T4 - rose, and Tx - brown (information about the stage is not available).OPLS-DA for males measured a with UHPSFC/MS and c with shotgun MS for the training set (104 T + 30 N). OPLS-DA for females measured with b UHPSFC/MS and d shotgun MS for the training set (157 T + 49 N). ROC curves for males (M) and females (F) in training (Tr.) and validation (Va.) sets: e UHPSFC/MS, and f shotgun MS. Box plots for molar concentration in human serum from PDAC patients (T) and healthy controls (N) for males (M) and females (F): g SM 41:1 measured by UHPSFC/MS, h SM 41:1 measured by shotgun MS (LR), for both box plots for males (104 T and 30 N) and females (109 T and 49 N), i SHexCer 41:1(OH) measured by MALDI-MS, and j SHexCer 40:1(OH) measured by MALDI-MS, for both box plots for males (15 T and 14 N) and females (18 T and 19 N). In each box plot, the centerline represents the median, the bounds represent the 1st and 3rd quartile and whiskers span 1.5 fold inter-quartile range from the median. Source data are provided as a Source Data file. OPLS-DA for males measured a with UHPSFC/MS and c with shotgun MS for the training set (104 T + 30 N). OPLS-DA for females measured with b UHPSFC/MS and d shotgun MS for the training set (157 T + 49 N). ROC curves for males (M) and females (F) in training (Tr.) and validation (Va.) sets: e UHPSFC/MS, and f shotgun MS. Box plots for molar concentration in human serum from PDAC patients (T) and healthy controls (N) for males (M) and females (F): g SM 41:1 measured by UHPSFC/MS, h SM 41:1 measured by shotgun MS (LR), for both box plots for males (104 T and 30 N) and females (109 T and 49 N), i SHexCer 41:1(OH) measured by MALDI-MS, and j SHexCer 40:1(OH) measured by MALDI-MS, for both box plots for males (15 T and 14 N) and females (18 T and 19 N). In each box plot, the centerline represents the median, the bounds represent the 1st and 3rd quartile and whiskers span 1.5 fold inter-quartile range from the median. Source data are provided as a Source Data file. The goal of Phase II was to confirm that a similar differentiation of case and control groups can be achieved by other experienced lipidomic laboratories, which should exclude a possible hidden bias for measurements in the single laboratory. Cooperating labs 2 and 3 (details in “Methods” section) had no prescriptions concerning the analytical method used for lipidomic quantitation, so they followed their own established protocols for sample preparation, MS-based measurements, and data processing. The new sample set for Phase II consisted of 554 samples, whereby 344 samples of newly obtained aliquots were from the same volunteers included in Phase I, and 210 samples were from new subjects. The extended cohort was measured by four different MS-based methods (UHPSFC/MS, shotgun MS with low resolution (LR) and high resolution (HR), and reversed-phase ultrahigh-performance liquid chromatography (RP-UHPLC)/MS) (Supplementary Fig. 1). RP-UHPLC/MS allowed the quantitation of 431 lipids (Supplementary Data 4, 5), whereby the lipid species separation is applied due to the hydrophobic interactions of fatty acyls with the nonpolar stationary phase. Shotgun MS is based on the direct sample infusion into MS using specific scan events in case of LR or combined with tandem mass spectrometry (MS/MS) in case of HR. 232 lipids were quantified with shotgun LR-MS (Supplementary Data 6) and 183 lipids with shotgun HR-MS (Supplementary Data 7). For UHPSFC/MS, the lipid class separation was applied, which results in the quantitation of 202 lipid species (Supplementary Data 8). NIST 1950 reference plasma was measured with all methods as well and used for the normalization of lipid concentrations obtained by individual methods separately for males (Supplementary Fig. 2) and females (Supplementary Fig. 3). The box plots of some of the most dysregulated lipid species (Fig. 4a–c, Supplementary Fig. 2i, j, 3i, and 4a–l) reveal the same pattern and similar normalized concentrations for all methods. The RSD of concentrations of selected lipid species for each sample obtained by four methods (Fig. 4d–f) illustrate the acceptable reproducibility of different quantitation approaches. RSD < 40% for the majority of samples was observed, regardless of the use of different approaches for the sample preparation, IS mixtures, randomization, and lipidomic analysis. The future harmonization of analytical protocols planned within the International Lipidomics Society should further improve the correlation among different laboratories. MDA for individual method data sets from Phase II, such as the ROC curves (Fig. 4g–j), OPLS-DA score plots, and the evaluation of sensitivity, specificity, and accuracy prepared separately for males (Supplementary Fig. 2a–h) and females (Supplementary Fig. 3a–h) were performed. Statistical results show similar outcomes regarding the discrimination of case and control groups for all methods.Fig. 4Comparison of Phase II results obtained at three different laboratories using four mass spectrometry-based approaches.Box plots of lipid concentrations normalized to the NIST reference material for samples obtained from PDAC patients (443 T) and healthy controls (95 N) of both genders including both validation and training sets: a SM 41:1, b LPC 18:2, and c Cer 41:1 for UHPSFC/MS (Method 1), shotgun MS (LR) (Method 2), shotgun MS (HR) (Method 3), and RP-UHPLC/MS (Method 4). In each box plot, the centerline represents the median, the bounds represent the 1st and 3rd quartile and whiskers span 1.5 fold inter-quartile range from the median. The RSD of the concentrations obtained by four methods for each sample d SM 41:1, e LPC 18:2, and f Cer 41:1. Color annotation: light blue—control females, blue—control males, red—cancer females, and dark red—cancer males. ROC curves for males (M) and females (F) in training (Tr.) and validation (Va.) sets: g UHPSFC/MS, h shotgun MS (LR), i shotgun MS (HR), and j RP-UHPLC/MS. Source data are provided as a Source Data file. Box plots of lipid concentrations normalized to the NIST reference material for samples obtained from PDAC patients (443 T) and healthy controls (95 N) of both genders including both validation and training sets: a SM 41:1, b LPC 18:2, and c Cer 41:1 for UHPSFC/MS (Method 1), shotgun MS (LR) (Method 2), shotgun MS (HR) (Method 3), and RP-UHPLC/MS (Method 4). In each box plot, the centerline represents the median, the bounds represent the 1st and 3rd quartile and whiskers span 1.5 fold inter-quartile range from the median. The RSD of the concentrations obtained by four methods for each sample d SM 41:1, e LPC 18:2, and f Cer 41:1. Color annotation: light blue—control females, blue—control males, red—cancer females, and dark red—cancer males. ROC curves for males (M) and females (F) in training (Tr.) and validation (Va.) sets: g UHPSFC/MS, h shotgun MS (LR), i shotgun MS (HR), and j RP-UHPLC/MS. Source data are provided as a Source Data file. In the timeframe between Phase II and Phase III, UHPSFC/MS method and the sample preparation protocol for the lipidomic analysis were optimized and validated. Furthermore, the influence of preanalytical and analytical aspects, such as the blood collection tubes, lipidomic profile stability in the period of one year for the same volunteers, and the influence of the mass spectrometer was systematically investigated. All investigations led to an improved understanding of capabilities of lipidomic profiling for clinical sample screening and illustrated the high reproducibility of UHPSFC/MS for the lipidomic analysis. Phase III aimed at the verification of the applicability of lipidomic profiling for the differentiation of control and cancer samples using the optimized and validated UHPSFC/MS method for the lipidomic analysis of 830 samples (Supplementary Data 9). The sample set consisted of various sample groups obtained from four different blood collection sites, whereby 554 of 830 samples from clinic 1 correspond to samples from Phase II. The effects of various factors were investigated in addition to PDAC vs. control differentiation, such as pancreatitis, diabetes mellitus, age, cancer stage, and treatment. The training set included 341 male samples (122 controls and 219 cases, Fig. 5) and 335 female samples (124 controls and 211 cases, Fig. 6). The minor group differentiation was observed in PCA score plots (Figs. 5a and 6a), but OPLS-DA (Figs. 5b and 6b) showed a clear group clustering of PDAC and controls. The influence of the cancer stage was visualized by color codes of samples. No clustering depending on the cancer stage was visible, which indicated that the lipidomic profiling may have a potential for early PDAC detection. The sensitivity, specificity, and accuracy values were overall >94% for the training set and >80% for the validation set (Figs. 5c, 6c, and Supplementary Data 10). The lipid species with the highest concentration differences between case and control samples were visualized by S-plots (Figs. 5d and 6d) and heat maps (Figs. 5e and 6e), whereby lipid concentrations downregulated in case samples are marked in blue, and upregulated lipid species are in red color. Furthermore, statistical tests were performed and lipid species with fold change ≥20%, p-value <0.05 according to the Welch test, and variable importance in the projection (VIP) values >1 were defined as statistical relevant and summarized in Supplementary Data 11–13. Lipid species with p-value < the Bonferroni correction are additionally highlighted and considered as especially statistically significant for the lipidomic differentiation, such as selected sphingolipids, glycerophospholipids, and glycerolipids. However, glycerolipid concentrations may be affected by dietary intake, and therefore may be prone to misinterpretation. Consequently, considering statistical parameters (fold change, p-value, and VIP) and excluding exogenous interference, the lipid species SM 41:1, SM 42:1, Cer 41:1, Cer 42:1, SM 39:1, LPC 18:2, and PC O-36:3 were of the highest relevance for the differentiation, which is in accordance with results from Phase I and Phase II.Fig. 5Results for the lipidomic profiling of male serum samples from PDAC patients (T) and healthy controls (N) in Phase III.a PCA for the training set (219 T + 122 N). b OPLS-DA for the training set (219 T + 122 N). Individual samples are colored according to tumor (T) stage: T1 - yellow, T2 - orange, T3 - red, T4 - rose, and Tx - brown (information about the stage is not available). c Sensitivity (red), specificity (blue), and accuracy (green) for the training (219 T + 122 N) and validation (56 T + 6 N) sets. d S-plot for the training set with the annotation of most upregulated (red) and downregulated (blue) lipid species. e Heat map for both training and validation sets (275 T + 128 N) using the lipid species concentrations [nmol/mL]. f OPLS-DA for early stages T1 + T2, age aligned (mean age is 65 ± 4 years for N and 67 ± 4 for T), and number aligned (39 T + 39 N). This graph includes both genders. Source data are provided as a Source Data file.Fig. 6Results for the lipidomic profiling of female serum samples from PDAC patients (T) and healthy controls (N) in Phase III.a PCA for the training set (211 T + 124 N). b OPLS-DA for the training set (211 T + 124 N). Individual samples are colored according to their tumor (T) stage: T1 - yellow, T2 - orange, T3 - red, T4 - rose, and Tx - brown (information about the stage is not available). c Sensitivity (red), specificity (blue), and accuracy (green) for training and validation sets. d S-plot for the training set with the annotation of most upregulated (red) and downregulated (blue) lipid species. e Heat map for both training and validation sets (271 T + 134 N) using the lipid species concentrations [nmol/mL]. Source data are provided as a Source Data file. a PCA for the training set (219 T + 122 N). b OPLS-DA for the training set (219 T + 122 N). Individual samples are colored according to tumor (T) stage: T1 - yellow, T2 - orange, T3 - red, T4 - rose, and Tx - brown (information about the stage is not available). c Sensitivity (red), specificity (blue), and accuracy (green) for the training (219 T + 122 N) and validation (56 T + 6 N) sets. d S-plot for the training set with the annotation of most upregulated (red) and downregulated (blue) lipid species. e Heat map for both training and validation sets (275 T + 128 N) using the lipid species concentrations [nmol/mL]. f OPLS-DA for early stages T1 + T2, age aligned (mean age is 65 ± 4 years for N and 67 ± 4 for T), and number aligned (39 T + 39 N). This graph includes both genders. Source data are provided as a Source Data file. a PCA for the training set (211 T + 124 N). b OPLS-DA for the training set (211 T + 124 N). Individual samples are colored according to their tumor (T) stage: T1 - yellow, T2 - orange, T3 - red, T4 - rose, and Tx - brown (information about the stage is not available). c Sensitivity (red), specificity (blue), and accuracy (green) for training and validation sets. d S-plot for the training set with the annotation of most upregulated (red) and downregulated (blue) lipid species. e Heat map for both training and validation sets (271 T + 134 N) using the lipid species concentrations [nmol/mL]. Source data are provided as a Source Data file. The effects of the cancer stage and age on the differentiation of case and control samples were further investigated by age-matched controls and early stage PDAC samples classified as T1 and T2 (Fig. 5f). The OPLS-DA model was created for 39 control samples with the average age of 65 ± 4 years and 39 case samples with the average age of 67 ± 4 considering both genders, because age-matched and gender-separated models would result in the insufficient number of samples. The sensitivity, specificity, and accuracy were 97.4% for the differentiation of early cancer stages from control samples, which supports the previous claim on the suitability for early stage PDAC detection and excludes possible bias due to the fact that cancer patients are typically older than healthy controls in many reported studies including this work. Furthermore, the box plots of control, early stage (T1 and T2), late stage (T3 and T4), and pancreatitis (Pan) samples were prepared for statistically most significant lipid species SM 41:1 and Cer 41:1 (Fig. 7b, c). Concentrations measured in cancer samples are downregulated in comparison to control samples independent of cancer stages, and concentrations in pancreatitis samples are similar to control samples. These results suggest that lipidomic profiling may be applicable for differentiation of pancreatitis from PDAC samples, but the confirmation with a higher sample number of samples within the frame of a prospective study is certainly required. ROC curves for males and females for training and validation sets provided AUC values over 0.90 (Fig. 7a). The effect of age on the most dysregulated lipid species was also visualized for all 830 samples in Phase III (Fig. 7d, e). Lipid concentrations were overall similar for individual age groups, with the exception of slightly elevated lipid concentrations for SM 41:1 and LPC 18:2 for cancer patients younger than 39 years old (Fig. 7d, e), but this observation could be influenced by the smaller number of subjects in this age group. The diabetes mellitus is connected with a dysfunction of the pancreas, the effect of diabetes on the lipid profiles was investigated by the comparison of lipid concentrations of subjects with and without diabetes mellitus for case and control groups for SM 41:1 (Fig. 7f), where no visible effect was observed.Fig. 7Results for the lipidomic profiling in Phase III and investigating the influence of cancer stage, age, and diabetes mellitus.a ROC curves for males (M) and females (F) in training (Tr.) and validation (Va.) sets. Box plots of lipid molar concentrations normalized to the NIST reference material for: b SM 41:1 and c Cer 41:1. Only samples with known tumor (T) stage classification were included, where early stage (T1 and T2, 24 males and 30 females) and late stages (T3 and T4, 174 males and 176 females) are summarized and compared to samples of healthy controls (128 males and 134 females) and pancreatitis patients (13 males and 9 females). Comparison of age intervals for control (blue) and cancer (red) samples (d) SM 41:1 and e LPC 18:2. Box plot investigating the influence of diabetes (f) SM 41:1. In each box plot, the centerline represents the median, the bounds represent the 1st and 3rd quartile and whiskers span 1.5 fold inter-quartile range from the median. Comparison of the ROC curves for the samples investigated in Phase III for both genders (g) red - CA 19-9, blue - lipidomics, green - combination of CA 19-9 and lipidomics, and purple - CancerSeek. h Sensitivity (red) and specificity (blue) for CA 19-9, lipidomics, combination of CA 19-9 and lipidomics, and CancerSeek. Source data are provided as a Source Data file. a ROC curves for males (M) and females (F) in training (Tr.) and validation (Va.) sets. Box plots of lipid molar concentrations normalized to the NIST reference material for: b SM 41:1 and c Cer 41:1. Only samples with known tumor (T) stage classification were included, where early stage (T1 and T2, 24 males and 30 females) and late stages (T3 and T4, 174 males and 176 females) are summarized and compared to samples of healthy controls (128 males and 134 females) and pancreatitis patients (13 males and 9 females). Comparison of age intervals for control (blue) and cancer (red) samples (d) SM 41:1 and e LPC 18:2. Box plot investigating the influence of diabetes (f) SM 41:1. In each box plot, the centerline represents the median, the bounds represent the 1st and 3rd quartile and whiskers span 1.5 fold inter-quartile range from the median. Comparison of the ROC curves for the samples investigated in Phase III for both genders (g) red - CA 19-9, blue - lipidomics, green - combination of CA 19-9 and lipidomics, and purple - CancerSeek. h Sensitivity (red) and specificity (blue) for CA 19-9, lipidomics, combination of CA 19-9 and lipidomics, and CancerSeek. Source data are provided as a Source Data file. The overall performance of lipidomic profiling for PDAC screening was compared to the clinical established CA 19-9 biomarker for monitoring the progress of PDAC. Cut-off values for CA 19-9 do not differ for genders, therefore the MDA for lipidomic profiling was performed for both genders as well. Moreover, the comparison was also done for the combination of lipidomic profiling and CA 19-9 to predict sample groups as well as for the recently published CancerSeek method combining the analysis of proteins and ctDNA for cancer screening including PDAC. The ROC curves showed the best performance for the combination of lipidomic profiling and CA 19-9, followed by the lipidomic profiling, the CancerSeek method, and finally the determination of only CA 19-9 (Fig. 7g). The comparison of sensitivity and specificity values for various methods (Fig. 7h) showed that CA 19-9 and CancerSeek yielded higher specificity than sensitivity values. The opposite was observed for lipidomic profiling yielding higher sensitivity than specificity values. The combination of lipidomic profiling and CA 19-9 resulted in increased specificity. The influence of cancer treatment on the lipidomic profiling was investigated for a small subgroup within the sample set with blood collection before and several days after surgery. MDA plot does not show any return to control group (Fig. 8a), which indicates that PDAC might be a systemic disease with a strong influence on the metabolism, and the tumor removal does result in immediate recovery of lipidomic profile. The box plots for SM 41:1 and LPC 18:2 (Fig. 8b, c) showed that the concentrations mainly decrease after surgery in contrary to control samples (Figs. 5d and 6d). Furthermore, some patients received medical treatment (e.g., chemotherapy), and samples were collected before and after treatment. No statistically significant effects due to the medical treatment on the lipid profiles were observed (Fig. 8d, e). Furthermore, OPLS-DA models (Fig. 8f, g) were prepared for patients before any treatment, and groups of age-matched healthy controls to exclude any possible biases caused by treatment. The accuracy over 90% and the same patterns of dysregulated lipids show that the actual treatment did not cause relevant changes in lipid profiles.Fig. 8Results for the lipidomic profiling of human serum samples for PDAC patients (T) and healthy controls (N) including both genders in Phase III.Influence of surgery on the lipidomic profile: a OPLS-DA for females (211 T + 124 N) using the training set with highlighted samples before (green, n = 13) and after (orange, n = 10) surgery. Box plots of molar lipid concentrations for paired samples collected before and after surgery for both genders (2 males and 10 females): b SM 41:1, and c LPC 18:2. Box plots for paired samples collected before (n = 22) and after treatment (n = 22 for collection 1, n = 12 for collection 2, n = 7 for collection 3, n = 4 for collection 4) for both genders using molar concentrations: d SM 41:1, e LPC 18:2. In each box plot, the centerline represents the median, the bounds represent the 1st and 3rd quartile and whiskers span 1.5 fold inter-quartile range from the median. OPLS-DA models only for subjects before any treatment or surgery separately for f males (83 T + 122 N) and g females (72 T + 124 N). Source data are provided as a Source Data file. Influence of surgery on the lipidomic profile: a OPLS-DA for females (211 T + 124 N) using the training set with highlighted samples before (green, n = 13) and after (orange, n = 10) surgery. Box plots of molar lipid concentrations for paired samples collected before and after surgery for both genders (2 males and 10 females): b SM 41:1, and c LPC 18:2. Box plots for paired samples collected before (n = 22) and after treatment (n = 22 for collection 1, n = 12 for collection 2, n = 7 for collection 3, n = 4 for collection 4) for both genders using molar concentrations: d SM 41:1, e LPC 18:2. In each box plot, the centerline represents the median, the bounds represent the 1st and 3rd quartile and whiskers span 1.5 fold inter-quartile range from the median. OPLS-DA models only for subjects before any treatment or surgery separately for f males (83 T + 122 N) and g females (72 T + 124 N). Source data are provided as a Source Data file. The potential of lipids for prognostic purposes was investigated using the data from Phase II for different methods. Lipid concentrations for all samples and the lifetime data were processed with Kaplan–Meier survival analysis for individual methods (Supplementary Data 14 and Supplementary Table 5). Several lipid species showed a statistically significant correlation (p < 0.05) with the overall survival (Fig. 9a–c,), such as LPC 18:2, Cer 36:1, PC 32:0, and PC O-38:5. Whereby LPC 18:2 was positively correlated with the survival, in agreement with the previous work. In contrast, Cer 36:1, Cer 38:1, Cer 42:2, PC 32:0, PC O-38:5, and SM 42:2 were negatively correlated with the survival. The gender and treatment did not show any statistically significant effect on the survival probability (Supplementary Fig. 5), but concentrations of CA 19-9 had a strong negative correlation with the survival function (Supplementary Fig. 5). Cox proportional-hazards model was another regression tool used for the visualization of associations among survival time and predictor variables (Fig. 9d), which demonstrated that the concentration of LPC 18:2 higher than median was positively correlated with survival, while the opposite trend was observed for CA 19-9 and PC O-38:5.Fig. 9Potential of selected lipids for the survival prognosis in Phase II measured by UHPSFC/MS.Kaplan–Meier Survival plots for: a LPC 18:2 (n = 128 for binary code 0, and n = 72 for binary code 1), b Cer 36:1 (n = 89 for 0, and n = 111 for 1), and c PC 32:0 (n = 99 for 0, and n = 101 for 1) together with the two-sided log-rank test p-value. d Cox proportional-hazards model for CA 19-9, PC 32:0, PC O-38:5, LPC 18:2, Cer 36:1, Cer 38:1, Cer 42:2, and SM 42:2. The forest plot illustrates the 95% confidence intervals of the hazard ratios and the corresponding log rank test p values for every parameter are presented. Hazard ratios > 1 indicate poorer survival. Lipid species concentrations normalized to the NIST reference material obtained for all samples in Phase II were converted into the binary code, whereby 0 was set for c < median and 1 for c > median (the median of concentrations was calculated for each lipid species including all samples). Source data are provided as a Source Data file. Kaplan–Meier Survival plots for: a LPC 18:2 (n = 128 for binary code 0, and n = 72 for binary code 1), b Cer 36:1 (n = 89 for 0, and n = 111 for 1), and c PC 32:0 (n = 99 for 0, and n = 101 for 1) together with the two-sided log-rank test p-value. d Cox proportional-hazards model for CA 19-9, PC 32:0, PC O-38:5, LPC 18:2, Cer 36:1, Cer 38:1, Cer 42:2, and SM 42:2. The forest plot illustrates the 95% confidence intervals of the hazard ratios and the corresponding log rank test p values for every parameter are presented. Hazard ratios > 1 indicate poorer survival. Lipid species concentrations normalized to the NIST reference material obtained for all samples in Phase II were converted into the binary code, whereby 0 was set for c < median and 1 for c > median (the median of concentrations was calculated for each lipid species including all samples). Source data are provided as a Source Data file. Accurate cancer screening using peripheral blood as a minimally invasive and standardized method is desired in medical healthcare, assuming that early detection of cancer may improve patient outcome. Recently, liquid biopsy based on the analysis of genetic mutations, ctDNA, and proteins in serum or plasma for cancer diagnosis has been intensively investigated, and the results are promising for cancer screening. Furthermore, metabolomics belongs to one of the hot research topics with high expectations in clinical diagnostics. However, the reproducible and comprehensive analysis of small molecules is challenging and often not achieved in highly complex human body fluid samples, which is reflected by the overall poor acceptance of metabolomics in comparison to genomics and proteomics despite the enormous research interest. This work is based on a rather complex three-phase concept with the goal to exclude any possible hidden biases and to confirm that the observed changes in serum lipid concentration are really connected to PDAC, and not influenced by other interfering factors. Our measurements and evaluation consisted of different laboratories (groups in Pardubice, Regensburg, and Singapore), different MS-based workflows (UHPSFC/MS, shotgun LR-MS, shotgun HR-MS, RP-UHPLC/MS, and MALDI-MS), different collection sites (clinics in Brno, Prague, Olomouc, and Pilsen), and sample preparation, IS used for the quantitation, and data processing were done independently by individual laboratories. Regardless of this considerable heterogeneity, we can conclude that reported lipid dysregulation are really statistical relevant for PDAC patients in comparison to healthy controls, and should be reproducible by any laboratory experienced in the quantitative lipidomic analysis. Furthermore, all obtained data sets allow the preparation of MDA statistical models applicable for the differentiation of PDAC patients from controls with relatively high accuracy including early stage PDAC patients. The final confirmation of the applicability of lipidomic profiling to differentiate case from control samples was performed with UHPSFC/MS. The influence of cancer stage, age, diabetes, treatment, and pancreatitis was investigated for 830 samples. 67% of these samples were also included in Phase II. We believe that MS-based lipidomic profiling indicates the potential for early detection of PDAC, but the follow-up confirmatory study and the verification of the clinical utility of such screening are essential before possible implementation into screening programs in individual countries. UHPSFC/MS was selected as the method of choice for further investigations of PDAC screening, but the simple shotgun LR-MS setup may be also considered for future screening because this configuration is well established in newborn screening. The comparison of diagnostic performance results revealed that lipidomic profiling can compete with the clinically established method for the monitoring of PDAC progress, CA 19-9, and with CancerSeek (Supplementary Table 6), one of the most promising cancer screening tests published in recent years based on the analysis of ctDNA and proteins. However, CA 19-9 and CancerSeek perform better regarding specificity, important for general population screening, whereby lipidomic profiling performs better regarding sensitivity independent of the cancer stage, which may be of interest for the screening of high-risk individuals. The combination of lipidomic profiling and CA 19-9 improves the diagnostic performance, especially regarding the specificity in comparison to only lipidomic profiling, may be of interest for the establishment of the universal blood screening test. From the biological point of view, the altered lipid metabolism may originate from tumor cells, tumor stroma, and apoptotic cells. An immune response of the organism may also be involved. All these processes can naturally contribute to the observed cancer lipidomic phenotype. In measurements from all involved laboratories and phases, we observed a clear downregulation of multiple lipid species in the serum of PDAC patients (Fig. 10), such as decreased levels of most very long chain monounsaturated sphingomyelins and ceramides. These changes could be linked to the KRAS-driven metabolic switch. In this context, alterations in sphingolipids concentrations deserve attention, as the normal metabolism of sphingomyelins might be necessary to maintain KRAS function. Targeted biological investigations are needed to explain the mechanism of lipid alterations in the serum of PDAC patients, but it will require the development of suitable animal models in the future.Fig. 10Network visualization of the most dysregulated lipid species in PDAC for data from Phase III.Graphs show lipidomic pathways with clustering into individual lipid classes for a males, and b females using the Cytoscape software (http://www.cytoscape.org). Circles represent the detected lipid species, where the circle size expresses the significance according to p-value, while the color darkness defines the degree of up/downregulation (red/blue) according to the fold change. The most discriminating lipids are annotated. Source data are provided as a Source Data file. Graphs show lipidomic pathways with clustering into individual lipid classes for a males, and b females using the Cytoscape software (http://www.cytoscape.org). Circles represent the detected lipid species, where the circle size expresses the significance according to p-value, while the color darkness defines the degree of up/downregulation (red/blue) according to the fold change. The most discriminating lipids are annotated. Source data are provided as a Source Data file. In summary, we developed a reproducible, robust, and high-throughput lipidomic profiling approach for the detection of PDAC in human serum, which is applicable for the screening of at least 2000 samples per month on one MS system. In lab 1 (University of Pardubice, Czech Republic), solvents for sample preparation and analysis, such as acetonitrile, 2-propanol, methanol (HPLC/MS grade), hexane, and chloroform stabilized with 0.5–1% ethanol (both HPLC grade), were purchased from either Sigma-Aldrich (St. Louis, MO, USA) or Merck (Darmstadt, Germany), respectively. Mobile phase additives (ammonium acetate, ammonium formate, and acetic acid) were purchased from Sigma-Aldrich. Deionized water was obtained from a Milli-Q Reference Water Purification System (Molsheim, France). Carbon dioxide of 4.5 grade (99.995%) was purchased from Messer Group (Bad Soden, Germany). Non-endogenous lipids used as IS for the quantitative lipidomic analysis were purchased either from Avanti Polar Lipids (Alabaster, AL, USA), Nu-Chek (Elysian, MN, USA), or Merck. Lipid concentrations used for the IS mixture are provided in Supplementary Tables 1 and 2 depending on the employed method, further details for the preparation and dilution of the IS mixture used for UHPSFC/MS measurements were previously published. The NIST SRM 1950 metabolite reference plasma was used as QC sample and for normalization of concentrations between different MS-based methods. Furthermore, a pooled serum sample of PDAC patients and healthy controls were used as QC samples. The lipid annotation used in this manuscript is according to the recommendations of the Lipidomics Standard Initiative (LSI) and given in Supplementary Data 15. The chemicals and standards mentioned above were used for the sample preparation and measurements performed in lab 1. In lab 2 (University Hospital of Regensburg, Germany), chloroform and 2-propanol were purchased from Roth (Karlsruhe, Germany) and methanol from Merck (Darmstadt, Germany). All solvents were HPLC grade. Ammonium formate and cholesteryl ester (CE) standards were purchased from Sigma-Aldrich (Taufkirchen, Germany). Triacylglycerol (TG) and diacylglycerol (DG) standards were purchased from Larodan (Solna, Sweden) and dissolved in 2,2,4-trimethylpenthane/2-propanol (3:1, v/v). Phosphatidylcholine (PC), ceramide (Cer), sphingomyelin (SM), lysophosphatidylcholine (LPC), and lysophosphatidylethanolamine (LPE) standards were purchased from Avanti Polar Lipids (Alabaster, Alabama, USA), and dissolved in chloroform. In lab 3 (National University of Singapore), chemicals and reagents were obtained from the following sources: ammonium formate, acetic acid, and butanol from Sigma-Aldrich or Merck (Darmstadt, Germany); MS-grade acetonitrile and methanol from Fisher Scientific (Waltham, MA, USA); lipid standards from Avanti Polar Lipids (Alabaster, AL, USA). Ultrapure water (18 MΩ·cm at 25 °C) was obtained from an Elga Labwater system (Lane End, UK). The study is categorized into three phases in line with the recommendation in the literature: Phase I (discovery), Phase II (qualification), and Phase III (verification). In Phase I, 364 samples were investigated for the lipidomic serum profile differentiation of PDAC patients from healthy controls in the main laboratory (lab 1 - Pardubice) using UHPSFC/MS. For confirmation of results, the samples were again randomly processed and measured with shotgun MS and, for a smaller subset, with MALDI-MS in lab 1. For Phase II, new sample aliquots (554 samples) from the Masaryk Memorial Cancer Institute in Brno were obtained, further re-aliquoted, and distributed among the laboratory at University of Pardubice, Czech Republic (lab 1), the laboratory at University Hospital of Regensburg, Germany (lab 2), and the laboratory at National University of Singapore (lab 3). Each laboratory processed the sample set independently according to their preferred sample preparation method. For the quantitative lipidomic serum profile analysis in all three laboratories, no specifications of the applied MS-based method were provided. The purpose was that the individual laboratories should apply their preferred, optimized, and validated methods for lipidomic analysis. This experimental design is purposely selected to rule out that PDAC differentiation from controls and dysregulation of specific lipids is method-or laboratory-dependent. The following MS-based analytical methods were used for Phase II: UHPSFC/MS (lab 1), shotgun MS with low- and high-resolution (lab 2), and RP-UHPLC/MS (lab 3). The sample preparation protocol and lipidomic analysis were further developed and validated in lab 1 between Phase I and Phase II, and the optimized and validated conditions were applied for Phase II and III. Phase III was performed in lab 1 using UHPSFC/MS for the serum lipidomic analysis of samples obtained from different collection sites to verify that lipidomics profiling is diagnostically conclusive and independent of the sample collection site. 554 samples from Phase II are included in 830 samples of Phase III in lab 1. Blood samples were drawn after overnight fasting. For Phase I (364 samples) and Phase II (554 samples), all human serum samples and clinical data were obtained from the Bank of Biological Material (BBM) in Masaryk Memorial Cancer Institute in Brno, approved by the institutional ethical committee, and all blood donors signed informed consent. The sample selection was based on the availability of stored serum samples. The only exclusion criterion for healthy controls (normal, N) was the absence of malignant disease in the life-time history without any other exclusion criteria for other diseases. For all PDAC patients, the disease was confirmed by abdominal computed tomography and/or endoscopic ultrasound followed by needle biopsy or surgical resection. All PDAC patients and healthy controls were of Caucasian ethnicity. The samples were collected from 2013 to 2015. For Phase III (830 samples), serum samples and clinical data were provided by the BBM of Masaryk Memorial Cancer Institute in Brno (554 samples, see Phase II), by the First and Third Faculty of Medicine at the Charles University in Prague (147 samples), by the University Hospital in Pilsen (31 samples) and by the Palacký University and University Hospital in Olomouc (98 samples). All involved institutes provided the ethical approval and signed informed consent for blood collections. Participants did not obtain any compensation for their blood donation. 22 patients with chronic pancreatitis (9 females and 13 males) treated at two outpatient departments were enrolled in this study. The etiology of pancreatitis was either ethanol-induced or recurrent acute pancreatitis. The diagnosis was confirmed by imaging methods (endoscopic ultrasound or endoscopic retrograde cholangio-pancreatography). The overview and detailed description of clinical data and patient characteristics are provided in Supplementary Data 16 and 17. The samples were independently processed for each method used in the study. To avoid biases due to sample collection, sample preparation, and measurements, all samples within the particular phase were processed and measured in the randomized order. The operator had no information about the sample classification during the sample preparation and measurements. The sample sets in all phases were divided into training and validation sets to determine the assay performance using the rigid rule defined before the study that each 6th sample belongs to the validation set, and the rest constitutes the training set. The sample classification for the training set was disclosed for MDA. The classification of the validation set was disclosed after the final prediction of the validation set. Briefly, the whole blood was drawn into tubes containing no anticoagulant (Sarstedt S-Monovette, Germany) and incubated at room temperature for 60 min. Then, the samples were centrifuged at 1500 × g for 15 min, the serum was isolated, immediately frozen, and stored at −80 °C until extraction. The final lipid extraction protocol in lab 1 represents a modified Folch procedure published earlier. Human serum (25 µL) and 20 µL of the IS mixture (Supplementary Tables 1 and 2) were homogenized in 3 mL of chloroform/methanol (2:1, v/v) for 15 min in an ultrasonic bath (40 °C). When the samples reached ambient temperature, 600 µL of ammonium carbonate buffer (250 mM) was added, and the mixture was ultrasonicated for 15 min. After 3 min of centrifugation (886 × g), the organic layer was removed, and 2 mL of chloroform was added to the aqueous phase. After 15 min of ultrasonication and 3 min of centrifugation, the organic layers were combined and evaporated under a gentle stream of nitrogen. The residue was dissolved in a mixture of 500 µL of chloroform/methanol (1:1, v/v) and vortexed. The sample preparation protocol in Phase I was slightly different because only a single extraction was employed without any buffer, with different IS concentrations, and only vortexing instead of ultrasonication. Finally, the extract was diluted 1:5 with chloroform/methanol (1:1, v/v) or 1:20 with the mixture of hexane/2-propanol/chloroform (7:1.5:1.5, v/v/v) (Phase I) for the UHPSFC/MS analysis, 1:8 with chloroform/methanol/2-propanol (1:2:4, v/v/v) mixture containing 7.5 mM of ammonium acetate and 1% of acetic acid for the shotgun MS analysis, and 1:1 (v/v) with methanol for the MALDI-MS. The lipid extraction in lab 2 was performed according to the Bligh and Dyer protocol in the presence of exogenous lipid species as IS (Supplementary Table 3) using 10 µL of human serum for the extraction. Chloroform phase was recovered by the pipetting robot (Tecan Genesis RSP 150) and vacuum dried. Residues were dissolved in either 7.5 mM ammonium acetate in methanol/chloroform (3:1, v/v) (for LR-MS) or chloroform/methanol/2-propanol (1:2:4, v/v/v) with 7.5 mM ammonium formate (for HR-MS). The lipid extraction in lab 3 was performed in a randomized order using the stratified randomization based on the sample group, age, gender, and BMI. The sample extraction was done over three days (~230 samples/day). Human serum samples (~100 μL each) were taken out of −80 °C freezer into a biosafety cabinet and thawed on ice. 10 μL of each serum sample was transferred into 1.5 mL Eppendorf tubes. In addition, 5 μL of each serum sample was pooled together, mixed, and then 10 µL was aliquoted in 59 Eppendorf tubes to constitute batch quality control (BQC) samples. Process blanks (PBLK 1-4) were prepared by aliquoting 10 µL of water into 1.5 mL Eppendorf tubes for extraction control. 10 µL of commercial human plasma was pipetted into 1.5 mL Eppendorf tubes as reference samples (LTR 1-4). 10 µL of NIST SRM 1950 plasma was pipetted into 1.5 mL Eppendorf tubes as additional reference samples (NIST 1-4). The extraction was done on all above-mentioned samples as follows: Add 190 µL of chilled butanol/methanol (1:1, v/v) containing IS (Supplementary Table 4) to the samples. Vortex each sample for 10 s and sonicate in ice water for 30 min. Centrifuge at 14,000 relative centrifugal force for 10 min at 4 °C to pellet insoluble. Transfer 140 µL of supernatant into clean vials. Pool 30 µL of lipid extract from each vial (only from samples, not including BQC, NIST, and LTR), mix, and aliquot into 59 vials as technical quality control (TQC) samples. The TQC extract was diluted with chilled butanol/methanol (1:1, v/v) to prepare 80, 60, 40, and 20% diluted TQC solutions, which were used to assess the instrument response linearity. The lipid extracts in LC/MS vials were kept in the −80 °C freezer until LC/MS/MS analysis. On the day of analysis, LC/MS vials were taken out of the freezer, thawed at room temperature for 30 min, sonicated in ice-cold water for 15 min, and injected into LC/MS/MS. CA 19-9, a mucin corresponding to the sialylated Lewis (Le) blood group antigen, was quantitatively determined using the electro-chemoluminescence immunoassay Elecsys (Roche, Rotkreuz, Switzerland) according to manufacturer instructions. The CA 19-9 test was performed for all 830 samples from Phase III, whereby 2 outliers were observed and excluded from the study. The repetition of the CA 19-9 measurements for that outlier was not possible due to the limited sample amount. The cut-off value for the CA 19-9 test is 37 U/mL, therefore all values over 37 U/mL were classified as PDAC. UHPSFC/MS measurements were carried out on the Acquity Ultra Performance Convergence Chromatography (UPC) system coupled to the hybrid quadrupole-traveling wave ion mobility time-of-flight mass spectrometer Synapt G2-Si from Waters by using the commercial interface kit (Waters, Milford, MA, USA). The chromatographic settings were used with minor improvements from previously published methods. The main difference is that the data were recorded in the continuum mode. The peptide leucine enkephalin was used as the lock mass with the scan time of 0.1 s and the interval of 30 s. The lock mass was scanned but not automatically applied. The noise reduction was performed on raw files using the Waters compression tool (v4.1), and then data were lock mass corrected as well as converted into centroid data using the exact mass measure tool from Waters. For data preprocessing, the MarkerLynx software (v4.1) from Waters was used. First, the time scan range of each lipid class peak was determined from the base peak intensity (BPI) chromatogram using MassLynx (v4.1), afterwards individual scans were combined from the predefined scan range for each lipid class within a mass range of 50 mDa. Only m/z values with intensities higher than the threshold of 3000 counts were extracted. Obtained tables in MarkerLynx plotting m/z vs. intensity for each lipid class were exported and further processed using LipidQuant software for the identification and quantitation of lipids. LipidQuant (v1.0) is a laboratory-made Excel macro script written in Visual Basic, which helps with the identification of lipid species via comparing measured m/z values with exact m/z values from the embedded database with a mass tolerance of 5 mDa. The identified species were isotopic corrected and quantified by calculating the concentration (nmol/mL) of the lipid species based on the comparison of the intensity of a particular lipid with the intensity of IS of the same lipid class of known concentration. To facilitate the statistical analysis without manipulating the outcome, lipids present in less than 25% of the samples were excluded from the data set, and zero values were replaced by 80% of the minimum for all samples of the corresponding lipid species. The LipidQuant software, instructions on how to use it, details on the scan ranges, exported txt files, as well as raw data, are provided on figshare (https://figshare.com/s/cc087785ca362af7118e). Shotgun experiments were performed on the quadrupole-linear ion trap mass spectrometer 6500 QTRAP (Sciex, Concord, ON, Canada) equipped with the ESI probe. The AB Sciex Analyst software (v1.6.2) was used for the data acquisition. Characteristic precursor ion (PIS) and neutral loss (NL) scan events were used for the detection of individual lipid classes and previously reported MS settings applied. Then, the data were transferred to the LipidView software (v1.2) for further processing and alignment. For the data analysis, all observed ions in the positive ion mode characterized by m/z values, type of scan, and ion intensities were exported as.txt data file and further processed using the LipidQuant software (v1.0) available on figshare (https://figshare.com/s/b28049603a4f361c818b). MALDI matrix 9-aminoacridine (Sigma-Aldrich, St. Louis, MO, USA) was dissolved in methanol-water mixture (4:1, v/v) to the concentration of 5 mg/mL. Diluted lipid extracts of serum were mixed with matrix (1:1, v/v). The volume of 1 µL of extract/matrix mixture was deposited on the target plate using the dried droplet crystallization. A small aliquot of chloroform was applied onto MALDI plate spots before the application of the diluted extract/matrix mixture to avoid the drop spreading. Mass spectra were measured on the high resolution MALDI mass spectrometer LTQ Orbitrap XL (Thermo Fisher Scientific, Waltham, MA, USA) equipped with the nitrogen UV laser (337 nm, 60 Hz) with a beam diameter of about 80 µm × 100 µm. The LTQ Orbitrap instrument was operated in the negative ion mode over a normal mass range m/z 400−2000 with the mass resolution 100,000 (full width at half-maximum definition at m/z 400). The zig-zag sample movement with 250 µm step size was used during the data acquisition. The laser energy corresponds to 15% of the maximum, and 2 microscans/scan with 2 laser shots per microscan at 36 different positions were accumulated for each measurement to achieve a reproducible signal. Each sample (spotted matrix and lipid extract mixture) was spotted five times. The total acquisition time for one sample, including measurements of five consecutive spots, was 10 min. Each measurement was represented by one average MALDI-MS spectrum with thousands of m/z values. The automatic peak assignment was subsequently performed, and m/z peaks were matched with deprotonated molecules from a database created during the identification procedure using the LipidQuant (v1.0) software available on figshare (https://figshare.com/s/cb071be45cd91a7c90e2). This peak assignment resulted in the generation of the list of present m/z of studied lipids with average intensities over all spectra, which was used for further IS or relative normalization and statistical evaluation. The analysis of lipids was performed by direct flow injection analysis (FIA) using a triple quadrupole (QqQ) mass spectrometer (FIA-MS/MS) and a Fourier Transform (FT) hybrid quadrupole—Orbitrap mass spectrometer (FIA-FTMS). FIA-MS/MS was performed in the positive ion mode using the analytical setup and the strategy described previously. The fragment ion of m/z 184 was used for phosphatidylcholines (PC), sphingomyelins (SM), and lysophosphatidylcholines (LPC). The following neutral losses were applied for: phosphatidylethanolamines (PE) – 141, phosphatidylserines (PS) – 185, phosphatidylglycerols (PG) – 189, and phosphatidylinositols (PI) – 277 (ref. ). PE-based plasmalogens (PE-P) were analyzed according to the principles described by Zemski-Berry. Sphingosine-based ceramides (Cer) and hexosylceramides (HexCer) were analyzed using the fragment ion of m/z 264 (ref. ). FIA-FTMS setup was described in detail in previous work. Triacylglycerols (TG), diacylglycerols (DG), and cholesteryl esters (CE) were recorded in the positive ion mode in m/z range 500−1000 for 1 min with a maximum injection time (IT) of 200 ms, an automated gain control (AGC) of 1·10, 3 microscans, and a target resolution of 140,000 (at 200 m/z). The mass range of the negative ion mode was split into two parts. LPC and lysophosphatidylethanolamines (LPE) were analyzed in the range m/z 400–650. PC, PE, PS, SM, and ceramides were measured in m/z range 520–960. Multiplexed acquisition (MSX) was used for [M + NH4] of free cholesterol (FC) (m/z 404.39) and cholesterol D7 (m/z 411.43) using 0.5 min of acquisition time with the normalized collision energy of 10%, IT of 100 ms, AGC of 1·10, the isolation window of 1 Da, and the target resolution of 140,000. Data processing details were described in Höring et al. using the ALEX software, which includes the peak assignment procedure and intensity picking. The extracted data were exported to Microsoft Excel 2016 (v16.0.5239.1001) and further processed by the self-programmed Macros available on figshare (https://figshare.com/s/e336bdf3a52f04c2de1f). Lipid species were annotated according to the shorthand notation of lipid structures derived from MS. For QqQ glycerophospholipid species, the annotation was based on the assumption of even numbered carbon chains only. SM species annotation is based on the assumption that a sphingoid base with two hydroxyl groups is present. The RP-UHPLC/MS/MS analysis was performed on the Agilent UHPLC 1290 liquid chromatography system connected to the Agilent QqQ 6495 A mass spectrometer. For the data acquisition, the MassHunter software was used (vB.09.00 – B9037.0) The Agilent Eclipse Plus C18 column (100 mm × 2.1 mm, 1.8 µm) was used for the LC separation. The mobile phases A (30% acetonitrile—20% isopropanol—50% 10 mM ammonium formate in H2O, v/v/v + 0.1% formic acid) and B (90% isopropanol—9% acetonitrile—1% 10 mM ammonium formate in H2O, v/v/v + 0.1% formic acid) were used for both positive and negative ionization. The following gradient was applied: 0 min 15% B, 2.5 min 50% B, 2.6 min 57% B, 9 min 70% B, 9.1 min 93% B, 11 min 96% B, 11.1 min 100% B, 11.9 min 100% B, and 12.0 min 15% B, held for 3 min (total runtime of 15 min). The column temperature was maintained at 45 °C. The flow rate was set to 0.4 mL/min and the sample injection volume was 2 µL. The spray voltage was set to 3.5 kV in the positive ionization mode and 3 kV in the negative ionization mode. The nozzle voltage was set at 1 kV. The drying gas temperatures were kept at 150 °C. The sheath gas temperature was 250 °C. The drying gas and sheath gas flow rates were 14 and 11 L/min, respectively. The nebulizer gas setting was 20 psi. The iFunnel high- and low-pressure RF were 180 and 160 V, respectively, in the positive ionization mode and 90 and 60 V, respectively, in the negative ionization mode. The MRM list is provided in Supplementary Data 18. Quantitative data were extracted by using the Agilent MassHunter Quantitative Analysis (QqQ) software (vB.10.00). The data were manually curated to ensure that the software integrated the right peaks. Peak areas of the extracted ion chromatograms peaks for each MRM transition were exported to Microsoft Excel (v1808). Peak areas were normalized to the peak areas of IS using an in-house R (v4.0.0) script and the following packages: here (v0.1), dplyr (v0.8.5), tidyr (v1.0.3), purrr (v0.3.4), readr (v1.3.1), lubridate (v1.7.8), and stringr (v1.4.0). The data quality was assessed using the following criteria, MRM transitions kept for the analysis had to satisfy: coefficient of variation (CoV) measured across the QC injections < 20%, linearity TQC dilution series Pearson R > 0.80, signal in processed blanks < 10% of the signal observed in the QC. Data are available at figshare: https://figshare.com/s/1fd10f273b049b93fa24 The UHPSFC/MS method was validated in line with FDA and EMA guidelines, as previously published. Solvent blanks and QC samples were regularly measured after each 40 samples. For the QC samples, a pooled serum sample and the NIST SRM reference plasma sample were extracted and aliquoted. Furthermore, a mixture of naturally occurring lipid species was used as a system suitability standard. In order to assess the instrumental state, the instrument stability and sample preparation quality, the signal response of selected endogenous lipids, and the IS in all samples were monitored during the whole sequence. The signal responses of selected lipids were plotted against the number of measured samples, which allows the visualization of outliers due to sample preparation or instrumental failures. Typically, a gradual signal drop is observed for the IS caused by contamination of the mass spectrometer over time. Furthermore, PCA for the lipidomic profiles in all samples was performed to review for outliers and clustering of QC samples. SIMCA software, version 13.0.3 (Umetrics, Umeå, Sweden) was used to perform the unsupervised PCA with unclassified samples, and the supervised OPLS-DA with the known sample classification. Only scatter plots of the first and second components are presented in PCA score plots. OPLS-DA separates samples based on the known classes and can be used for prediction. Differences in lipid profiles between genders were observed in the Phase I (Fig. 2), therefore data sets for males and females were handled separately. The variables were log-transformed, centered, and scaled (unit variance (UV) or Pareto (Par) scaling) to achieve better performance and model stability. The outliers were evaluated and checked for potential data-entry errors. Logarithmic transformation was applied for each lipid species. Centering relates the relative changes of a lipid species to the average, where UV or Pareto scaling compensates the concentration variance differences for lipid species. The scaling was chosen regarding improved separation of PDAC patient and control samples and reduced number of outliers without using class information employing PCA. Pareto scaling was superior for UHPSFC/MS, MALDI-MS, low- and high-resolution shotgun MS (lab 2) and RP-UHPLC/MS (lab 3) measurements, and UV scaling for shotgun MS measurements in lab 1 during Phase I. For PCA and OPLS-DA, the number of components was assessed by model fit and prediction ability. In the case of too few components, the differentiation of classes (i.e., health state) is insufficient, while in the case of too many components, the model may be overfitted, resulting in diminished prediction power. The model fit was determined by the evaluation of R2, which describes the variation of variables (lipid species) explained by the model. The insight into the prediction ability of the model is described by Q2 and is estimated using 7-fold cross-validation. PCA plot was evaluated for outliers, errors in measurements, clustering of QC samples as well as for the separation of sample types, i.e., PDAC patients vs. healthy controls. Afterwards, OPLS-DA was performed to discriminate between PDAC patients and healthy controls. The number of predictive and orthogonal components for all methods is provided in Supplementary Table 7. OPLS models were built for the training set for individual methods and validated by the prediction of the validation set using predicted response values. The unpredicted original value of Y is 0, if a human subject is without cancer, and 1 in case of PDAC (binary variable). The predicted response value is continuous and computed using the last model component. Based on the predicted value of Y, the sample is classified as non-cancer subject (if predicted Y ≤ 0.5) or cancer subject (if predicted Y > 0.5). A summary of the predicted response values obtained for training and validation sets with the various methods at different clinical phases is provided in Supplementary Data 19 and 20. Depending on the correctly identified healthy and cancer samples, the selectivity, specificity, and accuracy of the model for the training and validation samples were determined (Supplementary Data 10). To evaluate lipids of statistical significance, a two-sided two sample T-test assuming unequal variances (Welch test) was performed for healthy and cancer samples. P-values < 0.05 were considered to indicate the statistical significance. The Bonferroni approach was applied to all p-values for the multiple testing correction. The Microsoft Excel Professional Plus 2016 software (v16.0.5239) was used for these calculations. The summary statistics and average molar lipid concentrations for healthy and cancer samples are summarized in Supplementary Data 11–13 for all methods and study phases. Furthermore, the parameter of variable influence of projection (VIP) was evaluated for each statistical OPLS-DA model using the SIMCA software. Finally, only lipid species with p-values < 0.05, VIP values >1, and fold changes ≥20% for molar concentrations were considered as statistically important and reported in Supplementary Data 11–13. For the visualization of differences in lipid concentrations (up and downregulation) between cancer and control samples, box plots were constructed in R free software environment (v3.6.2) (https://www.r-project.org) using ggplot2 (v3.3.3), ggpubr (v0.4.0), readxl (v1.3.1), dplyr (v1.0.2), and rstatix (v0.6.0) packages. In each box plot, the median was presented by a horizontal line, the box represented 1st and 3rd quartile values, and whiskers stood for 1.5*IQR from the median. Each measurement was plotted as a jittered point value. The receiver operating characteristics (ROC) curves were generated by using the package AUC (v0.3.0) in R. For verification of the data processing, statistical analysis, and results, data were cross-checked and independently reprocessed or evaluated by applying the online metabolomics platform MetaboAnalyst (v4.0). The Kaplan–Meier plots and Cox proportional hazards analysis was performed for each of the lipids. The groups of patients with values of particular lipid below the median and above median were compared in terms of survival. The Kaplan–Meier survival analysis plot and the Cox proportional-hazard analysis plots were generated by using the packages survival (v3.2-7), dplyr (v1.0.2), readxl (v1.3.1), and survminer (v0.4.8) in R software. The Cytoscape software (v3.8.0) was used to prepare Fig. 10. The Adobe Illustrator CC 2018 (v22.1, 64 bit) was applied to process graphics and prepare final figures. The QC system and the PCA analysis revealed outliers. In Phase I, sample No. 355 was excluded from the UHPSFC/MS data set, and sample No. 210 for the shotgun MS data set, due to the sample preparation failure. The repetition of the sample preparation was not possible due to insufficient serum volume. In Phase II, samples No. 246 and 500 were excluded from the low resolution shotgun MS data set, and samples No. 246 and 409 from the high resolution shotgun MS data set. Further information on research design is available in the Nature Research Reporting Summary linked to this article.
|
PMC10175254
|
Glucocerebrosidase activity and lipid levels are related to protein pathologies in Parkinson’s disease
|
Parkinson’s disease (PD) and dementia with Lewy bodies (DLB) are progressive neurodegenerative diseases characterized by the accumulation of misfolded α-synuclein in the form of Lewy pathology. While most cases are sporadic, there are rare genetic mutations that cause disease and more common variants that increase incidence of disease. The most prominent genetic mutations for PD and DLB are in the GBA1 and LRRK2 genes. GBA1 mutations are associated with decreased glucocerebrosidase activity and lysosomal accumulation of its lipid substrates, glucosylceramide and glucosylsphingosine. Previous studies have shown a link between this enzyme and lipids even in sporadic PD. However, it is unclear how the protein pathologies of disease are related to enzyme activity and glycosphingolipid levels. To address this gap in knowledge, we examined quantitative protein pathology, glucocerebrosidase activity and lipid substrates in parallel from 4 regions of 91 brains with no neurological disease, idiopathic, GBA1-linked, or LRRK2-linked PD and DLB. We find that several biomarkers are altered with respect to mutation and progression to dementia. We found mild association of glucocerebrosidase activity with disease, but a strong association of glucosylsphingosine with α-synuclein pathology, irrespective of genetic mutation. This association suggests that Lewy pathology precipitates changes in lipid levels related to progression to dementia.Parkinson’s disease (PD) presents clinically as a movement disorder and is confirmed post-mortem by the loss of dopaminergic neurons in the substantia nigra and the presence of Lewy pathology throughout the brain. Parkinson’s disease progresses to dementia (PDD) in up to 80% of cases and is pathologically nearly identical to dementia with Lewy bodies (DLB), suggesting that PD, PDD, and DLB are on a disease spectrum. Lewy pathology is composed of misfolded α-synuclein protein and additional lipids and organelles. These cytoplasmic inclusions are hypothesized to be the result of cells’ inability to clear these toxic proteins. It is not known what precipitates the initial misfolding of α-synuclein, and most PD cases are sporadic, without a known genetic cause. However, common genetic risk variants and rare familial mutations give insight into the development of PD. The most common genetic risk variants for PD lie in the GBA1 gene. GBA1 encodes the lysosomal lipid hydrolase, glucocerebrosidase (GCase). Homozygous mutations in GBA1 can lead to the lysosomal storage disease, Gaucher disease, due to the accumulation of GCase lipid substrates, glucosylceramide (GlcCer) and glucosylsphingosine (GlcSph), in the lysosome. While heterozygous carriers of the mutations will not develop Gaucher disease, they show an approximately fivefold elevated risk of developing PD. Remarkably, GBA1 variant carriers also have an 8-fold elevated risk of developing DLB, making GBA1 variants the most common risk factor for PD and DLB. The relationship between GBA1 variants and elevated risk for both PD and DLB suggests that disease risk is directly related to α-synuclein pathology development or clearance. Indeed, neuropathologically, idiopathic and GBA1-linked PD and DLB present similarly with extensive Lewy pathology. GCase activity is reduced in GBA1-PD/PDD/DLB (Supplementary Fig. 1), but to a much lesser extent than in Gaucher disease. The retained GCase activity in GBA1-PD seems sufficient to keep lipid substrates largely within normal levels, although elevated GlcCer and GlcSph have been reported in some regions. Together, these data suggest that decreased GCase activity and elevated glycosphingolipid levels may exacerbate Lewy pathology. However, the relationship between GCase and α-synuclein is not unidirectional. Work in cell and animal models has suggested that total α-synuclein levels or misfolded α-synuclein may reduce GCase activity as well. Indeed, several studies have found that GCase activity is reduced and glycosphingolipid substrates are elevated in the brains of idiopathic PD patients (Supplementary Fig. 1). However, these GCase activity and lipid changes have only been observed in certain regions, at certain ages, and have not been found consistently across cohorts. Most previous studies have focused on either GCase activity or glycosphingolipid analyses; idiopathic or GBA1-PD, making it difficult to draw conclusions across studies. Only two studies have examined the relationship of GCase activity to α-synuclein pathology, and both studies found mild negative correlations of pathological α-synuclein with GCase activity. GCase activity has also been explored in the context of other genetic risk factors for PD that impact lysosomal function. Among these is the most common genetic cause of familial PD, mutations of the LRRK2 gene. Pathogenic LRRK2 mutations, including the most prevalent G2019S, increase protein kinase activity and have been associated with altered lysosomal morphology, pH, impaired autophagic flux, and most recently, GCase dysregulation. An initial study on dried blood spots found that GCase activity was increased in LRRK2 mutation carriers manifesting PD compared to non-carriers. Similarly, elevated GCase activity was identified in PBMCs from LRRK2 carriers manifesting PD relative to healthy controls and subjects with idiopathic PD. However, another study reported reduced GCase activity in patient fibroblasts and iPSC-derived dopaminergic neurons from LRRK2 mutation carriers, an effect that was reversed with LRRK2 kinase inhibition. Overall, the influence of LRRK2 kinase activity on GCase activity has varied by cell type and methodologies applied, and GCase activity has not been assayed in brain tissue from LRRK2 mutation carriers. Together, there is substantial evidence for a role of GCase activity and glycosphingolipid substrates in the etiology of PD and DLB. However, there is still an incomplete understanding of the relationship between genetic status, disease state, GCase activity, lipid levels, and protein pathologies across brain regions. Here, we aimed to gain a systematic understanding of how each of these factors relate by examining neuropathology, GCase activity, and glycosphingolipid levels in parallel across four brain regions of idiopathic PD/PDD/DLB, GBA1-PD/PDD/DLB, LRRK2-PD/PDD and matched controls. We found that GCase activity was reduced in GBA1-PD, but not in idiopathic or LRRK2-PD. GlcSph was elevated in GBA1 and idiopathic cases, especially in individuals with dementia. Importantly, we found that GlcSph was highly correlated with both α-synuclein and tau pathologies, which themselves are highly inter-correlated, suggesting that glycosphingolipid accumulation may occur downstream of protein pathology. We sought to examine the relationships of neuropathology to GCase activity and related lipid levels in genetic and idiopathic α-synucleinopathies. We identified 28 GBA1 mutation carriers and 7 LRRK2 mutation carriers with available frozen tissue. To enable the best comparison between idiopathic and genetic cases, we selected 37 idiopathic cases that were matched for age, post-mortem interval (PMI), sex, and disease to the genetic cases. We also identified 19 non-neurologically impaired controls that were matched as closely as possible following the same criterion (Fig. 1a, Supplementary Table 1). We collected frozen tissue from four regions of each brain—cingulate cortex, frontal cortex, putamen, and cerebellum. Cingulate cortex, frontal cortex, and putamen were selected as regions that each exhibit substantial Lewy pathology, but without extensive neurodegeneration that could impact readouts. The cerebellum was chosen as a comparator region lacking Lewy pathology. Each of these regions has also been examined in previous studies, enabling direct comparisons (Supplementary Fig. 1). Tissues were chipped frozen, but fine dissections were done on thawed tissue to collect parallel pieces for histology, GCase activity and lipid analysis by mass spectrometry. For tissue not used for histology, gray matter was carefully resected from white matter to avoid contamination of myelin, which has different sphingolipid content than gray matter.Fig. 1Parallel assessment of neuropathology, GCase activity, and lipid levels in idiopathic and genetic PD.a Study design. Post-mortem brain tissue was taken from four groups (LRRK2-PD/PDD, GBA-PD/PDD/DLB, idiopathic PD/PDD/DLB and matched controls. Four regions of brain were microdissected, leaving parallel sections for histology, lipid analysis or GCase activity analysis. Samples used for lipid or GCase activity analysis had white matter carefully resected away. b Ages of subjects at death. c Post-mortem interval of tissues. d Percentages of total group represented by each sex. Male (M) are blue-green and female (F) are orange. Bars represent mean ± S.E.M. with individual values plotted. One-way ANOVA; Tukey’s multiple comparison test. *p < 0.05, **p < 0.01. a Study design. Post-mortem brain tissue was taken from four groups (LRRK2-PD/PDD, GBA-PD/PDD/DLB, idiopathic PD/PDD/DLB and matched controls. Four regions of brain were microdissected, leaving parallel sections for histology, lipid analysis or GCase activity analysis. Samples used for lipid or GCase activity analysis had white matter carefully resected away. b Ages of subjects at death. c Post-mortem interval of tissues. d Percentages of total group represented by each sex. Male (M) are blue-green and female (F) are orange. Bars represent mean ± S.E.M. with individual values plotted. One-way ANOVA; Tukey’s multiple comparison test. *p < 0.05, **p < 0.01. Age for control cases was significantly lower than several of the disease groups (Fig. 1b). This is a general feature of control tissue in the brain bank and suitable tissue from control cases of older ages could not be identified. No other major differences were observed between disease cohorts, although LRRK2 cases were on the older range. Post-mortem interval was well-matched between groups (Fig. 1c). The majority of GBA1 cases were male and sex was well-matched for all cohorts (Fig. 1d). White matter was retained on tissue used for histology to enable proper orientation for histological examination. Tissue was fixed in 10% NBF, paraffinized, mounted on slides, and stained for the four major aggregating proteins associated with neurodegenerative disease—pS129 α-synuclein, pS202/T205 tau, Aβ, and pS409/410 TDP-43. Gray matter was annotated on each slide (Fig. 2a) to enable a direct comparison to GCase activity and lipid levels. Antibodies were selected due to the high signal and low background (Fig. 2b) which enabled automated pixel thresholding to quantify area occupied by each of the stains (Fig. 2c). While we validated the presence of pS409/410 TDP-43 in control tissue, none of the experimental slides had positive TDP-43 stain, so remaining quantification was limited to α-synuclein, tau and Aβ. Quantification of the percentage of area occupied with pathology spanned several log-fold and enabled pathological comparisons between cohorts (Fig. 2d). Control tissues had minimal pathology, other than several cases that had substantial Aβ pathology. Cerebellum also served as an appropriate outgroup, as almost no pathology was observed in this region. Idiopathic and genetic PD/PDD/DLB cases mostly had substantial Lewy pathology in the regions examined, with the exception of a couple LRRK2-PD cases, which have been reported to exhibit variable Lewy pathology.Fig. 2Quantitative neuropathology of idiopathic and genetic PD.a Representative staining and annotation of gray matter from the cortex of a subject with abundant Lewy, tau, and Aβ pathology. This individual, as for all cases tested, had no apparent TDP-43 pathology. Scale bar = 1 mm. b A zoomed in image of gray matter showing abundant Lewy bodies, tau tangles, and Aβ plaques. A positive control tissue also shows abundant TDP-43 pathology. Scale bar = 50 μm. c The same images as in panel B but overlaid with a pixel detection classifier in red at the optimized threshold settings. Scale bar = 50 μm. d A heatmap of pathology measures from all assayed tissue. Tissues that were not available for assessment are indicated in gray. a Representative staining and annotation of gray matter from the cortex of a subject with abundant Lewy, tau, and Aβ pathology. This individual, as for all cases tested, had no apparent TDP-43 pathology. Scale bar = 1 mm. b A zoomed in image of gray matter showing abundant Lewy bodies, tau tangles, and Aβ plaques. A positive control tissue also shows abundant TDP-43 pathology. Scale bar = 50 μm. c The same images as in panel B but overlaid with a pixel detection classifier in red at the optimized threshold settings. Scale bar = 50 μm. d A heatmap of pathology measures from all assayed tissue. Tissues that were not available for assessment are indicated in gray. To further examine the relationship between pathology and disease cohort, we compared pathology in the cingulate cortex across patient groups. Lewy pathology (pSyn) was elevated in every group except LRRK2-PD, compared to control tissues (Fig. 3a). Further, in the idiopathic group, there was elevated Lewy pathology in iPDD and iDLB, compared to iPD (Fig. 3a). Tau pathology was also elevated in idiopathic and GBA1 groups (Fig. 3b). Aβ exhibited a striking bimodal distribution with an increased prevalence of high Aβ cases in groups with dementia (Fig. 3c). We also assessed the relatedness of each pathology type to the other in all examined tissues. pSyn and pTau pathology were highly correlated, with a few notable regions with high pTau and low pSyn, largely from the GBA1-DLB/AD group (Fig. 3d). The burden of pSyn pathology was also highly correlated with Aβ (Fig. 3e), although regions with low pSyn/high Aβ or high pSyn/low Aβ were observed.Fig. 3Neuropathological correlations.a pS129 α-synuclein (pSyn) levels in the cingulate cortex. b pS202/T205 (AT8, pTau) levels in the cingulate cortex. c Aβ levels in the cingulate cortex. Bars represent mean ± S.E.M. with individual values plotted. d pSyn and pTau levels are highly correlated across all tissues measured. Outliers with high pTau and lower pSyn are largely GBA1-DLB/AD cases. e pSyn and Aβ levels are also correlated across all tissues, but with a bimodal distribution of Aβ. Lines represent linear regression line of best-fit and shaded area is the 95% confidence interval. a, b Welch’s ANOVA test; Dunnett’s T3 multiple comparisons test. c One-way ANOVA; Tukey’s multiple comparison test. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. a pS129 α-synuclein (pSyn) levels in the cingulate cortex. b pS202/T205 (AT8, pTau) levels in the cingulate cortex. c Aβ levels in the cingulate cortex. Bars represent mean ± S.E.M. with individual values plotted. d pSyn and pTau levels are highly correlated across all tissues measured. Outliers with high pTau and lower pSyn are largely GBA1-DLB/AD cases. e pSyn and Aβ levels are also correlated across all tissues, but with a bimodal distribution of Aβ. Lines represent linear regression line of best-fit and shaded area is the 95% confidence interval. a, b Welch’s ANOVA test; Dunnett’s T3 multiple comparisons test. c One-way ANOVA; Tukey’s multiple comparison test. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. We next assayed GCase activity in all tissues. GCase activity was determined via the 4-methylumbelliferyl-β-D-glucopyranoside (4-MUG) assay using lysed tissue. GCase activity across patient groups was normalized to control brain tissue levels for ease of interpretation, and when individual regions were analyzed, values were normalized to control brain of only that region. For GCase activity and lipid analyses, two cases were removed that had either homozygous (N370S) or compound heterozygous (N370S, R463C) GBA1 genotypes due to the dramatically different GCase activity and lipid levels for these tissues. We first tested whether GCase activity differed between regions (Fig. 4a). There were large differences by region, the highest GCase activity in the cingulate, moderate activity in the putamen and cerebellum and the lowest activity in frontal cortex. GCase activity was significantly lower in GBA1 cases, irrespective of brain region (Fig. 4b–e). While there were other changes in mean GCase activity levels, there were no significant differences in the idiopathic group, and LRRK2 cases had no apparent change in GCase activity relative to controls (Fig. 4b–e). We next evaluated the relationship of GCase activity to pSyn pathology. Overall, there was no correlation of GCase activity with pSyn levels in all samples (Fig. 4f). There was a slight negative correlation in individual regions other than the cerebellum (Fig. 4g), however there was substantial variability in individual residuals. Variation was not well-explained by GBA1 mutation (Supplementary Fig. 3a, b).Fig. 4GCase activity in genetic and idiopathic PD.a GCase activity for all cases, normalized to all control case measures, separated by region. GCase activity is subsequently normalized by region and broken down by neuropathological disease and genetics for each of the four regions: (b) cingulate, (c) frontal, (d) putamen, and (e) cerebellum. Bars represent mean ± S.E.M. with individual values plotted. f Log normalized pSyn pathology plotted against normalized GCase activity for all samples. g Log normalized pSyn pathology plotted against normalized GCase activity but normalized and broken down by brain region. Lines represent linear regression line of best-fit and shaded area is the 95% confidence interval. c Welch’s ANOVA test; Dunnett’s T3 multiple comparisons test. a, b, d, e One-way ANOVA; Tukey’s multiple comparison test. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. a GCase activity for all cases, normalized to all control case measures, separated by region. GCase activity is subsequently normalized by region and broken down by neuropathological disease and genetics for each of the four regions: (b) cingulate, (c) frontal, (d) putamen, and (e) cerebellum. Bars represent mean ± S.E.M. with individual values plotted. f Log normalized pSyn pathology plotted against normalized GCase activity for all samples. g Log normalized pSyn pathology plotted against normalized GCase activity but normalized and broken down by brain region. Lines represent linear regression line of best-fit and shaded area is the 95% confidence interval. c Welch’s ANOVA test; Dunnett’s T3 multiple comparisons test. a, b, d, e One-way ANOVA; Tukey’s multiple comparison test. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. While GlcCer is one of the main substrates of GCase, it does not accumulate to a substantial degree in heterozygous GBA1 mutation carriers, or in idiopathic PD. We first sought to determine how GlcCer is distributed across brain regions. We found that GlcCer was significantly lower in the cerebellum than in the cingulate, frontal, or putamen (Fig. 5a). We observed no major accumulation of GlcCer in any region or disease group compared to controls, although the GBA1-PDD group did have a small, but significant elevation of GlcCer (Fig. 5b–e). Interestingly, there was an overall positive correlation of GlcCer and pSyn load (Fig. 5f), which seemed largely related to a correlation in the cingulate and frontal cortices (Fig. 5g). No similar relationship was observed for the stereoisomeric species GalCer (Supplementary Fig. 2). There was a substantial degree of variation in GlcCer measures, even within GBA1 mutation carriers. To determine if this was related to the mutation carried, we separated measured GlcCer by mutation (Supplementary Fig. 3c). The N370S group showed a wide range of GlcCer levels, while A456P and S196P cases showed the highest GlcCer levels. Finally, we assessed individual isoforms of GlcCer to determine if there were disease-related changes in select isoforms (Supplementary Figs. 4–9). GlcCer isoform (d18:1/22:0) was significantly elevated in the idiopathic DLB group in cingulate cortex (Supplementary Fig. 7b) and showed the highest correlation with pSyn pathology (Supplementary Fig. 7f). GlcCer isoform (d18:1/24:1) seemed to drive most of the elevation of total GlcCer in GBA1-PDD and was also elevated in idiopathic DLB cingulate, although this elevation was not statistically significant (Supplementary Fig. 9).Fig. 5GlcCer in genetic and idiopathic PD.a Total GlcCer measures for all cases, separated by brain region. GlcCer levels are subsequently broken down by neuropathological disease and genetics for each of the four regions: (b) cingulate, (c) frontal, (d) putamen, and (e) cerebellum. Bars represent mean ± S.E.M. with individual values plotted. f Log normalized pSyn pathology plotted against GlcCer levels for all samples. g Log normalized pSyn pathology plotted against GlcCer levels but broken down by brain region. Lines represent linear regression line of best-fit and shaded area is the 95% confidence interval. a Welch’s ANOVA test; Dunnett’s T3 multiple comparisons test. b–e: One-way ANOVA; Tukey’s multiple comparison test. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. a Total GlcCer measures for all cases, separated by brain region. GlcCer levels are subsequently broken down by neuropathological disease and genetics for each of the four regions: (b) cingulate, (c) frontal, (d) putamen, and (e) cerebellum. Bars represent mean ± S.E.M. with individual values plotted. f Log normalized pSyn pathology plotted against GlcCer levels for all samples. g Log normalized pSyn pathology plotted against GlcCer levels but broken down by brain region. Lines represent linear regression line of best-fit and shaded area is the 95% confidence interval. a Welch’s ANOVA test; Dunnett’s T3 multiple comparisons test. b–e: One-way ANOVA; Tukey’s multiple comparison test. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. GlcSph is present at much lower levels than GlcCer, but GlcSph levels have been reported to be increased in GBA1-PD and idiopathic PD, albeit to different levels for different regions and dependent on age (Supplementary Fig. 1). In contrast to GlcCer, GlcSph is highest in the cerebellum, with moderate levels in the cingulate cortex and putamen, and lowest levels in the frontal cortex (Fig. 6a). Individuals with a GBA1 mutation had higher levels of GlcSph, independent of region (Fig. 6b–e). Within individuals carrying GBA1 mutations, those with PDD or DLB had GlcSph statistically higher than controls. The GBA1-PD group was small and had a lower abundance of the N370S mutation than PDD or DLB/AD groups (Supplementary Fig. 3a). The low abundance of mutations other than N370S makes it difficult to make any major conclusions related to specific mutations, and major differences were not observed when GCase activity, GlcSph levels, or pSyn pathology were separated by GBA1 mutation type (Supplementary Fig. 3b–d) LRRK2 mutation carriers appeared similar to controls (Fig. 6b–e). In the idiopathic groups, there was a trend for increased GlcSph in the cingulate and frontal cortex associated with increased disease progression. This reached statistical significance in the cingulate cortex, where the iPD group had similar GlcSph levels to controls, but iDLB had significantly higher GlcSph levels than either control or iPD tissue (Fig. 6b–e). This association of GlcSph levels with progression to dementia may be related to the burden of pathology in those regions. Considering all samples together, there was no correlation between pSyn pathology and GlcSph (Fig. 6f), but this is largely due to the high GlcSph and low pSyn pathology in the cerebellum, as there was a strong correlation between pSyn pathology and GlcSph in the cingulate, frontal, and putamen (Fig. 6g). No similar relationship was observed for the stereoisomeric species GalSph (Supplementary Fig. 10).Fig. 6GlcSph in genetic and idiopathic PD.a GlcSph measures for all cases, separated by region. GlcSph levels are subsequently broken down by neuropathological disease and genetics for each of the four regions: (b) cingulate, (c) frontal, (d) putamen, and (e) cerebellum. Bars represent mean ± S.E.M. with individual values plotted. f Log normalized pSyn pathology plotted against GlcSph levels for all samples. g Log normalized pSyn pathology plotted against GlcSph levels but broken down by brain region. Lines represent linear regression line of best-fit and shaded area is the 95% confidence interval. e Welch’s ANOVA test; Dunnett’s T3 multiple comparisons test. a–d One-way ANOVA; Tukey’s multiple comparison test. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. a GlcSph measures for all cases, separated by region. GlcSph levels are subsequently broken down by neuropathological disease and genetics for each of the four regions: (b) cingulate, (c) frontal, (d) putamen, and (e) cerebellum. Bars represent mean ± S.E.M. with individual values plotted. f Log normalized pSyn pathology plotted against GlcSph levels for all samples. g Log normalized pSyn pathology plotted against GlcSph levels but broken down by brain region. Lines represent linear regression line of best-fit and shaded area is the 95% confidence interval. e Welch’s ANOVA test; Dunnett’s T3 multiple comparisons test. a–d One-way ANOVA; Tukey’s multiple comparison test. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. Given the relationships of protein pathologies with each other and relationships of pathology to GCase activity and lipids, we sought to understand overall relationships across all features measured in this study (Fig. 7, Supplementary Fig. 11). Importantly, we also wanted to know if factors such as PMI and age were related to other measures. We found that PMI had no significant correlation with any other measure. GCase activity has previously been reported to decrease with age, with a commensurate increase in lipid substrates. We found no significant correlation of age with GCase activity, or glycosphingolipid levels (Fig. 7, Supplementary Fig. 11). We looked more closely in controls and idiopathic PD at the relationship between age and GCase activity, GlcCer, and GlcSph levels (Supplementary Fig. 12). No significant correlation was observed between age and any of these measures for either group. We also assessed correlations with reported disease duration in one of the most impacted regions—cingulate cortex. We found no substantial correlation of pSyn pathology, GCase activity, or GlcCer levels with disease duration (Supplementary Fig. 13a–c). There was a slight negative correlation of GlcSph levels with disease duration (Supplementary Fig. 13d). Together, these analyses suggest that the relationship between lipids and pathology is more related to disease state than age or disease duration.Fig. 7Overall relationships between neuropathology, GCase activity, and lipid levels.Pearson’s correlations between each of the different measure factors, including age, post-mortem interval (PMI), pathology, GCase activity and lipids are plotted here separately for cingulate (a), frontal (b), putamen (c), and cerebellum (d). Several of the protein pathologies correlate with each other; several lipid levels also correlate with each other. Of note (highlighted in orange boxes) is the positive correlation between GlcSph and protein pathologies, especially in the cingulate and frontal cortices, and to a lesser extent in the putamen. GCase activity also negatively correlated with GlcSph levels, although this was only significant in putamen and cerebellum. Plots display Pearson’s correlations with Holm’s correction for multiple comparisons. Correlations are noted also by number and those comparisons with p > 0.05 have an “X” over the intersecting box. Pearson’s correlations between each of the different measure factors, including age, post-mortem interval (PMI), pathology, GCase activity and lipids are plotted here separately for cingulate (a), frontal (b), putamen (c), and cerebellum (d). Several of the protein pathologies correlate with each other; several lipid levels also correlate with each other. Of note (highlighted in orange boxes) is the positive correlation between GlcSph and protein pathologies, especially in the cingulate and frontal cortices, and to a lesser extent in the putamen. GCase activity also negatively correlated with GlcSph levels, although this was only significant in putamen and cerebellum. Plots display Pearson’s correlations with Holm’s correction for multiple comparisons. Correlations are noted also by number and those comparisons with p > 0.05 have an “X” over the intersecting box. As previously noted, there were strong correlations of each pathology with each other. Several lipids also showed high correlations with other lipids. The strongest correlation was between GalCer and GalSph. The relationship between GlcCer and GlcSph was notably weaker, only reaching significance in the cingulate and frontal cortex. The strongest correlation with pathology burden was GlcSph levels (Fig. 7, Supplementary Fig. 11). Due to the apparent associations of GlcSph with pathology, we sought to further delineate the relationships of GlcSph to brain region, GCase activity and pSyn pathology. To examine the contribution of each of these variables in an unbiased manner, we applied a regression decision tree algorithm. A decision tree is a non-parametric supervised learning algorithm used to predict a target variable by learning decision rules from predictor variables. The tree begins with all samples (i.e., root node) for a target variable, GlcSph, and splits on an independent variable that results in most homogeneous sub-nodes (i.e., leaf nodes) of GlcSph values. This partition process is continued recursively. Effectively, each split selects a variable among all variables with the lowest error in predicting GlcSph. The variable importance is then calculated based on the reduction of squared error attributed to each variable at each split and is placed on a scale of 0–100% for each independent variable. Tree models were built for healthy controls, GBA1 mutation carriers and idiopathic cases separately. Tree models help capture how variables affected GlcSph regulation. GlcSph levels were associated only by brain regions in healthy, aged controls (Fig. 8a, Supplementary Fig. 14a), consistent with the clear regional differences in GlcSph levels (Fig. 6a). However, in GBA1-PD cases, GCase activity was the primary differentiator of GlcSph levels (Fig. 8a, Supplementary Fig. 14b), consistent with GBA1 mutations driving decreased GCase activity and increased glycosphingolipid levels. This was especially true when GCase activity levels were very low. Interestingly, when GCase activity was within a moderate range, GlcSph was co-modulated by pSyn pathology. In idiopathic cases, GlcSph regulation varied regionally. GlcSph was regulated by pSyn pathology, independent of GCase activity, in the frontal cortex only. In the putamen and cingulate cortex, GlcSph was jointly regulated by both GCase activity and pSyn pathology in a complex non-linear fashion (Fig. 8a, Supplementary Fig. 14c). As expected, no significant associations were observed between GlcSph and pSyn pathology in the cerebellum in all PD cases. Together, these analyses support the influence of pSyn on GlcSph levels in the presence and absence of GBA1 mutation and reduced GCase activity.Fig. 8Variables influencing GlcSph levels in PD.a The importance of variables indicated on the left side were determined and placed within a scale of 0–100 within each group. The R-Squared (R² or the coefficient of determination) is a measure of how well the data fit the regression tree model. Variable importance was determined by calculating the relative influence of each variable contribution to increasing R², i.e., the accuracy of GlcSph prediction. GCase activity (4-MUG) is the most important variable predicting GlcSph in GBA1-PD; pSyn is the most important variable predicting GlcSph in iPD; Cerebellum region is the most important variable predicting GlcSph in controls. b GCase activity for all idiopathic PD, PDD and DLB cases are plotted against GlcSph levels, separated by region. c Log normalized pSyn pathology for all idiopathic PD, PDD and DLB cases are plotted against GlcSph levels, separated by region. Lines represent linear regression line of best-fit and shaded area is the 95% confidence interval. a The importance of variables indicated on the left side were determined and placed within a scale of 0–100 within each group. The R-Squared (R² or the coefficient of determination) is a measure of how well the data fit the regression tree model. Variable importance was determined by calculating the relative influence of each variable contribution to increasing R², i.e., the accuracy of GlcSph prediction. GCase activity (4-MUG) is the most important variable predicting GlcSph in GBA1-PD; pSyn is the most important variable predicting GlcSph in iPD; Cerebellum region is the most important variable predicting GlcSph in controls. b GCase activity for all idiopathic PD, PDD and DLB cases are plotted against GlcSph levels, separated by region. c Log normalized pSyn pathology for all idiopathic PD, PDD and DLB cases are plotted against GlcSph levels, separated by region. Lines represent linear regression line of best-fit and shaded area is the 95% confidence interval. The finding that pSyn pathology is a driver of GlcSph levels in idiopathic cases in the tree models prompted us to examine correlations specifically within idiopathic cases. First, we examined the relationship between GCase activity and GlcSph levels. Outside of the putamen, there is minimal predictivity of GlcSph levels by GCase activity measurements (Fig. 8b). An important caveat of this finding is that the assay for GCase activity used in this study measures whole tissue GCase activity. It is possible that lysosomal GCase activity is a better predictor of GlcSph levels. However, pSyn pathology showed strong predictivity of GlcSph levels outside of the cerebellum where there is no pathology (Fig. 8c). This is especially prominent in the frontal cortex, as predicted by the tree model. Together, these findings suggest that glycosphingolipids levels are driven by GCase activity in GBA1 mutation cases, while they are driven by pSyn pathology and GCase activity in idiopathic cases. GBA1 variants are the strongest genetic risk factor for developing PD, PDD, or DLB. Yet, most GBA1-PD patients retain a wildtype GBA1 allele and resulting in only modest reductions in GCase activity compared to Gaucher patients DLB. Fewer than 10% of individuals carrying a GBA1 mutation will develop PD, suggesting that factors outside of GBA1 influence incidence of disease. Multiple studies since 2012 have investigated GCase and its lipid substrates in GBA1-PD, and idiopathic PD, to better understand how GBA1 mutations impact enzyme activity and lipid status, and determine if this enzyme is also impacted in individuals without mutations (Supplementary Fig. 1). Most studies have individually examined GCase activity, lipid content, idiopathic PD, or GBA1-PD. Cohorts have also been differentially segmented by age, mutation, or disease duration, which precludes global summary of these data. However, some major themes have emerged. GCase activity is reduced in GBA1-PD brains, but this is variable and independent of region. This reduction in activity therefore seems related to the mutation itself and less related to disease status. We would hypothesize that healthy GBA1 mutation carriers would have subtle reductions in GCase activity in the brain though this has not been examined, to our knowledge. Our data are consistent with a reduction in GCase activity across brain regions in GBA1 mutation carriers, even in regions such as the cerebellum that are largely unaffected by protein pathologies. Reduced GCase activity has been reported in idiopathic PD, but the reported reduction is dependent on brain region and age. If GCase activity is related to Lewy pathology, it would be expected to be reduced only in those regions bearing pathology. However, the cerebellum showed a similar reduction of GCase activity in those studies even though cerebellum is typically devoid of α-synuclein pathology. We find no significant reduction of GCase activity in idiopathic PD, although there are certain individuals with low GCase activity, and these tend to be iPDD/DLB groups, consistent with a mild negative correlation between α-synuclein pathology and GCase activity. Lipid substrates of GCase have been examined in fewer studies, partially owing to the specialized expertize and equipment necessary to isolate GlcCer and GlcSph from their stereoisomers, GalCer and GalSph. In most studies, GlcCer is either unchanged or slightly elevated. Consistent with the literature, we found no consistent change in GlcCer levels in idiopathic or GBA1 cohorts, suggesting that GlcCer levels are only weakly linked to GCase activity or neuropathology. GlcSph, while it is less abundant than GlcCer, is much more affected in disease. All studies that have examine GlcSph levels in PD brains have found an elevation in GlcSph in some of the regions assayed. Our study expanded on these earlier findings, showing that GlcSph levels were elevated in GBA1 mutation carriers, independent of region, but partially dependent on disease (PD, PDD, DLB). In idiopathic patients, GlcSph showed a strong relationship to disease (PD < PDD < DLB), especially in cortical regions, and a strong correlation with α-synuclein pathology. Further non-parametric analyses identified pSyn pathology burden as the sole regulator of GlcSph levels in the frontal cortex. In the cingulate cortex and putamen, GlcSph was jointly regulated by pSyn and GCase activity (as determined by 4-MUG). Consistent with this analysis, pSyn pathology showed high correlations with GlcSph levels in idiopathic cases in all regions with pathology, while GCase activity only showed high correlations with GlcSph levels in the putamen. Finally, our study extended the examination of GCase activity and lipid substrate levels to PD subjects carrying LRRK2 mutations. These tissues were included due to the previous literature showing that LRRK2 kinase activity may be related to GCase activity. Despite this compelling literature, we found that the LRRK2 group had similar GCase activity and lipid levels to control brains, suggesting that LRRK2 mutations are not driving disrupted GCase activity in the tissues examined. It should be noted that both LRRK2 and GCase are highly expressed in peripheral tissues and may have a stronger interaction there. In addition, only 7 LRRK2 cases were examined, so future work with larger cohorts will be important to test this relationship. An additional item often reported in the literature which our study hoped to address is the negative feedback loop between GCase activity and pathological α-synuclein accumulation. It can be difficult to resolve where feedback loops begin once they are in place. While the current study cannot show whether a change in GlcSph or α-synuclein pathology came first, it supports that the loop may start with pathological α-synuclein. 90% of GBA1 mutation carriers never develop PD. Therefore, carrying a GBA1 mutation does not necessarily precipitate α-synuclein pathology or PD. However, all patients with GBA1-PD/PDD/DLB have α-synuclein pathology, suggesting that it is an integral feature of GBA1-PD, and that those individuals who have a GBA1 mutation will be less protected in the event of α-synuclein accumulation. But the data on GCase activity in idiopathic PD has been variable with some studies showing reduction in certain regions, and others showing no change. In the current study, GCase activity trended downward in PDD and DLB, but did not reach significance. However, GlcSph showed an increase in idiopathic patients that corresponded with progression to dementia. Across cohorts, GlcSph showed a strong correlation with α-synuclein pathology. Interestingly, GlcSph levels were highest in the cerebellum despite the fact that this region does not accumulate α-synuclein pathology. This suggests that increased levels of GlcSph alone are insufficient to drive α-synuclein pathology, at least in cerebellar neurons. To determine the mechanistic relationship between Lewy pathology, GCase activity, and glycosphingolipid accumulation, it is useful to focus on PD/PDD/DLB without GBA1 mutations. In these subjects, GCase is a poor predictor of GlcSph levels, especially in cortical regions. α-Synuclein pathology burden shows a much stronger predictivity of GlcSph levels. These data suggest that GCase itself may not be the direct driver of this relationship. Instead, α-synuclein pathology may disrupt GlcSph degradation, or change its distribution. While GlcSph accumulation in PD cases is mild compared to Gaucher disease, it may be sufficient to drive formation of more pathogenic conformations of α-synuclein. This study has several limitations. The use of post-mortem tissue precludes the ability to study disease longitudinally. While this is a general difficulty with studying the brain, there are substantial efforts underway to measure GCase activity and lipid levels from blood and cerebrospinal fluid, in the hopes that these more accessible biofluids may reflect changes in the brain. Another limitation is the number of cases. While this is one of the largest cohorts collected to date, future studies with larger numbers of cases may enable further subtype analysis to see if there are specific patients that are more likely to respond positively to GCase-targeted therapies. Another important consideration is the sampling strategy. Sampling methods are not well-reported across the field, but sampling needs to be done carefully to enable comparisons across groups. White matter, for example, has dramatically different lipid content than gray matter, necessitating careful removal of white matter, as possible. White matter was carefully resected in the current study. We also avoided the substantia nigra pars compact due to the likely shift of cell type and phenotype associated with disease. The dramatic degeneration of this region is associated not only with loss of dopaminergic neurons, but also with gliosis. These factors are likely to shift the GCase and lipid content in severely degenerated regions. However, it is also a limitation of this study that we cannot compare more brain regions. A final related limitation of all studies, including our own, is the bulk nature of the data. Different cells likely have different GCase and glycosphingolipid levels, and by taking all these cells indiscriminately, we may miss important cell-specific effects with analysis of the whole tissue. Use of 4-MUG in an acidic lysate is also limited in its ability to specifically capture lysosomal GCase activity. Future studies would benefit from analyses that retain spatial localization of GCase activity and glycosphingolipid levels. This study provides a comprehensive assessment of GCase activity, lipid substrates, and neuropathological assessment from adjacent tissues in PD. We examined this in idiopathic, GBA1, and LRRK2 PD/PDD/DLB. One of the advantages of having a large cohort for most groups in this study is that they could be stratified by progression to dementia (PD/PDD/DLB). When stratified in this way, there are clear differences in the groups, both in terms of neuropathology, as would be expected, but also in GlcSph levels, which accumulate to a greater extent in PDD/DLB than in PD. There are important remaining questions related to how each of these variables interact. While GlcSph accumulates in idiopathic and GBA1-PD, the amount of accumulation is minimal compared to that seen in individuals with homozygous GBA1 mutations. Is this accumulation sufficient to keep a negative feedback loop going? Is GlcSph accumulation a biproduct of lysosomal dysfunction? Does GlcSph accumulate to a much higher level, but only in specific cells that bear the burden of disease? Future studies in human tissue and preclinical models will help clarify the association of GCase, lipid levels and Lewy body disease. A total of 90 brains were used in this study, 18 healthy matched controls, 37 idiopathic PD, 28 GBA1-PD and 7 LRKK2-PD. All brain tissues were obtained from the Center of Neurodegenerative Disease Research (CNDR) at the University of Pennsylvania. The study protocol was approved by the University of Pennsylvania ethics committee and written informed consent was obtained from next of kin. Large brain regions were removed directly from each brain while still frozen. Each of the four brain regions was then thawed on wet ice, and carefully prosected to include only the region of interest (cingulate cortex, putamen, frontal cortex, or cerebellum). A fine slice was transferred to 10% neutral buffered formalin (NBF) for overnight fixation. The remainder of the tissue had white matter carefully removed, and ~100–200 mg of gray matter allocated into tubes for subsequent GCase activity or lipid analysis. The fixed tissue was embedded in paraffin after 24 h for further histological examination. Genomic DNA was extracted from brain tissues using QIAamp DNA mini kit (Qiagen, Germantown, MD). Mutations and variants in GBA1 and LRRK2 were identified by targeted next generation sequencing (NGS) on a neurodegenerative disease-focused panel, which includes genes associated with Parkinson’s disease, as previously described and alignment of sequence reads and variant calling from NGS were assessed by SureCall software (Agilent, Santa Clara, CA). Identified mutations and variants of interest in GBA1 and LRRK2 were confirmed by Sanger sequencing or TaqMan assay (Thermo Fisher Scientific). After fixation, brains were embedded in paraffin blocks, cut into 6 μm sections and mounted on charged glass slides. All slides were de-paraffinized with 2 sequential xylene baths (5 min each) and then incubated for 1 min in a descending series of ethanol baths: 100%, 100%, 95%, 80%, 70%. After a rinse in distilled water, antigen retrieval was performed by using either formic acid for 5 min at room temperature or citric acid pH = 6 (Vector Laboratories; Cat# H-3300) for 15 min at 95 °C. Slides were allowed to cool for 20 min at room temperature and washed in running tap water for 10 min. To quench the endogenous peroxidase, slides were incubated in 7.5% hydrogen peroxide for 30 min at room temperature. Slides were washed for 10 min in running tap water, placed for 5 min in 0.1 M Tris buffer pH = 6 and then blocked for 1 h at RT in 0.1 M Tris/2% fetal bovine serum (FBS). Slides were incubated in primary antibody in 0.1 M Tris/2% FBS in a humidified chamber overnight at 4 °C. The following antibodies were used: rabbit polyclonal anti-pS129 α-synuclein (EP1536Y) (1:20,000; Abcam, Cat# ab51253), mouse monoclonal anti-pSer202/Thr205 tau (AT8) (1:10,000; life technologies, Cat# MN1020), anti-β-Amyloid (1:200,000, NAB228, Center for Neurodegenerative Disease Research, University of Pennsylvania), rabbit monoclonal anti-pS409/410 TDP-43 (1:20,000, Proteintech, Cat# 80007-1RR). Primary antibodies were rinsed off with 0.1 M tris for 5 min and then incubated with the appropriate secondary antibody: goat anti-rabbit (1:1000, Vector, Cat# BA1000) or horse anti-mouse (1:1000, Vector, cat# BA2000) biotinylated IgG in 0.1 M tris/2% FBS for 1 h at room temperature. Slides were rinsed using 0.1 M tris for 5 min, then incubated with avidin-biotin solution (Vector, Cat#PK-6100) for 1 h. Slides were then rinsed for 5 min with 0.1 M tris, then developed with ImmPACT DAB peroxidase substrate (Vector, Cat# SK-4105) and counterstained for 15 s with Harris Hematoxylin (Fisher, Cat# 67-650-01). Slides were washed in running tap water for 5 min, dehydrated in ascending ethanol baths for 1 min each (70%, 80%, 95%, 100%, 100%) and incubated in 2 sequential xylene baths (5 min each). Slides were mounted with coverslip using Cytoseal Mounting Media (Fisher, Cat# 23-244-256). Slides were scanned at ×20 magnification using an Aperio ScanScope XT. The digitized images were then used for quantitative pathology. All sections, staining, annotation, and quantification were done blinded to disease and genotype. The digitized images were imported into QuPath software, where gray matter was manually annotated for frontal cortex, cingulate cortex, and cerebellum). Putamen contained interspersed white matter tracts that were not removed. Optical density thresholds of 0.35 were set for each protein (α-synuclein, tau, and β-amyloid) immunostaining so only pathological signal was detected. The percentage of positive area occupied was then measured for each stain. For TDP-43, pathology was detected in the positive control tissue, but not in any of the cases, so no quantification was performed. Linear regressions and one-way ANOVA test followed by Dunn’s post hoc were performed in GraphPad Prism 9. Tissue lysates were prepared in 300 µL ice-cold lysis buffer containing 50 mM Tris–HCl, pH 7.4, 1% (by vol) Triton X-100, 10% (by vol) glycerol, 0.15 M NaCl, 1 mM sodium orthovanadate, 50 mM NaF, 10 mM 2-glycerophosphate, 5 mM sodium pyrophosphate, 1 µg/ml microcystin-LR, and complete EDTA-free protease inhibitor cocktail (Roche, 11836170001). The tissue and buffer were placed in a 2 mL round bottom tube (Eppendorf) with a steal bead (Biospec Products, 6.35 mm, 11079635 C) and homogenized for 90 s at 30% amplitude in a Qiagen TissueLyser at 4 °C. Lysates were centrifuged at 20,000 x g for 30 min and supernatants were collected. Total protein was determined with a Micro BCA Protein Assay per kit protocol (Thermo Scientific, 23235). To determine Gcase activity, lysates were diluted 1:5 (cingulate cortex, cerebellum, putamen) or 1:10 (frontal cortex) into assay buffer containing 0.25% sodium taurocholate (Cayman Chemical, 16215), 0.25% triton-X-100 (Thermo Fisher, A16046.AP), 1 mM EDTA (Millipore, 150-38-9) and citric acid sodium phosphate buffer (pH 5.4). Lysate were incubated shaking, in the presence or absence of conduritol B-epoxide (CBE, Cayman Chemical, 15216), at room temperature for 30 min. Seventy-five microliters of 1.25 mM 4-Methylumbelliferyl β-D-glucopyranoside prepared in 1% BSA (Research Products Inc., A30075) assay buffer was added to 25 µL of lysate. Samples were incubated, protected from light, for 60 min shaking at 37 °C. The reaction was stopped by adding 150 µL ice cold 1 M glycine (pH 12.5). Plates were read (Ex 355, Em 460) on a BioTek Citation5. Two technical replicates were performed for each sample, with and without CBE, and corrected by background subtraction. GCase activity was quantified as [(Sample – Sample)/total protein] and normalized to the control group. Extraction and quantification of tissue glucosylceramides and glucosylsphingosine was performed as previously established. Briefly, frozen brain specimens supplied as resected gray matter isolates were homogenized in MeOH:H2O (1:1) and normalized across samples by weight. Lipid extraction was performed at room temperature by combining 50 μL of homogenate in with an additional 150 μL of MeOH spiked with d3 glucosylceramide d18:1/16:0 and C6-glucosylsphingosine (d18:1) standards (Matreya, LLC, State College, PA) at 200 ng/mL and 4 ng/mL respectively prior. Solutions were mixed and 200 μL Acetone:MeOH (1:1) added before brief centrifugation and resuspension with 100 μL of H2O. Extract solutions were then subjected to centrifugation at 10,000 × g, and supernatants (250 μL × 2) were transferred to new 96-well plates containing 200 μL MeOH:H2O (1:1) per well. Lipid analytes were isolated from supernatants and preconcentrated via either C18 solid phase extraction (isolute C18, Biotage AB, Uppsala, Sweden) for glucosylceramides or strong cation exchange (Oasis MCX, Waters Corp. Inc., Milford, MA) for glucosylsphingosine as previously reported. Eluates were evaporated to dryness under gentle N2 gas and reconstituted in 50 μL DMSO and 200 μL mobile phase B liquid chromatography buffer (see below). Specimens were processed within the same batch run and analyzed by LC-MS/MS overnight. Multiple reaction monitoring targeted LC-MS/MS quantification of selected glycosphingolipids was performed using a Waters Acquity UPLC (Waters Corp., Inc.) and SCIEX 5500 QTRAP mass spectrometer (Sciex LLC, Framingham, MA) running in positive ion electrospray mode. Separation was performed using a HALO HILIC 2.7 mm column (Advanced Materials Technology, Inc., Wilmington, DE) and 10 min normal phase LC gradient (mobile phase A: 0.1% formic acid in H2O; mobile phase B: 95% acetonitrile, 2.5% MeOH, 2.0% H2O, 0.5% formic acid and 5 mM ammonium formate). Transitions for selected endogenous glucosylceramide fatty acid chain length variants were as follows: C16:0 m/z 700.5 > 264.2, C18:0 m/z 728.6 > 264.2, C20:0 m/z 756.6 > 264.2, C22:0 m/z 784.6 > 264.2, C23:0 m/z 798.6 > 264.2, C24:1 m/z 810.6 > 264.2, C24:0 m/z 812.6 > 264.2 and d3 glucosylceramide d18:1/16:0 reference standard m/z transition was 703.5 > 264.2. Linear calibration curves using d5 labeled glucosylceramide d18:1/18:0 standard (Avanti Polar Lipids, Inc., Alabaster, Alabama; m/z 733.6 > 269.4) were used to estimate concentrations of each of the targeted glucosylceramide fatty acid variants; total glucosylceramide values were represented as the sum concentrations of the C16:0 through C24:1 fatty acid isoforms. Glucosylsphingosine was monitored as a single analyte (m/z 462.4 > 282.1), and concentrations were determined using linear calibration curves of glucosylsphingosine and C6-glucosylsphingosine (m/z 467.4 > 282.1) synthetic standards (Matreya). Peak area and curve fit quantification were performed using SciEx MultiQuant software. Single trees were built based on CART algorithm using R language package, rpart. We built three separate trees for healthy controls, GBA1 mutation carriers and idiopathic PD. The tree model performance and variable importance are evaluated via tenfold cross-validation using ‘train’ and ‘vmrImp’ functions from caret package. R-squared was calculated to estimate the prediction accuracy for each tree. Variable importance is a relative measure of each variable contribution to accuracy of prediction. It was scaled to a maximum value of 100. GraphPad Prism software version 9.3.1 (GraphPad Software Inc., La Jolla, CA, USA) was used for pair-wise statistical analysis. The data shown in this study are mean ± standard error of the mean (SEM). For comparison of groups, a Brown-Forsythe test was first applied to test if variances were significantly different. If group variances were not different, a one-way ANOVA was applied with Tukey’s multiple comparison test to determine differences between any groups. If group variances were different, Welch’s ANOVA test was applied with Dunnett’s T3 multiple comparisons to determine if there were differences between groups. Linear regressions and correlation coefficients were all calculated in R (https://www.R-project.org/). Correlation matrices were generated using the ‘ggcorrmat’ function in the ‘ggstatsplot’ package in R. Further information on research design is available in the Nature Research Reporting Summary linked to this article.
|
PMC5779697
|
Adipose tissue ATGL modifies the cardiac lipidome in pressure-overload-induced left ventricular failure
|
Adipose tissue lipolysis occurs during the development of heart failure as a consequence of chronic adrenergic stimulation. However, the impact of enhanced adipose triacylglycerol hydrolysis mediated by adipose triglyceride lipase (ATGL) on cardiac function is unclear. To investigate the role of adipose tissue lipolysis during heart failure, we generated mice with tissue-specific deletion of ATGL (atATGL-KO). atATGL-KO mice were subjected to transverse aortic constriction (TAC) to induce pressure-mediated cardiac failure. The cardiac mouse lipidome and the human plasma lipidome from healthy controls (n = 10) and patients with systolic heart failure (HFrEF, n = 13) were analyzed by MS-based shotgun lipidomics. TAC-induced increases in left ventricular mass (LVM) and diastolic LV inner diameter were significantly attenuated in atATGL-KO mice compared to wild type (wt) -mice. More importantly, atATGL-KO mice were protected against TAC-induced systolic LV failure. Perturbation of lipolysis in the adipose tissue of atATGL-KO mice resulted in the prevention of the major cardiac lipidome changes observed after TAC in wt-mice. Profound changes occurred in the lipid class of phosphatidylethanolamines (PE) in which multiple PE-species were markedly induced in failing wt-hearts, which was attenuated in atATGL-KO hearts. Moreover, selected heart failure-induced PE species in mouse hearts were also induced in plasma samples from patients with chronic heart failure. TAC-induced cardiac PE induction resulted in decreased PC/ PE-species ratios associated with increased apoptotic marker expression in failing wt-hearts, a process absent in atATGL-KO hearts. Perturbation of adipose tissue lipolysis by ATGL-deficiency ameliorated pressure-induced heart failure and the potentially deleterious cardiac lipidome changes that accompany this pathological process, namely the induction of specific PE species. Non-cardiac ATGL-mediated modulation of the cardiac lipidome may play an important role in the pathogenesis of chronic heart failure.The development of chronic systolic heart failure is marked by a continuous incline in adrenergic activity that results in the typical clinical signs of this disease, including tachycardia and increased peripheral vascular resistance [1, 2]. In addition to these cardiovascular effects, chronic sympathetic activation exerts crucial effects on adipose tissue metabolism, including a marked induction of adipose tissue lipolysis via β-receptor stimulation . Adipose tissue lipolysis is further enhanced by natriuretic peptides such as atrial natriuretic peptide (ANP) and brain natriuretic peptide (BNP), both of which are elevated in chronic heart failure . Together, these processes lead to a higher lipolytic rate in patients with heart failure compared to control subjects . Increased adipose tissue lipolysis is associated with elevated plasma levels of free fatty acids (FFA) in patients with chronic heart failure [5, 6]. This ultimately leads to systemic metabolic disturbances such as insulin resistance and glucose intolerance, which further perpetuate the cycle of impaired left ventricular dysfunction [7, 8]. However, the immediate consequences of stimulated adipose tissue lipolysis on cardiac morphology and function during heart failure development are largely unknown. Triacylglycerol (TAG) hydrolysis in adipose tissue occurs in a stepwise process and involves three major lipases: adipose triglyceride lipase (ATGL), hormone sensitive lipase (HSL), and monoglyceride lipase (MGL), the cumulative effect of which is an efflux of glycerol and FFA from adipocytes into the circulation . ATGL, also referred to as patatin-like phospholipase containing A2 (PNPLA2), catalyzes the first step of TAG hydrolysis to DAGs . HSL has broad substrate specificity and mediates the hydrolysis of TAGs to diacylglycerols (DAGs), and DAGs to monoacylglycerols (MAGs), as well as the cleavage of FAs from MAGs . Of these three functions, HSL is most efficient at regulating DAG hydrolysis . Finally, MGL mediates the hydrolysis of MAGs to glycerol and FAs . Mice carrying a constitutive deletion of ATGL showed a marked reduction of FA release from white adipocytes and of TAG accumulation in adipose tissue . Similar findings were observed in mice with conditional ATGL-deletion in adipose tissue demonstrating reduced adipocyte lipolysis, decreased FA release from white adipose tissue associated with decreased non-esterified FA- and TAG-blood levels, and impaired brown adipose tissue function [10, 11]. Moreover, the deletion of ATGL in adipocytes markedly improved hepatic insulin signaling and attenuated inflammatory reactions in the liver, confirming the presence of crosstalk between adipose tissue and the liver . These data corroborate the important role of adipose tissue ATGL in the regulation of blood lipid levels, and the potential of this enzyme to affect the function of non-adipose tissue organs, particularly under conditions of increased lipolysis (e.g. fasting, exercise, heart failure). Along this line, we recently demonstrated that deletion of ATGL in adipose tissue attenuated the development of exercise-induced physiological cardiac hypertrophy in female mice . More importantly, perturbation of adipose tissue lipolysis by ATGL deficiency modified the circulating lipid profile, and a selected FA, namely C16:1 palmitoleic acid, was identified as a new molecular mediator of training-induced cardiac responses . Interorgan communication between adipose tissue and other organs has been recently described as a crucial regulator of insulin action. In accordance with our results, Cao and colleagues identified C16:1 palmitoleic acid as a lipokine released from adipose tissue modulating insulin sensitivity in muscle and liver . Thus, communication between adipose tissue and other organs appears to be a common process during the initiation and progression of different diseases . In this study, we report that perturbation of adipose tissue lipolysis by conditional deletion of ATGL protects against pressure-induced cardiac failure. By applying an untargeted shotgun lipidomics approach, we identified cardiac lipidome changes significantly associated with LV failure and demonstrated that these were prevented by ATGL deficiency in adipose tissue. In particular, failing wild-type (wt) hearts exhibited a marked increase of multiple phosphatidylethanolamines (PE) (PE16:0–18:1, PE16:0–18:2, PE16:0–20:4, PE17:0–20:4, PE18:0–20:4, PE18:0–22:4, PE18:1–18:1, PE18:2–18:0, PE18:2–19:0, PE20:4–20:0) whereas ceramides (Cer) (Cer36:1; Cer38:1; Cer38:2), and cardiolipin (CL76:12) were decreased. These changes were prominently attenuated in atATGL-KO hearts. Moreover, distinct heart-failure-regulated PE species in mouse hearts were modified in a similar manner in plasma samples from patients with chronic heart failure. These findings strongly suggest that adipose tissue lipolysis determines cardiac morphology and function during heart failure development. This inter-organ crosstalk likely involves an alteration of the cardiac lipidome which may mediate functional impairment by mechanisms such as induction of apoptosis or mitochondrial dysfunction. We previously described that fabp4-Cre-driven ATGL deletion resulted in a lack of ATGL in white adipose tissue but not in the heart or in bone marrow-derived macrophages, and in an attenuation of adipose tissue lipolysis . To exclude that the fabp4-Cre mediates deletion of ATGL in endothelial cells, as previously described , we analyzed ATGL expression in cardiac endothelial cells. ATGL expression did not significantly differ between wt- and atATGL-KO mice (S1A Fig). Mice were challenged with transverse aortic constriction (TAC) to induce pressure-mediated cardiac enlargement or failure. Pressure gradients (pre-/ post-stenotic) in TAC mice assessed by Doppler echocardiography did not differ between wt- and atATGL-KO mice, indicating similar degree of stenosis, LV-pressure load and aortic vascular function in both genotypes (wt-TAC: 54.2±2.5mmHg vs. atATGL-KO-TAC: 48.1±3.4mmHg; p = n.s.). Eleven weeks post TAC, wt hearts increased markedly in size and weight compared to wt-sham mice (Fig 1A and 1B). TAC-induced cardiac enlargement was attenuated in atATGL-KO mice (Fig 1A and 1B). Accordingly, the myocardial area in cardiac cross section increased significantly in wt-mice after TAC whereas this effect was diminished in atATGL-KO mice (Fig 1C and 1D). Echocardiographic analysis revealed a significant rise in left ventricular mass (LVM) in wt-mice 11 weeks after TAC (Fig 1E–1G and S1 Table). The induction of LVM by pressure overload was clearly suppressed in atATGL-KO mice (Fig 1E–1G and S1 Table). In addition, pressure-induced LV enlargement in wt-mice shown by the increased diastolic LV inner diameter (LVIDd) was completely prevented in atATGL-KO mice (Fig 1H and S1 Table). To assess LV function after TAC, LV ejection fraction (EF) and fractional shortening (FS) were analyzed in wt and atATGL-KO mice (Fig 1I and 1J). In wt-mice, pressure-overload led to a significant reduction in EF and FS after 11 weeks (Fig 1I and 1J), but this was not observed in atATGL-KO mice (Fig 1I and 1J). In keeping with the morphological changes described above, marker genes for cardiac failure such as the beta-myosin heavy chain (β-Myhc) were also markedly induced by TAC in wt-hearts but not in atATGL-KO hearts (Fig 1K). Cardiac fibrosis was significantly induced during pressure overload in wt-mice as shown by collagen fiber staining (Fig 1L–1N) and mRNA expression of the fibrotic marker genes, collagen (Col) 1a1 and 3 (Fig 1O and 1P). In atATGL-KO mice we observed a moderate induction of collagen fibers (Fig 1L–1N) and fibrotic marker genes (Fig 1O and 1P) by TAC, however, this did not reach statistical significance (Fig 1N) suggesting attenuated pressure-induced fibrotic remodeling in the absence of atATGL. Together, these data show that perturbation of ATGL in adipose tissue protects the heart against pressure-induced LV enlargement and systolic failure. A: Representative images of the hearts. B: Heart weight (HW)/ body weight (BW) ratio (mean and SEM, n = 5–6). C: Representative microscopic cross-sections of the hearts stained with hematoxylin/ eosin (H/E). D: Myocardial area, calculated based on microscopic sections of heart tissue, stained with H/E, analogue to the images presented in C (mean and SEM, n = 5–6). E: Representative M-Mode images of the echocardiographic analysis. F-J: cardiac echocardiographic analysis of mice (mean and SEM, n = 7): F: Left-ventricular mass (LVM). G: LVM relative to tibia length (LVM/TL). H: Left-ventricular internal diameter in diastole (LVID-d). I: Ejection fraction [%] (EF). J: Fractional shortening [%] (FS). K: Analysis of mRNA expression of beta-cardiac myosin heavy chain isogene (βMHCH), qRT-PCR studies were carried out using total RNA isolated from LV tissue. Data are presented as x-fold over wt-sham mice (mean and SEM, n = 5–6). L: Representative microscopic cross-sections of the hearts stained with picrosirius red. M: Representative high magnification images from picrosirius red-stained sections. N: Cardiac fibrosis calculated based on microscopic sections of the heart tissue, stained with picrosirius red, analogue to the images presented in L: 0 = no fibrosis, 1 = mild fibrosis, 2 = moderate fibrosis, 3 = severe fibrosis (mean and SEM, n = 5–6). O and P: Analysis of mRNA expression of collagen (Col) 1a1 (O) and Col3 (P), qRT-PCR studies were carried out using total RNA isolated from LV tissue. Data are presented as x-fold over wt-sham mice (mean and SEM, n = 5–6).*p<0.05 vs. wt sham, **p<0.01 vs. wt sham, ***p<0.001 vs. wt sham, ****p<0.0001 vs. wt sham, $ p<0.05 vs. wt TAC, $ $ p<0.01 vs. wt TAC, $ $ $ p<0.001 vs. wt TAC, $ $ $ $ p<0.0001 vs. wt TAC, ##p<0.01 vs. atATGL-KO sham; 2-way ANOVA (Bonferroni post-test). Reduced adipose tissue lipolysis in atATGL-KO mice led to improved insulin sensitivity and glucose tolerance compared to their wt littermates (Fig 2A–2D). In addition, perturbation of adipose ATGL abrogated TAG-hydrolysis leading to a diminished release of glycerol and FAs in the circulation . Accordingly, multiple TAG species were significantly increased in white adipose tissue from atATGL-KO mice compared to wt littermates (S1B Fig). Primarily, TAGs with shorter FA chain length and lower desaturation grade were significantly regulated in atATGL-KO mice. This is in accordance with ATGL´s substrate selectivity which declines with increasing FA chain length and desaturation . To identify potential factors involved in the regulation of cardiac function in atATGL-KO mice, we next examined whether the lack of ATGL regulates the release of preferential FAs from adipose tissue in a selective manner. Our analysis measured the total amount of FAs present in serum and not the amount of free circulating FAs. After alkaline hydrolysis, the total serum concentration of distinct FAs (independent of their head group) was analyzed using HPLC coupled with a triple quad mass spectrometer with electrospray ionization. As depicted in Fig 2E (TAC) and S1C Fig (sham), lack of ATGL in adipose tissue resulted in a significant reduction of selected circulating FAs including C16:0 (palmitic acid), C16:1 (palmitoleic acid), C18:1 (oleic acid), C18:2 (linoleic acid), and C20:5 (eicospentanoic acid) in TAC mice (Fig 2E). In accordance, non-esterified FA (NEFA) and TAG serum levels were also significantly lower in atATGL-KO mice compared to wt-mice (Fig 2F and 2G). A: Intraperitoneal Glucose Tolerance Test (ipGTT), (n = 4–6, 2-way ANOVA (Bonferroni post-test) from AUC). B: Area under the curve (AUC) of ipGTT, (mean and SEM, n = 4–6, 2-way ANOVA (Bonferroni posttest)). C: Insulin Tolerance Test (ITT), (n = 7–8, 2-way ANOVA (Bonferroni posttest) from AUC). D: Area under the curve (AUC) of ITT, (mean and SEM, n = 7–8, 2-way ANOVA (Bonferroni posttest)). E: Profile of selected serum FAs in TAC-operated mice analyzed by rapid resolution HPLC/ Tandem MS. F: Serum level of non-esterified fatty acids (NEFAs) in wt-TAC and atATGL-KO-TAC mice. G: Serum level of triacylglyerols (TAGs) in wt-TAC and atATGL-KO-TAC mice. (mean and SEM, n = 5, or as indicated, unpaired t-test). ***p<0.001 vs. wt sham, $ p<0.05 vs. wt TAC, $ $ $ p<0.001 vs. wt TAC, $ $ $ $ p<0.001 vs. wt TAC. We hypothesized that dysregulated serum FAs may serve as the biochemical bridge between the lack of ATGL in adipose tissue and the observed cardiac phenotype. Thus, we next investigated whether modifications of the circulating FA profile are accompanied by changes in the cardiac lipidome using MS-based shotgun lipidomics. A total of 225 lipid species from 18 different lipid classes were analyzed in LV samples, and included glycerolipids, glycerophospholipids, sterol lipids and sphingolipids (S2 Table). The most abundant lipid classes in mouse hearts were PCs, PEs, and CLs (Fig 3A). Comparison of lipid classes in each genotype revealed significant changes in levels of the low-abundant classes of Cers, LPIs, PC O-s, and PIs in failing wt-hearts vs. sham controls (Fig 3A). These changes were not significant in atATGL-KO mice (Fig 3A). All other classes were not significantly altered by the intervention (Fig 3A). To compare the abundance of individual lipid species between the 4 groups, we used a linear model and performed three tests to capture the effects of the genotype (wt; atATGL-KO), the intervention (sham; TAC) and the sum of both and their interaction. Multiple testing correction was applied, and all lipid species with a significant p-value (adjusted) in at least one test are shown in the heat map (Fig 3B). Data for each lipid species are presented as row z-scores of the mean log2-transformed relative values. The complete data set of the analysis is shown in S2 Table. MS-based shotgun lipidomics analysis of heart tissue samples (LV) isolated 11 weeks after intervention (sham or TAC) from wild-type (wt) or atATGL-KO mice. A: For each lipid class, mean total mole percent values ± SEM are shown on a logarithmic scale. To test for differential changes, a Mann-Whitney U test between sham and TAC in wt- and atATGL-KO mice, respectively, was performed. Adjusted p-values are indicated: *p<0.05 vs. wt sham, **p<0.01 vs. wt sham. B: Differentially regulated lipid species were filtered by testing for the effect of the genotype (wt; atATGL-KO), the intervention (sham; TAC) and the sum of both plus their interaction using a linear model. Species with an FDR adjusted p-value < 0.1 in at least one test were retained. Hierarchical clustering of the corresponding t-statistics yielded four distinct groups: lipid species upregulated in TAC (blue), upregulated in atATGL-KO (red), downregulated in atATGL-KO (green), downregulated in TAC (black). Data for each lipid species are presented as row scaled z-scores of mean log-transformed relative changes. Lipid classes: Cer: ceramide, Chol: cholesterol, CL: cardiolipin, DAG: diacylglycerol, LPC: lyso-phosphatidylcholine, LPE: lyso-phosphatidylethanolamine, LPE O-: lyso-phosphatidylethanolamine-ether, LPI: lyso-phosphatidylinositol, LPS: lyso-phosphatidylserine, PC: phosphatidylcholine, PC O-: phosphatidylcholine-ether, PE: phosphatidylethanolamine, PE O-: phosphatidylethanolamine-ether, PG: phosphatidylglycerol, PI: phosphatidylinositol, PS: phosphatidylserine, SM: sphingomyelin, TAG: triacylglycerol. Among the identified lipid species, the most significant change was an increase in individual PE species in failing wt-hearts (TAC) compared to wt-sham hearts (Fig 3B). This increase was clearly attenuated in mice lacking adipose ATGL comparing atATGL-KO TAC mice and their sham controls (Fig 3B). A similar effect was detected for two sphingomyelins (SM), SM34:1 and SM42:2, for phosphatidylglycerol (PG), PG16:0–20:4, and for phosphatidylinositol (PI), PI20:4–18:0 (Fig 3B). In contrast, other lipid species were markedly downregulated in failing hearts from wt-mice, an effect broadly extenuated in atATGL-KO mice (Fig 3B). The specific lipid species reduced by pressure-overload in wt-mice included PC O 16:1–22:6, Cers (Cer36:1; Cer38:1; Cer38:2), and the CL76:12 (Fig 3B). This downregulation was attenuated in atATGL-KO mice (Fig 3B). Comparison of wt- and atATGL-KO mice independent of the intervention revealed primarily changes of a small number of lipid species including the species from the PC class: upregulation of PC20:4–16:1 and PC18:0–18:1 and downregulation of PC18:2–15:0 and PC18:2–17:1 (Fig 3B). In summary, lack of ATGL in adipose tissue and reduced lipolysis associated with the reduction of selected circulating FAs mediates pressure-induced modifications of distinct cardiac lipid species among which PEs display the most apparent regulation. In order to better understand the degree and relevance of cardiac lipid changes during LV failure, we performed an in-depth analysis of all lipid classes in which the log2-fold change (sham vs. TAC), and the concentration of each lipid species (mean mole %) was included in a pairwise comparison (sham vs. TAC) in wt and atATGL-KO mice. Fig 4A and 4B provides an overview of all lipid classes and species in wt and atATGL-KO mice, respectively. The overall extent of regulation of cardiac lipid species in sham vs. TAC mice was obviously greater in wt-mice compared to atATGL-KO mice, as reflected by the vertical distribution pattern on the scatter plot (Fig 4A and 4B). TAC-induced regulation of lipid species in wt-hearts (Fig 4A, triangles) was robustly attenuated in atATGL-KO mice (Fig 4B). Detailed evaluation of each regulated lipid class revealed a total of 10 PE-species significantly upregulated during LV failure in wt-hearts containing mostly saturated, mono- or polyunsaturated long-chain FAs (Fig 4C). The PE with the highest concentration in wt LV samples (PE18:0–20:4) exhibited an induction of 1.4-fold over sham controls (Fig 4C). The maximum PE induction by TAC intervention was 2.1-fold over sham for PE16:0–20:4. Some PE-species in the higher concentration range were not significantly regulated by TAC (Fig 4C, circles), likely explaining the absence of a statistical significant regulation of the total PE class, as shown in Fig 3A. Induction of significantly regulated PEs during heart failure development was absent in atATGL-KO mice (Fig 4D). In addition, we identified 4 ether-PEs (PE O-) containing a fatty alcohol instead of a fatty acid at the sn1-position which were upregulated in failing wt-hearts but not in atATGL-KO hearts (Fig 4C and 4D). Among the PE O-s, PE O- 16:0–20:4 showed the most dramatic induction over sham controls in wt mice (3.1-fold over sham) (Fig 4C). In accordance to the diacyl PE regulation, hearts from atATGL-KO mice lacked the significant PE O- induction after TAC (Fig 4D). The majority of the differentially regulated PE- and PE O- species in failing wt-hearts contained at least one FA from the depleted blood FA pool (C16:0, C16:1; C18:1, C18:2, C20:5, Fig 2E) in atATGL-KO mice (Fig 4C), suggesting a link between blood FAs and cardiac lipid species in the present model. In contrast to the PE- and PE O species, PC- and PC O- species were not significantly regulated in failing wt-hearts, or in atATGL-KO hearts (Fig 4E and 4F). MS-based shotgun lipidomics analysis of heart tissue samples (LV) isolated 11 weeks after intervention (sham or TAC) from wild-type (wt) or atATGL-KO mice. Lipid class denotations see Fig 3. A-F: Mean log2-fold change (TAC vs. sham) vs. mean mole percent of lipid species. Triangles represent significantly changed lipid species (FDR adjusted p-value < 0.1 and absolute value of log2-fold change ≥ 0.5); bubbles represent those which are not significantly changed; size indicates log-transformed adjusted p-values. A, C, E: wt-mice. B, D, F: atATGL-KO-mice. G+H: Significantly changed PC-PE ratios of matched FAs in wt-mice (G) or atATGL-KO mice (H). The mean ratio ± SEM is shown on a logarithmic scale, Mann-Whitney U test for TAC vs. sham: *p<0.05, **p<0.01 (adjusted) in wt-mice, no significant changes were found in at ATGL-KO mice. I: upper panels: WB analysis of heart lysates using antibodies against cleaved caspase 3 and Bcl-associated X protein (Bax); lower panel: WB analysis of HL-1 cardiomyocytes lysates from cells stimulated with vehicle (Veh) or fatty acid (FA) mix (C16:0, C18:1,C18:2 in different concentrations) using antibodies against cleaved caspase 3; loading control: β–actin, glyceraldehyde 3-phosphate dehydrogenase (GAPDH). Recently, a decrease of the PC/ PE ratio has been identified to regulate cell membrane integrity in hepatocytes, and was accompanied by pronounced cell damage . Thus, we calculated the ratio of the identified PCs and PEs matched for bound FAs in our model. In agreement with the decline of LV function in TAC-operated wt-mice, cardiac PC/ PE-ratios were reduced in failing wt-hearts compared to wt-sham hearts (Fig 4G). In atATGL-KO mice, PC/ PE ratios did not significantly differ between sham- and TAC-mice (Fig 4H). The highest ratio was detected in wt-sham hearts (S3 Table). Since the decrease of PC/ PE ratios has been associated with cell damage, we evaluated the expression of the pro-apoptotic markers cleaved caspase 3 and Bax in LV samples from sham- and TAC-operated wt- and atATGL-KO mice. As shown in Fig 4I (upper panels), both proteins were induced in failing LVs from wt-mice but not atATGL-KO mice. In accordance, when HL-1 cardiomyocytes were stimulated with FAs (C16:0, C18:1, C18:2) regulated by ATGL in-vivo, we detected increased apoptosis by cleaved caspase 3 protein expression (Fig 4I lower panel). Cleaved caspase 3 was not induced in HL-1 cells under physiological high serum conditions (S1D Fig) suggesting that increased FA-levels are required for the induction of apoptosis in these cells. In summary, these data demonstrate that pressure-mediated LV failure is associated with an increase of distinct cardiac PE-species leading to a reduction of cardiac PC/PE ratios. The decline of PC/ PE ratios is accompanied by the enhanced expression of cardiac apoptosis markers. Deficiency of adipose tissue ATGL prevents cardiac PE upregulation in an FA-specific manner and protects the heart against LV failure. To explore the relevance of our findings to LV failure in humans, we performed shotgun lipidomics analysis using plasma samples obtained from male patients with HFrEF (n = 13) and male control subjects with normal LV function (n = 10). The clinical characteristics of both groups are outlined in Table 1 and S4 Table. Patients with HFrEF had a mean LVEF of 25 ± 4.8% compared to 60.9 ± 3.3% in the control group (Table 1). HFrEF patients were on average older than subjects in the control group and had a slightly higher BMI (Table 1). Shown are means ± SD, [min.-max.], or [%]; MI: myocardial infarction; LV: left ventricular We analyzed 147 lipid species from 13 lipid classes in human plasma (S5 Table and Fig 5A). PCs, TAGs, sterols (ST) and sterol esters (SE) were found in the highest concentration in human plasma (Fig 5A). HFrEF patients had significantly higher plasma levels of PEs and PIs, and lower plasma levels of SE compared to controls (Fig 5A). We subsequently analyzed the individual lipid species within a lipid class and detected 10 plasma lipid species which were significantly differentially regulated between controls and HFrEF patients using a linear model to account for age and BMI (Fig 5B, triangles, and Fig 5C). Among the regulated lipid species, 3 PE species (PE16:0–18:1, PE16:0–18:2, PE18:2–18:0) were significantly higher in blood plasma from HFrEF patients compared to controls (Fig 5B and 5C, S5 Table). Interestingly, all of the PE species with increased levels in blood plasma from HFrEF patients were also upregulated in the hearts of wt mice with LV failure (Fig 4C), suggesting a link between circulating and cardiac PE levels. The strongest rise was observed for PE16:0–18:2, with 5.6-fold higher levels in HFrEF compared to controls (S5 Table and Fig 5B and 5C). In line with the increase in PEs, distinct DAG species (which are known precursors of PEs) were higher in HFrEF patients including DAG16:0–18:2 and DAG18:2–18:2 (S5 Table and Fig 5B and 5C). In contrast to the mouse data, we discovered two PC species (PC16:0–16:0, PC17:1–16:0) and Cer38:1 upregulated, and 2 PC O- species (PC O- 16:1–16:0, PC O- 18:2–18:2) downregulated in plasma from HFrEF patients (S5 Table and Fig 5B and 5C). PC O- species were depleted from the same FAs (C16:0; C16:1, and C18:2) which were concurrently increased in PE species, suggesting a possibility of selective FA partitioning to distinct glycerophospholipid classes in HFrEF (S5 Table and Fig 5B and 5C). A: Box plots show distribution of total mole percent values for each lipid class corrected for age and BMI (see supplementary methods). To test for differential changes a Mann-Whitney U test between HFrEF-patients and controls was performed. Adjusted p-values are indicated: *p<0.05, **p<0.01, **p<0.001. B: Estimated mean log2 fold change (HFrEF vs. control) vs. estimated mean mole percent of lipid species in the control group (see supplementary methods). Triangles represent significantly changed lipid species (FDR adjusted p-value < 0.1 and absolute value of log2-fold change ≥ 0.5), bubbles show those which are not significantly changed, size indicates log-transformed adjusted p-values. C: Bar graph shows the estimated log2-fold change (HFrEF vs. control) ± regression standard error of differentially changed lipid species (see B.). Colors represent log10 adjusted p-values as indicated. Lipid classes: Cer: ceramide, DAG: diacylglycerol, LPC: lyso-phosphatidylcholine, LPE: lyso-phosphatidylethanolamine, PC: phosphatidylcholine, PC O-: phosphatidylcholine-ether, PE: phosphatidylethanolamine, PE O-: phosphatidylethanolamine-ether, PI: phosphatidylinositol, SE: sterol ester, SM: sphingomyelin, ST: sterols, TAG: triacylglycerol. Taken together, our data from human plasma samples demonstrate that the presence of HFrEF is accompanied by a dysregulation of individual lipid classes and lipid species. Interestingly, comparable changes in PE species were observed in failing mouse hearts and in plasma from HFrEF patients. In this study, we demonstrated for the first time that pressure-induced cardiac failure is accompanied by marked changes in the cardiac lipidome. The most robust alteration in lipid species was an upregulation of distinct PE species in failing wt-hearts. Perturbation of adipose tissue lipolysis in atATGL-KO mice led to an improvement of pressure-induced heart failure accompanied by the prevention of cardiac PE induction. Pressure-induced cardiac PE induction resulted in a decreased PC/ PE-ratio and a concurrent increase in apoptotic marker expression in failing wt-hearts, a process which was absent in atATGL-KO hearts. Taken together, this study demonstrates that the cardiac lipidome undergoes profound changes during heart failure development. The absence of these changes in atATGL-KO mice suggests that non-cardiac tissues such as adipose tissue, and in particular ATGL-mediated adipose tissue lipolysis, participate in the regulation of the cardiac lipidome. The role of ATGL in heart failure development and cardiac metabolism has been previously studied, with most studies focusing mainly on cardiac ATGL function [18–20]. In this study, we concentrated on adipose ATGL and the impact of adipose tissue lipolysis on the development of pressure-induced LV failure. Whereas previous studies demonstrated that perturbation of cardiac ATGL resulted in the accumulation of TAG and impaired systolic LV function [9, 18, 21], we showed that the deletion of ATGL in adipose tissue prevents pressure-induced LV failure. This finding suggests that adipose tissue lipolysis has a deleterious effect during heart failure development. These data are in accordance with the positive association between augmented adipose tissue lipolysis and impairment of LV function in heart failure [5, 6]. The relevance of adipose tissue lipolysis in heart failure development has recently been supported by a study in mice deficient for the adipose-specific perilipin-1 isoform . Perilipin-1 has been shown to prevent TAG hydrolysis in adipocytes [22, 23]. The lack of perilipin-1 in these mice resulted in enhanced fat tissue lipolysis associated with the development of hypertrophic cardiomyopathy and LV failure . Together, these data support a role for adipose tissue lipolysis in the pathogenesis of LV failure. Communication between the heart, the vasculature and other organs has been recently described as an important pathophysiological mechanism involved in the development of cardiovascular disease . Cross-talk between skeletal muscle—heart, or between kidney—heart/ vasculature has gained increasing attention in the treatment of diseases such as cardiac cachexia or the cardiorenal syndrome [24, 25]. Adipose tissue communicates with multiple organs such as liver and skeletal muscle , or as described in this study with the heart. Whether this cross-talk will result in new therapies for heart failure requires additional studies. Recently, Schweiger and colleagues characterized a new small molecule inhibitor of ATGL, named Atglistatin, in a mouse model of diet-induced obesity . They described that Atglistatin preferentially inhibits lipolysis in adipose tissue thereby improving weight gain, insulin resistance and liver steatosis, demonstrating that an adipose tissue-based pharmacological therapy of non-adipose organs might be feasible. . Lipidomic analysis using advanced MS-technologies has been recently utilized in cardiovascular research for the identification of new biomarkers in the context of risk prediction . Stegemann and colleagues identified plasma TAGs and CEs with short acyl chains and low degree of unsaturation, such as TAG54:2 and CE16:1, as strong predictors for cardiovascular disease in the Bruneck cohort. In contrast to the analysis in blood plasma, data on lipid profiling in cardiac tissue are very limited and have focused primarily on lipid class analysis. These data have shown that in patients with heart failure and obesity or diabetes, cardiac TAG levels are increased . In contrast, in the absence of metabolic disease, cardiac TAG levels were significantly decreased in patients with advanced heart failure undergoing left ventricular assist device (LVAD) implantation, while Cers and DAGs were increased . In a recent review, Schulze and colleagues concluded from this that cardiac lipotoxicity in heart failure is primarily driven by Cers and DAGs and not TAGs. In the present study, Cers, DAGs, and TAGs were not significantly increased in failing hearts from wt-mice or in human plasma. On the contrary, cardiac Cer levels were mildly reduced in failing LVs from wt-mice. One explanation for this discrepancy might be the absence of systemic insulin resistance in our study. Systemic insulin resistance was significantly induced in patients with heart failure in the LVAD study , whereas in our model, pressure overload-mediated LV failure was not accompanied by an impairment of insulin sensitivity or glucose tolerance. This supports the notion, also described by Chokshi and colleagues that the occurrence of systemic insulin resistance in heart failure is linked to the regulation of cardiac Cer- and DAG-levels, a process not present in our model. In addition, we found an increase of cardiac PIs in wt-TAC mice compared to sham mice, a regulation absent in atATGL-KO mice. PIs are important precursors of PI (4,5) biphosphate (PIP2), a crucial generator of two second messenger molecules, DAGs and inositol 1,4,5-triphosphate (IP3) . Both signaling molecules have been shown to be involved in the pathogenesis of cardiac hypertrophy and heart failure [32, 33]. Thus, the upregulation of PIs in failing wt-hearts together with the absence of the regulation in atATGL-KO mice, may provide one potential additional underlying mechanism of the observed cardiac phenotype in our model. The results described in this study are the first to demonstrate an induction of cardiac PE species during the development of LV failure. Furthermore, we showed that these cardiac lipidome changes were dependent on adipose tissue lipolysis. PEs are the most abundant phospholipids in mammalian cells, and are predominantly located in plasma membranes and mitochondrial membranes . Disruption of whole-body PE synthesis in CTP:phosphoethanolamine cytidylyltransferase (Pcyt2) +/- mice resulted in myocardial hypertrophy and cardiac dysfunction in male mice . However, these mice also developed a pronounced metabolic syndrome with abdominal obesity, hyperlipidemia, insulin resistance and liver steatosis which likely mediated the cardiac phenotype, thus making a direct comparison with our model difficult . In this study, we found increased levels of PEs in human plasma samples and significantly higher levels of individual cardiac PE species during heart failure. Our data point towards a functional role of selected PE species in the pathogenesis of LV failure. Moreover, the reduction in circulating levels of distinct FAs in atATGL-KO mice predominantly affected the level of cardiac PE species, suggesting on the one hand that cardiac lipid species concentration is determined by plasma FA levels, and on the other hand that the FA distribution into different lipid classes occurs in a head group-specific manner. This process of cardiac FA partitioning into different cardiac lipid classes, and its functional relevance has recently gained greater attention . In the context of pathological cardiac hypertrophy and failure, incorporation of oleate (C18:1) and linoleate (C18:2), but not other FAs, into cardiac DAGs has been associated with profound LV hypertrophy and systolic failure [36, 37]. In our analysis, DAG species were not significantly altered in failing wt-hearts, but two DAG species (16:0–18:2; 18:2–18:2) were elevated in plasma from HFrEF patients. Since DAGs are required for PE synthesis via the Kennedy pathway , the increase in cardiac tissue PE species observed in our study may be a potential consequence of higher circulating DAG levels. The fact that cardiac DAGs were not elevated in our model may suggest that our analysis was performed at a different time point during the pathogenesis of LV hypertrophy or failure compared to previous studies . How does the increase in PE species affect the development of LV failure during pressure overload, and can the absence of this regulation explain why atATGL-KO mice are protected from LV failure? PEs are major constituents of plasma membranes . Previous work in primary rat cardiomyocytes demonstrated that an increase of membrane PE content induces sarcolemmal destabilization and destruction . Moreover, lowering the PE levels in cardiomyocytes, as seen for distinct PE-species in atATGL-KO mice, attenuated cardiomyocyte damage induced by ischemic or metabolic stress by maintaining the physicochemical properties of the sarcolemma . Similarly, the ratio of PCs to PEs has been identified as a determinant of membrane integrity . Decreased PC/ PE ratios induced membrane permeability in hepatocytes resulting in cell damage . We calculated species-specific PC/PE ratios and discovered significant decreases accompanied with an increased Bax and cleaved caspase 3 expression only in failing wt-hearts but not in atATGL-KO hearts. Since we analyzed whole heart tissue we can currently not exactly define the cellular/ subcellular localization of the observed PC/PE ratio changes. There are several potential scenarios connecting the decreased PC/PE-ratio with the increased pro-apoptotic level of Bax/ cleaved caspase 3. Bax is known to control the intrinsic pathway of apoptosis promoting mitochondrial outer membrane (MOM) permeabilization and the release of pro-apoptotic factors . In addition, changes of MOM´s lipid composition including a dysregulation of PE content or a decrease of the PC/PE ratio can affect the biogenesis of MOM proteins as well as Bax function [39, 40]. Thus, these data suggest that high cardiac PE levels and subsequent low PC/ PE ratios may affect MOM integrity thereby promoting Bax-mediated apoptosis. An alternative way may involve the extrinsic arm of apoptosis by which extracellular stress signals via death receptors initiate intracellular pro-apoptotic signaling. Death receptor signaling involves receptor redistribution into membrane lipid rafts, and lipid raft assembly is, at least in part, controlled by membrane lipid composition [41, 42]. Thus, one may hypothesize that changes of plasma membrane PC/PE content possibly augment death receptor signaling thereby promoting cardiomyocyte apoptosis, a process absent in atATGL-KO mice. However, future experiments are required to delineate the exact link between the decreased PC/PE ratios, increased Bax/ cleaved caspase 3 expression, and the impact on myocardial apoptosis. An additional mechanism of how enhanced ATGL-activity/ AT-lipolysis and increased cardiac PE-levels may impair cardiac function might be the modulation of β-adrenergic signaling. Cardiac β2-adrenergic stimulation has been recently shown to be protective in CHF . In parallel, phospholipids have been identified as G-protein-coupled receptor modulators, and in particular, increased levels of membrane PEs stabilize the β2-receptor in an inactive state . Thus in addition to pro-apoptotic effects, the cardiac PE-increase in our study may have affected cardiac function by inhibiting protective cardiac β2-signaling. We also identified changes in cardiac sphingolipid and CL levels. In particular, CLs have been implicated in the regulation of cardiac energy metabolism and function [45, 46]. A decrease in CLs, which are major mitochondrial membrane phospholipids, results in mitochondrial dysfunction with impaired oxidative phosphorylation, and has been associated with LV dysfunction . In our study, we observed a reduction in CL 76:12 and it is possible that CL depletion may also have contributed to pressure-induced LV failure in our model. Finally, changes in distinct plasma lipid classes and individual lipid species were shown in a small group of HFrEF patients when compared to a non-HFrEF control group. These data allude to the presence of lipidomic plasma changes in human heart failure which have been previously described in a metabolomics study . In our study, the changes in selected PE species in human plasma corresponded to the changes in cardiac PE species observed in mice with LV failure, and this observation may be the first to suggest that PE dysregulation also occurs in human HFrEF pathogenesis. However, correlations between plasma/ serum and cardiac lipid compositions need to be judged very cautiously. For instance, the phospholipid composition varies markedly between plasma/ serum and tissues . PEs, highly abundant in cardiac tissue, represent only 4% of the phospholipid fraction in plasma . Tonks and colleagues investigated the correlation between plasma and muscle glycerolipids, sphingolipids, and cholesterol esters . Of the 40 detected sphingolipid species in plasma and muscle only 8 correlated significantly . This is further supported by a recent work from Ji and colleagues investigating the concentration of Cer species in serum and cardiac tissue samples from control and heart failure patients . From 7 Cer species significantly regulated in serum only 3 exhibited significant changes in myocardial samples . Together these data indicate a more complex relationship between tissue and plasma/ serum lipid composition. It is very likely that during the process of FA partitioning secreted FAs from adipose TAGs are incorporated into distinct plasma/ serum lipid classes which differ from the acceptor lipid class in target organs such as the heart . Taken together, the results of this study demonstrate for the first time that perturbation of adipose tissue lipolysis by deletion of AT-specific ATGL leads to an improvement in pressure-induced heart failure, and can prevent potentially deleterious cardiac lipidome changes such as PE induction. Furthermore, we showed that the development of LV failure is associated with a distinct lipidomic profile in human plasma and mouse hearts, suggesting a process of cardiac FA partitioning into different cardiac lipid classes. Further studies are required to improve our understanding of the underlying mechanisms and functional consequences of cardiac FA partitioning during heart failure, and to identify novel therapeutic targets for the treatment of this disease. Male adipose tissue specific adipose triglyceride lipase (ATGL) deficient mice (atATGL-KO) and control littermates (wt) were generated, as described previously . Mice were randomized to sham or transverse aortic constriction (TAC) operated groups. The TAC procedure was performed as previously described . The detailed TAC protocol is provided in the S1 Supporting Information. Eleven weeks after surgery, mice were euthanized under isoflurane anesthesia by cervical dislocation. All animal procedures were performed according to the guidelines of the Charité Universitätsmedizin Berlin and were approved by the Landesamt für Gesundheit und Soziales (Berlin, Germany) for the use of laboratory animals and according to the current version of the German Law on protection of animals. Echocardiographic analysis was performed 11 weeks after sham/TAC-intervention, using Vevo 770 high-resolution imaging system with a 30-MHz transducer (RMV-707B; VisualSonics, Toronto, Canada) as described previously . Animals were metabolically phenotyped after 10 weeks. Intraperitoneal glucose tolerance tests (ipGTT) and intraperitoneal insulin tolerance tests (ipITT) were performed using a dose of 1 g/kg body weight (BW) glucose and by injecting 0.25 U/kg BW insulin (ActrapidHM, Novo Nordisk), respectively, as described previously . Tail vein blood was used for glucose quantification during ipGTT and ipITT using a Glucometer (Precision Xtra, Abbott). Mouse HL-1 cardiomyocytes, kindly provided by W.C. Claycomb (Louisiana State University, LA) were cultivated and stimulated, as described previously . Briefly, cells were starved for 24h and afterwards stimulated for 6 h with a mix of C16:0, C18:1 and C18:2, dissolved in 10% FFAs-free BSA. For experiments FAs were used in both equimolar- or 5x lower- serum concentration (FAmix and FAmix –low, respectively). mRNA analysis and western immunoblotting were performed as previously described . The details of these procedures are outlined in the S1 Supporting Information. Cardiac tissue samples were formalin-fixed, paraffin-embedded and stained with Hematoxylin/Eosin, and Picrosirius red, and analyzed, as previously described [12, 56]. Cardiac myocyte cross sectional area was assessed using H/E-stained sections of myocardium. Perimeter of the cell borders of 50 myocytes cut transversally were measured using cellSens (Olympus). Myocytes were selected in up to ten 400x microscopic fields each separated into 20 equally sized squares. Only one transversally myocyte per square was measured to increase representativeness. Fibrotic content (collagen fibers stained with picrosirius red) was assessed from the total cross section by a blinded expert in veterinary pathology using an analysis software (Olympus) and a semi-quantitative scoring system (0 = no fibrosis, 1 = occasionally collagen fibers (mild), 2 = several collagen fibers (moderate), 3 = massive accumulation of collagen fibers (severe)). Human fasting plasma samples were collected from healthy subjects and patients with systolic heart failure (HFrEF) at the Division of Cardiology, University Hospital RWTH Aachen, Germany. Human plasma was isolated from EDTA blood samples by centrifugation at 2000rpm for 10min. Supernatant was collected and stored at -80°C as 200μL aliquots. Left ventricular (LV) ejection fraction was assessed by standard echocardiography. Blood sampling and clinical data collection and analysis was approved by an institutional review board (Ethics committee University Hospital RWTH, Aachen, Germany), and all participants provided written informed consent. Mouse serum was isolated from whole blood. Blood samples were placed for 60min at room temperature for induction of coagulation. Afterwards, samples were centrifuged at 5000rpm for 5min at 4°C, and supernatant (serum) was collected. For FA serum profiling, 40μl samples of murine serum were hydrolyzed under alkaline-methanolic conditions, neutralized and diluted 1:10 in methanol containing internal FA standards. The HPLC measurement (Agilent 1200 HPLC system), coupled with an Agilent 6460 triple quad MS was performed, as described previously . Serum concentration of FFAs was measured using a HR-NEFA kit (WAKO), as described previously . TAGs in serum were determined using the DiaSys Triglycerides FS kit (DiaSys GmbH), according to the manufacturer’s instructions. Lipids from adipose tissue were extracted using a two-step chloroform/ methanol procedure . Samples were analyzed by direct infusion on a QExactive mass spectrometer (Thermo Scientific) equipped with a TriVersa NanoMate ion source (Advion Biosciences). Further details see section Lipidomic Analysis. Lipid extraction and analysis of mouse LV samples (3mg per sample) and human plasma (15μl per sample) were performed as described previously [58, 59] at Lipotype GmbH: Lipids were extracted using a two-step chloroform/methanol procedure . Samples were spiked with internal lipid standard mixture containing: cardiolipin 16:1/15:0/15:0/15:0 (CL), ceramide 18:1;2/17:0 (Cer), diacylglycerol 17:0/17:0 (DAG), hexosylceramide 18:1;2/12:0 (HexCer), lyso-phosphatidate 17:0 (LPA), lyso-phosphatidylcholine 12:0 (LPC), lyso-phosphatidylethanolamine 17:1 (LPE), lyso-phosphatidylglycerol 17:1 (LPG), lyso-phosphatidylinositol 17:1 (LPI), lyso-phosphatidylserine 17:1 (LPS), phosphatidate 17:0/17:0 (PA), phosphatidylcholine 17:0/17:0 (PC), phosphatidylethanolamine 17:0/17:0 (PE), phosphatidylglycerol 17:0/17:0 (PG), phosphatidylinositol 16:0/16:0 (PI), phosphatidylserine 17:0/17:0 (PS), cholesterol ester 20:0 (CE), sphingomyelin 18:1;2/12:0;0 (SM), triacylglycerol 17:0/17:0/17:0 (TAG). After extraction, the organic phase was transferred to an infusion plate and dried in a speed vacuum concentrator. 1st step dry extract was re-suspended in 7.5 mM ammonium acetate in chloroform/methanol/propanol (1:2:4, V:V:V) and 2nd step dry extract in 33% ethanol solution of methylamine in chloroform/methanol (0.003:5:1; V:V:V). All liquid handling steps were performed using Hamilton Robotics STARlet robotic platform with the Anti Droplet Control feature for organic solvents pipetting. Samples were analyzed by direct infusion on a QExactive mass spectrometer (Thermo Scientific) equipped with a TriVersa NanoMate ion source (Advion Biosciences). Samples were analyzed in both positive and negative ion modes with a resolution of Rm/z = 200 = 280000 for MS and Rm/z = 200 = 17500 for MSMS experiments, in a single acquisition. MSMS was triggered by an inclusion list encompassing corresponding MS mass ranges scanned in 1 Da increments . Both MS and MSMS data were combined to monitor CE, DAG and TAG ions as ammonium adducts; PC, PC O-, as acetate adducts; and CL, PA, PE, PE O-, PG, PI and PS as deprotonated anions. MS only was used to monitor LPA, LPE, LPE O-, LPI and LPS as deprotonated anions; Cer, HexCer, SM, LPC and LPC O- as acetate adducts. Data were analyzed with in-house developed lipid identification software based on LipidXplorer [60, 61]. Only lipid identifications with a signal-to-noise ratio >5, and a signal intensity 5-fold higher than in corresponding blank samples were considered for further data analysis. Measured lipid amounts were normalized to total lipid amount in samples and log-transformed to approach a symmetric and approximative normal distribution. Lipid species with missing values across many samples were removed from further analysis, and in total, we evaluated the results of 225 lipid species in the mouse and 147 in the human data set. Normal distribution was checked visually by using Q-Q plots. Depending on their scale and distribution, results are given in proportions, mean, standard deviation or standard error of the median with 25% to 75% quartiles as indicated. Statistical tests for significance were performed by using the tow-tailed t-test or Mann-Whitney-U-test as appropriate. Multivariable analysis was done by linear and robust linear regression with log-transformed relative lipid amounts as dependent and experimental or clinical conditions, age and BMI, as independent variables. For class level comparisons only, we used regression or median imputation of missing values. To avoid alpha inflation, p-values were adjusted by the Benjamini-Hochberg procedure . Adjusted p-values of less than <0.1 were considered to be significant. All analyses were performed with R version 3.3.1. See S1 Supporting Information for further details regarding data analysis. To compare differences between groups of mice phenotypes, we performed two-way ANOVA (Bonferroni posttest), two-way ANOVA with repeated measures (Bonferroni posttest) or t-test, as appropriate, and data was analyzed with GraphPad Prism Software. Statistical significance was assumed at p<0.05. Vertical lines in the histograms indicate means ± standard error of the mean (SEM). The n-number is indicated for each experiment.
|
PMC11413214
|
Deciphering metabolomics and lipidomics landscape in zebrafish hypertrophic cardiomyopathy model
|
To elucidate the lipidomic and metabolomic alterations associated with hypertrophic cardiomyopathy (HCM) pathogenesis, we utilized cmybpc3-/- zebrafish model. Fatty acid profiling revealed variability of 10 fatty acids profiles, with heterozygous (HT) and homozygous (HM) groups exhibiting distinct patterns. Hierarchical cluster analysis and multivariate analyses demonstrated a clear separation of HM from HT and control (CO) groups related to cardiac remodeling. Lipidomic profiling identified 257 annotated lipids, with two significantly dysregulated between CO and HT, and 59 between HM and CO. Acylcarnitines and phosphatidylcholines were identified as key contributors to group differentiation, suggesting a shift in energy source. Untargeted metabolomics revealed 110 and 53 significantly dysregulated metabolites. Pathway enrichment analysis highlighted perturbations in multiple metabolic pathways in the HM group, including nicotinate, nicotinamide, purine, glyoxylate, dicarboxylate, glycerophospholipid, pyrimidine, and amino acid metabolism. Our study provides comprehensive insights into the lipidomic and metabolomic unique signatures associated with cmybpc3-/- induced HCM in zebrafish. The identified biomarkers and dysregulated pathways shed light on the metabolic perturbations underlying HCM pathology, offering potential targets for further investigation and potential new therapeutic interventions.Hypertrophic Cardiomyopathy (HCM) is a disorder with an estimated prevalence of 1 in 500, clinically characterized by left ventricular myocardial wall hypertrophy in the absence of other systemic, metabolic, or cardiac diseases. It has a critical complication of sudden death due to cardiac failure, especially in asymptomatic athletes and younger patients. HCM is a heterogeneous disease caused by pathogenic variation in genes encoding sarcomeric, Z-disc, and calcium-controlling proteins. The most prevalent genetic variants impact sarcomeric-encoding proteins (40–60%), affecting the contractility of the heart. These include myosin heavy chain 7 (MYH7), myosin binding protein C3 (MYBPC3), cardiac troponin T2 (TNNT2) and cardiac troponin I3 (TNNI3). With a frequency of ~ 30–40%, pathogenic variants in MYBPC3 are the most common cause of HCM. This is attributed to its vital role in sarcomeric structural configuration and function through modulation of myofilament sliding velocity and its key contribution to the actin-myosin interaction, which is responsible for contractility. Pathogenic MYBPC3 variants are typically frameshift, non-sense, or splice site variants, which lead to premature termination codons (PTC). Consequently, PTCs are degraded through non-sense mediated RNA decay, which could lead to allelic loss of function, impacting the function of the protein and interaction with other associated proteins. Despite the significant contribution of genetic variants in understanding the etiology of HCM, they have been proven insufficient in fully capturing the overall associated clinical and pathological features related to clinical presentations and severity. Thus, the clinical picture of ~ 40% of HCM patients remains unresolved when assessed through clinical examination and genetic testing. Changes in cardiac structure and function in HCM often correlate with changes in energy consumption as a compensation mechanism in HCM. In the heart, ATP is the primary source of energy and adaptation to energy demands throughout the day due to physiological and nutritional changes is crucial. ATP can be sourced from metabolites including amino acids, lipids, and carbohydrates. Disturbances in ATP supply could thus lead to alterations in cardiac activity and eventual cardiovascular diseases (CVDs). Cellular studies have depicted that HCM-associated variants increase the demand for utilizing ATP, thus emphasizing a higher requirement for myocardial energy in HCM patients. As there is insufficient ATPase activity in such patients, it is hypothesized that the higher demand for energy leads to energy stress and adverse remodeling within the heart. Thus, understanding the changes in cardiac metabolic activity could provide an avenue for further understanding of CVDs, including HCM. Through the integration of metabolomics and lipidomics profiling, the identification of distinct metabolite features and changes associated with a genotype, environmental factors, including diet and physical activity could lead to a wider understanding of HCM etiology. The zebrafish model has been reported to possess a similar metabolism to humans, making it an increasingly common target for metabolomic studies. Zebrafish cmybpc3 knockout line has been previously employed and proven as an efficient model to study the association of genetic variants and clinical presentation. To shed more light on the relationship between metabolic activity and HCM, with the aim to improve diagnosis efficiency and patient stratification, lipidomics, metabolomics, and fatty acid signatures, analysis of heterozygous and homozygous established zebrafish MYBPC3-HCM models is performed. The ultimate goal is to tackle the pathophysiological interactions of biomarkers related to HCM progression. We analyzed 21 zebrafish homogenate samples through lipidomics and metabolomics to study HCM signatures in order to identify, relatively quantify, and statistically compare lipids and metabolites between the experimental groups of transgenic reporter Tg myl7:eGFP line as a control (CO), cmybpc3+/-::myl7:eGFP heterozygous (HT), and cmybpc3-/-::myl7:eGFP homozygous (HM) (Fig. 1) using liquid chromatography-high resolution mass spectrometry. In total, we have analyzed 7 replicates for each group of CO, HT and HM; each replicate included pooled 150 larvae. Each sample’s performance was monitored with the addition of isotopically labeled internal standards through analysis. These internal standards were monitored for mass accuracy, peak area response, and retention time variability. All parameters were within accepted criteria of 3ppm, +/- 30% of the median and +/- 2% of the median, respectively. The selection of features for analysis was designed to capture the metabolic and lipidomic alterations associated with HCM. Diverse data sources, including an in-house MS2 library and mzCloud for metabolomics and LipidSearch for lipidomics, were incorporated to provide wide coverage of potential biomarkers and metabolic pathways. Fig. 1Representative images of the zebrafish larvae at 5 days old. (A) Representative lateral view of zebrafish larvae (control (CO), cmybpc3+/-::myl7:eGFP heterozygous (HT), and cmybpc3-/-::myl7:eGFP homozygous (HM))Transgenic reporter line myl7:egfp expressing green fluorescent two-chambered heart (500 μm scale bar). (B) A ventral view of HM heart at 168x high-magnification image captured using Axiozoom Zeiss stereomicroscope equipped with Axiocam camera, scale bar 50 μm. Representative images of the zebrafish larvae at 5 days old. (A) Representative lateral view of zebrafish larvae (control (CO), cmybpc3+/-::myl7:eGFP heterozygous (HT), and cmybpc3-/-::myl7:eGFP homozygous (HM))Transgenic reporter line myl7:egfp expressing green fluorescent two-chambered heart (500 μm scale bar). (B) A ventral view of HM heart at 168x high-magnification image captured using Axiozoom Zeiss stereomicroscope equipped with Axiocam camera, scale bar 50 μm. Fatty acids, acylcarnitines and lipidomic changes associated with cmybpc3-HCM zebrafish model. A total of 10 fatty acids were detected in the zebrafish lysate that showed variability in fatty acids profile between the different groups. Hierarchical cluster analysis revealed that both HT and HM groups showed variability in concentrations of detected fatty acids (Fig. 2). Fig. 2Hierarchical cluster analysis of fatty acids. Z-score clustering results for fatty acids are shown as a heatmap for all samples using MetaboAnalyst 5.0. The distance measure was Euclidean, and clustering method was Ward. The rows and columns are ordered so that similar groups and similar metabolites are close to one another. The fatty acids are present in the rows, and the sample groups in columns. Dark blue to dark red color gradient denotes lower to higher levels. On the top of the heatmap, the homozygous sample group is represented in blue, the heterozygous group in green and the control group in red. Hierarchical cluster analysis of fatty acids. Z-score clustering results for fatty acids are shown as a heatmap for all samples using MetaboAnalyst 5.0. The distance measure was Euclidean, and clustering method was Ward. The rows and columns are ordered so that similar groups and similar metabolites are close to one another. The fatty acids are present in the rows, and the sample groups in columns. Dark blue to dark red color gradient denotes lower to higher levels. On the top of the heatmap, the homozygous sample group is represented in blue, the heterozygous group in green and the control group in red. PCA showed the clustering of the three examined groups with the first two principal components explaining 81.8% of the variability. The HM group showed a relatively distinct separation in comparison to the CO and HT groups, despite the 95% confidence region overlapping between HM and the other groups (Fig. 3A). Fig. 3Principal components analysis (PCA) graph. The percentages on the axes represent the variation in the data set of the first two components. The points in the shaded regions represent the 95% confidence interval of each group and any overlaps between groups. The PCAs were generated using MetaboAnalayst 5.0 comparing between sample groups (control (CO), cmybpc3+/-::myl7:eGFP heterozygous (HT), and cmybpc3-/-::myl7:eGFP homozygous (HM). (A) Fatty Acids, (B) Acycarnitines, (C) Lipids and (D) Metabolites. Principal components analysis (PCA) graph. The percentages on the axes represent the variation in the data set of the first two components. The points in the shaded regions represent the 95% confidence interval of each group and any overlaps between groups. The PCAs were generated using MetaboAnalayst 5.0 comparing between sample groups (control (CO), cmybpc3+/-::myl7:eGFP heterozygous (HT), and cmybpc3-/-::myl7:eGFP homozygous (HM). (A) Fatty Acids, (B) Acycarnitines, (C) Lipids and (D) Metabolites. The PLS-DA differentiated metabolic phenotypes with the first component explaining 67.5% of the variability (Supplementary Figure S1). The results highlight 9,12,15-octadecatrienoic acid as the variable showing the highest VIP score of 1.95 for component 1, followed by hexadecanoic acid (Supplementary Figure S1). Particularly, 13-Octadecanoic acid (C18:1), 9,12-Octadecadienoic acid, 9,12,15-octadecatrienoic acid, Hexadecanoic acid, 14-methyl-, and 5,8,11,14,17-Eicosapentaenoic acid, were significant between the control group (CO) and cmybpc3-/- homozygous group (HM) (Supplementary Figure S1). We analyzed the total lipids to identify lipidomic profiling to relate key lipid biomarkers in the cmybpc3-HCM model. Lipid profiling determined a total of 221 lipid annotations where fragmentation spectra were matched against LipidSearch software (Supplementary Table S1). The annotated lipid compounds covered several lipid classes: acylcarnitines, carnitine esters, ceramides, diacylglycerols, hexosylceramides, lysophosphotidylcholines, lysophophatidylethanolamines, phosphatidylcholines, phosphatidylethanolamines, phophatidylglycerol phosphatidylinositol, phosphatidylmethanol, sphingomyelin, triacylglycerol and coezyme-Q10. A total of two lipids were significantly dysregulated between the CO and HT; and 48 lipids between CO and HM groups using One way ANOVA analysis with Tukey as post-hoc test (Supplementary Table S2 and S3) with false discovery rate (FDR) < 0.05. The first two principal components explained 54.3% of the variability in cmybpc3-HCM models (Fig. 3B). Sparse Partial Least Squares-Discriminant Analysis (sPLS-DA) using five components with 10 variables per component established a clear separation between the three groups based on lipidomic profiles (Supplementary Figure S2). The results demonstrate that acylcarnitines and phosphatidylcholines contribute the most to the differentiation between the three examined groups (Supplementary Figure S2). The results of the hierarchical clustering indicated that within the homozygous group, there is observed variability in the concentrations of lipids that form a distinct signature in the mutant HCM model (Supplementary Figure S2). To discriminate between the three groups, we compared the groups using acylcarnitines profiles. The first two components, acylcarnitines (18:3 and 20:4), explained 88.2% of this variability between CO, HT and HM samples. The HM samples clustered separately from the CO, HT groups (Fig. 3C). The hierarchical cluster analysis demonstrated that the HM group clusters separately with an elevation in acylcarnitine levels (Supplementary Figure S3). The PLS-DA model revealed significant discrimination in the metabolic phenotypes of acylcarnitines. The acylcarnitines with the highest VIP scores, contributing to the differentiation of the HM group from the other HT and CO groups, were highlighted in Supplementary Figure S3B and S3C. To understand the underlying metabolic perturbation of cmybc3-/- related pathological HCM, untargeted metabolomics was employed to characterize the differences between the examined groups. A total of 110 and 53 metabolites were annotated under positive and negative ESI modes, respectively with an MS2 fragmentation pattern match in an in-house library or mzCloud software (Supplementary Table S4). A total of 13 metabolites were significantly dysregulated between HM and CO (Supplementary Table S5), while none were significant between HT and CO, with FDR < 0.05. PCA did not show clear clustering between the groups with the first two components explaining 48.4% of the variability (Fig. 3D). Further analysis using sPLS-DA with 10 variables per component on the top 5 components showed a more distinct separation of the HM samples and a unique metabolic signature related to both HM and HT groups (Supplementary Figure S4). To explore the significant metabolites distinguishing HCM phenotype, we compared between two groups: HM and CO groups, HM and HT groups, HT and CO groups. Metabolites were considered significant if they had a log2 fold change > 1 and FDR < 0.05. Comparing HM and CO groups, among the 163 identified metabolites, 16 were significant, with 6 showing downregulation and 10 upregulation (Supplementary Figure S4). The OPLS-DA revealed a clear and distinct clustering of the HM and CO groups with RY (model interpretation rate) of 87.7% and Q (model predictive ability) of 74.5% (Supplementary Figure S4). Metabolites with VIP > 1 were selected as important metabolites with asparagine showing the highest VIP score, followed by stachydrine (Supplementary Figure S4). Furthermore, Random Forest model demonstrates robust discriminatory power with AUROC curve (Area Under the Receiver Operating Characteristic curve) of 0.995 (CI 1–1) using 15 features (Fig. 4A) and predictive accuracy of 93%. The metabolites were mapped to pathways to determine pathways enriched in the HM group compared to CO group. The enrichment analysis shows that 8 pathways were significantly dysregulated with FDR < 0.05 (Fig. 5A and Supplementary Table S6). Fig. 4Receiver Operating Characteristics (ROC) curve for significant metabolites. ROC was generated by the Random Forest model using MetaboAnalyst 6.0, employing different number of features (5, 10, 15, 25, 50 and 100). The corresponding Area Under the Curve (AUC) value and confidence interval (CI) were determined for each set of features. (A) ROC curves of the cmybpc3-/-::myl7:eGFP homozygous (HM) vs. control (CO) model. (B) ROC curves of the cmybpc3+/-::myl7:eGFP heterozygous (HT) vs. HM model. Receiver Operating Characteristics (ROC) curve for significant metabolites. ROC was generated by the Random Forest model using MetaboAnalyst 6.0, employing different number of features (5, 10, 15, 25, 50 and 100). The corresponding Area Under the Curve (AUC) value and confidence interval (CI) were determined for each set of features. (A) ROC curves of the cmybpc3-/-::myl7:eGFP homozygous (HM) vs. control (CO) model. (B) ROC curves of the cmybpc3+/-::myl7:eGFP heterozygous (HT) vs. HM model. Fig. 5Pathway enrichment analysis. Pathway enrichment analysis was done using MetaboAnalyst 5.0 and 6.0 using KEGG for species zebrafish (Danio rerio). The Y-axis shows the matched pathways according to the p-values from the pathway enrichment analysis and X-axis pathway impact values from the pathway topology analysis. The node color of each pathway is determined by the p-value (red = lowest p-value and highest statistical significance), and the node radius (size) is based on the pathway impact factor, with the biggest indicating the highest impact. Pathway enrichment analysis was done comparing (A) Homozygous and Control groups. (B) Homozygous and Heterozygous groups. Nicotinate and nicotinamide metabolism was significant in both Homozygous compared to Control and Heterozygous groups with p-values of 6.28E-6 and 2.65E-4, respectively. Pathway enrichment analysis. Pathway enrichment analysis was done using MetaboAnalyst 5.0 and 6.0 using KEGG for species zebrafish (Danio rerio). The Y-axis shows the matched pathways according to the p-values from the pathway enrichment analysis and X-axis pathway impact values from the pathway topology analysis. The node color of each pathway is determined by the p-value (red = lowest p-value and highest statistical significance), and the node radius (size) is based on the pathway impact factor, with the biggest indicating the highest impact. Pathway enrichment analysis was done comparing (A) Homozygous and Control groups. (B) Homozygous and Heterozygous groups. Nicotinate and nicotinamide metabolism was significant in both Homozygous compared to Control and Heterozygous groups with p-values of 6.28E-6 and 2.65E-4, respectively. The pathways involved the metabolism of nicotinate, nicotinamide, purine, glyoxylate, dicarboxylate, glycerophospholipid, pyrimidine, pantothenate, CoA, sphingolipid, alanine, aspartate, and glutamate metabolism. Comparing HM to HT group, 26 metabolites displayed significantly different levels, 24 upregulated and 2 downregulated (Supplementary Figure S4). The Random Forest model effectively differentiates between these groups with AUROC of 1 (1–1) using 10 features (Fig. 4B) and a predictive accuracy of 98%. Pathway enrichment analysis revealed significant pathways involving cysteine, methionine, nicotinate, nicotinamide, phenylalanine, tyrosine, and tryptophan metabolism (Fig. 5B and Supplementary Table S7). Comparison between HT and CO groups revealed that only 3 metabolites showed significantly different levels of nicotinic acid, ornithine, and prolinamide. Between these two groups, no pathways were highlighted with an FDR < 0.05. There is limited research on HCM metabolic alterations in cardiac remodeling. Herein, we performed a comprehensive multi-omics assessment at the early stages of cardiac remodeling in HCM utilizing the zebrafish model. Our integrative omics analysis reveals unique metabolic and lipidomic alterations implying the impact on critical molecular and biochemical pathways contributing to HCM pathophysiology. We collected 150 larvae in 7 replicates of cmybpc3 HCM zebrafish model at 5 days post-fertilization (dpf), as the heart is fully developed by 3dpf. In the zebrafish, at 1dpf, the heart is a valveless tube, at 1.5dpf cardiac looping and chamber ballooning occur, followed by the formation of the two-chamber heart at 2dpf, and finally at 3dpf ventricular trabeculation occurs. This study revealed several significant differences in fatty acid, lipid and metabolite profiles detected in zebrafish homogenate samples between the examined CO, HT, and HM groups. The main alterations were seen with acylcarnitines, redox regulation via NAD/NADH, pyrimidine, and purine metabolism. These differences further prove the pathogenicity of MYBPC3 genetic variation. The rate of acylcarnitine metabolism plays an essential role in maintaining a balance between lipid metabolism and levels of intracellular glucose. Acylcarnitines function as activated long-chain fatty acid carriers, transporting fatty acids into the mitochondria for beta-oxidation, producing acetyl CoA, and subsequently providing cellular energy through the citric acid cycle. As well, acylcarnitines have been associated with the production of ketone bodies and the peroxidation of fatty acids. Furthermore, acylcarnitines have been linked as potential biomarkers for HCM severity and other cardiovascular diseases. Fatty acids and associated lipids play an important role in cardiomyocyte structure and function. Changes occur in fatty acids regulation during cardiac growth and development. In HCM patients, reports of impaired myocardial fatty acid oxidation were identified through various multi-omics analyses. This was linked to a decrease in acylcarnitine levels that affects the supply of efficient cellular energy in the heart. Studies in patients with HCM have identified an inverse relationship between myocardial and circulating acylcarnitines levels. In our zebrafish HCM model, a higher abundance of acylcarnitines was noted in HM group compared to HT and CO, suggesting higher use of beta-oxidation of fatty acids as an energy source. Moreover, plasma acylcarnitines were associated as one of the top metabolites that can draw the distinction between mild and severe forms of HCM in age and sex-matched patients, highlighting the relevance of monitoring their levels in efficient HCM diagnosis. Cardiac hypertrophy is induced by pressure overload with changes in fatty acids that were seen in our HM HCM model. The 9,12,15-octadecatrienoic acid, also known as alpha-linolenic acid (ALA), was highlighted as the highest VIP score distinguished among the three examined zebrafish groups with a higher level in HM and a lower level in HT compared to the control group. ALA is a polyunsaturated fatty acid that was found to play a crucial role as a structural component in cell membranes, affecting permeability, flexibility and fluidity along with other omega-3 fatty acid. Also, these fatty acids have been shown to have highly protective effects in cardiovascular disease-associated risks by suppressing pro-inflammatory cytokines. In regard to saturated fatty acids, our analysis demonstrated that hexadecenoic (palmitic acid) was highlighted as the second highest VIP score fatty acids distinguishing the examined profiles. Long fatty acid chains, such as palmitic acid and S-59-adenosyl-methionine through the catechol-O-methyltransferase biosynthesize palmitic acid methyl ester, have been shown earlier to activate voltage-dependent Ca channels that are associated with HCM. The fatty acid levels showed opposite trends in our zebrafish model, with higher levels in HM group and lower levels in HT group in comparison to the control group. This might be due to the fact that HT group resembles the human autosomal dominant HCM disease, while truncating variants were reported to be lethal. Cardiomyocytes generate two-thirds of ATP by the oxidation of fatty acids and one-third by glucose. In end-stage heart failure patients, a decrease in fatty acid oxidation and a shift towards increased glucose metabolism leading to fatty acids accumulation was reported, where as in our zebrafish HM group, a similar signature was observed. In contrast, our results suggest that the HT group fatty acid levels appear to be compensating for energy demands through the oxidation of fatty acids. Importantly, oxidative stress is implicated in various cardiac diseases, including HCM and heart failure. It plays a key role in HCM remodeling and dysfunction. The initiation of HCM occurs via the Ca-dependent pathway, activated by mitochondrial reactive oxygen species (ROS), triggering the cardiac Na+/Ca exchanger. Additionally, serum biomarkers indicative of oxidative stress were observed in patients with HCM. Similar to previous findings, metabolic pathways associated with oxidative stress were significantly dysregulated in our zebrafish cmybpc3-/- HM compared to cmybpc3+/- HT group. Methionine, S-adenosylmethionine and cysteine were significantly dysregulated in the HM, suggesting an imbalance between the production of ROS and the ability to detoxify them within the model. Notably, nicotinate, nicotinamide and tryptophan metabolites were upregulated in cmybpc3-/- HM compared to CO and HT groups. NAD + is involved in many metabolic processes, including energy metabolism and stress response, in which significant pathophysiological alterations have been observed in HCM patients. NAD + can be synthesized either through the de novo pathway using tryptophan or through the salvage pathway using nicotinamide riboside, nicotinamide mononucleotide, nicotinic acid, nicotinamide, or nicotinamide ribose. Our observation is consistent with earlier studies that reported a depletion of cellular NAD + levels in HCM models and patients, suggesting an increased effort to synthesize more NAD + in response to severe oxidative stress. Furthermore, we identified disruptions in pathways involving purines, pyrimidines, and glycerophospholipids, comparing the HM to CO zebrafish groups. Consistent with previous reports of HCM models and patients’ data, a link between purine and pyrimidine metabolite alterations was demonstrated in our zebrafish model. In addition, according to the New York Heart Association categorization, a significant relationship was identified between purine and pyrimidine metabolites indicative of a worse prognosis in HCM patients. As well, HCM patients’ lipidomics analysis showed altered levels of glycerophospholipids and triglycerides. The present study, while providing valuable insights into cardiac remodeling in HCM model, is not without its limitations. While the zebrafish model is advantageous for its genetic tractability and experimental manipulability, caution must be exercised when extrapolating findings to the complexities of the human cardiovascular system. One notable limitation pertains to the intrinsic anatomical disparity between the zebrafish heart, with its two-chambered structure, and the human heart, which is characterized by a more complex four-chambered configuration. Additionally, using laboratory animal models, while beneficial for controlled experimental conditions, introduces some limitations related to these optimal conditions. The multifactorial nature of cardiovascular health, influenced by lifestyle, dietary habits, and environmental factors, is inherently challenging to fully replicate in a controlled laboratory setting. Furthermore, we utilized whole larvae lysate samples as the process of dissecting the zebrafish hearts while preserving sample integrity poses a technical challenge that could impact the reliability of results. To address this, we opted to analyze the entire zebrafish lysate, believing this approach may unveil unique and significant differences in metabolomics and lipidomics profiles, providing a comprehensive perspective on cardiac physiology in this model organism. Overall, while this study provides valuable insights into the metabolic perturbations underlying HCM, addressing these limitations for future research directions will not only enhance our understanding of HCM but also pave the way for developing more effective therapeutic interventions. This potential impact should instill hope in the future of HCM research and treatment. In conclusion, our study provides a comprehensive multi-omics assessment of hypertrophic cardiomyopathy (HCM) using the cmybpc3-/- zebrafish model, highlighting significant alterations in lipidomic and metabolomic profiles associated with the disease. The findings underscore the importance of acylcarnitines, phosphatidylcholines, and various metabolic pathways, including nicotinate, nicotinamide, purine, pyrimidine, and amino acid metabolism, in HCM pathophysiology. These findings are significant and should be a cornerstone for future research in this field. Future research should investigate the mechanistic roles of these identified metabolites and lipids in HCM progression, validate our findings in mammalian models, and conduct longitudinal studies to monitor changes over time. Additionally, exploring the dysregulated pathways as potential therapeutic targets and developing drug models to test in zebrafish and other systems could offer new treatment strategies. Zebrafish (Danio rerio) Tg myl7:eGFP line as a control (CO), cmybpc3+/-::myl7:eGFP heterozygous (HT) and cmybpc3-/-::myl7:eGFP homozygous (HM) adults were maintained in standard conditions according to the Ministry of Public Health (MOPH), Qatar animal research guidelines and under an approved protocol by the Institutional Animal Care Committee (Qatar Foundation, protocol IACUC 2020 − 1132) and adhered to the ARRIVE guidelines. The adult zebrafish were setup for breeding, the embryos were collected and incubated in E3 egg water media (5.0mM NaCl, 0.17mM KCl, 0.16mM MgSO4-7H2O, 0.4mM CaCl2-2H2O in ddH2O) at 28.5 °C. For each experimental group, CO, HT, and HM at 5 days old, 150 larvae were pooled in one tube with a total of 7 replicates for each group to obtain more information on fatty acids, metabolomic and lipidomic profiles. All deformed larvae were excluded prior to collection; no criteria were set for exclusion during the experiment. Excess media in the pooled larvae tubes was removed, then the tubes were snap frozen in liquid nitrogen then stored in -80 °C freezer until further processing. All samples were processed as previously described with some modifications; the samples were homogenized in chilled 400 µL methanol (Fisher A456-500) using a Bel-Art pestle tube homogenizer (SCIENCEWARE F19923-0000)., and kept on ice for 20 min; then vortexed at 2000 rpm, 4 °C for three cycles of 10 s with a 15 s break between each cycle using an Eppendorf thermomixer. Then, it was centrifuged at 14,000 g for 5 min at 4 °C, and 400 µL of the supernatant was collected in glass vials (Waters 186007201 C). The supernatant was further processed and analyzed by three methods for fatty acid profiling, metabolomics and lipidomics. The different groups were de-identified throughout analysis. Fatty acid analysis was conducted on 50 µL of the zebrafish homogenate supernatant after converting total lipids to fatty acid methyl esters through transesterification. The samples were transferred to 10 mL Reacti-vial vials (Thermo Scientific TS13225) and 2 mL of 25mL/L H2SO4 (Sigma Aldrich 339741) in methanol was added. The mixture was mixed and then placed in an oven for 2 h at 60 °C. Once the tubes were cooled at room temperature, 2 mL of saturated NaCl (Sigma Aldrich 31434) in H2O was added and mixed, followed by 1 mL of hexane (Scharlau HE02342500). Following centrifugation, the hexane layer was transferred to a glass vial and any trace amounts of water were removed by mixing a small amount of Na2SO4 (Sigma Aldrich 798592) to the solution. The sample was then transferred to an autosampler vial and injected into a gas chromatograph coupled to a single quadrupole mass spectrometer in Scan mode (Agilent Technologies, Santa Clara, CA). GC-MS analyses were carried out on an Agilent 7890 C gas chromatograph with a 5977 A EI-MSD (Agilent Technologies, Wilmington, DE). Chromatography was carried out using a 60 m x 0.25 mm TR-FAME capillary column with a film thickness of 0.25 μm (Thermo Scientific, Waltham, MA). Helium was used as the carrier gas at a flow rate of 1.0 mL/min with constant flow compensation. The GC inlet were held at 290 °C, and the MS transfer line was maintained at 260 °C. Sample injections of 0.5 µL were performed with a split ratio of 5:1. After a solvent delay of 8.7 min, the oven temperature was programmed from 30 to 200 °C at a rate of 5 °C/min and then to 220 °C at 1 °C/min and to 260 °C at 6.5 °C/min with a final hold of 2 min. Full-scan acquisition with electron impact ionization carried out using 70 eV was used for quantification based on the extracted ion chromatogram. All mass spectra were acquired over the m/z range of 30–500. Spectra were matched with a FAME standard mix (37 Component FAME Mix, Supelco, Whippany, NJ) or putatively annotated by comparison with NIST EI library (2017 release). An aliquot of the zebrafish homogenate supernatant (125 µL) was transferred to glass vials and 20 µL of isotopically labelled QC standard mix (MSK-QC-KIT, Metabolomics QC Kit, Cambridge Isotope Laboratories, Tewsbury, MA) was added. The mixture was evaporated in a vacuum centrifuge (Genevac EZ-2 Plus, Genevac Ltd, Ipswich, UK) to dryness. The residue was then stored in − 80 °C freezer until analysis. After reconstitution in 90 µL of 1% methanol in water (Fisher W6500), 20 µL of each sample was mixed to prepare a sample mix used for compound identification. Each sample was injected individually (2 µL injection volume) on an ultra-high pressure liquid chromatograph coupled to an orbitrap tribrid mass spectrometer (Vanquish LC and Fusion Lumos, Thermo Scientific, Waltham, MA). The column utilized was an Accucore C18 + reverse phase column (27101–152130, 2.1 × 150 mm, 1.5 μm, Thermo Scientific, Waltham, MA) held at 45 °C and using a mobile phase gradient of H2O and methanol with 0.1% formic acid (Fisher A1171-AMP) additive (MPA and MPB respectively) at a flow rate of 0.2 mL/min. Initial mobile phase conditions were 1% MPB rising to 100% MPB over 12 min. The heated electrospray ion source (NGS-HESI) was operated with a sheath gas flow of 40 AU, auxiliary gas flow rate of 8 AU, sweep gas flow of 1 AU and spray voltage of 3.5 kV or 3 KV in positive and negative mode respectively. The vaporizer and ion transfer tube temperature settings were 320 °C and 275 °C respectively. Automatic gain control target was set to 1E5 ions and with an auto setting of maximum injection time and RF lens setting of 35%. Each sample was injected four times with varied method parameters: (1) positive ESI polarity with low mass range (59–250 m/z), (2) positive ESI polarity with high mass range (250–1200 m/z), (3) negative ESI polarity with low mass range, and (4) negative ESI polarity with high mass range, all in full scan mode with orbitrap mass resolution of 240 K (FWHM at m/z 200) with internal calibration of all scans with fluoranthene ion. The sample mix was injected numerous times with AcquireX deep scan mode to enable maximal MS2 fragmentation with HCD fragmentation mode of detected peaks above an intensity threshold of 2E4. The data-dependent MS/MS mode utilized an orbitrap mass resolution of 60 K in MS1 and 30 K in MS2 and stepped collision energies of 20, 35 and 50%. Dynamic exclusion was utilized with isotope exclusion and feature exclusion for 4 s after one MS2 scan with a mass tolerance setting of 3ppm. Another aliquot of the zebrafish homogenate supernatant (125 µL) was transferred to glass vials and 20 µL of isotopically labelled lipid QC mix (330707, Splash Lipidomix, Avanti Polar Lipids, Alabaster, AL) and 805 µL of H2O: CHCl3:methanol (24.8:31.1:44.1) was added and the mixture mixed. Another 250 µL of CHCl3 was then added and the mixture was mixed, followed by an aliquot of 250 µL of H2O and mixing. The mixture was centrifuged at 4000 rpm at 15 °C for 15 min to produce a bilayer. The bottom layer (400 µL) was transferred to a glass vial and evaporated in a vacuum centrifuge and the residue stored at − 80 °C until analysis. On the day of analysis, the sample was reconstituted in 100 µL of isopropanol: methanol: CHCl3 (1:1:1, isopropanol Fisher lA461-500, CHCl3 Sigma Aldrich 650498) and 20 µL of each sample mixed to make a sample mix. Each sample (1 µL injection volume) was injected as above with a column temperature setting of 70 °C using mobile phases acetonitrile: water (6:4, MPA) and acetonitrile: isopropanol (1:9, MPB), each with 0.1% formic acid and 10 mM ammonium acetate as additives. The initial mobile phase condition was 10% MPB and rose to 100% over 32 min. The HESI source was set to sheath gas flow 40 AU, auxiliary gas flow 7 AU, sweep gas flow 1 AU and spray voltages 3.5 kV and 2.8 kV in positive and negative modes, respectively, with vaporizer temperature of 300 °C and ion transfer tube temperature 350 °C. Automatic gain control target was set to 2E5 ions, with an auto setting of maximum injection time and RF lens setting of 60%. Each sample was injected twice in ESI positive and negative scan mode at 240 K orbitrap mass resolution with a full scan mass range of 200–2000 m/z. The sample mix was injected several times with AcquireX deep scan mode to enable MS2 fragmentation in HCD fragmentation mode utilizing an orbitrap mass resolution of 120 K in MS1 and 15 K in MS2 with stepped collision energies of 25, 30 and 35% and a method that triggered CID fragmentation at 32% and 35% collision energies after the detection of the phosphocholine headgroup (m/z 184.0733) and the fragment ions formed by the neutral loss of M + NH4 adducts of common fatty acids respectively. The processing of data files was conducted using Compound Discoverer (version 3.3), where features were aligned and grouped with a mass tolerance of 3 and 5 ppm for metabolomics and lipidomics samples, respectively, with a retention time maximum shift of 0.2 min. Lipidomics samples were annotated with LipidSearch (version 4.23), allowing for [M + H], [M + NH4], [M + Na], [M + H-H2O], [M-H], [M + HCOO] and [M + CH3COO] ions with a mass tolerance of 5 ppm and fragment mass tolerance of 10 ppm. Features in the metabolomics data files were annotated with the following online databases: (1) in-house MS2 library of over 950 metabolite compounds where match factors were elemental composition, retention time and MS2 spectrum and (2) mzCloud where match factors were elemental composition and MS2 spectrum, both with a parent mass tolerance of 5 ppm and fragment mass tolerance of 10 ppm. All discovered compounds were integrated for peak area in each sample and used for statistical analysis. The added isotopically-labeled internal standards were monitored in each sample for retention time fluctuations, peak abundance and mass accuracy. The monitored peaks were within acceptable limits of ± 2% of median retention time, ± 30% of median peak abundance and ± 3 ppm mass accuracy. MetaboAnalyst v5.0 and v6.0Superscript> were used for multivariate statistics. Log10 data transformation and Pareto scaling were applied to all data. Initially, the variance in the data was analysed using Principal Components Analysis (PCA). Following this, a Partial Least Squares - Discriminant Analysis (PLS-DA) was applied and variable importance was assessed using VIP (variable importance in projection) scores. Sparse PLS-DA was used for metabolomics and lipidomics data to effectively reduce the number of variables. In this analysis, five components were used with 10 variables per component. Hierarchical clustering heatmaps were generated using Euclidean distance measure and Ward clustering method. Pathway enrichment analysis was performed with global Ancova with topology analysis set to relative-betweeness centrality. ROC curve analysis was performed using Random Forests with Monte-Carlo cross-validation where within each iteration, the data was split randomly into two-thirds training and one-third validation sets. This multivariate study on lab-grown zebrafish in a highly controlled environment where comparisons were made with an experimental control group has minimal confounders.
|
PMC9301695
|
Are n-3 PUFAs from Microalgae Incorporated into Membrane and Storage Lipids in Pig Muscle Tissues?—A Lipidomic Approach
|
For the study of molecular mechanisms of to lipid transport and storage in relation to dietary effects, lipidomics has been rarely used in farm animal research. A feeding study with pigs (German Landrace sows) and supplementation of microalgae (Schizochytrium sp.) was conducted. The animals were allocated to the control group (n = 15) and the microalgae group (n = 16). Shotgun lipidomics was applied. This study enabled us to identify and quantify 336 lipid species from 15 different lipid classes in pig skeletal muscle tissues. The distribution of the lipid classes was significantly altered by microalgae supplementation, and ether lipids of phosphatidylcholine (PC), phosphatidylethanolamine (PE), and phosphatidic acid (PA) were significantly decreased. The total concentration of triacylglycerides (TAGs) was not affected. TAGs with high degree of unsaturation (TAG 56:7, TAG 56:6, TAG 54:6) were increased in the microalgae group, and major abundant species like TAG 52:2 and TAG 52:1 were not affected by the diet. Our results confirmed that dietary DHA and EPA are incorporated into storage and membrane lipids of pig muscles, which further led to systemic changes in the lipidome composition.Dietary strategies to modify the fatty acid profile of pig muscles by enhancing polyunsaturated fatty acid (PUFA) contents, predominantly n-3 PUFA, were very successful. In contrast to the ruminants, it was shown in the pigs that dietary PUFAs could be incorporated into muscles and other tissues with only minor biochemical modifications of PUFAs. Intervention studies on PUFA-supplemented pig diets containing linseed, rapeseed, sea buckthorn, pomace cakes/oils/meals resulted in an increase in n-3 PUFA content and a decrease in the n-6/n-3 fatty acid ratio in intramuscular fat. In contrast, sunflower seed/oil/meal-supplemented diets revealed an increase in n-6 PUFA content in pig muscle tissues. The supplementation with microalgae in the pig diet is a potential alternative to improve the lipid/fatty acid profile of pig muscle with respect to human nutrition compared to vegetable oils/press cakes/meals. Besides essential amino acids, vitamins, polysaccharides, microalgae contain long-chain n-3 PUFAs, primary docosahexaenoic acid (C22:6n-3, DHA) or eicosapentaenoic acid (C20:5n-3, EPA). A large amount of evidence suggests that EPA and DHA have stronger beneficial effects on human health compared to C18:3n-3. Furthermore, the conversion of C18:3n-3 to EPA and DHA is limited in humans (<10%); therefore, direct consumption of foods rich in EPA and DHA is required to reach the recommended daily n-3 PUFA intake. First lipidomic approaches were performed in pigs and beef cattle primarily using electrospray ionization-tandem mass spectrometry (UHPLC-ESI-MS/MS), shotgun lipidomics, or MALDI-TOF MS. Only a small number of lipidomic studies on pig muscles are available. Recently, lipidomics was applied to identify different muscles of pigs and to differentiate/authenticate raw pork species. The comparative lipidomic analysis of selected local pork in China led to the definition of a lipid marker panel that could classify different pork cuts and geographical origins. Another study with a targeted lipidomic approach using HPLC-ESI-MS/MS was conducted to investigate pig muscle phospholipids (PLs) and variations of phospholipid hydrolysis products at different aging periods. As the main phospholipid classes in pig muscles, phosphatidylcholine (PC), ether-linked PC (PC-O), phosphatidylethanolamine (PE), and ether-linked PE (PE-O) were identified as comprising up to 70% of the overall PL content. Very recently, Meyer et al. investigated the replacement of soybean extraction meal with insect meal in the diet of growing pigs using transcriptomics, metabolomics, and lipidomics. A 4 week insect meal-based diet (Tenebrio molitor L.) in growing pigs revealed only weak changes in the lipid metabolism in the plasma and liver. The concentrations of main PL classes, such as PC, PE, phosphatidylinositol (PI), lysophospholipids, and sphingolipids, were not affected by the insect-based diet; however, no further lipid analysis of pig muscle tissue was performed. The first study investigating the changes of lipid profiles in skeletal muscle of Landrace pigs fed with n-3 or n-6 PUFA-rich diets indicated large differences between the diet groups. The results showed that dietary and de novo synthesized n-3 PUFAs were predominantly incorporated into muscle PL species PE and cardiolipins (CL); however, the distribution pattern of different PL classes in pig muscle was unchanged. In addition, alkenyl-acyl and alkyl-acyl phospholipids (ether-linked PLs) were elevated in muscle of pigs fed with n-3 PUFA-based diets. The occurrence, molecular, and physiological impact of ether-linked PLs, primarily plasmalogens, in pig tissues is still enigmatic. In general, the application of microalgae as a diet supplement for pigs opens up the opportunity to improve growth and meat quality in pigs and also in ruminants; however, the results are affected by the supplemented microalgae species. Recently, our group investigated the effects of long-term microalgae supplementation (Schizochytrium sp.) on muscle microstructure, meat quality, and fatty acid composition in growing German Landrace pigs. The samples of pig muscles, highly accumulated in docosahexaenoic acid (DHA) and eicosapentaenoic acid (EPA), were used in this nontargeted shotgun lipidomic study. The objective of the present study was to investigate which lipid species predominantly influence the dietary changes with increased levels of long-chain n-3 PUFAs in pig skeletal muscle (longissimus thoracis) and their incorporation in membrane and/or storage lipids by dietary microalgae supplementation. The study was conducted at the experimental pig facilities of the Research Institute for Farm Animal Biology (FBN) in Dummerstorf, Germany. All procedures, including the use and treatment of animals, were performed in accordance with the German animal protection law and approved by the relevant authorities (Landesamt für Landwirtschaft, Lebensmittelsicherheit und Fischerei Mecklenburg-Vorpommern, Germany; 7221.3-2-051/15). The experimental details of the dietary study of microalgae supplementation on meat quality and muscle microstructure in growing pigs were recently described. Briefly, the dietary pig study (German Landrace sows) was performed with the supplementation of microalgae (Schizochytrium sp.). The piglet diet was fed until day 95 (∼13.7 MJ of ME/kg) and followed the fattening diet (∼12.9 MJ of ME/kg). The microalgae (Schizochytrium sp.) diet was supplemented with 7% (piglet diet) and 5% (fattening diet) DHA Gold (DSM, Bramsche, Germany), whereas the control diet was adjusted to a total fat of 5.4% (piglet diet) and 3.2% (fattening diet) using soybean oil plus lard as lipid sources. The major difference between diets consisted in the proportion of DHA (37% in the piglet diet and 33% in the fattening diet of the microalgae group and 0.05% in the piglet diet and 3.1% in the fattening diet of the control group). The female piglets were allocated to the control group (n = 15) or the microalgae group (n = 16) at day 28 of age. After one week of adaptation, the microalgae-supplemented diet was initiated on day 35 of age and was fed until pigs were slaughtered at days 145/146 of age. After slaughtering, muscle samples were immediately collected from the right side of the carcass. Longissimus thoracis (LT) muscle samples of pigs obtained from the 12/14th rib were used for lipidomic investigation. Immediately after sampling, the muscle samples were cut into small pieces, deep-frozen in liquid nitrogen, and homogenized under liquid nitrogen using a stainless steel grinding mill (mill M20, IKA, Staufen, Germany). After homogenization, the muscle samples were stored at −80 °C until total lipid extraction. The total lipids of muscle samples (2 g muscle powder) were extracted using an Ultra Turrax (T25, IKA, Staufen, Germany), 3 × 15 s, 15,777g using chloroform/methanol (ratio 2:1) at room temperature. All solvents contained 0.005% (w/v) of t-butylhydroxytoluene (BHT) to prevent PUFA oxidation. The details of the lipid extraction procedure were already described. The final extraction mixtures were stored at 5 °C for 18 h in the dark and subsequently washed with a 0.02% CaCl2 solution. The organic phase was separated and dried with a mixture of Na2SO4 and K2CO3 (10:1, w/w), and the solvent was subsequently removed using a ScanSpeed 40 (LaboGene, Allerød, Denmark) vacuum centrifuge at 438g and 30 °C for 30 min. In total, 31 pig muscle lipid extracts (control group, n = 15; and microalgae group, n = 16) were stored at −20 °C until lipidomic analysis. The muscle lipid extracts were dissolved in chloroform/methanol/water (60:30:4.5, v/v/v) with BHT at 0.05% (w/v). The individual samples were in a second dilution step normalized to 8.38 mg/mL of total fat. Afterward, all samples were mixed with an internal standard solution (Supporting Table S1) and the ESI spray solution as reported earlier with a final dilution factor of 1100 from the stock solution. Shotgun lipidomics was performed by a Q Exactive Plus (Thermo, Bremen, Germany) using TriVersa NanoMate (Advion, Ithaca) as a nanoelectrospray source as reported earlier. For lipid identification, LipidXplorer 1.2.7 was utilized, and quantification was achieved using responses of the respective lipid class-specific standard as reported earlier. Lipidomic data processing details are available on the Lipidomics Informatics for Life Science (LIFS) web portal (https://lifs-tools.org/). Cholesterol concentrations were determined based on an approach reported earlier. The reported lipidomes will be made available under the preliminary LipidCompass accession number LCE9. All used solvents and chemicals were obtained in the highest purity grade (ROTISOLV, HPLC grade) from Carl Roth GmbH (Karlsruhe, Germany). For lipidomic experiments, all used solvents and additives were of LC–MS quality and obtained from Sigma-Aldrich (Deisenhofen, Germany). Lipid standards were purchased from Avanti Polar Lipids (Alabaster). A total of 336 lipid species quantified using shotgun lipidomics were grouped in samples of the control group (n = 15) and microalgae group (n = 16). Multiple t-tests between control and microalgae groups were performed using log-transformed data. To adjust for multiple comparisons, we calculated q-values and limited the false discovery rate to 0.01 using the R package q value (v2.22.0). In addition, the log2-fold change (log2FC) between microalgae and control lipid species concentrations (nmol/mg total lipids) was calculated to quantify the variations. We applied partial least-squares discriminant analysis (PLS-DA) to identify the key variables of the 15 lipid classes (including two subclasses) and the sparse variant (sPLS-DA) for the 336 lipid species that drive the discrimination of the two investigated groups, i.e., control vs microalgae-supplemented group. Both methods are implemented in the mixOmics package (v6.14.1). We used scaled data to analyze the lipid classes, while the individual lipid species were not scaled as we aimed to keep the information of the lipid concentration and distinguish between the relevance of minor and major lipids. Sample plots are presented to visualize the discriminatory ability of the lipid classes and individual lipid species in the space spanned by the first two latent variables. Loading plots show the importance of the 15 lipid classes and 15 lipid species, which had the strongest impact on group separation in PLS-DA or sPLS-DA, respectively. We used repeated 5-fold cross-validation to evaluate the performance of the fitted PLS-DA models. The models have a very good performance in discriminating the two treatment groups with a stabilized balanced error rate of 0.015 (lipid classes) and an error rate close to zero (lipid species) after two components. Statistical data analyses and data visualization were performed using R 4.0.3 (R Core Team, 2020). In this study, 336 lipids of 15 different classes were identified and quantified in total lipid extracts of skeletal muscle tissues of pigs fed either with a control died or supplemented with microalgae (Supporting Table S2). The highest number of species in pig muscle extracts were identified for TAGs (57 lipid species), alkyl/alkenyl-phosphatidylethanolamines (PE-O, 52 lipid species), PEs (47 lipid species), alkyl/alkenyl-phosphatidylcholines (PC-O, 43 lipid species), and PCs (30 lipid species). The concentration range for a single lipid covered more than 5 orders of magnitude with TAG 52:2 as the major abundant component (210 nmol/mg total lipids) and LPE 22:6 (0.002 nmol/mg total lipids, Supporting Table S2). TAG was found as the most abundant lipid class with 65–68% of total lipids in a concentration of 833.3 ± 75.3 nmol/mg total lipids (control group) and 871.9 ± 49.8 nmol/mg total lipids (microalgae group). The overall TAG content was not affected by microalgae supplementation. In the case of membrane lipids, our lipidomic analysis demonstrated that PC and PE (including the ether-linked PLs—PC-O and PE-O) were the most abundant PL classes and represent approximately 80% of total PL content independent of the dietary intervention. Both subclasses PC-O and PE-O were significantly reduced by supplementation with microalgae (Table 1). Furthermore, phosphatidic acid, comprising only two quantifiable lipid species, showed significantly lower concentration in the microalgae group with 0.38 ± 0.03 nmol/mg total lipids. No further changes were observed for the remaining lipid classes (Table 1). It is further noteworthy that free cholesterol levels were not altered due to microalgae supplementation (Supporting Table S2). Differences are classified as being significant and have a type 1 error p < 0.05 and a log2-fold change > 0.5. Abbreviations of lipid classes: triacylglyceride, TAG; phosphatidylcholine, PC; ether-linked PC, PC-O; ether-linked PE, PE-O; phosphatidylethanolamine, PE; sphingomyelin, SM; phosphatidylinositol, PI; phosphatidylserine, PS; lyso-phosphatidylinisitole, LPI; diacylglyceride, DAG; lyso-phosphatidylcholine, LPC; lyso-phosphatidylethanolamine, LPE; cardiolipin, CL; phosphatidic acid, PA; phosphatidylglycerol, PG. To gain further insight into the metabolic imprint caused by microalgae supplementation, partial least-squares discriminant analysis (PLS-DA) was applied (Figure 1). Partial least-squares discriminant analysis (PLS-DA) for the lipidome of pig muscle tissue (longissimus thoracis). (A, C) PLS-DA sample plot and loading plot for the total concentration of the analyzed 13 lipid classes and 2 subclasses and (B, D) PLS-DA sample plot and loading plot for the concentration of 15 important lipid species, referring to a cutoff of 0.785. The data set comprised lipidome data of the control group (n = 15—blue) and microalgae group (n = 16—orange). In the loading plot (D), 15 lipid species were selected, which had the strongest impact on group separation (the complete data set is listed in Supporting Table S2). The PLS-DA analysis based on concentrations of lipid classes already revealed a clear separation of the two diet groups (Figure 1A). However, the analysis of the 336 lipid species in muscles of both diet groups (Figure 1C) resulted in a much clearer separation. This analysis indicated that PLS-DA with all species in the muscle of the two groups provides a much more differentiated separation of the two groups when comparing the sum concentration of the lipid classes. This is an indication that specific lipid species were strongly affected by the diet. In addition, the loading plots of total concentrations of lipid classes (Figure 1B) and concentrations of 15 important lipid species (Figure 1D) (referring to a cutoff of 0.785) were presented. This analysis further underlined the strong influence of TAG species on the separation of the diet groups shown in PLS-DA (Figure 1C). TAGs with a high degree of unsaturation were increased in the microalgae group, which were in contrast with some of the major abundant species like TAG 52:2 and TAG 52:1 that were not affected by supplementation. This illustrates that the use of all 336 single lipid species identified in pig muscle for PLS-DA lead to a stronger separation of both diet groups compared to the use of 15 lipid classes, only. Further pairwise comparison of the lipidomes revealed that 199 out of 336 lipid species showed significant concentration changes that represent a prevalence of 66% (q < 0.01). Using more stringent cutoff criteria, still 113 of the 336 lipid species (34%) showing significant differences (q < 0.01) also have substantial changes in concentration (abs [log2FC] > 1.0). This result was visualized in a volcano plot (Figure 2 and Supporting Table S2) and underlines that microalgae supplementation is strongly reflected in the muscle lipidome. Comparison of altered lipid quantities in muscle tissue (longissimus thoracis) of pigs fed according to the control diet and with microalgae supplementation. q-values were calculated from p-values of multiple t-tests between control and microalgae groups using log-transformed data, limiting the false discovery rate to 0.01. Differences are classified as being substantial (gray shaded area) if they are both significant with q < 0.01 and have a log2-fold change > 1. Selected lipid species are annotated according to the complete list of quantified lipids (Supporting Table S2). Next, we investigated in detail which lipids were mostly affected with a focus on two questions. First, do lipids comprising DHA and EPA show increased abundance after microalgae supplementation? Second, are there compensational effects detectable? For the top 20 nutritional most affected lipids, representing 63.5% of overall concentration change (Table 2), one can recognize that lipids with more than four double bonds were increased in abundance. The strongest gain in abundance was observed for highly unsaturated TAGs that are likely to comprise DHA and EPA, the main n-3 PUFAs in lipids of microalgae (Schizochytrium sp.). At the same time, it can be observed that the six PE-O and PC-O lipids of the top twenty list were decreased in concentration compared to the control (Table 2). Furthermore, it can be assumed that for all PE-O and PC-O species comprising aliphatic chains with double bonds ≤ 4 a reduction in concentration was observed (Figure 3A,B). Membrane lipids comprising DHA (n = 18/19) and EPA (n = 13/14) showed in the majority of cases increased abundance in the microalgae group. With particular interest, PE-O species using a specific mass spectrometric fragmentation mechanism for identification of PE-plasmalogens were analyzed. This analysis further supported the general observation that DHA and EPA comprising lipids were increased in concentration in the microalgae group, while main components with double bonds ≤ 4 were generally reduced when compared to the control (Figure 3C and Supporting Table S3). A detailed analysis of TAG profiles was performed, which confirmed that most of the abundant species with fewer double bonds were not affected by microalgae intervention. TAG 52:2 and TAG 52:1 in pig muscle had concentrations of 247 nmol/mg total lipids (control) vs 232 nmol/mg total lipids (microalgae) and 144 nmol/mg total lipids (control) vs 137 nmol/mg total lipids (microalgae), respectively (Figure 4A and Supporting Table S2). Concentrations (nmol/mg total lipid) of most abundant alkyl/alkenyl ether of PC-O (A) and PE-O (B) in muscle (longissimus thoracis) of pigs fed with control vs microalgae-based diet (*significant differences log2-fold change > 1, q value < 0.01), sorted by the highest mean concentrations. (C) PE-plasmalogen analysis derived from the positive ion mode tandem mass spectrometric analysis according to its specific fragmentation. The molar percentage was computed on basis of the PE-plasmalogen fragment ion intensities for all identified 49 identified species in control (C) and microalgae (M) groups (Supporting Table S3). Concentrations (nmol/mg total lipid) of TAG species containing up to four double bonds (A) and ≥ 4 double bonds (B) in muscles (longissimus thoracis) of pigs fed with control vs microalgae-based diet (*significant differences log2-fold change > 1, q value < 0.01), sorted by the highest mean concentrations. (C) Analysis of the fatty acid composition of TAG 52:2, TAG 56:6, and TAG 54:6. [TAG + NH4] adduct ions undergo a neutral loss (NL) of the fatty acid and ammonia and the contribution of single fatty acids can be estimated from the intensities of resulting fragment ions (Supporting Figures SF1–SF4). (D) Overall FA content of TAG determined from MS analysis. Presented profiles are the mean values for the complete data set, control (n = 16) and microalgae (n = 15). The table is sorted by the highest group differences. They were not diet-affected, also reflected in the sum TAG concentrations (Table 1). However, 10 of the 57 TAG species (e.g., TAG 56:6, TAG 56:8, TAG 54:7, Figure 4B) comprising at least 5 double bonds showed significantly increased concentration with log2FC > 1. MS analysis of these TAG species revealed the increased incorporation of DHA and EPA for the microalgae group. This was exemplarily shown for TAG 56:6 and TAG 54:6 in Figure 4C, while the FA composition for TAG 52:2 was not changed. Subsequently, the FA compositional changes between both groups for all TAG species were determined (Figure 4D). Besides the expected increase for DHA and EPA, the significantly increased abundance of fatty acids (FAs) 16:2, 16:3, 18:4, 20:4, 22:3, and 22:5 could further be observed. Some of these FAs were not analyzed in standard assays, and at this level of analysis, the identity of isomers cannot be determined resulting from double bond position and configuration. Noteworthy is also the detection of the highly unsaturated FAs 20:6 and 18:5 that were not affected by microalgae intervention. Our results confirm that dietary DHA and EPA, highly enriched in lipids of microalgae (Schizochytrium sp.), were incorporated into storage and membrane lipids in pig muscles. Thus, supplementation with microalgae offers a unique opportunity to increase the levels of essential n-3 LC-PUFAs (DHA, EPA) in pork and thus contributes to the recommended intake of long-chain n-3 PUFAs by the consumers. Dietary n-3 PUFAs incorporated in complex lipids can affect a range of metabolic and physiological functions, such as energy storage, membrane organization, and signal transductions via lipid mediators synthesized by cyclooxygenase (COX), lipoxygenase (LOX), and cytochrome P450 (CYP). These physiological important processes are modulated in regard to different key lipids in which n-3 PUFAs are bound. Accordingly, lipidome analysis was performed to catalogue-induced compositional changes by n-3 PUFA supplementation in pig muscle tissues. It has been shown that dietary n-3 PUFAs can inhibit the transcription of lipogenic genes by suppressing sterol regulatory element-binding protein 1c (SREBP-1c) gene expression or by inhibiting the proteolytic release of nuclear SREBP-1c in pig muscles and adipose tissues. It appears that n-3/n-6 PUFAs act as ligands/modulators for nuclear receptors, thereby suppressing de novo fatty acid synthesis, and thus, lipogenesis and n-3 PUFAs appear to be more potent than n-6 PUFAs. Lipidomic approaches on pig muscle tissues to study dietary effects and lipid metabolism have so far only been conducted very rarely. Our present study using microalgae (Schizochytrium sp.) clearly demonstrated the incorporation of dietary DHA and EPA (up to 37% DHA and 1% EPA in total diets) in membrane and storage lipids of pig muscle tissues (Table 2 and Supporting Table S2). Microalgae supplementation led to an almost two times higher number of lipid species containing ≥ 5 double bonds in pig muscles compared with the control. This incorporation of DHA/EPA should lead to the reorganization of membrane composition in pig muscles, which was indicated by the changed meat quality in our study as published before. Microalgae supplementation increased the water holding capacity (WHC) and the protein proportion in Landrace pig muscles compared to the control group. The DHA/EPA incorporation into membrane lipids could be predominantly shown for PC16:0_22:6, PE18:0_22:6, PI18:0_22:6, and EPA-containing species (PC16:0_20:5, PE18:0_20:5) with the highest species concentrations in muscle of pigs fed with microalgae-supplemented diets. This supports the hypothesis that primarily n-3 PUFAs seem to enable the muscle fibers to build a more flexible lipid bilayer membrane associated with higher WHC in pigs. Moreover, the observed compositional changes of membranes by the incorporation of EPA/DHA in skeletal muscle lipid species appear to stimulate the synthesis of proteins, resulting in higher total protein contents in pig longissimus thoracis muscle. DHA/EPA incorporation into the membrane lipid species was contrasted by the displacement of ether lipid species containing n-6 PUFAs with double bonds ≤4 in the acyl chains (e.g., PC-O-16:0_18:2, PC-O-16:1_18:2, PE-O-18:1_20:4, Figure 3), resulting in lower species concentrations. This is in line with our results of total fatty acid analysis showing lower 18:2n-6, 20:4n-6, and 22:4n-6 concentrations in muscles of pigs fed with microalgae compared to the control group. This finding suggests that systemic changes have occurred because of nutritional intervention with microalgae. Generally, it is known that ether-linked phospholipids represent about 20% of total PLs in mammalian cells. These ether lipids comprise two structurally different types, defined by either alkyl or alkenyl linkage of the aliphatic chain on glycerol, which have different physicochemical properties and most likely different functions. In recent studies, the occurrence of ether-linked phospholipids in farm animal tissues has been described; however, the physiological effects on the muscle of pigs fed with different PUFA-based diets were not investigated. Generally, plasmalogens are doubted to function as endogenous antioxidants in tissues. We assume that ether-linked phospholipids in the microalgae group may have been consumed to maintain antioxidant status/membrane homeostasis in the muscle. However, the complex interaction of lipid synthesis pathways and nutritional intervention is interconnected by lipid signaling events and crosstalk with other lipid classes. We likely detect effects of ether lipids signaling and PUFA-containing PLs that are known precursors of signaling molecular species like lipid mediators. This should be considered in further dietary lipid studies on pigs to cover such secondary up- or downregulation of other lipids occurring because of the primary effect of the diet. Based on the current knowledge, plasmalogens should play an important role in the regulation of membrane homeostasis, in particular membrane trafficking. Microalgae-supplemented diet led to increased concentrations of LPC 20:5, LPI 20:5, and LPI 22:6 in pig muscle and reduced concentrations of n-6 PUFA-containing lysophospholipid species (LPC 22:4, LPI 22:4). The overall low concentration level of these lysolipids points rather to changed lipid signaling than structural changes. Lysolipids are products of phospholipase A2 (PLA2), and for instance, LPC 22:6 shows anti-inflammatory properties compared to other n-6 PUFA-substituted LPCs. It is also noteworthy that despite the modifications in the acyl groups of individual species of LPC, LPE, and LPI in response to the microalgae diet, there was no significant difference in the sum concentrations of these PL classes between both diets groups (Table 1). This observation suggests that the functional properties of individual species of LPE, LPC, and LPI are potentially modulated by the incorporation of long-chain n-3 PUFAs, however without affecting sum PL class concentrations in pig muscle. The DHA/EPA incorporation in membrane lipid species in pig muscles should originate almost completely from microalgae supplementation, although pigs are able to synthesize these long-chain n-3 PUFA de novo. Our recent results of fatty acid analysis of pig muscle tissues confirmed the high accumulation rate of DHA/EPA showing significantly higher total DHA and EPA concentrations in microalgae group, for DHA (97.4 mg/100g muscle) compared to the control group (14.7 mg DHA/100g muscle). However, it is not possible to infer the origin of DHA and EPA from microalgae supplementation or de novo synthesized using the available data. A number of pig diet intervention studies revealed that feeding of vegetable-based PUFA-rich supplements—lipids high in 18:3n-3 and 18:2n-6 (linseed or rapeseed, cakes/oils/meals)—did not result in enhanced DHA/EPA accumulation in muscle tissues. One reason seems to be that n-3 PUFA supplementation of pig diet inhibits the expression of transcription factors and genes encoding lipogenic enzymes (SREBP-1, ELOVL5, FADS1, and FADS2) and caused the endogenous de novo synthesis of DHA comprising lipids to remain unchanged. In particular, dietary n-3 PUFAs have been shown to inhibit the transcription of lipogenic genes by suppressing SREBP-1c gene expression in porcine muscle and adipose tissue. Based on current knowledge, it appears that n-3/n-6 PUFAs act as modulators for nuclear receptors and consequently suppress de novo fatty acid synthesis, so lipogenesis and n-3 PUFAs appear to be more effective than n-6 PUFAs. A recent pig dietary intervention study applying supplements of 5–10% insect meal in the diet indicated decreased incorporation of de novo synthesized DHA into hepatic phospholipid species using a lipidomic approach; however, no specific investigation into the skeletal muscle tissues was performed. For the endogenous synthesis of DHA, two pathways are assumed, one is the elongation of 22:5n-3 (DPA) to 24:5n-3 subsequently desaturated into 24:6n-3 (Δ6 desaturase) and backconverted into 22:6n-3 (DHA) via peroxisomal β-oxidation. The second potential pathway can be the direct desaturation of 22:5n-3 into 22:6n-3 (Δ4 desaturase). Thus, further detailed lipidomic studies combined with transcriptomic and proteomic studies on pigs fed with different diets are necessary to investigate the putative pathways of DHA incorporation into skeletal muscle tissues. In conclusion, lipidomic analysis can improve our understanding of lipid metabolism and its influence on skeletal muscle physiology. Further investigation of dietary n-3 PUFA incorporation into storage and membrane lipids is required to gain insight into functional and physiological consequences of microalgae supplementation and other n-3 PUFA natural sources.
|
PMC11269705
|
Spatially and temporally probing distinctive glycerophospholipid alterations in Alzheimer’s disease mouse brain via high-resolution ion mobility-enabled sn-position resolved lipidomics
|
Dysregulated glycerophospholipid (GP) metabolism in the brain is associated with the progression of neurodegenerative diseases including Alzheimer’s disease (AD). Routine liquid chromatography-mass spectrometry (LC-MS)-based large-scale lipidomic methods often fail to elucidate subtle yet important structural features such as sn-position, hindering the precise interrogation of GP molecules. Leveraging high-resolution demultiplexing (HRdm) ion mobility spectrometry (IMS), we develop a four-dimensional (4D) lipidomic strategy to resolve GP sn-position isomers. We further construct a comprehensive experimental 4D GP database of 498 GPs identified from the mouse brain and an in-depth extended 4D library of 2500 GPs predicted by machine learning, enabling automated profiling of GPs with detailed acyl chain sn-position assignment. Analyzing three mouse brain regions (hippocampus, cerebellum, and cortex), we successfully identify a total of 592 GPs including 130 pairs of sn-position isomers. Further temporal GPs analysis in the three functional brain regions illustrates their metabolic alterations in AD progression.Glycerophospholipids (GPs) are the building blocks of cell membranes and play many critical roles in a wide variety of physiological processes, including energy storage, signal transduction, cell proliferation, and apoptosis. Consequently, GP composition is carefully regulated to ensure proper cellular functions. A growing number of studies have demonstrated that dysregulation of lipid metabolism is associated with various pathologies, including diabetes, cancers, and neurodegenerative diseases. In particular, mounting evidence has implicated that the aberrant alteration of stereospecific numbering (sn)-position-specific GPs in terms of abundance or the ratio of GP/lyso-GP is closely related to a variety of cancers and neurodegenerative diseases, owing to homeostatic disruption of sn-position selective phospholipase and lysophospholipid acyltransferase enzymes that are involved in GP remodeling. In the central nervous system, GPs are responsible for the proper functions of synapses, receptors, transporters, neurotransmitters, and signaling processes. Accordingly, numerous studies have revealed that altered GP compositions and phospholipase enzyme activities in the brain region could link to neurodegeneration including Alzheimer’s disease (AD). Heterogenous spatial distribution and alteration of GPs have been observed across different functional regions in accordance with aging and AD progression, which could be attributed to region-specific pathologies. Thus, performing the in-depth characterization of spatial and temporal sn-position-resolved GP profiles across complex functional brain regions would help to reveal and provide a more precise interpretation of the molecular mechanisms of AD progression underpinned by GP metabolism. Demonstrating high sensitivity, resolution, and the capacity for structural characterization, mass spectrometry (MS) based strategies, either alone or coupled with liquid chromatography (LC-MS), have become popular choices for lipid identification and quantification. However, lipidome-wide analyses with complete structural characterization still represent a long-standing analytical challenge in lipidomics. Currently, routine lipidomics analysis workflows using collision-induced dissociation (CID)-based MS/MS can only identify GPs at the level of lipid fatty acyl compositions, but fail to reveal subtle yet important structural features such as C═C position/geometry and sn-position. As sn-isomeric GPs are commonly present in biological mixtures in a wide dynamic range and are commonly co-eluted in LC, although we could assign the sn-position for the dominant GP component by the relative intensities of the two fatty acyl chain carboxylate anions, it would still be difficult to prove or rule out the presence of its sn-isomer. Providing comprehensive lipid structural characterization hinges on the development of alternative analytical approaches that overcome the technical obstacles of traditional methods. In recent years, significant advances have been achieved in elucidating GP sn- and C═C location isomers by chemical derivatization strategies and/or different ion activation/dissociation techniques to generate structure-specific fragment ions. These include electron-induced dissociation (EID), electron impact excitation of ions from organics (EIEIO), ozone-induced dissociation (OzID), ultraviolet photodissociation (UVPD), CID and UVPD of lipid−metal ion complexes, and Paternò–Büchi (PB) reaction followed by CID. However, it is still challenging for these strategies to precisely quantify lipid isomers, primarily due to the different dissociation efficiencies between isomers and interfering diagnostic ions from co-eluting lipid species. Therefore, powerful separation techniques at intact lipid level are needed to complement ion activation and provide unequivocal structural assignment and quantification. Ion mobility spectrometry (IMS) is a rapid gas-phase separation technique in which ions are separated by size, shape, and charge. Collision cross section (CCS), originating from rotationally averaged measurements of the cross-sectional area of an analyte ion in the gas phase, can be directly correlated with the structure and conformation of the gas-phase ion. Coupling ion mobility (IM)-MS with LC shows promise in lipid isomer separation and annotation by providing high selectivity across four dimensions: m/z, retention time (RT), CCS, and MS/MS spectra. More importantly, recent reports have demonstrated that precise and reproducible lipid CCS measurement together with machine learning algorithms could enable large-scale lipid CCS prediction. For example, Zhu and coworkers utilized support vector regression (SVR) models to predict the CCS values of a variety of lipid classes. McLean and colleagues presented a large, unified CCS compendium to predict CCS values of compounds including 810 lipid species according to the relationship between CCS and m/z. Recently, Zhu and co-workers reported the strategy that integrated four-dimensional (4D) (m/z, RT, CCS, and MS/MS) library-based matching and rule-based refinement to reduce over-report and improve the accuracy of lipid identification. These current 4D lipidomics studies greatly improved the depth in lipid structure characterization. However, in these studies that utilized rule-based refinement for in-depth lipid analysis, only a part of GP species was roughly identified at fatty acyl sn-position level and the existence of sn-position isomers was largely ignored. Specifically, in the rule set for lipid identification, GP species from each extracted peak was assigned to only one of the sn-isomer pairs if the intensity ratio of sn-1/sn−2 was <0.9 in negative ion mode, while the existence of corresponding sn-isomers was ignored. Additionally, the identification of GP was still at fatty acyl level when the intensity ratio of sn-1/sn−2 was ≥ 0.9. Moreover, the quantification could only be achieved for the most abundant isomer or fatty acyl sum composition. Nevertheless, these pioneering studies motivated the lipid structural annotation and quantification to the next level for providing unequivocal structural assignment of lipid isomers. However, robust implementation of such an approach for lipids with a higher level of structural elucidation requires the availability of high-resolution IM measurements. Typically, sn-position isomers exhibit around 1% differences in CCS values and therefore require IM resolving power (Rp) over 100 for separation. This high Rp was out-of-reach in previous IM-based lipidomic analyses, eliminating the capacity for any of these former methodologies to accurately measure and predict the CCS values of lipid sn-position isomers. Recently, using trapped ion mobility spectrometry (TIMS)-MS with Rp up to 410 using ultra-low scanning rate, Fernandez-Lima and his team have shown the potential of high-resolution IM in discriminating phosphatidylcholine (PC) sn-isomers in human plasma. High-resolution demultiplexing (HRdm) strategy was first reported in 2020, which is an extended Hadamard multiplexing and post-acquisition data processing strategy for improving the sensitivity and resolution of IM measurements without the need for instrument modifications. In the past four years, a series of isomeric biomolecules in complex samples were investigated with HRdm, including glycans, peptides, metabolites, and oxidized lipids. However, utilizing the strategy to systematically separate subtle yet crucial structural sn-position isomers of GPs in biological samples has not been widely explored in existing studies. In addition to requiring high IM Rp, large-scale GP sn-isomer distinguishment is also hindered by insufficient commercially available pure lipid standards with determined sn-position, thus reducing the confidence of empirical assignments. And finally, even though predictive models may be considered a worthwhile approach, very few molecular descriptors (MDs) in common simulation packages reflect differences of GP sn-position isomers. This limitation essentially eliminates the ability to distinguish species in silico. To overcome these challenges, in this work, we develop a high-resolution IM-MS-based 4D lipidomics strategy that leverages a machine learning-empowered library for large-scale, in-depth structural analysis of GP sn-position isomers. The use of HRdm strategy provides an increase of drift tube ion mobility spectrometry (DTIMS) Rp from ∼50 to 250 while still allowing millisecond IMS separation of GP sn-isomers without any instrumental modifications. We further construct a comprehensive experimental 4D GP database of 498 GPs identified from pooled mouse brain lipid extracts. These empirical measurements are redeployed to facilitate the creation of an in-depth, extended 4D library of 2500 GPs by machine learning-based prediction. With both the experimental database and the extended library, a significantly high number (>540) of GP species with sn-position information are identified and quantified from each of the three functional brain regions, revealing the spatial and temporal GP alterations in the brains of wild-type (WT) and APP/PS1 AD mouse model. By revealing significant changes in either abundance or ratios of sn-isomers in a set of GPs during aging and AD progression, we demonstrate that this developed strategy is powerful to uncover potential biomarkers for AD progression associated with dysregulated GP metabolism. Distinguishing the minute structural differences between GP sn-position isomers is a long-standing analytical challenge by IM-MS methodologies. Recently, HRdm strategy was put forth to overcome the known limitations in IM Rp and duty cycle. By using multiplexed ion injection and post-acquisition data processing, this technological development has been shown to significantly improve the sensitivity and resolution of IM-MS measurements without the need for instrument modifications. Taking the lipid standard PC 18:1(11Z)/18:1(11Z) as an example, Fig. 1a illustrates characteristic HRdm multiplexing – the injection of 16 ion packets at predetermined intervals across the drift time window. This multiplexing strategy results in an increased duty cycle and a decreased onset of detection saturation. The raw, multiplexed ion packets in the drift spectrum are then deconvoluted into one, resulting in a spectrum reminiscent of single pulse ion injection and bearing no difference in measured drift time (Fig. 1b). Although standard demultiplexing itself only slightly improves the IM Rp (55 in standard demultiplexed mode vs 52 in single pulse) due to reduced space charge effects, the signal intensity increased by more than 12-fold higher when compared to single pulse mode (Fig. 1c). Further processing of the demultiplexed data by Hadamard Transform and data post-processing (see Methods), HRdm processed data revealed IM peak widths significantly narrower than those observed in either single pulse or standard demultiplexed spectra, achieving high IM Rp up to 250 across these standard trials. Because the total signal is preserved, the reduction in peak width results in an increase in peak height, providing an additional ~ 5-fold increase in sensitivity (Fig. 1c).Fig. 1Improved sensitivity and separation of GP sn-isomers with high-resolution HRdm IM-MS.a Raw IM-MS heatmap of PC 18:1(11Z)/18:1(11Z) obtained from 5-bit multiplexing injection mode. b Demultiplexed IM-MS heatmap of PC 18:1(11Z)/18:1(11Z). c Overlaid drift spectra of the single pulse, 5-bit standard demultiplexing, and 5-bit HRdm. Overlaid drift spectra of PE 16:0/18:1(9Z), PE 18:1(9Z)/16:0, and equimolar mixture processed by standard demultiplexing (d) and HRdm (e), and all drift spectra are normalized to maximum peak height. Overlaid drift spectra of PS 16:0/18:1(9Z), PS 18:1(9Z)/16:0, and equimolar mixture processed by standard demultiplexing (f) and HRdm (g). Mass spectra and HRdm IM-MS heatmaps of the equimolar mixture of PE 16:0/18:1(9Z) and PE 18:1(9Z)/16:0 standards (h) and PS 16:0/18:1(9Z) and PS 18:1(9Z)/16:0 standards (i). Source data are provided as a Source Data file. a Raw IM-MS heatmap of PC 18:1(11Z)/18:1(11Z) obtained from 5-bit multiplexing injection mode. b Demultiplexed IM-MS heatmap of PC 18:1(11Z)/18:1(11Z). c Overlaid drift spectra of the single pulse, 5-bit standard demultiplexing, and 5-bit HRdm. Overlaid drift spectra of PE 16:0/18:1(9Z), PE 18:1(9Z)/16:0, and equimolar mixture processed by standard demultiplexing (d) and HRdm (e), and all drift spectra are normalized to maximum peak height. Overlaid drift spectra of PS 16:0/18:1(9Z), PS 18:1(9Z)/16:0, and equimolar mixture processed by standard demultiplexing (f) and HRdm (g). Mass spectra and HRdm IM-MS heatmaps of the equimolar mixture of PE 16:0/18:1(9Z) and PE 18:1(9Z)/16:0 standards (h) and PS 16:0/18:1(9Z) and PS 18:1(9Z)/16:0 standards (i). Source data are provided as a Source Data file. To ascertain whether the high IM Rp from HRdm could benefit sn-isomer separation on a large scale, a series of isomer-pure GP standards were evaluated through both standard demultiplexing and HRdm processing with their sn-isomers (Fig. 1d–i). Phosphatidylethanolamine (PE) and phosphatidylserine (PS) standards were examined in their most abundant ion forms in positive mode ESI, [M + H]. PE 16:0/18:1(9Z) and PE 18:1(9Z)/16:0 exhibited a small CCS or drift time difference of around 1%, precluding any mobility separation of equimolar mixtures through the modest resolution found in standard demultiplexing (Rp~50, Fig. 1d). Achieving Rp of over 200 via HRdm, the enhanced Rp readily facilitated almost baseline separation of these PE sn-position isomers when analyzed separately and as a mixture of two (Fig. 1e), thus increasing peak-to-peak resolution (Rpp) from 0.28 to 1.4. These results were rearticulated during our analysis of PS sn-isomers, PS 16:0/18:1(9Z) and PS 18:1(9Z)/16:0. Failing to differentiate equimolar mixtures of isomers through standard demultiplexing (Fig. 1f), the pair of sn-position isomers were successfully separated using HRdm (Fig. 1g). Generally, sn-position isomers, as constitutional isomers, exhibit CCS difference of approximately 1%, which needs more than 100 resolving power to achieve 10% separation and more than 200 resolving power to achieve 90% separation in low-field DTIMS instruments. Owing to the high resolution afforded by HRdm, most of GP sn-position isomers could be separated sufficiently. The quality of a separation is quantified in terms of Rpp with peaks deemed resolved when Rpp exceeds 0.5. To provide a comprehensive evaluation of the separation efficiency, a series of GP sn-position isomers were assessed using HRdm to evaluate the effectiveness of separation by this technique (Supplementary Fig. 1). The Rpp for GPs with different fatty acyl chain compositions from different classes were at a range from 0.78 to 1.4, indicating that each pair of isomers was successfully separated. Phosphatidylglycerol (PG), phosphatidylinositol (PI), and phosphatidic acid (PA) were predominately present as [M + NH4] in positive ion mode. The sn-isomers from these classes were also separated by HRdm in ion forms of [M + NH4] (Supplementary Fig. 1). The IM separation of GP sn-isomers with sodium adduction in positive mode (Supplementary Note 1 and Supplementary Fig. 2) and their deprotonated form, [M-H], in negative mode (Supplementary Note 2 and Supplementary Fig. 3) was also investigated. Examining HRdm IM-MS heatmaps of the equimolar mixtures of the PE (Fig. 1h) and PS (Fig. 1i) sn-position isomers, our data reveal excellent drift time alignment across all isotope envelopes, highlighting accuracy, precision, and reproducibility of CCS measurements within HRdm. HRdm separation of GP sn-position isomers at various molar ratios was also demonstrated quantitatively (Supplementary Fig. 4). A mixture of isomer-pure PE 16:0/18:1(9Z) and PE 18:1(9Z)/16:0 standards at the ratios ranging from 1 to 10 were adequately separated with Rpp of 1.01-1.46, demonstrating the consistency of separation in the complex mixture. We also validated that the relative abundance of each isomer could be truthfully reflected by the ratio of the peak areas (Supplementary Fig. 4), indicating the quantitative accuracy of HRdm. Additionally, the high accuracy in quantifying GP sn-isomers across a wide range of GP classes, using ion abundance in HRdm drift spectra, was validated by the well-established phospholipase A2 (PLA2) digestion method. The information on GP standards was listed in Supplementary Table 1. The results indicated that the isomeric abundance obtained using IM-MS aligns consistently with that obtained via PLA2 digestion (Supplementary Note 3, Supplementary Fig. 5, and Supplementary Table 2). In order to further demonstrate the reproducibility of HRdm technique in complex biological samples, we compared the drift spectra of isotope-encoded lipid standards spiked in pure solvent (isopropanol) and complex biological samples (mouse brain lipid extract) acquired with HRdm. As shown in Supplementary Fig. 6, all the drift spectra of deuterium-labeled lipids from the complex biological matrices could still align well with these in pure solvent, indicating that HRdm processing did not generate artificial peaks due to matrix interferences. Deuterium-labeled triacylglycerol (TG), diacylglycerol (DG), monoacylglycerol (MG), PC, PE, PS, PG, PI, PA, and sphingomyelin (SM) showed a single peak with high-resolution IM technique (Rp up to 217) in both pure solvent and complex biological samples. Meanwhile, small peaks from isomers were observed in the drift spectra of LPC 18:1 (d7)/0:0 and LPE 18:1 (d7)/0:0 (Supplementary Fig. 6j-k). In order to validate that these small peaks are results from true isomers rather than artificial signals, we performed the PLA2 digestion of isopure standards to obtain the pure lyso-GP standards. As shown in Supplementary Fig. 7, all isopure lyso-GPs including LPC 18:1/0:0, LPE 18:1/0:0 appeared as a single peak and aligned well with deuterium-labeled standards. As not all these deuterium-labeled lyso-GP standards are isopure standards, we deduced that these two lyso-GPs contain small amounts of corresponding sn-isomers with fatty acyl at sn−2. This also accords with previous studies that lyso-GPs with fatty acyl at sn−2 have the larger CCS values. The peaks of deuterium-labeled lyso-GP in both pure solvent and biological samples were also all aligned well. Moreover, it is important to note that our study utilized the officially released stable version 2.0 of the HRdm software, which includes several key improvements to largely reduce or eliminate artifacts compared with the initial report of HRdm in 2020. Taken together, we are confident that lipid species identified by HRdm from complex biological samples could largely reflect true lipid signals rather than artificial peaks. To provide reasonable comparison and demonstrate the advantages of HRdm strategy, we also evaluated other types of commonly used high-end IM paradigms, including TIMS and Traveling Wave IMS (TWIMS)-based cyclic IM (cIM), for large-scale GP profiling at fatty acyl sn-position level. The IM-MS data acquisition parameters for TIMS and cIM-MS were described in the Supplementary Information. As shown in Supplementary Fig. 8, in TIMS, sn-isomers could not be separated when we used the default 4D lipidomics instrument parameters with 100 ms ramping time. It required a prolonged ramping time (up to 1000 ms) to achieve a slight separation. The sensitivity remarkably decreased as only ~10% ion intensities were preserved with the long ramping time. In cIM-MS (Supplementary Fig. 9), an increase in a number of cyclic passes of up to 20 passes (around 400 ms arrival time) was needed to achieve a similar degree of separation as that shown in HRdm. It is important to note that this approach is more suitable for targeted analysis as it requires a carefully calculated narrow IM selection window to prevent “wrap-around” in the cIM where high-mobility ions catch up with low-mobility ions in multipass experiments. The ion intensity was also significantly decreased to ~10% when increasing the number of cyclic passes to 20. These compromises, including prolonged drift/ramping time and/or narrowed IM selection window, would come at cost of sacrificing either sensitivity, throughput, or IM coverage for large-scale untargeted lipidome profiling. In contrast, this DTIMS-based HRdm strategy demonstrated unique advantages for large-scale lipidome profiling in LC-IM-MS workflows as enhanced IM resolving power and sensitivity could still be achieved in a typical 60 ms IM scanning window for comprehensive ion collection. In addition, DTIMS provides reproducible first-principle CCS measurement, and CCSN2 measurement is considered the gold standard and is most widely reported in data repositories like PubChem, LIPIDMPAS, and MSDIAL. For these reasons, DTIMS-based HRdm IMS is currently the most suitable paradigm for sn-position-resolved lipidomic analysis and CCS database construction. Degree of GP structural identity can be categorized into five levels: lipid class, sum composition, fatty acyl composition, fatty acyl sn-positions, and C═C location/geometry (cis/trans) in the unsaturated fatty acyls (Fig. 2a). Separator “_” means unspecified sn-position, and “/” means confirmed sn-position for acyl/alkyl constituents. As current routine LC-MS-based lipidomic analysis often identifies only lipid fatty acyl composition, the elucidation of fatty acyl sn-positon through HRdm IM pushes the lipidomic analysis to the next stage. After examining multiple GP standards, including PC 16:0/18:1(9Z) vs PC 16:0/18:1(11Z) and PC 18:1(9Z)/18:1(9Z) vs PC 18:1(11Z)/18:1(11Z) (Supplementary Fig. 10), we find such species usually exhibit a CCS or drift time difference of less than 0.2% and require IM Rp ~ 1000 for baseline separation. As the relatively small CCS differences among C═C bond position isomers are negligible compared to sn-position isomers (~1%), the differentiation of GP sn-position isomers is not compromised. In most mammals, as fatty acyl composition of GPs in trans C═C configuration is rare, only cis C═C configuration is considered in this study. Knowing this, four-dimensional analysis of GPs obtained from LC-HRdm IM-MS (RT, CCS, precursor m/z, and MS/MS spectrum) provides an avenue towards unambiguous GP identification and structural characterization. For example, as shown in Fig. 2b, a representative GP precursor with a nominal m/z of 788 could be eventually classified into 8 species with sn-position information, twice the number of species as that may be resolved on high-resolution MS alone. However, though automatic sn-isomer-resolved GP identification may be facilitated through a 4D CCS library, de novo construction of such a library is too costly and not feasible given the lack of suitable standards. To remedy this shortcoming, we extracted GPs from mouse brain – a biological source known to be rich in GPs and sn-isomers for library construction.Fig. 2Comprehensive GP analysis at sn-position resolved level.a Hierarchy of GP identification and characterization using the identification of PC 16:0/18:1(9Z) at each level as an example. b Illustrating the identification of isobaric GPs detected at nominal m/z = 788 as an example with required molecular information for each level of identification. c MS2 and drift spectra of PC 16:0/18:1(11Z). d MS2 and drift spectra of PC 16:0_20:4. e MS2 and drift spectra of PE 40:8. The number (f) and proportion (g) of GP molecular species identified by LC–MS/MS (IM off) and LC–HRdm IM–MS/MS. Source data are provided as a Source Data file. a Hierarchy of GP identification and characterization using the identification of PC 16:0/18:1(9Z) at each level as an example. b Illustrating the identification of isobaric GPs detected at nominal m/z = 788 as an example with required molecular information for each level of identification. c MS2 and drift spectra of PC 16:0/18:1(11Z). d MS2 and drift spectra of PC 16:0_20:4. e MS2 and drift spectra of PE 40:8. The number (f) and proportion (g) of GP molecular species identified by LC–MS/MS (IM off) and LC–HRdm IM–MS/MS. Source data are provided as a Source Data file. To determine universal rules for assigning GP sn-position isomers based on MS/MS fragmentation patterns and CCS differences, we tested all 6 GP classes, each class containing at least 2 standards with different degrees of unsaturation together with the GPs from pooled mouse brain samples (Supplementary Table 1). The 1st rule was summarized from MS/MS fragmentation patterns; the fatty acyl chain fragment ions with higher intensities are at the sn−2 position in negative mode, as shown in Fig. 2c using PC 16:0/18:1(11Z) standard as an example. Many studies have also demonstrated that the peak intensity of the carboxylate anion from the sn−2 chain is approximately 3 times the intensity of the sn−1 chain of GPs due to sterically favorable release of the fatty acyl chain from the sn−2 position in negative ion mode. Although the assignment of fatty acyl sn-positions could be achieved by examining fragmentation patterns, there are instances where the assignment is obfuscated through similar abundances of GP sn-position isomers. Furthermore, if the sn−2 chain is polyunsaturated, the intensity ratio could be compromised due to the partial loss of CO2 during fragmentation. The loss of CO2 is observed in negative-ion mode fragmentation but is not observed when using metal-adducts in positive-ion mode. These commonly occurring cases may require further evidence for more confident sn-position assignment. Although most of GP sn-isomers could be assigned by matching the estimated abundance of sn-position isomers from the carboxylate anion intensity in MS/MS spectra and abundance of IM-resolved peaks (1st rule), there were circumstances when the two GP sn-isomers were of similar abundance. Based on the CCS values from GP standards and GPs successfully assigned through the fragmentation rule, we also concluded the 2nd rule to be that GPs with smaller fatty acyl chains at sn−1 position generally showed lower CCS values than their sn-position isomers, which was also reported previously. For example, PE 16:0/18:1 and PS 16:0/18:1 showed smaller CCS values than their sn-position isomers, PE 18:1/16:0 and PS 18:1/16:0, respectively (Fig. 1e, g). As this prevailing trend has not been widely reported due to the limitations in IM Rp, we reasoned that this finding could be attributed to the higher gas-phase flexibility and freedom of the fatty acyl chain at the sn−1 located at the terminus of the glycerol backbone. We also performed in silico simulations of several pairs of representative GP sn-isomers in the gas phase to prove this finding. The CCS values of the simulated GPs were consistent with our experimental measurement, both obeying the 2nd rule (Supplementary Fig. 11). Their coordination information (Z-matrix) is also included in Supplementary Data 1. With that, as shown in Fig. 2d, the similar intensities of C16:0 and C20:4 in the MS/MS spectrum of PC 36:4 existed together with the peak of C20:4 with CO2 loss. As the fragmentation pattern was not sufficient to assign the position to the 2 peaks shown in the drift spectrum, the 2nd rule was applied to determine that the peak at 40.54 ms was PC 16:0/20:4 and the peak at 40.87 ms was PC 20:4/16:0, respectively. In terms of lyso-GPs, we found, from commercial standards, that lyso-GPs with acyl chains at sn−1 position generally showed smaller CCS values (Supplementary Fig. 12), which was consistent with previous report. Combining both rules, identification of GPs at the sn-position level could be achieved within a complex mixture of isomers beyond solely sn-isomers. Taking PE 40:8 as an example (Fig. 2e), the fatty acyl composition could be concluded as PE 18:2_22:6 and PE 20:4/20:4 from the three carboxylate anions detected in the MS/MS spectrum. From the high intensities of C18:2 and C22:6 carboxylate anions at m/z 279.23 and 327.23 respectively, PE 18:2_22:6 could be concluded as the main constituents. Drift times of the two most abundant peaks, 39.55 ms and 39.94 ms, could be assigned to PE 18:2/22:6 and PE 22:6/18:2 according to the 2nd rule. The remaining peak at 40.85 ms could be subsequently annotated as PE 20:4/20:4. In the negative MS/MS spectra of this study, we found that fatty acyl composition of most GPs in mouse brain is not too complicated and almost all isomers at fatty acyl composition level are no more than 2. This is also reported by other literature. In this way, there could be no more than 4 isomers if we take sn-positional isomers into account. That is the main reason that we annotated no more than 4 isomers from HRdm spectra. According to these negative MS/MS spectra and drift spectra, the annotation workflow was illustrated in Supplementary Fig. 13. Additionally, a series of GPs from mouse brain extracts were also used to comprehensively evaluate the consistency of drift spectra and negative MS/MS spectra from the complex biological matrix (Supplementary Fig. 14). Specifically, to enhance the identification accuracy and confidence, we set the minimum peak height at 3000 counts in MS1 as a criterion for identification. Together with fatty acyl composition information of lipid isomers from DDA spectra, we follow the decision tree to conduct the annotation: If HRdm spectra show the same number of peaks as the number of compositional isomers, when GP contains one fatty acyl composition, it indicates sn-position isomers does not exist. We can assign the sn-connectivity of the fatty acyl chains by rule 1 (Supplementary Fig. 14a–c). For instance, as shown in Supplementary Fig. 14a, there was only a single peak in the HRdm drift spectrum of corresponding to the m/z of PC 44:10. The only two fatty acyl fragment ions, corresponding to C22:6 and C22:4 at a ratio of C22:4/C22:6 ≈ 3, indicated single fatty acyl composition (PC 22:6_22:4), and the connectivity could be further assigned as PC 22:6/22:4. When GP contains two fatty acyl compositions, it would be annotated as two sn-resolved fatty acyl compositions or two sn-isomers of major composition according to rule 1 and rule 2. For instance, as shown in Supplementary Fig. 14h, from the MS/MS spectra, PE 16:0_18:1 was the predominant fatty acyl chain composition, and the very low abundance of PE 16:1_18:0 indicated it would not be the second relatively high peak in drift spectrum. The intensity ratio of fragment ions from PE 16:0_18:1 accords with their abundance in drift spectrum, the connectivity could be further assigned as PE 16:0/18:1 and PE 18:1/16:0. It was rarely observed that the number of IM peaks less than fatty acyl compositions (means one IM peak, two fatty acyl compositions), if existed, it would be annotated as the major composition with sn-position. Another case is that the peak number is higher than the number of lipid isomers at the fatty acyl composition level. If the maximum possible number of sn-isomers matches the number of peaks from HRdm spectra (Supplementary Fig. 14 d–g, i, l), we would use rule 1 (and rule 2) to assign the sn-resolved GPs. There are also cases where the maximum possible number of sn-isomers is higher than the number of peaks from HRdm spectra. For example, two compositional isomers were identified in negative MS/MS spectra, which could lead to four possible sn-position resolved isomers, but only three peaks were observed in HRdm spectra (Supplementary Fig. 14j–k). If this happens, we will use rule 1 to pick the top 3 abundant sn-position resolved GP isomers. The assignment of the three peaks in HRdm spectra is based on peak intensities and rule 2. In our data, we did not observe the case that the number of peaks in HRdm spectra is higher than the maximum possible number of sn-isomers from negative MS/MS spectra after we carefully eliminate artifacts if any exist. In the very rare cases when more than two compositional isomers are identified in DDA spectra, we pick the top 4 abundant sn-position resolved GP isomers and the top 4 abundant peaks in HRdm spectra and then assign them based on rule 1 and rule 2. For the very low abundance isomers, although they could be identified in our workflow from MS/MS spectra in negative mode, they were not annotated from the HRdm drift spectra to enhance the identification accuracy and confidence. It is worth noting that this potential limitation also exists in current mainstream label-free lipidomics methods when quantification is the major focus. Through comparison with standards, the identification of GPs in mouse brains was further validated (Supplementary Fig. 15). Representative IM-MS heatmaps of GPs sn-isomers identified from mouse brains were shown in Supplementary Fig. 16. Hence, the combination of the carboxylate anion intensity in MS/MS spectra together with the abundance and CCS distribution in the IM spectra enables unambiguous GP isomer annotation in the drift spectrum. Thus, the abundance ratio of isomer pairs could be subsequently concluded. After analyzing pooled mouse brain lipid extracts by data-dependent LC-MS/MS acquisition using QTOF-only mode without IM measurement, a total of 318 molecular GP species with fatty acyl composition were identified (Supplementary Fig. 17a, Supplementary Data 2) – a number that is similar to the number of brain GPs reported using other methods. When operating on multiplexed IM-QTOF mode with the same LC gradient, integration of data-independent LC-IM-MS/MS and data-dependent LC-MS/MS acquisition (Supplementary Fig. 17b), 498 distinct GP molecular species were identified from the same sample, a ~57% increase in identification numbers as a result of sn-position annotation with the 1st and 2nd rules while a total of 452 GP could been annotated only through 1st rule (Fig. 2f and Supplementary Data 3). One of the concerns is that infrequent artifacts might affect the identification accuracy. Notably, the infrequent artifacts could be easily filtered out according to their features and would not affect spectral interpretation. Specifically, the artifacts are usually of low abundance (<10% of the primary signal within each m/z extraction window) and are sufficiently distant from the signal region of interest to not affect spectral interpretation. Moreover, IM drift times of genuine ion signal align across isotopic peaks due to the negligible contribution of mass to the ion mobility, whereas the isotopic peaks of artifacts are not aligned. For instance, as shown in Supplementary Fig. 18 a1, a2, the abundance of artifactual peak (at 41.69 ms) in IM-MS heatmap and extracted drift spectrum were much lower than true ions, and the isotopic peaks were not aligned. Thus, artifactual peak (at 41.69 ms) co-existing with PE 18:0/20:4 (40.38 ms) and PE 20:4/18:0 (40.72 ms) in QC sample were easily distinguished from genuine ions due to their long distance from the true ions, their low abundance, and non-aligned isotope peaks. The existence of PE 18:0/20:4 (40.38 ms) and PE 20:4/18:0 (40.72 ms) in QC sample also could be validated with the negative MS2 spectrum (Supplementary Fig. 18a3). On the one hand, the low possibility of the artificial peaks could be easily discriminated from our genuine ions and would not affect the identification accuracy. On the other hand, we also validated the drift spectra of lipid extracts from QC samples with that from most of investigated biological samples, the artificial peak, which only randomly appeared in certain spectra but not all spectra, could be distinguished from consistently appeared genuine peaks. For instance, as shown in Supplementary Fig. 18, one peak in drift spectrum of PE 18:0_20:4 happened to exist in the QC sample at a very low abundance and with an isotope of inconsistent drift time (Supplementary Fig. 18 a1&a2), but did not appear in other samples. The two major drift peaks aligned well across all drift spectra and their isotopes, indicating they were genuine peaks, PE 18:0/20:4 and PE 20:4/18:0. The very small peak (at 41.69 ms), with not aligned isotopic peaks, and far from the major peaks in the drift spectrum of QC sample, was the artificial peak. Therefore, checking drift time alignment of isotopic peaks and validation across replicates are recommended to determine the genuine peaks when using HRdm to acquire high-resolution IM data for unknown biomolecules and new database. From the 318 extracted drift spectra of 498 identified lipids in pooled mouse brain samples, artificial peaks were only observed in 15 extracted drift spectra, indicating that less than 5% of lipid ions co-existed with artifacts. Drift times of all artificial peaks from pooled mouse brain samples were labeled and supplemented in the notes of Supplementary Data 3. Herein, the low ratio of the artificial peaks could be easily discriminated from our genuine ions and would not affect the identification accuracy. To clearly highlight the obvious improvement in identification number benefits from high-resolution IM, a comparison of identification numbers from different 4D lipidomics strategies has been conducted. we performed the analysis of the lipid extract from mouse brain using 4D library-based match and rule-based refinement without additional HRdm strategy as Zhu and co-workers reported. GPs were determined as one major component of sn-position isomers if the intensity ratio of sn−1/sn−2 was <0.9, and the existence of their corresponding sn-isomers was ignored. As shown in Supplementary Fig. 19, a total of 326 GPs including 241 GPs at sn-resolved level and 85 GPs at fatty acyl level in mouse brain were identified. The identification result is comparable with studies that annotation of GPs at fatty acyl level using LC-MS/MS. In our study, using the 4D lipidomics with HRdm, a total of 498 lipid species have been identified with additional sn-position isomers from sn-position isomer pairs. To reduce the incidence of false positive identifications due to potential artificial peaks, the annotated GPs were manually assessed to ensure that their CCS values remained consistent across all replicate measurements. The robustness and reliability of our presented methodology were demonstrated by the precision and consistency of our measurements. Specifically, we observed an average MS error of 0.2 ppm. The average coefficients of variation (CV) of the RT was 2.1%, and the average CV of CCS was 0.1% (Supplementary Data 3). Proportions of each GP class were summarized in Fig. 2g. All six identified GP classes demonstrated an increase in identification numbers, owing to the significant portion of GPs containing sn-position isomers. As shown, the improved sensitivity and resolution of HRdm enabled the construction of a large and in-depth 4D GP library (Supplementary Data 4) that may be employed for automated data analysis. Compared to GP identification by only MS/MS in negative mode, although HRdm method has potential limitations in annotating very low abundance isomers when, in rare cases, multiple fatty acyl compositional isomers co-eluting, it increased annotation of high abundance fatty acyl composition isomers with additional sn-position information, expanding the quantifiable GP isomers. To note, current routine untargeted LC-MS/MS or LC-IM-MS/MS lipidomics methods only enable quantification at the level of GP sum composition or for the predominant composition among co-eluting GPs. In contrast, benefiting from high-resolution IM, separating GP sn-position isomers in the drift spectra enables the unequivocal quantification of both sn-position isomers. As demonstrated, coupling LC with a high-resolution IM-MS/MS platform allows for comprehensive, in-depth GP profiling. However, to facilitate the automated identification of sn-position-resolved GPs beyond those GPs from our manually validated 4D GP library, it is necessary to obtain CCS values accurate enough to differentiate sn-position isomers. Broadly speaking, CCS measurement of GP isomers is found to be severely lacking as few GP-rich biological sources are known and commercial standards are often limited, indicating that greater analytical understanding of GPs is challenging yet in demand. To address these limitations, we implemented a machine learning-based workflow, already validated in previous studies, for large-scale CCS prediction of GPs. This allowed us to expand on the experimental data generated from our analyses of brain lipid extracts (Fig. 3a).Fig. 3Machine learning-enabled CCS prediction for sn-position-resolved GPs.a Schematic illustration of the workflow to curate the extended GP CCS library. The internal validation of the machine learning algorithm predicted CCS values of GPs (n = 335) using optimized MDs selected from CDK (b), Mordred (c), and pool of the whole CDK and part of Mordred relevant to GP sn-position (d). e The prediction performance evaluated by external validation (n = 84). Source data are provided as a Source Data file. a Schematic illustration of the workflow to curate the extended GP CCS library. The internal validation of the machine learning algorithm predicted CCS values of GPs (n = 335) using optimized MDs selected from CDK (b), Mordred (c), and pool of the whole CDK and part of Mordred relevant to GP sn-position (d). e The prediction performance evaluated by external validation (n = 84). Source data are provided as a Source Data file. In the prediction workflow (Fig. 3a), the chemical structure of each GP was first converted to a simplified molecular-input line-entry specification (SMILES) format, from which molecular descriptors (MDs) of the molecule were calculated. MDs can be defined as mathematical representations of the molecule. By converting the chemical structures into numeric values of various MDs, they are used to quantitatively describe the physical and chemical information of molecules. Prediction was performed on singly protonated PC, PE, and PS to reduce systematic errors, largely due to insufficient experimental data input from PG, PI, PA, and their preference towards ammonia adduct formation (see Methods). The selection of appropriate MDs to effectively distinguish sn-position isomers in the prediction model was of vital importance and yet presented a significant challenge due to the subtle differences in GP structures and physicochemical properties. In our first trial for prediction model construction, MDs were calculated by the most widely used “rcdk” package (CDK) for lipid CCS prediction. Using the least absolute shrinkage and selection operator (LASSO) algorithm (see Methods), 24 MDs out of the 221 MDs from CDK were selected (Supplementary Table 4). With these 24 MDs, CCS prediction was carried out using an SVR-based machine learning algorithm, giving a linear fit with R = 0.9844 for the training set (Fig. 3b, Supplementary Data 5). While the high regression fit of our model is competitive when benchmarked against previous studies, higher prediction accuracy would still be needed for more confident GP sn-position assignment. As such, we also considered the 1824 two- and three-dimensional MDs calculated via Mordred for testing and optimization. Again, using LASSO, 21 MDs from Mordred (Supplementary Table 5) were used within the machine learning prediction model and gave a linear fit with R = 0.9806 for the training set (Fig. 3c and Supplementary Data 5). We reason that the unsatisfactory fits from both sets of MDs were due to the poor differentiation ability of sn-position isomers in the prediction models - only 15 out of 221 MDs from CDK and 500 out of 1824 MDs from Mordred software showed differences between sn-position isomers. To further improve the accuracy and precision of the prediction model and reinforce the influence of sn-position on the entire model, we constructed the prediction model with MDs from both CDK and Mordred. As initial input, a total of 253 MDs were selected. These included 32 MDs from Modred, which demonstrated > 0.1% difference between sn-position isomers (Supplementary Table 6), and all 221 MDs from CDK. Thirty-seven final MDs from the 253 MDs were selected using LASSO (Supplementary Table 7) to build an SVR-based prediction model, demonstrating excellent precision with R = 0.9922 in the internal validation (Fig. 3d and Supplementary Data 5). The external validation results also showed that the CCS prediction had a good linear fit with R = 0.9925 and a median relative error (MRE) of 0.34% (Fig. 3e and Supplementary Data 5). Beyond CCS prediction, a prediction model for RTs was also built and the predicted RTs were added to the extended library to improve the identification accuracy. Since sn-isomers almost have the same RT on a short time C18 gradient, MDs calculated by “rcdk” were used for retention time prediction. Using the LASSO algorithm, 51 MDs out of 221 MDs were selected and using an SVR-based machine learning algorithm to build the prediction model and gave a linear fit with R = 0.9622 for the training set (Supplementary Fig. 20 and Supplementary Data 6). Overall, the results demonstrated that the prediction model has an excellent capability for precisely predicting the physicochemical properties of sn-position-resolved GPs. Finally, using the optimized prediction model, CCS values and RTs of 2500 GPs with sn-position information were predicted – 870 PCs, 883 PEs, and 747 PSs (Supplementary Data 7). AD is a progressive neurodegenerative disorder with histopathological hallmarks of β-amyloid (Aβ) plaques and neurofibrillary tangles in the brain. Extensive studies have demonstrated that altered GP compositions and enzyme activities involved in the GP remodeling in the brain were associated with AD progression. Although many studies on GPs in the AD mouse brain have been reported, the detailed spatial distribution of GP sn-position across different brain regions and their functions remains unexplored. In this study, we used APPswe/PS1dE9 transgenic mice to investigate GP alteration in AD progression. The animals express mutant forms of APP (Mo/HuAPPswe with the K595N/M596L mutation) and PS1 (deletion of exon 9). They are characterized by an early-onset of AD and age-associated increase in Aβ-levels followed by Aβ deposition, morphological alterations, and are also widely used to investigate lipid alterations in AD. Extensive studies in patients with advanced AD have demonstrated that the hippocampus and cortex are more heavily affected by Aβ pathology than the cerebellum. Therefore, hippocampus, cortex, and cerebellum were chosen for our examination. Following extraction of GPs from both 3-month-old wild type (WT) and APPswe/PS1dE9 mice (simply referred to as AD mice thereafter) (Fig. 4a), we performed quantification analysis of GPs using the validated LC-HRdm IM-MS method in tandem with our experimental 4D library and extended CCS library (Fig. 4b; detailed schematic is shown in Supplementary Fig. 17). A total of 592 GPs were identified across the 3 regions, including 130 pairs of sn-position isomers (Supplementary Data 8). Specifically, 498 GPs identified by our experimental 4D library were classified as level 1 annotations in this context, and the 94 GPs identified from matching with the extended CCS library were classified as level 2 annotations in this context. The cortex was found to contain the greatest diversity of GPs with 571 GPs identified, while 551 and 547 GP species were identified in the hippocampus and cerebellum, respectively (Fig. 4c, Supplementary Data 8). Similar findings regarding the number of identified lipid species in the hippocampus and cerebellum have been reported in previous studies. In terms of the number of species identified in each GP class, PC and PE were dominant across all 3 regions, consistent with findings from previous studies. Further cross-regional analysis revealed the diversity of GP distributions among different brain regions with a total of 517 GPs identified in all three regions, though each region did reveal unique species (Fig. 4d).Fig. 4Spatial characterization of GPs in the AD and WT mouse brain.a Illustration of three functional regions (hippocampus (HP), cerebellum (CB), and cortex (CTX)) of mouse brains (age of 3 months; n = 3) selected for spatial characterization of GPs between WT and AD groups. Figure 4a was created with BioRender.com, released under a Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International license. b Graphic illustration of the workflow for GP identification from mouse brain. Figure 4b was created with BioRender.com, released under a Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International license c Numbers of GPs identified in the three brain regions. d Overlap of GPs identified from 3 brain regions. e Total abundance of 6 GP classes quantified in all brain regions of WT and AD groups. The GP abundance was normalized to wet tissue weight (mg). f HCA discrimination of the AD and WT group of 3 mouse brain functional regions by quantitative results of GPs at sn-position resolved level. g Percentages of GP classes in clusters 1–5. Abundance distributions of representative GPs in AD and WT mouse brain regions, PE 16:0/22:6 (p = 0.0004 for HP, 0.0001 for CB, and 0.0005 for CTX) (h), PE 18:0/22:6 (p = 0.0140 for HP, 0.0003 for CB) (i), LPE 16:0/0:0 (p = 1.4E-5 for HP, 7.0E-5 for CB, and 8.5E-5 for CTX) (j), LPE 18:0/0:0 (p = 0.0030 for HP, 0.0035 for CTX) (k). The ratio of sn-1 C16:0-containg PE/LPE 16:0/0:0 (p = 0.0031 for HP, 0.0472 for CB) (l) and sn-1 C18:0-containg PE/LPE 18:0/0:0 (p = 0.0007 for HP, 0.0007 for CB, and 0.0006 for CTX) (m) in AD and WT mouse brain regions. Data are presented as mean values +/−SD (n = 3, *p < 0.05, **p < 0.01, ***p < 0.001 (one-way ANOVA with correction for multiple comparisons using the two-stage linear step-up procedure of Benjamin, Krieger, and Yekutieli at a 0.05 FDR)). Source data are provided as a Source Data file. a Illustration of three functional regions (hippocampus (HP), cerebellum (CB), and cortex (CTX)) of mouse brains (age of 3 months; n = 3) selected for spatial characterization of GPs between WT and AD groups. Figure 4a was created with BioRender.com, released under a Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International license. b Graphic illustration of the workflow for GP identification from mouse brain. Figure 4b was created with BioRender.com, released under a Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International license c Numbers of GPs identified in the three brain regions. d Overlap of GPs identified from 3 brain regions. e Total abundance of 6 GP classes quantified in all brain regions of WT and AD groups. The GP abundance was normalized to wet tissue weight (mg). f HCA discrimination of the AD and WT group of 3 mouse brain functional regions by quantitative results of GPs at sn-position resolved level. g Percentages of GP classes in clusters 1–5. Abundance distributions of representative GPs in AD and WT mouse brain regions, PE 16:0/22:6 (p = 0.0004 for HP, 0.0001 for CB, and 0.0005 for CTX) (h), PE 18:0/22:6 (p = 0.0140 for HP, 0.0003 for CB) (i), LPE 16:0/0:0 (p = 1.4E-5 for HP, 7.0E-5 for CB, and 8.5E-5 for CTX) (j), LPE 18:0/0:0 (p = 0.0030 for HP, 0.0035 for CTX) (k). The ratio of sn-1 C16:0-containg PE/LPE 16:0/0:0 (p = 0.0031 for HP, 0.0472 for CB) (l) and sn-1 C18:0-containg PE/LPE 18:0/0:0 (p = 0.0007 for HP, 0.0007 for CB, and 0.0006 for CTX) (m) in AD and WT mouse brain regions. Data are presented as mean values +/−SD (n = 3, *p < 0.05, **p < 0.01, ***p < 0.001 (one-way ANOVA with correction for multiple comparisons using the two-stage linear step-up procedure of Benjamin, Krieger, and Yekutieli at a 0.05 FDR)). Source data are provided as a Source Data file. For the abundance distribution of GPs in each brain region, as shown in Fig. 4e, the total abundance of each GP class showed vast differences, ranging across 3 orders of magnitude in 3 brain regions. PE was the GP class of the highest abundance within each region with a minimum abundance of 29.30 ± 2.86 µmol/g in the AD mouse cortex and a maximum value of 49.12 ± 7.57 µmol/g in the WT mouse hippocampus. In contrast, PA was the GP class that demonstrated the lowest abundance in each region with a minimum abundance of 92.25 ± 10.13 nmol/g in the WT cortex and a maximum value of 324.70 ± 36.22 nmol/g in the WT hippocampus. It is worth noting that a decreased abundance of all GPs was observed for AD groups compared to WT groups in each of the 3 brain regions, indicating the disruption of GP homeostasis of AD which has also been reported by other studies. To comprehensively understand the AD-induced alteration of GPs within 3 distinct brain regions, abundances of sn-position-resolved GPs were used to perform hierarchical clustering analysis (HCA) (Fig. 4f), which successfully discriminated AD and WT groups within all three regions. The successful clustering of the three brain regions also revealed sn-position-defined GPs had spatial heterogeneity, suggesting their distinct functions. Within HCA, GPs could be grouped into five clusters (Fig. 4f, g, and Supplementary Data 9), among which Cluster 2 and Cluster 5 contained the greatest number of GP species across all six GP classes. In terms of GP numbers, PE class is the most dominant in both of these two clusters (Fig. 4g). When comparing proportion of molecular abundance to number of GP species in each GP class, an orthogonal perspective emerges on the percentage in each cluster (Supplementary Fig. 21). For instance, PC is dominant in terms of GP numbers (>50%) in Cluster 1, yet it is minor (~30%) in terms of molecule percentage. GPs in Cluster 2 were expressed at the highest levels in the WT hippocampus and the lowest level in the AD cerebellum. GPs in Cluster 5 exhibited very high levels in the WT cerebellum and the lowest levels in the AD hippocampus. Quantitative analysis across three regions revealed a marked disparity in the abundance of GPs in the hippocampus and cerebellum. Moreover, the observed changes in the abundance of individual GPs between AD and WT mice across the three regions indicate that GP profiling could serve as a sensitive marker for AD. The altered abundance of multiple prominent GPs demonstrated differences between WT and AD mice in the three brain regions, especially polyunsaturated fatty acids (PUFA)-containing GPs (Supplementary Data 10). Previous studies on fatty acid composition demonstrated a progressive decline of the omega-3 fatty acid, especially docosahexaenoic acid (DHA, C22:6), in late-stage AD brains. PE constitutes the major storage form of DHA in the brain and can facilitate the signal transduction of bioactive mediators. It has been reported that DHA-containing PE could protect PUFAs from oxidation in gray matter. DHA-containing PE, especially in the forms of PE 16:0/22:6 and PE 18:0/22:6, were also detected in high abundance in all three investigated brain regions. The abundance of PE 16:0/22:6 and PE 18:0/22:6 was significantly lower in AD compared to WT mice across all three regions except for PE 18:0/22:6 displaying no obvious difference in the cortex (Fig. 4h, i). Previous studies have reported a significant difference in the ratio of GP/lyso-GP between AD patients and healthy controls, with an accuracy of 82%-85% using this ratio as an additional biomarker for neuropathological diagnosis of AD. However, GP molecules investigated in those prior studies were not annotated with sn-position specificity, the limitation directly addressed in our investigations. LPE 16:0/0:0 and LPE 18:0/0:0 were taken as examples because GPs containing C16:0 and C18:0 fatty acyl at sn-1 are the most abundant. We observed that LPE 16:0/0:0 and LPE 18:0/0:0 decreased in AD compared to WT mice in all three regions (Fig. 4j, k). We further evaluated the ratio of sn-1 C16:0-containing PE/LPE 16:0/0:0 and sn-1 C18:0-containing PE/LPE 18:0/0:0 in all regions (Fig. 4l, m). Our data revealed the ratio of sn-1 16:0-containing PE/LPE 16:0/0:0 was only significantly elevated in the hippocampus, and the ratio of sn-1 C18:0-containing PE/LPE 18:0/0:0 was likewise increased in hippocampus and cortex. The increased levels of PE/LPE suggested that dysregulated lipid metabolism occurred in the AD mouse brain. Furthermore, the heterogeneous alterations indicated that enzymatic activity involved GP metabolism may not be conserved across all brain regions. We also observed spatial alterations of GP sn-position isomers between WT and AD brain regions (Supplementary Fig. 22). As a representative example, PS 18:1/20:4 showed similar abundance in WT and AD across 3 regions. However, PS 20:4/18:1 in hippocampus, but not the cortex and cerebellum, showed a significant decrease in AD compared to WT. As AD is commonly considered as an age-related neurodegenerative disease, unraveling the potential molecular mechanisms of AD progression is of vital importance. To better understand GP alteration in AD progression, we analyzed the GPs in the same three brain regions of the mice at the age of 3 and 8 months (Fig. 5a, Supplementary Fig. 23). Detailed analysis of GPs in the hippocampus is presented here as an example. Based on the abundance of sn-position-resolved GPs in the hippocampus, all samples were correctly classified by HCA into 4 groups, and each correctly matched on age and genotype (Fig. 5b and Supplementary Data 11). At the outset, these results indicated that the abundance of sn-position resolved GPs could sufficiently reflect neurodegeneration and could be used to classify normal and diseased samples. In particular, the clusters representing WT at 8 months and AD at 3 months were clustered adjacent to one another, further highlighting the age-related nature of AD. Our results demonstrate that mice of different ages and genotypes can be distinctly differentiated based on quantified GPs in the hippocampus (Fig. 5b, Supplementary Fig. 23a) and cortex (Supplementary Fig. 23b). However, the PCA of GPs in the cerebellum (Supplementary Fig. 23c) showed overlap between WT mice at 8 months and AD mice at 3 months, indicating that GP profiles in the hippocampus and cortex are more sensitive markers for monitoring age-related AD progression than those in the cerebellum. In order to provide a comprehensive comparison of the GP changes of different genotype and age in three mouse brain regions, we performed volcano plot analysis based on GP abundance to evaluate the alterations (Supplementary Fig. 24, Supplementary Data 12). Substantial numbers of GP species exhibited a significant decrease in AD compared to WT at the age of 3 months in all 3 regions. Among the large number of significantly decreased GPs in AD, the obvious decrease of many PUFA-containing GPs in all 3 regions were noted. A more detailed analysis was also performed to elucidate the age‑dependent changes in PUFA-containing GP abundance between WT and AD groups. Notably, compared to the WT 3-month group, most identified GPs with C20:4 and DHA at sn-2 decreased in the AD 3-month group and the WT 8-month group (Fig. 5c–l). The abundances of many GPs in AD both at 3 and 8 months were close to WT at the 8 months, indicating that our GP profiling could also reflect the fact that AD is an age-related disease. These findings still hold true in the cerebellum and cortex (Supplementary Figs. 25–28). This alteration trend was consistent with previous reports that PUFA compositions were decreased in aging and AD.Fig. 5Age-associated temporal diversity of GPs in the mouse brain.a Illustration of three functional regions of mouse brains (age of 3 months and 8 months; n = 3) selected for investigation of temporal diversity of GPs in aging and AD progression. Figure 5a was created with BioRender.com, released under a Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International license. b HCA of GPs in the hippocampus. Compositional variations of GPs with C20:4 at sn-2 position (c–h), PC 16:0/20:4 (p (WT 3 vs. AD 3) = 5.6E-5, p (WT 3 vs. WT 8) = 5.6E-5) (c), PE O-16:0/20:4 (p (WT 3 vs. AD 3) = 1.9E-5, p (WT 3 vs. WT 8) = 1.2E-4, and p (WT 8 vs. AD 8) = 0.0027) (d), PS 16:0/20:4 (p (WT 3 vs. AD 3) = 0.0068, p (WT 3 vs. WT 8) = 0.0036) (e), PC 18:0/20:4 (p (WT 3 vs. AD 3) = 0.0095) (f), PE 18:0/20:4 (p (WT 3 vs. AD 3) = 0.0008, p (WT 3 vs. WT 8) = 0.0018, and p (WT 3 vs. WT 8) = 0.0157) (g), PS 18:0/20:4 (p (WT 3 vs. AD 3) = 0.0026, p (WT 3 vs. WT 8) = 0.0211) (h). Compositional variations of GPs with DHA at sn-2 position (i–l) among the AD and WT mouse at 3 and 8-month age in hippocampus, PE 18:0/22:6 (p (WT 3 vs. AD 3) = 0.0330) (i), PS 18:0/22:6 (p (WT 3 vs. AD 3) = 0.0067, p (WT 3 vs. WT 8) = 0.0051) (j), PG 18:0/22:6 (p (WT 3 vs. AD 3) = 0.0133) (k), PI 18:0/22:6 (p (WT 3 vs. AD 3) = 0.0027, p (WT 3 vs. WT 8) = 0.0032) (l). m The sn-isomer ratios for representative GPs among the AD and WT mice at 3 and 8 months (p values are shown in graphs). n Biosynthetic pathways of glycerophospholipids, including de novo synthesis and the fatty acyl remodeling of GPs. The ratio variations of representative GP/lyso-GP among the AD and WT mouse at 3 and 8-month age in hippocampus, PC 18:0/X / LPC 18:0/0:0 (p (WT 3 vs. AD 3) = 0.0014, p (WT 3 vs. WT 8) = 0.0018) (o), PE 16:0/X / LPE 16:0/0:0 (p (WT 3 vs. AD 3) = 0.0318) (p), PE 18:0/X / LPE 18:0/0:0 (p (WT 3 vs. AD 3) = 0.0070, p (WT 3 vs. WT 8) = 0.0176) (q). Data are presented as mean values +/−SD (n = 3, *p < 0.05, **p < 0.01, ***p < 0.001 (one-way ANOVA with correction for multiple comparisons using the two-stage linear step-up procedure of Benjamin, Krieger, and Yekutieli at a 0.05 FDR). Source data are provided as a Source Data file. a Illustration of three functional regions of mouse brains (age of 3 months and 8 months; n = 3) selected for investigation of temporal diversity of GPs in aging and AD progression. Figure 5a was created with BioRender.com, released under a Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International license. b HCA of GPs in the hippocampus. Compositional variations of GPs with C20:4 at sn-2 position (c–h), PC 16:0/20:4 (p (WT 3 vs. AD 3) = 5.6E-5, p (WT 3 vs. WT 8) = 5.6E-5) (c), PE O-16:0/20:4 (p (WT 3 vs. AD 3) = 1.9E-5, p (WT 3 vs. WT 8) = 1.2E-4, and p (WT 8 vs. AD 8) = 0.0027) (d), PS 16:0/20:4 (p (WT 3 vs. AD 3) = 0.0068, p (WT 3 vs. WT 8) = 0.0036) (e), PC 18:0/20:4 (p (WT 3 vs. AD 3) = 0.0095) (f), PE 18:0/20:4 (p (WT 3 vs. AD 3) = 0.0008, p (WT 3 vs. WT 8) = 0.0018, and p (WT 3 vs. WT 8) = 0.0157) (g), PS 18:0/20:4 (p (WT 3 vs. AD 3) = 0.0026, p (WT 3 vs. WT 8) = 0.0211) (h). Compositional variations of GPs with DHA at sn-2 position (i–l) among the AD and WT mouse at 3 and 8-month age in hippocampus, PE 18:0/22:6 (p (WT 3 vs. AD 3) = 0.0330) (i), PS 18:0/22:6 (p (WT 3 vs. AD 3) = 0.0067, p (WT 3 vs. WT 8) = 0.0051) (j), PG 18:0/22:6 (p (WT 3 vs. AD 3) = 0.0133) (k), PI 18:0/22:6 (p (WT 3 vs. AD 3) = 0.0027, p (WT 3 vs. WT 8) = 0.0032) (l). m The sn-isomer ratios for representative GPs among the AD and WT mice at 3 and 8 months (p values are shown in graphs). n Biosynthetic pathways of glycerophospholipids, including de novo synthesis and the fatty acyl remodeling of GPs. The ratio variations of representative GP/lyso-GP among the AD and WT mouse at 3 and 8-month age in hippocampus, PC 18:0/X / LPC 18:0/0:0 (p (WT 3 vs. AD 3) = 0.0014, p (WT 3 vs. WT 8) = 0.0018) (o), PE 16:0/X / LPE 16:0/0:0 (p (WT 3 vs. AD 3) = 0.0318) (p), PE 18:0/X / LPE 18:0/0:0 (p (WT 3 vs. AD 3) = 0.0070, p (WT 3 vs. WT 8) = 0.0176) (q). Data are presented as mean values +/−SD (n = 3, *p < 0.05, **p < 0.01, ***p < 0.001 (one-way ANOVA with correction for multiple comparisons using the two-stage linear step-up procedure of Benjamin, Krieger, and Yekutieli at a 0.05 FDR). Source data are provided as a Source Data file. Subsequently, although most GPs were downregulated in AD groups, including the simultaneous decrease of many pairs of GP isomers, illustrating the differences in their ratio of change would facilitate the more detailed interpolation of the underlying mechanism. In terms of GP sn-position isomer ratios, definitive alteration of multiple GPs was observed among the 4 different mouse groups (Fig. 5m), marking the importance of this overlooked field in previous lipidomic studies. At the sn-position isomer level, the ratio of some sn-isomer pairs also showed significant alterations in AD groups compared to WT groups both at 3 months and 8 months. For example, the ratios of LPC 18:0/0:0 to LPC 0:0/18:0, were significantly elevated in AD mice compared to WT mice at 3 months in hippocampus. A portion of lyso-GP is the degradation product of GP catalyzed by phospholipase A (PLA). The increased ratio of LPC (18:0/0:0)/LPC (0:0/18:0) suggested that the sn-2 PUFA-containing GPs were more degraded to produce LPC 18:0/0:0 than their corresponding isomers. Thus, the significant deviation in ratios of sn-isomers was potentially correlated with the increased activities of PLA2, which was responsible for cleaving the acyl groups at the sn-2 position of GPs and resulting in the production of lyso-GPs and free fatty acids (Fig. 5n). To validate this finding, we further evaluated the cytosolic phospholipase A2 (cPLA2) activities in WT and AD mice across three brain regions as cPLA2 is one of the major PLA2 in the brain. Congruous with the previous reports, this assay demonstrated significantly increased cPLA2 activity in AD compared to WT at 3 months in the hippocampus, which was also consistent with the results obtained through our LC-IM-MS/MS strategy (Supplementary Fig. 29). On the contrary, the ratios of some 1,2-diacyl GP isomer pairs, including PC 18:0/20:4, PE 18:0/20:1, PS 18:0/22:4 to their sn-isomers, were significantly elevated in 3-month-old AD mice compared to 3-month-old WT mice. The results indicated that, although increased cPLA2 activities in early AD may enhance the degradation of sn-2 PUFA-containing 1,2-diacyl GPs, the sn-1 PUFA-containing 1,2-diacyl GPs were actually more degraded than their corresponding isomers in early AD. Moreover, most of these ratios also showed significant changes at 8 months. Together, these data suggested that, in the complex lipid metabolism processes of AD and aging mouse brain, the activities of lipid metabolic-related enzymes, besides cPLA2, were likely also altered and might require a more systematic evaluation. In addition, we also mapped altered ratios of GP/lyso-GP for multiple species (Fig. 5o–q). The ratio of sn-1 C18:0-containing PC/LPC 18:0/0:0 was significantly decreased in the AD hippocampus at 8 months (Fig. 5o). The results are consistent with a previous study showing that the decreased ratio of PC/lyso-PC in AD might be highly useful as a potential plasma biomarker for the diagnosis of early dementia. However, the ratio of PE/LPE has rarely been reported. We found the ratios of sn-1 C16:0-containing PE/LPE 16:0/0:0 (Fig. 5p), and sn-1 C18:0-containing PE/LPE 18:0/0:0 (Fig. 5q) were significantly higher in 3-month AD compared to 3-month WT in the hippocampus. The ratio of sn-1 C18:0-containing PE/LPE 18:0/0:0 was also significantly increased in the cortex of 3-month-old WT mice compared to that of 3-month-old AD mice, but this difference was not observed in the cerebellum. However, a significant increase in this ratio could be observed in the cerebellum of older, 8-month-old mice (Supplementary Fig. 26, 28). The cPLA2 activities in WT and AD mice across 3 brain regions were also varied (Supplementary Fig. 29). The results, together with previous studies, suggested the potential heterogeneous enzyme activities in different ages and functional brain regions of AD. Taken together, our results detail the spatial and temporal diversity of GPs in the mouse brain and highlight putative molecular signatures of aging and AD progression. Lipids play important roles in cell signaling, cell structure, and energy storage, however detailed lipid structures remain largely underexplored in routine lipidomics owing to its complexity. Over the past decade, significant research efforts have been devoted to developing analytical methods for the detailed structural elucidation of lipids sn-position, C═C location and geometry isomers. Recently, IM-MS has emerged as a promising technology for lipidomics to facilitate the separation and identification of complex mixtures of analytes. Studies have demonstrated the power of IM-MS for enhancing identification accuracy of lipidome in complex biological samples. Moreover, studies have coupled LC-IM-MS/MS analysis and rule-based refinement for improving accuracy of lipid identification and achieving the partial annotation of GPs with the main composition of sn-position isomers. Furthermore, recent research has highlighted the enormous potential of high-resolution IM-MS for in-depth lipid isomer separation. Nevertheless, to the best of our knowledge, the use of label free LC-IM-MS/MS for large-scale GP sn-isomers identification and quantification analysis in biological samples has not been widely explored. In this work, we develop an LC-HRdm IM-MS/MS-based 4D lipidome profiling method that allows large-scale and sn-position-resolved GP identification and quantification in complex biological samples. With advanced multiplexed ion injection, both increased sensitivity and enhanced IM Rp up to 250 have been achieved following HRdm data processing. These improvements enable efficient separation of co-eluting GP sn-position isomers (~1% CCS differences) in the IM dimension at large-scale without sacrificing throughput or mobility coverage. Together with acyl chain information obtained from MS/MS analysis, our strategy provides the ability to identify GPs from the fatty acyl composition level (e.g., PC 16:0_18:1) to the sn-position level (e.g., PC 16:0/18:1 and/or PC 18:1/16:0). This strategy makes it possible for GP identification through intact analysis without further metal adduction, enabling highly confident quantification. In addition to developing a strategy for in-depth GP analysis, we isolated GPs from mouse brains and compiled the empirical findings into an sn-position-resolved GP 4D database to benefit the broader community with an automated GP analysis pipeline that does not require prior knowledge or tedious manual curation. The 4D database contains a total of 498 GPs with acyl chain sn-position information which were identified from the LC-HRdm IM-QTOF platform. These results represent more than 50% increase in the number of GP identification, which is 318 without sn-position assignment obtained through LC-QTOF (IM off), and 326 using 4D library-based match and rule-based refinement without additional HRdm strategy. As accurate and reliable CCS values serve as the key for the identification of GP sn-position isomers, we used a machine learning-based prediction approach to construct an extended 4D library comprised of 2500 GPs for sn-position isomer identification. Our extended 4D library, with the high prediction accuracy (R > 0.992) of the CCS values, is more in-depth than other currently available predicted CCS libraries which do not differentiate sn-position isomers. Many studies have demonstrated significant changes of GP in AD and aging mouse brain. However, these studies only revealed GP alteration at fatty acyl sum composition or fatty acyl level across different brain regions with aging, while the alteration of detailed GPs at sn-position resolved level remains unexplored. With both the experimental database and the extended library developed in this study, we investigated the spatial and temporal GP alterations in the brains from a mouse model of AD. Through our analyses, we found that GPs profiling at sn-position-resolved level allowed for correct clustering of spatial, age, and genotype-matched samples. In addition, significant changes in either abundance or ratios of sn-isomers in a set of GPs over aging and AD progression have been revealed. In this study, our research demonstrated a significant decrease in the total abundance of all 6 GP classes in AD mice across all three functional brain regions compared to WT at 3 months of age, indicating early temporal aberrant GP homeostasis occurred in all 3 functional regions of our AD model mice. This is in line with a body of evidence linking GP depletion with AD. Moreover, notable reduction of GP with PUFAs (e.g., DHA), at the sn-2 position was observed in AD and aged mice across all 3 brain regions. Therefore, there seems to be a temporal and genotype-dependent shift in the progression of changes. Herein, our results are consistent with previous studies and suggest that a drastic decrease in DHA-containing GPs might be an indicator of AD and aging. Obviously, it is often difficult to claim biological implications of MS-based findings and further studies are required to clarify the currently unclear molecular mechanisms of GP metabolism with aging and AD. However, it is tempting to connect our present data with independent studies that link depletion of DHA-containing GP to changes in membrane fluidity and flexibility as well as regulation of neuroinflammation. In addition to reveal significantly altered GPs in AD and aging mice, we found notable alterations in relative abundance ratios of GP sn-isomeric pairs as well as GP/lyso-GPs, and their potential correlation to related enzyme activities in GP remodeling pathways. In this study, with the capability to discriminate sn-position, the ratio of sn-isomer pairs showed significant alterations between WT and AD groups (Fig. 5m). Intriguingly, although many pairs of the sn-position isomer simultaneously decreased in AD and aged mice, more significant changes in the ratio of many sn-isomer pairs were observed in AD genotype than aging. Therefore, the ratio of sn-isomer pairs may show distinct associations with AD pathogenesis. In addition, our study revealed a significant elevation of the ratios of LPC 18:0/0:0 to LPC 0:0/18:0 in AD mice compared to WT mice at 3 months in hippocampus. These changes might be linked to lipid biosynthetic pathways of GPs, which comprise two components: the de novo pathway and the remodeling pathway (Fig. 5n). In the de novo pathway, PA is initially synthesized from glycerol-3-phosphate (G3P) and then converted to diacylglycerol (DG) or cytidine diphosphate-DG (CDP-DG), which will be ultimately converted into cardiolipin (CL), PG, PI, and PE, PC, PS, respectively. Subsequently, in the remodeling pathway, GP acyl chains are remodeled by orchestrated reactions of PLAs, acyl-CoA synthases, acyltransferase, and lysophospholipid acyltransferases (LPLATs). The increase of the ratios of LPC 18:0/0:0 to LPC 0:0/18:0 suggested that the sn-2 PUFA-containing GPs might be more degraded than their corresponding isomers, potentially correlating with the increased PLA2 activities. We further verified that the cPLA2 activities in AD mice at 3 months were remarkedly higher compared to WT mice at 3 months in hippocampus (Supplementary Fig. 29). This finding is also aligned with existing literature reporting increased cPLA2 activity in AD brains. Interestingly, we observed a significant increase of the ratio of sn-2 PUFA-containing 1,2-diacyl GPs to their corresponding isomers in AD mice compared to WT mice, and most of the ratios also showed significant changes at 8 months, indicating that the sn-1 PUFA-containing 1,2-diacyl GPs might be more degraded than their corresponding isomers in AD as the abundance of both sn-isomers decreased. Previous reports also demonstrated that the activities of various enzymes were significantly changed in AD patients, including acyltransferase, phosphodiesterase, LPLATs, and other enzymes, which were involved in the remodeling process of acylating lyso-GP into 1,2-diacyl GP to maintain the composition of neuronal membranes. The activities of these enzymes were likely also altered in the complex lipid metabolism processes of AD mouse brains considering both increased cPLA2 activities and the elevated ratio of sn-2 PUFA-containing 1,2-diacyl GPs to their corresponding isomers in AD groups. Moreover, we found that the ratios of GP/lyso-GP across different GP classes showed varied alteration trends between WT and AD mouse brains. Therefore, it is possible that some other enzymes involved in the lipid metabolic pathway of AD mouse brains were also altered. The exact molecular mechanisms underlying aging and AD disease pathologies correlated with lipid metabolism warrant comprehensive research. Overall, the present results, together with previous findings from literature, suggested that it is important to investigate the dysregulated lipid metabolism in AD pathology at GP sn-position resolved level. Beyond the sn-position resolved GP quantification results, orthogonal enzyme activity assay also disclosed altered cPLA2 activity in different brain regions in aging and AD progression, suggesting potential alteration of enzyme activities in the GP remodeling pathways. We recognize that in the absence of clear mechanistic validation, our MS data should not be overinterpreted. However, the fact that they are consistent with the above-mentioned independent studies suggests that the MS-based results provide promising candidates of GP isomers for AD pathology and a more systematic investigation of GP-related pathways in AD is warranted. In summary, we developed a 4D lipidomics strategy that enables large-scale, label-free, robust, and rapid GP profiling at sn-position resolved level without additional derivatization and instrument modification. Delving deeper, we have constructed a pioneering and valuable 4D database with 498 GPs at sn-position resolved level, sourced from both standards and mouse brain lipid extracts. This was complemented by an expansive 4D library of 2500 GPs, constructed via a robust machine learning-based predictive approach. This database and library would be groundbreaking contributions to the field and are instrumental for researchers in biology and analytical sciences, offering valuable resources for detailed lipid structural identification. Our study not only delves into the spatial and temporal diversity of GPs in the mouse brain but also highlights potential molecular signatures of AD progression. This might offer deeper insights into disease pathology associated with dysregulated lipid metabolism and aid in uncovering potential lipid biomarkers for diseases. However, several limitations exist in our study. Firstly, for the very low abundance isomers of GPs with complicated fatty acyl compositions, although they might be identified in our workflow from MS/MS spectra in negative mode, they were not annotated from the HRdm drift spectra to enhance the identification accuracy and confidence. This limitation would be addressed by employing high-sensitivity nano-ESI-MS and next-generation IM-MS with higher resolution. The challenges of annotating GPs with multiple major fatty acyl compositional isomers could potentially be overcome by combining high-resolution IM-MS with advanced chromatographic separation, including extended chromatographic gradient and multidimensional chromatographic separation. Secondly, lipid C═C positional isomers could not be differentiated on this system since it requires IM Rp ~ 1000 for baseline separation. Incorporating additional sample treatment steps such as cutting-edge derivatization/fragmentation strategies to pinpoint C═C information will ultimately achieve complete in-depth structural profiling of the lipidome. Thirdly, in our CCS prediction model, although small variances are still inevitably present between the experimental and predicted CCS values, we reason that involving more stereospecific 3D MDs would provide a more accurate prediction model. Additionally, we envision that a joint effort from the whole lipidomics community to expand the number of empirical measurements within the CCS database will be an invaluable addition. Another limitation of our study is that, although we performed orthogonal enzyme activity analysis to facilitate further exploration of the lipidomics data, it did not comprehensively cover all possible related enzymes. For example, we tested cPLA2 activities, but it is crucial to note that, besides cPLA2, various other isoforms of PLA2 including secretory phospholipase A2 (sPLA2), calcium-independent phospholipase A2 (iPLA2), also exist in brain regions. Given the diversity of PLA2 subtypes and additional enzymes, such as LPLATs, phospholipase A1s, acyl-CoA synthases, and transacylases, involved in lipid remodeling, further investigation into the changes in the activities of multiple enzymes is necessary in the context of AD. Nevertheless, our data provides insights into possible enzyme activity alterations in aging and AD progression. It is worth mentioning that although we tried to minimize postmortem lipid changes that may occur in the brain by minimizing the euthanasia to tissue collection time, the possible oxidation of GPs is unavoidable. Of additional note, it remains to be seen if the dynamics of bioactive lipid mediators, including eicosanoids and resolvins which are potentially evoked by the release of PUFAs from GPs, could offer deeper insights for early diagnosis and therapeutic strategies. Additionally, we plan to utilize imaging mass spectrometry in conjunction with high-resolution ion mobility, along with our established database, to map the spatial distribution of sn-position resolved GPs in a region-specific manner across the entire brain. Collective studies may be possible to provide a deeper perspective for early diagnosis, and preventive or therapeutic options for AD. In the future, together with biological validation, our strategy could serve as an enabling tool not only providing in-depth lipid structural characterization but also sensitively monitoring the differential expression of enzymes involved in GP remodeling to ultimately provide crucial mechanistic insights into many disease pathologies. APP695/swe/PS1-dE9 (APP/PS1) double transgenic male mice were obtained from Jackson Laboratory (MMRRC Stock No. 34832-JAX). Genotyping from tail DNA was performed at weaning by Transnetyx (Cordova, TN). APP/PS1 mice were studied at the age of three months (n = 3) and eight months (n = 3). Male mice were studied with wild‐type (WT) littermates used as controls. WT mice were also studied at the age of three months (n = 3) and eight months (n = 3). All male mice were housed in standard cages provided by the University Laboratory Animal Resources and grouped with littermates with 1–5 mice per cage. Sex and gender-based analysis is not considered in this study. Animals were housed in facilities with a standard 12-hour light-dark cycle, humidity of 50% at 24 °C, and were provided standard chow and water ad libitum. Animal experiments were performed in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals and were approved by the Institutional Animal Care and Use Committee of the University of Wisconsin-Madison (protocol #M005120). For tissue collection, mice were euthanized per protocol in a CO2 chamber, and brains were extracted from the skull and rinsed in ice-cold 1X PBS solution. The brains were then dissected to separate out the hippocampus, cerebellum, and cortex according to the established protocol, the details were shown in Supplementary Fig. 30. The tissue was directly frozen in liquid nitrogen and kept at −80 °C before further experimentation. Lipid standards were purchased from Avanti Polar Lipids (Alabama, USA). ESI low-concentration tune mix was obtained from Agilent Technologies (Santa Clara, CA). Phospholipase A2 from porcine pancreas (P6534, ≥ 600 U/mg) was purchased from Aladdin (St. Louis, MO). All other chemicals (LC-MS-grade) were purchased from Fisher Scientific (NJ, USA) and used without further purification. Lipid standard stock solutions were prepared in chloroform: methanol (v/v, 1:1). A modified Folch method was employed for lipid extraction from brain hippocampus, cerebellum, and cortex samples. Approximately 10 mg of brain tissue from wild-type (WT) and APP/PS1 (AD) mice were weighed and dissolved in 0.5 mL of cold PBS. Each sample was spiked with 3 μL of EquiSPLASH (100 µg/ml of each isotope-labeled lipid standard) before lipid extraction. Brain tissues were homogenized with a probe sonicator in an ice water bath with a pulse of 10 s on and 10 s using 60 W energy for 20 cycles and subsequently mixed with 0.5 mL cold methanol and 1 mL cold chloroform for liquid–liquid extraction. After 10 min of vortexing, the mixture was centrifuged at 4 °C for 15 min at 10, 000 × g followed by collecting the bottom layer of chloroform. The above liquid–liquid extraction procedure was repeated twice. The chloroform layers from the three extractions were then combined and evaporated to dryness in a vacuum concentrator. The extracted samples were stored at −20 °C prior to MS analysis. The PLA2 digestion was performed using the reported protocol with some modifications. Briefly, 10 nmol dried GP standard was resuspended in a mixture of 480 μL PBS buffer and 10 μL 100 mM CaCl2. The mixture was thoroughly vortexed for 2 min. Subsequently, 10 μL of the enzyme solution was added. The resulting mixture was vortexed for 2 min, then incubated at 37 °C for 4 h. Lipids were extracted by adding 500 μL of chloroform and 500 μL of methanol. The mixture was again vortexed for 2 min and centrifuged at 10, 000 × g at 4 °C for 5 min. The bottom phase was collected and evaporated to dryness in a vacuum concentrator. Dried samples were reconstituted in methanol for LC-IM-MS/MS analysis. cPLA2 activity in brain tissue was determined using the cytosolic phospholipase A2 assay kit from Abcam (ab133090) according to manufacturer’s instructions. Briefly, approximately 5 mg of mouse brain tissue were homogenized in 200 μL HEPES/EDTA cold buffer (50 mM HEPES, pH 7.4, 1 mM EDTA) with a probe sonicator in an ice water bath. After centrifugation at 10,000 × g for 15 min at 4 °C, 200 μl of each supernatant was concentrated using 10 K molecular weight cut-off concentrators (Thermo Fisher Scientific, San Jose, CA). sPLA2 and iPLA2 have been removed by membrane filter and inhibited by bromoenol lactone respectively according to manufacturer’s instructions. Each sample (10 μl) was used to determine cPLA2 activity after 60 min of reaction. The final absorbance at 414 nm was read using a plate reader. The cPLA2 activity was normalized by protein concentration. Protein concentrations were measured using a BCA protein assay kit (Thermo Fisher Scientific, San Jose, CA). The LC-IM-MS/MS analyses were performed using an Agilent 1290 UHPLC system coupled with Agilent 6560 IM-QTOF platform (Agilent, Santa Clara, CA). LC separations were performed on a C18 column (Phenomenex Kinetex, 150 mm × 2.1 mm, 1.7 µm particle size) with column temperature maintained at 50 °C. The LC separation was performed with mobile phases A of 5 mM ammonium acetate in methanol/acetonitrile/water (1:1:1, v/v/v) and mobile phase B of 5 mM ammonium acetate in isopropanol/acetonitrile (5:1, v/v) at a flow rate of 0.3 mL/min. The 25 min elution gradient was set as follows: 0–0.75 min: 20% B; 0.75–2.5 min: 20% B to 40% B; 2.5–4.5 min, 40% B to 60% B; 4.5–19.5 min: 60% B to 98% B; 19.5–21 min: 98% B to 20% B; 21–25 min: 20% B. The sample was maintained at 4 °C during the whole analysis. The MS parameters for both positive and negative ESI modes were set as follows: mass range, 50−1700; sheath gas temperature, 350 °C; sheath gas flow, 12 L/min; dry gas temperature, 325 °C; drying gas flow, 8 L/min; nebulizer pressure, 35 psi; capillary voltage, 3500 V. The maximum drift time was set as 60 ms and the scan rate was set to 0.9 frames per second. Trap filling and trap release times were set as 1800 µs, and 200 µs, respectively. Pulsing sequence length was set as 5-bit. The pressure of the drift tube was set at 3.95 Torr. Capillary voltage was set at 3500 V, or −3500 V in positive and negative modes, respectively. The “Alternating frames” mode of IM-MS was used for the data-independent MS/MS acquisition. The collision energy in frame 2 was set as 25 V in positive mode or 30 V in negative mode. The entrance and exit voltages of the drift tube were set as 1500 V and 250 V, respectively. The “Auto MS/MS” of QTOF-only mode was used for the data-dependent MS/MS acquisition. All data acquisitions were carried out using Mass Hunter Workstation Data Acquisition Software (v 11.0). Multiplexing acquisition mode was utilized to achieve the high resolving power of DTIMS. Before HRdm and CCS analyses, multiplexed raw data files were firstly demultiplexed using the PNNL PreProcessor (v 4.0 build 2022.02.17, https://omics.pnl.gov/software/pnnl-preprocessor). Demultiplexing was performed using the data interpolation feature to facilitate the inherent reduction of data points across the drift peak observed with HRdm analyses. Parameters for data deconvolution were set as follows, interpolate drift bins: 1 drift bin becomes 3 drift bins; Demultiplexing: chromatography/infusion; (moving) average: 3; Signal Intensity lower threshold: counts 20; Spike Removal: Require 1 adjacent points per dimension. Following demultiplexing, feature finding was performed on MassHunter IM-MS Browser (v 10.0, Agilent) using the ion mobility feature extraction (IMFE) algorithm. Parameters used for IMFE were listed as follows, Processing: Chromatography; isotopic model: Common organic molecules; Limit charge state: z < =2; Ion intensity >=100. HRdm processing was then performed using High-Resolution Demultiplexer (v 2.0, Agilent), which performs a post-processing enhancement to the IM Rp. The detailed settings are listed as followers, HR processing level: High; m/z width multiplier: 6; IF multiplier: 1. CCS values were subsequently determined using single-field CCS method in MassHunter IM-MS Browser with the Agilent tune mix, which has been described in detail previously. In this study, the CCS value obtained for each isomeric lipid ion was put together with its m/z, retention time (RT), and fragment ions and then imported as a transition list into Skyline-daily (MacCoss Lab Software, v. 22.2.1), which was subsequently used for 4D lipid library construction and peak interpolation and integration for the LC-IM-MS/MS analyses. Initially, we determine the protonation state of the gas-phase molecule and generate 500 different conformer structures using the RDKit toolkit. Each conformer is then optimized with the Universal Force Field (UFF) and divided into several clusters for structurally distinct conformations. Potential structures of compounds are generated from the clusters and are optimized in Gaussian 16, using B3LYP Density Functional Theory (DFT) and the 3-21 G basis set. The 6-31 G + (d) basis set is then used to generate the energy-minimized structures. Theoretical CCS values for the compounds are calculated from their DFT-optimized geometries using IMoS (v. 1.13) software, employing the trajectory method (TM) models with 3 orientations and 300000 nitrogen gas molecules per orientation. Other settings in IMoS, including the use of nitrogen as the drift gas, are kept default. We used the LC-IM-MS/MS method to acquire the m/z, RT, CCS, and MS/MS spectra of lipid extract from pooled mouse brain samples. The GP library was developed within Skyline software according to the previous report with light modifications. The workflow for LC-high resolution-IM-MS/MS-based 4D sn-resolved GP library construction was shown in Supplementary Fig. 31. The schematics of the GP identification in as shown in Supplementary Fig. 17b. Lipid subclass information was obtained from the distinct head group related fragmentation ions in positive mode. The fatty acyl composition for each lipid species was obtained from LC–MS/MS (DDA) in negative ion mode. The MS/MS spectra acquired under DDA mode in negative mode were subsequently assigned to the IM peaks processed by HRdm. The transition lists were generated manually or using LipidCreator and MS DIAL (v 4.80). MS/MS spectra of mouse brain samples were validated by “Auto MS/MS” (data-dependent acquisition, DDA) at CE = 25 V in positive and CE = 30 V in negative mode. The library was populated with lipid class, name, RT, precursor formula, precursor m/z, adducts, product and neutral loss formulas, adducts, precursor drift time, precursor CCS, and high-energy drift offset values. All lipids in the library were manually validated according to the quality control requirements and met the confidence criteria of presence in more than 5 pooled sample runs, ± 5 ppm mass measurement accuracy, 2 min retention window, and 0.3% CCS deviation. iRT calculators were built using 10 isotope-labeled lipid standards from EquiSPLASH spanning the LC gradient. In addition, GP standards from 6 classes were used to measure their RT and CCS to validate the accuracy of this method (Supplementary Table 1). In total, 498 GP species were incorporated into the Experimental GP library including calculated m/z, RTs, CCS values, and MS/MS spectra. To develop a machine learning-based prediction model of GP CCS values and retention times, we used 419 GPs from three GP classes (205 PCs, 173 PEs, and 41 PSs) from the experimental GP database and randomly divided them into a training data set (80%; n = 335) and an external validation data set (20%; n = 84). Since the SMILESs of lipids in LIPID MAPS do not contain the adducts, in this study, SMILES of adducted GPs were manually generated by combing the SMILES of head groups, adducts, fatty acyl chains in sn-1 and sn-2 in a uniformed format. The protonation sites are determined to be the quaternary amine for PC, and the primary amine for PE and PS. For each GP, 221 MDs were first calculated by R package “rcdk” (R version 4.2.1) using the SMILES structure. For CCS prediction, another 32 MDs, which demonstrated more than 0.1% differences between sn-position isomers, were selected from 1824 MDs calculated by Mordred software and added to each GP. Next, MDs were optimized using the LASSO algorithm (R package “glmnet”) in the training set. 10-fold cross-validation was used to optimize the “lambda” value in LASSO, and the “lambda” with the lowest mean squared error (mse) was selected as the best model. From this optimization, 37 MDs were selected for the final CCS prediction. Employing the same methods, 51 MDs from “rcdk” were selected and used for retention time prediction. SVR algorithm (R package “e1071”) was used to build the prediction model. Two parameters, cost of constraints violation (C) and gamma (γ) were first optimized from 168 parameter combinations via 10-fold cross-validation with 100 repeats to achieve the best prediction performance. The optimized parameters were optimized as follows: C: 36, γ: 0.001 for CCS prediction, and C: 8, γ: 0.001 for retention time prediction, respectively. In the end, the SVR-based CCS and retention time prediction were further validated using the external validation data set. To supplement the experimental GP library, we developed an extended sn-position resolved GP library. Each GP class was considered in combination with the 35 most common fatty acyl chains in mammals; all unsaturated FAs were considered in their most common C═C double bond composition since C═C double bond isomers are indistinguishable in our strategy. We calculated the MDs and predicted the CCS values of GPs with the C═C bonds in the cis (Z) configuration, as this configuration is the most dominant in naturally occurring lipids in mammals. The 35 most common fatty acids are listed in Supplementary Table 3. Among all 2919 GPs in the class of PC, PE, and PS in LIPID MAPS, 419 GPs were already incorporated into the experimental GP library of mouse brain extract. Consequently, the rest 2500 GPs were used for prediction. Precursor m/z values of the 2500 GPs were calculated using the monoisotopic formula of [M + H]. Their MS/MS fragmentation ions were primarily generated and modified by using LipidCreator. Their CCS and RT values were generated by the above-established SVR-based prediction model. The transition list of this extended library was built with the predicted CCS value, RT, precursor m/z, and fragment ions of each GP to be compatible with automated analysis by Skyline. The extended GP library was provided in Supplementary Data 7. MS and CCS values were single-field calibrated using the Agilent tune mix under the same IM-MS parameters as other data files. After HRdm deconvolution, the CCS calibration coefficients (TFix and Beta) of calibrants were calculated in IM-MS Browser and applied to data files in the same batch of experiments. The GPs in the experimental data, after CCS calibration, were annotated using Skyline with the experimental mouse brain GP library and extended GP library. We set the minimum peak height at 3000 counts in MS1 as a criterion for identification. The MS1 match tolerance was set as 20 ppm. Retention time tolerance was set as 2 min window. Drift time filtering was used with a resolving power of 160. All annotations were manually validated. The semi-quantification of a class of lipid molecular species was achieved through the normalization of the individual molecular ion peak area to its respective deuterium-labeled internal standard. Statistical analyses were conducted in R (v.4.2.1), Origin 2020, and GraphPad Prism 9.0. Abundances of GPs were compared using one-way ANOVA with correction for multiple comparisons using the two-stage linear step-up procedure of Benjamin, Krieger, and Yekutieli at a 0.05 false discovery rate (FDR). Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
|
PMC11458896
|
Glycerophospholipid remodeling is critical for orthoflavivirus infection
|
Flavivirus infection is tightly connected to host lipid metabolism. Here, we performed shotgun lipidomics of cells infected with neurotropic Zika, West Nile, and tick-borne encephalitis virus, as well as dengue and yellow fever virus. Early in infection specific lipids accumulate, e.g., neutral lipids in Zika and some lysophospholipids in all infections. Ceramide levels increase following infection with viruses that cause a cytopathic effect. In addition, fatty acid desaturation as well as glycerophospholipid metabolism are significantly altered. Importantly, depletion of enzymes involved in phosphatidylserine metabolism as well as phosphatidylinositol biosynthesis reduce orthoflavivirus titers and cytopathic effects while inhibition of fatty acid monounsaturation only rescues from virus-induced cell death. Interestingly, interfering with ceramide synthesis has opposing effects on virus replication and cytotoxicity depending on the targeted enzyme. Thus, lipid remodeling by orthoflaviviruses includes distinct changes but also common patterns shared by several viruses that are needed for efficient infection and replication.Orthoflaviviruses are emerging arthropod-borne pathogens belonging to the Orthoflavivirus genus within the family of Flaviviridae. Infections with orthoflaviviruses that are pathogenic to humans, including dengue virus (DENV), Zika virus (ZIKV), yellow fever virus (YFV), West Nile virus (WNV), Japanese (JEV), and tick-borne encephalitis virus (TBEV), substantially contribute to global morbidity and mortality rates. Rising global temperature and geographical factors such as increasing human population densities and urbanization are widening the ranges of the transmitting arthropods, thus increasing the risk of epidemic outbreaks. In addition, several orthoflaviviruses have just recently emerged as human pathogens such as Powassan virus in the USA, or increased in its range like Omsk hemorrhagic fever virus that circulates beyond Russia. While potent vaccines have been developed to prevent YFV, TBEV, JEV, and DENV infection, none are available for other pathogenic orthoflaviviruses such as WNV and ZIKV in humans, for WNV a vaccine for horses exists. As there are no specific antiviral treatments for any of these orthoflavivirus infections, only symptomatic treatment is available. The orthoflavivirus genome is a single-stranded, positive-sense RNA of ~11,000 nucleotides in length. The viral genomic RNA consists of a single open-reading-frame (ORF) flanked by 5′ and 3′ untranslated regions (UTRs). The ORF encodes a single polyprotein that is processed by viral and cellular proteases into three structural proteins (capsid [C], precursor membrane [prM], and envelope [E]) and seven nonstructural proteins (NS1, NS2A, NS2B, NS3, NS4A, NS4B, and NS5). While the structural proteins are important for viral attachment, entry, fusion, virus assembly, and virion secretion, the nonstructural proteins contribute to viral RNA (vRNA) replication, virion assembly and release, and evasion of innate immunity. Orthoflavivirus replication complexes reside in replication organelles (ROs) that are convoluted invaginations of the endoplasmic reticulum (ER) membrane. Accordingly, orthoflaviviruses rely on cellular lipid metabolism and usurp several metabolic pathways to foster efficient RO formation and vRNA replication and in addition, as these viruses are enveloped, virion assembly. Interestingly, the formation of the ROs requires several different host lipid classes, including fatty acids (FAs), cholesterol, glycerophospholipids, and sphingolipids. Most studies investigated single orthoflaviviruses but some already noted differences in host cell metabolic requirements, e.g., the dependency of WNV and DENV replication on ceramides differs remarkably with attenuation or enhancement of viral replication, respectively. But differences may also relate to the host cell or organism that is infected, as in mosquito cells inhibition of sphingolipid Δ4 desaturase, which synthesizes ceramide, reduces DENV multiplication. In addition, DENV infection causes an increased abundance of monounsaturated FAs and inhibition of the enzyme catalyzing FA monodesaturation decreases DENV replication. Further, DENV infection leads to an autophagy-dependent degradation of lipid droplets (LDs) called lipophagy to release FAs, which are then used to increase cellular ß-oxidation. This orthoflavivirus-induced lipophagy depends on NS4A/B and host cell ancient ubiquitous protein 1 and is conserved among several orthoflaviviruses. Triglycerides (TAGs) that are stored in LDs increase after infection with ZIKV, and inhibition of the triglyceride-synthesizing enzyme diacylglycerol-O-acyltransferase 1 reduces ZIKV infection in human placenta and human neural cells. Overall, ZIKV infection significantly alters host lipid composition, and in addition to TAGs, particularly relies on a functional sphingolipid metabolic pathway. With regard to orthoflavivirus-induced remodeling of cellular lipids, different lipidomic studies of mosquito cells have demonstrated changes in the total lipid profile that correlate with intracellular membrane alterations induced by ZIKV and DENV infection. To extend the knowledge about orthoflavivirus-induced changes of cellular lipid profiles, we performed lipidomic studies of orthoflaviviral-infected human hepatoma cells using high-resolution mass spectrometry. Here, we observe alterations in several lipid classes and lipid species with greater effects on the host cell lipid profile of orthoflaviviruses that are more pathogenic in humans. Many of the altered lipids have characteristic roles in influencing membrane architecture as well as in cellular signal transduction pathways. Specifically, orthoflavivirus infections cause an accumulation of ceramides and decrease levels of total TAGs and diacylglycerols (DAGs). In addition, we identify virus-induced changes in FA desaturation and in glycerophospholipid pathways. Using inhibitors and RNAi, we confirm the requirement of different lipid remodeling enzymes for DENV-, ZIKV-, WNV-, TBEV-, and YFV-infection, indicating that these pathways are essentially required for efficient orthoflavivirus replication. To investigate if different members of the Orthoflavivirus genus induce distinct changes in the lipidome of infected cells, we performed a semi-targeted shotgun lipidomic analysis over the course of the infection. We decided to analyze orthoflavivirus family members that are most important to human health: the neurotropic ZIKV, WNV, and TBEV and the viscerotropic DENV as well as YFV-17D, the apathogenic vaccine strain of YFV (Fig. 1a). As host cells we chose Huh7 cells, as they are permissive for all viruses we were interested in. We tested different multiplicities of infection (MOI, titrated on Huh7 cells or BHK21 cells in case of TBEV) to attain 80% virus-positive cells without apparent cytopathic effects, as confirmed by immunofluorescence analysis 48 h post infection (hpi). In order to reach equal infection levels, we had to inoculate the cells with different MOIs: ZIKV, TBEV, and DENV replicate slightly slower in Huh7 cells, thus we needed higher MOIs of 1, 1, and 0.5, respectively, WNV displayed intermediate replication kinetics (MOI 0.1), and YFV-17D replicates fastest (MOI 0.005). These differences may be due to inherent differences as, for example, YFV-17D is adapted to replicate in cell culture, or may be specific for the cell type.Fig. 1Experimental setup and quality control of the lipidomic analysis of cells infected with different human pathogenic orthoflaviviruses.a Huh7 cells were seeded 1 day prior to infection with the different orthoflaviviruses (ZIKV MOI 1, WNV MOI 0.01, TBEV MOI 1, DENV MOI 0.5, and YFV-17D MOI 0.005). 12, 24, and 48 h post infection (hpi) samples were taken for protein and RNA extraction, and fixed for immunofluorescence analysis. b Virus genome equivalents (GE) per µg total cellular RNA normalized to 18S rRNA measured via RT-qPCR (mean ± SEM, n = 3 independent experiments). c Cells were stained with antibodies against orthoflavivirus E protein (magenta), and BODIPY493/503 (green) and Hoechst (blue) were used to visualize lipid droplets (LDs) and nuclei, respectively. Scale bar 10 µm. d Quantification of lipid droplets from more than 45 individual cells from three independent experiments (median ± 95% CI, *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001, unpaired two-tailed Mann–Whitney U test). e Samples were analyzed by immunoblotting with orthoflavivirus E protein and GAPDH antibodies (shown is one representative experiment of three independent experiments). Source data are provided as a Source Data file. a Huh7 cells were seeded 1 day prior to infection with the different orthoflaviviruses (ZIKV MOI 1, WNV MOI 0.01, TBEV MOI 1, DENV MOI 0.5, and YFV-17D MOI 0.005). 12, 24, and 48 h post infection (hpi) samples were taken for protein and RNA extraction, and fixed for immunofluorescence analysis. b Virus genome equivalents (GE) per µg total cellular RNA normalized to 18S rRNA measured via RT-qPCR (mean ± SEM, n = 3 independent experiments). c Cells were stained with antibodies against orthoflavivirus E protein (magenta), and BODIPY493/503 (green) and Hoechst (blue) were used to visualize lipid droplets (LDs) and nuclei, respectively. Scale bar 10 µm. d Quantification of lipid droplets from more than 45 individual cells from three independent experiments (median ± 95% CI, *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001, unpaired two-tailed Mann–Whitney U test). e Samples were analyzed by immunoblotting with orthoflavivirus E protein and GAPDH antibodies (shown is one representative experiment of three independent experiments). Source data are provided as a Source Data file. Next, we infected cells and harvested them at 12, 24, and 48 hpi in three independent experiments. For quality control of the samples, we first characterized replication kinetics by RT-qPCR of viral genomes, immunofluorescence analysis, and immunoblotting of cells infected with the different viruses using the predetermined MOIs. WNV, TBEV, and DENV showed similar infection kinetics by RT-qPCR reaching almost equal plateaus of ~10 genome copies per µg total cellular RNA (Fig. 1b). ZIKV only reached ~10 viral genomes but with a similar kinetic as the other three viruses while YFV-17D reached the highest genome numbers and replicated fastest (Fig. 1b). Similar results were obtained by immunofluorescence analysis of viral E protein in infected cells (Fig. 1c, shown is one representative experiment). Here, only few cells were positive for YFV-17D at 24 hpi compared to the other viruses, but all five viruses analyzed reached ~80% of E-positive cells at 48 hpi. This high infection rate reduces the contribution of uninfected neighboring cells to a minimum thus allowing a detailed analysis of lipid profiles, while at earlier timepoints changes due to signaling events may become detectable. As we are interested in lipid metabolism, we additionally stained LDs which are the neutral lipid storage organelles in cells. In infected E-positive cells we detected fewer LDs than in uninfected controls or neighboring E-negative cells (Fig. 1c). This effect was most striking 48 hpi and confirms the lipophagic degradation of LDs observed by others. Indeed, quantification of LDs revealed that all orthoflaviviruses except for TBEV reduced the number and total amount of LDs in infected cells, while especially ZIKV and DENV infections also slightly increased the size of individual LDs (Fig. 1d and Fig. S1). Finally, we compared E protein levels by immunoblot analysis. Of note, the detection sensitivity of the anti-orthoflavivirus E protein is not equal for all viruses. Still, in immunoblot analysis we detected a relative increase of E levels over the course of infection that substantiated the previous results (Fig. 1e, shown is one representative experiment). To study the impact of viral infection on the cellular lipidome, we quantified 322 lipids of 16 classes (Supplementary Data 1). Of note, due to different biosafety requirements, BSL2 for ZIKV and YFV-17D, and BSL3 for TBEV, WNV, and DENV, we harvested two mock-infected controls in each independent experiment, but the cells were split just prior to infection from the same culture. A high correlation of all quantified lipids was observed between the BSL2 and BSL3 control samples (r > 0.98 for all time points and independent experiments, Fig. S2) indicating only negligible changes for these two laboratories. Nevertheless, principal component analysis (PCA) of the identified lipids revealed that 12 and 24 hpi infected samples overlapped with each other and the respective BSL2 and BSL3 controls, indicating that the majority of lipids identified stemmed from the uninfected cells within the culture (Fig. 2a). In contrast, the lipidomic samples taken at 48 hpi clearly clustered according to the virus they were infected with and the mock-infected controls clustered with a small distance from each other (Fig. 2a). ZIKV and YFV-17D-infected cells least deviated from their respective controls and showed largest overlap in the lipidome, while WNV and DENV clustered furthest from the controls. Therefore, an influence of the location of cell cultivation can be seen at 12 and 24 hpi but at 48 hpi the influence of viral infection is clearly observed in the PCA analysis.Fig. 2Pathogenic orthoflaviviruses induce distinct changes in the lipidome of infected cells.a Huh7 cells infected with the different orthoflaviviruses were subjected to lipid and protein extraction and analyzed by direct infusion tandem mass spectrometry. Principal component analysis (PCA) of the lipidomics dataset. Relative lipid species abundance (Supplementary Data 1, sheet mol%) was used for the analysis. Single dots represent the different independent experiments. b Hierarchical clustering of lipid species (mol%) and samples of the 48 hpi dataset. Lipid species concentrations were log2-transformed and centered to the mean. Only lipid species that were detected in 90% of the samples (277 lipids) were used for clustering. Hierarchical clustering was calculated for individual samples as well as for lipid species using Euclidean distance metric and complete linkage clustering method. a Huh7 cells infected with the different orthoflaviviruses were subjected to lipid and protein extraction and analyzed by direct infusion tandem mass spectrometry. Principal component analysis (PCA) of the lipidomics dataset. Relative lipid species abundance (Supplementary Data 1, sheet mol%) was used for the analysis. Single dots represent the different independent experiments. b Hierarchical clustering of lipid species (mol%) and samples of the 48 hpi dataset. Lipid species concentrations were log2-transformed and centered to the mean. Only lipid species that were detected in 90% of the samples (277 lipids) were used for clustering. Hierarchical clustering was calculated for individual samples as well as for lipid species using Euclidean distance metric and complete linkage clustering method. We then performed hierarchical clustering of the lipid species identified 48 hpi and displayed them as heatmaps to identify lipid species that might distinguish the different orthoflaviviruses and infected cells from controls. As already observed in the PCA analysis, DENV-, WNV-, and TBEV-infected cells clustered according to the virus they were infected with, while ZIKV- and YFV-17D-infected as well as BSL2 and BSL3 controls clustered together (Fig. 2b). Certain molecular species of cholesterol ester (CE) showed a similar pattern and were elevated in DENV, WNV, and TBEV compared to ZIKV, YFV-17D, and the mock-infected controls, clearly setting them apart (Fig. 2b). DENV and WNV infection additionally led to increased levels of ceramides (Cer). Interestingly, DENV infection led to unique elevation of some TAG and DAG species as well as hexosyl-ceramide (HexCer) species containing monounsaturated FAs and a decrease in phosphatidylcholine (PC) species. TBEV infection caused a decrease in certain sphingomyelin (SM) and lyso-phospholipid (PL) species compared to all other samples (Fig. 2b). We next analyzed the lipid class and species content following infection with the different orthoflaviviruses relative to the combined controls to minimize variation among control samples (Fig. 3a, b). One striking phenotype we observed, and that has been reported by others for ZIKV infection, is the increased abundance of Cer species after 48 hpi, most pronounced for WNV, DENV, and YFV-17D infection, less for ZIKV, and almost absent in TBEV infection (Fig. 3b). Interestingly, a simultaneous decrease in abundance of SM species at 48 hpi was also observable for all orthoflaviviruses including TBEV, but absent in ZKV infection (Fig. 3b). This finding might indicate an increased SMase activity during the viral infection, specifically at later time points. As all orthoflaviviruses analyzed here except for TBEV are cytopathic in Huh7 cells, the increased levels of Cer species likely reflect activated cell death pathways and do not correlate with the ability of the virus to replicate as TBEV replication levels in Huh7 cells are high (Fig. 1). At 48 hpi we observed a virus-specific phenotype for HexCer species for all investigated viruses. For DENV we recognized an increased abundance for a number of species like HexCer 18:2/16:0;0, which was not observed for the other viral infections, and for WNV we observed reduced levels of HexCer 18:1/24:1;0 (Fig. 3b).Fig. 3Orthoflaviviruses remodel the lipidome of infected cells.a Bar graphs depict the lipid class amount (pmol) per µg protein (mean ± SD, n = 3, *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001, FDR-adjusted p value, unpaired two-tailed Welch’s t-test). b Shown is the lipid amount (pmol) per µg protein of all species in lipid class as mean log2 fold change between infected cells and mock controls at 12, 24, and 48 hpi (n = 3 independent experiments). FDR-adjusted p values (unpaired two-tailed Welch’s t-test) are indicated by the size of the points. Lipid species are ordered according to molecular mass from left to right. hpi hours post infection, Chol cholesterol, PC phosphatidylcholine, LPC lyso-PC, PE phosphatidylethanolamines, LPE lyso-PE, PI phosphatidylinositol, LPI lyso-PI, PS phosphatidylserine, PG phosphatidylglycerol, PA phosphatidic acid, SM sphingomyelins, Cer ceramide, HexCer hexosyl-ceramide, DAG diglycerides, TAG triglycerides, CE cholesterol ester. Source data are provided as Supplementary Data S1, sheet pmol_ug. a Bar graphs depict the lipid class amount (pmol) per µg protein (mean ± SD, n = 3, *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001, FDR-adjusted p value, unpaired two-tailed Welch’s t-test). b Shown is the lipid amount (pmol) per µg protein of all species in lipid class as mean log2 fold change between infected cells and mock controls at 12, 24, and 48 hpi (n = 3 independent experiments). FDR-adjusted p values (unpaired two-tailed Welch’s t-test) are indicated by the size of the points. Lipid species are ordered according to molecular mass from left to right. hpi hours post infection, Chol cholesterol, PC phosphatidylcholine, LPC lyso-PC, PE phosphatidylethanolamines, LPE lyso-PE, PI phosphatidylinositol, LPI lyso-PI, PS phosphatidylserine, PG phosphatidylglycerol, PA phosphatidic acid, SM sphingomyelins, Cer ceramide, HexCer hexosyl-ceramide, DAG diglycerides, TAG triglycerides, CE cholesterol ester. Source data are provided as Supplementary Data S1, sheet pmol_ug. Early in orthoflavivirus infection neutral lipids tend to accumulate. At 12 hpi, ZIKV induced a significant increase in total TAG and CE content affecting nearly all lipid species (Fig. 3a, b). Later in infection, especially at 48 hpi, TAG levels strongly decreased in WNV-, TBEV-, and DENV-, and to a lesser extent in ZIKV- and YFV-17D-infected cells (Fig. 3a), which correlates to reduced LD content observed by microscopy (Fig. 1c, d). This decrease was observed for nearly all TAG species (Fig. 3b). Concomitant with the decreased levels of TAGs late in infection, DAG species and total levels were similarly affected (Fig. 3a, b), as for example DAG 14:0_18:1, DAG 14:1_16:0, and DAG 14:1_16:1, that were significantly decreased in DENV, TBEV, and WNV at 48 hpi. In contrast, levels of CE, the other main component of LDs, remained rather stable during infection with the exception for DENV where at 24 and 48 hpi CE 18:1, CE 20:1, CE 20:3, and CE 22:5 were significantly increased (Fig. 3b). The detailed analysis of membrane lipids revealed that lysophosphatidylcholine (LPC) and lysophosphatidylethanolamine (LPE) are increased in abundance at 12 hpi for ZIKV, WNV, and YFV-17D (Fig. 3b). Individual species like LPE 18:2, LPC 16:1, and LPC 20:4 for ZIKV, LPC 15:0, 16:0, 18:0, and 20:3 for WNV, as well as LPE 16:1, LPE 18:2, LPE 20:4, LPC 16:1, and LPC 20:4 for YFV-17D (Fig. 3b), which were elevated at 12 hpi showed no further alteration later in infection. DENV exclusively showed increased amounts for LPC 18:0 at 48 hpi. Overall, a large portion of the glycerophospholipidome is affected by orthoflavivirus infection. Specifically, WNV, TBEV, and DENV infection led to decreased levels of PC, PE, and PS at 48 hpi (Fig. 3b). While ZIKV induced changes on PE and PC were negligible it was the only case with a slight increase in PS 36:4 at 48 hpi. For all other viral infections, we observed decreased amounts for some PS species at 48 hpi like PS 32:1 and PS 34:1 (Fig. 3b). Of note, the modulation of the membrane phospholipid composition affected cholesterol only for WNV and TBEV infection at 48 hpi (Fig. 3a, b). To identify lipid metabolic pathways that are dysregulated by infection we next performed lipid ontology analysis utilizing lipid ontology (LION) enrichment analysis web application (LION/web). At 48 hpi clear patterns of dysregulated ontology terms were identified (Fig. S3). The top downregulated terms related to the storage of TAG in LDs while sterol and sterol ester terms were upregulated. In contrast, membrane lipids, especially endoplasmic membrane lipid terms like PC and overall glycerophospholipids, were increased for all viruses tested, while DAG terms were significantly less (Fig. S3). This corresponds to the membrane requirement for the formation of ROs in orthoflavivirus-infected cells. In addition, and as already clearly visible in the lipid species quantification (Fig. 3), Cer lipid annotations were enriched (Fig. S3). As glycerophospholipid abundances and corresponding ontology terms were affected by orthoflavivirus infection (Fig. 3 and Fig. S3), we next analyzed glycerophospholipid remodeling using BioPAN that enables the identification of the enzyme activities that may be involved. The BioPAN analysis is based on the conversion of individual lipid species and Z scores of reactions between lipid classes are then calculated. Arrow thickness represents the reaction activity of the individual class conversions while the color of the lipid class circles indicates log2 fold change of lipid class abundance (additionally displayed in a heatmap, Fig. S4). BioPan analysis revealed an early deacylation of diacylglycerophospholipids to form the lyso species which later in infection reverted (Fig. 4a). Accordingly, phospholipase A2 (PLA2) proteins were predicted to have increased activity early in infection (Fig. 4b, upper triangles), mostly without concomitant changes in expression level (Fig. 4b, lower triangles). Of note, all enzymes predicted to catalyze the same enzymatic reaction are listed. An additional characteristic suggested by the BioPAN analysis was the flow of lipids from DAG to phosphatidylethanolamine (PE) and PC (Fig. 4a), which was predicted to be the result of increased activity of choline phosphotransferase 1 (CHPT1) and choline/ethanolamine phosphotransferase 1 (CEPT1) (Fig. 4b). Phosphatidylinositol (PI) levels were salvaged from phosphatidic acid (PA) via CDP-diacylglycerol synthase (CDS)1 and CDS2 as well as CDP-diacylglycerol inositol 3-phosphatidyltransferase (CDIPT) (Fig. 4b). Other reactions were more distinct between the individual orthoflaviviruses. For example, ZIKV progressively displayed increased phosphatidylserine (PS) levels through the conversion of PE to PS by phosphatidylserine synthase 2 (PTDSS2), PC/PE to PS by PTDSS1, and from PA to PS through CDS1 and PTDSS1 (Fig. 4a, b). In contrast WNV, DENV, and YFV-17D caused increased decarboxylation of PS to PE likely via phosphatidylserine decarboxylase (PISD) (Fig. 4a, b). Interestingly, the enzymes that are predicted to have altered activities in orthoflavivirus-infected cells are not differently expressed in transcriptome analysis that was performed on samples from the same experiments as the lipidomics analysis (Fig. 4b, compare triangles). Only PLA2G4C was expressed higher in orthoflavivirus-infected cells at 48 hpi when lysolipids were not elevated any more but single lipid species conversion was still detected.Fig. 4Profound glycerophosholipid remodeling in orthoflavivirus infection.a Analysis of the lipidomics dataset using BioPan pathway analysis. Lipid species abundance (Supplementary Data 1, sheet pmol_ug) was used for the analysis. Predicted reaction activities between lipid classes are indicated by arrows (red indicates increased and blue decreased reaction activities). Color of lipid classes indicate log2 fold change in abundance per µg protein of infected versus uninfected cells (n = 3 independent experiments). b Upper left triangle: BioPan prediction of Z scores of active (red) and suppressed (blue) enzyme activities using the standard settings (Z > 1.645 at p < 0.05). Lower right triangle: log2 fold change of expression level determined by transcriptome analysis of the same samples. Source data are provided as a Source Data file. a Analysis of the lipidomics dataset using BioPan pathway analysis. Lipid species abundance (Supplementary Data 1, sheet pmol_ug) was used for the analysis. Predicted reaction activities between lipid classes are indicated by arrows (red indicates increased and blue decreased reaction activities). Color of lipid classes indicate log2 fold change in abundance per µg protein of infected versus uninfected cells (n = 3 independent experiments). b Upper left triangle: BioPan prediction of Z scores of active (red) and suppressed (blue) enzyme activities using the standard settings (Z > 1.645 at p < 0.05). Lower right triangle: log2 fold change of expression level determined by transcriptome analysis of the same samples. Source data are provided as a Source Data file. To analyze if the FA composition of membrane and storage lipids species changes following infection, we determined the FA profile of PC, PE, PS, and TAG from the associated tandem mass spectrometric experiments (Supplementary Data 2). Depicted is the difference in relative amounts of FAs within the lipids class (i.e., ∆ lipid abundance (mol% of infected cells − mol% of mock control, normalized to [−1, 1] in each lipid class), Fig. 5). The most striking effect of infection we observed was changes in the content of monounsaturated vs. saturated FA in the analyzed lipid classes. Distinct modulation of FA 16:1, FA 16:0, FA 18:1, and FA 18:0 was observed for PE with strongest changes in YFV-17D-, TBEV-, and DENV-infected cells, while the response in ZIKV-infected cells was much less pronounced. The increased portion of FA 16:1 and FA 18:1 at 48 hpi may indicate increased activity of stearoyl-CoA desaturases (SCDs) or de novo synthesis of PE 34:1 for membranes required for virus replication, while the opposing effect was observed in PC, PS, and TAGs. The second phenotype was the increased utilization of arachidonic acid (FA 20:4) in PE and PS species. This was more pronounced at later time points after infection and might reflect altered activity of one or more phospholipase A2 (cPLA2) enzymes (Fig. 5).Fig. 5Fatty acid composition of membrane phospholipids and triglycerides changes following orthoflavivirus infection.We determined the relative abundance (mol%) of FAs in PC, PE, PI, PS, and TAG and calculated the difference between infected and uninfected control cells for each lipid class (Δ relative abundance: mol% (infected) − mol% (mock), normalized to [−1, 1]). Heatmaps illustrate the Δ relative abundances of FAs in PC, PE, PI, PS, and TAG. Shown is the mean of three independent experiments. Red and blue depict increases and decreases, respectively, and gray indicates NA. Source data are provided as Supplementary Data 2. We determined the relative abundance (mol%) of FAs in PC, PE, PI, PS, and TAG and calculated the difference between infected and uninfected control cells for each lipid class (Δ relative abundance: mol% (infected) − mol% (mock), normalized to [−1, 1]). Heatmaps illustrate the Δ relative abundances of FAs in PC, PE, PI, PS, and TAG. Shown is the mean of three independent experiments. Red and blue depict increases and decreases, respectively, and gray indicates NA. Source data are provided as Supplementary Data 2. For follow-up studies to investigate potential function in viral replication, we focused on enzymes that were predicted to have increased activity (Fig. 4b), Cer metabolism that was identified through the quantification of lipid classes and species (Figs. 2 and 3), and the changes in the FA profile (Fig. 5). For those enzymes that are targetable using inhibitors, we next performed infection studies. For all other enzymes with altered activities, we used gene knockdown approaches. A common phenotype in orthoflavivirus infection with the exception of TBEV is the increase in Cer during the course of infection. This is in line with studies of ZIKV infection where an essential role of Cer biosynthesis for ZIKV replication was delineated. To determine if replication of other orthoflaviviruses is affected by inhibition of sphingolipid biosynthesis, we made use of the small-molecule inhibitors fumonisin B1 and myriocin to block the activity of ceramide synthase (CerS) and serine palmitoyltransferase (SPT1), respectively (Fig. 6a). We treated cells 48 h prior infection and added inhibitor again to the inoculum during infection (Fig. 6b). We first verified that the inhibitors had no effect on cell viability in concentrations that we used (Fig. 6c). None of the inhibitors affected the viability of Huh7 cells. In immunoblot analysis of infected inhibitor-treated cells, SPTi had minor effects on orthoflavivirus E protein levels (Fig. 6d), but reduced DENV titers (Fig. 6e). Interestingly, for all orthoflaviviruses that cause a cytopathic effect (CPE) in Huh7 cells, SPTi-treated cells displayed less CPE than DMSO-treated control-infected cells with the most significant decrease for DENV (Fig. 6f and Fig. S5). In contrast, CerSi treatment that blocks Cer synthesis both in the de novo and salvage pathway had also little impact on E protein level (Fig. 6d), but significantly enhanced ZIKV and YFV-17D titers (Fig. 6e), and increased cell death in ZIKV, WNV, and YFV-17D infection (Fig. 6f and Fig. S5). These findings imply an opposing effect of de novo Cer synthesis and its salvage pathway on virus-induced cell death pathways and an antiviral function of Cer in ZIKV and YFV infections.Fig. 6Pro- and antiviral function of ceramide biosynthesis and fatty acid desaturation.a Scheme of the target enzymes and respective inhibitors. b Huh7 cells were treated with the inhibitors 48 h prior to infection with the different orthoflaviviruses (ZIKV MOI 0.1, WNV MOI 0.001, TBEV MOI 0.05, DENV MOI 0.05, and YFV-17D MOI 0.005). After infection, the respective inhibitors were added again, followed by analysis of viral protein expression level by immunoblot, determination of viral titers (TCID50), or analysis of cytopathic effects (CPE). c Cell viability was analyzed 72 h post treatment with different inhibitors. Treatment with 10% DMSO served as dead control (mean ± SEM, n = 3 independent experiments). d Intracellular levels of orthoflavivirus E protein were assessed by immunoblotting at 2 dpi. Shown is one representative blot (n = 4 (ZIKV, YFV-17D, WNV) or n = 5 (TBEV, DENV) independent experiments). e Virus titers in inhibitor-treated infected cells were determined by TCID50 titration at 2 dpi (mean ± SEM, n = 5 (DENV, TBEV, YFV-17D, ZIKV) or n = 4 (WNV) independent experiments, except for SPTi (all viruses), DENV (cPLA2i) and TBEV (SPTi): one experiment excluded due to contamination, *p ≤ 0.05, **p ≤ 0.01, unpaired two-tailed Mann–Whitney U test). f Cells were fixed at 3 dpi, surviving cells stained with crystal violet, and CPE was quantified using ImageJ. Box-and-whisker plot indicates CPE as log2 fold change infected over mock, normalized to DMSO control (center line: median, box limits: upper and lower quartiles, whiskers: 1.5 × interquartile range, points: outliers, n = 4 (ZIKV, DENV, WNV) or n = 5 (YFV-17D) independent experiments, *p ≤ 0.05, **p ≤ 0.01, unpaired two-tailed one sample Student’s t-test). dpi days post infection. Source data are provided as a Source Data file. a Scheme of the target enzymes and respective inhibitors. b Huh7 cells were treated with the inhibitors 48 h prior to infection with the different orthoflaviviruses (ZIKV MOI 0.1, WNV MOI 0.001, TBEV MOI 0.05, DENV MOI 0.05, and YFV-17D MOI 0.005). After infection, the respective inhibitors were added again, followed by analysis of viral protein expression level by immunoblot, determination of viral titers (TCID50), or analysis of cytopathic effects (CPE). c Cell viability was analyzed 72 h post treatment with different inhibitors. Treatment with 10% DMSO served as dead control (mean ± SEM, n = 3 independent experiments). d Intracellular levels of orthoflavivirus E protein were assessed by immunoblotting at 2 dpi. Shown is one representative blot (n = 4 (ZIKV, YFV-17D, WNV) or n = 5 (TBEV, DENV) independent experiments). e Virus titers in inhibitor-treated infected cells were determined by TCID50 titration at 2 dpi (mean ± SEM, n = 5 (DENV, TBEV, YFV-17D, ZIKV) or n = 4 (WNV) independent experiments, except for SPTi (all viruses), DENV (cPLA2i) and TBEV (SPTi): one experiment excluded due to contamination, *p ≤ 0.05, **p ≤ 0.01, unpaired two-tailed Mann–Whitney U test). f Cells were fixed at 3 dpi, surviving cells stained with crystal violet, and CPE was quantified using ImageJ. Box-and-whisker plot indicates CPE as log2 fold change infected over mock, normalized to DMSO control (center line: median, box limits: upper and lower quartiles, whiskers: 1.5 × interquartile range, points: outliers, n = 4 (ZIKV, DENV, WNV) or n = 5 (YFV-17D) independent experiments, *p ≤ 0.05, **p ≤ 0.01, unpaired two-tailed one sample Student’s t-test). dpi days post infection. Source data are provided as a Source Data file. In our analysis of lipid species (Fig. 2) and of the FA profile of glycerophospholipids (Fig. 5), we detected features of enhanced SCD activity. Earlier reports showed that SCD activity is required for DENV, ZIKV, and JEV RNA replication and DENV virion production as well as HCV replication. In line we confirmed the antiviral activity of SCDi on orthoflaviviruses (Fig. 6d–e and Fig. S5). Viral titers were slightly reduced for WNV and TBEV, and CPE reduced in ZIKV, WNV, DENV, and YFV-17D. In our analysis of membrane lipid remodeling, increased activities of PLA2 enzymes and decreased activities of PLD1/2 were predicted. As several PLA2 isoforms are expressed in Huh7 cells, we addressed if formation of lyso-PL blocks orthoflavivirus infection by using an inhibitor targeting cPLA2 enzymes, and investigated if pre-treatment with PLDi can boost infection. However, PLDi as well as cPLA2 inhibitor treatment had no effect on orthoflavivirus infection, replication, or virion production (Fig. 6d–e and Fig. S5). For other enzymes that were identified in the BioPAN analysis to have altered activity in orthoflavivirus infection (Fig. 4), no specific inhibitors are available. Therefore, we designed small hairpin (sh)RNAs to downregulate target mRNAs and proteins (Fig. 7a). We isolated total cellular RNA from shRNA-transduced cells 5 days post transduction (dpt) to quantitate the knockdown and performed viability assays (Fig. 7b). All shRNAs reduced mRNA expression levels of the target enzymes more than 50% compared to non-targeting shRNA (shNT) control cells (Fig. 7c) and transduced cells displayed viability within the range of the controls (Fig. 7d). We then analyzed the impact of knocking down enzymes involved in glycerophospholipid remodeling on viral protein expression in infected cells but did not observe strong differences on E protein levels between knockdown and control cells (Fig. 7e).Fig. 7Phosphatidylserine turnover and phosphatidylinositol biosynthesis are required for orthoflavivirus infection and replication.a Scheme of the synthesis of glycerolipids and glycerophospholipids. Target enzymes are indicated in blue. b Huh7 cells were transduced with lentiviruses carrying shRNAs targeting the different enzymes or a non-targeting (NT) shRNA followed by an infection with different orthoflaviviruses 4 dpt (ZIKV MOI 0.1, WNV MOI 0.001, TBEV MOI 0.05, DENV MOI 0.05, and YFV-17D MOI 0.005). Cells were used for validation of the lentiviral constructs and for analysis of viral protein expression level by immunoblot, determination of viral titers (TCID50), or analysis of cytopathic effects (CPE). c shRNA-knockdown efficiency of the targets was determined by RT-qPCR 4 dpt. Shown are mRNA expression levels of the corresponding enzymes relative to shNT and normalized to 18S rRNA (mean ± SEM, n = 3 independent experiments). d Cell viability was determined 5 dpt. Treatment with 10% DMSO served as non-viable control (mean ± SEM, n = 3 independent experiments). e Intracellular levels of viral E protein were assessed by immunoblotting at 2 dpi. Shown is one representative immunoblot (n = 3 (TBEV) or n = 4 (ZIKV, WNV, DENV, YFV-17D) independent experiments). f Virus titers in shRNA-transduced infected cells were determined by TCID50 titration at 2 dpi (mean ± SEM, n = 3 (ZIKV, TBEV, YFV-17D, WNV) or n = 4 (DENV) independent experiments, *p ≤ 0.05, unpaired two-tailed Mann–Whitney U test). g Cells were fixed at 3 dpi and cells were stained with crystal violet staining. CPE was quantified using ImageJ. Box-and-whisker plot indicates CPE as log2 fold change of infected over mock control, normalized to shNT (center line: median, box limits: upper and lower quartiles, whiskers: 1.5 × interquartile range, points: outliers, n = 2 (DENV), n = 4 (WNV, YFV-17D), or n = 6 (ZIKV) independent experiments, *p ≤ 0.05, unpaired two-tailed one sample Student’s t-test). dpt days post transduction, dpi days post infection. Source data are provided as a Source Data file. a Scheme of the synthesis of glycerolipids and glycerophospholipids. Target enzymes are indicated in blue. b Huh7 cells were transduced with lentiviruses carrying shRNAs targeting the different enzymes or a non-targeting (NT) shRNA followed by an infection with different orthoflaviviruses 4 dpt (ZIKV MOI 0.1, WNV MOI 0.001, TBEV MOI 0.05, DENV MOI 0.05, and YFV-17D MOI 0.005). Cells were used for validation of the lentiviral constructs and for analysis of viral protein expression level by immunoblot, determination of viral titers (TCID50), or analysis of cytopathic effects (CPE). c shRNA-knockdown efficiency of the targets was determined by RT-qPCR 4 dpt. Shown are mRNA expression levels of the corresponding enzymes relative to shNT and normalized to 18S rRNA (mean ± SEM, n = 3 independent experiments). d Cell viability was determined 5 dpt. Treatment with 10% DMSO served as non-viable control (mean ± SEM, n = 3 independent experiments). e Intracellular levels of viral E protein were assessed by immunoblotting at 2 dpi. Shown is one representative immunoblot (n = 3 (TBEV) or n = 4 (ZIKV, WNV, DENV, YFV-17D) independent experiments). f Virus titers in shRNA-transduced infected cells were determined by TCID50 titration at 2 dpi (mean ± SEM, n = 3 (ZIKV, TBEV, YFV-17D, WNV) or n = 4 (DENV) independent experiments, *p ≤ 0.05, unpaired two-tailed Mann–Whitney U test). g Cells were fixed at 3 dpi and cells were stained with crystal violet staining. CPE was quantified using ImageJ. Box-and-whisker plot indicates CPE as log2 fold change of infected over mock control, normalized to shNT (center line: median, box limits: upper and lower quartiles, whiskers: 1.5 × interquartile range, points: outliers, n = 2 (DENV), n = 4 (WNV, YFV-17D), or n = 6 (ZIKV) independent experiments, *p ≤ 0.05, unpaired two-tailed one sample Student’s t-test). dpt days post transduction, dpi days post infection. Source data are provided as a Source Data file. Next, transduced Huh7 cells were infected with orthoflaviviral stocks at 4 dpt. 48 hpi, we analyzed viral titers by TCID50 titration on BHK21 cells. Overall, titers of ZIKV and YFV-17D were less affected in the different knockdown cells than WNV, TBEV, and DENV titers (Fig. 7f). Despite the variability between independent experiments, we identified proviral enzymes important for the replication of several orthoflaviviruses. For instance, knockdown of PTDSS2, but not PTDSS1, reduced WNV, DENV, and to a lesser extent TBEV and YFV-17D virus titers (Fig. 7f). While PTDSS1 catalyzes the conversion of PC and PE to PS in the ER, PTDSS2 is specific for PE and has a high affinity for docosahexaenoic acid (DHA) and thus creates DHA-containing PS. Likewise, the CPE of orthoflavivirus infection was reduced in cells lacking PTDSS2 (Fig. 7g), even though the knockdown itself caused some growth defect (Fig. S6). Depletion of PISD that decarboxylates PS at the inner mitochondrial membrane to produce PE led to reduced titers of WNV, TBEV, and DENV, but only slightly rescued cells from CPE induced by the infection (Fig. 7e, g). Finally, de novo synthesis of PI from CDP-DAG by CDIPT is required for WNV, TBEV, and, to a lower extent, for ZIKV and DENV infection, as both virus titers and virus-induced CPE in CDIPT-knockdown cells were reduced (Fig. 7e, g). CDS1 and 2 that catalyze the upstream reaction from PA to CDP-DAG also negatively affected viral replication and CPE, with the broadly acting CDS1 having a slightly stronger effect on CPE. Increased PA to PI reaction activities were found for all orthoflaviviruses at some point after infection (Fig. 4a). Regarding unique dependencies, DENV is the only analyzed orthoflavivirus that requires PLD2, but not PLD1 expression. PLD1 is active in the ER, Golgi, and endosomal membranes while PLD2 localizes to the plasma membrane. The dependency fits to the uniquely increased PA levels 24 and 48 hpi in DENV-infected cells (Fig. 4 and Fig. S4). The CEPT/CHPT-dependent synthesis of PE and PC as well as the methylation of PE to PC by PEMT are not required for efficient orthoflavivirus infection. In order to decipher which step of the viral replication cycle is affected by altered glycerophospholipid metabolism, we first investigated viral entry. We focused on six of our initial targets that showed the most prominent effect on viral infectivity and altered CPE, namely PTDSS1 and 2, PISD, PLD1 and 2 as well as CDIPT. We infected Huh7 cells expressing the different shRNAs with a high MOI of the different orthoflaviviruses and determined viral genome equivalents (GE) via RT-qPCR at 4 hpi. Interestingly, in contrast to a strong effect on viral titers, viral entry was unaffected in knockdown cells (Fig. 8a). Next, we analyzed vRNA replication and protein translation by quantitating intracellular vRNA copy numbers by RT-qPCR and intracellular viral protein levels by immunoblot analysis. We only observed minor changes in intracellular vRNA levels for the knockdown of PLD2 in WNV infection and slightly decreased viral protein levels for the knockdown of PTDSS2 in DENV infection, but overall, depletion of these lipid remodeling enzymes had no effect on vRNA replication and translation (Fig. 8b). To test the hypothesis of altered virus assembly or release, we investigated vRNA as well as viral E protein in the supernatant of infected knockdown cells. Correlating to minor effects on viral titers in YFV-17D and ZIKV infection, we did not observe major changes in vRNA and E protein levels in the supernatant for these viruses (Fig. 8c). In comparison, we detected significantly lower viral E protein in WNV-infected knockdown cells for shPTDSS1 and shPTDSS2 as well as shPLD2 and shCDIPT, while the extracellular viral copy numbers were only slightly altered. Even though the variability of vRNA in TBEV-infected knockdown cells is relatively high, E protein levels of shPTDSS2 and shPLD2 cells were drastically reduced, leading to the conclusion that defects in assembly or release might be the cause for altered TBEV infectivity in the culture supernatants.Fig. 8Impaired phospholipid remodeling reduces viral particle production but does not affect virus entry or RNA replication and protein expression.a shRNA-transduced Huh7 cells were infected at 4 dpt with high MOIs (ZIKV MOI 2, WNV MOI 5, DENV MOI 2, TBEV MOI 10, and YFV-17D MOI 5). Virus entry was determined by measuring viral genome equivalents (GE) 4 h post infection using RT-qPCR (mean ± SEM, technical duplicates from 2 to 3 independent experiments, non-significant in unpaired two-tailed Welch’s t-test). b, c Transduced Huh7 cells were infected with low MOIs 4 dpt (ZIKV MOI 0.1, WNV MOI 0.001, TBEV MOI 0.05, DENV MOI 0.05, and YFV-17D MOI 0.005) and analyses were performed at 2 dpi. b Intracellular viral RNA (vRNA) was quantified using RT-qPCR (mean ± SEM, n = 4 (DENV), or n = 5 (ZIKV, TBEV, WNV, YFV) independent experiments, non-significant in unpaired two-tailed Welch’s t-test) and the viral envelope/E protein level was determined by immunoblot analysis. Box-and-whisker plot indicates E protein expression as log2 fold change over shNT (center line: median, box limits: upper and lower quartiles, whiskers: 1.5 × interquartile range, points: outliers, n = 3 (TBEV, YFV-17D) or n = 4 (ZIKV, WNV, DENV) independent experiments, non-significant in unpaired two-tailed one sample Student’s t-test). c Extracellular vRNA was quantified by RT-qPCR (mean ± SEM, n = 4 (ZIKV, YFV-17D) or n = 5 (TBEV, WNV, DENV) independent experiments, non-significant in unpaired two-tailed Welch’s t-test) and E protein levels in the supernatant by immunoblot analysis. Box-and-whisker plot indicates E protein levels as log2 fold change over shNT (center line: median, box limits: upper and lower quartiles, whiskers: 1.5 × interquartile range, points: outliers, n = 2 (YFV-17D) or n = 5 (ZIKV, WNV, TBEV) independent experiments, *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001 unpaired two-tailed one sample Student’s t-test). dpt days post transduction, dpi days post infection. Source data are provided as a Source Data file. a shRNA-transduced Huh7 cells were infected at 4 dpt with high MOIs (ZIKV MOI 2, WNV MOI 5, DENV MOI 2, TBEV MOI 10, and YFV-17D MOI 5). Virus entry was determined by measuring viral genome equivalents (GE) 4 h post infection using RT-qPCR (mean ± SEM, technical duplicates from 2 to 3 independent experiments, non-significant in unpaired two-tailed Welch’s t-test). b, c Transduced Huh7 cells were infected with low MOIs 4 dpt (ZIKV MOI 0.1, WNV MOI 0.001, TBEV MOI 0.05, DENV MOI 0.05, and YFV-17D MOI 0.005) and analyses were performed at 2 dpi. b Intracellular viral RNA (vRNA) was quantified using RT-qPCR (mean ± SEM, n = 4 (DENV), or n = 5 (ZIKV, TBEV, WNV, YFV) independent experiments, non-significant in unpaired two-tailed Welch’s t-test) and the viral envelope/E protein level was determined by immunoblot analysis. Box-and-whisker plot indicates E protein expression as log2 fold change over shNT (center line: median, box limits: upper and lower quartiles, whiskers: 1.5 × interquartile range, points: outliers, n = 3 (TBEV, YFV-17D) or n = 4 (ZIKV, WNV, DENV) independent experiments, non-significant in unpaired two-tailed one sample Student’s t-test). c Extracellular vRNA was quantified by RT-qPCR (mean ± SEM, n = 4 (ZIKV, YFV-17D) or n = 5 (TBEV, WNV, DENV) independent experiments, non-significant in unpaired two-tailed Welch’s t-test) and E protein levels in the supernatant by immunoblot analysis. Box-and-whisker plot indicates E protein levels as log2 fold change over shNT (center line: median, box limits: upper and lower quartiles, whiskers: 1.5 × interquartile range, points: outliers, n = 2 (YFV-17D) or n = 5 (ZIKV, WNV, TBEV) independent experiments, *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001 unpaired two-tailed one sample Student’s t-test). dpt days post transduction, dpi days post infection. Source data are provided as a Source Data file. Although DENV showed a great loss of infectivity due to the knockdown of proteins involved in glycerophospholipid remodeling, we did not detect changes in viral copy numbers in the supernatant. Due to the size of DENV E protein and low levels of infectious virus in the supernatant, we were unable to quantify viral protein E levels. Since hepatoma cells such as Huh7 are not the ideal model for neurotropic orthoflaviviruses, we aimed to confirm our results in a more relevant model system and chose immortalized human microglial cells (HMC3). We first transduced HMC3 cells with our shRNA-encoding lentiviral particles, validated knockdowns via RT-qPCR, and confirmed viability of the transduced cells (Fig. 9a–c). Unfortunately, we were unable to obtain a decent knockdown of both PLD isoforms, which is reflected by their limited effects on virus replication as determined by immunoblot of orthoflavivirus E protein and by TCID50 analysis (Fig. 9d, e). Whereas in Huh7 cells ZIKV viral titers were only marginally affected by the knockdown of PTDSS1 and PTDSS2 (Fig. 7f), CPE was significantly reduced by the depletion of these enzymes (Fig. 7g). In HMC3 cells, we detected an even more prominent effect on viral titers as well as viral E protein levels (Fig. 9d, e), indicating the importance of PS synthesis for ZIKV infection. In addition, PTDSS knockdown negatively impacted WNV and TBEV infectivity and intracellular protein level. Similar to Huh7 cells knockdown of CDIPT strongly affected WNV and TBEV infection but only had a minor impact on ZIKV.Fig. 9Enzymes of the glycerophospholipid pathways are required for efficient infection of neurotropic orthoflaviviruses in human microglia cells.a HMC3 cells were transduced with shRNAs targeting enzymes of the phospholipid pathway followed by infection with neurotropic orthoflaviviruses. b Knockdown efficiency was validated by RT-qPCR (mean ± SEM, n = 2 (shPTDSS1), n = 3 (shPLD1), n = 4 (shCDIPT, shPISD, shPTDSS2), or n = 5 (shPLD2) independent experiments). c Cell viability of shRNA-transduced HMC3 cells was determined at 4 dpt. shNT served as control and 10% DMSO-treated cells as non-viable control (mean ± SEM, n = 3). d Transduced HMC3 cells were infected at 4 dpt with neurotropic orthoflaviviruses (ZIKV MOI 0.005, WNV MOI 0.25, TBEV MOI 0.5) and E protein levels were determined by immunoblotting at 2 dpi. Shown is one representative immunoblot of three (TBEV, WNV) or four (ZIKV) independent experiments. e Determination of viral titers by TCID50 assays (mean ± SEM, n = 3–4, *p ≤ 0.05, unpaired two-tailed Mann–Whitney U test). dpt days post transduction, dpi days post infection. Source data are provided as a Source Data file. a HMC3 cells were transduced with shRNAs targeting enzymes of the phospholipid pathway followed by infection with neurotropic orthoflaviviruses. b Knockdown efficiency was validated by RT-qPCR (mean ± SEM, n = 2 (shPTDSS1), n = 3 (shPLD1), n = 4 (shCDIPT, shPISD, shPTDSS2), or n = 5 (shPLD2) independent experiments). c Cell viability of shRNA-transduced HMC3 cells was determined at 4 dpt. shNT served as control and 10% DMSO-treated cells as non-viable control (mean ± SEM, n = 3). d Transduced HMC3 cells were infected at 4 dpt with neurotropic orthoflaviviruses (ZIKV MOI 0.005, WNV MOI 0.25, TBEV MOI 0.5) and E protein levels were determined by immunoblotting at 2 dpi. Shown is one representative immunoblot of three (TBEV, WNV) or four (ZIKV) independent experiments. e Determination of viral titers by TCID50 assays (mean ± SEM, n = 3–4, *p ≤ 0.05, unpaired two-tailed Mann–Whitney U test). dpt days post transduction, dpi days post infection. Source data are provided as a Source Data file. To gain more insight into the physiological relevance of glycerophospholipid remodeling for YFV-17D and DENV replication, we next transduced THP-1 cells and treated them with phorbol-12-myristate-13-acetate (PMA) and interleukin 4 (IL-4) to obtain CD209/DC-SIGN-positive differentiated THP-1 cells (Fig. 10a). Differentiated THP-1- cells can be used as model systems for DENV and YFV-17D infection in immune cells. In comparison to THP-1 cells, which have limited susceptibility for DENV and YFV-17D infection, induced expression of the CD209/DC-SIGN receptor can facilitate orthoflavivirus infection. Validation of shRNA-mediated knockdown efficacy by RT-qPCR revealed that we were not able to achieve sufficient reduction of PISD and PLD 1 and 2 mRNAs (Fig. 10b), which may in part be due to overall low expression levels of these enzymes in THP-1 cells. For the well-expressed PTDSS1, PTDSS2, and CDIPT we were able to substantially reduce expression levels using the lentiviral shRNA delivery system. Next, we verified the induction of CD209/DC-SIGN expression through differentiation by treatment with IL-4 and PMA by flow cytometry. As reported by others, more than 90% of differentiated THP-1 cells were positive for CD209/DC-SIGN compared to virtually no expression in undifferentiated THP-1 cells (Fig. 10c). Viability assays of the knockdown cells indicated lower metabolic activity of PTDSS1 and PTDSS2 knockdown cells, whereas shPISD-expressing cells displayed higher metabolic activity (Fig. 10d). Viral protein levels of YFV-17D were undetectable in PTDSS1 and 2 and CDIPT-knockdown cells and strongly reduced by PISD and PLD1 knockdown. Correlating to the reduced protein levels, we observed a strong reduction in YFV-17D titers for all knockdown cells except for PLD2 (Fig. 10f). Although DENV E protein was not detectable in immunoblot analysis, we were able to analyze viral titers. Here we detected a strong reduction of DENV titers for all target shRNAs tested, but titers were close to the limit of detection (Fig. 10f).Fig. 10Knockdown of enzymes of the glycerophospholipid pathway drastically reduces viral titers in CD209-positive THP-1 cells.a THP-1 monocytes were transduced with shRNAs targeting proteins of the glycerophospholipid pathway. 4 dpt, cells were differentiated using IL-4 (20 ng/ml) and PMA (20 ng/ml) for 4 days prior to infection. b Knockdown efficiency was validated using RT-qPCR (mean ± SEM, n = 2 (shPTDSS1, shPTDSS2), n = 3 (shPISD, shPLD2), n = 4 (shCDIPT), or n = 5 (shPLD1) independent experiments). c Differentiation into CD209-positive THP-1 cells was confirmed 4 days after IL-4 and PMA treatment by flow cytometry using a CD209-specific antibody. Shown is one representative experiment. d Cell viability was analyzed 8 dpt and 4 days post differentiation (mean ± SEM, n = 3 independent experiments). e Knockdown cells were infected with YFV-17D (MOI 1) and the amount of E protein was analyzed by immunoblotting at 2 dpi. Shown is one representative blot of three independent experiments. f Infectivity in the supernatant of DENV (MOI 2) and YFV-17D (MOI 1) infected cells was determined using TCID50 assays at 2 dpi (mean ± SEM, n = 4 (YFV-17D) or n = 3 (DENV) independent experiments, except for shCDIPT: one experiment excluded due to contamination, *p ≤ 0.05, unpaired two-tailed Mann–Whitney U test). dpt days post transduction, dpi days post infection. Source data are provided as a Source Data file. a THP-1 monocytes were transduced with shRNAs targeting proteins of the glycerophospholipid pathway. 4 dpt, cells were differentiated using IL-4 (20 ng/ml) and PMA (20 ng/ml) for 4 days prior to infection. b Knockdown efficiency was validated using RT-qPCR (mean ± SEM, n = 2 (shPTDSS1, shPTDSS2), n = 3 (shPISD, shPLD2), n = 4 (shCDIPT), or n = 5 (shPLD1) independent experiments). c Differentiation into CD209-positive THP-1 cells was confirmed 4 days after IL-4 and PMA treatment by flow cytometry using a CD209-specific antibody. Shown is one representative experiment. d Cell viability was analyzed 8 dpt and 4 days post differentiation (mean ± SEM, n = 3 independent experiments). e Knockdown cells were infected with YFV-17D (MOI 1) and the amount of E protein was analyzed by immunoblotting at 2 dpi. Shown is one representative blot of three independent experiments. f Infectivity in the supernatant of DENV (MOI 2) and YFV-17D (MOI 1) infected cells was determined using TCID50 assays at 2 dpi (mean ± SEM, n = 4 (YFV-17D) or n = 3 (DENV) independent experiments, except for shCDIPT: one experiment excluded due to contamination, *p ≤ 0.05, unpaired two-tailed Mann–Whitney U test). dpt days post transduction, dpi days post infection. Source data are provided as a Source Data file. Taken together, we observed the dependency of orthoflavivirus infection and replication on monounsaturated FA production through SCD, PS turnover, and de novo synthesis of PI. Inhibiting the synthesis of Cer, which are accumulating during infection, has opposing effects depending on the target enzyme: SPTi that targets the rate-limiting step of sphingolipid de novo synthesis impairs virus-induced cell death but has only antiviral activity against DENV, while CerSi, that blocks both the de novo and the salvage pathway, boosts ZIKV and YFV-17D replication and the cytopathic effect of both viruses and WNV. Effects of diminished PS and PI metabolism were substantiated by experiments in HMC3 microglia and differentiated THP-1 cells. Previous studies investigated the lipidome of flavivirus infection using unbiased mass spectrometry but mostly focused on one prototype virus in one cell line. Here we compared the lipidome in a time course of infection with five different pathogenic orthoflaviviruses that cover viscerotropic (DENV and YFV) and neurotropic (WNV, TBEV, and ZIKV) viruses transmitted by mosquitoes or ticks. In order to allow for comparison, we chose Huh7 cells that are permissive for all these orthoflaviviruses. We observed a strong remodeling of lipid metabolism during infection with some trends observed in all viruses, and other changes being specific to certain viruses. Fourty-eight hours post infection, when more than 80% of cells were infected, principal component as well as hierarchical cluster analysis clearly indicated distinct lipidome changes in cells infected with DENV, WNV, and TBEV, whereas ZIKV and YFV-17D diverged less from the controls and overlapped with each other. One of the most striking phenotypes we observed was the increase in Cer and some HexCer species at late time points after infection. This phenomenon has been observed previously and linked to a proviral role of Cer during ZIKV infection. Interestingly, we only detected the accumulation in Cer for viruses that are cytopathic in Huh7 cells (ZIKV, WNV, DENV, and YFV) but not for TBEV, which replicates to high levels, but is not cytotoxic. In addition, in follow-up experiments, the inhibition of Cer synthesis was not broadly antiviral. In detail, SPT inhibition that targets the rate-limiting step of sphingolipid de novo synthesis reduced virus cytotoxicity and DENV infection. In contrast, CerS inhibition that blocks both the de novo and the salvage pathway boosted viral replication and cell death, especially for ZIKV and YFV-17D. These results are contrary to what has been shown in ZIKV infection in earlier studies despite similar experimental conditions. One difference however relates to the ZIKV strain, i.e., the Brazilian isolate strain FSS1302536 compared to the PRVABC59 strain produced using reverse genetics. Inhibition of Cer biosynthesis has been additionally investigated for WNV and DENV replication in Vero cell lines. Here, opposing effects were observed with lower concentrations of the SPT inhibitor fumonisin B1 slightly increasing DENV titers and not affecting WNV. However, in higher concentrations, like the ones we used, SPTi had no proviral function. In our hands high concentrations decreased DENV titers and cytopathic effects, clearly indicating an antiviral effect in DENV infection that was not observed for WNV. Inhibition of CerS increased WNV-induced CPE but had no effect on DENV in contrast to earlier reports where CerSi increased DENV but decreased WNV titers. Interestingly, it has also been shown that SMs are enriched in WNV virions and inhibition of knockdown of SMase reduced viral titers. As we did detect remarkable differences between the chosen orthoflaviviruses, the strain used for analysis as well as the cell lines could of course influence the results. For DENV infection some studies analyzed lipid metabolism in mosquito cells, the midgut of mosquitoes, and complete mosquitoes. The increase in Cer and lyso-PL as well as decrease in TAG is a phenotype that is not restricted to mammalian cells, but in a similar manner observed in mosquito cells. In contrast, in Huh7 cells SM and DAG species tend to decrease in abundance in response to DENV infection while they increase in mosquito cells. Thus, lipidomic remodeling induced by DENV infection is in part independent of the host cells. In whole mosquitoes, metabolomics and lipidomics indicated that DENV replication causes phospholipid remodeling. Early in infection several PL and lyso-PL species increased while later in infection those levels tended to decrease, a phenomenon similarly noted in our study in Huh7 cells. On a molecular level this was linked to decreased 1-acylglycerol-3-phosphate O-acyltransferase 1 (AGPAT1) expression in DENV-infected mosquitoes. In our analysis, we noticed a slight ~0.5-fold decrease in AGPAT1 expression 24 and 48 hpi and a slight increase in abundance of some PA species but were unable to detect lyso-PA species. However, the dependency on PLD2 expression that might also lead to increased PA levels points to similarities between mammalian cells and mosquitoes. In general, lyso-PL species, especially LPI levels, were elevated early in infection but decreased during the course of infection for all orthoflaviviruses analyzed, suggesting elevated PLA2 activities required for initiation of replication. For WNVKUN a similar accumulation of LPC was detected 24 hpi and treatment with a broad PLA2 inhibitor decreased WNVKUN replication, likely through inhibition of RO formation. In our study treatment with a broad cPLA2 inhibitor slightly increased ZIKV and decreased DENV titers but had no effect on WNV NY99, TBEV, or YFV-17D, indicating the effects observed for WNVKUN might not be generalizable for other orthoflaviviruses. Inhibition of cPLA2 or knockdown of PLA2G4C also inhibited HCV replication and morphogenesis while knockdown of PLA2G4A or treatment with a different broad inhibitor essentially blocked HCV virion formation, and this block was rescued by addition of arachidonic acid. Interestingly, in the same study decreased DENV titers were noted after cPLA2 inhibition, but the rescue of virion production was only minor, indicating a specific function of the lysolipids in DENV infection. SCD inhibitors potently block HCV replication by inhibiting the formation of ROs. Regarding orthoflaviviruses, inhibition of SCD was also linked to lower levels of DENV, ZIKV, and JEV replication and the formation of DENV particles. While in our hands the titers were marginally affected by SCD inhibition, the cell pathogenicity of all viruses was strongly reduced. As for most of the enzymes that were identified using the BioPan analysis no specific inhibitors are readily available, knockdown approaches were used to identify common or distinct virus dependencies. Knockdown of PTDSS2, but not PTDSS1, reduced viral titers most strongly for WNV and DENV and reduced the CPE of all investigated orthoflaviviruses. Both enzymes synthesize PS but have different substrate specificities, as PTDSS1 prefers PC over PE whereas PTDSS2 is specific for PE and in addition has a high affinity for DHA. Simultaneous depletion of PTDSS1 and PTDSS2 has been previously shown to negatively impact DENV infection and linked to the entry process, although lowered PS levels could additionally contribute to decreased infectivity of viral particles as they mediate particle entry by apoptotic mimicry. We demonstrate that interference with PS synthesis in the host cells through PTDSS depletion does not impact orthoflavivirus entry or vRNA replication and translation but virion morphogenesis or secretion. It remains to be elucidated if the incorporation of PS in the lipid envelope is altered by the knockdown of PTDSS1 or 2 and thereby reduces virus binding to TIM and TAM receptors, which are known to facilitate orthoflavivirus infection. In our study, not only inhibition of PS synthesis, but also prevention of PS turnover catalyzed by PISD led to reduced viral titers, although only slightly rescuing cells from CPE induced by the infection. These data indicate that the balance of synthesis and turnover is critical to successful infection and replication. Again, WNV and DENV were the most susceptible orthoflaviviruses. De novo synthesis of PI by CDIPT and CDS1 and 2 is required for WNV, TBEV, ZIKV, and DENV infection, as both virus titers and virus-induced CPE in CDIPT-, and to a lower extent and less broadly, in CDS1- and CDS2-knockdown cells decreased. For the knockdown of CDIPT we also detected significantly lower viral E protein in the supernatant of WNV-infected cells implying the importance for the late steps in orthoflavivirus replication cycle. These findings correlate with increased PA to PI reaction activities in all orthoflavivirus infections. While neither CDIPT nor CDS enzymes have been implicated in orthoflavivirus infection, PI and its phosphorylated forms, phospho-inositides (PIPs), have plenty of different functions in replication cycles of a broad range of viruses. One limitation of our lipidomic study is that we did not analyze PIPs and thus cannot correlate the requirement of intact PI biosynthesis to an increase in different PIP species. One phenotype we observed, however, was the enrichment of LPI species common to all investigated viruses early in infection indicating that LPI might be important to establish infection. We validated our findings in other cell types. ZIKV, WNV, and TBEV have a neurotropism, therefore we used immortalized human microglia cells (HMC3), a cell type that is important for orthoflavivirus infection. Depletion of PTDSS1 and PTDSS2 in HMC3 cells decreased orthoflavivirus replication to an even stronger extent compared to Huh7 cells, highlighting a greater dependency of the neurotropic viruses we investigated on PS in more relevant non-cancer cells. Furthermore, our data indicate that the role of these enzymes is cell type independent. Correlating to the role of CDIPT in Huh7 cells, defects in PI synthesis impair WNV and TBEV infection, whereas ZIKV replication is not affected. It has been shown that ZIKV infection in microglia leads to altered levels of numerous metabolites, including lyso-phospholipids and phospholipids. Here we demonstrate not only the importance of lipid metabolizing enzymes for ZIKV infection but also for WNV and TBEV infection in microglia cells. For YFV-17D and DENV we confirmed our results in differentiated THP-1 cells that express DC-SIGN. Overall, depletion of glycerophospholipid remodeling enzymes remarkably reduced DENV and YFV-17D infection rates in our differentiated THP-1 cells, confirming the relevance of a functional glycerophospholipid metabolism in orthoflavivirus infection. Taken together, lipid remodeling by different orthoflaviviruses includes distinct changes but also common themes shared by a variety of viruses that are needed for efficient infection and replication. Inhibition of monounsaturation of FAs rescued from virus-induced cell death and depletion of enzymes involved in PS and PI biosynthesis, as well as PS turnover was detrimental to orthoflavivirus infection in different cell types. Interestingly, inhibition of Cer synthesis had opposing effects depending on the target enzyme. HEK293T and Vero E6 were obtained from the American Type Culture Collection, BHK21 from C. Munoz-Fontela, Huh7 from R. Bartenschlager, HMC3 from A. Slowik, and THP-1 from the German Collection of Microorganisms. All cells were grown under standard cell culture conditions in either RPMI (THP-1) or high glucose DMEM (all other cells) supplemented with 10% FCS and GlutaMax, as well as Pen/Strep (Gibco). Cell cultures were frequently tested for mycoplasma contaminiation. Cell viability was determined by CellTiter96 Aqueous One Solution Reagent (Promega). Transfection for lentivirus production was performed by calcium phosphate precipitation. The following plasmids were described previously: pFK-DVs encoding DENV-2 16681, pACNR-FLYFV-17Da encoding YFV-17D, pWN-AB and pWN-CG encoding WNV NY99, pJW231 and pJW232 encoding ZIKV PRVABC59, and pTNd/c encoding TBEV Neudörfl strain. Lentiviral shRNA constructs were cloned into pSicoR-MS1 as described using the primers listed in Supplementary Data 3 as described previously. The following antibodies were obtained commercially: orthoflavivirus group antigen D1-4G2-4-15 (Novus Biologicals, 1:100 IF, 1:1000 WB), GAPDH (Santa Cruz, 1:500 WB), Alexa 555-conjugated IgG (donkey, (H + L), Invitrogen, 1:2000 IF), HRP-labeled secondary antibodies (Jackson Laboratories, 1:10000 WB), and FITC-conjugated CD209 antibody (Thermo Fisher, 1:20). FIPI (PLDi, 7.5 mM stock in DMSO, f.c. 150 nM) was purchased from Merck, fumonisin B1 (CAY62580-1, CerSi, 5 mM stock in DMSO, f.c. 10 µM) and myriocin (CAY63150-1, SPTi, 15 mM stock in DMSO, f.c. 30 µM) from Biomol, ASB14780 (cPLA2i 20 mM stock in DMSO, f.c. 5 µM) from Biotechne, and A939572 (SCDi,10 mM stock in DMSO, f.c. 5 µM) from MedChem. Of note, as myriocin is not fully soluble at 15 mM in DMSO, stocks were thawed at room temperature, then heated for 15 min at 55 °C immediately prior to addition to media. Enzymes for molecular cloning were from New England Biolabs, cell culture reagents from Gibco/Life Technologies, and fine chemicals, if not noted otherwise, from AppliChem or Sigma. LDs were stained with BODIPY493/503 (Life Technologies). For the preparation of DENV, TBEV, and YFV-17D viral stocks, respective plasmids were digested for linearization and purified by phenol-chloroform extraction. For WNV and ZIKV a two-plasmid system with an overlap PCR strategy was used with the primers listed in Supplementary Data 3. The overlap PCR products were purified by gel extraction. For in vitro transcription we used the MegaScript SP6 Transcription Kit and Cap Analog (m7G(5′)ppp(5′)G) (Thermo Fisher) for DENV and YFV-17D, or HiScribe T7 ARCA mRNA Kit (NEB) for all other viruses. Virus stocks were produced in BHK21 cells (TBEV) or Vero E6 cells (all other viruses). 4 × 10 cells were washed in Opti-MEM (Life Technologies), and resuspended in 400 µl cytomix buffer (120 mM KCl, 5 mM MgCl2, 0.15 mM CaCl2, 2 mM EGTA, 1.9 mM ATP, 4.7 mM GSH, 25 mM HEPES, 10 mM potassium phosphate buffer, pH 7.6). The cell suspension was added to 10 µg in vitro transcribed orthoflavivirus RNA and pulsed at 260 V and 950 µF using the Gene Pulser II (Biorad). For P1 virus stock production, naïve BHK21 or Vero E6 cells were infected with P0 virus stock. For infection of Huh7, virus stocks were diluted in DMEM containing 10% FBS and added to cells for 1 h if not stated otherwise. Viral titers were determined with tissue culture infectious dose titration (TCID50) using Huh7 cells for virus stocks (except TBEV) and BHK21 cells for cell supernatants in infection experiments and TBEV viral stocks. Briefly, cells were seeded in 96-well plates and 1 day later infected with viral stocks or infectious supernatant in a serial dilution of 1:10–1:10. Five days post infection cells were fixed and stained with crystal violet. TCID50/ml was calculated with the Reed and Muench calculator. For each experiment 2.1–2.7 × 10 cells were harvested and inactivated with 640 µl acidified methanol (3% acetic acid, 0.3 μg/ml butylated hydroxytoluene) and 320 µl chloroform. The samples were dried and afterwards a mix of internal standards (Table S1) was added prior to a methyl tert-butyl ether extraction. Organic phases were pooled, and lipids were dried using a vacuum concentrator and resuspended in 100 µL MS-storage solution chloroform/methanol/water (60/30/4.5; v/v/v) until data acquisition. Cholesterol was determined after acetylation as described earlier. Shotgun lipidomics measurements were performed as described earlier using a Q Exactive Plus (Thermo Fisher Scientific) mass spectrometer coupled with the TriVersa NanoMate (Advion). For the chip-based nano-ESI measurements, a flow rate of ~300 nl/min is achieved using 1.1 kV spray voltage and back pressure of 1.1 psi. Data acquisition for one sample was performed over 10 min starting with positive ion mode and switching to negative ion mode after 5 min. Tandem mass spectrometric experiments were performed with a precursor selection in unit resolution using a predefined inclusion list. MS experiments were recorded with a resolution of 280,000 (at m/z 200, FWHM) and MS with a resolution of 70,000 (at m/z 200, FWHM) in both ionization modes. For tandem mass spectrometric experiments, a normalized collision energy of 30 was utilized (details on scan events and inclusion list are summarized in Supplementary Data 4). Lipid identification was performed with LipidXplorer 1.2.8 and post processing including quality control and quantitation was executed with lxPostman. The data processing pipeline including LipidXplorer import settings, MFQL-queries and post-processing parameters will be made available via LipidCompass (accession number LCE00000015) and are also summarized in the report file (Supplementary Data 5) according to the minimal reporting checklist. All lipid amounts were either normalized to the total lipid content (mol%) or protein content of the extracted cells (pmol/µg protein) as indicated in the text and figure legends and are the basis for further multivariate data analyses (Supplementary Data 1). FA content of PE, PC, PS, and TAG were determined with a dedicated data analysis strategy using LipidXplorer and lxPostman. FA content of TAGs was determined by the specific neutral loss observed for ammonium adducts in positive ion mode as described earlier. For PE, PC, and PS the fragmentation in negative ion mode was utilized to determine lipid species that were evaluated with lxPostman according to the intensity ratio of sn 1/2 FA signals of the respective internal standard. In the next step, the responses of the FA were normalized according to the lipid class-specific internal standard and the FA 18:1D7 (Table S1). Finally, all FA of the specific class were summed up and calculated in mol% (Supplementary Data 2). Total protein content was determined from the dried pellet of the water phase after extraction using the bicinchoninic acid assay (Thermo Fisher Scientific). Pellets were resuspended in 150 μl sample buffer, mixed well, and sonicated for 2 min. After a short centrifugation step, the samples were incubated for 30 min at 99 °C and shaking at 1400 rpm (Thermomixer, Eppendorf) and were sonicated every 5 min for 30 s. After cooling to 25 °C, the assay was carried out according to the manufacturer’s instructions and absorbance was measured at 562 nm with an Infinite M200 Pro plate reader (Tecan). For transcriptomic analysis, total RNA was isolated using TriReagent (Sigma-Aldrich) following the manufacturer’s protocol followed by DNA digestion with the TURBO DNA-free DNase kit (Ambion). Concentration of RNA was measured on an Invitrogen Qubit 3 Fluorometer using Qubit RNA High Sensitivity (HS) Kit (Thermo Fisher Scientific Inc.). Quality was assessed using the Bioanalyzer 2100 instrument with an RNA Nano 6000 Kit (Agilent), followed by depletion of ribosomal rRNA with the NEBNext rRNA Depletion Kit v2 (Human/Mouse/Rat) (New England Biolabs) and Agencourt RNAClean XP beads (Beckman Coulter). Library preparation was performed according to the manufacturer’s protocol by using NEBNext Ultra II Directional RNA Library Prep Kit for Illumina. To barcode the libraries and therefore facilitate multiplex sequencing, Multiplex Oligos for Illumina (96 Unique Dual Index Primer Pairs) (New England Biolabs) were used. Libraries were further purified with Agencourt AMPure XP Reagent (Beckman Coulter), and quality and quantity of the amplified cDNA were assessed with a DNA High Sensitivity Kit (Agilent). After pooling, denaturing, and diluting the libraries, 150 bp long paired-end reads were sequenced on the NovaSeq system (Illumina). The generated sequence reads were basecalled, and demultiplexed prior to analysis. All sequence reads obtained were filtered by mapping against their respective virus genome (DENV 2 (NC_001474.2); TBEV (NC_001672.1); WNV isolate NY99 (MZ605381.2); YFV-17D/Tiantan (FJ654700.1); ZIKV strain PRVABC59 (KU501215.1)), before aligning them to the human genome sequence GRCh38.p14 (RefSeq GCF_000001405.40) using STAR. Aligned sequence reads were sorted by position using SAMtools, and sequence read counts per gene or transcript were calculated for each library using the subread function featureCounts. Differential gene expression analyses were performed with the R-package DESeq2. Data were normalized per regularized log transformation and shrunk using the adaptive shrinkage estimator for the ashr package. Genes were considered differentially expressed when the adjusted p value was below 0.05 and the log2 fold change was above 1 or below −1. Lentiviral particles were produced as previously described. Briefly, 293T cells were co-transfected with the pSicoR-MS1 shRNA constructs, a packaging construct (pCMVΔR8.91), and a construct expressing the glycoprotein of vesicular stomatitis virus (VSV-G) (pMD.G). Pseudotyped lentiviral particles were concentrated using ultracentrifugation. Lentiviral transductions were carried out in the presence of 4 µg/ml polybrene (Sigma). Lentiviral stocks were titrated on Huh7 cells using flow cytometry (Guava easyCyte HT Cytometer) or microscopy. Huh7 cells were either transduced with constructs for shRNA targets or treated with inhibitors. Four days post transduction (shRNAs) or 2 d post treatment (inhibitors) cells were infected with the different orthoflaviviruses. Cells were fixed 3 dpi and stained with crystal violet to visualize CPE. Plates were scanned on a ChemiDoc imaging system (BioRad) and CPE quantified using ImageJ. Cells were lysed in RIPA buffer (150 mM NaCl, 50 mM Tris-HCl pH 7.6, 1% NP-40, 0.5% sodium deoxycholate, 5 mM EDTA, protease inhibitor cocktail (Sigma), PMSF (AppliChem)) for 1 h on ice. Clarified lysates or culture supernatants were subjected to SDS–PAGE followed by blotting onto a nitrocellulose membrane (Amersham Protran, Cytiva). Following staining with antibodies, protein bands were detected by chemiluminescence using Lumi-Light substrate (Roche) or SuperSignal West Femto (Thermo Fisher) on a ChemiDoc imaging system (BioRad). Band signal intensities were quantified using the quantification function of Image Lab (BioRad). vRNA from the cultures supernatant was isolated using NucleoSpin RNA Virus Kit (Machery Nagel) according to the manufacturer’s protocol. Total cellular RNA isolation was performed with TriReagent (Sigma-Aldrich). Cellular DNA was digested with rDNaseI using the TURBO DNA-free DNase kit (Ambion). RNA was transcribed into cDNA using random hexamers (Qiagen), SuperScript III reverse transcriptase (Thermo Fisher), and RNaseOut (Thermo Fisher). cDNA was subjected to qPCR using the Luna Universal qPCR master mix (New England Biolabs). For quantitative PCR, we used the primers listed in Supplementary Data 3 and Luna Universal qPCR master mix (New England Biolabs) on a StepOnePlus Real Time—qPCR Thermal Cycler (Applied Biosystems). For immunofluorescence microscopy, cells seeded on coverslips were infected and fixed at different time points in 4% paraformaldehyde. After permeabilization for 5 min in 0.1% Triton X-100 in PBS and incubation in blocking solution (5% BSA, 1% fish skin gelatin, 50 mM Tris in PBS), cells were stained overnight with primary antibodies, followed by Alexa Fluor-conjugated secondary antibodies. LDs were stained with BODIPY493/503, nuclei with Hoechst, and coverslips were embedded in Mowiol mounting media. Microscopy was performed on a Leica TCS SP5 II confocal laser scanning microscope or Leica Thunder DMI8 widefield microscope. For the quantification of LDs, we used cells that expressed a cell-based orthoflavivirus reporter to monitor infection. LDs were quantified using the particle analyzer function of Fiji. Viral entry efficacy was determined as described with minor modifications: Huh7 cells were transduced with lentiviral particles and seeded at 4 dpt. At 5 dpt, the media was changed to DMEM supplemented with 3% FCS and the cultures were pre-cooled for 10 min at 4 °C prior to infection with ZIKV (MOI 2), WNV (MOI 5), DENV (MOI 2), TBEV (MOI 10), and YFV-17D (MOI 5). The inoculum was removed after 1 h, cells were washed with cold DMEM containing 10% FCS and shifted to 37 °C. Four hours post infection, the cells were washed 1× with DMEM and 1× with PBS. Following incubation with pre-warmed trypsin for 2 min at RT the cells were washed again with DMEM and PBS prior to lysis with TriReagent (Sigma-Aldrich). RNA Isolation and RT-qPCR were performed as described above. THP-1 were spin-infected in 24-well plates in RPMI supplemented with polybrene with lentiviral particles by centrifugation at 500 × g for 1 h followed by incubation with the inoculum o/n. At 4 dpt cells were seeded and IL-4 and PMA were added to initiate differentiation as described (final concentration: IL-4 (20 ng/ml), PMA (20 ng/ml)). After 4 days, the presence of the receptor DC-SIGN/CD209 was confirmed by flow cytometry and cells were used for infection experiments with YFV-17D and DENV. For flow cytometry, cells were trypsinized and fixed with 4% paraformaldehyde, washed twice with PBS/glycine, blocked with 5% BSA, 1% fish skin gelatin, 50 mM Tris in PBS, and stained with FITC-conjugated CD209 antibody (Thermo Fisher) for 1 h at 4 °C while rotating. Analysis was performed on a Guava easyCyte HT flow cytometer and data was analyzed using FlowJo. R and RStudio were used to visualize data and for statistical analysis. Statistical analysis was performed using unpaired two-tailed Welch’s unequal variances t-test, Mann–Whitney U test, or, in case of normalized data, one sample Student’s t-test, and adjusted for multiple comparison with the false discovery rate (FDR) as indicated in the figure legends. Sample sizes (n) represent independent experiments, if not stated otherwise. The data were analyzed and visualized using several packages available for R, namely, gdata (version 3.0.0), dplyr (version 1.1.4), tidyr (version 1.3.0), factoextra (version 1.0.7), ggplot2 (version 3.4.4), pheatmap (version 1.0.12), cowplot (version 1.1.1), ggrepel (version 0.9.4), purrr (version 1.0.2), coin (version 1.4-3), ggsignif (version 0.6.4), ggbeeswarm (version 0.7.2), as well as BioPan (https://www.lipidmaps.org/biopan/), LipidLynxX, and LION (http://www.lipidontology.com) with missing data imputed with MetImp 1.2 (https://metabolomics.cc.hawaii.edu/software/MetImp/) using MCAR/MAR method RF.
|
PMC10491837
|
Differential intracellular management of fatty acids impacts on metabolic stress-stimulated glucose uptake in cardiomyocytes
|
Stimulation of glucose uptake in response to ischemic metabolic stress is important for cardiomyocyte function and survival. Chronic exposure of cardiomyocytes to fatty acids (FA) impairs the stimulation of glucose uptake, whereas induction of lipid droplets (LD) is associated with preserved glucose uptake. However, the mechanisms by which LD induction prevents glucose uptake impairment remain elusive. We induced LD with either tetradecanoyl phorbol acetate (TPA) or 5-aminoimidazole-4-carboxamide-1-β-D-ribofuranoside (AICAR). Triacylglycerol biosynthesis enzymes were inhibited in cardiomyocytes exposed to FA ± LD inducers, either upstream (glycerol-3-phosphate acyltransferases; GPAT) or downstream (diacylglycerol acyltransferases; DGAT) of the diacylglycerol step. Although both inhibitions reduced LD formation in cardiomyocytes treated with FA and LD inducers, only DGAT inhibition impaired metabolic stress-stimulated glucose uptake. DGAT inhibition in FA plus TPA-treated cardiomyocytes reduced triacylglycerol but not diacylglycerol content, thus increasing the diacylglycerol/triacylglycerol ratio. In cardiomyocytes exposed to FA alone, GPAT inhibition reduced diacylglycerol but not triacylglycerol, thus decreasing the diacylglycerol/triacylglycerol ratio, prevented PKCδ activation and improved metabolic stress-stimulated glucose uptake. Changes in AMP-activated Protein Kinase activity failed to explain variations in metabolic stress-stimulated glucose uptake. Thus, LD formation regulates metabolic stress-stimulated glucose uptake in a manner best reflected by the diacylglycerol/triacylglycerol ratio.Type 2 Diabetes Mellitus is a serious metabolic syndrome, the incidence of which is dramatically increasing worldwide. It makes up about 90–95% of total diabetes cases and is caused by insulin resistance combined with compromised activity of the pancreatic β-cells, which leads to hyperglycemia and subsequent glucose toxicity. Diabetic patients have increased myocardial infarction incidence and severity. Stimulation of myocardial glucose uptake during ischemia improves the post-ischemic recovery of function and survival of the myocardium. Thus, impairment of myocardial metabolic stress-stimulated glucose uptake could jeopardize the myocardial response to ischemic event and reperfusion injury. A hallmark of type 2 Diabetes Mellitus is hyperlipidemia. Chronic exposure of cardiomyocytes to fatty acids (FA) or to triglyceride-rich lipoproteins particles results in marked inhibition of basal, insulin-stimulated and metabolic stress-stimulated glucose uptake. The mechanisms involved in the reduction of metabolic stress-stimulated glucose uptake by FA exposure involve inactivation of focal adhesion kinase (FAK) and chronic activation of protein kinase C δ (PKCδ). Chronic treatment with tetradecanoyl phorbol acetate (TPA) restores FAK activity and completely restores metabolic stress-stimulated glucose uptake in cardiac myocytes exposed to FA. TPA is a diacylglycerol analog able to activate classic and novel PKC; however chronic exposure to TPA also triggers a marked downregulation of several PKC isoforms (PKCα, PKCδ and PKCε), and impairs the translocation of PKCδ to the membrane fraction. TPA treatment also increases the biogenesis of lipid droplets. In the same way, chronic treatment with 5-aminoimidazole-4-carboxamide 1-β-D-ribofuranoside (AICAR), an AMP analog precursor capable of activating AMP-activated Protein Kinase (AMPK), results in an improvement of metabolic stress-stimulated glucose metabolism associated with a redirection of FA away from fatty acid oxidation and towards incorporation into lipid droplets. Lipid droplets are intracellular organelles that store neutral lipids within cells. They are mainly composed of a neutral lipid core containing mostly triacylglycerol and cholesteryl esters, surrounded by a phospholipid monolayer and lipid droplet-coating membrane proteins. Skeletal muscle exhibits the “athlete’s paradox” as increased intramyocellular lipid concentration results in increased insulin responsiveness in trained skeletal muscle. This phenomenon can be explained by the incorporation of fatty acid moieties in triacylglycerol within lipid droplets, instead of being turned into potentially toxic lipid derivatives. Triacylglycerol biogenesis is regulated by glycerol-3-phosphate acyl transferases (GPAT), which catalyze the acylation of glycerol-3-phosphate (G3P) into lysophosphatidic acid (LPA), and by diacylglycerol acyltransferases (DGAT), which converts diacylglycerol into triacylglycerol. To understand the mechanisms for the protective effect of lipid droplets on glucose uptake, in this study we chronically inhibited lipid droplet biogenesis upstream (GPAT inhibition with FSG67) and downstream (DGAT inhibition with either A922500 or T863) of the diacylglycerol synthesis step in cardiomyocytes exposed to either FA only or to FA with lipid droplets induction (TPA or AICAR). In this way, we expected to either decrease (GPAT inhibition) or increase (DGAT inhibition) the intracellular diacylglycerol concentration. We thus investigated how the effects of GPAT or DGAT inhibition on intracellular lipid management in cardiomyocytes influence glucose uptake, intracellular diacylglycerol concentration and PKCδ activation. We started by testing the effects of different perturbations in triacylglycerol metabolism on the biogenesis of lipid droplets. To do this, we inhibited enzymes acting either upstream (GPAT) or downstream (DGAT) of the diacylglycerol step in triacylglycerol biosynthesis and measured the effect on lipid droplet biogenesis (Fig. 1). We estimated the lipid droplet area as a percentage of total cell surface area. As previously observed, chronic treatment with either TPA or AICAR induced lipid droplet biogenesis in cardiomyocytes exposed to FA, whereas very few lipid droplets were observed in cardiomyocytes exposed to FA alone. Chronic inhibition of either GPAT with FSG67 (Fig. 1a) or DGAT with A922500 or T863 (Fig. 1b) significantly reduced lipid droplet biogenesis in cardiomyocytes exposed to FA and lipid droplets inducers (TPA or AICAR), thus showing that triacylglycerol synthesis was required for lipid droplets induction and confirming the efficacy of the selected GPAT and DGAT inhibitors.Figure 1DGAT or GPAT inhibition reduce lipid droplet biogenesis. Primary cardiomyocytes were cultured for 7 days with 0.4 mM fatty acids (FA), 100 nM TPA or 0.2 mM AICAR and the GPAT inhibitor FSG67 (30 µM; panel a) or the DGAT inhibitors A922500 (1 µM) or T863 (100 nM) panel (b) or dimethyl sulfoxide as the vehicle control. Cells were fixed, permeabilized and stained for F-actin (red), DNA (blue) and neutral lipids (green) and the relative density of lipid droplets was measured (right panels). Scale bar 20 µm. Left panel shows the position of the inhibited enzymes in the triacylglycerol biosynthesis pathway. Results are shown as mean ± SEM; n = 6–32. †: significant effect (q < 0.05) of either the GPAT or the DGAT inhibitors. DGAT or GPAT inhibition reduce lipid droplet biogenesis. Primary cardiomyocytes were cultured for 7 days with 0.4 mM fatty acids (FA), 100 nM TPA or 0.2 mM AICAR and the GPAT inhibitor FSG67 (30 µM; panel a) or the DGAT inhibitors A922500 (1 µM) or T863 (100 nM) panel (b) or dimethyl sulfoxide as the vehicle control. Cells were fixed, permeabilized and stained for F-actin (red), DNA (blue) and neutral lipids (green) and the relative density of lipid droplets was measured (right panels). Scale bar 20 µm. Left panel shows the position of the inhibited enzymes in the triacylglycerol biosynthesis pathway. Results are shown as mean ± SEM; n = 6–32. †: significant effect (q < 0.05) of either the GPAT or the DGAT inhibitors. We next sought to evaluate whether and how inhibition of lipid droplet biogenesis could influence glucose uptake. We first tested whether the two triacylglycerol synthesis inhibitors could influence glucose uptake in cardiomyocytes in the absence of FA or lipid droplets. Neither the GPAT inhibitor (FSG67), nor the DGAT inhibitors (A922500 and T863) modified basal, insulin- or metabolic stress-stimulated glucose uptake in cardiomyocytes not exposed to FA (Fig. S1). Having thus established that inhibitors of triacylglycerol biosynthesis had no impact on glucose uptake in the absence of FA, we next investigated the effect of GPAT inhibition on glucose uptake in cardiomyocytes exposed to FA, with or without induction of lipid droplets. Chronic exposure to FA alone markedly impaired basal and stimulated glucose uptake (Fig. 2a), which is consistent with our previous observations. GPAT inhibition in cardiomyocytes exposed to FA alone improved metabolic stress-stimulated glucose uptake (Fig. 2a). In cardiomyocytes with induced lipid droplets GPAT inhibition did not affect the glucose uptake in response to either insulin or oligomycin (Fig. 2b,c), although a trend towards an increase in metabolic stress-stimulated glucose uptake was observed. Basal and insulin-stimulated glucose uptake were not affected by GPAT inhibition. Overall, this demonstrates that inhibition of triacylglycerol synthesis upstream of the diacylglycerol step improved metabolic stress-stimulated glucose uptake in cardiomyocytes.Figure 2GPAT inhibition improves glucose uptake in cardiomyocytes exposed to FA: Primary cardiomyocytes were cultured for 7 days with or without 0.4 mM fatty acids (FA); lipid droplets were induced with either 100 nM TPA or 0.2 mM AICAR. Additional treatments were the GPAT inhibitor FSG67 (30 µM) or dimethyl sulfoxide as the vehicle control. Glucose uptake was then measured during 1 h exposure to either 1 µM insulin (blue bars), 1 µM oligomycin (red bars) or control (black bars). (a): effect of GPAT inhibition on glucose uptake in cardiomyocytes exposed to FA only. (b): effect of GPAT inhibition on glucose uptake in cardiomyocytes exposed to FA + TPA. (c): effect of GPAT inhibition on glucose uptake in cardiomyocytes exposed to FA + AICAR. Results are shown as mean ± SEM; n = 4–7. *: Significant effect (q < 0.05) of insulin or oligomycin; #: significant effect of FA. †: significant effect of the GPAT inhibitor FSG67. GPAT inhibition improves glucose uptake in cardiomyocytes exposed to FA: Primary cardiomyocytes were cultured for 7 days with or without 0.4 mM fatty acids (FA); lipid droplets were induced with either 100 nM TPA or 0.2 mM AICAR. Additional treatments were the GPAT inhibitor FSG67 (30 µM) or dimethyl sulfoxide as the vehicle control. Glucose uptake was then measured during 1 h exposure to either 1 µM insulin (blue bars), 1 µM oligomycin (red bars) or control (black bars). (a): effect of GPAT inhibition on glucose uptake in cardiomyocytes exposed to FA only. (b): effect of GPAT inhibition on glucose uptake in cardiomyocytes exposed to FA + TPA. (c): effect of GPAT inhibition on glucose uptake in cardiomyocytes exposed to FA + AICAR. Results are shown as mean ± SEM; n = 4–7. *: Significant effect (q < 0.05) of insulin or oligomycin; #: significant effect of FA. †: significant effect of the GPAT inhibitor FSG67. Since GPAT inhibition influenced only metabolic stress-stimulated glucose uptake, we focused our investigation on AMP-dependent Protein Kinase (AMPK) signaling. Indeed, AMPK is the energy gauge responsible for triggering the stress signal pathway leading to GLUT4 translocation. Therefore, AMPK is activated in response to the intracellular ATP deprivation that occurs during ischemia in vivo or during oligomycin exposure in our in vitro model. We assessed AMPK activation by evaluating the phosphorylation of AMPKα subunit on the T residue and the phosphorylation of the AMPK substrate raptor on the S residue using phosphorylation-specific antibodies. Oligomycin-induced metabolic stress significantly increased both TAMPKα and Sraptor phosphorylation in all experimental conditions (Fig. 3). GPAT inhibition in cardiomyocytes exposed to FA only increased TAMPKα phosphorylation (Fig. 3a). Chronic treatment with either TPA or AICAR, which restored glucose transport and induced lipid droplet formation, significantly reduced TAMPKα phosphorylation (Fig. 3c,e). In contrast, inhibition of lipid droplet biogenesis by GPAT inhibition restored TAMPKα phosphorylation. These effects on ThrAMPKα phosphorylation were not entirely reflected in its activity, as neither lipid droplet induction nor GPAT inhibition had any effect on Sraptor phosphorylation cardiomyocytes (Fig. 3b,d,f). The Rab GTPase-activating protein AS160 participates in the regulation of GLUT4 translocation and is also a substrate for AMPK, which phosphorylates AS160 on several residues recognized by the S/T-phosphorylated Akt substrate antibody. Phosphorylation of AS160 was very variable, but generally followed a pattern similar to that of TAMPKα (Fig S2a). Therefore, these data indicated that changes in AMPK phosphorylation and activity did not readily explain the improved metabolic stress-stimulated glucose uptake upon GPAT inhibition.Figure 3The presence of lipid droplets induced by TPA or AICAR reduces AMPK phosphorylation but not its activity. Primary cardiomyocytes were cultured for 7 days with or without 0.4 mM fatty acids (FA); lipid droplets were induced with either 100 nM TPA or 0.2 mM AICAR. Additional treatments were the GPAT inhibitor FSG67 (30 µM) or dimethyl sulfoxide as the vehicle control. Cardiomyocytes were then stimulated for 10 min with 1 µM insulin, for 20 min with 1 µM oligomycin or left unstimulated. Cells were extracted and submitted to western blot analysis to measure the expressions of phosphorylated (T) and total AMPKα phosphorylation (panels a,c,e) and of phosphorylated (S) and total raptor (panels b,d,f). The ratios of phosphorylated/total protein were than calculated and are displayed in the graphs. Right panels show western blots representative of the quantitative analysis in the graphs; the dashed line indicates where an image has been cut and spliced. Results are shown as mean ± SEM; n = 4–7. *: significant effect (q < 0.05) of oligomycin; §: significant effect of TPA or AICAR. †: significant effect of the GPAT inhibitor FSG67. The presence of lipid droplets induced by TPA or AICAR reduces AMPK phosphorylation but not its activity. Primary cardiomyocytes were cultured for 7 days with or without 0.4 mM fatty acids (FA); lipid droplets were induced with either 100 nM TPA or 0.2 mM AICAR. Additional treatments were the GPAT inhibitor FSG67 (30 µM) or dimethyl sulfoxide as the vehicle control. Cardiomyocytes were then stimulated for 10 min with 1 µM insulin, for 20 min with 1 µM oligomycin or left unstimulated. Cells were extracted and submitted to western blot analysis to measure the expressions of phosphorylated (T) and total AMPKα phosphorylation (panels a,c,e) and of phosphorylated (S) and total raptor (panels b,d,f). The ratios of phosphorylated/total protein were than calculated and are displayed in the graphs. Right panels show western blots representative of the quantitative analysis in the graphs; the dashed line indicates where an image has been cut and spliced. Results are shown as mean ± SEM; n = 4–7. *: significant effect (q < 0.05) of oligomycin; §: significant effect of TPA or AICAR. †: significant effect of the GPAT inhibitor FSG67. In a second approach we blocked lipid droplet biogenesis by inhibiting DGAT, the most downstream enzyme in the biosynthesis of triacylglycerol, with two structurally unrelated pharmacological inhibitors A922500 and T863. In contrast to GPAT inhibition, chronic DGAT inhibition significantly impaired the metabolic stress-stimulated, but not the insulin-stimulated, glucose uptake in cardiomyocytes exposed to either FA + TPA (Fig. 4a) or FA + AICAR (Fig. 4b). In cardiomyocytes exposed to FA only, chronic DGAT inhibition did not show any significant effect glucose uptake (Fig. 4c). Thus, in cardiomyocytes with induced lipid droplets biogenesis, DGAT inhibition has an effect on metabolic stress-stimulated glucose uptake opposite that of GPAT inhibition.Figure 4DGAT inhibition reduces glucose uptake in cardiomyocytes displaying lipid droplets. Primary cardiomyocytes were cultured for 7 days with 0.4 mM fatty acids (FA) in the presence or absence of 100 nM TPA or 0.2 mM AICAR. Additional treatments were the DGAT inhibitors A922500 (1 µM) or T863 (100 nM) or dimethyl sulfoxide as the vehicle control. Glucose uptake was then measured during 1 h exposure to either 1 µM insulin (blue bars), 1 µM oligomycin (red bars) or control (black bars). (a) Effect of DGAT inhibition on glucose uptake in cardiomyocytes exposed to FA + TPA. (b) Effect of DGAT inhibition on glucose uptake in cardiomyocytes exposed to FA + AICAR. (c) Effect of DGAT inhibition on glucose uptake in cardiomyocytes exposed to FA alone. Results are shown as mean ± SEM; n = 3–11. *: significant effect (q < 0.05) of insulin or oligomycin; #: significant effect of FA. †: significant effect of either DGAT inhibitor. DGAT inhibition reduces glucose uptake in cardiomyocytes displaying lipid droplets. Primary cardiomyocytes were cultured for 7 days with 0.4 mM fatty acids (FA) in the presence or absence of 100 nM TPA or 0.2 mM AICAR. Additional treatments were the DGAT inhibitors A922500 (1 µM) or T863 (100 nM) or dimethyl sulfoxide as the vehicle control. Glucose uptake was then measured during 1 h exposure to either 1 µM insulin (blue bars), 1 µM oligomycin (red bars) or control (black bars). (a) Effect of DGAT inhibition on glucose uptake in cardiomyocytes exposed to FA + TPA. (b) Effect of DGAT inhibition on glucose uptake in cardiomyocytes exposed to FA + AICAR. (c) Effect of DGAT inhibition on glucose uptake in cardiomyocytes exposed to FA alone. Results are shown as mean ± SEM; n = 3–11. *: significant effect (q < 0.05) of insulin or oligomycin; #: significant effect of FA. †: significant effect of either DGAT inhibitor. As again only metabolic stress-stimulated glucose uptake was affected by chronic DGAT inhibition, we assessed AMPK signaling. Although GPAT inhibition and DGAT inhibition had similar effects on lipid droplets biogenesis, they showed opposite effects on metabolic stress-stimulated glucose uptake in cardiomyocytes with induced lipid droplets. Nevertheless, similar to the results obtained with chronic GPAT inhibition, we observed a reduction of TAMPKα phosphorylation in response to oligomycin in lipid droplet-displaying cardiomyocytes, which was partially restored by chronic DGAT inhibition (Fig. 5a,c). Again, none of the effects on ThrAMPKα phosphorylation were reflected in its activity, as neither lipid droplet induction nor DGAT inhibition had any effect on Sraptor phosphorylation in FA + TPA or FA + AICAR-treated cardiomyocytes (Fig. 5b,d). Phosphorylation of AS160 was very variable, but generally followed a pattern similar to that of TAMPKα (Fig. S2b). Thus, as with GPAT inhibition, changes in AMPK phosphorylation and activity upon DGAT inhibition did not correlate with metabolic stress-stimulated glucose uptake.Figure 5DGAT inhibition improves AMPK phosphorylation but not its activity. Primary cardiomyocytes were cultured for 7 days with or without 0.4 mM fatty acids (FA), 100 nM TPA (panels a,b) or 0.2 mM AICAR (panels c,d). Additional treatments were the DGAT inhibitors A922500 (1 µM) or T863 (100 nM) or dimethyl sulfoxide as the vehicle control. Cardiomyocytes were then stimulated for 10 min with 1 µM insulin, for 20 min with 1 µM oligomycin or left unstimulated. Cells were extracted and submitted to western blot analysis to measure the expressions of phosphorylated (T) and total AMPKα phosphorylation (panels a,c) and of phosphorylated (S) and total raptor (panels b,d). The ratios of phosphorylated/total protein were than calculated and are displayed in the graphs. Right panels show western blots representative of the quantitative analysis in the graphs; the dashed line indicates where an image has been cut and spliced. Results are shown as mean ± SEM; n = 6–12. *: significant effect (q < 0.05) of oligomycin; §: significant effect of TPA. †: significant effect of the DGAT inhibitor T863. DGAT inhibition improves AMPK phosphorylation but not its activity. Primary cardiomyocytes were cultured for 7 days with or without 0.4 mM fatty acids (FA), 100 nM TPA (panels a,b) or 0.2 mM AICAR (panels c,d). Additional treatments were the DGAT inhibitors A922500 (1 µM) or T863 (100 nM) or dimethyl sulfoxide as the vehicle control. Cardiomyocytes were then stimulated for 10 min with 1 µM insulin, for 20 min with 1 µM oligomycin or left unstimulated. Cells were extracted and submitted to western blot analysis to measure the expressions of phosphorylated (T) and total AMPKα phosphorylation (panels a,c) and of phosphorylated (S) and total raptor (panels b,d). The ratios of phosphorylated/total protein were than calculated and are displayed in the graphs. Right panels show western blots representative of the quantitative analysis in the graphs; the dashed line indicates where an image has been cut and spliced. Results are shown as mean ± SEM; n = 6–12. *: significant effect (q < 0.05) of oligomycin; §: significant effect of TPA. †: significant effect of the DGAT inhibitor T863. Novel PKCs are kinases which are completely dependent on diacylglycerol for their activity; once activated they translocate to the membrane fraction where they can phosphorylate several proteins. In our previous study we identified the novel PKCδ as the isoform activated upon chronic FA exposure and, at least in part, involved in the reduction of glucose uptake. We analyzed association of PKCδ with the membrane fraction of cardiomyocytes, characterized by the enrichment of the membrane-associated protein connexin 43. Connexin 43 was selected as a convenient marker of the membrane fraction, being present in cardiomyocytes in both the plasma membrane and mitochondria, two compartments to which PKCδ translocates upon activation. We focused our analysis on a subset of conditions showing significant differences in metabolic stress-stimulated glucose uptake in response to inhibition of lipid droplet biogenesis, i.e. GPAT inhibition in cardiomyocytes exposed to FA, and DGAT inhibition (T863) in cardiomyocytes exposed to FA + TPA (Fig. 6). PKCδ expression (left panel) remained unchanged upon chronic FA exposure with or without GPAT inhibition. In contrast, as previously observed, chronic TPA treatment almost abolished PKCδ expression (Fig. 6). Chronic FA exposure induced translocation of PKCδ to the membrane fraction (right panel). However, GPAT inhibition significantly blunted the FA-induced translocation of PKCδ. Chronic TPA treatment abolished PKCδ translocation in cardiomyocytes exposed to FA; DGAT inhibition in addition to chronic TPA treatment had no effect. Thus, the improvement of metabolic stress-stimulated glucose uptake by GPAT inhibition in FA-exposed cardiomyocytes (Fig. 2a) was associated with prevention of PKCδ translocation.Figure 6GPAT inhibition prevents PKCδ activation by FA. Primary cardiomyocytes were cultured for 7 days with or without 0.4 mM fatty acids (FA), 100 nM TPA and the GPAT inhibitor FSG67, the DGAT inhibitor T863 or dimethyl sulfoxide. The membrane fraction was then isolated, and whole cell extracts and membrane fractions submitted to western blot analysis of PKCδ and connexin 43 expression. Top panels show western blots representative of the quantitative analysis in the graphs below; the dashed line indicates where an image has been cut and spliced; for connexin 43 expression, membranes had been cut horizontally below the 55 kDa molecular weight marker before incubation with the primary antibody. Left: whole cell extracts; right: membrane fractions. Results are shown as mean ± SEM; n = 4–5. #: significant effect of FA. §: significant effect of TPA; †: significant effect of the GPAT inhibitor. GPAT inhibition prevents PKCδ activation by FA. Primary cardiomyocytes were cultured for 7 days with or without 0.4 mM fatty acids (FA), 100 nM TPA and the GPAT inhibitor FSG67, the DGAT inhibitor T863 or dimethyl sulfoxide. The membrane fraction was then isolated, and whole cell extracts and membrane fractions submitted to western blot analysis of PKCδ and connexin 43 expression. Top panels show western blots representative of the quantitative analysis in the graphs below; the dashed line indicates where an image has been cut and spliced; for connexin 43 expression, membranes had been cut horizontally below the 55 kDa molecular weight marker before incubation with the primary antibody. Left: whole cell extracts; right: membrane fractions. Results are shown as mean ± SEM; n = 4–5. #: significant effect of FA. §: significant effect of TPA; †: significant effect of the GPAT inhibitor. To determine how changes in intracellular diacylglycerol and triacylglycerol contents could explain changes in glucose uptake and PKCδ translocation, we performed high-performance thin-layer chromatography (HPTLC) and shotgun mass spectrometry (MS) analyses of total cellular lipids with the same set of conditions as the PKCδ translocation experiments. Diacylglycerols were barely detectable, and not quantifiable by HPTLC. In MS analyses, exposure to FA led to an increased intracellular content of total diacylglycerol as compared to cardiomyocytes exposed to BSA alone, while GPAT inhibition reduced diacylglycerol content during FA exposure (Fig. 7a,b) Surprisingly, TPA treatment failed to reduce diacylglycerol content despite induced lipid droplet biogenesis. Furthermore, DGAT inhibition did not increase intracellular amount of diacylglycerol in FA + TPA-treated cardiomyocytes. Looking into the different diacylglycerol (DAG) species, we discovered that DAG(34:1) (comprising one palmitoyl and one oleyl chain) and DAG(36:2) (comprising two oleyl chains) were the most predominant in cardiomyocytes chronically exposed to FA, which were oleate and palmitate. In contrast, the most prominent diacylglycerol in control cardiomyocytes was DAG(38:4) (Fig. 7b). Both DAG(34:1) and DAG(36:2) were significantly increased in FA-exposed cardiomyocytes, and both were significantly reduced when GPAT activity was chronically inhibited in FA + FSG67-treated cardiomyocytes (Fig. 7b). Again, chronic TPA treatment failed to reduce different diacylglycerol species concentrations. Remarkably, when DGAT activity was inhibited in FA + TPA + T863-treated cardiomyocytes, thus preventing lipid droplets biosynthesis, the palmitate- and oleate-based diacylglycerols were significantly reduced (Fig. 7b), although there was no difference in total diacylglycerol amounts (Fig. 7a). In this condition we observed a significant increase in palmitate- and oleate-based phosphatidyl choline (Fig. S3), reflected by a similar trend in total palmitate- and oleate-based phospholipids. Thus, prevention of diacylglycerol accumulation by GPAT inhibition in FA-exposed cardiomyocytes correlated with reduced PKCδ translocation (Fig. 6) and improved metabolic stress-stimulated glucose uptake (Fig. 2a).Figure 7Analysis of cellular diacylglycerol and triacylglycerol levels and of FA metabolism in cardiomyocytes under GPAT or DGAT inhibition. Primary cardiomyocytes were cultured for 7 days with or without 0.4 mM fatty acids (FA), 100 nM TPA and the GPAT inhibitor FSG67, the DGAT inhibitor T863 or dimethyl sulfoxide. Panels (a–e) Total cellular lipids were then extracted and analyzed by mass spectrometry. (a) Total diacylglycerol (DAG). (b) Diacylglycerol species identified in all samples by MS. Species are identified by total acyl chain length and total number of double bonds; bold characters identify palmitate and/or oleate-based species. (c) Total triacylglycerol (TAG). (d) Detail of the 10 most prominent triacylglycerol species. Species are identified by total acyl chain length and total number of double bonds; bold characters identify palmitate and/or oleate-based species. (e): Exogenous diacylglycerol/triacylglycerol ratios (DAG/TAG). Panels (f–g) After the 7 days culture period as above, palmitate uptake and oxidation were measured. Results are shown as mean ± SEM; n = 3–5. #: significant effect of FA. §: significant effect of TPA; †: significant effect of the GPAT or the DGAT inhibitor. Analysis of cellular diacylglycerol and triacylglycerol levels and of FA metabolism in cardiomyocytes under GPAT or DGAT inhibition. Primary cardiomyocytes were cultured for 7 days with or without 0.4 mM fatty acids (FA), 100 nM TPA and the GPAT inhibitor FSG67, the DGAT inhibitor T863 or dimethyl sulfoxide. Panels (a–e) Total cellular lipids were then extracted and analyzed by mass spectrometry. (a) Total diacylglycerol (DAG). (b) Diacylglycerol species identified in all samples by MS. Species are identified by total acyl chain length and total number of double bonds; bold characters identify palmitate and/or oleate-based species. (c) Total triacylglycerol (TAG). (d) Detail of the 10 most prominent triacylglycerol species. Species are identified by total acyl chain length and total number of double bonds; bold characters identify palmitate and/or oleate-based species. (e): Exogenous diacylglycerol/triacylglycerol ratios (DAG/TAG). Panels (f–g) After the 7 days culture period as above, palmitate uptake and oxidation were measured. Results are shown as mean ± SEM; n = 3–5. #: significant effect of FA. §: significant effect of TPA; †: significant effect of the GPAT or the DGAT inhibitor. MS and HPTLC analyses of cellular triacylglycerol yielded similar results (Fig. 7c and S4). Triacylglycerol increased upon chronic FA exposure but remained unexpectedly high following GPAT inhibition. TPA co-treatment with FA pushed the triacylglycerol levels significantly higher, whereas DGAT inhibition in this condition markedly reduced triacylglycerol levels. Moreover, we discovered that among the most abundant triacylglycerol (TAG) isoforms, TAG 50:1 (2 palmitoyl + 1 oleyl chains), TAG 52:2 (1 palmitoyl + 2 oleyl chains) and TAG 54:3 (3 oleyl chains) are the ones whose concentration was significantly influenced by lipid droplet formation (Fig. 7d). Because neither diacylglycerol nor triacylglycerol contents readily explained the observed changes in glucose uptake, we surmised that the differential management of exogenous FA, expressed as the ratio of diacylglycerol to triacylglycerol, might rather be a determinant of glucose uptake. The ratio of exogenous diacylglycerol/triacylglycerol was calculated for cardiac myocytes exposed to FA; we observed that the conditions that reduced the diacylglycerol/triacylglycerol ratio upon FA exposure (GPAT inhibition or TPA chronic treatment) improved the metabolic stress-stimulated glucose uptake. Conversely, the increased diacylglycerol/triacylglycerol ratio resulting from DGAT inhibition is associated with impaired glucose uptake during metabolic stress in cardiomyocytes exposed to FA + TPA (Fig. 7e). Therefore, metabolic stress-stimulated glucose uptake was reduced when the diacylglycerol/triacylglycerol ratio was increased and conversely. To gain further insight into possibly relevant changes in FA metabolism, we measured palmitate uptake and oxidation in conditions showing significant differences in metabolic stress-stimulated glucose uptake in response to inhibition of lipid droplet biogenesis, i.e. GPAT inhibition in cardiomyocytes exposed to FA, and DGAT inhibition (T863) in cardiomyocytes exposed to FA + TPA. As previously observed, chronic FA exposure tended to increase both palmitate uptake (Fig. 7f) and oxidation (Fig. 7g), although the data failed to reach statistical significance. GPAT inhibition however did not modify either rate of FA metabolism. Chronic TPA treatment further increased palmitate uptake in cardiomyocytes exposed to FA, consistent with the marked accumulation of lipid droplets. However, DGAT inhibition, which markedly reduced lipid droplets, did reduce neither palmitate uptake nor its oxidation. In previous studies we observed that the induction of lipid droplets was associated with the restoration of stimulated glucose uptake in cardiomyocytes exposed to FA. We hypothesized that it was the accumulation of diacylglycerol in FA-exposed cardiomyocytes that triggered PKCδ activation and thereby inhibited glucose uptake. In this context, accumulation of triacylglycerol-rich lipid droplet would lead to diacylglycerol consumption for triacylglycerol synthesis. This should prevent PKC activation and thus indirectly preserve glucose uptake. In the present study, we investigated whether lipid droplets directly could contribute to protect glucose uptake in cardiomyocytes exposed to fatty acids. For this purpose, we prevented their constitution through inhibition of enzymes of the triacylglycerol biosynthesis pathway, GPAT and DGAT. By acting at two distinct levels of the triacylglycerol biosynthetic pathway, we expected to differentially modify diacylglycerol accumulation: GPAT inhibition should prevent, but DGAT inhibition worsens, diacylglycerol accumulation. The main findings of this study can be summarized in 6 parts:GPAT and DGAT inhibitions showed different effects on glucose uptake between cardiomyocytes exposed to FA only and cardiomyocytes displaying LD in presence of FAGPAT inhibition in cardiomyocytes exposed to FA alone reduces cellular diacylglycerol, prevents PKCδ activation and improves metabolic stress-stimulated glucose uptake, thus partially supporting the hypothesis formulated above and in our previous paper.Lipid droplet biogenesis in cardiomyocytes does not reduce diacylglycerol levels, suggesting that additional mechanisms are at play. In cardiomyocytes with lipid droplets, GPAT inhibition has little effect on glucose uptake while DGAT inhibition impairs metabolic stress-stimulated glucose uptake, again suggesting the existence of additional mechanisms.Metabolic stress-stimulated glucose uptake seems to be negatively impacted by an increased diacylglycerol/triacylglycerol ratio, and vice versa.Manipulations of lipid droplet biogenesis have a specific impact on metabolic stress-stimulated glucose uptake, without affecting the response to insulin, showing that the responses to insulin and metabolic-stress are differentially affected by alterations of intracellular fatty acids management.Changes in AMPK activity, if any, do not seem to explain the variations in metabolic stress-stimulated glucose uptake, as readouts of AMPK activity vary in opposite directions to metabolic stress-stimulated glucose uptake when lipid droplets biosynthesis is induced by TPA or AICAR or reduced by DGAT inhibition. GPAT and DGAT inhibitions showed different effects on glucose uptake between cardiomyocytes exposed to FA only and cardiomyocytes displaying LD in presence of FA GPAT inhibition in cardiomyocytes exposed to FA alone reduces cellular diacylglycerol, prevents PKCδ activation and improves metabolic stress-stimulated glucose uptake, thus partially supporting the hypothesis formulated above and in our previous paper. Lipid droplet biogenesis in cardiomyocytes does not reduce diacylglycerol levels, suggesting that additional mechanisms are at play. In cardiomyocytes with lipid droplets, GPAT inhibition has little effect on glucose uptake while DGAT inhibition impairs metabolic stress-stimulated glucose uptake, again suggesting the existence of additional mechanisms. Metabolic stress-stimulated glucose uptake seems to be negatively impacted by an increased diacylglycerol/triacylglycerol ratio, and vice versa. Manipulations of lipid droplet biogenesis have a specific impact on metabolic stress-stimulated glucose uptake, without affecting the response to insulin, showing that the responses to insulin and metabolic-stress are differentially affected by alterations of intracellular fatty acids management. Changes in AMPK activity, if any, do not seem to explain the variations in metabolic stress-stimulated glucose uptake, as readouts of AMPK activity vary in opposite directions to metabolic stress-stimulated glucose uptake when lipid droplets biosynthesis is induced by TPA or AICAR or reduced by DGAT inhibition. Chronic FA exposure presumably induces PKCδ activation by a mechanism involving the generation of diacylglycerol, known to be the physiological activator of classic and novel PKC. In several models of cultured cardiomyocytes, provision of FA results in increased cellular contents of diacylglycerol and activation of PKCδ. We observed that chronic GPAT inhibition, by reducing intracellular diacylglycerol concentration, prevented PKCδ activation and improved metabolic stress-stimulated glucose uptake in cardiomyocytes exposed to FA alone. In our previous study we showed that chronic treatment with the PKCδ specific inhibitor rottlerin was able to improve insulin and metabolic stress-stimulated glucose uptake in cardiomyocytes exposed to FA. Inhibition of PKCδ translocation has been shown to reduce ischemia and reperfusion-induced myocardial dysfunction and resulted in an improved regeneration of intracellular ATP, while the specific PKCδ inhibitor KAI-9803 ameliorated injury associated with ischemia and reperfusion in animal models of acute myocardial infarction. However, PKCδ activation only partially explains the effect of FA, as we did not observe a complete restoration of glucose uptake despite normalization of PKCδ translocation. This suggests that PKCδ activation represents only a part of the complex mechanisms by which FA impair glucose uptake in cardiomyocytes. The primary aim of this study was to evaluate if lipid droplet biogenesis protected glucose uptake against inhibition by FA by reducing the intracellular amount of bioactive lipid intermediates such as diacylglycerol, thereby preventing PKCδ activation. In our previous studies, we showed that chronic TPA treatment induced an abrogation of several PKC isoforms and a significant inhibition of PKCδ translocation to the membrane fraction. However, chronic AICAR treatment, which also induced lipid droplet formation, failed to inhibit PKCδ translocation. Lipidomics analysis indicated that lipid droplet formation was not accompanied by a significant reduction of intracellular diacylglycerol. This ruled out lipid droplet biogenesis itself as the driver of PKCδ activity prevention in cardiomyocytes exposed to FA + TPA. Instead, TPA treatment on these cells had the effect of abrogating PKCδ expression. Therefore, this indicates that lipid droplet formation occurring as a result of chronic AICAR and TPA treatments on cardiomyocytes has a protective role on glucose uptake not linked to PKCδ inhibition. Lipidomics analysis showed that exogenous FA increased intracellular diacylglycerol content; specifically, the main diacylglycerol species found in FA-exposed cardiomyocytes were oleate- and palmitate-based. As expected, chronic GPAT inhibition significantly reduced oleate- and palmitate-derived diacylglycerols while other diacylglycerol species were reduced without reaching statistical significance, confirming that exogenous FA might be responsible for PKCδ activation. Surprisingly, lipid droplet biogenesis did not reduce diacylglycerol content in cardiomyocytes exposed to FA + TPA; possibly the increased FA uptake induced by TPA (Fig. 7f) is not entirely compensated by increased incorporation of fatty acyl moieties into triacylglycerol. Also, no increase in the amount of cellular diacylglycerol was observed when DGAT activity was chronically inhibited. Taken together, these results suggest that the protective effect of lipid droplets on glucose uptake was not directly related to diacylglycerol accumulation and consumption. Similar observations had been made in vivo in mice, wherein myocardial diacylglycerol was increased following high fat diet but not further increased by DGAT1 deficiency. Conversely, myocardial overexpression of DGAT1 was shown to reduce diacylglycerol concentration and protected the heart from lipotoxicity. DGAT1 overexpression in skeletal muscles protected mice from high fat diet-induced insulin resistance, while DGAT1 deficiency diminished insulin sensitivity. However, in our model, chronic GPAT or DGAT inhibition in cardiomyocytes displaying lipid droplets did not influence insulin-stimulated glucose uptake and only DGAT inhibition impaired metabolic stress-stimulated glucose uptake. This discrepancy might pertain to different metabolic properties of skeletal muscle vs. myocardium. To the best of our knowledge, no study investigated the involvement of DGAT activity in insulin sensitivity of the myocardium. As expected, cardiomyocytes displaying lipid droplets also showed increased intracellular triacylglycerol concentration. However, both mass spectrometry- and chromatography-based lipidomics showed significant levels of triacylglycerol were also present in FA-exposed cardiomyocytes even though very few lipid droplets were observed. In addition, with chronic GPAT inhibition in FA-exposed cardiomyocytes, triacylglycerol level remained elevated despite a decrease of diacylglycerol concentration. These observations suggest that significant triacylglycerol amounts may exist in cardiomyocytes outside of lipid droplets as identified by our staining method. How could substantial amount of triacylglycerol accumulate in cardiomyocytes exposed to FA despite GPAT inhibition ? In the heart, triacylglycerol synthesis is mainly regulated by GPAT1 and GPAT3, localized in the mitochondria and in the endoplasmic reticulum, respectively; both GPAT isoforms are inhibited by FSG67. Since the other triacylglycerol biosynthesis intermediate phosphatidic acid, which is also downstream of GPAT, did not show much variation, one hypothesis could be that an alternative pathway bypassing GPAT, such as acylation of dihydroxyacetone phosphate followed by the reduction to lysophosphatidic acid, can induce triacylglycerol formation even when GPAT activity is inhibited. Indeed, Kojta et al. showed how GPAT silencing in skeletal muscle markedly reduced diacylglycerol content but with a much lesser effect on triacylglycerol. Significant changes in cellular diacylglycerol and triacylglycerol concentrations were only observed for the palmitate- and oleate-based species, which were dominant in FA-containing conditions, but minorized in the BSA control condition. This suggests that exogenous FA and their derivatives are the principal effectors of lipid-induced metabolic modulation in cardiomyocytes, in line with the very limited capacity of cardiomyocytes for the biosynthesis of fatty acids. Similar to the situation with the amounts of diacylglycerol, we could not discern a direct correlation, positive or negative, between cellular triacylglycerol contents and glucose uptake. However, there seems to be a good inverse correlation between metabolic stress-stimulated glucose uptake and the exogenous diacylglycerol/triacylglycerol ratio, suggesting that this ratio might be more important than the actual concentration of either of its determinants in regulating glucose uptake. In a previous study we observed that either stimulation or complete abolition of FA oxidation could improve metabolic stress-stimulated glucose uptake. In the present study however, neither DGAT nor GPAT inhibition had any effect on FA oxidation, arguing against modulation of FA oxidation playing a role in the effects of these maneuvers on glucose uptake. Extracellular lipids may interact with glucose uptake regulation by inducing disassembly and inhibition of the vacuolar-type H-ATPase (v-ATPase). This results in further stimulation of FA uptake and inhibition of insulin-stimulated GLUT4 translocation both in cultured cardiomyocytes and in ex vivo cardiomyocytes obtained from rats fed a high-fat diet. Although LD biogenesis could prevent v-ATPase by sequestering intracellular FA, we do not think that this mechanism is involved in the protective effect of LD on metabolic stress-stimulated glucose uptake for two reasons. First, v-ATPase reactivation restored insulin-stimulated glucose uptake in palmitate-exposed cardiomyocytes, whereas manipulation of LD biosynthesis in our study only impacted metabolic stress-stimulated, not insulin-stimulated glucose uptake. Second, induction of LD biosynthesis resulted in a stimulation of FA uptake rather than an inhibition (Fig. 7f), as should have been the case with v-ATPase reactivation. Activation of the AMPK complex mediates the stimulation of myocardial glucose uptake by metabolic stress in vivo, ex vivo and in vitro. Thus, we investigated AMPK signaling, assessed at two levels: phosphorylation of AMPKα on T and phosphorylation of raptor (an AMPK substrate not involved in the regulation of glucose uptake) on S. We have found an interesting relationship between lipid droplet biogenesis and AMPKα phosphorylation, since we observed that both chronic treatments inducing lipid droplet formation (TPA and AICAR) reduced T AMPKα phosphorylation, while chronic GPAT or DGAT inhibition restored it. Nevertheless, these effects were not reflected on S raptor phosphorylation, used as a readout for AMPK activity. Raptor phosphorylation was previously shown to better correlate with metabolic-stress stimulated glucose uptake than AMPKα phosphorylation. This may suggest that the regulatory mechanisms of glucose uptake impacted by variations in lipid droplet metabolism are either downstream or independent of AMPK activation. However, as far as we know, raptor is not directly involved in the regulation of glucose uptake in cardiomyocytes and we cannot exclude that other pathways may be differentially activated by AMPK as compared to raptor phosphorylation. In contrast to raptor, the phosphorylation of AS160 is known to be required for translocation of GLUT4 to occur in response to metabolic stress, and AS160 is a direct target of AMPK. Indeed, the variations of AS160 phosphorylation with variations in lipid droplet metabolism generally matched the changes in AMPKα phosphorylation. However, these variations go in opposite directions as compared to glucose uptake, and therefore cannot explain the effects of lipid metabolism on metabolic stress-stimulated glucose uptake. Being essential for the maintenance of cellular energy homeostasis, AMPK is also involved in the regulation of lipid metabolism in cardiomyocytes. AMPK can modulate lipid droplet recycling and dispersion via perilipin phosphorylation. AMPK, by regulating hormone-sensitive lipase and perilipin activity within lipid droplets, induces triacylglycerol hydrolysis in adipocytes and L6 myoblasts. However, we are not aware of any evidence regarding a reverse direct effect of lipid droplet metabolism on AMPK regulation. The only observation in the literature similar to ours was made in hepatocytes of rats on a high-fat diet, wherein increases in both GPAT expression and lipid droplet density were also associated with reduced AMPK phosphorylation, suggesting that GPAT biosynthetic activity could directly influence AMPK activation. Some limitations of this study deserve comments. Due to technical reasons, the primary limitation of this study was the low number of conditions and replicates used for mass spectrometry-based lipid analysis. We focused on the most significant conditions pertaining to changes in metabolic stress-stimulated glucose uptake. Nevertheless, a clear signal was obtained in these albeit limited conditions. Another limitation was our inability to identify mechanisms downstream of AMPK involved in glucose metabolism that are affected by alterations of lipid droplet metabolism. The last limitation was the use of the ATPase inhibitor oligomycin as a surrogate for ischemia in vitro. Ischemia in vivo impacts the mitochondrial function in a much more complex manner than simply shutting off ATPase activity. However simulated ischemia in vitro takes a much longer time to act than in vivo, long enough that changes in protein expression, such as overexpression of GLUT1, can occur and confound the interpretation of glucose uptake data. Therefore, oligomycin remains a convenient way to achieve ischemia-like energy depletion and AMPK activation without changing protein expression. The experimental model used herein also deserves a comment. Adult rat cardiomyocytes in long-term primary culture under high FCS conditions undergo dedifferentiation and redifferentiation, resulting after 7–8 days in a phenotype with high myofibrillar organization, contractility and a glucose transport response to insulin and metabolic stress almost identical to that observed in freshly isolated cardiomyocytes despite extensive cytoskeletal reorganization leading to the loss of the rod-like shape. In addition, inclusion of 9-cis retinoic acid to the culture medium markedly reduces the extent of transient dedifferentiation and of subsequent hypertrophy. We therefore believe this experimental model to be valid for investigations on the impact of chronic metabolic interventions on glucose metabolism in cardiomyocytes. In conclusion (Fig. 8), we showed that chronic FA exposure on cardiomyocytes increased intracellular diacylglycerol concentration, thereby inducing PKCδ activation and impairing metabolic stress-stimulated glucose uptake. Instead, boosting lipid droplet biogenesis protects glucose uptake in FA-exposed cardiomyocytes in a mechanism independent from diacylglycerol levels. Critically, we found that the intracellular diacylglycerol/triacylglycerol ratio of oleate- and palmitate-derived FA is inversely related to glucose uptake.Figure 8Conclusions. Chronic fatty acids (FA) exposure of cardiomyocytes increases intracellular diacylglycerol (DAG) concentration, thereby inducing PKCδ activation and impairing metabolic stress-stimulated glucose uptake. Lipid droplet formation reduces the diacylglycerol/triacylglycerol (DAG/TAG) ratio, thus preserving glucose uptake despite diacylglycerol accumulation, suggesting a protective effect by lipid droplets independent of PKCδ activity. Red arrows: mechanisms impairing metabolic stress-stimulated glucose uptake; green arrows: mechanisms preserving metabolic stress-stimulated glucose uptake. Conclusions. Chronic fatty acids (FA) exposure of cardiomyocytes increases intracellular diacylglycerol (DAG) concentration, thereby inducing PKCδ activation and impairing metabolic stress-stimulated glucose uptake. Lipid droplet formation reduces the diacylglycerol/triacylglycerol (DAG/TAG) ratio, thus preserving glucose uptake despite diacylglycerol accumulation, suggesting a protective effect by lipid droplets independent of PKCδ activity. Red arrows: mechanisms impairing metabolic stress-stimulated glucose uptake; green arrows: mechanisms preserving metabolic stress-stimulated glucose uptake. Male Sprague Dawley rats (100–200 g) obtained from Janvier Labs (France) were deeply anesthetized by an intraperitoneal injection of ketamine (100 mg/kg) and xylazine (10 mg/kg) and the hearts were harvested under deep coma. All the methods were carried out in accordance with relevant institutional guidelines and regulations. The protocol was approved by the Canton of Geneva Committee for Animal Experimentation (authorizations # GE/210/17, GE/60/19 and GE153). All the experiments were carried out in accordance with ARRIVE guidelines. Cardiomyocytes were isolated as previously described by retrograde perfusion of the hearts with collagenase (type II; Worthington; 120 IU/ml) and hyaluronidase (1% w/v). Cardiomyocytes were separated from non-myocyte cardiac cells by pre-plating the whole cell suspension for 90 min on untreated plastic, to which non-myocyte cells, but not cardiomyocytes, readily adhere. Cardiomyocytes were plated in M199 medium containing 5.5 mM glucose supplemented with 5 mM creatine, 2 mM L-carnitine, 5 mM taurine, 100 μM cytosine-β-D- arabinofuranoside, 100 nM 9-cis retinoic acid, 10 nM triiodothyronine and 20% fetal calf serum. Dishes were previously coated with 0.1% gelatin for 4 h and incubated overnight with complete culture medium. For confocal microscopy, cardiomyocytes were plated on laminin-coated glass coverslips. Free fatty acids (FA) consisted of a 1:1 mix of palmitate (C16:0) and oleate (C18:1 n-9) bound to bovine serum albumin (BSA), at a final total concentration of 0.4 mM FA; TPA or AICAR were added at the time of plating. Small molecules GPAT and DGAT inhibitors were also added at the time of plating. The GPAT inhibitor FSG67 was used at a final concentration of 30 µM, the DGAT inhibitor A922500 at a final concentration of 1 µM and the DGAT inhibitor T863 at a final concentration of 100 nM. The culture medium was renewed every 2–3 days, and subsequent analyses were performed on day 7. At this time point, control cardiomyocytes display a well-differentiated phenotype with stable insulin responsiveness. For confocal fluorescent microscopy, cardiomyocytes cultured on laminin-coated glass coverslips were washed with ice-cold PBS and fixed with 4 mM paraformaldehyde in PBS for 20 min at room temperature. Fixation was quenched with 200 mM glycine in PBS and cardiomyocytes permeabilized with 0.3% Triton X-100 in PBS for 3 min. Non-specific dye or antibody binding was reduced by preincubation with 3% BSA and 0.1% Tween-20 in PBS. Neutral lipids were stained with 2.5 μg/ml Bodipy 493/503 for 30 min; F-actin counterstaining was obtained with AlexaFluor 543-labeled phalloidin (2 U/ml). Following washes with PBS and H2O, coverslips were mounted on glass slides with ProLong Diamond antifade containing DAPI for DNA staining. Cardiomyocytes were examined with a Zeiss LSM800 confocal microscope, using a × 63 oil immersion objective. One-micrometer-thick confocal slices were acquired throughout the thickness of the cardiomyocytes and Z-stack projections obtained with the ImageJ software in maximum intensity projection mode. Image luminosity and contrast were digitally enhanced, taking care to apply the same linear adjustments to images from different experimental groups. Glucose transport was estimated by measuring 2-deoxyglucose (2-DG) uptake, as previously described. Briefly, cardiomyocytes were incubated in M199 containing 10 nM [2,6-H]-2-DG (ANAWA Clinisciences Group; 1–2 μCi/ml) and 5.5 mM cold glucose at 37 °C for 1 h, in the presence or absence of glucose transport agonists. Glucose transport agonists used were insulin (10 M) or oligomycin (10 M), a mitochondrial FO-ATPase inhibitor, to induce metabolic stress. 2-DG uptake was halted by three washes with ice-cold PBS before lysis in 1 ml 0.1 M NaOH. Two 20-μl aliquots were taken for protein content determination and the remaining NaOH lysate assayed for radioactivity in a TriCARB 1900 TR liquid scintillation analyzer (Packard). Following stimulation with insulin or oligomycin, incubations were terminated by three washes in ice-cold PBS before solubilizing cells in 200 μl lysis buffer containing 150 mM NaCl, 50 mM Tris–HCl (pH 7.5), 1 mM EDTA, 0.5% sodium deoxycholate, 1% Igepal CA 630, Halt protease, and phosphatase inhibitor Cocktail (Pierce, Thermo Scientific). Proteins (30 μg) from each sample were separated on SDS-PAGE gels and transferred onto polyvinylidene difluoride membranes. Primary and secondary antibodies used for western blot analysis are listed in Supplemental Table. For all blots, incubation with the primary antibody was overnight at 4 °C, incubation with the secondary antibody for 1 h at room temperature. Densitometric analysis of chemiluminescent signals captured with a LAS-4000 Luminescent Image Analyzer (Fujifilm) was performed using the ImageJ software (National Institutes of Health, http://rsb.info.nih.gov/ij). For each separate membrane, the data were normalized such that the highest intensity, regardless of treatment, was equal to one. Cytosol and membrane fractions of cultured cardiomyocytes were obtained as previously described. Cardiomyocytes were washed in ice-cold PBS and scraped in buffer A containing 20 mM Tris–HCl pH 7.5, 2.5 mM EGTA, 1 mM EDTA, 100 mM NaF, 2 mM dithiothreitol, and Halt protease and phosphatase inhibitor cocktail. The suspension was sonicated with four 5-s bursts on ice and then centrifuged at 1500 × g for 10 min. The supernatant was then ultracentrifuged at 100,000 × g for 45 min at 4 °C. The pellet containing the membrane fraction was solubilized in buffer A containing 1% Triton X-100, sonicated, and centrifuged at 15,000 × g for 15 min at 4 °C to retain the supernatant. Cardiomyocytes were washed three times in ice-cold PBS without Ca and Mg and detached with trypsin/EDTA for 5 min. Trypsin activity was stopped by M199 medium + 20% fetal calf serum and the suspension was centrifugated for 1 min at 500 rpm. The pellet was washed three times in ice-cold PBS without Ca and Mg, resuspended in ice-cold PBS without Ca and Mg and sonicated. Shotgun lipidomics analysis of the samples was outsourced to Lipotype GmbH (www.lipotype.com; Dresden, Germany). Lipids were extracted using chloroform and methanol. Samples were spiked with lipid class-specific internal standards prior to extraction. After drying and re-suspending in mass spectrometry (MS) acquisition mixture, lipid extracts were subjected to mass spectrometric analysis. Mass spectra were acquired on a hybrid quadrupole/Orbitrap mass spectrometer (Q Exactive Plus; Thermo Scientific) equipped with an automated nano-flow electrospray ion source in both positive and negative ion mode. Lipid identification using LipotypeXplorer was performed on unprocessed (*.raw format) mass spectra. For MS-only mode, lipid identification was based on the molecular masses of the intact molecules. MS/MS mode included the collision-induced fragmentation of lipid molecules and lipid identification was based on both the intact masses and the masses of the fragments. Prior to normalization and further statistical analysis, lipid identifications were filtered according to mass accuracy, occupation threshold, noise and background. Lists of identified lipids and their intensities were stored in a database optimized for the particular structure inherent to lipidomic datasets. The intensity of lipid class-specific internal standards was used for lipid quantification. High-Performance Thin-Layer Chromatography of neutral lipids was performed as recently described. Briefly, lipids from cell pellets were extracted using the protocol of Bligh and Dyer and spotted on HPTLC Silica Gel60 plates (Merck, 1.05631.0001) using an automated sampler (CAMAG ATS 4). The HPTLC plate was pre-washed in chloroform:methanol (1:1) and dried in a vacuum chamber. For neutral lipids separation, lipid standards (cholesterol, cholesteryl ester, sphingomyelin, phosphatidylcholine, phosphatidylethanolamine and triacylglycerol) and cell lipid extracts resuspended in chloroform:methanol:water 20:9:1 were spotted on the plate using the automated system. The plate was developed first in chloroform:methanol:ammonium hydroxide (65:25:4) for 5 cm, dried briefly and further developed in hexane:diethyl ether:acetic acid (80:20:2) for 9 cm. After separation, the HPTLC plate was dried in a vacuum chamber for 30 min. Lipids were visualized using a method adapted from Churchward. Briefly, a CuSO4 staining solution was prepared: 5 g of CuSO4 dissolved in 40 ml water, filtered, mixed with 4.7 ml of 85% ortho-phosphoric acid and filled up to 50 ml with water. 10 ml of freshly prepared staining solution was poured on the TLC plate, incubated for 1 min, decanted and the plate was dried in the vacuum chamber for 15 min. Lipids were then charred at 145 °C for 7.5 min. The plate was visualized in fluorescence light at 488 nm (ChemiDoc MP, Bio-Rad). Lipid spots were quantified from the fluorescent-inverted pictures using densitometry with the ImageJ software. Palmitate oxidation was estimated based on the rate of transfer of H from [9,10-H]palmitate to H2O. Cardiomyocytes were incubated for 60 min in medium containing palmitate (0.05 mM), oleate (0.05 mM), and 1 μCi/ml [9,10-H]palmitate (ANAWA Clinisciences Group) complexed to bovine serum albumin (0.2 mM). At the end of the incubation period, the incubation medium (1 ml) was retrieved, immediately mixed with 1 ml of ice-cold 10% trichloracetic acid and centrifuged at 2,200 g for 10 min at 4 °C. The supernatant was neutralized with 250 µl of NaOH 6 M. H2O in the supernatant was separated from [9,10-H]palmitate by anion exchange chromatography in Dowex 1 × 4. H2O eluted from the Dowex column was counted by liquid scintillation. Cardiomyocytes were washed twice with ice-cold PBS, dissolved in 0.1 M NaOH. Twenty-µl aliquots were taken for protein content determination and the remaining NaOH lysate assayed for radioactivity in a TriCARB 1900 TR liquid scintillation analyzer (Packard). Palmitate uptake was estimated from the sum of H label transferred to H2O and H label remaining in the cardiomyocytes. Data are presented as mean ± SEM obtained from replicated experiments. Data were compared by one-way or two-way ANOVA (Prism 7, GraphPad Software) followed by post hoc testing for false discovery rates by the method of Benjamini and Yekutieli. Post hoc testing indicated a positive discovery when the false discovery rate q was < 0.05. Throughout the article, the following symbols are used for positive discoveries (q values): * indicates a significant effect of insulin or oligomycin stimulation as compared with unstimulated cardiomyocytes having received the same chronic treatment; # indicates a significant effect of chronic FA exposure, as compared with cardiomyocytes not exposed to FA undergoing the same acute stimulation; § indicates a significant effect of chronic TPA or AICAR exposure as compared with cardiomyocytes not exposed to TPA or AICAR undergoing the same acute stimulation; † indicates a significant effect of chronic DGAT or GPAT inhibition (but with the same exposure—or absence thereof—to FA, TPA or AICAR), as compared with cardiomyocytes not exposed to DGAT or GPAT inhibitor undergoing the same acute stimulation.
|
PMC8259984
|
LAMP3 deficiency affects surfactant homeostasis in mice
|
Lysosome-associated membrane glycoprotein 3 (LAMP3) is a type I transmembrane protein of the LAMP protein family with a cell-type-specific expression in alveolar type II cells in mice and hitherto unknown function. In type II pneumocytes, LAMP3 is localized in lamellar bodies, secretory organelles releasing pulmonary surfactant into the extracellular space to lower surface tension at the air/liquid interface. The physiological function of LAMP3, however, remains enigmatic. We generated Lamp3 knockout mice by CRISPR/Cas9. LAMP3 deficient mice are viable with an average life span and display regular lung function under basal conditions. The levels of a major hydrophobic protein component of pulmonary surfactant, SP-C, are strongly increased in the lung of Lamp3 knockout mice, and the lipid composition of the bronchoalveolar lavage shows mild but significant changes, resulting in alterations in surfactant functionality. In ovalbumin-induced experimental allergic asthma, the changes in lipid composition are aggravated, and LAMP3-deficient mice exert an increased airway resistance. Our data suggest a critical role of LAMP3 in the regulation of pulmonary surfactant homeostasis and normal lung function.The lysosome-associated membrane glycoprotein (LAMP) family consists of five members (LAMP1, LAMP2, LAMP3 / DC-LAMP, CD68 / macrosialin, and LAMP5 (BAD-LAMP)). All family members are heavily N- and O-glycosylated type I transmembrane proteins localized to the limiting membrane of lysosomes and lysosome-related organelles . LAMP1 and LAMP2 are ubiquitously expressed , while LAMP3 (synonymously called DC-LAMP), CD68, and LAMP5 show cell-type-specific or tissue-specific expression. CD68 expression is restricted to macrophages, monocytes, and microglia , whereas LAMP5 is expressed exclusively in the brain . The LAMP3-expression pattern between humans and mice differs: LAMP3 is expressed in humans in alveolar type II (AT2) cells and dendritic cells (DC) (eponymously for its alternative name "DC-LAMP" or CD208). In mice,LAMP3 is found exclusively in AT2 cells but not in DCs [4–7]. These findings suggest a conserved and critical function of LAMP3 in AT2 cells. On the subcellular levels, LAMP3 is localized in lamellar bodies (LBs) . However, the function of LAMP3 in these cells remains enigmatic until now; if and how LAMP3 affects surfactant homeostasis, e.g., mediating surfactant release via LBs, is also poorly understood. The primary function of AT2 cells is mainly attributed to pulmonary surfactant release to lower surface tension at the lung’s air/liquid interface. Pulmonary surfactant is a mixture of lipids and proteins : The major hydrophobic protein components, surfactant proteins B (SP-B) and C (SP-C), are small proteins deeply embedded into the phospholipids of the surfactant. Notably, deficiency in SP-B or SP-C due to genetic mutations in SFTPB or SFTPC cause hereditary forms of childhood interstitial lung disease . Sftpb gene-targeted mice die of respiratory failure after birth, associated with irregular and dysfunctional LBs and tubular myelin . SP-C deficient mice have a regular life span and only show subtle deficits in lung mechanics and surfactant stability . The composition of surfactant lipids has been analyzed extensively [13–17]. The most abundant and characteristic lipid class is phosphatidylcholine (PC). PC makes up approximately 70% of the total surfactant mass with dipalmitoylphosphatidylcholine (DPPC; PC 16:0/16:0) as the primary lipid species . An overall decrease of PC lipids is coupled to higher surface tension and lower surfactant protein inclusion . Other highly abundant and functionally essential phospholipid classes in surfactant include phosphatidylglycerol (PG) and phosphatidylinositol (PI), while phosphatidylethanolamine (PE), and sphingomyelin (SM) are reported as minor components of surfactant and most likely derived from other cell membranes. In addition, lysophosphatidylcholine (LPC) can be found in low abundance in pulmonary surfactant . Preassembled surfactant is stored in secretory organelles known as LB that fuse with the cell membranes and release pulmonary surfactant. LBs are multivesicular body / late endosome-derived specialized organelles filled with surfactant characterized by a typical "myelin"-like appearance by electron microscopy. Phospholipids are directly transferred from the endoplasmic reticulum to LB through a vesicle-independent mechanism by the phospholipid transporter "ATP binding cassette subfamily A member 3" (ABCA3) . Like SFTPB or SFTPC, loss-of-function mutations in ABCA3 were identified in full-term infants who died from unexplained fatal respiratory distress syndrome . ABCA3 mutations represent the most frequent form of congenital surfactant deficiencies . A presumably loss-of-function mutation in LAMP3 (p.(E387K)) was recently identified in an Airedale Terrier dog breed, leading to severe symptoms and pathology similar to clinical symptoms of the most severe neonatal forms of human surfactant deficiency . The affected puppies’ symptoms included lethal hypoxic respiratory distress and occurred within the first days or weeks of life . LBs were smaller, contained fewer lamellae, and occasionally revealed a disrupted common limiting membrane . To conclusively address the physiological role of LAMP3 in surfactant homeostasis using a genetically defined animal model, we generated Lamp3 knockout mice (KO, Lamp3) by CRISPR/Cas9. In contrast to dogs bearing a natural mutation in LAMP3, Lamp3 mice are vital, display a regular lung function at the basal level, and show no increased prenatal lethality. The lungs of Lamp3 animals did neither macroscopically nor microscopically differ from those of wild type littermates. However, biochemical analysis of broncho-alveolar lavage (BAL) fluid clearly revealed increased pro-SP-C levels and altered lipid composition in the Lamp3 mice. These differences are further pronounced under diseased conditions as evoked by allergen-induced experimental asthma, and most notably, diseased Lamp3 mice revealed significantly increased airway resistance, when stressed during methacholine provocation testing. In conclusion, our data suggest a critical but not essential role of LAMP3 in pulmonary surfactant homeostasis associated with normal lung function. All animal studies were approved by the local animal ethics committe (Ministerium für Energiewende, Landwirtschaft, Umwelt, Natur und Digitalisierung; AZ V242-4255/2018, and V244-2826/2017 (32-3/17)). Analytical grade chemicals were purchased, if not stated otherwise, from Sigma-Aldrich (MO., USA). Antibodies: The following antibodies were used in the study: Rat monoclonal LAMP3, clone 1006F7.05 (Dendritics), SP-C, rabbit polyclonal (Abcam), rabbit polyclonal antibodies against mature SP-B and Pro-SP-B (Seven Hills Bioreagents) were a generous gift from Jeffrey A. Whitsett, polyclonal β-Actin (Santa Cruz Biotechnology, Inc), polyclonal rabbit antiserum against ABCA3 was a kind gift from Michael L. Fitzgerald . Fluorophore-conjugated secondary antibodies against the corresponding primary antibody species (AlexaFluor 488) were purchased from Invitrogen / Molecular Probes and were diluted 1:500. Mice with a null mutation in the Lamp3 gene were generated in a C57BL/6N background using a CRISPR genome-editing system. For this purpose, Cas9 was combined with two single guide RNAs (sgRNAs) targeting the second exon in the Lamp3 gene with the following spacer sequences: sgRNA_Lamp3 F1: TCATCTACTGACGATACCAT and sgRNA_Lamp3 R1: GCTAGACTAGCTCTGGTTGT microinjected into fertilized zygotes. Genome editing was confirmed by PCR amplification in the resulting founder mice with the primers Lamp3F 5´-GATGGGGGAGGGATCTTTTA-3’ and Lamp3R 5´-GTTGGCCTCTGATTGGTTGT-3’. A founder harboring a 40 bp deletion within the second exon leading to a frameshift mutation was selected for further breeding. Mice were housed under specific pathogen-free conditions (12 hours’ light/dark cycle, constant room temperature, and humidity). Food and water were available ad libitum. Female, 6- to 8-week-old Lamp3 and wildtype littermates mice (CAU, Kiel, Germany) were used for the allergic asthma model. They received ovalbumin (OVA)-free diet and water ad libitum. Mice were sensitized to OVA by three intraperitoneal (i.p.) injections of 10 μg of OVA (OVA grade VI, Sigma-Aldrich, St. Louis, MO, USA) adsorbed to 150 μg of aluminum hydroxide (imject alum, Thermo Fisher Scientific, Waltham, MA, USA) on days 1, 14 and 21. To induce acute allergic airway inflammation, mice were exposed three times to an OVA (OVA grade V, Sigma-Aldrich) aerosol (1% wt/vol in PBS) on days 26, 27, and 28. Control animals were sham sensitized to PBS and subsequently challenged with OVA aerosol. Steady-state lung parameters midexpiratory flow (EF50), frequency (f), functional residual capacity (Frc), minute volume (MV), peak expiratory flow (PEF), peak inspiratory flow (PIF), expiration time (Te), inspiration time (Ti), and tidal volume (TV) were measured in conscious, spontaneously breathing mice using FinePointe non-invasive airway mechanics double-chamber plethysmography (NAM, Data Science International, St. Paul, MN, USA). Steady-state airway resistance (RI) and dynamic compliance (Cdyn) were measured in anesthetized and ventilated mice using FinePointe RC Units (Data Science International). Airway responsiveness to methacholine (MCh, acetyl-β-methylcholine chloride; Sigma-Aldrich) challenge was invasively assessed on day 29 using FinePointe RC Units (Data Science International, St. Paul, MN, USA) by continuous measurement of RI. Animals were anesthetized with ketamine (90 mg/kg body weight; cp-pharma) and xylazine (10 mg/kg BW; cp-pharma) and tracheotomized with a cannula. Mechanical ventilation was previously described . Measurements were taken at baseline (PBS) and in response to inhalation of increased concentrations of aerosolized methacholine (3.125; 6.25; 12.5; 25; 50; and 100 mg/mL). After assessment of lung function, all animals were sacrificed by cervical dislocation under deep anesthesia. For bronchoalveolar lavage, lungs were rinsed with 1 ml of fresh, ice-cold PBS containing protease inhibitor (Complete, Roche, Basel, Switzerland) via a tracheal cannula. Cells in BAL fluid were counted in the Neubauer chamber. Fifty microliter aliquots of lavage fluids were subjected to centrifugation (5 minutes at 320 g) (Zytozentrifuge Shandon Cytospin 4, Thermo Fisher Scientific), and cells were microscopically differentiated according to morphologic criteria. Animals were anesthetized and lungs were fixed by perfusion through the right ventricle with HEPES buffer (pH 7.35) containing 1.5% glutaraldehyde and 1.5% formaldehyde, at a fixed positive inflation pressure of 25 cm H2O. After storage of the lungs in the same solution overnight, 3 mm sized pieces of lung tissue were postfixed first in 1% osmium tetroxide for 2 hours and then in 4% uranyl acetate overnight (both aqueous solutions). Samples were dehydrated in acetone and embedded in Epon. Ultrathin sections of 60 nm were poststained with aqueous uranyl acetate and lead citrate and imaged in a Morgagni TEM (FEI, Eindhoven, NL). Lungs were dissected from mice, and the left lung lobe was snap-frozen in liquid nitrogen. The other lobes were fixed ex-situ with 4% (wt/vol) formaldehyde via the trachea under constant pressure, removed and stored in 4% formaldehyde. Volume of these lobes was determined based on Archimedes’ principle. In brief, the remaining lung lobes were placed in a water-filled container set on electronic scales. The weight of water displaced by the submerged lung equals the volume of the lung. Then lung tissues were embedded in paraffin. For analysis of lung inflammation, 2 μm sections were stained with periodic acid-Schiff (PAS). Photomicrographs were recorded by a digital camera (DP-25, Olympus, Tokyo, Japan) attached to a microscope (BX-51, Olympus) with a 20-fold magnification objective using Olympus cell^A software. Mucus quantification was performed as previously described . Apoptotic cells were stained using ApopTag Peroxidase In Situ Apoptosis Detection Kit (Merck Millipore, MA, USA) according to manufacturer’s instructions. 4% Formaldehyde-fixed lungs were incubated in cryoprotectant solution (30% w/v sucrose in 0.1 M phosphate buffer, pH 7.4) over night. 35 μm thin sections were cut with a Leica 9000s sliding microtome (Leica, Wetzlar, Germany) and collected in cold 0.1 M phosphate buffer (PB). Free floating sections were stained by blocking in blocking solution (0.5% Triton-X 100, 4% normal goat serum in 0.1 M PB pH 7.4) for 1 hour at room temperature. Subsequently, sections were incubated in blocking solution containing the primary antibody at 4°C over night. After washing three times with wash solution (0.1 M PB pH 7.4 containing 0.25% Triton-X 100), sections were incubated for 2 hours in secondary antibody in solution, washed again three times in wash solution containing 4′,6-Diamidin-2-phenylindol (DAPI) and finally brought on glass slides and embedded in Mowiol/DABCO. Cells were imaged using a Zeiss confocal microscope equipoped with a 63x objective (Zeiss LSM980 AiryScan 2). Deep frozen lungs were homogenized with mortar and pestle. An aliquot of 30 mg of lung powder was transferred into RLT buffer (Qiagen), and RNA was isolated with RNeasy Mini Kit (Qiagen) according to the manufacturer’s guidelines. 1 μg of RNA was used for cDNA synthesis (First Strand cDNA Synthesis Kit, Thermo Fisher Scientific, Waltham, MA, USA). Another 30 mg aliquot of lung powder was transferred into radioimmunoprecipitation assay buffer (RIPA) buffer, and proteins were isolated. Protein concentration was determined by bicinchoninic acid (BCA) assay (Thermo Fisher Scientific, Waltham, MA, USA). Western analysis was performed on lung tissue homogenates from wildtype and LAMP3-deficient mice. Protein content was measured by the BCA assay (Thermo Fisher Scientific, Waltham, MA, USA). Defined amounts of protein were separated on SDS-PAGE and transferred to nitrocellulose membrane. Membranes were incubated with anti-pro-SP-C (rabbit monoclonal, 1:1000, Abcam, Cambridge, UK), anti-pro-SP-B (rabbit polyclonal, 1:2000, Seven Hills Bioreagents, CA, USA), anti-mature SP-B (rabbit polyclonal, 1:1000, Seven Hills Bioreagents), and anti-β-actin (mouse monoclonal, 1:200, Santa Cruz Biotechnology, Heidelberg, Germany). Donkey anti-rabbit IgG HRP (1:2000, Santa Cruz) and donkey anti-mouse IgG HRP (1:2000, Cell Signaling) served as secondary antibodies Immunoreactive proteins were visualized using the ECL Western blotting detection system (Bio-Rad Laboratories GmbH, Feldkirchen, Germany), band intensity was quantified with Image J 1.52 (NIH), and data were normalized to β-actin levels. Quantitative RT-PCR was performed on the Roche LightCycler 480 Instrument II system using the LightCycler 480 SYBR Green I Master (Roche Applied Science, Mannheim, Germany). Template cDNA was diluted 1:10. RT-PCR was performed in triplicate in a total volume of 10 μl according to the manufacturer’s instruction with a final primer concentration of 0.5 μM. Cycling conditions were as follows: 45 cycles at 95°C for 10 s, touch-down annealing temperature (63 to 58°C, temperature reduction 0.5°C per cycle) for 10 s and 72°C for 8 s. For each primer pair, a standard curve was established by serial cDNA dilutions. The following primer sequences for reference gene RPL32 were used: forward 5‘-AAAATTAAGCGAAACTGGCG-3’, reverse 5‘-ATTGTGGACCAGGAACTTGC-3’ (NM_172086.2,156 bp). QuantiTect Primer Assays from Qiagen (Venlo, The Netherlands) were used for Lamp3 (QT00157031, NM_177356, 105 bp), Sftpb (QT00124908, NM_011359, 125 bp) and Sftpc (QT00109424, NM_011359, 125 bp). Negative controls (reverse transcriptase/RNA template) were included to detect possible contaminations. Amplification specificity was checked using the melting curve and agarose gel (2%) electrophoreses with a Gene Ruler 50bp DNA ladder (Thermo Scientific). Specific mRNA levels were normalized to the level of the reference gene Rpl32 in the same sample. Data were analyzed employing the ‘advanced relative quantification’ and ‘standard curve method’. A calibrator cDNA was included in each run to correct for run-to-run differences. All solvents, reagents, and lipid standards were used in the highest available purity. Internal standards were acquired from Avanti Polar Lipids (Avanti Polar Lipids, Alabama, USA). Storage solution was created by combining chloroform, methanol, and water (20/10/1.5; v/v/v). ESI spray mixture was created by combining chloroform, methanol with added 0.1% (wt/v) ammonia acetate and 2-propanol (1:2:4; v/v/v). A customized methyl-tert-butyl ether (MTBE)-based lipid extraction method was applied for BAL and lung tissue homogenates. Details can be found in the S1 Text. Lipid extracts were stored in a storage solution at -20°C until further usage. Cholesterol determination was performed as described earlier (S1 Text). Aliquots of 10 μL of the lipid extract and cholesterol derivatization were diluted in 190 μL ESI spray mixture for each measurement, vigorously mixed, and centrifuged prior to loading into a 96 well plate (Eppendorf, Hamburg, Germany). The well plate was sealed with aluminum sealing foil and kept at 15°C during the measurement process. All measurements were performed in duplicate. All mass spectrometric acquisitions were performed using a Triversa Nanomate (Advion, Ithaca, USA) as an autosampler and nano-ESI source, applying a spray voltage of 1.1 kV and backpressure of 1.1 psi in ionization modes. For the acquisition of tandem mass spectrometric data, a Q Exactive Plus was used (Thermo Fisher Scientific, Bremen, Germany; methodological details are found in the S1 Text). For data interpretation, the software pipeline described earlier was utilized (see details S1 Text) . Briefly, mass spectra raw files were converted to mzML format with the software module "msconvert" from ProteoWizard version 3.0.18212-6dd99b0f6 . Converted files were then imported into LipidXplorer version 1.2.8.1 . For each ESI mode, a different set of MFQL files is used. The files generated in LipidXplorer were used in lxPostman for further post-processing, including quality control and quantitation. All details on how the lipidome data (S1 Data) was generated for visualization and statistical analysis is available at LIFS webportal https://lifs-tools.org/publications/99-shotgun-lipidomics-data-set-lamp3-is-critical-for-surfactant-homeostasis-in-mice.html Two-way ANOVA was performed with the parameters matching time points for each row, mixed-effects with full fitting, and within each row comparing columns with correcting for multiple comparisons via Tukey. Lipid values are expressed as mean ± standard deviation and significance is designated as *, P < 0.05; **, P < 0.01; ***, P < 0.001. For easier visualization of significant lipid results, lipids from two-way ANOVA were autoscaled. In this procedure, data are centered on the means of groups and then divided by the standard deviation, removing the impact of differences between lipid abundance. Cell free BAL was weighed (= 400–800 μl) and ultra-centrifuged (J2-MC, Beckman Coulter, Krefeld, Germany) for 1h at 4°C at 38,730 g. The supernatant was decanted. The pellet was weighed and resuspended using saline containing 1.5 mmol/L CaCL2 equivalent to a concentration of 1:50 (w/w) compared to the basic BAL. Two μl of the concentrated BAL were analyzed for content of choline-containing phospholipids (DPPC, lysolecithin, sphingomyelin; LabAssay Phospholipids, FUJIFILM WAKO). Then, samples were normalized to final phospholipids-, i.e. choline containing phospholipids-, concentration of 1.5 mg/ml and analyzed in the Pulsating Bubble Surfactometer . A Pulsating Bubble Surfactometer -sample chamber was prefilled using degassed 10% sucrose in saline plus 1.5 mmol/l CaCL2. Then, 5 μl of each normalized BAL sample were pipetted on the top of the sucrose prefill close to the chimney, where it remained by buoyancy. Subsequently, to mounting of the chamber in the Pulsating Bubble Surfactometer, an air bubble was created by aspiration of ambient air through the chimney and phospholipids were allowed to adsorb for 30s to the air/liquid interface. Then, the bubble was cyclically expanded and compressed at a frequency of 20/min, a compression rate of 50% surface area and at 37° C. Surface tension at maximum (γmax) and minimum (γmin) bubble size was calculated by the Pulsating Bubble Surfactometer and recorded for a 300s period. Statistical analyses. If not stated otherwise, a two-tailed unpaired t-test was performed using GraphPad Prism Software Version 5.03. Significant values were considered at P < 0.05. Values are expressed as mean ± standard error of the mean (SEM) and significance is designated as *, P < 0.05; **, P < 0.01; ***, P < 0.001. Single datapoints for plotting data are available as S2 Data. LAMP3 is expressed highest in humans and mice in the lung’s AT2 cells, where it localizes to LBs . Notably, a naturally occurring recessive missense LAMP3 variant has recently been associated with a fatal neonatal interstitial lung disease in Airedale Terrier dogs . Therefore, we aimed to analyze the function of LAMP3 in the lung in a genetically more defined animal model and generated Lamp3 knockout mice (Lamp3). CRISPR/Cas9 editing of the Lamp3 allele in mice with guide RNAs targeting exon 2 (Fig 1A and 1B) yielded a 40 bp deletion resulting in an early Stop-codon in one founder, and the mutation was transmitted in the germline (Fig 1C). PCR with primers spanning this deletion was used for genotyping (Fig 1D). Reverse transcriptase PCR and quantitative realtime PCR with total lung RNA as a template with primers specific for LAMP3 revealed a complete absence of LAMP3 transcript, indicating nonsense-mediated decay (Fig 1E and 1F). LAMP3 was readily detectable by immunohistochemistry on lung sections with a monoclonal antibody in AT2 cells of wildtype mice but was completely absent in Lamp3 mice, validating the knockout at the protein level (Fig 1G). We hypothesized that knockout of Lamp3 in mice causes abnormal surfactant homeostasis and postnatal death. However, in contrast to dogs with a LAMP3 mutation, homozygous Lamp3 mice were surprisingly born according to the expected Mendelian distribution and survived the weaning phase without increased mortality (Fig 1H). No premature death was observed until 15 months-of-age, indicating that LAMP3 is not essential for survival in mice. Macroscopically, Lamp3 mice were indistinguishable from wildtype littermates with no obvious phenotype. Histology and analysis of semi-thin sections revealed normal micro-anatomy of the lung with a typical distribution of AT2 cells, type I pneumocytes, alveolar macrophages, and regular capillaries (Fig 1I). Macroscopic signs of multifocal emphysema or edema (as previously described in dogs with a mutation in LAMP3, ) were absent in Lamp3 mice (Fig 1J). The lung volume, normalized to the body weight in 3-month-old animals, was comparable between wildtype and Lamp3 mice (Fig 1K). (A) CRISPR guide-RNA sequences used for targeted editing of the murine Lamp3 allele. The Protospacer Adjacent Motif (PAM) is underlined. (B) Schematic drawing of CRISPR/Cas9-mediated targeting of exon 2 of the Lamp3 locus. (C) Sanger sequencing chromatogram of a PCR-product covering the targeted segment with genomic tail-DNA of one wild type and one homozygous Lamp3 mouse as a template. CRISPR target site one is underlined. The frameshift resulting from endogenous repair mechanisms results in an early stop codon. (D) Agarose gel with PCR-products of a PCR covering parts of exon 2 of the Lamp3 locus used for genotyping. (E) Representative agarose gel with the PCR products of reverse transcriptase reactions with total lung RNA as a template and primers specific for Lamp3 from two wildtype and two Lamp3 mice. (F) Quantitative real-time PCR of total RNA from wildtype and Lamp3 mice for LAMP3. The expression is was normalized to the housekeeping gene RPL32. n = 4 (per genotype); *** = p ≤ 0.001. (G) Immunohistochemistry staining of lung sections of wildtype and Lamp3 mice with an antibody specific for LAMP3. (H) Genotype distribution of the litters from heterozygote Lamp3breeding pairs after weaning (3–4 weeks after birth). N = 208 individuals. (I) Left: Hematoxylin / Eosin staining of lung sections from adult wildtype and Lamp3 mice. Right: Toluidine blue staining of plastic-embedded semi-thin sections from adult wildtype and Lamp3 mice. (J) Photos of the inflated lungs of adult wildtype and Lamp3 mice. (K) Lung volume/body weight ratio of wild type and Lamp3 mice. n = 8 (per genotype), ns = not significant. (L) The number of inflammatory cells (lymphocytes, neutrophils, eosinophils, and macrophages) in the BAL of wildtype and Lamp3 mice. n = 8 (per genotype), ns = not significant; 0.05; * = p ≤ 0.05. Furthermore, lung function analysis revealed no abnormalities in the breathing pattern of Lamp3 mice regarding the frequency, flow, volume, and time parameters as non-invasively assessed in spontaneously breathing animals as well as for airway resistance and compliance as recorded invasively in relaxed, ventilated animals (Table 1). Analysis of the BAL cellular composition revealed a tendency towards a higher number of lymphocytes and neutrophils in Lamp3 mice, which, however, did not reach statistical significance. The number of macrophages was moderately increased in Lamp3 mice (Fig 1L). Airway resistance (RI) and dynamic compliance were measured in anesthetized and ventilated mice using FinePointe RC Units. Midexpiratory flow (EF50), frequency (f), functional residual capacity (Frc), minute volume (MV), peak expiratory flow (PEF), peak inspiratory flow (PIF), expiration time (Te), inspiration time (Ti), and tidal volume (TV) were measured in conscious, spontaneously breathing mice using FinePointe non-invasive airway mechanics (NAM) double chamber plethysmography. Values displayed as mean ± SEM, n = 7 per genotype. We reasoned that a lack of LAMP3 might affect the levels of the surfactant proteins SP-B and SP-C, major components of the LB-stored- and secreted surfactant. Indeed, quantitative PCR analysis revealed reduced levels of Sftpb transcript and a similar trend (though not statistically significant) for Sftpc in total lung RNA (Fig 2A) of Lamp3 knockout mice. In contrast, immunoblot analysis of lung tissue lysates from wildtype and Lamp3 mice for SP-B and SP-C showed normal levels of pro-SP-B and a striking increase in the levels of pro-SP-C, the immature pro-form of SP-C, indicating a critical role of LAMP3 in the biosynthesis or turnover of SP-C (Fig 2B and 2C). The levels of mature SP-B were also significantly increased, though only to a moderate level compared to the increase in SP-C. (A) Quantitative real-time PCR of total RNA from wildtype and Lamp3 mice for SP-B and SP-C. The expression was normalized to the housekeeping gene RPL32. n = 8 (per genotype), ns = not significant; ** = p ≤ 0.01. (B) Representative immunoblots of lung tissue lysates from wildtype and Lamp3 mice for pro-SP-C, pro-SP-B, and mature SP-B. Actin is depicted as a loading control. (C) Quantification of immunoblots for pro-SP-C, pro-SP-B, and mature SP-B normalized to Actin as a loading control. The average of the wildtype group was set as “1”. n = 8 (per genotype); ns = not significant; *** = p ≤ 0.001. (D) Immunofluorescence staining of lung sections from adult wildtype and Lamp3 mice for ABCA3 and mature SP-B (green). Nuclei are stained with DAPI (blue). Immunofluorescence analysis on lung sections from wildtype and Lamp3 mice revealed no major differences in the staining pattern of the essential LB proteins ABCA3 or mature SP-B. Both proteins localized in wildtype and Lamp3 mice to ring-like structures of AT2 cells, representing LBs as determined by high-resolution fluorescence microscopy (Fig 2D). Given the apparent differences in SP-C levels and mature SP-B in protein extracts of lung tissue lysates, we next analyzed the lipid composition of the BAL and total lung homogenates by a semi-targeted shotgun lipidomics approach. After shotgun lipidomics, a total of 208 different lipid species in the BAL were quantified. We first determined if the two groups (wildtype and Lamp3 BAL lipids) could be separated by analyzing the lipidomics dataset by unsupervised principal component analysis (PCA) (Fig 3A). The apparent separation was characterized by increased PGs and LPG levels in the wildtype while SMs are increased in the Lamp3 mice. To gain better insight into the perturbations on the lipid species level, we further applied hierarchical clustering (Fig 3B). In conjunction with the PCA, wildtype and Lamp3 mouse BAL lipid profiles were clearly distinguished by the underlying Spearman Rank correlation. According to the genotype, the abundance of lipid species correlated for PG, LPG, and DAG, showing a relative increase in wildtype mice, which was also confirmed by statistical analysis for the selected indicated lipids species (Fig 3C). In contrast, PI and LPI were decreased in Lamp3 mice. Notably, PG is an essential component of the pulmonary surfactant, and as multiple lipids of the same class were affected, a disruption of the normal BAL physical properties is likely. In contrast to BAL’s lipid profiles, no statistical differences were identified between wild type and Lamp3 mice by lipidomic analysis of the perfused total lung tissue (S1A–S1C Fig). These findings, in summary, indicate that LAMP3 deficiency leads to subtle but statistically significant changes in the lipidome and particularly surfactant-relevant lipids of the BAL. (A) Three-dimensional principal component analysis (PCA) loadings plot of the BAL lipidome of wild type (black) and Lamp3 (green) mice. Each point represents one individual animal. Arrows show the 10 lipid species with the strongest effect across all principal components. Ellipses show the 95% confidence interval of the group. PC1 to PC3 explain 65% of the variability in the data set. (B) Hierarchical clustering of 140 lipid species (of 208 after application of a 90% occupation threshold) identified in eight wild type and seven Lamp3 mice BAL fluid samples. Each row represents a lipid species and each column a sample. Red boxes indicate enlarged selections (C2) and (C3). (C1) The hierarchical clustered tree of sample groups correlates both groups into two major branches. (C2) The Upper selected segment of the HCA with 31 lipids shown with a relative decrease to wildtype mice. (C3) The lower selected segment of the HCA was shown with a relative increase of 7 lipids to the control. (D) Differences in the BAL lipidome of wild type and Lamp3 mice mutant using autoscaled ranges. Samples are color-coded according to their group, black for Lamp3 and green for Lamp3 mice. Boxplots show the median with the interquartile range. Whiskers show maximum and minimum with outliers in the respective colors. Data is on the basis of mole percentage. Only lipid species are shown with a significant difference computed with a two-way ANOVA. Defects in surfactant homeostasis are typically accompanied by a change of the ultrastructure of LBs . This finding prompted the analysis of the Lamp3 lungs by transmission electron microscopy, focusing on LBs in the AT2 cells. Lamellar bodies from wildtype mice had the typical lamellar appearance (Fig 4A). In one of the three Lamp3 mice that were analyzed by transmission electron microscopy, conspicuous lamellar bodies could occasionally be observed within AT2 cells (Fig 4A). They consisted of several clews of lipid lamellae and seemingly homogenous areas that were surrounded by a common limiting membrane. At higher magnification, the homogenous areas also revealed lamellae that were not immediately obvious at lower magnification (Fig 4A). However, lamellar bodies in the other two animals had a regular appearance comparable to the wildtype. Accordingly, we tested the surfactant’s biophysical properties and functionality obtained from BAL of wildtype and Lamp3 mice to address any functional relevance of such morphological alterations. Surfactant samples were investigated during dynamic compression of an air bubble in a pulsating bubble . The surface tension (γ) of wildtype and Lamp3 mice is summarized at minimum (γmin) (Fig 4B) and maximum (γmax) bubble size (Fig 4C). Surfactant standardized at a phospholipid concentration of 1.5 mg/ml of both wildtype and Lamp3 mice reached a low physiological γmin close to 0 mN/m during the observational period of 300s, respectively (Fig 4B). During the compression period, γmin (0.95 vs 3.6 mN/m, median, p<0.01) and γmax (29.2 vs. 29.9 mN/m, median, p<0.0001) in Lamp3 mice were found to be moderatly lower compared to wild type surfactant. Thus initially, γmin of samples from Lamp3 mice decreased faster, reaching <10 mN/m after 40s compared to wildtype surfactant after 80s (mean) (Fig 4C), indicating a slightly increased adsorption property of surfactant from Lamp3 mice. (A) Transmission electron microscopy of LB in an alveolar AT2 cells of a wildtype and a Lamp3 mouse. Wildtype: Individual Lamellar Bodies are clearly discernible and separated (left), containing continuous lipid lamellae (right). Lamp3: Left: Multiple clews of lipid lamellae (white asterisks) are located within shared limiting membranes (white arrowheads). The space between the clews (black asterisks) is entirely filled with light grey material. Right: enlargement of the boxed area indicated left. Higher magnification reveals lamellae also inside the light grey material, which in some cases obviously are continuations of the darker stained lamellae inside the clew (top) or the stack of lamellae bottom left. (B, C) Surface tension at minimum (γmin; C)) and maximum (γmax; B) bubble size, derived from a pulsating bubble surfactometer during a pulsating period of 300s. Eight pellets of BAL from Lamp3 and wild type mice were prepared by ultracentrifugation and resuspended in saline CaCl2 to standardize phospholipids to 1.5 mg/ml each. Phospholipid concentration was determined using a commercial kit. Data are plotted as mean +/-SD. Statistical analysis was performed using an unpaired t-test. The overall phenotype of Lamp3 mice was mild and did not affect lung function to the extent that functional lung tests could reveal alterations in healthy, spontaneously breathing animals (Table 1). We, therefore, wondered if LAMP3 is more critical under challenged, i.e., pathological conditions. We applied a well-described model of experimental allergic asthma (Fig 5A) that is characterized by impaired lung physiology arising on the basis of allergic airway inflammation, mucus hyperproduction, and airway hyperresponsiveness (AHR), but provides the advantage that LAMP3-expressing AT2 cells remain largely untouched by the induced disease. Analysis of the different immune-cell populations (macrophages, lymphocytes, neutrophils, eosinophils) infiltrating the bronchoalveolar lumen in diseased animals did not unveil any differences between wildtype and Lamp3 mice (Fig 5B). The number of mucus-producing goblet cells and the amount of mucus stored in the airway mucosa also did not differ (S3 Fig). The degree of apoptotic cell death as revealed by terminal deoxynucleotidyl transferase dUTP nick end labeling (TUNEL) staining between wildtype and Lamp3 mice remained unchanged (S4 Fig). Remarkably, when animals were subjected to MCh-provocation testing to determine the degree of AHR, Lamp3 mice displayed significantly increased airway resistance in response to MCh-inhalation (Fig 5C). In contrast, baseline airway resistance values did not differ between OVA-sensitized and challenged wildtype and Lamp3 mice (S5 Fig). (A) Experimental setup of the stress-challenged model of allergic asthma and the experimental groups. (B) Number of inflammatory cells in the BAL of wildtype, Lamp3, wildtype + OVA, Lamp3 +OVA mice. N = 8 (per genotype), ns = not significant; 0.05; * = p ≤ 0.05. (C) Increase in airway resistance relative to baseline (%) of the four experimental groups (wildtype, Lamp3, wildtype + OVA, Lamp3 +OVA). N = 8 (per genotype), significances calculated for methacholine provocation test after 100 mg/ml methacholine exposure, ns = not significant; 0.05; **** = p ≤ 0.0001. (D) Two-dimensional PCA loadings plot of the BAL lipidome of wild type (black), Lamp3 (green), wild type +OVA (blue), and Lamp3 +OVA (red) mice. The datasets for wild type (black), Lamp3 (green) are identical to (Fig 3A). Each point represents one individual. Arrows show the 10 lipid species with the strongest effect on all principal components. Ellipses indicate the 95% confidence interval of the experimental groups, where PC1 and PC2 describe 48.9% of the variability in the data set. The separation of the groups is in PC1, and PC2 is not complete for all groups. A clear separation between the Lamp3 +PBS and Lamp3 +OVA can be seen. The separation between untreated Lamp3 and treated Lamp3 is not complete. (E) Hierarchical clustering of 140 lipid species (of 208 after application of a 90% occupation threshold) identified in 8 wild type +OVA and 8 Lamp3 +OVA mice BAL fluid samples. Each row represents a lipid species and each column a sample. Red boxes indicate selections shown in (F). The two major branches of the tree represent the genotypes without false assignments. Selected segments of the hierarchical clustered tree where a relative decrease for 21 lipids for Lamp3, mainly DAG and PG species, compared to the Lamp3 mice are observed. The lower selection shows, in contrast, a cluster of 26 lipids with a relative increase in quantity in the mutant compared to Lamp3. (G) Differences in the BAL lipidome of wild type +OVA and Lamp3 +OVA applying autoscaling. Samples are color-coded according to their group, red for Lamp3 +OVA and blue for Lamp3 +OVA. Boxplots show the median with the interquartile range, and whiskers show maximum and minimum with outliers in the respective colors. Data is based on mole percentage. Selected lipid species showed a significant difference by two-way ANOVA. (H) Bar plots of absolute lipid values normalized on protein content. Error bars signify standard deviations. To further investigate the impact of experimental asthma in mice, the lipid composition of the BAL and total lung tissue was determined (Fig 5D–5G). The response to experimentally induced asthma led to further divergence in the BAL lipid composition between Lamp3 mice and wildtype. Further increases in PG, LPG, and DAG abundance were observed in wildtype mice while levels of SM and PI increased in Lamp3. In the lipidomes of the perfused total lung tissue, no further changes were observed (S2 Fig). Hierarchical cluster analysis further revealed that the opposite response in the BAL lipidome of the wildtype and Lamp3 mice affected a wider range of lipid species, compared to the healthy animals (Figs 5E and 3B–3D). Spearman Rank correlation between the two distinct groups (wildtype and Lamp3) further confirmed this lipid phenotype by high correlation factor for the two major branches comprising lipid compositions of 8 animals in each group (Fig 5F). Lipid species of mainly PG, LPG, and DAG classes showed a relative increase in Lamp3 mice, which were statistically significant for indicated lipid species. In contrast, PI, LPE, Cer, and SM showed increased abundance in Lamp3 mice (Fig 5G). In summary, the lipid homeostasis for surfactant production in BAL of Lamp3 mice was more disturbed under the pathophysiologic conditions of experimental asthma when compared to the unchallenged animals. The two best-characterized family members of the LAMP family, LAMP1 and LAMP2, are ubiquitously expressed and quantitatively account for a substantial fraction of the lysosomal membrane proteome . Though the function of LAMP1 and LAMP2 are still not fully understood, they are supposed to play pivotal roles in the fusion of membranes, autophagy, microtubule-dependent positioning of lysosomes, and especially as structural components of the lysosomal membrane . In contrast, the expression of the structurally related LAMP3 protein is limited to specific cell types, implicating a more defined and tissue/cell-type-specific function. Together with the fact that LAMP3 localizes to LBs rather than lysosomes, a critical function of LAMP3 in LB function and surfactant homeostasis is suggested. In fact, the recent finding that a natural mutation (p.(E387K) in LAMP3 leads to clinical symptoms and pathology similar to the most severe neonatal forms of surfactant deficiency in an Airedale Terrier dog breed suggested a critical role in surfactant biology and lung physiology for LAMP3. Therefore, we aimed to investigate the role of LAMP3 in lung physiology by using a genuine knockout model. To our surprise, our findings obtained from Lamp3 mice revealed striking differences to those found in dogs with the recessive missense LAMP3 variant: While lethal hypoxic respiratory distress and failure lead to the premature death of dog puppies within the first days or week of living , Lamp3 mice developed normally and did not display signs of lung pathology and dysfunction after birth. We can only speculate about the apparent difference between the different animal models: On the one hand, in contrast to laboratory mouse strains, dog breeds like the Airedale Terrier were generated by inbreeding that aimed to pronounce or produce distinct anatomical features. Thus, it could not be excluded that other genetic factors, possibly homozygous in the inbred dogs, impact this phenotype. On the other hand, the point mutation (p.(E387K)) found in the dogs resulted in the expression of a LAMP3 variant that could still be detected with a LAMP3-specific antibody in lung tissue , suggesting that the overall protein structure of the LAMP3 variant is at least in part untouched by the mutation. This finding leaves the possibility of an impaired function of the mutated variant, which is not the case when LAMP3 is completely knocked-out by CRISPR/Cas9. Another possibility potentially explaining the differences between the complete knockout and the point-mutation is genetic compensation, a process of transcriptional adaptation which acts independently of protein feedback loops . While the knockout might lead to the upregulation of compensatory mechanisms, attenuating the phenotype in mice, the point mutation might not induce such compensatory pathways. These findings, however, should be considered when human patients suffering from surfactant deficiencies are analyzed: Our data suggest that LAMP3 mutations might also present with a mild clinical phenotype, perhaps not immediately presenting with the typical signs of a full surfactant deficiency in human patients suffering from childhood interstitial lung disease. Since neither snap-frozen tissue samples nor BAL from the affected Airedale Terrier puppies could be obtained, characterization of the phenotype caused by a mutation of the LAMP3 gene in dogs is quite limited in relation to the lipidome profile of BAL and the function of the surfactant. The lipidomics analyses performed in this study are in good agreement with earlier reported quantities of main lipid species and classes by Prueitt et al. . We find PC at 67.1 mol% in wildtype mice (16 mol% neutral lipids excluded as in study) compared to the previously reported 75.2 mol% . The main contributing PC class lipid species is DPPC (PC 16:0/16:0) with 43.9% compared to the 30–60% reported in the literature . PG was found at 9 mol% in agreement with the values reported by others . Our analyses show a clear separation of the genotypes according to different lipid compositions and upon induction of experimental asthma. In the lipidomes of the perfused lung tissues, only minor Lamp3 genotype-dependent changes were observed, which underlines the importance of LAMP3 to regulate the lipid secretion in the surfactant. It is further noteworthy that the BAL lipid composition was mostly affected by the reduction in the abundance of PGs, LPG, and DAG in the Lamp3 mice. In transmission electron microscopy, phospholipids typically appear as bilayer lamellae, whereas neutral lipids typically form homogenous lipid droplets. Thus, the occurrence of conspicuous lamellar bodies in one out of three transgenic mice may indicate altered metabolism and/or trafficking of lipids in AT2 cells; however, the finding that LBs in the other two animals had a regular shape comparable to the wildtype puts this interpretation into perspective. Remarkably, the alterations of the surfactant lipid and protein composition were not associated with pronounced changes of its adsorption properties, which even appeared to be slightly increased in Lamp3 mice. Indeed, this is quite in line with our findings on the lung structure and function of these mice under healthy conditions. Nevertheless, it does not exclude the possibility that an altered surfactant composition could have an impact on lung physiology under disease conditions, when the functional reserves of the respiratory system are exhausted. The altered lipid composition and particularly the increase in SP-C could cause enhanced adsorption of surfactant, and a more detailed functional analysis of surfactant is warranted. Hence, experimentally induced asthma further reduced the levels of essential surfactant lipids and, in parallel, led to an increase in lipids like SM, PI, and PS. Interestingly lipids containing arachidonic acid (AA, FA 20:4) were elevated in Lamp3 mice. Further effects on signaling were not investigated in this study, although AA plays a crucial role as educt of numerous lipid mediators . The increase in SM, PI and PS might indicate a higher degree of cell lysis in conjunction with increased number of immune cells in BAL after OVA treatment and an excessive immune reaction. However, it should be noted that we did not observe major differences in either the numbers, the numbers of apoptotic cells in lung tissues or infiltrating immune cells (e.g. eosinophils, lymphocytes, etc.) in BAL, which indicates that the sensitization and subsequent inflammation of the airways in reaction to allergen inhalation are not impacted by the knockout of Lamp3. In turn, these findings suggest that the changes in surfactant composition and its function give rise to an altered inflammatory response in the airways or eventually related apoptosis of cells. While our data support a role of LAMP3 in surfactant homeostasis, we can only speculate about the mechanisms of how LAMP3 might regulate surfactant levels secreted by AT2 cells. Different scenarios are feasible: On the one hand, it has been speculated previously that LAMP3 plays a role in surfactant recycling by retrieving remnants of secreted LB via re-endocytosis . On the other hand, other LAMP family members act as accessory subunits of multi-transmembrane spanning proteins like LAMP1/LAMP2 for the lysosomal polypeptide transporter TAPL . Likewise, UNC-46, the C. elegans orthologue of LAMP5, acts as a trafficking chaperone, essential for the correct targeting of the nematode vesicular GABA-transporter UNC-47 . LAMP3 could take over a similar function, e.g., by modulating or regulating the trafficking or stability of one of the integral LB proteins like SP-C (which is synthesized as a type II transmembrane protein). Interestingly, the transcript-levels of Sftpb and Sftpc were significantly reduced or showed a trend towards a decrease, while protein levels were distinctively increased, suggesting posttranslational mechanisms governing the increased stability thereby leading to higher steady-state protein levels: a finding that could hint towards a stabilizing effect of LAMP3.However, such a mechanism might not be essential and become relevant only under conditions where surfactant levels need to be regulated more tightly. It is interesting to note that an aggravation of the lipid changes under the pathological conditions of experimental allergic asthma was associated with the lung function differences observed after MCh-provocation, when lung function was pushed to the limit. In sensitized animals, inhalation of allergen aerosols results in an acute inflammatory reaction with infiltration of T helper 2 (TH2) cells and eosinophils into the airways and the release of a plethora of cytokines, chemokines, and cytotoxic metabolites . Consequently, this causes activation of submucosal glands and differentiation of goblet cells, ultimately resulting in increased mucus production. TH2-type cytokines, together with eosinophil products, trigger the development of AHR that leads to an increased airway resistance in response to non-allergic stimuli and the formation of the typical asthma attack . Therefore, in mice with experimental allergic asthma, the degree of allergic airway inflammation is closely correlated with the degree of mucus production, and AHR. Remarkably, this was not the case in Lamp3 mice: Numbers of inflammatory cells such as eosinophils and lymphocytes as well as the number of goblet cells and amount of mucus stored in the airway mucosa are not different between wildtype and Lamp3 mice. The lack of any differences between Lamp3 and wildtype animals could be explained by the fact that the pathologic features of experimental allergic asthma are restricted to the airways, while the knock-out of Lamp3 affects AT2 cells, which are located in the alveoli, which remain largely untouched. Nevertheless Lamp3 mice displayed a markedly increased airway resistance in response to MCh inhalation. Since this response appeared to be independent of both the degree of airway inflammation and mucus production, respectively, it is possible that the increase in airway resistance was not compelling due to an increased AHR. Whether the alterations of the surfactant lipid and protein composition, which we found in healthy Lamp3 mice, could be causative for the observed impairment of lung function remains speculative. To our knowledge alteration of surfactant lipid or SP-B/C composition have not been connected to increased MCh-response of the airways yet, the surfactant film lining the alveoli and conducting airways has an impact on airway resistance. Hence, during expiration this film can develop high surface pressure and thus low surface tension, respectively, which in turn counteracts the tendency of liquid to accumulate in the airway’s most narrow section . At this point, alterations of the surfactant composition may result in an increased liquid accumulation and consequently blocking of terminal airways, which further leads to an increased airway resistance. Taken together, whereas the lack of LAMP3 appears to be compensable under physiological conditions, it impacts basal surfactant homeostasis and lung lipid composition and airway resistance in experimental allergic asthma.
|
PMC7449968
|
Maternal lipid levels across pregnancy impact the umbilical cord blood lipidome and infant birth weight
|
Major alterations in metabolism occur during pregnancy enabling the mother to provide adequate nutrients to support infant development, affecting birth weight (BW) and potentially long-term risk of obesity and cardiometabolic disease. We classified dynamic changes in the maternal lipidome during pregnancy and identified lipids associated with Fenton BW z-score and the umbilical cord blood (CB) lipidome. Lipidomics was performed on first trimester maternal plasma (M1), delivery maternal plasma (M3), and CB plasma in 106 mother-infant dyads. Shifts in the maternal and CB lipidome were consistent with the selective transport of long-chain polyunsaturated fatty acids (PUFA) as well as lysophosphatidylcholine (LysoPC) and lysophosphatidylethanolamine (LysoPE) species into CB. Partial correlation networks demonstrated fluctuations in correlations between lipid groups at M1, M3, and CB, signifying differences in lipid metabolism. Using linear models, LysoPC and LysoPE groups in CB were positively associated with BW. M1 PUFA containing triglycerides (TG) and phospholipids were correlated with CB LysoPC and LysoPE species and total CB polyunsaturated TGs. These results indicate that early gestational maternal lipid levels influence the CB lipidome and its relationship with BW, suggesting an opportunity to modulate maternal diet and improve long-term offspring cardiometabolic health.The Developmental Origins of Health and Disease (DOHaD) theory describes how insults during early life can permanently program the fetus/offspring, altering their risk of adult chronic disease. To adapt to the intrauterine environment (i.e. maternal nutrient supply), structural changes and functional modifications occur in fetal organs and tissues to ensure survival of the newborn, indicative of their developmental plasticity. Barker and Osmond developed this hypothesis from observations in England and Wales, confirmed by studies from the Dutch Famine, a well-documented famine in the Netherlands during World War II. Within this cohort, nutrient restriction in utero in the first trimester was associated with increased risk of obesity, dyslipidemia and cardiovascular disease, while restriction in the third trimester was associated with decreased risk of metabolic disease independent of birth weight (BW), affirming existence of distinct susceptibility windows for producing differing outcomes. Infant BW is the most common health outcome studied to determine if the maternal intrauterine environment influences the development of the fetus. Both low and high BW have been associated with an increased risk of obesity and cardiometabolic diseases, including type 2 diabetes (T2D). In recent years, profiling of small molecular weight compounds in a biological sample by metabolomics has been used to obtain an objective measurement of the metabolic environment to which the developing fetus is exposed. Metabolite levels are influenced by dietary intake and can both reflect and influence changes in metabolism. The application of metabolomics to developmental studies can identify changes in maternal nutrient availability across pregnancy. Using a targeted metabolomics approach, we previously reported an increase in plasma long-chain fatty acids and corresponding long-chain acylcarnitines (AC) in maternal plasma from first trimester to term. These results reflect the increase in lipolysis during late-gestation to fuel rapid fetal growth. Using a larger cohort and a comprehensive lipidomics platform, we aim to expand our previous findings and detail changes in phospholipids, ceramides (CER), cholesteryl esters (CE), and triglycerides (TG) during pregnancy. Metabolomics analyses have been applied to identify trimester-specific maternal metabolites associated with infant BW and adiposity, accounting for maternal characteristics such as pre-pregnancy BMI. Maternal metabolite levels, placental transfer, and interaction with other metabolites, such as inhibition by competition or metabolite-induced changes in placental transport activity, can all affect the relative levels of the umbilical cord blood (CB) metabolome. The CB metabolome is influenced by maternal characteristics such as pre-pregnancy BMI. The metabolome from CB has been correlated with BW finding positive associations with branched chain amino acids, AC, and phosphatidylcholines (PC) along with inverse associations with TGs. Other studies have found that CB lysophosphatidylcholine (LysoPC) metabolites with varying chain length and saturation are positively associated with newborn BW, as well as with weight at 6 months of age. However, what remains unclear is how the maternal lipidome influences that establishment of CB lipids related to BW. In this study we profiled 573 lipid species in 106 mother-infant dyads from maternal first trimester (M1) and delivery blood (M3) as well as umbilical CB. Women were recruited during the first trimester and followed through delivery (Fig. 1a). Average maternal age at their first trimester visit was 32.1 ± 3.6 years and the majority of women had a lean baseline BMI (62%) (Table 1) and were white, non-Hispanic (89%). On average, women gained 13.2 ± 5.2 kg between the baseline visit and term and 72% of infants were born vaginally. Offspring analysis was stratified by males (n = 51) and females (n = 55). We observed no significant difference in Fenton BW percentile between males and females.Figure 1Study design and lipidomics analysis strategy. (a) Women were recruited into the Michigan Mother-Infant Pairs Study at their first prenatal appointment between 8 and 14 weeks pregnancy (M1). Venous blood samples were collected at this visit, along with weight, height, and demographic information. At delivery, maternal venous blood samples (M3) and cord blood samples via venipuncture from the umbilical cord (CB) were collected, as well as birth weight. (b) Untargeted shotgun lipidomics was conducted on maternal first trimester, delivery, and cord blood plasma. After peak detection and data normalization, 573 lipids of multiple classes were identified. (c) Statistical analysis for this manuscript classified (1) relative differences in the lipidome across pregnancy (paired t-test), (2) differences in the connections of lipid groups between M1, M3, and CB (debiased sparse partial correlations), (3) lipid groups associated to birth weight (linear regression), and (4) M1 and M3 lipids that may modulate CB lipids significantly related to birth weight (Pearson’s correlations).Table 1Characteristics of the study populations among 106 MMIP participants, stratified by sex.Categorical variablesAllMalesFemalesp valuen (%)n (%)n (%)Parity033 (31%)16 (31%)17 (31%)0.988145 (42%)21 (41%)24 (44%)221 (20%)11 (22%)10 (18%) ≥ 37 (7%)3 (6%)4 (7%)Baseline BMI (kg/m) (n = 103)18.5–24.964 (62%)30 (61%)34 (63%)0.50425.0–29.918 (17%)11 (22%)7 (13%)30.0–34.912 (12%)4 (8%)8 (15%) ≥ 35.09 (9%)4 (8%)5 (9%)Delivery modeVaginal76 (72%)35 (69%)41 (75%)0.499Caesarean Section30 (28%)16 (31%)14 (25%)Continuous variablesAllMalesFemalesp valuemean ± SD (n)mean ± SD (n)mean ± SD (n)Maternal baseline characteristicsAge (years)32.1 ± 3.632.4 ± 4.131.7 ± 3.20.328Weight (kg)71.1 ± 17.669.4 ± 13.972.6 ± 20.40.347Height (cm)165 ± 7 (103)165 ± 6 (49)165 ± 7 (54)0.858BMI (kg/m)25.8 ± 6.0 (103)25.4 ± 5.3 (49)26.2 ± 6.6 (54)0.519Maternal and newborn delivery characteristicsGestation age (days)278 ± 7278 ± 8278 ± 60.981GWG (kg)13.2 ± 5.213.9 ± 4.512.6 ± 5.80.202Birth weight (g)3,510 ± 436 (105)3,638 ± 4673,390 ± 370 (54)0.003Fenton BW z score0.09 ± 0.87 (105)0.20 ± 0.95− 0.01 ± 0.78 (54)0.208Fenton BW percentile53.1 ± 26.1 (105)56.0 ± 28.250.4 ± 23.8 (54)0.273Head circumference (cm)35 ± 1 (101)35.2 ± 1.2 (48)34.7 ± 1.3 (53)0.048Fenton HC z score0.10 ± 0.86 (101)0.16 ± 0.72 (48)0.04 ± 0.96 (53)0.485Fenton HC percentile52.5 ± 25.5 (101)54.9 ± 22.9 (48)50.4 ± 27.6 (53)0.371Values are n (%) or mean ± SD. Stratified by males (n = 51) and females (n = 55). Represents Pearson's chi-square test for categorical variables. Represents unpaired t test for continuous variables. If sample size differs, mean ± SD (n). Study design and lipidomics analysis strategy. (a) Women were recruited into the Michigan Mother-Infant Pairs Study at their first prenatal appointment between 8 and 14 weeks pregnancy (M1). Venous blood samples were collected at this visit, along with weight, height, and demographic information. At delivery, maternal venous blood samples (M3) and cord blood samples via venipuncture from the umbilical cord (CB) were collected, as well as birth weight. (b) Untargeted shotgun lipidomics was conducted on maternal first trimester, delivery, and cord blood plasma. After peak detection and data normalization, 573 lipids of multiple classes were identified. (c) Statistical analysis for this manuscript classified (1) relative differences in the lipidome across pregnancy (paired t-test), (2) differences in the connections of lipid groups between M1, M3, and CB (debiased sparse partial correlations), (3) lipid groups associated to birth weight (linear regression), and (4) M1 and M3 lipids that may modulate CB lipids significantly related to birth weight (Pearson’s correlations). Characteristics of the study populations among 106 MMIP participants, stratified by sex. Values are n (%) or mean ± SD. Stratified by males (n = 51) and females (n = 55). Represents Pearson's chi-square test for categorical variables. Represents unpaired t test for continuous variables. If sample size differs, mean ± SD (n). Maternal BMI was inversely associated with gestational weight gain (GWG) (beta = − 0.360, r = 0.18, p < 0.001), indicating that leaner women gained more weight during pregnancy (Supplementary Figure S1a). In contrast, maternal BMI was positively associated with Fenton BW percentile (beta = 0.011, r = 0.06, p = 0.014) (Supplementary Figure S1b); and while not significant, maternal GWG trended towards a positive association with Fenton BW percentile (beta = 0.009, r = 0.03, p = 0.070) (Supplementary Figure S1c). We next examined the changes in the plasma lipidome across pregnancy (M1–M3) and between maternal delivery plasma and cord blood (M3–CB), the latter as a potential surrogate for placental transfer and fetal exposure. Figure 2a displays a Heatmap of lipids organized by class with individual lipid peak intensities normalized across M1, M3, and CB. Figure 2b,c displays fold changes (FC) of lipids from M1 to M3 and between M3 and CB, respectively (quantified in Supplementary Table S1). Most classes of lipids increased in levels from M1 to M3, consistent with previous observations of generalized increase in lipoproteins and associated lipids with advancing pregnancy. A subset of lipid species was reduced in M3 compared to M1 including polyunsaturated CEs and a variety of LysoPC and lysophosphatidylethanolamine (LysoPE) species. Most of the polyunsaturated fatty acid (PUFA) containing lipids that were lower in M3 were significantly increased in CB plasma (Fig. 2a,c, Supplementary Table S1). In addition, several lipid species that are statistically unchanged from M1 to M3 were higher in CB compared to M3, including sphingomyelins (SM) and TG enriched in PUFAs. These results are consistent with the known specific transfer of PUFAs across the placenta to the fetus, perhaps mediated by lysophospholipids (LysoPLs) (see below).Figure 2Dynamic changes in maternal and cord blood plasma lipidome during pregnancy. (a) Heatmap of standardized peak intensity for individual lipids (mean 0, standard deviation 1). Lipids are grouped by lipid class with increasing total chain length and double bond from left to right in each lipid class. (b) Differences between M1 and M3 lipidome using log2FC. Lipids that are greater in M1 than M3 have a positive log2FC. Of the 573 lipid species analyzed, 35% were significantly increased and 10% were decreased (Bonferroni adjusted, α = 0.05/573). (c) Differences between M3 and CB lipidome using log2FC. Of the 573 lipid species analyzed, 10% were significantly increased and 61% were decreased (Bonferroni adjusted, α = 0.05/573). Lipids that are greater in M3 than CB have a positive log2FC. Dynamic changes in maternal and cord blood plasma lipidome during pregnancy. (a) Heatmap of standardized peak intensity for individual lipids (mean 0, standard deviation 1). Lipids are grouped by lipid class with increasing total chain length and double bond from left to right in each lipid class. (b) Differences between M1 and M3 lipidome using log2FC. Lipids that are greater in M1 than M3 have a positive log2FC. Of the 573 lipid species analyzed, 35% were significantly increased and 10% were decreased (Bonferroni adjusted, α = 0.05/573). (c) Differences between M3 and CB lipidome using log2FC. Of the 573 lipid species analyzed, 10% were significantly increased and 61% were decreased (Bonferroni adjusted, α = 0.05/573). Lipids that are greater in M3 than CB have a positive log2FC. Given the specific patterns within lipid classes containing PUFAs, lipid clusters were created by grouping lipids based on a priori knowledge of lipid class and the number of double bonds in each lipid species, resulting in 41 groups at each time point (Supplementary Table S2). Using a Debiased Sparse Partial Correlation algorithm, we estimated partial correlation networks for lipid groups within M1, M3, and CB, using an adjusted p-value less than 0.1 as a threshold for including edges in the network (Fig. 3). At each time point, the majority of correlations were positive, signifying the connectivity of lipids across classes. In particular, multiple positive correlations were observed between (1) TGs, diacylglycerols (DG), and phospholipids; (2) CERs and SMs; (3) plasmenyl-phosphatidylcholine (PL-PC) and plasmenyl-phosphatidylethanolamine (PL-PE); and (4) LysoPCs and LysoPEs. Most positive correlations occurred between lipid groups containing lipids with similar number of double bonds. CEs are correlated with a variety of lipid classes including LysoPCs, phospholipids, DGs, SMs, and TGs, suggesting their interaction with a variety of lipid metabolic pathways. Of note, in M1, M3, and CB, CE-poly is inversely associated with DG-mono and, excluding M1, with TG-mono. M3 and CB have more significant correlations with each other than with M1 (Fig. 3), likely due to maternal lipolysis and delivery of fatty acids to the fetal circulation during late gestation resulting in proportional distribution to different lipids species. Additionally, eight partial correlations, all positive, were significant between lipid groups in M1 and M3, including between (1) M1 and M3 PE-poly, (2) M1 and M3 PC-poly, (3) M1 and M3 PLPE-poly, and (4) M1 and M3 SM-poly, suggesting that maternal polyunsaturated fatty acid levels track across pregnancy (Supplementary Figure S2). Partial correlations with an adjusted p-value less than 0.1 are reported in Supplementary Table S3.Figure 3Correlation between lipid groups differs between maternal and cord blood samples. The relationship between lipid groups was estimated using debiased sparse partial correlations within (a) M1 (45 edges), (b) M3 (55 edges), and (c) CB (47 edges). Colors of nodes depict class of lipid. Shapes of nodes depict number of double bonds. Positive correlations are depicted using red edges. Inverse correlations are depicted using green edges. Weight of the edges signifies a more significant correlation. Edges represented have an adjusted p-value < 0.1. Correlation between lipid groups differs between maternal and cord blood samples. The relationship between lipid groups was estimated using debiased sparse partial correlations within (a) M1 (45 edges), (b) M3 (55 edges), and (c) CB (47 edges). Colors of nodes depict class of lipid. Shapes of nodes depict number of double bonds. Positive correlations are depicted using red edges. Inverse correlations are depicted using green edges. Weight of the edges signifies a more significant correlation. Edges represented have an adjusted p-value < 0.1. The next objective was to assess the relationship between maternal characteristics, baseline BMI and GWG, and the lipidome at M1, M3, and CB. At each time point, multiple lipid classes were associated with BMI and GWG, adjusting for sex, maternal age, parity, and gestational age. These associations were largely in opposite directions, due to the inverse correlation between BMI and GWG (Supplementary Figure S1a). Maternal BMI was inversely associated with PCs and PL-PEs at both M1 and M3, while GWG was positively associated with LysoPC and LysoPE clusters at both M1 and M3, independent of the number of double bonds in the fatty acid tails, with stronger associations evident in M3 (FDR < 0.1) (Supplementary Figure S4; Supplementary Table S5). On the other hand, maternal BMI was positively associated with SM-mono and SM-poly at both M1 and M3, while GWG showed inverse associations with SM-mono and SM-poly, again at both M1 and M3 (Supplementary Figure S3, Supplementary Table S4). Finally, maternal BMI was positively associated with saturated phosphatidylglycerols (PG) and SM-poly in CB, whereas few associations were observed between GWG and lipids groups in CB. We identified maternal and CB lipid groups associated with Fenton BW z score (Fig. 4, Supplementary Table S6). In the combined model, the early gestation lipidome (M1) was not associated with infant BW, however, in the sex-stratified models, M1 lipid groups were associated with BW in males (unadjusted p value < 0.05) (Fig. 4a), including positive associations with saturated free fatty acids (FFA) and negative associations with CER-mono, CER-poly, PC-mono, SM-sat, and SM-mono. Within M3, significant positive correlations were observed between BW and DG-sat, LysoPE-sat, PL-PC-sat, SM-mono, SM-poly, and TG-sat (unadjusted p value < 0.05) (Fig. 4b). The direction of these associations was consistent within males and females, however in the sex-stratified models, the relationship reached significance only in females. Of interest, in females, M3 SM-mono and SM-poly were positively associated and CE-poly was inversely associated with BW (FDR < 0.1).Figure 4Distinct maternal and cord blood lipid groups are related to birth weight. Regression models estimated the linear relationship between lipid groups from (a) maternal first trimester, (b) maternal term, and (c) cord blood and Fenton BW z score, adjusting for sex, maternal age, parity, gestational age, baseline BMI, and GWG. For the combined model, beta coefficients plotted as gray bars (β ± SE) with significance depicted as gray striped bars (α = 0.05). Sex stratified models were run. For males, beta coefficients are plotted as blue dots (β ± SE) with significance depicted as blue “X” (α = 0.05). For females, beta coefficients are plotted as red dots (β ± SE) with significance depicted as red “X” (α = 0.05). Distinct maternal and cord blood lipid groups are related to birth weight. Regression models estimated the linear relationship between lipid groups from (a) maternal first trimester, (b) maternal term, and (c) cord blood and Fenton BW z score, adjusting for sex, maternal age, parity, gestational age, baseline BMI, and GWG. For the combined model, beta coefficients plotted as gray bars (β ± SE) with significance depicted as gray striped bars (α = 0.05). Sex stratified models were run. For males, beta coefficients are plotted as blue dots (β ± SE) with significance depicted as blue “X” (α = 0.05). For females, beta coefficients are plotted as red dots (β ± SE) with significance depicted as red “X” (α = 0.05). Within the CB, all LysoPC and LysoPE lipids groups were positively associated with BW (FDR < 0.1), independent of the number of double bonds, consistent in both male and female infants (Fig. 4c). Additionally, CB DG-poly and TG-poly were inversely associated with BW (FDR < 0.1), driven by female infants, and CE-poly was inversely associated with BW (unadjusted p-value < 0.05), driven by male infants. In response to the unique relationship between CB TG-poly and BW, sex-stratified regression models determined the relationship between individual CB TGs (97 triglycerides) and Fenton BW z score, adjusting for maternal age, parity, gestational age, GWG, and BMI (Supplementary Table S7). From the sex-stratified models, beta coefficients were plotted by the number of double bonds in the TGs (Fig. 5). Observing non-linear trends, we used non-parametric regression to fit polynomial lines to each curve, emphasizing the evident sex differences in the relationship between polyunsaturated TGs and BW, again driven by females.Figure 5Association between triglycerides and birth weight in male and female infants. Regression models estimated the linear relationship between cord blood individual triglycerides and Fenton BW z score in (a) males and (b) females, adjusting for maternal age, parity, gestational age, baseline BMI, and GWG. Beta coefficients are plotted by the number of double bonds in the fatty acid tails of the TGs. Significant TGs are marked in red (α = 0.05). Non-parametric regression fit polynomial lines for each plot. Association between triglycerides and birth weight in male and female infants. Regression models estimated the linear relationship between cord blood individual triglycerides and Fenton BW z score in (a) males and (b) females, adjusting for maternal age, parity, gestational age, baseline BMI, and GWG. Beta coefficients are plotted by the number of double bonds in the fatty acid tails of the TGs. Significant TGs are marked in red (α = 0.05). Non-parametric regression fit polynomial lines for each plot. Our results indicated that CB LysoPLs of varying chain length and number of double bonds are elevated in the CB compared to M3 (Fig. 2) and are significantly associated with BW, independent of sex (Fig. 4). To assess if maternal lipids during pregnancy might influence the level of CB LysoPLs, the M1 and M3 lipidome was correlated with individual CB LysoPCs and LysoPEs, as well as the unsupervised Principal Component Analysis scores of LysoPC and LysoPE subgroups (Sat, Mono and Poly) and all LysoPLs (LysoPL-total). Lipids from M1 and M3 that were significantly correlated with the CB LysoPL-total group are displayed in Fig. 6. In M1, 13 lipids positively correlated and 47 lipids inversely correlated with CB total LysoPL levels. Lipids with a positive correlation include AC 18:0 and LysoPC species (16:0, 17:0, 18:0, 19:0, and 20:0), as well as the saturated LysoPC lipid group. Lipids with an inverse correlation mainly include individual lipids and lipid groups with PUFA containing DG, PC, phosphatidylethanolamine (PE), and TGs, highlighting the association between early maternal levels of PUFAs within multiple lipid classes and CB LysoPL-total.Figure 6Maternal lipids correlate with cord blood lysophospholipids. The association between maternal individual lipids and lipid groups with CB LysoPCs, LysoPEs, and total LysoPLs was classified using Pearson’s correlations. Maternal lipids listed were significantly correlated with the lipid group CB LysoPL-total (unadjusted p-value α = 0.05). Maternal lipids (rows) are broken down by positive and inverse associations. Cord blood lysophospholipids (columns) are ordered by increasing chain length and number of double bonds. CB lipids groups LysoPC-sat, LysoPC-mono, LysoPC-poly, LysoPE-sat, LysoPE-poly, LysoPC-total, LysoPE-total, and LysoPL-total are listed in the last columns. Maternal lipids correlate with cord blood lysophospholipids. The association between maternal individual lipids and lipid groups with CB LysoPCs, LysoPEs, and total LysoPLs was classified using Pearson’s correlations. Maternal lipids listed were significantly correlated with the lipid group CB LysoPL-total (unadjusted p-value α = 0.05). Maternal lipids (rows) are broken down by positive and inverse associations. Cord blood lysophospholipids (columns) are ordered by increasing chain length and number of double bonds. CB lipids groups LysoPC-sat, LysoPC-mono, LysoPC-poly, LysoPE-sat, LysoPE-poly, LysoPC-total, LysoPE-total, and LysoPL-total are listed in the last columns. In M3, 34 lipids were positively correlated, and 7 lipids were inversely correlated with CB LysoPL levels. M3 saturated LysoPC and LysoPE lipid groups, as well as individual LysoPL, were positively associated with CB LysoPL-total, implying that higher M3 levels results in higher CB levels. CER and SM containing saturated and monounsaturated TGs were positively associated with CB LysoPL-total. Interestingly, inverse associations were observed between M3 LysoPE 20:3 and 22:5 with CB LysoPL-total, suggesting differences based on the fatty acid tail. Most correlations between the significant M1 and M3 individual lipids and individual CB LysoPCs and LysoPEs were in the same direction as CB LysoPL-total. Subtle variations in the correlation patterns are observed for LysoPC 18:2, 18:3, 20:5, and 26:4 and LysoPE 18:2, 20:3, 20:4, 22:6, and 24:0, potentially suggesting differences in how the maternal lipidome modulates these CB polyunsaturated LysoPLs. The apparent preferential transfer of lipids containing PUFAs into the CB (Fig. 2) and an inverse association between Fenton BW z score and CB lipid groups containing PUFAs (Figs. 4, 5) prompted us to investigate which maternal lipids are associated with the levels of PUFAs in CB. The M1 and M3 lipidome was correlated with CB TGs. Lipids from M1 and M3 that were significantly correlated with the CB TG-poly group are displayed in Fig. 7. A variety of M1 PUFA-containing lipids, including PC, PL-PC, DG, and TG, were positively associated with CB TG-poly. In M3, there were positive correlations between PCs and PEs, 75% of which contained PUFAs, with CB TG-poly, suggesting the mobilization of PUFAs within PCs and PEs to the placenta for transport. Lastly, M3 monounsaturated and polyunsaturated SM were inversely associated with CB TG-poly. Note that these lipids were found to be positively associated with maternal BMI (Supplementary Figure S3) and Fenton BW z score (Fig. 4).Figure 7Maternal lipids correlate with cord blood polyunsaturated triglycerides. The association between maternal individual lipids and lipid groups with CB TGs was classified using Pearson’s correlations. Maternal lipids listed were significantly correlated with the lipid group CB TG-poly (unadjusted p-value α = 0.05). Maternal lipids (rows) are broken down by positive and inverse associations. Cord blood triglycerides (columns) are ordered by increasing chain length and number of double bonds. CB lipids groups TG-sat, TG-mono, and TG-poly are listed as the last three columns. Maternal lipids correlate with cord blood polyunsaturated triglycerides. The association between maternal individual lipids and lipid groups with CB TGs was classified using Pearson’s correlations. Maternal lipids listed were significantly correlated with the lipid group CB TG-poly (unadjusted p-value α = 0.05). Maternal lipids (rows) are broken down by positive and inverse associations. Cord blood triglycerides (columns) are ordered by increasing chain length and number of double bonds. CB lipids groups TG-sat, TG-mono, and TG-poly are listed as the last three columns. Using a comprehensive lipidomics platform including almost 600 lipid species, this study objectively classified the maternal metabolic environment during the first trimester and at delivery as well as the infant lipidome, reflecting infant metabolism. We found dynamic shifts in the maternal lipidome across pregnancy and evidence for selective transfer of lipids containing PUFA from maternal to fetal circulation. We further identified associations between the lipidome and BW, emphasizing the inverse correlations observed with CB DGs and TGs containing PUFAs and positive correlations observed with CB LysoPL. Lastly, we identified maternal lipids which may modulate CB lipid levels that are influential in growth. Early in gestation, insulin sensitivity is enhanced to increase maternal energy storage. However, as pregnancy advances, maternal insulin resistance increases with elevations in lipid parameters such as lipoproteins, TGs, and total cholesterol to support the necessary metabolic changes for pregnancy maintenance and fetal growth. Of the 573 lipids analyzed, 35% were significantly increased while only 10% were decreased in M3 compared to M1 (Fig. 2b, Supplementary Table S1). Between M1 and M3, we observed higher levels of PC, PE, DG, CER, SM, TG, and FFA, many containing fatty acids with 16–18 carbons. The majority of lipid species reduced in M3 are increased in CB and are enriched in long-chain PUFAs, suggesting specific transfer of these entities across the placenta to the fetus to provide LC-PUFAs to support brain development. While several studies have demonstrated the apparent transfer of individual PUFAs into the fetal circulation during the later stages of pregnancy, our study is the first to document the transfer of a wide variety of lipid species containing LC-PUFA including CE, LysoPL, TG, and SM as well as free LC-PUFA. The transport of PUFA from maternal circulation to the fetus is dependent upon fetal requirements of PUFAs, which increase during late gestation. PUFAs such as arachidonic acid (20:4n-6) (AA) and docosahexaenoic acid (22:6n-3) (DHA) are essential for establishing cell membranes in the brain and retina. Within the maternal liver, these fatty acids are derived from the essential fatty acids (EFA) linoleic acid and alpha-linolenic acid via elongation by liver Δ5 and Δ6 desaturases. Fetal supply of very LC-PUFAs is dependent on placental transfer because placenta lacks the enzymes Δ5 and Δ6 desaturases and the fetus only has limited desaturase activity. Correlations between the maternal and CB lipidome can indicate the role of transport proteins. Our results suggest that starting early in gestation, M1 PUFA levels within PC, PL-PC, DG, and TG are positively associated with CB TG-poly (Fig. 7). Therefore, maternal PUFA levels early in gestation are crucial to establish infant PUFA levels. Since there is no evidence of direct placental transport of TG from maternal to fetal circulation, these results suggest a diversion of PUFAs from the maternal TG pool with re-esterification in CB TGs. This diversion could occur during hepatic TG elongation and desaturation of EFAs, with increased esterification of PUFA or by TG lipolysis in placenta and selective uptake of PUFA at the maternal–fetal interface. Maternal fatty acids are transported to the fetus across the placenta via a variety of fatty acid transport proteins as well as passive diffusion. Using C-labeled fatty acids administered to women 12-h prior to scheduled Cesarean delivery, Gil-Sánchez et al. demonstrated that the esterification of fatty acids into different lipid fractions influences placental transfer to the fetus. In maternal plasma at the time of delivery, C-fatty acid enrichment was detected with the incorporation of saturated fatty acid C-palmitate and monounsaturated fatty acid C-oleate in TGs while the PUFAs C-linoleic and C-DHA were found in both TGs and phospholipids. Within the placenta, 90% of labeled FA were found in the phospholipid fraction with significant enrichment of C-DHA in CB compared to both maternal and placental samples. These results suggest that the incorporation of essential PUFA into maternal phospholipids may be important for placental transfer at the end of gestation. Our results found M3 polyunsaturated FA within PCs and PEs, rather than DGs and TGs, are positively associated with CB TG-poly (Fig. 7), supporting the importance of the phospholipid fraction in the transfer of PUFA late in gestation. Our results indicated a sexual dimorphic effect of CB DG- and TG-poly with BW, with a larger effect seen in the females (Figs. 4, 5). Previous studies that have looked at the effect of PUFA supplementation during pregnancy on infant BW have yielded mixed results, potentially due to differences in gestation length, baseline characteristics of the participants, and source, timing, and dose of n-3 PUFA supplement. PUFA supplementation has been associated with longer gestation and an increase in fetal weight, especially in women at risk for intrauterine growth retardation. However, O’Tierney-Ginn, et al. found that total CB saturated fatty acids was positively associated with high skinfold thickness at birth (but not birthweight BMI z score), as well as increased BMI-z trajectory in early infancy, while PUFA levels were inversely correlated with BMI-z trajectory. Hauner, et al. demonstrated that dietary n-3 LC-PUFA supplementation of women had no effect on BW in a randomized trial of 208 women but had a larger effect on the gene expression profile in female placentas. Sedlmeier et al. observed placental gene expression to be more responsive to maternal omega-3 supplementation in females than males. The reason for the sexually dimorphic differences is not yet clear. We observed a positive association between CB LysoPE and LysoPCs and Fenton BW z score (Fig. 4), independent of the chain length and the number of double bonds. This confirms findings from other studies that CB LysoPCs with varying chain length and saturation are positively associated with newborn BW, as well as with weight at age 6 months. The associations between LysoPL and BW were not as apparent in M1 and M3, with an exception being M3 LysoPE-sat being positively associated with BW (Fig. 4). Additionally, we found that maternal GWG is positively associated with M1 and M3 LysoPC and LysoPE lipid groups of varying number of double bonds (Supplementary Table S4). Recently, the Major Facilitator Superfamily Domain Containing 2a (MFSD2a) protein has been proposed as a NA-dependent LysoPL transporter in the brain and placenta. This transporter provides a mechanism for the active transport of LysoPL from maternal plasma, consistent with our study findings that (1) LysoPC and LysoPE are depleted in maternal plasma and elevated in CB (Fig. 2) and (2) M1 and M3 LysoPLs, mainly containing saturated FA, are positively correlated with total CB LysoPL levels (Fig. 6). Since the sn-2 position of phospholipids tends to be occupied by PUFAs, it has been hypothesized that specific placental endothelial lipases cleave sn-1 fatty acids, allowing transport of the PUFA-containing LysoPL to the fetal circulation, perhaps enhancing delivery to the developing fetal brain. An intriguing possibility suggested by our data is that there is an interaction between PUFA during development and later transfer of LysoPL to the fetus. We see a significant inverse association of PUFA-containing lipids M1 and LysoPC levels in the CB of infants and conversely, a positive association of M1 (and M3) lipids enriched in saturated fatty acids (Fig. 6). This might imply early programming of LysoPL transport capacity. Control of MFSD2a expression in placenta has not been established, but is potentially regulated by diet. In mice, cortical and subcortical brain MFSD2a levels increased in rodents fed a high fat diet, with PUFA-containing diets having a greater effect at lower concentrations than a lard-containing diet. Additionally, MFSD2a expression in placenta is positively correlated with CB DHA levels, supporting its role in PUFA transport and implying it responds to the levels of DHA. Of interest, polyunsaturated SM are decreased between M3 and CB (Fig. 2), suggesting that maternal late gestation SM may be (1) directly transferred to the fetus to provide PUFAs or (2) converted into LysoPL, the preferential transfer method of PUFAs. Multiple significant correlations were observed between M3 SM and the CB lipidome. CB polyunsaturated TGs are inversely correlated with M3 SM-mono and SM-poly individual lipids and lipid groups (Fig. 7), suggesting that elevated late gestation polyunsaturated SM may result in lower levels of polyunsaturated containing TG in the CB. Supporting these results, we found that M3 SM-mono and M3 SM-poly are positively associated with maternal baseline BMI (Supplementary Figure S4) and positively associated with newborn BW (Fig. 4). These results suggest that inability to convert SM to transfer the monounsaturated and polyunsaturated fatty acids across the placenta results in less favorable fetal development, potentially linked to maternal BMI. Lastly, M3 saturated and monounsaturated SM were positively associated with CB LysoPL-total levels. These results highlight the importance of the SM lipid fraction during gestation, which potentially is related to the pathogenesis of obesity, T2D, and various metabolic diseases. The novel finding that the lipids that show the greatest differences between maternal and fetal circulation at term are also associated with BW suggests a complex interaction between PUFA intake, incorporation into specific lipid subtypes, and transport to the developing fetus. Future longitudinal studies with a larger sample size are desired to assess if changing the maternal metabolic environment to a more favorable lipidome will improve infant outcomes. It would be of particular interest to further explore the sex-specific influence of the maternal lipidome on the developing fetus. Collection of maternal dietary intake can help elucidate the relationship between nutrient intake and blood concentrations of very long-chain FA, as the metabolome does not solely represent dietary intake, but rather a combination of genetics, diet, and phenotypes. Maternal blood collection at delivery may not be representative of the metabolic states during the third trimester due to the stress of parturition, although we observed no differences in lipidome components based on route of delivery. While our results are consistent with previous studies, a potential weakness of our work is the fact that the pragmatic recruitment of women during their first prenatal visit (M1) resulted in variable post-prandial times. In addition, it is obviously difficult to control the fed/fasted state in women being admitted into the hospital with spontaneous labor. In M1, the main associations of interest included lipid classes with polyunsaturated fatty acid tails (Figs. 6, 7). The levels of PUFAs in peripheral blood reflect the chronic intake of PUFAs and many of the lipid fractions with PUFAs were correlated between M1 and M3 (Supplementary Fig. S2), suggesting a relatively constant intake of PUFAs across pregnancy. Finally, while small, the meal-induced changes that can alter the levels of lipid species in the blood along with the variable post-prandial time would be expected to weaken any association. Despite these confounding factors, we do see robust association of PUFA-containing TGs, PCs and PEs and CB lipid related to birth weight. Collecting multiple fasting blood samples throughout pregnancy will help to further classify lipidome changes throughout pregnancy. Collection of placenta samples and quantifying gene expression of key transport proteins (i.e. MFSD2a) would provide details on preferential transfer of lipid species. The results of this study indicate that starting in early gestation, the maternal metabolite environment may have an important effect on offspring weight and potentially, long-term cardiometabolic risk, consistent with developmental origins of health and disease. The subjects that participated in this study are part of the Michigan Mother-Infant Pair (MMIP) birth cohort study (2010–present). Briefly, pregnant women were recruited at their first prenatal appointment between 8 and 14 weeks of gestation. Eligibility criteria for MMIP include: age between 18 and 42 years old, had a spontaneously conceived singleton pregnancy, and intended to deliver at the University of Michigan Hospital. A subset of MMIP participants were chosen for lipidomics measures. Criteria for inclusion for the current study include the mother-infant pairs having complete demographic, survey and health information at their initial study visit and availability of all biospecimen at all-time points from mother and child. The MMIP study was approved by the University of Michigan Institutional Review Board (HUM00017941). The study was carried out in accordance with the approved guidelines and regulations. Written informed consent was received from all participants and all plasma samples and metadata were deidentified prior to analysis. During the initial study visit at 8–14 weeks of pregnancy, participants provided a blood sample (M1) and weight and height were collected by a clinician to calculate their baseline BMI (kg/m). Maternal weight at delivery was measured to calculate gestational weight gain (GWG) between the initial visit and delivery. Maternal venous blood samples (M3) and cord blood (CB) samples via venipuncture from the umbilical cord were collected. Women were not required to be fasted prior to blood sampling (M1 and M3). Physician-measured infant BW and head circumference were adjusted for gestation age and infant sex using Fenton growth curves for BW and growth curves developed by the Canadian Institute of Health Research for head circumference. Additional maternal information collected includes race/ethnicity, parity, marital status, and delivery route (vaginal and planned or unplanned Cesarean section). All 106 mother-infant pairs met additional inclusion criteria for this study including infant BW greater than 2,500 g, maternal BMI > 18.5, full term, and no pregnancy complications (i.e. gestational diabetes mellitus). Plasma samples were stored at − 80 °C prior to lipidomics analysis. Untargeted shotgun lipidomics was performed as previously detailed using reconstituted lipid extract from M1, M3, and CB plasma samples. Samples were ionized in positive and negative ionization model using a Triple TOF 5600 (AB Sciex, Concord Canada). Chromatographic peaks that represent features were detected using a modified version of existing commercial software (Agilent MassHunter Qualitative Analysis). Pooled human plasma samples and pooled experimental samples were randomized and run for quality control across batches. Data normalization followed a recently described drift removal method. Detected features were excluded with (1) a relative standard deviation greater than 40% in the pooled samples and/or (2) less than 70% presence in the samples. Missing data were imputed using the K-nearest neighbor method. Lipids were classified using LIPIDBLAST, a computer-generated tandem mass spectral library containing 119,200 compounds from 26 lipid classes. All lipids are annotated by the nomenclature X:Y, where X is the sum length of the fatty acid carbon chain and Y, the number of double bonds. After data processing, 573 biological lipids were identified, consisting of free fatty acids (FFA), cholesteryl esters (CE), acylcarnitines (AC), ceramides (CER), diacylglycerols (DG), lysophosphatidylcholine (LysoPC), lysophosphatidylethanolamine (LysoPE), phosphatidic acid (PA), phosphatidylcholine (PC), phosphatidylethanolamine (PE), plasmenyl-phosphatidylcholine (PL-PC), plasmenyl-phosphatidylethanolamine (PL-PE), phosphatidylglycerol (PG), phosphatidylinositol (PI), phosphatidylserine (PS), sphingomyelins (SM), and triglycerides (TG). Lipidomics methodology depicted in Fig. 1b. Prior to the main analysis, we examined differences in maternal and newborn characteristics stratified by sex of the newborn. Pearson chi-square tests were used to identify sex differences in categorical variables. Unpaired t-tests were used to identify sex differences in continuous variables. Scatterplots depicted the linear relationship between maternal baseline BMI, gestational weight gain, and Fenton BW percentile, highlighting the slope of the best-fit line (beta coefficient) and the tightness (r) of the trend. Step 1: To observe changes in the lipid profile across pregnancy, raw peak intensities of each lipid were standardized (mean = 0, standard deviation = 1) across M1, M3, and CB. Paired t-tests were used to identify differences in raw peak intensities of individual lipids between M1 and M3, to represent change in maternal metabolome during pregnancy, and M3–CB, to represent transfer of lipids through the placenta. We calculated FC for M1–M3 (log2[M3/M1]) and for M3–CB (log2[CB/M3]) to quantify change between time points. Differential lipids were classified using a Bonferroni p-value adjustment accounting for the total number of lipids (α = 0.05/573). Step 2: Due to biological constraints on metabolism, many metabolites are highly correlated. In response, we used prior knowledge to create lipid groups. Lipids were grouped by their class and number of double bonds, yielding 41 groups (Supplementary Table S2). For each lipid group, scores were created using unsupervised principal component analysis. The first principal component was retained, accounting for the largest portion of variance using PROC FACTOR in SAS. Within the principal component, each lipid receives a factor loading, which is the correlation coefficient between the lipid and the component. Scores were created by (1) multiplying the lipid factor loading by the lipid standardized peak intensity and (2) adding together these values for all lipids within a lipid group, using PROC SCORE in SAS. Mother-infant dyads received lipid group scores for M1, M3, and CB, separately. Using Debiased Sparse Partial Correlations, the correlations between lipid groups within and between time points were measured. Significantly correlated lipid groups have a false discovery rate (FDR) adjusted p value < 0.1. Step 3: Linear regression was used to identify the relationship between maternal baseline BMI and GWG with M1, M3, and CB lipid groups, adjusting for sex, parity, maternal age, and gestational age. Linear regression was used to identify M1, M3, and CB lipid groups associated with Fenton BW z score, adjusting for sex, parity, maternal age, gestational age, maternal baseline BMI, and GWG. Sex-stratified models were run. Associations were labeled using a FDR adjusted p value < 0.1 and, for exploratory analyses, an unadjusted p value < 0.05. Furthermore, linear models were used to explore sex differences between individual CB triglycerides (mean 0, standard deviation 1) and Fenton BW z score, stratified by infant sex and adjusted for parity, maternal age, gestational age, maternal baseline BMI, and GWG. Non-parametric regression curves, classified using varying coefficient models, were plotted to classify how the relationship between TGs and BW is related to the number of double bonds in TGs (R package ‘np’). Step 4: Lastly, we sought to identify maternal lipids related to infant CB lipid groups associated with Fenton BW. CB lysophospholipids and triglycerides were correlated with M1 and M3 individual lipids and lipid groups to determine if maternal lipids modulate the levels with CB, using Pearson's correlations (unadjusted p value < 0.05). Statistical methodology depicted in Fig. 1c. Unless otherwise stated, all statistical analyses were performed using SAS 9.4 (Cary, North Carolina). Figures were created using GraphPad Prism version 7.4 (La Jolla, California). Heatmaps were created using an in-house software package, CoolMap.
|
PMC5339821
|
Large-scale human skin lipidomics by quantitative, high-throughput shotgun mass spectrometry
|
The lipid composition of human skin is essential for its function; however the simultaneous quantification of a wide range of stratum corneum (SC) and sebaceous lipids is not trivial. We developed and validated a quantitative high-throughput shotgun mass spectrometry-based platform for lipid analysis of tape-stripped SC skin samples. It features coverage of 16 lipid classes; total quantification to the level of individual lipid molecules; high reproducibility and high-throughput capabilities. With this method we conducted a large lipidomic survey of 268 human SC samples, where we investigated the relationship between sampling depth and lipid composition, lipidome variability in samples from 14 different sampling sites on the human body and finally, we assessed the impact of age and sex on lipidome variability in 104 healthy subjects. We found sebaceous lipids to constitute an abundant component of the SC lipidome as they diffuse into the topmost SC layers forming a gradient. Lipidomic variability with respect to sampling depth, site and subject is considerable, and mainly accredited to sebaceous lipids, while stratum corneum lipids vary less. This stresses the importance of sampling design and the role of sebaceous lipids in skin studies.The primary function of the largest organ of the human body - skin - is to keep in what is inside and keep out what is outside. This functionality, its physiology and pathophysiology are derivative of structure and composition of its topmost layer, the stratum corneum. The SC consists of a layered meshwork of corneocytes sealed with lipids arranged in lamellar fashion. The most abundant constituents of the lipid matrix are ceramides (Cer), cholesterol (Chol) and free fatty acids1. Other classes found in SC include triacylglycerol (TAG), diacylglycerol (DAG) and cholesterol esters (CE)2. Lipidomic studies of skin have demonstrated that age, gender, ethnicity, and season of the year affect the skin lipid composition3. Likewise, alterations of lipid profiles were linked to dermatological and systemic diseases, like atopic dermatitis45, hereditary ichthyosis6 or Netherton syndrome7. Early pioneering thin layer chromatography studies discerned 8 ceramide sub-classes8. The advent of mass spectrometry coupled with chromatographic separation of analytes (i.e. LC-MS) offered much increased sensitivity. This allowed further differentiation of ceramides into 12 ceramide sub-classes and identification of other lipid classes9101112. These 12 ceramide sub-classes are defined by a combination of four different types of sphingoid bases (dehydrosphingosine, sphingosine, phytosphingosine and 6-hydroxy sphingosine), referred to as long chain base (LCB) with three different types of acyl chains (non-hydroxy, alpha-hydroxy and esterified omega-hydroxy fatty acid (FA))1. A more recent approach to skin lipidomic analysis utilizes shotgun mass spectrometry, where the sample extract is subjected to mass spectra acquisition without prior chromatographic separation (direct infusion)1314. This technology was applied for the analysis of non-hydroxy and alpha-hydroxy ceramides, but it did not allow for absolute quantification or for the distinction between all ceramide sub-classes. In comparison to shotgun, mass spectrometry involving chromatographic separation of analytes is characterized by increased sensitivity due to the lowered spectra complexity, which together with additional information about retention times makes them suitable for the discovery of new, even low abundant lipids10. The lack of chromatographic separation prior to analysis in shotgun mass spectrometry requires high-resolution instruments, as spectra are very complex, which also necessities refined data processing15. On the other hand, it makes this technique more suitable to cover a higher number of lipid classes, which would otherwise necessitate different separation conditions and longer analysis time in chromatographic approaches. However, the central advantage of shotgun mass spectrometry resides in its high-throughput capabilities. We previously reported the development of a shotgun lipidomics platform for the analysis of blood plasma permitting the comprehensive quantification of lipids in hundreds of samples16. The capacity to analyze a high number of samples allows screening, biomarker, intervention, and mode-of-action studies, facilitating a much-needed increase in the statistical power of lipidomic results. High throughput is also a prerequisite for any prospective clinical setting of skin lipidomics as well as for derma-pharmacological and cosmetic applications. In order to exploit fully the high-throughput potential of shotgun lipidomics for skin samples, the sampling procedure itself has to fulfil certain requirements. Firstly, it needs to deliver reproducibly a sufficient amount of material to allow comprehensive lipid coverage. Secondly, the variation of lipid composition with respect to sampling site as well as sampling depth should be controlled, especially if insight into physiological processes is sought. Furthermore, the sampling procedure should neither introduce exogenous substances into the sample, nor remove (or suppress) analytes. Several skin-sampling techniques are employed so far, most fulfilling only some of the above criteria at a time. Scrape biopsy for instance delivers abundant amounts of sample with minimal contaminants, but neither the exact sample amount, nor sampling depth can be easily controlled. In situ extraction (the so-called “cup method”17) offers clean and direct extraction, but does not have high-throughput capabilities and is potentially taxing for the subject. Tape-stripping offers control of sampling position, depth and sample amount. However, the extraction of polymeric adhesive tape with organic solvents can potentially introduce background. In this paper we present a shotgun mass spectrometry-based lipidomic method compatible with the most convenient sampling method – tape stripping – and characterized by broad lipid coverage, including unambiguous identification of lipid species belonging to all 12 ceramide classes, as well as di- and triacylglycerols, cholesterol and its esters. All lipids are absolutely quantified by inclusion of internal standards. In combination with automated lipid extraction and acquisition, this method achieves unprecedented high-throughput. With this method, within days we performed to our knowledge the largest (104 subjects; more than 268 individual samples analyzed in total) stratum corneum lipidomic study to date. We investigated (1) the dependence between SC lipid profiles and sampling depth; (2) determined intra-individual variability of the SC lipidome by analyzing 14 different sampling sites; and (3) assessed inter-individual variability by analyzing SC lipidomes of 39 males and 65 females of different age. Lipid species are annotated according to their molecular composition as follows: [lipid class]-[sum of carbon atoms in LCB and FAs]:[sum of double bonds in LCB and FAs];[sum of hydroxyl groups in LCB and FA] (e.g., EOS 70:3;2 denotes omega-hydroxy-shpingosine with a total length of its LCB and FA of 70; with 3 double bonds and 2 hydroxylations in total). For lipid sub-species, the individual acyl chain composition according to the same rule is given (e.g. 18:1;0–24:2;0), with the first entity denoting a sphingoid base (LCB) and the second a fatty acid (FA), in case of ceramides. Ceramides naming convention was adopted from18. Methanol, propan-2-ol, chloroform, acetyl chloride and ammonium acetate were of analytical grade. Deuterated NS D3 (36:1;2) (cat# 2201) and EOS D9 (68:3;2) (cat# CUS9530) were purchased from Matreya LLC. Deuterated TAG D5 (cat# 110544) and DAG D5 (cat# 110538), and CE(20:0;0) (cat# 110870) were purchased from Avanti Polar Lipids. All sampling from human subjects occurred with informed consent in accordance with the Declaration of Helsinki and was approved by the Bioethical Committee of the Wroclaw Medical University (decision number 406/2015). Scrape biopsy samples were collected from the calves of one subject by careful scraping of the topmost skin layers with a scalpel directly into a tube. The sampling site for tape-stripping was prepared by removing the topmost SC layer with a large CUDERM D-Squame sampling disc (D100). A discs was pressed on the sampling site for 15 s with a D-Squame pressure instrument (D500) providing uniform pressure (225 g/cm), and removed with one fluent motion. Next, in order to collect the actual sample, a small D-Squame stripping disc (D101) was pressed with the pressure instrument to the prepared site again for 15 s, removed and placed into a tube. Samples were stored at −20 °C until extraction. The sample amount collected with stripping discs was determined gravimetrically using a Radwag Microbalance Type MYA 5.3Y. To this end, 42 volar forearm samples of 6 subjects were collected according to above protocol, while stripping discs were weighed directly before and after sampling, the difference being equal to the weight of skin collected. For the investigation of lipidome variability with respect to sampling depth, triplicate samples from adjacent positions on left volar forearm from one female and one male subject were collected layer by layer – including the topmost – until the 20. After the 10 layer every second layer was sampled, and the remaining layers were discarded. Thus, 90 samples were collected in total. All these sequentially tape-stripped samples belonged to the stratum corneum proper consisting of corneocytes embedded in a lipid matrix19. For the assessment of intra-individual skin lipidome variation triplicate samples from adjacent positions on skin were collected from each sampling site. Samples were collected from one male and one female at 14 different sites (forehead, cheek, pectoral region, abdomen, groin, thigh, heel sole, central calf, shoulder blade, buttock, palm, outside of hand, top side of foot and volar forearm). Laterally symmetric sites were all sampled from the left side of the body. For the inter-individual lipidome variation 65 females and 39 males of different age (20–89) were sampled from the volar forearm, totalling 104 samples. Each analytical batch of samples was complemented by blank samples each containing a sampling disc without skin. Lipid extraction of tape-stripping samples containing one stripping disc each and scrape biopsy samples alike was carried out in 2 mL polypropylene tubes where 900 μL of methanol including internal standards was added to each sample. The samples were shaken at 1400 rpm at 4 °C for one hour. Thereafter, extracts were transferred to a multi-well plate and dried in a speed vacuum concentrator. Dried extracts were re-suspended in an acquisition mixture of 7.5 mM ammonium acetate in chloroform:methanol:propan-2-ol (1:2:4, V:V:V). For cholesterol, the dried extract was acetylated20, then dried again and re-suspended in the above acquisition mixture. All liquid handling steps were performed using Hamilton Robotics STARlet robotic platform with Anti Droplet Control for pipetting of organic solvents. Samples were analyzed by direct infusion with a QExactive mass spectrometer (Thermo Scientific) equipped with a TriVersa NanoMate ion source (Advion Biosciences) in a single acquisition for both positive and negative ion modes with a resolving power of Rm/z = 200 = 280000 for MS and Rm/z = 200 = 17500 for MSMS. MSMS fragmentation was performed at normalized collision energy of 35% and was triggered by an inclusion list encompassing corresponding MS mass ranges16. Both MS and MSMS data were combined to monitor CE, DAG and TAG ions as ammonium adducts and all ceramide sub-classes as acetate adducts. Precursor ions and confirmatory MSMS fragments as reported previously101314 are summarized in Supplementary Table 1. Cholesterol was identified in a separate acquisition as cholesterol-acetate after a derivatization procedure20. The resolving power used allows for identification of lipids based on precursor masses21. Combining MSMS fragmentation data with the high resolution MS precursor data followed by isotopic correction (type I and II according to the strategy described previously1622), permitted structural elucidation of lipid molecular species. It increased the identification specificity, but also made it possible to unambiguously distinguish all 12 ceramide sub-classes; including NP/AdS and NH/AS pairs. Internal standards used in this study were chosen not to be natively present in skin samples. Per sample 42 pmol of EOS D9 68:3;2 (18:1;2, 32:0;0, 18:2;0), 14 pmol NS D3 36:1;2 (18:1;2, 18:0;0), 50 pmol DAG D5 34:0;0 (17:0;0, 17:0;0), 100 pmol CE 20:0;0, 1000 pmol cholesterol D6 and 100 pmol TAG D5 51:0;0 (17:0;0, 17:1;0, 17:0;0) were delivered as internal standards. Quantification was conducted via normalization of the isotopically corrected intensity of the monoisotopic peak of each native species to the isotopically corrected intensity of the monoisotopic peak of the internal standard. The quantities of lipid molecular species were calculated from ratios between their respective characteristic MSMS fragments, as described previously16. Non-hydroxy and alpha-hydroxy ceramide sub-classes were normalized to deuterated NS, whereas omega-hydroxy ceramides were normalized to deuterated EOS. Quantitative lipidomic data of the study can be found in Supplementary Table 2. Data were analyzed with in-house developed lipid identification software based on LipidXplorer2324. Only lipid identifications with measured mass deviations below 3 ppm from the theoretical mass for MS and 8 ppm for MSMS peaks in scans where lock mass was available, and 5 and 12, respectively, where it was not; signal-to-noise ratio greater than 5; and an amount at least 5-fold higher than in corresponding blank samples were considered as positive hits, yielding 862 unique lipids across all samples from all experiments. Further, unless stated otherwise, only lipids present in at least two replicates and above 2 pmol per sample were included in subsequent data analysis, resulting in 509 unique lipids. Data visualization, linear regression (linear least squares method), and correlation (two-tailed Pearson and Spearman correlation) calculations were performed on Prism6.0 h software (GraphPad Software, Inc.). Where it does not lower figure clarity, individual data points are shown25. Statistical models were trained with the R Environment for Statistical Computing (R version 3.3.2 (2016-10-31)), by the caret package (version 6.0–73)26, which in turn uses the randomForest package (version 4.6–12)27. For statistical modeling only lipids present in at least 50% of samples per cohort were used. If a lipid was present in one cohort, it was also included for other cohorts, even if its occurrence rate was lower. During model training with 5 times repeated 10 times cross validation, the resampled data were preprocessed by centering and scaling, missing values were imputed with the median, and near zero variance predictors were removed. In order to take advantage of the quick spectra acquisition capabilities of shotgun mass spectrometry, the sampling procedure and all other sample-handling steps were optimized for speed, throughput and convenience utilizing automation whenever possible. Automation of procedures also benefits in higher reproducibility28. Tape-stripping was used as it was simultaneously the most convenient and non-invasive sampling method (Fig. 1a). It also had the additional advantage of allowing for control over the sampling depth (by collecting the appropriate stratum corneum layer by sequential stripping) and collecting comparable sample amounts. As surface furrows might affect the amount of material collected29, we gravimetrically determined the reproducibility of sample amount collected via tape-stripping of the second layer with one stripping disc to be 62 ± 14 μg (mean ± s.d.) (Fig. 1b). Extraction of lipids with commonly used organic solvents from a polymeric tape with adhesive (i.e. stripping disc) was complicated because established methods utilizing chloroform3031 or methyl tert-butyl ether32 interfered with tape constituents, physically dissolving them. We found, however, that methanol extraction successfully used for skin lipidomics in previous studies11 allowed for efficient extraction of skin lipids from stripping discs without compromising analysis quality, even without prior chromatographic separation of extracts. Next we investigated whether the sample amount delivered by one stripping disc is suitable for lipid analysis. We found that the lipid amounts quantified from 0.5, 1, 1.5 and 2 discs stripped from the SC from one subject were linearly proportional to the number of discs used (Fig. 1c). Some of the ceramide sub-classes (EOdS, AS or EOH) were not proportionally represented when only half a disc was used. This was due to low abundant species falling below the limit of detection, as also reflected by a lower number of lipids detected in total (222 in comparison to 250–253 in other sampling amounts). However, the sample amount collected by one stripping disc ensured that all lipid classes were measured proportionally and maximum coverage of lipid species was achieved. Therefore, one stripping disc per sample was used for all other experiments. In order to rule out a negative influence of stripping disc constituents on the analysis, we compared the results obtained from skin samples analyzed with and without a stripping disc present. To this end, we collected skin by scraping and divided it into individual samples of a weight similar to the amount collected with one stripping disc (77 ± 2.5 μg per sample, mean ± s.d.). To 5 samples one clean stripping disc was added and to the remaining 5 none. All ten were extracted and analyzed as described, and the lipid quantities were compared (Fig. 1d). Evidently, even in the presence of a stripping disc, lipid amounts were virtually identical to amounts in samples without disc (Pearson correlation coefficient = 0.99, P < 0.0001). This result shows that stripping discs do not influence lipid quantification. Due to the wide concentration range of analytes encountered in lipidomics analyses, it is essential that the analytical method spans a large dynamic range and offers sufficient sensitivity to allow for the quantitative analysis of molecules in the low and high concentration regime. To determine these parameters, we added increasing amounts of lipid class-specific reference lipids into tape-stripped skin samples, and we recorded how the acquired signal correlates with the amount (Fig. 1e). Lipid references used for this assessment needed to be distinguishable from endogenous lipids; therefore this analysis was limited to classes for which a proper reference lipid was available. Linear dynamic range was defined as the concentration range at which the linearity of signal-to-noise values to pmol amount and the slope of the resulting function were close to 1 (slope from 0.85 to 1.21, depending on the lipid class (Supplementary Table 1)). This implies that changes in concentration of a given lipid will result in directly proportional changes in signal. The limit of quantification of most lipid classes lies in the low pmol range. Only for cholesterol it is about 50 pmol. However, this is sufficient since cholesterol is typically in the range of hundreds of pmol per skin sample. The linear dynamic range spans 1.8 (for NS, EOS, DAG and cholesterol) and 2.5 (for TAG and CE) orders of magnitude. Thus, the method is capable of measuring at least a hundred-fold change reliably. Reproducibility of the method was assessed by analyzing lipids in samples collected from three different subjects and pooled after extraction. Pooled extracts were analyzed and quantified independently (n = 10), and the coefficient of variation (CV) of all quantified lipid sub-species was plotted against their respective amounts (Fig. 1f). Median CV was 7.37% with 86% of all lipid subspecies varying by less than 15%. As expected, there is an inverse correlation between CV and lipid amount1633. Obtained reproducibility parameters place majority of quantified lipids well below the 20% CV threshold that is commonly used for in vitro diagnostic assays34. Together with the method’s sensitivity and dynamic ranges reported above, shows that skin lipids can be reliably and reproducibly measured in tape-stripped samples, meeting the criteria of a good analytical method34. To check how skin lipidome changes with the depth, we collected samples from the left volar forearm from one female and one male through sequential layer by layer stripping – including the topmost – until the 20 layer and measured their lipid composition. In both subjects we observed the total lipid amount continually decreasing by 85 and 87% until the 5 and 7 layer in female and male respectively, reaching a plateau thereafter (Fig. 2a). This could be explained either by a higher mass of skin sampled from the topmost layers19, by a higher proportion of lipids to other constituents within these layers or by a combination of both. A previous study on three subjects tape-stripped sequentially until the 20 layer showed that sample weight collected by CUDERM tapes within the first layers is doubled in comparison to deeper layers35. However, the observed surplus of lipid amount in the uppermost layers compared to deeper layers is greater than would have resulted from the reported differences in sample weight. This suggests that the uppermost layers, while more abundant in mass, are enriched in lipids compared to deeper samples. The topical excretion of sebaceous lipids could account for this observation as sebum lipids were shown to surpass stratum corneum lipids on the skin surface over tenfold by weight3637. This raises the question, whether lipid classes representative of sebum penetrate into deeper stratum corneum layers. In both subjects, we observe a marked decline in the relative amount of TAGs and DAGs, jointly accounting for half of the known sebum lipids on average38, following the decrease in total lipid amount (Fig. 2c and d). This supports the notion of sebum lipid penetration into the SC down to the 5–7 layer, where TAGs and DAGs levels reach their plateaus. The relative amount of cholesterol and all ceramide sub-classes increased with sampling depth. Ceramides are known to be synthesized deep within SC39 and are not a sebum constituent, which fully accounts for the observed ceramide sampling depth gradient direction. Likewise, cholesterol constitutes as much as one third of the SC lipids, and only a small fraction of the sebum lipids38, which is reflected in our data in its lower relative amount at the skin surface. Cholesterol esters were previously reported in stratum corneum40 as well as in sebum38 and this is corroborated by our measurements, where the relative amount of cholesterol esters decreased in the topmost layers and increases again below the 5–7 stripping layer. Investigating the relationship between individual species of cholesterol esters and sampling depth we observed short-length fatty acid species being more abundant at or exclusive to the topmost layers (Fig. 2b). This indicates that different, preferentially shorter cholesterol ester species are of sebaceous origin, whereas others, mostly longer species stem from the stratum corneum proper. With the principal component analyses (PCA) using all lipids as input, we were able to separate samples from the first five (female) and six layers (male) from deeper samples. Separation along principal component 1 (PC 1) corresponds to sampling depth, but samples from deeper layers do not follow this trend (Fig. 2e and f). The previously observed sebum lipid enrichment in the upper layer samples led us to postulate that mostly sebum lipids are driving the separation with respect to sampling depth. To test this hypothesis we again performed the PCA but including only ceramides as input (Fig. 2g and h). Here, the distinction of the samples by layer was limited to the first two (female) or three (male) tape-strippings. Therefore, the above data permits the conclusion that in both subjects skin lipid composition at the volar forearm varies discernably with sampling depth. The discriminating lipids originate from the skin surface and are of sebaceous origin. The ceramide profile is less variable with respect to sampling depth. To assess intra-individual lipidome variability, samples were collected in triplicate from 14 body sites from one male and one female. Mean triplicate total lipid amount variation was 15 and 12%, for male and female respectively (Fig. 3a), which is below the variation of weight of tape-stripping as determined before (23%, Fig. 1b), indicating a general robustness of triplicate sampling across different sampling sites. Across all sampling sites we observed an intra-individual variability of total lipid amount of 104 and 81% for male and female respectively. The most distinct samples for both subjects were from forehead where the total lipid amount was as high as 21801 pmol and from heel where it was as low as 321 pmol on average. Such differences in a lipid content of skin from various body parts were observed previously3741. Notably, the site-specific total lipid amounts accessible to tape-stripping exhibit a significant correlation between the two subjects (Spearman correlation coefficient = 0.74, P = 0.0027). Since the likely impact of site-specific sample amounts on lipid content cannot be excluded36, relative lipid amounts were used to conduct PCA (Fig. 3b and c). This analysis illustrated that in both subjects the lipid profiles at any particular sampling site (i.e. within a triplicate) are more similar to one another than to lipid profiles at other sampling sites. It also indicates that skin on the forehead, cheek and to some degree from the pectoral region and shoulder blade is distinct from skin sampled from any of the other ten body sites. To confirm this distinction a hierarchical clustering was performed on averaged replicate values from all sampling sites. It shows that for both subjects cheek, forehead, pectoral region and shoulder blade lipidomes form the most closely related cluster (Fig. 3d and f). As face and back were shown to contain a high amount of triglycerides36 and the face features the highest density in sebum glands42, this led us to propose, that the facial samples, together with samples along the neckline (pectoral region and shoulder blade) are distinct from all other body sites primarily because of sebaceous lipids. Indeed, a cluster analysis of the samples considering only sebum lipids (TAGs, DAGs) confirms the same tight association of the facial samples and to some degree also the neckline samples (Fig. 3e and g). This finding holds true when investigating the similarity of both subjects jointly. Facial and neckline samples from male and female (except from the male shoulder blade) all fall into one super-cluster (Fig. 3h). Again, this similarity can be attributed fully to sebaceous lipids (Fig. 3i). In summary, we showed the applicability of our method to skin of the entire human body. We observed site-specific differences in total lipid amounts as well as distinct lipid profiles. Skin at the investigated body sites differs predominantly by the content of sebaceous lipids, which are the major factor responsible for sample separation and clustering. To assess inter-individual variability 65 females and 39 males of different age were sampled at the volar forearm. In both sexes, the total lipid amount varied considerably (male: 57%, female: 90%) (Fig. 4a and b), as observed in other studies43. In our study, the total lipid amount in males was not correlated with age, but females, exhibited a negative correlation between age and total lipid amount (Pearson correlation coefficient = −0.40, P = 0.0009). The relative amounts of TAG and DAG in female skin markedly decreased with age, while the relative amount of cholesterol rose. This circumstance was less pronounced in the male population (Fig. 4c and d). A previous investigation of human epidermis also discovered a decline of triglycerides in older subjects, as compared with younger ones2, which was further corroborated by the study on 110 women where it was shown that the sebum amount decreases with age44. We observed relative cholesterol ester and ceramide levels remaining constant. This indifference of ceramide content to age in skin was observed before45. A possible explanation for gender-specific, age-dependent depletion of sebaceous lipids might be connected to hormonal regulation of sebum secretion46. It was shown, that at increasing age eccrine glands are either reduced in number or in functional capacity47. This could account for the pronounced decline of both absolute and relative amounts of sebaceous lipids among older women. This raised the question, whether the volar forearm lipidomes of males and females are unique enough to allow sex-specific differentiation of samples. The principal component analysis showed no differences between the male and female cohorts (Fig. 4e). This finding is in line with a recent study, which did not find distinguishing features between skin ceramides of adult men and women48. Also, another statistical algorithm (random forest) trained for sex differences and fitted on all samples with all lipids included, resulted in an accuracy of sex prediction based on skin lipidomes of 0.7 ± 0.073 (mean ± s.d.), while the same fitting without sebum lipids achieved only 0.58 ± 0.099 (Fig. 4f). Therefore, neither approach could generate a reliable sex prediction based on the lipidome, as a Null Error Rate is was 0.65 (where 65% of all samples are female). In summary, our study revealed large inter-individual variability of skin lipidomes in samples collected from the volar forearm, comparable in magnitude to the intra-individual variability as measured at 14 sites on one male and one female. Both absolute and relative amounts of sebaceous lipids were observed to decrease with increasing age, especially among females. In this study, we developed a method for analysis skin lipids, which combines advantages of shotgun lipidomics with benefits of tape-stripping, to produce quantitative skin lipidomes of unprecedented coverage. For the first time, stratum corneum lipids and lipids representing the majority of the sebum were analyzed simultaneously down to the level of sub-species. The linear dynamic range of the method proved sufficient to accommodate for the variability of all analyzed samples. To illustrate the capabilities of the method, we conducted a large-scale survey addressing three questions: how do skin lipid profiles change with respect to sampling depth; how variable are lipidomes of skin collected at various sampling sites of a body; and how do lipid profiles vary in volar forearm samples depending on age and gender of different subjects. Sampling via consecutive stripping gave insight into the depth profile of sebaceous lipids down to the 20 tape-stripping layer of the volar forearm skin. We observed pronounced depth dependent gradients in total lipid amounts and lipid composition. Most surprisingly, up to 5–7 topmost tape-stripping layers, lipids of sebaceous origin constitute a larger part of the total lipidome than ceramides and cholesterol. In turn, in deeper samples cholesterol increases to become the most abundant lipid. Re-evaluation of the relation between lipid content and corresponding morphological features and functional parameters (transepidermal water loss, hydration), including different sampling sites, should prove insightful for establishing a role of individual lipids in skin functions. The intra-individual variability between the 14 sites analyzed on one male and one female subject was large, both in terms of absolute lipid amounts and lipid composition. The variability within triplicate samples on the other hand was comparatively small, illustrating both method reproducibility and a local biological lipidome stability sufficient to establish lipid profiles characteristic for the surveyed sites. In both subjects, we observed that facial and neckline samples sites are more similar to one another than samples from other body regions. Similarities and differences between sampling sites were accredited to sebaceous rather than to stratum corneum lipids. The assessment of inter-individual lipidome variability in volar forearm samples collected from 104 subjects revealed a negative correlation between total lipid amount and age for females. With increasing age the fraction of sebaceous lipids decreased, while the cholesterol fraction rose and this phenomenon was more pronounced among females then among males. However male and female samples could not be distinguished by their lipidomes. This study raises important issues for future comparative skin lipidomic study design. Firstly, our study demonstrates the importance of the choice of sampling site and sampling depth. Depending on the objective of the study, these parameters along with age of the subjects must be controlled because their influence on measured lipidomes is dominant. Secondly, lipidome variability precludes the establishment of a general skin lipidome baseline, and comparisons should rather be made between adjacent sampling sites of the same subject. Therefore, in intervention studies topical drugs or agents should be applied to one skin site, while preferentially an adjacent, untreated site should be chosen as a negative control. These sites would then be tape-stripped and their lipidomes analyzed. Finally, the broad lipid coverage, absolute quantification and high-throughput makes shotgun mass spectrometry based skin lipidomics a well-suited tool for rigorous and systematic studies of various topics, such as: influence of drugs on the skin lipidome; the action of substances influencing skin lipid metabolism or the skin microbiome-lipidome relation; impact of cosmetic substances on a skin lipidome with respect to their efficacy claims; and many others. How to cite this article: Sadowski, T. et al. Large-scale human skin lipidomics by quantitative, high-throughput shotgun mass spectrometry. Sci. Rep. 7, 43761; doi: 10.1038/srep43761 (2017). Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
|
PMC10469212
|
The elucidation of plasma lipidome profiles during severe influenza in a mouse model
|
Although influenza virus infection has been shown to affect lipid metabolism, details remain unknown. Therefore, we elucidated the kinetic lipid profiles of mice infected with different doses of influenza virus A/Puerto Rico/8/34 (H1N1) (PR8) by measuring multiple lipid molecular species using untargeted lipidomic analysis. C57BL/6 male mice were intranasally infected with PR8 virus at 50 or 500 plaque-forming units to cause sublethal or lethal influenza, respectively. Plasma and tissue samples were collected at 1, 3, and 6 days post-infection (dpi), and comprehensive lipidomic analysis was performed using high-performance liquid chromatography–linear trap quadrupole–Orbitrap mass spectrometry, as well as gene expression analyses. The most prominent feature of the lipid profile in lethally infected mice was the elevated plasma concentrations of phosphatidylethanolamines (PEs) containing polyunsaturated fatty acid (PUFA) at 3 dpi. Furthermore, the facilitation of PUFA-containing phospholipid production in the lungs, but not in the liver, was suggested by gene expression and lipidomic analysis of tissue samples. Given the increased plasma or serum levels of PUFA-containing PEs in patients with other viral infections, especially in severe cases, the elevation of these phospholipids in circulation could be a biomarker of infection and the severity of infectious diseases.Accumulating evidence has revealed the role of lipids in the progression and suppression of diverse diseases. In particular, polyunsaturated fatty acids (PUFAs) have received considerable attention for their pro- and anti-inflammatory actions. Among PUFAs, arachidonic acid is one of the most important lipids because it is a precursor of downstream molecules, such as prostaglandins, leukotrienes, and thromboxanes. These eicosanoids significantly regulate inflammatory responses by causing vasocontraction/vasodilation, leukocyte migration, platelet activation, etc. Arachidonic acid is hydrolyzed from intracellular phospholipids by phospholipases including cytoplasmic phospholipase 2 (cPLA2). Notably, cPLA2 activation by infection with viruses, such as human immunodeficiency virus, human cytomegalovirus, respiratory syncytial virus, and influenza virus, can induce increased production of free arachidonic acid. Since cPLA2 deficiency prevents pneumonia caused by severe acute respiratory syndrome coronavirus (SARS-CoV) infection and experimental autoimmune encephalomyelitis in mouse models, the arachidonic acid cascade may be critically involved in the pathogenesis of various inflammatory diseases. Unlike arachidonic acid, docosahexaenoic acid (DHA) and eicosapentaenoic acid (EPA) have been studied for their anti-inflammatory effects; both DHA and EPA decrease the production of inflammatory cytokines and prostaglandins and inhibit the endotoxin-induced inflammatory cascade. Because free arachidonic acid is further converted into prostaglandins and other molecules that induce various inflammatory responses, stimulation of arachidonic acid production in infectious diseases has been extensively studied, as described above. In addition, early studies have already demonstrated that viral infections can affect lipid composition, such as the ratios of phospholipids and cholesterol esters to total lipids, in the blood and tissues. Furthermore, our recent study revealed modulation of the signaling pathway of sphingosine-1-phosphate by influenza virus infection. These findings demonstrate that lipid metabolism is one of the target functions significantly affected by viral infections in the host. However, information on the comprehensive and kinetic profiles of lipid molecules during infection is limited. To better understand host responses to viral infections, we applied an untargeted lipid measurement strategy to an established severe influenza mouse model. In the analysis, high-performance liquid chromatography (HPLC)–linear trap quadrupole (LTQ)–Orbitrap mass spectrometry (MS), which enables the simultaneous measurement of multiple lipids from different classes in a small amount of plasma samples, was used. The results demonstrated a unique lipid profile in biological samples from the lethally infected mice. Liquid chromatography/mass spectrometry (LC–MS) grade methanol, isopropanol, chloroform, and 1 M ammonium acetate solution were obtained from Wako Pure Chemical Industries, Ltd. (Osaka, Japan). The EquiSPLASH Lipidomix quantitative standard (100 µg/mL) and oleic acid-d9 for MS was purchased from Avanti Polar Lipids (Alabaster, AL, USA). The influenza virus A/Puerto Rico/8/34 (H1N1) (PR8) was kindly provided by the National Institute of Infectious Diseases (Tokyo, Japan). The virus was propagated in 10-day-old embryonated chicken eggs at 35 °C for 48 h, and aliquots of the collected allantoic fluids were stored at − 80 °C until further analysis. Male C57BL/6 mice purchased from Hokudo (Sapporo, Japan) were kept in a BSL-2 laboratory at the International Institute for Zoonosis Control, Hokkaido University, under standard laboratory conditions (room temperature, 22° ± 2 °C; relative humidity, 50 ± 10%) and a 12/12 h light/dark cycle. The mice were administered a standard CE-2 chow diet purchased from CLEA Japan (Sapporo, Japan) with water ad libitum. Experiments were performed on 9–12-week-old male mice. Virus infection and sample collection were performed as previously reported. PR8 virus at 50 or 500 plaque-forming units (PFU) in 50 µL phosphate-buffered saline (PBS) or PBS only (control) were intranasally inoculated into mice under inhalation anesthesia with isoflurane. Body weight was monitored daily. At 1, 3, or 6 days post-infection (dpi), the mice were euthanized by an overdose of isoflurane followed by cervical dislocation, and their blood, lungs, and liver were collected. Blood samples were centrifuged at 2000×g for 10 min at room temperature in the presence of heparin sodium (10 U/mL), and supernatants were collected as plasma and stored at − 80 °C until further analysis. Tissues for lipidome analyses were washed in cold PBS to remove excess blood and stored at − 80 °C until sample preparation. Tissue samples for gene expression analyses were homogenized in TRIzol reagent (Thermo Fisher Scientific, Waltham, MA, USA) and stored at − 80 °C until further analyses. Lung samples for virus titer measurement were homogenized in 1 mL RPMI-anti medium [RPMI-1640 (Thermo Fisher Scientific) with 100 U/mL penicillin (Sigma-Aldrich), 100 µg/mL streptomycin (Sigma-Aldrich), and 20 µg/mL gentamicin (Thermo Fisher Scientific)]. After centrifugation at 1750×g for 10 min, the supernatants were collected and stored at − 80 °C until further analysis. This study followed the Animal Research: Reporting of In Vivo Experiments guidelines, with the exception of blinding. Because of the requirement to clearly indicate viral infection and treatment with any reagents on the cage cards, the investigators could not be blinded. To measure lung virus titers, a plaque assay was performed using lung homogenates, as previously reported. The plasma levels of interferon-γ (IFN-γ), interleukin-6 (IL-6), IFN-γ-induced protein-10 (IP-10), monocyte chemoattractant protein-1 (MCP-1), macrophage inflammatory protein-1β (MIP-1β), and tumor necrosis factor-α (TNF-α) were determined using a MAGPIX Milliplex kit (Merck, Darmstadt, Germany) following the manufacturer’s instructions, as reported previously. Total RNA was extracted from the tissue samples using TRIzol (Thermo Fisher Scientific, Waltham, MA, USA), and cDNA synthesis was performed using High-Capacity cDNA Reverse Transcription Kits (Thermo Fisher Scientific) following the manufacturer’s instructions. Gene expression of ethanolamine kinase 1 (Etnk1; Mm07299373_m1), (Ept1; Mm01210813_m1), phospholipase A2 group IVA (Pla2g4a; Mm00447040_m1), acyl-CoA synthetase long-chain family member 4 (Acsl4; Mm00490331_m1), and lysophosphatidylcholine acyltransferase 3 (Lpcat3; Mm00520147_m1) was quantified using real-time polymerase chain reaction (PCR) with a StepOne Real-Time PCR System (Applied Biosystems, Foster City, CA, USA) with the indicated TaqMan probes (Applied Biosystems). Gene expression was normalized to 18S (Mm03928990_g1) from the same samples, and relative expression was calculated using the comparative Ct method (ddCt). Total lipids were extracted from the liver, lungs, and plasma using Folch’s method with minor modifications. Briefly, weighed liver or lung tissue was transferred into a 1.5 mL Eppendorf tube and homogenized in 10 volumes of methanol and five to six 1.4 mm ceramic beads (Fisherbrand, Pittsburgh, PA, USA) using a Bead Mill 4 (Fisherbrand) homogenizer (30 s, 2 cycles); for plasma samples, 50 µL of plasma was mixed with 100 µL of methanol. Then, 100 µL of the methanolic homogenate containing 10 mg of each tissue or 150 µL of the plasma–methanol mixture was transferred to a new Eppendorf tube, after which 100 µL of a pre-mixed internal standard solution (10 ng of EquiSPLASH + 100 ng of oleic acid-d9) prepared in methanol was added. The mixture was vortexed at 3500 rpm for 30 s. After adding 400 µL of chloroform, the mixture was vortexed at 3500 rpm for 5 min, and 100 µL of Milli-Q water was added with an additional vortex for approximately 30 s. The biphasic extracts were centrifuged at 20,630×g for 10 min at 4 °C, and the lower chloroform layer was transferred to a new vial. The lipids contained in the upper aqueous layer were re-extracted with 400 µL of chloroform. The organic extracts were combined, and the organic solvent was evaporated under vacuum. The dried total lipids were re-dissolved in 100 µL of methanol with gentle vortexing and centrifuged at 20,630×g for 10 min at 4 °C, after which the supernatant was collected as the LC–MS sample. The untargeted lipidomic analysis was performed using a prominence UFLC system (Shimadzu Corp., Kyoto, Japan) connected to an LTQ Orbitrap MS (Thermo Fisher Scientific Inc., San Jose, CA) and an Atlantis T3 C18 column (2.1 × 150 mm, 3 µm, Waters, Milford, MA, USA). The flow rate was 200 μL/min with linear flow of the mobile phase, the injection volume via an autosampler was 10 µL, and the column temperature was 40 °C. The mobile phase comprised 10 mM CH3COONH4 (A), isopropanol (B), and methanol (C) in negative mode with a gradient of 30% B and 35% C (0–1 min); 80% B and 10% C (1–14 min); 85% B and 10% C (14–27 min) or in positive mode with a gradient 6% B and 90% C (0–1 min); 83% B and 15% C (1–10 min); 83% B and 15% C (10–19 min); 6% B and 90% C (19–19.5 min); 6% B and 90% C (19.5–22 min). Mass spectrometric data were acquired in negative mode with electron spray ionization (ESI) under the conditions: capillary temperature (330 °C), sheath gas flow (50 units), and auxiliary gas flow (20 units). The source voltage value was set to 3 kV. The Fourier transform full scan range was set to m/z 100–1750 to acquire MS spectra for high-resolution masses. Tandem MS (MS/MS) was performed at a collision energy of 40 V in the ion-trap mode to obtain low-resolution MS/MS spectra for identifying lipid molecular species. Raw data were processed using MS-DIAL (version 4.2) software (RIKEN, Wako, Japan) for the alignment, identification, and peak processing of lipid species. The semi-quantification of each lipid molecule was performed with the deuterated internal standard of the same lipid sub-classes or representative lipid class category, following the instructions of Lipidomics Standards Initiative (https://lipidomicstandards.org/) level 2 and level 3. The concentration of the lipid molecular species was calculated by taking the peak intensity ratios of each analyte to those of the internal standard and normalized by the weight or volume of the tissue or plasma, respectively. The concentrations of individual lipid molecules in the plasma, liver, and lungs are provided in Supplementary Tables S1, S2, and S3, respectively. The concentrations of 297 molecules were determined in the plasma samples, and the characteristics of lipid profiles in each biological sample were analyzed with MetaboAnalyst 5.0 (http://www.metaboanalyst.ca). Features with more than 25% relative standard deviation were removed (data filtering). The remaining 222 values were then log-transformed and normalized by the autoscaling function. Within them, 169 features were found with false discovery rate (FDR) < 0.01 by one-way analysis of variance (ANOVA). Sparse partial least squares discriminant analysis (sPLS-DA) was performed with 20 variables to calculate variable importance in projection scores for components 1 and 2. Utilizing the top 100 features with lowest FDR values, heatmap analysis with data clustering was performed using Euclidean distance measuring and a Ward clustering algorithm. All mouse experiments were performed with approval (approval# 17-003) from the Animal Care and Use Committee of Hokkaido University following the Fundamental Guidelines for Proper Conduct of Animal Experiment and Related Activities in Academic Research Institutions under the jurisdiction of the Ministry of Education, Culture, Sports, Science, and Technology in Japan. Body weight loss was monitored daily after infection, and mice were humanely euthanized when weight loss reached 25%. Statistical analyses were performed using Prism 9 software (GraphPad Software, San Diego, CA, USA). Differences were identified using two-way ANOVA or mixed-effects analysis with a correction for multiple comparisons, if necessary, and considered significant when p < 0.05. Data are represented as the mean ± SEM. Our previous studies defined severe symptoms of mice infected with the PR8 virus infection at 500 PFU, significant weight loss reaching over 25% at 3–6 dpi, and blood coagulation abnormalities such as increased thrombin production and intravascular blood clotting in the lungs. Therefore, this infection condition was employed to cause lethal disease in mice. In addition, infection at 50 PFU, which does not cause these pathological events, was used as a sublethal disease condition. Based on the observations in the previous studies, the plasma and lungs were collected from mice at 1, 3, and 6 dpi for representing samples at the very early, symptom-onset, and lethal stages during influenza, respectively. Upon intranasal infection with 50 or 500 PFU of PR8 virus, the mice showed significant body weight loss starting at 3 dpi, whereas the control mice showed no weight loss at any time point (Fig. 1a). Plaque assays were conducted to determine the virus titers in the lungs. In mice infected with 50 PFU, the highest virus titer was observed at 6 dpi, whereas in those infected with 500 PFU, the titer peaked at 3 dpi and decreased at 6 dpi (Fig. 1b). At 3 dpi, the virus titers between the mice in the 500 PFU group was significantly higher than those in the 50 PFU group.Figure 1Body weight change, virus titers in the lungs, and plasma concentrations of cytokines/chemokines. Male C57BL/6 mice were intranasally inoculated with PBS control or PBS comprising PR8 virus (50 or 500 PFU), and (a) body weight change was monitored. At 1, 3, and 6 dpi, lung and plasma samples were collected to evaluate (b) virus titers in the lungs and (c–h) cytokine/chemokine concentrations in the plasma. (a) The body weight change of the mice was calculated as a percentage of the original weight. Symbols represent mean ± SEM (n = 5–15 mice; 5 mice in each group were euthanized at 1, 3, and 6 dpi). White, gray, and black symbols indicate data from PBS control, and 50 or 500 PFU of PR8 virus-infected mice, respectively. Mixed-effect analysis using a multiple comparison correction. (b) Virus titers in the lungs were determined by a plaque assay on MDCK cells. Dots represent individual values, and the mean values with SEM are indicated by lines (n = 5). Gray and black symbols indicate data from mice infected with 50 and 500 PFU of the PR8 virus, respectively. (c–h) Concentrations of cytokines and chemokines in the plasma, (c) IFN-γ, (d) IL-6, (e) IP-10, (f) MCP-1, (g) MIP-1β, and (h) TNF-α, were determined by a multiplex ELISA kit. In each panel, dots represent individual values, and bars represent the mean ± SEM (n = 5 mice). White, light gray, and dark gray bars indicate data from PBS control, and 50 or 500 PFU of PR8 virus-infected mice, respectively. Two-way ANOVA using a multiple comparison correction. (a–h) PR8 influenza virus A/Puerto Rico/8/34, PFU plaque-forming unit, dpi day-post-infection, ND not detected. Body weight change, virus titers in the lungs, and plasma concentrations of cytokines/chemokines. Male C57BL/6 mice were intranasally inoculated with PBS control or PBS comprising PR8 virus (50 or 500 PFU), and (a) body weight change was monitored. At 1, 3, and 6 dpi, lung and plasma samples were collected to evaluate (b) virus titers in the lungs and (c–h) cytokine/chemokine concentrations in the plasma. (a) The body weight change of the mice was calculated as a percentage of the original weight. Symbols represent mean ± SEM (n = 5–15 mice; 5 mice in each group were euthanized at 1, 3, and 6 dpi). White, gray, and black symbols indicate data from PBS control, and 50 or 500 PFU of PR8 virus-infected mice, respectively. Mixed-effect analysis using a multiple comparison correction. (b) Virus titers in the lungs were determined by a plaque assay on MDCK cells. Dots represent individual values, and the mean values with SEM are indicated by lines (n = 5). Gray and black symbols indicate data from mice infected with 50 and 500 PFU of the PR8 virus, respectively. (c–h) Concentrations of cytokines and chemokines in the plasma, (c) IFN-γ, (d) IL-6, (e) IP-10, (f) MCP-1, (g) MIP-1β, and (h) TNF-α, were determined by a multiplex ELISA kit. In each panel, dots represent individual values, and bars represent the mean ± SEM (n = 5 mice). White, light gray, and dark gray bars indicate data from PBS control, and 50 or 500 PFU of PR8 virus-infected mice, respectively. Two-way ANOVA using a multiple comparison correction. (a–h) PR8 influenza virus A/Puerto Rico/8/34, PFU plaque-forming unit, dpi day-post-infection, ND not detected. Proinflammatory cytokines and chemokines, IFN-γ (Fig. 1c), IL-6 (Fig. 1d), IP-10 (Fig. 1e), MCP-1 (Fig. 1f), MIP-1β (Fig. 1g), and TNF-α (Fig. 1h), were measured in the plasma samples collected from PR8 virus-infected mice at 1, 3, and 6 dpi. Among them, IFN-γ, IL-6, and TNF-α were similarly increased at 3 and 6 dpi in both infected groups; however, MCP-1, IP-10, and MIP-1β were higher induced in the lethally infected mice than in those infected with a sublethal dose, especially at 3 dpi. These results indicate that viral infection under both conditions caused systemic inflammation, but it was more severe in lethally infected mice than in the sublethal condition. Under these experimental conditions, comprehensive lipidome analyses were further performed to explore lipid profiles associated with influenza virus infection, particularly in the severe case. Untargeted lipidome analyses were performed on the plasma samples collected at 1, 3, and 6 dpi by the HPLC–LTQ–Orbitrap MS, and 297 lipid molecular species were identified after confirming their MS/MS spectra. The lipid molecular species from major lipid classes included 148 phospholipids [25 phosphatidylcholines (PC), 50 phosphatidylethanolamines (PE), 2 phosphatidic acids (PA), 6 phosphatidylethanols (PEtOH), 1 phosphatidylmethanol (PmeOH), 17 phosphatidylglycerols (PG), 22 phosphatidylinositols (PI), and 15 phosphatidylserines (PS), 8 cardiolipins (CL) and 2 monolysocardiolipins (MLCL)], 45 lysophospholipids [18 lysophosphatidylcholines (LPC), 13 lysophosphatidylethanolamines (LPE), 1 lysophosphatidic acid (LPA), 2 lysophosphatidylglycerols (LPG), 8 lysophosphatidylinositols (LPI), and 3 lysophosphatidylserines (LPS)], 23 free fatty acids (FAs), 45 sphingolipids [24 ceramides (CER), 7 hexosylceramides (HexCER), and 14 sphingomyelins (SM)], 6 cholesteryl esters (CE), 20 triacylglycerols (TG), and 10 diacylglycerols (DG). Elution profiles and plasma concentrations of the annotated lipids are provided in Supplementary Fig. S1 and Supplementary Table S1. The plasma lipid concentrations of each category are shown in Supplementary Fig. S2. Infected mice demonstrated increased PC, PE, and PG at 3 dpi (Supplementary Fig. S2a), increased lysophospholipids at 3 and 6 dpi (Supplementary Fig. S2b), increased FA at 3 dpi and decreased TG and DG at 3 and 6 dpi (Supplementary Fig. S2c). While increased LPC and LPI, and decreased TG and DG were also observed in sublethally infected mice, the elevation of phospholipids at 3 dpi appeared to be specifically associated with lethal infection. Furthermore, sPLS-DA revealed the characteristics of the effect of disease stages and infection conditions, sublethal (50 PFU) and lethal (500 PFU), on the lipid profile in the plasma (Fig. 2a). At 6 dpi, comparable effects of PR8 virus infection under both conditions on plasma lipids were observed, as indicated by similar high values of component 1 scores of the groups in the plots. On the other hand, at 3 dpi, an increase in the component 2 score was observed uniquely in the mice infected under the lethal condition. Top 20 features based on variable importance for prediction (VIP) scores for components 1 and 2 are listed in Fig. 2b and c, respectively, and the individual concentrations of top 5 lipids are shown in Supplementary Fig. S3.Figure 2sPLS-DA of plasma lipidome data. Male C57BL/6 mice were intranasally inoculated with PBS control or PBS comprising PR8 virus (50 or 500 PFU), and lipidome analyses determined the concentrations of 297 molecules in the plasma samples collected at 1, 3, and 6 dpi (n = 5 mice). After data filtering on the basis of relative standard deviation, the remaining 222 values were then log-transformed and normalized by the autoscaling function. The sparse partial least squares discriminant analysis (sPLS-DA) was performed with 20 normalized lipid levels in the plasma as variables to calculate variable importance in projection scores for components 1 and 2. (a) The score plot between component 1 and component 2. Dots represent samples and circles represent the 95% confidence region for each treatment group. (b,c) Variable importance for prediction (VIP) scores for (b) component 1 and (c) component 2. The colored boxes on the right indicate relative values of the features in each group. PR8 influenza virus A/Puerto Rico/8/34, PFU plaque-forming-unit, dpi day-post-infection, PA phosphatidic acid, PC phosphatidylcholine, PE phosphatidylethanolamine, PI phosphatidylinositol, PS phosphatidylserine, LPE lysophosphatidylethanolamine, Cer ceramide, TG triacylglycerol. sPLS-DA of plasma lipidome data. Male C57BL/6 mice were intranasally inoculated with PBS control or PBS comprising PR8 virus (50 or 500 PFU), and lipidome analyses determined the concentrations of 297 molecules in the plasma samples collected at 1, 3, and 6 dpi (n = 5 mice). After data filtering on the basis of relative standard deviation, the remaining 222 values were then log-transformed and normalized by the autoscaling function. The sparse partial least squares discriminant analysis (sPLS-DA) was performed with 20 normalized lipid levels in the plasma as variables to calculate variable importance in projection scores for components 1 and 2. (a) The score plot between component 1 and component 2. Dots represent samples and circles represent the 95% confidence region for each treatment group. (b,c) Variable importance for prediction (VIP) scores for (b) component 1 and (c) component 2. The colored boxes on the right indicate relative values of the features in each group. PR8 influenza virus A/Puerto Rico/8/34, PFU plaque-forming-unit, dpi day-post-infection, PA phosphatidic acid, PC phosphatidylcholine, PE phosphatidylethanolamine, PI phosphatidylinositol, PS phosphatidylserine, LPE lysophosphatidylethanolamine, Cer ceramide, TG triacylglycerol. The clustering analysis in the heatmap demonstrated that lipid profiles of plasma samples from lethally and sublethally infected mice at 6 dpi were in the same cluster and that lethally infected mice showed a distinct tendency of the profile at 3 dpi (Fig. 3a). In addition to multivariate analysis, univariate analysis on lipidome data also suggested a unique tendency of the lipid profile of lethally infected mice at 3 dpi (Supplementary Fig. S4). Considering the alteration pattern by infection severity and disease stage, lipid molecules were classified to 3 groups; Group 1, increased at 6 dpi under both infection conditions; Group 2, increased uniquely at 3 dpi by lethal infection and decreased at 6 dpi under both conditions; Group 3, decreased under both infection conditions at 6 dpi. While Group 1 mainly consisted of lysophospholipids, most of the lipids in Group 2 were PEs containing polyunsaturated fatty acid (PUFA). Group 3 included TG and DG as well as PE, PC, and PI containing monounsaturated fatty acid (MUFA) containing phospholipids.Figure 3Heatmap analysis of plasma lipidome data. Male C57BL/6 mice were intranasally inoculated with PBS control or PBS comprising PR8 virus (50 or 500 PFU), and lipidome analyses determined the concentrations of 297 molecules in the plasma samples collected at 1, 3, and 6 dpi (n = 5 mice). After data filtering on the basis of relative standard deviation, the remaining 222 values were then log-transformed and normalized by the autoscaling function. Within them, 169 significant features were found; false discovery rate (FDR) < 0.01, one-way analysis of variance (ANOVA). (a) Utilizing the top 100 lipids with the lowest FDR values, heatmap analysis with data clustering was performed using Euclidean distance measuring and a Ward clustering algorithm. The colored boxes indicate the relative values of the features in each group. The three major clusters are shown in gray, red, and blue. (b) Concentrations of PEs containing monounsaturated fatty acid (MUFA; PE-MUFA) and polyunsaturated fatty acid (PUFA; PE-PUFA) and (c) the ratios of PE-PUFA to PE-MUFA are shown separately. (b,c) In each panel, dots represent individual values, the box shows the interquartile range, the horizontal line within the box shows the median value, and the whiskers/vertical lines show maximum (top) and minimum (bottom) values. White, light gray, and dark gray boxes indicate data from PBS control, and 50 or 500 PFU of PR8 virus-infected mice, respectively. Two-way ANOVA using a multiple comparison correction. (a–c) PR8 influenza virus A/Puerto Rico/8/34, PFU plaque-forming unit, dpi day-post-infection. Heatmap analysis of plasma lipidome data. Male C57BL/6 mice were intranasally inoculated with PBS control or PBS comprising PR8 virus (50 or 500 PFU), and lipidome analyses determined the concentrations of 297 molecules in the plasma samples collected at 1, 3, and 6 dpi (n = 5 mice). After data filtering on the basis of relative standard deviation, the remaining 222 values were then log-transformed and normalized by the autoscaling function. Within them, 169 significant features were found; false discovery rate (FDR) < 0.01, one-way analysis of variance (ANOVA). (a) Utilizing the top 100 lipids with the lowest FDR values, heatmap analysis with data clustering was performed using Euclidean distance measuring and a Ward clustering algorithm. The colored boxes indicate the relative values of the features in each group. The three major clusters are shown in gray, red, and blue. (b) Concentrations of PEs containing monounsaturated fatty acid (MUFA; PE-MUFA) and polyunsaturated fatty acid (PUFA; PE-PUFA) and (c) the ratios of PE-PUFA to PE-MUFA are shown separately. (b,c) In each panel, dots represent individual values, the box shows the interquartile range, the horizontal line within the box shows the median value, and the whiskers/vertical lines show maximum (top) and minimum (bottom) values. White, light gray, and dark gray boxes indicate data from PBS control, and 50 or 500 PFU of PR8 virus-infected mice, respectively. Two-way ANOVA using a multiple comparison correction. (a–c) PR8 influenza virus A/Puerto Rico/8/34, PFU plaque-forming unit, dpi day-post-infection. Given the results of sPLS-DA and heatmap analysis, PE molecules were suggested to change significantly by infection depending on infection dose and disease stage. Moreover, the increasing and decreasing tendency of PE appeared to be associated with the degree of unsaturation of the fatty acid contained. The concentrations of MUFA- and PUFA-containing PE (PE-MUFA and PE-PUFA, respectively) are shown in Fig. 3b. The ratio of PE-PUFA to PE-MUFA significantly increased at 3 and 6 dpi, with a higher increase in lethally infected mice compared to those in the sublethal group at 3 dpi (Fig. 3c). Among PE-PUFA, especially arachidonic acid (20:4), DHA (22:6), and EPA (20:5) showed a marked increase in concentrations upon severe infection at 3 dpi, and this trend was also observed in other phospholipids (Fig. 4). The total concentration of PE, PC, and PI containing arachidonic acid in the plasma increased by 2.20-, 2.41-, and 1.59-fold, respectively (Fig. 4a). In addition to arachidonic acid, the concentrations of phospholipids containing DHA were elevated during severe influenza, by 4.23-fold in PC and 1.69-fold in PE (Fig. 4b); PS, PE, PC, and PI containing EPA increased by 2.23-, 1.95-, 1.80-, and 1.77-fold (Fig. 4c). Therefore, an increase in PE and PC containing PUFA, particularly arachidonic acid and DHA, was considered the most prominent feature of the lipid profile during severe influenza.Figure 4Concentrations of phospholipids containing 20:4, 22:6, or 20:5 in the plasma. Male C57BL/6 mice were intranasally inoculated with PBS control or PBS comprising PR8 virus (50 or 500 PFU), and lipidome analyses were performed with plasma samples collected at 1, 3, and 6 dpi (n = 5 mice). Log-transformed concentrations of phospholipids containing (a) 20:4 (arachidonic acid), (b) 22:6 (DHA), and (c) 20:5 (EPA) in the plasma samples are shown here. In each panel, dots represent individual values, the box shows the interquartile range, the horizontal line within the box shows the median value, and the whiskers/vertical lines show maximum (top) and minimum (bottom) values. White, light gray, and dark gray boxes indicate data from PBS control, and 50 or 500 PFU of PR8 virus-infected mice, respectively. Two-way ANOVA using a multiple comparison correction. PR8 influenza virus A/Puerto Rico/8/34, PFU plaque-forming unit, dpi day-post-infection, DHA docosahexaenoic acid, EPA eicosapentaenoic acid, PC phosphatidylcholine, PE phosphatidylethanolamine, PG phosphatidylglycerol, PI phosphatidylinositol, PS phosphatidylserine. Concentrations of phospholipids containing 20:4, 22:6, or 20:5 in the plasma. Male C57BL/6 mice were intranasally inoculated with PBS control or PBS comprising PR8 virus (50 or 500 PFU), and lipidome analyses were performed with plasma samples collected at 1, 3, and 6 dpi (n = 5 mice). Log-transformed concentrations of phospholipids containing (a) 20:4 (arachidonic acid), (b) 22:6 (DHA), and (c) 20:5 (EPA) in the plasma samples are shown here. In each panel, dots represent individual values, the box shows the interquartile range, the horizontal line within the box shows the median value, and the whiskers/vertical lines show maximum (top) and minimum (bottom) values. White, light gray, and dark gray boxes indicate data from PBS control, and 50 or 500 PFU of PR8 virus-infected mice, respectively. Two-way ANOVA using a multiple comparison correction. PR8 influenza virus A/Puerto Rico/8/34, PFU plaque-forming unit, dpi day-post-infection, DHA docosahexaenoic acid, EPA eicosapentaenoic acid, PC phosphatidylcholine, PE phosphatidylethanolamine, PG phosphatidylglycerol, PI phosphatidylinositol, PS phosphatidylserine. To find the source for the PE-PUFA elevated in the plasma, we conducted another infectious experiment with the lethal condition and measured lipid contents in lung and liver samples, which are the infection site and the primary tissue that determines plasma lipid profiles, respectively. Although not all lipid species showed the same trend, in addition to the plasma, the elevation of PE-20:4 and PE-22:6 and that of PE-22:6 and PC-22:6 were confirmed in the lungs and livers, respectively, of severely infected mice at 3 dpi (Fig. 5). Since PC-PUFA was not detected in the lung samples, PE metabolic pathway was further focused on.Figure 5Concentrations of PEs and PCs containing 20:4, 22:6, or 20:5 in the lungs and the liver. Male C57BL/6 mice were intranasally inoculated with PBS control or PBS comprising 500 plaque-forming unit of PR8 virus, and lipidome analyses were performed with tissue samples collected at 1, 3, and 6 dpi (n = 5 mice). Log-transformed concentrations of phosphatidylethanolamines (PEs) and phosphatidylcholines (PCs) containing 20:4 (arachidonic acid), 22:6 (DHA), or 20:5 (EPA) in (a) the lungs and (b) the liver are shown here. In each panel, dots represent individual values, the box shows the interquartile range, the horizontal line within the box shows the median value, and the whiskers/vertical lines show maximum (top) and minimum (bottom) values. White, light gray, and dark gray boxes indicate data from PBS control, and PR8 virus-infected mice, respectively. Two-way ANOVA using a multiple comparison correction. PR8 influenza virus A/Puerto Rico/8/34, dpi day-post-infection, DHA docosahexaenoic acid, EPA eicosapentaenoic acid. Concentrations of PEs and PCs containing 20:4, 22:6, or 20:5 in the lungs and the liver. Male C57BL/6 mice were intranasally inoculated with PBS control or PBS comprising 500 plaque-forming unit of PR8 virus, and lipidome analyses were performed with tissue samples collected at 1, 3, and 6 dpi (n = 5 mice). Log-transformed concentrations of phosphatidylethanolamines (PEs) and phosphatidylcholines (PCs) containing 20:4 (arachidonic acid), 22:6 (DHA), or 20:5 (EPA) in (a) the lungs and (b) the liver are shown here. In each panel, dots represent individual values, the box shows the interquartile range, the horizontal line within the box shows the median value, and the whiskers/vertical lines show maximum (top) and minimum (bottom) values. White, light gray, and dark gray boxes indicate data from PBS control, and PR8 virus-infected mice, respectively. Two-way ANOVA using a multiple comparison correction. PR8 influenza virus A/Puerto Rico/8/34, dpi day-post-infection, DHA docosahexaenoic acid, EPA eicosapentaenoic acid. As shown in Fig. 6a, there are two main pathways for PE-PUFA production; de novo synthesis from ethanolamine and remodeling of PUFA hydrolyzed from lipid membrane. Since viral infection causes the alteration of lipid metabolism, the effect of PR8 virus infection on the gene expression of enzymes related to these pathways was investigated in the lungs and the liver (Fig. 6b). Gene expression levels of Etnk1 and Ept1, rate-limiting enzymes in the reaction of converting ethanolamine to PE, were increased in the lungs at 3 dpi, whereas the liver of infected mice showed the increase and the decrease in those of Etnk1 and Ept1, respectively at 3 dpi. The expression of PC-metabolizing enzymes was also increased at 3 dpi, particularly in the lung (Supplementary Fig. S5).Figure 6Gene expression of PUFA-containing PE metabolizing enzymes. (a) Schematic representation of PE containing PUFA (PE-PUFA) biosynthesis. PE phosphatidylethanolamine, PUFA polyunsaturated fatty acids, DG diacylglycerol, LPE lysophosphatidylethanolamine, ETNK1 ethanolamine kinase 1, PCYT2 phosphate cytidylyltransferase 2, ethanolamine, EPT1 ethanolamine phosphotransferase 1, PLA2 phospholipase A2, ACSL4 acyl-CoA synthetase long chain family member 4, LPLAT lysophospholipid acyltransferase. (b) Male C57BL/6 mice were intranasally inoculated with PBS control or PBS comprising 500 plaque-forming unit of PR8 virus, and relative gene expression levels of Etnk1, Pcyt2, Ept1, phospholipase A2 group IVA (Pla2g4a) encoding cPLA2, Acsl4, and lysophosphatidylcholine acyltransferase 3 (Lpcat3) were measured in lung and liver samples collected at 1, 3, and 6 dpi by real-time PCR. Bars represent the mean ± SEM (n = 4 mice). In each panel, dots represent individual values, and white and gray bars indicate data from PBS control and PR8 virus-infected mice, respectively. Two-way ANOVA using a multiple comparison correction on ddCt values. PR8 influenza virus A/Puerto Rico/8/34, dpi day-post-infection. Gene expression of PUFA-containing PE metabolizing enzymes. (a) Schematic representation of PE containing PUFA (PE-PUFA) biosynthesis. PE phosphatidylethanolamine, PUFA polyunsaturated fatty acids, DG diacylglycerol, LPE lysophosphatidylethanolamine, ETNK1 ethanolamine kinase 1, PCYT2 phosphate cytidylyltransferase 2, ethanolamine, EPT1 ethanolamine phosphotransferase 1, PLA2 phospholipase A2, ACSL4 acyl-CoA synthetase long chain family member 4, LPLAT lysophospholipid acyltransferase. (b) Male C57BL/6 mice were intranasally inoculated with PBS control or PBS comprising 500 plaque-forming unit of PR8 virus, and relative gene expression levels of Etnk1, Pcyt2, Ept1, phospholipase A2 group IVA (Pla2g4a) encoding cPLA2, Acsl4, and lysophosphatidylcholine acyltransferase 3 (Lpcat3) were measured in lung and liver samples collected at 1, 3, and 6 dpi by real-time PCR. Bars represent the mean ± SEM (n = 4 mice). In each panel, dots represent individual values, and white and gray bars indicate data from PBS control and PR8 virus-infected mice, respectively. Two-way ANOVA using a multiple comparison correction on ddCt values. PR8 influenza virus A/Puerto Rico/8/34, dpi day-post-infection. In the lungs and the liver of mice severely infected with PR8 virus, the gene expression of Pla2g4a, which encodes cPLA2, a well-characterized PLA2, that hydrolyses arachidonic acid from the lipid membrane, was significantly increased, suggesting that arachidonic acid hydrolysis from the lipid membrane is facilitated in these tissues upon virus infection. In addition to Pla2g4a, Acsl4 was significantly increased only in the lungs at 3 dpi, whereas no increase in the expression of Lpcat3, a well-characterized lysophospholipid acyltransferase (LPLAT) contributing to the production of PE-20:4, PE-22:6, and PC-20:4, was observed. Since ACSL4 critically contributes to the acylation of various PUFA including 20:4, 20:5, and 22:6 in the remodeling pathway of free arachidonic acid, the induction of Acsl4 observed in the lungs may explain the increase in PE-PUFA at least partially. In this study, the kinetic lipid profiles during sublethal and lethal influenza in a mouse model were obtained by untargeted lipidome analyses with the HPLC–LTQ–Orbitrap MS, which determined concentrations of 297 lipid molecular species in each plasma sample. Although infected mice showed clear changes in the concentrations of various types of lipids, not all of which may be due to infection. It should be noted that the mice were not starved in this study, so there could be differences in the effects of feeding conditions on lipid profiles between infected and control mice. For example, it is suspected that the decreased TG and DG levels observed in infected mice may have been due to decreased food intake in the animals. This speculation is based on a previous study that most of the lipids including lysophospholipids, phospholipids, CE, and TG were decreased in serum from Sprague–Dawley rats after fasting for 22 h. However, in contrast to simple fasting, influenza virus infection resulted in an increase in lysophospholipids (LPC, LPE, LPG, LPS, and LPI) and phospholipids (PC, PE, PG, and PS) in this study (Supplementary Fig. S2). Particularly, our multivariate and univariate analyses on lipidome data in the plasma demonstrated that the most prominent feature of the lipid profile in the plasma of the lethally infected mice was the elevated concentrations of PC- and PE-PUFA at 3 dpi, when the virus titer in the lungs was the highest. Therefore, the increase in these identified phospholipids is considered to have resulted from lethal infection. Gene expression analyses demonstrated a significant increase in Acsl4 as well as PE-biosynthesizing enzymes at 3 dpi in the lungs of infected mice (Fig. 6), which could explain the elevation of PE-20:4 and PE-22:6 in the tissue. A previous study demonstrated higher levels of PE-20:4 and PC-20:4 in the lungs in an influenza ferret model. Although PC-20:4 was reported to be abundant even in the airway cells of uninfected mouse lungs by matrix-assisted laser desorption/ionization (MALDI) imaging MS, it was not detected in our analysis, probably due to differences in lipid detection methods. However, given that ACSL4 is also involved in the PC-20:4 synthesis, and that gene expression of PC-metabolizing enzymes, such as CHK and PCYT1, was significantly increased (Supplementary Fig. S5), PC-20:4 production was thought to be increased in the lungs of infected mice at 3 dpi in the present study. Another possibility is that PC-PUFA abundantly detected in the plasma at 3 dpi was supplied from tissues other than the lung. To address this point, a comprehensive lipidome analysis of plasma and multi-organs from infected mice will be needed. In addition, the critical involvement of ACSL4 in the alteration of phospholipids observed in our study needs to be elucidated in the future, for example by using ACLS4 knockout mice or a specific inhibitor. At 6 dpi, the expression levels of several enzymes involved in the PE and PC synthesis decreased in the lungs (Fig. 6 and Supplementary Fig. S5), suggesting that phospholipid biosynthesis decreases in the later stage of infection. This notion is supported by the significant decrease in total PEs and PCs in the lungs at 6 dpi (Supplementary Fig. S6), and the same tendency was observed in the plasma (Supplementary Fig. S2). In a previous study, the significant decrease of PC (16:0/16:0), PC (16:0/16:1), PC (16:0/18:2), PE (16:0/18:2), PG (16:0/16:0), and PG (16:0/18:1) was demonstrated in alveolar type II (AT2) cells isolated from influenza virus-infected mice at 6 dpi. All these lipids decreased by 0.47-, 0.39-, 0.69-, 0.76-, 0.33-, and 0.45-fold, respectively, in the lungs of infected mice at 6 dpi also in the present study (Supplementary Table S3). PR8 virus infects and induces cell death in bronchial Clara cells and AT2 cells in mouse lungs. Given that AT2 cells actively produce surfactant lipids and highly express ACSL4, both infection and infection-induced cell death should have affected lipid metabolism in AT2 cells. Furthermore, since PC (16:0/16:0) and PC (16:0/16:1) are major components of lung surfactant, the significant decrease in the PCs may be associated with the dysregulation of surfactant, which contributes to the host defense against virus infection. These results suggest that altered gene expression of phospholipid-metabolizing enzymes after infection determined the phospholipid profile in the lungs, which could have affected those in the blood probably through extracellular vesicles secreted from lung cells, such as alveolar epithelial cells, macrophages, and vascular endothelial cells, as reported in lung injury and inflammation. In addition to increased cytokines, the phospholipids increased in the plasma may also transduce inflammatory signals to tissues distant from the site of infection. Interestingly, the liver of infected mice, which is not an infection target of the PR8 virus, also showed an increase of PC- and PE-PUFA, particularly PC-22:6 and PE-22:6, at 3 dpi (Fig. 5). Although the exact mechanisms under which respiratory infection induced these changes need to be elucidated in the future, the altered lipid profiles in extrapulmonary organs may be associated with systemic dysfunction in severe influenza. The elevation of PC- and PE-20:4 level may be a counteraction against the increase of free arachidonic acid to reduce the production of its metabolites prostaglandins by incorporating the lipid into phospholipids. This is called the arachidonic acid remodeling pathway, where ACSL4 plays a pivotal role. Interestingly, the protein expression of ACSL4 has been reported to be induced by the infection with SARS-CoV-2 and influenza virus, which may explain an increase in PE-20:4 and PE-22:6 reported in the plasma of COVID-19 patients in addition to that in PE-20:4, PE-22:6, PC-20:4, and PC-22:6 in influenza studies in animal models including the present study. Clinical studies on patients with Ebola and zika virus infectious diseases also reported elevated PE-PUFA including PE-20:4 and PE-22:6, but not PC-PUFA. This difference may be associated with the tropism of viruses that seems to affect the phospholipid metabolism in each tissue differently. There is a possibility that an increase in PE containing PUFA occurs in various infectious diseases, with induction of ACSL4 as a common mechanism. Although further investigations are needed, PE-PUFA levels in circulation and the ratio of PE-PUFA to PE-MUFA (Fig. 3b) could be a biomarker of diverse infectious diseases, particularly respiratory viral infections. Furthermore, PE-20:4 and/or its oxidized products may be involved in the pathogenicity of infectious diseases. PC and PE containing PUFA, particularly arachidonic acid and DHA, are highly susceptible to oxidation, causing ferroptosis, a new class of programmed cell death if lipid peroxide reduction is insufficient. Given the increase in malondialdehyde or 4-hydroxynonenal, lipid peroxide metabolites, in patients infected with influenza virus, hepatitis C virus, dengue virus, and SARS-CoV-2, the lipid oxidation–reduction balance is thought to incline toward oxidation by viral infections. Notably, direct evidence of phospholipid peroxides-associated cell death has been recently shown in A549 cells infected with influenza virus. Therefore, elucidating the biological significance of altered phospholipid metabolism and oxidized derivatives in the hosts will provide insights into pathogenesis of viral infections. This study has several limitations. First, it was difficult to compare concentrations between different lipid molecules because relative, but not absolute, quantification was conducted for each lipid. Second, plasma and tissue samples were prepared in different sets of experiments, so that correlation analysis between lipidome data of plasma and tissues could not be performed. In future studies, these points need to be improved. In summary, we elucidated the plasma lipid profile of mice in sublethal and lethal influenza by untargeted lipidome analyses using HPLC–LTQ–Orbitrap MS. Of note, the increase in plasma concentrations of PE-PUFA was observed only in the middle stage of severe influenza but not in mice with mild disease, when gene expression analyses and lipid profiles in the lungs suggested that the production of PE-PUFA was increased in the lungs. In the future, we will further investigate the biological significance of the alteration in phospholipid metabolism associated with severe influenza in regulating host immune responses, especially immune cell recruitment and infected cell clearance. In addition, the sharp decrease in CE at the early stage of infection in the lungs (Supplementary Fig. S6) is also an attractive research target. These investigations will provide novel insight into the pathogenicity and therapeutic strategy of infectious diseases.
|
PMC9546883
|
Deficiency of the frontotemporal dementia gene GRN results in gangliosidosis
|
Haploinsufficiency of GRN causes frontotemporal dementia (FTD). The GRN locus produces progranulin (PGRN), which is cleaved to lysosomal granulin polypeptides. The function of lysosomal granulins and why their absence causes neurodegeneration are unclear. Here we discover that PGRN-deficient human cells and murine brains, as well as human frontal lobes from GRN-mutation FTD patients have increased levels of gangliosides, glycosphingolipids that contain sialic acid. In these cells and tissues, levels of lysosomal enzymes that catabolize gangliosides were normal, but levels of bis(monoacylglycero)phosphates (BMP), lipids required for ganglioside catabolism, were reduced with PGRN deficiency. Our findings indicate that granulins are required to maintain BMP levels to support ganglioside catabolism, and that PGRN deficiency in lysosomes leads to gangliosidosis. Lysosomal ganglioside accumulation may contribute to neuroinflammation and neurodegeneration susceptibility observed in FTD due to PGRN deficiency and other neurodegenerative diseases.About half of the human brain mass is composed of lipids. Lysosomes are crucial in degrading and recycling cellular lipids, and accumulation of lipids and other macromolecules due to lysosomal dysfunction is linked to numerous neurodevelopmental and neurodegenerative diseases broadly classified as lysosomal storage disorders. Granulins are polypeptides produced from progranulin (PGRN), a precursor protein that is cleaved in the lysosome. Deficiency of granulins due to homozygous mutations in the GRN gene lead to neuronal ceroid lipofuscinosis, a severe neurodevelopmental disease, in humans and neuroinflammation in mice. Haploinsufficiency of GRN almost invariably causes frontotemporal dementia (FTD). How granulins function in lysosomes and why their absence causes neurodegeneration is unclear. Inasmuch as granulin-deficiency is associated with lipofuscin accumulation, we tested the hypothesis that PGRN deficiency results in detrimental lysosomal lipid abnormalities. Consistent with this notion, a previous lipidomic study showed that PGRN deficiency in humans or mice alters levels of brain triglycerides (TAG), sterol esters (SE), and phosphatidylserine (PS). However, this study did not examine gangliosides, which are sialic-acid-containing glycosphingolipids that are highly abundant in the nervous system. Ganglioside catabolism occurs in the lysosome, and defects in the abundance or activity of different enzymes that catabolize gangliosides result in severe neurological diseases, such as Tay-Sachs or Sandhoff diseases (Fig. 1a).Fig. 1Deficiency of progranulin leads to ganglioside accumulation in mouse and human brain tissues.a Ganglioside degradation pathway in the lysosome. The names of glycosyl hydrolases (green), activator proteins (purple), and associated metabolic diseases (red) are indicated in the scheme. b Quantification of mono-sialyated and di-sialyated ganglioside species isolated from Grn (gray) (n = 6), Grn (blue) (n = 4), and Grn (purple) (n = 4) mouse brains. c Quantification of mono-sialyated and di-sialyated ganglioside species isolated from the frontal lobes of control (pink), FTD-TDP43-A (sporadic-non-GRN) (green), and FTD-TDP43-A (GRN) (blue) human brains. Box plots display mean ± the minimum and maximum number in the data set of control (n = 3), FTD-TDP43-A (sporadic-non-GRN) (n = 6) or FTD-TDP43-A (GRN) (n = 12). Box plots display mean ± the minimum and maximum number. One-way ANOVA, followed by multigroup comparison (Dunn’s) test, was performed. *p < 0.05, **p < 0.01. G is for ganglioside; M/D/T are for monosialic, disialic, or trisialic; and the number refers to the order of discovery. GM2-AP GM2 ganglioside activator protein, SAP-B/C/D saposin-B/C/D, GLB1 galactosidase beta 1, HEXA beta-hexosaminidase subunit alpha, NEU3/4 neuraminidase 3/4, GALC galactosylceramidase, GCase glucosylceramidase beta, ASAH1 N-acylsphingosine amidohydrolase 1, SAP-C/D saposin-C/D. a Ganglioside degradation pathway in the lysosome. The names of glycosyl hydrolases (green), activator proteins (purple), and associated metabolic diseases (red) are indicated in the scheme. b Quantification of mono-sialyated and di-sialyated ganglioside species isolated from Grn (gray) (n = 6), Grn (blue) (n = 4), and Grn (purple) (n = 4) mouse brains. c Quantification of mono-sialyated and di-sialyated ganglioside species isolated from the frontal lobes of control (pink), FTD-TDP43-A (sporadic-non-GRN) (green), and FTD-TDP43-A (GRN) (blue) human brains. Box plots display mean ± the minimum and maximum number in the data set of control (n = 3), FTD-TDP43-A (sporadic-non-GRN) (n = 6) or FTD-TDP43-A (GRN) (n = 12). Box plots display mean ± the minimum and maximum number. One-way ANOVA, followed by multigroup comparison (Dunn’s) test, was performed. *p < 0.05, **p < 0.01. G is for ganglioside; M/D/T are for monosialic, disialic, or trisialic; and the number refers to the order of discovery. GM2-AP GM2 ganglioside activator protein, SAP-B/C/D saposin-B/C/D, GLB1 galactosidase beta 1, HEXA beta-hexosaminidase subunit alpha, NEU3/4 neuraminidase 3/4, GALC galactosylceramidase, GCase glucosylceramidase beta, ASAH1 N-acylsphingosine amidohydrolase 1, SAP-C/D saposin-C/D. Here we discover that PGRN deficiency in cells, murine brains, or human frontal lobes of subjects with FTD due to GRN mutations, results in gangliosidosis. By studying this phenotype in human cells lacking PGRN, we uncover a mechanism for the gangliosidosis due to deficiency of lysosomal lipids that are required for ganglioside degradation. Our findings suggest a model for an adult neurodegenerative disease that may result from defective clearance of lysosomal lipids. We utilized lipidomics to examine glycosphingolipids in PGRN-deficient tissues and cells. We first analyzed brains isolated from 18-month-old Grn mice, a murine model of PGRN deficiency. These mice harbor the murine equivalent of the most prevalent human GRN mutation that causes FTD (R493X). They phenocopy Grn knockout mice, exhibiting CNS microgliosis, cytoplasmic TDP-43 accumulation, reduced synaptic density, lipofuscinosis, and excessive grooming behavior. Lipidomics revealed increased levels of mono-sialylated GM1 species and a two- to four-fold increase in di-sialylated GD3 species in mouse whole cortex brains of Grn mice (Fig. 1b, Supplementary Data 1). Levels of these gangliosides trended higher in Grn heterozygous mice, but did not reach significance. GM2 levels also trended higher in Grn mice, but were not significantly changed in Grn mice. Di-sialylated GD1 species were increased in brains of Grn brains and trended higher in Grn brains (Fig. 1b, Supplementary Fig. 1a, and Supplementary Data 1). We also found modestly lower levels of long-chain bases (sphingosine and sphinganine) in Grn brains than in control brains (Supplementary Fig. 1a). Because sphingosines are generated by degradation of more complex sphingolipids (Fig. 1a), their reduced levels suggest that degradation of sphingolipids is impaired in PGRN-deficient brains. However, the levels of hexosylceramides (glucosylceramide and galactosylceramide) were similar in mouse brains for all genotypes (Supplementary Fig. 1a). Also, levels of the phospholipids phosphatidylethanolamine (PE), phosphatidylcholine (PC), PS, and of neutral lipids were comparable among genotypes (Supplementary Fig. 1a). Similar to the findings in the brain, deficiency of PGRN in the kidney also led to elevated levels of gangliosides (Supplementary Fig. 1b and Supplementary Data 1). However, ganglioside levels in rodent peripheral tissues are 2–10% of those in the brain, so the amount of gangliosides that accumulate in this tissue is considerably smaller. To test whether PGRN deficiency’s impact on the brain lipidome was also present in patients with GRN-related FTD, we analyzed the lipid composition of postmortem human frontal and occipital lobe brain tissue from control (n = 3), sporadic FTD (sporadic FTD-TDP, Type A, (n = 6), and GRN mutation-related FTD (GRN FTD-TDP, Type A, (n = 12) subjects (The number of available healthy control samples was unfortunately limited and therefore a limitation of these analyses). As reported for PGRN-deficient fibroblasts and murine brain, we found increased levels of SEs and TAGs in the frontal lobes of GRN FTD-TDP subjects; in contrast, TAGs were unchanged, and SEs were below the detection limit in the occipital lobes of the same subjects (Supplementary Fig. 1c and Supplementary Data 2). Additionally, we found reductions in PE and cardiolipins (Supplementary Fig. 1c) and increases in sphingomyelin, particularly in the frontal lobes of the patients with GRN FTLD-TDP. Human brains are abundant in a variety of gangliosides, including GM1, GD1a/b, GD3, and GT1b. In a pattern that was similar to the changes in murine brain, we detected increased levels of mono-sialylated GM1 and di-sialylated GD3 and GD1 species in GRN FTD-TDP subjects (Fig. 1c and Supplementary Data 2). Some of these ganglioside species also trended higher in sporadic FTD-TDP subjects. The abundance of GT1, which can be catabolized at the plasma membrane, was lower in GRN FTD-TDP subjects and unchanged in sporadic FTD-TDP subjects (Supplementary Fig. 1c). In contrast to the findings in the frontal lobes, we detected no increase in the levels of gangliosides in the occipital lobes of either FTD group (Supplementary Fig. 1c and Supplementary Data 2). To address the mechanism of ganglioside accumulation, we established HeLa TMEM192-3xHA cell lines (lysosomal tagged, ref. 14) with PGRN deficiency and such cells with PGRN expression restored (GRN + PGRN-addback) via lentiviral transduction with untagged human PGRN cDNA (Fig. 2a and Supplementary Fig. 2a). In HeLa cells, the most abundant ganglioside class is GM2, and the gangliosides GM3, GD3, and GD1a are present in lesser amounts. Levels of GM2 were ~two-fold higher in PGRN-knockout cells than control cells, and normal levels were restored by PGRN-addback. GD1 and GD3 levels were unchanged in PGRN-knockout cells, but GD3 levels were reduced in the PGRN-addback cells (Fig. 2b and Supplementary Data 3). TAG levels were increased in PGRN-knockout cells and restored by PGRN-addback, whereas PC, PE, and PS levels were unaffected upon PGRN depletion. For unclear reasons, diacylglycerol (DAG) levels (Fig. 2c) and several sphingolipid catabolic products were unchanged in PGRN-knockout cells but were increased in PGRN-addback cells (Supplementary Fig. 2b and Supplementary Data 3). The levels of unesterified cholesterol and cholesterol esters were similar in control, PGRN-knockout, and PGRN-addback cells, in contrast to the increased cholesterol and reduced cholesterol esters in NPC1- or NPC2-knockout cell lines that are deficient in cholesterol export from lysosomes (Supplementary Fig. 2c).Fig. 2Lipidomic and immunofluorescence analyses of HeLa cells reveals GM2 accumulation upon loss of progranulin that is restored by PGRN-addback.a Western blot of full-length progranulin protein levels in GRN, GRN, and GRN + PGRN-addback HeLa cell lines. HeLa cells were gene-edited to contain TMEM192-3xHA. PCNA, proliferating cell nuclear antigen. b Quantification of gangliosides (GM2, GD3, GD1) and c quantification of phospholipids (PC, PE, PS) and neutral lipids (DAG, TAG, SE) isolated from GRN (green) (n = 7), GRN (orange) (n = 7), and GRN + PGRN-addback (blue) (n = 6) HeLa cell lines. d Representative confocal images of fixed HeLa cells stained with anti-GM2 antibody (magenta), anti-LAMP1 antibody (green) and Hoechst (blue). Scale bar, 50 μm. Bar graphs display number of GM2 puncta per cell and the Pearson’s correlation coefficient between GM2/LAMP1. The numbers of cells used to calculate the GM2 puncta per cell were n = 67, n = 69, n = 63, n = 42 for GRN, GRN, GRN + PGRN-addback, and HEXA, respectively. Box plots display mean ± the minimum and maximum number. One-way ANOVA, followed by multigroup comparison (Dunn’s) test, was performed. *p < 0.05 or ***p < 0.001. PC phosphatidylcholine, PE phosphatidylthanolamine, PS phosphatidylserine, DAG diacylglycerol, TAG triacylglycerol, SE sterol esters. a Western blot of full-length progranulin protein levels in GRN, GRN, and GRN + PGRN-addback HeLa cell lines. HeLa cells were gene-edited to contain TMEM192-3xHA. PCNA, proliferating cell nuclear antigen. b Quantification of gangliosides (GM2, GD3, GD1) and c quantification of phospholipids (PC, PE, PS) and neutral lipids (DAG, TAG, SE) isolated from GRN (green) (n = 7), GRN (orange) (n = 7), and GRN + PGRN-addback (blue) (n = 6) HeLa cell lines. d Representative confocal images of fixed HeLa cells stained with anti-GM2 antibody (magenta), anti-LAMP1 antibody (green) and Hoechst (blue). Scale bar, 50 μm. Bar graphs display number of GM2 puncta per cell and the Pearson’s correlation coefficient between GM2/LAMP1. The numbers of cells used to calculate the GM2 puncta per cell were n = 67, n = 69, n = 63, n = 42 for GRN, GRN, GRN + PGRN-addback, and HEXA, respectively. Box plots display mean ± the minimum and maximum number. One-way ANOVA, followed by multigroup comparison (Dunn’s) test, was performed. *p < 0.05 or ***p < 0.001. PC phosphatidylcholine, PE phosphatidylthanolamine, PS phosphatidylserine, DAG diacylglycerol, TAG triacylglycerol, SE sterol esters. Next, we used immunostaining of PGRN-knockout cells with an antibody that detects GM2. In agreement with the increase in GM2 gangliosides detected by lipidomic analyses, we found that PGRN-knockout cells had more GM2 puncta than control cells, although less GM2 puncta than in positive control HEXA-deficient cells (Fig. 2d). These GM2 puncta partially co-localized with the lysosomal marker LAMP1, and their accumulation was reversed by PGRN-addback (Fig. 2d). Gangliosides are catabolized by lysosomal enzymes, and deficiency of these enzymes in abundance or activity leads to lysosomal lipid accumulation. We, therefore, tested lysosomal function with a number of assays. Using quantitative whole-cell and lysosomal (isolated using Lyso-IP) tandem mass tag (TMT) proteomics, we found no major differences in the abundances of lysosomal proteins or glycosphingolipid-metabolizing enzymes in PGRN-knockout and PGRN-addback cells (Fig. 3a–d, Supplementary Fig. 3a, and Supplementary Data 4 and 5). Moreover, the activity of the glycosphingolipid catabolism enzyme β-hexosaminidase subunit α (HEXA) was unchanged in PGRN-knockout and PGRN-addback genotypes when incubated with artificial substrates (Fig. 3e). The activity of glucosylceramidase β (GCase), another glycosphingolipid catabolism enzyme, was more variable in PGRN-knockout cell lysate than in control, and the average trended lower (~15%). Most of the lysosomal proteome was unaffected in mouse brains deficient for PGRN, but we found a modest upregulation of several enzymes that mediate glycosphingolipid degradation (Supplementary Fig. 3b and Supplementary Data 4). Furthermore, we found normal enzymatic activity of HEXA in murine and human brain protein lysates in vitro. Similar to cells, the activity of GCase measured in mouse brain lysates derived from Grn and Grn genotypes trended lower than in control brain lysate (Supplementary Fig. 3c). Notably, human brain lysate of the frontal lobe had less GCase activity when incubated with artificial substrates (Supplementary Fig. 3c). In contrast, brain lysate of the occipital lobe showed no differences in GCase activity (Supplementary Fig. 3c).Fig. 3TMT-quantitative proteomic and in vitro analyses show no major differences in abundances or activities of glycosphingolipid metabolic enzymes in cells with PGRN depletion.a Volcano plot representation of whole-cell proteomes of GRN (left) and GRN + PGRN-addback (right) plotted against GRN with log2-fold-change (ratio of relative abundance, x-axis) and -log10 p-value (y-axis). All proteins (gray) and lysosomal proteins (black) quantified are shown. A corrected p-value < 0.05 (Welch’s test, two-sided) was used to calculate statistically significant differences between genotypes. b Heat map analysis of the relative abundance of a subset of proteins from whole-cell extracts that are involved in glycosphingolipid biosynthesis and degradation. c Western blotting analysis of proteins that are involved in glycosphingolid degradation (n = 3). d Heat-map analysis of the relative abundance of a subset of proteins that are involved in glycosphingolipid degradation from isolated lysosomal extracts. e HEXA and GCase activities were assessed in GRN (green), GRN (orange), and GRN + PGRN-addback (blue) cells when incubated with artificial substrates (n = 4 with three technical replicates each, mean ± SD). One-way ANOVA, followed by multigroup comparison (Dunn’s) test, was performed. *p < 0.05, **p < 0.01, or ***p < 0.001. f Electron micrographs from GRN and GRN cells with lysosomes and ILVs indicated by arrows. Scale bar, 500 nm. g Volcano plot representation of lysosomal proteomes of GRN plotted against GRN with log2-fold-change (ratio of relative abundance, x-axis) and −log10 p-value (y-axis). All proteins (gray), lysosomal proteins (black), and ILV-associated proteins (red) quantified are shown. A corrected p-value < 0.05 (Welch’s test, two-sided) was used to calculate statistically significant differences between genotypes. h, i Western blotting analysis of abundance of LAPTM4B in whole-cell extracts and mouse brains from different genotypes (n = 3). a Volcano plot representation of whole-cell proteomes of GRN (left) and GRN + PGRN-addback (right) plotted against GRN with log2-fold-change (ratio of relative abundance, x-axis) and -log10 p-value (y-axis). All proteins (gray) and lysosomal proteins (black) quantified are shown. A corrected p-value < 0.05 (Welch’s test, two-sided) was used to calculate statistically significant differences between genotypes. b Heat map analysis of the relative abundance of a subset of proteins from whole-cell extracts that are involved in glycosphingolipid biosynthesis and degradation. c Western blotting analysis of proteins that are involved in glycosphingolid degradation (n = 3). d Heat-map analysis of the relative abundance of a subset of proteins that are involved in glycosphingolipid degradation from isolated lysosomal extracts. e HEXA and GCase activities were assessed in GRN (green), GRN (orange), and GRN + PGRN-addback (blue) cells when incubated with artificial substrates (n = 4 with three technical replicates each, mean ± SD). One-way ANOVA, followed by multigroup comparison (Dunn’s) test, was performed. *p < 0.05, **p < 0.01, or ***p < 0.001. f Electron micrographs from GRN and GRN cells with lysosomes and ILVs indicated by arrows. Scale bar, 500 nm. g Volcano plot representation of lysosomal proteomes of GRN plotted against GRN with log2-fold-change (ratio of relative abundance, x-axis) and −log10 p-value (y-axis). All proteins (gray), lysosomal proteins (black), and ILV-associated proteins (red) quantified are shown. A corrected p-value < 0.05 (Welch’s test, two-sided) was used to calculate statistically significant differences between genotypes. h, i Western blotting analysis of abundance of LAPTM4B in whole-cell extracts and mouse brains from different genotypes (n = 3). Upon screening for other lysosome-mediated processes, we found no differences for mTOR signaling, autophagic flux, or phosphorylation of microphthalmia/transcription factor E (MiT/TFE) proteins in PGRN-deficient cells or tissues (Supplementary Fig. 3d, e). These results suggest that PGRN depletion has minimal effects on lysosomal composition and function in HeLa cells under basal conditions. Intralumenal vesicles (ILVs) are sites of lipid degradation within lysosomes. Electron microscopy used to analyze the ultrastructure of lysosomes revealed ILVs were present in control and PGRN-deficient cells (Fig. 3f). However, lysosomes isolated from PGRN-deficient cells by Lyso-IP had fewer ILV-associated proteins, including LAPTM4A, LAPTM4B, SDCBP, CD81, and CD9 (Fig. 3g). Further, the modest reduction of LAPTM4B in PGRN-knockout cells was reversed by the expression of exogenous PGRN (Fig. 3h). Reduced levels of LAPTM4B were also found in brains of Grn knockout mice (Fig. 3i). Because the levels of the enzymes that catabolize gangliosides were not changed in PGRN-deficient cells or tissues, and lysosomes appeared mostly intact and functional, we searched for another cause of gangliosidosis. A recent study found that bis(monoacylglycero)phosphate (BMP) levels are reduced in Grn-deficient mouse brains. BMP is crucial in glycosphingolipid and ganglioside degradation in lysosomes, and its levels are altered in many lysosomal storage diseases. BMP is found in ILVs where its negatively charged phosphate headgroup is thought to enable binding of lysosomal hydrolases. We hypothesized that reduced BMP levels underlie the gangliosidosis we found in progranulin deficiency. We measured BMP levels in PGRN-knockout HeLa cells and found it ~50% reduced, whereas levels of the BMP isomer (and presumptive precursor) PG were unchanged. The reductions in BMP levels were restored in PGRN-addback cells (Fig. 4a and Supplementary Data 4). We also examined BMP accumulation in the HeLa cell model system with radioactive tracers. Metabolic labeling studies utilizing C-arachidonic acid confirmed the reduction of BMP levels in PGRN-knockout HeLa cells (Fig. 4b), suggesting alterations in the synthesis or degradation of BMP with polyunsaturated fatty acids in PGRN deficiency. Similarly, brains of PGRN-deficient mice showed a 50–60% reduction in BMP levels, and all detected BMP species were significantly reduced in Grn brains (Fig. 4c, Supplementary Fig. 4a, and Supplementary Data 2).Fig. 4BMP levels are reduced in progranulin-deficient cells or brain tissues.a Quantification of PG and BMP isolated from GRN (green), GRN (orange), and GRN + PGRN-addback (blue) HeLa cell lines (n = 7) reveals a ~50% reduction in BMP levels in PGRN-knockout cells. b Labeling of cellular lipids by feeding C-arachidonic acid–albumin complex to GRN and GRN HeLa cell lines (60 min). Inset highlights reduced levels of BMP in Grn HeLa cells, compared to control cells. c Quantification of PG isolated from Grn (gray) (n = 8), Grn (blue) (n = 8), and Grn (purple) (n = 8) and quantification of BMP isolated from Grn (gray) (n = 8), Grn (blue) (n = 6), and Grn (purple) (n = 6) mouse brains reveals a decrease in BMP levels upon loss of PGRN in mouse brains (~50%). d Quantification of PG and BMP isolated from the frontal lobes of control (pink) (n = 3), FTD-TDP43-A(sporadic-non-GRN) (green) (n = 6), and FTD-TDP43-A(GRN) (blue) (n = 11) human brains. BMP species with mono- or di-unsaturated fatty acid moieties are not different, whereas BMP species containing two docosahexanoic acid moieties (22:6/22:6) are reduced in the frontal and occipital lobes of all FTD subjects. e Quantification of GM2 species isolated from Grn (green) (n = 6) and Grn (orange) (n = 6) HeLa cells after feeding no lipids, di-oleoyl PC or di-oleoyl BMP. f Model of the role of progranulin in the degradation of gangliosides. PGRN/granulin deficiency leads to reduced BMP levels through unclear mechanisms. Reduced BMP levels contribute to impaired ganglioside degradation. Eventually, this leads to lysosomal dysfunction and downstream consequences, including neuroinflammation and neurodegeneration. Box plots display mean ± the minimum and maximum number. One-way ANOVA, followed by multigroup comparison (Dunn’s) test, was performed. *p < 0.05, **p < 0.01. a Quantification of PG and BMP isolated from GRN (green), GRN (orange), and GRN + PGRN-addback (blue) HeLa cell lines (n = 7) reveals a ~50% reduction in BMP levels in PGRN-knockout cells. b Labeling of cellular lipids by feeding C-arachidonic acid–albumin complex to GRN and GRN HeLa cell lines (60 min). Inset highlights reduced levels of BMP in Grn HeLa cells, compared to control cells. c Quantification of PG isolated from Grn (gray) (n = 8), Grn (blue) (n = 8), and Grn (purple) (n = 8) and quantification of BMP isolated from Grn (gray) (n = 8), Grn (blue) (n = 6), and Grn (purple) (n = 6) mouse brains reveals a decrease in BMP levels upon loss of PGRN in mouse brains (~50%). d Quantification of PG and BMP isolated from the frontal lobes of control (pink) (n = 3), FTD-TDP43-A(sporadic-non-GRN) (green) (n = 6), and FTD-TDP43-A(GRN) (blue) (n = 11) human brains. BMP species with mono- or di-unsaturated fatty acid moieties are not different, whereas BMP species containing two docosahexanoic acid moieties (22:6/22:6) are reduced in the frontal and occipital lobes of all FTD subjects. e Quantification of GM2 species isolated from Grn (green) (n = 6) and Grn (orange) (n = 6) HeLa cells after feeding no lipids, di-oleoyl PC or di-oleoyl BMP. f Model of the role of progranulin in the degradation of gangliosides. PGRN/granulin deficiency leads to reduced BMP levels through unclear mechanisms. Reduced BMP levels contribute to impaired ganglioside degradation. Eventually, this leads to lysosomal dysfunction and downstream consequences, including neuroinflammation and neurodegeneration. Box plots display mean ± the minimum and maximum number. One-way ANOVA, followed by multigroup comparison (Dunn’s) test, was performed. *p < 0.05, **p < 0.01. We also measured BMP levels in brain tissues from patients with GRN FTD-TDP and compared them with sporadic FTD-TDP and control subjects. We found marked decreases in a BMP species containing two docosahexanoic acid moieties (22:6/22:6) in the frontal and occipital lobes of all FTD subjects (Fig. 4d, Supplementary Fig. 4b, and Supplementary Data 1). BMP with two 22:6 polyunsaturated fatty acids is one of the most abundant BMP species in human brain. BMP species with mono- or di-unsaturated fatty acid moieties trended lower in some samples but were not overall different among the study groups. To further test if BMP deficiency is responsible for ganglioside accumulation in PGRN-knockout cells, we assayed whether adding exogenous BMP to HeLa cells lacking PGRN could normalize ganglioside levels. We incubated cells with di-oleoyl-PC or di-oleoyl-BMP and measured ganglioside levels by lipidomics. BMP but not PC was sufficient to normalize GM2 ganglioside levels in PGRN-knockout cells to levels similar to those in control cells (Fig. 4e). This finding was confirmed by feeding BMP to cells and measuring the number of GM2 puncta by immunofluoresence (Supplementary Fig. 4c, d). These experiments establish a causal link between BMP deficiency and ganglioside accumulation in PGRN-deficient cells. Our experiments lead to a model for PGRN function in which lysosomal granulin peptides maintain lysosomal function and homeostasis, including the levels of BMP, that are crucial for ganglioside catabolism (Fig. 4f). In the setting of PGRN deficiency, reduced granulins in lysosomes lead to reduced BMP levels on intralumenal vesicles through an unclear mechanism. Low BMP levels, in turn, impair ganglioside catabolism and result in gangliosidosis. The resulting gangliosidosis may also compromise certain lysosomal functions, such as the response to membrane perturbation and eventually lead to other hallmarks of the disease (e.g., protein aggregation of TMEM106B, TDP-43). In agreement with an important role of lysosomal degradation in FTD, mutation of CHMP2B, a component of the ESCRT machinery of ILV formation, also leads to FTD. Our model predicts that other lipids, such as neutral lipids or other sphingolipids, may also be poorly catabolized due to changes in lysosomal lipid degradation. Indeed, glucosylsphingosine, whose degradation is thought to require BMP, was reported to be increased in PGRN-deficient mice. It is currently unclear how PGRN deficiency leads to low BMP levels. One possibility is that granulins may interact directly with BMP to modify its levels. A recent study reported that His-tagged PGRN bound to liposomes containing BMP, suggesting it may directly influence its abundance. We reproduced this finding for full-length progranulin in our binding assay (Supplementary Fig. 4e), but believe that further investigation for binding of native, untagged and functional granulin peptides is needed to understand this potential interaction and its implications. An alternative is that granulins indirectly affect BMP levels, for example by altering the ion environment of lysosomes or by changing synthesis or degradation rates of BMP through pathways that are currently not understood. Our results are consistent with several reports linking PGRN deficiency to altered glycosphingolipid catabolism. Specifically, PGRN has been linked to the function of the HEXA enzyme, and GCase activity is compromised in PGRN-deficient mice and neurons. Although we found that HEXA activity was not changed, GCase activity was decreased in the frontal lobes (but not the occipital lobes) of GRN FTD-TDP patients. Finally, and of particular relevance to our findings of PGRN deficiency causing gangliosidosis, a recent report found reduced BMP levels in brains of PGRN-deficient mice. Despite the widespread expression of PGRN in tissues, PGRN deficiency results primarily in a disease of the central nervous system (CNS). Our findings suggest that this may be because gangliosides are found at much higher levels in the CNS than in other tissues. In the brain, neuronal gangliosidosis may trigger the activation and recruitment of microglia to phagocytose and process the excess gangliosides. In particular, GM2 gangliosides may incite TNF-α expression and inflammation in monocyte-derived cells. A hallmark of PGRN deficiency in murine brain is microgliosis and neuroinflammation, and microglial cells (and macrophages) are hyperactivated in the setting of PGRN deficiency. The accumulated effects of long-term defects in lysosomal ganglioside metabolism and neuroinflammation may, therefore, contribute to GRN FTD-TDP. With respect to therapeutic implications, increased ganglioside levels or reduced BMP levels may serve as biomarkers for PGRN-deficient FTD or other neurodegenerative disorders. It may also be of benefit to determine if drugs that lower ganglioside production are beneficial in GRN-FTD-TDP. Finally, it will be of interest to analyze ganglioside levels in other chronic adult neurodegenerative diseases. The following reagents were purchased from commercial vendors: acetonitrile, methanol, water (all HPLC/MS grade), chloroform (HPLC grade, stabilized by 0.5–1% ethanol), ammonium formate, ammonium acetate, formic acid, and acetic acid were purchased from Sigma-Aldrich. N-Omega-CD3-octdecanoyl GM1 (Matreya LLC, Cat#2050), SPLASH® LIPIDOMIX® Mass Spec Standard (Avanti Polar Lipids, Cat# 330707-1EA). TMTpro was purchased from Thermo Fisher (Cat#A44520). Primary antibodies against the following targets were used in the present study: GM2 monoclonal antibody (clone MK1-16) (TCI America, Cat#A2576), human progranulin (R&D systems, Cat#AF2420), anti-mouse progranulin polyclonal antibody that recognizes an epitope at amino acids 198–214, PCNA (Santa Cruz Biotechnology Cat#sc56), GALC (Proteintech Cat#11991-1-AP), GLA (Proteintech Cat#66121-1-Ig), GM2A (Proteintech Cat#10864-2-AP), NEU1 (Santa Cruz Biotechnology Cat#sc166824), PSAP (Proteintech Cat#10801-1-AP), GLB1 (Proteintech Cat#15518-1-AP), HEXA (Proteintech Cat#11317-1-AP), HSP90 (Proteintech Cat#60318-1-Ig), TFEB (Cell Signaling Technology Cat#4240), TFE3 (Proteintech Cat#14480-1-AP), SQSTM1 (Proteintech Cat#18420-1-AP), CALCOCO2 (Proteintech Cat#12229-1-AP), MAP1LC3B (Cell Signaling Technology Cat#2775), GABARAP (Proteintech Cat#18723-1-AP), MTOR (Cell Signaling Technology Cat#2983), MTOR (S2448) (Cell Signaling Technology Cat#2971), P70S6K (Cell Signaling Technology Cat#2708), P70S6K (T389) (Cell Signaling Technology Cat#9234), ULK1 (Cell Signaling Technology Cat#8054), ULK1 (S757) (Cell Signaling Technology Cat#14202), ASAH1 (Proteintech Cat#11274-1-AP), HEXB (Proteintech Cat#16229-1-AP) and BETA-ACTIN (Santa Cruz Biotechnology Cat#69879). The entry vector pDONR223 containing the full-length GRN (1–1179 base pairs) from the human orfeome collection was used to engineer a stop codon by site-directed mutagenesis (New England Biolabs) at the end of the open-reading frame sequence (ORF). Gateway technology (Thermo Fisher) was used to transfer the GRN ORF with LR cloning from the entry vector to the pHAGE lentiviral destination expression vector. sgRNA sequences for editing the TMEM192, GRN, HEXA, NPC1 and NPC2 loci were cloned into the pX459 V2.0 vector (Addgene Cat#62988) as described. CRSIPR/Cas9-mediated gene editing of HeLa cells (ATCC Cat# CCL-2) was performed as described (10.17504/protocols.io.4r3l2oxqqv1y/v1). The following sgRNA sequences were used: TMEM192 (5′-AGTAGAACGTGAGAGGCTCA-3′) GRN (5′-ATCGACCATAACACAGCACG-3′) HEXA (5′-CGGCCGAGCTGACATCGTAC-3′) NPC1 (5′-TACCTGGACAGAAACTGTAG-3′) NPC2 (5′-AGCTGCCAGGAAACGCATCG-3′) To engineer the lyso-IP tag into HeLa cells by homology-directed repair, a gene block encoding a 3xHA epitope tag, a puromycin cassette, and homology arms on either side of the sgRNA cleavage site was synthesized by Integrated DNA Technologies to edit the TMEM192 locus, similar to what was reported. This sequence was cloned into the pSmart (Lucigen Cat#40041-2) donor vector using Gibson assembly (New England Biolabs). The donor vector along with the TMEM192 sgRNA sequence was transiently transfected into HeLa cells and puromycin selection was performed 5 days post-transfection for 7–8 days. The mixed pool of cells that were puromycin resistant were plated as single cells, and clonal lines of homozygous HeLa TMEM192-3xHA were isolated. The HeLa TMEM192-3xHA cell line was used for engineering all subsequent gene deletions. The lentiviral vector was packaged in HEK293T (ATCC Cat#CRL-3216) by co-transfection of psPAX2, pMD2.G (Addgene Cat#12260 Cat#12259) and pHAGE-GRN in a 4:2:1 ratio using polyethyleneimine. Virus-containing supernatant was harvested 2 days after transfection and filtered through a 0.45-micron syringe filter. Polybrene was added to a final concentration of 8 μg/ml to the viral supernatant. HeLa Tmem192-3xHA GRN KO cells were infected with 50 μL of viral supernatant, and stable cell lines were selected 48 h post-infection using hygromycin at a concentration of 100 μg/mL. HeLa TMEM192-3xHA and HEK293T cells were grown at 37 °C in Dulbecco’ Modified Eagles Medium (DMEM) (Invitrogen Cat#11995-073), supplemented with 10% fetal bovine serum (FBS) (HyClone Cat#SH30910.03) and 1% penicillin-streptomycin (Thermo Fisher Cat#15140163). Cells were plated on to 24-well glass-bottom dish (Cellvis Cat#P24-1.5H-N). All immunofluorescence experiments were performed at room temperature. Cells at 70% confluence were washed twice in PBS and fixed with 4% paraformaldehyde in PBS for 20 min. Cells were solubilized in 0.02% Saponin detergent in PBS for 15 min and then blocked with 2% BSA in 0.02% Saponin-PBS (blocking buffer) for 30 min. For the GM2-LAMP1 co-stain experiment, the cells were fixed and blocked, but not treated with detergent. Immunostaining for 2 h was performed with the following primary antibodies (1:100 dilution in blocking buffer): GM2 (Tokyo Chemical Industry Cat#2576), LAMP1 (Cell Signaling Technology Cat#9091), and HA-tag (Cell Signaling Technology Cat#3724). Cells were washed 3×5 min with PBS. Cells were incubated with the appropriate Alexa Fluor-conjugated secondary antibodies (Thermo Fisher Cat#A32731 Cat#A21203) (1:400 dilution in blocking buffer) for 1 h. Cells were washed 3 × 5 min with PBS. Cells were stained with Hoechst (1 μg/mL in blocking buffer) for 5 min. Cells were washed 3 × 5 min with PBS. Cells were imaged using a Yokogawa CSU-X1 spinning-disk confocal on a Nikon Ti-E inverted microscope at the Nikon Imaging Center in Harvard Medical School. The microscope is equipped with a Nikon Plan Apo 40x/1.30 NA objective lens and 445 nm (75 mW), 488 nm (100 mW), 561 nm (100 mW), and 642 nm (100 mW) laser lines controlled by AOTF. All images were collected with a Hamamatsu ORCA-ER cooled CCD camera (6.45 µm photodiode) with MetaMorph image acquisition software. Z series are displayed as maximum z-projections and brightness and contrast were adjusted for each image equally and then converted to RGB. Image analysis was performed using Fiji. General protocols for western blotting performed here can be found at: 10.17504/protocols.io.kxygxzr94v8j/v1. Cell pellets or mouse tissues were resuspended in ice-cold 8 M urea buffer (8 M urea, 50 mM Tris pH 7.4, 50 mM NaCl) supplemented with protease and phosphatase inhibitors (Roche). The resuspended samples were sonicated, and the lysates were clarified at 17000 x g for 10 min at 4 °C. A Bradford assay was performed, and equal amounts of lysate were boiled in LDS-Laemmli buffer supplemented with 50 mM DTT for 10 min at 95 °C. Lysates were run on 4–20% Tris glycine gels (BioRad) and transferred on to PVDF membranes (Millipore), which were blocked for 1 h at room temperature in 2% BSA-0.1% TBS-tween (blocking buffer). Immunoblotting with primary antibodies was performed overnight at 4 °C (1:1000 dilution in blocking buffer). Immunoblots were incubated with the appropriate secondary antibodies (1:5000 dilution in blocking buffer) for 1 h at room temperature. Images of blots were acquired using Enhanced-Chemi luminescence on a BioRad ChemiDoc imager. Cells pellets were harvested, washed twice in ice-cold PBS, and resuspended either in ice-cold neutral lysis buffer (50 mM Tris pH 7.4, 150 mM NaCl, 0.5% NP-40, pH 7.4) or acidic lysis buffer (50 mM NaOAc pH 5.3, 150 mM NaCl, 0.5% NP-40, pH 5.2) supplemented with protease and phosphatase inhibitors (Roche). All subsequent steps were performed at 4 °C. Resuspended cells were incubated on a rotator for 30 min, centrifuged at 17,000 × g for 10 min, and the supernatant was collected. In all, 1 mg of supernatant from each sample was incubated with 20 μL of BMP-beads, LPA-beads, or control beads (Echelon Biosciences Cat#P-BLBP-2 Cat#L-6101 Cat#P-B000) for 3 h on a rotator. The beads were pelleted by centrifugation at 500 × g for 1 min, the supernatant aspirated, and washed twice with either neutral or acidic lysis buffer. The beads were boiled in LDS-Laemmli buffer supplemented with 50 mM DTT for 10 min at 95 °C, and western blotting was performed as described earlier. Cells were plated on to six-well dish to 70% confluence. Next, medium was changed to DMEM supplemented with 10 μM 18:1/18:1-PC or 18:1/18:1-BMP. Lipid-supplemented medium was prepared by drying the lipids in a glass vial under N2 stream, and dried lipids were resuspended in complete DMEM medium using water bath sanctum (~30 min) to get the final lipid concentration to 10 μM lipids in the medium. Cells were maintained in lipid-supplemented medium for 24 h, and medium was changed every 6 h to ensure lipid supplementation. Cells were washed with cold PBS and processed for lipid extraction or fixed with 4% (w/v) of formaldehyde in PBS for anti-GM2 ganglioside immunofluorescence assay. Electron microscopy imaging was performed in the Harvard Medical School Electron Microscopy Facility. Fixative solution containing 2.5% glutaraldehyde, 1.25% paraformaldehyde, 0.03% picric acid in 0.1 M sodium cacodylate buffer, pH 7.4, was added in a 1:1 ratio to cells grown to 70% confluency for 1 h at room temperature. The cells were then postfixed for 30 min in 1% osmium tetroxide (OsO4)/1.5% potassium ferrocyanide (KFeCN6), washed in water 3x, and incubated in 1% aqueous uranyl acetate for 30 min, followed by two washes in water and subsequent dehydration in grades of alcohol (5 min each; 50, 70, 95%, 2 × 100%). Cells were removed from the dish in propyleneoxide, pelleted at 500×g for 3 min, and infiltrated for 2 h to overnight in a 1:1 mixture of propyleneoxide and TAAB Epon (Marivac Canada Inc., St. Laurent, Canada). The samples subsequently embedded in TAAB Epon and polymerized at 60 °C for 48 h. Ultrathin sections (~60 nm) were cut on a Reichert Ultracut-S microtome, picked up on to copper grids stained with lead citrate and examined in a JEOL 1200EX Transmission electron microscope or a TecnaiG² Spirit BioTWIN and images were recorded with an AMT 2k CCD camera. Lysosome immunoprecipitation was carried out as described with a few modifications. All steps of the process were carried out at 4 °C with cold solutions. Briefly, HeLa TMEM192-3xHA endogenously tagged cells that were grown to 80% confluency in 150-mm plates were washed twice with PBS, scraped into tubes, and then pelleted at 500×g for 5 min. The cells were re-suspended in 2 mL of lysoIP buffer (50 mM KCl, 100 mM KH2PO4, 100 mM K2HPO4, pH 7.2) supplemented with protease and phosphatase inhibitors (Roche). The cells were transferred to a glass homogenizer and dounced using 25 strokes. The lysed cells were centrifuged at 1000×g for 10 min, and the post-nuclear supernatant (PNS) was collected. The concentration of the PNS was determined by Bradford assay. Lysosomes from normalized amounts of PNS were immunoprecipitated by incubation with 50 μL of magnetic HA-beads (Thermo Scientific Cat#88837) for 1 h on a rotator. The beads were sequentially washed once with lysoIP buffer containing 300 mM NaCl and once with lysoIP buffer. The lysosomes were solubilized and eluted off the beads by incubating the beads with 150 μL of lysoIP buffer with 0.5% NP-40 in a thermomixer (1000 rpm) for 30 min. The eluates were snap frozen in liquid nitrogen and stored in −80 °C. Animal procedures were approved by the Institutional Animal Care and Use Committee of the Harvard Medical Area Standing Committee on Animals and followed NIH guidelines. All mouse experiments were performed under the oversight and ethical guidelines from the Harvard Center for Comparative Medicine. Mice were housed in a pathogen-free barrier facility with a 12 h light/12 h dark cycle and allowed food and water ad libitum. Grn mice and Grn mice were on the C57BL/6 J background (backcrossed more than eight generations). Mice used in this study were aged 18–20 months and of both sexes. Postmortem brain samples were provided by the University of California, San Francisco (UCSF) Neurodegenerative Disease Brain Bank. Brains were donated with the consent of the participants or their surrogates in accordance with the Declaration of Helsinki, and the research was approved by the University of California, San Francisco Committee on Human Research. Tissue blocks were dissected from the middle frontal gyrus and lateral occipital cortex of three controls, as well as six patients with sporadic FTLD-TDP and 13 with GRN-FTLD-TDP. All patients with FTD‐GRN carried a pathogenic variant in GRN and had FTLD‐TDP, Type A, identified at autopsy (Table see below). Clinical and neuropathological diagnoses were made using standard diagnostic criteria. Characteristics of each group were as follows:Median age (years)Interquartile ranges (years)Control (n = 3)8611Sporadic FTD-TDP (n = 6)715GRN FTD-TDP (n = 13)682.5 Patients in the GRN-FTLD and sporadic FTLD groups showed a range of FTD syndromes (bvFTD, nfvPPA, etc.) and all showed FTLD-TDP Type A pathology (with one showing advanced comorbid AD). HeLa cells were grown in a 10-cm culture dish until they reached ~80% confluence. Tissue samples were obtained as described, and mouse whole cortex brains were collected and immediately snap frozen in liquid nitrogen. Cell and tissue homogenates were obtained by snap-freezing (in liquid nitrogen)/thawing (using an ultrasound water bath for 3 min) repeatedly, and finally extracted according to Folch’s method. Internal standard SPLAH mix and deuterated ganglioside standard spiked in prior to extraction were used for normalization. The organic phase of each cell-culture sample was normalized by total soluble protein amounts and measured by BCA assay (Thermo Scientific, 23225, Waltham, MA), whereas tissue samples were normalized according to dry weight measurements. Samples were routinely subjected to two rounds of extraction. The HPLC-mass spectroscopy (MS) method was adopted from. Briefly, HPLC analysis of the organic phases was performed employing a C30 reverse-phase column (Thermo Fisher Scientific, Acclaim C30, 2.1 × 250 mm, 3 μm, operated at 55 °C; Bremen, Germany) connected to a Dionex UltiMate 3000 HPLC system and a QExactive orbitrap mass spectrometer (Thermo Fisher Scientific, Bremen, Germany) equipped with a heated electrospray ionization (HESI) probe. Dried lipid samples were dissolved in appropriate volumes of 2:1 MeOH:CHCl3 (v/v), and 5 μL of each sample was injected, with separate injections for positive and negative ionization modes. Mobile phase A consisted of 60:40 can:H2O, including 10 mM ammonium formate and 0.1% formic acid, and mobile phase B consisted of 90:10 2-propanol:ACN, also including 10 mM ammonium formate and 0.1% formic acid. The elution was performed with a gradient of 90 min; for 0–7 min, elution started with 40% B and increased to 55%; from 7 to 8 min, increased to 65% B; from 8 to 12 min, elution was maintained with 65% B; from 12 to 30 min, increased to 70% B; from 30 to 31 min, increased to 88% B; from 31 to 51 min, increased to 95% B; from 51 to 53 min, increased to 100% B; during 53 to 73 min, 100% B was maintained; from 73 to 73.1 min, solvent B was decreased to 40% and maintained for another 16.9 min for column re-equilibration. The flow-rate was set to 0.2 mL/min. The column oven temperature was set to 55 °C, and the temperature of the autosampler tray was set to 4 °C. The spray voltage was set to 4.2 kV, and the heated capillary and the HESI were held at 320 °C and 300 °C, respectively. The S-lens RF level was set to 50, and the sheath and auxiliary gas were set to 35 and 3 units, respectively. These conditions were held constant for both positive and negative ionization mode acquisitions. External mass calibration was performed using the standard calibration mixture every 7 days. MS spectra of lipids were acquired in full-scan/data-dependent MS2 mode. For the full-scan acquisition, the resolution was set to 70,000, the AGC target was 1e6, the maximum injection time was 50 msec, and the scan range was m/z = 133.4–2000. For data-dependent MS2, the top 10 ions in each full scan were isolated with a 1.0 Da window, fragmented at a stepped normalized collision energy of 15, 25, and 35 units, and analyzed at a resolution of 17,500 with an AGC target of 2e5 and a maximum injection time of 100 msec. The underfill ratio was set to 0. The selection of the top 10 ions was subject to isotopic exclusion with a dynamic exclusion window of 5.0 sec. Processing of raw data was performed using LipidSearch software (Thermo Fisher Scientific/Mitsui Knowledge Industries). The aqueous phase was desalted by applying to a C18 cartridge (Waters) equilibrated with 2:43:55 (chloroform:methanol:water) and eluted with 1:1 CHCl3:MeOH. The eluates were dried down again and resuspended in chloroform:methanol:water (600:425:75, v/v/v). The HILIC-MS method was adopted from. HPLC analysis was performed employing a Phenomenex (Thermo Fisher Scientific, CAT#, 2.0 × 150 mm, operated at 60 °C; Bremen, Germany). Dried lipid samples were dissolved in appropriate volumes of 2:1 MeOH: CHCl3 (v/v) and 5 μL of each sample was injected and acquired in negative ionization mode. Mobile phase A consisted of acetonitrile with 0.2% formic acid and mobile phase B consisted of 10 mM aqueous ammonium acetate, pH 6.1, adjusted with formic acid. Column equilibration was performed using 12.3% B for 5 min prior to each run. Chromatographic condition: mobile-phase gradient as follows: 0 min: 87.7% A + 12.3% B; and 15 min: 77.9% A + 22.1% B. The re-equilibration time between runs was 5 mins. The flow rate for the separation was set to 0.6 mL/min. The column oven temperature was set to 40 °C, and the temperature of the autosampler tray was set to 4 °C. The spray voltage was set to −4.5 kV, and the heated capillary and the HESI were held at 300 °C and 250 °C, respectively. The S-lens RF level was set to 50, and the sheath and auxiliary gas were set to 40 and 5 units, respectively. These conditions were held constant during the acquisitions. External mass calibration was performed using the standard calibration mixture every 7 days. MS spectra of lipids were acquired in full-scan/data-dependent MS2 mode. For the full-scan acquisition, the resolution was set to 70,000, the AGC target was 1e6, the maximum injection time was 50 msec, and the scan range was m/z = 700–2500 in the negative ion mode. For data-dependent MS2, the top 10 ions in each full scan were isolated with a 1.0 Da window, fragmented at a stepped normalized collision energy of 25, 35, and 50 units, and analyzed at a resolution of 17,500 with an AGC target of 2e5 and a maximum injection time of 100 msec. The underfill ratio was set to 0. The selection of the top 10 ions was subject to isotopic exclusion with a dynamic exclusion window of 5.0 sec. Processing of raw data was performed in Xcalibur™ software (Thermo Fisher Scientific). MS-based lipid analysis was performed by Lipotype GmbH (Dresden, Germany) as described. If not indicated otherwise, 500 µg of tissue were used per extraction. Lipids were extracted using a two-step chloroform/methanol procedure. Samples were spiked with internal lipid standard mixture containing: cardiolipin 16:1/15:0/15:0/15:0 (CL, 50 pmol per extraction), ceramide 18:1;2/17:0 (Cer, 30 pmol), diacylglycerol 17:0/17:0 (DAG, 100 pmol), hexosylceramide 18:1;2/12:0 (HexCer, 30 pmol), lyso-phosphatidate 17:0 (LPA, 30 pmol), lyso-phosphatidylcholine 12:0 (LPC, 50 pmol), lyso-phosphatidylethanolamine 17:1 (LPE, 30 pmol), lyso-phosphatidylglycerol 17:1 (LPG, 30 pmol), lyso-phosphatidylinositol 17:1 (LPI, 20 pmol), lyso-phosphatidylserine 17:1 (LPS, 30 pmol), phosphatidate 17:0/17:0 (PA, 50 pmol), phosphatidylcholine 17:0/17:0 (PC, 150 pmol), phosphatidylethanolamine 17:0/17:0 (PE, 75 pmol), phosphatidylglycerol 17:0/17:0 (PG, 50 pmol), phosphatidylinositol 16:0/16:0 (PI, 50 pmol), phosphatidylserine 17:0/17:0 (PS, 100 pmol), cholesterol ester 20:0 (CE, 100 pmol), sphingomyelin 18:1;2/12:0;0 (SM, 50 pmol), triacylglycerol 17:0/17:0/17:0 (TAG, 75 pmol), GM1-D3 18:1;2/18:0;0 (200 pmol), and cholesterol D6 (Chol, 300 pmol). After extraction, the organic phase was transferred to an infusion plate and dried in a speed vacuum concentrator. First, dry extract was re-suspended in 7.5 mM ammonium acetate in chloroform/methanol/propanol (1:2:4, V:V:V), and second, dry extract in 33% ethanol solution of methylamine in chloroform/methanol (0.003:5:1; V:V:V). All liquid handling steps were performed using Hamilton Robotics STARlet robotic platform with the Anti Droplet Control feature for organic solvents pipetting. Samples were analyzed by direct infusion on a QExactive mass spectrometer (Thermo Scientific), equipped with a TriVersa NanoMate ion source (Advion Biosciences). Samples were analyzed in both positive and negative ion modes with a resolution of Rm/z = 200 = 280,000 for MS and Rm/z = 200 = 17,500 for MS/MS experiments, in a single acquisition. MS/MS was triggered by an inclusion list encompassing corresponding MS mass ranges scanned in 1 Da increments. Both MS and MSMS data were combined to monitor CE, DAG and TAG ions as ammonium adducts; PC, PC, and O-, as acetate adducts; and CL, PA, PE, PE O-, PG, PI, and PS as deprotonated anions. MS only was used to monitor LPA, LPE, LPE O-, LPI, LPS, and GM2 as deprotonated anions; Cer, HexCer, SM, LPC, and LPC O-as acetate adducts; and cholesterol as an ammonium adduct of an acetylated derivative. Ganglioside classes GM1, GD1, GD2, GD3, GT1, GT2, GT3, and GQ1 were extracted and analyzed as follows. Gangliosides in the remaining water phase of the two-step chloroform:methanol procedure were subjected to purification using solid-phase extraction (Thermo Scientific SOLA SPE plates, 10 mg/2 mL). The water phase was loaded on columns pre-washed with chloroform:methanol (2:1, V:V), methanol and methanol:water (1:1, V:V); with the input flow through re-applied three times. Then, columns were washed with water, and the elution was carried out two times with methanol and one time with chloroform:methanol (1:1, V:V). Washing and elution steps were carried using a vacuum manifold. Pooled eluates were dried in a speed vacuum concentrator and re-suspended in 33% ethanol solution of methylamine in chloroform:methanol (0.003:5:1; V:V:V). Ganglioside extracts were analyzed by direct infusion on a QExactive MS (Thermo Scientific) equipped with a TriVersa NanoMate ion source (Advion Biosciences). Samples were analyzed in negative ion modes with a resolution of Rm/z = 200 = 140,000; AGC target of 1e6; maximum injection time of 500 ms and 3 microscans. Data were analyzed with in-house developed lipid identification software, based on LipidXplorer. Data post-processing and normalization were performed using an in-house developed data management system. Only lipid identifications with a signal-to-noise ratio >5, and a signal intensity five-fold higher than in corresponding blank samples were considered for further data analysis. Lipids were normalized to lipid class–specific internal standards. In case of ganglioside classes for which no suitable lipid class–specific internal standards are available, spectral intensities were normalized to the internal standard GM1-D3 18:1;2/18:0;0, and the normalized intensities further normalized to total lipid content (in pmol) of the sample. Only results with >1.5-fold-change and a p-value < 0.05 were selected from the web browser-based data visualization tool (LipotypeZoom) for levels of lipid species as the percentage of the total lipid amount for each sample. Protein estimation of the lysate was done by BCA assay. The activities of β-hexosaminidase and glucosylceramidase β were measured by a fluorometric method using 4-methylumbelliferyl N-acetyl-β-D-glucosaminide (Sigma, Cat#69585) and 4-methylumbelliferyl β-d-glucopyranoside (Sigma, Cat#M3633) as substrates, respectively. The reaction mixture consisted of either of the substrates (5 mM), enzyme fraction, and 100 mM citrate buffer, pH 5.0, in a final volume of 200 µL were incubated at 45 °C. After 1 h, the reaction was stopped with 2.5 mL of 0.5 M Na2CO3 buffer, pH 10.4. The 4-methylumbelliferone released was measured in a Tecan Spark multimode microplate reader with excitation and emission set at 350 and 440 nm, respectively. Detailed methods for proteomic analysis are provided at 10.17504/protocols.io.ewov14b7kvr2/v3. In all, 50 μg of protein extracts from cell pellets or mouse tissue were subjected to disulfide bond reduction with 5 mM TCEP (room temperature, 10 min) and alkylation with 25 mM chloroacetamide (room temperature, 20 min), followed by methanol–chloroform precipitation. For lysosomal samples, disulfide bond reduction and alkylation were performed using 20 μg of extracts, followed by trichloroacetic acid precipitation. Samples were resuspended in 50 μL of 200 mM EPPS, pH 8.5, and digested at 37 °C for 2 h with LysC protease at a 200:1 protein-to-protease ratio. Trypsin was then added at a 100:1 protein-to-protease ratio, and the reaction was incubated for 6 h at 37 °C. Tandem mass tag labeling of each sample was performed by adding 10 μL each of the 20 ng/μL stock of TMTpro reagent along with acetonitrile to achieve a final acetonitrile concentration of approximately 30% (v/v). After incubation at room temperature for 1 h, the labeling efficiency of a small aliquot was tested, and the reaction was then quenched with hydroxylamine to a final concentration of 0.5% (v/v) for 15 min. The TMTpro-labeled samples were pooled together at a 1:1 ratio. The sample was vacuum centrifuged to near dryness, resuspended in 5% formic acid and subjected to C18 solid-phase extraction (SPE) (Sep-Pak, Waters). Dried TMTpro-labeled sample was resuspended in 100 μl of 10 mM NH4HCO3, pH 8.0, and fractionated using BPRP HPLC. Briefly, samples were offline fractionated over 90 min and separated into 96 fractions by high pH reverse-phase HPLC (Agilent LC1260) through an Agilent ZORBAX 300Extend C18 column (3.5-μm particles, 4.6-mm ID and 250 mm in length) with mobile phase A containing 5% acetonitrile and 10 mM NH4HCO3 in LC-MS grade H2O, and mobile phase B containing 90% acetonitrile and 10 mM NH4HCO3 in LC-MS grade H2O (both pH 8.0). The 96 resulting fractions were then pooled in a non-continuous manner into 24 fractions, and 12 fractions (non-adjacent) were used for subsequent MS analysis. For lysosomal extracts, the dried TMTpro-labeled sample was resuspended in 300 μL of 0.1% trifluoroacetic acid and then fractionated into six fractions, using the high pH reversed-phase peptide fractionation kit (Thermo Fisher Cat#84868). Fractions were vacuum centrifuged to near dryness. Each consolidated fraction was desalted via StageTip, dried again via vacuum centrifugation, and reconstituted in 5% acetonitrile, 1% formic acid for LC-MS/MS processing. MS data were collected using an Orbitrap Fusion Lumos MS (Thermo Fisher) coupled to a Proxeon EASY-nLC1200 LC pump (Thermo Fisher). Peptides were separated on a 100-μm inner diameter microcapillary column packed in house with ~35 cm of Accucore150 resin (2.6 μm, 150 Å, Thermo Fisher) with a gradient consisting of 5–21% (0–125 min), 21–28% (125–140 min) (ACN, 0.1% FA) over a total 150 min run at ~500 nL/min. For analysis, we loaded 2–3 μg of each fraction onto the column. Each analysis used the Multi-Notch MS-based TMT method, to reduce ion interference compared to MS quantification. The scan sequence began with an MS spectrum (Orbitrap analysis; resolution 120,000 at 200 Th; mass range 400–1400 m/z; automatic gain control (AGC) target 5 × 10; maximum injection time 50 ms). Precursors for MS analysis were selected using a Top10 method. MS analysis consisted of collision-induced dissociation (quadrupole ion trap analysis; Turbo scan rate; AGC 2.0 × 10; isolation window 0.7 Th; normalized collision energy (NCE) 35; maximum injection time 35 ms). Monoisotopic peak assignment was used, and previously interrogated precursors were excluded using a dynamic window (120 s ± 10 ppm). After acquisition of each MS spectrum, a synchronous-precursor-selection MS scan was collected on the top 10 most intense ions in the MS spectrum. MS precursors were fragmented by high-energy collision-induced dissociation and analyzed using the Orbitrap (NCE 55; AGC 3 × 10; maximum injection time 100 ms, resolution was 50,000 at 200 Th). Raw mass spectra obtained were processed using Sequest. Mass spectra were converted to mzXML using a version of ReAdW.exe. Database searching included all entries from the Human Reference Proteome. Searches were performed with the following settings: (1) 20 ppm precursor ion tolerance for total protein level analysis, (2) Product ion tolerance was set at 0.9 Da, (3) TMTpro on lysine residues or N-termini at +304.207 Da, and (4) Carbamidomethylation of cysteine residues (+57.021 Da) as a static modification and oxidation of methionine residues (+15.995 Da) as a variable modification. Peptide-spectrum matches (PSMs) were adjusted to a 1% false discovery rate PSM filtering was performed using a linear discriminant analysis. To quantify the TMTpro-based reporter ions in the datasets, the summed signal-to-noise (S:N) ratio for each TMTpro channel was obtained and found the closest matching centroid to the expected mass of the TMTpro reporter ion (integration tolerance of 0.003 Da). Proteins were quantified by summing reporter ion counts across all matching PSMs, as described PSMs with poor quality, or isolation specificity <0.7, or with TMTpro reporter summed signal-to-noise ratio that were less than 100 or had no MS spectra were excluded from quantification. Values for protein quantification were exported and processed using Perseus to calculate Log fold-changes and p-values. Volcano plots using these values were plotted in Excel. Analytic thin-layer chromatography (TLC) was performed on 10-cm high-performance thin-layer chromatography (HPTLC) plates (Sigma, Cat# 1056310001). The organic fractions of samples were dried down and analyzed by 2D TLC with chloroform:methanol:water-concentrated ammonia 70:30:3:2 (by vol) used as the first dimension and chloroform:methanol:water 65:35:5 (by vol) used as the second dimension as described. Standard lipids (10 µg) dissolved in methanol or chloroform-methanol (1:1, v/v) were used as a reference. All statistical analysis was performed using GraphPad Prism 8. Information about significance test is provided in the respective figure legends. All multiple comparisons were performed with the Dunn multiple comparisons correction. Further information on research design is available in the Nature Research Reporting Summary linked to this article.
|
PMC7897626
|
Diacylglycerol kinase and phospholipase D inhibitors alter the cellular lipidome and endosomal sorting towards the Golgi apparatus
|
The membrane lipids diacylglycerol (DAG) and phosphatidic acid (PA) are important second messengers that can regulate membrane transport by recruiting proteins to the membrane and by altering biophysical membrane properties. DAG and PA are involved in the transport from the Golgi apparatus to endosomes, and we have here investigated whether changes in these lipids might be important for regulation of transport to the Golgi using the protein toxin ricin. Modulation of DAG and PA levels using DAG kinase (DGK) and phospholipase D (PLD) inhibitors gave a strong increase in retrograde ricin transport, but had little impact on ricin recycling or degradation. Inhibitor treatment strongly affected the endosome morphology, increasing endosomal tubulation and size. Furthermore, ricin was present in these tubular structures together with proteins known to regulate retrograde transport. Using siRNA to knock down different isoforms of PLD and DGK, we found that several isoforms of PLD and DGK are involved in regulating ricin transport to the Golgi. Finally, by performing lipidomic analysis we found that the DGK inhibitor gave a weak, but expected, increase in DAG levels, while the PLD inhibitor gave a strong and unexpected increase in DAG levels, showing that it is important to perform lipidomic analysis when using inhibitors of lipid metabolism.The cell itself and all of its organelles are surrounded by a lipid bilayer, made up of phospholipids, sphingolipids and sterols. It is becoming increasingly clear that lipids play an active role in numerous cellular processes, including transport events and signaling. Both diacylglycerol (DAG) and phosphatidic acid (PA) are important membrane lipids and second messengers that contribute to cellular processes either by their biophysical effect on the membrane or by recruiting proteins to the membrane. Due to their small head groups, DAG and PA have conical shapes, and local accumulation of DAG or PA can impart negative curvature to membranes thereby facilitating membrane budding, and fusion- and fission reactions during vesicular transport . DAG- or PA-mediated protein recruitment and activation of kinases can also contribute to modulation of vesicular trafficking . DAG interacts with proteins containing a C1-domain, such as the protein kinase C (PKC) and protein kinase D (PKD) families . A broad range of PA-binding proteins have been identified [4, 5]. PA-protein binding is thought to be mediated by interaction between positively charged amino acids in target proteins and the negatively charged head group of PA, but other factors, such as the membrane packing and curvature, are also important determinants for PA-protein interaction . PA and DAG can be formed by de novo lipid biosynthesis, and are intermediates for the production of all the glycerophospholipids [1, 7]. DAG and PA are also produced during receptor signaling which activates a number of different signaling pathways, and the levels of DAG and PA must therefore be strictly controlled. DAG and PA are produced by phospholipid hydrolysis mediated by phospholipase C (PLC) and phospholipase D (PLD), respectively. Sequence analyses have revealed six mammalian PLD isoforms, but only three (PLD1, PLD2 and PLD6, also called mitoPLD) have so far been shown to have enzymatic activity resulting in PA production . Importantly it was recently demonstrated that PLD3 is localized on endosomal structures that are positive for retromer components; it plays a role for intracellular sorting and it causes changes in accordance with the idea that it has enzymatic activity . DAG and PA signaling are terminated by the action of DAG kinases (DGKs) and PA phosphatases, respectively, interconverting the two lipids. In mammalian cells, 10 DGK isoforms have been identified, with different tissue expression and subcellular localization . DAG is also formed as a byproduct in sphingomyelin synthesis while PA can also be formed from lysophosphatidic acid (LPA) by the action of LPA acyltransferase (LPAAT). Protein toxins can be used as tools to study intracellular transport and both the bacterial toxin Shiga toxin (an AB5 toxin), as well as the plant toxin ricin (an AB toxin), have during the years proven useful to discover pathways in cells (for reviews, see [11–13]). The ricin B-chain binds to N-terminal galactoses on glycoproteins and glycolipids on the cell membrane and is therefore likely to be transported along the various pathways that exist in cells. Actually, in 1979, ricin was used to demonstrate that endocytosed molecules could be recycled , and some of the early evidence for clathrin-independent endocytosis was obtained using ricin [15, 16]. The ricin A-chain has enzymatic activity that removes a specific adenine from the 60 s ribosomal subunit, thereby preventing protein synthesis. To reach its cellular target, ricin must be transported from endosomes, via the Golgi apparatus, to the endoplasmic reticulum (ER). Here, the A chain is released from the B-chain and translocated into the cytosol where it exerts its toxic action. In the present study we have examined how modulation of the DAG and PA balance affects retrograde transport. We have found that inhibitors of DGK and PLD strongly increase ricin transport from the endosomes to the Golgi and as a consequence of this upregulation, more toxin is able to reach the cytosol. To study the different steps we have used native ricin as well as I-labeled ricin and genetically modified ricin that can be sulfated in the Golgi apparatus and mannosylated in the ER . Lipidomic analysis showed that DGK inhibition as expected gave an increase in DAG levels, whereas PLD inhibition surprisingly increased the levels of both DAG and phosphatidylglycerol (PG), a lipid which has so far not been extensively studied and where specific roles in intracellular transport has not, to our knowledge, been reported. Chemicals were purchased from Sigma-Aldrich unless stated otherwise. Ricin holotoxin was purchased from Sigma-Aldrich. Ricin-sulf1 and ricin-sulf2 were prepared as previously described . The plasmids expressing StxB-sulf2 and the Shiga toxin 1 mutant (Stx1m) were kind gifts from Prof. B Goud (Institut Curie, Paris, France) and Prof. A. D. O’Brien (Uniformed Services University of the Health Sciences, Bethesda, MD, USA), respectively, and StxB-sulf2 and Stx1m were purified as previously described. H2SO4 (S-RA-1) was from Hartmann Analytic. The iodine-125 radionuclide (I, NEZ033A010MC) for radioactive protein labeling, l-[3,4,5-H(N)]-leucine (NET460005MC) and d-[2-H(N)]-mannose (NET570A005MC) were from Perkin Elmer. The following compounds were used: R59022 (D5919), R59949 (D5794), Bisindolylmaleimide I hydrochloride (BIM I; B6292), Phorbol 12-myristate 13-acetate (PMA; 79346) and Swainsonine (S9263) from Sigma-Aldrich, CAY10593 (13206-10) and CAY10594 (13207-10) from Cayman Chemical, FIPI (3600) and VU 0364739 hydrochloride (4171) were from Tocris, Sotrastaurin (S2791) and SAR405 (S7682) from Selleckchem. The environment-sensitive probe NR12S was a kind gift from Prof. A. Klymchenko (University of Strasbourg, Strasbourg, France). The following Ambion Silencer® Select siRNAs were purchased from Thermo Fisher Scientific: Negative Control No. 1 (4390844), DGKA (4427038 s3913), DGKD (4427037 s224987), DGKE (4427038 s16208), DGKH (4427037 s46229), DGKZ (4427037 s16205), PLD1 (4427038 s10638), PLD2 (4427038 s10642), PLD3 (4427037 s24272). The following antibodies were used: homemade rabbit anti-ricin antibody, mouse anti-Golgin-97 (A-21270, Molecular Probes), mouse anti-Shiga toxin 1 (STX1-3C10, Toxin Technology), rabbit anti-Golgin-97 (13192S, Cell Signaling Technology), rabbit anti-EEA1 (2411, Cell Signaling Technology), rabbit anti-Phospho-(Ser/Thr) PKD substrate (4381, Cell Signaling Technology), rabbit anti-cofilin (ab42824, Abcam), goat anti-Vps35 (ab10099, Abcam), mouse anti-SNX2 (611308, BD Transduction Laboratories). Secondary antibodies for immunofluorescence were from Jackson ImmunoResearch and Molecular Probes, IRDye secondary antibodies for Western blotting were from LICOR Biosciences, GmbH. Ricin holotoxin was labeled with Alexa Fluor 555 using the Alexa Fluor™ 555 Microscale Protein Labeling Kit (A30007, Invitrogen) according to the manufacturer’s protocol. HEp-2 cells (human cervical adenocarcinoma, CCL-23, ATCC), and U-2 OS cells (human bone osteosarcoma, HTB-96, ATCC) were grown in Dulbecco’s Modified Eagle Medium (DMEM, D0819, Sigma-Aldrich) with 10% fetal bovine serum (FBS, F7524, Sigma-Aldrich). PC-3 cells (human prostate adenocarcinoma, CRL-1435, ATCC) were grown in DMEM/F-12 (1:1) (31331–093, Thermo Fisher Scientific) with 7% FBS. CaCo2 cells (human colorectal adenocarcinoma, HTB-37, ATCC) were grown in DMEM with 15% FBS. hTERT-RPE1 cells (retinal pigment epithelia, CRL-4000, ATCC) stably expressing GFP-WDFY2 was a kind gift from Prof. H. Stenmark, Oslo, and were grown in DMEM/F-12 (1:1) with 10% FBS. All cell lines were maintained in medium supplemented with 100 U/ml penicillin and 100 µg/ml streptomycin (P4333, Sigma-Aldrich) at 37 °C and 5% CO2. HEp-2 cells were plated at 1 × 10 cells/well (6 well plate) or 2 × 10 cells/well (24 well plate) the day before transfection. Cells were transfected with Ambion Silencer® Select siRNAs of each target gene or control at a concentration of 10 nM using Lipofectamine® RNAiMAX Transfection Reagent (13778, Thermo Fisher Scientific) according to the manufacturer’s protocol. The transfection medium was replaced with growth medium after 24 h, and assays were carried out 48 h post transfection. Knockdown efficiencies were determined by measuring mRNA levels using qPCR. Total RNA was isolated using the RNeasy Plus Mini Kit (74134, Qiagen) according to the manufacturer’s protocol. cDNA was synthesized from 1 µg RNA using the iScript cDNA synthesis kit (1708891, Bio-Rad Laboratories Inc) according to the manufacturer’s protocol. qPCR was performed using the LightCycler® 480 SYBR Green I Master kit (04707516001, Roche Diagnostics) in combination with QuantiTect Primer Assays (all from Qiagen) for DGKA (QT00091112), DGKD (QT00068894), DGKE (QT00090720), DGKH (QT00092939), DGKQ (QT00005348), DGKZ (QT00071176), PLD1 (QT00085512), PLD2 (QT00017682), PLD3 (QT00029239) and TBP (reference gene; QT00000721). Serial dilutions of control samples were used to plot standard curves and to quantify primer efficiencies in each qPCR run. The qPCR was carried out using a LightCycler® 480 Instrument (Roche Diagnostics). The reactions were run in duplicates, and the samples were first denatured at 95 °C for 10 min, followed by 45 cycles of denaturation at 95 °C for 10 s, primer annealing at 55 °C for 20 s and primer extension at 72 °C for 20 s. Upon completion of the cycling steps, a melting curve analysis was done. Cp values and primer efficiencies were determined using the LightCycler® 480 software (Roche Diagnostics). For sulfation experiments, the following cell numbers were seeded in 6 well plates one or two days prior to experiments: HEp-2: 2–2.5 × 10 cells/well (1 day) and 1 × 10 cells/well (knockdown experiments), U-2 OS: 2 × 10 cells/well (1 day), PC-3: 3 × 10 cells/well (1 day), CaCo-2: 2 × 10 cells/well (2 days), RPE GFP-WDFY2: 2 × 10 cells/well (1 day). Cells were washed twice with sulfate-free medium (1.0 mM CaCl2, 5.4 mM KCl, 1.0 mM MgCl2, 0.12 M NaCl, 26.2 mM NaHCO3, 10.9 mM NaH2PO4, 1 g/l D-glucose, 0.01 g/l phenol red, supplemented with MEM vitamin solution, MEM amino acids and MEM non-essential amino acid; pH 7.4) and subsequently incubated with 0.1–0.2 mCi/ml SO4 for 2 h at 37 °C, before inhibitors were added and the incubation continued for 1 h. For the knockdown experiments, the cells were incubated with 0.2 mCi/ml SO4 for 3 h. Subsequently, 4 µg/ml ricin-sulf1, 15 µg/ml ricin-sulf2 or 2 µg/ml StxB-sulf2 was added for 1.5 h, 3 h or 1 h, respectively. Samples treated with ricin-sulf1 or ricin-sulf2 were washed twice with 0.1 M lactose in HEPES-buffered medium (MEM without sodium bicarbonate supplemented with 20 mM HEPES, 2 mM l-alanyl-l-glutamine, 100 U penicillin and 100 µg/ml streptomycin) for 5 min at 37 °C to remove surface-bound ricin, then washed with ice-cold PBS (1.1 mM NaH2PO4, 5.5 mM Na2HPO4, 138.6 mM NaCl; pH 7.4) and lysed in lysis buffer (0.1 M NaCl, 10 mM Na2HPO4, 1 mM EDTA, 1% Triton X-100, supplemented with cOmplete™ Protease inhibitor Coctail (05056489001, Roche Diagnostics) and 60 mM n-octyl-β-glucopyranoside; pH 7.4). Samples treated with StxB-sulf2 were washed with PBS and lysed. Lysates were cleared by centrifugation and ricin-sulf1, ricin-sulf2 or StxB-sulf2 were immunoprecipitated overnight at 4 °C from cleared lysates using Protein A Sepharose® beads (17-0963-03, GE Healthcare) with the appropriate antibody adsorbed. The immunoprecipitate was washed twice with 0.35% TritonX-100 in PBS, resuspended in Laemmli sample buffer (161–0747, Bio-Rad Laboratories Inc) with 100 mM DTT and boiled for 5 min. The immunoprecipitate was separated by SDS-PAGE, blotted onto a PVDF membrane and visualized by digital autoradiography using a phosphor imaging screen (Imaging Screen-K (Kodak), Bio-Rad Laboratories Inc) and the Molecular Imaging PharosFX system (Bio-Rad Laboratories Inc). Band intensities were quantified using the Quantity One 1-D Analysis Software (Bio-Rad Laboratories Inc). To determine the total amount of protein sulfation, proteins from the supernatants after immunoprecipitation were precipitated with 5% trichloroacetic acid (TCA; 100807, Merck). The TCA precipitate was dissolved in 0.1 M KOH and measured by liquid scintillation counting using a Tri-Carb 2100TR Liquid Scintillation Analyzer (Packard). Ricin was I-labelled using Pierce Iodination Tubes (cat. no 28601, Thermo Fisher Scientific) according to the manufacturer’s protocol. HEp-2 cells were seeded at a concentration of 5 × 10 cells/well (for inhibitor treatment) or 2 × 10 cells/well (for knockdown) in a 24 well plate one day prior to the experiment. For the treatment with inhibitors, the cells were washed once with HEPES-buffered medium before being incubated with inhibitors for 1 h at 37 °C. Subsequently, ~ 60 ng/ml of I-labeled ricin was added and the incubation continued for 20 min at 37 °C. The cells were incubated with or without 0.1 M lactose in HEPES-buffered medium for 5 min to remove surface-bound ricin and then washed three times with PBS or 0.1 M lactose. The cells were dissolved in 0.1 M KOH and the radioactivity in the solution was determined by gamma-counting using a Hidex Automatic Gamma Counter (Hidex). Ricin endocytosis was expressed as the amount of I-labeled ricin in lactose-washed cells divided by the amount of I-labeled ricin in PBS-washed cells. HEp-2 cells were seeded at a concentration of 5 × 10 cells/well in a 24 well plate one day prior to the experiment. The cells were washed once with HEPES-buffered medium before being incubated with inhibitors for 1 h at 37 °C. Then, ~ 60 ng/ml of I-labeled ricin was added and the incubation was continued for 20 min at 37 °C. Surface-bound I-labeled ricin was removed by incubating the cells with 0.1 M lactose in HEPES-buffered medium for 5 min and washing three times with the same solution. The I-labeled ricin was then chased in the presence of inhibitors for 2 h at 37 °C in 1 mM lactose in HEPES-buffered medium to prevent further uptake of the toxin. To determine the amount of degraded and recycled I-labeled ricin, the medium was collected and proteins were precipitated with a final concentration of 4% TCA and 0.4 mg/ml bovine serum albumin (BSA) and pelleted. The supernatant contains I released from cells after I-labeled ricin degradation, whereas the pellet contains recycled I-labeled ricin. To determine the amount of cell-associated I-ricin, cells were dissolved in 0.1 M KOH. The amount of I in the supernatant, pellet and cells was measured by gamma-counting and divided by the total count to give the amount of degradation, recycling and cell association, respectively. Images were acquired using Zeiss LSM 780 or LSM 880 laser scanning confocal microscopes equipped with an Ar-Laser Multiline (458/488/514 nm), a DPSS-561 10 (561 nm), a Laser diode 405–30 CW (405 nm), and a HeNe-laser (633 nm); the objective used was Zeiss plan-Apochromat 63x/1.40 Oil DIC M27 (all from Carl Zeiss MicroImaging GmbH). RPE-GFP-WDFY2 cells were seeded onto coverslips immersed in growth media-filled wells. Next day, the cells were washed once with warm HEPES buffered medium and incubated with indicated inhibitors in HEPES buffered medium for 1 h at 37 °C. The cells were fixed in 4% formaldehyde/PBS solution for 20 min at room temperature. Finally, the coverslips were washed 3 times in PBS, rinsed with water and mounted to microscope slides using ProLong™ Diamond antifade mountant (ThermoFisher Scientific). Images were acquired using Zeiss LSM 780 or LSM 880 microscopes with pixel size 0.071 µm and a pinhole of 50 µm. The focus was set to the plane with the majority of the large WDFY2-positive endosomes. Image pre-processing was performed using Fiji software . The background was subtracted using rolling ball with the size 50 and then images were exported to HDF5 format using the Ilastik plugin. Endosome size was analyzed using Ilastik software using the pipeline for the pixel classification followed by object classification (Fig. 6d). The training was performed using 1–2 images from each condition. First, the software was trained to classify pixels based on GFP signal: negative (background), GFP-positive or lumen (within the lumen of large endosomes). Then, GFP-positive pixels were further classified into different objects: small, medium, large/cluster and giant/cluster. This classification failed to discriminate between large endosomes and the clusters, thus it was used only for quantifying small endosomes (GFP-positive structures without lumen). To quantify medium and large endosomes (with visible lumen), pixels classified as luminal in the pixel classification step were used for the object classification. The luminal pixels were separated into three object classes: medium, large and giant. Finally, to quantify cell-covered area in each image, a separate pixel classification was performed on the same images and the pixels were classified as positive (cells; included both strong WDFY2-GFP staining on the endosomes and the weak WDFY2-GFP staining in the cytosol) or negative for WDFY2-GFP signal. RPE-GFP-WDFY2 cells were grown and treated with the inhibitors as described in the section “Endosome size analysis”. To preserve tubules during cell fixation, 16% formaldehyde (18814, Polysciences) was warmed to 37 °C and added drop-wise directly to the cell medium to a final concentration of 4%. The cells were then fixed for 20 min at room temperature, washed three times with PBS and the coverslips were mounted to microscope slides using ProLong™ Diamond antifade mountant. The images were acquired using similar settings as for the “Endosome size analysis”, but using 3D scanning with the step size of 0.37 µm. Images of the maximum intensity projections from the 3D scans were used to analyze the number and length of the tubules in the cells. The number of the tubules per cell and the length of the tubules were quantified by manually marking and measuring all the visible tubules using Fiji software . For the cells that had no visible tubular structures, a value of “0” was added to the measured tubule length. Because it was difficult to distinguish between short tubules and small elongated endosomes, only tubules of at least 0.7 µm lengths (called long tubular structures) were included in the counting for mean length and number of long tubules per cell. RPE-GFP-WDFY2 cells were seeded into glass-bottom microwell dishes (MatTek Corporation). Next day, the cells were washed twice with FluoroBrite™ DMEM medium (Gibco™, Thermo Fisher Scientific) and incubated with indicated inhibitors prepared in FluoroBrite™ DMEM medium. The imaging was started after 1 h and continued for up to 30 min. The images were acquired with a Zeiss LSM 880 microscope using Airyscan detector in a fast scanning mode. The images were taken each 2 s for 2 min and processed using ZEN 2.3 Blue software (Carl Zeiss). For toxin transport to the Golgi and endosome size based on EAA1 staining, U-2 OS or HEp-2 cells grown on coverslips were treated as indicated in figure legends and subsequently fixed in 4% formaldehyde/PBS for 20 min at room temperature, followed by permeabilization in 0.1% Triton X-100 for 5 min and blocking in 5% FBS in PBS for 30 min. The samples were labeled with primary antibodies in 5% FBS for 1 h and with secondary antibodies in 5% FBS for 30 min, with 3 × 5 min wash in PBS after each antibody labeling. Then the coverslips were rinsed with water and mounted to microscope slides using ProLong™ Gold or ProLong™ Diamond antifade mountant with DAPI (ThermoFisher Scientific). Images were acquired using a Zeiss LSM 780 microscope with pixel size 0.071 µm and the pinhole of 50 µm. Individual cells were manually marked in the images and the toxin labeling intensity within the Golgin-97-positive structures was analyzed using Fiji software . To determine the size of EEA1-positive structures, individual cells were marked and the endosome area was measured using the particle analyzer in the Fiji software . For analyzing ricin transport via the endosomal system, RPE-GFP-WDFY2 cells were grown and treated with the inhibitors as described in “Endosome size analysis”. Ricin-Alexa555 was then added to the cell medium (final conc. 1 µg/ml). After 20 min, lactose solution was added to the cell medium (final conc. 0.1 M) to remove surface exposed ricin-Alexa555. After 10 min, the cells were fixed by adding 16% paraformaldehyde (37 °C) directly to the cell medium (final conc. 4%) and incubated for 20 min at room temperature. The cells were washed 3-times with PBS and then permeabilized and blocked using 0.05% saponin/5% FBS in PBS (blocking solution) for 20 min at room temperature. The antibodies were prepared in blocking solution, and the samples were labeled with primary antibodies for 1 h and with the secondary antibodies for 30 min, with 3 × 5 min wash in 0.05% saponin in PBS after each antibody labeling. Then the coverslips were rinsed with water and mounted to microscope slides using ProLong™ Diamond antifade mountant. Images were acquired with a Zeiss LSM 880 microscope using Airyscan detector with a voxel size of 0.035 × 0.035 × 0.16 µm for high resolution imaging and super-resolution processing. Image acquisition, processing and analysis were performed with ZEN 2.3 Blue software (Carl Zeiss MicroImaging GmbH), while 3D visualization of z-stacks was done using Imaris 9.2.1 (Bitplane AG). HEp-2 cells were seeded at a concentration of 2.5 × 10 cells/well in 6 well plates one day prior to the experiment. The cells were washed and serum-starved in sulfate-free medium for 2 h at 37 °C to mimic the conditions of the sulfation experiments. Inhibitors were added and the incubation continued for 1 h at 37 °C. The cells were washed in TBS (20 mM Tris, 150 mM NaCl, pH 7.6) and lysed in lysis buffer supplemented with PhosStop (04906837001, Roche Diagnostics). Cleared lysates were mixed with Laemmli sample buffer supplemented with 100 mM M DTT and boiled for 5 min, before proteins were separated by SDS-PAGE. Proteins were transferred onto a low-fluorescent PVDF membrane that was subsequently blocked in 5% skim milk in TBS for 40–60 min. The membrane was washed in 0.1% TBS-Tween (TBST), cut and incubated with primary antibodies in 5% BSA in TBST overnight at 4 °C. The membrane was washed and incubated with secondary antibodies in 5% BSA in TBST for 1 h in the dark. After washing and drying, proteins were visualized by scanning the membrane on an Odyssey imaging system (LICOR Biosciences, GmbH). Band intensities were quantified using the Quantity One 1-D Analysis Software (Bio-Rad Laboratories Inc.). HEp-2 cells were seeded at a concentration of 1.2 × 10 in 10 cm dishes one day prior to experiments. The cells were washed with HEPES-buffered medium and incubated with inhibitors for 1 or 3 h at 37 °C. The cells were then washed with Dulbecco’s PBS without Ca and Mg (14190094, Thermo Fisher Scientific) and detached with Accutase (A6964, Sigma-Aldrich) at 4 °C. The cells were washed twice in PBS and counted using a Countess II Automated Cell Counter (Thermo Fisher Scientific). The cell concentration was adjusted to a final concentration of 1.5 × 10 cells/ml and samples were stored at -80 °C until shipment to Lipotype GmbH for analysis using mass spectrometry (MS). MS-based lipid analysis was performed by Lipotype GmbH (Dresden, Germany) as described . Lipids were extracted using a two-step chloroform/methanol procedure . Samples were spiked with internal lipid standard mixture containing: cardiolipin 16:1/15:0/15:0/15:0 (CL), ceramide 18:1;2/17:0 (Cer), diacylglycerol 17:0/17:0 (DAG), hexosylceramide 18:1;2, 12:0 (HexCer), lyso-phosphatidate 17:0 (LPA), lyso-phosphatidylcholine 12:0 (LPC), lyso-phosphatidylethanolamine 17:1 (LPE), lyso-phosphatidylglycerol 17:1 (LPG), lyso-phosphatidylinositol 17:1 (LPI), lyso-phosphatidylserine 17:1 (LPS), phosphatidate 17:0/17:0 (PA), phosphatidylcholine 17:0/17:0 (PC), phosphatidylethanolamine 17:0/17:0 (PE), phosphatidylglycerol 17:0/17:0 (PG), phosphatidylinositol 16:0/16:0 (PI), phosphatidylserine 17:0/17:0 (PS), cholesterol ester 20:0 (CE), sphingomyelin 18:1;2/12:0;0 (SM), triacylglycerol 17:0/17:0/17:0 (TAG). After extraction, the organic phase was transferred to an infusion plate and dried in a speed vacuum concentrator. 1st step dry extract was re-suspended in 7.5 mM ammonium acetate in chloroform/methanol/propanol (1:2:4, v:v:v) and 2nd step dry extract in 33% ethanol solution of methylamine in chloroform/methanol (0.003:5:1; v:v:v). All liquid handling steps were performed using Hamilton Robotics STARlet robotic platform with the Anti Droplet Control feature for organic solvents pipetting. Samples were analyzed by direct infusion of a QExactive mass spectrometer (Thermo Fisher Scientific) equipped with a TriVersa NanoMate ion source (Advion Biosciences). Samples were analyzed in both positive and negative ion modes with a resolution of Rm/z=200 = 28,000 for MS and Rm/z=200 = 17,500 for MSMS experiments, in a single acquisition. MSMS was triggered by an inclusion list encompassing corresponding MS mass ranges scanned in 1 Da increments . Both MS and MSMS data were combined to monitor CE, DAG and TAG ions as ammonium adducts; PC, PC-O (ether-linked PC; alkyl or alkenyl), as acetate adducts; and CL, PA, PE, PE-O (ether-linked PE; alkyl or alkenyl), PG, PI and PS as deprotonated anions. MS only was used to monitor LPA, LPE, LPE-O, LPI and LPS as deprotonated anions; Cer, HexCer, SM, LPC and LPC-O as acetate adducts. Data were analyzed with in-house developed lipid identification software based on LipidXplorer [28, 29]. Data post-processing and normalization were performed using an in-house developed data management system. Only lipid identifications with a signal-to-noise ratio > 5, and a signal intensity fivefold higher than in corresponding blank samples were considered for further data analysis. The acquired data was then post-processed to remove minor species and possible miss-detections based on these criteria: (1) keep only species that are above 0 in at least 9 samples and are above 0 in both biological replicates in at least 2 conditions; (2) remove all species in the whole sample if there is a clear misdetection (completely different species distribution compared to the rest of the samples). The remaining lipid values were used for figure plotting and further data quantification. Glycerolipid species are listed with the fatty acyl groups separated with a hyphen as the sn-position of the fatty acids was not identified. The differences between means for two groups were determined using two-tailed Student’s t-test. The level of significance was set as follows: *p ≤ 0.05, ** p ≤ 0.01, *** p ≤ 0.005 compared to control unless otherwise stated. To modulate the DAG and PA levels in the cells, we have used the DGK inhibitors R59022 (RI) and R59949 (RII) and the PLD inhibitors CAY10593 (CAY93) and CAY10594 (CAY94) (Fig. 1a). These inhibitors have been reported to have different selectivities for DGK and PLD isoforms. Both DGK inhibitors strongly inhibit the DGKα isoform; RI moderately inhibits DGKε and DGKθ, whereas RII strongly inhibits DGKγ and moderately inhibits DGKδ and DGKκ . CAY93 primarily inhibits PLD1, whereas CAY94 primarily inhibits PLD2 . To determine the effect of DAG and PA modulation on transport to the Golgi apparatus, we have taken advantage of the sulfation process that occurs specifically in the trans-Golgi network (TGN). Using modified protein toxins containing sulfation sites that can be labeled with radioactive sulfate, we can monitor retrograde transport to the Golgi. Treatment with both DGK and PLD inhibitors strongly increased sulfation of ricin-sulf1 (Fig. 1b) in HEp-2 cells, suggesting that retrograde transport is increased when DAG and PA levels are altered. Combination of the DGK and PLD inhibitors gave an additive effect, indicating that DGK and PLD inhibition may increase transport by different mechanisms. The inhibitors did not affect the sulfation process per se, as total protein sulfation was not affected. Since RI and CAY94 had a tendency for higher effect on ricin sulfation compared to RII and CAY93, respectively, we decided to focus on the DGK inhibitor RI and the PLD inhibitor CAY94 for the rest of the study.Fig. 1Ricin transport to the Golgi is strongly increased upon DGK and PLD inhibitor treatment. a Schematic presentation of the effect of DGK and PLD inhibitors on DAG and PA metabolism. b HEp-2 cells were subjected to sulfation assay with ricin-sulf1 (RS1) after treatment with the 10 µM of the indicated inhibitors for 1 h. The upper figure shows an autoradiogram and the western blot against ricin for the same membrane from one representative experiment and the lower figure shows the quantification of RS1 sulfation expressed as percent of control, n ≥ 3. c RS1 sulfation was measured as described above after treatment with RS1 for the indicated time points. The upper figure shows an autoradiogram from one representative experiment and the lower figure shows the quantification of RS1 sulfation, n = 3. d RS1 sulfation was measured as described above in CaCo-2, PC-3 and U-2 OS cells, n = 3. e U-2 OS cells were treated with 10 µM of the indicated inhibitors for 1 h, followed by a 30 min pulse with 4 µg/ml RS1 and a 1 h chase. Cells were fixed, permeabilized, stained with antibodies against ricin (red) and Golgin97 (green) and mounted in ProLong Diamond with DAPI (blue). Left: Representative confocal images; scale bar, 10 µm. Right: To quantify RS1 transport to the Golgi, at least 29 individual cells from 3–6 confocal images for each condition were manually marked and analyzed using Fiji for each experiment. The upper figure shows the mean intensity of ricin in the Golgin97 mask in individual cells from one representative experiment, the lower figure shows the summary of means from three independent experiments Ricin transport to the Golgi is strongly increased upon DGK and PLD inhibitor treatment. a Schematic presentation of the effect of DGK and PLD inhibitors on DAG and PA metabolism. b HEp-2 cells were subjected to sulfation assay with ricin-sulf1 (RS1) after treatment with the 10 µM of the indicated inhibitors for 1 h. The upper figure shows an autoradiogram and the western blot against ricin for the same membrane from one representative experiment and the lower figure shows the quantification of RS1 sulfation expressed as percent of control, n ≥ 3. c RS1 sulfation was measured as described above after treatment with RS1 for the indicated time points. The upper figure shows an autoradiogram from one representative experiment and the lower figure shows the quantification of RS1 sulfation, n = 3. d RS1 sulfation was measured as described above in CaCo-2, PC-3 and U-2 OS cells, n = 3. e U-2 OS cells were treated with 10 µM of the indicated inhibitors for 1 h, followed by a 30 min pulse with 4 µg/ml RS1 and a 1 h chase. Cells were fixed, permeabilized, stained with antibodies against ricin (red) and Golgin97 (green) and mounted in ProLong Diamond with DAPI (blue). Left: Representative confocal images; scale bar, 10 µm. Right: To quantify RS1 transport to the Golgi, at least 29 individual cells from 3–6 confocal images for each condition were manually marked and analyzed using Fiji for each experiment. The upper figure shows the mean intensity of ricin in the Golgin97 mask in individual cells from one representative experiment, the lower figure shows the summary of means from three independent experiments Next, by measuring ricin sulfation at different time points ranging from 30 to 90 min, we investigated how the inhibitors affect the kinetics of the retrograde transport of ricin. We found that ricin was more rapidly transported to the Golgi after both DGK and PLD inhibition, and detectable levels of sulfated toxin could be seen already after 45 min in inhibitor-treated cells, whereas only faint bands were visible after 60 min in control-treated cells (Fig. 1c). Stronger band intensities also indicate that more ricin was able to reach the Golgi after DGK and PLD inhibition. The increase in retrograde transport of ricin is not restricted to HEp-2 cells, as we also saw a similar, although lower, increase in ricin sulfation after DGK and PLD inhibition in Caco-2, PC-3 and U-2 OS cells (Fig. 1d). To visualize ricin transport to the Golgi, cells were treated with a 30 min ricin-sulf1-pulse and chased for 60 min in the presence of 1 mM lactose to prevent reuptake of recycled toxin. We then looked at ricin colocalization with the TGN marker Golgin-97 by immunofluorescence confocal microscopy. As the Golgi morphology is highly variable in HEp-2 cells, we used U2-OS cells for this assay. After ricin-sulf1 pulse-chase, we could see a clear perinuclear staining partially overlapping with the TGN marker (Fig. 1e), and quantification of the ricin intensity/area in the TGN (Golgin-97 positive structures) showed an increase after inhibitor treatment, in agreement with the sulfation data. To further support the notion that retrograde transport of ricin is increased after DGK and PLD inhibition, we measured ricin transport to the ER and into the cytosol. ER transport was measured using a modified ricin molecule containing both sulfation and glycosylation sites (ricin-sulf2). N-linked glycosylation occurs in the ER and controls proper folding of proteins, and in the presence of radioactive mannose, ricin-sulf2 will be radioactively labeled when it reaches the ER. As shown in Fig. S1a, ricin mannosylation was increased after treatment with DGK and PLD inhibitors, corroborating the idea that retrograde transport is increased. Since ricin-sulf2 also contains a sulfation site, this molecule can be used to investigate transport from the Golgi to the ER when incubated in the presence of radioactive sulfate. Ricin-sulf2 will then be radioactively labeled in the Golgi and upon reaching the ER, glycosylation will increase the size of ricin-sulf2 which can be visualized by autoradiography. Comparison of the intensity of the two ricin-sulf2 bands shows that DGK and PLD inhibition has no effect on transport between the Golgi and the ER (Fig. S1b). Ricin exerts its toxic action by removing an adenine from the 60S ribosomal subunit, thereby preventing protein synthesis. We found that in the presence of DGK and PLD inhibitors, less ricin is needed to inhibit protein synthesis, resulting in a two-fold sensitization towards ricin (Fig. S1c,d). Overall, these experiments clearly show that DGK and PLD inhibition significantly increase the retrograde transport of ricin to the Golgi, the ER and the cytosol. The increase in retrograde transport can be caused by a change in endosomal sorting or by increased internalization, and we therefore measured the endocytic uptake after DGK and PLD inhibition using I-labelled ricin. As shown in Fig. 2a, equal amounts of ricin were internalized, suggesting that DGK and PLD inhibition increases retrograde transport by altering endosomal sorting. We next investigated if sorting into the degradative and recycling pathways was also affected by DGK and PLD inhibition. We treated cells with a 20 min pulse of I-labelled ricin, followed by a 2 h chase in medium containing lactose to prevent reuptake of ricin and determined the amount of cell-associated, recycled and degraded ricin. There were no significant changes in ricin degradation or recycling after DGK or PLD inhibition (Fig. 2b). Thus, DGK and PLD inhibition seems to increase the endosomal sorting of ricin into the Golgi without affecting the recycling or degradation of the toxin. We have previously shown that transport of ricin to the Golgi apparatus is dependent on Vps34 and formation of PI3P . Thus, we tested whether the Vps34 inhibitor SAR405 could prevent RI- and CAY94-induced increase in ricin transport (Fig. S2). Indeed, SAR405 gave a significant reduction in ricin sulfation when added in combination with RI or CAY94, indicating that the inhibitor-mediated increase in ricin sulfation is dependent on Vps34 activity.Fig. 2DGK and PLD inhibition do not inhibit the endocytosis, degradation or recycling of ricin. HEp-2 cells were treated with 0.1% DMSO, 10 µM RI and/or 10 µM CAY94 for 1 h followed by incubation with I-ricin for a 20 min to measure ricin endocytosis, n = 3 or b 20 min plus 2 h chase to measure the amount of ricin degradation and recycling, n = 3 DGK and PLD inhibition do not inhibit the endocytosis, degradation or recycling of ricin. HEp-2 cells were treated with 0.1% DMSO, 10 µM RI and/or 10 µM CAY94 for 1 h followed by incubation with I-ricin for a 20 min to measure ricin endocytosis, n = 3 or b 20 min plus 2 h chase to measure the amount of ricin degradation and recycling, n = 3 Next, we investigated if the increase in retrograde transport is specific for ricin or if other retrograde cargos are similarly affected. To this end, we studied the Golgi transport of the bacterial toxin Shiga toxin. Shiga toxin binds to the glycosphingolipid Gb3 on the cell surface and is also transported retrogradely from endosomes to the Golgi and the ER [11, 13]. We used a modified Shiga toxin molecule containing sulfation sites (StxB-sulf2) to investigate transport to the Golgi and measured sulfation in a similar manner as described above for ricin. We found that treatment with RI gave a clear increase in StxB-sulf2 sulfation, whereas treatment with CAY94 had a variable effect. In most experiments, CAY94 increased StxB-sulf2 sulfation, but generally had less effect than the DGK inhibitor (Fig. S3a). Colocalization studies between a non-toxic version of Shiga toxin (Stx1m) and the TGN marker Golgin-97 showed a similar trend of increased transport, without reaching statistical significance (Fig. S3b,c). These data suggest that the increase in retrograde transport mediated by DGK and PLD inhibitors is not restricted to ricin, but that not all pathways are changed to the same extent. Inhibition of DGK is expected to increase DAG levels while decreasing PA levels, whereas PLD inhibition can be expected to decrease PA and perhaps also DAG (Fig. 1a). To test whether this was indeed the case, we treated cells with inhibitors for 1 or 3 h and measured the changes in the cellular lipidome by MS lipidomics. As expected, after both 1 and 3 h, the DGK inhibitor RI gave a weak, but consistent, increase in DAG levels without affecting any of the other lipid classes measured (Fig. 3 and Supplementary material 2). The PLD inhibitor CAY94, however, gave a surprising transient increase in PA that was reversed after 3 h and a persistent increase in DAG and PG (Fig. 3 and Supplementary material 2). The inhibitors increased the levels of several DAG, PA and PG lipid species, with the most prominent effect on 16:0–18:1 (Fig. 4). DAG 16:0–16:0 was also increased after 1 h treatment with the PLD inhibitor and PA was increased after 1 h treatment with the PLD inhibitor. The DAG, PA and PG species that were most affected by inhibitor treatment share the same fatty acyl composition as the most abundant PC species. There were no major changes in the 18:0–20:4 species of DAG, PA and PG, which is the most abundant PI species, suggesting that PC is the source of the increased DAG, PA and PG levels (Fig. 4).Fig. 3Changes in lipid composition after inhibitor treatment. HEp-2 cells were treated with 0.1% DMSO, 10 µM RI, 10 µM CAY94 or with the combination of the inhibitors for 1 h or 3 h, and whole-cell lysates were analyzed by MS. The figure shows the amount of different lipid classes in the samples expressed as percentage of the total lipid for each sample. The error bars show mean deviation, n = 2. For detailed description of the calculations performed on MS data see section “Data analysis and post-processing” in the “Materials and methods” section and the abbreviations of all lipid classes are given under the section “Lipid extraction for MS lipidomics” in the Materials and methods sectionFig. 4Species composition of DAG, PA, PG, PC and PI after inhibitor treatment. HEp-2 cells were treated with 0.1% DMSO, 10 µM RI, 10 µM CAY94 or with the combination of the inhibitors for 1 h (a) or 3 h (b), and whole-cell lysates were analyzed by MS. The figure shows relative amount of major DAG, PA and PG species and the equivalent PC and PI species. The error bars show mean deviation, n = 2. For detailed description of the calculations performed on MS data see section “Data analysis and post-processing” in the Materials and methods section and the abbreviations of all lipid classes are given under the section “Lipid extraction for MS lipidomics” in the Materials and methods section Changes in lipid composition after inhibitor treatment. HEp-2 cells were treated with 0.1% DMSO, 10 µM RI, 10 µM CAY94 or with the combination of the inhibitors for 1 h or 3 h, and whole-cell lysates were analyzed by MS. The figure shows the amount of different lipid classes in the samples expressed as percentage of the total lipid for each sample. The error bars show mean deviation, n = 2. For detailed description of the calculations performed on MS data see section “Data analysis and post-processing” in the “Materials and methods” section and the abbreviations of all lipid classes are given under the section “Lipid extraction for MS lipidomics” in the Materials and methods section Species composition of DAG, PA, PG, PC and PI after inhibitor treatment. HEp-2 cells were treated with 0.1% DMSO, 10 µM RI, 10 µM CAY94 or with the combination of the inhibitors for 1 h (a) or 3 h (b), and whole-cell lysates were analyzed by MS. The figure shows relative amount of major DAG, PA and PG species and the equivalent PC and PI species. The error bars show mean deviation, n = 2. For detailed description of the calculations performed on MS data see section “Data analysis and post-processing” in the Materials and methods section and the abbreviations of all lipid classes are given under the section “Lipid extraction for MS lipidomics” in the Materials and methods section It should be noted that after 3 h of inhibitor treatment, one of the major DAG species, 16:0–16:0, is completely lost from control and RI-treated samples. During inhibitor treatment, cells are serum-starved. Our preliminary data show that incubations in the absence of serum reduced the cellular DAG levels in a time-dependent manner and that 1 h and 3 h in the absence of serum gave similar DAG levels as found in the cells treated with DMSO for 1 and 3 h (Fig. 3 and Supplementary material 2 and 3). Interestingly, treatment with the PLD inhibitor CAY94 gave similar DAG levels as in cells grown in complete medium. The DAG species most affected by removal of serum were 16:0–16:0 and 16:0–18:1 (Supplementary material 3), the same species that were most changed by CAY94 treatment (Fig. 4). Given the surprising effects of the PLD inhibitor on the cellular PA level, we wanted to verify that the inhibitor actually inhibits PLD activity. To this end, we used a modified version of the recently described IMPACT (Imaging Phospholipase D Activity with Clickable alcohols via Transphosphatidylation) method to measure PLD activity in living cells . Basal PLD activity depends both on PLD1 and PLD2, while phorbol ester PMA stimulates mainly PLD1 . We tested the inhibitor efficiency on both basal and PMA-stimulated PLD activity and found that at a concentration of 10 µM, CAY94 inhibited both the basal and PMA-stimulated PLD activity, indicating that at this concentration, CAY94 inhibits both PLD1 and PLD2 (Fig. S4). Thus, the increase in PA seen in the lipidomics assay is not due to inefficient PLD inhibition. Based on our sulfation data, both CAY93 and CAY94 increased ricin transport (Fig. 1b). The IC50 values of CAY93 in cells are 11 nM for PLD1 and 1.8 µM for PLD2, and for CAY94, the IC50 values are 1 µM for PLD1 and 110 nM for PLD2 . The 10 µM concentration of CAY93 and CAY94 should therefore inhibit both PLD1 and PLD2, while at 1 µM they should be isoform specific. In agreement with the reported IC50 values, 1 µM CAY93 efficiently blocked PMA-induced PLD activity, but had only a slight reduction in basal PLD activity, whereas 1 µM CAY94 efficiently reduced basal PLD activity and had no effect on PMA-induced PLD activity (Fig. S4). At 1 µM concentrations, CAY93 and CAY94 gave only a slight increase (15–40%) in ricin-sulf1 sulfation, suggesting that both PLD1 and PLD2 need to be inhibited to strongly stimulate ricin transport to the Golgi (Fig. S5). To investigate whether PLD1 and PLD2 need to be inhibited, we also tested the dual PLD1 and PLD2 inhibitor FIPI. At 1 µM concentration, FIPI strongly inhibited both basal and PMA-stimulated PLD activity (Fig. S4), in agreement with published IC50 values in a nanomolar range for both PLD1 and PLD2 . In ricin-sulf1 sulfation assays, FIPI gave similar results as CAY93 and CAY94 when used at 1 or 10 µM (Fig. S5). The CAY inhibitors and FIPI inhibits PLD via binding to its HKD domain. We also tested a new PLD inhibitor VU 0364739 (VU036), which in addition to binding to the HKD domain, also interacts with an allosteric site of PLD and has IC50 values 1.5 µM for PLD1 and 20 nM for PLD2 in cells . At 1 µM concentration, VU036 effectively inhibited basal PLD activity but had no effect on PMA-induced PLD activity, while at 10 µM concentration, VU036 blocked both basal and PMA-induced PLD activity (Fig. S6a,b) which is in agreement with the published IC50 values . When tested in the ricin sulfation assay, VU036 gave very similar results as CAY94, with no or very little effect at 1 µM concentration and a high increase in ricin sulfation at 10 µM (Fig. S6c). Based on The Human Protein Atlas, HeLa cells, which have a similar karyotype as HEp-2 cells , express four isoforms of PLD (PLD1, PLD2, PLD3 and PLD6), and seven isoforms of DGK (α, β, δ, ε, ζ, η and θ) [36, 37]. Therefore, we chose to test the expression levels of three PLD isoforms, PLD1, PLD2 and PLD3 (we did not include the mitochondrial PLD6), and six DGK isoforms, α, δ, ε, ζ, η and θ (we did not include β, which is enhanced in brain tissue and has very low expression in HeLa cells ). Gene expression was analyzed by qPCR and we found that HEp-2 cells have high expression of PLD3 and lower expression of PLD1 and PLD2 (Fig S7), which is similar to the expression pattern of PLDs in HeLa cells. It should be mentioned that the primer efficiency for PLD2 was lower than for the other primers, leading to higher Cp values and underestimation of the expression level. For the DGK expression in HEp-2 cells, the highest mRNA level was observed for DGKδ (Fig S7), which is similar to HeLa cells. However, the expression of some DGK isoforms seems to differ between HEp-2 and HeLa cells: DGKθ is one of the abundant DGK isoforms in HeLa cells, while it was least expressed of all of the analyzed DGKs in HEp-2 cells (Fig. S7). On the contrary, we found DGKε to be the second most abundant of the six DGK isoforms tested in HEp-2 cells, while in HeLa cells, it has lowest expression of the six . Based on the expression pattern of PLD and DGK isoforms in HEp-2 cells, we chose to knock down PLD1, PLD2, PLD3, DGKα, DGKδ, DGKε, DGKζ and DGKη by siRNA for 48 h and then analyzed ricin transport in these cells. Knockdown of PLD1 or PLD3 did not have a significant effect on ricin sulfation, while knockdown of PLD2 led to a significant reduction in the sulfation of ricin (Fig. 5a), without affecting its binding or uptake (Fig S8). The knockdown efficiency was more than 90% for all three siRNAs, and the downregulation of one isoform also affected the mRNA levels of other isoforms but to a lower extent than the targeted isoform (Fig. 5b). Since our inhibitor data indicated that more than one PLD needs to be inhibited to increase ricin transport to the Golgi, we also tested double and triple knockdown of the PLDs. Similar to knockdown of PLD2 only, combined knockdown of PLD1 and PLD2 gave a significant reduction in ricin transport, while combined knockdown of PLD1 and PLD3 led to varying increase in ricin sulfation (Fig. 5c). Importantly, the increase in ricin sulfation after the double knockdown of PLD1 and PLD3 correlates with the knockdown efficiency of PLD1 and PLD3: higher knockdown efficiency of PLD1 and PLD3 led to higher increase in ricin sulfation (Fig. 5d,e), which, together with the lack of effect on ricin sulfation after single knockdown of PLDs, shows that both of these enzymes need to be efficiently knocked down to increase ricin transport to the Golgi.Fig. 5Ricin transport to the Golgi after knockdown of different PLD isoforms. a, b HEp-2 cells were transfected with 10 nM siRNA against PLD1, PLD2 o PLD3, or with 10 nM control siRNA for 48 h and they were then subjected to sulfation assay with RS1 or analyzed for mRNA levels using qPCR. a The upper figure shows an autoradiogram from one representative experiment, and the lower figure shows the quantification of RS1 sulfation and total protein sulfation expressed as percent of control (n = 3). b mRNA levels of different PLD isoforms normalized to reference gene (TBP) and expressed as percent of control. Red color indicates the isoform which was targeted by the siRNA used in the sample. c–e HEp-2 cells were transfected with combination of either two siRNAs against PLD1, PLD2 o PLD3 or all three siRNAs (10 nM each), or with 20 nM or 30 nM control siRNA for 48 h and they were then subjected to sulfation assay with RS1 or analyzed for mRNA levels using qPCR. c The upper figure shows an autoradiogram from one experiment, and the lower figure shows the quantification of RS1 sulfation and total protein sulfation expressed as percent of control (n = 3 for double knockdown and n = 2 for triple knockdown). d mRNA levels of different PLD isoforms normalized to reference gene (TBP) and expressed as percent of control. Red color indicates the isoforms which were targeted by the siRNAs used in the sample. e The figure shows the dependence between mRNA levels for PLD1 and PLD3 and RS1 sulfation in the cells treated with the combination of siRNA against PLD1 and PLD3. As shown, the best knockdown of PLD1 and PLD3 gives the highest RS1 sulfation Ricin transport to the Golgi after knockdown of different PLD isoforms. a, b HEp-2 cells were transfected with 10 nM siRNA against PLD1, PLD2 o PLD3, or with 10 nM control siRNA for 48 h and they were then subjected to sulfation assay with RS1 or analyzed for mRNA levels using qPCR. a The upper figure shows an autoradiogram from one representative experiment, and the lower figure shows the quantification of RS1 sulfation and total protein sulfation expressed as percent of control (n = 3). b mRNA levels of different PLD isoforms normalized to reference gene (TBP) and expressed as percent of control. Red color indicates the isoform which was targeted by the siRNA used in the sample. c–e HEp-2 cells were transfected with combination of either two siRNAs against PLD1, PLD2 o PLD3 or all three siRNAs (10 nM each), or with 20 nM or 30 nM control siRNA for 48 h and they were then subjected to sulfation assay with RS1 or analyzed for mRNA levels using qPCR. c The upper figure shows an autoradiogram from one experiment, and the lower figure shows the quantification of RS1 sulfation and total protein sulfation expressed as percent of control (n = 3 for double knockdown and n = 2 for triple knockdown). d mRNA levels of different PLD isoforms normalized to reference gene (TBP) and expressed as percent of control. Red color indicates the isoforms which were targeted by the siRNAs used in the sample. e The figure shows the dependence between mRNA levels for PLD1 and PLD3 and RS1 sulfation in the cells treated with the combination of siRNA against PLD1 and PLD3. As shown, the best knockdown of PLD1 and PLD3 gives the highest RS1 sulfation The single knockdown of all tested DGK isoforms had a significant effect on ricin sulfation: the knockdown of DGKα, DGKδ and DGKε gave an increase in ricin sulfation, while the knockdown of DGKζ and DGKη led to a reduction in ricin sulfation (Fig. 6a). The knockdown of DGKα also gave a significant reduction in total protein sulfation (Fig. 6a) and also led to increased expression of all other isoforms of the DGKs (Fig. 6b), making it difficult to say whether DGKα has a direct effect on ricin transport. Since DGKα and DGKε are the main targeted isoforms by RI , we also tested whether combined knockdown of the two could lead to an even higher increase in ricin transport to the Golgi. However, the combined knockdown of DGKα and DGKε did not give significantly higher increase in ricin sulfation than the knockdown of DGKε alone (Fig. 6c).Fig. 6Ricin transport to the Golgi after knockdown of different DGK isoforms. a, b HEp-2 cells were transfected with 10 nM siRNA against DGKα (DGKA), DGKδ (DGKD), DGKε (DGKE), DGKη (DGKH), DGKζ (DGKZ) or with 10 nM control siRNA for 48 h and they were then subjected to sulfation assay with RS1 or analyzed for mRNA levels using qPCR. a The upper figure shows an autoradiogram from one experiment, and the lower figure shows the quantification of RS1 sulfation and total protein sulfation expressed as percent of control (n = 3; except for DGKA: n = 5). b mRNA levels of different DGK isoforms normalized to reference gene (TBP) and expressed as percent of control. Red color indicates the isoform which was targeted by the siRNA used in the sample. c–e HEp-2 cells were transfected with 10 nM siRNA against DGKα (DGKA), DGKε (DGKE), or the combination of the two. As a control, the cells were transfected with 20 nM control siRNA and 10 nM of control siRNA was also added to single treatment samples to have 20 nM siRNA in all wells. After 48 h, the cells were subjected to sulfation assay with RS1 or analyzed for mRNA levels using qPCR. c The upper figure shows an autoradiogram from one experiment, and the lower figure shows the quantification of RS1 sulfation and total protein sulfation expressed as percent of control (n = 3). d mRNA levels of different DGK isoforms normalized to reference gene (TBP) and expressed as percent of control. Red color indicates the isoforms which were targeted by the siRNAs used in the sample Ricin transport to the Golgi after knockdown of different DGK isoforms. a, b HEp-2 cells were transfected with 10 nM siRNA against DGKα (DGKA), DGKδ (DGKD), DGKε (DGKE), DGKη (DGKH), DGKζ (DGKZ) or with 10 nM control siRNA for 48 h and they were then subjected to sulfation assay with RS1 or analyzed for mRNA levels using qPCR. a The upper figure shows an autoradiogram from one experiment, and the lower figure shows the quantification of RS1 sulfation and total protein sulfation expressed as percent of control (n = 3; except for DGKA: n = 5). b mRNA levels of different DGK isoforms normalized to reference gene (TBP) and expressed as percent of control. Red color indicates the isoform which was targeted by the siRNA used in the sample. c–e HEp-2 cells were transfected with 10 nM siRNA against DGKα (DGKA), DGKε (DGKE), or the combination of the two. As a control, the cells were transfected with 20 nM control siRNA and 10 nM of control siRNA was also added to single treatment samples to have 20 nM siRNA in all wells. After 48 h, the cells were subjected to sulfation assay with RS1 or analyzed for mRNA levels using qPCR. c The upper figure shows an autoradiogram from one experiment, and the lower figure shows the quantification of RS1 sulfation and total protein sulfation expressed as percent of control (n = 3). d mRNA levels of different DGK isoforms normalized to reference gene (TBP) and expressed as percent of control. Red color indicates the isoforms which were targeted by the siRNAs used in the sample Both the DGK and the PLD inhibitor resulted in increased DAG levels, and thus we next investigated whether membrane recruitment of DAG-binding proteins could play a role in increasing sorting into the retrograde pathway. PKC is the most characterized DAG-binding protein and it has previously been shown that downregulation of PKCδ strongly reduced the retrograde transport of Shiga toxin . PKD, another DAG-binding protein, has been shown to be essential in the secretory pathway, where it activates PI4K in the Golgi apparatus and contributes to the generation of cell surface specific transport carriers [2, 40]. Activation of PKD was tested by Western blotting using an antibody that detects proteins containing phosphorylated Ser/Thr at the PKD consensus sequence. Treatment with the DGK inhibitor gave an increase in phosphorylated proteins (Fig. 7a), suggesting an increase in PKD membrane translocation and activity, in agreement with the increase in DAG seen in the lipidomics assay. Although treatment with the PLD inhibitor gave a stronger relative increase in DAG levels, it gave only a small increase in phosphorylated PKD substrates (Fig. 7a). PKC can act as an upstream activator of PKD and treatment with the PKC inhibitors bisindolylmaleimide I (BIM) and sotrastaurin (Sotra) strongly inhibited the RI-induced phosphorylation of PKD substrates (Fig. 7a). To check whether PKC activity was important for the increased retrograde transport of ricin after DGK or PLD inhibition, we measured ricin-sulf1 sulfation after treatment with RI or CAY94 in the presence or absence of PKC inhibitors. The PKC inhibitors did not affect the increase in ricin-sulf1 sulfation after DGK or PLD inhibition (Fig. 7b,c), indicating that it is not caused by increased PKC activity. Since the PKC inhibitors also prevented RI-mediated PKD activation, we reason that the DAG effectors PKC and PKD are dispensable for the up-regulation of ricin-sulf1-sulfation after DGK- and PLD inhibition.Fig. 7The RI- and CAY94-mediated increase in ricin transport is not dependent on the DAG substrate PKC. a The phosphorylation of PKD substrates was measured by Western blot after treating HEp-2 cells with 10 µM RI or 10 µM CAY94 in the presence or absence of the PKC inhibitors BIM I (BIM; 10 µM) and Sotrastaurin (Sotra; 1 µM) for 1 h. Cofilin was used as a loading control. One representative experiment is shown. b, c HEp-2 cells were subjected to sulfation assay with ricin-sulf1 (RS1) after treatment with 10 µM RI or 10 µM CAY94 in the presence or absence of the PKC inhibitors b BIM (10 µM) and c Sotra (1 µM) for 1 h. The upper figures show representative autoradiograms and the lower figures show quantifications of RS1 sulfation expressed as percent of control (n = 4) The RI- and CAY94-mediated increase in ricin transport is not dependent on the DAG substrate PKC. a The phosphorylation of PKD substrates was measured by Western blot after treating HEp-2 cells with 10 µM RI or 10 µM CAY94 in the presence or absence of the PKC inhibitors BIM I (BIM; 10 µM) and Sotrastaurin (Sotra; 1 µM) for 1 h. Cofilin was used as a loading control. One representative experiment is shown. b, c HEp-2 cells were subjected to sulfation assay with ricin-sulf1 (RS1) after treatment with 10 µM RI or 10 µM CAY94 in the presence or absence of the PKC inhibitors b BIM (10 µM) and c Sotra (1 µM) for 1 h. The upper figures show representative autoradiograms and the lower figures show quantifications of RS1 sulfation expressed as percent of control (n = 4) By altering the DAG levels, the inhibitors may also change the biophysical properties of the membrane, and a local increase in the concentration of DAG has been shown to affect both fission- and fusion processes [2, 42]. We therefore studied how treatment with the DGK and PLD inhibitors affect endosomal morphology. HEp-2 cells were treated with inhibitors and stained with antibodies against the endosomal marker EEA1 and investigated by immunofluorescence confocal microscopy. Both RI and CAY94 gave a significant increase in the size of EEA1-positive structures, and we also noticed a more irregular shape of the endosomes (Fig. 8a,b).Fig. 8Treatment with DGK and PLD inhibitors alters endosome morphology. a, b HEp-2 cells were treated with 10 µM RI or 10 µM CAY94 for 1 h and stained with antibodies against EEA1. a Representative images of EEA1-positive endosomes in fixed cells from confocal microscopy; scale bar, 10 µm. b Quantification of endosome size from three independent experiments. c–f RPE GFP-WDFY2 cells were treated with 10 µM RI, 10 µM CAY94 or with a combination of 10 µM RI and 10 µM CAY94 for 1 h. c Representative images of GFP-WDFY2 positive endosomes in fixed cells from confocal microscopy; scale bar, 10 µm. d The pipeline used for image analysis. e The number of small endosomes (classified as small in the object classification of GFP-positive pixels) and the number of endosomes with detected lumen (total number of objects in the object classification of luminal pixels) per cell area. f Distribution of medium, large and giant endosomes within the subpopulation of endosomes with detected lumen; mean values ± STD, n = 5 for DMSO 0.1%, RI and CAY94, n = 2 for DMSO 0.2% and RI + CAY94 Treatment with DGK and PLD inhibitors alters endosome morphology. a, b HEp-2 cells were treated with 10 µM RI or 10 µM CAY94 for 1 h and stained with antibodies against EEA1. a Representative images of EEA1-positive endosomes in fixed cells from confocal microscopy; scale bar, 10 µm. b Quantification of endosome size from three independent experiments. c–f RPE GFP-WDFY2 cells were treated with 10 µM RI, 10 µM CAY94 or with a combination of 10 µM RI and 10 µM CAY94 for 1 h. c Representative images of GFP-WDFY2 positive endosomes in fixed cells from confocal microscopy; scale bar, 10 µm. d The pipeline used for image analysis. e The number of small endosomes (classified as small in the object classification of GFP-positive pixels) and the number of endosomes with detected lumen (total number of objects in the object classification of luminal pixels) per cell area. f Distribution of medium, large and giant endosomes within the subpopulation of endosomes with detected lumen; mean values ± STD, n = 5 for DMSO 0.1%, RI and CAY94, n = 2 for DMSO 0.2% and RI + CAY94 To investigate whether inhibitors affect biophysical properties of endosomal membranes, we used an environment-sensitive probe NR12S, which has been proven useful in analyzing plasma membrane order in living cells [43, 44]. Recently, the NR12S probe has also been employed to probe the membrane packing in the endocytic recycling compartment, as it is taken up by non-selective endocytosis and delivered to the endocytic recycling compartment . Thus we employed NR12S to analyze whether PLD or DGK inhibition affects membrane packing in the endosomes. HEp-2 cells were incubated with Alexa647-labeled transferrin to mark early and recycling endosomes and then stained with the NR12S probe and imaged using a Zeiss LSM 780 confocal microscope (Fig. S9a). The mean generalized polarization (GP) value was quantified within the plasma membrane and endosome (transferrin-positive) mask as described in the Supplementary material 1. The GP value in the transferrin-positive endosomes was higher than the GP value in the plasma membrane (Fig. S9b), indicating higher lipid packing in the endosome membrane, which is in agreement with results in . Since the NR12S probe has a slow flipping rate across the bilayer , it should mainly be localized at the outer leaflet of the plasma membrane and the luminal leaflet of the endosomes in the time frame of the experiment. As shown, inhibitor treatment did not affect the GP value at the plasma membrane (at least at the outer leaflet, which is probed by NR12S), but there seemed to be an increase in the GP value at the endosomes after PLD inhibition (Fig. S9b) possibly mediated via interaction of the two leaflets of the membrane and/or flipping of lipids, such as DAG, from the cytosolic to the luminal side of the endosomal membrane. The absolute value of GP depends on the imaging settings and cannot be compared between individual experiments without calibration; therefore we calculated the difference between the GP value at the endosomes and the GP value at the plasma membrane for each cell (delta GP). The deltaGP values represent the difference in membrane packing between the plasma membrane and the endosomes, and can be used to combine data from individual experiments. Using this approach, we saw a significant increase in deltaGP in the PLD inhibitor treated cells, but no significant change after DGK inhibitor treatment (Fig. S9c). To study the endosomal morphology in more detail, we used an RPE1 cell line stably expressing GFP-WDFY2. WDFY2, which interacts with VAMP3, was recently described as an endosomal protein localizing to endosomal tubules and found to regulate endosomal sorting . Under conditions of mild overexpression, WDFY2 also labels the limiting membrane of the endosome, allowing us to study endosomal morphology. After treatment with RI and CAY94, we saw an increase in endosome size and clustering (Fig. 8c). To quantify this increase, Ilastik software was used for automated pixel and object classification, allowing us to quantify the number of small, medium, large/clustered and giant/clustered endosomes, as shown in Fig. 8d. The number of small endosomes was similar in control- and CAY94-treated cells, but was reduced after RI treatment (Fig. 8e), whereas the number of endosomes with visible lumens was strongly increased after both RI and CAY94 treatment (Fig. 8e). The distribution of medium, large and giant endosomes within the subpopulation of endosomes with visible lumens was also changed after RI and CAY94 treatment. CAY94-treatment gave a larger proportion of giant endosomes than RI and control treatment, whereas both RI and CAY94-treatment increased the proportion of large endosomes (Fig. 8f). Interestingly, cells treated with the combination of RI and CAY94 seem to have the combination of the morphological changes induced by the inhibitors individually: many large endosomes that are more clustered than in cells treatment with CAY94 alone (Fig. 8c). To test whether there is a possible link between changes in the endosome morphology and the increase in ricin transport, we analyzed WDFY2-positive endosome size in the cells treated with two different concentrations (1 µM and 10 µM) of CAY94 and FIPI. As shown in Fig. S10, at 1 µM concentration, PLD inhibitors did neither affect the number of endosomes with detected lumen, nor the endosomes size distribution, while at 10 µM concentration, both CAY94 and FIPI gave a clear increase in the number of endosomes with lumen as well as an increase in their size. It seems that the changes in the endosomal size correlate with the increase in ricin sulfation, as 10 µM is required to give a strong increase in the ricin transport (Fig. S10). Retrograde cargo is sorted towards the Golgi from tubular endosomal structures, therefore we also investigated whether the inhibitors increased endosomal tubulation. To this end, we looked at tubule dynamics by live-cell imaging of RPE GFP-WDFY2 cells. Tubulation events were more frequent in RI and CAY94-treated cells than in control-treated cells and we could observe long tubular carriers being released from endosomes travelling through the cytosol (Fig. 9a and Supplementary material 4). For quantification, we fixed cells and performed 3D scans of the cells to capture all visible tubules (Fig. 9b). First, we determined the proportion of cells with visible tubular structures by manual inspection. After CAY94 treatment, all cells were positive for tubular structures, and these cells also had a higher number of long tubular structures (> 0.7 µm) and a higher mean length of the tubules, while treatment with RI increased only the number of long tubular structures per cell without affecting the mean length (Fig. 9c). The combined treatment of RI and CAY94 resembled treatment with CAY94 alone, but induced more clustered endosomes and thus slightly shorter tubules (Fig. 9c). Treatment with DGK and PLD inhibitors increases endosome size and tubulation, but to a different extent and lead to different morphology, indicating that these inhibitors affect the endosome morphology via different pathways.Fig. 9Treatment with DGK and PLD inhibitors increases endosome tubulation. a RPE GFP-WDFY2 cells were treated with 10 µM RI or 10 µM CAY94 for 1 h and imaged each 2 s for 2 min. Example still images are shown with zoomed-in time series underneath; the arrow heads indicate forming WDFY2-positive tubules; scale bars, 10 µm (main images) and 2 µm (zoomed-in images). Example videos are given in Supplementary file 4. b, c RPE GFP-WDFY2 cells were treated with 10 µM RI, 10 µM CAY94 or with a combination of 10 µM RI and 10 µM CAY94 for 1 h, fixed, mounted and imaged using 3D scanning. b Representative maximum intensity projections from the 3D scans are shown with zoomed-in example images; the arrow heads indicate WDFY2-positive tubules; the star (*) indicates clustered tubules (thick WDFY-2 positive tubules) characteristic for the CAY94 treated cells. c All visible tubules were measured and classified as long if were of at least 0.7 µm. The figures show the percentage of cells with visible long tubules, the number of long tubules per cell and the mean length of the long tubules. The empty shapes show the means for each independent experiment (n = 4 for DMSO 0.1%, RI and CAY, n = 2 for DMSO 0.2% and RI + CAY94); the box plots show pulled data of all independent experiments (at least 25 cells were analyzed for each condition in each experiment) Treatment with DGK and PLD inhibitors increases endosome tubulation. a RPE GFP-WDFY2 cells were treated with 10 µM RI or 10 µM CAY94 for 1 h and imaged each 2 s for 2 min. Example still images are shown with zoomed-in time series underneath; the arrow heads indicate forming WDFY2-positive tubules; scale bars, 10 µm (main images) and 2 µm (zoomed-in images). Example videos are given in Supplementary file 4. b, c RPE GFP-WDFY2 cells were treated with 10 µM RI, 10 µM CAY94 or with a combination of 10 µM RI and 10 µM CAY94 for 1 h, fixed, mounted and imaged using 3D scanning. b Representative maximum intensity projections from the 3D scans are shown with zoomed-in example images; the arrow heads indicate WDFY2-positive tubules; the star (*) indicates clustered tubules (thick WDFY-2 positive tubules) characteristic for the CAY94 treated cells. c All visible tubules were measured and classified as long if were of at least 0.7 µm. The figures show the percentage of cells with visible long tubules, the number of long tubules per cell and the mean length of the long tubules. The empty shapes show the means for each independent experiment (n = 4 for DMSO 0.1%, RI and CAY, n = 2 for DMSO 0.2% and RI + CAY94); the box plots show pulled data of all independent experiments (at least 25 cells were analyzed for each condition in each experiment) Retrograde cargo destined to the Golgi can be segregated into tubular structures by retromer and the SNX-BAR complexes . To analyze whether ricin is transported via WDFY2-positive endosomes and/or tubules, we treated RPE-GFP-WDFY2 cells with the inhibitors and then added ricin-Alexa555 prior to cell fixation. The cells were immunolabeled for the retromer component Vps35 and the SNX-BAR component SNX2 and imaged using fluorescence super-resolution imaging. Indeed, we clearly saw ricin inside the WDFY2-positive endosomes in control and in inhibitor treated cells. In addition, some of the WDFY2-positive tubules were also positive for ricin, although the signal for ricin was always strongest within the endosomal lumen (Fig. 10). Interestingly, CAY94 treatment often led to endosomes with thick tubular tails positive for WDFY2. Based on the 3D reconstruction, these structures look as multiple tubular structures extending from the endosome. Worth noticing, ricin was often accumulated at the base of such thick tubular structures (Fig. 10). Finally, both Vps35 and SNX2 decorated WDFY2-positive endosomes and tubules and they were often found located at the same position or close to the ricin signal within the endosome (Fig. 10).Fig. 10Ricin is in WDFY2-positive structures in close proximity to Vps35 and SNX2. RPE GFP-WDFY2 cells were treated with 10 µM RI or 10 µM CAY94 for 1 h followed by 20 min pulse with 1 µg/ml ricin-Alexa555. Then, lactose was added (final conc. 0.1 M) to remove surface bound ricin-Alexa555 and the cells were fixed, permeabilized, stained with antibodies against a Vps35 or b SNX2 and mounted in ProLong Diamond with DAPI. High resolution 3D images were acquired using Zeiss LMS 880 Airyscan microscope. Left: Representative 2D images after super-resolution processing; scale bar, 10 µm. Middle: Zoom in on endosomal structures; scale bar, 1 µm. Right: 3D visualization of the endosomal structures shown in the first row for each condition. Colors: green: WDFY2, red: ricin-Alexa555, blue: Vps35/SNX2, white: DAPI Ricin is in WDFY2-positive structures in close proximity to Vps35 and SNX2. RPE GFP-WDFY2 cells were treated with 10 µM RI or 10 µM CAY94 for 1 h followed by 20 min pulse with 1 µg/ml ricin-Alexa555. Then, lactose was added (final conc. 0.1 M) to remove surface bound ricin-Alexa555 and the cells were fixed, permeabilized, stained with antibodies against a Vps35 or b SNX2 and mounted in ProLong Diamond with DAPI. High resolution 3D images were acquired using Zeiss LMS 880 Airyscan microscope. Left: Representative 2D images after super-resolution processing; scale bar, 10 µm. Middle: Zoom in on endosomal structures; scale bar, 1 µm. Right: 3D visualization of the endosomal structures shown in the first row for each condition. Colors: green: WDFY2, red: ricin-Alexa555, blue: Vps35/SNX2, white: DAPI One important finding in the present study is that retrograde transport from endosomes to the Golgi apparatus of the protein toxin ricin can be strongly upregulated in HEp-2 cells by affecting lipid composition with inhibitors of DGK and PLD. A similar regulation of ricin transport was seen also in other cell lines, demonstrating that this is a general phenomenon. DGK depletion showed that different DGK isoforms affect ricin transport differently. Knockdown of the α, δ and ε isoforms increased retrograde ricin transport, whereas η and ζ knockdown led to a reduction. This is line with the results obtained with the DGK inhibitors, which mainly targets the α isoform. The data obtained with the different PLD inhibitors together with the siRNA knockdown of different PLD isoforms indicate that it is not sufficient to inhibit a single PLD isoform to get the increase in ricin transport. Like PLD1 and PLD2, PLD3 also has an HKD domain and we speculate that it could potentially be targeted by all of the PLD inhibitors used. It is not known yet whether PLD3 has enzymatic activity, but data based on siRNA knockdown of PLD3 indicate that it plays a role in intracellular sorting and might have enzymatic activity . However, the IC50 value for PLD3 might be higher than PLD1 or PLD2, thus requiring a higher inhibitor concentration. The combination of inhibitors of DGK and PLD had an even stronger effect than either one added alone, suggesting that different mechanisms are regulated by interfering with these lipid processing enzymes. The transport to the Golgi was monitored both by sulfation of genetically modified ricin and by immunofluorescence microscopy, and as expected the increase in retrograde transport was associated with an increased toxicity. Remarkably, the transport from the TGN to the ER did not seem to be affected by the inhibitors as the fraction of sulfated ricin that was mannosylated in the ER was unchanged. Since endocytosis, recycling and degradation of ricin were essentially unaffected, it seems that there is a selective effect on ricin transport between endosomes and the Golgi apparatus. We also investigated whether we could see a similar regulation on Golgi transport of Shiga toxin which binds to the glycosphingolipid Gb3, and interfering with DGK did have a significant effect on Golgi transport of Shiga toxin (about twofold), whereas there was essentially no change in transport after PLD inhibition. Thus, there is clearly selectivity in which pathway(s) that are changed. We have here chosen to focus on ricin transport not only because of the large change upon treatment with the different inhibitors, but also since ricin, in contrast to Shiga toxin, does not seem to induce signaling in cells . Shiga toxin, which crosslinks glycolipids at the cell surface, is able to induce signaling that in itself can change intracellular sorting . There are a number of pathways leading from endosomes to the Golgi apparatus, and several factors involved in sorting have been characterized [48, 49]. For many years there has been a discussion about whether Golgi transport has to go via recycling endosomes or whether it can occur by direct transport from early endosomes, and to which extent this might be cargo- and cell type- dependent . It has previously been reported that two proteins localized to the tubular recycling endosomes, EHD1 and EHD3, are involved in Shiga transport to the Golgi. These proteins interact with the PA-binding protein MICAL-L1, and it has been shown that both PLD inhibitors and knockdown of DGKα destroy these structures [50, 51]. The strong increase seen in Golgi transport of ricin after treatment with DGK and PLD inhibitors suggests that ricin does not depend on recycling endosomes for retrograde transport, and supports the idea that there is a regulation of transport from early endosomes . The lipid PI3P, which can be formed by the action of the kinase Vps34, is central for recruitment of several proteins to endosomes, including sorting nexins, such as SNX2 and SNX4. Transport of ricin to the Golgi is reduced after inhibition of Vps34 with inhibitors such as wortmannin and LY294002, by expression of Vps34 mutants and after knockdown of SNX2 and SNX4 , but it should be noted that in all these cases retrograde transport is not completely inhibited, a finding which may be due to the multiple pathways which ricin can utilize. Importantly, the upregulated transport of ricin after incubation with RI and CAY94 is to a large extent also inhibited by blocking PI3 kinase, indicating that PI3P is involved also in facilitating the strong increase in toxin going to the Golgi. Since PI3 kinase is important also for upregulated Golgi transport, we decided to study cells which express fluorescently labeled WDFY2, a PI3P binding protein that was recently shown to restrain matrix metalloproteinase secretion in RPE1 cells and to be enriched in actin-stabilized endosomal tubules. Such tubules are certainly candidates for mediating Golgi transport, since WDFY2 was also shown to interact with VAMP3 , a v-SNARE that has been implicated in Shiga toxin transport to the Golgi . By microscopy we could show that the inhibitors of DGK and PLD may affect both endosome size and tubulation, and ricin was found to be present in endosomal tubules with Vps35, a component of the retromer, as well as with SNX2, which is involved in Golgi transport of ricin . It has been suggested that actin-stabilization of tubules is required for efficient retrieval of slow-diffusing cargo from endosomes by allowing sufficient time for cargo accumulation in the tubules before they are pinched off . We observed more and longer tubules after inhibitor treatments, especially after PLD inhibition, which may give ricin more time to enter the tubules. Interestingly, it has been shown that cargo destined for the Golgi apparatus accumulate in retromer-positive subdomains of the early endosomes, whereas recycling cargo labeled the entire limiting membrane. Enrichment to these domains did not affect cargo itinerary, but instead increased the rate of endosome exit and subsequent delivery to the appropriate destination . We observed strong ricin staining at the base of WDFY2 tubules, which is in agreement with a large fraction of ricin being recycled. Ricin was also seen in the WDFY2-positive tubules, and since DGK and PLD inhibition increased the number of tubules, more ricin is likely to associate with retromer endosome subdomains, which may improve transport kinetics. It has been published that ligands destined for different locations can exit the endosome in the same tubule , and we cannot exclude that sorting at a later stage than tubule formation is also affected. In addition, the detailed mechanism for scission of tubules destined for the Golgi is still unknown and alterations in lipid composition may change scission kinetics and have different impact on Golgi-retrieval pathways. To investigate whether the inhibitors of DGK and PLD actually had an effect on the lipid composition of the cells, we analyzed the lipidome after similar incubations as when we studied toxin transport. It should be noted that our lipidomic studies both in earlier articles and as here shown, reveal that one cannot necessarily predict the changes that occur in cells upon a given treatment. We have previously by performing lipidomics discovered that lipid synthesis is regulated in ways that have still not been characterized. For instance, by adding a precursor to plasmalogens, the cells responded by also changing synthesis of glycosphingolipids . The data obtained here show that whereas the DGK inhibitor RI as expected gave a small increase in DAG levels without affecting any of the other lipid classes measured, surprising effects on the lipidome were obtained when using the PLD inhibitor CAY94: DAG was strongly increased after CAY94 treatment, and the PLD product PA, as well as PG, were also increased. DAG, PA and PG all had a similar lipid species distribution as PC, suggesting that PC is more likely than PI to be the source of these lipids. Importantly, both inhibitors increased the levels of DAG, which has a conical shape and thus might promote tubulation. The reason for the observed increase in PA after the treatment with CAY94 is not known, but we speculate that it is important for the cell to keep PA as constant as possible and that the cells compensate by increasing the activity of PC-PLC, thus generating more DAG which in turn in transferred to PA. An increase from 0.81 to 1.57% DAG is associated with an increase in PA from 0.20 to 0.33% during the first hour, but the PA level is normalized within the next two hours. PG increases from 0.44 to 0.83% during the first hour and increases further to 1.15% during the next two hours. PG is synthesized from PA via CDP-DAG (CDP: cytidine diphosphate) to PGP (phosphorylated PG) to PG . Thus, we speculate that the cell compensates for the inhibition of PLD by making PA via DAG and then transfers the extra PA formed to make PG to normalize the level of PA after 3 h. The combination of CAY94 and RI gives similar increase in PA as CAY94 alone, however, the speculations about the effects of CAY94 above may still be correct since RI is not targeting all DGK isoforms . If DGK inhibition is not complete, it seems likely that one can obtain data such as those described above, i.e. that an increase from 0.75 to 1.5% DAG can give an increase of PA from 0.15 to 0.30%. To test whether the increase in DAG, PA and PG upon CAY94 treatment comes from PC, one should specifically block PC-PLC activity in combination with CAY94. Although the PC-PLC activity has been known for years, the mammalian PC-PLC has not been purified and the gene sequence has not yet been resolved . The only commercially available pharmacological inhibitor of PC-PLC (D609) has several off-target effects , including inhibition of sphingomyelin synthase at similar concentrations as used to inhibit PC-PLC , thus at this point, we are not able to test our hypothesis. It should be noted that a challenge of using lipidomics to study endosomal sorting is that we get the sum of lipids in all cellular compartments. Lipids are sorted along the endosomal pathway and the lipid composition in different organelles may be different from the overall cellular composition. Lipid-modulating enzymes may also have subcellular specificity, and local changes in the lipid composition of endosomes may not be detectable at the cellular level. As discussed above, one role of the various membrane lipids is to recruit proteins, and there may be a change in membrane association of sorting nexins, the WASH complex and actin. We have previously described how inhibition of glycosphingolipid synthesis may change the localization of SNX1 and SNX2 . However, also the lipids themselves have been shown to be important for fusion; that is lipids that are prone to non-bilayer structure such as DAG . One common effect of RI and CAY94 treatment is that there is an increase in DAG 16:0–18:1. However, with the limited knowledge about the role of different lipid species in cells, more lipidomics has to be included in future studies to understand the complexity of membrane lipids in cell physiology.
|
PMC8481471
|
Mouse lipidomics reveals inherent flexibility of a mammalian lipidome
|
Lipidomics has become an indispensable method for the quantitative assessment of lipid metabolism in basic, clinical, and pharmaceutical research. It allows for the generation of information-dense datasets in a large variety of experimental setups and model organisms. Previous studies, mostly conducted in mice (Mus musculus), have shown a remarkable specificity of the lipid compositions of different cell types, tissues, and organs. However, a systematic analysis of the overall variation of the mouse lipidome is lacking. To fill this gap, in the present study, the effect of diet, sex, and genotype on the lipidomes of mouse tissues, organs, and bodily fluids has been investigated. Baseline quantitative lipidomes consisting of 796 individual lipid molecules belonging to 24 lipid classes are provided for 10 different sample types. Furthermore, the susceptibility of lipidomes to the tested parameters is assessed, providing insights into the organ-specific lipidomic plasticity and flexibility. This dataset provides a valuable resource for basic and pharmaceutical researchers working with murine models and complements existing proteomic and transcriptomic datasets. It will inform experimental design and facilitate interpretation of lipidomic datasets.Lipidomics is the quantitative and comprehensive analysis of lipids in biological samples. By means of automated sample extraction, state-of-the-art mass spectrometry, sophisticated and validated spectra annotation and data analysis processes as well as modern statistical methods (machine learning), biologically relevant information can readily and reliably be obtained based on thousands of data points. Therefore, lipidomics has become an indispensable tool for understanding (lipid) metabolism at the molecular level in basic, clinical, and pharmaceutical research. In the clinical context, the analysis of human blood samples in large population studies have revealed novel lipid signatures for a variety of indications such as: diabetes type 2, obesity, cardiovascular diseases and neurological disorders. Furthermore, in nutritional research lipidomics has proven a powerful read-out in dietary intervention studies. Lipidomics has successfully been applied to study disease and disease-related mechanisms in many different indications. Among these are neurological disorders, liver disease and cancer, resulting in the identification of potential drug targets involved in lipid metabolism. A primary model for these studies is the mouse Mus musculus. Typically, lipidomic changes in blood plasma and a variety of tissues or organs are assessed to understand disease mechanisms or modes of drug action. Several studies have shown that different cells, tissues, and organs exhibit highly specific lipid composition. However, the systematic knowledge of the natural, biological lipidomic variation of different organs and bodily fluids is lacking, hindering interpretation of lipidomic data. To fill this gap, in the present study, the effect of diet, sex, and genotype on the lipidomes of mouse tissues, organs, and bodily fluids has been investigated. Baseline quantitative lipidomes consisting of more than 796 individual lipid molecules for 10 different sample types (full blood, blood plasma, liver, skeletal muscle, brain, kidney, adipose tissue, small intestine, lung, and spleen) are provided. Furthermore, susceptibility of lipid levels to the tested parameters is assessed, providing insights into the organ-specific lipidomic plasticity and flexibility of a mammalian organism. This dataset provides a valuable resource for basic, pharmaceutical, and clinical researchers using mouse as model system and complements existing proteomic and transcriptomic datasets. It will inform experimental design and facilitate interpretation of lipidomic datasets. The aim of the present study is to investigate the lipidomes of mouse tissues, organs, and bodily fluids and how they are affected by different diets in mice of different genotype and sex. To this end, two standard laboratory mouse strains were selected, representing genetically outbred and inbred populations. Parental mice were allowed to breed, and females (already during pregnancy) and the consecutive litter mice were fed with a high protein (18 weight%) or a low protein (14 weight%) diet; both diets representing healthy and standard compositions and commonly used in nutrition studies. For each combination of conditions, three mice of each sex were sacrificed, and 10 sample types collected, prepared for lipid extraction, and analysed using a quantitative, high-throughput shotgun lipidomics platform. Sample types included: full blood, blood plasma, liver, skeletal muscle, brain, kidney, adipose tissue, small intestine, lung, and spleen; in total 240 samples. For details see Materials & Methods. The present study design enabled a factorial analysis of the influence of the different conditions (i.e., genotype, diet, and sex) on each individual lipid in every sample type. The comprehensive set of sample types allowed for a detailed description of mouse organ lipidomes regarding abundances of lipid classes and individual lipid molecules, fatty acid saturation/unsaturation, chain length and hydroxylation profiles as well as their susceptibility to differences in the experimental conditions. Furthermore, changes of lipids in different organs or tissues across the different conditions could be correlated with changes in blood lipidomes (full blood or blood plasma). This allowed to investigate, whether blood lipidomes are useful proxies for various organ and tissue lipidomes. We have published detailed methods for mass spectrometry-based shotgun lipidomics of blood plasma and adipose tissue. We have shown that for these two matrices the analytical precision (repeatability) is well below 15% (relative standard deviation, RSD), sensitivity is in the nM range and the linear dynamic range spans four orders of magnitude. To further extend method validation for the present study, liver, brain, and full blood were chosen as representative material to assess repeatability, sensitivity, and dynamic range of the analytical method. These sample types are the most complex and exemplary regarding lipid composition and quantities to be encountered across various tissues, organs, and other samples. More specifically, brain is a very lipid-rich organ with high amounts of (glyco-)sphingolipids. Liver, as a metabolically highly active organ, is rich in apolar lipids like triglycerides and cholesterolesters in addition to polar membrane lipids, making liver an especially complex matrix for lipidomics analyses. Full blood represents a lipid-rich liquid sample containing large amounts of membrane lipids (mainly derived from erythrocytes) and lipoprotein-derived storage lipids such as triglycerides and cholesterolesters. The method sensitivity assessed for these sample types was in the nM range and linearity (dynamic range) was achieved over four orders of magnitude. Repeatability of the method was 5.6%, 6.3%, and 8.3% RSD for liver, brain, and full blood samples, respectively (for details see Supplemental Text). These values were similar to the repeatability achieved by shotgun lipidomics for other sample types: plasma, skin stratum corneum and adipose tissue, with 11.3% (mean), 7.4% (median) and 6.8% (median) RSD, respectively. Such concurrence of reproducibility values for the representative samples of this study and for other sample types in previously published data supports the conclusion that such reproducibility is representative for a shotgun lipidomics method and can be expected for other sample types as well. The achieved technical reproducibility placed the measurements well below the 20% RSD threshold that is commonly used for in vitro diagnostic assays. The sample set was analyzed in four analytical batches, according to the required extraction and mass spectrometry methods (see Materials & Methods). The analytical performance was assessed by including reference samples for each analytical batch. The median RSD for the analytical batches on the level of individual lipid (sub)species for plasma was 6.6%, for full blood 15.7%, for adipose tissue 4.9%, and for the remaining tissues 12.3%. For the study samples, a total of 24 lipid classes comprising a total of 796 lipid species could be quantified across all sample types (Table 1). Among these, spleen and lung exhibited the broadest coverage (24 lipid classes), while in adipose tissue the lowest number (11) of lipid classes could be detected. The reproducibility as assessed by the median RSD for lipid (sub)species of the biological triplicates of each organ and combination of experimental conditions was between 12.5% and 25.8%. Reproducibility was highest for brain and full blood and lowest for kidney and spleen. Note, that these reproducibility values include both technical variation of the analytical method and biological variation between the triplicate mice.Table 1Lipidome coverage and reproducibility of the measurements. Median RSD is provided for individual lipid (sub)species in the biological triplicates.Sample typeLipid classesLipid speciesMedian RSD (%)Adipose tissue1115016.5Brain2236312.5Full blood1841012.5Intestine2342020.2Kidney2352125.8Liver2040516.4Lung2444217.6Muscle2144721.6Plasma1730614.1Spleen2453123.0 Lipidome coverage and reproducibility of the measurements. Median RSD is provided for individual lipid (sub)species in the biological triplicates. To obtain an overview of the similarities between different sample types, a principal component analysis (PCA) was performed based on amounts of lipid species normalized to total lipid for all experimental conditions (molar fractions expressed as mol%, Fig. 1A). Unsurprisingly, almost every sample type forms a separate cluster, indicating a distinct, organ-specific lipid composition. Notable exceptions are adipose tissue and muscle, which are not discriminated well in the first two principal components of the PCA. This lipidomic similarity of adipose and muscle tissues, despite their apparent biological difference, is caused by the fact that both samples contained high ( mol%) amounts of triglycerides (TAG). As this observation is obvious for adipose tissue, for the muscle it is most likely caused by not differentiating between muscle tissue and the intermuscular fat during sample collection. Brain appears to have the most distinct lipid composition of all organs and tissues analyzed, as its samples are clustered well separated from other sample types, in both principal component dimensions (principal components 1 and 2). Furthermore, along principal component 1, brain and adipose tissue/muscle are most distant to each other, indicating distinct compositional differences due to the high concentrations of TAG in muscle and adipose tissue, which on the other hand is almost entirely absent in brain. When the storage lipids (cholesterolester (CE), diacylglycerol (DAG), TAG) are excluded from the PCA, the tissue-specific clusters are in most cases preserved (Fig. 1B). This is the case for brain, kidney, intestine, spleen, and lung. For the remaining tissues, the clusters become even more distinct arguing for a tissue-specific composition of the membrane lipids in particular. This effect is most pronounced for adipose and muscle tissue which overlap entirely in the presence of storage lipids but form separate clusters based on membrane lipids. On the lipid class level, brain exhibits high concentrations of cholesterol (ST), phosphatidylserine (PS), and most importantly, phosphatidylethanolamine ether (PE O-) and hexosyl ceramide (HexCer) (Fig. 1C). Intestine samples form a distinct cluster separated from the other sample types in principal component 2. They show high concentrations of phosphatidylcholine ether (PC O-), lyso-phosphatidylcholine (LPC), ceramides (Cer), lyso-phosphatidylethanolamine (LPE), and other lyso-lipid classes. With around 5 mol%, the LPE concentration in the intestine is by far the highest among all tissues and organs. Lung samples form another well-separated cluster. They are rich in sterols, sphingomyelin (SM) and in particular phosphatidylglycerol (PG). Like lung, spleen is rich in cholesterol (ca. 25 mol%, Fig. 1C). In addition, in spleen there are significant amounts of PE O-, PC O-, phosphatidylinositol (PI), PS, and SM. The blood samples (full blood and plasma) obviously share characteristic similarities, such as high concentrations of CE and LPC. Full blood differs from plasma because it contains higher concentrations of phosphatidylethanolamine (PE), PE O-, PI, and PS, which are contributed by the cellular components of blood, mainly erythrocytes. Interestingly, kidney and liver samples are forming partially overlapping clusters. Liver and kidney, both share a comparably high concentration of DAG and intermediate concentrations of TAG and cholesterol. Furthermore, liver has the highest PI concentration among all sample types. The tissue-specific lipid (sub-)species composition is based in part on a distinct fatty acid profile of the different sample types (Fig. 1D). In general, the three most abundant fatty acids are 16:0 (palmitic acid), 18:1 (likely oleic acid), and 18:2 (likely linoleic acid). Palmitic acid is most abundant in lung tissue, where it is a major component of surfactant lipids. Plasma and full blood samples contain the highest amounts of the poly-unsaturated fatty acid (PUFA) arachidonic acid (20:4) while brain is rich in docosahexaenoic acid (22:6). Brain also contains, together with intestine, comparably high amounts of stearic acid (18:0). Complete lipidomic data are provided in Supplemental Table s1.Figure 1Lipid composition of mouse organs. (A) Principal component analysis (PCA) based on lipid species in mol%. Axis labels indicate principal component 1 and 2, including % variance explained. (B) PCA as in A, but excluding storage lipids (DAG, TAG, CE). (C) Lipid class composition. (D) Profiles of fatty acids derived from complex lipids. Shown are mean values for all combinations of experimental conditions (biological triplicates for all combinations of diet, sex, and genotype; ). Error bars indicate standard deviations. Lipid composition of mouse organs. (A) Principal component analysis (PCA) based on lipid species in mol%. Axis labels indicate principal component 1 and 2, including % variance explained. (B) PCA as in A, but excluding storage lipids (DAG, TAG, CE). (C) Lipid class composition. (D) Profiles of fatty acids derived from complex lipids. Shown are mean values for all combinations of experimental conditions (biological triplicates for all combinations of diet, sex, and genotype; ). Error bars indicate standard deviations. Lipidome data can be presented in many ways. This is due to the fact, that lipids are built from structural entities that are shared among all lipid classes. This allows for grouping the lipidomic data according to different structural features. These structural features can be headgroups, the defining characteristic of lipid classes, or fatty acids with varying length (number of carbon atoms in an acyl moiety) and number of double bonds (unsaturation). Therefore, the lipid class abundance (as the molar fraction of the total measured lipidome, expressed in mol%), the weighted mean of the number of double bonds, and the weighted mean of the number of carbon atoms in the fatty acid moiety of the lipids belonging to a given lipid class are three lipidomic features that enable a quantitative as well as qualitative description of a given lipid class. This results in a reduced number of lipidomic features. For example, the most complex sample type in the present study (spleen) contains 531 lipid molecules belonging to 24 lipid classes. By describing a lipid class by the three parameters, the number of features (the data granularity) to be analyzed is reduced to 72 (24 lipid classes ), a seven-fold decrease. The advantage of condensing the lipidome data in this way is, that the number of features to be analyzed statistically can be decreased without losing the most important biological information, conveyed by lipid classes content and their length/saturation indices. This condensed qualitative view on the lipid class composition can be displayed by plotting the weighted mean of the number of double bonds and the weighted mean of the number of carbon atoms in the fatty acid moiety of the lipids belonging to a given lipid class against each other (Fig. 2). Unsurprisingly, for most lipid classes, there is a correlation between the carbon chain length and the degree of unsaturation: the longer the fatty acids, the more double bonds are to be found. However, this phenomenon is more pronounced in some classes than in others. For phosphatidylcholine (PC) and PG there seems to be a perfectly linear relationship between chain length and unsaturation. However, the sphingolipids (ceramide, SM and hexosyl ceramide) do not follow this correlation but exhibit rather organ-specific length and unsaturation profiles due to the differential expression of ceramide synthases. For example, SM has a very distinct profile in intestine samples with short (35–36 carbons) and saturated (1–1.1 double bond) SM species. In plasma however, SM is longer (ca. 38 carbons) and more unsaturated (ca. 1.5 double bond). A striking feature of the lipid class length and unsaturation profiles is that they appear to be more conserved in certain organs than in others. For example, especially brain exhibits a very distinct lipid class profile for the sphingolipids, PC, PE, PE O-, PI and PS. Similarly, lung shows a very tightly controlled lipid profile in PC and PG. Of note, lung PC exhibits the shortest and most saturated profile across all sample types analyzed. Intestine is another organ with rather specific lipid class profiles: its PI and PS are comparably short and saturated, while PE O- in intestine samples shows longer carbon chains (ca. 38) than in any other sample type. Sphingolipids like ceramide, SM, and hexosyl ceramide show a specific chain length and double bond profile in basically every organ. Other lipid classes have a more distinct lipid class composition in some organs but can vary broadly in terms of fatty acid chain length and degree of unsaturation in others. A noteworthy example is PC, which in brain and lung has a very specific lipid profile, while especially in spleen, but also in plasma, its lipid class composition is rather variable. Similar observations can be made for other lipid classes, such as PE, PE O-, and PS. One may generalize these observations and say that across all organs and tested conditions certain lipid classes exhibit a greater heterogeneity in their fatty acid composition (profile) than others. For example, PS shows a very broad double bond range (from 0 to almost 7) while PI shows a very narrow double bond range (from 3.2 to 4.1). Similarly, PC O- has a very broad carbon length range (from 32 to almost 38) while PE shows a very narrow carbon length range (from 36 to 39).Figure 2Feature profiles of the major lipid classes. Shown is the weighted mean number of double bonds per lipid class in a given sample and the weighted mean number of carbons in the hydrocarbon chain moiety per lipid class in a given sample (each dot represents an individual sample). Minor lipid classes (lyso-lipids) are omitted for clarity. Feature profiles of the major lipid classes. Shown is the weighted mean number of double bonds per lipid class in a given sample and the weighted mean number of carbons in the hydrocarbon chain moiety per lipid class in a given sample (each dot represents an individual sample). Minor lipid classes (lyso-lipids) are omitted for clarity. To express these observations in a simple parameter, we propose the concept of lipidomic plasticity. We define lipidomic plasticity as the ability of a lipid class to undergo changes in the composition of its molecular lipid species, that is changes in their features, such as fatty acid length, unsaturation or abundance within the class, in response to varying conditions. For example, a lipid class that is observed to assume a very broad range of double bond numbers, carbon chain length, or abundances across a variety of experimental conditions, has a high degree of lipidomic plasticity. On the contrary, a lipid class which exhibits hardly any variation in abundance, double bond profile or carbon chain length has a low degree of lipidomic plasticity. Hence, plasticity is the degree of structural heterogeneity/variation within a given lipid class. Therefore, lipidomic plasticity expresses the capability of an organism (such as mouse) or its part (an organ, a tissue, a cell), to adopt the amount and/or composition of lipidomic entities (such as lipid classes) in response to internal and external stimuli. The plasticity of a lipid class was estimated as the product of the range of the weighted mean of the number of double bonds and the weighted mean of the number of carbon atoms in the hydrocarbon moiety of a lipid class (Fig. 3A, for details of the calculation see Materials & Methods). As defined above, a lipid class that exhibits a wide range of unsaturation and/or carbon chain length will have a high plasticity, while a lipid class with low variability will have a low plasticity. In this way, it is possible to easily compare the capability for adjusting the species profile of a lipid class across the different organs. For example, PC has a high plasticity in spleen, while very low in intestine (Fig. 3A). This is explained by the fact that PC clusters tightly around 35–35.5 carbons in the fatty acid moiety and 2.5 double bonds in intestine, while in spleen the ranges are from ca. 33–34.5 for carbons in fatty acids and double bond numbers between 1 and 2 (Fig. 3A). The analysis of lipid class plasticity therefore reveals, that in some organs the fatty acid composition of certain classes is more constrained (i.e., less affected by diet, sex, genotype) than in others (Fig. 3B). In brain, but also in lung, the lipid class profiles are the least plastic. Other organs, like spleen, tend to have more plastic lipid classes profiles. Liver shows a similar tendency for many lipid classes, but not for all. In summary, along the parameters of lipid class abundance, carbon chain length and double bond profile, each lipid class exhibits a distinct quantitative and qualitative pattern in every organ. Here we introduce plasticity as an additional defining characteristic of each lipid class.Figure 3Lipid class plasticity across all combinations of experimental conditions. (A) Plasticity is calculated per class and organ by multiplying the scaled ranges of values of weighted mean number of double bonds and the weighted mean number of carbons in the hydrocarbon chain moiety per lipid class. The wider the ranges, the higher the plasticity. (B) Lipid class plasticity for the different mouse organs. For details see Materials & Methods. Lipid class plasticity across all combinations of experimental conditions. (A) Plasticity is calculated per class and organ by multiplying the scaled ranges of values of weighted mean number of double bonds and the weighted mean number of carbons in the hydrocarbon chain moiety per lipid class. The wider the ranges, the higher the plasticity. (B) Lipid class plasticity for the different mouse organs. For details see Materials & Methods. As shown above, there is significant variation of the organ lipidomes across the tested conditions. An analysis of variance based on multiple linear regression with lipid species concentrations in mol% as outcome and sex, diet, genotype, and sample type as covariates shows that the major source of variance in the dataset is the sample type, accounting for about 60% of overall variance (not shown). This confirms the observations in the PCA (Fig. 1). Therefore, organ identity is the most determining factor for lipid composition. Diet, sex, and genotype contribute to only about 5% of overall variance each, indicating that, in the range tested, their influence is more subtle. To investigate effects of diet, sex, and genotype on the mouse organ lipidomes, multiple linear regression with lipidomic features (i.e., lipid class concentration in mol%, weighted mean of lipid class unsaturation, weighted mean of lipid class chain length) as outcome was performed. Out of 68 lipidomic features included in the analysis, 56 (82%) were significantly affected by any of the tested conditions in at least one of the organs. Those lipidomic features that were not affected represent only minor, less abundant, lipid classes, such as lyso-lipids. More specifically, every organ is affected by diet, sex, and genotype in at least one lipidomic feature. The only exception is brain, which is not affected by diet and sex. However, genotype has a highly specific effect on the brain sphingolipids and cholesterol (Fig. 4B). Cholesterol, SM, and hexosyl ceramide content is increased in outbred mice, while ceramide content is decreased. Moreover, unsaturation and carbon chain length are reduced for ceramides and SM in outbred mice, while the degree of unsaturation of hexosyl ceramide is increased. Complete results for the regression analysis are provided in Supplemental Table s3. To summarize the lipidomic changes induced by the experimental conditions and to quantitatively assess their impact on the mouse organ lipidomes, we apply the concept of lipidomic flexibility. We define lipidomic flexibility as a quantitative measure for the magnitude of changes in a lipidome (or parts of it) induced by certain experimental conditions. It can be calculated as sum of changes on various levels (for example degree of unsaturation, fatty acid chain length or lipid amounts) within an organ or tissue. Highly flexible organs or tissues display stronger changes in their lipidomes in different conditions than less flexible organs or tissues. Here, lipidomic flexibility is calculated by summing the absolute effect sizes ( coefficients of the linear regression) for significantly affected features () of the different organs (Fig. 4A). In this way, the highly specific effect of the genotype on the brain lipidome becomes obvious. Furthermore, genotype and sex exert organ-specific effects. While plasma, full blood and brain are strongly affected by genotype, adipose tissue and lung are only weakly affected. Similarly, sex strongly influences the full blood lipidome (Fig. 4B), while brain, adipose tissue, but also lung and intestine are hardly affected. Also, the kidney and liver lipidomes exhibit sex-specific differences. In liver, diet has a significant impact on TAG and cardiolipin (CL) composition (Fig. 4B). Of note, full blood and liver usually show the highest flexibility, irrespective of the tested condition. Full blood (and in a similar way blood plasma) is strongly affected throughout the entire lipidome. There are significant changes not only in the storage lipids TAG and CE, but also in the major phospholipid classes PC, PI, PS as well as in the sphingolipid SM and cholesterol (see Supplemental Table s3). Hence, full blood can serve as a rich source of information on systemic changes in lipid metabolism.Figure 4Flexibility of the mouse lipidome. (A) Analysis of lipidomic flexibility in mouse organs based on the tested experimental conditions. For details see main text. Mouse plots were created with the R library gganatogram (v.2). (B) Analysis of lipidomic responses by multiple linear regression. Shown are coefficients. Error bars indicate standard error. Examples shown are: Genotype effects for brain; diet effects for liver; sex effects for full blood. Significantly affected features () are displayed non-transparently. Flexibility of the mouse lipidome. (A) Analysis of lipidomic flexibility in mouse organs based on the tested experimental conditions. For details see main text. Mouse plots were created with the R library gganatogram (v.2). (B) Analysis of lipidomic responses by multiple linear regression. Shown are coefficients. Error bars indicate standard error. Examples shown are: Genotype effects for brain; diet effects for liver; sex effects for full blood. Significantly affected features () are displayed non-transparently. In this study, mice of each genotype and sex were fed two different diets: high (18%) and low (14%) protein (for details see Materials & Methods). This factorial design allows for the analysis of differential effects of a given condition depending on the status of the other conditions, in particular diet-induced lipidomic changes in dependence of genotype or sex. For example, we observed that ceramide levels in kidney samples are increased in inbred mice (female and male) fed a high protein diet as compared with low protein diet while they are not affected in kidneys of outbred mice (Fig. 5A). Similarly, the degree of unsaturation of PE is negatively affected by high protein diet in inbred mice but not significantly affected outbred mice. We therefore performed in systematic analysis of sex- or genotype-specific effects of diet based on interaction terms in a multiple linear regression model. In kidney, differential effects can be observed throughout many lipid classes, in particular phospholipids (Fig. 5B). Especially the degree of unsaturation and hydrocarbon chain length of ether lipids (PC O- and PE O-) change in opposite directions depending on the genotype. Sex-specific diet effects could be mostly detected in intestine and muscle. Of note, in muscle, differential effects could mostly be observed for the sphingolipids SM, hexosyl ceramide, and ceramide (Fig. 5C). Genotype-specific diet effects could be detected mostly in brain, intestine and kidney. Similar effects could be observed in all organs to different degrees (Supplemental Table s3). Therefore, dietary effects in an organ depend on sex and the genetic context.Figure 5(A) Genotype-specific effects of diet on ceramide levels and degree of PE unsaturation in kidney. (B) Genotype- and sex-specific (C) and effects of diet based on a multiple linear regression including interaction between diet and genotype and diet and sex, respectively. Shown are coefficients. Error bars indicate standard error. Significantly affected features ( for the interaction term) are displayed non-transparently. Lipidomic features are depicted on the y axis (with “_db” denoting the weighted mean double bond number and “_c” denoting the weighted mean of the carbon chain length of a lipid class; when neither is specified, the feature name refers to lipid class abundance). (A) Genotype-specific effects of diet on ceramide levels and degree of PE unsaturation in kidney. (B) Genotype- and sex-specific (C) and effects of diet based on a multiple linear regression including interaction between diet and genotype and diet and sex, respectively. Shown are coefficients. Error bars indicate standard error. Significantly affected features ( for the interaction term) are displayed non-transparently. Lipidomic features are depicted on the y axis (with “_db” denoting the weighted mean double bond number and “_c” denoting the weighted mean of the carbon chain length of a lipid class; when neither is specified, the feature name refers to lipid class abundance). We wondered whether the full blood (or blood plasma) itself, as an easily accessible sample, is a good proxy for lipidomic changes occurring in other tissues and organs. To investigate whether lipidomic changes in mouse organs are reflected in blood samples, we performed a correlation analysis of amounts of individual lipids in the different organs with the respective lipids in bodily fluids. We consider only significant () positive correlations as indicative for corresponding changes in amounts. To identify which organs are best reflected by blood samples, these positive correlations were used as input for hierarchical clustering (Fig. 6A). Correlations in blood plasma are used as a reference point. The closer the organs cluster with blood plasma, the better their lipid composition is mirrored in the circulation. It appears that positive correlations of amounts of organ lipids with amounts of plasma lipids can be found across the entire lipidome and for all organs. Expectedly, full blood clusters closest with blood plasma since they are sharing the storage lipid complement (TAG and CE) organized in lipoprotein particles. Interestingly, it is liver that clusters closest with blood samples (blood plasma or full blood), followed by adipose tissue and muscle, while the brain and spleen lipidomes are most distant. Hence, changes in the liver lipidome are best reflected in blood samples, while lipidomic changes in brain are only weakly reflected by the blood lipidome. In liver, mostly changes in PC species correlate well with plasma samples. Additionally, changes in liver PI and TAG species are reflected in plasma samples. In muscle and adipose tissue, mostly TAG species are reflected in the plasma lipidome. It should be noted that organs were not subjected to perfusion after dissection. Therefore, any residual blood in these organs could affect correlations of their lipidomes with blood lipidomes. However, this does not seem to be the case, as lipidomes of organs well supplied with blood (such as spleen, intestine or lung) do not correlate well with blood lipidomes. Considering correlations based on lipidomic features, TAG appears to be the class that is best reflected in the blood lipidome for various organs. For adipose tissue, muscle, liver, and kidney there are significant positive correlations of TAG unsaturation and chain length indices in both blood plasma and full blood (Fig. 6B). Additionally, there are robust correlations for PI features in liver, intestine, and kidney (not shown). Even though there are numerous organ lipids whose abundance is reflected in blood samples, the blood lipidome is not a comprehensive mirror of lipidomic changes in organs. In most cases, not more than 5% of lipids in an organ are positively correlated with their amounts in the circulation. The exception is liver, with ca. 15% of lipids correlating with blood. Therefore, organ lipidomes respond in a specific manner to different stimuli and the responses are not necessarily reflected in blood.Figure 6Correlations of organ lipidomes with blood lipidomes. (A) Hierarchical clustering based on correlation coefficients for diet effects on individual lipids in different organs with lipids in blood plasma as reference point. Colour scale indicates correlation coefficient . Non-significant positive correlations and negative correlations are shown in blue. (B) Scatter plots showing correlations for TAG lipidomic features in organs with counterparts in blood plasma and full blood. Dashed lines show a linear regression of the data (with grey areas indicating the 95% confidence interval). In both panels only significant () positive correlations are shown and correlations were adjusted for sex and genotype. Correlations of organ lipidomes with blood lipidomes. (A) Hierarchical clustering based on correlation coefficients for diet effects on individual lipids in different organs with lipids in blood plasma as reference point. Colour scale indicates correlation coefficient . Non-significant positive correlations and negative correlations are shown in blue. (B) Scatter plots showing correlations for TAG lipidomic features in organs with counterparts in blood plasma and full blood. Dashed lines show a linear regression of the data (with grey areas indicating the 95% confidence interval). In both panels only significant () positive correlations are shown and correlations were adjusted for sex and genotype. Lipidomics has become an indispensable technology in basic, clinical, and pharmaceutical research to investigate lipid metabolism quantitatively at the molecular level. Therefore, a thorough characterization of experimental model systems is required to provide a reference for future studies, facilitating experimental design and data interpretation. Here, we present a comprehensive and quantitative lipidomic atlas of mouse organs, full blood, and blood plasma. This study confirms a remarkable specificity of the organ’s lipid composition both at the level of lipid classes as well as for the individual lipid molecules. In fact, each organ shows a unique and unambiguous lipid fingerprint. Including mice of different genotypic background, sex and feeding of two different diets allowed for an analysis of lipidomic flexibility for each organ and lipid class. This analysis reveals that the lipidome of certain organs is more susceptible to variation of genotype, sex, and diet while other organs exhibit a remarkably robust lipid composition unimpeded by these perturbations. The multi-factorial study design further enabled the analysis of genotype- and sex-specific diet effects for each organ. Again, when specificity could be demonstrated, it was organ- and lipid class-dependent, highlighting the complexity of an individual’s lipid metabolism. Moreover, correlations of organ lipidomes with full blood and plasma lipidomes revealed that lipidomic changes in the brain, spleen, and intestine are only poorly mirrored in the circulation, while the liver lipidome is well reflected in the blood. Most importantly, the factorial design of this study enabled the analysis of lipidomic plasticity of the different lipid classes in different organs (Fig. 3) and the lipidomic flexibility of different organs in response to diet, sex, and genotype (Fig. 4). We define lipidomic plasticity as the degree of structural heterogeneity/variation within a given lipid class, i.e., the possible range of the degree of unsaturation (double bond numbers) and hydrocarbon chain length. Lipidomic flexibility of an organ, on the other hand, is the magnitude of lipidomic changes induced by experimental conditions. Both parameters are characteristic features of lipidomes and specific for organs, tissues, and even cell types. The specific lipid composition of organs has been documented in detail previously. The lipid composition obviously reflects histological and cellular structures, which are the foundation of the organ’s functions. For example, adipose tissue with its triglyceride-laden lipid droplets in their adipocytes is composed almost exclusively of triglycerides (ca. 99 mol%). Lung samples are characterized by high amounts of surfactant components, the most prominent being dipalmitoyl PC (PC 16:0/16:0), PG, and SM. Brain samples are rich in PE O-, hexosyl ceramide, ceramide, and cholesterol. PE O-, hexosyl ceramide and cholesterol are major components of myelin sheaths, structures insulating the axons in neurons. High concentrations of ceramides likely serve as precursor for the synthesis of complex glycolipids such as sulfatide and gangliosides, which are major lipid components of the mammalian brain. These lipids were not included in our analysis. Furthermore, high amounts of CL in both liver and kidney reflect the large numbers of mitochondria in these organs. Compositional specificity not only at the lipid class level, but also at the level of individual lipid molecules, is achieved by organ-specific expression of lipid biosynthetic genes. This is best exemplified for ceramide synthases, which show highly tissue-specific expression profiles. In intestine, ceramide synthase 6 is expressed highest with a preference for myristic (C14:0) and palmitic (C16:0) acid, resulting in ceramides with a total carbon number of 32-34. In brain, ceramide synthase 1 is expressed at the highest levels with a preference for stearic acid (18:0), resulting in ceramides with a total carbon number of 36. At the other end of the scale is liver, in which ceramide synthase 2 is expressed abundantly. Ceramide synthase 2 prefers fatty acids with 20 to 26 carbons, giving rise to ceramide species with total carbon numbers ranging from 38 to 44. These expression patterns are reflected in the ceramide profiles for the various tissues (Fig. 2). Similarly, organ-specific species profiles for other lipid classes are achieved by distinct expression and activity profiles of fatty acid desaturases, elongases and acyl transferases. However, there is much to be learned about the molecular mechanisms that give rise to the perplexing complexity of lipid metabolism and composition for different organs, tissues, and cell types. We propose lipidomic plasticity and flexibility as intrinsic properties of an organ or bodily fluid. While lipidomic plasticity is the ability of a lipid class to assume varying degrees of fatty acid unsaturation or chain length, lipidomic flexibility provides a quantitative measure for the magnitude of changes in a lipidome induced by certain experimental conditions. Lipidomic flexibility of an organ is directly derived from the lipidomic plasticity of its lipid constituents. Here, we investigate these parameters at the level of organs and bodily fluids. Future studies will likely necessitate the extension of this concept to tissues and even cell types. The largest variation in the data set is caused by organ-specific differences of the lipidomes. Diet, genotype, and sex each contribute to about 5% of overall variance. Nevertheless, for each organ a significant fraction of the lipidome is affected by the tested conditions. However, some organs are more responsive than others. For example, the brain lipidome is only affected by genotype while diet and sex hardly have any significant effect. Therefore, the brain appears to have a very robust lipid composition. This is confirmed by the very low internal plasticity of the major brain lipid classes (like for example PC, PE O-, hexosyl ceramide, Fig. 3). Other sample types such as liver and full blood are affected by all the tested conditions. One may interpret a low lipidomic flexibility as a consequence of a tight coupling of composition, structure, and function. A very specific lipid composition (i.e., a low lipidomic plasticity) with particular structural features such as degree of unsaturation and fatty acid chain length are required to fulfil a specific function. In brain that would be the formation of the myelin sheath and in the lung the coating of the alveolar interior with surfactant lipids like dipalmitoyl PC. Liver and full blood on the other hand fulfil rather metabolic than structural functions, which requires/results in a greater flexibility to serve as buffer against metabolic perturbations such as the compositional differences related to lipids in the diets tested here. Lipidomic plasticity and flexibility of an organ have important implications for experimental design: If an organ with high flexibility (resulting in large effect sizes) is under investigation, lower numbers of biological replicates might be sufficient to observe robust phenotypes. Furthermore, changes observed in organs that are characterized by low plasticity and/or flexibility might be more meaningful and hint at a severe phenotype. Another level of complexity is added by the fact that intervention effects (here simulated by diet) might depend on genotype or sex of the animal. This might be generalized to other backgrounds, like for example diet- or genotype-specific drug effects in pharmaceutical time-dose studies. Blood samples (in most cases blood plasma) are usually used to study intervention effects in model systems. However, and perhaps not surprisingly, our results suggest that not every organ’s lipid metabolism is reflected in the circulation. As the liver plays a central role in lipoprotein metabolism, the liver lipidome correlates strongest with blood plasma, confirming previous studies. Lipid metabolism of other organs is not well reflected in blood, especially brain, spleen, and intestine. However, more research is obviously needed to understand how the blood and plasma lipidome reflects systemic effects on lipid metabolism. Nevertheless, if organ-specific effects of interventions are eventually desired or expected, it is recommended to not only rely on an initial screening of blood samples as important phenotypes might be missed. This study presents a complex analysis of mouse organ lipidomes and their dependence on different experimental conditions/factors. However, some important organs or bodily fluids were not included, such as reproductive organs, pancreas, heart muscle, brown adipose tissue, cerebrospinal fluid, or urine. Analysis of additional tissue and cell types from various organs will certainly increase the degree of complexity. A higher number of biological replicates () would increase statistical power and likely allow for the identification of more significant lipidomic changes. Moreover, additional analytical approaches (for example based on liquid chromatography coupled to mass spectrometry) would allow for the quantification of low abundant signalling lipids such as phosphoinositides, long chain bases or oxylipins. Furthermore, the analysis of complex glycolipids would be insightful, as would the inclusion of additional mouse genotypes. Finally, the dietary intervention used here is rather subtle. Treatment with a high fat diet or different drugs might lead to more pronounced effects and hence different conclusions regarding organ flexibility and dependence on genotype and sex. Here we present a systematic analysis of the lipid composition of mouse organs and blood samples and an assessment of overall plasticity and flexibility of the mouse lipidome induced by diet, genotype, and sex. This dataset provides a valuable resource for basic and pharmaceutical researchers using mouse as a model system and complements existing proteomic and transcriptomic datasets. More basic and quantitative characterizations of model systems are required for other omics technologies as they will aid standardization of research areas and inform experimental design and facilitate interpretation of lipidomic datasets. Two standard laboratory mouse wild-type strains were selected: Hsd:ICR (CD-1) and C57BL/6JOlaHsd, genetically representing an outbred and inbred population, respectively. The mice were allowed to breed within the strains, and once pregnancy was confirmed, the pregnant females were put on two standard, healthy diets differing by protein content: Teklad global 14% protein (Envigo #2014S, low protein) and Teklad global 18% protein (Envigo #2018S, high protein). The high protein diet delivered 24% of total calories from proteins and 18% from fats, whereas the low protein diet delivered 20% of total calories from proteins and 13% from fats. Qualitatively, the lipid compositions of high and low protein diets are comparable (Supplemental Table s2). After birth, mothers were kept at the same diet and the litter mice were introduced to the same diet as their mothers after weaning. At the age of ca. 8.5 weeks, mice were weighted and then sacrificed, to yield three females and three males from each strain and each diet ( representing the minimal sample number for meaningful statistical analysis). Hsd:ICR (outbred) animals came from different litters. On the contrary, C57BL/6JOlaHsd (inbred) animals came from the same litter, that is all six mice (three males and three females) for each diet were siblings. All protocols were approved by the institutional Animal Welfare Officer of the Max Planck Institute of Molecular Cell Biology and Genetics (Dresden, Germany), and the necessary licenses were obtained by the governmental veterinary authority, as dictated by the German Welfare Legislation regulating the use of animals for scientific purposes. All animal housing, handling and experimental techniques were in accordance with the principles set out in the Declaration of Helsinki, as well as in accordance with the ethical standards of the European and German Animal Welfare legislation. The study was carried out in compliance with the arrive guidelines. Mice were sacrificed and immediately decapitated, and the full blood (F) samples were collected directly from necks into EDTA-tubes (Sarstedt, #41.1395.105). For plasma (P) preparation, tubes were centrifuged for 10 minutes at 2000 g and blood plasma was collected within 1 hour after blood collection. After diluting full blood and plasma with water, samples were frozen at C until lipid extraction. Organs and tissues were dissected from fresh cadavers after all blood was drained (no perfusion was performed), resulting in the following samples: liver-H (a section of the middle part of the left lateral lobe); skeletal muscle-M (left soleus); brain-B (section of left hemisphere); kidney-K (whole left kidney, without fat tissue); adipose tissue-A (from the abdominal region); small intestine-I (whole duodenum); lung-L (section of middle region of left lung) and spleen-S (whole spleen). These organs and tissues were put into 2 mL microcentrifuge tubes, immediately frozen on dry ice, weighed using a Radwag Microbalance Type MYA 5.3Y and stored at C until homogenization. Homogenization was performed after samples were thawed on ice and resuspended in 1.5 mL of C cold 150 mM ammonium bicarbonate buffer (except for adipose tissue samples for which ammonium bicarbonate buffer was mixed 1:1 with ethanol to facilitate homogenization as described previously), by shaking in 2 ml tubes for 15 min ( minutes with 2 minutes cool down periods on ice) at C with several 3.1 mm stainless steel beads using a Qiagen TissueLyser II. Homogenized samples were diluted, and volumes corresponding to the following amounts were used for lipid extraction: 80 g brain, 100 g liver, 270 g intestine, 200 g kidney, 250 g lung, 650 g muscle, 400 g spleen, 100 g adipose tissue. For blood plasma and full blood, volumes corresponding to 1 l of the undiluted sample were pipetted from the dilutions. Samples were divided into four separate analytical batches: plasma, adipose tissue and two batches for the remaining samples. Each batch was accompanied by a set of blank samples (pure buffer) and identical reference samples: human plasma for the plasma batch, full blood for the remaining sample types, that were aliquoted to provide identical technical replicates. These control samples were distributed evenly across each batch, extracted, and processed together with study samples to control for background signals, technical variation, and overall performance of the method. All batches were measured on 4 consecutive days. Lipid extraction for lipidomic analysis was performed as described for plasma and adipose tissue. For all other types of samples, the procedure was as follows. Lipids were extracted using a two-step chloroform/methanol procedure with chloroform:methanol 10:1 (V:V) and 2:1 (V:V) in the first and second step, respectively. Prior to extraction, samples were spiked with a sample type-specific internal lipid standard mixture (for plasma, full blood, adipose tissue, and the remaining tissue samples; for compositional details see Supplemental Table s4). After extraction, part of the organic phase was transferred to an infusion plate and dried in a speed vacuum concentrator. For mass spectrometry acquisition, 1st step dry extract was re-suspended in 7.5 mM ammonium acetate in chloroform/methanol/propanol (1:2:4, V:V:V) and 2nd step dry extract in 33% ethanol solution of methylamine in chloroform/methanol (0.003:5:1; V:V:V) by vigorous shaking for 1 minute to ensure complete dissolution of extracts. All liquid handling steps were performed using Hamilton Robotics STARlet robotic platform with the Anti Droplet Control feature for pipetting organic solvents. All solvents and chemicals used were of analytical grade. Mass spectrometry analysis was performed as described previously for plasma and adipose tissue. Samples were analyzed by direct infusion on a QExactive mass spectrometer (Thermo Scientific) equipped with a TriVersa NanoMate ion source (Advion Biosciences). Samples were analyzed in both positive and negative ion modes with a resolution of for MS and for MSMS experiments, in a single acquisition. MSMS was triggered by an inclusion list encompassing corresponding MS mass ranges scanned in 1 Da increments. Both MS and MSMS data were combined to monitor CE, DAG and TAG ions as ammonium adducts; PC, PC O-, as acetate adducts; and CL, phosphatidate (PA), PE, PE O-, PG, PI and PS as deprotonated anions. MS only was used to monitor lyso-phosphatidate (LPA), LPE, lyso-phosphatidylethanolamine ether (LPE O-), lyso-phosphatidylglycerol (LPG), lyso-phosphatidylinositol (LPI), and lyso-phosphatidylserine (LPS) as deprotonated anions; Cer, HexCer, SM, LPC and lyso-phosphatidylcholine ether (LPC O-) as acetate adducts and cholesterol as ammonium adduct of an acetylated derivative. Data were analyzed with in-house developed lipid identification software based on LipidXplorer. Data post-processing and normalization were performed using an in-house developed data management system. Only lipid identifications with a signal-to-noise ratio , and a signal intensity 5-fold higher than in corresponding blank samples were considered for further data analysis. Lipid molecules are identified as species or subspecies. Fragmentation of the lipid molecules in MSMS mode delivers subspecies information, i.e., the exact acyl chain (e.g., fatty acid) composition of the lipid molecule. MS only mode, acquiring data without fragmentation, cannot deliver this information and provides species information only. In that case, the sum of the carbon atoms and double bonds in the hydrocarbon moieties is provided. Lipid species are annotated according to their molecular composition as lipid class <sum of carbon atoms>:< sum of double bonds>;< sum of hydroxyl groups>. For example, PI 34:1;0 denotes phosphatidylinositol with a total length of its fatty acids equal to 34 carbon atoms, total number of double bonds in its fatty acids equal to 1 and 0 hydroxylations. In case of sphingolipids, SM 34:1;2 denotes a sphingomyelin species with a total of 34 carbon atoms, 1 double bond, and 2 hydroxyl groups in the ceramide backbone. Lipid subspecies annotation contains additional information on the exact identity of their acyl moieties and their sn-position (if available). For example, PI 18:1;0_16:0;0 denotes phosphatidylinositol with octadecenoic (18:1;0) and hexadecanoic (16:0;0) fatty acids, for which the exact position (sn-1 or sn-2) in relation to the glycerol backbone cannot be discriminated (underline “_” separating the acyl chains). On contrary, PC O- 18:1;0/16:0;0 denotes an ether-phosphatidylcholine, in which an alkyl chain with 18 carbon atoms and 1 double bond (O-18:1;0) is ether-bound to sn-1 position of the glycerol and a hexadecanoic acid (16:0;0) is connect via an ester bond to the sn-2 position of the glycerol (slash “/” separating the chains signifies that the sn-position on the glycerol can be resolved). Lipid identifiers of the SwissLipids database (http://www.swisslipids.org) are provided in Supplemental Table s1. Data were analyzed with R version 4.0.3 using tidyverse (v.1.2.1) packages. Data were corrected for batch effects and analytical drift based on reference samples. Only lipids with amounts pmol and identified in at least two out of three biological replicates in at least one experimental condition are reported, resulting in 796 lipid (sub-)species included in the final dataset (of a total of 1377 identified lipid (sub-)species). Molar amount values (in pmol) of individual lipid (sub-)species were normalized to total lipid content per sample, yielding molar fraction values, expressed in mol%. A weighted mean length per total fatty acids or weighted mean saturation per total fatty acids per lipid classes is calculated. This returns an estimate of the mean total fatty acid length or mean total number of fatty acid double bonds in this lipid class:[12pt] $$ }}_}= }_}} _} }} ( } _}=1^}_}}\; }_}} ) $$¯f=1nc∑j∈q(j·∑i=1kjnji)where is the weighted mean length/saturation per hydrocarbon moiety, is the feature of a lipid with a list of instances: e.g. total double bonds or total length , is the total molar amount of the respective lipid class, is the molar amount of lipid species i with the respective matching the feature instance j and is the number of species within feature instance j.Principal component analysis (PCA) was performed using the stats::prcomp() function on centered and scaled data. is the weighted mean length/saturation per hydrocarbon moiety, is the feature of a lipid with a list of instances: e.g. total double bonds or total length , is the total molar amount of the respective lipid class, is the molar amount of lipid species i with the respective matching the feature instance j and is the number of species within feature instance j. The internal lipid class plasticity was calculated based on the minimum-maximum range of values for weighted mean of lipid class unsaturation and the weighted mean of lipid class carbon chain length. The ranges for each feature across different sample types was scaled to assume values between 1 and 10, meaning that the class with the widest range across organs receives a value of 10, the class with narrowest range receives a value of 1. Features with intermediate ranges receive values between 1 and 10, according to scale. The scaled ranges were then used to calculate the internal lipid class plasticity by multiplying the scaled carbon chain length range with the scaled double bond number range (Fig. 3). Because of the limited lipid composition, adipose tissue was omitted from the plasticity calculations. Sex, diet and genotype effect sizes ( coefficients) were determined for centered and scaled lipidomic features (z scores for lipid class amount in mol%, weighted mean of lipid class unsaturation and weighted mean of lipid class carbon chain length) by linear regression using the stats::lm() function adjusting for the other covariates. For the calculation of sex- and genotype-specific diet effects, the respective interaction terms were included in the linear regression. Mouse plots were created with the R library gganatogram (v.2). Diet-dependent Spearman correlation coefficients () between blood and tissue samples were calculated using the function RVAideMemoire::pcor.test (v.0.9-78) adjusting for sex and genotype. Hierarchical clustering (Euclidian distance, complete linkage) heatmap generation was performed using the pheatmap::pheatmap() (v.1.0.12) function, which was also used for heatmap generation.
|
PMC11564093
|
Nonlinear dynamics of multi-omics profiles during human aging
|
Aging is a complex process associated with nearly all diseases. Understanding the molecular changes underlying aging and identifying therapeutic targets for aging-related diseases are crucial for increasing healthspan. Although many studies have explored linear changes during aging, the prevalence of aging-related diseases and mortality risk accelerates after specific time points, indicating the importance of studying nonlinear molecular changes. In this study, we performed comprehensive multi-omics profiling on a longitudinal human cohort of 108 participants, aged between 25 years and 75 years. The participants resided in California, United States, and were tracked for a median period of 1.7 years, with a maximum follow-up duration of 6.8 years. The analysis revealed consistent nonlinear patterns in molecular markers of aging, with substantial dysregulation occurring at two major periods occurring at approximately 44 years and 60 years of chronological age. Distinct molecules and functional pathways associated with these periods were also identified, such as immune regulation and carbohydrate metabolism that shifted during the 60-year transition and cardiovascular disease, lipid and alcohol metabolism changes at the 40-year transition. Overall, this research demonstrates that functions and risks of aging-related diseases change nonlinearly across the human lifespan and provides insights into the molecular and biological pathways involved in these changes.Aging is a complex and multifactorial process of physiological changes strongly associated with various human diseases, including cardiovascular diseases (CVDs), diabetes, neurodegeneration and cancer. The alterations of molecules (including transcripts, proteins, metabolites and cytokines) are critically important to understand the underlying mechanism of aging and discover potential therapeutic targets for aging-related diseases. Recently, the development of high-throughput omics technologies has enabled researchers to study molecular changes at the system level. A growing number of studies have comprehensively explored the molecular changes that occur during aging using omics profiling, and most focus on linear changes. It is widely recognized that the occurrence of aging-related diseases does not follow a proportional increase with age. Instead, the risk of these diseases accelerates at specific points throughout the human lifespan. For example, in the United States, the prevalence of CVDs (encompassing atherosclerosis, stroke and myocardial infarction) is approximately 40% between the ages of 40 and 59, increases to about 75% between 60 and 79 and reaches approximately 86% in individuals older than 80 years. Similarly, also in the United States, the prevalence of neurodegenerative diseases, such as Parkinson’s disease and Alzheimer’s disease, exhibits an upward trend as well as human aging progresses, with distinct turning points occurring around the ages of 40 and 65, respectively. Some studies also found that brain aging followed an accelerated decline in flies and chimpanzees that lived past middle age and advanced age. The observation of a nonlinear increase in the prevalence of aging-related diseases implies that the process of human aging is not a simple linear trend. Consequently, investigating the nonlinear changes in molecules will likely reveal previously unreported molecular signatures and mechanistic insights. Some studies examined the nonlinear alterations of molecules during human aging. For instance, nonlinear changes in RNA and protein expression related to aging have been documented. Moreover, certain DNA methylation sites have exhibited nonlinear changes in methylation intensity during aging, following a power law pattern. Li et al. identified the 30s and 50s as transitional periods during women’s aging. Although aging patterns are thought to reflect the underlying biological mechanisms, the comprehensive landscape of nonlinear changes of different types of molecules during aging remains largely unexplored. Remarkably, the global monitoring of nonlinear changing molecular profiles throughout human aging has yet to be fully used to extract basic insights into the biology of aging. In the present study, we conducted a comprehensive deep multi-omics profiling on a longitudinal human cohort comprising 108 individuals aged from 25 years to 75 years. The cohort was followed over a span of several years (median, 1.7 years), with the longest monitoring period for a single participant reaching 6.8 years (2,471 days). Various types of omics data were collected from the participants’ biological samples, including transcriptomics, proteomics, metabolomics, cytokines, clinical laboratory tests, lipidomics, stool microbiome, skin microbiome, oral microbiome and nasal microbiome. The investigation explored the changes occurring across different omics profiles during human aging. Remarkably, many molecular markers and biological pathways exhibited a nonlinear pattern throughout the aging process, thereby providing valuable insight into periods of dramatic alterations during human aging. We collected longitudinal biological samples from 108 participants over several years, with a median tracking period of 1.7 years and a maximum period of 6.8 years, and conducted multi-omics profiling on the samples. The participants were sampled every 3–6 months while healthy and had diverse ethnic backgrounds and ages ranging from 25 years to 75 years (median, 55.7 years). The participants’ body mass index (BMI) ranged from 19.1 kg m to 40.8 kg m (median, 28.2 kg m). Among the participants, 51.9% were female (Fig. 1a and Extended Data Fig. 1a–d). For each visit, we collected blood, stool, skin swab, oral swab and nasal swab samples. In total, 5,405 biological samples (including 1,440 blood samples, 926 stool samples, 1,116 skin swab samples, 1,001 oral swab samples and 922 nasal swab samples) were collected. The biological samples were used for multi-omics data acquisition (including transcriptomics from peripheral blood mononuclear cells (PBMCs), proteomics from plasma, metabolomics from plasma, cytokines from plasma, clinical laboratory tests from plasma, lipidomics from plasma, stool microbiome, skin microbiome, oral microbiome and nasal microbiome; Methods). In total, 135,239 biological features (including 10,346 transcripts, 302 proteins, 814 metabolites, 66 cytokines, 51 clinical laboratory tests, 846 lipids, 52,460 gut microbiome taxons, 8,947 skin microbiome taxons, 8,947 oral microbiome taxons and 52,460 nasal microbiome taxons) were acquired, resulting in 246,507,456,400 data points (Fig. 1b and Extended Data Fig. 1e,f). The average sampling period and number of samples for each participant were 626 days and 47 samples, respectively. Notably, one participant was deeply monitored for 6.8 years (2,471 days), during which 367 samples were collected (Fig. 1c). Overall, this extensive and longitudinal multi-omics dataset enables us to examine the molecular changes that occur during the human aging process. The detailed characteristics of all participants are provided in the Supplementary Data. For each participant, the omics data were aggregated and averaged across all healthy samples to represent the individual’s mean value, as detailed in the Methods section. Compared to cross-sectional cohorts, which have only a one-time point sample from each participant, our longitudinal dataset, which includes multiple time point samples from each participant, is more robust for detecting complex aging-related changes in molecules and functions. This is because analysis of multi-time point samples can detect participants’ baseline and robustly evaluate individuals’ longitudinal molecular changes.Fig. 1Most molecules and microbes undergo nonlinear changes during human aging.a, The demographics of the 108 participants in the study are presented. b, Sample collection and multi-omics data acquisition of the cohort. Four types of biological samples were collected, and 10 types of omics data were acquired. c, Collection time range and sample numbers for each participant. The top x axis represents the collection range for each participant (read line), and the bottom x axis represents the sample number for each participant (bar plot). Bars are color-coded by omics type. d, Significantly changed molecules and microbes during aging were detected using the Spearman correlation approach (P < 0.05). The P values were not adjusted (Methods). Dots are color-coded by omics type. e, Differential expressional molecules/microbes in different age ranges compared to baseline (25–40 years old, two-sided Wilcoxon test, P < 0.05). The P values were not adjusted (Methods). f, The linear changing molecules comprised only a small part of dysregulated molecules in at least one age range. g, Heatmap depicting the nonlinear changing molecules and microbes during human aging. a, The demographics of the 108 participants in the study are presented. b, Sample collection and multi-omics data acquisition of the cohort. Four types of biological samples were collected, and 10 types of omics data were acquired. c, Collection time range and sample numbers for each participant. The top x axis represents the collection range for each participant (read line), and the bottom x axis represents the sample number for each participant (bar plot). Bars are color-coded by omics type. d, Significantly changed molecules and microbes during aging were detected using the Spearman correlation approach (P < 0.05). The P values were not adjusted (Methods). Dots are color-coded by omics type. e, Differential expressional molecules/microbes in different age ranges compared to baseline (25–40 years old, two-sided Wilcoxon test, P < 0.05). The P values were not adjusted (Methods). f, The linear changing molecules comprised only a small part of dysregulated molecules in at least one age range. g, Heatmap depicting the nonlinear changing molecules and microbes during human aging. We included samples only from healthy visits and adjusted for confounding factors (for example, BMI, sex, insulin resistance/insulin sensitivity (IRIS) and ethnicity; Extended Data Fig. 1a–d), allowing us to discern the molecules and microbes genuinely associated with aging (Methods). Two common and traditional approaches, linear regression and Spearman correlation, were first used to identify the linear changing molecules during human aging. The linear regression method is commonly used for linear changing molecules. As expected, both approaches have very high consistent results for each type of omics data (Supplementary Fig. 1a). For convenience, the Spearman correlation approach was used in the analysis. Interestingly, only a small portion of all the molecules and microbes (749 out of 11,305, 6.6%; only genus level was used for microbiome data; Methods) linearly changed during human aging (Fig. 1d and Supplementary Fig. 1b), consistent with our previous studies (Methods). Next, we examined nonlinear effects by categorizing all participants into distinct age stages according to their ages and investigated the dysregulated molecules within each age stage compared to the baseline (25–40 years old; Methods). Interestingly, using this approach, 81.03% of molecules (9,106 out of 11,305) exhibited changes in at least one age stage compared to the baseline (Fig. 1e and Extended Data Fig. 2a). Remarkably, the percentage of linear changing molecules was relatively small compared to the overall dysregulated molecules during aging (mean, 16.2%) (Fig. 1f and Extended Data Fig. 2b). To corroborate our findings, we employed a permutation approach to calculate permutated P values, which yielded consistent results (Methods). The heatmap depicting all dysregulated molecules also clearly illustrates pronounced nonlinear changes (Fig. 1g). Taken together, these findings strongly suggest that a substantial number of molecules and microbes undergo nonlinear changes throughout human aging. Next, we assessed whether the multi-omics data collected from the longitudinal cohort could serve as reliable indicators of the aging process. Our analysis revealed a substantial correlation between a significant proportion of the omics data and the ages of the participants (Fig. 2a). Particularly noteworthy was the observation that, among all the omics data examined, metabolomics, cytokine and oral microbiome data displayed the strongest association with age (Fig. 2a and Extended Data Fig. 3a–c). Partial least squares (PLS) regression was further used to compare the strength of the age effect across different omics data types. The results are consistent with the results presented above in Fig. 2a (Methods). These findings suggest the potential utility of these datasets as indicators of the aging process while acknowledging that further research is needed for validation. As the omics data are not accurately matched across all the samples, we then smoothed the omics data using our previously published approach (Methods and Supplementary Fig. 2a–c). Next, to reveal the specific patterns of molecules that change during human aging, we then grouped all the molecules with similar trajectories using an unsupervised fuzzy c-means clustering approach (Methods, Fig. 3b and Supplementary Fig. 2d,e). We identified 11 clusters of molecular trajectories that changed during aging, which ranged in size from 638 to 1,580 molecules/microbes (Supplementary Fig. 2f and Supplementary Data). We found that most molecular patterns exhibit nonlinear changes, indicating that aging is not a linear process (Fig. 2b). Among the 11 identified clusters, three distinct clusters (2, 4 and 5) displayed compelling, straightforward and easily understandable patterns that spanned the entire lifespan (Fig. 2c). Most molecules within these three clusters primarily consist of transcripts (Supplementary Fig. 2f), which is expected because transcripts dominate the multi-omics data (8,556 out of 11,305, 75.7%). Cluster 4 exhibits a relatively stable pattern until approximately 60 years of age, after which it shows a rapid decrease (Fig. 2c). Conversely, clusters 2 and 5 display fluctuations before 60 years of age, followed by a sharp increase and an upper inflection point at approximately 55–60 years of age (Fig. 2c). We also attempted to observe this pattern of molecular change during aging individually. The participant with the longest follow-up period of 6.8 years (Fig. 1c) approached the age of 60 years (range, 59.5–66.3 years; Extended Data Fig. 1g), and it was not possible to identify obvious patterns in this short time window (Supplementary Fig. 2g). Tracking individuals longitudinally over longer periods (decades) will be required to observe these trajectories at an individual level.Fig. 2Clustering reveals nonlinear changes in multi-omics profiling during human aging.a, Spearman correlation (cor) between the first principal component and ages for each type of omics data. The shaded area around the regression line represents the 95% confidence interval. b, The heatmap shows the molecular trajectories in 11 clusters during human aging. The right stacked bar plots show the percentages of different kinds of omics data, and the right box plots show the correlation distribution between features and ages (n = 108 participants). c, Three notable clusters of molecules that exhibit clear and straightforward nonlinear changes during human aging. The top stacked bar plots show the percentages of different kinds of omics data, and the top box plots show the correlation distribution between features and ages (n = 108 participants). The box plot shows the median (line), interquartile range (IQR) (box) and whiskers extending to 1.5 × IQR. Bars and lines are color-coded by omics type. Abs, absolute.Fig. 3Functional analysis of nonlinear changing molecules in each cluster.a, Pathway enrichment and module analysis for each transcriptome cluster. The left panel is the heatmap for the pathways that undergo nonlinear changes across aging. The right panel is the pathway similarity network (Methods) (n = 108 participants). b, Pathway enrichment for metabolomics in each cluster. Enriched pathways and related metabolites are illustrated (Benjamini–Hochberg-adjusted P < 0.05). c, Four clinical laboratory tests that change during human aging: blood urea nitrogen, serum/plasma glucose, mean corpuscular hemoglobin and red cell distribution width (n = 108 participants). The box plot shows the median (line), interquartile range (IQR) (box) and whiskers extending to 1.5 × IQR. a, Spearman correlation (cor) between the first principal component and ages for each type of omics data. The shaded area around the regression line represents the 95% confidence interval. b, The heatmap shows the molecular trajectories in 11 clusters during human aging. The right stacked bar plots show the percentages of different kinds of omics data, and the right box plots show the correlation distribution between features and ages (n = 108 participants). c, Three notable clusters of molecules that exhibit clear and straightforward nonlinear changes during human aging. The top stacked bar plots show the percentages of different kinds of omics data, and the top box plots show the correlation distribution between features and ages (n = 108 participants). The box plot shows the median (line), interquartile range (IQR) (box) and whiskers extending to 1.5 × IQR. Bars and lines are color-coded by omics type. Abs, absolute. a, Pathway enrichment and module analysis for each transcriptome cluster. The left panel is the heatmap for the pathways that undergo nonlinear changes across aging. The right panel is the pathway similarity network (Methods) (n = 108 participants). b, Pathway enrichment for metabolomics in each cluster. Enriched pathways and related metabolites are illustrated (Benjamini–Hochberg-adjusted P < 0.05). c, Four clinical laboratory tests that change during human aging: blood urea nitrogen, serum/plasma glucose, mean corpuscular hemoglobin and red cell distribution width (n = 108 participants). The box plot shows the median (line), interquartile range (IQR) (box) and whiskers extending to 1.5 × IQR. Although confounders, including sex, were corrected before analysis (Methods), we acknowledge that the age range for menopause in females is typically between 45 years and 55 years of age, which is very close to the major transition points in all three clusters (Fig. 2c). Therefore, we conducted further investigation into whether the menopausal status of females in the dataset contributed to the observed transition point at approximately 55 years of age (Fig. 2c) by performing separate clustering analyses on the male and female datasets. Surprisingly, both the male and female datasets exhibited similar clusters, as illustrated in Extended Data Fig. 4a. This suggests that the transition point observed at approximately 55 years of age is not solely attributed to female menopause but, rather, represents a common phenomenon in the aging process of both sexes. This result is consistent with previous studies, further supporting the notion that this transition point is a major characteristic feature of human aging. Moreover, to investigate the possibility that the transcriptomics data might skew the results toward transcriptomic changes as age-related factors, we conducted two additional clustering analyses—one focusing solely on transcriptomic data and another excluding it. Interestingly, both analyses yielded nearly identical three-cluster configurations, as observed using the complete omics dataset (Extended Data Fig. 4b). This reinforces the robustness of the identified clusters and confirms that they are consistent across various omics platforms, not just driven by transcriptomic data. To gain further insight into the biological functions associated with the nonlinear changing molecules within the three identified clusters, we conducted separate functional analyses for transcriptomics, proteomics and metabolomics datasets for all three clusters. In brief, we constructed a similarity network using enriched pathways from various databases (Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG) and Reactome) and identified modules to eliminate redundant annotations. We then used all modules from different databases to reduce redundancy further using the same approach and define the final functional modules (Methods, Extended Data Fig. 4c and Supplementary Data). We identified some functional modules that were reported in previous studies, but we defined their more accurate patterns of change during human aging. Additionally, we also found previously unreported potential functional modules during human aging (Supplementary Data). For instance, in cluster 2, we identified a transcriptomic module associated with GTPase activity (adjusted P = 1.64 × 10) and histone modification (adjusted P = 6.36 × 10) (Fig. 3a). Because we lack epigenomic data in this study, our findings should be validated through additional experiments in the future. GTPase activity is closely correlated with programmed cell death (apoptosis), and some previous studies showed that this activity increases during aging. Additionally, histone modifications have been demonstrated to increase during human aging. In cluster 4, we identified one transcriptomics module associated with oxidative stress; this module includes antioxidant activity, oxygen carrier activity, oxygen binding and peroxidase activity (adjusted P = 0.029) (Fig. 3a). Previous studies demonstrated that oxidative stress and many reactive oxygen species (ROS) are positively associated with increased inflammation in relation to aging. In cluster 5, the first transcriptomics module is associated with mRNA stability, which includes mRNA destabilization (adjusted P = 0.0032), mRNA processing (adjusted P = 3.2 × 10), positive regulation of the mRNA catabolic process (adjusted P = 1.46 × 10) and positive regulation of the mRNA metabolic process (adjusted P = 0.00177) (Fig. 3a). Previous studies showed that mRNA turnover is associated with aging. The second module is associated with autophagy (Fig. 3a), which increases during human aging, as demonstrated in previous studies. In addition, we also identified certain modules in the clusters that suggest a nonlinear increase in several disease risks during human aging. For instance, in cluster 2, where components increase gradually and then rapidly after age 60, the phenylalanine metabolism pathway (adjusted P = 4.95 × 10) was identified (Fig. 3b). Previous studies showed that aging is associated with a progressive increase in plasma phenylalanine levels concomitant with cardiac dysfunction, and dysregulated phenylalanine catabolism is a factor that triggers deviations from healthy cardiac aging trajectories. Additionally, C-X-C motif chemokine 5 (CXCL5 or ENA78) from proteomics data, which has higher concentrations in atherosclerosis, is also detected in cluster 2 (Supplementary Data). The clinical laboratory test blood urea nitrogen, which provides important information about kidney function, is also detected in cluster 2 (Fig. 3c). This indicates that kidney function nonlinearly decreases during aging. Furthermore, the clinical laboratory test for serum/plasma glucose, a marker of type 2 diabetes (T2D), falls within cluster 2. This is consistent with and supported by many previous studies demonstrating that aging is a major risk factor for T2D. Collectively, these findings suggest a nonlinear escalation in the risk of cardiovascular and kidney diseases and T2D with advancing age, particularly after the age of 60 years (Fig. 2c). The identified modules in cluster 4 also indicate a nonlinear increase in disease risks. For instance, the unsaturated fatty acids biosynthesis pathway (adjusted P = 4.71 × 10) is decreased in cluster 4. Studies have shown that unsaturated fatty acids are helpful in reducing CVD risk and maintaining brain function. The pathway of alpha-linolenic acid and linolenic acid metabolism (adjusted P = 1.32 × 10) can reduce aging-associated diseases, such as CVD. We also detected the caffeine metabolism pathway (adjusted P = 7.34 × 10) in cluster 4, which suggests that the ability to metabolize caffeine decreases during aging. Additionally, the cytokine MCP1 (chemokine (C-C motif) ligand 2 (CCL2)), a member of the CC chemokine family, plays an important immune regulatory role and is also in cluster 4 (Supplementary Data). These findings further support previous observations and highlight the nonlinear increase in age-related disease risk as individuals age. Cluster 5 comprises the clinical tests of mean corpuscular hemoglobin and red cell distribution width (Fig. 3c). These tests assess the average hemoglobin content per red blood cell and the variability in the size and volume of red blood cells, respectively. These findings align with the aforementioned transcriptomic data, which suggest a nonlinear reduction in the oxygen-carrying capacity associated with the aging process. Aside from these three distinct clusters (Fig. 2c), we also conducted pathway enrichment analysis across all other eight clusters, which displayed highly nonlinear trajectories, employing the same method (Fig. 2b and Supplementary Data). Notably, cluster 11 exhibited a consistent increase up until the age of 50, followed by a decline until the age of 56, after which no substantial changes were observed up to the age of 75. A particular transcriptomics module related to DNA repair was identified, encompassing three pathways: positive regulation of double-strand break repair (adjusted P = 0.042), peptidyl−lysine acetylation (adjusted P = 1.36 × 10) and histone acetylation (adjusted P = 3.45 × 10) (Extended Data Fig. 4d). These three pathways are critical in genomic stability, gene expression and metabolic balances, with their levels diminishing across the human lifespan. Our findings reveal a nonlinear alteration across the human lifespan in these pathways, indicating an enhancement in DNA repair capabilities before the age of 50, a marked reduction between the ages of 50 and 56 and stabilization after that until the age of 75. The pathway enrichment results for all clusters are detailed in the Supplementary Data. Altogether, the comprehensive functional analysis offers valuable insights into the nonlinear changes observed in molecular profiles and their correlations with biological functions and disease risks across the human lifespan. Our findings reveal that individuals aged 60 and older exhibit increased susceptibility to CVD, kidney issues and T2D. These results carry important implications for both the diagnosis and prevention of these diseases. Notably, many clinically actionable markers were identified, which have the potential for improved healthcare management and enhanced overall well-being of the aging population. Although the trajectory clustering approaches described above effectively identify nonlinear changing molecules and microbes that exhibit clear and compelling patterns throughout human aging, it may not be as effective in capturing substantial changes that occur at specific chronological aging periods. In such cases, alternative analytical approaches may be necessary to detect and characterize these dynamics. To gain a comprehensive understanding of changes in multi-omics profiling during human aging, we used a modified version of the DE-SWAN algorithm, as described in the Methods section. This algorithm identifies dysregulated molecules and microbes throughout the human lifespan by analyzing molecule levels within 20-year windows and comparing two groups in 10-year parcels while sliding the window incrementally from young to old ages. Using this approach and multiomics data, we detected changes at specific stages of lifespan and uncovered the sequential effects of aging. Our analysis revealed thousands of molecules exhibiting changing patterns throughout aging, forming distinct waves, as illustrated in Fig. 3a. Notably, we observed two prominent crests occurring around the ages of 45 and 65, respectively (Fig. 4a). Notably, too, these crests were consistent with findings from a previous study that included only proteomics data. Specifically, crest 2 aligns with our previous trajectory clustering result, indicating a turning point at approximately 60 years of age (Fig. 2c).Fig. 4Waves of molecules and microbes during aging.a, Number of molecules and microbes differentially expressed during aging. Two local crests at the ages of 44 years and 60 years were identified. b,c, The same waves were detected using different q value (b) and window (c) cutoffs. d, The number of molecules/microbes differentially expressed for different types of omics data during human aging. a, Number of molecules and microbes differentially expressed during aging. Two local crests at the ages of 44 years and 60 years were identified. b,c, The same waves were detected using different q value (b) and window (c) cutoffs. d, The number of molecules/microbes differentially expressed for different types of omics data during human aging. To demonstrate the significance of the two crests, we employed different q value cutoffs and sliding window parameters, which consistently revealed the same detectable waves (Fig. 4b,c and Supplementary Fig. 4a,b). Furthermore, when we permuted the ages of individuals, the crests disappeared (Supplementary Figs. 3a and 4c) (Methods). These observations highlight the robustness of the two major waves of aging-related molecular changes across the human lifespan. Although we already accounted for confounders before our statistical analysis, we took additional steps to explore their impact. Specifically, we investigated whether confounders, such as insulin sensitivity, sex and ethnicity, differed between the two crests across various age ranges. As anticipated, these confounders did not show significant differences across other age brackets (Supplementary Fig. 4d). This further supports our conclusion that the observed differences in the two crests are attributable to age rather than other confounding variables. The identified crests represent notable milestones in the aging process and suggest specific age ranges where substantial molecular alterations occur. Therefore, we investigated the age-related waves for each type of omics data. Interestingly, most types of omics data exhibited two distinct crests that were highly robust (Fig. 3b and Supplementary Fig. 4). Notably, the proteomics data displayed two age-related crests at ages around 40 years and 60 years. Only a small overlap was observed between our dataset and the results from the previous study (1,305 proteins versus 302 proteins, with only 75 proteins overlapping). The observed pattern in our study was largely consistent with the previous findings. However, our finding that many types of omics data, including transcriptomics, proteomics, metabolomics, cytokine, gut microbiome, skin microbiome and nasal microbiome, exhibit these waves, often with a similar pattern as the proteomics data (Fig. 4d), supports the hypothesis that aging-related changes are not limited to a specific omics layer but, rather, involve a coordinated and systemic alteration across multiple molecular components. Identifying consistent crests across different omics data underscores the robustness and reliability of these molecular milestones in the aging process. Next, we investigated the roles and functions of dysregulated molecules within two distinct crests. Notably, we found that the two crests related to aging predominantly consisted of the same molecules (Supplementary Fig. 6). To focus on the unique biological functions associated with each crest and eliminate commonly occurring molecules, we removed background molecules present in most stages. To explore the specific biological functions associated with each type of omics data (transcriptomics, proteomics and metabolomics) for both crests, we employed the function annotation approach described above (Methods). In brief, we constructed a similarity network of enriched pathways and identified modules to remove redundant annotations (Supplementary Fig. 6 and Extended Data Fig. 5a,b). Furthermore, we applied the same approach to all modules to reduce redundancy and identify the final functional modules (Methods and Extended Data Fig. 6a). Our analysis revealed significant changes in multiple modules associated with the two crests (Extended Data Fig. 6b–d). To present the results clearly, Fig. 5a displays the top 20 pathways (according to adjusted P value) for each type of omics data, and the Supplementary Data provides a comprehensive list of all identified functional modules.Fig. 5Functional analysis of aging-related waves of molecules across the human lifespan.a, Pathway enrichment and biological functional module analysis for crests 1 and 2. Dots and lines are color-coded by omics type. b, The overlapping of molecules between two crests and three clusters. a, Pathway enrichment and biological functional module analysis for crests 1 and 2. Dots and lines are color-coded by omics type. b, The overlapping of molecules between two crests and three clusters. Interestingly, the analysis identifies many dysregulated functional modules in crests 1 and 2, indicating a nonlinear risk for aging-related diseases. In particular, several modules associated with CVD were identified in both crest 1 and crest 2 (Fig. 5a), which is consistent with the above results (Fig. 3b). For instance, the dysregulation of platelet degranulation (crest 1: adjusted P = 1.77 × 10; crest 2: adjusted P = 1.73 × 10), complement cascade (crest 1: adjusted P = 3.84 × 10; crest 2: adjusted P = 2.02 × 10), complement and coagulation cascades (crest 1: adjusted P = 1.78 × 10; crest 2: adjusted P = 2.02 × 10), protein activation cascade (crest 1: adjusted P = 1.56 × 10; crest 2: adjusted P = 1.61 × 10) and protease binding (crest 1: adjusted P = 2.7 × 10; crest 2: adjusted P = 0.0114) have various effects on the cardiovascular system and can contribute to various CVDs. Furthermore, blood coagulation (crest 1: adjusted P = 1.48 × 10; crest 2: adjusted P = 9.10 × 10) and fibrinolysis (crest 1: adjusted P = 2.11 × 10; crest 2: adjusted P = 1.64 × 10) were also identified, which are essential processes for maintaining blood fluidity, and dysregulation in these modules can lead to thrombotic and cardiovascular events. We also identified certain dysregulated metabolic modules associated with CVD. For example, aging has been linked to an incremental rise in plasma phenylalanine levels (crest 1: adjusted P = 9.214 × 10; crest 2: adjusted P = 0.0453), which can contribute to the development of cardiac hypertrophy, fibrosis and dysfunction. Branched-chain amino acids (BCAAs), including valine, leucine and isoleucine (crest 1: adjusted P: not significant (NS); crest 2: adjusted P = 0.0173), have also been implicated in CVD development and T2D, highlighting their relevance in CVD pathophysiology. Furthermore, research suggests that alpha-linolenic acid (ALA) and linoleic acid metabolism (crest 1: adjusted P: NS; crest 2: adjusted P = 0.0217) may be protective against coronary heart disease. Our investigation also identified lipid metabolism modules that are associated with CVD, including high-density lipoprotein (HDL) remodeling (crest 1: adjusted P = 1.073 × 10; crest 2: adjusted P = 2.589 × 10) and glycerophospholipid metabolism (crest 1: adjusted P: NS; crest 2: adjusted P = 0.0033), which influence various CVDs. In addition, the dysregulation of skin and muscle stability was found to be increased at crest 1 and crest 2, as evidenced by the identification of numerous modules associated with these processes (Fig. 5a,b). This suggests that the aging of skin and muscle is markedly accelerated at crest 1 and crest 2. The extracellular matrix (ECM) provides structural stability, mechanical strength, elasticity and hydration to the tissues and cells, and the ECM of the skin is mainly composed of collagen, elastin and glycosaminoglycans (GAGs). Phosphatidylinositols are a class of phospholipids that have various roles in cytoskeleton organization. Notably, the dysregulation of ECM structural constituent (crest 1: adjusted P = 3.32 × 10; crest 2: adjusted P = 1.61 × 10), GAG binding (crest 1: adjusted P = 1.805 × 10; crest 2: adjusted P = 4.093 × 10) and phosphatidylinositol binding (crest 1: adjusted P = 3.391 × 10; crest 2: adjusted P = 7.832 × 10) were identified. We also identified cytolysis (crest 1: adjusted P = 2.973 × 10; crest 2: adjusted P: NS), which can increase water loss. The dysregulated actin binding (crest 1: adjusted P = 3.536 × 10; crest 2: adjusted P = 3.435 × 10), actin filament organization (crest 1: adjusted P = 8.406 × 10; crest 2: adjusted P = 1.157 × 10) and regulation of actin cytoskeleton (crest 1: adjusted P = 0.00090242; crest 2: adjusted P = 6.788 × 10) were identified, which affect the structure and function of various tissues. Additionally, cell adhesion is the attachment of a cell to another cell or to ECM via adhesion molecules. We identified the positive regulation of cell adhesion (crest 1: adjusted P = 3.618 × 10; crest 2: adjusted P = 8.272 × 10) module, which can prevent or delay skin aging. Threonine can affect sialic acid production, which is involved in cell adhesion. We also identified the glycine, serine and threonine metabolism (crest 1: adjusted P: NS; crest 2: adjusted P = 0.00506). Additionally, scavenging of heme from plasma was identified (crest 1: adjusted P = 1.176 × 10; crest 2: adjusted P = 1.694 × 10), which can modulate skin aging as excess-free heme can damage cellular components. Rho GTPases regulate a wide range of cellular responses, including changes to the cytoskeleton and cell adhesion (RHO GTPase cycle, crest 1: adjusted P = 9.956 × 10; crest 2: adjusted P = 1.546 × 10). In relation to muscle, previous studies demonstrated that muscle mass decreases by approximately 3–8% per decade after the age of 30, with an even higher decline rate after the age of 60, which consistently coincides with the observed second crest. Interestingly, we identified dysregulation in the module associated with the structural constituent of muscle (crest 1: adjusted P = 0.00565; crest 2: adjusted P = 0.0162), consistent with previous findings. Furthermore, we identified the pathway associated with caffeine metabolism (crest 1: adjusted P = 0.00378; crest 2: adjusted P = 0.0162), which is consistent with our observations above (Fig. 2b) and implies that the capacity to metabolize caffeine undergoes a notable alteration not only around 60 years of age but also around the age of 40 years. In crest 1, we identified specific modules associated with lipid and alcohol metabolism. Previous studies established that lipid metabolism declines with human aging. Our analysis revealed several modules related to lipid metabolism, including plasma lipoprotein remodeling (crest 1: adjusted P = 2.269 × 10), chylomicron assembly (crest 1: adjusted P = 9.065 × 10) and ATP-binding cassette (ABC) transporters (adjusted P = 1.102 × 10). Moreover, we discovered a module linked to alcohol metabolism (alcohol binding, adjusted P = 8.485 × 10), suggesting a decline in alcohol metabolization efficiency with advancing age, particularly around the age of 40, when it significantly diminishes. In crest 2, we observed prominent modules related to immune dysfunction, encompassing acute-phase response (adjusted P = 2.851 × 10), antimicrobial humoral response (adjusted P = 2.181 × 10), zymogen activation (adjusted P = 4.367 × 10), complement binding (adjusted P = 0.002568), mononuclear cell differentiation (adjusted P = 9.352 × 10), viral process (adjusted P = 5.124 × 10) and regulation of hemopoiesis (adjusted P = 3.522 × 10) (Fig. 5a). Age-related changes in the immune system, collectively known as immunosenescence, have been extensively documented, and our results demonstrate a rapid decline at age 60. Furthermore, we also identified modules associated with kidney function (glomerular filtration, adjusted P = 0.00869) and carbohydrate metabolism (carbohydrate binding, adjusted P = 0.01045). Our previous findings indicated a decline in kidney function around the age of 60 years (Fig. 3c), as did the present result of this observation. Previous studies indicated the influence of carbohydrates on aging, characterized by the progressive decline of physiological functions and increased susceptibility to diseases over time. In summary, our analysis identifies many dysregulated functional modules identified in both crest 1 and crest 2 that underlie the risk for various diseases and alterations of biological functions. Notably, we observed an overlap of dysregulated functional modules among clusters 2, 4 and 6 because they overlap at the molecular level, as depicted in Fig. 5b. This indicates that certain molecular components are shared among these clusters and the identified crests. However, it is important to note that numerous molecules are specific to each of the two approaches employed in our study. This suggests that these two approaches complement each other in identifying nonlinear changes in molecules and functions during human aging. By using both approaches, we were able to capture a more comprehensive understanding of the molecular alterations associated with aging and their potential implications for diseases. Analyzing a longitudinal multi-omics dataset involving 108 participants, we successfully captured the dynamic and nonlinear molecular changes that occur during human aging. Our study’s strength lies in the comprehensive nature of the dataset, which includes multiple time point samples for each participant. This longitudinal design enhances the reliability and robustness of our findings compared to cross-sectional studies with only one time point sample for each participant. The first particularly intriguing finding from our analysis is that only a small fraction of molecules (6.6%) displayed linear changes throughout human aging (Fig. 1d). This observation is consistent with previous research and underscores the limitations of relying solely on linear regression to understand the complexity of aging-related molecular changes. Instead, our study revealed that a considerable number of molecules (81.0%) exhibited nonlinear patterns (Fig. 1e). Notably, this nonlinear trend was observed across all types of omics data with remarkably high consistency (Fig. 1e,g), highlighting the widespread functionally relevant nature of these dynamic changes. By unveiling the nonlinear molecular alterations associated with aging, our research contributes to a more comprehensive understanding of the aging process and its molecular underpinnings. To further investigate the nonlinear changing molecules observed in our study, we employed a trajectory clustering approach to group molecules with similar temporal patterns. This analysis revealed the presence of three distinct clusters (Fig. 2c) that exhibited clear and compelling patterns across the human lifespan. These clusters suggest that there are specific age ranges, such as around 60 years old, where distinct and extensive molecular changes occur (Fig. 2c). Functional analysis revealed several modules that exhibited nonlinear changes during human aging. For example, we identified a module associated with oxidative stress, which is consistent with previous studies linking oxidative stress to the aging process (Fig. 3a). Our analysis indicates that this pathway increases significantly after the age of 60 years. In cluster 5, we identified a transcriptomics module associated with mRNA stabilization and autophagy (Fig. 3a). Both of these processes have been implicated in the aging process and are involved in maintaining cellular homeostasis and removing damaged components. Furthermore, our analysis uncovered nonlinear changes in disease risk across aging. In cluster 2, we identified the phenylalanine metabolism pathway (Fig. 3b), which has been associated with cardiac dysfunction during aging. Additionally, we found that the clinical laboratory tests blood urea nitrogen and serum/plasma glucose increase significantly with age (cluster 2; Fig. 3c), indicating a nonlinear decline in kidney function and an increased risk of T2D with age, with a critical threshold occurring approximately at the age of 60 years. In cluster 4, we identified pathways related to cardiovascular health, such as the biosynthesis of unsaturated fatty acids and caffeine metabolism (Fig. 3b). Overall, our study provides compelling evidence for the existence of nonlinear changes in molecular profiles during human aging. By elucidating the specific functional modules and disease-related pathways that exhibit such nonlinear changes, we contribute to a better understanding of the complex molecular dynamics underlying the aging process and its implications for disease risk. Although the trajectory clustering approach proves effective in identifying molecules that undergo nonlinear changes, it may not be as proficient in capturing substantial alterations that occur at specific time points without exhibiting a consistent pattern in other stages. We then employed a modified version of the DE-SWAN algorithm to comprehensively investigate changes in multi-omics profiling throughout human aging. This approach enabled us to identify waves of dysregulated molecules and microbes across the human lifespan. Our analysis revealed two prominent crests occurring around the ages of 40 years and 60 years, which were consistent across various omics data types, suggesting their universal nature (Fig. 4a,e). Notably, in the proteomics data, we observed crests around the ages of 40 years and 60 years, which aligns approximately with a previous study (which reported crests at ages 34 years, 60 years and 78 years). Due to the age range of our cohort being 25–75 years, we did not detect the third peak. Furthermore, the differences in proteomics data acquisition platforms (mass spectrometry versus SomaScan) resulted in different identified proteins, with only a small overlap (1,305 proteins versus 302 proteins, of which only 75 were shared). This discrepancy may explain the age variation of the first crest identified in the two studies (approximately 10 years). However, despite the differences in the two proteomics datasets, the wave patterns observed in both studies were highly similar (Fig. 4a). Remarkably, by considering multiple omics data types, we consistently identified similar crests for each type, indicating the universality of these waves of change across plasma molecules and microbes from various body sites (Fig. 4e and Supplementary Fig. 3). The analysis of molecular functionality in the two distinct crests revealed the presence of several modules, indicating a nonlinear increase in the risks of various diseases (Fig. 5a). Both crest 1 and crest 2 exhibit the identification of multiple modules associated with CVD, which aligns with the aforementioned findings (Fig. 3b). Moreover, we observed an escalated dysregulation in skin and muscle functioning in both crest 1 and crest 2. Additionally, we identified a pathway linked to caffeine metabolism, indicating a noticeable alteration in caffeine metabolization not only around the age of 60 but also around the age of 40. This shift may be due to either a metabolic shift or a change in caffeine consumption. In crest 1, we also identified specific modules associated with lipid and alcohol metabolism, whereas crest 2 demonstrated prominent modules related to immune dysfunction. Furthermore, we also detected modules associated with kidney function and carbohydrate metabolism, which is consistent with our above results. These findings reinforce our previous observations regarding a decline in kidney function around the age of 60 years (Fig. 3c) while shedding light on the impact of dysregulated functional modules in both crest 1 and crest 2, suggesting nonlinear changes in disease risk and functional dysregulation. Notably, we identified an overlap of dysregulated functional modules among clusters 2, 4 and 6, indicating molecular-level similarities between these clusters and the identified crests (Fig. 5b). This suggests the presence of shared molecular components among these clusters and crests. However, it is crucial to note that there are also numerous molecules specific to each of the two approaches employed in our study, indicating that these approaches complement each other in identifying nonlinear changes in molecules and functions during human aging. The present research is subject to certain constraints. We accounted for many basic characteristics (confounders) of participants in the cohort; but because this study primarily reflects between-individual differences, there may be additional confounders due to the different age distributions of the participants. For example, we identified a notable decrease in oxygen carrier activity around age 60 (Figs. 2c and 3a) and marked variations in alcohol and caffeine metabolism around ages 40 and 60 (Fig. 3a). However, these findings might be shaped by participants’ lifestyle—that is, physical activity and their alcohol and caffeine intake. Regrettably, we do not have such detailed behavioral data for the entire group, necessitating validation in upcoming research. Although initial BMI and insulin sensitivity measurements were available at cohort entry, subsequent metrics during the observation span were absent, marking a study limitation. A further constraint is our cohort’s modest size, encompassing merely 108 individuals (eight individuals between 25 years and 40 years of age), which hampers the full utilization of deep learning and may affect the robustness of the identification of nonlinear changing features in Fig. 1e. Although advanced computational techniques, including deep learning, are pivotal for probing nonlinear patterns, our sample size poses restrictions. Expanding the cohort size in subsequent research would be instrumental in harnessing the full potential of machine learning tools. Another limitation of our study is that the recruitment of participants was within the community around Stanford University, driven by rigorous sample collection procedures and the substantial expenses associated with setting up a longitudinal cohort. Although our participants exhibited a considerable degree of ethnic age and biological sex diversity (Fig. 1a and Supplementary Data), it is important to acknowledge that our cohort may not fully represent the diversity of the broader population. The selectivity of our cohort limits the generalizability of our findings. Future studies should aim to include a more diverse cohort to enhance the external validity and applicability of the results. In addition, the mean observation span for participants was 626 days, which is insufficient for detailed inflection point analyses. Our cohort’s age range of 25–70 years lacks individuals who lie outside of this range. The molecular nonlinearity detected might be subject to inherent variations or oscillations, a factor to consider during interpretation. Our analysis has not delved into the nuances of the dynamical systems theory, which provides a robust mathematical framework for understanding observed behaviors. Delving into this theory in future endeavors may yield enhanced clarity and interpretation of the data. Moreover, it should be noted that, in our study, the observed nonlinear molecular changes occurred across individuals of varying ages rather than within the same individuals. This is attributed to the fact that, despite our longitudinal study, the follow-up period for our participants was relatively brief for following aging patterns (median, 1.7 years; Extended Data Fig. 1g). Such a timeframe is inadequate for detecting nonlinear molecular changes that unfold over decades throughout the human lifespan. Addressing this limitation in future research is essential. Lastly, our study’s molecular data are derived exclusively from blood samples, casting doubt on its direct relevance to specific tissues, such as the skin or muscles. We propose that blood gene expression variations might hint at overarching physiological alterations, potentially impacting the ECM in tissues, including skin and muscle. Notably, some blood-based biomarkers and transcripts have demonstrated correlations with tissue modifications, inflammation and other elements influencing the ECM across diverse tissues. In our future endeavors, the definitive confirmation of our findings hinges on determining if nonlinear molecular patterns align with nonlinear changes in functional capacities, disease occurrences and mortality hazards. For a holistic grasp of this, amalgamating multifaceted data from long-term cohort studies covering several decades becomes crucial. Such data should encompass molecular markers, comprehensive medical records, functional assessments and mortality data. Moreover, employing cutting-edge statistical techniques is vital to intricately decipher the ties between these nonlinear molecular paths and health-centric results. In summary, the unique contribution of our study lies not merely in reaffirming the nonlinear nature of aging but also in the depth and breadth of the multi-omics data that we analyzed. Our study goes beyond stating that aging is nonlinear by identifying specific patterns, inflection points and potential waves in aging across multiple layers of biological data during human aging. Identifying specific clusters with distinct patterns, functional implications and disease risks enhances our understanding of the aging process. By considering the nonlinear dynamics of aging-related changes, we can gain insights into specific periods of significant changes (around age 40 and age 60) and the molecular mechanisms underlying age-related diseases, which could lead to the development of early diagnosis and prevention strategies. These comprehensive multi-omics data and the approach allow for a more nuanced understanding of the complexities involved in the aging process, which we think adds value to the existing body of research. However, further research is needed to validate and expand upon these findings, potentially incorporating larger cohorts to capture the full complexity of aging. The participant recruitment, sample collection, data acquisition and data processing were documented in previous studies conducted by Zhou et al., Ahadi et al., Schüssler-Fiorenza Rose et al., Hornburg et al. and Zhou et al.. Participants provided informed written consent for the study under research protocol 23602, which was approved by the Stanford University institutional review board. This study adheres to all relevant ethical regulations, ensuring informed consents were obtained from all participants. All participants consented to publication of potentially identifiable information. The cohort comprised 108 participants who underwent follow-up assessments. Exclusion criteria encompassed conditions such as anemia, kidney disease, a history of CVD, cancer, chronic inflammation or psychiatric illnesses as well as any prior bariatric surgery or liposuction. Each participant who met the eligibility criteria and provided informed consent underwent a one-time modified insulin suppression test to quantify insulin-mediated glucose uptake at the beginning of the enrollment. The steady-state plasma glucose (SSPG) levels served as a direct indicator of each individual’s insulin sensitivity in processing a glucose load. We categorized individuals with SSPG levels below 150 mg dl as insulin sensitive and those with levels of 150 mg dl or higher as insulin resistant. Thirty-eight participants were missing SSPG values, rendering their insulin resistance or sensitivity status undetermined. We also collected fasting plasma glucose (FPG) data for 69 participants at enrollment. Based on the FPG levels, two participants were identified as having diabetes at enrollment, with FPG levels exceeding 126 mg dl (Supplementary Data). Additionally, we measured hemoglobin A1C (HbA1C) levels during each visit, using it as a marker for average glucose levels over the past 3 months: 6.5% or higher indicates diabetes. Accordingly, four participants developed diabetes during the study period. At the beginning of the enrollment, BMI was also measured for each participant. Participants received no compensation. Comprehensive sample collection was conducted during the follow-up period, and multi-omics data were acquired (Fig. 1b). For each visit, the participants self-reported as healthy or non-healthy. To ensure accuracy and minimize the impact of confounding factors, only samples from individuals classified as healthy were selected for subsequent analysis. Transcriptomic profiling was conducted on flash-frozen PBMCs. RNA isolation was performed using a QIAGEN All Prep kit. Subsequently, RNA libraries were assembled using an input of 500 ng of total RNA. In brief, ribosomal RNA (rRNA) was selectively eliminated from the total RNA pool, followed by purification and fragmentation. Reverse transcription was carried out using a random primer outfitted with an Illumina-specific adaptor to yield a cDNA library. A terminal tagging procedure was used to incorporate a second adaptor sequence. The final cDNA library underwent amplification. RNA sequencing libraries underwent sequencing on an Illumina HiSeq 2000 platform. Library quantification was performed via an Agilent Bioanalyzer and Qubit fluorometric quantification (Thermo Fisher Scientific) using a high-sensitivity dsDNA kit. After normalization, barcoded libraries were pooled at equimolar ratios into a multiplexed sequencing library. An average of 5–6 libraries were processed per HiSeq 2000 lane. Standard Illumina pipelines were employed for image analysis and base calling. Read alignment to the hg19 reference genome and personal exomes was achieved using the TopHat package, followed by transcript assembly and expression quantification via HTseq and DESeq2. In the realm of data pre-processing, genes with an average read count across all samples lower than 0.5 were excluded. Samples exhibiting an average read count lower than 0.5 across all remaining genes were likewise removed. For subsequent global variance and correlation assessments, genes with an average read count of less than 1 were eliminated. Plasma sample tryptic peptides were fractionated using a NanoLC 425 System (SCIEX) operating at a flow rate of 5 μl min under a trap-elute configuration with a 0.5 × 10 mm ChromXP column (SCIEX). The liquid chromatography gradient was programmed for a 43-min run, transitioning from 4% to 32% of mobile phase B, with an overall run time of 1 h. Mobile phase A consisted of water with 0.1% formic acid, and mobile phase B was formulated with 100% acetonitrile and 0.1% formic acid. An 8-μg aliquot of non-depleted plasma was loaded onto a 15-cm ChromXP column. Mass spectrometry analysis was executed employing SWATH acquisition on a TripleTOF 6600 system. A set of 100 variable Q1 window SWATH acquisition methods was designed in high-sensitivity tandem mass spectrometry (MS/MS) mode. Subsequent data analysis included statistical scoring of peak groups from individual runs via pyProphet, followed by multi-run alignment through TRIC60, ultimately generating a finalized data matrix with a false discovery rate (FDR) of 1% at the peptide level and 10% at the protein level. Protein quantitation was based on the sum of the three most abundant peptide signals for each protein. Batch effect normalization was achieved by subtracting principal components that primarily exhibited batch-associated variation, using Perseus software v.1.4.2.40. A ternary solvent system of acetone, acetonitrile and methanol in a 1:1:1 ratio was used for metabolite extraction. The extracted metabolites were dried under a nitrogen atmosphere and reconstituted in a 1:1 methanol:water mixture before analysis. Metabolite profiles were generated using both hydrophilic interaction chromatography (HILIC) and reverse-phase liquid chromatography (RPLC) under positive and negative ion modes. Thermo Q Exactive Plus mass spectrometers were employed for HILIC and RPLC analyses, respectively, in full MS scan mode. MS/MS data were acquired using quality control (QC) samples. For the HILIC separations, a ZIC-HILIC column was used with mobile phase solutions of 10 mM ammonium acetate in 50:50 and 95:5 acetonitrile:water ratios. In the case of RPLC, a Zorbax SBaq column was used, and the mobile phase consisted of 0.06% acetic acid in water and methanol. Metabolic feature detection was performed using Progenesis QI software. Features from blanks and those lacking sufficient linearity upon dilution were excluded. Only features appearing in more than 33% of the samples were retained for subsequent analyses, and any missing values were imputed using the k-nearest neighbors approach. We employed locally estimated scatterplot smoothing (LOESS) normalization to correct the metabolite-specific signal drift over time. The metid package was used for metabolite annotation. A panel of 62 human cytokines, chemokines and growth factors was analyzed in EDTA-anticoagulated plasma samples using Luminex-based multiplex assays with conjugated antibodies (Affymetrix). Raw fluorescence measurements were standardized to median fluorescence intensity values and subsequently subjected to variance-stabilizing transformation to account for batch-related variations. As previously reported, data points characterized by background noise, termed CHEX, that deviate beyond five standard deviations from the mean (mean ± 5 × s.d.) were excluded from the analyses. The tests encompassed a comprehensive metabolic panel, a full blood count, glucose and HbA1C levels, insulin assays, high-sensitivity C-reactive protein (hsCRP), immunoglobulin M (IgM) and lipid, kidney and liver panels. Lipid extraction and quantification procedures were executed in accordance with established protocols. In summary, complex lipids were isolated from 40 μl of EDTA plasma using a solvent mixture comprising methyl tertiary-butyl ether, methanol and water, followed by a biphasic separation. Subsequent lipid analysis was conducted on the Lipidyzer platform, incorporating a differential mobility spectrometry device (SelexION Technology) and a QTRAP 5500 mass spectrometer (SCIEX). Immediately after arrival, samples were stored at −80 °C. Stool and nasal samples were processed and sequenced in-house at the Jackson Laboratory for Genomic Medicine, whereas oral and skin samples were outsourced to uBiome for additional processing. Skin and oral samples underwent 30 min of beads-beating lysis, followed by a silica-guanidinium thiocyanate-based nucleic acid isolation protocol. The V4 region of the 16S rRNA gene was amplified using specific primers, after which the DNA was barcoded and sequenced on an Illumina NextSeq 500 platform via a 2 × 150-bp paired-end protocol. Similarly, stool and nasal samples were processed for 16S rRNA V1–V3 region amplification using a different set of primers and sequenced on an Illumina MiSeq platform. For data processing, the raw sequencing data were demultiplexed using BCL2FASTQ software and subsequently filtered for quality. Reads with a Q-score lower than 30 were excluded. The DADA2 R package was used for further sequence data processing, which included filtering out reads with ambiguous bases and errors, removing chimeras and aligning sequences against a validated 16S rRNA gene database. Relative abundance calculations for amplicon sequence variants (ASVs) were performed, and samples with inadequate sequencing depth (<1,000 reads) were excluded. Local outlier factor (LOF) was calculated for each point on a depth-richness plot, and samples with abnormal LOF were removed. In summary, rigorous procedures were followed in both the collection and processing stages, leveraging automated systems and specialized software to ensure the quality and integrity of the microbiome data across multiple body sites. For all data processing, statistical analysis and data visualization tasks, RStudio, along with R language (v.4.2.1), was employed. A comprehensive list of the packages used can be found in the Supplementary Note. The Benjamini–Hochberg method was employed to account for multiple comparisons. Spearman correlation coefficients were calculated using the R functions ‘cor’ and ‘cor.test’. Principal-component analysis (PCA) was conducted using the R function ‘princomp’. Before all the analyses, the confounders, such as BMI, sex, IRIS and ethnicity, were adjusted using the previously published method. In brief, we used the intensity of each feature as the dependent variable (Y) and the confounding factors as the independent variables (X) to build a linear regression model. The residuals from this model were then used as the adjusted values for that specific feature. All the omics data were acquired randomly. No statistical methods were used to predetermine the sample size, but our sample sizes are similar to those reported in previous publications, and no data were excluded from the analyses. Additionally, the investigators were blinded to allocation during experiments and outcome assessment to the conditions of the experiments. Data distribution was assumed to be normal, but this was not formally tested. The icons used in figures are from iconfont.cn, which can be used for non-commercial purposes under the MIT license (https://pub.dev/packages/iconfont/license). The ‘cross-sectional’ dataset was created by briefly extracting information from the longitudinal dataset. The mean value was calculated to represent each molecule’s intensity for each participant. Similarly, the age of each participant was determined by calculating the mean value of ages across all sample collection time points. We detected linear changing molecules during human aging using Spearman correlation and linear regression modeling. The confounders, such as BMI, sex, IRIS and ethnicity, were adjusted using the previously published method. Our analysis revealed a high correlation between these two approaches in identifying such molecules. Based on these findings, we used the Spearman correlation approach to showcase the linear changing molecules during human aging. The permutation test was also used to get the permutated P values for each feature. In brief, each feature was subjected to sample label shuffling followed by a recalculation of the Spearman correlation. This process was reiterated 10,000 times, yielding 10,000 permuted Spearman correlations. The original Spearman correlation was then compared against these permuted values to obtain the permuted P values. To depict the dysregulated molecules during human aging compared to the baseline, we categorized the participants into different age stages based on their ages. The baseline stage was defined as individuals aged 25–40 years. For each age stage group, we employed the Wilcoxon test to identify dysregulated molecules in comparison to the baseline, considering a significance threshold of P < 0.05. Before the statistical analysis, all the confounders were corrected. Subsequently, we visualized the resulting dysregulated molecules at different age stages using a Sankey plot. The permutation test was also used to get the permutated P values for each feature. In brief, we shuffled the sample labels and recalculated the absolute mean difference between the two groups, against which the actual absolute mean difference was benchmarked to derive the permuted P values. To identify the molecules and microbes that exhibited significant changes at any given age stage, we adjusted the P values for each feature by multiplying them by 6. This adjustment adheres to the Bonferroni correction method, ensuring a rigorous evaluation of statistical significance. To assess whether each type of omics data accurately reflects the ages of individuals in our dataset, we conducted a PCA. Subsequently, we computed the Spearman correlation coefficient between the ages of participants and the first principal component (PC1). The absolute value of this coefficient was used to evaluate the degree to which the omics data reflect the ages (Fig. 2a). PLS regression was also used to compare the strength of the age effect to the different omics data types. In brief, the ‘pls’ function from the R package mixOmics was used to construct the regression model between omics data and ages. Then, the ‘perf’ function was used to assess the performance of all the modules with sevenfold cross-validation. The R was extracted to assess the strength of the age effect on the different omics data types. To accommodate the varying time points of biological and omics data, we employed the LOESS approach. This approach allowed us to smooth and predict the multi-omics data at specific time points (that is, every half year). In brief, for each molecule, we fitted a LOESS regression model. During the fitting process, the LOESS argument ‘span’ was optimized through cross-validation. This ensured that the LOESS model provided an accurate and non-overfitting fit to the data (Supplementary Fig. 2a,b). Once we obtained the LOESS prediction model, we applied it to predict the intensity of each molecule at every half-year time point. To conduct trajectory clustering analysis, we employed the fuzzy c-means clustering approach available in the R package ‘Mfuzz.’ This approach was previously described in our publication. The analysis proceeded in several steps. First, the omics data were auto-scaled to ensure comparable ranges. Next, we computed the minimum centroid distances for a range of cluster numbers, specifically from 2 to 22, in step 1. These minimum centroid distances served as a cluster validity index, helping us determine the optimal cluster number. Based on predefined rules, we selected the optimal cluster number. To refine the accuracy of this selection, we merged clusters with center expression data correlations greater than 0.8 into a single cluster. This step aimed to capture similar patterns within the data. The resulting optimal cluster number was then used for the fuzzy c-means clustering. Only molecules with memberships above 0.5 were retained within each cluster for further analysis. This threshold ensured that the molecules exhibited a strong association with their assigned cluster and contributed considerably to the cluster’s characteristics. Pathway enrichment analysis was conducted using the ‘clusterProfiler’ R package. The GO, KEGG and Reactome databases were used. The P values were adjusted using the Benjamini–Hochberg method, with a significance threshold set at <0.05. To minimize redundant enriched pathways and GO terms, we employed a series of analyses. First, for enriched GO terms, we used the ‘Wang’ algorithm from the R package ‘simplifyEnrichment’ to calculate the similarity between GO terms. Only connections with a similarity score greater than 0.7 were retained to construct the GO term similarity network. Subsequently, community analysis was performed using the ‘igraph’ R package to partition the network into distinct modules. The GO term with the smallest enrichment adjusted P value was chosen as the representative within each module. The same approach was applied to the enriched KEGG and Reactome pathways, with one slight modification. In this case, the ‘jaccard’ algorithm was used to calculate the similarity between pathways, and a similarity cutoff of 0.5 was employed for the Jaccard index. After removing redundant enriched pathways, we combined all the remaining GO terms and pathways. Subsequently, we calculated the similarity between these merged entities using the Jaccard index. This similarity analysis aimed to capture the overlap and relationships between the different GO terms and pathways. Using the same approach as before, we performed community analysis to identify distinct biological functional modules based on the merged GO terms and pathways. First, we used the ‘Wang’ algorithm for the GO database and the ‘jaccard’ algorithm for the KEGG and Reactome databases to calculate the similarity between pathways. The enriched pathways served as nodes in a similarity network, with edges representing the similarity between two nodes. Next, we employed the R package ‘igraph’ to identify modules within the network based on edge betweenness. By gradually removing edges with the highest edge betweenness scores, we constructed a hierarchical map known as a dendrogram, representing a rooted tree of the graph. The leaf nodes correspond to individual pathways, and the root node represents the entire graph. We then merged pathways within each module, selecting the pathway with the smallest adjusted P value to represent the module. After this step, we merged pathways from all three databases into modules. Subsequently, we repeated the process by calculating the similarity between modules from all three databases using the ‘jaccard’ algorithm. Once again, we employed the same approach described above to identify the functional modules. To perform pathway enrichment analysis for metabolomics data, we used the human KEGG pathway database. This database was obtained from KEGG using the R package ‘massDatabase’. For pathway enrichment analysis, we employed the hypergeometric distribution test from the ‘TidyMass’ project. This statistical test allowed us to assess the enrichment of metabolites within each pathway. To account for multiple tests, P values were adjusted using the Benjamini–Hochberg method. We considered pathways with Benjamini–Hochberg-adjusted P values lower than 0.05 as significantly enriched. The DE-SWAN algorithm was used. To begin, a unique age is selected as the center of a 20-year window. Molecule levels in individuals younger than and older than that age are compared using the Wilcoxon test to assess differential expression. P values are calculated for each molecule, indicating the significance of the observed differences. To ensure sufficient sample sizes for statistical analysis in each time window, the initial window ranges from ages 25 to 50. The left half of this window covers ages 25–40, whereas the right half spans ages 41–50. The window then moves in one-year steps; this is why Fig. 4 displays an age range of 40–65 years. To account for multiple comparisons, these P values are adjusted using Benjamini–Hochberg correction. To evaluate the robustness and relevance of the DE-SWAN results, the algorithm is tested with various parcel widths, including 15 years, 20 years, 25 years and 30 years. Additionally, different q value thresholds, such as <0.0001, <0.001, <0.01 and <0.05, are applied. By comparing the results obtained with these different parameters to results obtained by chance, we can assess the significance of the findings. To generate random results for comparison, the phenotypes of the individuals are randomly permuted, and the modified DE-SWAN algorithm is applied to the permuted dataset. This allows us to determine whether the observed results obtained with DE-SWAN are statistically significant and not merely a result of chance. Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
|
PMC10902711
|
Lipotype acquisition during neural development is not recapitulated in stem cell–derived neurons
|
We report a time-resolved resource of lipid molecule abundance across brain development. Using the resource as a reference, we find that stem cell–derived neurons do not acquire a neural lipotype in vitro, unless supplemented with adequate metabolic precursors.Lipids are a diverse category of biomolecules that constitute and functionalize membranes, serve as energy reservoirs, and function as signaling molecules. Although different tissues and cell types produce and maintain characteristic lipid compositions, termed lipotypes (Hicks et al, 2006; Harayama et al, 2014; Capolupo et al, 2022), how they are acquired during the course of development is largely unknown. Furthermore, the functional implications of maintaining a distinct lipotype are not understood, although the existence of cell type– and tissue-specific lipid homeostasis across organisms points to conserved mechanisms governing critical functions (Yamashita et al, 2014; Bozek et al, 2015). Lipotype acquisition and homeostasis are of particular interest in neuroscience because of the enrichment of distinct lipid molecules in the brain, which is functionally coupled to the sophisticated membrane trafficking system of neural cell types (Davletov & Montecucco, 2010; Puchkov & Haucke, 2013; Lauwers et al, 2016; Ingólfsson et al, 2017). Studies have shown severe impairments of neuronal development in mice lacking the lipid binding protein SCAP (Verheijen et al, 2009), as well as the lipid metabolic enzymes CerS2 (Imgrund et al, 2009) and DAGL (Gao et al, 2010), which are part of the cholesterol biosynthesis, ceramide (Cer) biosynthesis, and diacylglycerol breakdown pathways, respectively. Defects in sphingolipid metabolism have also been associated with neural tube defects in embryos (Stevens & Tang, 1997; Missmer et al, 2006), the degeneration of motor neurons (Bejaoui et al, 2001; Dawkins et al, 2001), and myelination defects (Fewou et al, 2005). Moreover, the loss of ether-linked phosphatidylethanolamines (PE O-, i.e., plasmalogens) and sulfatides has been linked to age-related neurodegeneration in diseases such as Alzheimer’s disease (Han et al, 2001, 2002; Tu et al, 2017). Biochemical studies have also shed light on the various roles of sphingomyelin (SM) and the ganglioside GM1 (Tettamanti et al, 1996; Fan et al, 2021), and potential lipid signaling molecules such as retinoids, terpenoids, steroids, and eicosanoids, in triggering differentiation programs in neural stem cells (Bieberich, 2012). Furthermore, studies performed in vitro (Cao et al, 2009; Pinot et al, 2014) and in vivo (Janssen et al, 2015) have demonstrated that docosahexaenoic acid (DHA, 22:6), found in membrane glycerophospholipids (GPL), is important for neurogenesis and the formation of synapses (Salem et al, 2001; Innis, 2007). Thus, lipids have been identified as key players across neurological development, physiology, and disease. Decades of research on brain lipids have established that the mammalian brain has a unique lipotype characterized by high levels of 22:6-containing GPL, 18:0-containing sphingolipids, and cholesterol (O’Brien et al, 1964; O’Brien & Sampson, 1965a; Fitzner et al, 2020). These canonical lipid biomarkers are conserved between mammalian species, ranging from rodents to humans (Bozek et al, 2015), and present in all neural tissues and most of the cell types (O’Brien et al, 1964; O’Brien & Sampson, 1965a, 1965b; Fitzner et al, 2020) (Fig S1A–J). In particular, these three categories of lipid biomarkers are enriched in neuron-rich gray matter (O’Brien et al, 1964; O’Brien & Sampson, 1965a) (Fig S1A and B), as well as neurons isolated from E16.5 mouse embryos (Fitzner et al, 2020) (Fig S1E). These same markers are also prevalent in synaptic vesicles (Takamori et al, 2006) and the postsynaptic plasma membrane (Tulodziecka et al, 2016) (Fig S1H–J). Because these lipids are primarily membrane constituents, it is likely that the regulation of their levels is functionally coupled to membrane trafficking events and activities of cell surface receptors of especially mature neurons. (A, B) Mol% of DHA (22:6) among fatty acyls present in phosphatidylethanolamine (PE) and phosphatidylserine (PS) in the human brain tissue from humans varying in age (m = months, y = years). Data from O’Brien and Sampson (1965a) (A) and O’Brien et al (1964) (B). (C) Mol% of cholesterol among all lipids extracted from the human brain tissue (O’Brien & Sampson, 1965b). (D) Mol% of 16:0 and 18:0 among fatty acyl chains in ceramides (Cer) and sphingomyelins (SMs) from the brain tissue of cows (O’Brien et al, 1964). (E) Mol% of 22:6 among fatty acyl chains present in phosphatidic acid (PA), PE, PS, and ether PE (PE O-) classes in individual cell types isolated from E16.5 (neurons) and P1 (glia) mouse brain and cultured in vitro (DIV = days in vitro). (F) Mol% of 18:0-containing (36:1; 2) sphingolipids within the respective lipid class and across cell types. (G) Mol% of cholesterol among all lipids across cell types. Data in panels (E, F, G) are from the lipidomic data resource generated by Fitzner et al (2020). (H) Mol% of 22:6-containing PE, PS, and PE O- species in postsynaptic plasma membranes isolated from rat brain at P2, P7, P14, P21, and P60. (I) Mol% of the 18:0-containing (36:1; 2) sphingolipids within their respective lipid class in the postsynaptic plasma membrane. (J) Mol% of cholesterol in the postsynaptic plasma membrane. Data in (H, I, J) are from Tulodziecka et al (2016). Although the canonical lipid biomarkers of neurons are well established, it is still largely unknown how the neuronal lipotype is acquired during development. Unraveling this can in principle come from studies using in vitro neuronal differentiation of stem cells. This, however, presents the conundrum that fully mature neurons in vivo are not able to de novo synthetize the canonical 22:6ω3-containing lipids themselves (Moore et al, 1991; Kim, 2007) and that cell culture media used for in vitro neuronal differentiation are devoid of the essential fatty acid 22:6ω3 and are instead supplemented with the fatty acid 18:3ω3 (Bardy et al, 2015). Moreover, stem cell–derived neurons also do not down-regulate the expression of ceramide synthases 5 and 6, which produce 16:0-containing sphingolipids instead of the neuron-specific 18:0-containing sphingolipids (Fig S10). Thus, albeit stem cell–derived neurons acquire unique morphological hallmarks and express specific protein markers, it is at the present unclear whether the lipotype of stem cell–derived neurons in vitro is comparable to that of mature neurons and brain tissues. To shed light on how the neural lipotype is acquired during development, and to assess the extent to which this can be recapitulated during in vitro neuronal differentiation, we generated a detailed lipidomic resource for early brain development in mice, starting from the embryonic stage of E10.5, where the brain tissue first becomes accessible for dissection, and up to the postnatal stage P21. This revealed that the high levels of canonical neural lipid biomarkers 22:6-GPL and 18:0-sphingolipids begin to be established already at the embryonic stage in utero, coinciding with extensive neurogenesis (E10.5–E15.5) (Caviness, 1982; Finlay & Darlington, 1995), whereas the increase in cholesterol occurs postnatally. Using the resource as a reference, we examined to which extent this can be recapitulated by commonly used protocols for in vitro neuronal differentiation of mouse embryonic stem cells (mESCs) and human induced pluripotent stem cells (hiPSCs). Here, we found that stem cell–derived neurons could only recapitulate a partial neuronal lipotype, and only upon supplementing the culture medium with brain-specific lipid metabolic precursors. Taken together, our resource shows that early mouse brain development coincides with extensive lipidome remodeling, starting already at the embryonic stage, and that in vitro neuronal differentiation using standard protocols yields immature neurons that are only partially committed at the lipid metabolic level. To identify lipid hallmarks associated with neural development, we compiled a comprehensive lipidomic resource of mouse brain development. To this end, we microdissected the whole mouse brain region at the developmental stage E10.5, as well as excised cerebral hemispheres from E15.5, P2, and P21 mice (Fig 1A). These biopsies were analyzed by high-resolution MS lipidomics (Almeida et al, 2015; Sprenger et al, 2021). Overall, this analysis identified and quantified 1,488 lipid molecules encompassing 26 lipid classes (Supplemental Data 1). Statistical analysis revealed that 451 distinct lipid molecules in the cerebral hemispheres (30% of detected lipids) were significantly changed in abundance across the developmental timeline with a twofold or greater difference between E10.5 and P21 (ANOVA followed by multiple hypothesis correction using the Benjamini–Hochberg procedure, P ≤ pc = 0.04) (Fig 1B), of which 133 (9%) and 318 (21%) lipids increased and decreased, respectively. Principal component analysis showed a clustering of samples according to the time point and tissue, with only a weak distinction between the E10.5 brain tissue and the rest of the embryo (Figs 1C and S2). Similarly, the distinction between cerebral hemispheres and the rest of the brain tissues was not pronounced at E15.5 but clearly increased over the course of development as each brain region acquired a more specialized lipotype. (A) Mouse brain tissue was collected at the indicated time points and analyzed by in-depth MS lipidomics. The colors denote distinct brain regions that were sampled at each developmental stage. (B) Volcano plot of molecular lipid species. The fold change of each lipid within its lipid class is calculated between the E10.5 whole brain and the P21 cerebral hemisphere. (C) Principal component analysis of lipid abundance across samples, colored as per the schematic in (A). (D) Profile (mol%) of different lipid classes across developmental time. Bars representing the mean value for each sample group are colored in accordance with panel (A). (E) Lipid feature ENrichment Analysis of features among decreasing and increasing lipids, where −log2 (odds ratio) is plotted in the y-axis, and the color and size correspond to the −log10 (P-value). (F) Profile of the most abundant Cer and SM species across samples, containing an 18:1;2 sphingoid chain and either a 16:0 or an 18:0 acyl chain. (G) Abundance of 22:6 (DHA) among fatty acyls constituting phosphatidic acid (PA), PE, PE O-, and PS species. Data were collected on n = 5 replicates. P-values were obtained using a one-way ANOVA test, and changes were considered significant if P ≤ pc = 0.04, as determined by the Benjamini–Hochberg procedure. PC1 and PC2 loadings of lipids where either loading was greater than 0.1 in magnitude are shown in the plot. Assessing the bulk abundance of lipid classes in the brain tissue across development, we observed a 1.5-fold increase in cholesterol and a twofold increase in PE O-, as well as a corresponding decrease in phosphatidylcholines (PC) in the postnatal phase between P2 and P21. Low levels of cholesterol esters (CE) and triacylglycerol present at E10.5 were further diminished thereafter. The total level of phosphatidylinositols (PI), precursors of signaling phosphoinositides, gradually decreased in abundance throughout development (Fig 1D). To systematically examine whether the lipids increasing in abundance have common molecular traits, we carried out Lipid feature ENrichment Analysis (LENA) (Sprenger et al, 2021), akin to gene ontology analysis of gene transcripts and proteins. This demonstrated that membrane GPL (P = 8 × 10) and sphingolipids (SP, P = 7 × 10) are enriched among the increasing lipid molecules, whereas storage glycerolipids (GL, P = 1 × 10) are depleted from this pool. Moreover, we found that the lysolipids LPC, LPE, LPS, and LPI are enriched in the pool of increasing lipids, in addition to hexosylceramides (HexCer) and PE O- lipids. The polyunsaturated structural attribute 22:6 (P = 6 × 10), corresponding to DHA, is significantly enriched among the increasing lipids, as are sphingolipids with a sphingosine 18:1;2 chain (P = 2 × 10) (Fig 1E). Inspecting the molecular timelines of the canonical brain lipid biomarkers showed that the sphingolipids Cer 18:1;2/18:0 and SM 18:1;2/18:0, as well as 22:6-containing phosphatidylserine (PS), PE, and PE O- species, already reach their expected high levels within the early developmental period studied here (Fig 1F and G). For all time points, we also subjected the remainder of the brain (after removal of the cerebral hemispheres) to lipidomic analysis and found similar lipidomic changes as in the cerebral hemisphere (Fig S3A–E). In addition, the rest of the E10.5 embryos after removal of the brain region were also analyzed (Fig S4A–C), revealing a lipotype similar to the brain region at this time point (with only 13 of 1,219 [1%] analyzed lipids showing a significant twofold or larger change in abundance between the E10.5 brain and the remaining tissue). This suggests that the neural lipotype acquisition has not yet begun as of E10.5, making it a suitable starting point for profiling the lipidomic changes that concur with neural development. (A) Brain regions apart from cerebral hemispheres were collected separately at the indicated time points and analyzed by MS lipidomics. (B) Abundance of different lipid classes across developmental time. Bars representing the mean value for each sample group are colored in accordance with panel (A). (C) Volcano plot of molecular lipid species. The fold change in mol% of each lipid within its lipid class is calculated between the E10.5 brain region and the rest of the P21 brain after removal of the cerebral hemispheres. (D) Most abundant ceramide (Cer) and sphingomyelin (SM) species across samples, containing an 18:1; 2 sphingoid chain and either a 16:0 or an 18:0 acyl chain. (E) Abundance of 22:6 (DHA) among fatty acyls constituting phosphatidic acid (PA), phosphatidylethanolamine (PE and PE O-), and phosphatidylserine (PS). Data were collected on n = 5 replicates. P-values were obtained using a one-way ANOVA test and considered significant if P ≤ pc = 0.041, as determined by the Benjamini–Hochberg procedure. The rest of the E10.5 embryo after excision of the brain region was collected and analyzed by lipidomics. (A) Schematic showing the dissection and resulting portions of the E10.5 embryo in the two sample groups. (B) Volcano plot where each circle is one lipid species and its abundance is compared between the E10.5 brain region and the rest of the E10.5 embryo. Lipids with P ≤ pc = 0.0025 (t test followed by the Benjamini–Hochberg procedure) and fold changes greater than twofold are colored red. (C) Lipid class profile of the two tissue samples collected at E10.5. Data were collected on n = 5 replicates. Taken together, our lipidomic resource provides data on the molar abundance of individual lipid species with annotation of individual fatty acyl chains, thereby providing a comprehensive compendium of over a 1,000 molecular lipid species. We observed that neural development coincides with a stepwise lipotype acquisition. Specifically, our results reveal that 18:0-sphingolipids and 22:6-GPL (i.e., PS, PE, and PE O-) already become enriched in utero, which coincides with the onset of neurogenesis (Caviness, 1982; Finlay & Darlington, 1995), whereas enrichment of cholesterol occurs postnatally. The comparison with published data from primary neurons and glial cells further demonstrates that the lipid biomarkers observed early in development by us are present in most of the cell types isolated from the brain and are strongly enriched in neurons (Fig S1E and F). Given that stem cell–derived neurons are commonly used for studying neuronal fate determination and acquisition of specific functions, we next examined to what extent in vitro neuronal differentiation yields neurons with lipidomic hallmarks akin to the neural lipotype observed in vivo. To this end, we made use of a commonly used protocol (Bibel et al, 2007) where mESCs are grown as suspended aggregates and differentiated into neuronal progenitors (i.e., immature neurons) over a time period of 12 d (Fig 2A) (Tiwari et al, 2012, Gehre et al, 2020; Song et al, 2021; Muckenhuber et al, 2023). The rationale for not going beyond 12 d is that longer culture time concurs with a loss in overall cell viability. Notably, in our hands, 84% ± 1% cells were positive for β-tub III, indicating the high purity of the neuronal progenitors on day 12 (Fig 2B). (A) Schematic of the neuronal differentiation protocol for mouse embryonic stem cells with representative bright-field images at stages in the differentiation protocol where samples were collected for lipidomics. (B) Representative fluorescence images of cells on day 12 stained for DAPI and the neuronal marker βTub-III are shown. Quantification of βTub-III–positive cells is shown on the right, based on 16 images from n = 2 biological replicates. (C) Volcano plot of molecular lipid species. The fold change in mol% of each molecular lipid species is calculated between day 0 and day 12 of differentiation. (D) Profile of different lipid classes across time. (E) Lipid feature ENrichment Analysis of the pool of decreasing and increasing lipids. (F) Profile of sphingolipids of interest within their respective lipid classes. (G) Profile of 22:6 (DHA) among fatty acyls constituting PA, PE, PE O-, and PS species. Data were collected on n = 4 replicates. P-values were obtained using a one-way ANOVA test, and changes were considered significant if P ≤ pc = 0.023 as determined by the Benjamini–Hochberg procedure. Along this timeline, we analyzed cells by MS lipidomics at 0, 4, 8, and 12 d after the onset of differentiation. Overall, the lipidomic analysis afforded quantitative monitoring of 1,355 lipid molecules encompassing 26 lipid classes (Supplemental Data 1). Of these, the levels of 431 lipids (32%) were significantly altered over the course of differentiation, with a twofold or greater difference between day 0 and day 12 (ANOVA followed by a Benjamini–Hochberg procedure, P-value ≤ pc = 0.023), with 200 (15%) and 231 (17%) increasing and decreasing, respectively (Fig 2C). At the bulk lipid class level, we observed a significant increase in the storage lipids CE and triacylglycerol on day 4 and day 8, indicating that the cells were storing fat at this stage (Fig 2D). LENA demonstrated that the structural features GPL, 16:1, PE O-, and PS were enriched in the pool of increasing lipids (P = 8 × 10, 1 × 10, 1 × 10, 1 × 10). The increase in 16:1 is indicative of de novo lipid synthesis (Lin & Smith, 1978; Chakravarty et al, 2004; Freyre et al, 2019), which coincides with the removal of serum from the culture medium and a switch to using glucose as the primary substrate for de novo lipogenesis on day 8 (Supplemental Data 1). Furthermore, the polyunsaturated features 20:4 and 22:4 were depleted among the pool of increasing lipids (P = 4 × 10, 3 × 10) (Fig 2E). Lastly, we inspected the temporal dynamics of the canonical neural lipid biomarkers during the in vitro neuronal differentiation. We found that only cholesterol is significantly increased, from 19 mol% on day 0 to 22 mol% on day 12 (Fig 2D), reaching similar values as in the developing embryonic brain (21 mol% at E10.5 and 23 mol% at E15.5), but far from those observed in neural tissue (32 mol% at P21). Notably, all 22:6-GPL are relatively low in abundance at all time points and most of these are also progressively reduced during the in vitro differentiation (Fig 2G). A similar trend was observed for the sphingolipids Cer 18:1;2/18:0 and SM 18:1;2/18:0, which was offset by an increase in Cer 18:1;2/16:0 and SM 18:1;2/16:0 (Fig 2F). Notably, 16:0-sphingolipids were found to decrease considerably during brain development in vivo, while being replaced by 18:0-sphingolipids (Fig 1F). Apart from the lack of lipid remodeling, we also observed that the molecular composition of PIs in the cells differed from that of the brain tissue, with elevated levels of 18:1 chains and reduced levels of 20:4 chains in the PI molecules (Fig S5). Comparison of the mol% abundance of various fatty acyl chains among PIs in mouse embryonic stem cells being differentiated in vitro into neurons and mouse brain tissues over development. Data were collected on n = 4 replicates of cells and n = 5 replicates of the brain tissue. In summary, our analysis shows that although the differentiated mouse stem cells resemble neurons by morphology and express specific protein markers on day 12 of differentiation, they do not acquire a unique lipotype akin to the brain tissue or primary neurons (Figs 1 and S1). In fact, the lipidome of cells at day 12 is most similar to that of the E10.5 brain (Fig S6), where the neural lipotype acquisition has not yet begun (Fig S4). The Euclidean distance between samples based on mol% abundance of all detected lipids. The labels “h” and “r” denote the hemisphere and rest of the brain tissue, respectively. Prompted by our finding that neuronal progenitors generated by culturing mESCs in embryoid bodies (Bibel et al, 2007) do not acquire a neural lipotype, we examined lipotype acquisition in two other lineages of stem cell–derived neurons: (1) in vitro differentiation of mESCs cultured in an adherent monolayer and stimulated with retinoic acid (Ying et al, 2003) (Fig S7A) and (2) differentiation of human iPSCs to a population of predominantly dopaminergic neurons (Bogetofte et al, 2019), because the canonical brain lipid biomarkers are known to be conserved among mammals, ranging from mice to humans (Fig S1). Briefly, the human iPSCs are differentiated into neural stem cells (NSCs) through a neural rosette-based protocol (Swistowski et al, 2009). The NSCs are then further differentiated for up to 25 d to postmitotic neurons with Sonic hedgehog (Shh) stimulation to induce dopaminergic specification (Bogetofte et al, 2019). (A) Neuronal differentiation of adherent mouse embryonic stem cells. (B) Relative abundance of different lipid classes across time. Samples from days −2, −1, and 0 are all in N2B27 + 2iLIF and are collected 2, 1, and 0 d before the removal of 2iLIF and the start of differentiation. (C) Relative abundance of 22:6-glycerophospholipids PE 40:6, PE O-40:7, and PS 40:6; 22:6-containing PA was not detected in this dataset and is therefore not shown. (D) Mol% abundance of sphingolipids of interest within their respective lipid classes is plotted across samples. As the data are from a single replicate of the differentiation protocol, results were not tested for significance and are meant to be indicative rather than conclusive. Mol% abundance of 22:6-glycerophospholipids and 18:0-sphingolipids in cerebral hemispheres across development is shown on the right panel in (C, D) for comparison. The neuronal lineage of these cells was confirmed by immunostaining of β-tub III and Microtubule Associated Protein 2 (MAP2) on day 25, where 86.4% ± 1.0% and 67.9% ± 0.9% of the cells were positive for β-tub III and MAP2, respectively (Fig S8A and B). In both cases, the differentiated cells failed to acquire the characteristic lipotype of neural tissue (Figs S7B–D and S8C–E), despite presenting morphological features of neurons, expressing neuron-specific mRNA and protein biomarkers (e.g., β-tub III, MAP2, Sox1, NeuN), and having synaptic activity (Bibel et al, 2007; Bogetofte et al, 2019). These findings suggest a general failure of in vitro neuronal differentiation models in prompting cells to acquire the characteristic lipotype of the brain tissue and especially neurons in vivo. This discrepancy highlights a challenge in using in vitro differentiated neuronal progenitors for mechanistic studies of lipid metabolic programming and lipotype acquisition. Moreover, it prompts the need to develop new in vitro differentiation protocols that can adequately commit stem cells to acquire a neural lipotype. We note here that the standard culture media for in vitro neuronal differentiation of mESCs, containing the B27 supplement, are devoid of the essential polyunsaturated fatty acids DHA (22:6ω3) and arachidonic acid (20:4ω6). Instead, the B27 supplement contains their respective precursors, linolenic acid (18:3ω3) and linoleic acid (18:2ω6) (Brewer & Cotman, 1989; Brewer et al, 1993). The lack of 22:6ω3 is particularly surprising given that 22:6ω3 deficiency in mothers during pregnancy and in children is known to negatively impact growth and cognitive development (Lauritzen et al, 2016). (A) Immunofluorescence images of human induced pluripotent stem cell–derived neurons on day 25 of differentiation from the neural stem cell stage, labeled for the neuronal markers, β-tub III and Microtubule Associated Protein 2. Regions in dashed white lines are shown below. (B) Percentage of cells positive for β-tub III and Microtubule Associated Protein 2 as quantified from 45 images per group, from n = 3 biological replicates. (C) Abundance of lipid classes across time from days 0, 10, and 25 of differentiation. (D) Abundance of the 22:6-GPL PE 18:0-22:6, PS 18:0-22:6, and PE O-18:1/22:6; 22:6-containing PA was not detected in this dataset and is therefore not shown. (E) Abundance of 16:0- and 18:0-sphingolipids; the most abundant Cer and SM species across samples contain a 16:0 fatty acyl chain on days 0 and 10 but are replaced by 18:0-containing Cer and SM, respectively, on day 25. Data were obtained from n = 2 independent differentiations. Mol% abundance of 22:6-GPL and 18:0-sphingolipids in cerebral hemispheres across development is shown on the right panel in (C, D) for comparison. Studies using primary rat neurons and astrocytes have shown that the polyunsaturated fatty acids 22:6ω3 and 20:4ω6, found to be abundant in neural lipids, are produced by astrocytes from the precursors, 18:3ω3 and 18:2ω6, and are thereafter salvaged by neurons (Moore et al, 1991; Kim, 2007). This indicates that the external supply of fatty acids to in vitro differentiated stem cells is important for lipid metabolic remodeling and accretion of the neural lipotype. Thus, we attempted to recapitulate the lipotype acquisition observed in vivo by supplementing the culture media of differentiating mESCs with relevant neural lipid precursors. Specifically, we supplemented the cells with a mix of 20 μM DHA (for the production of 22:6-PE, PE O-, and PS lipids), 10 μM arachidonic acid (for the biosynthesis of 20:4-PI lipids), and 30 μM stearic acid (for the production of 18:0-sphingolipids). We added the fatty acid mixture to the cell culture media from day 8 of differentiation (Fig 3A). We then collected the fatty acid–supplemented cells, as well as untreated control cells on day 12 for lipidomics. (A) Schematic of the differentiation protocol where cells were supplemented with fatty acids. (B) Volcano plot of molecular lipid species. The fold change in mol% abundance of each molecular lipid species was calculated between the FA-supplemented and control cells. Red circles correspond to lipid species containing four or more double bonds. (C) Profile of Cer and SM 18:1;2/16:0 and 18:1;2/18:0 in control (cyan bar) and FA-supplemented (orange bar) cells on day 12 of differentiation. (D) Profile of 22:6 among fatty acyls constituting PA, PE, PE O-, and PS species. Data were collected on n = 4 replicates. P-values were obtained using a t test and considered significant if P ≤ pc = 0.029, as determined by the Benjamini–Hochberg procedure. The lipidomic analysis identified and quantified 1,200 lipid molecules from 26 lipid classes (Supplemental Data 1). We found that among the 310 lipid molecules (26%) significantly altered in abundance by twofold or more (a t test followed by the Benjamini–Hochberg procedure, P-value ≤ pc = 0.029), 124 (10%) were increased and 186 (16%) were reduced. 112 (90%) of the increased lipids likely featured a polyunsaturated chain (using the criteria that at least one of the acyl chains had ≥4 double bonds or that the total number of double bonds in the lipid ≥4 for lipids whose acyl chain composition could not be determined) (Fig 3B). The incorporation of 22:6 into PE, PE O-, and PS closely resembled that seen in the P21 brain, with the molar abundance of 22:6 reaching 20 mol% in PE, 25 mol% in PE O-, and 31 mol% in PS (as compared to 22, 22, and 33 mol% in the P21 cerebral hemisphere) (Fig 3D). The incorporation of 20:4 into PI also resembled that seen in the brain, with the molar abundance of 20:4 reaching 38 mol% (as compared to 43 mol% in the P21 cerebral hemisphere) (Fig S9B). The stearic acid supplementation promoted only a modest increase in 18:0-containing sphingolipid species, reaching 28 mol% for ceramides and 12 mol% for SMs (as compared to 76 mol% for both ceramides and SMs in the P21 cerebral hemisphere) (Fig 3C). We note that fatty acid supplementation did not alter the bulk level of individual lipid classes (Fig S9A). (A) Profile of lipid classes in control versus FA-supplemented in vitro differentiated neurons. (B) Mol% abundance of various fatty acyl chains among PIs in the developing mouse brain and FA-supplemented in vitro differentiated neurons. Data were collected on n = 4 replicates. In summary, it appears that the lipid metabolic machinery underpinning the neural lipotype is partially established in the in vitro differentiated neuronal progenitors. On the one hand, it is correctly programmed to be able to take up and incorporate the essential polyunsaturated 22:6 and 20:4 into membrane GPL. On the other hand, the metabolic branch responsible for sphingolipid production appears to require more rewiring as it fails to produce the high levels of 18:0-containing sphingolipids seen in the developed brain tissue. Together, our results pinpoint two key factors responsible for lipotype acquisition, namely, cell-intrinsic enzymatic activities of the underlying lipid metabolic machinery and cell-extrinsic lipid building blocks such as polyunsaturated fatty acids (i.e., derived from the cell culture medium in vitro and the neighboring cell types in tissues in vivo). In this study, we generated a comprehensive lipidomic resource of mouse brain development, starting at the early embryonic stage E10.5. This resource compliments previous lipidomic investigations of brain regions, cell types, and membrane fractions obtained from postnatal pups and adolescent animals (Breckenridge et al, 1972; Dawson, 2015; Lauwers et al, 2016; Tulodziecka et al, 2016; Fitzner et al, 2020), and adds to the emerging multi-omics picture of neural development (Yousefi et al, 2021). Our time series analysis allowed us to identify neural lipid biomarkers that increase in abundance throughout development, starting in utero (22:6-glycerophospholipids and 18:0-sphingolipids), whereas cholesterol increases postnatally (Fig 1). The specific lipotype acquisition that we observe in the developing mouse brain underscores the important role of lipids and their metabolism in the acquisition of neural functions in vivo. Moreover, the early onset of lipotype acquisition evident at E15.5 indicates that this process is closely coupled to cell differentiation in the brain. Using the lipidomic resource as a reference, we examined to what extent commonly used protocols for in vitro neuronal differentiation of mESCs and human iPSCs concur with extensive lipid metabolic remodeling and acquisition of a neural lipotype. Despite the evident acquisition of neuronal morphology and the expression of neuronal markers in vitro (Figs 2, S7, and S8) (Izant & McIntosh, 1980; Haendel et al, 1996; Katsetos et al, 2003; Daubner et al, 2011), we did not find lipidomic changes similar to those seen across brain development. Two factors could contribute to this incomplete lipotype acquisition, namely, insufficient programming of the underlying lipid metabolic machinery required for the biosynthesis of neural lipids, the absence of required metabolic precursors, or a combination thereof. It has previously been shown that DHA (22:6) is produced in astrocytes and is thereafter salvaged by neurons (Moore et al, 1991; Kim, 2007). Our fatty acid supplementation confirms that the lack of 22:6-GPL is indeed due to a deficiency in the culture medium, as supplementation with 20 μM DHA allows cells to produce 22:6-GPL to a level comparable with the neural lipotype (Fig 3). On the contrary, supplementing the cells with stearic acid (18:0) does not result in high levels of 18:0-sphingolipids. It is known that the Cer synthase CerS1 is specific for stearoyl-CoA (Venkataraman et al, 2002), which results in the production of 18:0-sphingolipids, whereas the synthases CerS5 and CerS6 are responsible for 16:0-sphingolipid production. During brain development, one observes a 35-fold increase in the expression of Cers1 and a down-regulation of Cers5 and Cers6 compared with embryonic tissue (Sladitschek & Neveu, 2019) (Fig S10A). In contrast, during in vitro neuronal differentiation, between days 8 and 12, Cers1 expression increases only by fivefold, and contrary to expectation, Cers6 expression is up-regulated and Cers5 expression is unchanged (Gehre et al, 2020) (Fig S10B). This could underpin the observation that 16:0-sphingolipids remain elevated, whereas brain-specific 18:0-sphingolipids only increase marginally, despite supplementation with stearic acid. Overall, this suggests that appropriate programming of the sphingolipid metabolic machinery is not fully established in stem cell–derived neuronal progenitors. (A) Expression levels are compared between E7 mouse embryos and pooled brain samples from P56-P84 mice. The data are from n = 1 replicate and taken from RNA-seq (Sladitschek & Neveu, 2019). (B) Expression levels are compared across the in vitro differentiation of mouse embryonic stem cells into neurons in embryoid bodies. The data are from n = 5 replicates and taken from RNA-seq (Gehre et al, 2020). CerS1 and CerS4 are responsible for incorporating C18 chains into sphingolipids, whereas CerS5 and CerS6 incorporate C16 chains (Levy & Futerman, 2010). Previous work has indicated that homeoviscous adaptation can lead to an increase in cholesterol levels to preserve membrane packing when cells of various cell types, including neurons, are supplemented with high levels of PUFAs (Sinensky, 1974; Ernst et al, 2016; Levental et al, 2020). This finding is of particular interest in the context of the brain and neurons, as they have relatively high levels of cholesterol and PUFAs (Figs 1 and S3). However, our data show that cholesterol increases only postnatally and not concomitantly with the increase in PUFAs seen throughout development. Moreover, fatty acid supplementation did not lead to an increase in cholesterol in the differentiating mESCs (Fig S9A), in line with some other cell types (Zech et al, 2009). Although the cholesterol level increases in some of the standard protocols used in this study (Figs 2D and S7B), the final levels are far from those observed in vivo. Previous work has shown that astrocytes are also producers of cholesterol and neurons are recipients (Pfrieger & Ungerer, 2011; Valenza et al, 2015; Ferris et al, 2017), again highlighting the importance of lipid exchange between cell types in the brain. In summary, here we have outlined the lipidomic landscape of brain development in mice by tracing the abundance of more than 1,000 lipids over developmental time. This has allowed us to identify two key changes that begin in the embryonic phase of development, namely, an increase in 22:6-GPL and the replacement of 16:0-sphingolipids by 18:0-sphingolipids. In contrast, we find that the cholesterol level in the brain increases only postnatally. Such orchestrated lipidome remodeling suggests that the differentiation and maturation of brain cells in vivo are coupled to lipid metabolic changes. Although primary neurons from E16.5 embryos show these lipid hallmarks (Fitzner et al, 2020), our attempt to recapitulate this in vitro using neuronal differentiation models to study the molecular mechanisms responsible for lipid metabolic commitment has demonstrated that the acquisition of neural lipid hallmarks is lacking in neuronal progenitors generated in vitro. To address this deficit, we added metabolic precursors for producing the canonical neural lipids, but this only partially established the neural lipotype. Thus, future work is needed to systematically improve in vitro neuronal differentiation protocols, and perhaps consider the need for co-cultures, if they are to be used for mechanistic investigations of lipid biochemistry, membrane biology and biophysics. Our approach provides an important framework for further optimization of differentiation protocols to make in vitro neural cells that more closely recapitulate their in vivo counterparts. Mice were housed in the Laboratory Animal Resources Facility at EMBL Heidelberg in accordance with the guidelines of the European Commission, revised Directive 2010/63/EU and AVMA Guidelines 2007, under veterinarian supervision. No procedure was performed on live animals. The mothers carrying E10.5 and E15.5 embryos, as well as P2 and P21 pups, were euthanized following a protocol approved by the EMBL Institutional Animal Care and Use Committee. Chloroform, methanol, and 2-propanol (Rathburn Chemicals), and ammonium formate (Fluka Analytical) were all of HPLC-grade. Lipid standards were purchased from Avanti Polar Lipids and Larodan Fine Chemicals. mESCs (129XC57BL/6J generated from male 129-B13 agouti mice) were initially cultured on a feeder layer of mouse fibroblast cells from CD1 mice in ESC media containing Knockout DMEM (Cat. #10829018; Gibco) with 15% EmbryoMax FBS (Cat. #ES009-M; Merck) and 20 ng/ml leukemia inhibitory factor (LIF from EMBL Protein Expression and Purification Core Facility), 1% non-essential amino acids (Cat. #11140050; Gibco), 1% GlutaMAX (Cat. #35050061; Gibco), 1 mM sodium pyruvate (Cat. #11360070; Gibco), 1% (50 U/ml) Pen/Strep (Cat. #15070063; Gibco), and 143 μM β-mercaptoethanol (Cat. #21985023; Gibco). They were cultured over three passages after which the feeder cells were selectively depleted from culture by allowing them to adhere to tissue culture dishes for 10 min. The unadhered mESCs were replated for further propagation where they were differentiated into neuronal progenitor cells over the course of 12 d according to the protocol in Bibel et al (2007) (Fig 2A). For this, the mESCs were grown in suspension in non-adherent dishes containing differentiation media with high glucose DMEM (Cat. #11965092; Gibco), 10% FBS (Cat. #26140079; Gibco), and no LIF, but otherwise identical in composition to the ES medium. This results in the formation of embryoid bodies that grow larger over the course of 8 d. On day 4, 5 μM retinoic acid was added to induce neuronal differentiation. On day 8, embryoid bodies were dissociated with trypsin (Cat. #25300054; Gibco) and cells were plated at a density of 140,000 cells/cm on plates precoated with poly-D-lysine and laminin-511 (Cat. #LN511; BioLamina) in N2B27 media (high glucose DMEM supplemented with 1% N2 [Cat. #17502048; Gibco], 1% B27 [without vitamin A, Cat. #17504044; Gibco], 1% [50 U/ml] Pen/Strep, and 1 mM sodium pyruvate [Cat. #11360070; Gibco]). On day 10, cells were treated with 1 μM cytosine arabinoside (AraC, Cat. #C1768; Merck) to kill proliferating, non-differentiated cells. Media were changed every day before day 0 because of rapid cell proliferation in the stem cell state and every 2 d thereafter, with special care to prevent dissociation of the embryoid bodies because of shear and care to avoid exposing the plated neuronal progenitors to air. Cells were placed at 37°C with 5% CO2 under all medium conditions. Stock solutions of fatty acids to be supplemented were prepared in ethanol (40 mM 18:0, 32.8 mM 20:4, and 60 mM 22:6) and pipetted dropwise into N2 media containing 3 mM fatty acid–free BSA (Cat. #A8806; Merck) at 37°C with constant stirring to obtain final concentrations of 3 mM 18:0, 1 mM 20:4, and 2 mM 22:6. This 100x stock of the FA-supplemented N2 media was aliquoted and stored at −20°C. On day 8 of the differentiation protocol, it was added to the N2 culture medium in a 1:100 ratio. mESCs were maintained as undifferentiated stem cells in an adherent monolayer in N2B27 media (a 1:1 mixture of DMEM/F12 [without Hepes and with glutamine, Cat. #131331028; Gibco] and Neurobasal medium [Cat. #21103049; Gibco] supplemented with 0.5x N2 [17502001; Gibco], 0.5x B27 [without vitamin A], 0.25 mM L-glutamine [Cat. #25030149; Gibco], 0.1 mM β-mercaptoethanol [Cat. #21985023; Gibco], 10 mg/ml BSA fraction V [Cat. # 10735078001; Merck], 10 mg/ml Human recombinant Insulin [Cat. #91077C; Merck], and 1% Pen/Strep) +2iLIF (10 ng/ml leukemia inhibitory factor + 3 μM CHIR99021 + 1 μM PD0325901 from Tocris, Cat. # 4423 and 4192, respectively). The cells were grown on 0.1% gelatin-coated dishes and usually seeded at a density of 0.5–1.5 × 10/cm. The medium was replaced every day, and the cells were trypsinized and reseeded every 2 d in their undifferentiated state. For monolayer differentiation, the cells were similarly seeded and kept in N2B27 without the 2iLIF for 24 h, after which the media were supplemented with 1 μM retinoic acid and refreshed every 24 h. The human iPSC line XCL-1 (XCell Science Inc.) was differentiated into NSCs by XCell Science Inc. using a 14-d protocol where iPSCs were initially differentiated in suspension as embryoid bodies followed by the formation of attached neural rosettes. These NSCs were then isolated and expanded (Swistowski et al, 2009). The NSCs were propagated using Geltrex (Cat. #A1413202; Thermo Fisher Scientific)-coated plates in Neurobasal medium (Cat. #21103049; Thermo Fisher Scientific) supplemented with NEAA (Cat. #11140050; Thermo Fisher Scientific), GlutaMAX-I (Cat. #35050038; Thermo Fisher Scientific), B27 supplement (Cat. #17504044; Thermo Fisher Scientific), penicillin–streptomycin (Cat. #15140; Thermo Fisher Scientific), and bFGF (Cat. #233-FB; R&D Systems). Cells were enzymatically passaged with Accutase (Cat. #A1110501; Thermo Fisher Scientific) when they were 80–90% confluent. Neuronal differentiation was achieved by culturing NSCs in DOPA Induction and Maturation Medium (Cat. #D1-011; XCell Science) according to the manufacturer’s instructions by supplementing with 200 ng/ml human recombinant Sonic hedgehog (Cat. #100-45; PeproTech) from days 0–10 and passaging cells at days 0, 5, and 10 onto poly-L-ornithine (Cat. #3655; Sigma-Aldrich)– and laminin (Cat. #23017015; Thermo Fisher Scientific)-coated plates at a density of 50,000 cells/cm. mESCs differentiated into neurons for 12 d were fixed in 4% (vol/vol) paraformaldehyde (Cat. #28908; Thermo Fisher Scientific) for 20 min. Excess paraformaldehyde was quenched with 30 mM glycine for 5 min, and coverslips were washed three times with PBS. Cells were permeabilized with 0.1% Triton X-100 (Cat. # 3051.4; Carl Roth) and blocked with 0.5% BSA (Cat. # 8076.4; Carl Roth) for 30 min at RT. Cells were incubated with a primary antibody against β-tubulin III (Cat. # ab78078; Abcam) diluted in 0.5% BSA 1:200 and incubated for 1 h at RT. A secondary antibody conjugated to an Alexa Fluor 594 dye (Cat. # A-11005; Invitrogen) was used for detection. Cells were stained with DAPI 5 μg/ml (Cat. # D9542; Sigma-Aldrich) for 5 min, and coverslips were mounted onto glass slides with ProLong Gold (Cat. # P36934; Invitrogen). Images were acquired with a Nikon Eclipse Ti fluorescence microscope using the 20x objective. Human iPSC-derived neurons, on differentiation day 25 from NSCs, were fixed for 15 min at RT in 4% (wt/vol) paraformaldehyde (Cat. #158127; Sigma-Aldrich) and permeabilized with 0.1% Triton X-100 (Cat. #9002-93-1; Sigma-Aldrich). Unspecific binding was blocked with 10% goat serum (Cat. #S26; Millipore), and cultures were incubated overnight at 4°C with primary antibodies diluted in TBS/10% goat serum: mouse anti-β-tubulin-III (#T8660; Sigma-Aldrich) or mouse anti-MAP2 (Cat. # M1406; Sigma-Aldrich). Incubation with secondary antibodies Alexa Fluor 555 goat anti-mouse IgG (Cat. #A21422; Molecular Probes) was done at 1:500 in TBS/10% goat serum for 2 h at RT. Cell nuclei were counterstained with 10 μM 4″, DAPI dihydrochloride (Cat. #D9542; Sigma-Aldrich). Coverslips were mounted onto glass slides with ProLong Diamond Mounting Medium (Cat. #P3690; Molecular Probes). To estimate the proportion of differentiated neurons, the ratio of neuronal marker–positive cells to the total number of cells (DAPI) in each field of view was determined and averaged over several images. Brain tissue samples for lipidomics were collected from WT CD-1 mice from the EMBL breeding colonies at four different time points in development, namely, the embryonic stages E10.5 and E15.5 and the postnatal stages P2 and P21 (Fig 1A). No procedure was performed on live animals. The mothers carrying E10.5 and E15.5 embryos, and P2 and P21 pups were euthanized by IACUC-approved methods, after which tissues were collected for lipidomics. At E10.5, each brain sample consisted of a pooled brain region from five embryos to obtain enough samples for lipidomics. At all remaining time points, the brain from a single mouse embryo or pup was dissected to obtain two samples: one with cerebral hemispheres and another with the rest of the brain. All tissues were washed with PBS, flash-frozen in liquid nitrogen, and stored at −80°C until lipid extraction. Mouse cells were collected for lipidomics on days 0, 4, 8, and 12 of differentiation in embryoid bodies of mESCs. On day 0, cells were collected after trypsinization. On days 4, 8, and 12, suspended embryoid bodies from one 6-cm dish each were collected by a cell scraper. In the case of monolayer differentiation of mESCs, cells in one well of a six-well plate were collected per sample. In each case, the cells were pelleted, washed twice in Hepes–KOH buffer, suspended in 200–300 μl 155 mM ammonium formate, flash-frozen in liquid nitrogen, and stored at −80°C until lipid extraction. Human iPSC-derived neurons were collected after treatment with Accutase on differentiation days 0, 10, and 25 (from the NSC stage). Cells were collected in ice-cold PBS, pelleted, washed with 155 mM ammonium acetate (Cat. #A7330; Sigma-Aldrich), flash-frozen in liquid nitrogen, and stored at −80°C until lipid extraction. Lipids were extracted from the samples as described previously (Almeida et al, 2015). In short, all samples were thawed at 4°C and homogenized by sonication (and mechanically using an ultra-turrax in addition to this, in the case of brain samples). Samples were spiked with an internal standard (IS) mix prepared in-house and containing known amounts of synthetic lipid standard (Table S1). Lipid extraction was performed by two-step extraction (Sampaio et al, 2011), first by partitioning the sample between 155 mM aqueous ammonium formate and chloroform/methanol (10:1, vol/vol) and then using the aqueous fraction to partition against chloroform/methanol (2:1, vol/vol). The solutions were shaken using a ThermoMixer (Eppendorf) set at 1,400 rpm and 4°C during both steps (for 2 and 1.5 h, respectively) and their organic phases collected. Solvents were then removed by vacuum evaporation, leaving deposits of lipids extracted in the 10:1- and 2:1-extracts, respectively. The 10:1- and 2:1-extracts were dissolved in chloroform/methanol (1:2, vol/vol). MS lipidomic analysis was performed as previously described (Almeida et al, 2015). In short, mass spectra of the lipid extracts were recorded in both positive and negative ion modes using an Orbitrap Fusion Tribrid mass spectrometer (Thermo Fisher Scientific) equipped with a robotic nanoflow ion source, TriVersa NanoMate (Advion Biosciences). Aliquots of 10:1-extracts were diluted with 2-propanol to yield an infusate composed of chloroform/methanol/2-propanol (1:2:4, vol/vol/vol) and 7.5 mM ammonium formate for positive ion mode analysis. The 10:1-extracts were also diluted with 2-propanol to yield an infusate composed of chloroform/methanol/2-propanol (1:2:4, vol/vol/vol) and 0.75 mM ammonium formate for negative ion mode analysis. Aliquots of the 2:1-extracts were diluted with methanol to yield an infusate composed of chloroform/methanol (1:5, vol/vol) and 0.005% methylamine for analysis in the negative ion mode. High-resolution mass spectra of intact ions (MS) were recorded using an Orbitrap mass analyzer, precursor ions from each 1-Da window within the detected range were fragmented, and the mass spectra of the fragments (MS) were recorded using the Orbitrap mass analyzer. Lipid molecules and fragment ions were identified using ALEX (Pauling et al, 2017; Ellis et al, 2018). The molar abundance of lipid molecules was quantified by normalizing their MS intensities to that of spiked-in internal lipid standards and by scaling the known concentration of the internal standard (Table S1). Lipid quantification and downstream analysis of fatty acid composition and LENA were carried out on the SAS9.4 platform (SAS) as described previously (Sprenger et al, 2021). In the analysis of time-course data, replicates from identical conditions were considered to constitute a sample group (resulting in nine sample groups among the mouse brain samples and four sample groups among cell culture samples). ANOVA testing with these sample groups was performed to identify significant changes in the mol% abundance of lipids. An unpaired t test was used to test for significant differences in the lipidome of day 12 neurons between control and fatty acid–supplemented conditions. To minimize false positives because of multiple hypothesis testing, we calculated a Benjamini–Hochberg critical value q for every P-value with a false-positive rate of 0.05, and changes were considered significant for all P ≤ pc, where pc is the largest P-value that is smaller than its corresponding q-value. The statistical tests were performed using R, and the principal component analysis was done using ClustVis (Metsalu and Vilo, 2015). Tableau Desktop (Tableau Software) was used for data visualization. RNA-seq data from Gehre et al (2020) and Sladitschek and Neveu (2019) were accessed on the ArrayExpress repositories with accession numbers E-MTAB-6821 and E-MTAB-4904, respectively. The lipidomic dataset generated in this work can be found as Supplemental Data 1, and analysis codes are available on request.
|
PMC8755711
|
A multi-omics study of circulating phospholipid markers of blood pressure
|
High-throughput techniques allow us to measure a wide-range of phospholipids which can provide insight into the mechanisms of hypertension. We aimed to conduct an in-depth multi-omics study of various phospholipids with systolic blood pressure (SBP) and diastolic blood pressure (DBP). The associations of blood pressure and 151 plasma phospholipids measured by electrospray ionization tandem mass spectrometry were performed by linear regression in five European cohorts (n = 2786 in discovery and n = 1185 in replication). We further explored the blood pressure-related phospholipids in Erasmus Rucphen Family (ERF) study by associating them with multiple cardiometabolic traits (linear regression) and predicting incident hypertension (Cox regression). Mendelian Randomization (MR) and phenome-wide association study (Phewas) were also explored to further investigate these association results. We identified six phosphatidylethanolamines (PE 38:3, PE 38:4, PE 38:6, PE 40:4, PE 40:5 and PE 40:6) and two phosphatidylcholines (PC 32:1 and PC 40:5) which together predicted incident hypertension with an area under the ROC curve (AUC) of 0.61. The identified eight phospholipids are strongly associated with triglycerides, obesity related traits (e.g. waist, waist-hip ratio, total fat percentage, body mass index, lipid-lowering medication, and leptin), diabetes related traits (e.g. glucose, insulin resistance and insulin) and prevalent type 2 diabetes. The genetic determinants of these phospholipids also associated with many lipoproteins, heart rate, pulse rate and blood cell counts. No significant association was identified by bi-directional MR approach. We identified eight blood pressure-related circulating phospholipids that have a predictive value for incident hypertension. Our cross-omics analyses show that phospholipid metabolites in the circulation may yield insight into blood pressure regulation and raise a number of testable hypothesis for future research.Long-term high blood pressure, of which 90–95% essential hypertension, is a major risk factor for cardiovascular diseases, e.g. coronary artery disease, stroke, heart failure, atrial fibrillation, etc. Pervious study showed that the patients with essential hypertension have abnormal sodium-lithium counter transport across the red cell membrane, and that the level of transport is heritable. Phosphatidylcholines (PC), phosphatidylethanolamines (PE), lysophosphatidylcholines (LPC), PE-based plasmalogens (PLPE), ceramides (CERs) and sphingomyelin (SPM) are groups of phospholipids that have a key function in the bilayer of (blood) cell membranes. Although changes of membrane phospholipids in essential hypertension have been recognized and studied for a long time, these previous studies either focused on animal models or overall phospholipid groups with limited resolution in the measurement. More detailed characterization of phospholipids in relation to hypertension at the population level is lacking. In recent decades, high-throughput mass spectrometry (MS) has offered the opportunity to determine phospholipids on the chemical molecular level with high resolution at a low price. Thus, phospholipid panels with detailed characterisation are increasingly adopted in large epidemiological studies. Despite these developments, the number of studies on the role of phospholipid profiles in hypertension and blood pressure is still limited. Very few studies of blood pressure and hypertension have investigated phospholipid profile in high resolution. The study by Kulkarni et al. examined 319 (phospho)lipids in 1192 Mexican-Americans and found that diacylglycerols (DG) in general and DG 16:0/22:5 and DG 16:0/22:6 in particular are significantly associated with systolic (SBP), diastolic (DBP) and mean arterial pressures as well as the risk of incident hypertension. Stefan et al. studied 135 cases and 981 non-cases of incident hypertension in a European study and identified four phospholipids and two amino acids which could improve the predictive performance of hypertension in addition to the known risk markers. To our knowledge, no large-scale epidemiological study of blood pressure and/or hypertension with high-throughput measured phospholipids has been performed with replication in an independent study or has studied in detail the mechanism of the associations. The aim of this study was to conduct an in-depth multi-omics study of the associations and causality of the associations of phospholipids with SBP and DBP, which are the diagnostic variables of hypertension, through metabolomics, genomics and phenomics. To this end, we investigated the association of blood pressure and 151 quantified phospholipids including 24 SPMs, 9 CERs, 57 PCs, 15 LPCs, 27 PEs, and 19 PLPEs, in 3971 individuals from five European populations. Using Mendelian Randomization (MR), we further investigated the causality in these relationships. The potential genetic pleiotropy between phospholipids and blood pressure was also explored. This study was conducted using five populations throughout Europe. The individuals with both blood pressure and phospholipid measure available were included: (1) the CROATIA-Vis study conducted on the island of Vis, Croatia (n = 710), (2) the Erasmus Rucphen Family (ERF) study, conducted in the Netherlands (n = 717), (3) the Northern Swedish Population Health Survey (NSPHS) in Norrbotten, Sweden (n = 678), (4) the Orkney Complex Disease Study (ORCADES) in Scotland (n = 681), and finally (5) the MICROS study from the South Tyrol region in Italy (n = 1185) which was included for replication. Fasting blood samples were collected for the biochemical measurements. All studies were approved by the local ethics committees and all participants gave their informed consent in writing. The association tests of phospholipids and blood pressure were performed on the same baseline data, for each of the five studies. The predictive analysis was performed in ERF study in which we collected follow-up data from March 2015 to May 2016 (9–14 years after baseline visit). During the follow-up, a total of 572 participants’ records from the 717 individuals included in the baseline analysis were scanned for common diseases in general practitioner’s databases. As part of the European Special Populations Research Network (EUROSPAN) project, the absolute concentrations (µM) of 151 lipid traits in plasma were centrally measured by electrospray ionization tandem mass spectrometry (ESIMS/MS), including 24 SPMs, 9 CERs, 57 PCs, 15 LPCs, 27 PEs and 19 PLPEs. The methods used have been validated and described previously and a brief description can be found in the Supplementary Material. For each lipid molecule, we adopted the naming system where lipid side chain composition is abbreviated as x:y, where x denotes the number of carbons in the side chain and y the number of double bonds. For example, PC 34:4 denotes an acyl-acyl PC with 34 carbons in the two fatty acid side chains containing four double bonds. Supplementary Table 1 describes how SBP and DBP, T2D status, total cholesterol (TC), high-density-lipoprotein cholesterol (HDL-C), lipid-lowering medication usage, body mass index (BMI) and antihypertensive medication are measured or defined in the cohorts. In all cohorts, blood pressure was measured by automated reading in the sitting position after a rest. The medication information was collected during the personal interview. For the additional analysis in ERF only (described below), we imputed the missing values by multiple imputation in R package ‘mice’ and followed the Rubin’s rules. Lipids were natural log-transformed and standardized (mean-centered and divided by their standard deviation). We corrected blood pressure levels for antihypertensive medication use by adding 15 mmHg to the SBP and 10 mmHg to the DBP of users of antihypertensive medication. As all of the five cohorts included closely related individuals, family relationship based on the genotype was adjusted for in the analysis by extracting polygenic residuals for the phenotypic traits, by using the polygenic option in GenABEL package in R. In each study, we used linear regression to examine the association between each of the phospholipids and blood pressure individually. Blood pressure variables were used as dependent variables and phospholipids were used as independent variables. We performed a discovery analysis in CROATIA-Vis, ERF, NSPHS, and ORCADES, adjusting for age and sex (model 1). Results from the four discovery populations were meta-analyzed with inverse-variance weighted fixed-effects model using the METAL software. To correct for multiple testing, we used Bonferroni correction using the number of 70 independent components extracted from the 151 directly measured phospholipids (P-value < 7.1 × 10, 0.05/70). Matrix Spectral Decomposition was separately used to calculate the number of independent equivalents in each of the four discovery studies. Bonferroni correction was done for 70 tests which was the highest number obtained in CROATIA-Vis. We did not correct for the number of blood pressure variables as SBP and DBP are highly correlated (R = 0.65, P-value < 2.2 × 10 in ERF, n = 2802). We replicated our findings in MICROS (n = 1185) using the same statistical framework as in the discovery analysis and using a Bonferroni correction for the independent number of tested associations, i.e., equivalents of the significant phospholipids. In a combination of all five cohorts, we examined a further model (model 2) to assess the impact of potential confounders and mediators by additionally adjusting for BMI, HDL-C, TC, lipid-lowering medication and type 2 diabetes (T2D) status. We checked the pairwise Pearson's correlation matrices of the blood pressure related phospholipids in ERF adjusting for age, sex and family relationships. The phospholipids significantly associated with SBP or DBP were tested for the association with the occurrence of hypertension during the follow-up in ERF. The incident cases were defined as the participants free of hypertension at baseline who were diagnosed with hypertension at follow-up by general practitioners. Time-to-event was defined as the time from the enrollment date at baseline to either the onset date of disease, date of death, date of censoring (moving away) or date of follow-up collection. Cox proportional regression analysis was used to evaluate the individual effect of phospholipids considering of the follow-up time (time-to-event). To determine the joint effect of the phospholipids on the discrimination of future hypertension patients, we calculated the area under the receiver operator characteristics (ROC) curve (AUC). We further determined whether the addition of the listed phospholipids increase the AUC value of the factors in the Framingham risk score for hypertension which includes age, sex, SBP, DBP, BMI, cigarette smoking and parental hypertension (Framingham model). Integrated Discrimination Improvement test (IDI) and continuous Net Reclassification Improvement test (NRI) were performed to compare different joint models. In ERF, we further examined the association of these identified phospholipids with known cardiometabolic traits, including adiponectin, albumin, alcohol consumption, anti-diabetic and anti-hypertensive medications, BMI, creatinine, C-reactive protein, glucose, HDL-C, insulin, intima-media thickness, low-density-lipoprotein cholesterol (LDL-C), heart rate, homeostatic model assessment-insulin resistance (HOMA-IR), leptin, lipid-lowering medication, metabolic syndrome, plaque score, pulse wave velocity, resistin, smoking status, TC, total fat percentage, triglycerides, T2D, waist circumference and waist-to-hip ratio. The description and measurement methods of the above mentioned cardiometabolic traits can be found in our previous reports. The distributions of adiponectin, insulin, leptin, triglycerides, C-reactive protein, HOMA-IR and resistin are skewed and therefore were log-transformed before performing the analysis. We used the standardized residuals of natural-log-transformed phospholipid levels as the dependent variable, adjusted for age, sex and family relationship. A hierarchical clustering approach was used to cluster the cardiometabolic traits. We estimated the false discovery rate less than 0.05 by Benjamini & Hochberg method considering of the gathering of categorical and continuous variables. MR is a statistical method which uses the effect of genetic variants determining an exposure and test its association with the outcome under study, based on the assumption that the genetic variant is inherited independent of the confounding variables. We performed a two-sample bi-directional MR of the 11 significant associations of phospholipids and SBP or DBP. We used summary statistics level data of blood pressure and phospholipids utilizing the pipeline in the R-package TwoSampleMR. In brief, the genetic instrument was based on the top genetic determinant SNPs with linkage disequilibrium R < 0.05 within 500kbps clumping distance. The proportion of variance in the exposure explained by the genetic variance (R) and F statistics were calculated to estimate the statistic power of MR. As the sample size in the phospholipid GWAS in the ESIMS/MS platform is small (n = 4034), to increase the explained variance of the instrumental variable, P-value < 1.0 × 10 was used to define the genetic determinants of phospholipids. The GWAS summary statistics of the same phospholipids available in Biocrates metabolomics platform were also used additionally to increase the statistical power (n = 7478). For genetic determinants of blood pressure, the genome-wide significance level (P-value < 5 × 10) was used. Inverse-variance weighted MR was used with weighted median, sample mode and weighted mode methods as sensitivity to investigate pleiotropy. MR-Egger regression was used to control the directional horizontal pleiotropy, and the Egger estimates on the intercept was used for the heterogeneity tests. The Bonferroni corrected P-value with independent equivalents of phospholipids was used as the significance level. We further studied the pleiotropic effect of the genetic determinants of the identified phospholipids using phenome-wide association study (Phewas) by data-mining from previous publications. For the top SNPs of either phospholipids used in MR, we looked up their pheno-wide associations in GWAS ATLAS. We estimated the false discovery rate less than 0.05 by Benjamini & Hochberg method. CROATIA-VIS: All subjects were asked to provide written consent, after being informed on the study goals and main approaches, in accordance with the Declaration of Helsinki. The study was approved by the ethics committees of the University of Zagreb (No. 018057) and the University of Split School of Medicine (No. 2181-198-A3-04110-11-0008), Croatia and the Multi-Centre Research Ethics Committee for Scotland (No. 01/0/71). ERF: The study protocol was approved by the medical ethics board of the Erasmus Medical Center Rotterdam, the Netherlands. All participants gave their informed consent in writing. MICROS: A detailed information sheet and a form for the written informed consent were provided to each prospective participant to approve. The study was approved by the Ethics Committee of the Autonomous Province of Bolzano. NSPHS: The NSPHS study was approved by the local ethics committee at the University of Uppsala (Regionala Etikprövningsnämnden, Uppsala, Dnr 2005:325) in compliance with the Declaration of Helsinki. All participants gave their written informed consent to the study. For participants of under legal age, a legal guardian also signed. The procedure used to obtain informed consent and the respective informed consent form has been recently discussed according to current ethical guidelines. ORCADES: ORCADES received ethical approval from the appropriate research ethics committees in 2004. Data collection was carried out in Orkney between 2005 and 2007. Informed consent and blood samples were provided by 1019 Orcadian volunteers who had at least one grandpa rent from the North Isles of Orkney. Figure 1 shows the flow chart of the study design and the main results.Figure 1Flow chart of the current study design and its main results. SBP: systolic blood pressure; DBP: diastolic blood pressure; PC: phosphatidylcholines; PE: phosphatidylethanolamines; MR: Mendelian randomization; AUC: area under the ROC curve; Phewas: phenome-wide association study; * Bonferroni correction for the independent number of tests. Flow chart of the current study design and its main results. SBP: systolic blood pressure; DBP: diastolic blood pressure; PC: phosphatidylcholines; PE: phosphatidylethanolamines; MR: Mendelian randomization; AUC: area under the ROC curve; Phewas: phenome-wide association study; * Bonferroni correction for the independent number of tests. Baseline characteristics of the five participating cohorts are shown in Table 1.Table 1Baseline characteristics of the study population in the association analysis.DiscoveryReplicationCROATIA-VisERFNSPHSORCADESMICROSN7107176786811,185Age (y)56.6 (15.6)51.9 (14.2)47.1 (20.8)57.0 (13.9)45.7 (16.4)Sex (% women)57.4059.0053.257.956.0Body mass index (kg/m)27.4 (4.3)27.0 (4.4)26.4 (4.8)27.7 (4.8)25.6 (4.8)HDL-C (mmol/L)1.10 (0.16)1.29 (0.36)1.60 (0.41)1.57 (0.42)1.68 (0.38)TC (mmol/L)5.12 (0.99)5.67 (1.08)5.86 (1.33)5.56 (1.14)5.87 (1.19)Lipid-lowering medication use (%)3.0016.6010.474.105.47Systolic blood pressure (mmHg)137.8 (24.45)141.7 (22.0)122.8 (18.6)130.1 (18.46)132.6 (20.5)Diastolic blood pressure (mmHg)80.5 (11.47)80.6 (10.24)74.1 (7.8)76.0 (9.6)79.6 (11.3)Antihypertensive medication use (%)24.125.220.539.114.6Type 2 diabetes (%)4.26.04.12.83.1Values are mean (SD) or percentages. N refers to the largest sample size used in this study. Baseline characteristics of the study population in the association analysis. Values are mean (SD) or percentages. N refers to the largest sample size used in this study. The mean age ranged from 47.1 (with standard deviation 20.8) years old in NSPHS to 56.6 (with standard deviation 15.6) years old in CROATIA-Vis. and the proportion of females ranged from 53.2% in NSPHS to 57.0% in ORCADES. The means and standard deviations of the concentration of the 151 phospholipids across the five cohorts are shown in Supplementary Table 2 and Supplementary Figure 1. Most of the phospholipids have similar concentrations across cohorts, except for PLPE 18:1, PLPE 18:0 and PLPE 16:0. The associations of all 151 phospholipids with blood pressure in the discovery panel, replication panel and combination are shown in in Supplementary Table 3. Volcano plots in Fig. 2A and B show the meta-analysis results of the discovery panel in a J shape. Five phospholipids (PC 32:1, PC 40:5, PE 38:4, PE 40:5 and PE 40:6) were significantly associated with SBP, and seven phospholipids (PE 38:3, PE 38:4, PE 38:6, PE 40:4, PE 40:5, PE 40:6 and LPC 22:4) were associated with DBP using Bonferroni corrected significance threshold. For the significant phospholipids found in the discovery analysis, only LPC 22:4 was associated inversely to DBP. Eleven of the 12 significant associations from the discovery were replicated in MICROS based on the following adjusted P-value thresholds: 0.017 for SBP and 0.013 for DBP (Fig. 2). Only LPC 22:4 did not replicate in MICROS. Further, we focused on the 11 significant associations with eight unique phospholipids which were replicated.Figure 2Association of phospholipids and blood pressure in model 1 in the discovery meta-analysis. (A): phospholipids associated with SBP. (B) phospholipids associated with DBP. Age, sex and family relationship were adjusted for in the regression analysis. Red: Lipids are significantly associated with blood pressure and replicated. Blue: lipid LPC 22:4 significantly associated with blood pressure but failed in the replication. Association of phospholipids and blood pressure in model 1 in the discovery meta-analysis. (A): phospholipids associated with SBP. (B) phospholipids associated with DBP. Age, sex and family relationship were adjusted for in the regression analysis. Red: Lipids are significantly associated with blood pressure and replicated. Blue: lipid LPC 22:4 significantly associated with blood pressure but failed in the replication. The significant associations were generally attenuated upon adjustment for BMI, HDL-C, TC, lipid-lowering medication and T2D status in model 2, with the proportion of the effect estimate decreased ranged from 4.5% for the association between PC 32:1 and SBP to 46.2% for the association between PE 40:6 and DBP, but all of the associations remained significant (Table 2). All the identified phospholipids were highly correlated with each other, while the correlation among the PEs was obviously higher than with the PCs or between the PCs (Supplementary Figure 2).Table 2Effect of adjustments on the association between selected lipids and blood pressure.NameModel 1Model 2Discovery(N = 2786)Replication(N = 1185)Combined(N = 3971)Combined(N = 3937)EffectP-valueEffectP-valueEffectP-valueEffectP-valueSystolic blood pressurePC 32:12.71.5 × 101.82.4 × 102.22.6 × 102.19.6 × 10PC 40:52.06.7 × 102.01.0 × 102.23.2 × 101.62.7 × 10PE 38:32.17.0 × 102.22.6 × 102.16.8 × 101.54.0 × 10PE 40:52.14.9 × 102.31.7 × 102.23.2 × 101.62.7 × 10PE 40:62.42.2 × 101.76.7 × 102.08.6 × 101.21.9 × 10Diastolic blood pressurePE 38:41.59.7 × 101.27.6 × 101.33.7 × 100.91.3 × 10PE 40:51.31.4 × 101.46.7 × 101.43.5 × 101.03.4 × 10PE 40:61.44.4 × 101.33.5 × 101.35.7 × 100.73.4 × 10PE 38:31.13.0 × 101.31.8 × 101.27.0 × 100.84.0 × 10PE 38:61.16.5 × 100.82.5 × 100.91.2 × 100.62.7 × 10PE 40:41.13.9 × 101.93.1 × 101.63.8 × 101.27.2 × 10Table shows the identified lipids through discovery and replication (Fig. 2). Model 1 was performed in discovery, replication and combined data with age and sex as covariates; Model 2 was performed in the combined data with additional adjustment for BMI, HDL-C, TC, lipid-lowering medication and type 2 diabetes status based on model 1. Effect of adjustments on the association between selected lipids and blood pressure. Table shows the identified lipids through discovery and replication (Fig. 2). Model 1 was performed in discovery, replication and combined data with age and sex as covariates; Model 2 was performed in the combined data with additional adjustment for BMI, HDL-C, TC, lipid-lowering medication and type 2 diabetes status based on model 1. We studied the relationship between the identified phospholipids and incident hypertension in 447 available participants from the ERF study, including 92 patients with incident hypertension. None of the eight identified phospholipids (six PEs and two PCs) were individually significantly associated with incident hypertension in our study (Supplementary Table 4). But the joint effect of the eight phospholipids was significantly associated with incident hypertension (P-value = 5.0 × 10, Fig. 3). Although in the phospholipids-only model, the discrimination between those with and without future hypertension is limited (AUC = 0.61) and significantly lower than that of the Framingham model, adding the eight phospholipids significantly improved the AUC on top of the Framingham model from 0.75 to 0.76 (PIDI = 0.02, PNRI = 0.06, Fig. 3).Figure 3AUC comparison between eight phospholipids associated with either SBP or DBP, the factors included in the Framingham risk score and their combination. Model 1 Novel: The model includes phospholipids associated with either SBP or DBP only: PC 32:1, PC 40:5, PE 38:3, PE 40:6, PE 40:5, PE 38:4, PE 38:6, PE 40:4. Model 2 FHS: The model includes the factors from the Framingham risk scores of incident hypertension: age, sex, SBP, DBP, BMI, cigarette smoking and parental hypertension. Model 3 FHS + Novel: the advanced model adding factors in model 1 and Model 2. * PIDI < 0.05. IDI: Integrated Discrimination Improvement test. AUC comparison between eight phospholipids associated with either SBP or DBP, the factors included in the Framingham risk score and their combination. Model 1 Novel: The model includes phospholipids associated with either SBP or DBP only: PC 32:1, PC 40:5, PE 38:3, PE 40:6, PE 40:5, PE 38:4, PE 38:6, PE 40:4. Model 2 FHS: The model includes the factors from the Framingham risk scores of incident hypertension: age, sex, SBP, DBP, BMI, cigarette smoking and parental hypertension. Model 3 FHS + Novel: the advanced model adding factors in model 1 and Model 2. * PIDI < 0.05. IDI: Integrated Discrimination Improvement test. Figure 4 shows the association of the blood pressure-related phospholipids with the classical/clinical cardiometabolic traits measured in ERF (N = 818 analytical sample size). Triglycerides were strongly associated with all of the eight blood pressure-related phospholipids and form the first cluster themselves. Although the direction and strength of association are very similar to triglycerides, the association of triglycerides appears to be independent of the second cluster. The second cluster involved waist, waist-hip ratio, glucose, total fat percentage, BMI, leptin, use of anti-hypertensives, T2D, lipid-lowering medication, C-reactive protein, HOMA-IR, insulin and TC. Most of the significant associations were in the same (positive) direction of the association between blood pressure and related phospholipids. The third cluster had much fewer significant associations. We found associations between PCs and environmental exposures such as smoking and alcohol intake, but also heart rate, albumin, HDL and LDL-C and adiponectin. No significant association was found between the phospholipids and vascular-related variables, including pulse wave velocity, intima-media thickness and plaque score.Figure 4Association of blood pressure related phospholipids and cardiometabolic traits in ERF. Hierarchical clustering approach was used for the clustering. Red: positive association; blue: negative association. The depth of the color displays the strength of z score in the regression. * FDR < 0.05. · P-value < 0.05. Association of blood pressure related phospholipids and cardiometabolic traits in ERF. Hierarchical clustering approach was used for the clustering. Red: positive association; blue: negative association. The depth of the color displays the strength of z score in the regression. * FDR < 0.05. · P-value < 0.05. The MR framework resulted in two to six independent SNPs included in the genetic risk score as instrumental variables for each phospholipid (R range from 2.7 to 5.2%), 471 SNPs for SBP (R = 4.0%) and 506 SNPs for DBP (4.3%). The F-statistics ranged from 55.1 for the MR in PE 40:4 to DBP to 105.7 in PE 40:5 to SBP and DBP. The two PCs, i.e. PC 32:1 and PC 40:5 were also performed using the summary statistics of Biocrates platform. However, no significant results were found in either MR test (Supplementary Table 5). The top SNPs which were associated at genome-wide significance with blood pressure related phospholipids are rs174576, rs10468017, rs261338, rs12439649, rs740006 and rs7337573 and located in the protein-coding genes TMEM258, FADS2, ALDH1A2, LIPC, and antisense gene RP11-355N15.1, after considering for the linkage disequilibrium. In total, 1513 SNP-trait associations were identified from the Phewas database after controlling for false discovery rate (Supplementary Table 6). The most highly significant related traits were in metabolic domain which are mainly lipoproteins, and blood cell counts. The next highly related traits are heart rate and pulse rate in the cardiovascular domain. Other highly significant related traits including male pattern baldness, height, glucose, etc. (Fig. 5; Supplementary Table 6).Figure 5Results of the phenome-wide association study of the six genetic determinants of the blood pressure related phospholipids. In each domain (x-axis), the top trait was annotated in the figure. The traits with P-value level less than 1.0 × 10 were annotated as 1.0 × 10 in the figure. The dots depict the P-values from either rs174576, rs10468017, rs261338, rs12439649, rs740006 and rs7337573 which are located in the protein-coding genes TMEM258, FADS2, ALDH1A2, LIPC, and antisense gene RP11-355N15.1. Results of the phenome-wide association study of the six genetic determinants of the blood pressure related phospholipids. In each domain (x-axis), the top trait was annotated in the figure. The traits with P-value level less than 1.0 × 10 were annotated as 1.0 × 10 in the figure. The dots depict the P-values from either rs174576, rs10468017, rs261338, rs12439649, rs740006 and rs7337573 which are located in the protein-coding genes TMEM258, FADS2, ALDH1A2, LIPC, and antisense gene RP11-355N15.1. The current study identified and replicated the association of eight phospholipids with either SBP or DBP. These phospholipids jointly associated with incident hypertension and improved the discrimination model of incident hypertension. Strong associations were identified of these phospholipids with triglycerides, but also with obesity related traits (e.g., waist, waist-hip ratio, total fat percentage, BMI and leptin), T2D and related traits (e.g. glucose, HOMA-IR and insulin). Meanwhile, the genetic determinants of these phospholipids also highly and genome-widely associated with lipoproteins, blood cell counts, heart rate, pulse rate, glucose and many potential pleiotropic traits, e.g. male pattern baldness, height, etc. No significant association was identified between the genetic susceptibility of blood pressure and phospholipids by MR approach, in either direction. We found a predictive effect of the joint phospholipids on future hypertension, which was consistent with the associations of these phospholipids and incident hypertension in the Mexican–American population; among the 11 replicated associations in the current study, ten associations were replicated in the Mexican–American population using the current Bonferroni P-value adjustment (0.017 for SBP and 0.013 for DBP, Supplementary Table 7). Moreover, PC 40:5, PE 38:3, PE 38:4, PE 38:6, PE 40:5 and PE 40:6 were also associated with incident hypertension in their study. The similarity of the findings in populations of different ancestries suggests a probability of a generalizable biological process. This is in line with our finding that the significant associations of the identified phospholipids and blood pressure remain after adjustment for various potential confounders or mediators, suggesting that the associations are independent of these covariates. The association of the phospholipids with incident hypertension and the improved predictive performance with adding these phospholipids onto the Framingham model also implies that the phospholipid level may be a predictor of the occurrence of diagnosed hypertension. However, we could not confirm any causation by current MR approach. Further research is required to investigate this hypothesis. The identified phospholipids are strongly associated with triglycerides, which share 1,2- diglyceride with PCs and PEs as a substrate in their biosynthesis. This is in line with the previous finding that blood pressure is related to DG in general and DG 16:0/22:5 and DG 16:0/22:6. We also found that DBP is related to PE 38:6 which also includes a fraction of PE 16:0/22:6. These blood pressure related phospholipids are also associated with obesity and diabetes related traits (Fig. 4). This is consistent with our results that after adjustment for BMI, HDL-C, TC, lipid-lowering medication and T2D status in model 2, the effect estimate generally attenuated. It implies that these factors may act as mediators in the associations. However, as the associations were still statistically significant, we could not completely exclude the direct association of the abnormal phospholipid levels and blood pressure. A one sample based formal causal mediation analysis in a large-scale cohort is suggested to confirm their mediating effect on the association of these phospholipids and blood pressure. All the blood pressure-related phospholipids identified in the current study have side chains that include poly-unsaturated fatty acids. For example, phospholipid identified with PE 40:5 as for the number of carbon: double bond includes a fraction of 18:0 (sn-1)/22:5(sn-2) with docosapentaenoic acid (DPA) in the sn-2 position. The same PE 40:5 measurement with our MS method also includes a fraction of 18:0/22:4, but also 18:1/20:4 which is commonly found as arachidonic acid (omega-6) and associated with unfavorable cardiovascular outcomes. In addition, we did not see a significant association between PC32:1/PC40:5 ratio and high blood pressure (Supplementary Table 8). This is consistent with our findings that the effect estimates of PC32:1 and PC40:5 are both positively and significantly associated with blood pressure. We have also seen a high correlation between PC32:1 and PC40:5 (Supplementary Figure 2) which have a very consistent association with most of the cardiometabolic traits (Fig. 4). Besides PCs, we also found all polyunsaturated PEs highly correlated (Supplementary Figure 2) and consistently associated with blood pressure and the cardiometabolic traits (Fig. 4), thus indicating potentially distinct pathways other than involving omega classification /unsaturation. Our lipid measuring method does not differentiate between omega statuses, nor can identify the separate fatty acid chains in the phospholipids. Thus, our results are complementary to the well-established theory on the association of the omega-6 to omega-3 ratio and the increased risk of cardiovascular events. We suggest further studies on the link between the levels of the current identified PCs and PEs and the omega-6 and omega 3 fatty acids, which will provide more insights into the association of the omega-6 to omega-3 ratio and the risk of cardiovascular events. Moreover, the circulating fatty acid levels are also determined by the degradation of fatty acids, while the individual capability of degradation is partially determined by heritability. If the degradation is abnormal, this will cause high level of exogenous unsaturated fatty acids in circulation, which subsequently leads to high level of phospholipids which contain these unsaturated fatty acid chains. This is consistent with the J shape (Fig. 2) of the associations between blood pressure and phospholipids in fasting blood samples, most of which have a positive direction. A recent study reported a higher heritability in the phosphatidylcholines with a high degree of unsaturation than phosphatidylcholines with low degrees of unsaturation. Among our study, six of the eight identified phospholipids with polyunsaturated fatty acid chain (PC 40:5, PE 38:3, PE 38:4, PE 38:6 and PE 40:6) are replicated but also validated by a previous study. This provides evidence that the associations of the specific phospholipids and blood pressure are genetically driven. FADS1, FADS2 and TMEM258 are all located on chromosome 11 and band q12.2 (11q12.2) in linkage disequilibrium. Our findings that they are also genome-widely significantly associated with heart and pulse rate raise the chance that phospholipids metabolism may be implicated in the relationship with blood pressure through the pleiotropic effect of genes located in 11q12.2. An in-depth study in the (pleiotropic) role of gene FADS/TMEM258 on the association of phospholipids, blood pressure and these traits is highly suggested. The strengths of this study include the use of detailed characterized phospholipid data in a large sample size, as well as the use of replication panels. A multi-omics approach and the integration of genomic, metabolomic and epidemiologic data were performed to maximize the in-depth research of the mechanism. One of the limitations is the small number of incident hypertension cases in the current study. However, the integration of genetic data has raised an interesting hypothesis to be tested in future pathophysiological studies, in human beings and animals. To our knowledge, this is the first study performing MR on phospholipids and blood pressure. Though no significant findings were identified in the current study, the development of high-throughput technology on lipidomics will facilitate the discovery of more genetic determinants for the phospholipids and improve the strength of the instrumental variables. Previous studies reported that some anti-hypertensive drugs may have an effect on metabolism as well. In the current study, we found a significant association of anti-hypertensive drug intake and the blood pressure related phospholipids. Following the route of the previous large GWAS study of blood pressure which adjusted for anti-hypertensive drugs intake and using MR to overcome confounders in the association of blood pressure and phospholipids, we still cannot fully exclude the effect of anti-hypertensive drugs on phospholipids. Indeed, one of the genes we identified, ALDH1A2 has been implicated in coronary artery calcification and is known to interact with atenolol, a beta blocker that is prescribed to treat high blood pressure and irregular heartbeats (arrhythmia).This asks for more careful exploration of the difference between the effect of hypertension and the effect of anti-hypertensives. In conclusion, we show eight phospholipids in the circulation that significantly associate with blood pressure and show strong clustering with components of cardiometabolic disease. These phospholipids collectively associate with incident hypertension and improve the discrimination effect of previous prediction model. Our cross-omics analyses show that phospholipid metabolites in circulation may yield insight into blood pressure regulation and raise a number of testable hypothesis for future research. All methods were carried out in accordance with relevant guidelines and regulations in the method section.
|
PMC6461633
|
Changes in the Canine Plasma Lipidome after Short- and Long-Term Excess Glucocorticoid Exposure
|
Glucocorticoids (GCs) are critical regulators of metabolic control in mammals and their aberrant function has been linked to several pathologies. GCs are widely used in human and veterinary clinical practice as potent anti-inflammatory and immune suppressive agents. Dyslipidaemia is a frequently observed consequence of GC treatment, typified by increased lipolysis, lipid mobilization, liponeogenesis, and adipogenesis. Dogs with excess GC show hyperlipidaemia, hypertension, and a higher risk of developing type 2 diabetes mellitus, but the risk of developing atherosclerotic lesions is low as compared to humans. This study aimed to examine alterations in the canine plasma lipidome in a model of experimentally induced short-term and long-term GC excess. Both treatments led to significant plasma lipidome alterations, which were more pronounced after long-term excess steroid exposure. In particular, monohexosylceramides, phosphatidylinositols, ether phosphatidylcholines, acyl phosphatidylcholines, triacylglycerols and sphingosine 1-phosphates showed significant changes. The present study highlights the hitherto unknown effects of GCs on lipid metabolism, which will be important in the further elucidation of the role and function of GCs as drugs and in metabolic and cardiovascular diseases.Glucocorticoids (GCs) are highly effective anti-inflammatory and immunosuppressant drugs, commonly used to treat acute and chronic inflammatory and immune-mediated diseases. Besides their positive therapeutic effects, GCs also have many metabolic side effects, which are typically more severe after high doses and chronic applications. These adverse effects include hypertension, thin and fragile skin, wound healing disturbances, osteoporosis, muscle atrophy, weight gain, steroid diabetes, and venous thromboembolism. The clinical signs of chronic GC applications are summarised in iatrogenic Cushing’s syndrome (CS), where patients present with increased morbidity and mortality primarily due to cardiovascular, thrombotic, and metabolic complications. CS can also arise from endogenous cortisol overproduction but both forms lead to indistinguishable clinical signs. Dyslipidaemia is a frequent feature in patients with CS, resulting from a GC-induced increase in lipolysis, lipid mobilization, liponeogenesis, and adipogenesis. Hypercholesterolemia, increased triglyceride levels, and alterations in HDL cholesterol and LDL/HDL ratios are also typically observed. An aggressive lipid-lowering management is recommended to lower the cardiovascular disease (CVD) risk in these patients. Dogs have been used as models to study metabolic obesity, diabetes, dyslipidaemia and response to pharmacologic interventions. In dogs, treatment with GCs causes many of the side effects seen in humans, and extended GC therapy can induce iatrogenic CS. Dogs with endogenous or iatrogenic CS show hyperlipidaemia, hypertension, and have a higher risk of developing type 2 diabetes mellitus. However, CS in dogs does not appear to increase the risk of developing atherosclerotic lesions. This may be explained by difference in their lipid metabolism: unlike humans, dogs have very low cholesteryl ester transfer protein (CETP) activity, an enzyme that mediates the transport of triglycerides (TG) from LDL and VLDL to HDL2 and of cholesteryl esters from HDL2 to VLDL and LDL. Dogs carry most of their plasma cholesterol on HDL particles, resulting in an athero-protective profile, as high HDL-cholesterol levels are associated with a low risk of CVD. In recent years, lipidomics has emerged as a new field involving the large-scale study of novel lipids as functional elements and markers of disease. Although little is known about the plasma lipidome composition in dogs, a recent study suggests that dogs show the highest overall similarity to patients with dyslipidaemia in terms of their lipid profiles and responses to the statin, simvastatin. Therefore, dogs may represent a relevant model to evaluate the effects of GCs on lipid metabolism. Thus, the aim of this study was to examine changes in the canine plasma lipidome after short-term and long-term exposure to excess GC. In the first experiment, healthy Beagle dogs received a long-term treatment with tetracosactide, a synthetic peptide, comprising the first 24 amino acids of the adrenocorticotropic hormone (ACTH), which induced chronic endogenous cortisol secretion (experimentally induced CS model). In the second experiment, dogs received a short-term treatment of a high-dose of prednisolone to study the acute effects of GC use. Selected clinical chemistry parameters were measured before and after treatment to monitor the efficacy of prednisolone and tetracosactide treatments and the general health of the animals during the experiments (Fig. 1 and Supplementary Table S1). Short-term prednisolone treatment led to a decrease in serum cortisol levels in six of eight dogs (median of all dogs: −3.2-fold; Fig. 1a). However, dogs P4 and P6 had increased cortisol levels (1.4-fold). By comparison, long-term tetracosactide treatment significantly increased plasma cortisol levels (median: 3.1-fold; Fig. 1b). Plasma lipase activities significantly increased after both treatments, with more relevant changes observed for dogs receiving tetracosactide (median prednisolone: 1.9-fold, tetracosactide: 3.9-fold; Fig. 1c,d). Dog T3 showed a considerably higher lipase activity compared to the others. Plasma triglycerides (TG) were significantly increased after long-term tetracosactide treatment (median: 2.2-fold) but not after short-term prednisolone treatment (Fig. 1e,f). Dog T6 showed a higher increase in TG levels compared to the other dogs after tetracosactide treatment. Neither treatment had a significant effect on total cholesterol levels (Fig. 1g,h). Alanine aminotransferase and alkaline phosphatase activities were significantly increased in all dogs after both treatments (median prednisolone: 2.2 and 1.8-fold, respectively, tetracosactide: 5.3 and 7.3-fold, respectively; Fig. 1i-l).Figure 1Changes in clinical plasma markers of lipid metabolism. Levels for each dog before (0 d and 0 w, respectively) and after (3 d and 25 w, respectively) treatment with short-term prednisolone (left) or long-term tetracosactide (right). Values from the same dog, before and after treatment, are connected with dotted lines. The P values of paired two-tailed t tests comparing log-transformed values before and after treatment are indicated at the top of each panel. Small inserts illustrate data at a magnified scale. ALAT, alanine aminotransferase activity; ALP alkaline phosphatase activity; TG triglycerides. Changes in clinical plasma markers of lipid metabolism. Levels for each dog before (0 d and 0 w, respectively) and after (3 d and 25 w, respectively) treatment with short-term prednisolone (left) or long-term tetracosactide (right). Values from the same dog, before and after treatment, are connected with dotted lines. The P values of paired two-tailed t tests comparing log-transformed values before and after treatment are indicated at the top of each panel. Small inserts illustrate data at a magnified scale. ALAT, alanine aminotransferase activity; ALP alkaline phosphatase activity; TG triglycerides. To further test the efficacy of long-term tetracosactide treatment, we performed low-dose dexamethasone suppression (LDDS) and ACTH stimulation tests. We found that all dogs were positive for both tests (see Supplementary Table S1). Furthermore, long-term tetracosactide treatment resulted in a significant body weight loss in all dogs (range: −2.1 to −4.4 kg, median: −3.0 kg, P = 0.0004; see Supplementary Table S1). The body weights were not recorded after short-term prednisolone treatment, as no relevant changes were expected after such a short treatment. All dogs showed polyuria, polydipsia and polyphagia after prednisolone treatment. After long-term tetracosactide, all but one (T6) dog showed polyuria and polydipsia. Muscle wasting was detected in all tetracosactide-treated dogs, except for dogs T1 and T6, and increased central obesity was observed for dogs T2 and T3 (Supplementary Table S1). Hierarchical clustering based on the abundance of all measured lipid species (Supplementary Table S2) could clearly discriminate plasma samples taken before and after treatment with prednisolone and tetracosactide (Fig. 2a,b). Principal component analysis (PCA) based on the abundance of all measured lipid species separated the samples taken before and after treatments, and showed the distinct effects of long-term tetracosactide treatment, mostly evidenced by principle components PC 1 (Fig. 2c,d) and of short-term prednisolone treatment, evidenced by PC 3 (Fig. 2d). PCA plots obtained from fold changes of all quantified lipid species (Supplementary Table S3) showed a separation between prednisolone- and tetracosactide-treated dogs, mostly evidenced by PC 1 (Fig. 2e) and a separation between the sexes, evidenced by PC 3 and PC 4 (Fig. 2f).Figure 2Principal component (PCA) and clustering analyses of the plasma lipidomic datasets before and after treatments. All plots are based on concentrations of all quantified lipid species in all plasma samples before and after short-term (3 days) prednisolone (dogs P1-P8) or long-term (25 weeks) tetracosactide (dogs T1-T6) treatments. Female dogs have underlined IDs. (a,b) Heatmaps showing the full datasets with abundance from all quantified lipid species in the plasma samples before (empty bar) and after (filled bar) treatment with short-term prednisolone (A, blue) or long-term tetracosactide (B, red). Colour scale represents normalized and centered lipid concentrations per lipid species (in rows). Dog IDs in grey font indicate before, and blue and red, after prednisolone and tetracosactide treatment, respectively. (c,d) PCA showing the specific effects of long-term tetracosactide (red squares) treatment in principal components 1 (PC 1) and of short-term prednisolone treatment (blue triangles) in principal components 3 (PC 3). Samples taken before treatment are indicated with empty symbols and samples after treatment with filled symbols. (e,f) PCA of log2-fold changes before and after short-term prednisolone (dogs P1-P8, blue triangles) and long-term tetracosactide (dogs T1-T6; red squares) treatments, respectively. PC 1 vs. PC 2 highlights distinct plasma lipidome changes between the two treatments (e), whereas PC 3 vs. PC 4 indicates sex-specific effects of the treatments on the plasma lipidome, separating female and male dogs into two clusters ((e); purple and grey ellipses, respectively). Principal component (PCA) and clustering analyses of the plasma lipidomic datasets before and after treatments. All plots are based on concentrations of all quantified lipid species in all plasma samples before and after short-term (3 days) prednisolone (dogs P1-P8) or long-term (25 weeks) tetracosactide (dogs T1-T6) treatments. Female dogs have underlined IDs. (a,b) Heatmaps showing the full datasets with abundance from all quantified lipid species in the plasma samples before (empty bar) and after (filled bar) treatment with short-term prednisolone (A, blue) or long-term tetracosactide (B, red). Colour scale represents normalized and centered lipid concentrations per lipid species (in rows). Dog IDs in grey font indicate before, and blue and red, after prednisolone and tetracosactide treatment, respectively. (c,d) PCA showing the specific effects of long-term tetracosactide (red squares) treatment in principal components 1 (PC 1) and of short-term prednisolone treatment (blue triangles) in principal components 3 (PC 3). Samples taken before treatment are indicated with empty symbols and samples after treatment with filled symbols. (e,f) PCA of log2-fold changes before and after short-term prednisolone (dogs P1-P8, blue triangles) and long-term tetracosactide (dogs T1-T6; red squares) treatments, respectively. PC 1 vs. PC 2 highlights distinct plasma lipidome changes between the two treatments (e), whereas PC 3 vs. PC 4 indicates sex-specific effects of the treatments on the plasma lipidome, separating female and male dogs into two clusters ((e); purple and grey ellipses, respectively). After short-term prednisolone treatment, only Cer d18:2/18:0 was significantly increased among the 16 measured ceramides (2.7-fold; Fig. 3), whereas tetracosactide induced an increase in Cer d18:1/18:0 (3.3-fold). As an exception, dog T4 showed decreased levels in almost all other Cer species. Excluding dog T4, there was an increasing trend in the levels of Cer d18:1 Cer species, even though most of these changes were not statistically significant (see Supplementary Fig. S4 and Supplementary Table S5). Nine of the ten quantified monohexosylceramides (Hex1Cer) were significantly increased after prednisolone treatment (1.7 to 2.2-fold; Fig. 3). After tetracosactide treatment, five Hex1Cer species were significantly increased (1.6 to 3.5-fold; Fig. 3). None of the quantified Hex2Cer species were significantly changed after either treatment (Fig. 3). Monosialodihexosylgangliosides (GM3) were the only gangliosides measured in this study. After prednisolone, GM3 d18:2/18:0 was significantly decreased (1.6-fold), whereas, after tetracosactide treatment, GM3 d18:1/18:0 and GM3 d18:2/18:0 were significantly decreased (2.2 and 3.8-fold; Fig. 3).Figure 3Individual changes in plasma levels of measured plasma sphingolipid species after short-term (3 days) prednisolone (left panels) and long-term (25 weeks) tetracosactide (right panels) treatments. The first column (P; green white colour scale) indicates the FDR-adjusted P values of the paired t-test comparing log2-transformed concentrations before and after treatment (*P ≤ 0.05, **P ≤ 0.01). The second column (FC) indicates average fold changes after vs. before treatment (color scale corresponds to log2-fold changes (log2FC), the value in the fields to fold changes). Subsequent columns indicate log2-fold changes for the individual dogs (P1-P8 for prednisolone, T1-T6 for tetracosactide; female dogs are underlined) after vs. before treatment. |FC| > 4 are capped at the colour scale with the maximum red and blue colour. Dark grey rows (NQ) indicate lipid species that were not detectable or did not meet the quality control criteria in the corresponding treatment group. Individual changes in plasma levels of measured plasma sphingolipid species after short-term (3 days) prednisolone (left panels) and long-term (25 weeks) tetracosactide (right panels) treatments. The first column (P; green white colour scale) indicates the FDR-adjusted P values of the paired t-test comparing log2-transformed concentrations before and after treatment (*P ≤ 0.05, **P ≤ 0.01). The second column (FC) indicates average fold changes after vs. before treatment (color scale corresponds to log2-fold changes (log2FC), the value in the fields to fold changes). Subsequent columns indicate log2-fold changes for the individual dogs (P1-P8 for prednisolone, T1-T6 for tetracosactide; female dogs are underlined) after vs. before treatment. |FC| > 4 are capped at the colour scale with the maximum red and blue colour. Dark grey rows (NQ) indicate lipid species that were not detectable or did not meet the quality control criteria in the corresponding treatment group. Prednisolone did not induce any significant changes in the levels of 20 measured SM species, except for SM 32:2 (Fig. 3). However, dog P5 showed higher, and dog P6, lower, levels in almost all SM species compared to the average trend in the prednisolone group. Long-term tetracosactide treatment resulted in significant increases in SM 34:0 and SM 36:1 (1.7 and 2.2-fold). In particular, dogs T4 and T5 showed decreased levels in all other SM species, whereas the remaining dogs showed variable, non-significant changes. The major plasma S1P species, S1P d16:1, S1P d18:0, S1P d18:1 and S1P d18:2, were all significantly increased after short-term prednisolone treatment (1.3 to 1.8-fold; Fig. 3). However, in dog P6, the levels of all S1P species were decreased or only slightly increased. The same dog also showed decreased levels of most of the measured ceramides and sphingomyelin species, all belonging to the sphingolipid pathway. After long-term tetracosactide treatment, only S1P d16:1 was significantly decreased (2.1-fold) in all dogs; none of the other S1P species showed a consistent change (Fig. 3). After short-term prednisolone treatment, none of the measured five LPC-O, 15 PC-O, one PE-O and 9 PE-P species were changed, and only two of the 12 PC-Ps were significantly increased (Fig. 4). On the other hand, long-term tetracosactide treatment resulted in a significant decrease in four of the five LPC-O species, 13 of the 15 PC-O species and all 12 PC-P species (−1.5 to −3.5-fold; Fig. 4). PE-O 18:1/20:3 was also significantly decreased (2.4-fold), whereas among all nine PE-P species, PE-P 18:0/18:2 was significantly increased (2.1-fold) and the others were unchanged (Fig. 4).Figure 4Individual changes in plasma levels of measured plasma ether-linked and plasmalogen phospholipid species after short-term (3 days) prednisolone (left panels) and long-term (25 weeks) tetracosactide (right panels) treatments. See Fig. 3 for further descriptions. Individual changes in plasma levels of measured plasma ether-linked and plasmalogen phospholipid species after short-term (3 days) prednisolone (left panels) and long-term (25 weeks) tetracosactide (right panels) treatments. See Fig. 3 for further descriptions. Among the 21 measured LPC species, LPC 19:0 was significantly decreased (2.2-fold) and LPC 20:5 significantly increased (4.5-fold) after short-term prednisolone treatment (Fig. 5a). However, in dogs P2 and P3, LPC 20:5 and also LPC 22:6 were unchanged or decreased, while in other six dogs they were strongly increased (5.3 to 12.8-fold, and 1.5 to 6.0-fold, respectively; see Supplementary Table S3). After long-term tetracosactide treatment, there was a significant decrease in the levels of five out of 11 LPC species bearing fatty acid chain-lengths longer than 18 carbons (Fig. 5a). In contrast, there was a significant increase in LPC 20:3 (2.4-fold) and in several LPCs bearing chain lengths shorter than 19 carbons, i.e. LPC 16:0, 17:1, LPC 18:2 and LPC 18:3 (1.5 to 1.7-fold).Figure 5Individual changes in plasma levels of measured plasma acyl phospholipids after short-term (3 days) prednisolone (left panels) and long-term (25 weeks) tetracosactide (right panels) treatments. (a) Changes in individual acyl phospholipid species. See Fig. 3 for further descriptions. (b) Changes in total levels of PC, PE and PI lipid species with either 0 to 3 (≤3, filled circles) or 4 to 8 (≥4, open circles) fatty acid double bonds after long-term tetracosactide treatment. Values originating from the same dog are linked by grey dashed lines; the red dotted lines represent averages. P values were calculated from paired two-tailed t tests comparing log-transformed total abundances before and after treatment (shown above the points), and comparing log2FC of lipids with ≤3 and ≥4 double bonds, respectively (shown on the top). Individual changes in plasma levels of measured plasma acyl phospholipids after short-term (3 days) prednisolone (left panels) and long-term (25 weeks) tetracosactide (right panels) treatments. (a) Changes in individual acyl phospholipid species. See Fig. 3 for further descriptions. (b) Changes in total levels of PC, PE and PI lipid species with either 0 to 3 (≤3, filled circles) or 4 to 8 (≥4, open circles) fatty acid double bonds after long-term tetracosactide treatment. Values originating from the same dog are linked by grey dashed lines; the red dotted lines represent averages. P values were calculated from paired two-tailed t tests comparing log-transformed total abundances before and after treatment (shown above the points), and comparing log2FC of lipids with ≤3 and ≥4 double bonds, respectively (shown on the top). Short-term prednisolone treatment resulted in a significant decrease in 10 of the 37 measured PC species (Fig. 5a). In comparison, long-term tetracosactide treatment led to a significant decrease in the concentration of 10 PC species containing ≥4 double bonds, and in a significant increase in the concentration of eight PC species containing ≤3 double bonds (Fig. 5a,b). The ratio between the total concentration of PC species containing ≥4 and species with ≤3 double bonds was also significantly decreased (Fig. 5b). Prednisolone treatment had no effect on any of the five measured LPE species, whereas LPE 18:2 was increased after tetracosactide treatment (Fig. 5a). Of the 12 PE species tested, prednisolone only significantly increased the levels of PE 34:1 (Fig. 5a). Tetracosactide had no significant effect on the individual and total levels of the 13 measured PE species (Fig. 5a,b); however, there was a considerable increase in all measured PE species in samples from dog T6. Excluding dog T6, a possible trend in decreased levels of individual PE species with ≥4 double bonds, and a trend in increased levels of PE species with ≤3 double bonds can be seen (Supplementary Fig. S4 and Supplementary Table S6). The ratio between the total level of PE species containing ≥4 and species with ≤3 double bonds was significantly decreased, independently of whether dog T6 was excluded or not (Fig. 5b and Supplementary Fig. S4). When excluding dog T6, a significant decrease in the total level of PE species with ≥4 double bonds and a significant increase in the total level of PE species with ≤3 double bonds was found (Supplementary Fig. S4). Short-term prednisolone treatment led to a significant decrease in 11 of the 12 quantified PI species (−1.9 to −2.8-fold; Fig. 5a). PI 38:2 was not quantified in short-term prednisolone treatment samples, as in this group it did not meet the analytical QC criteria. After tetracosactide treatment, a significant decrease in the levels of three molecular species (PI 36:4, PI 38:4, PI 38:5; 1.6 to 2.3-fold) and in the total levels of PIs with ≥4 double bonds was observed. PI 38:2 was significantly (3.3-fold) and the total levels of PI species with ≤3 double bonds were non-significantly increased (Fig. 5a,b). The ratio between the total concentration of PI species containing ≥4 and ≤3 double bonds was also significantly decreased (Fig. 5b). The only quantifiable PS species (PS 38:4) did not significantly change after the treatments (Fig. 5a). Prednisolone treatment led to a statistically significant decrease in only one (TG 48:3) of the 25 quantified TG species (Fig. 6). None of the six measured DGs were significantly changed. However, one dog (P3) showed increased levels of DG and TG species following prednisolone treatment (Fig. 6). When P3 was excluded, a significant decrease in one DG and nine TG species was observed (see Supplementary Fig. S4 and Table S7). Pronounced changes were observed after long-term tetracosactide treatment, with a significant increase in 10 of the 31 measured TGs (3.0 to 19.9-fold); other species showed non-significant trends in increased levels (Fig. 6). For some TG species, one or two dogs (T4 and T5) had decreased, while the other dogs had increased levels. Among the six measured DGs species, only DG 18:0_18:2 was significantly elevated (4.4-fold), however a trend in increased levels in three additional DG species was observed (Fig. 6).Figure 6Individual changes in plasma levels of measured plasma glycerolipids and cholesteryl esters after short-term (3 days) prednisolone (left panels) and long-term (25 weeks) tetracosactide (right panels) treatments. See Fig. 3 for further descriptions. Individual changes in plasma levels of measured plasma glycerolipids and cholesteryl esters after short-term (3 days) prednisolone (left panels) and long-term (25 weeks) tetracosactide (right panels) treatments. See Fig. 3 for further descriptions. Short-term prednisolone exposure did not lead to any significant changes in the 17 quantified CE species, except of CE 24:4, which was increased (2.8-fold; Fig. 6). Tetracosactide induced a significant decrease in the levels of CE 20:4, CE 22:4 and CE 22:5 (2.3 to 2.6-fold; Fig. 6). Dyslipidemia is a common consequence of exposure to excess endogenous or exogenous GC. Dogs may represent a relevant model to study dyslipidemia and to evaluate the effects of GCs on lipid metabolism. Here, we sought to examine how short-term and long-term GC excess alters the canine plasma lipidome. Short-term prednisolone treatment resulted in typical clinical signs (polyuria, polydipsia and polyphagia) and increased serum lipase activity. The decrease in serum cortisol concentrations during prednisolone treatment is expected as a result of the negative feedback of the exogenously administered GC. In two dogs, however, serum cortisol levels were slightly increased, which may be ascribed to the cross-reactivity of the assay with prednisolone that was still present in the serum. The weight loss, the increased serum cortisol, lipase activity, and TG levels, and the positive adrenal function tests of the long-term tetracosactide-treated dogs are consistent with hypercortisolism and the symptoms of CS. The increased serum alanine aminotransferase and alkaline phosphatase activities after both treatments are likely the result of GC-induced hepatopathies and dog-specific GC-induced over-expression of alkaline phosphatase isoforms. Considering that a substantial percentage of circulating lipids are transported by lipoproteins synthesised in the liver, it is plausible that hepatopathy could contribute to plasma lipidome changes. The treatments with short-term prednisolone and long-term tetracosactide had distinct effects on the dog plasma lipidomes (Fig. 2). Of note, there were variabilities among individual dogs in the responses of the clinical chemistry parameters and the plasma lipidome to the treatments, which cannot be attributed to the specific dog. Two limitations in this study were the small sample sizes of the experimental groups and the heterogeneity in age of the individual dogs. Despite these age differences, the body weights of the dogs were comparable and were indicative of fully grown dogs in all groups. The lipid profiles diverged more between individual dogs after the treatments, indicating individuality in the response to the treatments (Fig. 2c,d). Biological and experimental factors, such as differences in resorption, pharmacokinetics, adrenal response, and GC sensitivity may explain such variabilities and have been reported previously for both humans and dogs. Biological variability is expected within dog cohorts, which are usually more heterogeneous than, for example, mice cohorts. A sex-specific response to GC, affecting plasma lipids, have been observed in humans. Each treatment group in our study included about two-thirds male and one-third female dogs. We observed sex-specific effects on the plasma lipidome in both treatments; however, these effects were weaker than the effects of the treatment (Fig. 2e,f). Given the already small sample size, we have not further investigated which lipids contribute to these sex-specific differences. In summary, the differences in overall plasma lipid profiles induced by the treatments were larger than the effects of gender, age, weight, and other possible confounding factors. The most substantial changes in a lipid class occurring in both short-term prednisolone and long-term tetracosactide exposures were found for Hex1Cer levels (Fig. 3). The increased synthesis of Hex1Cer species seen in our study could be attributed to an increase in their precursor ceramide or could be the result of an altered catabolism of complex glycosphingolipids. Interestingly, no significant changes to Hex2Cer were found, which derives from Hex1Cer; yet, a trend towards increased ceramide synthesis was present in response to long-term tetracosactide treatment (Fig. 3). This corresponds to findings in mice, where a seven days exposure to the GC dexamethasone resulted in an increased synthesis of specific plasma ceramides. However, Cer levels did not change after short-term prednisolone treatment, which may have been too short to induce the ceramide production. After long-term tetracosactide treatment, only one ceramide species was significantly elevated, however, other d18:1 ceramides tended to be elevated. A biologically important class of sphingolipids is S1P, which is implicated in various biological processes, including immune reactions, vascular integrity, blood coagulation, tissue growth, apoptosis and human diseases. After short-term prednisolone treatment, we measured an increase in the four most abundant S1P molecular species, which are also the most abundant in human plasma. Vettorazzi et al. showed that dexamethasone treatment led to a significant increase in plasma S1P levels in mice, which they ascribed to an induction in the expression of sphingosine kinase 1 (SphK1), an enzyme that phosphorylates sphingosine to S1P. It is possible that the short-term prednisolone treatment in our study led to an increased synthesis of S1P, which might be implicated in the regulation of inflammatory processes. Long-term tetracosactide treatment significantly decreased the levels of the low-abundant S1P d16:1. No data is available yet in the literature on the functions of S1P d16:1, probably because only recently it was possible to quantify this species in plasma. However, it has been shown that sphingolipids containing the d16:0 and d16:1 sphingoid bases have, compared with the more abundant d18-containing sphingolipids, different biophysical and biological properties and seem to be important in cardiac functions. The most relevant change induced by tetracosactide treatment was measured for ether-linked PCs. Long-term tetracosactide exposure led to a significant reduction in the plasma levels of 29 of the 32 measured ether-linked PC species, while short-term prednisolone did not affect this lipid class (Fig. 4). Ether-linked PCs are synthesised via a distinct pathway compared to acyl PCs. LPC-Os (also named Lyso-PAF) are the precursors for platelet-activating factors (PAF). Lyso-PAF and PAF are signalling molecules implicated in platelet reactivity, endothelial function, and various inflammatory processes. The decreased levels of all measured Lyso-PAF species after long-term tetracosactide treatment—which might consequently reflect altered levels of PAF—may therefore influence inflammatory processes. PC-Ps are normally enriched in the polyunsaturated fatty acids (PUFAs) arachidonic acid (AA) and docosahexaenoic acid (DHA), which are precursors of eicosanoids and related compounds. Reduced circulating PC-P levels may also affect the protective function of lipoproteins, as ether- and plasmalogen phospholipids can act as scavengers of reactive oxygen species. In contrast to our results, most of the ether-linked PC species were unchanged in human CS patients compared to healthy individuals. This may indicate either a difference between dogs and humans in the effects of GCs on this lipid class, and/or a difference in tetracosactide-induced hypercortisolism in contrast to endogenous hypercortisolism in humans. The significance and biological role of these changes in ether-linked PCs in dogs after long-term tetracosactide will have to be better investigated in the future. PE-Ps were not affected by the tetracosactide treatment, suggesting different functions or regulations of this lipid class compared to ether-linked PCs. Another distinct and consistent effect of short-term prednisolone treatment—besides the effects on Hex1Cer—concerns the PI species, which were significantly decreased (Fig. 5a). PIs are sources of arachidonic acid and precursors of phosphatidylinositol phosphates (PIPs), important signalling molecules in inflammatory processes. Phosphoinositide 3-kinases (PI3K) catalyse the synthesis of PIP from PI and interact with GCs during the innate immune response. It will be interesting to further investigate the mechanism and physiological implication of these significant changes in PI levels. After long-term tetracosactide treatment, a significantly decreased ratio between PCs, PEs and PIs with ≥4 and species with ≤3 double bonds was observed (Fig. 5a,b). This difference is mainly caused by a decrease in species with ≥4 and increases in those with ≤3 double bonds (Fig. 5a,b). Others have shown that steroids and ACTH can inhibit Δ5- and Δ6-desaturases in the liver and adrenal glands of rats, resulting in reduced synthesis of 18:3n-6, 18:4n-3 and 20:4n-6 fatty acids. GCs may therefore reduce the synthesis of PUFA-containing phospholipids and hence lower the availability of the precursor pool for eicosanoids, important factors in inflammatory processes. In human patients with CS, plasma levels of PC 38:3, PC 38:4, and PC 40:4 were all reduced as compared with healthy subjects. In our study, however, PC 38:3 was significantly increased. In human plasma, PC 38:3 species mainly consists of PC 18:0/20:3. Interestingly, LPC 20:3 was also strongly increased. A more detailed analysis of the fatty acid components will be helpful in future to better understand the effects of excess GC on the fatty acid metabolism. After short-term prednisolone treatment, approximately one-fourth of the measured PC species were significantly decreased, whereby no effects based on acyl chain-length or saturation could be identified (Fig. 5a). These findings overlap with published data on reduced plasma PC levels after a single-dose of dexamethasone in humans, which presumably result from GC-induced lipolysis. In agreement with elevated plasma TG levels previously reported in humans and dogs with CS, we observed significantly elevated levels of various TG species after long-term tetracosactide treatment (Fig. 6). These increases were reflected in the total TG levels reported by clinical chemistry measurements (Fig. 1f). However, in contrast to the typically elevated total cholesterol levels in dogs and human with CS, we did not observe any significant increases in serum total cholesterol and plasma CE levels (Figs 1h and 6). Three PUFA-containing CE species were also significantly decreased, which may reflect the same observation made for PUFA-containing acyl-glycerophospholipids. The short-term prednisolone treatment led to a trend in reduced levels of many TG species (Fig. 6 and Supplementary Fig. S4). This observation matches previous findings in dogs and humans, where a single dose of dexamethasone resulted in decreased plasma TG levels. However, our TG lipidomic results are not reflected by the total TG levels measured by clinical chemistry (Fig. 1e). This discrepancy may be explained by the overall small effects on total TGs and significant differences in the levels of specific TG species only. CE and total cholesterol levels did not change, in agreement with previously reported effects of a single dose of dexamethasone in human and dogs. Our study revealed specific effects of both treatments on the plasma lipidome that were consistently observed across all individuals: (i) the increase in Hex1Cer after both short-term prednisolone and long-term tetracosactide treatments; (ii) the decrease in PIs after short-term prednisolone treatment; and after long-term tetracosactide, (iii) the decrease in ether-phospholipids, (iv) the increase of TGs and, (v) the decrease in the ratio of acyl phospholipids with 4 or more unsaturations compared with acyl phospholipids with a lower number of double bonds. In summary, our data reveal wide-spread changes in the plasma lipidome—some for the first time—that occur in response to GC treatment, and highlight the similarities and differences between short- and long-term GC excess exposure in dogs. This study represents a first step toward a detailed characterisation of the influence of GCs on the lipid metabolome in dogs and other organisms. These data indicate that our dog model of tetracosactide-induced CS is a promising approach to study the effects of hypercortisolism on lipid metabolism in dogs. The relevance of our study can also be translated to human diseases, as many of the enzymes in lipid pathways that are regulated by GCs are dysregulated in chronic conditions of GC excess; as well as in e.g., metabolic syndrome and CS. Our findings support the potential use of these plasma lipid classes as markers to monitor the effects of short- and long-term GC use. Finally, our results highlight that co-medication with a GC can be a significant confounding factor when analysing the plasma lipidome of patients, such as in those with auto-immune diseases (e.g., rheumatoid arthritis and lupus erythematosus). Beyond the scope of the present study, the biological role and significance of these findings will need to be evaluated in future investigations to provide deeper and novel insights of GC signalling and the impact of GCs on metabolism in health and disease. Purpose-bred male and female Beagle dogs were used in both experiments. The dogs were kept in groups at the research unit of the Vetsuisse Faculty of the University of Zurich and fed a standard commercial maintenance (pellet/kibble) diet (Josera, Adult Sensitive, Kleinheubach, Germany). The experimental group treated with short-term prednisolone included 8 dogs (5 male and 3 females, 8–58 months old, 12.7 to 16.0 kg body weight). The long-term tetracosactide-treated group included 6 dogs (4 male and 2 female, 23 to 83 months old, 12.2 to 18.9 kg body weight) (see Supplementary Table S1). The study was approved by the Cantonal Veterinary Office of Zurich, Switzerland (TVB 133/2013; 276/2014) and conducted in accordance with guidelines established by the Animal Welfare Act of Switzerland. Eight Beagle dogs were treated with 50 mg prednisolone (Streuli Pharma AG, Uznach, Switzerland) orally twice daily for 3 consecutive days (prednisolone group). Blood samples were collected before (control) and at 12 h after the last prednisolone treatment. Six Beagle dogs were infused subcutaneously (Alzet osmotic pump, Durect Corporation, Cupertino, CA, USA) for 25 weeks with tetracosactide (synthetic ACTH; Bachem AG, Bubendorf, Switzerland) (tetracosactide group). Every 4 weeks, new pumps were implanted subcutaneously into the dorsolateral neck under general anaesthesia to deliver increasing doses of tetracosactide. Dogs received a starting dose of 1.3–1.95 µg/kg/d tetracosactide, increasing the concentration to a final dose of 6–10 µg/kg/d. Blood samples were collected before (control) and at 25 weeks after treatment commenced. The successful induction of hypercortisolism was tested with an ACTH stimulation test and a low-dose dexamethasone suppression test at 25 weeks after the start of treatment. For the ACTH stimulation test, 5 μg/kg synthetic tetracosactide (Synacthen, Novartis Pharma Schweiz AG, Bern, Switzerland, Pharmacode: 6748610) was injected intravenously, and plasma cortisol measured at 0 and 1 h after injection. For the low-dose dexamethasone suppression (LDDS) test, 0.01 mg/kg dexamethasone (Dexadreson, MSD Animal Health GmbH, Luzern, Switzerland, ATCvet number: QH02AB02) was injected intravenously, and plasma cortisol measured at 0, 4, and 8 h after injection. Blood samples were collected from the jugular vein after overnight fasting before (control) and at the end of the treatment (see above). For lipidomics analyses, blood was collected into lithium heparin tubes (Vacuette Greiner Bio-One, Frickenhausen, Germany) and centrifuged immediately after collection (1,862 g, 10 min, 4 °C). Obtained plasma was stored at −80 °C until analysis. For clinical chemistry analyses, serum was harvested after clot retraction at room temperature. All clinical blood serum chemistry parameters were determined at the Clinical Laboratory, Vetsuisse Faculty (University of Zurich) on a Cobas Integra 800 instrument (Roche Diagnostics AG, Rotkreuz, Switzerland). The clinical chemistry parameters included triglycerides, cholesterol, alanine aminotransferase (ALAT), alkaline phosphatase (ALP), lipase and other metabolites (Fig. 1 and Supplementary Table S1). Serum cortisol concentrations were measured by a competitive immunoassay validated for dogs (DPC Immulite 1000, Siemens Schweiz AG, Zurich, Switzerland). The intra-assay coefficients of variation were 10.0% and 6.3% at cortisol levels of 2.7 and 18.9 µg/dL, respectively. The sensitivity of the assay was 0.2 µg/dL. For the lipidomics analyses, Process Quality Control (PQC) samples were generated for each experimental group (prednisolone and tetracosactide, respectively) by pooling equal volumes of each plasma sample within an experimental group. Aliquots (10 µL) of corresponding pooled PQC samples were then tested alongside samples of each experimental group. Blank samples, which did not contain any plasma, were also prepared and tested (see “Lipid extraction”). Plasma lipids were extracted using a modified version of the butanol/methanol extraction method. Briefly, plasma samples were thawed on ice, and 10 µL of each sample (or PQC or control) was then transferred into a 2-mL Eppendorf tube (Eppendorf, Germany). Nothing was added to the tubes to be used as Blank samples. To prevent lipid oxidation, 1 µL of 2,6-di-tert-butyl-4-methylphenol (BHT, 10 mmol/L; Sigma-Aldrich; St Louis, MO, USA; B1378) in ethanol was added to each sample (including PQCs and Blanks). To this, 90 µL of 1-butanol:methanol (1:1, v/v) containing the internal standard lipids (Supplementary Table S8) was added. The samples were vortexed for 30 sec and sonicated in an ultrasound water bath at 20 °C for 30 min. After centrifugation (14,000 g, 10 min, 4 °C), 90 µL of supernatant were transferred into 1.5-mL tubes (Sarstedt Nümbrecht, Germany) and dried under a nitrogen stream at 37 °C. Samples were stored at −80 °C. Before use, samples were reconstituted in 90 µL of 1-butanol:methanol (1:1, v/v), sonicated for 10 min at room temperature, and centrifuged at 20,800 g, for 10 min at 4 °C. Then, 80 µL supernatant were transferred into autosampler vials with glass inserts (Agilent Technologies, Santa Clara, CA, USA) for LC-MS analysis. Phospholipids, diacylglycerols, and cholesteryl esters were measured based on a published method, with some modifications. Mobile phase A consisted of acetonitrile:water 4:6 (v:v) with 10 mmol/L ammonium formate (Sigma-Aldrich, 78314) and mobile phase B of acetonitrile:2-propanol 1:9 (v:v) with 10 mmol/L ammonium formate. An Agilent Zorbax RRHD Eclipse Plus C18 (2.1 × 50 mm, 1.8 µm, 95 Å) reversed-phase column maintained at 40 °C was used as the stationary phase. The gradient was composed as follows: 20% B for 2 min, 20% to 60% B from 2 to 7 min, 60% to 100% B from 7 to 9 min, back to 20% B from 9 min until the end (total runtime of 10.8 min). The flow rate was set to 0.4 mL/min and 2 µL sample was injected for analysis. The LC-MS system consisted of an Agilent 1290 infinity UHPLC pump and the Agilent 6460 triple quadrupole (QQQ) mass spectrometer. The ESI source parameters are detailed in Supplementary Table S9. All analyses were performed in dynamic MRM (multiple reaction monitoring) mode with unit resolution. Monitored transitions are listed in Supplementary Table S10. Sphingolipids were measured using the same LC method described above. The injected sample volume was 1 µL. An Agilent QQQ 6495 was used as a mass spectrometer, with distinct ESI source settings (see Supplementary Table S9) and MRM transition list (see Supplementary Table S11). For triacylglycerols, the stationary phase consisted of an Agilent Zorbax Eclipse XDB-C18 Silica, 3 × 150 mm, 1.8 μm, 80 Å column maintained at 25 °C. LC separation was performed isocratically for 25 min at a flow rate of 128 μL/min with chloroform:methanol 1:1 (v:v) containing 2 mmol/L ammonium acetate (Sigma-Aldrich, 17843) as the mobile phase. The LC-MS system consisted of an Agilent 1100 HPLC pump and a Sciex 4000 QTrap mass spectrometer (SCIEX, Framingham, MA, USA) operated in single-ion monitoring (SIM) mode at unit resolution to measure TG precursor ions (see Supplementary Table S12). The ESI source settings are described in Supplementary Table S9. The injected sample volume was 10 µL. Sphingosine-1-phosphate (S1P) analysis was performed according to the method described by Narayanaswamy et al.. To 50 µL of reconstituted lipid extract, 50 µL of C2D2–S1P d18:1 internal standard (20 ng/mL; Toronto Research Chemicals, Toronto, Canada) in 1-butanol:methanol (1:1, v/v) was added, and the sample was derivatized by the addition of 20 µL of TMS-diazomethane (2 mol/L in hexanes; Acros Organics, Thermo Fisher Scientific, New Jersey, USA) for 20 min at 25 °C and 700 rpm (Thermomixer, Eppendorf, Germany). The reaction was stopped by the addition of 1 µL of 100% acetic acid. Samples were centrifuged at 20,800 g for 10 min at 7 °C, and supernatants transferred into autosampler vials for subsequent LC-MS analysis. The mobile phase A consisted of 1:1 (v/v) acetonitrile:25 mmol/L ammonium formate solution (in water, adjusted to pH 4.6 with formic acid), and the mobile phase B of 95:5 (v:v) acetonitrile: 25 mmol/L ammonium formate solution. Analytes were eluted with the following gradient: 99.9% B from 0 to 5 min; 40% B 5 to 5.5 min; 10% B 5.5 to 6.6 min and 99.9% B 6.6 to 9.6 min (total run time 9.6 min). The flow rate was 0.4 mL/min. Samples (5 µL) were injected onto a Waters (Milford, USA) ACQUITY UPLC BEH HILIC (2.1 × 100 mm, 1.7 µm, 130 Å) analytical column maintained at 60 °C. The MS system consisted of an Agilent QQQ 6490. The ESI source parameters are indicated in Supplementary Table S9, and the monitored MRM transition list in Supplementary Table S13. Data acquisition was performed using Agilent MassHunter software (version B.06). The TG raw data was processed with Analyst (Version 1.6.2, SCIEX), and all other data from Agilent instruments with MassHunter QQQ Quantitative software (version B.08). The retention time and, if available, qualifier transitions were used to assign peaks to corresponding lipids. For sphingolipids, transitions of precursor ions with water loss were used as qualifiers. For S1P, the m/z 60 transitions were used as quantifiers and the interference from the S1P d18:1 M + 2 isotope was subtracted from S1P d18:0. For plasmalogen PE (PE-P), transitions with the fatty acid as product were used as quantifiers, and those with the head group as qualifiers. Plasmalogen PCs (PC-P), ether PCs (PC-O) and odd-chain fatty acid PCs were distinguished based on retention time. Normalised peak areas were calculated by dividing the peak areas of the analyte with the corresponding internal standard (ISTD; see Supplementary Table S14). Relative abundance was obtained by multiplying the normalised peak areas with the molar concentration of the corresponding ISTD (see Supplementary Table S8). Lipid species with a median peak area in the PQC samples below 250 or less than 5 times of the Blank samples were excluded. Additionally, the coefficient of variation (CV) of the normalised peak area was calculated for each lipid species in the PQC samples of each experimental group. Species with a CV higher than 25% in any of the two groups were excluded from subsequent evaluation, except of Cer d18:1/18:0 and Hex1Cer d18:1/24:1, which had CVs of 27.8% and 30.1%, respectively, in the tetracosactide group, and 7.3% and 10.9% in the prednisolone group. These species were kept in the dataset to allow comparison between the two groups. The final filtered dataset included 262 lipid species quantified in all samples of both experimental groups (see Supplementary Table S2). The median %CVs of lipids species were 5.3% and 4.5% in the phospholipid/CE/DG, 12.1% and 10.9% in the sphingolipid, 8.5% and 13.1% in the S1P, and 18.1% and 18.4% in the TG panel analyses of the prednisolone and tetracosactide group, respectively. All calculations were performed using R (see Data Availability). Statistical significance of changes in measured clinical chemistry parameters, body weights and lipid abundances between dogs before and after treatment were determined by paired, two-tailed t-tests from log2-transformed values. FDR (false discovery rate)-adjusted P values were calculated using the Benjamini–Hochberg procedure. All calculations and figures were generated using R scripts using R and following described R packages (see Data Availability). Heatmaps of the lipid abundances from samples before and after treatments (Fig. 2a,b) were generated using the heatmap.2 function of the gplots R package with Pearson’s distance and Ward clustering algorithms. The scales indicate autoscaled log2-fold changes. PCA plots were generated with the R packages FactoMineR and factoextra from scaled, centred log2-transformed lipid abundances (Fig. 2c,d) and log2-folds changes (Fig. 2e,f). The heatmaps of fold changes and FDR-adjusted P values (Figs 3 to 6) were generated using ComplexHeatmap. Other plots were generated using ggplot2. Figures were scaled and further annotated in Adobe Illustrator CC (Adobe Systems, San Jose, CA).
|
PMC4906355
|
Gender, Contraceptives and Individual Metabolic Predisposition Shape a Healthy Plasma Lipidome
|
Lipidomics of human blood plasma is an emerging biomarker discovery approach that compares lipid profiles under pathological and physiologically normal conditions, but how a healthy lipidome varies within the population is poorly understood. By quantifying 281 molecular species from 27 major lipid classes in the plasma of 71 healthy young Caucasians whose 35 clinical blood test and anthropometric indices matched the medical norm, we provided a comprehensive, expandable and clinically relevant resource of reference molar concentrations of individual lipids. We established that gender is a major lipidomic factor, whose impact is strongly enhanced by hormonal contraceptives and mediated by sex hormone-binding globulin. In lipidomics epidemiological studies should avoid mixed-gender cohorts and females taking hormonal contraceptives should be considered as a separate sub-cohort. Within a gender-restricted cohort lipidomics revealed a compositional signature that indicates the predisposition towards an early development of metabolic syndrome in ca. 25% of healthy male individuals suggesting a healthy plasma lipidome as resource for early biomarker discovery.Blood plasma analysis is a cornerstone of clinical chemistry. Plasma is an abundant, readily available clinical resource whose composition is reflective of basic merits of metabolism and homeostasis. It contains informative molecular markers of basic pathophysiological processes such as inflammation, atherosclerosis or metabolic syndrome, to mention only a few. A typical blood test to diagnose metabolic syndrome or type 2 diabetes mellitus may report more than 30 clinically relevant indices, however only four of them, total triacylglycerols (TAG), total cholesterol (Chol) and the cholesterol content in HDL and LDL fractions, are directly reflecting the status of lipid homeostasis. Since recently, human blood plasma is being extensively studied by lipidomics (reviewed in1). An inter-laboratory effort spearheaded by the LIPID MAPS consortium quantified 588 individual lipids from 21 major lipid classes2. Other plasma lipidome studies pinpointed individual molecules or entire lipid classes whose abundance was specifically altered in obesity3, type 14 and type 25 diabetes, insulin resistance6, hypertension7, cardiovascular disease89, Alzheimer’s disease10 and schizophrenia1112. Associating lipidome changes with diseases progression shed light on their molecular mechanisms and metabolic consequences and lead to the identification of promising biomarkers13, means of dietary intervention14, or tools for monitoring the efficacy of lipid homeostasis correction through therapeutic or surgical treatments1516. Clinical lipidomics is an emerging field (reviewed in17) and standard operation procedures for quantifying lipids in biofluids and biopsies, as well as general guidelines for recruiting representative patient cohorts are yet to be established. One common approach is to determine relative (fold) changes between the abundance of lipid species in samples from disease and control cohorts and use statistical corrections to adjust for differences in age, BMI or common comorbidities (reviewed in1819). While this may foster the discovery of disease-specific biomarkers, it neither makes the results of independent studies comparable nor improves our understanding of how complex pathologies (e.g. metabolic syndrome) impact the whole lipidome. Nowadays, lipids can be quantified by different means of mass spectrometry (reviewed in202122) and accurate measurements should afford consistent molar values. However, the concordance of lipid concentrations determined by mass spectrometry and by common methods of clinical chemistry has so far received little attention. Plasma lipidome varies between healthy individuals of different ethnic origin and is influenced by circadian rhythm23 and diet24. Lipid metabolism is also gender-dependent (reviewed in25), however it remains unclear how the molecular composition of plasma lipidome is affected by gender and if it is influenced by the level of sex hormones26. While numerous epidemiological screens compared plasma lipidomes of healthy and sick individuals in population-wide cohorts (reviewed in27), no reference values of lipid concentrations and their natural biological variance were established. We applied shotgun lipidomics and liquid chromatography tandem mass spectrometry (LC-MS/MS) to quantify the molar concentrations of 281 molecules from 27 major lipid classes in the plasma lipidome of 36 male and 35 female healthy young Caucasians. We established that gender is a major lipidomic factor that is independent of major clinical and anthropometric indices and whose impact is strongly enhanced by hormonal contraceptive medication in females. Within a gender-restricted group, lipidomics revealed compositional trends indicating metabolic syndrome predisposition in currently healthy individuals. We used shotgun mass spectrometry and LC-MS/MS to determine absolute (molar) concentrations of 281 lipids from 27 major lipid classes. The accuracy and consistency of the lipid quantification was validated in two ways. First, we compared lipid concentrations determined in two independent series of experiments performed with a time gap of two months and each time using two independent internal standards for each lipid class (Fig. 1A). Second, for each member of the study cohort we summed up the concentrations of glycerolipids (48 TAG and 12 diacylglycerol (DAG) species) and compared it with the total concentration of TAG determined by the clinical blood test. In the same way, we summed up the concentration of free Chol and 15 cholesterol ester (CE) species and compared it with the cholesterol concentration from the blood test. Molar concentrations of glycerolipids and cholesterol determined by mass spectrometry and by clinical chemistry were concordant. On average, the difference was −10.4% (r = 0.98) for TAG (Fig. 1B) and 4.1% (r = 0.89) for Chol (Fig. 1C). Next, we assembled a representative anthropometrically homogenous cohort of locally recruited medical students consisting of 36 male and 35 female Caucasians under the age of 33 years. According to the collected anamnesis, each individual had a clean medical record, received no pharmacological treatment at the time of investigation and further examination by a physician revealed no factors commonly comorbid with metabolic disorders. For each individual all 35 clinical indices reported by the blood test, blood pressure and anthropometric indices, including body mass index (BMI) and waist-to-hip ratio (WHO) were within ranges generally accepted as a gender-dependent medical norm. As expected, mean values of these indices also differed between the male and female sub-cohorts. For example, mean BMI differed by 1.2-fold between males and females. However, within each sub-cohort these values varied by less than 10% (Supplementary Tables S1 and S2) suggesting their anthropometric and physiological homogeneity. Therefore, stringent recruitment criteria and focus on young individuals with a clinically documented health status alleviated the need to recruit a larger study cohort without compromising the interpretation confidence. The analysis of healthy plasma lipidomes by mass spectrometry revealed pronounced differences in their molecular composition (Supplementary Table S2). Out of 281 quantified lipids, the abundance of 112 species was significantly (p < 0.01) different (Fig. 2A). Consistently with the clinical blood test, the total lipid content (Fig. 2B) and the total abundance of 21 out of 27 lipid classes were elevated in females with the notable exception of lyso-lipids, ether lipids and ceramides (Cer) (Fig. 2A). The magnitude of concentration differences was lipid class-dependent. It was as high as 50% for phosphatidylethanolamines (PE) and lyso-phosphatidylcholine ethers (LPC O-); at the same time, the difference in sphingomyelin (SM) concentration was small (<20%) yet highly significant (p < 0.001). Cer, PE O-, PC O- and, interestingly, also DAG and TAG, were least affected by gender (Fig. 2A). We wondered whether differences between male and female plasma lipidomes were common to all or to only some members of the cohort? We therefore analysed the lipidomics dataset by unsupervised principal component analysis (PCA) (Fig. 3A) and by non-centred Minimum Curvilinear Embedding (ncMCE)2829 analysis (Fig. 3B). These methods rely on different, yet complementary computational principles30. PCA reflects relations based on the sample variance in the high-dimensional space, whereas ncMCE captures relations based on the hierarchical organization of the samples. Although the distribution of data points (here reflecting full individual lipidomes) may look different, the similarity of their clustering patterns provides an independent evidence of their compositional likeness. Within the female cohort PCA and ncMCE distinguished two partially intermingling sub-groups reflecting the use of hormonal contraceptives (CC) by 19 participants (CC-females). Their lipidomes were separated with high statistical confidence from the lipidomes of males and females not taking CC (nonCC-females), whereas lipidomes of nonCC-females and males partially overlapped (Fig. 3A,B). At the same time, we observed no clear impact of the type of CC-medication. The 4 females who used vaginal rings (Nuvaring) did not group separately from females taking CC as pills. We reasoned that both the intake of CC and inherent gender-related metabolism could be responsible for the observed compositional differences. To delineate the basal differences we first compared the lipidomes of nonCC-females and males. Then, by comparing nonCC-females against CC-females, we identified the differences that arose from or were enhanced by pharmacological interference (Fig. 4). A hallmark difference between the lipidomes of males and nonCC-females was the enrichment of glycosphingolipids (GSL) and SM (Fig. 4A; Supplementary Figs S1 and S2, Supplementary Table S3) in nonCC-females. Four out of 8 classes of GSL were significantly higher (p < 0.01) and also stood out in the comparison between males vs all females (Fig. 2A). At the same time, among GPL the abundance of PE was significantly increased in females. Interestingly, glycerolipid (GL) concentrations were hardly influenced by gender. No DAG or TAG species was significantly different (Supplementary Table S3). Mean values of anthropometric (BMI, WHR, body weight) and clinical (blood pressure, LDL, HDL) indices differ between male and female sub-cohorts (Supplementary Fig. S3, Supplementary Table S1) and we wondered if they were also covariate to gender discriminative lipids? Do lipid concentrations only reflect apparent anthropometric differences, or are they neatly associated with differently regulated lipid metabolism? To answer this question we built a correlation network representing significant associations between gender-discriminative abundances of lipid classes and individual species with anthropometric and clinical indices (Fig. 5A). We observed that clinical indices only associated with each other within a few closed clusters with no evidence of covariate relationship to gender discriminating lipids. The comparison of lipidomes of nonCC-females against CC-females, on the other hand, suggested that hormonal medication shifted the female lipidome composition further away from the male lipidome. However, it did not markedly affect GSL concentrations (Figs 2A and 4A). At the same time, the impact of CC on GPL and GL classes was remarkable: concentrations of 5 out of 8 GPL classes were significantly different (p < 0.01). CC increased plasma concentrations of phosphatidylcholines (PC), PE, and phosphatidylinositols (PI), and lowered the concentration of lyso-lipids (LPC and LPC O-). Concentrations of PE and PC were already higher in the lipidome of nonCC-females compared to males (Fig. 2A), and they were even stronger enriched in CC-females (Fig. 4A). The same trend was observed for lipid classes enriched in males. Concentrations of LPC, LPC O- and LPE were lower in nonCC-females than in males, and they were even stronger depleted from the plasma of CC-female (Fig. 4A). If compared to males and nonCC-females, plasma of CC-females accumulated more lipids in total (Fig. 4B), including Chol, GL and GPL (Fig. 4C–E, respectively). In contrast, plasma of nonCC-females was enriched in GSL compared to males and the difference remained unaffected (and even slightly reduced) by CC (Fig. 4F). Concomitantly, neither sphingolipids (Fig. 4G), nor ether-lipids (PC O- and PE O-), Cer, and eicosanoids were affected by gender alone or by CC (Fig. 4A). These compositional relationships were also reflected in the association network (Fig. 5B). The elevated content of TAG (according to mass spectrometry and clinical chemistry) correlated positively with unsaturated GPL and inversely with lyso-lipids, while showing no association with anthropometric and clinical indices, with the interesting exception of SHBG. Altogether, gender and CC impact plasma lipidomes in different ways. While the concentration of GPL and GL was strongly influenced both by gender and CC (Fig. 4D,E), the latter did not affect the concentration of GSL (Fig. 4F). Overall, CC medication further contrasted the native gender-related differences in lipid metabolism. Lipids already enriched in nonCC-females plasma are becoming even more abundant, while lipids depleted in nonCC-females plasma are depleted to an even larger extent. What biochemical mechanisms could underlie the impact of CC on the lipidome? We noticed that intake of CC reduced the concentration of endogenous estradiol by ca. 3-fold (Supplementary Table S2). This might decrease the expression of the phosphatidylethanolamine-N-methyltransferase (PEMT) and, in turn, decelerate the conversion of polyunsaturated PE (such as PE 38:6 and PE 40:6) into corresponding PC3132 (Supplementary Fig. S4). Although the comparison of plasma of CC- and nonCC-females shows trends consistent with this notion (Supplementary Fig. S4), the variation of levels of endogenous sex hormones, such as estradiol in nonCC-females or (free-) testosterone in males, did not translate into significant and consistent perturbation of the plasma lipidome composition (Supplementary Table S4). It also showed no association with both gender-discriminating lipids (Fig. 5A) and lipids strongly affected by CC (Fig. 5B). Concomitantly, the absence of strong correlation with the levels of other reproductive hormones in females (luteinizing hormone, thyroid-stimulating hormone, and follicle-stimulating hormone) indicates that the menstrual cycle might not be affecting the plasma lipidome composition. Interestingly, CC-intake reduced the level of lyso-lipids, including LPC O-. LPC are inflammation responsive molecules (reviewed in33), whose content is reduced in plasma of patients suffering from e.g. liver inflammation. This corroborates with significantly (p < 0.001) elevated levels of the generic inflammation marker, C-reactive protein (hsCRP), in CC-females plasma (Supplementary Fig. S3), consistently with previous reports3435. Conceivably, lower concentrations of lyso-lipids in the plasma of CC-females together with higher hsCRP suggests that low-grade inflammation induced by CC-intake might “stress” hepatocytes and enhance lipid biosynthesis (Fig. 6A,B). Sex hormone-binding globulin (SHBG) is a plasma glycoprotein produced and secreted by the liver that regulates the bioavailability of sex hormones in circulation (reviewed in36). A lower plasma concentration of SHBG correlates with a higher risk for developing metabolic syndrome, type-2 diabetes, and cardiovascular disease. SHBG was also regarded as a major determinant of lipid plasma profile. This notion, however, was only supported by comparative analyses of common clinical indices of lipid metabolism, such as total concentration of TAG and HDL-cholesterol37. We observed that SHBG concentrations were strongly elevated in CC-females compared to both nonCC-females and males (Fig. 7A). Interestingly, in respect to SHBG CC-females responded to CC in a very different way. SHBG concentrations exceeded 130 nM in a major fraction (12 out of the total of 19) of CC-females. Within this group, the average SHBG concentration was 2.5-fold higher compared to other CC-females (n = 7) and 4.5-fold higher compared to nonCC-females (n = 16). The increase of SHBG concentrations could not be associated with a specific type of CC and seemed to exclusively depend on the individual response. We then asked if the difference in SHBG concentration is associated with a specific lipid profile or trend in clinical indices related to lipid metabolism? We therefore considered the two sub-groups of CC-females having SHBG concentration above (n = 12) and below (n = 7) an arbitrary threshold of 130 nM (Fig. 7A). Individuals from the “high SHBG” sub-group had lower concentrations of CE, Cer, and GM3. The same group also had lower concentrations of LDL, proinsulin, uric acid, thrombocytes, and free fatty acids (FFA) (Fig. 7F–J), while HDL, insulin and hsCRP were slightly increased (Fig. 7K–M) and there was no difference in BMI (Fig. 7N). Taken together, CC could lead to a significant increase of SHBG in plasma, while the response was strongly individual and driven by a yet unknown factor. Higher SHBG concentration might decrease the risk of metabolic syndrome since several lipid classes (e.g. Cer, CE, GM3) and clinical parameters typically associated with its development were reduced638. Statistical analyses indicated that lipidomes of both males and nonCC-females were compositionally heterogeneous (Fig. 3B). We therefore wondered if this heterogeneity only reflected an inherent biological variability between individual lipidomes, or was due to intermingling of compositionally distinct sub-groups? Since our study encompassed only 16 nonCC-females, here we only considered the compositional differences within the larger male cohort (n = 36). ncMCE (Fig. 3B) split the male cohort into two clusters termed Cluster I (n = 27) and Cluster II (n = 9). In comparison to Cluster I, plasma of Cluster II members contained ca 20% more lipids in total (Fig. 8A) along with particularly strong enrichment of TAG (77%) and DAG (63%) and moderate increase of CE and Chol (15%) (Fig. 8B–E). Concentration of other lipid classes increased to a variable extent with a clear tendency towards enrichment of species with unsaturated fatty acid moieties among glycero- (TAG; short and middle chain length DAG) and glycerophospholipids (PC, PI). The abundance of lyso-lipids and ceramides was practically unchanged (Supplementary Tables S2 and S3). Interestingly, the mean abundance of the ether lipids (PC O- and PE O-) (Fig. 8F,G) was reduced in Cluster II compared to Cluster I, consistently with the trend that was previously observed in hypertensive patients7. We reasoned that the lipidome of Cluster II might already show early indication of dyslipidemia or other manifestations of metabolic syndrome, such as hypertension in otherwise healthy individuals. We next asked if anthropometric and clinical indices of members of Cluster II might be having a similar trend towards worsening lipid homeostasis, albeit their absolute values still remained within the normal range. Indeed, indices significant for metabolic syndrome were unfavourably altered in members of Cluster II: increased concentration of insulin, proinsulin and C-peptide, along with higher LDL and lower HDL, testosterone and free testosterone (Fig. 9A,B and D–H). At the same time, common anthropometrical indices (BMI, WHR) (Fig. 9I,J), and other general indices of homeostasis and metabolism (Supplementary Tables S2 and S3) remained unchanged. Interestingly, the inflammation related index hsCRP (Fig. 9C) was substantially reduced in Cluster II, although the major eicosanoids 12-HEPE and 12-HETE were both increased (Supplementary Tables S2 and S3). TAG and DAG strongly differed between the two clusters (Fig. 8A,B). However, selecting the equivalent (n = 9) number of individuals with the highest absolute levels of TAG did not reproduce Cluster II (only 6 from 9 subjects were common) and lead to less pronounced differences between clinical indices (Supplementary Table S2). Although 8 out of 36 males (19.4%) had a BMI between 25.0 and 29.9 kg/m (overweight) only one of them was in Cluster II. Hence, monitoring the level of GL or BMI alone, either by clinical analyses or mass spectrometry, could not reveal perturbed metabolism trends. We therefore concluded that, within the male cohort, lipidomics identified two sub-groups of otherwise healthy individuals. One subgroup (Cluster II; n = 9) showed trends that were reminiscent of those common to developed metabolic syndrome. We observed similar, yet less pronounced, trends also within the nonCC-female cohort. Statistical analyses defined a small cluster (n = 5) of lipidomes having decreased concentration of ether- lipids, yet in contrast to males, their TAG and CE levels were practically unchanged. These individuals showed similar trends towards worsening metabolic indices compared to other nonCC-females, however the small number of selected individuals did not allow us to reach a definitive conclusion. We established the reference values and biological variance of molar concentrations of individual plasma lipids for an age-restricted, ethnically and anthropometrically homogeneous cohort of male and female individuals having no noticeable health and, particularly, metabolism abnormalities. The health status of each study subject was established by 35 indices of a clinical blood test, anthropometric indices, anamnesis and the examination by a physician. Stringent recruitment of healthy individuals alleviated the need to adjust for common comorbidities and allowed us to reduce the cohort size without compromising the interpretation confidence. This notion corroborates the recent report by Begum et al. that showed that statistically confident comparison of healthy lipidomes is also possible using very small study cohorts24. We underscore that this specially recruited healthy, young, ethnically and socially homogeneous cohort does not reflect the diversity of an “average” local population. The reported reference values may therefore serve as a resource to explore the “multidimensionality” of the lipidome variability under medical, ethnical and social contexts. Our interpretation of lipidome compositions solely relied upon molar concentrations of lipid molecules consistently detected throughout the cohort. While this was hardly possible in the past, now a wide and constantly expanding palette of high quality lipid standards is becoming available39. The reference values we reported here could be directly compared with the results of other studies, independently of their design or employed analytical methods. In this way, quantification discrepancies could be spotted and addressed through targeted validation procedures, including the use of multiple lipid standards and independent determinations by conventional methods of clinical chemistry (Fig. 1). We argue that reporting absolute quantities of lipids (rather than their fold changes relative to some arbitrary baseline values) should be generally adopted as a prerequisite clause for clinical lipidomics studies. Our study revealed two major trends in gender-related lipidome differences: the content of GSL and SM (but not Cer) was significantly different between males vs nonCC-females, while they were not (or considerably less) different within the female-restricted cohort. Strong gender-related differences in Gb3 levels were previously observed in mice tissues40 and were thought to reflect metabolism differences associated with Fabry disease. However, we found that in human plasma these differences spanned the entire pool of gangliosides, lactosyl- and glucosylceramides, i.e. they were independent of both glycosylation type and exact structure of the sphingosine backbone. The second trend encompassed GPL (most remarkably PE, PC, PI and corresponding lyso-lipids). While these differences were already apparent in the comparison against male lipidomes, they were strongly enhanced by CC medication, but not affected by the endogenous levels of sex hormones in both females (estradiol) and males (testosterone). One interesting practical consequence is that the exact positioning of female subjects along their menstrual cycle might not be required for clinical lipidomics screens. We argue that CC is a major and often underestimated factor that should be carefully considered when screening lipidomes of females of the reproductive age. The observed differences may mostly reflect the impact of CC, rather than of unfolding metabolic disorders. We are unaware of any published lipidomic screen in which CC-females were either excluded from the study or considered as a separate sub-group. Along the same line, the lipidomics community might need to re-evaluate the possible impact of other life-quality or performance-enhancing medication. While often not considered as drugs and prescribed to otherwise healthy individuals, they may have an unexpectedly strong impact on lipid metabolism. Plasma lipidomics profiling of healthy male individuals revealed that they were already split into two groups. Although indistinguishable by their anthropometric indices, the smaller group (ca 25% of the male cohort) showed a clear trend consistent with unfavourable lipid metabolism, although their clinical indices remained within the acceptable range. We speculate that elevated levels of glycerolipids (TAG and DAG) and CE, along with reduced levels of PE O- and PC O-, relative enrichment of lipid species with polyunsaturated fatty acid moieties, along with unperturbed levels of lyso-lipids (LPE, LPC, LPC O-) are contributing to a collective signature (“lipotype”) of early metabolism disturbance. It is too early to tell if this compositional trend may have a diagnostic value. These studies should continue on a larger population cohort and over an extended period of time. It would also be important to asses if changes in diet and lifestyle may overt or delay the appearance of first clinically recognizable manifestations of metabolic syndrome among subjects of the risk cluster. However, the very existence of a risk cluster indicates that we might need to reconsider a paradigm of biomarker discovery by focusing on early metabolic trends in otherwise healthy individuals. This should now become increasingly possible since omics technologies are able of identifying minor changes in complex molecular composition in any diagnostically promising biological media. Common chemicals and solvents of ACS or LC–MS grade from Sigma–Aldrich Chemie (Munich, Germany) and methanol (LiChrosolv grade) from Merck (Darmstadt, Germany). The study was approved by the local competent authority the Ethik-Kommission an der Technischen Universität Dresden (protocol EK328092011). All approached subjects submitted their informed written consent. 103 individuals (48 males; 55 females), all under 33 years of age, were recruited locally. Each participant completed a questionnaire according to http://www.adipositas-portal.de. Clinical and anthropometric indices covered by the medical examination and by a clinical blood test are in Supplementary Table S2. Upon pre-screening completion, 12 males and 20 females did not meet the selection criteria (Supplementary Table S5) and were excluded. The exception was made for 8 out of 36 males (19.4%) and 1 out of 35 females (2.9%) with BMI between 25.0 and 29.9 kg/m (overweight), because their blood test results were within the requested range. Out of 35 female participants, 19 used hormonal CC. Out of these 19 CC-females, 4 and 15 used hormonal vaginal rings and oral medication, respectively. Within the latter group, 6 persons were taking daily ethinylestradiol (0.02–0.035 mg)/drospirenone (3 mg); 3 persons: ethinylestradiol (0.03 mg)/dienogest (2 mg); 2 persons: ethinylestradiol (0.03 mg)/levonorgestrel (0.125 mg); 1 person: ethinylestradiol (0.035 mg)/cyproteronacetate (2 mg); 1 person: perethinylestradiol (0.02 mg)/desogestrel (0.15 mg); 1 person: dienogest (2 mg). Human blood plasma samples were collected in accordance to ethical guidelines and approved standard clinical protocol after overnight fasting. EDTA-plasma was prepared by 10 min centrifugation at 4 °C and 3000 g. Upon collection, plasma samples were immediately shock-frozen in liquid nitrogen and stored at −80 °C until analysed. Synthetic lipid standards were purchased from Avanti Polar Lipids, Inc. (Alabaster, AL, USA) or Sigma–Aldrich Chemie (Munich, Germany). Stocks of internal standards were stored in glass ampoules at −20 °C until used for plasma analysis. Actual concentrations of standard lipids in stocks were independently validated by direct mass spectrometric analysis using “Quantitative LIPID MAPS standards” (QLMS): high quality reference stock solutions produced, aliquoted, quantified and shipped in sealed glass vials by Avanti Polar Lipids. QLMS validation covered TAG, diacylglycerols (DAG), PC, PE, PI, lyso-phospatidylcholines (LPC), Cer, sphingomyelins (SM). Frozen plasma samples were thawed and lipids were extracted with methyl-tert-butyl ether (MTBE) as described41. Briefly, 5 μl of EDTA plasma was placed in a 2 ml vial (Eppendorf, Hamburg, Germany). Then, we added 700 μl of a mixture of internal standards in MTBE/methanol (5:1.5; v/v) containing: 6199 pmol of CholE 12:0; 4743 pmol of CholD7; 1720 pmol of TAG 36:0; 366 pmol of DAG 24:0; 165 pmol of Cer 30:1:1; 1987 pmol of PC 25:0; 272 pmol of PE 25:0; 195 pmol of PI 32:0; 712 pmol of SM 30:1:1, 440 pmol of LPC 12:0; 487 pmol of LPC 13:0; 425 pmol of LPE 13:0; and 308 pmol LPE 14:0. The samples were vortexed at 4 °C for 1 h. Then, 140 μl of water were added, and the tubes were thoroughly vortexed for 15 min at 4 °C. After centrifuging for 15 min at 13,400 rpm on a Minispin centrifuge (Eppendorf, Hamburg, Germany), the upper organic phases were transferred into glass vials and measured the same day. 10 μl of total lipid extract were finally diluted in 90 μl isopropanol/methanol/chloroform 4:2:1 (v/v/v) containing 7.5 mM ammonium formate and used for mass spectrometric analysis. Mass spectrometric analyses were performed on a Q Exactive instrument (Thermo Fisher Scientific, Bremen, Germany) equipped with a robotic nanoflow ion source TriVersa NanoMate (Advion BioSciences, Ithaca NY, USA) using nanoelectrospray chips with the diameter of spraying nozzles of 4.1 μm. The ion source was controlled by the Chipsoft 8.3.1 software (Advion BioSciences). Ionization voltage was +0.96 kV in positive and −0.96 kV in negative mode; backpressure was set at 1.25 psi in both modes by polarity switching41. The temperature of the ion transfer capillary was 200 °C; S-lens RF level was set to 50%. Each sample was analyzed for 5.7 min. FTMS spectra were acquired within the range of m/z 400–1000 from 0.02 min to 1.5 min in positive and within the range of m/z 400–1000 from 4.2 min to 5.7 min in negative mode at a mass resolution of Rm/z200 = 140000 and automated gain control (AGC) of 10. Free cholesterol was quantified by targeted FT MS/MS within 1.5–4.0 min runtime. For FT MS/MS micro scans were set to 1; isolation window to 0.8 Da; normalized collision energy to 12.5%, AGC to 5 × 10. Lipids were identified by LipidXplorer software42. Molecular Fragmentation Query Language (MFQL) queries were compiled for PC, PC O-, LPC, LPC O-, PE, PE O-, LPE, PI, SM, TAG, DAG, Cer, Chol, CE lipid classes and are available at LipidXplorer wiki site: https://wiki.mpi-cbg.de/wiki/lipidx/index.php/Main_Page. The identification relied on accurately determined intact lipid masses (mass accuracy better than 3 ppm). Lipids were quantified by comparing the isotopically corrected abundances of their molecular ions with the abundances of internal standards of the same lipid class. Only lipids whose monoisotopic peaks were detected with a signal-to-noise ratio above the value of 10 were quantified. Frozen plasma samples were thawed followed by robotic assisted 96-well sample extraction using a Hamilton Microlab Star system (Hamilton Robotics). Simple glycolipids, ranging from hexosylceramide up to Gb3 were extracted from 10 μL plasma as described43. Sphingosines and spingosine-1-phosphates were extracted from 25 μL plasma by 1.1 mL of ice-cold methanol containing 0.1% BHT. Gangliosides were extracted from 100 μL plasma as described by44 with minor modifications. Eicosanoids were extracted from 150 μL plasma as in45. Internal standards were spiked into plasma samples prior extraction. LC-MS/MS analyses were performed on Eksigent XL-100 UHPLC system (AB Sciex) coupled to a QTRAP 5500 triple quadrupole or 6500 QTRAP mass spectrometers (AB Sciex). Hexosylceramides, lactosylceramides and globotriaosylceramides were analysed as described46; sphingosines and sphingosine-1-phosphates as in47; eicosanoides as in45; gangliosides as in48. MRM data were processed by MultiQuant software (Sciex). Lipid species quantified in the plasma of less than 6 out of 71 (~10%) study participants were discarded. Mann-Whitney nonparametric test with Benjamini-Hochberg multiple testing correction was used to determine the significance of changes in lipid abundances. To reveal the similarity between multidimensional lipidome compositions the full dataset was analyzed by juxtaposing two complementary unsupervised and parameter-free pattern visualisation techniques: linear dimensionality reduction principal component analysis (PCA) and nonlinear dimensionality reduction non-centred Minimum Curvilinear Embedding (ncMCE)282930. MATLAB and R code for performing ncMCE is available at: https://sites.google.com/site/carlovittoriocannistraci/5-datasets-and-matlab-code/minimum-curvilinearity-ii-april-2012 We first compared lipidomics datasets nonCC-females vs males and nonCC-females vs CC-females and determined p-values for the abundances of lipid species, lipid classes and also for the clinical variables listed in Supplementary Table 1S and then adjusted them by Benjamini correction. Then only the significant features (with adjusted p-value < 0.05) were selected and then a correlation network was constructed on the significant features, choosing only the significant interactions (with p-value < 0.05), which were characterized by a largely significant correlation, i.e. |r| > 0.7, where r is the Pearson correlation coefficient. These steps were repeated in order to construct the networks that represent significant associations between significantly discriminative (across cohorts) molecular and clinical variables. How to cite this article: Sales, S. et al. Gender, Contraceptives and Individual Metabolic Predisposition Shape a Healthy Plasma Lipidome. Sci. Rep. 6, 27710; doi: 10.1038/srep27710 (2016).
|
PMC8521299
|
Shortening of membrane lipid acyl chains compensates for phosphatidylcholine deficiency in choline‐auxotroph yeast
|
Phosphatidylcholine (PC) is an abundant membrane lipid component in most eukaryotes, including yeast, and has been assigned multiple functions in addition to acting as building block of the lipid bilayer. Here, by isolating S. cerevisiae suppressor mutants that exhibit robust growth in the absence of PC, we show that PC essentiality is subject to cellular evolvability in yeast. The requirement for PC is suppressed by monosomy of chromosome XV or by a point mutation in the ACC1 gene encoding acetyl‐CoA carboxylase. Although these two genetic adaptations rewire lipid biosynthesis in different ways, both decrease Acc1 activity, thereby reducing average acyl chain length. Consistently, soraphen A, a specific inhibitor of Acc1, rescues a yeast mutant with deficient PC synthesis. In the aneuploid suppressor, feedback inhibition of Acc1 through acyl‐CoA produced by fatty acid synthase (FAS) results from upregulation of lipid synthesis. The results show that budding yeast regulates acyl chain length by fine‐tuning the activities of Acc1 and FAS and indicate that PC evolved by benefitting the maintenance of membrane fluidity.The glycerophospholipid phosphatidylcholine (PC) is an essential membrane lipid accounting for at least 50% of total phospholipids in most eukaryotes (van Meer et al, 2008). The exception is presented by several species of green algae that often contain the phosphorus‐free betaine lipid diacylglyceryl‐N, N, N‐trimethylhomoserine (DGTS) instead of PC (Sato & Furuya, 1985; Giroud et al, 1988). DGTS and PC both carry a quaternary amine‐containing zwitterionic head group and share similar biophysical properties (Sato & Murata, 1991). PC is also present in more than 10% of Bacteria, however, bacterial PC has not been assigned any essential function (Geiger et al, 2013). Besides its role as a building block of lipid bilayers, PC has regulatory functions in signal transduction and metabolic regulation in eukaryotes. For example, specific molecular species of PC serve as endogenous ligands for peroxisome proliferator‐activated receptor‐α (PPARα) and liver receptor homologue 1 (LRH1), respectively (Chakravarthy et al, 2009; Lee et al, 2011). Loss‐of‐function mutations of key PC biosynthetic enzymes cause a wide spectrum of human pathologies (van der Veen et al, 2017). Furthermore, alterations in PC metabolism have been implicated in cancer (Ridgway, 2013). By their sheer abundance, PC and its metabolic precursor phosphatidylethanolamine (PE) are important players in determining physical membrane properties such as membrane fluidity and intrinsic curvature that impact the function of membranes and membrane proteins (de Kroon et al, 2013; Covino et al, 2018). Whereas PC has an overall cylindrical molecular shape that makes it ideally suited to build the membrane bilayer matrix, PE is a lipid with non‐bilayer propensity that can adopt a conical shape depending on its acyl chain composition (Renne & de Kroon, 2018). The increased PC/PE ratio induced by obesity in mouse liver was found to inhibit Ca transport by SERCA, causing ER stress (Fu et al, 2011). A decrease in PC/PE ratio in mouse liver induces steatohepatitis and ultimately causes liver failure due to loss of membrane integrity (Li et al, 2006). The tolerance of the model eukaryote Saccharomyces cerevisiae towards variation in membrane lipid composition, makes it ideally suited for addressing the functions of lipid classes in membrane lipid homeostasis (de Kroon et al, 2013). The yeast double deletion mutant cho2opi3 lacking the methyltransferases converting PE to PC, relies on supplementation with choline for the synthesis of PC by the CDP‐choline route (Fig 1A) (Summers et al, 1988; Kodaki & Yamashita, 1989), and has been used to manipulate cellular PC content. Both DGTS and phosphatidyldimethylethanolamine (PDME), a lipid containing two instead of three N‐methyl groups with physical properties similar to PC, can substitute for PC in cho2opi3 (McGraw & Henry, 1989; Boumann et al, 2006; Riekhof et al, 2014), demonstrating that PC is dispensable for yeast growth. ACartoon depicting the biosynthetic pathways producing PC in yeast.B10‐fold serial dilutions of 1 OD600 unit/ml of the indicated strains were spotted on SD plates containing 0 (C) or 1 mM choline (C) and 0 (I) or 75 µM inositol (I) and incubated at 30°C for 3 days. A representative experiment is shown (from n = 5).C2D‐TLC analysis of total lipid extracts of coS#2 cells cultured in SD CI and CI; ori, origin; NL, neutral lipids.DRead‐depth analysis indicating monosomy of chromosome XV in coS#3, S#4, and S#5, but not in coS#2. Each data point represents the median chromosome copy number per 5‐kb bin plotted over the genome, with alternating colours for each successive chromosome and the mitochondrial DNA.ERepresentative DNA content profiles of haploid and diploid cho2opi3 controls (cultured in C) and the indicated cho2opi3 suppressor strains. Cartoon depicting the biosynthetic pathways producing PC in yeast. 10‐fold serial dilutions of 1 OD600 unit/ml of the indicated strains were spotted on SD plates containing 0 (C) or 1 mM choline (C) and 0 (I) or 75 µM inositol (I) and incubated at 30°C for 3 days. A representative experiment is shown (from n = 5). 2D‐TLC analysis of total lipid extracts of coS#2 cells cultured in SD CI and CI; ori, origin; NL, neutral lipids. Read‐depth analysis indicating monosomy of chromosome XV in coS#3, S#4, and S#5, but not in coS#2. Each data point represents the median chromosome copy number per 5‐kb bin plotted over the genome, with alternating colours for each successive chromosome and the mitochondrial DNA. Representative DNA content profiles of haploid and diploid cho2opi3 controls (cultured in C) and the indicated cho2opi3 suppressor strains. Here, we report the isolation and characterization of cho2opi3 suppressor mutants that exhibit sustained growth in the absence of choline. As the suppressors do not contain PC or a PC substitute, elucidation of the mechanism of suppression provides an unbiased route to address PC function. The choline auxotrophy of cho2opi3 is suppressed by 2n‐1 monosomy of chromosome XV or by a point mutation in the ACC1 gene encoding acetyl‐CoA carboxylase. The genetic changes in both suppressors shorten average acyl chain length due to reduced activity of Acc1. Inhibition of Acc1 is sufficient for suppressing choline auxotrophy as evidenced by the rescue of cho2opi3 by soraphen A, a specific inhibitor of Acc1. The results indicate that the suppression by chromosome XV monosomy relies on inhibition of Acc1 by accumulating acyl‐CoA, providing novel clues about the regulation of acyl chain length by the interplay between Acc1 and the fatty acid synthase complex (FAS). Based on the compensatory changes in the PC‐free lipidomes, we propose that the acquisition of PC during evolution provided selective advantage in maintaining membrane physical properties, membrane fluidity in particular. After incubating the cho2opi3 mutant on choline‐free agar plates at 30°C for 14 days, cho2opi3 suppressor clones were obtained. Most of the clones exhibit sustained growth in the absence of choline, and can be stored as and revived from −80°C glycerol stocks in choline‐free SD medium (SD C). A subset of four cho2opi3 (co) suppressor clones, coS#2‐S#5, was characterized in detail (Fig 1). In contrast to their choline auxotroph cho2opi3 parent, coS#2‐#5 grow robustly in the absence of choline, albeit slower than the corresponding WT, irrespective of supplementation with inositol (Fig 1B). The effect of inositol was examined because of its key role in the phosphatidic acid (PA)‐mediated transcriptional regulation of phospholipid biosynthesis genes containing UASINO (Henry et al, 2012). Remarkably, in the absence of inositol, coS#3‐#5 grow slightly better without than with choline present (Fig 1B), suggesting a choline‐sensitive requirement for inositol. The doubling times observed in the corresponding liquid media (Fig EV1A) are consistent with the growth phenotypes on agar plates. ADoubling times of WT (BY4742), cho2opi3, and co S#3, S#4 and S#2, cultured in SD medium supplemented with 1 mM choline and/or 75 µM inositol as indicated. Data are presented as the mean of 2 (wild type) or 3 (other strains) biological replicates, with the individual values indicated.B9 out of 10 independent cho2opi3 suppressor strains exhibit monosomy of chr XV. Copy numbers of chr I, IV, VI, IX and XV derived from qPCR and FACS analysis of cellular DNA content (lower panel and Fig 1E) are shown for the haploid co parent and 10 suppressor strains compared to wild type.CSerial dilution experiment (10–10) comparing haploid co MAT a and co MATα to co diploid on SD C after incubation for 14 days at 30°C, and absolute copy numbers of chr I, IV, VI, IX and XV in a co diploid and 4 derived suppressor strains as determined by qPCR and FACS.DThe cho2opi3 suppressor strains retain the α‐mating type. Suppressor and control strains were mated with SH85 (MAT a ) and SH80 (MATα) as indicated and then streaked on SD C plates without amino acids (AA). The amino acid containing plate (AA) serves as control. The reduced mating efficiency of coS#4 may be due to haploinsufficiency originating from the loss of one copy of the MAT locus on chr III. The ability of BY4743 (MAT a /MAT α) to mate with SH85 is attributed to loss of heterozygosity by homologous recombination (Harari et al, 2018).ECharacterization of 3 cho2opi3 suppressor clones generated on SD CI plates supplemented with 1 mM propanolamine (Prn). Growth phenotype on SD with 1 mM Prn or without supplement after 5 days at 30°C, DNA content by FACS, and absolute copy numbers of chr I, IV, VI, IX and XV in co SPrn#1, #2, #3 and haploid co parent compared to wild type. Doubling times of WT (BY4742), cho2opi3, and co S#3, S#4 and S#2, cultured in SD medium supplemented with 1 mM choline and/or 75 µM inositol as indicated. Data are presented as the mean of 2 (wild type) or 3 (other strains) biological replicates, with the individual values indicated. 9 out of 10 independent cho2opi3 suppressor strains exhibit monosomy of chr XV. Copy numbers of chr I, IV, VI, IX and XV derived from qPCR and FACS analysis of cellular DNA content (lower panel and Fig 1E) are shown for the haploid co parent and 10 suppressor strains compared to wild type. Serial dilution experiment (10–10) comparing haploid co MAT a and co MATα to co diploid on SD C after incubation for 14 days at 30°C, and absolute copy numbers of chr I, IV, VI, IX and XV in a co diploid and 4 derived suppressor strains as determined by qPCR and FACS. The cho2opi3 suppressor strains retain the α‐mating type. Suppressor and control strains were mated with SH85 (MAT a ) and SH80 (MATα) as indicated and then streaked on SD C plates without amino acids (AA). The amino acid containing plate (AA) serves as control. The reduced mating efficiency of coS#4 may be due to haploinsufficiency originating from the loss of one copy of the MAT locus on chr III. The ability of BY4743 (MAT a /MAT α) to mate with SH85 is attributed to loss of heterozygosity by homologous recombination (Harari et al, 2018). Characterization of 3 cho2opi3 suppressor clones generated on SD CI plates supplemented with 1 mM propanolamine (Prn). Growth phenotype on SD with 1 mM Prn or without supplement after 5 days at 30°C, DNA content by FACS, and absolute copy numbers of chr I, IV, VI, IX and XV in co SPrn#1, #2, #3 and haploid co parent compared to wild type. Data information: Chromosome copy numbers in panels B, C and E are presented as mean values from 2 assays using primers complementary to non‐coding regions on the left and right arm of each chr, respectively, with the individual values indicated. Analysis by thin layer chromatography (TLC) of total lipid extracts of the suppressors cultured in SD C‐ indicated that the suppressors are devoid of PC, leaving PE as the predominant membrane lipid (Fig 1C, Appendix Fig S1A). MS analysis corroborated this result. PC could not be detected in negative ion mode as acetate‐adduct, nor in positive ion mode as H‐adduct. Fragmentation in the positive ion mode did not reveal the phosphocholine head group. To elucidate the nature of the adaptation, coS#2‐#5 were subjected to whole genome sequencing (WGS). WGS did not reveal single nucleotide polymorphisms (SNPs), insertions or deletions shared by the four suppressors (Table EV1). However, analysis of chromosome copy number by WGS and fluorescence‐activated cell sorting (FACS) revealed changes in ploidy (Fig 1D and E). Suppressors coS#3, #4 and #5 exhibit 2n‐1 aneuploidy, by losing a copy of chromosome XV (chr XV) after genome duplication. In addition, coS#4 lost part of the right arm of one copy of chr III, whereas coS#5 gained an extra copy of chr IX and lost its mitochondrial DNA. Ploidy changes, including aneuploidy with gain or loss of chromosomes, are common in adaptive evolution of yeast mutants lacking (non‐)essential genes (Storchova, 2014; Szamecz et al, 2014; Liu et al, 2015). Partial karyotype analysis by FACS analysis and a quantitative polymerase chain reaction (qPCR)‐based assay (Pavelka et al, 2010) addressing chr XV with chr I, IV, VI and IX as controls, was applied to an extended set of suppressor clones. Like coS#3‐#5, suppressors coS#6‐#11 exhibit chr XV monosomy, and similar to coS#5, coS#8 and S#9 gained extra copies of chr IX (Fig EV1B). Generation of (2n‐1) suppressors from a diploid co strain proceeds more readily than from its haploid counterparts (Fig EV1C), suggesting that genome duplication is limiting. The odd one out is coS#2 that turned diploid and retained both copies of chr XV (Fig 1D and E). WGS of coS#2 revealed a homozygous point mutation in the ACC1 gene encoding acetyl‐CoA carboxylase, catalysing the rate limiting step of FA synthesis (Tehlivets et al, 2007). Adenosine at position 657039 of both copies of chr XIV is replaced by cytosine, resulting in the substitution of asparagine at position 1446 of Acc1 by histidine (N1446H; Table EV1). Suppressors coS#2‐S#5 retained the MAT α mating type as shown by their ability to mate with a threonine‐auxotrophic MATa strain (Fig EV1D), indicating that genome duplication happened through endoreduplication rather than mating preceded by mating type switch (Harari et al, 2018). Previous research showed that propanolamine (Prn) could substitute for choline in supporting growth of cho2opi3 (Choi et al, 2004). This finding was unexpected, since the physical properties of phosphatidylpropanolamine (PPrn) resemble those of PE rather than PC (Storey et al, 2001). In our hands, Prn does not support growth of cho2opi3 cells. However, suppressors generated on choline‐free agar plates supplemented with 1 mM Prn also grow without supplements and exhibit chr XV monosomy (Fig EV1E). In retrospect, our data suggest that Choi et al, (2004) may have studied cho2opi3 suppressors. We conclude that PC biosynthesis is essential in yeast. However, the requirement for PC can be overcome by adaptive evolution. Electron microscopy for morphological examination of PC‐free cells revealed that compared to WT and the cho2opi3 parent cultured with choline (Fig 2A and B), PC‐free coS#3, S#4 and, to a lesser extent, coS#2 (Fig 2E, C and J) show accumulation of lipid droplets (LD). Quantitation of the area occupied by LD in 2D projection images shows a nearly 3‐fold increase in coS#3 and S#4 compared to WT and parent (Appendix Fig S1B). Other salient features of PC‐free cells include the “spikes” of ER often surrounding LD (Fig 2D), in agreement with LD being formed at and staying connected to the ER (Jacquier et al, 2011). In PC‐free coS#3 and S#2, proliferation of the ER is apparent from protrusions in the nuclear envelope, adopting a “brass‐knuckles” shape that occasionally pushes into the vacuole (Fig 2F, I and J). These structures are reminiscent of the nuclear envelope morphology of a temperature‐sensitive acc1 mutant at the restrictive temperature (Schneiter et al, 1996). In coS#2 unidentified vacuolar structures accumulate at the limiting membrane of the star‐shaped vacuole (Fig 2K). Some mitochondria have aberrant morphology with sheet‐like cristae membranes, often detached from the inner boundary membrane (Fig 2G). Given the defects in mitochondrial structure, it is not surprising that cho2opi3 suppressors do not grow on the non‐fermentable carbon source glycerol without choline (Appendix Fig S1C). Upon supplementing choline, the PC‐free cells return to wild‐type morphology after 3 doublings (Fig 2H and L), with smaller LD (Appendix Fig S1B) often found anchored to the vacuole, suggesting that removal of LD involves lipophagy (van Zutphen et al, 2014). A–LWild type cultured in C (A), cho2opi3 cultured in C (B), coS#4 (C, D), coS#3 (E‐G) and coS#2 (I‐K) cultured in C were analysed by electron microscopy. In addition, coS#3 (H) and coS#2 (L) are shown after culture in C for 3 generations. The arrow heads (F, I, J) point to protrusions in the nuclear envelope. CW, cell wall; ER, endoplasmic reticulum; PM, plasma membrane; M, mitochondria; N, nucleus; V, vacuole; *, lipid droplet. Scale bars correspond to 200 nm (A, B, D, F, G) or 500 nm (C, E, H, I, J, K, L). Wild type cultured in C (A), cho2opi3 cultured in C (B), coS#4 (C, D), coS#3 (E‐G) and coS#2 (I‐K) cultured in C were analysed by electron microscopy. In addition, coS#3 (H) and coS#2 (L) are shown after culture in C for 3 generations. The arrow heads (F, I, J) point to protrusions in the nuclear envelope. CW, cell wall; ER, endoplasmic reticulum; PM, plasma membrane; M, mitochondria; N, nucleus; V, vacuole; *, lipid droplet. Scale bars correspond to 200 nm (A, B, D, F, G) or 500 nm (C, E, H, I, J, K, L). To investigate whether chr XV monosomy is sufficient to suppress choline auxotrophy, a 2n‐1 cho2opi3 strain was constructed by counter selection against a conditionally stable copy of chr XV as described (Reid et al, 2008). Insertion of the GAL1 promoter and a URA3 marker adjacent to the centromere (CEN15) enabled CEN destabilization on galactose‐containing medium, and counter selection against URA3 with 5‐fluoroorotic acid (5‐FOA), respectively. 5‐FOA‐induced loss of the destabilized copy of chr XV conferred uracil‐auxotrophy while suppressing choline auxotrophy (Fig 3A), unequivocally demonstrating that chr XV monosomy rescues the choline auxotrophy of cho2opi3. In the absence of FOA, suppressors of choline auxotrophy appear more frequently with than without uracil present (Fig 3A), as expected based on probability theory. Engineered co S(2n‐1) and evolved coS#3 and S#4 exhibit similar growth phenotypes in the presence or absence of choline and/or inositol (Fig 3B). AInduction of chr XV loss in three independent diploid co/co clones containing a conditionally stable chr XV centromere (left panel), suppresses choline auxotrophy. After culture on solid YPGal, cell patches were replica‐plated on SD plates with or without choline, uracil and 5‐FOA, as indicated (right panel) and incubated at 30°C for 4 days.BGrowth of 10‐fold serial diluted engineered co S(2n‐1) and evolved co S#3 and S#4 on C I SD at 30°C for 4 days.CCholine supplementation induces endoduplication of chr XV in aneuploid cho2opi3 suppressors. Growth phenotype on SD C and C (4 days at 30°C) and absolute copy number of chr I, IV, VI, IX and XV as determined by qPCR and FACS after culturing co S#4 in SD C for the number of days indicated with daily passage to fresh medium at OD600 0.05. Representative data are shown with chromosome copy number presented as mean value from 2 assays using primers complementary to non‐coding regions on the left and right arm of each chr, respectively, with the individual values indicated.DSerial dilutions of the strains indicated were spotted on SD C I and incubated at 30°C for 4 days. Induction of chr XV loss in three independent diploid co/co clones containing a conditionally stable chr XV centromere (left panel), suppresses choline auxotrophy. After culture on solid YPGal, cell patches were replica‐plated on SD plates with or without choline, uracil and 5‐FOA, as indicated (right panel) and incubated at 30°C for 4 days. Growth of 10‐fold serial diluted engineered co S(2n‐1) and evolved co S#3 and S#4 on C I SD at 30°C for 4 days. Choline supplementation induces endoduplication of chr XV in aneuploid cho2opi3 suppressors. Growth phenotype on SD C and C (4 days at 30°C) and absolute copy number of chr I, IV, VI, IX and XV as determined by qPCR and FACS after culturing co S#4 in SD C for the number of days indicated with daily passage to fresh medium at OD600 0.05. Representative data are shown with chromosome copy number presented as mean value from 2 assays using primers complementary to non‐coding regions on the left and right arm of each chr, respectively, with the individual values indicated. Serial dilutions of the strains indicated were spotted on SD C I and incubated at 30°C for 4 days. Upon culture in SD C, 2n‐1 suppressors gradually lose the ability to grow in SD C (and improve growth in SD C) over a period of 6–10 days (Fig 3C, Appendix Fig S1D). This is accompanied by gain of a second copy of chr XV by endoduplication, in agreement with restoration of euploidy after removal of selection pressure (Reid et al, 2008; Chen et al, 2012). The N1446H mutation was introduced in Acc1 in the co background by CRISPR‐Cas9. The engineered haploid co acc1 mutant recapitulates the growth phenotype of coS#2 (Fig 3D), proving that a single point mutation in ACC1 renders PC redundant. By crossing co acc1 to co, a co ACC1/acc1 heterozygous diploid was generated that shows intermediate growth in SD C. Since Acc1 activity is directly linked to changes in lipid metabolism, we subjected the engineered suppressor strains, parent and WT to mass spectrometry‐based shotgun lipidomics (Fig 4, Dataset EV1). After culture in choline‐free medium, co acc1 and co S(2n‐1) show an almost twofold increase in membrane lipid content compared to WT and parent strain, accompanied by 3‐ and 10‐fold increases in triacylglycerol (TAG) in co acc1 and co S(2n‐1), respectively, reflecting increased FA and glycerolipid synthesis in PC‐free suppressors (Fig 4A). Supplementation of choline reduces the level of membrane lipids and TAG in co acc1 to and below WT level, respectively. In co S(2n‐1), lipid content is also reduced by choline but stays above WT/parent levels (Fig 4A). Compared to TAG, the ergosterolester content shows a modest increase in the suppressors that in co S(2n‐1) is not affected by choline (Fig 4A). AMembrane lipid and TAG content, and ergosterolester content (EE, inset) per OD600 unit of the yeast strains indicated after culture to mid‐log phase in SD with or without 1 mM choline (C); *P < 0.05, **P < 0.01, unpaired two‐tailed t‐test of the indicated bar compared to the C condition.BMembrane lipid class composition of classes contributing at least 1% of total membrane lipids of the indicated strains cultured to mid‐log phase in SD C, the inset shows CDP‐DAG and the separate lyso(L)‐phospholipids (lyso‐PL).CFatty acyl chain profiles of the total lipid, the membrane glycerolipid (ML) and the TAG fraction of the indicated strains cultured to mid‐log phase in SD C, showing acyl chains that contribute at least 1% of total, with the C10‐C14 acyl chains in the insets.DPE molecular species profile (sum of carbon atoms in the acyl chains: sum of double bonds in the acyl chains) of the indicated strains cultured to mid‐log phase in SD C, showing species that contribute at least 2% of total PE.E, FPercentage of molecular species containing more than 32 carbon atoms in both acyl chains (C34+C36) (E) and of saturated acyl chains (SFA) (F) in the membrane glycerolipids (ML) and the major membrane lipids, of the indicated strains cultured to mid‐log phase in SD C. Membrane lipid and TAG content, and ergosterolester content (EE, inset) per OD600 unit of the yeast strains indicated after culture to mid‐log phase in SD with or without 1 mM choline (C); *P < 0.05, **P < 0.01, unpaired two‐tailed t‐test of the indicated bar compared to the C condition. Membrane lipid class composition of classes contributing at least 1% of total membrane lipids of the indicated strains cultured to mid‐log phase in SD C, the inset shows CDP‐DAG and the separate lyso(L)‐phospholipids (lyso‐PL). Fatty acyl chain profiles of the total lipid, the membrane glycerolipid (ML) and the TAG fraction of the indicated strains cultured to mid‐log phase in SD C, showing acyl chains that contribute at least 1% of total, with the C10‐C14 acyl chains in the insets. PE molecular species profile (sum of carbon atoms in the acyl chains: sum of double bonds in the acyl chains) of the indicated strains cultured to mid‐log phase in SD C, showing species that contribute at least 2% of total PE. Percentage of molecular species containing more than 32 carbon atoms in both acyl chains (C34+C36) (E) and of saturated acyl chains (SFA) (F) in the membrane glycerolipids (ML) and the major membrane lipids, of the indicated strains cultured to mid‐log phase in SD C. Data information: All data were obtained by mass spectrometry and are presented as mean ± SD (n = 3 biological replicates); *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001, unpaired two‐tailed t‐test of the indicated bar compared to the cho2opi3 parent unless indicated otherwise. Lipidomics analysis of the membrane lipid class distribution in co acc1 and co S(2n‐1) cultured in SD C showed that PE and PI take over as major membrane lipids in the absence of PC (Fig 4B). CDP‐DAG and PS, the metabolic precursors of membrane lipids and PE, respectively, are depleted, reflecting upregulated lipid synthesis, in agreement with cho2opi3 mutants derepressing UASINO genes in the absence of choline (McGraw & Henry, 1989; Boumann et al, 2006). Derepression of the INO1 gene, the UASINO gene with the highest repression/derepression ratio (Henry et al, 2012), results in increased synthesis of inositol that accounts for the increased PI content. The levels of lyso‐PE (LPE) and lyso‐PI (LPI) rise under choline‐free conditions, while the ergosterol (Erg) content is remarkably constant (Fig 4B). Changes in the relative abundance of the sphingolipid M(IP)2C relative to its precursors IPC and MIPC follow that of PI (Fig 4B), as reported previously (Jesch et al, 2010). Under choline‐replete conditions, membrane lipid composition of co S(2n‐1) is restored to that of the parent (Fig 4B), that in turn is similar to WT cultured in SD C (Fig EV2A). Suppressor co S(2n‐1) has lower PI and M(IP)2C levels than co acc1 , and like S#3‐5 above grows slower in CI than in CI (Fig 3B and D), indicating that contrary to co acc1 , 2n‐1 suppressors exhibit a choline‐sensitive requirement for inositol. AMembrane lipid class composition of cho2opi3 cells cultured in SD C compared to WT (BY4742) cultured in SD C, as analysed by MS (mean ± SD, n = 3 biological replicates). The inset shows the distribution of the separate lyso‐phospholipids (lyso‐PL).BPhospholipid class composition of the indicated strains cultured to mid‐log phase in SD CI as indicated. Total lipid extracts were separated by TLC and phospholipid classes quantified by phosphate content. Data are presented as the mean of 2 (cho2opi3 and co S(2n‐1)) or 3 (coS#2) biological replicates, with the individual values indicated.CPhospholipid synthesis in cho2opi3 and derived co S#4 cultured in SD C as indicated. Mid‐log phase cells were pulse‐labelled with [P]‐orthophosphate for 30 min as detailed in the experimental section. Total lipid extracts were analysed by 2D‐TLC followed by phosphor imaging. The label incorporated per phospholipid class has been depicted as mean percentage of the total label incorporated in the total phospholipid fraction of 2 biological replicates with the individual values indicated for classes that contain at least 2% of the label in one of the strains/conditions.DFatty acyl chain composition of total lipid extracts of the indicated strains cultured to mid‐log phase in SD CI as indicated, analysed by GC. Data are presented as the mean of 3 (C) or 4 (C) biological replicates, with the individual values indicated.EMolecular species profile of PE showing individual acyl chain combinations that contribute at least 1% of total PE with the inset showing the most abundant species, of the indicated strains cultured to mid‐log phase in in SD C, as determined by MS (mean ± SD, n = 3 biological replicates). Membrane lipid class composition of cho2opi3 cells cultured in SD C compared to WT (BY4742) cultured in SD C, as analysed by MS (mean ± SD, n = 3 biological replicates). The inset shows the distribution of the separate lyso‐phospholipids (lyso‐PL). Phospholipid class composition of the indicated strains cultured to mid‐log phase in SD CI as indicated. Total lipid extracts were separated by TLC and phospholipid classes quantified by phosphate content. Data are presented as the mean of 2 (cho2opi3 and co S(2n‐1)) or 3 (coS#2) biological replicates, with the individual values indicated. Phospholipid synthesis in cho2opi3 and derived co S#4 cultured in SD C as indicated. Mid‐log phase cells were pulse‐labelled with [P]‐orthophosphate for 30 min as detailed in the experimental section. Total lipid extracts were analysed by 2D‐TLC followed by phosphor imaging. The label incorporated per phospholipid class has been depicted as mean percentage of the total label incorporated in the total phospholipid fraction of 2 biological replicates with the individual values indicated for classes that contain at least 2% of the label in one of the strains/conditions. Fatty acyl chain composition of total lipid extracts of the indicated strains cultured to mid‐log phase in SD CI as indicated, analysed by GC. Data are presented as the mean of 3 (C) or 4 (C) biological replicates, with the individual values indicated. Molecular species profile of PE showing individual acyl chain combinations that contribute at least 1% of total PE with the inset showing the most abundant species, of the indicated strains cultured to mid‐log phase in in SD C, as determined by MS (mean ± SD, n = 3 biological replicates). Data information: *P < 0.05, **P < 0.01, ***P < 0.001, unpaired two‐tailed t‐test of the indicated bar compared to the cho2opi3 parent unless indicated otherwise. Source data are available online for this figure. Conventional TLC analysis of phospholipid composition qualitatively confirmed the lipidomics data and revealed consistent differences between strains when inositol was supplied in the medium except for the PI level remaining largely unchanged (Fig EV2B). Possible causes of differences in lipid class levels as determined by MS and TLC, of PI in particular, have been discussed elsewhere (de Kroon, 2017). Phospholipid biosynthesis was examined by pulse labelling coS#4 cells with [P]‐orthophosphate for 30 min. The reduced percentage of label in PA and the appearance of labelled LPE in the absence of choline (Fig EV2C) indicate increased rates of glycerolipid synthesis and PE turnover, respectively. Analysis of fatty acid content by lipidomics and gas chromatography revealed that both cho2opi3 suppressors exhibit choline‐dependent changes that are hardly affected by supply of inositol (Figs 4C and EV2D). Most conspicuously, the relative C18 FA content of co acc1 is reduced by 50% compared to the parent, an effect observed in both the TAG and the membrane glycerolipid fraction (ML, comprising PC, PE, PI, PS, PA, DAG) with concurrent increases in C10‐C14 (Fig 4C). The changes induced by choline deprivation are largely traced back to the TAG fraction with rises in C16:1 at the expense of C16:0 and C18:0 in both suppressors (Fig 4C). The ML of co acc1 and co S(2n‐1) shows decreases in C18:1 content of 50 and 20%, respectively, that are compensated for by rises in C10‐C14, but otherwise resembles the choline‐supplemented parent. Zooming in on the molecular composition of the individual membrane glycerolipid classes unveils class‐ and choline‐dependent variation between parent and suppressors (Fig 4D and Appendix Fig S2A). Both suppressors show increases in PE 32:1 at the expense of PE 34:2 enhanced by choline deprivation. Moreover, both exhibit a drop in PE 34:1 and rises in C28‐32 species, which in co S(2n‐1) is induced by the absence of choline (Fig 4D). Of note, in choline‐free medium a small but significant fraction of PE 34:1 contains C16:1 and C18:0 next to the dominating PE C16:0_C18:1 species (Fig EV2E). The shortening of acyl chains, i.e. the decrease in the proportion of C34 and C36 lipids (Fig 4E) that is stronger in co acc1 than in co S(2n‐1), is much more pronounced in PE than in PI in PC‐free cells. Choline exerts opposite effects on acyl chain length of PE and PI in the suppressors (Fig 4E), by increasing the proportions of C34 in PE and C26‐28 in PI (Fig 4D and Appendix Fig S2A). The variation in lipid unsaturation is limited, except for increased saturation of PE in co S(2n‐1) and in PC‐free co acc1 (Fig 4F). The sphingolipid profiles of co acc1 are similar to those of the parent strain, whereas co S(2n‐1) shows an increase in C44 species at the expense of C46, accompanied by an increase in hydroxylation that is stronger and enhanced by choline supply in IPC and MIPC compared to M(IP)2C (Appendix Fig S2B). In conclusion, rewiring of lipid synthesis in PC‐free suppressors shortens the average acyl chain length and increases the saturation of PE, adaptations consistent with homeostatic control of membrane fluidity and intrinsic curvature (Ernst et al, 2016; Renne & de Kroon, 2018). Lro1 is a phospholipid:diacylglycerol acyltransferase that converts DAG to TAG by taking an acyl chain from the sn‐2 position of glycerophospholipids (Fig 5A) (Dahlqvist et al, 2000; Oelkers et al, 2000), with substrate preference for PE (Dahlqvist et al, 2000; Ghosal et al, 2007; Horvath et al, 2011). The involvement of Lro1 in the rapid turnover of PE into LPE (Fig EV2C) and the accumulation of TAG under PC‐free conditions (Fig 4A) was investigated. A triple cho2opi3lro1 deletion mutant does not yield suppressors of choline auxotrophy, in contrast to cho2opi3 (Fig 5B). Aneuploid suppressors lacking a copy of chr XV generated from cho2opi3lro1 with LRO1 episomally expressed from the GAL promoter lose the ability to grow without choline upon counter selection against the plasmid (Fig 5B and C, Appendix Fig S3). Moreover, deletion of LRO1 in co acc1 abolishes growth in SD C (Fig 5D). Galactose‐induced expression of Lro1 in co acc1 lro1 transformed with pLRO1 showed that growth and LPE content increase with the galactose concentration in the medium (Fig 5E). In conclusion, Lro1 accounts for the turnover of PE and is essential in PC‐free yeast. AScheme showing the reaction catalysed by Lro1.BGeneration of suppressors of choline auxotrophy of cho2opi3, cho2opi3lro1 and cho2opi3lro1 pLRO1 on choline‐free medium as indicated at 30°C for 7 days.CThree suppressor clones derived of cho2opi3 and cho2opi3lro1 transformed with plasmid pLRO1 were spotted on SGal ura plates without choline, incubated at 30°C for 4 days, replica‐plated onto SGal C and C plates containing 5‐FOA and uracil and incubated at 30°C for 7 days.DSerial dilutions of the strains indicated were spotted on SD C and incubated at 30°C for 6 days.EGrowth (30°C for 6 days) and phospholipid composition of co acc1 lro1 pLRO1 cultured in SD C ura, containing glucose/galactose mixtures (2%, w/v) as carbon source with the percentage of galactose indicated. Phospholipid composition analysed by TLC is presented as mean percentage of total phospholipid of 3 biological replicates with the individual values indicated. Scheme showing the reaction catalysed by Lro1. Generation of suppressors of choline auxotrophy of cho2opi3, cho2opi3lro1 and cho2opi3lro1 pLRO1 on choline‐free medium as indicated at 30°C for 7 days. Three suppressor clones derived of cho2opi3 and cho2opi3lro1 transformed with plasmid pLRO1 were spotted on SGal ura plates without choline, incubated at 30°C for 4 days, replica‐plated onto SGal C and C plates containing 5‐FOA and uracil and incubated at 30°C for 7 days. Serial dilutions of the strains indicated were spotted on SD C and incubated at 30°C for 6 days. Growth (30°C for 6 days) and phospholipid composition of co acc1 lro1 pLRO1 cultured in SD C ura, containing glucose/galactose mixtures (2%, w/v) as carbon source with the percentage of galactose indicated. Phospholipid composition analysed by TLC is presented as mean percentage of total phospholipid of 3 biological replicates with the individual values indicated. Acc1 activity is known to regulate acyl chain length (Hofbauer et al, 2014). The effect on Acc1 activity of the N1446H substitution rescuing choline auxotrophy of cho2opi3 was examined by testing sensitivity to soraphen A (SorA), a high affinity (K d 1 nM) inhibitor of Acc1 (Vahlensieck et al, 1994; Weatherly et al, 2004). Already at 0.05 µg/ml, SorA abolishes the growth of co acc1 and coS#2 irrespective of the presence of choline, demonstrating that the point mutation reduces Acc1 activity (Fig 6A). Accordingly, the lipid‐incorporation of [1‐C]acetate in co acc1 and coS#2 during a 1 h pulse is reduced by 40% compared to the cho2opi3 parent (Fig 6B). A10‐fold serial dilutions of 1 OD600 unit/ml of the strains indicated were spotted on SD C containing 0.05 or 0.25 μg/ml SorA and incubated at 30°C for 3 days. Control plates contain 0.02% (v/v) ethanol.BLipid‐incorporation of [1‐C]acetate after a 1 h pulse in coS#2 and co acc1 relative to the cho2opi3 parent strain during culture on SD C medium. Data are shown as mean of 2 biological replicates.C ACC1 transcript levels after switching the strains indicated from SD CI to the medium indicated at OD600 0.02 or 0.2 (co diploid in C), and culture for 24 h at 30°C. Data were normalized to ACT1 and expressed as means of 3 biological replicates relative to the corresponding strain cultured in CI, with the individual values indicated.D, EFatty acyl chain profile of the total lipid fraction (D) and PE molecular species profile showing species that contribute at least 1% of total PE (E) of cho2opi3 transferred to SD C and SD C containing 0.05 μg/ml SorA as indicated and cultured from starting OD600 0.02 to mid‐log phase. Data obtained by mass spectrometry are presented as mean ± SD (n = 3 biological replicates).F, GEM analysis of cho2opi3 cells that after preculture in SD C, were transferred to OD600 0.05 and subsequently cultured to mid‐log phase in SD C (F) or SD C (G) both containing 0.05 μg/ml SorA. CW, cell wall; ER, endoplasmic reticulum; M, mitochondria; N, nucleus; V, vacuole; *, lipid droplet. Scale bars correspond to 500 nm. 10‐fold serial dilutions of 1 OD600 unit/ml of the strains indicated were spotted on SD C containing 0.05 or 0.25 μg/ml SorA and incubated at 30°C for 3 days. Control plates contain 0.02% (v/v) ethanol. Lipid‐incorporation of [1‐C]acetate after a 1 h pulse in coS#2 and co acc1 relative to the cho2opi3 parent strain during culture on SD C medium. Data are shown as mean of 2 biological replicates. ACC1 transcript levels after switching the strains indicated from SD CI to the medium indicated at OD600 0.02 or 0.2 (co diploid in C), and culture for 24 h at 30°C. Data were normalized to ACT1 and expressed as means of 3 biological replicates relative to the corresponding strain cultured in CI, with the individual values indicated. Fatty acyl chain profile of the total lipid fraction (D) and PE molecular species profile showing species that contribute at least 1% of total PE (E) of cho2opi3 transferred to SD C and SD C containing 0.05 μg/ml SorA as indicated and cultured from starting OD600 0.02 to mid‐log phase. Data obtained by mass spectrometry are presented as mean ± SD (n = 3 biological replicates). EM analysis of cho2opi3 cells that after preculture in SD C, were transferred to OD600 0.05 and subsequently cultured to mid‐log phase in SD C (F) or SD C (G) both containing 0.05 μg/ml SorA. CW, cell wall; ER, endoplasmic reticulum; M, mitochondria; N, nucleus; V, vacuole; *, lipid droplet. Scale bars correspond to 500 nm. Importantly, SorA at 0.05 and 0.25 µg/ml rescues the choline auxotrophy of cho2opi3, unequivocally demonstrating that reduced Acc1 activity is sufficient for suppression (Fig 6A). In contrast to WT and the parent strain, the co S(2n‐1) suppressor is sensitive to SorA at 0.25 µg/ml (Fig 6A), which, together with the observed shortening of average acyl chain length (Fig 4C), indicates that reduced Acc1 activity is key to the mechanism of suppression conferred by chr XV monosomy. The mRNA level of ACC1 goes up 2‐ to 3‐fold in response to inositol or choline deprivation in both suppressors (Fig 6C) in line with UASINO regulation of ACC1 (Chirala, 1992; Henry et al, 2012), and indicating that the lower Acc1 activity in the suppressors is not due to reduced expression. The lipidome of cho2opi3 cultured in medium containing 0.05 µg/ml SorA (Dataset EV1) shows a strongly reduced C18 content and a choline deprivation‐dependent reduction of C34 species in the PE species profile (Fig 6D and E), similar to co acc1 (Fig 4C and D). Except for a much lower total lipid content under choline‐free conditions (Fig EV3A versus Fig 4A), lipidome features of cho2opi3 cultured in the presence of SorA bear strong resemblance to those of co acc1 (compare Fig EV3B, C and D with Fig 4B, E and F). EM analysis shows that the cellular ultrastructure of cho2opi3 cultured with SorA and choline is indistinguishable from that of choline‐deprived cho2opi3 cultured with SorA (Fig 6F and G), and similar to WT cells (Fig 2A). The normal morphology of the SorA‐rescued, PC‐depleted cho2opi3 cells suggests that the morphological changes in the PC‐free suppressor strains (Fig 2) are due to a combination of no PC, short acyl chains and lipid overproduction. AMembrane lipid and TAG content, and EE content (inset) per OD600 unit of cho2opi3 cultured in SD C and in the presence or absence of 0.05 μg/ml SorA, as indicated.BMembrane lipid class composition of lipid classes contributing at least 1% of total membrane lipids, with the inset showing CDP‐DAG and the separate lyso‐phospholipids (lyso‐PL) of cho2opi3 cultured in SD C and in the presence or absence of 0.05 μg/ml SorA, as indicated.CPercentage of molecular species containing more than 32 carbon atoms in both acyl chains (C34+C36) of cho2opi3 cultured in SD C and in the presence or absence of 0.05 μg/ml SorA, as indicated.DPercentage of saturated acyl chains (SFA) in the membrane glycerolipid fraction (ML), and the major membrane lipid classes, of cho2opi3 cultured in SD C and in the presence or absence of 0.05 μg/ml SorA, as indicated.EGrowth phenotype of the single lro1 mutant and cho2opi3lro1 on SD C with and without 0.05 µg/ml SorA as indicated at 30°C for 3 days.FGrowth of the suppressor strains indicated in the presence and absence of 0.25 µg/ml SorA on CI at 30°C for 6 days. Control plates contain 0.02% (v/v) ethanol.GMembrane lipid and TAG content per OD600 unit of co S#4 compared to co S(2n‐1) after culture to mid‐log phase in SD C. Data of co S(2n‐1) taken from Fig 4A.HFatty acyl chain composition of total lipid extracts of co S#4 compared to co S(2n‐1) cultured under the conditions indicated. Acyl chains that contribute at least 1% of total are depicted with the C10‐C14 acyl chains in the inset. Data of co S(2n‐1) taken from Fig 4C. Membrane lipid and TAG content, and EE content (inset) per OD600 unit of cho2opi3 cultured in SD C and in the presence or absence of 0.05 μg/ml SorA, as indicated. Membrane lipid class composition of lipid classes contributing at least 1% of total membrane lipids, with the inset showing CDP‐DAG and the separate lyso‐phospholipids (lyso‐PL) of cho2opi3 cultured in SD C and in the presence or absence of 0.05 μg/ml SorA, as indicated. Percentage of molecular species containing more than 32 carbon atoms in both acyl chains (C34+C36) of cho2opi3 cultured in SD C and in the presence or absence of 0.05 μg/ml SorA, as indicated. Percentage of saturated acyl chains (SFA) in the membrane glycerolipid fraction (ML), and the major membrane lipid classes, of cho2opi3 cultured in SD C and in the presence or absence of 0.05 μg/ml SorA, as indicated. Growth phenotype of the single lro1 mutant and cho2opi3lro1 on SD C with and without 0.05 µg/ml SorA as indicated at 30°C for 3 days. Growth of the suppressor strains indicated in the presence and absence of 0.25 µg/ml SorA on CI at 30°C for 6 days. Control plates contain 0.02% (v/v) ethanol. Membrane lipid and TAG content per OD600 unit of co S#4 compared to co S(2n‐1) after culture to mid‐log phase in SD C. Data of co S(2n‐1) taken from Fig 4A. Fatty acyl chain composition of total lipid extracts of co S#4 compared to co S(2n‐1) cultured under the conditions indicated. Acyl chains that contribute at least 1% of total are depicted with the C10‐C14 acyl chains in the inset. Data of co S(2n‐1) taken from Fig 4C. Data information: Data in A‐D and G‐H were obtained by mass spectrometry and are presented as mean ± SD (n = 3 biological replicates); *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001, unpaired two‐tailed t‐test comparing the indicated bars. Consistent with the essential role of Lro1 under PC‐free conditions, SorA does not rescue the choline auxotrophy of cho2opi3lro1 (Fig EV3E). Of note, evolved coS#3 and S#4 are more sensitive to SorA than engineered co S(2n‐1) and coS#5 (Fig EV3F), indicating that the acquired point mutation in one ACC1 allele (Dataset EV1) reduces enzyme activity. Accordingly, compared to engineered co S(2n‐1) the lipidome of coS#4 shows less accumulation of TAG and a stronger drop in C18:1 in the absence of choline (Fig EV3G and H, Dataset EV1). Taken together, these results show that the shortening of average acyl chain length induced by inhibiting Acc1 is sufficient for rendering PC redundant. As attempts to identify the gene(s) on chr XV requiring lower dosage in co S(2n‐1) suppressors by complementation with a genomic library failed, whole genome transcript profiling was applied to obtain clues about the mechanism of suppression. The transcriptomes of coS#3, S#4 and S#5 cultured without choline and inositol and that of the cho2opi3 parent cultured in CI for up to 3 generations, i.e. still in log phase (Boumann et al, 2006), were compared to WT (Dataset EV2). The clustered heat map (Fig 7A) reveals differences between the transcript profiles of the 3 suppressors that in turn differ dramatically from that of the parent. The transcriptome of the choline‐deprived parent shows correlation with the slow growth signature (Fig EV4A) that is similar to the environmental stress response (Gasch et al, 2000; O’Duibhir et al, 2014), with increased transcription of stress response genes, accompanied by drops in ribosome biogenesis and amino acid metabolism (Fig 7A, Appendix Table S1). In the three suppressors, levels of stress induction are restored to WT (Figs 7A and EV4B). Changes in expression of genes governing FA and phospholipid synthesis are modest in parent strain and suppressors (Fig EV4B). This also applies to the derepression of UASINO genes including their most highly regulated representative INO1, as expected based on the absence of inositol from the culture medium (Jesch et al, 2006). After preculture in SD CI, deprivation of choline causes a stronger derepression of INO1 in the cho2opi3 mutants than deprivation of inositol (Fig EV4C), in agreement with previous reports (Summers et al, 1988; McGraw & Henry, 1989). Derepression is stronger in coS#2 than in co S(2n‐1) and cho2opi3. AHeat map showing log2 fold changes of mRNA expression in the cho2opi3 parent strain cultured in SD CI for 13 h versus wild type, and in co S#3, #4 and #5 cultured to mid‐log phase in CI versus wild type and versus parent. Transcripts changing more than 1.7‐fold (P < 0.01) in at least one of the comparisons are depicted. Hierarchical clustering was by average linkage (cosine correlation); (functional) categories of enriched transcripts were assigned per cluster.BTranscript profiles of the PC‐depleted parent and evolved suppressors versus wild type of UPR‐induced genes (Travers et al, 2000; Kimata et al, 2006) changing more than 1.7‐fold (P < 0.01) in the parent, and of ORE genes from the cluster in Fig 7A. Heat maps show log2 fold changes according to the colour scales; asterisk (*) indicates gene on chromosome XV.C POX1 transcript levels after switching the strains indicated from SD CI to the medium indicated at OD600 0.02 or 0.2 (co diploid in C) and culture for 24 h at 30°C. Data were normalized to ACT1 and expressed as means of 3 biological replicates relative to the corresponding strain cultured in CI, with the individual values indicated. Heat map showing log2 fold changes of mRNA expression in the cho2opi3 parent strain cultured in SD CI for 13 h versus wild type, and in co S#3, #4 and #5 cultured to mid‐log phase in CI versus wild type and versus parent. Transcripts changing more than 1.7‐fold (P < 0.01) in at least one of the comparisons are depicted. Hierarchical clustering was by average linkage (cosine correlation); (functional) categories of enriched transcripts were assigned per cluster. Transcript profiles of the PC‐depleted parent and evolved suppressors versus wild type of UPR‐induced genes (Travers et al, 2000; Kimata et al, 2006) changing more than 1.7‐fold (P < 0.01) in the parent, and of ORE genes from the cluster in Fig 7A. Heat maps show log2 fold changes according to the colour scales; asterisk (*) indicates gene on chromosome XV. POX1 transcript levels after switching the strains indicated from SD CI to the medium indicated at OD600 0.02 or 0.2 (co diploid in C) and culture for 24 h at 30°C. Data were normalized to ACT1 and expressed as means of 3 biological replicates relative to the corresponding strain cultured in CI, with the individual values indicated. AScatter plot comparing the transcript profile versus WT (BY4742) of a cho2opi3 mutant deprived of choline for 13 h in inositol‐free medium to the slow growth signature (taken from O’Duibhir et al, 2014). Transcripts with a change in expression level with P < 0.01 were included.BTranscript profiles of the PC‐depleted parent strain and evolved suppressors versus wild type, showing the 20 stress response genes (GO‐BP:0006950) exhibiting the strongest increase in expression, genes involved in fatty acid and phospholipid biosynthesis, and autophagy that change more than 1.4‐fold (P < 0.01) in at least one of the comparisons (not caused by their location on chromosome XV), and genes governing sulphur amino acid biosynthesis changing more than 1.7‐fold (P < 0.01) in parent or suppressor versus wildtype. Heat maps show log2 fold changes versus wild type; asterisks (*) indicate genes on chromosome XV; genes in bold contain UASINO.C, D INO1 (C) and KAR2 (D) transcript levels after switching the strains indicated from SD CI to the medium indicated at OD600 0.02 or 0.2 (co diploid in C), and culture for 24 h at 30°C. Data were internally normalized to ACT1 and are expressed as means of 3 biological replicates relative to the corresponding strain cultured in CI, with the individual values indicated.EGrowth on SD C (30°C for 3 days) of two co pox1 suppressor clones obtained after 10 days incubation of co pox1 on SD C plates, compared to that of the strains indicated, and absolute copy numbers of chr I, IV, VI, IX and XV in co pox1 and two independent co pox1 suppressor clones, as determined by qPCR and FACS. Data are presented as mean values from 2 assays using primers complementary to non‐coding regions on the left and right arm of each chr, respectively, with the individual values indicated. Scatter plot comparing the transcript profile versus WT (BY4742) of a cho2opi3 mutant deprived of choline for 13 h in inositol‐free medium to the slow growth signature (taken from O’Duibhir et al, 2014). Transcripts with a change in expression level with P < 0.01 were included. Transcript profiles of the PC‐depleted parent strain and evolved suppressors versus wild type, showing the 20 stress response genes (GO‐BP:0006950) exhibiting the strongest increase in expression, genes involved in fatty acid and phospholipid biosynthesis, and autophagy that change more than 1.4‐fold (P < 0.01) in at least one of the comparisons (not caused by their location on chromosome XV), and genes governing sulphur amino acid biosynthesis changing more than 1.7‐fold (P < 0.01) in parent or suppressor versus wildtype. Heat maps show log2 fold changes versus wild type; asterisks (*) indicate genes on chromosome XV; genes in bold contain UASINO. INO1 (C) and KAR2 (D) transcript levels after switching the strains indicated from SD CI to the medium indicated at OD600 0.02 or 0.2 (co diploid in C), and culture for 24 h at 30°C. Data were internally normalized to ACT1 and are expressed as means of 3 biological replicates relative to the corresponding strain cultured in CI, with the individual values indicated. Growth on SD C (30°C for 3 days) of two co pox1 suppressor clones obtained after 10 days incubation of co pox1 on SD C plates, compared to that of the strains indicated, and absolute copy numbers of chr I, IV, VI, IX and XV in co pox1 and two independent co pox1 suppressor clones, as determined by qPCR and FACS. Data are presented as mean values from 2 assays using primers complementary to non‐coding regions on the left and right arm of each chr, respectively, with the individual values indicated. PC‐depletion in the parent strain induces the unfolded protein response (UPR, Fig 7B) in agreement with previous reports (Fu et al, 2011; Thibault et al, 2012). Remarkably, the PC‐free suppressors do not show strongly increased transcription of UPR‐induced genes, including PBI2 and PIR3 that are specifically induced by lipid bilayer stress (Ho et al, 2020). Accordingly, RT–qPCR showed modest increases of the well‐characterized UPR‐induced KAR2 transcript in WT, co S(2n‐1) and coS#2 upon deprivation of choline, whereas KAR2 is strongly upregulated in the choline‐deprived parent (Fig EV4D). Likewise, transcripts related to autophagy upregulated in the choline‐deprived parent (Vevea et al, 2015) return to wild‐type levels in the suppressors (Fig EV4B). Since Cho2 and Opi3 are major consumers of S‐adenosyl methionine (Hickman et al, 2011; Sadhu et al, 2014; Ye et al, 2017), their inactivation impacts the transcription of genes controlling the biosynthesis of sulphur amino acids in both parent and suppressors (Figs 7A and EV4B, Appendix Table S1). Zooming in on the transcript levels increased in the evolved 2n‐1 suppressors but not in the parent, a cluster enriched in genes containing oleate responsive elements (ORE) was identified, including genes encoding enzymes catalysing β oxidation (Fig 7A and B, Appendix Table S1). ORE genes are induced by the transcription factors Oaf1 and Pip2 upon activation by free FA, most strongly by C16:1 (Phelps et al, 2006; Karpichev et al, 2008). Induction of ORE genes in the presence of the preferred carbon source glucose is unprecedented, raising the question whether β oxidation is required for suppression, e.g. by degrading C18 FA. RT–qPCR confirmed that the transcript of the POX1 gene encoding acyl‐CoA oxidase, the rate limiting enzyme of peroxisomal β oxidation, is upregulated in co S(2n‐1) and coS#4, under choline‐free conditions, but not in coS#2 (Fig 7C). A cho2opi3pox1 triple‐mutant plated on choline‐free medium yields suppressors with chr XV monosomy (Fig EV4E) like cho2opi3, suggesting that the induction of ORE genes in 2n‐1 suppressors is a side effect, induced by the high intracellular concentration of free FA. TLC analysis indeed showed that the free FA content of co S(2n‐1) cultured without choline is increased along with the rise in TAG (Appendix Fig S4A). In addition to yielding high levels of free FA and TAG, increased FA synthesis is expected to increase the intracellular concentration of acyl‐CoA, a known inhibitor of Acc1 activity (Kamiryo et al, 1976). To test the hypothesis that inhibition of Acc1 activity by acyl‐CoA accounts for the suppression of choline auxotrophy in co S(2n‐1), Dga1 and Gpt2, which consume acyl‐CoA in synthesizing TAG and lyso‐PA, respectively (Henry et al, 2012), were overexpressed in the suppressors. The ability of co S(2n‐1) to grow without choline was impaired by overexpressing DGA1 or GPT2, whereas growth of co acc1 was not affected (Fig 8), supporting the hypothesis. Moreover, growth of co S(2n‐1) overexpressing DGA1 or GPT2 was restored by SorA, excluding indirect effects on Acc1. Since DGA1 is localized on chr XV, we verified that loss of 1 copy of DGA1 in a diploid cho2opi3 strain is insufficient for suppression of choline auxotrophy (Appendix Fig S4B). In conclusion, feedback inhibition of Acc1 by the product of FAS is crucial for sustained growth of PC‐free co S(2n‐1). Growth of co S(2n‐1) in the absence of choline is lost when GPT2 or DGA1 is overexpressed, and restored by SorA. Plasmids pHEYg‐1‐DGA1, pHEYg‐1‐GPT2 and the empty vector control (pEV) were transformed into the strains indicated. After preculture in SD C, 10‐fold serial dilutions were spotted on SD C with or without 0.1 μg/ml SorA and incubated at 30°C for 3 days. The availability of PC‐free yeast strains provides the opportunity to address PC‐function in vivo. The contribution of PC to the physical state of yeast membranes was investigated using the membrane permeable, polarity‐sensitive fluorescent membrane probe C‐Laurdan. The emission spectrum of C‐Laurdan shifts in response to changes in lipid packing, which is quantified by a parameter denoted generalized polarization (GP) (Gaus et al, 2006; Sezgin et al, 2014). In agreement with previous data from mammalian cells (Lorent et al, 2020), we observed that the GP‐value of plasma membranes is generally higher than that of internal, organellar membranes (Fig 9A, white arrows), particularly in cho2opi3 and co acc1 cells cultured without choline. Since plasma membranes only represent a small part of the stained membranes in yeast cells, changes in average GP per image (Fig 9A, histograms and Fig 9B), are mainly due to changes in lipid packing of internal membranes. AIntensity encoded GP‐images of the indicated C‐Laurdan‐stained yeast strains cultured with or without choline, accompanied by the corresponding intensity weighted GP‐histogram fitted to a single Gaussian function (light blue line), and transmission image. White arrows point to plasma membranes exhibiting higher lipid packing than internal, organellar membranes; blue horizontal arrows indicate highly packed lipid material in co S(2n‐1). To deplete the cho2opi3 parent of PC, cells were transferred to SD C at OD600 0.1 and cultured into mid‐log phase. White scale bars correspond to 10 µm; red scale bars in transmission images correspond to 5 µm.BComparison of average GP‐values ± SD from at least 7 different images per strain and condition from two independent experiments, with the data points representing the average GP‐values of the separate images. Statistical comparison by GraphPad Prism one‐way ANOVA: **P < 0.01, ****P < 0.0001. Intensity encoded GP‐images of the indicated C‐Laurdan‐stained yeast strains cultured with or without choline, accompanied by the corresponding intensity weighted GP‐histogram fitted to a single Gaussian function (light blue line), and transmission image. White arrows point to plasma membranes exhibiting higher lipid packing than internal, organellar membranes; blue horizontal arrows indicate highly packed lipid material in co S(2n‐1). To deplete the cho2opi3 parent of PC, cells were transferred to SD C at OD600 0.1 and cultured into mid‐log phase. White scale bars correspond to 10 µm; red scale bars in transmission images correspond to 5 µm. Comparison of average GP‐values ± SD from at least 7 different images per strain and condition from two independent experiments, with the data points representing the average GP‐values of the separate images. Statistical comparison by GraphPad Prism one‐way ANOVA: **P < 0.01, ****P < 0.0001. Whereas WT and the cho2opi3 parent strain cultured with choline show similar GP‐values, PC‐depletion in cho2opi3 cells drastically increased the GP‐value of C‐Laurdan, reflecting decreased membrane polarity due to increased lipid packing/reduced membrane fluidity (Fig 9A and B) as a result of increased PE content (cf. Dawaliby et al, 2016; Ballweg et al, 2020). In both suppressor strains cultured without choline, the GP‐value returns to values intermediate between choline‐deprived and choline‐supplied cho2opi3 cells (Fig 9A and B), in agreement with acyl chain shortening restoring membrane fluidity in the absence of PC. These results indicate that PC plays a crucial role in maintaining membrane fluidity. Growth of PC‐free cho2opi3acc1 and coS(2n‐1) is reduced and ablated, respectively, at both 20 and 37°C (Fig EV5A), further supporting this concept. AGrowth phenotype of the strains indicated after 4 days of incubation on SD C at 20, 30 and 37°C.BEfficiency of sporulation of MATα/MATa cho2opi3 acc1 homozygous diploids pre‐cultured in SD with or without choline (C) compared to diploid wild type BY4743. The % sporulation was estimated by dividing the number of asci containing two, three or four spores by the total number of cells. At least 300 cells were examined for each strain in subgroups of around 30, with each data point reflecting the % sporulation in one subgroup. Growth phenotype of the strains indicated after 4 days of incubation on SD C at 20, 30 and 37°C. Efficiency of sporulation of MATα/MATa cho2opi3 acc1 homozygous diploids pre‐cultured in SD with or without choline (C) compared to diploid wild type BY4743. The % sporulation was estimated by dividing the number of asci containing two, three or four spores by the total number of cells. At least 300 cells were examined for each strain in subgroups of around 30, with each data point reflecting the % sporulation in one subgroup. Another interesting observation was the presence of highly packed lipid material in PC‐free co S(2n‐1) cells (Fig 9A, blue horizontal arrows). In the transmission channel, the red GP‐dots seemed mostly to coincide with transparent dots, which probably correspond to the large lipid droplets observed by electron microscopy (Fig 2). Reportedly, PC is required in yeast sporulation. The phospholipase D Spo14 specifically hydrolyses PC in vitro, and its activity is essential in sporulation (Rose et al, 1995). Accordingly, sporulation of a diploid homozygous cho2opi3acc1 mutant was found to depend on the presence of choline during preculture (Fig EV5B). However, since sporulation also requires mitochondrial function (Treinin & Simchen, 1993), which is impaired in PC‐free cells (Appendix Fig S1C), the prime cause of the failure of PC‐free cells to form spores remains unknown. The choline auxotrophy of cho2opi3, i.e. the requirement for PC, is suppressed by reduced activity of Acc1, as shown by the acc1 and 2n‐1 aneuploid suppressor mutants, and by the rescue by SorA (Fig 10). The sustained growth of the PC‐free suppressors implies that PC or PC substitutes, such as DGTS or PDME, their biosynthesis and turnover do not fulfil essential functions during mitotic fermentative growth of yeast. In contrast, growth on non‐fermentable carbon source is abolished, indicating that full mitochondrial function is incompatible with the absence of PC (Griac et al, 1996), and/or incompatible with the adaptation. Furthermore, the PC‐free cho2opi3acc1 diploid is not capable of sporulation. It will be interesting to address other processes in which PC or PC metabolism has been implicated in a PC‐free background, such as intracellular vesicle trafficking (Bankaitis et al, 2010), and mRNA localization (Hermesh et al, 2014). PC and PC biosynthesis have become obsolete (blurred). The inhibition of Acc1 (highlighted in red) is essential for shortening average acyl chain length, which in turn is required for maintaining the physical properties of membranes faced with excess PE. The decreased non‐bilayer propensity of PE is indicated by its reduced conical shape. See text for details. Our findings qualify PC biosynthesis as an evolvable essential process (Liu et al, 2015). As frequently observed in suppression of severe growth defects due to genetic or environmental perturbations (Chen et al, 2012; Liu et al, 2015; van Leeuwen et al, 2020), a change in ploidy is the prevalent genetic adaptation suppressing PC deficiency. The stress caused by PC‐depletion probably induces chromosome instability resulting in aneuploidy. Aneuploidy drives rapid adaptation by changing the stoichiometry of gene(s) and by further increasing chromosome instability, imparting fitness gains under adverse conditions (Torres et al, 2007; Pavelka et al, 2010; Zhu et al, 2012; Liu et al, 2015). The phenotypic variation introduced by aneuploidy is immediately clear from the different transcript profiles of coS#3‐5, while the extra copies of chr IX in some suppressors underscore chromosome instability. After genome doubling, both coS#3 and S#4 acquired point mutations in one allele of ACC1 that reduce Acc1 activity, illustrating that aneuploidy facilitates genetic adaptation (Yona et al, 2012; Szamecz et al, 2014). The acc1 mutation in suppressor coS#2 preceded genome duplication, suggesting a fitness advantage of the diploid over the haploid PC‐free state (cf. Harari et al, 2018). Yeast Acc1 is a 500‐kDa homodimer, of which the crystal structure has been solved (Wei & Tong, 2015). Asparagine 1446 that is mutated to histidine in coS#2, is one of the few conserved amino acids in the Acc1 central (AC) region and located close to the catalytic carboxyltransfer (CT) domain. The AC region is thought to properly position the biotin carboxylase (BC) and CT dimers for catalysis (Wei & Tong, 2015). We speculate that subtle changes in positioning caused by the N/H substitution account for the reduced activity of Acc1. Acc1 activity is rate limiting fatty acid synthesis, affecting both amount and length of the acyl‐CoA’s produced by FAS (consisting of six Fas1 and six Fas2 subunits). Higher concentrations of malonyl‐CoA result in the synthesis of longer acyl‐CoA’s in vitro (Lynen et al, 1964; Hori et al, 1987). Accordingly, the hyperactive Acc1 mutant lacking a phosphorylation site for the Snf1 kinase displays a shift to longer average acyl chain length, along with increased FA synthesis and TAG accumulation (Hofbauer et al, 2014). Conversely, conditional acc1 mutants exhibit diminished average acyl chain length under restrictive conditions (Schneiter et al, 1996, 2000), similar to co acc1 . Since the sphingolipid species profiles are similar between co acc1 and parent, the supply of malonyl‐CoA probably does not limit the activities of the acyl‐CoA elongases Elo1, 2, 3 (Tehlivets et al, 2007). It is important to realize that ACC1 and FAS1/2 are UASINO genes co‐regulated at the transcriptional level by the ER‐associated transcription factor Opi1 that senses changes in PA concentration. As the PA level drops, Opi1 translocates into the nucleus to repress the Ino2/4 transcriptional activator complex (Henry et al, 2012). Under choline‐free conditions, the derepression of UASINO genes in cho2opi3 strains is much stronger than in inositol‐deprived WT, which is attributed to the enhanced binding of Opi1 to PA in membranes containing increasing levels of PE (Young et al, 2010; Putta et al, 2016), and to PA with shorter acyl chains (Hofbauer et al, 2014). The derepressed state of UASINO genes likely explains the increased lipid production in PC‐free co acc1 , and contributes to that in co S(2n‐1). Derepression in co acc1 is stronger than in co S(2n‐1), which in addition to shorter average acyl chain length, may be due to haploinsufficiency of INO4 in the latter. Similarly, INO4 haploinsufficiency may account for the choline‐sensitive inositol requirement of co S(2n‐1), induced by the CDP‐choline pathway depleting PA levels via DAG (Gaspar et al, 2017). In the PC‐free aneuploid co S(2n‐1) suppressor, the shift to shorter average acyl chain length and the increased sensitivity to SorA are modest, whereas FAS activity is exceedingly high, compared to co acc1 . In fact, the reduced length of sphingolipids in co S(2n‐1) may be explained by FAS activity competing with the elongase Elo3 for malonyl‐CoA (Al‐Feel et al, 2003). We propose that the increased FA synthesis conferred by chr XV monosomy is required for inhibition of Acc1 by acyl‐CoA (Fig 10), as depletion of acyl‐CoA by overexpression of Dga1 or Gpt2 abolishes suppression. These results identify acyl‐CoA as feedback regulator of Acc1 in vivo. Inhibition of yeast and mammalian Acc1 by acyl‐CoA was previously demonstrated ex vivo and in vitro, respectively (Kamiryo et al, 1976; Ogiwara et al, 1978). Importantly, the results show that the ratio of Acc1‐to‐FAS activity rather than Acc1 activity per se determines acyl chain length, in agreement with loading of malonyl‐CoA competing with release of the mature acyl‐CoA at the malonyl/palmitoyl transferase (MPT) domain of FAS, a concept based on in vitro studies (Heil et al, 2019). Whether the aneuploidy‐induced reduction of the Acc1‐to‐FAS activity ratio is sufficient to account for suppression in co S(2n‐1), awaits identification of the haploinsufficient genes on chr XV (likely 2 or more) responsible for the suppression. The overall shift to shorter acyl chains in the PC‐free suppressors compensates for the decrease in membrane fluidity conferred by increased PE content (Dawaliby et al, 2016; Renne & de Kroon, 2018), as confirmed by the C‐Laurdan GP measurements. Concomitantly, it keeps membrane intrinsic curvature in check by reducing PE’s non‐bilayer propensity (Fig 10) (de Kroon et al, 2013). Indeed, responding to increased negative intrinsic curvature by decreasing unsaturation would have aggravated the drop in membrane fluidity, while enhancing acyl chain unsaturation in response to decreasing membrane fluidity would have jeopardized membrane integrity by increasing the non‐bilayer propensity of PE. In agreement with this latter notion, rising PE levels do not affect the Mga2 sensor that activates transcription of OLE1 encoding the yeast desaturase (Ballweg et al, 2020). The average acyl chain length of PE in the suppressors is reduced by choline‐starvation, which is offset by increased acyl chain length of PI. Similar changes in PE profile occur during PC depletion for four generations in the parent strain (Boumann et al, 2006) and presumably proceed by a similar mechanism. The relatively mild UPR induction in PC‐free cells indicates that the physicochemical properties of the adapted PC‐free membranes are close to wild type (Halbleib et al, 2017). PC biosynthesis controls the level of PE directly by N‐methylation and indirectly by shunting PA via DAG into the CDP‐choline pathway (Fig 10). Choline‐deprived single cho2 and opi3 mutants show a rise in TAG (Fei et al, 2011; Thibault et al, 2012), and growth of a cho2lro1dga1 mutant in the absence of choline is severely compromised (Garbarino et al, 2009; Vevea et al, 2015), indicating that TAG synthesis is required to buffer defective PC synthesis. Under PC‐free conditions, Lro1 becomes essential by degrading PE, compensating for the lack of PC biosynthesis. Whether Lro1 degrades specific PE molecular species, whether its activity is regulated, and whether the LPE produced is just a metabolic intermediate or contributes to bilayer stability by virtue of its more cylindrical molecular shape (Tilcock et al, 1986), are important questions for future research, as is the metabolic fate of LPE. Bacteria with a high content of unsaturated acyl chains often contain PC (Goldfine, 1984). Inactivation of PC biosynthesis abolishes their growth at higher temperatures, arguing that PC evolved to neutralize the tendency of unsaturated PE to adopt non‐bilayer structure (Geiger et al, 2013). In the reverse evolution of cho2opi3 yeast reported here, the lack of PC is compensated for by acyl chain shortening, in line with PC restraining negative intrinsic curvature conferred by PE. The nature of the compensatory response, i.e. shortening rather than decreased unsaturation of acyl chains, argues that PC became crucial in maintaining fluidity of eukaryotic membranes during evolution. The temperature‐sensitive growth of the PC‐free suppressors (Fig EV5A), and the increased membrane order detected by C‐Laurdan in PC‐depleted cells (Fig 9) support this notion. The PC‐free cho2opi3 suppressor strains add a new model system for use in research aimed at understanding membrane lipid homeostasis and lipid function. Reagent or ResourceSourceIdentifier Yeast strains BY4742 (MATα; his3Δ1; leu2Δ0; lys2Δ0; ura3Δ0)EuroSCARFY10000BY4741 (MATa ; his3Δ1; leu2Δ0; met15Δ0; ura3Δ0)EuroSCARFY00000BY4743 (MATa /α; his3Δ1/his3Δ1; leu2Δ0/leu2Δ0; LYS2/lys2Δ0; MET15/met15Δ0; ura3Δ0/ura3Δ0)EuroSCARFY20000 cho2 (BY4742 cho2::KanMX4)EuroSCARFY14787 cho2 (BY4741 cho2::KanMX4)EuroSCARFY04787 cho2opi3 (co MATα) (BY4742 cho2::KanMX4; opi3::LEU2)Boumann et al (2004)N/A cho2opi3 (co MATa ) (BY4741 cho2::KanMX4; opi3::LEU2)This paperN/A co diploid (co 2n) (MATa /α; his3Δ1/his3Δ1; leu2Δ0/leu2Δ0; LYS2/lys2Δ0; MET15/met15Δ0; ura3Δ0/ura3Δ0; cho2::KanMX4/cho2::KanMX4; opi3::LEU2/opi3::LEU2)This paperN/A co lro1 (co MATα lro1::HIS3Sp )This paperN/A coS#2, S#3, S#4, S#5This paperN/A co MATα conditional CEN15 (co MATα CEN15::pGAL1‐CEN15‐URA3Kl )This paperN/A co diploid conditional CEN15 (co diploid CEN15/CEN15::pGAL1‐CEN15‐URA3Kl )This paperN/A co S(2n‐1) (co diploid CHR XV/chr XV)This paperN/A co acc1 (co MATα ACC1::acc1 )This paperN/A co acc1 lro1 (co MATα ACC1::acc1 lro1::HIS3Sp )This paperN/A co diploid ACC1/acc1 (co diploid ACC1/ACC1::acc1 )This paperN/A co diploid acc1 /acc1 (co diploid ACC1::acc1 /ACC1::acc1 )This paperN/A co dga1 (co MATα dga1::HIS3Sp )This paperN/A co pox1 (co MATα pox1::HIS3Sp )This paperN/ASH80 (MATα thr)Henry LabN/ASH85 (MATa thr)Henry LabN/A Recombinant DNA Plasmid: pCEN15‐UG (AmpR, left of CEN15‐URA3Kl ‐pGAL1‐right of CEN15)Reid et al (2008)N/APlasmid: pCRCT (2μ, AmpR, URA3, pTEF1‐iCas9)Addgene, Bao et al (2015)#60621Plasmid: pMEL16 (2μ, AmpR, HIS3, gRNA‐CAN1.Y)EuroSCARF, Mans et al (2015)P30785Plasmid: pFA6a‐HIS3MX6 (pBR322 origin, AmpR, HIS3Sp , T7 promoter)Longtine et al (1998)N/APlasmid: pBY011‐LRO1 (pLRO1) (2μ, AmpR, URA3, pGAL1‐10‐LRO1)DNASUScCD00094595Plasmid: pHEYg‐1 (pEV) (2μ, AmpR, URA3, KanMX4, TEF1 promoter)Natter LabN/APlasmid: pHEYg‐1‐GPT2 (pGPT2) (2μ, AmpR, URA3, KanMX4, pTEF1‐GPT2 [codon‐optimized])Natter LabN/APlasmid: pHEYg‐1‐DGA1 (pDGA1) (2μ, AmpR, URA3, KanMX4, pTEF1‐DGA1 [codon‐optimized])Natter LabN/A Oligonucleotides Primers for PCR see Appendix Table S2 This paperN/A Isotope labeled chemicals [P]orthophosphatePerkin ElmerNEX053[1‐C] acetatePerkin ElmerNEC084 Chemicals, Peptides, and Recombinant Proteins Choline‐ and inositol‐free agarSigma‐Aldrich05038Choline chlorideSigma‐AldrichC1879Myo‐inositolSigma‐Aldrich57570Boric acidSigma‐AldrichB7901HexaneHoneywell6505523‐amino‐1‐propanol (Prn)Sigma‐Aldrich239844Paraformaldehyde (PFA)Sigma‐Aldrich441244Glutaraldehyde (GA)Polyscience Inc00216Ergosterolester (C13:0)Thiele LabN/ARNAse ASigma‐Aldrich10109169001Proteinase KSigma‐Aldrich1245680100Power SYBR Green PCR Master MixThermoFisher Scientific1901521TaqMan Universal PCR Master Mix, no AmpErase UNGThermoFisher Scientific4324018Oligo(dT)12‐18 primerThermoFisher Scientific18418012Sytox GreenThermoFisher ScientificS70205‐fluoroorotic acid (5‐FOA)MELFORDF5001C‐LaurdanTOCRIS7273Soraphen AHelmholtz Centre for Infection Research Braunschweig‐FRGN/A Psp1406I/AclIThermoFisher ScientificER0942DpnIThermoFisher ScientificER1702RNase‐Free DNase SetQIAGEN79254Genomic DNA Buffer SetQIAGEN19060Qiaquick Gel Extraction KitQIAGEN28706SuperScript III Reverse TranscriptaseThermoFisher Scientific18080085RNeasy Mini KitQIAGEN74106TaqMan Gene Expression Assay (FAM) ACT1 ThermoFisher ScientificSc04120488_s1TaqMan Gene Expression Assay (FAM) INO1 ThermoFisher ScientificSc04136910_s1TaqMan Gene Expression Assay (FAM) KAR2 ThermoFisher ScientificSc04135107_s1TaqMan Gene Expression Assay (FAM) ACC1 ThermoFisher ScientificSc04161658_s1TaqMan Gene Expression Assay (FAM) POX1 ThermoFisher ScientificSc04110462_s1FAME standardNu‐Chek‐Prep63‐B Software and Algorithms ViiA7ThermoFisher ScientificViiA 7 Real‐Time PCR SystemFlowJoTree Star Inc https://www.flowjo.com/ ImageQuant TL8.1GE Healthcare Life Sciences https://www.gelifesciences.com/en/us/shop/protein‐analysis/molecular‐imaging‐for‐proteins/imaging‐software/imagequant‐tl‐8‐1‐p‐00110 IMODUniversity of Colorado https://bio3d.colorado.edu/imod/ REVIGOSupek et al (2011) http://revigo.irb.hr/ Freec v7.2 https://github.com/UMCUGenetics/IAP/releases/tag/v2.5.1 Prism 8GraphPad https://www.graphpad.com/ All yeast strains used in this study are listed in the Reagents and Tools Table. Synthetic defined medium (Griac et al, 1996) was supplemented with or without 1 mM choline (C) as indicated. Inositol (I) was supplemented at 75 μM inositol where indicated. 2% (w/v) glucose (SD), 3% (v/v) glycerol (SG), 2% (w/v) galactose (SGal) or mixtures of glucose and galactose were added as carbon source as indicated. YPD and YPGal medium contained 10 g/l yeast extract, 20 g/l bacto‐peptone and 20 g/l glucose and galactose, respectively. Solid media contained 2% (w/v) choline‐ and inositol‐free agar (Sigma‐Aldrich). 3‐amino‐1‐propanol (Prn) was added from a 1 M stock solution in water adjusted to pH 7.4, to a final concentration of 1 mM. Soraphen A was dissolved in ethanol (2 mg/ml) and added at the concentrations indicated. Strains were cultured at 30°C unless indicated otherwise. Optical density at 600 nm (OD600) was measured with a Hitachi U‐2000 double‐beam spectrophotometer. The WT strain was pre‐cultured in SD without choline (C), whereas cho2opi3 and derived triple mutants were pre‐cultured in SD with choline (C) to the mid‐log phase of growth (OD600 ˜0.45–1.2). Cells transformed with plasmids were grown in selective SD drop‐out media. To deplete phosphatidylcholine, cho2opi3 cells were collected by centrifugation or filtration where indicated, washed thoroughly with pre‐warmed SD C (30°C), and transferred to fresh SD C at the initial OD600 indicated (Boumann et al, 2006). SH80 and SH85 were pre‐cultured in YPD. Doubling time (DT) was determined based on OD600 values at different time points according to: DT = ln2/μmax with μmax the growth rate during exponential growth, μmax = (lnOD600 ‐ lnOD600 )/(t 2‐t 1). The cho2opi3 double mutant was pre‐cultured in SD C to late log phase (OD600 ˜1.5), washed twice with sterile water and resuspended in water at OD600 1. 100 μl of the suspension was spread on a choline‐free SD plate. After 14 days at 30°C, 10 cho2opi3 suppressor strains, cho2opi3 S#2 to S#11, were obtained. After double confirmation of growth on C plates at 30°C, the suppressor strains were frozen as glycerol stocks in SD C. cho2opi3S strains were maintained in SD C medium. After pre‐culture in choline‐free SD to OD600˜1, cho2opi3 S#4 was transferred to SD medium with choline (C) at 30°C with daily transfer to fresh medium for up to 40 days. Samples were taken every day and frozen as glycerol stocks. After collecting all glycerol stocks, cells were streaked on C and incubated at 30°C for 3 days. Single colonies were inoculated in C liquid medium and tested as indicated. Cells were pre‐cultured to mid‐log phase (OD600 ˜0.45–1.2) in SD as above, harvested by centrifugation and washed twice with sterile MQ water. The cells were adjusted to OD600 1.0, and serially diluted in 10‐fold increments to 10. 8 μl aliquots of each dilution were spotted onto agar plates containing the medium and supplements indicated. Plates were incubated at 30°C for the number of days indicated. Microbiological techniques followed standard procedures. Diploid cho2opi3 (co) strains were obtained by crossing co MAT a with co MATα obtained by deleting the OPI3 gene in cho2 strains as described (Boumann et al, 2004) and selection on lys met drop‐out medium. The LRO1, DGA1 and POX1 genes were deleted by standard PCR‐based homologous recombination replacing the respective ORFs with the Sp_HIS3 cassette from the plasmid pFA6a‐HIS3MX6 (Sikorski & Hieter, 1989), using the primers listed in Appendix Table S2. Correct integration was verified by colony PCR using primers A and D flanking the regions of LRO1, DGA1 and POX1 homology, and primers B and C internal to the Sp_HIS3 coding region (Appendix Table S2). The genomic single nucleotide mutation acc1 was introduced by CRISPR/Cas9 according to published methods (Mans et al, 2015). The pMEL16 backbone containing the 2μ origin of replication and HIS3 marker was amplified by PCR using conditions and primers as described (Mans et al, 2015). The gRNA insert primers listed in Appendix Table S2 were designed with the Yeastriction tool (yeastriction.tnw.tudelft.nl). To obtain the double strand gRNA insert, the complementary gRNA insert primers were dissolved in water to a final concentration of 100 μM, mixed in a 1:1 volume ratio, heated at 95°C for 5 min and cooled down to room temperature. The complementary 120 bp repair DNA primers (Appendix Table S2) were designed to replace adenosine (A) at position 657039 of chromosome XIV with cytosine (C) and converted to double stranded repair DNA as above. co MATα transformed with pCRCT was co‐transformed with 100 ng linearized pMEL16 backbone, 300 ng dsgRNA and 1 µg repair DNA fragment as described (Mans et al, 2015), followed by selection on a ura his plate. The correct genetic modification was verified by restriction analysis and by partial sequencing of a 1,725 bp DNA fragment (Chr XIV 656333‐658057) containing the mutation, that was obtained by PCR amplification of genomic DNA using the primers listed (Appendix Table S2). Digestion of the PCR mixture with Psp1406I/AclI according to the manufacturer’s instructions confirmed the loss of a restriction site. The purified PCR product was partially sequenced by Eurofins (Ebersberg, Germany) with the forward primer (Appendix Table S2) that starts at position 657469 of the Crick chain of chromosome XIV. All primers were obtained from IDT (Leuven, Belgium). Plasmids were removed by culturing the cells in liquid SD medium containing 20 μg/ml uracil and 20 μg/ml histidine for 7 days/passages and verified by growth on the non‐selective (ura his) plate but not on selective (ura and his) medium. The loss of one copy of chromosome XV from a diploid cho2opi3 strain was induced using a conditional centromere as described (Reid et al, 2008). Briefly, plasmid pCEN15‐UG was digested with NotI (Thermo Fisher Scientific) to liberate the integrating fragment containing the CEN locus of Chr XV interrupted by the K. lactis URA3 gene and the GAL1 promoter. The fragment was transformed into co MATα, generating co MATα CEN15::pGal1‐CEN15‐Kl_URA3 by homologous recombination (Guthrie & Fink, 1991). Transformants containing the conditional centromere were verified by PCR using the primers listed in Appendix Table S2. co MATα CEN15::pGal1‐CEN15‐Kl_URA3 was crossed with co MATa and diploids were selected on lys met medium as above. The conditional centromere was destabilized by plating on YPGal, and subsequently, a copy of chromosome XV was removed by counter‐selection against the URA3 gene by replica‐plating on SD medium containing 1 mg/ml 5‐fluoroorotic acid (5‐FOA) from a 100 mg/ml stock in DMSO. Plasmids were transformed according to the high efficiency transformation protocol (Guthrie & Fink, 1991). When removing the URA3 gene by 5‐FOA, strain growth was rescued with 20 μg/ml uracil. Diploid strains, cultured to late log phase in SD C or C, were transferred to 1% (w/v) potassium acetate, 0.005% (w/v) zinc acetate, and incubated in a shaking incubator at room temperature. After 7 days, cultures were examined by phase contrast microscopy. Samples for Fluorescence‐Activated Cell Sorting (FACS) were prepared according to published protocols (Haase & Reed, 2002; Pavelka et al, 2010). 1 × 10 yeast cells (corresponding to 0.5 OD600 units) were harvested and fixed with 1 ml 70% cold ethanol while spinning at 4°C overnight. After washing with 200 mM Tris–HCl pH 8.0, 2 mM EDTA and a second wash with 100 mM Tris–HCl pH 8.0, 2 mM EDTA, the cells in 1 ml 100 mM Tris–HCl pH 8.0, 2 mM EDTA were sonicated for 10 min in an ice bath (Branson 3800), and then resuspended in 1 ml RNAse solution (1 mg/ml RNAse A in 50 mM Tris–HCl pH 8.0, boiled for 15 min and allowed to cool to room temperature) for 6 h at 37°C at 800 rpm (shaking incubator, Eppendorf). After RNAse A was removed by centrifugation, samples were collected and incubated in 500 μl of proteinase K solution (100 µg/ml proteinase K in 50 mM Tris–HCl pH 8.0, 2 mM CaCl2) at 55°C at 800 rpm overnight. The samples were washed with 1 ml 50 mM Tris pH 8.0, 10 mM EDTA, followed by another washing step with 1 ml 50 mM Tris pH 8.0, 1 mM EDTA and finally resuspended in 1 ml 1 µM Sytox Green in 50 mM Tris–HCl pH 7.5. After sonication for 10 min in an ice bath, samples were analysed with a FACSCalibur flow cytometer (Becton Dickinson, NJ) with excitation at 488 nm and sorted based on area (DNA‐A) of the Sytox Green fluorescence signal, which was collected in the FL1 channel (530 ± 15 nm). Data were plotted as histograms showing the fluorescence distribution with FlowJo Software. Genomic DNA was isolated from 2 clones of each strain corresponding to 50 OD600 units with the Genomic DNA Buffer Set and Genomic‐tip 100/G (QIAGEN) according to the manufacturer’s instructions. Briefly, cell pellets were resuspended in Y1 buffer containing zymolyase and incubated at 30°C at 200 rpm for 1 h. Spheroplasts were pelleted, resuspended with G2 buffer containing 2 μg/ml RNAse A and 20 μg/ml proteinase K, and incubated in a water bath at 50°C for 1 h. After centrifugation at 3,000 g, the supernatant was collected and loaded on Genomic‐tip 100/G which was equilibrated with buffer QBT. After washing twice with buffer QC, gDNA was eluted with 5 ml buffer QF and then precipitated with 3.5 ml cold isopropanol at −20°C overnight. The gDNA was spun down at 10,000 g at 4°C and washed with 70% cold ethanol. The gDNA dissolved in 50 μl 10 mM Tris–HCl pH 8.0, 1 mM EDTA was used for whole genome sequencing or qPCR analysis. The concentration of gDNA was determined by measuring absorbance using NanoDrop (Thermo Fisher Scientific). Sequencing libraries with a mean insert size of 350 bp were constructed from 50 ng of genomic DNA and sequenced with paired‐end (2 × 250 bp) runs using an Illumina MiSeq instrument and V2 reagent kit to a minimal depth of 25× base coverage. Final library concentrations were measured using Qubit (Thermo Fisher Scientific). The alignment of sequencing reads was done using Burrows Wheeler Alignment (bwa 0.7.5a) (Li & Durbin, 2009) with settings “bwa mem ‐c 100 ‐M” against the S. cerevisiae reference genome (S288C version R64‐1‐1). Mapped reads were sorted, and duplicate reads were marked using the Sambamba v0.5.8 toolkit. Single nucleotide variants (SNVs), insertions and deletions (InDels) were called using GenomeAnalysisTK v3.4.46 HaplotypeCaller. Variants present in the parent cho2opi3 strain were filtered out from all samples, as were variants between the two clones of one strain. SNVs were assigned a “FILTER” flag using GenomeAnalysisTK v3.4.46 VariantFiltration using settings "‐‐filterName SNP_LowQualityDepth ‐‐filterExpression 'QD < 2.0' ‐‐filterName SNP_MappingQuality ‐‐filterExpression 'MQ < 40.0' ‐‐filterName SNP_StrandBias ‐‐filterExpression 'FS > 60.0' ‐‐filterName SNP_HaplotypeScoreHigh ‐‐filterExpression 'HaplotypeScore > 13.0' ‐‐filterName SNP_MQRankSumLow ‐‐filterExpression 'MQRankSum< −12.5' ‐‐filterName SNP_ReadPosRankSumLow ‐‐filterExpression 'ReadPosRankSum < −8.0' ‐‐clusterSize 3 ‐clusterWindowSize 35". For further analysis, only SNVs with the “FILTER” flag set to “PASS” were considered. Copy number variation (CNV) was determined based on read depth using freec v7.2, the GC normalized ratios were used to construct karyotypes using a 5 kb sliding window. All code used for these analyses is freely and openly available on github (github.com/UMCUGenetics/IAP/release/tag/v2.5.1). gDNA was isolated from cells cultured to OD600˜1 as described above and diluted to 50 pg/μl. Primer pairs targeting intergenic regions proximal (≤ 25 kb) to the centromere within each arm of 5 chromosomes (chr01La, chr01Ra, chr04La, chr04Ra, chr06La, chr06Ra, chr09La, chr09Ra, chr15La, chr15Ra) designed as described (Pavelka et al, 2010), were obtained from IDT (Leuven, Belgium), and dissolved at 3.2 μM in water. The qPCR reactions were performed in 96‐well plates using a ViiA™ 7 Real‐Time PCR System (Applied Biosystems, Thermo Fisher Scientific) in 20 μl reaction volumes. All reactions were set up in technical duplicates. Each reaction contained 10 μl Power SYBR Green PCR Master Mix (Applied Biosystems), 2.5 μl 3.2 μM forward primer, 2.5 μl 3.2 μM reverse primer and 5 μl 50 pg/μL gDNA. The cycling conditions were 50°C for 2 min, 95°C for 10 min, followed by 40 cycles of 95°C for 15 s and 60°C for 1 min (Pavelka et al, 2010). Melting curves were recorded (15 s at 95°C, 1 min at 60°C, followed by a gradual increase to 95°C in 15 s) to verify that no side products had been amplified. Ct values were determined using ViiA™7 Software. Chromosome copy numbers were calculated as described previously using BY4742 as WT euploid reference strain (Livak & Schmittgen, 2001; Pavelka et al, 2010). Yeast total lipid extracts were prepared by the 2‐phase lipid extraction method (Houtkooper et al, 2006), unless indicated otherwise. Briefly, cells corresponding to 200 OD600 units were lyophilized. The dry cell powder was resuspended in 6 ml chloroform: methanol (2:1, v/v) with 0.1 ml 0.1 M HCl, and sonicated for 20 min in a Branson B1200 bath sonicator (Bransonic Ultrasonics, Danbury, CT) containing ice water. Subsequently, 1 ml water was added to induce phase separation. After vigorous vortexing, cell debris was removed by centrifugation at 3,000 g for 4 min. The organic phase was collected, and residual lipid in the aqueous phase was extracted again with 3 ml chloroform: methanol (2:1, v/v). After washing with water, the combined organic phase was dried under a stream of nitrogen. The phospholipid content of lipid extracts was determined as described (Rouser et al, 1970) using KH2PO4 as a standard after destruction in 70% HClO4 at 180°C for 2 h. Separation of phospholipids by 2D‐TLC Total lipid extracts corresponding to 200 nmol phospholipid were applied on silica gel plates (Merck 1.05641), freshly impregnated with 2.4% (w/v) boric acid (de Kroon et al, 1997). The eluent for the first dimension contained chloroform: methanol: 25% ammonia (71:30:4, v/v/v). After drying under a flow of nitrogen for 30 min, the plate was run in the second dimension using chloroform: methanol: acetic acid (70:25:10, v/v/v) as eluent. The lipid spots were visualized by iodine staining. Spots were scraped off, and phospholipid classes were quantified (Rouser et al, 1970). Separation of neutral lipids by 1D‐TLC Total lipid extracts corresponding to 10 nmol phospholipid were applied on a silica gel plate. The neutral lipids were separated using hexane: diethylether: acetic acid (35:15:1, v/v/v) as eluent (Schneiter & Daum, 2006). Ergosterol obtained from Sigma‐Aldrich (E‐6625), cholesterol ester, monoacylglycerol (MAG), diacylglycerol (DAG), triacylglycerol (TAG) and free fatty acid obtained from Nu‐Chek‐Prep (Elysian, MN) were used as standards. The lipid spots were visualized by MnCl2 charring. 10 OD600 units of mid‐log phase cells pre‐cultured as above were harvested, washed and resuspended in 2.5 ml SD C medium. After 30 min at 30°C in a shaking incubator, [P]orthophosphate was added to 100 μCi/ml. After 30 min, the incubation was ended by adding 5% (w/v) TCA and putting the samples on ice. Cells were washed with water twice, homogenized by vortexing in the presence of glass beads for three times 1.5 min with intermittent cooling on ice. After adding HCl to 0.1 M, lipids were extracted (Bligh & Dyer, 1959). To determine the distribution of the lipid‐incorporated [P]‐label over the phospholipid classes, the lipid extracts were analysed by 2D‐TLC as above. Radioactive spots were detected using a Typhoon FLA7000 PhosphorImager (GE Healthcare Life Sciences) and quantified by ImageQuant TL8.1 software. For labelling with [C]acetate, 10 OD600 units of mid‐log cells cultured in SD C medium were harvested, washed and resuspended in 5 ml of the corresponding medium. After 30 min at 30°C, [1‐C]acetate was added to 2 μCi/ml, and the incubation was continued for 60 min. Next, samples were processed and lipids extracted as above. The [C]‐label incorporated into lipids was quantitated by liquid scintillation counting and normalized to phospholipid‐phosphorus content. Total lipid extracts corresponding to 100 nmol of phospholipid phosphorus were transesterified in 3 ml methanol containing 2.5% (v/v) sulfuric acid at 70°C for 2.5 h. After cooling to room temperature, 2.5 ml water and 2.5 ml hexane were added. The hexane phase was collected, and the aqueous phase was washed with another 2.5 ml hexane. After washing the pooled hexane phase at least three times with water to remove residual sulfuric acid, 100 µl isopropanol was added, and the samples were dried under nitrogen gas. 100 µl of hexane was added to the fatty acid methyl esters (FAME). FAME were analysed by Gas Chromatography‐Flame Ionisation Detection (GC‐FID) on a Trace GC Ultra (Thermo Fisher Scientific) equipped with a biscyanopropyl polysiloxane column (Restek, Bellefonte PA) using nitrogen as carrier gas and a temperature gradient from 160 to 220°C. Peak identification and calibration of the integrated signal intensities were performed using a FAME standard. Cells were harvested at mid‐log phase (OD600 ˜0.3‐1) and washed twice with 150 mM ice cold ammonium bicarbonate (ABC) buffer. Cells corresponding to ˜10 OD600 units were resuspended in 0.5 ml ABC, and vortexed vigorously in the presence of 200 μl glass beads for 2 × 5 min at 4°C with intermittent cooling on ice for 3 min. The lysates were frozen directly in liquid nitrogen and stored at −80°C until further processing. Mass spectrometry‐based lipid analysis was performed by Lipotype GmbH (Dresden, Germany) as described (Ejsing et al, 2009; Klose et al, 2012). Lipids were extracted using a two‐step chloroform/methanol procedure (Ejsing et al, 2009). Samples were spiked with internal lipid standard mixture containing: CDP‐DAG 17:0/18:1, ceramide 18:1;2/17:0 (Cer), diacylglycerol 17:0/17:0 (DAG), lysophosphatidate 17:0 (LPA), lyso‐phosphatidylcholine 12:0 (LPC), lysophosphatidylethanolamine 17:1 (LPE), lyso‐phosphatidylinositol 17:1 (LPI), lysophosphatidylserine 17:1 (LPS), phosphatidate 17:0/14:1 (PA), phosphatidylcholine 17:0/14:1 (PC), phosphatidylethanolamine 17:0/14:1 (PE), phosphatidylglycerol 17:0/14:1 (PG), phosphatidylinositol 17:0/14:1 (PI), phosphatidylserine 17:0/14:1 (PS), ergosterol ester 13:0 (EE), triacylglycerol 17:0/17:0/17:0 (TAG), stigmastatrienol, inositolphosphorylceramide 44:0;2 (IPC), mannosylinositolphosphorylceramide 44:0;2 (MIPC) and mannosyl‐di‐(inositolphosphoryl)ceramide 44:0;2 (M(IP)2C). After extraction, the organic phase was transferred to an infusion plate and dried in a speed vacuum concentrator. 1 step dry extract was resuspended in 7.5 mM ammonium acetate in chloroform/methanol/propanol (1:2:4, v/v/v) and 2 step dry extract in 33% ethanol solution of methylamine in chloroform/methanol (0.003:5:1, v/v/v). All liquid handling steps were performed using Hamilton Robotics STARlet robotic platform with the Anti Droplet Control feature for organic solvents pipetting. Samples were analysed by direct infusion on a QExactive mass spectrometer (Thermo Scientific) equipped with a TriVersa NanoMate ion source (Advion Biosciences). Samples were analysed in both positive and negative ion modes with a resolution of Rm/z=200 = 280,000 for MS and Rm/z=200 = 17,500 for MSMS experiments, in a single acquisition. MSMS was triggered by an inclusion list encompassing corresponding MS mass ranges scanned in 1 Da increments (Surma et al, 2015). Both MS and MSMS data were combined to monitor EE, DAG and TAG ions as ammonium adducts; PC as an acetate adduct; and CL, PA, PE, PG, PI and PS as deprotonated anions. The absence of PC in cho2opi3 suppressor strains was confirmed in positive ion mode by MSMS scan for the phosphocholine headgroup fragment. MS only was used to monitor LPA, LPE, LPI, LPS, IPC, MIPC, M(IP)2C as deprotonated anions; Cer and LPC as acetate adducts and ergosterol as protonated ion of an acetylated derivative (Liebisch et al, 2006). Data were analysed with in‐house developed lipid identification software based on LipidXplorer (Herzog et al, 2011, 2012). Data post‐processing and normalization were performed using an in‐house developed data management system. Only lipid identifications with a signal‐to‐noise ratio > 5, and a signal intensity 5‐fold higher than in corresponding blank samples were considered for further data analysis. Since the TAG species cannot be unambiguously assigned because of the presence of 3 FAs (is combinatorically not possible), the acyl chain distribution of the TAG fraction was quantitated as follows. TAG precursors were fragmented and the FAs released identified. The profile of FAs determined for each TAG precursor was normalized to the abundance of the precursor, yielding pmol values for each FA, which were summed up to obtain the FA profile for TAG. In quantifying the lipidomics data, molecular species containing acyl chains with an odd number of C atoms or a number of C atoms larger than 18 that constituted < 3 and < 0.5% of total, respectively, were left out. The preparation of samples for EM analysis was performed as described previously (Griffith et al, 2008; Mari et al, 2014). Briefly, after harvest the cells were rapidly mixed with an equal volume of double strength fixative [4% (w/v) paraformaldehyde (PFA), 0.4% (v/v) glutaraldehyde (GA) in 0.1 M PHEM (20 mM PIPES, 50 mM HEPES, pH 6.9, 20 mM EGTA, 4 mM MgCl2)], and incubated for 20 min at room temperature on a roller bank. The mixture fixative‐media was replaced by 2% PFA, 0.2% GA in 0.1 M Phem pH 6.9 for 2 h at RT followed by an overnight incubation at 4°C. Fixative was removed by centrifugation, and cells were embedded in 12% gelatin. Blocks of 1 mm were obtained and incubated in 2.3 M sucrose overnight at 4°C before being mounted on pins. Cells sections were obtained using a LEICA cryo‐EM UC7. Membrane contrast was performed as described previously (Griffith et al, 2008; Mari et al, 2014). Thin sections (50 nm) were viewed in an electron microscope (1200 EX; JEOL). The 2D projection images (Tiff format) of non‐overlapping regions in the cryosection were imported into the IMOD software package. Cell area and lipid droplet area were determined in at least 15 2D projection images by counting the total number of pixels covering yeast cells and the total number of pixels covering the lipid droplets. Lipid droplet content is expressed as the area occupied by lipid droplets determined as a percentage of total cell area and was calculated using Excel. Significance was determined with Student’s t‐test. Wild‐type BY4742, co S#3, S#4 and S#5 were pre‐cultured in SD medium (C) to mid‐log phase (OD600 ˜1.0) and transferred to 15 ml of the corresponding medium at OD600 of 0.05. The co MATα parent strain was pre‐cultured to OD600 ˜1.0 in SD medium containing 1 mM choline; cells were collected by filtration, washed with choline‐free SD medium (30°C) and used to inoculate 15 ml fresh SD C medium to an OD600 of 0.1 (Boumann et al, 2006). All strains were cultured at 30°C in biological replicate and harvested in early mid‐log phase (OD600 of 0.55–0.65) by centrifugation at room temperature for 3 min. Time from removing culture from incubator until freezing pellet is maximally 5 min. Total RNA was isolated by phenol extraction and purified as described (Kemmeren et al, 2014). All subsequent procedures in expression‐profiling including RNA amplification, cRNA labelling, microarray hybridization, quality control and data normalization were carried out as described previously (Kemmeren et al, 2014). Two channel microarrays were used. RNA isolated from WT was used in this common reference design, in one of the channels for each hybridization. Two independent cultures were hybridized on two separate microarrays. For the first hybridization, the Cy5 (red) labelled cRNA from the mutant was hybridized together with the Cy3 (green) labelled cRNA from the common reference. For the replicate hybridization from the independent cultures, the labels were swapped. The reported fold change is the average of the four replicate mutant profiles versus the average of the WT controls. Genes were considered significantly changed when the fold change (FC) was > 1.7 and the P < 0.01. P values were obtained from the limma R package version 2.12.0 (Smyth, 2005) after Benjamini–Hochberg FDR correction. Hierarchical clustering of genes subject to significant expression changes was by average linkage (cosine correlation). Functional enrichment was by a hypergeometric test on Gene Ontology Biological Process (GO‐BP; P < 0.01, Bonferroni corrected). Enriched GO terms were summarized by the REVIGO software using a cut‐off value C of 0.5 (Supek et al, 2011). Diploid yeast strain BY4743, co diploid, co S(2n‐1) and co S#2 were pre‐cultured in SD CI to log phase (OD600 ˜0.45–1.2). After washing with pre‐warmed SD CI (30°C) by filtration, cells were rapidly transferred to CI and cultured for 24 h. Total RNA was isolated from 20 OD600 units after cell disruption by rapid agitation in the presence of glass beads and lysis buffer, using the RNeasy Mini Kit and RNase‐free DNase Set according to the manufacturer’s instructions (QIAGEN). RNA quality and quantity were checked on a 1% agarose gel and with a NanoDrop spectrophotometer (ND‐1000, Thermo Fisher Scientific), respectively. 1 µg RNA was converted to cDNA according to the first strand cDNA synthesis protocol of Invitrogen™ (Thermo Fisher Scientific) using SuperScript™ III reverse transcriptase. 1 µl oligo(dT)12–18 Primer, 1 µl 10 mM dNTP and 1ug RNA were mixed in a PCR tube, and water was added to adjust the total volume to 14 µl. Mixture was heated at 65°C for 5 min and incubated on ice for 1 min. 4 µl 5× First Strand Buffer, 1 µl 0.1 M DTT and 1 µl SuperScript III Reverse Transcriptase was added into the mixture gently, followed by an incubation at 50°C for 60 min. The reaction was inactivated by heating at 70°C for 15 min and 80 µl MQ water was added. qPCR was performed in technical duplicate as described above in 96‐well plates with 20 µl reactions containing 10 µl TaqMan Universal PCR Master Mix, no AmpEraseTM UNG (Applied Biosystems), 1 µl TaqMan™ probes and primers (ACT1, ACC1, POX1, KAR2, INO1, Reagents and Tools Table), 4 µl MQ water and 5 µl cDNA (from 50 ng RNA). Non‐template control (cDNA) and non‐reaction control were routinely performed. The data were analysed according to the ΔΔCt method (Livak & Schmittgen, 2001), normalized to the control gene ACT1 and expressed relative to the corresponding strain cultured in CI. 6 × 10 yeast cells, cultured in SD C medium until log phase, were incubated in 1 ml PBS buffer with 5 µM C‐Laurdan at 30°C for 1h. Cells were collected by centrifugation and resuspended in 1 ml PBS buffer. 150 µl of 1.5% pre‐heated agarose gel and 150 µl cell suspension were mixed in the wells of µ‐slide 8 well IBIDI® slides. The slide was centrifuged at 1,000 g for 5 min to assure that cells were on the bottom of the slide while the gel solidifies. Cells were then directly imaged by a Zeiss LSM880 airyscan confocal microscope at 405 nm excitation, and emission was recorded in two channels: I 1 = 420–475 nm and I 2 = 480–570 nm. Generalized polarization (GP) images were calculated by formula 1: GP=I1‐GI2I1+GI2, in which G is a calibration factor depending on acquisition settings. To display the morphology of the cells, signal was also recorded in the transmission channel at the same excitation wavelength. Since to our knowledge this is the first time that yeast cells have been imaged by C‐Laurdan, we calibrated the G‐factor with a 5 µM C‐Laurdan solution in DMSO. The actual GP value of this solution was determined by recording a fluorescence emission spectrum at an excitation wavelength of 405 nm in a Cary Eclipse fluorescence spectrometer. For GP determination, fluorescence intensities at 440 (I1 ) and 490 nm (I2 ) were used in formula 1 using a G‐factor of one (GP = 0.232, SD = 0.003, n = 3). The same solution was further imaged under the same conditions as yeast cells, and the G‐factor was determined according to Gaus et al, (2006) and using the GP measured by spectroscopy as reference GP. GP‐images and histograms were processed as described earlier (Owen et al, 2011). For each experiment, the number of biological replicates is indicated in the corresponding figure legend and/or methods section. GraphPad Prism 8 was used to determine means, standard deviations and statistical significance, with P‐values determined by multiple t‐test using the Holm–Sidak method with alpha = 0.05, or one‐way ANOVA; *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.
|
PMC10630469
|
Q-RAI data-independent acquisition for lipidomic quantitative profiling
|
Untargeted lipidomics has been increasingly adopted for hypothesis generation in a biological context or discovery of disease biomarkers. Most of the current liquid chromatography mass spectrometry (LC–MS) based untargeted methodologies utilize a data dependent acquisition (DDA) approach in pooled samples for identification and MS-only acquisition for semi-quantification in individual samples. In this study, we present for the first time an untargeted lipidomic workflow that makes use of the newly implemented Quadrupole Resolved All-Ions (Q-RAI) acquisition function on the Agilent 6546 quadrupole time-of-flight (Q-TOF) mass spectrometer to acquire MS2 spectra in data independent acquisition (DIA) mode. This is followed by data processing and analysis on MetaboKit, a software enabling DDA-based spectral library construction and extraction of MS1 and MS2 peak areas, for reproducible identification and quantification of lipids in DIA analysis. This workflow was tested on lipid extracts from human plasma and showed quantification at MS1 and MS2 levels comparable to multiple reaction monitoring (MRM) targeted analysis of the same samples. Analysis of serum from Ceramide Synthase 2 (CerS2) null mice using the Q-RAI DIA workflow identified 88 lipid species significantly different between CerS2 null and wild type mice, including well-characterized changes previously associated with this phenotype. Our results show the Q-RAI DIA as a reliable option to perform simultaneous identification and reproducible relative quantification of lipids in exploratory biological studies.Lipids are a group of organic compounds with high hydrophobicity that play important biological roles, including cellular signalling, energy metabolism and as a structural component of cell membranes. Lipidomics entails comprehensive identification and robust quantitation of lipids using analytical chemistry methods such as liquid chromatography mass spectrometry (LC–MS) and it has matured as an independent field of study underscoring the importance of metabolism in disease progression. Typically, untargeted lipidomics focuses on the identification and relative quantitation of lipids for discovery approaches, while targeted lipidomics serves to identify and quantify a group or panel of known lipids with high sensitivity. In untargeted lipidomics there are two common analytical approaches. The first is direct-infusion MS (shotgun lipidomics), often performed using a nanospray source and a high-resolution mass spectrometer. Due to its speed, this is suitable for high-throughput studies and it has the advantage of allowing a reliable quantification, due to the same matrix effect experienced by all ions. A disadvantage may be lower sensitivity and higher complexity of fragmentation spectra when compared to LC–MS. LC–MS based methods can perform better in terms of species coverage and sensitivity due to the chromatographic separation. Lipids with similar m/z values can be distinguished based on different retention times. In all these experiments, tandem MS (MS/MS) is essential for reliable identification. There are two ways to acquire MS/MS data in untargeted lipidomics: data dependent acquisition (DDA) and data independent acquisition (DIA). In DDA a full MS scan is conducted to obtain abundance-based precursor information of lipids in the sample, followed by fragmentation of the precursors to obtain MS/MS spectra of product ions. However, DDA approaches are sometimes ineffective in fragmenting low abundance compounds due to the preference given to the most abundant ones. In addition, peak intensities of product ions are not quantitatively accurate as fragmentation is often triggered before the apex of the peak and there are insufficient data points along the chromatograms. The disadvantages of DDA are then low reproducibility and incomplete coverage. Recent advances in MS technologies have led to the development of DIA for comprehensive MS/MS data generation and targeted quantification at both precursor and product ion levels, thereby addressing some of the limitations of DDA, including biased selection of abundant ions for fragmentation. In DIA, all precursors within a selected mass range are scanned and consistently fragmented regardless of abundance. The major limitation of this technique is the high complexity of the spectra generated during fragmentation of co-eluting species, which makes lipid identification difficult. Ways of reducing this complexity include the use of well characterised LC separations and the use of consecutive quadrupole isolation windows. A selective DIA methodology that was first reported by Gillet et al. utilizes sequential quadrupole isolation windows for acquisition of all theoretical mass spectra (SWATH) in proteomics applications, where the mass range of interest is divided into several subranges (or isolation windows) for fragmentation. Generally, an overlap is included between windows, to ensure that all precursors in the desired mass range are selected for fragmentation. When DIA is applied in direct infusion shotgun lipidomics, utilizing the MS/MS approach that is available in orbitrap mass spectrometers and latest hybrid quadrupole time-of-flight (Q-TOF) technologies, all precursors in the sample are essentially isolated across the entire mass range of interest in 1 m/z isolation windows, followed by fragmentation. Isolation window sizes can be customized depending on the application, with key considerations being the abundance of precursors, coelution of analytes and chromatographic peak widths. Since all precursors can be fragmented, larger window sizes would result in more precursors simultaneously entering the collision cell for fragmentation and hence resulting in increased complexity of MS2 spectra. This causes issues particularly in high-throughput analyses, where numerous analytes co-elute and chromatographic peaks are narrow, with ion suppression and convoluted MS/MS with additional interferences that results in insufficient data points for accurate peak area calculation. The isolation window size and the number of windows is inversely related, where smaller window sizes would result in a larger number of windows to cover the entire mass range. These parameters also determine the total cycle time, which is derived from the MS1 and the MS2 acquisition rates and the number of windows. A general consensus is that there should be at least 10 data points (10 cycles) across the narrowest chromatographic peak for reliable quantification in chromatographic analyses. For this reason, 25 m/z wide windows are commonly used in proteomics, and 10–20 m/z windows in lipidomics, to ensure a good balance between purity of MS/MS spectra and quality of quantification. While the early generation of DIA methods was performed with fixed size windows, the need for reliable quantification has also spurred on the development of windows of variable sizes, optimized based on the precursor density across the entire mass range. Smaller window sizes would be adopted for m/z regions with a higher density of precursors, while larger window sizes would be adopted for regions where precursors are less present, thereby reducing the loss of sensitivity. This allows for the same total number of windows to be maintained, thereby ensuring that the cycle time is not compromised and enough data points can be captured for all ions of interest. The choice between DDA and DIA methodologies also has implications on the downstream data analysis approach. For lipidomic data acquired in DDA mode, software with lipid identification capabilities such as MS-Dial, Lipid Annotator, LipidMatch, LipiDex, LipidXplorer, LipidBlast and others have been developed. Most of these tools enable lipid identification by searching against a known database or spectral library using a combination of features such as retention time (RT), isotopic distribution, precursor mass and MS2 spectra. Others may use alternative approaches: MS-Dial utilises a decision tree algorithm within a rule-based annotation system, LipidXplorer implements user-defined fragmentation pathways described using the molecular fragmentation query language and Lipostar can work with a database-free rule-based fragmentation system. Most software can read data from a variety of LC–MS instruments and methods. Recently, we have reported a new data processing software package, MetaboKit, to support seamless integration of DDA and DIA for identification and relative quantification, respectively with potential advantages over the alternatives, including the explicit MS/MS-based annotation of in-source fragments (ISFs) and the ability for the user to build and annotate an in-house spectral library based on product ion spectra and retention times. The DIA module of MetaboKit performs a targeted extraction of MS2 fragments using spectral libraries obtained from DDA, which facilitates users to gradually build a customized database of MS2 spectra and decreases reliance on commercially or publicly available libraries for benchmarking in the long term. Meanwhile, LC-based DIA methodologies for lipidomics have been most actively developed on SCIEX Triple-TOF instruments, with standardized workflows for identification and quantification. Software packages such as MetDIA and MS-Dial have been developed to analyse the acquired MS2 complex data, where the fragmentation spectra are deconvoluted and benchmarked to spectral libraries in-built within the software. Spectral deconvolution strategies are essential in DIA as the direct link between precursor and product ion spectra is lost and data become more complex due to higher background signals in the obtained spectra. To optimize analytical methods, tools such as the SWATH Acquisition Variable Window Calculator by SCIEX, that calculate the best sequence of variable windows based on the density of precursors have been released. However, these approaches have not been extensively validated for other instruments. Herein, we report the development and application of Quadrupole-Resolved All Ions (Q-RAI) data from the Agilent 6546 Q-TOF, in combination with DDA-based library construction, as a novel alternative DIA analysis platform on Agilent LC–MS instruments. The proposed workflow is seamlessly supported by the MetaboKit software package. The Q-RAI acquisition mode is a SWATH-like data independent acquisition mode newly developed on the Agilent 6546 LC/Q-TOF, first reported for quantitative analysis of per- and polyfluoroalkyl substances (PFAS). However, this acquisition mode has not been tested for high throughput lipidomic applications and could provide an additional tool for simultaneous untargeted lipid identification and quantification at both MS1 and MS2 levels without significant loss in lipid coverage, particularly for users with this line of instruments. This DIA approach would be advantageous over conventional untargeted DDA workflows where MS2-based quantification is not applicable, especially for compounds in which the product ions offer quality ion chromatograms as good as or better than that of the precursor ion. In addition, it will also enable easier development of targeted assays from the product ion spectra of specific lipids of interest. To properly test this newly developed instrumental and data analysis workflow, we first analysed commercially available human plasma and then conducted a differential abundance analysis of murine serum samples from wild type and a Ceramide Synthase 2 (CerS2) null model. 1-Butanol was purchased from Merck (Darmstardt, Germany). Methanol, isopropanol and acetonitrile were purchased from Thermo Fisher Scientific (Waltham, Massachusetts, United States). Ammonium Formate (10M in water) was purchased from Sigma-Aldrich (St. Louis, Missouri, United States). All solvents were of HPLC grade. Pooled commercial human plasma obtained via EDTA whole blood from healthy subjects of European descent and mixed gender was purchased from Sera Laboratories International (West Sussex, United Kingdom). MilliQ water used for mobile phase preparation had a resistivity of 18 mΩ. Most internal standards were purchased from AVANTI (Alabaster, Alabama, United States), including ceramide (Cer) d18:1/12:0, Cer m18:1/12:0, cholesterol (COH)-d7, glucosylceramide (HexCer) d18:1/12:0, dihexosylceramide (Hex2Cer) d18:1/12:0, lysophosphatidylcholine (LPC) 13:0, lysophosphatidylethanolamine (LPE) 14:0, phosphatidylcholine (PC) 13:0/13:0, phosphatidylethanolamine (PE) 17:0/17:0, phosphatidylglycerol (PG) 17:0/17:0, phosphatidylinositol (PI) 12:0/13:0, phosphatidylserine (PS) 17:0/17:0, plasmalogen phosphatidylcholine (PC-P) 18:0/18:1-d9, plasmalogen phosphatidylethanolamine (PE-P) 18:0/18:1-d9 and sphingomyelin (SM) d18:1/12:0. Other standards include acylcarnitine 16:0-d3, triacylglycerol (TG) 12:0/12:0/12:0 and TG 17:0/17:0/17:0 from Sigma-Aldrich, ganglioside GM3 d18:0/18:1-d3 and globotriaosylceramide (Hex3Cer) d18:0/18:1-d3 from Cayman (Ann Arbor, Michigan, United States), cholesteryl ester (CE) 18:0-d6 from CDN isotopes (Quebec, Canada) and diacylglycerol (DG) 15:0/15:0 from Santa Cruz Biotech (Dallas, Texas, United States). CerS2 null mice were generated as previously described, in accordance with the Animal Research: Reporting of In Vivo Experiments (ARRIVE) guidelines. The experimental protocols were approved by the Weizmann Institute of Science’s Institutional Animal Care and Use Committee (IACUC), with animals treated according to the IACUC Animal Care Guidelines and the National Institutes of Health's Guidelines for Animal Care. Serum was isolated from wild type and CerS2 null mice at 26 weeks. Lipid extraction from commercial human plasma and mouse serum samples was performed using a variation of the Butanol-Methanol extraction method. Briefly, 180 µL of Butanol-Methanol (1:1, v/v) spiked with internal standards were added to 20 µL of sample. The resultant mixture was vortexed for 10 s and then sonicated in a water bath at 4 °C for 30 min. Centrifugation was subsequently performed at 14,000 rcf at 4 °C for 10 min. The supernatant was transferred into MS vials for LC–MS analysis. To develop the DIA method, the lipid extracts from human plasma samples were pooled, mixed and aliquoted into MS vials. Serum samples from five wild-type mice and five CerS2 null mice were used for DIA analysis. The sequence of serum samples for extraction and subsequent LC–MS analysis were obtained via random sampling on Microsoft Office Excel 2016. Aliquots from each serum extract were pooled and split across the analytical run for equilibration of the system (5 samples) and as technical quality control (QC) samples (6 samples analysed at regular intervals). Lipid extracts were analysed on an Agilent 1290 Infinity II LC System coupled to the Agilent 6546 LC/Q-TOF system using a ZORBAX Eclipse Plus, C18, 95 Å, 1,8 µm, 2,1 × 50 mm (Agilent). Samples for MRM targeted analysis were measured with the same chromatographic system connected to the Agilent 6495 LC/TQ. The Agilent Data Acquisition Software Version 10.1 was used for all LC–MS analyses. For the chromatographic runs, Water/Acetonitrile (6:4, v/v) + 10 mM Ammonium Formate was used as Mobile Phase A and Isopropanol/Acetonitrile (9:1, v/v) + 10 mM Ammonium Formate was used as Mobile Phase B, with an elution gradient starting at 0 min with 80% A and including: 2 min 40% A, 12 min 0% A, 14 min 80% A. The stop time was set at 15.8 min with a flow rate of 0.4 mL/min and a column temperature of 40 °C. An injection volume of 4 µL was used in positive mode and 8 µL in negative mode. For the dilution series, injection volumes of 0.25, 0.5, 1, 2, 4 and 8 µL were used. The ion source parameters for DDA and DIA were as follows: Gas temperature 325 °C, Drying gas 10L/min, Nebulizer 20 psi, Sheath gas temperature 350 °C, Sheath gas flow 12L/min, Capillary voltage 3500 V, Nozzle voltage 1000 V, Fragmentor voltage 180 V, Skimmer voltage 45 V and Octopole RF 750 V. For DDA, acquisition was set to auto MS/MS mode with a mass range of 300–1000 m/z for MS1 and 50–1000 for MS2 at acquisition rates of 4 spectra/s (Cycle Time 1.6 s). The iterative MS/MS function was used to increase the number of fragmented species and was set at 20 ppm tolerance with RT exclusion tolerance of 0.2 min. A narrow (~ 1.3 m/z) isolation width, collision energy of 20 V, 5 maximum precursors per cycle, precursor abundance-based scan speed of 10,000 counts/spectrum, 200 counts at 0.01% as the threshold for MS/MS, purity stringency of 100% and purity cut off of 30% were also used. Reference masses at 121.0599 m/z and 922.0098 m/z were used in positive mode and 112.9856 m/z and 966.0007 m/z in negative mode. Active exclusion was enabled for one repeat for 0.08 min, corresponding to about half peak width. Abundance dependent accumulation was used with a value of 10,000 counts with MS/MS maximum accumulation time limit set to the method cycle time and precursors that cannot reach target abundance were not rejected. Acquisition was set to data independent acquisition mode with a mass range of 300–1000 m/z for MS1 and 50–1000 for MS2 at acquisition rates of 20 spectra/s for MS1 and 40 spectra/s for MS2. A fixed collision energy of 20 V was used. The variable windows were obtained by analysing the acquired DDA data on the DIAwindow_calc module in the in-house developed MetaboKit software. The module first applies a feature detection step on the MS1 map to get a list of MS1 features. For each of these features, the density (number of features nearby) was calculated. The summed density for all features was approximately equal among all windows, subject to minimum and maximum window size allowable by the mass spectrometer. Finally, each window was extended by 0.5 m/z on both ends to ensure a 1 m/z overlap. The analysis was done using a window width of 10–100, 20 windows and mass range of 300–1000 m/z, with all other parameters kept as default. Fixed and variable windows values used during method development are detailed in Tables S1 and S2 and the variable windows used for the mouse study are reported in Table S4. Variable windows in positive mode as depicted in Table S2 were used for the dilution series. The triple quadrupole scan type was set to dynamic MRM with a delta EMV of 200 V (+) and 0 V (−) for analysis in positive mode. Time filtering was enabled with a peak width of 0.07 min and the cycle time was 900 ms. Ion source parameters were as follows: Gas temperature 250 °C, Gas flow 14 L/min, Nebulizer 35 psi, Sheath gas temperature 250 °C, Sheath gas flow 11 L/min, Capillary voltage 3000 V (+) and 4000 V (−), Nozzle voltage 1000 V (+ , −), High pressure RF 150 V (+ , −), Low pressure RF 60 V (+ , −). Raw DDA data files were analysed for identification of all lipid classes using settings as recommended by Koelmel et al.: Q-Score ≥ 30, Mass deviation ≤ 10 ppm, Fragment Score ≥ 30, Total Score ≥ 60, Report dominant constituent if relative abundance differential ≥ 10%. All other parameters were kept as default. Agilent raw data files were first converted using MSConverter to mzML format. MSConverter is an open-source tool for conversion of raw data files to mzML format for data processing and is available at https://proteowizard.sourceforge.io/download.html. DDA data were analysed on MetaboKit (2023/07/19 release) using the following settings: ms1to1 = 0.005 (amu), ms2to1 = 0.01 (Da), MS2_score = 0.5, min_peaks = 1 and RT_shift = 10. All other settings were kept as default, with MS/MS fragment matching fixed at 0.01 Da. Agilent raw data files were first converted to abf format using AnalysisBaseFileConverter. DDA data were analysed on MS-Dial Version 4.72, with MS1 Tolerance 0.01 Da, MS2 Tolerance 0.025 Da, MS1 mass range 300–1000 m/z and MS2 mass range 50–1000 m/z. All other parameters were kept as default. Raw DDA data files were directly analysed on Lipostar 2.1.2. Peak detection was done using a signal filtering threshold of 200 and m/z tolerance of 0.02 amu while all other parameters were kept as default. Identification was done using MS tolerance of 10 ppm and MS/MS tolerance of 20 ppm with all other parameters kept as default. For the manual curation and filtering of software outputs, only [M + H] and [M + NH4] adducts were considered in positive mode and [M–H] and [M + HCOO] adducts in negative mode. Manual curation of all identified lipid species was done based on RT and manual inspection of the quality of spectral matching. Duplicate IDs for the same lipid species were also removed from the final set of readouts. For ambiguous lipid classes or species identified by the software but not previously observed in-house, quality of the spectral matching with the library was used as the main criterion for evaluation. We exercised a more conservative approach whereby at least two matching MS/MS peaks are required, followed by verification through in-house targeted tandem high-resolution MS/MS of these species using the same extracts. Lipids identified by the software but not present in the sample according to this manual inspection process were deemed to have failed filtering. For Lipostar outputs, identified lipids with an isotopic pattern score of 0 were also considered to have failed the filtering process. Agilent raw data files were first converted to mzML format using MSConverter. DIA data were analysed with MetaboKit using the following settings: ms1to1 = 0.005, ms2to1 = 0.01, min_peaks = 1, RT_shift = 20. All other settings were kept as default, with MS/MS fragment matching of the software fixed at 0.01 Da. Raw DIA data files were directly analysed on Lipostar 2.1.2. Parameters for peak detection and identification were the same as those used for DDA. Identified lipids with an isotopic pattern score of 0 were deemed to have failed the filtering process. For both software, only [M + H] and [M + NH4] adducts were considered in positive mode and [M–H] and [M + HCOO] adducts in negative mode. Ambiguous species that were not identified in all triplicate injections, not previously observed and verified through in-house DDA or have doubtful RT were deemed to have failed the filtering process. Dilution series data were analysed with MassHunter Quantitative Analysis, Version 10.1 using a quantitative method specific for MRM data. Agile2 was used as the default integrator while for transitions with close multiple peaks, either General or Spectral Summation integrators were used. Manual inspection was also performed for all data following peak integration. For data analysis performed with MetaboKit, the HMDB, NIST, LipidBlast, Lipid Atlas and all publicly available libraries in the MS-Dial repository were used, together with an RT list and spectral library obtained from prior in-house DDA analysis. These libraries contained a total of 892,093 publicly available spectra in positive mode, 845,599 publicly available spectra in negative mode and 729 in-house spectra. However, all possible constituents for the same lipid sum composition are listed in the Lipid Atlas library in ascending numerical order and MetaboKit gives priority to the molecular composition with the highest match score, followed by the last entry listed in the spectral library when generating the outputs (e.g. SM 32:1 can be identified in negative mode based on the matching of the 659.5148 ion from formate loss and the 168.0416 phosphatidylcholine head group fragments but it is shown as SM 30:1;2O/2:0 in the output, which is the last entry for this particular sum composition in the Lipid Atlas library). Hence, all readouts were expressed as sum composition in the data for ease of interpretation. Readouts identified against the Lipid Atlas library had the sum composition computed using an in-house R script while readouts identified against the other libraries had the sum composition computed manually. Automated generation of sum composition lipid names by MetaboKit would be implemented in the foreseeable future but in the meantime, it is advisable for users to report the results at both levels to facilitate decision making with regard to the level of structural detail. For MS-Dial analysis, all publicly available libraries in the MS-Dial repository were used for identification. In-silico libraries used in Lipid Annotator were taken from the MS-Dial repository using specific algorithms for transfer and assessment of quality control. All calculations of R, plotting of heatmaps and analysis for the mice study were done in RStudio (https://www.rstudio.com). Dot plots and visualization of lipid class distributions were generated using Graphpad Prism Version 8.4.3. Data for all other figures and tables were processed and analysed using Microsoft Office Excel 2016. MetaboKit is an open-source command line tool publicly available at https://github.com/MetaboKit/MetaboKit. MS-Dial is also an open-source software. Conventional DIA workflows begin with MS/MS spectral library construction via DDA analysis. After generating an in-house MS/MS spectral library of lipids based on DDA acquisition of a commercial human plasma sample, we first tested the performance of different software packages for identification and quantification of lipids in our sample. Prior to data acquisition, we tested the impact of different collision energies (10 V, 20 V and 30 V) and observed that 20 V provided the best lipid coverage in both DDA and DIA (data not shown). Considering that the Q-RAI acquisition method is only configured with a single collision energy at the time of this publication, 20 V was selected for all our LC–MS analyses to maximize the lipid species coverage. At this stage, the term quantification refers to peak area integration without normalisation by lipid standards, with the exception of the validation case study. In our experiment, ten iterative MS/MS injections of a pooled lipid extract were analysed using MetaboKit, Lipid Annotator, MS-Dial and Lipostar software in parallel, and the numbers of lipid species identified by individual tools were compared. MS-Dial was used for comparison as it is a commonly used software for processing of untargeted lipidomics data acquired using the DDA and DIA approaches and contains the spectral libraries used by MetaboKit. Lipid Annotator was used for this comparison as it was developed by Agilent specifically for the analysis of iterative DDA data files from Agilent Q-TOF instruments. Iterative MS/MS allows precursors selected for fragmentation in prior injections to be excluded from fragmentation in subsequent analysis of the same sample, allowing less abundant precursors to be detected and hence improving the overall coverage in DDA. On Agilent Q-TOF instruments, iterative selection of precursors can be enabled in the acquisition method whereas open-source iterative tools may be required for other platforms, such as AcquireX that is used for Thermofisher Orbitrap instruments. After manual curation, a total of 228 lipid species were identified with MetaboKit in the positive mode data, compared to 216 with MS-Dial, 177 with Lipid Annotator and 226 with Lipostar (Fig. 1a). Different tools showed variable coverage of molecular species identified in each class; more SM species were reported by MS-Dial, more DG species by Lipostar and more Hex2Cer, PC-Ps, PEs and ether lipids by MetaboKit. Across all software, TGs were the most commonly identified species, followed by PCs, SMs and LPCs.Figure 1Identification of lipid species using Lipid Annotator vs MetaboKit vs MS-Dial vs Lipostar for (a) Positive Mode and (b) Negative Mode when analysing samples with iterative MS/MS (10 iterative injections) and DDA. In positive mode, MetaboKit gave the highest coverage with 228 lipids. In negative mode, Lipid Annotator, MetaboKit and MS-Dial gave comparable coverage while Lipostar gave the most identifications. Identification of lipid species using Lipid Annotator vs MetaboKit vs MS-Dial vs Lipostar for (a) Positive Mode and (b) Negative Mode when analysing samples with iterative MS/MS (10 iterative injections) and DDA. In positive mode, MetaboKit gave the highest coverage with 228 lipids. In negative mode, Lipid Annotator, MetaboKit and MS-Dial gave comparable coverage while Lipostar gave the most identifications. In the negative mode data, Lipid Annotator, MetaboKit, MS-Dial and Lipostar gave a total of 125, 113, 126 and 163 species respectively (Fig. 1b). Similar to the positive mode analysis, some differences were observed in lipid class composition. MetaboKit was the only software that identified N-acyl-lysophosphatidylethanolamine (LNAPE) and lysophosphatidylglycerol (LPG) species, while Lipostar identified more ceramides, PCs, PEs and SMs. For Lipid Annotator, MetaboKit and MS-Dial in both polarities, a higher lipid coverage as compared to the traditional DDA approach was indeed observed with the iterative DDA approach and the use of five iterative injections was enough to reach the maximum number of species identified. (Fig. S1). Although Lipostar gave the best identification results, it also reported 21,186 outputs that were manually filtered out in positive mode, which was significantly higher than other software ( Supplementary File 1). MS-Dial performed well but its report also contained a higher number of identifications that were filtered out in positive mode (334), compared to 73 and 203 in Lipid Annotator and MetaboKit, respectively (Supplementary File 1). Similarly, 15,762 identifications from Lipostar were filtered out in negative mode, compared to 86, 65 and 69 with MS-Dial, Lipid Annotator and MetaboKit respectively. These differences in identification results could be due to the different identification algorithms, score thresholds, error rates and spectral libraries used by the different software. For instance, the Bayesian theorem algorithm combined with non-negative least squares for spectral deconvolution that is unique to Lipid Annotator could have resulted in a more conservative annotation, leading to lesser identified species in positive mode. Conversely, the rule-based annotation system adopted by MS-Dial leads to a number of identifications generated without MS/MS evidence, which may have resulted in more false positives and reduced annotation confidence. This would also imply that certain lipid classes are detected better by different tools. In turn, a balance needs to be achieved between identification numbers and stringency and any compounds of interest identified through this DDA approach would need to be further verified through tandem MS/MS and targeted analysis of the product ions. To further assess the identification performance in DDA for MetaboKit, an analysis was also conducted using the same search space (spectral libraries) as MS-Dial, which yielded comparable performances (Fig. S2). These results showed that MetaboKit delivers comparably reliable identification performances with a coverage equivalent to other widely used tools, with the additional benefit of generating a customized MS/MS spectral library that includes product ion peak intensities and retention times. Overall, these findings suggest that MetaboKit is suitable for untargeted lipidomic analysis when using DDA and facilitates the generation and customization of data resources also for DIA analysis. While initially developed as a quantitative tool for small molecules, we used the Agilent Q-RAI acquisition function as both a discovery and quantitation workflow for DIA analysis of lipids. Using the spectral library generated in DDA mode, a DIA approach was performed in both polarities on triplicates of the same pooled human plasma lipid extracts, using both fixed (20(+ 1) Da) and variable quadrupole windows. In this workflow, DDA was used as the primary mode of lipid identification while quantification in DIA was done only if tandem MS supported the lipid identification in DDA mode. Among the lipids quantified in DIA positive mode, MetaboKit was able to extract quantitative data for 215 and 220 lipid species using fixed and variable windows, respectively (Fig. 2a, Supplementary File 2). Though these numbers show a slight reduction when compared to the 228 species identified with DDA positive mode using iterative MS/MS, these observations are expected since the quantification module requires that the ion chromatograms for the fragment ions in the DDA spectral library be retrievable for peak identification in DIA. Both types of isolation windows gave similar lipid class distributions, with TGs being the most represented, followed by PCs, SMs and LPCs. This trend is similar to the class distribution observed in DDA.Figure 2Number of lipid species identified when using DIA with the MetaboKit and Lipostar software for (a) Positive Mode and (b) Negative Mode using either fixed (20 (+ 1) Da) or variable windows. Lipostar gave marginally improved identification results, with the variable windows setup providing a slightly higher lipid coverage for both software in both polarities with the exception of MetaboKit in negative mode. Number of lipid species identified when using DIA with the MetaboKit and Lipostar software for (a) Positive Mode and (b) Negative Mode using either fixed (20 (+ 1) Da) or variable windows. Lipostar gave marginally improved identification results, with the variable windows setup providing a slightly higher lipid coverage for both software in both polarities with the exception of MetaboKit in negative mode. Among the lipids quantified in DIA negative mode, MetaboKit was able to extract quantitative data for 112 and 108 lipids using fixed and variable windows respectively (Fig. 2b, Supplementary File 2). Similar to the observations in positive mode, these numbers show a slight reduction when compared to the 113 species identified with DDA negative mode using iterative MS/MS. Detailed curation resulted in 145 and 124 MS1 features being filtered out in positive mode in fixed and variable windows respectively and 42 and 45 MS1 features being filtered out in negative mode in fixed and variable windows respectively. These results indicate that peak areas were associated to most of the lipid entries in the spectral library built with the previous DDA analysis and the targeted DIA extraction of MS2 fragments by MetaboKit enables MS2 quantitation of most lipid species identified in DDA. When analysing the same data files with Lipostar, which also supports lipid identification using DIA data, 212 and 243 lipid species were identified in positive mode using fixed and variable windows respectively, while 127 and 144 lipid species were identified in negative mode (Fig. 2). Manual curation resulted in 12,662 and 14,313 MS1 features being excluded in positive mode in fixed and variable windows respectively and 4,505 and 5,327 MS1 features being excluded in negative mode. As this type of data generation is not yet widely adopted on Agilent instruments, the input file format was not compatible with MS-Dial (at least when this manuscript was generated) and we were not able to optimally process Q-RAI Agilent data files using this bioinformatic tool. In addition, Lipostar does not provide MS2 quantification in the software outputs and hence we were unable to compare MS2 quantification performances. In terms of quantification, when using MS1 data obtained with DIA in both polarities, 98–99% of identified features showed RSD < 30%. Even though the fixed and variable windows approaches gave similar lipid coverages, the variable windows approach provided more reproducible results at the MS2 level, as seen from a higher number of fragments with RSD < 30%. The variable windows setup also generated more quantifiable MS2 fragments, as expected and previously reported. More than 90% of MS1 signals had at least one corresponding MS2 fragment with RSD < 30%, confirming that a reliable quantification of lipids using their product ions is one of the advantages of the DIA approach. As expected, the RSD values at both MS1 and MS2 levels decreased exponentially with increasing abundance, with MS2 quantitation giving higher RSD than MS1 (Fig. S3). Table 1 summarizes the quantification data.Table 1Descriptive statistics for the comparison between MS1 and MS2 measurements in DIA mode.Positive ModeNegative ModeFixed Windows (20 (+ 1) Da)Variable WindowsFixed Windows (20 (+ 1) Da)Variable WindowsMS1 Total Number of Features215220112108 Number of Features with RSD < 30%209 (97.21%)217 (98.64%)112 (100%)107 (99.07%) Median RSD for Features with RSD < 30%2.432.373.042.06 Number of Features with RSD > 30%6 (2.79%)3 (1.36%)0 (0%)1 (0.93%)MS2 Total Number of Fragments1982203810781040 Number of Fragments with RSD < 30%1150 (58.02%)1317 (64.62%)471 (43.69%)523 (50.29%) Median RSD for Fragments with RSD < 30%8.447.1610.969.45 Number of MS1 Features with MS2 Fragments of RSD < 30%210 (95.45%)217 (98.64%)109 (97.32%)107 (99.07%) Number of Fragments with RSD > 30%573 (28.91%)540 (26.50%)316 (29.31%)301 (28.94%) Number of Non-Quantifiable Fragments259 (13.07%)181 (8.88%)291 (26.99%)216 (20.77%) Descriptive statistics for the comparison between MS1 and MS2 measurements in DIA mode. These results show that the Q-RAI DIA approach coupled with MetaboKit processing is a feasible approach for semi-targeted quantification of lipids in both MS1 and MS2. For this purpose, a variable windows approach may be more suitable, due to its higher reproducibility and sensitivity (Table 1). The improved identification performances using variable windows, as seen from the Lipostar readouts, also showcases an added benefit of this methodological approach but would require further testing of identification algorithms in MetaboKit and other tools. To better evaluate the quantification performance of the Q-RAI DIA methodology, a dilution series using pooled plasma extracts was analysed using variable windows in DIA positive mode with the Q-TOF and, at the same time, with a targeted MRM method using a QQQ. While our DIA approach was not expected to outperform MRM due to different sensitivity performances, the rationale was to use the MRM approach as a benchmark and compare different quantification strategies. Various lipid amounts on column were measured by injecting serially increasing volumes: 0.25, 0.5, 1, 2, 4 and 8 µL. Table 2 and Supplementary File 3 summarizes the obtained results.Table 2Descriptive statistics for the comparison between measurements acquired on a Q-TOF by MS1 (DIA) and MS2 (DIA), or on a QQQ by MRM, for a six-point dilution series of a lipid extract from commercially available human plasma.Injection Volume (μL)0.250.51248DIA MS1 Total Number of Features106 Median RSD (%)3.802.351.821.091.080.92 Number of Features with RSD > 30%1 (0.94%)0 (0%)0 (0%)3 (2.83%)2 (1.89%)1 (0.94%) Median R0.990 Number of Features with R < 0.85 (4.72%)DIA MS2 Total Number of Quantified Fragments651 out of 1007 (64.65% of all fragments) Median RSD (%)15.1511.028.276.065.424.74 Number of Fragments with RSD > 30%167 (25.65%)108 (16.59%)73 (11.21%)61 (9.37%)37 (5.68%)58 (8.91%) Median R0.957 Number of Fragments with R < 0.8143 (21.97%)MRM Total Number of Transitions287 Median RSD (%)5.741.711.411.681.385.51 Number of Transitions with RSD > 30%0 (0%)0 (0%)0 (0%)0 (0%)0 (0%)0 (0%) Median R0.997 Number of Transitions with R < 0.80 (0%) Descriptive statistics for the comparison between measurements acquired on a Q-TOF by MS1 (DIA) and MS2 (DIA), or on a QQQ by MRM, for a six-point dilution series of a lipid extract from commercially available human plasma. When using DIA, similar results were obtained for the RSD at MS1 and MS2 level, with 8 µL injection volume giving the lowest RSD values followed by 4 µL (Table 2). However, saturation was evident at injection volumes above 4 µL. In view of this, the R for all outputs was calculated only up to an injection volume of 4 µL. A good linear response was observed, whereby MS1 showed R of 0.988 and for MS2 this was found to be 0.978 (Table 2). For the 106 lipid species monitored in the DIA dilution series data, heatmaps of the normalized median-centred fold changes showed proportional increases in the abundance for all lipids according to the injection volume, both at the MS1 (Fig. 3a) and MS2 (Fig. 3b) level.Figure 3Heatmaps of normalized median-centred fold changes of DIA MS1 (a) and DIA MS2 (b) for 106 lipid species identified and quantified in dilution series of lipid extracts from commercial human plasma. A distinct increase in concentration, directly proportional to increasing injection volumes (grey to black), is shown for all lipids measured, for both DIA MS1 and DIA MS2 acquisitions. Heatmaps of normalized median-centred fold changes of DIA MS1 (a) and DIA MS2 (b) for 106 lipid species identified and quantified in dilution series of lipid extracts from commercial human plasma. A distinct increase in concentration, directly proportional to increasing injection volumes (grey to black), is shown for all lipids measured, for both DIA MS1 and DIA MS2 acquisitions. To further investigate the potential use of MS2 for quantification, we calculated the Pearson’s correlation between the 651 generated DIA MS2 fragments and their corresponding MS1 precursors, obtaining a median Pearson's R of 0.973 with only 126 fragments (19.35%) having R < 0.8 (Table S3). As the best candidates for quantification for each of the 106 identified lipid species, the DIA MS2 fragments with the highest R for each MS1 feature were chosen. When considering these 106 MS2 fragments, the median R increased to 0.997 with no fragments having R < 0.8 (Table S3). A similar comparison between MRM and DIA showed that across all injection volumes, the median RSD and the percentage of fragments with RSD > 30% were the lowest when using MRM and highest when using DIA MS2 (Table 2). As described for DIA, saturation was observed for injection volumes > 4 µL, while MRM gave the best linearity values, with no transitions having R < 0.8 and a median R of 0.997 (Table 2). In light of this, different injection volumes of 4 µL and 8 µL were deemed optimal for DIA MS2 quantification in positive and negative mode respectively, due to the different sensitivity for different polarities and based on the lowest median RSD calculated when injecting different volumes. When using only the lipid species commonly found in the DIA and MRM methods, similar results were obtained in terms of RSD and R (Figs. S4, S5). 4 signals out of 88 showed a Pearson’s R < 0.8 when considering MRM vs DIA MS2, higher than the 2 signals out of 106 when comparing MRM and DIA MS1 (Fig. S6). When selecting the DIA MS2 fragment that showed the best correlation with DIA MS1, a higher correlation value with the corresponding MRM data was also observed, improving the results obtained when using the same MS2 fragment for both DIA and MRM (Fig. S7). Comparing the lipid species in common between DIA MS1, DIA MS2 and MRM also saw a reproducible linear increase of signal with injection volume for all three approaches (Fig. S8). The differences observed in the reproducibility (RSD) and linearity suggest that quantification performances are better in MRM mode. Although we did not observe a clear superiority of the Q-RAI DIA workflow, our results show that DIA MS1 features, DIA MS2 fragments and MRM transitions are highly correlated. This result is somewhat expected, considering the high sensitivity and reproducibility of MRM methodologies. However, considering that prior knowledge of lipids of interest is required for MRM, together with the time and effort required to develop a suitable targeted panel, the Q-RAI DIA approach is a valid choice for simultaneous untargeted lipid identification without prior knowledge and reasonably precise MS2 quantification of identified lipid species. Due to the presence of multiple MS2 fragments for each lipid, this DIA workflow can also be used to identify the best quantifier ions for targeted lipidomics optimization. User discretion might be advisable in terms of selecting the most representative MS2 fragments to quantify each MS1 feature for different applications. As shown in our results, Pearson’s correlation may be used as a criterion for selection of the best performing MS2 fragments for improvement of existing MRM lists. In addition to this, a variety of other criteria such as RSD, dot product score and linearity assessment may be used for the selection of either MS1 or MS2 ions. In general, MS1 measurements may have higher precision while MS2 measurements would be more sensitive and with their fragmentation information could serve as a more accurate quantifier, but this would be dependent on the analytes. For instance, ionization efficiency is dependent on numerous factors such as basicity of the analytes and solvents, relative intensities of the analytes, type of LC–MS instrument and experimental conditions, which could all contribute to deviation in MS1 quantification. The fragmentation yield is largely dependent on collision energy and may not be the optimal one for all analytes when using a constant value. As such, one might expect to observe less background noise for MS2 and more variability in MS1 quantitation (although a larger cohort would have to be used to further evaluate this aspect). Ceramide Synthase 2 (CerS2) is one of the six Ceramide Synthase (CerS1-CerS6) enzymes, which are distinguished by their use of distinct sub-sets of acyl CoAs for N-acylation of the sphingoid base. CerS2 uses mainly C22 and C24 acyl CoAs, thus generating very long acyl chain (VLC)-dihydroceramides and ceramides—the backbones of all sphingolipids (SLs). Lack of CerS2 results in a significant reduction of all VLC-SLs and a compensatory higher production of long chain C16 and C18 SLs. The depletion of VLC-SLs changes membrane properties, disrupting membrane domains and their properties. CerS2 null mice show defects in insulin resistance (by inhibition of insulin receptor translocation), fatty acid uptake (by CD36 mislocalization and FATP5 down-regulation) and gap junction dysfunction. Increased levels of C16-ceramides and sphinganine inhibit the mitochondrial respiratory complex IV, causing chronic oxidative stress. Hence, perturbation of the expression of CerS2 results in a reduction of C22- and C24-containing SLs and could induce a possible significant change in the concentrations of other lipid classes, which has not been extensively studied, via a possible regulatory effect of the affected sphingolipid species. Application of our newly developed DIA workflow in this context would both confirm the effect of CerS2 suppression on the SL levels but also shed light on the influence of SL perturbations on other lipid pathways. The DIA workflow was applied to the identification and quantification of lipids in serum samples from wild type (WT) and CerS2 null male mice. Five iterative injections of a pooled extract of all the murine samples were analysed in both polarities in DDA mode for generation of an in-house spectral library of lipids in mouse serum. Study samples, from five WT and five CerS2 null mice, were then injected in triplicates in DIA acquisition mode, using variable windows calculated by MetaboKit (Table S4). Similar to what was found in human plasma, DIA MS1 provided a more reproducible quantification than MS2 on average, based on the metrics calculated for the pooled quality control (QC) samples. The median RSD associated with DIA MS1 measurements was 7.41 and 5.23 across all features in positive and negative mode respectively, while for DIA MS2 the median RSDs were 17.72 and 18.13 (Table 3). In total, 244 and 143 lipid species were quantified in DIA positive and negative mode in the QC samples (Table 3).Table 3Descriptive statistics of all transitions quantified in DIA mode for the QC samples (n = 6) of the CerS2 null mice study.Positive ModeNegative ModeDIA MS1DIA MS2DIA MS1DIA MS2Total Number of Quantified Features/Fragments2441844 out of 2505143755 out of 1461Median RSD (%)7.4117.725.2318.13Number of Features/Fragments with RSD > 30%4 (1.64%)552 (29.93%)1 (0.70%)195 (25.83%)Number of Features/Fragments with RSD 20–30%7 (2.87%)281 (15.24%)2 (1.40%)152 (20.13%)Number of Features/Fragments with RSD < 20%233 (95.49%)1011 (54.83%)140 (97.90%)408 (54.04%) Descriptive statistics of all transitions quantified in DIA mode for the QC samples (n = 6) of the CerS2 null mice study. Lipid species with (1) MS1 and at least one MS2 fragment with RSD < 30% in the QC samples and (2) Pearson’s R > 0.8 between MS2 and MS1 in all samples were retained for analysis. Considering both polarities, a total of 173 lipid species passed the filtering criteria in positive mode and 119 in negative mode (Table S5). For these lipids, missing values in the mouse serum samples were then replaced using the outputs in the quant_fill_all text file and differential abundance analysis was conducted (Supplementary File 4). Choosing either the MS1 or the MS2 readout with the lowest RSD across both polarities as a representative result for each lipid, we found that 88 lipids had a significantly different abundance between WT and CerS2 null serum (Fig. 4). For some of the lipid species, after MS2-based quantification their levels were significantly different between WT and CerS2 null sera even when MS1 measurements although showing the same trend, did not reach statistical significance (Fig. S9a,c, Supplementary File 5). This indicates that in some cases DIA MS2 may be more sensitive for quantification. Nevertheless, for both polarities, most lipids quantified in the two groups were confirmed to be significantly different at both MS1 and MS2 levels (Fig. S9a,c). The heatmap of normalized median-centred lipid fold changes in the samples also shows distinct clustering based on lipidomic expression profile at both MS1 and MS2 levels (Fig. S10). These clusters revealed the presence in the serum of CerS2 null mice of lower levels of VLC-SLs and TGs but higher levels of other species, such as LPCs, acylcarnitines and ether-PC and -PE. This approach also further emphasized the effectiveness of using Pearson’s correlation in addition to RSD as a criterion for the selection of either MS1 or representative MS2 fragments from our Q-RAI workflow for semi-quantification.Figure 4Volcano plot representing all lipids quantified in WT and CerS2 null mice serum using a Q-RAI DIA workflow. 88 lipids showed significantly different concentrations between WT and CerS2 null (FC > 2, FDR < 0.05). As expected, very long chain sphingolipids (C22-24) were present at significantly lower levels in CerS2 null mice, while acylcarnitines, phosphatidylcholines (PCs) and long chain sphingolipids (C16-18) were significantly higher in CerS2 null mice. Volcano plot representing all lipids quantified in WT and CerS2 null mice serum using a Q-RAI DIA workflow. 88 lipids showed significantly different concentrations between WT and CerS2 null (FC > 2, FDR < 0.05). As expected, very long chain sphingolipids (C22-24) were present at significantly lower levels in CerS2 null mice, while acylcarnitines, phosphatidylcholines (PCs) and long chain sphingolipids (C16-18) were significantly higher in CerS2 null mice. Reflecting the specificity of the CerS2 enzyme, the most significant changes were registered on SLs, such as ceramides and SMs carrying C22 and C24 fatty acyl chains, having significantly lower levels in CerS2 null mice (Fig. 5). At the same time, for the previously mentioned compensatory mechanism, the long chain C16- and C18-containing species were increased (Fig. 5). For most of these SL, significantly different concentrations were found both at MS1 level and for all fragments in MS2 (Supplementary File 5). These findings on SL are also consistent with the MRM data previously published by Pewzner-Jung et al. and confirmed the reliability of our DIA approach for quantification.Figure 5Dot plot of abundance values for long chain FA (C16 and C18)- and very long chain FA (C22 and C24)-containing sphingolipid species in WT (n = 5) and CerS2 null (n = 5) mice, obtained using the developed Q-RAI DIA workflow. Confirming what was previously reported, very long chain ceramides (a), hexosylceramides (b) and sphingomyelins (c) were significantly decreased in CerS2 null mice, while long chain species in the same lipid classes were significantly increased in CerS2 null mice (FC > 2, FDR < 0.05). Dot plot of abundance values for long chain FA (C16 and C18)- and very long chain FA (C22 and C24)-containing sphingolipid species in WT (n = 5) and CerS2 null (n = 5) mice, obtained using the developed Q-RAI DIA workflow. Confirming what was previously reported, very long chain ceramides (a), hexosylceramides (b) and sphingomyelins (c) were significantly decreased in CerS2 null mice, while long chain species in the same lipid classes were significantly increased in CerS2 null mice (FC > 2, FDR < 0.05). When considering other lipid species, an increase was observed for Coenzyme Q9 (CoQ9), several acylcarnitines and PCs, while TG species with saturated or monounsaturated fatty acyl chains decreased. Dysregulation of acylcarnitine levels suggests an impaired function of the mitochondria caused by incomplete β-oxidation of lipids, possibly leading to elevated oxidative stress, as previously shown in the same animal model. As a compensatory mechanism, increased production of CoQ9 might occur to coordinate the assembly of the electron transport chain and promote antioxidative function within the mitochondria. This informative finding was only allowed by the Q-RAI DIA untargeted approach, as this is not a molecule usually covered by lipid panels. An increased level of C16 sphingolipids can also induce similar mitochondrial defects, which could generate an increase in the TG levels as a buffering system for the increase in toxic free fatty acids. Conversely, C22-C24 ceramides can cause dysregulation of hepatic CD36/FAT expression, thereby resulting in reduced hepatic accumulation of TGs and reduced secretion into plasma as a complex with very low-density lipoproteins (VLDLs) or chylomicrons. Another interesting finding is the negative correlation existing between VLC-SL levels and ether lipids, which has been observed earlier by other investigators in cell models. This effect seems to be due to SL and ether lipids having similar functions in the cell membrane. Although the reason for this co-regulation is not known, from our results we can hypothesise that this might be mainly governed by the SL species containing VLC FA, as in our system the long chain SL changed in the same direction as the ether species. More in general, 38 and 44 lipids were found to have significantly different levels between WT and CerS2 null samples in positive mode and negative mode respectively (Fig. S9b,d). These findings indicate the success of the Q-RAI DIA approach in elucidating the possible pivotal role of CerS2 as a regulatory gene in liver homeostasis and general lipid metabolism, thereby illustrating the potential of this methodology for high-throughput lipidomic screening in biological applications. In summary, we developed for the first time a DIA workflow for lipidomics using Q-RAI acquisition on the Agilent 6546 LC/Q-TOF instrument. Following data acquisition, MetaboKit provides robust functionalities in untargeted lipid identification for DDA analysis and targeted quantification for DIA data, at both the MS1 and MS2 levels, using a customized spectral library. The availability of multiple MS2 fragments for each identified MS1 feature in the DIA readouts provides an option for the user to select representative MS2 fragments of interest for each compound, using pre-defined objective quality metrics, which will be particularly applicable as a digital footprint for development of in-house targeted methods or for small scale biological studies. This solution may potentially be extended to larger scale applications and analysis of polar metabolites in the near future.
|
PMC6760179
|
Genetic architecture of human plasma lipidome and its link to cardiovascular disease
|
Understanding genetic architecture of plasma lipidome could provide better insights into lipid metabolism and its link to cardiovascular diseases (CVDs). Here, we perform genome-wide association analyses of 141 lipid species (n = 2,181 individuals), followed by phenome-wide scans with 25 CVD related phenotypes (n = 511,700 individuals). We identify 35 lipid-species-associated loci (P <5 ×10), 10 of which associate with CVD risk including five new loci-COL5A1, GLTPD2, SPTLC3, MBOAT7 and GALNT16 (false discovery rate<0.05). We identify loci for lipid species that are shown to predict CVD e.g., SPTLC3 for CER(d18:1/24:1). We show that lipoprotein lipase (LPL) may more efficiently hydrolyze medium length triacylglycerides (TAGs) than others. Polyunsaturated lipids have highest heritability and genetic correlations, suggesting considerable genetic regulation at fatty acids levels. We find low genetic correlations between traditional lipids and lipid species. Our results show that lipidomic profiles capture information beyond traditional lipids and identify genetic variants modifying lipid levels and risk of CVD.Cardiovascular diseases (CVDs) encompass many pathological conditions of impaired heart function, vascular structure and circulatory system. CVDs are the leading cause of mortality and morbidity worldwide, necessitating the need for better preventive and predictive strategies. Plasma lipids, the well-established heritable risk factors for CVDs, are routinely monitored to assess CVD risk. Standard lipid profiling measures traditional lipids (referred to LDL-C, HDL-C, total triglycerides and total cholesterol), but does not capture the functionally and chemically diverse molecular components—the lipid species. These molecular lipid species may independently and specifically affect different manifestations of CVD, such as ischaemic heart disease and stroke. Lipid species including cholesterol esters (CEs), lysophosphatidylcholines (LPCs), phosphatidylcholines (PCs), phosphatidylethanolamines (PEs), ceramides (CERs), sphingomyelins (SMs) and triacylglycerols (TAGs) potentially improve CVD risk assessment over traditional lipids. Understanding of the genetic architecture and genetic regulation of these lipid species could help guide tool development for CVD risk prediction and treatment. Genetic studies of traditional lipids have identified over 250 genomic loci and improved our understanding of CVD pathophysiology. For the majority of the lipid loci, however, their effects on detailed lipidome beyond traditional lipids are unknown. Only a few studies have reported genetic associations for lipid species either through studies on subsets of the lipidome or GWASs on metabolome. In light of the limited information about the genetics of lipidomic profiles and their relationship with CVDs, we carried out a GWAS of lipidomic profiles of 2181 individuals using ~9.3 million genetic markers followed by PheWAS including 25 CVD-related phenotypes in up to 511,700 individuals (Fig. 1). We aimed to (1) determine heritability of lipid species and their genetic correlations; (2) identify genetic variants influencing the plasma levels of lipid species; (3) test the relationship between identified lipid–species-associated variants and CVD manifestations and (4) gain mechanistic insights into established lipid variants. We find that lipid species are heritable, suggesting a considerable role of endogenous regulation in lipid metabolism. We report association of new genomic loci with lipid species and CVD risk in humans. In addition to enhancing the current understanding of genetic regulation of circulating lipids, our study emphasises the need of lipidomic profiling in identifying additional variants influencing lipid metabolism.Fig. 1Study design and work flow. The figure illustrates the study design and key findings of the study Study design and work flow. The figure illustrates the study design and key findings of the study First, we determined SNP-based heritability for each of the lipid species and traditional lipids using genetic relationship matrix for all the study participants. The demographic characteristics of the study participants are provided in Supplementary Table 1. SNP-based heritability estimates ranged from 0.10 to 0.54 (Fig. 2a; Supplementary Table 2), showing considerable variation across lipid classes (Fig. 2b), with similar trends as reported previously. CERs showed the greatest estimated heritability (median = 0.39, range = 0.35–0.40), whereas phosphatidylinositols (PIs) showed the least heritability (median = 0.19, range = 0.11–0.31). Sphingolipids had higher heritability than glycerolipids ranging from 0.24 to 0.41 (Fig. 2b), which is similar to a previous study that reported higher heritability for sphingolipids ranging from 0.28 to 0.53 estimated based on pedigrees. Lipids containing polyunsaturated fatty acids, particularly C20:4, C20:5 and C22:6, had significantly higher heritability compared with other lipid species (Fig. 2c). For instance, PC (17:0;0–20:4;0) and LPC (22:6;0) had the highest heritability (> 0.50), whereas PC (16:0;0–16:1;0) and PI (16:0;0–18:2;0) had the lowest heritability estimates (< 0.12) (Supplementary Table 2).Fig. 2Heritability of lipidomic profiles and genetic correlations among the lipid species. a Histogram and kernel density curve showing the distribution of heritability estimates across all the lipid species. b Boxplot showing the heritability estimates in each lipid class. c Boxplot showing comparison of the median heritability estimates of lipid species containing C20:4, C20:5 and C22:6 acyl chains and all others. The P-values were calculated using the Wilcoxon rank-sum test. d Hierarchical clustering of lipid species based on genetic correlations among lipid species. Lipids containing polyunsaturated fatty acids C20:5, C20:4 and C22:6 are highlighted with black bars. The data presented in the boxplots represent the interquartile range (IQR) defined by the bounds of the box with the median (middle line of the box) and whiskers extending to the largest/smallest values no further than 1.5 times the IQR. CER ceramide, DAG diacylglyceride, LPC lysophosphatidylcholine, LPE lysophosphatidylethanolamine, PC phosphatidylcholine, PCO phosphatidylcholine-ether, PE phosphatidylethanolamine, PEO phosphatidylethanolamine-ether, PI phosphatidylinositol, CE cholesteryl ester, SM sphingomyelin, ST sterol, TAG triacyglycerol, Trad traditional lipids Heritability of lipidomic profiles and genetic correlations among the lipid species. a Histogram and kernel density curve showing the distribution of heritability estimates across all the lipid species. b Boxplot showing the heritability estimates in each lipid class. c Boxplot showing comparison of the median heritability estimates of lipid species containing C20:4, C20:5 and C22:6 acyl chains and all others. The P-values were calculated using the Wilcoxon rank-sum test. d Hierarchical clustering of lipid species based on genetic correlations among lipid species. Lipids containing polyunsaturated fatty acids C20:5, C20:4 and C22:6 are highlighted with black bars. The data presented in the boxplots represent the interquartile range (IQR) defined by the bounds of the box with the median (middle line of the box) and whiskers extending to the largest/smallest values no further than 1.5 times the IQR. CER ceramide, DAG diacylglyceride, LPC lysophosphatidylcholine, LPE lysophosphatidylethanolamine, PC phosphatidylcholine, PCO phosphatidylcholine-ether, PE phosphatidylethanolamine, PEO phosphatidylethanolamine-ether, PI phosphatidylinositol, CE cholesteryl ester, SM sphingomyelin, ST sterol, TAG triacyglycerol, Trad traditional lipids Longer, polyunsaturated lipids (those with four or more double bonds) had stronger genetic correlations with each other than with other lipid species (Supplementary Fig. 1, Supplementary Data 1). This can be seen in the hierarchical clustering based on genetic correlations that segregate TAG subspecies into two clusters based on carbon content and degree of unsaturation (Fig. 2d). These patterns were not seen in phenotypic correlations that were estimated based on the plasma levels of lipid species (Supplementary Fig. 2). We observed low phenotypic and genetic correlation between traditional lipids and molecular lipid species, except strong positive genetic correlations of triglycerides with TAGs and DAGs (average r = 0.88) (Fig. 3). However, triglycerides had low genetic correlation with other lipid species (average (abs) r = 0.26). HDL-C and LDL-C levels had low genetic and phenotypic correlations with most of the lipid species (Fig. 3; Supplementary Data 1). Consistently, all of the known lipid variants explained 2–21% of variances in plasma levels of various lipid species, with the least variance accounting for LPCs (Fig. 3). To rule out the possibility that lipid-lowering medications resulted in the observed low genetic correlations between traditional lipids and lipid species, we also calculated the genetic correlations after excluding the individuals using lipid lowering medications (N = 172). This re-analysis provided the similar results as the primary analysis (Supplementary Fig. 3). It is to be noted that this sample size might not provide sufficient power for heritability estimations in unrelated samples. Our study also included the family samples which provides higher statistical power in heritability estimation than unrelated samples.Fig. 3Lipidomic profiles capture information beyond traditional lipids. The genetic and phenotypic correlations between traditional lipids and molecular lipid species are shown in lower panel. The bar plot in the upper panel shows the heritability estimates of each lipid species (red bars) and the variance explained by all the known loci together (green bars). The lipid species are ordered based on the hierarchical clustering showing the correlations between the lipid species and traditional lipids. TC total cholesterol, TG triglycerides Lipidomic profiles capture information beyond traditional lipids. The genetic and phenotypic correlations between traditional lipids and molecular lipid species are shown in lower panel. The bar plot in the upper panel shows the heritability estimates of each lipid species (red bars) and the variance explained by all the known loci together (green bars). The lipid species are ordered based on the hierarchical clustering showing the correlations between the lipid species and traditional lipids. TC total cholesterol, TG triglycerides Next, we performed the genome-wide association analyses for 141 lipid species with ~9.3 million genetic markers. We identified 2817 associations between 518 variants located within 11 genomic loci (1MB blocks) and 42 lipid species from 10 lipid classes at study-wide significance (P < 1.5 × 10 accounting for 34 principal components that explain 90% of the variance in lipidome) (Table 1; Supplementary Data 2, 3). These included three new loci (ROCK1, MAF and SYT1) that are not previously reported for any lipid measure or related metabolite (Fig. 4). Among the new loci, the strongest association was at an intronic variant rs151223356 near ROCK1 with short acyl-chain LPC(14:0,0) (P = 1.9 × 10). ROCK1 encodes for a serine/threonine kinase that plays key role in glucose metabolism. In line with our observation of higher heritability for lipids with C20:4, C20:5 or C22:6 acyl chains, we detected associations for 15 out of 21 lipids with these acyl chains.Table 1Genomic loci associated with molecular lipid species at genome-wide significanceSNPPositionGeneChangeRefAltAFLipid speciesEffectSEPrs2013853661:897866KLHL17IntronicCT0.019LPE(22:6;0)−0.870.163.6 × 10rs1871639481:14399146KAZNIntronicGA0.011TAG(53:3;0)0.950.173.5 × 10rs768663862:44075483ABCG5/8IntronicTC0.077CE(20:2;0)−0.390.063.9 × 10rs580292412:98701245VWA3BIntergenicTA0.062TAG(50:1;0)0.370.071.9 × 10rs130701103:21393248ZNF385DIntergenicTC0.085Total CER0.330.063.9 × 10rs102124393:142655053PAQR9IntergenicTC0.602PI(18:0;0–18:1;0)0.180.033.1 × 10rs131513744:8122221ABLIM2IntronicGA0.153TAG(50:1;0)0.250.043.7 × 10rs1866894844:97033701PDHA2IntergenicGA0.051TAG(52:4;0)−0.400.074.2 × 10rs5438955016:74120350DDX43IntronicCT0.013Total LPC0.870.162.9 × 10rs48963076:138297840TNFAIP3IntergenicCT0.216PCO(16:1;0–16:0;0)−0.230.043.3 × 10rs5346931557:101081274COL26A1IntronicAG0.010LPC(16:1;0)1.240.233.9 × 10rs102817417:157793122PTPRN2IntronicGC0.225TAG(54:6;0)0.210.042.2 × 10rs14788988:11395079BLK IntronicGA0.440PC(16:0;0–16:0;0)0.170.032.5 × 10rs115708918:19822810LPLIntronicCT0.075TAG(52:3;0)−0.330.062.9 × 10rs1467177109:137549865COL5A1IntronicCT0.011PC(16:0;0–16:1;0)−1.030.192.8 × 10rs14064584710:118863255SHTN1IntronicGT0.101LPE(20:4;0)−0.320.063.3 × 10rs2845611:61589481FADS2IntronicAG0.405CE(20:4;0)−0.590.031.1 × 10rs96418411:116648917APOA5IntergenicGC0.855TAG(52:3;0)−0.2580.0459.5 × 10rs1079049511:122198706MIR100HGIntronicAG0.590TAG(56:4;0)−0.200.042.1 × 10rs11738857312:78980665SYT1IntergenicAG0.020LPC(14:0;0)−0.770.139.8 × 10rs51294813:52374489DHRS12IntronicTC0.225LPE(18:2;0)−0.220.041.4 × 10rs800807014:64233720SYNE2IntronicAT0.133SM(32:1;2)0.480.052.9 × 10rs390295114:69789755GALNT16IntronicTG0.361PEO(18:1;0–18:2;0)0.190.031.9 × 10rs3586193815:45637343GATMIntergenicTC0.398PCO(18:2;0–18:1;0)0.180.032.7 × 10rs26129015:58678720LIPCIntronicTC0.617PE(18:0;0–20:4;0)−0.370.034.0 × 10rs3522197716:79563576MAFIntronicGC0.054LPC(16:0;0)−0.460.081.3 × 10rs7920268017:4692640GLTPD2IntronicGT0.032SM(34:0;2)−0.850.093.4 × 10rs14320335217:77293933RBFOX3IntronicTC0.024PC(16:0;0–18:1;0)0.600.113.2 × 10rs15122335618:18627427ROCK1IntronicAC0.013LPC(14:0;0)0.970.151.9 × 10rs724661719:8272163CERS4IntergenicGA0.402SM(38:2;2)0.250.032.5 × 10rs245506919:51728641CD33MissenseAG0.383TAG(52:5;0)−0.190.039.3 × 10rs873619:54677189MBOAT7UTRCT0.388PI(18:0;0–20:4;0)−0.380.039.8 × 10rs437429819:55738746TMEM86BSynonymousGA0.166PEO(16:1;0–20:4;0)−0.250.042.3 × 10rs36458520:12962718SPTLC3IntergenicAG0.670Total CER−0.200.039.1 × 10rs18668000822:39754367SYNGR1IntronicAC0.015CE(20:3;0)−0.810.152.6 × 10Ref reference allele, Alt alternate allele, AF alternate allele frequency, SE standard error, UTR untranslated regionThe strongest association between SNP and lipid species in the genome-wide significant loci (P < 5.0 × 10) are presented. The P-values were calculated from the meta-analyses using the inverse variance weighted method for fixed effects. The study-wide significant associations are marked by hash symbol. The SNPs are annotated to the nearest gene if identified in this study (marked by asterisk symbol) or to previously known gene if in linkage disequilibrium with the known loci for any lipid measure. The effect sizes presented are change in standard deviation of the lipid species per alternate allele. Chromosomal positions are based on hg19 reference sequenceFig. 4Genetic architecture of the lipidome. a Manhattan plot showing associations for all 141 lipid species. Only the associations with P < 1.0 × 10 in the meta-analysis and consistent in directions in all three batches are plotted. The y-axis is capped at −log10 P-value = 30 for better representation of the data. The dotted line represents the threshold for genome-wide significant associations at P < 5.0 × 10. b Genome-wide significant associations between the identified lipid species-associated loci and lipid species showing effect of the loci on the lipidome. The plotted P-values were calculated from the meta-analyses using the inverse variance weighted method for fixed effects. New hits with P < 5.0 × 10 are shown as red dots, new independent hits in previously reported loci are presented as blue dots and hits in previously known loci are presented as black dots Genomic loci associated with molecular lipid species at genome-wide significance Ref reference allele, Alt alternate allele, AF alternate allele frequency, SE standard error, UTR untranslated region The strongest association between SNP and lipid species in the genome-wide significant loci (P < 5.0 × 10) are presented. The P-values were calculated from the meta-analyses using the inverse variance weighted method for fixed effects. The study-wide significant associations are marked by hash symbol. The SNPs are annotated to the nearest gene if identified in this study (marked by asterisk symbol) or to previously known gene if in linkage disequilibrium with the known loci for any lipid measure. The effect sizes presented are change in standard deviation of the lipid species per alternate allele. Chromosomal positions are based on hg19 reference sequence Genetic architecture of the lipidome. a Manhattan plot showing associations for all 141 lipid species. Only the associations with P < 1.0 × 10 in the meta-analysis and consistent in directions in all three batches are plotted. The y-axis is capped at −log10 P-value = 30 for better representation of the data. The dotted line represents the threshold for genome-wide significant associations at P < 5.0 × 10. b Genome-wide significant associations between the identified lipid species-associated loci and lipid species showing effect of the loci on the lipidome. The plotted P-values were calculated from the meta-analyses using the inverse variance weighted method for fixed effects. New hits with P < 5.0 × 10 are shown as red dots, new independent hits in previously reported loci are presented as blue dots and hits in previously known loci are presented as black dots We also replicated the previous associations of FADS2, SYNE2, LIPC, CERS4 and MBOAT7 with the same lipid species. The previously reported associations at the known loci identified in previous metabolomics GWASs are provided in Supplementary Data 4. This information was obtained from the databases-SNiPA (http://snipa.org) using block annotation and PhenoScanner v2 (http://www.phenoscanner.medschl.cam.ac.uk/), and were manually curated to include associations from literature search. In addition, we also identified new locus–lipid species associations at previously reported lipid loci including new associations of variants at ABCG5/8 with CE (20:2;0) (P = 3.9 × 10), MBOAT7 with PI (18:0;0–20:3;0) (P = 3.0 × 10) and GLTPD2 with SM (34:0;2) (P = 3.4 × 10) (Supplementary Data 2, 3). Further, we systematically evaluated the associations of variants previously identified in metabolomics GWAS (126 variants from 46 loci available in our data set out of 132 reported) with 141 lipid species. Of these known variants, 76 variants from 12 loci showed association with 98 different lipid species with P < 3.2 × 10 (correcting for 46 loci and 34 PCs for lipid species) (Supplementary Data 5). Of the 134 previously reported variant–lipid species pair associations that could be examined in our data set, 94 of such associations were replicated with the same direction of effect with P < 3.7 × 10 (accounting for 134 comparisons) in our study (Supplementary Data 6). In addition, 24 further loci were associated with at least one lipid species at regularly used genome-wide significance level (1.5 × 10>P < 5.0 × 10). Among these additional loci, 13 loci were located in genomic regions not previously reported for any lipid measure or related metabolite, and 8 loci were located near known loci for lipids but were independent of any previously reported variant (Table 1; Supplementary Data 3). The regional association plots for all 35 loci with P < 5.0 × 10 are presented in Supplementary Data 7, and the genotype–phenotype relationships for the lead variants in these 35 loci are provided in Supplementary Fig. 4. As many of the lipid species have previously been shown to predict CVD risk, we determined if the variants associated with lipid species affect individuals’ susceptibility to CVD-related phenotypes in FinnGen and UK Biobank cohorts. We identified 25 CVD-related phenotypes from the clinical outcomes derived from health registry data in the FinnGen and UK Biobanks (Supplementary Table 3). The follow-up PheWAS analyses included lead variants from all of the 35 independent loci that showed associations with P < 5.0 × 10 (Table 1). Overall, 10 of the 35 lipid–species variants (APOA5, ABCG5/8, BLK, LPL, FADS2, COL5A1, GALNT16, GLTPD2, MBOAT7 and SPTLC3) were associated with at least one of the CVD outcomes (FDR < 5%) (Fig. 5; Supplementary Data 8). These included novel associations of variants at COL5A1 with cerebrovascular disease (P = 4.6 × 10), GALNT16 with angina (P = 9.3 × 10), MBOAT7 with venous thromboembolism (P = 1.3 × 10), GLTPD2 with atherosclerosis (P = 5.3 × 10) and SPTLC3 with intracerebral haemorrhage (P = 1.0 × 10) (Fig. 5). FADS1-2-3 is a well-known lipid modifying locus; however, like many other known lipid loci, its effects on CVD risk has been unclear. We found an association of FADS2 rs28456-G with peripheral artery disease (P = 2.2 × 10) and aterial embolism and thrombosis (P = 2.5 × 10). BLK (rs1478898-A) was also found to be associated with decreased risk of obesity (OR = 0.97, P = 5.6×10) and type 2 diabetes (OR = 0.96, P = 4.5 × 10).Fig. 5Relationship between lipid species-associated variants and risk of CVDs. The upper panel shows the association of the identified variants with the strongest associated lipid species. Boxplots show the interquartile range (IQR) defined by the bounds of the box with the median (middle line of the box) of plasma levels of the respective lipid species for each genotype of the variants; whiskers extend to the largest/smallest values no further than 1.5 times the IQR. The lower panel depicts the relationship between the identified variants with CVD phenotypes. The effect sizes (odds ratio) with 95% confidence interval are plotted with respect to the alternate alleles. The associations with CVD phenotypes highlighted in red colour are significant at FDR <0.05 Relationship between lipid species-associated variants and risk of CVDs. The upper panel shows the association of the identified variants with the strongest associated lipid species. Boxplots show the interquartile range (IQR) defined by the bounds of the box with the median (middle line of the box) of plasma levels of the respective lipid species for each genotype of the variants; whiskers extend to the largest/smallest values no further than 1.5 times the IQR. The lower panel depicts the relationship between the identified variants with CVD phenotypes. The effect sizes (odds ratio) with 95% confidence interval are plotted with respect to the alternate alleles. The associations with CVD phenotypes highlighted in red colour are significant at FDR <0.05 Several studies have suggested a role for sphingolipids, including CERs and SMs, in the pathogenesis of CVDs. CER (d18:1/24:0) and CER (d18:1/24:1) have been reported to be associated with the increased risk of CVD events. We found that the CER (d18:1/24:1) decreasing variant SPTLC3 rs364585-G was associated with decreased risk of intracerebral haemorrhage, while CER (d18:1/24:0) increasing variant ZNF385D rs13070110-C was nominally associated with increased risk of intracerebral haemorrhage. Furthermore, consistent with the observation that elevated plasma SMs levels are atherogenic, we identified association of GLTPD2 rs79202680-T (associated with reduced levels of SMs) with reduced risk of atherosclerosis. Next, we determined if the detailed lipidomic profiles could provide new mechanistic insights into the role of known lipid variants in lipid biology. We present two examples of well-established lipid variants here. First is the fatty acid desaturase (FADS) gene cluster that has been consistently reported to be associated with omega-3 and omega-6 fatty acids levels with inverse effects on different PUFAs. Its mechanism, however, has not been fully deciphered. Here, we found that the FADS2 rs28456-G was associated with increased levels of lipids with a C20:3 acyl chain and decreased levels of lipids with C20:4, C20:5 and C22:6 acyl chains (Supplementary Fig. 5). The rs28456-G is also an eQTL that increases FADS2 expression while reduces the expression of FADS1 [GTEx v7]. These data together explain the inverse relationship of FADS2 variants with lipids containing different polyunsatureated fatty acids (PUFAs) (Fig. 6).Fig. 6Patterns in associations and proposed mechanisms for the effect of identified variants on lipid metabolism and clinical outcomes. a Associations of LPL rs11570891-T and LPL activity with TAGs. Change (beta and standard errors) in plasma levels of TAGs per increase in standard deviation of LPL activity with their corresponding P-values, as calculated using linear regression model, are plotted in lower panel. The upper panel shows change (beta and standard errors) in plasma levels of TAGs per T allele with their corresponding P-values, as obtained from meta-analyses of genome-wide association analysis. b Association of LPL variant rs11570891 with LPL activity. The effect size (beta in standardised units and standard error in parenthesis) and P-value were calculated using linear mixed model. Boxplot depicts the interquartile range (IQR) defined by the bounds of the box, median (middle line) and whiskers extending to the largest/smallest values no further than 1.5 times the IQR. c Based on the patterns of the association of lipid species-associated loci with different lipid species, we propose that: (1) LPL rs11570891-T and APOA5 rs964184-C might result in more efficient hydrolysis of medium length TAGs which might results in reduced CVD risk, (2) FADS2 rs28456-G may have observed effect on PUFA metabolism through its inverse effect on FADS2 and FADS1 expressions, (3) SYNGR1 rs18680008-C might have a role in the negative regulation of either desaturation of linoleic acid (C18:2,n-6) or elongation of gamma linoleic acid (C18:3,n-6). (4) PTPRN2 rs10281741-G and MIR100HG rs10790495-G, which have very similar patterns of association with reduced level of long polyunsaturated TAGs, might have a role in negative regulation of either elongation and desaturation of fatty acids or incorporation of long chain unsaturated fatty acids in glycerol backbone during TAG biosynthesis. The positive (+) and negative (−) signs indicate increase or decrease, respectively, in level of lipid species or risk of disease as observed in our study, with different colours for different genetic variant Patterns in associations and proposed mechanisms for the effect of identified variants on lipid metabolism and clinical outcomes. a Associations of LPL rs11570891-T and LPL activity with TAGs. Change (beta and standard errors) in plasma levels of TAGs per increase in standard deviation of LPL activity with their corresponding P-values, as calculated using linear regression model, are plotted in lower panel. The upper panel shows change (beta and standard errors) in plasma levels of TAGs per T allele with their corresponding P-values, as obtained from meta-analyses of genome-wide association analysis. b Association of LPL variant rs11570891 with LPL activity. The effect size (beta in standardised units and standard error in parenthesis) and P-value were calculated using linear mixed model. Boxplot depicts the interquartile range (IQR) defined by the bounds of the box, median (middle line) and whiskers extending to the largest/smallest values no further than 1.5 times the IQR. c Based on the patterns of the association of lipid species-associated loci with different lipid species, we propose that: (1) LPL rs11570891-T and APOA5 rs964184-C might result in more efficient hydrolysis of medium length TAGs which might results in reduced CVD risk, (2) FADS2 rs28456-G may have observed effect on PUFA metabolism through its inverse effect on FADS2 and FADS1 expressions, (3) SYNGR1 rs18680008-C might have a role in the negative regulation of either desaturation of linoleic acid (C18:2,n-6) or elongation of gamma linoleic acid (C18:3,n-6). (4) PTPRN2 rs10281741-G and MIR100HG rs10790495-G, which have very similar patterns of association with reduced level of long polyunsaturated TAGs, might have a role in negative regulation of either elongation and desaturation of fatty acids or incorporation of long chain unsaturated fatty acids in glycerol backbone during TAG biosynthesis. The positive (+) and negative (−) signs indicate increase or decrease, respectively, in level of lipid species or risk of disease as observed in our study, with different colours for different genetic variant Another example is lipoprotein lipase (LPL). LPL codes for lipoprotein lipase that is the master lipolytic factor of TAGs in TAG-enriched chylomicrons and VLDL particles. We found that LPL rs11570891-T was associated with reduced levels of medium length TAGs (C50–C56), with strongest associations with TAG (52:3;0). This suggested that LPL enzyme might have different efficiency in hydrolysis of TAGs of different length. We explored this possibility by evaluating (1) the effect of LPL rs11570891-T on LPL enzymatic activity and (2) the relationship between LPL activity and plasma levels of TAGs of different length, using post-heparin LPL measured in the EUFAM cohort. We found that LPL rs11570891-T (an eQTL increasing LPL expression) was associated with increased LPL activity, which in turn was associated with TAG species with stronger effect on medium length TAGs than other TAGs (Fig. 6). Consistent with a previous report by Rhee et al., variant rs964184-C at APOA5, which codes for the activator that stimulates LPL-mediated lipolysis of TAG-rich lipoproteins and their remnants, also showed association with medium length TAGs (Fig. 6). These results provide first clues to the probable variable role of LPL and APOA5 in the hydrolysis of different TAG species. Similarly, the association patterns of some of the newly mapped loci suggested their underlying functions. For example, SYNGR1 rs186680008-C showed strongest associations with decreased levels of lipid species with C20:3 acyl chain from different lipid classes, including CEs, PCs and PCOs (Supplementary Fig. 5), suggesting its role in PUFA metabolism (Fig. 6). PTPRN2 rs10281741-G and MIR100HG rs10790495-G showed associations with reduced levels of long polyunsaturated TAG species, suggesting their role in negative regulation of either elongation and desaturation of fatty acids or incorporation of long-chain unsaturated fatty acids during TAG biosynthesis. As intermediate phenotypes are known to provide more statistical power, we assessed whether the lipid species could help to detect genetic associations with greater power than traditional lipids using variants previously identified for traditional lipids (number of variants = 557; Supplementary Data 9). We found that molecular lipid species have much stronger associations than traditional lipids with the same sample size, except for well-known APOE and CETP (Fig. 7; Supplementary Data 10). The associations were several orders of magnitudes stronger for the variants in or near genes involved in lipid metabolism, such as FADS1-2-3, LIPC, ABCG5/8, SGPP1 and SPTLC3. This shows that the lipidomics provides higher chances to identify lipid-modulating variants, particularly the ones with direct role in lipid metabolism, with much smaller sample size than traditional lipids.Fig. 7Association of known variants for traditional lipids with lipid species and traditional lipids. The P-values for the associations of the lead SNPs (557 SNPs available in our data set) identified through different genome-wide or exome-wide studies of traditional lipids (HDL-C, LDL-C, TG and TC) with lipid species (upper panel) and traditional lipids (lower panel) are plotted. The y-axis in the upper panel is capped at −log10 P-value = 30 for better representation of the data. The SNPs on the x-axis are serially arranged based on their chromosomal positions and as listed in the Supplementary Data 8. The points on the plots are colour coded by the lipid classes in the upper panel and traditional lipid in the lower panel. CER ceramide, DAG diacylglyceride, LPC lysophosphatidylcholine, LPE lysophosphatidylethanolamine, PC phosphatidylcholine, PCO phosphatidylcholine-ether, PE phosphatidylethanolamine, PEO phosphatidylethanolamine-ether, PI phosphatidylinositol, CE cholesteryl ester, SM sphingomyelin, ST sterol, TAG triacyglycerol, TC total cholesterol, TG triglycerides Association of known variants for traditional lipids with lipid species and traditional lipids. The P-values for the associations of the lead SNPs (557 SNPs available in our data set) identified through different genome-wide or exome-wide studies of traditional lipids (HDL-C, LDL-C, TG and TC) with lipid species (upper panel) and traditional lipids (lower panel) are plotted. The y-axis in the upper panel is capped at −log10 P-value = 30 for better representation of the data. The SNPs on the x-axis are serially arranged based on their chromosomal positions and as listed in the Supplementary Data 8. The points on the plots are colour coded by the lipid classes in the upper panel and traditional lipid in the lower panel. CER ceramide, DAG diacylglyceride, LPC lysophosphatidylcholine, LPE lysophosphatidylethanolamine, PC phosphatidylcholine, PCO phosphatidylcholine-ether, PE phosphatidylethanolamine, PEO phosphatidylethanolamine-ether, PI phosphatidylinositol, CE cholesteryl ester, SM sphingomyelin, ST sterol, TAG triacyglycerol, TC total cholesterol, TG triglycerides We present findings from a large-scale study that integrate lipidome, genome and phenome revealing detailed description of genetic regulation of lipidome and its associations with CVD risk. In addition to enhancing the current understanding of genetic determinants of circulating lipids, our study highlights the potential of lipidomics in gene mapping for lipids and CVDs over traditional lipids. The study generates a publicly available knowledgebase of genetic associations of molecular lipid species and their relationships with thousands of clinical outcomes. Despite the expected influence of dietary intake on the circulatory lipids, plasma levels of lipid species are found to be heritable, suggesting considerable role of endogenous regulation in lipid metabolism. Importantly, genetic mechanisms do not seem to regulate all lipid species in a lipid class in the same way, as also observed in recent mice lipidomics studies. Longer and more unsaturated lipid species from different lipid classes clearly display stronger genetic correlations. These observations are consistent with a previous study based on family pedigrees. Our finding is important in the light of the proposed role of lipids containing PUFAs in CVDs, diabetes and other disorders. Identification of genetic factors regulating these particular lipids is important for understanding the subtleties of lipid metabolism and devising preventive strategies including dietary interventions. Our study provides multiple leads in this direction by identifying 11 genomic loci (KLHL17, APOA5, CD33, SHTN1, FADS2, LIPC, MBOAT7, MIR100HG, PTPRN2, PDHA2 and TMEM86B) associated with long, polyunsaturated lipids at genome-wide significance. Of these, FADS2, APOA5, LPL and MBOAT7 variants were also associated with risk of CVDs (Fig. 5). Further, we mapped genetic variants for lipid species from several lipid classes, including CERs, CEs, TAGs, SMs and PCs, that are shown to predict CVD risk. Our PheWAS analyses also suggested relationship between many of the mapped genetic variants and CVD outcomes. This knowledge can directly fuel studies on CVD prediction or drug target discovery. For instance, CERs and CEs have also been reported to associate with increased risk of CVD events. Our study revealed three loci associated with CEs, including FADS2 and two novel loci-ABCG5/8 and SYNGR1, and two loci for CERs (SPTLC3 and ZNF385D). CER species, particularly CER (d18:1/24:0) and CER (d18:1/24:1) are recently reported to be associated with the increased risk of CVD. We identified two variants near SPTLC3 and ZNF385D that modulate the plasma levels of CER (d18:1/24:1) and CER (d18:1/24:0), respectively, and risk for intracerebral haemorrhage. This information could also guide future studies to establish the causal relationship between lipid species and CVD. The detailed lipidomic profile also provided clues towards understanding the mechanisms of effects of well-established lipid loci like FADS2 and LPL on lipid metabolism and CVD risks. We show how the inverse effects of FADS2 rs28456-G on the expression of two desaturases (FADS2 and FADS1) could explain its opposite effects on lipids with different PUFAs. The delta-6 desaturation by FADS2 generates gamma-linolenic acid and stearidonic acid that by elongation yield dihomo-gamma-linolenic acid and eicosatetraenoic acid (Fig. 6). Further, delta-5 desaturation of dihomo-gamma-linolenic acid by FADS1 generates arachidonic acid and eicosapentaenoic acid. Thus, as depicted in Fig. 6, the inverse effects of FADS2 rs28456-G on FADS2 and FADS1 expressions explain its opposite effects on different PUFAs. The association of FADS2 rs28456-G with the reduced levels of lipids containing arachidonic acid may also explain its assocition with reduced risk of atherosclerotic CVD outcomes—peripheral artery disesae (PAD) and aterial embolism and thrombosis. LPL and APOA5 are the key players in TAG hydrolysis. Our integrated approach suggested that their activity could be different for different TAG species with higher efficiency for medium length TAGs (C50–C56). We show that an LPL variant increases the LPL activity resulting in decreased levels of medium length TAGs. The association of the LPL variant with reduced susceptibility to CVD and type 2 diabetes could be mediated through the decrease in medium length TAGs (Fig. 5). This is consistent with a previous report that showed a similar pattern of association of levels of TAG species with type 2 diabetes. Similarly, the patterns of assocations of newly mapped loci also suggested their involvement in the regulation of lipid metabolism. For example, rs10281741-G near PTPRN2 and rs10790495-G near MIR100HG showed distinct association patterns with TAGs, with strongest association with long polyunsaturated TAGs. PTPRN2 codes for protein tyrosine phosphatase receptor N2 with a possible role in pancreatic insulin secretion and development of diabetes mellitus, while MIR100HG rs10790495 is an eQTL for the heat-shock protein HSPA8 that has a role in cell proliferation. However, it is not known if PTPRN2 and MIR100HG or HSPA8 have any role in lipid metabolism. Finally, we show that lipidomic profiles capture information beyond traditional lipids and provide an opportunity to identify additional genetic variants influencing lipid metabolism and disease risk. Previously, Petersen et al. showed that lipoprotein subfractions correlate with traditional lipids and strengthen genetic associations at known lipid loci and that these loci explain more of the variance of lipoprotein subfractions than of serum lipids. Similarly, our study demonstrates that molecular lipid species have stronger statistical power compared with traditional lipids at known lipid loci using the same sample size. However, in contrast to Petersen et al., we found that many of the lipid species, including LPCs and PCs that have previously been associated with incident coronary heart disease risk, have low phenotypic and genotypic correlations with traditional lipids. We also show that the known lipid variants for traditional lipids explain less of the variance of lipid species than traditional lipids. Altogether, as expected these results suggest that lipidomic profiles could provide novel information that could not be captured by traditional lipids and lipoprotein measurements. Our study had some potential limitations. Though our study represents one of the largest genetic screen of lipidomic variation, larger cohorts are needed to achieve its full understanding. Blood samples for the EUFAM cohort were drawn after an overnight fast whereas the FINRISK cohort samples had varied fasting duration. This, however, does not seem to have substantial effect on the results and their interpretation as shown in Supplementary Data 11 and Supplementary Fig. 6. Moreover, a recent study by Rämö et al. also demonstrated similar lipidomic profiles for dyslipidemias from the EUFAM and FINRISK cohorts. The UK Biobank cohort is reported to have a “healthy volunteer” effect, which may affect the PheWAS results, however, given the large sample size, this is unlikely to have a substantial effect on genetic association analyses. Furthermore, lipidomic profiles were measured in whole plasma, which does not provide information at the level of individual lipoprotein subclasses and limits our ability to gain detailed mechanistic insights. We also excluded poorly detected lipid species to ensure high data quality that narrowed the spectrum of lipidomic profiles. Further advances in lipidomics platforms might help to capture more comprehensive and complete lipidomic profiles, including the position of fatty acyl chains in the glycerol backbone of TAGs and glycerophospholipids and detection of sphingosine-1-P species and several other species, that would allow to overcome these limitations. In conclusion, our study demonstrates that lipidomics enables deeper insights into the genetic regulation of lipid metabolism than clinically used lipid measures, which in turn might help guide future biomarker and drug target discovery and disease prevention. The study included participants from the following cohorts: EUFAM, FINRISK, FinnGen and UK Biobank. The EUFAM (The European Multicenter Study on Familial Dyslipidemias in Patients with Premature Coronary Heart Disease) study cohort is comprised of the Finnish familial combined hyperlipidemia families. The families in EUFAM study were identified via probands admitted to Finnish university hospitals with a diagnosis of premature coronary heart disease. The probands had premature coronary heart disease and high levels of the total cholesterol, triglycerides, or both (≥ 90th Finnish age-specific and sex-specific population percentile), or low HDL-C levels (≤ 10th percentile). Invitation was extended to all the family members and spouses of the probands if at least one first-degree relative of the proband had high levels of the total cholesterol, triglycerides, or both. Venous blood samples were obtained from all participants after overnight fasting. Triglycerides and total cholesterol were measured by enzymatic methods using an automated Cobas Mira analyser (Hoffman-La Roche, Basel, Switzerland). HDL-C was quantified by phosphotungstic acid/magnesium chloride precipitation procedures, and LDL-C was calculated using the Friedewald formula. The Finnish National FINRISK study is a population-based survey conducted every 5 years since 1972, and thus far samples have been collected in 1992, 1997, 2002, 2007 and 2012. Collections from the 1992, 1997, 2002, 2007 and 2012 surveys are stored in the National Institute for Health and Welfare /THL) Biobank. Lipidomic profiling was performed for 1142 participants that were randomly selected from the FINRISK 2012 survey (Supplementary Table 1). The participants were advised to fast for at least 4 h before the examination and to avoid heavy meals earlier during the day. Venous blood samples were obtained from all the participants and sera were separated. HDL-C, triglycerides and total cholesterol were measured with enzymatic methods (Abbott laboratories, Abbott Park, IL, USA) with Abbott Architect c8000 clinical chemistry analyser. The FinnGen data release 2 is composed of 102,739 Finnish participants. The phenotypes were derived from ICD codes in Finnish national hospital registries and cause-of-death registry as a part of FinnGen project. The quality of the CVD diagnoses in these registers has been validated in previous studies. The UK Biobank data is comprised of >500,000 participants based in UK and aged 40–69 years, annotated for over 2000 phenotypes. The PheWAS analyses in this study included 408,961 samples from white British participants. The study was conducted in accordance with the principles of the Helsinki declaration. Written informed consent was obtained from all the study participants. The study protocols were approved by the ethics committees of the participating centres (The Hospital District of Helsinki and Uusimaa Coordinating Ethics committees, approval No. 184/13/03/00/12). For the Finnish Institute of Health and Welfare (THL) driven FinnGen preparatory project (here called FinnGen), all patients and control subjects had provided informed consent for biobank research, based on the Finnish Biobank Act. Alternatively, older cohorts were based on study specific consents and later transferred to the THL Biobank after approval by Valvira, the National Supervisory Authority for Welfare and Health. Recruitment protocols followed the biobank protocols approved by Valvira. The Ethical Review Board of the Hospital District of Helsinki and Uusimaa approved the FinnGen study protocol Nr HUS/990/2017. The FinnGen preparatory project is approved by THL, approval numbers THL/2031/6.02.00/2017, amendments THL/341/6.02.00/2018, THL/2222/6.02.00/2018 and THL/283/6.02.00/2019. All DNA samples and data in this study were pseudonymized. Mass spectrometry-based lipid analysis of 2181 participants was performed in three batches-353 and 686 EUFAM participants in two batches and 1142 FINRISK participants in third batch at Lipotype GmbH (Dresden, Germany). Samples were analysed by direct infusion in a QExactive mass spectrometer (Thermo Scientific) equipped with a TriVersa NanoMate ion source (Advion Biosciences). The data were analysed using in-house developed lipid identification software based on LipidXplorer. Post processing and normalisation of data were performed using an in-house developed data management system. Only lipids with signal-to-noise ratio >5 and amounts at least fivefold higher than in the corresponding blank samples were considered for further analyses. Reproducibility of the assay was assessed by the inclusion of reference plasma samples (eight reference samples for EUFAM and three reference samples for FINRISK) per 96-well plate. Median coefficient of variation was <10% across all batches. The data were corrected for batch and drift effects. Lipid species detected in <80% of the samples in any of the batches and samples (N = 64) with low lipid contents were excluded. Among the lipid species which passed quality control, a total of 141 lipid species from 13 lipid classes (Supplementary Table 2) were detected consistently in all three batches and were included in all analysis. The total amounts of lipid classes were calculated by summing up the absolute concentrations of all lipid species belonging to each lipid class. The measured concentrations of the lipid species and calculated class total were transformed to normal distribution by rank-based inverse normal transformation. It is to be noted that Lipotype platform used in the study detected many additional lipid species (N = 83) that were not captured previously by other platforms. The list of the lipid species detected by different platforms and overlaps across the platforms are provided in the Supplementary Data 12 and Supplementary Fig. 7. Genotyping for both EUFAM and FINRISK cohorts was performed using the HumanCoreExome BeadChip (Illumina Inc., San Diego, CA, USA). The genotype calls were generated together with other available data sets using zCall at the Institute for Molecular Medicine Finland (FIMM). Genotype data underwent stringent quality control (QC) before imputation that included exclusion of samples with low call rate (<95%), sex discrepancies, excess heterozygosity and non-European ancestry. Variants with low call rate (<95%) and deviation from Hardy–Weinberg Equilibrium (HWE P < 1 × 10) were excluded. Imputation was performed using IMPUTE2, which used two population-specific reference panels of 2690 high-coverage whole-genome and 5093 high-coverage whole-exome sequence data. Variants with imputation info score <0.70 were filtered out. After QC on lipidomic profiles and imputed variants, all subsequent analyses included 2045 individuals and ~9.3 million variants with MAF >0.005 that were available in both cohorts. FinnGen samples were genotyped with Illumina and Affymetrix arrays (Thermo Fisher Scientific, Santa Clara, CA, USA). Genotype calls were made with GenCall and zCall algorithms for Illumina and AxiomGT1 algorithm for Affymetrix chip genotyping data. Genotyping data produced with previous chip platforms were lifted over to build version 38 (GRCh38/hg38) following the protocol described here: dx.doi.org/10.17504/protocols.io.nqtddwn. Samples with sex discrepancies, high genotype missingness (> 5%), excess heterozygosity (+-4SD) and non-Finnish ancestry were removed. Variants with high missingness (> 2%), deviation from HWE (P < 1e-6) and low minor allele count (MAC < 3) were removed. Pre-phasing of genotyped data was performed with Eagle 2.3.5 (https://data.broadinstitute.org/alkesgroup/Eagle/) with the default parameters, except the number of conditioning haplotypes was set to 20,000. Imputation was carried out by using the population-specific SISu v3 imputation reference panel with Beagle 4.1 (version 08Jun17.d8b, https://faculty.washington.edu/browning/beagle/b4_1.html) as described in the following protocol: [dx.doi.org/10.17504/protocols.io.nmndc5e]. SISu v3 imputation reference panel was developed using the high-coverage (25–30x) whole-genome sequencing data generated at the Broad Institute of MIT and Harvard and at the McDonnell Genome Institute at Washington University; and jointly processed at the Broad Institute. Variant callset was produced with GATK HaplotypeCaller algorithm by following GATK best-practices for variant calling. Genotype-, sample- and variant-wise QC was applied in an iterative manner by using the Hail framework v0.1 [https://github.com/hail-is/hail]. The resulting high-quality WGS data for 3775 individuals were phased with Eagle 2.3.5 as described above. Post-imputation quality control involved excluding variants with INFO score < 0.7. Genotyping for the majority of the UK Biobank participants was done using the Affymetrix UK Biobank Axiom Array, while a subset of participants was genotyped using the Affymetrix UK BiLEVE Axiom Array. Details about the quality control and imputation of UK Biobank cohort are described by Bycroft et al.. For heritability and genetic correlation estimation, rank-based inverse-transformed measures of lipid species, computed separately for the EUFAM and FINRISK cohorts, were combined to increase statistical power. The residuals of inverse-transformed measures after regressing for age, sex, first ten principal components (PCs) of genetic population structure, lipid medication, hormone replacement therapy, thyroid condition and type 2 diabetes were used as phenotypes. SNP-based heritability estimates were calculated using the variance component analysis using a genetic relationship matrix (GRM) as implemented in biMM. Only the good quality variants with missingness <10% and MAF >0.005 were used to generate the GRM. The GRM was generated using GCTA by setting the off-diagonal elements that are <0.05 to 0 as proposed by Zaitlen et al.. This allows to estimate SNP-based heritability in family data without removing closely related individuals. The heritability estimates of lipid species in different groups were compared using Wilcoxon rank-sum test. The genetic correlation between each pair of lipid species and between each lipid species and traditional lipids was determined using the generated GRM with bivariate linear mixed model as implemented in biMM. The correlations based on the plasma levels (termed as phenotypic correlations) between all the pairs of the lipid species and traditional lipids were calculated using Pearson’s correlation coefficient. The heatmaps and hierarchical clustering based on genetic and phenotypic correlations were generated using heatmap.2 in R. As lipid-lowering medications could affect the plasma levels of lipid species, all analyses were adjusted for the usage of lipid-lowering medications, and separate analyses were also performed after excluding individuals using lipid-lowering medications (N = 172). We performed univariate association tests for 141 individual lipid species, 12 total lipid classes and 4 traditional lipid measures (HDL-C, LDL-C, total cholesterol and triglycerides), in all batches to control for possible batch effects and combined the summary statistics by meta-analysis. The association analyses for the EUFAM cohort were performed using linear mixed models, including the above-mentioned covariates as fixed effects and kinship matrix as random effect as implemented in MMM. The kinship matrices for the GWAS analyses were computed separately for each chromosome to include the variants from the other chromosomes using directly genotyped variants with MAF >0.01 and missingness <2%. The FINRISK cohort was analysed with linear regression model adjusting for age, sex, first ten PCs, lipid medication and diabetes using SNPTEST v2.5. Meta-analyses were performed using the inverse variance weighted method for fixed effects adjusted for genomic inflation factor in METAL. In addition, analyses adjusting for the traditional lipids (in addition to above-mentioned covariates) were also performed for the identified variants to determine the independent effect on lipid species. Test statistics were adjusted for λ values if >1.0 before meta-analyses. Genomic inflation factor (λ) ranged from 0.98 to 1.19 across the batches whereas the final λ values for meta-analysis ranged from 0.998 to 1.045 (Supplementary Data 13). The P-values obtained from the meta-analysis were considered to determine the SNP–lipid species associations. To account for multiple tests, the study-wide P-value threshold was set at <1.5 × 10 after correcting for 34 principal components (PCs) that explain over 90% of the variance in lipidomic profiles. Only the associations consistent in effect direction in all three batches were considered significant. Variants were designated as new if not located within 1 Mb of any previously reported variants for lipids (any of the traditional lipids and molecular lipid species); and as independent signal in known locus if located within 1 Mb but r < 0.20 with the previous lead variants and confirmed by conditional analysis. Variants with the strongest association in the identified lipid species loci was identified as the lead variants, and were annotated to the nearest gene for the new loci. We identified 25 CVD-related outcomes from the derived phenotypes in the FinnGen and UK Biobanks (Supplementary Table 3). Associations between the 35 lead variants from the identified loci and 25 selected CVD phenotypes in FinnGen cohort were obtained from the ongoing analyses as a part of the FinnGen project. The associations were tested using saddle point approximation method adjusting for age, sex and first 10 PCs as implemented in SPAtest R package. Associations between selected binary phenotypes and 35 lead variants in UK Biobank were obtained from Zhou et al. that were tested using logistic mixed model in SAIGE with a saddle point approximation and adjusting for first four principal components, age and sex (https://www.leelabsg.org/resources). Data for four phenotypes were not available from Zhou et al. and hence were obtained from http://www.nealelab.is/uk-biobank/. Associations of quantitative traits were tested using linear regression models with the same covariates as mentioned above, both for Finnish and UK Biobank cohorts. Meta-analyses of both cohorts were performed using the inverse variance weighted method for fixed effects model in METAL. The P-values obtained from the meta-analyses of the two cohorts are reported for PheWAS associations. All the PheWAS associations with false discovery rate (FDR) <5% evaluated using the Benjamini–Hochberg method and consistent direction of effects were considered significant. To determine the variance explained by the known loci for traditional lipids, we included all the lead variants with MAF >0.005 in 250 genomic loci that have previously been associated with one or more of the four traditional lipids. Of the 636 reported variants, 557 variants with MAF >0.005 (including six proxies) were available in our QC passed imputed genotype data (Supplementary Data 10). A genetic relationship matrix (GRM) based on these 557 variants was generated using GCTA that was used to determine the variance in plasma levels of all lipid species explained by the known variants using variance component analysis in biMM. The post-heparin lipoprotein lipase (LPL) after 15 min of heparin load was measured for 630 individuals in the EUFAM cohort using the ELISA method developed by Antikainen et al.. The measured values were transformed using rank-based inverse normal transformation. Associations between the LPL activity and plasma levels of TAGs were determined using linear regression model adjusted for age, sex, lipid medication, hormone replacement therapy, thyroid condition and type 2 diabetes. Association between the LPL variant rs11570891 and LPL activity was tested using linear mixed model adjusted for age, sex, first ten PCs of genetic population structure, lipid medication, hormone replacement therapy, thyroid condition and type 2 diabetes as fixed effect and kinship matrix as random effect as implemented in MMM. Further information on research design is available in the Nature Research Reporting Summary linked to this article.
|
PMC9599481
|
Integrated Quantitative Targeted Lipidomics and Proteomics Reveal Unique Fingerprints of Multiple Metabolic Conditions
|
Aberrations in lipid and lipoprotein metabolic pathways can lead to numerous diseases, including cardiovascular disease, diabetes, neurological disorders, and cancer. The integration of quantitative lipid and lipoprotein profiling of human plasma may provide a powerful approach to inform early disease diagnosis and prevention. In this study, we leveraged data-driven quantitative targeted lipidomics and proteomics to identify specific molecular changes associated with different metabolic risk categories, including hyperlipidemic, hypercholesterolemic, hypertriglyceridemic, hyperglycemic, and normolipidemic conditions. Based on the quantitative characterization of serum samples from 146 individuals, we have determined individual lipid species and proteins that were significantly up- or down-regulated relative to the normolipidemic group. Then, we established protein–lipid topological networks for each metabolic category and linked dysregulated proteins and lipids with defined metabolic pathways. To evaluate the differentiating power of integrated lipidomics and proteomics data, we have built an artificial neural network model that simultaneously and accurately categorized the samples from each metabolic risk category based on the determined lipidomics and proteomics profiles. Together, our findings provide new insights into molecular changes associated with metabolic risk conditions, suggest new condition-specific associations between apolipoproteins and lipids, and may inform new biomarker discovery in lipid metabolism-associated disorders.Lipid metabolism plays a central role in maintaining the normal homeostasis of the human body. Aberration of lipid metabolism is a trigger for chronic diseases, including diabetes, neurological disorders, cancer, and cardiovascular disease (CVD) . Traditionally, screening among asymptomatic individuals for lipid disorders rests on consideration of age, gender, blood pressure, smoking status, and testing for cholesterol, triglycerides, and glucose levels in plasma. Cholesterol tests measure both endogenous free cholesterol (FC) and hydrolyzed cholesteryl esters (CEs). The cholesterol content of the density or size fractions of lipid-carrying lipoprotein particles in plasma allows estimation of the particle number of high-density (HDL), low-density (LDL), very low-density (VLDL) lipoproteins, and chylomicrons . In the case of CVD especially, these traditional lipid profile measures leave many individuals undiagnosed . The prevention and treatment of dyslipidemias require the development of alternative diagnostic tools that allow for the assessment of the lipid and protein constituents of lipoproteins that are more directly related to the underlying unique metabolic irregularities of individuals . The main lipid constituents of lipoproteins are CEs and triacylglycerols (TAGs) contained within their core, amphipathic phospholipids (PLs) and ceramides (CERs) on their surface, and FC distributed between the core and the surface. The main PLs are phosphatidylcholines (PCs), phosphatidylethanolamines (PEs), sphingomyelins (SMs), and lyso-derivatives (LPCs and LPEs). These lipid classes have high structural diversity. The cholesteryl, glycero-phosphatidyl, glyceryl, and sphingosine backbones carry fatty acyl moieties (FAs) that are linked through ester, ether, or amide bonds. FAs differ by carbon chain length (usually between 12 and 24) and their number of double bonds (depending on the chain length between 0 and 6). In addition, lipid molecules may differ by the position of FA groups on the backbone, carbohydrate modifications of the backbone, and the position of double bonds on the FA chain. With the polar PL head groups exposed to the aqueous plasma environment, the surface of the lipoproteins incorporates numerous types of proteins . An important class of lipoprotein binding proteins is apolipoproteins (apos), with amphipathic alpha-helical domains that have a unique affinity to phospholipid monolayers. Apos are essential for the biogenesis, structural integrity, and function of lipoprotein particles . Structurally, one of the most essential apo is apoA1 for HDL and apoB-100 for LDL and VLDL formation . Chylomicrons contain a truncated form of apoB (apoB-48) , and on the more atherogenic Lp(a) particles, apoB is extended through the S-S bond by apo(a) . Apos that interact with apoA1 and apoB containing lipoproteins, called exchangeable apos and have well-characterized roles as cofactors and inhibitors in lipoprotein remodeling processes include A2 , A4 , C1 , C2 , C3 , and E . Other exchangeable proteins are known as non-polar molecule carriers (apoD and apoM), lipid transfer proteins (CETP and PLTP), lipase enzymes (LCAT), or inflammation-related proteins (AACT, HP, PON1, SAA1, SAA4, and TF) . The lipid and protein constituents together affect the structural integrity as well as the metabolic fate and circulating plasma concentration of lipoproteins. The competitive binding of exchangeable apos is modulated by the fluidity of the phospholipid monolayer that is determined by species composition, studied mostly by using model membranes and purified or artificial lipoproteins . Advances in tandem mass spectrometry detection (MS/MS) techniques and the availability of stable isotope-labeled internal standards enable a steadily increasing number of both lipid species and proteins that can be quantified from one sample . The integration of these quantitative lipidomics and proteomics techniques brings promising opportunities not only in basic research but also in epidemiology and diagnostics . In this study, we leveraged a multi-omics approach to determine lipidomic and proteomic profile changes associated with different metabolic risk categories: hypercholesterolemic (HC), hypertriglyceridemic (HT), hyperlipidemic (HL), and hyperglycemic (HG), relative to a normolipidemic (NL) control group. By discriminant analysis, we have found unique concentration changes in both proteins and lipid species in each of the risk categories. Based on the concentration correlations between lipid species and proteins we constructed protein–lipid connectivity networks that provide new insights into lipid and protein constituents of lipoproteins unique to the individual metabolic categories. In support of our approach, the most significantly different lipid species and proteins and their concentration correlation patterns were consistent with known pathways of lipid synthesis and extracellular lipoprotein remodeling. Furthermore, using the machine learning artificial neural network (ANN) approach, known as “deep learning”, we have identified a set of lipids and proteins that can accurately distinguish among all four pathological conditions and controls. Together, our data provide new insights into molecular profiles of different metabolic conditions and demonstrate the potential of integrated multi-omics to improve the characterization and differentiation of metabolic disorders. HPLC grade methanol (MeOH), dichloromethane (DCM), 1-propanol, 2-propanol, sodium bicarbonate, sodium chloride, isopropanol, hexanes, ethanol, isopropanol, and water were purchased from Fisher Scientific (Waltham, MA, USA). Ammonium acetate (NH4AcO) was obtained from Millipore Sigma (St. Louis, MO, USA). Labeled d7-cholesterol was purchased from Sigma-Aldrich (St. Louis, MO, USA). Cholesteryl-d7-palmitate was purchased from Avanti Polar Lipids, (Alabaster, AL, USA). Labeled d98-tripalmitin was purchased from CDN Isotopes, (Pointe-Claire, QC, Canada). NIST standard reference materials were purchased from the National Institute of Standards and Technology (Gaithersburg, MD, USA). The internal standards kit for the Lipidyzer platform was purchased from AB SCIEX (Framingham, MA, USA). Labeled internal standard peptides for proteomics analysis were purchased from Biosynth (Gardner, MA, USA). Human plasma samples pooled for quality control (QC) were purchased from BioIVT, Inc. (Westbury, NY, USA). Serum samples from 146 specimens were purchased from BioIVT, Inc. (Westbury, NY, USA). The study population included 81 males (55%) and 65 females (45%), with a mean age of 60.6 ± 17.2 years. All samples were collected from individuals fasting for more than 8 h. Based on clinically measured levels of total cholesterol (Total-C), total triglycerides (Total-TAG), and glucose, the samples were ordered from five categories. The mean levels in the sample groups were as follows: hypercholesteremia, HC (n = 36, 274 (245–310) mg/dL Total-C and 93 (50–125) mg/dL Total-TAG); hypertriglyceridemia, HT (n = 32, 188 (136–236) mg/dL Total-C and 268 (161–380) mg/dL Total-TAG); hyperlipidemia, HL (n = 28, 279 (232–355) mg/dL Total-C and 299 (155–573) mg/dL Total-TAG); hyperglycemia HG (n = 29, >180 mg/dL glucose, 122 (82–164) mg/dL Total-C and 146 (26–459) mg/dL Total-TAG); and a normolipidemic, NL group (n = 21, Total-C < 200 mg/dL, Total-TAG < 150 mg/dL, glucose < 180 mg/dL). Of note, these Total-C (FC + CE) and Total-TAG measurements were obtained with our in-house developed LC-MS/MS method described in detail in ref. . Briefly, the sample extraction was conducted in 96-well plates. Each sample was extracted in a single well, “in one-pot”, without the need for manual liquid phase separation or sample transfer before LC-MS/MS analysis. For the simultaneous analysis of FC, CE, and TAG species, UHPLC separation and in-source collision-induced dissociation (CID) coupled MS/MS method was used. Aliquots of 50 μL of 1:100 dilute serum samples in 10 mm sodium bicarbonate and 75 mm sodium chloride pH = 7.4 buffer were placed on a 96-well plate. Cholesteryl palmitate was used as the external calibrator for the CE lipid class and the mixture of triolein, tripalmitin, and trilinolein in a ratio of 514:313:173, reflective of the typical ratio in humans, was used as an external calibrator for the TAG lipid class. QCs prepared from NIST SRM 1951c were analyzed with each plate. The internal standard (IS) spiking mix of stable isotope labeled analogs were prepared in ethanol, containing 0.033 mg/dL d7-cholesterol (IS for FC), 0.098 mg/dL cholesteryl-d7-palmitate (IS for CE), and 0.125 mg/dL d98-tripalmitin (IS for TAG). The UHPLC system Agilent 1290 (Agilent Technologies, Santa Clara, CA, USA) coupled to a hybrid triple quadrupole/linear ion trap Sciex 4000 QTrap (Sciex, Framingham, MA, USA) was used. The column was a Kinetex HILIC 1.7 μm, 2.1 × 50 mm (Phenomenex, Torrance, CA, USA). Mobile phase A was hexanes with 0.05% isopropanol. Mobile phase B was hexanes with 5% ethanol and 0.05% isopropanol. Class-specific fragments were generated in-source for CE and TAG prior to MS/MS. The multiple reaction monitoring (MRM) method in positive ion mode was used for data acquisition. Collected raw data were processed with Multiquant software. All samples were de-identified prior to shipment such that no personal identification was associated with any sample. The project was approved as research not involving identifiable human subjects under the U.S. Health and Human Services Department Policy for Protection of Human Research Subjects codified of Federal Regulations at 45 CFR part 46. Lipids were extracted using a modified Bligh and Dyer extraction protocol . Briefly, 2 mL methanol, 1 mL dichloromethane (DCM), and 1 mL water were added to 25 µL serum samples containing one or two internal standards for each lipid class. The list of deuterium-labeled internal standards spiked into all quality controls, and unknowns can be found in Supplementary Table S1. The generated monophase mixture was incubated at room temperature (20 ± 2 °C) for 30 min followed by the addition of 1 mL water and 0.9 mL DCM, gentle mixing, and 10-min centrifugation at 1200 RPM to assist in phase separation. The lower layer containing DCM was transferred to a separate tube, and the lipid extraction was repeated a second time. All collected lower phases containing lipids were evaporated under nitrogen to dryness and reconstituted with 250 µL buffer containing 50:50 (v:v) DCM:MeOH and 10 mM NH4AcO. In an earlier study using a similarly grouped sample set, we reported that HDL particles have lower SM/PL and higher PE/PL molar ratios than LDL and VLDL particles . Within HDL, LDL, and VLDL fractions, we found higher SM/PL and lower PE/PL ratios in HC and NL than in HT and HL samples. Consequently, the direction of these trends was similarly observed in unfractionated samples as well. Therefore, in this study, we analyzed lipids and proteins without fractionation and assumed that main lipid composition differences between sample groups would similarly apply to HDL, LDL, and VLDL fractions as well. The Lipidyzer platform (AB SCIEX, Framingham, MA, USA) was used to detect and quantify lipid concentrations in the serum extracts, as described in detail elsewhere . A 50 µL aliquot of the extracts was injected into a constant 50 µL/min flow of 50:50 (v:v) DCM:MeOH and 10mM NH4AcO buffer and directly infused into the triple quadrupole SCIEX QTRAP 5500 mass spectrometer. The infusion was repeated using two different acquisition methods, both containing polarity switching in positive and negative modes. The first method used the SelexION Differential Mobility Spectrometry (DMS) to analyze PC, PE, SM, LPC, and LPE species. The second method was run without the DMS to select for and analyze TAG, DAG, CER, CE, and FFA species. Each method cycles through its respective list of MRM scans twenty times, and all quantitation was accomplished using the average signal of the twenty cycles for both native lipid species and internal standards to calculate response ratios. The response ratios were multiplied with the respective spiked internal standard concentrations to obtain species concentration. All species except TAGs were quantified based on a single unique MS/MS signal relative to the analogous MS/MS signal of a labeled internal standard specie (Supplementary Table S1). TAGs were monitored by 1–3 MRM transitions. Lipid species had to have 10–20 %CVs and <30% missing values to be considered quantifiable. After applying these criteria, the list of species included 12 SMs, 9 LPCs, 4 LPEs, 22 CEs, 22 PCs, 20 PEs, 6 CERs, 4 HCERs, 23 FFAs, 17 DAGs, and 435 TAGs. The coefficient of variation (CV) for each lipid class was calculated using quality control (QC) samples (Supplementary Table S2). Summing the molar concentration of lipid species by class yielded lipid class concentrations (Table 1). Some species had a low %Abundance of 0.01–3% within the lipid class. All quantified species had at least pmol/mL level of absolute concentrations and were quantifiable with 10–20 %CVs and <30% missing values, sufficient to find statistically significant changes relative to controls and find correlations with proteins within confidence intervals around mean concentrations of individual lipid species. We categorized the lipid species according to their number of double bonds on FA carbon chains (Supplementary Table S3). The FA groups were annotated as odd chain, saturated or mono-unsaturated (SFA/MUFA), double-unsaturated (DUFA), and poly-unsaturated (PUFA). Lipid species that had an odd number of total FA carbons, generally containing FA15:0 or FA17:0, were categorized as odd, regardless of the other FAs on the molecule. PCs, PEs, and DAGs were categorized based on the annotation of the FA with the greater number of double bonds. TAGs with an even number of total FA carbons were categorized based on the FA group with the greatest number of double bonds. In this study, we conducted targeted proteomics analysis for a focused set of 20 apolipoproteins and proteins related to HDL, LDL, and VLDL remodeling. The proteomics data were acquired with a Perfinity IDP workstation (Shimadzu Scientific) using on-line protein digestion with an immobilized enzyme reactor (IMER) directly coupled to a HALO-C18 analytical column (Advanced Materials Technology, Wilmington, DE, USA). All samples were diluted 1:100 with buffer containing 10 mM NaHCO3 and 150 mM NaCl at pH 7.4. To a 100 μL aliquot from each diluted sample, a 50 μL of digest buffer containing 0.45% Zwittergent 3–12 was added. Then, samples were mixed on a shaker plate at 500 rpm for 5 min and placed directly into the autosampler at 8 °C for subsequent digestion and MRM analysis. A detailed protocol of the procedure can be found in Toth et al. . Labeled peptide internal standards were co-injected with the sample onto the IMER (Supplementary Figure S7). Peptides were trapped on a C18 trapping column, which is subsequently switched in-line with an analytical column using the same stationary phase. Eluted peptides from the analytical column were directly analyzed by MRM on a QTRAP 6500 (AB SCIEX, Framingham, MA, USA). A dilution series of plasma-based calibrators and QCs that had been previously value-assigned for target proteins were analyzed with each sample plate, and the calibrators were used to generate calibration curves of peptide area ratio versus protein concentration. Targeted protein analysis method reproducibility was established using QC samples from pooled human plasma. Protein concentration CVs for the QC samples calculated for each protein are shown in Supplementary Table S2. Targeted lipidomics and proteomics mass spectrometry raw data processing was performed with the Lipidomics Workflow Manager (AB SCIEX, USA) and Multiquant (AB SCIEX, USA), respectively. The lipid species and protein concentration quantification were performed based on the signal intensity relative to the corresponding internal standard (Supplementary Table S4). Further data processing and formatting were performed using JMP Pro software (SAS Institute, Cary, NC, USA). Prior to statistical analysis, the lipid species and protein concentration data with more than 30% missing values and CV for QC > 30% were removed. After applying these criteria, missing values for remaining lipid species (574) and proteins (20) were imputed with one-half of the minimum value for each variable. The non-parametric Wilcoxon and Kruskal–Wallis tests implemented in JMP Pro were used for the evaluation of absolute plasma concentration differences. The false discovery rate (FDR)-adjusted q-values were calculated with the Benjamini–Hochberg procedure. Means comparison analysis of lipid and protein concentrations, cluster analysis, and lipid–protein correlation network analysis were conducted using custom R and Python scripts. Namely, the MetaboAnalyst and ggvenn R packages were used to build Venn diagrams, heatmaps, and conduct the clustering. The Pearson’s correlations, p-values, and q-values for the protein-correlation networks were calculated using the scipy.stats and statmodels Python libraries. The volcano plots were visualized using the matplotlib and seaborn Python libraries. Networks were visualized using the Cytoscape software . Pathway enrichment analysis was performed based on the biological processes defined in Gene Ontology and signaling and metabolic pathways defined in Reactome databases using the Enrichr application programming interface (API) implemented in Python . From now on, we will refer to the analysis of both the Reactome pathways and GO biological processes as “pathway analysis”. The p-value < 0.01 and q-value < 0.01 were used as the thresholds for statistical significance (unless noted). Predictor Screening procedures and the ANN modeling were conducted using the JMP Pro software. First, we assessed the overall differences in the lipid class concentrations across different metabolic conditions. Using the Lipidyzer platform we have determined and quantified a total of 574 lipid species from 11 lipid classes (Table 1). The analysis of variance (ANOVA) showed that the average concentration of lipid classes varied significantly (p-value < 0.05) across the sample groups (Figure 1A, Table 1). Moreover, 8 classes, including CE, CER, DAG, FFA, LPC, PC, SM, and TAG, showed highly significant differentiation with p-values < 0.0001 (Table 1). Clustering analysis of concentrations by lipid classes revealed three lipid-class clusters: CER/HCER, TAG/PE/FFA/DAG/LPE, and CE/SM/PC/LPC (Figure 1A). The CER and HCER levels were highest in the HT and lowest in the NL group. The TAG containing cluster showed higher concentrations in the HT and HL than in the three other groups, while the concentrations of the CE containing cluster were higher for the HC and HL sample categories. Then, we performed means comparison analysis by lipid species, comparing mean concentrations in NL samples to the HC, HT, HL, and HG samples (Figure 1B). Out of 574 species, we found that the concentrations of 267 lipids in the HC, 547 lipids in the HL, 517 lipids in the HT, and 148 lipids in the HG group were significantly different (p-value < 0.05) when compared to the NL samples (Figure 1B, Supplementary Table S3). The false discovery rate (FDR) for more than 95% of the significantly different lipid species was low, FDR < 1%, otherwise the FDR was 1–18%. Most lipid species were up-regulated relative to NL, including 93 lipid species that were elevated in all four disorder groups (HT, HC, HL, and HG), 39 species in HT, HL, and HC, 26 species in HG along with HL and HT, 252 species in the groups with high triglyceride levels (HT and HL), and 22 in groups with high cholesterol (HC and HL) (Figure 1C, Supplementary Table S5). The number of species that uniquely increased in a single sample group was relatively small and worthy of mention, specifically: in the HC group PC(18:1/16:1), PC(18:1/18:1), SM(26:0), LCER(16:0), HCER(24:1), HCER(24:0), and HCER(22:0); in the HT group TAG(42:1-FA18:1); in the HL group PC(16:0/20:4), LPC(20:3), PE(P-16:0/18:1), PE(18:0/18:1), PE(18:0/18:2), PE(P-18:1/20:4), TAG(58:8-FA20:3), DAG(16:0/18:0), TAG(42:0-FA14:0), TAG(44:0-FA12:0), TAG(58:9-FA20:4), and PE(16:0/18:1); and in the HG group DAG(18:2/20:4), DAG(18:1/20:4), and DAG(16:0/20:4). The concentration of species that were down-regulated compared to NL included LPE(20:4), LPE(18:2), LPC(18:2) in all sample groups; PC(18:2/18:2), PE(P-16:0/22:4), and HCER(24:0) in HT and HG; and TAG(54:0-FA18:0) in HC and HG; and 5 CEs, 7 PCs, 2 CERs, 2 SMs, 1 LPC, 2 TAGs in only HG (Figure 1B,C, Supplementary Tables S3 and S5). Then, we sought to determine the common patterns in changes in lipid concentrations within the same lipid class. We assessed the lipid species abundance within lipid classes based on the FA group saturation. For each sample, we summed the lipid species concentrations within lipid classes by odd-chain FA, SFA/MUFA, DUFA, and PUFA containing sub-classes and divided by the total lipid class concentrations, obtaining %Abundance values (Figure 2A). Relative to NL, odd-chain FA containing CEs and LPCs were higher. LPCs and LPEs with SFA/MUFA were higher, and those with DUFA and PUFA were lower. Correspondingly, PC, PEs, and FFAs with PUFA were higher, but those with SFA/MUFA showed no significant difference. By closer examination of the individual species, we found that increase in FA(20:4) containing PEs corresponded with a lower abundance of FFA(20:4) (Figure 2B). This is consistent with the hydrolysis of PUFA-containing PCs and PEs, especially those with FA(20:4), to SFA/MUFA LPC/LPE and PUFA FFA products. Then, we determined changes in protein concentrations associated with different metabolic conditions. As with lipid classes, we performed means comparison and cluster analysis of the protein concentrations (Figure 3A). HG samples appeared most similar to the NL samples, while HC, HL, and HT formed a separate cluster, and three major protein clusters were identified. The first cluster included apo(a), AACT, PLTP, SAA1, SAA4, HP, and apoA4, which showed higher levels of protein concentrations in HG samples as compared to other sample groups. The second cluster of apoA1, apoA2, CETP, PON1, apoD, apoM, and TF appeared up-regulated in HC samples. The third cluster included apoC1, apoB, LCAT, apoC2, apoC3, and apoE, which had higher concentrations in the HL, HC, and HT groups than in the HG and NL groups. Using one-way ANOVA analysis, we found that the average concentrations of 13 out of 20 proteins were significantly different across the sample groups (p-value < 0.05, FDR < 5%) (Table 2). Compared to the NL samples, we found that 17 out of 20 proteins were significantly different (p-value < 0.05, FDR < 5%) in at least one of the other four sample categories (Supplementary Table S3). ApoE, apoA4, and HP were generally up-regulated relative to the NL group. Apos B, C1, C2, C3, and LCAT were up-regulated in HL, HC, and HT, but not in the HG group. SAA4 and AACT were elevated in HG and HL, and SAA1 appeared up-regulated in HL and HC groups. The HG group was significantly differentiated from the NL group by relative down-regulation of apos A1, A2, and C1. In addition, there was a unique up-regulation of PLTP and down-regulation of apoM in HG, and down-regulation of apoD in HT samples (Figure 3B,C). The correlations between lipid and protein concentrations within the same sample group may indicate functional or physical lipid–protein interaction. To determine such interactions, we closely examined the concentration correlations between proteins and lipids by sample group. We designated Pearson correlation coefficients of r > |0.3| and p-value < 0.05 as significant, and r > |0.5| and p-value < 0.002 as strong correlations (Supplementary Table S6). The strength of significant correlations between proteins and lipids did not correspond with their up- or down-regulation relative to the NL group (Supplementary Figure S1). Furthermore, the strengths of correlations did not follow ranks by %Abundance in the lipid classes (Supplementary Figure S2). Instead, the numbers and strengths of the correlations showed recognizable differences among species categories (odd-chain FA, SFA/MUFA, DUFA, and PUFA) for each lipid class and protein (Supplementary Figure S3). Therefore, we concluded that the strength of the lipid–protein correlation was not the result of simple coincidental associations and may indicate the participation of correlating lipids and proteins in the same lipoproteins remodeling processes. We found that, in the NL group, apoB and apoC2 correlated strongest with odd-chain FA and PUFA CEs, while apoC3 correlated with SFA/MUFA and DUFA CEs, PCs, and TAGs (Supplementary Figure S3). A unique feature of the HC group was the negative correlation of apoA1 and apoA2 with TAGs containing mostly DUFA, PUFA, and odd-chain FAs, while much fewer SFA/MUFA. Moreover, there were positive correlations of apos B, C2, C3, and E with TAGs and PCs, with a higher preference for SFA/MUFA and DUFA TAGs and PCs. In the HT group, apos A1, B, and C1 showed positive correlations with TAGs containing mostly SFA/MUFA, while apos C2, C3, and E correlated with TAGs with a preference toward PUFA TAGs. The HL group was unique in terms of few correlations of apoA1 and apoA2 with lipid species, and instead, apoA4 was found to be correlating negatively with many DUFA and PUFA CEs, and positively with a high number of TAGs with SFA/MUFA, DUFA, and odd-chain FA. In the HG group, lipid species from all classes correlated strongly and positively with apos B, C1, C2, C3, and M, especially PUFA containing CEs, PCs, and DAGs, as well as odd-chain FA containing TAGs and FFAs. To explore the differences in the protein–lipid association among metabolic conditions, we conducted a topological analysis of protein–lipid correlation networks (Figure 4). To build the networks, we used all strongly correlating lipid–protein pairs selected in each sample group based on significant up- or down-regulation compared to the control samples (p-value < 0.05) and strong Pearson correlations of r > |0.5| with corresponding p-value < 0.002 (Supplementary Table S6). The HL group samples gave a topological network where apoE appeared with the largest sub-network of 111 positively correlating TAG species (Figure 4A). A separate sub-network of CE, SM, and TAG species was around apoA4, including long-chain CE(20:2), CE(22:4), CE(20:3), and SM(26:1). There were also strong positive correlations between apoD and PE(18:0/18:1) and between SAA4 and LPE(10:4), as well as a negative correlation between apo(a) and CER(22:0). The network analysis performed for the HC group revealed apoC2 as a hub surrounded by a sub-network of 37 TAGs, many of which also connected with apoC3 and apoE (Figure 4B). Interestingly, TAG(52:2-FA16:0) was connected to apos B, C2, C3, and E, while PC(16:0/18:1) to both apoA1 and apoC2. For the HT group, the protein–lipid network revealed apoC3 as the central protein associated with various TAG, DAG, LPE, CER, and CE species (Figure 4C), and some of the TAGs were shared by apoA2 or apoC3. ApoB and HP showed strong positive correlations (r = 0.50–0.65, p-value < 0.005) with ten CE species, and some also connected to apos A2, C3, and C1. The most unique feature of the HG group (Figure 4D) was the high number of species around apoC1 from various lipid classes including TAG, CE, CER, DAG, SM, PC, LPC, and FFA. ApoC1 also shared PC, LPC, and SM species with apoA1 and apoA2. The latter two both correlated positively with the same set of CEs and SMs. Interestingly, several DAG, CE, LPC, and SM species around apoM in the HG network were also connected to apos C1, A1, and A2. In this study, we investigated the panel of 20 apolipoproteins and proteins involved in LDL, HDL, and VHDL remodeling. Since all 20 proteins measured in our study have defined functions in lipid metabolism, pathway enrichment analysis can uncover specific metabolic processes dysregulated in different sample categories. Furthermore, the biological functions defined for apolipoproteins allow us to link lipids to metabolic pathways through the constructed protein–lipid networks. The enrichment analysis was performed as described in the Methods section using the sets of proteins that strongly correlated (r > |0.5|) with the highest number of lipid species in each sample group (Figure 4E,F, Supplementary Table S7). Since a different set of proteins emerged in each metabolic group, the output of the pathway enrichment analysis was expected to be different as well. For the HC group, apos A1/B/C2/C3/E and LCAT were uniquely associated with HDL-mediated lipid transport, very-low-density lipoprotein particle remodeling, acylglycerol homeostasis, and triglyceride homeostasis. For the HT group, apos A2/B/C1/C2/C3 was related to negative regulation of cholesterol transport and lipoprotein particle clearance. For the HG group, A1/A2/C1/M and AACT indicated the cholesterol esterification and HDL particle assembly. Several other pathways were common to all sample groups, including signal transduction, cholesterol homeostasis, and cholesterol transport. Nonetheless, the pathways that were unique to a metabolic condition can be indirectly linked to a set of lipids through the same set of proteins. Recent studies have demonstrated the power of the deep learning approach to accurately predict dyslipidemic conditions. However, in most cases, only one condition, e.g., hyperlipidemia, or combined conditions as a single “dyslipidemia” condition, were investigated . Here, we explored the feasibility to build an ANN model to differentiate all five sample categories (HC, HG, HL, HT, and NL) based on a data-driven selection of both lipid species and proteins. First, we randomly split the 146 samples into a training set (n = 102, 70%) and a test set (n = 44, 30%). The training set included 26 HC, 20 HG, 18 HL, 23 HT, and 15 NL samples. Then, we conducted 100 repeated cycles of the predictor selection using a bootstrap forest-based Predictor Screening algorithm implemented in JMP software. The top 20 most frequently selected variables were prioritized as the predictor pool (p) to build ANN models. The selected 20 predictors included 2 proteins (AACT and apoC1) and 18 lipids (Supplementary Table S8). By evaluating different types of the hidden layer and ANN architecture, we found that an optimal ANN performance can be achieved by using a fully-connected multilayer perceptron (MLP) network with five Gaussian activation function nodes in the hidden layer (Figure 5A). To determine the set of predictors that gave the highest ANN model performance, we used a step-wise systematic evaluation of all 20 predictors. The predictors P = were sorted based on their relative contribution estimated by the predictor screening algorithm. Step 1, we built ANN models by using the first predictor, p_1 as “leading predictor”, and its combinations with every other predictor p_i∈P,i ∈ . Each resulting model was evaluated in terms of the accuracy for the training and test sets (Supplementary Figure S4). The predictor combination that provided a model with the highest accuracy was kept for the next step. In step 2, a third predictor was added to the predictor set, and as in step 1, the model construction and evaluation were repeated. For the given leading predictor, this process was repeated until the addition of one more variable did not improve the model accuracy. Then, the second and each consecutive predictor (e.g., p_2, … p_20) was selected as a “leading predictor” and steps 1 and 2 were repeated. After each cycle, the models were compared based on overall model accuracy and the area under the receiver operating characteristics (ROC) curve (AUC) by each sample group. Increasing the number of predictors improved model accuracy both for the training and for the test set, but it approached a plateau at 8 predictors, in the range of 0.90–0.95 and 0.71–0.79, respectively. A total of nine models with 8 predictors were constructed. In terms of inclusion frequency, the 8-predictor combinations included 2 proteins AACT > apoC1; 4 SMs, SM(14:0) > SM(22:1) > SM(22:0) > >SM(16:0); 4 CEs, CE(18:0) > CE(18:2) > CE(16:0) > CE(20:2); 2 DAGs, DAG(18:0/18:2) > >DAG(14:0/16:0); 3 TAGs; TAG(52:2) > >TAG(51:2) > TAG(52:7); and 2 LPCs, LPC(18:2) > >LPC(18:0); HCER(24:0) and CER(24:0). The combinations of the most frequently used predictors gave an accuracy of 0.92–0.95 derived from the confusion matrix (Figure 5B) and the areas under the ROC curves (AUC) > 0.98 (Figure 5C). For the test set, the overall accuracy was 0.74–79 (Figure 5D), and the ROC AUC was > 0.90 for all sample groups (Figure 5E). The maximum accuracy was achieved for the training (0.96) and test (0.80) sets using the following 8 variables AACT, SM(14:0), SM(22:1), CE(18:0), CE(18:2) CE(16:0), LPC(18:2), and TAG(52:2). In the present study, we applied combined targeted lipidomics and proteomics approaches and gained comparative insights into multiple metabolic disorders. The correlation analysis of protein and lipid concentrations allowed us to create protein–lipid interaction networks that provided new insights into functional and physical associations between lipid species and apolipoproteins in different metabolic groups. For biological interpretation of our data, we rely on established theories of lipid homeostasis . The lipid and protein compositions, as well as the relative particle numbers of HDL, LDL, and VLDL particles, are the result of their excretion rate from cells followed by extracellular remodeling. The continuous exchange of lipid species and proteins among all lipoprotein particles and their in vivo environment leads to a dynamic equilibrium concentration of individual lipid species and proteins in a fasting state. Therefore, lipid species and protein concentrations, measurable in whole plasma or serum samples, collectively characterize the lipid homeostasis of each individual person. The HDL, LDL, and VLDL lipid species composition is the result of interconnected intracellular lipid synthesis pathways . In our data, the interconnection of lipid pathways is evidenced by the concerted up- or down-regulation of lipid class concentrations (Figure 1A). For example, elevated TAG levels along with CERs and PEs were attributes of both HT and HL, but to a lower extent of HG or HC samples. This observation suggests a unique co-regulation of the TAG, CER, and PE de novo synthesis pathways in HT and HL patients. As another example, the upregulation of both CE and SM levels indicates an interplay between CE and SM synthesis pathways in HC and HL but not in other groups (Figure 1A). Additional evidence of interconnected lipid synthesis pathways was the corresponding abundance of DUFA and PUFA species within PC and PE classes, including plasmalogen PEs (Figure 2A,C). For instance, the overall decreased abundance of FA(18:2) across HC, HT, HL, and HG groups corresponded with an increased abundance of FA(20:4)-containing species (Figure 2B). The abundance shifts of FA(20:4)-containing species also corresponded with the increased abundance of shorter-FA-chain SM(14:0) and SM(18:0) species (Figure 2D). The plasma lipidome also reflects the activity of intra- and extracellular lipases that hydrolyze PCs, PEs, and TAGs to LPCs, LPEs, and DAGs, respectively. The PUFA groups in PC, PE, and TAG species are frequently paired with SFA/MUFA groups on the backbone. Thus, the hydrolysis of the PUFA group produces SFA/MUFA-containing LPCs, PEs, and DAGs, and vice versa. In the non-NL groups, we observed an elevated abundance of PUFA-containing PCs, PEs, and TAGs along with the reduced abundance of PUFA-containing and increased abundance of SFA/MUFA-containing LPCs, LPEs, and DAGs (Figure 2A). Therefore, there is a higher preference for PUFA group hydrolysis from PCs, PEs, and TAGs in the non-NL groups relative to the NL group. As an example, FA(20:4)-containing species abundances are shown in Figure 2B. The reduced abundance of FA(20:4)-containing LPEs corresponded with the increased abundance of FFA(20:4) (arachidonic acid), a precursor of both pro- and anti-inflammatory eicosanoids . The relative PC and LPC species composition is also affected by FC esterification, intracellularly by ACAT and extracellularly by LCAT ; both enzymes transfer FA groups from PC to FC while producing CEs and LPCs. Increased intracellular ACAT activity was linked to a higher abundance of CE(18:1) , while increased extracellular LCAT activity to a higher abundance of CE(18:2) . In the HT group, we found evidence for the latter, observing a decrease in the class abundance of DUFA CEs, with an increase in the class abundance of odd-chain FA and SFA/MUFA-containing CEs (Figure 2A). The protein composition of lipoproteins fractions and sub-fractions was characterized in numerous studies . On average, small HDL particles contain two apoA1 molecules while large HDL particles contain three apoA1. or 2–3 apoA1 and two apoA2 . LDL and VLDL particles are stabilized by a single apoB molecule per particle . Considering these stoichiometric and particle size/volume constraints, the combined concentration pattern of lipids, apoA1, apoA2, and apoB is expected to provide a fingerprint that corresponds with the relative particle number and size distribution of HDL, LDL, and VLDL particles. Particle concentration and distribution by size dictate total surface area for interaction with other exchangeable apos, i.e., C1, C2, C3, D, E, and M. These proteins are in dynamic exchange among HDL, LDL, and VLDL particles. The exchange is affected by the surface affinity and penetrability by these apos, attenuated by phospholipid species composition . Altogether, the relative concentration of exchangeable apos in plasma, along with those of apoA1, apoA2, apoB, and lipid species, contributes to a metabolic fingerprint that reflects the complexity of the within and between particle interactions. The inverse correlation between apoA1 (or HDL particle number) and TAG concentrations in plasma is widely reported . It is generally explained by the concerted actions of cholesteryl ester transferase (CETP) and lipase enzymes, resulting in the TAG transfer from VLDL to HDL particles, followed by hydrolysis of HDL TAGs, and delivery of the remaining CE content to the liver by HDL and LDL particles . According to our data, this pathway is up-regulated in HC and NL groups where we observed the strongest negative correlation of apoA1 and apoA2 with TAG species, despite normal or moderately elevated TAG levels overall (Figure S3A). Interestingly, these negative correlations were significant almost exclusively with odd-chain FA, DUFA, and PUFA-containing TAGs, while correlations were few with SFA/MUFA-containing TAG species. This apparent selectivity of the TAG lowering function of HDL corroborates studies showing the TAG lowering effect of diets rich in n-3 omega fatty acids . In theory, lipid–protein pairs are expected to correlate positively when they are simultaneously involved in the formation of a pool of particles that are similar in size. In other words, at similar core volume and surface area, both protein and lipid concentration are a function of particle number, thus the average protein and lipid concentrations show linear correlation. ApoB-containing LDL and VLDL particles collectively carry more TAG and CE molecules in plasma than apoA1-containing HDL particles. Therefore, some degree of positive correlation of TAG and CE species with apoB is expected. During extracellular remodeling, if the concentration of the lipid class and the abundance of specific species within the class change at the same time, the correlation of specific TAG or CE species with apoB may vary. The strongest apoB-TAG correlations were observed in the HC and HG group (Figure 4 and Figure S3A), in particular with higher abundant PUFA-containing TAGs (Figure 2). In the HT group, apoB also correlated strongly with CEs (Figure 4 and Figure S3A), mostly with higher abundant SFA/MUFA and DUFA-containing CEs (Figure 2). The number of apos C1, C2, C3, and E per particle is higher on LDL/VLDL (apoB containing) than on HDL (apoA1 containing) . Since LDL and VLDL particles also carry more TAGs, the increase in the LDL/VLDL particle numbers corresponds with the increases in both exchangeable apos and TAG species concentrations. In support, we found significant correlations of exchangeable apos with apoB (Figure S3B), and similarly of TAG and CE species with apoB (Figure S3A). As expected, we observed positive concentration correlations of many lipid species with exchangeable apos as well, mainly TAGs with apos C1, C2, C3, and E (Figure 4 and Figure S3A). We also found that the number and relative strength of correlations were unique to each sample group. In the HC and HT groups, apoC2 and apoC3 correlated with TAG species, but there were fewer and weaker correlations in HT (Figure 4 and Figure S3A). In the HL group, only apoE correlated strongly with TAGs. In the HG group, apoC1, C2, and apoC3 correlated strongly with nearly all monitored TAG species, however very few and weaker correlations were found between apoE and TAG species. The abovementioned differences in protein and lipid concentrations and strengths of correlations observed for different donor groups can be used as evidence of differences in metabolic remodeling pathways. Some of these pathways can be identified through pathway enrichment analysis by using the sets of proteins whose concentration was the most significantly changed in dyslipidemic samples as compared to the NL group (Figure 4D). In HC, apos A1/B/C2/C3/E and LCAT proteins were associated with the dysregulation of HDL-mediated lipid transport, very-low-density lipoprotein particle remodeling, acylglycerol homeostasis, and triglyceride homeostasis. In HG, apos A1/A2/C1/M and AACT were linked to cholesterol esterification and HDL particle assembly. In HT, apos A2/B/C1/C2/C3 were related to negative regulation of cholesterol transport and lipoprotein particle clearance. The concerted up- and down-shifts in lipids and protein concentrations and the relative strength of correlations are fingerprints of intertwining metabolic processes and functions for different metabolic categories. However, the construction of a predictive statistical model to evaluate the differences among complex lipidomic and proteomic fingerprints is a challenge. Traditional multivariate analysis tools, such as partial least square, principal component, and stepwise logistic regression are insufficient for the prediction of multi-level outcomes, especially in the case of a great number of intercorrelating measures . To overcome these limitations, we turned to a deep learning ANN approach preceded by systematic data-driven predictor screening. The resulting ANN models with the highest accuracy contained several CEs, SMs, TAGs, and DAGs as significant classification factors. Interestingly, SM(14:0) was the strongest predictor among SMs, consistent with the ANOVA analysis, which showed its greatest increase in the HG group when other SMs were reduced. The odd-chain FA15:0-containing TAG(51:2) also emerged as a top marker, consistent with its variation in a wide range of 0.8–2.3 fold in all four non-NL groups. Similar significant markers were LPC(18:2) and LPC(18:0), probably due to their link with obesity and type 2 diabetes . In addition, AACT emerged as the strongest protein marker. AACT is a serine protease inhibitor and inflammatory marker, also known as SerpinA3. We found that it was elevated the most in the HG group. Another protein that emerged as a putative marker was apoC1, a potent CETP inhibitor and LCAT activator, which was reduced in HG while increased in the other four groups. In this study, we applied targeted lipidomics and proteomics to the same human serum samples to determine molecular characteristics of different metabolic conditions. The quantitative data-driven analysis of absolute concentration differences and concentration correlations allowed us to establish protein–lipid connectivity networks unique to each sample category and link them to defined metabolic pathways. These data also suggest the changes in the composition of HDL, LDL, and VLDL particles under different pathological conditions. The integration of larger sample sets combined with detailed follow-up experimental studies is needed to further validate and refine the condition–protein–lipid associations observed in this work. Furthermore, inclusion of phenotypic characteristics, such as gender, age, race, body mass index (BMI) and other parameters, may further inform lipid and apolipoprotein-based biomarker discovery in metabolic disorders. Nonetheless, our study demonstrates the existence of unique molecular fingerprints for each condition that can be uncovered through systematic evaluation of proteomics and lipidomics profiles. Leveraging the power of the machine-learning approach, we demonstrated the feasibility of defining a small set of molecular features for simultaneous categorization of each metabolic condition investigated in this work. Together, the application of our approach may improve molecular classification of lipid metabolism-related chronic diseases to inform new effective individualized therapeutic interventions.
|
PMC6629680
|
Integrated targeted metabolomic and lipidomic analysis: A novel approach to classifying early cystic precursors to invasive pancreatic cancer
|
Pancreatic cystic neoplasms (PCNs) are a highly prevalent disease of the pancreas. Among PCNs, Intraductal Papillary Mucinous Neoplasms (IPMNs) are common lesions that may progress from low-grade dysplasia (LGD) through high-grade dysplasia (HGD) to invasive cancer. Accurate discrimination of IPMN-associated neoplastic grade is an unmet clinical need. Targeted (semi)quantitative analysis of 100 metabolites and >1000 lipid species were performed on peri-operative pancreatic cyst fluid and pre-operative plasma from IPMN and serous cystic neoplasm (SCN) patients in a pancreas resection cohort (n = 35). Profiles were correlated against histological diagnosis and clinical parameters after correction for confounding factors. Integrated data modeling was used for group classification and selection of the best explanatory molecules. Over 1000 different compounds were identified in plasma and cyst fluid. IPMN profiles showed significant lipid pathway alterations compared to SCN. Integrated data modeling discriminated between IPMN and SCN with 100% accuracy and distinguished IPMN LGD or IPMN HGD and invasive cancer with up to 90.06% accuracy. Free fatty acids, ceramides, and triacylglycerol classes in plasma correlated with circulating levels of CA19-9, albumin and bilirubin. Integrated metabolomic and lipidomic analysis of plasma or cyst fluid can improve discrimination of IPMN from SCN and within PMNs predict the grade of dysplasia.Pancreas cancer (PC) is expected to become the second most common cause of cancer related death within the next decade. Contrary to other cancer types such as breast and colorectal, whose prognosis has been progressively improving over the time, pancreatic cancer prognosis remains poor. This has in part been due to the lack of an effective screening method with the ability to identify pancreatic lesions that are at risk of progression and that appear before PC develops. Pancreatic cystic neoplasms (PCNs) are increasingly diagnosed and display a prevalence as high as 45% in the general population. Intraductal Papillary Mucinous Neoplasm (IPMN) account for half of all PCNs and are increasingly considered possible precursor lesions of PC. IPMNs can progress from low-grade dysplasia (LGD) through high-grade dysplasia (HGD) to invasive cancer. However, as the prevalence of IPMNs in the general population is higher than the incidence of PC, only a minority of patients affected by IPMNs will develop PC. Therefore, the detection of IPMNs by currently available imaging techniques is an opportunity for early diagnosis of neoplastic precursor lesions and prevention of PC. However, the development of a population-based screening program is challenged by two factors: on one hand the costs of lifelong surveillance, on the other hand the low pre-operative diagnostic accuracy for pancreatic cystic lesions. These two problems partly overlap, as due to the low diagnostic yield of conventional radiology, many patients will undergo unnecessary lifelong follow-up with magnetic resonance imaging and/or endoscopic ultrasound with associated high health care costs that might become particularly unsustainable in the near future. Current indications for surgery in IPMN patients are mainly based on the pre-operative radiological imaging that suffers from low accuracy (60–70%). Diagnostic yield can be slightly increased by adding endoscopic ultrasound-guided fine-needle aspiration (EUS-FNA), which allows for cytological and carcinoembryonic antigen (CEA) analyses. Nonetheless, fluid cytology does not allow differentiation between different types of mucinous cysts or between different grades of dysplasia in IPMN, and CEA is inaccurate to discriminate benign mucinous cysts and cysts with high-grade dysplasia or an associated invasive carcinoma. Thus, the pre-operative diagnostic accuracy to distinguish between the various benign or (pre-)malignant PCNs, such as IPMNs, is low, and there are no methods available to discriminate between the different grades of dysplasia associated with IPMNs. Correctly identifying PCNs and their risk for progression to cancer is clinically crucial; as such, novel biomarkers from blood or cyst fluid may allow for a more accurate definition of IPMNs and improve their management and treatment. Metabolic reprogramming is an established hallmark of cancer. In addition to the carbohydrate and amino acid nutrients required by growing cancer cells, the lipid-scavenger pathway and de novo fatty acid synthesis are important for maintaining cancer cell proliferation and survival in the tumor environment. Development and progression of PC is associated with alterations in circulating metabolic profiles. Whilst previous studies have compared the metabolic profiles of PC patients and healthy individuals, few have examined IPMN patients or considered the spectrum of IPMN severity, which is relevant for pancreatic surgery management. This study aimed at defining the metabolomic and lipidomic makeup of pancreatic cyst fluid and plasma in pancreas resection patients with IPMN and serous cystic neoplasm (SCN). This cohort study included 35 patients undergoing pancreas resection, from whom pre-operative blood plasma (n = 21) and peri-operative cyst fluid (n = 31) were collected (Supplementary Fig. S1). Following histological validation of resected tissues, four groups were assigned: serous cystic neoplasm (SCN), IPMN with low-grade dysplasia (LGD), IPMN with high-grade dysplasia (HGD), and invasive IPMN (Cancer) for which clinical parameters are summarized in Table 1. As expected, the IPMN group was older, of mixed gender, and had comparable BMI with SCN controls. Cardiovascular disease (CVD) and diabetes were more common in patients with IPMN. Compared to SCN, IPMN LGD and HGD showed no significant elevation of circulating CA19-9, HbA1c, amylase, albumin, bilirubin, or white blood cell count. Only Cancer had significantly increased circulating CA19-9 or HbA1c levels.Table 1Patient group characteristics.Cyst fluid (n = 31)Plasma (n = 21)SCNIPMNSCNIPMNLGDHGDCancerLGDHGDCancerPatients, % (n)16.1 (5)25.8 (8)22.6 (7)35.5 (11)23.8 (5)23.8 (5)28.6 (6)23.8 (5)Female, %10050*42.9*27.3**10020*33.3*40Alcohol use, %602528.618.2402016. 760Smokers, %40009.1400040CVD, %2071.471.454.620605060Statins use, %2012.514.39.0920016.740Diabetes, %012.542.936.402016.720Age, years4866**72***69***536572.5**69**median (range)(34–58)(56–81)(66–75)(46–83)(34–68)(56–76)(66–75)(65–83)BMI, kg/m29.6427.5127.2124.9728.0132.162425.69median (range)(24.1–32.0)(21.8–36.6)(23.4–28.3)(20.2–29.7)(24.1–31.0)(24.8–36.6)(21.5–28.3)(24.1–32.9)HbA1c, mmol/mol3142.53843*33443851.5median (range)(30–37)(35–48)(31–55)(31–67)(30–43)(37–48)(31–55)(31–81)S-CA 19–9, kE/L111811376*111116285**median (range)(6.8–62)(6.4–182)(<1–115)(<1–1040)(7.9–62)(6.4–182)(<1–115)(46–480)Serum amylase, µkat/L0.30.410.240.250.310.440.1950.27median (range)(0.19–1.64)(<0.13–0.65)(<0.13–0.93)(<0.13–0.87)(0.19–1.64)(<0.13–0.54)(<0.13–0.72)(<0.13–0.54)Albumin, g/L36363631383734.531.5median (range)(33–39)(26–38)(22–39)(19–38)(33–39)(36–39)(22–39)(28–34)Bilirubin, µmol/L66.552468830median (range)(3–18)(<3–13)(<3–315)(5–150)(3–7)(4–13)(4–315)(12–119)WBC, × 10/L6.37.457.89.86.37.58.311.2**median (range)(4.4–9.2)(5–9.4)(5.6–12.9)(5–13.9)(4.4–9.2)(5.3–9.4)(7.2–12.9)(8–13.9)Statistical comparisons between each group and the control group (SCN) were made using one-way ANOVA with Dunnett’s multiple comparisons test for quantitative parameters and chi-square test for qualitative values; *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001. Patient group characteristics. Statistical comparisons between each group and the control group (SCN) were made using one-way ANOVA with Dunnett’s multiple comparisons test for quantitative parameters and chi-square test for qualitative values; *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001. Cyst fluid and plasma were profiled single-blinded using a targeted and (semi)quantitative liquid chromatography-tandem mass spectrometry (LC-MS/MS) method. A total of 90 and 91 different metabolites were measured in cyst fluid and plasma, respectively. A hierarchical clustered heatmap of the metabolomic data showed no clear grouping of metabolite profiles according to diagnose group (Fig. 1A,B). However, principal component analysis (PCA) showed that cyst fluid but not plasma from SCN was dissimilar to all other groups (Fig. 1C,D).Figure 1Metabolomic profile. Heatmap of cyst fluid (A) and plasma (B) metabolite concentrations. Projection of patient samples on the first two principal components (PCA) for cyst fluid (C) and plasma (D) datasets. Data (cyst, n = 31; plasma, n = 21) were adjusted for confounding factors and features were subsequently standardized to have a mean of zero and unit variance. Dendrograms were built using the Euclidean distance matrix and Ward’s method. Metabolomic profile. Heatmap of cyst fluid (A) and plasma (B) metabolite concentrations. Projection of patient samples on the first two principal components (PCA) for cyst fluid (C) and plasma (D) datasets. Data (cyst, n = 31; plasma, n = 21) were adjusted for confounding factors and features were subsequently standardized to have a mean of zero and unit variance. Dendrograms were built using the Euclidean distance matrix and Ward’s method. We next applied quantitative metabolic pathway enrichment analysis, using metabolite identities (see Supplementary Table S1). Because HGD and Cancer are target groups for resection, these groups were combined (HGD/Cancer). Compared to SCN, 34 enriched pathways were found in cyst fluid from HGD/Cancer and 12 enriched pathways from LGD at a significance level of 0·05 (Supplementary Fig. S2). Among these, lipid pathways appeared to dominate, e.g. phosphatidylethanolamine biosynthesis, phosphatidylcholine biosynthesis, taurine and hypotaurine metabolism, phospholipid biosynthesis, beta oxidation of very long chain fatty acids, fatty acid metabolism, oxidation of branched chain fatty acids, and sphingolipid metabolism. Several of these were also significantly enriched in plasma samples of HGD/Cancer or LGD compared to SCN, including sphingolipid metabolism, phosphatidylethanolamine biosynthesis, phosphatidylcholine biosynthesis (Supplementary Fig. S2). As metabolic analysis of IPMN indicated altered lipid metabolism, and considering the pancreatic exocrine function of secreting lipases that might be affected in PCN patients, we next performed a high definition lipid profiling of paired aliquots using the SCIEX Lipidyzer™ technology, where 1100 lipid molecular species were measured in all the samples. Out of those 1100 measured lipid molecular species, we successfully detected and quantified a total of 430 in cyst fluid and 941 in plasma. Heatmap visualization showed that triacylglycerol (TAG)-related lipids are the most abundant type among all lipid classes in both cyst fluid and plasma (Fig. 2A,B). Lipidomic profiles of HGD and Cancer appear to be similar to each other, while LGD was clustered slightly differently. Although the lipidomic profiles of plasma and cyst fluid of the SCN group, as projected on the first two principal components (PCA), show a different direction and shape of the clouds of points when compared to IPMN groups, no evidently clear separation of diagnose groups can be observed from the 2D plot (Fig. 2C,D). The lipidomic analyses suggest alterations in lipid compound composition in cyst fluid and plasma, pointing to the possibility of discrimination of pancreatic disease severity, confirming the pathway enrichments we observed. Looking at fold change estimations (Fig. 3 and Supplementary Table S1) it is possible to observe a clear alteration of the TAG class in plasma samples in both LGD and HGD/Cancer compared to SCN. In IPMN cyst fluid samples, instead, free fatty acids (FFA) and ceramides (CER) appear to have, on average, higher concentrations than those of SCN samples while having lower amount of TAGs. Interestingly, the profile of the Cancer group is similar to that of LGD in plasma and to that of HGD in cyst fluid, while only TAGs differ significantly between HGD and LGD in both plasma and cyst fluid (Supplementary Fig. S3).Figure 2Lipidomic profile. Heatmap of cyst fluid (A) and plasma (B) lipid concentrations. Projection of patient samples on the first two principal components (PCA) for cyst fluid (C) and plasma (D) datasets. Data (cyst, n = 31; plasma, n = 21) were adjusted for confounding factors and features were subsequently standardized to have a mean of zero and unit variance. Dendrograms were built using the Euclidean distance matrix and Ward’s method.Figure 3Estimated fold changes of concentrations of all measured analytes (including metabolite and lipid molecular species) between selected groups. LGD compared to SCN in cyst fluid (A) and plasma (B). HGD/Cancer compared to SCN in cyst fluid (C) and plasma (D). HGD/Cancer compared to LGD in cyst fluid (E) and plasma (F). Filled dots are fold changes whose credibility interval does not overlap with the null reference value of one-fold change, or zero on the plotted log scale. Lipidomic profile. Heatmap of cyst fluid (A) and plasma (B) lipid concentrations. Projection of patient samples on the first two principal components (PCA) for cyst fluid (C) and plasma (D) datasets. Data (cyst, n = 31; plasma, n = 21) were adjusted for confounding factors and features were subsequently standardized to have a mean of zero and unit variance. Dendrograms were built using the Euclidean distance matrix and Ward’s method. Estimated fold changes of concentrations of all measured analytes (including metabolite and lipid molecular species) between selected groups. LGD compared to SCN in cyst fluid (A) and plasma (B). HGD/Cancer compared to SCN in cyst fluid (C) and plasma (D). HGD/Cancer compared to LGD in cyst fluid (E) and plasma (F). Filled dots are fold changes whose credibility interval does not overlap with the null reference value of one-fold change, or zero on the plotted log scale. Accurate classification of IPMN severity using novel biomarkers in cyst fluid or plasma may facilitate the discrimination of low-risk from high-risk patients. We therefore assessed the predictive capacity of the integrated metabolome and lipidome profiles to classify samples according to their corresponding disease group. As IPMN HGD and Cancer are considered as high-risk lesions, we combined these into a single group. Effects of clinical covariates (Table 1) were estimated and subtracted from the raw data prior to analysis, and the only covariates that improved the classification model were age and BMI. The result of binary classifications and the performance of the CPPLS-DA model are given in Table 2 and Supplementary Table S2. The model discriminated between SCN and IPMN with very high accuracy (100%) in both cyst fluid and plasma samples. Choline, 2-aminoisobutyrate, trimethylamine n-oxide, glycine, alanine, and glyceraldehyde were found to be essential discriminatory molecules in both cyst fluid and plasma. Furthermore, dimethylglycine was a discriminatory compound for cyst fluid while serine and GABA were for plasma. Overall, the two biofluids displayed similar predicting power, with cyst fluid-based classification performing slightly better (accuracy of approximately 90%) when classifying the three groups SCN, LGD and HGD/Cancer. Nevertheless, plasma-based classification could easily detect LGD samples (90% accuracy) while cyst fluid molecules’ concentrations were better for predicting HGD/Cancer samples (90% accuracy) (Table 2). The model had a low sensitivity in discriminating the three IPMN groups from each other when HGD and Cancer were considered separately. SCN samples form a distinct cluster from IPMN samples, whereas the other two groups overlap significantly in cyst fluid (Fig. 4A,B). The top 15 molecules in cyst fluid and plasma ranked by their VIP scores are presented in Fig. 4C,D. Interestingly, only a subset of metabolites, without lipids, were sufficient to achieve best performance. In particular, amino acids were the most important molecules for the classification of plasma samples.Table 2Performance measures of binary classifications with the chosen CPPLS-DA model.AUCSensitivitySpecificityBalanced accuracySCN vs AllCyst fluid1.0001.0001.0001.000Plasma0.9500.8000.8750.837LGD vs AllCyst fluid0.9350.8750.9130.894Plasma0.8251.0000.8120.906HGD-Cancer vs AllCyst fluid0.9490.8890.9230.906Plasma0.8540.6361.0000.818IPMN vs SCNCyst fluid1.0001.0001.0001.000Plasma1.0001.0001.0001.000Area Under the ROC Curve; (Sensitivity + Specificity)/2.Figure 4Canonical Powered Partial Least Squares and Discriminant Analysis (CPPLS-DA) results. Projection of patient samples on the first two principal components in cyst fluid (A) and plasma (B). Highest variable importance in projection (VIP) scores in cyst fluid (C) and plasma (D). Performance measures of binary classifications with the chosen CPPLS-DA model. Area Under the ROC Curve; (Sensitivity + Specificity)/2. Canonical Powered Partial Least Squares and Discriminant Analysis (CPPLS-DA) results. Projection of patient samples on the first two principal components in cyst fluid (A) and plasma (B). Highest variable importance in projection (VIP) scores in cyst fluid (C) and plasma (D). We next asked whether there was any significant correlation between the metabolites/lipids and the blood markers that are frequently used to define the IPMN patients. Pearson correlation analysis identified molecules in plasma showing strong correlations (r > 0.6 and adjusted p-value ≤ 0.05) with circulating CA19-9 and albumin, and to some extent with bilirubin (Supplementary Table S3). CA19-9 was found to be positively correlated with a total of 35 lipid molecules that were classified as CER (n = 1), FFA (n = 4), Phosphatidylcholines (PCs) (n = 10), Phosphatidylethanolamines (PE) (n = 10), Sphingomyelins (SM) (n = 4), and TAG (n = 5) (Supplementary Table S4a). In addition, albumin and bilirubin levels were positively correlated with 4-pyridoxate, adenine, carnitine, cysteine, and lipid molecules classified as CE (n = 1), FFA (n = 2), Lysophosphatidylcholines (LPC) (n = 2), Lysophosphatidylethanolamines (LPE) (n = 1), PE (n = 1), CER (n = 1), PC (n = 1), and TAG (n = 1) (Supplementary Table S4a). Negative correlations were noted between albumin and cystathionine, D-Ribose-5-P, inosinic acid or inosine monophosphate (IMP), and taurochenodeoxycholate and some lipids within the classes of PC (n = 9), TAG (n = 6) and CER (n = 6), while bilirubin negatively correlated with few metabolite and lipids (Supplementary Table S4b). Pancreatic IPMNs are common precancerous lesions. Today, only some radiological and clinical parameters are used to identify patients with high risk for cancer progression or malignancy, for example, main pancreatic duct dilatation, cyst diameter, rate of progression and elevated serum CA19-9. Unfortunately, no high-accuracy tools are available that determine the IPMN-associated grade of dysplasia or that offer accurate differential diagnosis from other benign and low-risk PCNs (i.e. SCNs). These diagnostic limitations negatively affect patient management and treatment. Until recently, metabolites involved in IPMN disease progression have been scarcely studied. A holistic view of the plasma and cyst fluid metabolic profile may aid the discovery of biomarkers capable of improving pre-operative PCN diagnosis. We have shown that an integrated metabolomics and lipidomics approach can be used to 1) discriminate between IPMN and SCN and 2) determine the IPMN-associated grade of dysplasia. Our analysis offered superior predictive accuracy compared to conventional cross-sectional imaging or EUS-FNA. The LOOCV balanced accuracy of discriminating Cancer/HGD from SCN and LGD was 90.6% for cyst fluid and 81.8% for plasma. When discriminating IPMN as a whole from SCN, accuracy reached 100% for both plasma and cyst fluid. The availability of accurate plasma-based tests could represent a major advantage for patients who do not require invasive procedures like EUS-FNA, which are associated with risk of complications and low-diagnostic accuracy. While previous PC metabolome studies measured around 50-100 metabolites per case, our high-definition integrated approach measured around 100 metabolites from 15 different biological classes and 1000 lipid molecular species from 13 different lipid classes, largely covering the important metabolome spectrum, i.e. sugars, nucleotides, amino acids, and lipids. Recent elegant metabolome studies pointed out that a number of compounds, including very long-chain fatty acids, phospholipids, and taurine, were differentially present in PC patients or PC tissue compared to healthy subjects or parenchyma tissue, respectively. This agrees with our findings of enriched taurine and fatty acid metabolism pathways and phospholipid biosynthesis pathways in cyst fluid of pre-malignant or early malignant cases, e.g. and LGD and Cancer/HGD, as compared to SCN. Moreover, our study also has shown significant alterations of different classes of molecules, mainly TAGs, detected in plasma of Cancer/HGD, and LGD, compared to SCN. We did not observe significant alterations of molecules in plasma when comparing HGD/Cancer with LGD, suggesting these disease groups display a more comparable plasma profile. The tumor marker CA19-9 is used for predicting malignancy of IPMN and monitoring PC progression, but its use as a definitive diagnostic marker, especially detecting IPMN HGD, is limited. Combining additional blood markers with CA19-9 was recently shown to improve early detection of PC, and building a broader molecular profile around CA19-9 in IPMN patients may enhance the diagnostic accuracy. The lipid metabolites strongly associated with CA19-9 in this cohort are therefore of interest and need to be examined further, possibly together with metabolites correlating with serum albumin and bilirubin. Our findings that plasma, but not cyst fluid metabolites, strongly correlated with these three markers, suggest that systemic, rather than local factors, may have an influence on development and progression of IPMN. A strength of this paper is the investigation of a surgical cohort of patients, with definitive histology and assessment of the grade of dysplasia. This allowed us to accurately match the actual IPMN-associated grade of dysplasia with our data, avoiding problems of misdiagnosis that occur in more than one third of the patients undergoing EUS-FNA with cyst fluid analysis. In addition, we used a validated targeted and (semi) quantitative analysis through a robust and reliable LC-MS/MS approach with strict quality management. We furthermore tested several linear mixed models to assess many covariate parameters (see Table 1), and while use of statins was considered as possible confounding factor, it did not improve the final classification model which only adjusted for age and BMI. However, this study has also some limitations. Firstly, there is possible selection bias because we analyzed a small and homogenous cohort of patients which might not be fully representative of the entire population and thus might potentially restrict the predictive power of lipidomic/metabolomics profiling to certain groups within the general population. In this relatively small cohort all SCNs cases are female, which is not surprising as the prevalence rate in the general population for SCNs is 9-16% of all cystic lesions while approximately 75% of patients with SCNs are women. Although this gender imbalance can be certainly considered a limitation we observed that, after adjusting for BMI and age, the effect of sex did not have any impact on out-of-sample model prediction accuracy. Therefore, we believe that our classification accuracy is mainly a consequence of differences between diseases while we expect unbalances in sex distribution to have only a very negligible influence. Moreover, we did not include other types of mucinous tumors of the pancreas, such as mucinous cystic neoplasm (MCN). However, such lesions are easier to recognize, due the peculiar epidemiological and radiological features (typically in body and tail of the pancreas, almost exclusively in young females) that make diagnosis not particularly challenging. Additionally, the current study might suffer from a possible sampling bias, considering that fluid was aspirated from one or two accessible cysts, despite IPMNs often being multifocal and occurring in different locations of the pancreas (head/body/tail). Therefore, one cannot exclude the possibility that metabolic profiles of the sampled cysts might not be representative of all cystic lesions. Lastly, we did not investigate possible factors that might have potential effects on (lipid) metabolism, such as specific genetic mutations (e.g GNAS or KRAS gene). This study has comprehensively mapped the metabolite and lipid makeup of cyst fluid and plasma from PCN patients with defined pathology, using integrated metabolomics and lipidomics. Our findings have clinical implications and may support assay development for differential diagnostics of PCNs to improve patient management. Future studies are needed to test larger patient cohorts using the proposed model, to better understand associations between metabotypes and IPMN malignancy risks. In this prospective cohort study, patients undergoing pancreatic surgery for suspected pancreatic cystic neoplasm (PCN) with post-surgically validated intraductal papillary mucinous neoplasm (IPMN) and serous cystic neoplasm (SCN) from February 2016 to January 2017 at Karolinska University Hospital, Sweden, were included. Excluded were cases without a cystic component, non-IPMNs, or those without cyst fluid in the resected pancreas (Supplementary Fig. S1). This study follows the Helsinki convention and good clinical practice with permission of the Ethical Review Board Stockholm and the Karolinska Biobank Board (Dnr 2015/1580-31/1). Written informed consent was obtained from all patients. Fresh resection specimens were received at the pathology laboratory within 20 minutes of surgical removal, in sterile conditions and on ice. Macroscopic assessment to identify the cystic lesion and main pancreatic duct was done by a specialist pancreatic pathologist. Fluid from the main pancreatic duct was collected using a syringe without needle. When the cystic lesion was readily identified in the intact specimen, the fluid was aspirated using a syringe with needle. For specimens in which the cystic lesion was not readily accessible from the surface the specimen was cut or when the cyst content was too viscous content was aspirated using a syringe without needle. Aspirated fluid was stored at −80 °C until further analysis. Venous whole blood was collected in K2 EDTA Vacutainer® tubes (BD, Stockholm, Sweden) immediately prior to surgery. Within four hours of collection, blood was processed using Ficoll Paque Plus (GE Life Sciences, Uppsala, Sweden) gradient-density centrifugation following manufacturer’s instructions and the plasma fraction was stored at −80 °C until further analyses. Resection specimens were fixed in 4% formaldehyde and processed for routine histopathological diagnosis. The cystic lesions were classified by light microscopic examination of hematoxylin-eosin stained slides by a specialized pancreatic pathologist as IPMN or SCN. The grade of dysplasia in IPMN was assessed using a 2-grade (high/low) scale, according to current international standard. To make the cyst fluid classification more representative of the neoplastic epithelium that produces it, specimens showing <5% high-grade dysplasia (HGD) were classified as low-grade dysplasia (LGD). Specimens with concomitant invasive carcinoma were classified as “Cancer” and considered as a separate class for further analyses. All metabolite standards used in the analysis were purchased from Sigma-Aldrich (Helsinki, Finland), while isotopically labelled metabolite internal standards (IS) were obtained from Cambridge Isotope Laboratory (Tewksbury, MA, USA). For lipidomics, kits containing 50 labelled internal standards across 13 lipids classes were purchased from SCIEX (Framingham, MA, USA). Ammonium formate, ammonium acetate, and ammonium hydroxide were obtained from Sigma-Aldrich (Helsinki, Finland). Formic acid (FA), acetonitrile (ACN), methanol (HiPerSolv CHROMANORM, LCMS grade), ethyl acetate (HPLC grade), 2-propanol, 1-propanol, and dichloromethane were purchased from VWR International (Helsinki, Finland). Deionized water, up to a resistivity of 18 MΩ⋅cm, was purified with a Barnstead Easypure RoDi water purification system (Thermo Scientific, Marietta, OH, USA). Whole blood was purchased from the Finnish Red Cross blood service (Helsinki, Finland) from which serum samples were prepared and used as internal quality control samples. Metabolomic analysis of samples was performed using liquid chromatography-mass spectrometry as previously described in the supplementary data of Nandania et al.. Briefly, 15 labeled internal standards were used to estimate quantitative levels of 100 metabolites. To 100 µL thawed sample (plasma or cyst fluid), 10 µL of labelled internal standard mixture was added, and then metabolites were extracted using protein precipitation by adding acetonitrile +1% formic acid (1:4, sample:solvent). The collected extracts were dispensed in Ostro 96-well plates (Waters Corporation, Milford, USA) and filtered by applying a vacuum at a delta pressure of 300–400 mbar for 2.5 min on robot’s vacuum station. Filtered sample extract (5 µL) was injected in an Acquity UPLC-system coupled to a Xevo TQ-S triple quadrupole mass spectrometer (Waters Corporation, Milford, MA, USA) which was operated in both positive and negative polarities with switching time of 20 milliseconds. Multiple Reaction Monitoring (MRM) acquisition mode was selected for the quantification of metabolites. MassLynx 4.1 software was used for data acquisition, data handling and instrument control. Data processing was done using TargetLynx 4.1 software. Lipids were extracted with liquid-liquid extraction (LLE) method using ethyl acetate and methanol. In borosilicate glass tubes, to 100 µL thawed sample (plasma or cyst fluid), 1 mL methanol and 1 mL water was added. Then, 100 μL of labelled internal standard mixture (prepared as per SCIEX LIPIDYZER manual’s instructions) was added and allowed to equilibrate with the samples. To each tube 3.5 mL of ethyl acetate was added after which tubes were put on a rotator shaker for 15 min at 30 RPM, followed by centrifugation at 3000 RPM for 10 min. After centrifugation, the upper layer of ethyl acetate was collected and dried under N2 gas. Dried samples were reconstituted with 250 μL of mobile phase (dichloromethane:methanol (50:50) containing 10 mM ammonium acetate) for injection. Lipid separation and quantitation was performed on the SCIEX Lipidyzer platform using a SCIEX 5500 QTRAP® mass spectrometer (SCIEX, Washington, D.C., USA) with SelexION® Differential ion mobility (DMS) technology by directly infusing 50 μL of extracted samples with a mobile phase at flow rate of 70 µL/min. Two acquisition methods, with and without SelexION® technology, were used to cover 13 lipid classes using flow injection analysis. The lipid molecular species were measured using MRM strategy in both positive and negative polarities. Positive ion mode was used for the detection of lipid classes – sphingomyelins (SM), diacylglycerols (DAG), cholesteryl esters (CE), ceramides (CER), triacylglycerols (TAG), and negative ion mode was used for the detection of lipid classes – lysophosphatidylethanolamines (LPE), lysophosphatidylcholines (LPC), phosphatidylcholines (PC), phosphatidylethanolamines (PE) and free fatty acids (FFA). Lipidomics Workflow Manager software was used for acquisition of samples, automated data-processing, signal detection and lipid species concentration calculations. All data analyses were performed with R 3.5.1 and Stan 2.17.1. Concentration values in μmol/L were assumed to follow a lognormal distribution and were therefore log-transformed as a preliminary normalization operation. Missing values were removed from the dataset following the “modified 80% rule”, according to which a variable is discarded if the relative frequency of missing values is more than 0.8 in all clinical groups. Remaining missing values were imputed with the QRILC function from R package imputeLCMD. Considering the lack of balance for basic clinical parameters (Table 1), all values were adjusted for the effects of confounding factors using a linear mixed model. Let be the log-concentration of molecule j observed at sample i belonging to phenotypic group g. The basic assumption of our model is that observations are conditionally independent and normally distributed, with same standard deviation σ but different mean value μ:[12pt] $$_|_,\, N(_,\,^)$$|μgij,σ~N(μgij,σ2) Covariates to include in the model were selected by the highest out-of-sample point-wise predictive accuracy. Parameter μ was then defined as the linear combination[12pt] $$_= +_+_+_+_ag_+_bm_$$μgij=θ+φi+λj+ηgj+γ1jagei+γ2jbmiiwhere θ is the grand mean, is the general effect of sample i, is the effect of molecule j, is the effect of phenotypic group g on molecule j, and are effects varying with molecules. Fold change of molecule j between groups g and h was therefore defined as . All coefficients associated with the same discrete category are constrained to sum to zero in order to make the model identifiable. We assigned to each unknown parameter weakly informative prior distributions as follows:[12pt] $$ t(3,\,0,\,10)$$θ~t(3,0,10)[12pt] $$_ N(0,\,^),\,i=1, ,n$$φi~N(0,α2),i=1,…,n[12pt] $$_ N(0,\,^),\,j=1, ,\,m$$λj~N(0,β2),j=1,…,m[12pt] $$_ N(0,_^),\,v=1,\,2,\,j=1, ,\,m$$γvj~N(0,ωv2),v=1,2,j=1,…,m[12pt] $$_ N(0,_^),g=1, ,\,k,\,j=1, ,\,m$$ηgj~N(0,ζg2),g=1,…,k,j=1,…,m[12pt] $$ halft(3,0,10)$$σ~halft(3,0,10)[12pt] $$ halft(3,0,10)$$α~halft(3,0,10)[12pt] $$ halft(3,0,10)$$β~halft(3,0,10)[12pt] $$_ halft(3,0,10),\,v=1,\,2$$ωv~halft(3,0,10),v=1,2[12pt] $$_ halft(3,0,10),\,g=1, ,\,k$$ζg~halft(3,0,10),g=1,…,kwhere t refers to the three-parameters Student’s t-distribution and halft to the same distribution but truncated at 0 and defined only on the positive values. Total number of phenotypic groups k depended on the particular statistical analysis being conducted. Estimated fold changes and their corresponding 95% credibility intervals, computed from 20,000 posterior samples, are available in Supplementary Table S1. Prior to visualization, classification, and enrichment analyses, the dataset was adjusted for the confounding covariates “age” and “BMI” and subsequently standardized to a mean of zero and unit variance. Principal Component Analysis (PCA) was applied to the adjusted data for exploratory data purposes. Heatmaps and data projection on the first two principal components were used to visualize the dataset. Classification was performed using a Canonical Powered Partial Least Squares Discriminant Analysis (CPPLS-DA), fitted with the pls R package. Classification performance was measured with a Leave-One-Out Cross Validation (LOO-CV) strategy, and balanced accuracy (average between sensitivity and specificity of the classifier) is reported. Best explanatory molecules were selected according to their Variable Importance in Projection (VIP) ranking scores according to the following iteration scheme. At each step of the algorithm the performance of the model was recorded with a LOO-CV strategy and the molecules were sorted according to their VIP score. Subsequently, 5% of the molecules with the lowest VIP score were discarded and this operation was repeated until the number of molecules allowed model identifiability. The model with the highest performance was ultimately selected. Pathway enrichment analysis (QEA) was performed with the free web service MetaboAnalyst 4.0. All figures were generated in R 3.5.1. This study follows the Helsinki convention and good clinical practice. This study was conducted at Karolinska University Hospital under permission of the Ethical Review Board Stockholm and the Karolinska Biobank Board (Dnr 2015/1580-31/1). Written informed consent was obtained from all patients.
|
PMC6128551
|
Quantitative lipidomic analysis of mouse lung during postnatal development by electrospray ionization tandem mass spectrometry
|
Lipids play very important roles in lung biology, mainly reducing the alveolar surface tension at the air-liquid interface thereby preventing end-expiratory collapse of the alveoli. In the present study we performed an extensive quantitative lipidomic analysis of mouse lung to provide the i) total lipid quantity, ii) distribution pattern of the major lipid classes, iii) composition of individual lipid species and iv) glycerophospholipid distribution pattern according to carbon chain length (total number of carbon atoms) and degree of unsaturation (total number of double bonds). We analysed and quantified 160 glycerophospholipid species, 24 sphingolipid species, 18 cholesteryl esters and cholesterol from lungs of a) newborn (P1), b) 15-day-old (P15) and c) 12-week-old adult mice (P84) to understand the changes occurring during postnatal pulmonary development. Our results revealed an increase in total lipid quantity, correlation of lipid class distribution in lung tissue and significant changes in the individual lipid species composition during postnatal lung development. Interestingly, we observed significant stage-specific alterations during this process. Especially, P1 lungs showed high content of monounsaturated lipid species; P15 lungs exhibited myristic and palmitic acid containing lipid species, whereas adult lungs were enriched with polyunsaturated lipid species. Taken together, our study provides an extensive quantitative lipidome of the postnatal mouse lung development, which may serve as a reference for a better understanding of lipid alterations and their functions in lung development and respiratory diseases associated with lipids.The lung is composed of more than 40 different pulmonary cell types, whose cellular membranes are enriched with lipids that perform a variety of functions including maintenance of the lung architecture [1–3]. In addition, alveolar epithelial type II cells of the pulmonary epithelium that are lining the alveolar surface synthesize and secrete surfactant into the alveolar space. Pulmonary surfactant is a complex mixture of lipids (phospholipids, triglycerides, fatty acids and cholesterol, etc.), surfactant proteins (A-D) and a small amount of carbohydrates. The majority of the pulmonary lipids comprise glycerophospholipids (GP) in which phosphatidylcholine (PC) is a predominant lipid class making up to 50% of the phospholipids. Phosphatidylethanolamine (PE) makes up to 20% of the lipids and phosphatidylserine (PS), phosphatidylinositol (PI) and phosphatidylglycerol (PG) constitute 12%-15% of the total phospholipid pool [4, 5]. Pulmonary lipids are important and diverse biomolecules that are involved in many biological processes. The thus far known functions of the major lung lipids include 1) prevention of alveolar collapse and preservation of bronchiolar patency , 2) improvement of mucociliary transport , 3) involvement in innate immunity and viral protection [8, 9], 4) action as potent intracellular signalling molecules in lung inflammation , 5) involvement of lipid mediators like leukotrienes, lipoxins and prostaglandins in specific reactions of inflammation and immunity [11, 12], 6) suppression of the proliferation, immunoglobulin production and cytotoxicity of lymphocytes . In fact, alterations in whole lung lipid composition and/or deficiency of pulmonary surfactant lipids are closely associated with a) respiratory distress syndrome (RDS) , b) bronchopulmonary dysplasia (BPD) [13, 14], c) asthma , d) chronic obstructive pulmonary disease (COPD) [15, 16], e) cystic fibrosis , f) pneumonia [17, 18], g) lung injury , h) cancer and in other lung diseases [6, 21]. High heterogeneity of lung tissue (e.g., bronchial versus alveolar regions) and differences in the lipid composition of individual pulmonary cell types create a complex mixture of lipid classes and molecular species associated with a set of potential complications. Despite these difficulties, several studies analyzed the lipid composition of lung tissue in different mammalian species such as in pig , rat , rabbit , monkey , dog , bovine , mice and human and also compared the lung lipid composition among different mammalian species [28, 29]. Actually, the major lipid classes were similar among different mammalian species and minor differences were observed only in the class of phospholipids. Interestingly, surfactant has proven to be highly diverse across species in its molecular design, especially in the concentration of individual surfactant proteins and its GP profile [30, 31]. Further, most of these studies focused on the analysis of PC, which are the prominent class of lipids in whole lung tissue and pulmonary surfactant. Furthermore, among the analyzed PC, dipalmitoylphosphatidylcholine (DPPC; PC 32:0), a species with two saturated acyl chains, is believed to be a major compound of the pulmonary surfactant . However, recent studies on homeothermic/heterothermic mammalian species surfactant showed that DPPC is not the only major surfactant phospholipid component. In addition to major PC lipid molecular species, there are minor lipid classes such as PG and PI, which also play an important role in lung biology [33, 34]. Lung development in the majority of mammalian species (e.g., rat, mice and humans) continues postnatally. One important aspect of postnatal lung development is alveolarization, a process, in which the total number of terminal gas exchange units increase total size and surface area of the lung [35–37]. To address the lipidomic changes in the fetal and postnatal lung development, various studies were conducted in several mammalian species (e.g., rat, rabbit, lamb, pig and human) and in birds (e.g., duck and chicken) by using morphometric and biochemical approaches [35–41]. However, the existing lipidomic studies of postnatal lung development primarily focused on the composition of surfactant PC and very few PG lipid species [31, 40, 41]. Dautel et al., recently reported postnatal developmental changes of mouse lung using a multi-omics approach . These data are consistent with our results, but in comparison to their study (day7, day14 and 6–8 week old animals), we have used a wider time frame between birth (new born) and the 12 week of life. Furthermore, the highly abundant cholesterol, as well as cholesteryl ester species were not analysed in the Dautel et al., study during postnatal development. Moreover, we performed direct infusion lipidomics using triple-quadrupole MS analytical setup and provided the quantitative information (nmol/mg wet weight) of individual lipid species of possible major lipid classes during postnatal development. In our study, we employed electrospray ionization tandem mass spectrometry (ESI-MS/MS) to 1) investigate total lipid quantity and 2) to perform a detailed analysis of lipid classes (PC, LPC, PG, PS, PE, PE P, PI, SM, Cer, HexCer, CE, and cholesterol), and 3) the composition of individual lipid species (e.g. PC 32:0, PC 32:1 etc.), and 4) to analyse their distribution pattern based on the carbon chain length (total number of carbon atoms). Additionally, we analysed the degree of unsaturation (total number of double bonds) of whole lung homogenates in newborn, 15-day-old and 12-week-old adult mice in order to provide a detailed lipidomic information during the postnatal development. Our current study provides an extensive quantitative lipidome of mouse whole lung, which may serve as a reference for a better understanding of the development of lung and molecular mechanisms underlying various pulmonary diseases associated with the lipid alterations. Unless otherwise mentioned, all chemicals were procured from Sigma-Aldrich (Deisenhofen, Germany). Phospholipid standards were obtained from Avanti Polar Lipids (Alabaster, AL, USA). Cholesterol and cholesteryl ester standards of purity greater than 95% were obtained from Sigma (Taufkirchen, Germany). High purity cholesterol-(25, 26, 26, 26, 27, 27, 27-d7) was purchased from Cambridge Isotope Laboratories (Andover, MA, USA). HPLC grade solvents methanol and chloroform were obtained from Merck (Darmstadt, Germany). Analytical grade ammonium acetate and acetyl chloride were obtained from Sigma-Aldrich (Buchs, Switzerland). All other reagents used were of high purity and analytical grade. Twelve-week-old adult male mice, 15-day-old males and pregnant females of C57BL/6J genetic background were obtained from Charles River, Sulzfeld, Germany. Mice were kept on a normal laboratory diet and water ad libitum and housed in cages under standardized environmental conditions (12 hours light/dark cycle, 23°C ± 1°C and 55% ± 1% relative humidity) at the central animal facility of the Justus Liebig University Giessen, Germany. After delivery of the newborn pups in the morning, they were taken directly out of the animal facility. 15-day-old and 12-week-old adult mice were killed by cervical dislocation and the newborn pups were killed by decapitation. All experiments with laboratory mice were approved by the governmental ethics committee for animal welfare (Regierungspräsidium Giessen, Germany, permit number: V 54–19 C 20/15 c GI 20/23). The newborn (P1), 15-day-old (P15) and 12-week-old adult (P84) male mice fur was vertically incised from the pelvis to the mandibles and removed to both sides. The abdomen was opened and a bilateral pneumothorax was produced by puncturing the abdominal surface of the diaphragm. The sternum was cut in the middle and the thorax was opened with a thorax spanner. The lungs were isolated carefully and snap-frozen immediately. Fresh snap-frozen lungs from P1, P15 and P84 male mice were homogenized with a Precellys homogenator (peQlab Biotech GmbH, Erlangen, Germany) at a concentration of 50 μg wet weight per μL. Homogenate corresponding to 2 mg wet weight was used for extraction, and lipids were extracted according to the procedure described by Bligh and Dyer . Upon phase separation, the chloroform phase was transferred to a fresh tube and dried under a stream of nitrogen gas. For each lipid class (except for SM and PE P), two naturally not occurring lipid species were added as internal standards, to compensate for variations in sample preparation and ionization efficiency. PC 14:0/14:0 (28:0), PC 22:0/22:0 (44:0) for PC and SM, PE 14:0/14:0 (28:0), PE 20:0/20:0 (40:0) for PE and PE based plasmalogens (PE P), PS 14:0/14:0 (28:0), PS 20:0/20:0 (40:0), PG 14:0/14:0 (28:0), PG 20:0/20:0 (40:0), PI 16:0/16:0 (32:0), LPC 13:0, LPC 19:0, Cer 14:0, Cer 17:0 (d18:1/17:0), cholesterol-d7, CE 17:0 and CE 22:0 internal standards were added for the analysed lipid classes. Lung homogenates were subjected to lipidome analysis by electrospray ionization-tandem mass spectrometry (ESI-MS/MS) in positive-ion mode as described [43–46]. Briefly, the samples were analyzed on a triple-quadrupole mass spectrometer (Quattro Ultima, Micromass, Manchester, UK) by direct flow injection analysis using an autosampler (HTS PAL, Zwingen, Switzerland) and a binary pump (Model 1100, Agilent, Waldbronn, Germany) with a solvent mixture of methanol containing 10 mM ammonium acetate and chloroform (3:1, v/v). A flow gradient was performed starting with a flow of 55 μL/min for 6 seconds followed by 30 μL/min for 1 minute and an increase to 250 μL/min for another 12 seconds. The mass spectrometer was equipped with electrospray ionization and operated in positive-ion mode using following tune parameters as capillary voltage, 3.5 kV; cone voltage, 110 V; collision energy, 30 V; collision gas, argon at a pressure of 0.13 Pa. A precursor ion of m/z 184, which is specific for phosphocholine-containing lipids, was used for the analysis of phosphatidylcholine (PC), lysophosphatidylcholine (LPC) and sphingomyelin (SM) lipid species [43, 44]. Neutral loss scans of 141 and 185 were used for the phosphatidylethanolamine (PE) and phosphatidylserine (PS) respectively . Fragment ions of m/z 364, 390, and 392 were used for the quantification of PE P-16:0, PE P-18:1 and PE P-18:0 plasmalogens according to Zemski et al. . Neutral loss scans of 189 and 277 were used for the ammonium adducts of phosphatidylglycerol (PG) and phosphatidylinositol (PI) respectively . Sphingosine (d18:1) based ceramides (Cer) were analysed using a fragment ion of m/z 264 . Cholesterol and cholesteryl esters (CE) were quantified using a fragment ion of m/z 369 after selective derivatization of cholesterol using acetyl chloride . After identification of relevant lipid species, selective ion monitoring analysis was performed to increase precision of the analysis of lipids. Quantification of different classes of lipid species was achieved by plotting the standard calibration curves of naturally occurring lipid species of PC 34:1, 36:2, 38:4, 40:0 and PC O-16:0/20:4; SM d18:1/16:0, 18:1, 18:0; LPC 16:0, 18:1, 18:0; PE 34:1, 36:2, 38:4, 40:6 and PE P-16:0/20:4; PS 34:1, 36:2, 38:4, 40:6; Cer d18:1/16:0, 18:0, 20:0, 24:1, 24:0; cholesterol, CE 16:0, 18:2, 18:1, 18:0. Correction of isotopic overlap of lipid species as well as data analysis were performed using self-programmed Excel Macros for all lipid classes according to principles described previously . In brief, data analysis was performed with MassLynx software, which included the NeoLynx tool (Micromass) for averaging the scans at half peak height of the total ion count. NeoLynx generates centroid peak data from the continuum spectra and allows selection of the intensities of certain peaks. Neolynx includes background subtraction and smoothing according to Savitzky Golay of the combined spectra. These NeoLynx results were exported to Excel spreadsheets and further processed by self-programmed Excel macros, which sort the results, correct for isotopic overlap, calculate the ratios to the internal standards, generate calibration curves, and calculate quantitative values . Lipid species were annotated according to the proposal for shorthand notation of lipid structures derived from mass spectrometry . Glycerophospholipid species annotation was based on the assumption of even numbered carbon chains only and presented as the sum of carbon chain length and degree of unsaturation, without specifying fatty acid location at the sn-1 or sn-2 position. SM species annotation was based on the assumption that sphingosine d18:1 is present . Final quantities of lipid species and total lipid (sum of analysed lipid species) were calculated and expressed in nanomoles per milligram wet weight of tissue. All data are expressed as mean ± standard deviation (SD) with at least three mice (n = 3) for each group. Two-way analysis of variance (ANOVA) was calculated using GraphPad Prism Version 5.4 (GraphPad Software, San Diego, CA). Statistical comparisons among the groups were performed by Bonferroni post-test using the same software. A p-value of 0.05 or lower was considered as significant. Significance is indicated as **** P < 0.0001, *** P < 0.001, ** P < 0.01, *P < 0.05. The current study presents an extensive quantitative lipidome analysis of the total mouse lung homogenates during the postnatal development performed with the help of ESI-MS/MS. In total, we performed quantitative analysis of 160 GP, 24 SP, 18 CE species and cholesterol in different stages of the postnatal lung development (lungs of newborn, 15-day-old and 12-week-old adult mice). The glycerophospholipids consist of 35 PC, 15 LPC, 16 PG, 23 PE, 33 PE P, 15 PI and 23 PS lipid species. Sphingolipids consist of 15 SM, 7 Cer and 2 cerebroside (HexCer) lipid species. The overview of total lipid analyses from mouse lung homogenates during development is depicted in Fig 1A and 1B. The numbers represent number of lipid species quantified for particular lipid class. To evaluate the total lipid content in mouse lungs during the postnatal development, lipid quantity of phospholipids (sum of all analyzed GP classes), cholesteryl esters and cholesterol was calculated and expressed as nmol/mg wet weight (Fig 2). Total phospholipid content significantly increased during the development from P1 (28.26 ± 3.08 nmol/mg) to P84 (35.20 ± 1.42 nmol/mg) and from P15 (30.68 ± 0.85 nmol/mg) to P84 (35.20 ± 1.42 nmol/mg). However, we did not observe a statistically significant increase between P1 (28.26 ± 3.08 nmol/mg) and P15 (30.68 ± 0.85 nmol/mg). In contrast, cholesterol was significantly increased during progressing from P1 (7.9 ± 0.83 nmol/mg) to P15 (13.16 ± 0.70 nmol/mg) and a significant increase of the values was observed during the development from P1 (7.9 ± 0.83 nmol/mg) to adult lung (12.81 ± 0.55 nmol/mg). However, while developing from P15 to P84, cholesterol remained constant in P84 (12.81 ± 0.55 nmol/mg). The esterified cholesterol species remained at the level of <0.5 nmol/mg wet weight. Displayed are nmol/mg wet weight of the lipid class of all analyzed lipid species. Values are represented as mean ± SD, p-value summary: **** P < 0.0001, *** P < 0.001, ** P < 0.01. Where significance is not mentioned, values are considered as being not significant. Phospholipids represents the sum of PC, LPC, PE, PE P, PG, PI, PS, and SM. Individual lipid class (sum of the analysed lipid species) quantities are presented in S1 Table. For instance, the PC species gradually increased from the stage of P1 (16.47 ± 2.23 nmol/mg) to P84 (18.03 ± 0.70 nmol/mg), but in P15 lung the PC species were lower (14.80 ± 0.57 nmol/mg) as compared to P1 and adult. In contrast, we observed a gradual increase in the total amount of PS, PI, PE, PE P, LPC and SM lipid classes during the postnatal development of the mouse lung. Among these, only PS (3.05 ± 0.19 to 5.35 ± 0.22 nmol/mg) and PE P (2.55 ± 0.23 to 3.82 ± 0.12 nmol/mg) showed statistical significance during the change from P1 to P84. To gain deeper insights into the postnatal developmental alterations, we evaluated the individual lipid species profile of respective lipid classes from mouse lung. The diacylglycerophospholipids were denoted with total number of carbon atoms and double bonds (C:D). LPC and CE species containing single fatty acids were denoted according to lipid species nomenclature. In total, 35 different PC (24 PC, 11 PC O) species with different chain length and degree of unsaturation were documented in the lung extracts and their composition is depicted in Fig 3A and 3B. As expected, PC 32:0, PC 32:1, PC 30:0, PC 34:1 and PC 34:2 were the most abundant phosphatidylcholine species in the tested stages in the mouse lung. The predominant PC 32:0 increased significantly while progressing from P1 (4.11 ± 0.6 nmol/mg) to P84 (6.27 ± 0.2 nmol/mg). Similarly, PC 34:2, PC 36:4, PC 38:6, PC 38:4 and PC 40:6 lipid species were significantly elevated in P84 as compared to P1 (Fig 3A & 3B). Values are represented as nmol/mg wet weight. A) PC B) PC and PC O. Values are mean ± SD, p-value summary: **** P < 0.0001, *** P < 0.001, ** P < 0.01, *P < 0.05. Where significance is not mentioned, values are considered as being not significant. In contrast, monounsaturated lipid species (total number of double bonds = 1) PC 34:1 showed a significant reduction during the development from P1 (2.23 ± 0.23 nmol/mg) to P15 (1.73 ± 0.06 nmol/mg) and remained constant in P84 (1.79 ± 0.09 nmol/mg). Similarly, the values of PC 32:1 also significantly dropped during the maturation from P1 (4.08 ± 0.6 nmol/mg) to P15 (2.21 ± 0.19 nmol/mg) and thereafter slightly increased from the stage P15 to P84 (2.74 ± 0.11 nmol/mg). The monounsaturated lipid species (PC 30:1, 32:1, 34:1) were elevated in P1 lungs (Fig 3A). The PC 30:0 gradually increased from the stage of P1 (1.3 ± 0.1 nmol/mg) to P15 (1.8 ± 0.07 nmol/mg) and then significantly decreased in P84 (0.91 ± 0.01 nmol/mg). Interestingly, higher levels of PC 30:0 were detected in P15 lungs in comparison to P1 and P84. The analyzed ether-phosphatidylcholine (PC O) lipid species were present at low concentrations and their quantitative information during postnatal pulmonary development is showed in Fig 3B. In total, 15 different lysophosphatidylcholine lipid species were analyzed and their composition pattern is depicted in Fig 4. The major LPC species detected were LPC 16:0 followed by 18:0, 18:1, 16:1, 18:2 and 20:4. The values of LPC 16:0 significantly increased during the maturation from P1 (0.103 ± 0.02 nmol/mg) to P15 (0.210 ± 0.01 nmol/mg) and P1 to P84 (0.212 ± 0.03 nmol/mg). Similarly, during development process from newborn to adult the amounts of all other LPC species including LPC 18:0, polyunsaturated LPC species 18:2 and 20:4 increased. Remaining LPC species occurred at lower concentrations and did not show any significant differences. Values are represented as nmol/mg wet weight. Values are mean ± SD, p-value summary: **** P < 0.0001, *** P < 0.001, ** P < 0.01, *P < 0.05. Where significance is not mentioned, values are considered as being not significant. We analyzed 16 individual lipid species of phosphatidylglycerol and their composition during the postnatal lung development is represented in Fig 5. As expected, PG 34:1, PG 34:2, PG 32:0 and PG 32:1 constituted highly abundant lipid species that were detected in all three groups. Interestingly, in contrast to PC 32:0, disaturated PG 32:0 decreased while progressing from P1 to P15. Notably, monounsaturated PG 32:1 significantly dropped in P84 (0.075 ± 0.002 nmol/mg) as compared to P1 (0.205 ± 0.04 nmol/mg). Similarly, the values of PG 34:1 dropped in P15 as compared to P1 and slightly increased while maturing from P15 to P84. Values are represented as nmol/mg wet weight. Values are mean ± SD, p-value summary: **** P < 0.0001, *** P < 0.001, ** P < 0.01, *P < 0.05. Where significance is not mentioned, values are considered as being not significant. In contrast, diunsaturated (total number of double bonds = 2) PG 34:2 shows a significant increase during the development from P1 (0.117 ± 0.023 nmol/mg)) to P15 (0.173 ± 0.020 nmol/mg) and from P15 to P84 (0.367 ± 0.016 nmol/mg). Similar to PC 30:0, the PG 30:0 was higher in P15 in comparison to P1 and P84. Interestingly, we detected higher levels of PG 36:4 in P15 lungs. Remaining PG lipid species did not show any statistically significant differences during development. Total 23 individual phosphatidylethanolamine lipid species were quantified and their composition is depicted in Fig 6. Interestingly, in comparison to the other lipid classes, PE lipids exhibited higher abundance of long chain polyunsaturated species. Strikingly, PE 38:4 showed higher abundance in all three groups in comparison to other PE species. Further, PE 38.4 significantly increased during maturation from P1 (0.50 ± 0.038 nmol/mg) to P15 (0.712 ± 0.046 nmol/mg) and remained constant in P84 (0.726 ± 0.066 nmol/mg). Similarly, PE 40:4, PE 40:5, PE 40:6 and PE 38:6 were relatively abundant in P84 and significantly increased during the development of the lung. In contrast, less abundant PE 34:1, 36:2 and 38:5 exhibited higher concentrations in P1 but then significantly decreased in P84. The concentration of other lipid species were low abundance and it did not show any significant differences in all three groups. Values are represented as nmol/mg wet weight. Values are mean ± SD, p-value summary: **** P < 0.0001, *** P < 0.001, ** P < 0.01, *P < 0.05. Where significance is not mentioned, values are considered as being not significant. PE P-16:0, PE P-18:0 and PE P-18:1 (sn-1) individual plasmalogen compositions were calculated and their values have been displayed in Fig 7. Interestingly, PE P-16:0 (sn-1 substituent) plasmalogens are present in higher amounts in all three groups. Regardless of the alkenyl chain in sn-1, the plasmalogens containing PUFAs in sn-2 position were the most abundant, represented mainly by 20:4 followed by 22:6, 22:4 and 22:5 in all tested groups. Values are represented as nmol/mg wet weight. Values are mean ± SD, p-value summary: **** P < 0.0001, *** P < 0.001, ** P < 0.01, *P < 0.05. Where significance is not mentioned, values are considered as being not significant. The PE P-16:0/22:4 and PE P-16:0/22:6 significantly rose during the postnatal lung development (from P1 to P15 and P1 to P84). In contrast, the predominant PE P-16:0/20:4 and PE P-16:0/22:5 also rose during development from P1 to P15, however, a slight decrease from P15 to P84 was noted. Similarly, PE P-18:0 based 20:4, 22:4, 22:6 and PE P-16:0/18:1 lipid species increased significantly during the development process (P1 to P84). However, the values of low abundant PE P-16:0/16:1 were significantly lower in P84 as compared to P1. Other individual ethanolamine based plasmalogens were not significant during the postnatal lung development. The individual composition of 15 species of phosphatidylinositol and 23 species of phosphatidylserine is depicted in Fig 8A and 8B respectively. Similarly, to PE and PE based plasmalogens, both PI and PS were found to be highly enriched with polyunsaturated species. PI 38:4 was the most abundant during all developmental stages. PI 38:4 (0.806 ± 0.065 nmol/mg to 1.173 ± 0.027 nmol/mg), PI 36:4 were significantly increasing while progressing from P1 to P84, whereas less abundant PI 36:2 and PI 38:5 species were higher in newborn and significantly gradually decreasing during the postnatal lung development (Fig 8A). Values are represented as nmol/mg wet weight. A) PI B) PS. Values are mean ± SD, p-value summary: **** P < 0.0001, *** P < 0.001, ** P < 0.01, *P < 0.05. Where significance is not mentioned, values are considered as being not significant. Among PS, polyunsaturated species (total number of double bonds ˃2) 40:4, 40:5, 40:6, 38:4 and the PS 36:1 were highly abundant in all stages of lung development. PS 40:4 was significantly gradually increasing from the phase of P1 (0.566 ± 0.04 nmol/mg) to P15 (1.218 ± 0.09) and from P15 to P84 (1.381 ± 0.06 nmol/mg). PS 38:4 and 36:1 species followed the same pattern (rise from P1 to P15 and P15 to P84). In addition, PS 36:2, 38:3 and 40:6 was detected in P1 and their increase was observed in P84. Among the PS species, PS 40:5 exhibited higher levels in P15 (0.827 ± 0.039 nmol/mg) as compared to P1 (0.335 ± 0.021 nmol/mg) and P84 (0.662 ± 0.031nmol/mg). The other analyzed PS species did not reach significance during lung development (Fig 8B). Within sphingolipids, individual 15 sphingomyelin, 7 ceramide and 2 cerebroside species compositions are depicted in Fig 9A and 9B respectively. SM 34:1 was the most dominant SM species in all developmental stages and gradually, but significantly increasing from the phase of P1 (0.46 ± 0.052 nmol/mg) to P15 (0.716 ± 0.05 nmol/mg), and thereafter slightly increasing from P15 to P84 (0.821 ± 0.057 nmol/mg). Similarly, SM 42:2 followed the same pattern during development. Data are represented as nmol/mg wet weight. A) SM B) Cer. Values are mean ± SD, p-value summary: **** P < 0.0001, *** P < 0.001, ** P < 0.01, *P < 0.05. Where significance is not mentioned, values are considered as being not significant. The very long chain SM lipid species 40:1, 42:1and 42:3 were significantly rising during the postnatal lung development and their elevated levels were detected in P84 (Fig 9A). Cer d18:1/16:0 was present in higher amounts in P15 lungs in comparison to P1 and P84. The very long chain fatty acid containing ceramide lipid species Cer d18:1/24:0 exhibited higher levels in P84. Interestingly, both of the analyzed HexCer d18:1/16:0 and HexCer d18:1/24:1 was present at higher levels in P1 and a significant gradual decrease was observed during the postnatal lung development up to the stage of P84 (Fig 9B). Eighteen esterified forms of cholesterol were analysed and their composition is depicted in Fig 10. Similar to LPC, the myristic acid (14:0) and palmitic acid (16:0) containing CE species were significantly elevated in the phase of P15 in comparison to P1 and P84. Oleic (18:1), palmitoleic (16:1) containing CE species were more elevated in P1 and significantly lowered during development. In contrast, linoleic (18:2) and arachidonic (20:4) acid containing CE species were significantly gradually increasing from the stage of P1 and reached higher levels in P84. Data are represented as nmol/mg wet weight. Values are mean ± SD, p-value summary: **** P < 0.0001, *** P < 0.001, ** P < 0.01, *P < 0.05. Where significance is not mentioned, values are considered as being not significant. Furthermore, the distribution pattern of GP according to their carbon chain length (total number of carbon atoms), degree of unsaturation (total number of double bonds) was elaborated and the results are shown in S2 Table, S1 Fig. Overall, the PC and PG comprised higher amounts of lipid species with carbon chain length ≤36. Whereas PE, PS and PI contained higher amount of lipid species with carbon chain length ˃36. Lipids constitute a diverse group of biomolecules, playing many roles in lung biology, especially in reducing the surface tension of alveoli to prevent the alveolar collapse and thereby stabilizing the lung parenchyma. The lung major lipid classes originating either from the pulmonary surfactant or from the bronchoalveolar lavage fluid (BALF) were already partly characterized, but a detailed distribution of total lipid classes and of their individual lipid molecular species composition in mouse lung during its postnatal development however, are not fully understood. In our current study, we used direct flow injection electrospray ionization tandem mass spectrometry to provide the total lipid quantity, and significant stage specific alterations of individual lipid species during the process of mouse postnatal pulmonary development. Furthermore, we showed the distribution pattern of lipid classes according to their carbon chain length and degree of unsaturation during development process. In the present study, we analyzed and quantified total 202 lipid species (GP, SP, and CE) and cholesterol of the pulmonary lipidome from the mouse lung homogenates of the P1, P15 and adult lungs (Fig 1A & 1B and Fig 2 & S1 Table). In general, an increase in the total lipid quantity (nmol/mg wet weight) of pulmonary tissue is a characteristic change during the lung development process. Our results revealed increased levels of the total phospholipid and cholesterol content during development. Indeed, these results are in consistence with the findings of Williams and colleagues as well as of Hahn, et al. in which rat lung and other organ lipids were studied during maturation [51, 52]. Furthermore, pulmonary lipid distribution patterns of the whole lung tissue in comparison to the pulmonary surfactant lipid composition in different mammalian species were studied (see book chapters in ref) [3, 53]. In this context, it is important to mention that the distribution pattern of the lipids in the lung tissue is heterogeneous between distinct species. In general, PC’s were the most predominant class of lipids of all tested species . The second most abundant lipid class of the surfactant GP was PG, however, this lipid class was detected in lower amounts in the whole lung tissue. In contrast, PE represented the minor class of lipids in the surfactant but constituted second major lipid class following PC in the whole lung tissue. Similarly, SM and PS lipid classes were detected in lower amounts in the pulmonary surfactant, however, they also were described as being abundant in the whole lung tissue. In fact, our results of the lipid class distribution of the mouse lung homogenates are in agreement with these findings [3, 53]. Our data clearly indicate that PC and cholesterol occupy major parts of the whole lung lipidome (Panel A in S2 Fig). Among the phospholipids, PC is the most dominant lipid class followed by PE (PE+PE P), PS, SM, PI and PG present during the postnatal mouse lung development (Panel B in S2 Fig). Based on these composition values, we calculated the molar ratios of lipid classes (S3 Table). For instance, the PC/LPC ratio is decreased from P1 to P15, and from P15 to P84. In contrast, SM/Cer molar ratio significantly increased from P1 to P15 and from P15 to P84. Interestingly, we observed a significant increase in the concentration of cholesterol from the phase of newborn to P15 (Fig 2, S1 Table & Panel A in S2 Fig), suggesting that cholesterol may play an important role in the process of alveolarization. Cholesterol is an integral component of various cell membranes, involved in the maintenance of membrane fluidity, membrane functions and signal transduction. In fact, cholesterol is the major neutral lipid component of the lung and up to 80% of the cholesterol present in the lung is in surfactant , and it is considered as a protosurfactant in immature lungs, lungs with lack of septation and in saccular lungs . Moreover, cholesterol enhances the adsorption of DPPC by increasing membrane fluidity and control the surface viscosity of the surfactant [56, 57]. In our study, we focused on the alterations of individual lipid species of mouse lungs during the postnatal development. Our results showed a significant increase in the abundance of PC 30:0 during the alveolarization (P1 to P15) (Fig 3A) process. These results are supported by previous observations on the postnatal development of the lung tissue and surfactant lipid analyses from 8-day-old mice and adult animals . Bernhard and co-workers reported about significant alterations in the abundance of major PC lipid molecular species in the pulmonary surfactant of different mammalian species during lung development [31, 40, 41]. In contrast to PC findings of surfactant lipidome, we showed a significant gradual increase of PC 32:0 (most likely to be DPPC) during the postnatal lung development. DPPC is the primary surface-active material found in majority of the mammalian species pulmonary surfactant. Maintenance of adequate DPPC within air space is essential for normal lung function . RDS is the major cause of mortality and morbidity in premature infants diagnosed with mainly DPPC deficiency in quantity and quality of pulmonary surfactant . Currently, surfactant replacement therapy with added products of DPPC is an effective therapeutic strategy available for RDS management . Interestingly, high contents of monounsaturated lipids of PC 32:1 and PC 34:1 was detected in P1 mice in comparison to P15 and adult lungs, suggesting that PC 32:1 might be involved in the establishment of the air-liquid interface in newborn animals. Furthermore, PC 34:1 is known to be crucial and plays an important role in the adsorption of DPPC immediately after birth [61, 62]. In contrast to mice, PC 30:0 is completely absent and PC 32:1 is minimal in nonalveolar species (birds), in which the lung contains capillaries instead of alveoli, suggesting that, PC 30:0 and PC 32:1 species are important and play an active dynamic role in the alveolarization process . In this context, it is important to mention the recent findings that PC 30:0 inhibited macrophage-triggered proliferation of T-lymphocytes and decreased the production of reactive oxygen species (ROS) during alveolarization [63, 64]. Moreover, PC 30:0 was significantly reduced in the emphysema patients and infants with BPD as well as in the neonatal rat models of reduced alveolarization suggesting that PC 30:0 may serve as a diagnostic marker for alveolar size during diseases . Similarly to these observations, in the pig model, PC 30:0 was found high in abundance in newborns and gradually decreased with age in adolescent pigs , whereas in humans and guinea pigs, PC 30:0 was increased during the lung development . The specific functions of PC 30:0, PC 32:1 and PC 34:1 lipid species during postnatal lung development are not clear yet. Also, the molecular mechanism and alterations of lipid species during the lung development are not clear yet. These alterations may be specific for individual mammalian species, a supposition that requires further investigation. Phosphatidylglycerol lipid concentrations are highly concentrated in the lung compared to other mammalian tissues . It is well documented that PG lipids are involved in the adsorption and, spreading of surfactant over the epithelial surface, as well as influence innate immunity and protect against viral infections [9, 67, 68]. Interestingly, surfactant deficiency in premature infants and also in mouse models of BPD showed the complete absence of PG lipids . Indeed, the presence of PG lipids in amniotic fluid is an indicator for the fetal lung maturity and PG lipids are known to be vital in the management of neonatal RDS and other obstetric conditions . It is known that PG lipid species are crucial for the lung function. There are, however, no reports on the composition of PG 30:0 during the postnatal lung development. In fact, Bernhard and colleagues were not able to measure PG 30:0 from rat surfactant during the postnatal lung development . Remarkably, in our study, we observed a significant increase in PG 30:0 abundance during alveolarization (P1 to P15), similar to PC 30:0, probably because of the high content of myristic acid during the postnatal lung development. In contrast to DPPC, we observed a significant decrease of abundance of DPPG during postnatal lung development. Recent findings suggest that DPPG interacts with vaccinia and variola virus strains and reduces the infection of pneumocytes in respiratory poxvirus infection . Moreover, several reports demonstrated that PG 34:1 (palmitoyl-oleoyl-phosphatidylglycerol, POPG) acts as a potent antiviral lipid against influenza A and respiratory syncytial virus [9, 67]. Interestingly, we observed a high content of this antiviral lipid PG 34:1 (POPG) in the newborn mouse lungs in comparison to P15 and P84 (Fig 5), suggesting that they might improve the innate immunity against viruses during the perinatal period. In addition to PC and PG, we measured other lipid species in the lung tissue such as PE, PI and PS (Figs 6, 8A & 8B). In contrast to PC and PG, we found that these lipid classes were found to be mostly enriched with long chain polyunsaturated species. Monounsaturated lipid species (PE 32:1) were found in high abundance in P1 in comparison to P15 and P84. In fact, long chain polyunsaturated species serve as substrates for the pro-inflammatory (leukotrienes, prostaglandins, etc.), as well as anti-inflammatory and pro-resolution (lipoxins, etc.) lipid mediators . Furthermore, mass spectrometry imaging of adult mouse lungs showed that these long chain polyunsaturated lipids are highly abundant at the epithelial lining of airways . We observed that 38:4 lipids were abundant in case of PE and PI lipid classes in all three age groups, which may serve as a source for arachidonic acid (AA) for the generation of lipid mediators. Proteomics data of a recent study revealed that proteins (cyclooxygenases, lipoxygenases, etc.) responsible for the generation of bioactive lipid mediators, are significantly upregulated in the adult mouse lungs, suggesting that these long chain polyunsaturated lipid species serve as a source for AA . Quantitative information of less abundant species of these lipid classes would help to understand postnatal developmental alterations in detail. However, the mechanism of alterations of individual lipid species of these lipid classes during lung development needs to be further explored. Plasmalogens are glycerophospholipids characterized by a vinyl ether linkage in sn-1 and an ester linkage in sn-2 position of the glycerol backbone. Plasmalogens are involved in the membrane dynamics, serve as an endogenous antioxidants, protect against ROS and prevent lipoprotein oxidation . Plasmalogen biosynthesis starts in the peroxisomes and deficiency of plasmalogens is associated with various peroxisomal disorders and other respiratory diseases like BPD , asthma and COPD . We observed that PE-based plasmalogens are much higher abundant compared to ether-phosphatidylcholines (PC O) during the postnatal mouse lung development. In ethanolamine plasmalogens, PE P-16:0 plasmalogens comprised the highest amount, whereas, PE P-18:0 and PE P-18:1 made up a smaller amount (Fig 7). Interestingly, we observed a high content of 20:4, 22:6, 22:5 and 22:4 (most likely to be arachidonic acid AA, docosahexaenoic acid DHA, docosapentaenoic acid DPA and adrenic acid)-rich plasmalogens (Fig 7). Further, the total quantities of plasmalogens (sum of all analyzed PE P species) were gradually increased from P1 to P84 during the postnatal lung development (S1 Table). These results are in consistent with the previous study, in which high abundance of PE P species were noted in adult mouse lungs during postnatal pulmonary developmental processes . Arachidonic acid enriched plasmalogens seem to play an important role in immune defence and normal lung physiology . Plasmalogens are reported to serve as a reservoir for the precursor molecules (e.g., AA, EPA, DHA, and DPA etc.) of eicosanoids, which are biologically active secondary lipid signalling messengers or for maresin and resolvins, lipid derivatives involved in the regulation of inflammation [77, 78]. Rüdiger and colleagues showed that addition of plasmalogens to surfactant-like phospholipid mixtures reduces surface tension and high content of plasmalogens in tracheal aspirate of preterm infants reduces the risk of respiratory diseases . Likewise, another study reported that high contents of plasmalogens protects the endothelial cells from hypoxia and ROS mediated stress . Sphingolipids are primarily found in cell membranes and are involved in diverse biologic processes such as migration, proliferation, differentiation, senescence, cell death, autophagy, and efferocytosis . In the lungs, sphingolipids are associated with cystic fibrosis, asthma, pulmonary edema, BPD, inflammation, lung injury and various types of lung cancers . Ceramides show both proliferative and apoptotic effects depending on their concentration and chain length . In analyzed SM lipid species, we detected that SM 34:1 lipid species as being predominant in all stages (Fig 9A). Both sphingomyelin 34:1 and ceramide species (Cer d18:1/16:0) showed high contents during alveolarization, especially in P15 (Fig 9A & 9B) mouse lungs, suggesting that, these lipid species are involved in the remodelling of tissue, also observed in the rat lung development . In contrast to our findings (P1 to P15), a recent study on mouse lungs using LC-MS/MS approach showed no significant alterations in the Cer d18:1/16:0 levels from P7 to P14 . The total content of sphingomyelin gradually increased with the age and the necessary transfer of biochemical substances across the semipermeable membranes (S1 Table). So far, no reports are available about the developmental changes of less abundant cholesteryl ester species in the mouse lung. Saturated fatty acids such as myristic acid (14:0)- and, palmitic acid (16:0)-containing lysophosphatidylcholines (Fig 4) and cholesteryl esters (Fig 10) were found highly abundant during P15 in comparison to P1 and P84. Monounsaturated fatty acid (MUFA) containing CE species were elevated in the newborn, whereas polyunsaturated fatty acid (PUFA) containing CE species were elevated in adult lungs. Physicochemical properties of lipids depend on their chain length and their degree of unsaturation. In this aspect, we calculated the distribution patterns of glycerophospholipids according to their carbon chain length (number of carbon atoms). We observed that, PC and PG glycerophospholipids are highly abundant of lipid species with carbon chain length C≤36, whereas PE, PS and PI glycerophospholipids are highly abundant with long chain lipid species (C˃36). The majority of the monounsaturated glycerophospholipids were found to be highly abundant in newborns, whereas polyunsaturated lipid species were highly abundant in adult lungs (S1 Fig). In contrary, P15 lungs exhibited high contents of myristic (14:0)- and palmitic (16:0)-acid containing lipid species. In the current study, we performed an extensive quantitative lipidomic analysis of P1, P15 and P84 mouse whole lung tissue homogenates to understand the changes occurring during postnatal development. The data provides lipidomic alterations in mouse lung during developmental process. However, at this stage we are not able to discriminate the lipidomic changes occurring specifically at cellular (membrane or intracellular) and extracellular (alveolar) level. Comprehensive comparative (quantitative) lipidomic analysis of bronchoalveolar lavage fluid (BALF) and whole lung tissue homogenates in mice and other mammalian species in which alveolarization continues beyond extra-uterine life (e.g. rats) needs to be investigated in the near future, which can provide deeper insights for a better understanding of pulmonary developmental process at molecular and cellular level. In our study, we have provided the total lipid quantity and given a detailed overview of lipid classes as well as absolute quantitative information on the individual lipid species and their distribution pattern according to carbon chain length and degree of unsaturation during postnatal mouse lung development using high-throughput tandem mass spectrometry. Our study provides an extensive quantitative lipidome of whole mouse lung tissue (including less abundant lipid species, neutral lipid components such as cholesterol and their esters), which may serve as reference for understanding the occurring lipid alterations, which in turn affect lung function during development or in pulmonary diseases.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.