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 decades1, with special focus on the detection of early cancer stages. Some cancer types, such as pancreatic cancer2, 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 cancers3. 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%, respectively4. Invasive procedures, i.e., biopsies, are performed only for the final confirmation of PDAC. Several types of blood tests were considered for PDAC screening5,6,7, 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 stages8, 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)9 and was evaluated as well for the diagnosis of PDAC using liquid biopsies10. However, the sensitivity for KRASmutation testing is low11, even though this mutation is encountered in more than 90% of PDAC12. 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 cancer13. Furthermore, the uptake of nutrients in KRAS mutated cells can include blood lipids for cell proliferation and survival14,15. KRAS mutation has been reported to be associated with lipid metabolism in pancreatic cancer cells16.

Lipids serve numerous functions in human metabolism17, such as cell membrane constituents, signaling molecules, energy supply, and storage. Changes in lipid concentrations were already reported in other cancer types18, mostly for cell lines19, tissues20, and less frequently for body fluids, too21.

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.

Study design

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 script22 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.

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.

Full size image

The study was divided into individual phases23 (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)/MS24. 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.

Phase I (discovery)

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 results25.

a Both genders. b Males. c Females. d Specificity, sensitivity, and accuracy for individual models. Source data are provided as a Source Data file.

Full size image

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 validation26 and the stability test for samples collected during one year24. The whole sample preparation protocol was optimized including the development of quality control (QC) system.

Phase II (qualification)

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)27. 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.

Phase III (verification)

In the timeframe between Phase II and Phase III, UHPSFC/MS method and the sample preparation protocol for the lipidomic analysis were optimized and validated26,28,29. Furthermore, the influence of preanalytical and analytical aspects, such as the blood collection tubes24, lipidomic profile stability in the period of one year for the same volunteers24, and the influence of the mass spectrometer30 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 intake31, 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.

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.

Full size image

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.

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 - CancerSeek32. 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.

Full size image

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 PDAC32. 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 values32. 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.

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.

Full size image

Potential for survival prognosis

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 work14. 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.

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.

Full size image

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 mutations11, ctDNA10, and proteins32 in serum or plasma for cancer diagnosis has been intensively investigated, and the results are promising for cancer screening33. 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 samples34,35, 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 switch36. In this context, alterations in sphingolipids concentrations deserve attention, as the normal metabolism of sphingomyelins might be necessary to maintain KRAS function37. 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.

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.