Pilot study of screening method for pancreatic cancer using lipidomic profiling of plasma or serum

Abstract

Background

Pancreatic cancer is one of the most aggressive malignancies with a 5-year relative survival rate of only 13%. Its poor prognosis is largely attributed to the lack of reliable tools for early detection. Current diagnostic standards rely on imaging methods that are invasive, costly, and often inadequate to detect early-stage disease. A noninvasive blood-based test with high sensitivity and specificity could substantially improve patient outcomes.

Methods

Lipid concentrations in plasma and serum samples were determined by ultrahigh-performance supercritical fluid chromatography–mass spectrometry, and multivariate statistical modeling was used to analyze lipid profiles and differentiate between groups.

Results

Here, we present results from a pilot study evaluating lipidomic profiling of prospectively collected plasma and serum samples from patients with pancreatic ductal adenocarcinoma (PDAC, n = 177), healthy controls (n = 218), and high-risk individuals for pancreatic cancer (n = 93). The lipidomic test distinguishes PDAC patients from healthy controls with an accuracy exceeding 95%, including robust detection of early-stage cases and even individuals with low CA 19-9 secretion. The sensitivity is approximately 30% higher than that of CA 19-9. In high-risk individuals, the method achieves a specificity of over 96% (95% CI, 89–99%), comparable to established imaging-based approaches.

Conclusions

This pilot study demonstrates the promising potential of lipidomic profiling as a noninvasive, blood-based screening tool for pancreatic cancer. The method outperforms current biomarkers, maintains high diagnostic accuracy in early-stage disease, and performs reliably in high-risk populations. These findings support the initiation of a clinical trial aimed at validating the lipidomic test for the early detection of PDAC.

Plain language summary

Pancreatic cancer is often diagnosed too late because early symptoms are vague and difficult to recognize. Current blood markers do not reliably detect early-stage disease, and imaging methods can be uncomfortable or invasive. Earlier detection is critical to improve patient outcomes. This study shows that a simple blood test measuring lipid molecules, combined with the statistical analysis, can identify pancreatic cancer, including early stages, with high accuracy and may serve as a future screening tool for individuals at high risk.

Peterka et al. develop a noninvasive blood-based lipidomic test that distinguishes pancreatic cancer patients from healthy controls. The test achieves >95% accuracy, detects early-stage disease, shows >96% specificity in high-risk individuals compared with imaging, and supports initiation of a multicenter clinical trial.

Introduction

Pancreatic cancer is among the deadliest malignancies, with a 5-year relative survival rate of only 13% (metastatic 3%, regional 17%, localized 42%, and unspecified 11%)¹. The most common subtype is pancreatic ductal adenocarcinoma (PDAC), accounting more than 90% of all pancreatic malignancies². The incidence and mortality caused by pancreatic cancer have been gradually rising, and it is predicted that pancreatic cancer will become the second leading cause of cancer-related deaths by 2030 and double the number of new cases in the United States³. The combination of chemotherapy and surgical techniques is the primary treatment options for PDAC, but fewer than 20% of patients are diagnosed at a resectable stage because PDAC is mostly asymptomatic or presents with vague symptoms, leading to detection at late-stage^4,5. Imaging modalities, such as endoscopic ultrasonography (EUS), magnetic resonance imaging (MRI), and computed tomography (CT), are the standard diagnostic tools for early diagnosis of PDAC⁶. However, these methods are time-consuming, technically demanding, and often uncomfortable or invasive for patients, which may expose individuals to radiation or contrast agents and can be contraindicated in subjects with certain medical devices, claustrophobia, or allergies⁷. The carbohydrate antigen (CA) 19-9 is the only biomarker routinely used in PDAC diagnostics, but it has low sensitivity for early-stage disease and lacks specificity, limiting its utility in screening programs^8,9. Similarly, despite considerable research into genomics, transcriptomics, proteomics^6,10–12, and their integration into multi-biomarker panels^13–15, no worldwide clinically validated biomarkers for early detection of pancreatic cancer have yet emerged¹⁴. However, a new blood test, PancreaSure (Immunovia), has been launched commercially in the US in 2025¹⁶. Lipidomics, a rapidly evolving field focused on the comprehensive analysis of lipids, offers a promising new avenue for biomarker discovery¹⁷. Lipids play essential roles in cellular processes, and alterations in lipid profiles have been observed in various cancers, including pancreatic cancer¹⁸.

Prospective studies that involve long-term surveillance of high-risk individuals (HRI) have demonstrated improved detection rates of resectable disease leading to median overall survival of 9.8 years compared to 1.5 years for patients diagnosed with PDAC outside surveillance^19,20. Although population-wide screening is not feasible, mainly due to cost and logistic limitations, targeted HRI screening is recommended²¹. In the general population, the lifetime risk of developing PDAC is approximately 1.5%²². However, conventional risk factors such as diet, obesity, alcohol consumption and smoking increase this risk^23,24. Additionally, individuals with genetic susceptibility syndromes, a familial pancreatic cancer, or hereditary pancreatitis exceed the risk threshold commonly used to define HRI^22,25. An additional potential target group for screening comprises individuals over 50 years of age newly diagnosed with type 2 diabetes mellitus, particularly type 3c. Approximately 1% of these individuals are diagnosed with pancreatic cancer within three years^26,27, although this group is not yet officially recognized as high-risk²². Despite recent advances in surgical and medical management, early detection remains the key limitation in improving PDAC prognosis²⁸. A noninvasive, high-throughput method capable of detecting reliable biomarkers in body fluids could be a breakthrough in the early diagnosis of PDAC.

This study builds on our previous work²⁹ that demonstrates the potential of lipidomic profiling to distinguish between healthy controls and PDAC patients with high sensitivity and specificity. In Phase 1, we refine the methodology and address the questions that arise from the earlier studies. In Phase 2, we apply the method to samples from HRI and compare the results with imaging data. The lipidomic profiling method shows high sensitivity and specificity, including for early-stage PDAC, and significantly outperforms carbohydrate antigen 19-9 (CA 19-9) and carcinoembryonic antigen (CEA) in diagnostic accuracy.

Methods

Human subjects

Three cohorts of subjects including both males and females were involved in this prospective sample collection: healthy controls, PDAC patients, and HRI for developing pancreatic cancer (Table 1). All subjects in the study were over 18 years of age, and blood samples were collected from each volunteer after an overnight fast. The inclusion criterion for healthy controls was the absence of any lifetime history of cancer; no restrictions were placed on other diseases. PDAC blood samples were collected from patients with the confirmed diagnosis of PDAC including all tumor stages. Patients in high-risk groups had no diagnosis of PDAC and were affected by at least one of the following factors: familial pancreatic cancer, chronic pancreatitis, or genetic susceptibility syndromes. A prospective sample collection involved 488 subjects, including 218 healthy controls, 177 PDAC patients, and 93 HRI. The final sample size was not predetermined, as this was an exploratory study, and the number of samples was based on the availability of volunteers. No formal power calculation was performed. The clinical information for each sample is summarized in Supplementary Data 1. This study was performed in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of the University Hospital Olomouc and Faculty of Medicine and Dentistry, Palacký University Olomouc (approval Nos. 147/20 and 117/23), the Ethics Committee of University Hospital Hradec Králové (approval No. 201903 S22P), and the Ethics Committee of the Masaryk Memorial Cancer Institute (approval No. 2021/1443/MOU), Czech Republic. Informed consent was obtained from all volunteers, and the ethical committee approved the collection of blood samples. All collected data were pseudonymized.

Table 1.

Characteristics of the subjects in this study

Parameter			Controls	PDAC patients	High-risk individuals
Subjects		Total	218	177	93
		Male	93	88	56
		Female	125	89	37
Agea			58 ± 11	67 ± 10	55 ± 9
Body mass indexa			27 ± 4	25 ± 5	28 ± 5
CA 19-9	<37 U/mL		215	60	86
	>37 U/mL		3	117	7
CEA	<5 ng/mL		216	122	89
	>5 ng/mL		2	55	4
Diabetes mellitus			12	67	4
T1			–	20	–
T2			–	76	–
T3			–	41	–
T4			–	16	–
Tx			–	24	–
Chronic pancreatitis			1	2	40
Familial pancreatic cancer			0	11	29
BRCA1			0	2	10
BRCA2			0	3	12
Peutz–Jeghers syndrome			0	0	2

^aParameters are presented as mean ± standard deviation.

Sample preparation

All samples were stored at −80 °C until the lipidomic extraction. The modified Folch extraction procedure³⁰ was used for the extraction of lipids. 25 μL of EDTA human plasma or serum spiked with 20 µL of IS-Mix (Table S1) was homogenized in 3 mL of chloroform/methanol (2:1, v/v) for 15 min in an ultrasonic bath at 30 °C. After cooling to ambient temperature, 600 µL of 250 mM ammonium carbonate was added and the mixture was stirred (560 rpm, IKA KS 130 shaker) at ambient temperature for 5 min. The mixture was centrifuged for 5 min at 867 × g. The aqueous phase was removed to the waste and the organic phase was evaporated under a stream of nitrogen at 35 °C. The residue was dissolved in 500 μL of mixture CHCl₃/MeOH (1:1, v/v). The extract was diluted 10-fold with mixture of CHCl₃/MeOH (1:1, v/v)) before the ultrahigh-performance supercritical fluid chromatography-mass spectrometry (UHPSFC/MS) analysis and 25-fold with 5 mM ammonium acetate in CHCl₃/MeOH/IPA (1:2:4, v/v/v) before the flow injection analysis tandem mass spectrometry (FIA-MS/MS) analysis.

Lipidomic analysis

Two validated and high‑throughput lipidomic quantitative methods using mass spectrometry, with and without the chromatographic separation, were used. The UHPSFC was connected to a Xevo G2-XS QTOF mass spectrometer (Waters, Milford, MA, USA) for high-resolution MS data acquisition. The UHPSFC/MS method³⁰ using lipid class separation approach was used for target lipidomic analysis under the following conditions: Viridis BEH column (100 × 3 mm, 1.7 μm, Waters), flow rate 1.9 mL/min, injection volume 1 μL, column temperature 60 °C, and the automatic backpressure regulator at 1800 psi. Mobile phase A was carbon dioxide and mobile phase B (modifier) was methanol containing 30 mM ammonium acetate and 1% of water. Gradient elution with total run time 8.0 min was set: 0 min—1% modifier; 1.5 min—16% modifier; 4 min—51% modifier; 7 min —51% modifier; 7.51 min, 1% modifier. Makeup solvent with the same composition as modifier with the flow rate 0.25 mL/min was used. The mass spectrometer was equipped with an electrospray ionization (ESI) source and all data were acquired in the positive ion mode.

The FIA-MS/MS method³¹ was employed for lipidomic profiling without the chromatographic separation, using specific precursor ion and neutral loss scans. The 6500 + QTRAP (quadrupole-linear ion trap, Sciex) mass spectrometer equipped with the Turbo V source was used for data acquisition. In FIA-MS/MS method³¹, sample introduction was performed using a Shimadzu Nexera LC system, with a mixture of CHCl₃/MeOH/IPA (1:2:4, v/v/v) as the pump solution. The flow rate was set to 50 µL/min between 0 and 0.55 min (for sample introduction into the MS source), 5 µL/min between 0.55 and 2.6 min (for MS/MS scan acquisition), and 300 µL/min between 2.6 and 4 min (for system washing). Between injections, the needle was rinsed thoroughly with 1000 µL of the CHCl₃/MeOH/IPA (1:2:4, v/v/v) mixture containing 0.5% water. All samples were analyzed in positive ion mode using electrospray ionization.

Quality control

Three types of quality control samples were used to control the system stability - pooled human plasma (QC-P), human serum (QC-S), and NIST SRM 1950 human plasma. The aliquots of analyzed human plasma or serum were mixed to create a representative sample for individual matrix. The lipidomic extracts of QC-P and QC-S were injected after every 20 injections and the signal response of exogenous lipids was monitored in all samples. The lipidomic extract of NIST SRM 1950 human plasma was injected after every 120 injections, which is a reference material and can be eventually used for inter-laboratory harmonization.

Data processing

The raw data from UHPSFC/MS analysis were subjected to noise reduction, the lock mass correction, and the data conversion from continuum to centroid mode using the Accurate Mass Measure tool in MassLynx 4.2. Text files including experimental m/z values with MS signal responses for individual lipid classes were generated by MarkerLynx. LipidQuant 2.1³² was used for the calculation of lipid concentrations in biological samples and type I and type II isotopic corrections were automatically applied. Generated lipid concentrations for UHPSFC/MS method are summarized in Supplementary Data 2 and data after linear normalization using quality control samples are shown in Supplementary Data 3.

The mass spectrometer was operated using Analyst software (version 1.6.3, Sciex). The raw data obtained by the FIA-MS/MS were processed using the LipidView (version 1.2) software and the text files were generated. These tables were manually cross-checked against the neutral loss and precursor ion spectra to ensure correct alignment of m/z intensities across samples. The LipidQuant 1.0³³ software was used for the calculation of lipid concentration with automatic type II isotopic correction and final data frames with concentrations were merged in R. Lipid concentrations determined by FIA-MS/MS method are summarized in Supplementary Data 4.

The lipid concentrations were determined using internal standards added to the sample before extraction. Lipid species with determined concentrations at least 75% of all samples were included in the dataset and zero filling was applied for missing values by setting 80% of the minimum measured concentration for a given lipid species from all samples.

Pancreatic cancer markers

The method for the quantitation of CA 19-9 is based on the reaction of monoclonal antibody 1116-NS-19-9 with carbohydrate antigen in plasma or serum. Chemiluminescent Microparticle Immunoassay Alinity i CA 19-9XR (Abbott, Wiesbaden, Germany) was used. The cutoff value for the CA 19-9 test is 37 U/mL, therefore, all values above 37 U/mL were classified as positive for pancreatic cancer.

The method for the quantitation of CEA is a two-step immunoassay based on the reaction of CEA antigen present in plasma or serum and paramagnetic microparticles coated with a layer of anti-CEA antibodies. Chemiluminescent Microparticle Immunoassay Alinity i CEA (Abbott, Wiesbaden, Germany) was used. The cutoff value for the CEA test is 5 ng/mL, therefore, all values above 5 ng/mL were classified as positive for pancreatic cancer.

Imaging examination

Endoscopic ultrasonography of the pancreas was performed under conscious sedation of the patient using Olympus Evis Exera III CV-190 Plus processor. EUS scope Olympus UCT180 was introduced through the patient’s mouth. Whole pancreas was visualized from standard stomach and duodenal scope positions. Still pictures and videos of the procedure were stored. In case of lesion suspicious for neoplasia, EUS-guided fine needle biopsy (EUS-FNB) was considered for tissue sampling and subsequent cytological or histological evaluation.

Magnetic resonance imaging of the pancreas was performed on Siemens 3T Magnetom Vida in accordance with Pancreatic Cancer Early Detection (PRECEDE) consortium guidelines³⁴. Each examination included native targeting the pancreatic region followed by diffusion-weighted imaging (DWI) b-1500 and cholangiography. For detailed imaging, contrast medium gadoteric acid (DOTAREM) was applied intravenously. MR cholangiography proved especially useful for imaging the pancreatic duct and cystic lesions.

Quantitation and statistics

The evaluation of statistical significance differences in lipid profiles was performed using a two-sided t-test (Welch test) and fold change (FC). P-values < 0.05 were considered to indicate the statistical significance. The Bonferroni approach was applied to verify statistical significance for all p-values by the multiple testing. Lipid species with the fold change (tumor/normal, female/male, or serum/plasma) exceeding ±20% in molar concentrations were considered to indicate the statistical significance. The box plots were generated for the most dysregulated lipid species, and the p-value < 0.0001 is indicated by (****) based on the Mann–Whitney U test. Receiver operating characteristic (ROC) curves were generated based on the predicted response values from orthogonal partial least squares discriminant analysis (OPLS-DA) models, and the corresponding area under curve (AUC) values were calculated. Visualizations were prepared in R software environment (ver. 4.4.1). Cytoscape software (v. 3.8.2) was used to build a network map for quantified lipids³⁵.

Multivariate data analysis (MDA) was performed by SIMCA 13.0.3 software (Umetrics, Sweden), and the concentrations were logarithmically transformed and Pareto scaled. Each model incorporated all quantified lipids summarized in Supplementary Data 3 and 4, 190 species for UHPSFC/MS and 75 species for FIA-MS/MS, without including any additional biomarkers, such as CA 19-9 or CEA. No sample was excluded from the study. Statistical models based on unsupervised principal component analysis (PCA) and supervised OPLS-DA were performed. The S-plot generated from OPLS-DA enabled the identification of the most dysregulated lipid species in the studied samples, with the area of interest located in the top-right (>0.1, >0.4) and bottom-left (<−0.1, <−0.4) regions of the plot. The variable importance in projection (VIP) scores were assessed for each OPLS-DA statistical model, and lipid species with VIP values greater than 1 were deemed statistically relevant. OPLS-DA models were built for both training and validation sets to obtain predicted response values. The training set was used to build statistical models, and seven-fold cross-validation approach was applied to prepare and evaluate the model. The overfitting in the presented models was investigated by comparing of R2Y and Q2 values. Q2 values being high and close to the corresponding R2Y values, indicating that the models were not overfitted. Additionally, permutation tests based on 100 permutations were used, and p-values were calculated. The validation set was used to independently assess the model by the predicting blinded samples that were not included in the model. For the validation model, samples were randomly divided for the initial model (75%) and blinded samples (25%) equivalent to the number of normal and tumor samples. Based on the predicted value, the samples were classified as normal (≤0.5) or tumor (>0.5). The 95% confidence intervals follow the Clopper–Pearson method.

Results

Study design

The present methodology for the early detection of pancreatic cancer is based on the measurement of lipid concentrations in human plasma or serum, followed by the statistical evaluation using multivariate data analysis. We previously demonstrated that lipidomic profiling can distinguish PDAC patients from healthy controls with sensitivity, specificity, and overall accuracy exceeding 94% in the training set and over 80% in the validation set²⁹. However, earlier and ongoing research raised several important questions that needed to be addressed prior to the clinical validation of this approach as a screening method could proceed. To address these questions, we conducted a prospective sample collection involving 488 subjects (Table 1). An overview of the study is presented in Fig. 1.

Fig. 1. Overview of the study design.

The figure illustrates the overall study workflow, including the methodology, primary research questions, and cohort structure across individual phases. A prospective sample collection involves 488 subjects, comprising 218 healthy controls, 177 PDAC patients, and 93 high-risk individuals.

In Phase 1, we investigated the effects of gender and sample matrix on lipid concentrations in defined lipid classes, as the literature does not offer a clear consensus on these factors^36–38. Furthermore, the methodology was upgraded with LipidQuant 2.1³² software and expanded to include phosphatidylethanolamines (PE), which showed a strong potential to enhance lipidomic profiling in a small cohort of PDAC patients (n = 25) and healthy controls (n = 25)³⁹. Unsupervised PCA revealed partial separation based on gender (Fig. S1A), whereas plasma and serum matrices showed no substantial differences in lipid profiles (Fig. S1B). On the contrary, clear and significant differences were observed between the lipidomic signatures of healthy controls and patients with PDAC (Fig. S1C). The tight clustering of QC samples (Fig. S1D) indicates excellent data consistency and analytical performance. By understanding the impact of individual variables and integrating methodological improvements, we finalized the analytical workflow and substantially increased the discriminatory power of the lipidomic profiling, which also improved the success rate of blind sample classification.

In Phase 2, we applied the final methodology to HRI predisposed to develop pancreatic cancer, with the aim of evaluating its suitability for screening purposes. HRI samples were prospectively collected in collaboration with two national pilot studies in the Czech Republic, SCREPAN and HEPACAS, both focused on the early detection of PDAC through regular imaging surveillance. Finally, the diagnostic performance of the lipidomic profiling method was compared with conventional imaging methods and tumor biomarkers CA 19-9 and carcinoembryonic antigen.

Effect of gender and matrix on the lipidome

Healthy control samples were used to compare gender-based lipid profiles in plasma (Fig. 2A–C) and serum (Fig. S2A–C). 93 males and 125 females with comparable age and body mass index (BMI) were included in the analysis (Table S2). The concentrations of 190 lipid species from 11 lipid subclasses were compared using univariate statistical methods (Supplementary Data 5) and multivariate data analysis (Figs. 2A and S2A). The supervised OPLS-DA revealed a clear separation between male and female lipidomes. The S-plot generated from OPLS-DA highlights the most dysregulated lipid species (Fig. 2A), which is further supported by volcano plots (Fig. 2B) that illustrate differences based on both p-values and fold change. Statistically significant higher concentrations were observed for nearly all sphingomyelins (SM) and plasmenyl/ether (P-/O-) phosphatidylcholines (PC) with shorter fatty acyl chains (C32–C34) in females. On the contrary, higher concentrations of triacylglycerols (TG) and diacylglycerols (DG) with shorter acyl chains (C32–C34) were identified in males. Slightly higher, though not statistically significant, concentrations were found for PE and PE P-/O- in women, and for LPC in men. Concentrations for other lipid classes, cholesteryl esters (CE), ceramides (Cer), and phosphatidylcholines (PC), were comparable between genders. Although several individual lipid species showed dysregulation within specific classes, no consistent pattern was observed with respect to the fatty acyl chain lengths or the degree of (un)saturation. Independent data evaluation confirmed the same trends for both plasma and serum matrices. Since the male group showed a slightly higher average BMI, the relationship between total TG concentrations and BMI was examined (Fig. 2C). However, statistical analysis did not confirm a significant difference and elevated TG levels in men reflect metabolic differences rather than BMI. However, higher TG levels were correlated with elevated BMI in both sexes, indicating that TG is not an optimal biomarker due to the high variability influenced by lifestyle factors.

Fig. 2. Comparison of lipid profiles obtained by UHPSFC/MS for following group pairs: females vs. males (plasma model), serum vs. plasma (female models), and PDAC patients vs. healthy controls (male models).

A Score plots of orthogonal partial least squares discriminant analysis (OPLS-DA) and the corresponding S-plots highlighting the most dysregulated lipid species. B Volcano plot showing the most dysregulated lipid species based on log₂-transformed fold changes and −log₁₀ p-values (two-sided t-test). C Relation between body mass index (BMI) and the sum of triacylglycerols (TG), where significance is determined by the Mann–Whitney U test; p-values > 0.05 correspond to non-significant (ns). BMI categories: 20–25 (22 males and 49 females), 25–30 (47 males and 46 females), and >30 (24 males and 25 females). D Visualization of differences between serum and plasma samples using multivariate and univariate approaches. S-plot is generated from OPLS-DA highlighting the contribution of individual lipid species to group separation. Volcano plot showing log₂-transformed fold changes and −log₁₀ p-values (two-sided t-test). E Supervised OPLS-DA score plots with cancer samples colored according to tumor stage (T1—yellow, T2—orange, T3—red, T4—rose, and Tx—brown, where information about the stage is not available), and the corresponding S-plots indicate the most upregulated (red) and downregulated (blue) lipid species for plasma and serum models. F Box plots of the most dysregulated lipid species showing differences in concentrations between healthy controls (N, blue; n = 93) and PDAC patients (T, red; n = 88). The number of significance symbols corresponds to p-value ranges from the Mann–Whitney U test; p < 0.0001 is indicated by ****. G Receiver operating characteristic (ROC) curves with the 95% confidence intervals for plasma and serum models in training and validation sets, combining male and female data (prediction scores derived from gender-specific models).

To assess the effect of sample matrix on the lipidome, lipid profiles were compared between plasma and serum using the same cohort of healthy volunteers. Blood collection tubes for the isolation of EDTA plasma and serum were used for each volunteer at the same time. Lipid concentrations were evaluated separately for males and females to eliminate the influence of gender. Significant matrix-related differences were observed for some DG (Figs. 2D and S2D), with approximately 10–30% higher concentrations in serum, particularly for DG 36:X species. No statistically significant differences between plasma and serum were found for other lipid classes (Supplementary Data 6). Minor trends were observed, including slightly lower concentrations of PE, PE P-/O-, and SM, and up to 5% higher concentrations for other lipid classes in serum for both genders, but these differences were not statistically significant.

Effect of gender and sample matrix on lipidomic profiling and model performance

Various types of blood collection tubes are used in cancer biomarker research, and discrepancies in the lipid dysregulation observed across studies are often attributed to differences in the sample matrix¹⁸. In our laboratory, lipidomic profiling has previously been performed for healthy controls and patients with PDAC using serum samples²⁹ and healthy controls, kidney, breast, and prostate cancers using plasma samples⁴⁰. While similar lipid dysregulations were observed, the most encouraging results were obtained in PDAC, raising the question of whether these outcomes were due to the cancer type or the sample matrix. To address this, we conducted a prospective collection of paired plasma and serum samples from the same individuals at the same time. In total, samples from 218 healthy controls and 177 PDAC patients were analyzed. The PDAC cohort included individuals classified as stage T1 (11%), T2 (43%), T3 (23%), T4 (9%), and unclassified (14%). This cohort was used to evaluate the performance of lipidomic profiling models in both plasma and serum (Figs. 2E–G and S2E–F).

Both plasma and serum models showed high efficiency in distinguishing healthy controls from PDAC patients, regardless of tumor stage, and identified the same dysregulated lipid species (Figs. 2E and S2E). No clustering of samples based on tumor stage was observed, indicating comparable sensitivity across all stages. On the other hand, it is not possible to determine the cancer stage based on changes in lipid profiles. The most downregulated lipid species included sphingolipids with very long N-acyl chains (C21–C26) and low number of double bonds (DB 1–2), all PE P-/O-, PC with C30–C34, PC P-/O- with C34–C36, and LPC 18:2. Among the most upregulated species were DG 36:X and PE 34:1. However, the statistical significance of DG was slightly lower in serum models compared to plasma, and higher in female models compared to male models. Supplementary Data 7 provides a summary of the statistical significance of all lipid species included in these models. Concentrations of the most dysregulated lipid species between healthy controls and PDAC patients are visualized using box plots (Figs. 2F, S2F, and S3), with significance assessed using the non-parametric Mann–Whitney U test. The box plots also confirm similar lipid concentrations in plasma and serum samples for both groups, with the exception of DG.

Differences between plasma and serum models were further explored by comparing model parameters (Supplementary Data 8) and prediction success rates in training and validation datasets (Supplementary Data 9). The validation dataset was selected randomly and consisted of ~25% of total samples (equivalent number of N and T), excluded from the model building process. Models were evaluated separately for each matrix, and either separately or combined for gender, with the sample distribution kept consistent across matrices. Combining both genders led to model parameters comparable to gender-specific models, although the sample sizes differed (395 total vs. 181 males and 214 females), which influenced model robustness. Models combining both genders showed a similar number of false predictions (8 vs. 7 for plasma and 11 vs. 10 for serum, Table 2). However, these misclassifications had prediction scores farther from the decision threshold (0.5) compared to gender-specific models (Supplementary Data 9). For the training datasets, sensitivity, specificity, and accuracy exceeded 96% in plasma models and 95% in serum models. In validation datasets, accuracy was over 95% across both matrices and genders, while sensitivity exceeded 94% (95% CI, 71–100%) for males and 89% (95% CI, 65–99%) for females. The lower success rate in females was influenced by the limited number of classified samples, whereas only two false negatives were observed. Cross-gender (Supplementary Data 9 and 10) and cross-matrix predictions (Fig. S4, Supplementary Data 11 and 12) were performed to investigate the robustness of the methodology. Nevertheless, accuracy remained above 94% and 93%, respectively (more details in Supplementary Results). However, all subsequent analyses in this study were therefore based on gender- and matrix-specific models.

Table 2.

Sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) (with 95% confidence interval) for the lipidomic profiling method in the training and validation sets separated by matrix and gender

Plasma
Gender	Both		Male		Female
Dataset	Tr.	Va.	Tr.	Va.	Tr.	Va.
Plasma
Sensitivity [%]	97 (93–99)	97 (85–100)	99 (94–100)	94 (71–100)	98 (92–100)	89 (65–99)
Specificity [%]	100 (97–100)	100 (92–100)	99 (94–100)	100 (82–100)	100 (97–100)	100 (86–100)
Accuracy [%]	98 (96–99)	99 (93–100)	99 (96–100)	97 (85–100)	99 (97–100)	95 (84–99)
PPV [%]	99 (96–100)	100 (90–100)	99 (93–100)	100 (79–100)	100 (96–100)	100 (79–100)
NPV [%]	97 (94–99)	98 (86–100)	99 (93–100)	95 (74–99)	98 (94–100)	93 (77–98)
Truea	171 / 217	34 / 44	87 / 92	16 / 19	87 / 125	16 / 25
Falsea	6 / 1	1 / 0	1 / 1	1 / 0	2 / 0	2 / 0
Serum
Sensitivity [%]	96 (92–98)	94 (81–99)	95 (89–99)	94 (71–100)	98 (92–100)	89 (65–99)
Specificity [%]	100 (97–100)	98 (88–100)	99 (94–100)	100 (82–100)	100 (97–100)	100 (86–100)
Accuracy [%]	98 (96–99)	96 (89–99)	97 (94–99)	97 (85–100)	99 (97–100)	95 (84–99)
PPV [%]	99 (96–100)	97 (83–100)	99 (92–100)	100 (79–100)	100 (96–100)	100 (79–100)
NPV [%]	97 (94–98)	96 (85–99)	96 (90–98)	95 (74–99)	98 (94–100)	93 (77–98)
Truea	170 / 217	33 / 43	85 / 91	16 / 19	87 / 125	16 / 25
Falsea	7 / 1	2 / 1	4 / 1	1 / 0	2 / 0	2 / 0

Concentrations of lipids were measured by UHPSFC/MS method.

^aNumber of samples is determined as true (true positive/true negative) and false (false negative/false positive).

A direct comparison of plasma and serum models showed better performance (Supplementary Data 8) and predictive ability (Table 2) for plasma models, particularly in males. Moreover, all misclassified samples in the plasma model had prediction values within the range of 0.4–0.6, indicating that they were close to the decision boundary. Model performance was visualized using ROC curves for the combined-gender dataset, with predictions derived from gender-specific models. The area under the ROC curve confirmed the superiority of plasma models for lipidomic profiling (Fig. 2G). This conclusion was independently verified in external laboratory using a FIA-MS/MS method, in which the smaller cohorts of plasma and serum samples derived from the same sample set were compared. Significant differentiation between healthy controls and PDAC patients was achieved (Fig. S5), primarily based on sphingolipids and phospholipids, and the observed dysregulations (Supplementary Data 13) were consistent with those found using UHPSFC/MS. The FIA-MS/MS results also confirmed the advantage of plasma models, both in terms of model parameters (Supplementary Data 8) and prediction values (Supplementary Data 9), as visualized by ROC curves (Fig. S5C). However, due to the lower number of lipid species analyzed, the overall model quality for FIA-MS/MS was inferior to that of UHPSFC/MS.

Lipidomic profiling of PDAC patients and healthy controls

The lipidomic profiling method demonstrates high consistency in the most dysregulated lipid species in nearly all samples, as visualized by the heatmap (Fig. 3A), with similar trends and levels of statistical significance observed between genders (Fig. 3C). A consistent pattern of dysregulation is evident, suggesting disruptions in certain metabolic pathways (Fig. 3B). All 190 quantified lipid species by UHPSFC/MS were used for lipidomic profiling, even though part of lipid species had only low statistical significance, but their cumulative effect on model performance was observed in our previous study⁴⁰. The downregulation is observed in ceramides and sphingomyelins with very long N-acyl chains (40:1, 40:2, 41:1, 41:2, and 42:1), as well as in plasmenyl/ether phospholipids (36:2, 36:3, 36:4, and 38:4). In contrast, several DG (DG 34:1, 34:2, 36:2, 36:3, and 36:4) are upregulated. These DG species lie at the metabolic crossroads of phospholipid synthesis, while phospholipids with the same acyl chain composition are downregulated. The behavior of PE also varies depending on their structural class—acyl-bonded PE are upregulated, whereas plasmenyl/ether-bound PE P-/O- are downregulated.

Fig. 3. Plasma lipidomic alterations in PDAC.

Results of the lipidomic profiling of PDAC patients (T, tumor) and healthy controls (N, normal) using plasma models. A Heatmap of the most dysregulated lipids, showing the concentrations of lipid species across individual samples. B Network mapping of lipid species, where the size of each circle reflects p-values (two-sided t-test) for individual lipids, and red/blue color saturation represents fold change (T/N). C Box plots of the most dysregulated lipid species visualizing differences in concentrations between PDAC patients (T, red) and healthy controls (N, blue). The number of significance symbols corresponds to p-value ranges from the Mann–Whitney U test; p < 0.0001 is indicated by ****. Study groups: 93 N vs. 88 T of males and 125 N vs. 89 T of females. D Sensitivity (red), specificity (blue), and accuracy (green) based on predictions using the lipidomic profiling method (lipids), carbohydrate antigen 19-9 (CA 19-9), and carcinoembryonic antigen (CEA) in the training and validation sets. E Sensitivity of individual methods according to tumor stages (T1—yellow, T2—orange, T3—red, T4—rose, and Tx—brown, where information about the stage is not available) for the lipidomic profiling method (Lipids), CA 19-9, and CEA in the training set.

The best performing model for differentiating healthy controls from PDAC patients is the gender-specific plasma model, which is used for all subsequent analyses. The lipidomic profiling results were compared with those of the established tumor markers CA 19-9 and CEA in the same set of samples (Fig. 3D). CA 19-9 is commonly used for the diagnosis of pancreatic cancer in symptomatic patients and for monitoring of therapy in these patients, while CEA for monitoring disease progression and tracking the response to treatment in patients with PDAC. Because lipidomic profiling requires prior model training, the dataset was divided into training and validation subsets. The results for CA 19-9 and CEA were assessed using standard threshold values (37 U/mL for CA 19-9 and 5 ng/mL for CEA) and divided in the same subset consistent with the lipidomic profiling. The lipidomic profiling method achieved 99% (95% CI, 97–100%) accuracy in the training dataset and 96% (95% CI, 89–99%) in the validation dataset (averaged across both genders). Specificity was 100% for both datasets (95% CI, 97–100% for training set and 92–100% for validation set), with sensitivity reaching 98% (95% CI, 95–100%) in the training set and 92% (95% CI, 77–98%) in the validation set. In comparison, both CA 19-9 and CEA showed high specificity above 98%, but substantially lower sensitivity—66% (95% CI, 59–73%) and 31% (95% CI, 24–38%) for the training set, and 60% (95% CI, 42–76%) and 20% (95% CI, 8–37%) for the validation set, respectively (Table 3).

Table 3.

Sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) (with 95% confidence interval) for the lipidomic profiling, CA 19-9, and CEA methods for the training set, validation set, and high-risk individuals (HRI) for developing pancreatic cancer

Method	Dataset	Sensitivity [%]	Specificity [%]	Accuracy [%]	PPV [%]	NPV [%]	Truea	Falsea
Lipidomic profiling	Tr.	98 (95–100)	100 (97–100)	99 (97–100)	99 (96–100)	99 (96–100)	174/217	3/1
	Va.	92 (77–98)	100 (92–100)	96 (89–99)	100 (89–100)	94 (83–98)	32/44	3/0
	HRI	–	96 (89–99)	–	–	100 (96–100)	–/89	– 4
CA 19-9	Tr.	66 (59–73)	99 (96–100)	84 (80–88)	98 (93–99)	78 (74–82)	117/215	60/3
	Va.	60 (42–76)	98 (88–100)	81 (71–89)	95 (75–99)	75 (67–82)	21/43	14/1
	HRI	–	93 (85–97)	–	–	100 (96–100)	–/86	–/7
CEA	Tr.	31 (24–38)	99 (97–100)	70 (64–73)	96 (87–99)	64 (62–66)	55/216	122/2
	Va.	20 (8–37)	98 (88–100)	63 (52–74)	88 (47–98)	61 (56–65)	7/43	28/1
	HRI	–	96 (89–99)	–	–	100 (96–100)	–/89	–/4

Concentrations of lipids were measured by UHPSFC/MS method.

^aNumber of samples is determined as true (true positive/true negative) and false (false negative/false positive).

Importantly, early-stage tumors (T1 and T2) remain a critical challenge for CA 19-9 and CEA, with markedly reduced sensitivity, whereas the lipidomic profiling method maintains comparable sensitivity across all stages (Fig. 3E). Results for each method are summarized in Supplementary Data 14, and stage-specific sensitivity is provided in Supplementary Data 15. The findings clearly demonstrate the superior diagnostic performance of the lipidomic profiling method, particularly in detecting early-stage of PDAC.

Lipidomic profiling of high-risk individuals

Although the lipidomic profiling method reliably distinguishes between healthy controls and PDAC patients, the lipid profiles of HRI may be influenced by other factors. Therefore, the validation of the method in this population is essential. The total of 93 HRI samples were analyzed as blinded samples using models built from 93 healthy controls (N) and 88 PDAC patients (T) for males, and 125 N and 89 T for females (Fig. 4A). Although the age distribution and the number of patients with diabetes mellitus in the individual subgroups are not perfectly matched (Table 1), the effect of these parameters on lipid profiling was examined in our previous study²⁹, and no influence on investigated lipid classes was observed. Therefore, this imbalance does not represent a limitation of the study. The visual inspection of the prediction results indicates that the lipid profiles of HRI without cancer closely resemble those of healthy controls, which is supported by box plots of concentrations of key lipids (Figs. 4B and S6). Prediction outcomes (Supplementary Data 14) identified 85 samples as negative (<0.4), 7 within a lower-confidence range (0.4–0.6), and 1 as positive (>0.6). Using a strict cutoff of 0.5, the results correspond to 89 correct and 4 incorrect predictions, which yields the specificity of 96% (95% CI, 89–99%) compared with outcomes from periodic EUS and/or MRI examinations. The participants were examined using at least one of the imaging techniques at the time of blood sample collection, and they were then invited each year for a follow-up examination. However, not all participants continued in the study or attended the scheduled yearly examinations. The results of imaging examinations for individual participants are summarized in Supplementary Data 1. Although pancreatic abnormalities, 2 IPMNs, 1 benign cyst, 3 unspecified cystic lesions, and 29 cases of chronic pancreatitis, were detected, no participants developed pancreatic cancer during the study period, and all were considered PDAC-negative based on imaging data. CA 19-9 and CEA were also measured in all HRI samples, providing specificities of 93% (95% CI, 85–97%) and 96% (95% CI, 89–99%), respectively (Table 3). However, the positive findings from lipidomic profiling, CA 19-9, and CEA did not occur in the same individuals (Supplementary Data 14). HRI were stratified by inclusion criteria into groups with chronic pancreatitis, familial pancreatic cancer, BRCA1/BRCA2 mutations, and Peutz–Jeghers syndrome. The lipidomic profiling method did not show increased misclassification in any specific group (Fig. 4C). Follow-up blood collection after one year was available for 23 HRI subjects. Most follow-up predictions remained close to the original classification, typically below the 0.5 cutoff, indicating consistent negative results (Fig. 4D). However, prediction scores in 4 individuals shifted from negative (< 0.4) to the lower-confidence range (0.4–0.5), potentially suggesting early metabolic changes. Imaging methods revealed no abnormalities in these cases, and continued long-term monitoring is warranted. The collaboration with HEPACAS and SCREPAN, two national pilot studies in the Czech Republic focused on early PDAC detection through annual imaging-based surveillance of HRI, is ongoing. Participants continue to be monitored at one-year intervals.

Fig. 4. Application of the lipidomic profiling method for patients from high-risk (HR) groups for developing pancreatic cancer (plasma models).

Study groups: 93 N, 56 HRI, and 88 T of males and 125 N, 37 HRI, and 89 T of females. A Prediction of for high-risk individuals (HRI) using a model trained on PDAC patients (tumor, red) and healthy controls (normal, blue). Prediction scores ≤ 0.5 are classified as normal; >0.5 as tumor. Gray dashed lines at 0.4 and 0.6 denote the low-confidence region. B Box plots of selected dysregulated lipids comparing concentrations among healthy controls (N, blue), high-risk individuals (HRI, light blue—chronic pancreatitis, orange—familial pancreatic cancer, brown—BRCA1 mutation, dark blue—BRCA2 mutation, and pink—Peutz–Jeghers syndrome), and PDAC patients (T, red). C Highlighting of subjects colored according to individual high-risk groups (light blue—chronic pancreatitis, orange—familial pancreatic cancer, brown—BRCA1 mutation, dark blue—BRCA2 mutation, and pink—Peutz–Jeghers syndrome). D Highlighting of subjects from high-risk groups with repeated blood collection after one year.

Discussion

The detection in early stage can be considered as crucial first step to improve PDAC patient prognosis. This study shows that a blood test analyzing lipid molecules can detect PDAC with high accuracy. Physiological and biological factors are often discussed in the context of their influence on lipid profiles, as they can negatively impact biomarker discovery or distort study outcomes. Present results show excellent agreement for gender differences with the findings of Tabassum et al.³⁶, who summarized conclusions from multiple studies and proposed relevant biological hypotheses, as well as good consistency with Sales et al.³⁷ with the exception of differences in PE P-/O- dysregulation and the significance of some lipid classes. The most critical gender-based differences were observed in sphingomyelins, which belong to the most dysregulated lipid classes in the lipidomic profiling of healthy controls and PDAC patients. The impact of gender on lipid profiles was confirmed, leading us to implement gender-separated models in order to minimize this confounding factor. Although these models require more samples for reliable training, the benefit of gender-specific modeling was demonstrated by a lower rate of false predictions and improved prediction accuracy.

Numerous studies have compared lipidomic profiles in plasma and serum^41–47, but the conclusions are not always consistent, as highlighted by Lehmann³⁸. Many of these investigations do not distinguish between metabolites and lipids, instead presenting general conclusions for both together, which can lead to ambiguity. In our study, no statistically significant differences were observed between plasma and serum lipid profiles for analyzed lipid classes, except for diacylglycerols. Differences in lipid concentrations between plasma and serum can be expected, especially due to pre-analytical conditions, such as the room temperature isolation of serum for 20–30 min, during which the enzymatic activity can persist. Even minor degradation of phospholipids or triacylglycerols can disproportionately influence lipid classes with lower concentrations, such as DG and lysophospholipids. Overall, plasma is the preferred matrix for lipidomic analysis, as pre-analytical variability may also affect other lipid classes, including oxidized lipids⁴⁸, endocannabinoids⁴⁹, or oxylipins⁵⁰. However, the pre-analytical part is the critical point and sample mishandling leads to changes in the lipid profile, especially incorrect storage of samples or freeze−thaw cycles resulting in the lipid degradation. Generally, phospholipids are decreased over time, while DG, free fatty acids, and lysophospholipids are increasing due to hydrolysis. In contrast, sphingolipids are more stable and less prone to oxidative degradation, as they predominantly contain saturated or monounsaturated fatty acyl chains⁵¹. The general recommendation is to store samples at least at −80 °C, but extremely long-term storage can also lead to degradations⁵². However, Ryan et al.⁵³ reported the stability of lipid profiles for at least one year in samples stored at −80 °C without any freeze−thaw cycles.

The lipid species dysregulations identified in this study are consistent with our previous work focused on PDAC biomarker discovery^29,39. Here, we validated these findings in new, prospectively collected cohorts using both plasma and serum matrices. The consistent downregulation of ceramides and sphingomyelins suggests metabolic disturbances in ceramide biosynthesis, supported by similar patterns observed in glycosphingolipids³⁹. While it remains unclear whether these metabolic alterations are a cause or consequence of cancer, similar dysregulation has been observed across hundreds of samples in multiple independent studies. Sphingolipids, especially ceramides, play a key role in regulating apoptosis and cell signaling⁵⁴, processes that may influence cancer cell survival and proliferation. However, the confirmation of these mechanistic hypotheses requires additional research using biological models and multiomics approaches. The primary objective of this study was to apply our refined methodology to HRI of developing pancreatic cancer. Present findings indicate that the lipidomic profiles of HRI closely resemble those of healthy controls, with no detectable metabolic alterations for investigated lipid classes. These results suggest that lipidomic profiling can distinguish HRI and healthy controls from patients with PDAC, supporting its potential use as a tool for detection of PDAC patients in high-risk cohorts.

Due to the high biological variability of the human lipidome, the identification of a single lipid species as a robust biomarker is unlikely because concentration ranges of dysregulated lipids often partially overlap across groups. In contrast to current diagnostic markers that typically rely on a single compound and a defined threshold, the lipidomic profiling method evaluates a broad range of lipid species using MDA. This approach enables the modeling of complex relationships among multiple variables, resulting in more accurate, robust, and comprehensive diagnostic performance than univariate methods⁵⁵. In screening applications, the balance between specificity and sensitivity can be adjusted by modifying the decision cutoff to reduce false negatives, although this may lead to an increased number of follow-up imaging examinations in healthy individuals. While omics-based approaches are not yet standard in clinical practice, they are gaining ground in the biomarker discovery research. One example is the use of circulating ceramides as biomarkers for cardiovascular disease^56,57. Current trends also include the implementation of multi-biomarker panels¹⁴, as well as the use of machine learning^58,59, and artificial intelligence⁶⁰, to enhance predictive accuracy. Nevertheless, machine learning methods may pose challenges, such as overfitting and poor interpretability, which could hinder clinical translation and lead to misleading conclusions.

This study builds on our long-term research focused on early detection of PDAC through lipidomic analysis of plasma and serum. We present the robust methodology optimized for the clinical application. The same dysregulated lipid patterns were confirmed in newly collected samples, with 54% of early stages (T1 and T2). The diagnosis of early-stage PDAC remains particularly challenging in Lewis-negative individuals⁶¹. Although these subjects were not explicitly evaluated in this study, our cohort contained 103 individuals with CA 19-9 values at the instrumental threshold (Supplementary Data 1), comprising 66 healthy controls, 19 HRI, and 18 PDAC patients. 99 of them were correctly classified using the lipidomic test (Supplementary Data 14), particularly all PDAC patients (1× T1, 8× T2, 3× T3, 1× T4, and 5× Tx) who were included in the training set and three of whom also in the validation set, which demonstrates its strong performance for subjects with low CA 19-9 secretion. Moreover, the method achieved an average sensitivity of 99% in the training set and 96% in the validation set for early-stage cases for the plasma models. Across all stages, the average sensitivity, specificity, and accuracy were 98% (95% CI, 95–100%), 100% (95% CI, 97–100%), and 99% (95% CI, 97–100%) in training, and 92% (95% CI, 77–98%), 100% (95% CI, 92–100%), and 96% (95% CI, 89–99%) in validation sets, which is an improvement of 4–10% compared to our previous method²⁸ and approximately 30% higher sensitivity than CA 19-9. The methodology also demonstrated 96% (95% CI, 89–99%) specificity for HRI when benchmarked against results from EUS and MRI. The method is noninvasive, high-throughput, inexpensive, and shows excellent diagnostic performance, which are key attributes for a viable screening tool. However, a limitation of this study is the incomplete assessment of method specificity compared to other cancer types. The study was designed to evaluate the distinction of PDAC from healthy controls, and HRI, and no additional cancer types were included. Therefore, the cancer specificity of the method compared to other cancer types cannot yet be determined. In this study, no medications were used as an exclusion criterion and their effect was not investigated. Although some medications can affect the lipid profile⁶², we did not observe any outliers in the models. Considering the high prediction performance of the method, commonly used medications in the population, do not appear to have a limiting effect. However, the effect of medications and other factors is the subject of the follow-up research. Another limitation of this study is retrospective measurement of the samples, the relatively small number of samples, particularly in the HRI group for pancreatic cancer and their short follow-up time. However, the participants monitoring continue at one-year intervals, and the results of this long-term study require prolonged duration. Based on these findings, a multicenter clinical trial has been initiated (registered at ClinicalTrials.gov, NCT6549725) in collaboration with 14 clinical centers in the Czech Republic. The trial aims to validate the method on a large cohort of HRI and PDAC patients with resectable tumor with comparison to results of imaging methods.

Author contributions

Conceptualization: O.P., R.J., and M.H.; formal analysis: O.P., Z.D., M.D., J.I., and O.S.; funding acquisition: M.H.; investigation: Z.D., M.D., J.I., and J.B.; methodology: O.P., R.J., Z.D., M.D., J.I., and J.B.; project administration: M.H.; resources: M.V., B.M.-D., I.K., M.L., B.M. O.U., J.T., and P.K.; supervision: R.J., K.P., B.M., and M.H.; visualization: O.P., R.J., and J.I.; writing original draft: O.P.; all authors read and approved the final manuscript.

Peer review

Peer review information

Communications Medicine thanks Alcibiade Athanasiou, Vijayasarathy Ketavarapu and Laura E. Kane for their contribution to the peer review of this work. A peer review file is available.

Data availability

The lipidomics datasets are provided for UHPSFC/MS measurement in Supplementary Data 2 and 3 and for FIA-MS/MS measurement in Supplementary Data 4. All figures were prepared from the data available in Supplementary Data. The raw data are available at 10.5281/zenodo.15409052⁶³. Additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Code availability

The instructions for software handling and source data for the software are available at 10.5281/zenodo.15409052⁶³.

Competing interests

O.P., Z.D., and M.D. are employees of Lipidica, a.s. M.H., O.U., and B.M. are members of the scientific council of Lipidica. M.H. and R.J. are listed as inventors on the patents EP3514545B1 and US 12,247,981 B2 related to this work. All other authors declare no conflict of interest.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-026-01445-5.