Inference of upstream-mutation and metabolomic-signature causality identifies prognostic biomarkers and therapeutic targets in pancreatic cancer Authors: Feng Chen, Xiaoying Lou, Xingting Guo, Ye Guo, Yuan Huang, Kexin Li, Yan Song, Jiangtao Li, Jing Wang, Kai Lin, Ruijun Tian, Min Wang, Xudong Dai, Hezhi Fang, Wei Cui Published: 2026 ABSTRACT Genomic mutations in pancreatic ductal adenocarcinoma (PDAC) are hypothesized to drive poor prognosis and low response rates to targeted therapy through crosstalk among downstream regulatory networks. Here, we apply a causal inference-based approach, Mutation-Upstream-of-Metabolomic-Signature (MUMS), to show that prognostic serum metabolomic signatures can capture such crosstalk and reflect the collective impact of mutation-driven networks on tumor progression. We identify a panel of nine serum metabolites that predicts survival outcomes across multiple independent PDAC cohorts. MUMS analysis further identifies and functionally validates GRPEL1 as a tumor-promoting gene whose downstream metabolic signature converges with that of the mTOR/PI3K/Akt signaling pathway. Consistently, GRPEL1 sensitizes PDAC cells to proliferation arrest induced by mTOR inhibition. Together, our findings provide proof-of-concept evidence that serum metabolic signatures can reflect crosstalk within the tumor mutational landscape. These co-regulatory patterns offer a framework for uncovering new therapeutic targets and guiding the design of rational combination therapies. Introduction Pancreatic ductal adenocarcinoma (PDAC) is one of the most lethal malignancies and has an exceptionally high mortality rate1, with a 5-year overall survival rate of less than 8% for all patients with PDAC2,3. Proper prognostic stratification of patients will facilitate better management and improve the clinical outcome of patients with PDAC. Serum-based biomarker is readily available, and the collection method is noninvasive and generally acceptable. Serum-based biomarker is an ideal way to predict prognostic survival in patients with PDAC. Whereas few serum-based biomarkers are available to inform prognostic outcomes in patients with PDAC. Thus, it is urgent to develop a powerful approach for the identification of a noninvasive biomarker for prognostic stratification of patients with PDAC. Among candidate markers suggested to improve management of PDAC by helping to predict PDAC progression are driver gene mutations, which have been reported as commonly occurring and vital molecular events during the onset and progression of tumorigenesis4–6. Mutations in KRAS, TP53, SMAD4, and CDKN2A were predominant5,7. Clinical research has revealed that patient outcomes are associated with alterations in driver genes in pancreatic adenocarcinoma, and patients with a greater number of altered driver genes exhibit worse disease-free survival and overall survival (OS)8. Thus, understanding genetic variances and their downstream events, especially in metabolism reprogram, that contribute to PDAC would provide an opportunity to find sensitive biomarkers for the prediction of PDAC progression. Mutations in KRAS and other canonical oncogenes (e.g., AKT, MYC, and PI3K) and tumor suppressors (e.g., TP53, RB, and PTEN) that drive accelerated growth also directly reprogram cellular metabolism to fuel the elevated demand for energy and macromolecular biosynthesis during PDAC progression9,10. In addition, the progression of PDAC has a profound influence on whole-body metabolism11,12. Thus, the metabolic alterations induced by driver gene mutation may serve as an important clue for the identification of biomarkers for the prediction of PDAC progression. Based on the role of driver genes in the progression of PDAC and their profound effect on whole-body metabolism, combining genetic and metabolic information provides a unique opportunity to gain further insights into how genetic mutations carry out their biological function by modulating metabolites and how alterations in the genome are associated with the outcome of diseases. These so-called genetically influenced metabotypes have provided functional insights into the serum concentrations of individual clinically relevant metabolites, such as glucose, insulin, lipids, and uric acid13–17. Thus, metabolomics provides a direct functional readout of activity and phenotype encoded in the genome. As the pancreas is a major organ involved in metabolic regulation, multi-omics analysis of alterations in serum metabolome profiles induced by tumor genomic variations may find sensitive biomarkers for the prediction of prognostic survival in patients with PDAC. Here, we performed an integrated analysis of serum metabolomics and whole-exome sequencing (WES) of matched PDAC tissue samples and identified causal inference links among the tumor genetic variation landscape, serum metabolome, and PDAC progression (OS). Through this multi-omics analysis, we identified that serum metabolome heterogeneity in patients with PDAC exhibits distinct prognostic outcomes, and that defined mutations upstream of metabolomic signature (MUMS) related to these metabolic subtype signatures demonstrate a prognostic ability. Based on this, we developed a 9-serum metabolite panel that can efficiently predict PDAC progression. Moreover, we identified multiple PDAC-promoting genes based on MUMS, and functionally validated that GRPEL1 could sensitize PDAC cells to mTOR inhibition-induced proliferation arrest. Our study provides proof-of-concept evidence for the approach of identifying blood metabolites as noninvasive biomarkers and establishes the potential value of causal links between tumor genetic mutations and serum metabolites as resources for mechanistic analysis. Results Heterogeneity in the serum metabolome of patients with PDAC was associated with distinct prognostic outcomes To explore causal links between the tumor mutational landscape, the serum metabolome, and PDAC progression, we first examined heterogeneity in the serum metabolome of patients with PDAC and assessed whether metabolomic profiles were associated with survival outcomes (Fig. 1a). We performed untargeted metabolomic profiling using a tissue–serum matched cohort of 41 patients, each followed for at least 2 years to record OS, with a median OS of 17 months (Supplementary Table 1). A total of 36,118 metabolic features (CV < 30% in QC samples) were detected in serum (Supplementary Fig. 1a), of which 5,057 were annotated metabolites previously cataloged in public databases (Metlin, HMDB, and BioML). To characterize distinct serum metabolic patterns, we performed unbiased clustering of these metabolites, which classified the 41 patients into three metabolic groups (Fig. 1b). These groups were not significantly associated with clinical variables including age (P = 0.712), sex (P = 0.438), clinical stage (P = 0.586), smoking (P = 0.15), alcohol consumption (P = 0.351), or CA19-9 levels (P = 0.871) (Supplementary Table 2). We next evaluated whether these metabolic patterns stratified survival using Kaplan–Meier analysis. The results revealed significant differences in OS across groups, with Group 1 exhibiting the shortest and Group 3 the longest survival (Fig. 1c), suggesting that serum metabolomic patterns may be informative of prognosis in PDAC.Fig. 1Serum metabolome heterogeneity in patients with pancreatic ductal adenocarcinoma exhibited distinct prognostic outcomes.a Schematic of the experimental design with the matched tissue–serum cohort. All individuals in the matched tissue‒serum cohort (n = 41patients) were patients diagnosed with PDAC. FFPE slides of paired tumor and para-tumor tissues from each individual was used for WES at 200× coverage. Matched serum samples within this cohort were used for untargeted metabolomics analysis. The genetic variations, serum metabolome, and prognostic outcome were then integrated for multi-omics analysis. Created in BioRender. Lou, X. (2026) https://BioRender.com/zgpsf2k. FFPE formalin-fixed and paraffin-embedded, PDAC pancreatic ductal adenocarcinoma, WES whole exome sequencing. b Unbiased clustering of the serum metabolome reveals serum metabolome heterogeneity. Individuals within the matched tissue–serum cohort (n = 41 patients) were classified into three groups accordingly. In the column of months, blue color represents the overall survival ≥ 17 months, yellow color represents the overall survival < 17months. c Kaplan–Meier survival-curve analysis of the three subtypes classified by the unbiased clustering of serum metabolite landscape. P = 0.025 for overall comparison among groups 1, 2, and 3; P-value was calculated by two-tailed log-rank test. HR hazard ratio. High: the prognostic model score above the cut-off value is designated as high; low: the score below the cut-off value is designated as low. d Venn diagram showing the overlap of signature metabolites in each subtype. e KEGG metabolic pathway enrichment analysis of serum metabolite signatures for each subtype. Statistical significance was assessed using the hypergeometric test, with one-tailed P values adjusted for multiple comparisons using the Benjamini–Hochberg procedure. KEGG Kyoto Encyclopedia of Genes and Genomes. Source data are provided as a Source data file. Serum metabolome heterogeneity in patients with pancreatic ductal adenocarcinoma exhibited distinct prognostic outcomes. a Schematic of the experimental design with the matched tissue–serum cohort. All individuals in the matched tissue‒serum cohort (n = 41patients) were patients diagnosed with PDAC. FFPE slides of paired tumor and para-tumor tissues from each individual was used for WES at 200× coverage. Matched serum samples within this cohort were used for untargeted metabolomics analysis. The genetic variations, serum metabolome, and prognostic outcome were then integrated for multi-omics analysis. Created in BioRender. Lou, X. (2026) https://BioRender.com/zgpsf2k. FFPE formalin-fixed and paraffin-embedded, PDAC pancreatic ductal adenocarcinoma, WES whole exome sequencing. b Unbiased clustering of the serum metabolome reveals serum metabolome heterogeneity. Individuals within the matched tissue–serum cohort (n = 41 patients) were classified into three groups accordingly. In the column of months, blue color represents the overall survival ≥ 17 months, yellow color represents the overall survival < 17months. c Kaplan–Meier survival-curve analysis of the three subtypes classified by the unbiased clustering of serum metabolite landscape. P = 0.025 for overall comparison among groups 1, 2, and 3; P-value was calculated by two-tailed log-rank test. HR hazard ratio. High: the prognostic model score above the cut-off value is designated as high; low: the score below the cut-off value is designated as low. d Venn diagram showing the overlap of signature metabolites in each subtype. e KEGG metabolic pathway enrichment analysis of serum metabolite signatures for each subtype. Statistical significance was assessed using the hypergeometric test, with one-tailed P values adjusted for multiple comparisons using the Benjamini–Hochberg procedure. KEGG Kyoto Encyclopedia of Genes and Genomes. Source data are provided as a Source data file. To further elucidate the biological relevance of serum metabolic patterns to survival outcomes, we next characterized subtype-specific metabolic signals and identified enriched metabolic pathways. Metabolites with significantly altered abundance in a given subtype compared to the other two were defined as subtype-specific (Fig. 1d and Supplementary Data 1). Subsequent metabolic pathway enrichment analysis of the top two pathways revealed that Group 1, the subtype with the poorest prognosis, exhibited overrepresentation of purine metabolism and inositol phosphate metabolism (Fig. 1e). This finding aligns with previous studies implicating purine pathway activation in supporting tumor growth18,19, and the inositol phosphate pathway as a key regulatory hub associated with multiple proliferation-related signaling cascades, including CK2, AKT, mTOR, and MYC20. In contrast, Groups 2 and 3, which were associated with more favorable prognostic outcomes, were enriched in metabolite pathways not linked to canonical cancer-associated signaling. This finding aligns with the absence of the activated anabolic metabolic phenotypes observed in Group 1 (Fig. 1e). Together, these results underscore the substantial heterogeneity of serum metabolomic profiles in patients with PDAC and demonstrate that serum metabolome-based subtyping is significantly associated with distinct prognostic outcomes. Prognostic diverse metabolic subtypes reflect cross-talk of distinct mutational signatures Distinct mutational profiles can contribute to different metabolic subtypes with varied prognostic outcomes21. To investigate how genetic mutational diversity in patients with PDAC contributes to the three distinct prognostic metabolic subtypes and, in turn, whether such prognosis-related serum metabolomic heterogeneity could serve as a noninvasive tool to reflect the driver mutational status in patients, we performed WES of tissue samples and an integrated genetic–metabolic–survival analysis within the matched tissue–serum cohort (Fig. 1a and Supplementary Fig. 1b). At the transcription level, previous findings has identified that rather than a single cis-regulation of downstream events, tumor genetic mutations lead to alterations in multiple gene expression variations, which are termed signatures of the related driver genes22. Moreover, downstream events are also subjected to integrative regulation of multiple upstream driver genes23. We also observed cross-regulation of the four main driver genes (KRAS, TP53, CDKN2A, and SMAD4) on the respective downstream gene signatures at the transcriptional level in PDAC (Supplementary Fig. 2). Such cross-talk between driver mutations in the regulation of downstream signatures was consistent at the metabolic level (Fig. 2a). Thus, we hypothesized that each distinctive metabolic subtype represents the aggregating effects of upstream mutations that can drive the corresponding alteration of these metabolites through modulating the activities of downstream signaling pathways and inter-pathway cross-talk.Fig. 2Diverse prognostic metabolic subtypes reflect diverse oncogene signatures.a Circular plot showing cross-talk of the four main driver mutations with downstream metabolic signatures. b Sankey plot showing cross-talk between genes, showing MUMS and serum metabolome signatures allocated to each metabolic subtype. MUMS mutations upstream of the metabolomic signature. c Plot showing the ranking of genes showing MUMS for each subtype. P values were calculated by Fisher’s exact test (one-sided). MUMS mutations upstream of the metabolomic signature. d Heatmap displaying the mutational profiles and enriched gene ontology pathways of metabolic subtype-associated genetic mutations in the matched tissue–serum cohort (n = 41 patients). In the column of months, blue color represents the overall survival ≥ 17 months, yellow color represents the overall survival < 17months. Gene groups (Group 1–3) were derived through an unsupervised clustering analysis based solely on the mutation status of all MUMS genes. Group 1, Group 2, or Group 3 MUMS was divided by the enrichment patterns of their associated metabolites, which were previously identified in Fig. 1c, e. e The prognostic value of the subtype metabolites-enriched MUMS gene signature was assessed in the matched tissue-serum cohort (n = 41 patients) using Kaplan–Meier survival curves, with statistical significance determined by a two-tailed log-rank test. MUMS mutations upstream of the metabolomic signature. f Expression profile of metabolic subtype-associated genes showing MUMS in the TCGA and CTPAC cohort (n = 275 patients) and unbiased subtyping of the TCGA and CTPAC cohort into three subtypes based on the expression patterns of these genes. MUMS mutations upstream of the metabolomic signature. g Validation of the enriched gene panel for predicting survival using the transcriptome dataset of the TCGA and CTPAC PDAC cohort (n = 275 patients) was performed by Kaplan–Meier survival-curve analysis, with P-values calculated by the two-tailed log-rank test. Source data are provided as a Source data file. Diverse prognostic metabolic subtypes reflect diverse oncogene signatures. a Circular plot showing cross-talk of the four main driver mutations with downstream metabolic signatures. b Sankey plot showing cross-talk between genes, showing MUMS and serum metabolome signatures allocated to each metabolic subtype. MUMS mutations upstream of the metabolomic signature. c Plot showing the ranking of genes showing MUMS for each subtype. P values were calculated by Fisher’s exact test (one-sided). MUMS mutations upstream of the metabolomic signature. d Heatmap displaying the mutational profiles and enriched gene ontology pathways of metabolic subtype-associated genetic mutations in the matched tissue–serum cohort (n = 41 patients). In the column of months, blue color represents the overall survival ≥ 17 months, yellow color represents the overall survival < 17months. Gene groups (Group 1–3) were derived through an unsupervised clustering analysis based solely on the mutation status of all MUMS genes. Group 1, Group 2, or Group 3 MUMS was divided by the enrichment patterns of their associated metabolites, which were previously identified in Fig. 1c, e. e The prognostic value of the subtype metabolites-enriched MUMS gene signature was assessed in the matched tissue-serum cohort (n = 41 patients) using Kaplan–Meier survival curves, with statistical significance determined by a two-tailed log-rank test. MUMS mutations upstream of the metabolomic signature. f Expression profile of metabolic subtype-associated genes showing MUMS in the TCGA and CTPAC cohort (n = 275 patients) and unbiased subtyping of the TCGA and CTPAC cohort into three subtypes based on the expression patterns of these genes. MUMS mutations upstream of the metabolomic signature. g Validation of the enriched gene panel for predicting survival using the transcriptome dataset of the TCGA and CTPAC PDAC cohort (n = 275 patients) was performed by Kaplan–Meier survival-curve analysis, with P-values calculated by the two-tailed log-rank test. Source data are provided as a Source data file. To test this hypothesis, we first determined whether the metabolomic signature of each subtype could be predicted by those significantly associated with driver mutations in the tissue–serum matched cohort. We performed a correlation analysis using census genetic variations detected in this cohort with all serum metabolites, and calculated the enrichment score of each gene co-related metabolite within each subtype signature metabolite (Fig. 2b). One genetic mutation was defined as specifically overrepresented for certain individual subtypes when the number of its correlated serum metabolites was significantly enriched within the signature metabolites of this subtype, and the enrichment was calculated using Fisher’s exact test, with a threshold of P < 0.05. Such overrepresented genes based on the enrichment of co-related metabolites within a certain metabolite set were termed MUMS. Based on this analysis, we identified lists of enriched MUMS for the metabolic signatures of each subtype (Fig. 2c). To experimentally test the causal links between genetic mutations and metabolic signatures, top three contributor genes for Group 1 (MAP2K2/MEK2, PDGFRB, and FAT4), Group 2 (LATS2, FANCE, and PCM1), and Group 3 (CNTNAP2, MYD88, and BCL2L12) were called out based on their rank of the direct corelated metabolite numbers (Supplementary Table 3). Attempts to generate PANC-1 cells with deletion of PDGFRB (Group 1), FANCE and PCM1 (Group 2), and CNTNAP2 and BCL2L12 (Group 3) were unsuccessful, while the knockouts proved lethal. In contrast, PANC-1 cells with deletion of MEK2 and FAT4 (Group 1), LATS2 (Group 2), and MYD88 (Group 3) were successfully generated and subjected to metabolic profiling (Supplementary Fig. 3a). Orthogonal partial least-squares discriminant analysis (OPLS-DA) revealed distinct metabolic profiles in PANC-1 cells with deletion of MEK2, FAT4, LATS2, or MYD88. Clear separation was observed between control cells and those with deletion of MEK2 or FAT4 (Group 1), LATS2 (Group 2), or MYD88 (Group 3) (Supplementary Fig. 3b). Notably, MEK2- and FAT4-deleted PANC-1 cells exhibited metabolic profile changes that were similar to one another, but distinct from those observed in LATS2- or MYD88-deleted cells and controls (Supplementary Fig. 3c). Together, these findings provide strong evidence for causal links between specific genetic mutations and serum metabolic signatures in patients with PDAC. To better understand the biological processes and pathways through which these MUMS contributed to the specification of these subtype-enriched metabolome signatures, we performed gene ontology (GO) analysis using MUMS for each of these three subtypes. Consistent with the enrichment of glycolysis/gluconeogenesis pathway-related metabolites in Group 1 (Fig. 1e), the genes overrepresented in this group also showed enrichment of cellular responses to hypoxia and positive regulation of the glycolytic process. Such metabolic activity-enriched PDAC subtypes have been defined as immunosuppressive and chemotherapy-resistant24,25, which explains the poor prognostic outcome in this group of study patients (Fig. 2d). In contrast, Group 3, which had the best prognostic outcome, was enriched with genes belonging to multiple interleukin signaling (Fig. 2d), suggesting the infiltration of immune cells. Such immune-active signature has also been reported to indicate better survival25. Moreover, we further determined whether these MUMS for each subtype could independently predict prognostic outcomes within this tissue–serum matched cohort, as corresponding metabolomic signatures for each individual subtype. For each individual, we calculated a signature index for each subtype-enriched MUMS gene. Based on the highest mutational signature score, we then categorized the 41 patients from the matched tissue–serum cohort into one of the three subtypes. Despite being not significant, this MUMS gene-based subtyping showed a consistent trend of prognostic prediction with metabolite-based stratification, with Group 1 MUMS-enriched patients showing the poorest survival, while Group 3 MUMS-enriched patients showed the best survival (Fig. 2e). Using expression levels of MUMS genes within the TCGA and CPTAC cohort of PDAC (Fig. 2f), Kaplan–Meier analysis also revealed significant specification of good or poor prognostic outcomes (Fig. 2g). Collectively, these results suggest that cross-talk of genetic mutational diversity in PDAC contributes to the three metabolic subtypes with distinct prognoses and that genetic mutations associated with distinct serum metabolomic signatures also showed similar prognostic power to the corresponding metabolomic subgroups. A prognostic model established on nine genetic-variations-induced serum metabolites showed reliable prediction ability for the survival of patients with PDAC We therefore aimed to establish a noninvasive prognostic prediction model based on the serum metabolite alterations driven by tumor genomic variations. To this end, we first identified these metabolite alterations in the matched tissue–serum cohort and then evaluated the model’s prognostic performance in an independent validation cohort. To identify candidate serum metabolic biomarkers, we applied the least absolute shrinkage and selection operator (LASSO) algorithm (Fig. 3a) to the untargeted metabolomic data obtained from the tissue–serum matched cohort. This analysis yielded a nine-metabolite panel comprising lysophosphatidylcholine (LysoPC) species (14:0, 18:2, and 18:1), lysophosphatidylethanolamine (LysoPE) species (20:5, 22:5, and 16:1), androsterone sulfate, α-carboxyethyl hydroxychroman (α-CEHC), and glycodeoxycholic acid sulfate. To facilitate clinical utilization, we subsequently developed a multiple reaction monitoring (MRM)-based pseudo-targeted metabolomics detection method for these nine metabolites. The specific parent and fragment ion pairs for each biomarker were acquired from the MS/MS spectrum of their respective pure standards. These nine metabolites can be detected with reliability (Supplementary Table 4). Moreover, when assessed using a pseudo-targeted metabolomics approach in the tissue–serum matched cohort, this nine-metabolite panel effectively stratified patients by survival outcome (Fig. 3b; HR 0.4096, 95% CI 0.1941–0.8643; P = 0.015; C-index = 0.6793). Within this panel, LysoPC(14:0) and LysoPE(16:1) emerged as the primary contributors associated with better and worse prognosis, respectively. Importantly, the panel demonstrated improved prognostic performance compared to CA19-9, the classic prognostic and predictive marker for PDAC, yielding higher C-index values in the matched tissue–serum cohort (0.6793 vs. 0.5883), the modeling cohort (0.5811 vs. 0.5298), and the validation cohort (0.5902 vs. 0.5196) (Fig. 3c and Supplementary Fig. 4)26.Fig. 3Genetic variations triggered serum metabolites predict PDAC survival.a Schematic diagram of the serum metabolite biomarker selection process for the pancreatic ductal adenocarcinoma prognosis model. b The 9-metabolite panel was detected by untargeted metabolomics. A prognostic Cox model was established within the matched tissue–serum cohort (n = 41 patients). The cutoff value was set using the median point of the composite scores calculated by the Cox model within this cohort, leading to clear separation between good and poor survival (HR 0.4096, 95% CI 0.1941–0.8643; P = 0.015). CI confidence interval, HR hazard ratio. High: the prognostic model score above the cut-off value is designated as high; low: the score below the cut-off value is designated as low. c The 9-metabolite panel was detected by pseudo-targeted metabolomics. A prognostic Cox model was established with the modeling cohort (n = 151 patients), and the cutoff was set to the median point of the composite score of the Cox model within this cohort, leading to clear separation between good and poor survival (HR 0.5053, 95% CI 0.3151–0.8104; P = 0.0042). Validation of the prognostic model and the cutoff value using the independent validation cohort (n = 79 patients). This model also showed clear separation (HR 0.5415, 95% CI 0.3014–0.9729; P = 0.038). CI confidence interval, HR hazard ratio. High: the prognostic model score above the cut-off value is designated as high; low: the score below the cut-off value is designated as low. The HR was calculated using the Cox proportional hazards model, with HR = exp(β) and 95% CI derived from the model’s Wald test. C-index was computed using the survConcordance function from the survival package in R (version 3.8-6). P-values were calculated using the two-tailed log-rank test. The fitted regression line is displayed with its 95% confidence bands (shaded area) (b, c). Source data are provided as a Source data file. Genetic variations triggered serum metabolites predict PDAC survival. a Schematic diagram of the serum metabolite biomarker selection process for the pancreatic ductal adenocarcinoma prognosis model. b The 9-metabolite panel was detected by untargeted metabolomics. A prognostic Cox model was established within the matched tissue–serum cohort (n = 41 patients). The cutoff value was set using the median point of the composite scores calculated by the Cox model within this cohort, leading to clear separation between good and poor survival (HR 0.4096, 95% CI 0.1941–0.8643; P = 0.015). CI confidence interval, HR hazard ratio. High: the prognostic model score above the cut-off value is designated as high; low: the score below the cut-off value is designated as low. c The 9-metabolite panel was detected by pseudo-targeted metabolomics. A prognostic Cox model was established with the modeling cohort (n = 151 patients), and the cutoff was set to the median point of the composite score of the Cox model within this cohort, leading to clear separation between good and poor survival (HR 0.5053, 95% CI 0.3151–0.8104; P = 0.0042). Validation of the prognostic model and the cutoff value using the independent validation cohort (n = 79 patients). This model also showed clear separation (HR 0.5415, 95% CI 0.3014–0.9729; P = 0.038). CI confidence interval, HR hazard ratio. High: the prognostic model score above the cut-off value is designated as high; low: the score below the cut-off value is designated as low. The HR was calculated using the Cox proportional hazards model, with HR = exp(β) and 95% CI derived from the model’s Wald test. C-index was computed using the survConcordance function from the survival package in R (version 3.8-6). P-values were calculated using the two-tailed log-rank test. The fitted regression line is displayed with its 95% confidence bands (shaded area) (b, c). Source data are provided as a Source data file. Subsequently, this nine-metabolite panel was used to construct a prognostic Cox model in a modeling cohort enrolling 151 patients with PDAC (baseline information in Supplementary Table 1). The cutoff value was set using the median point of the composite scores calculated by the Cox model within this cohort, which led to a clear separation between good and poor survival groups. (HR 0.5053, 95% CI 0.3151–0.8104; P = 0.0042) (Fig. 3c, left panel). Validation of the prognostic model and the cutoff value was then performed using an independent validation cohort that enrolled 79 patients with PDAC (baseline information shown in Supplementary Table 1), which also showed significant separation (HR 0.5415, 95% CI 0.3014–0.9729; P = 0.038) (Fig. 3c, right panel). Although two of the nine metabolites were undetectable in an external validation cohort from Southern China, likely due to prolonged sample storage, the remaining seven metabolites retained significant prognostic performance (Supplementary Fig. 5; HR = 2.0737, 95% CI: 1.164–3.6944, P = 0.011, C-index = 0.5884). Furthermore, no significant differences were observed in clinical variables, including age, sex, clinical stage, smoking, alcohol consumption, or CA19-9 levels, between the good- and poor-survival groups (Fig. 3b, c, and Supplementary Table 5), suggesting that these factors did not confound the metabolism-based survival stratification. Collectively, these findings indicate that serum metabolites influenced by PDAC-associated genetic alterations may serve as noninvasive biomarkers for predicting survival outcomes in patients with PDAC. Causal inference analysis identified PDAC-promoting genes that could be independently validated While gene mutations and serum metabolites are interconnected, the metabolites selected by causal inference analysis in our study could be considered to affect tumor progression, which can manifest as a distinctive prognostic outcome. As described above, the MUMS analysis of cancer census genes interfering survival-related serum metabolites demonstrated prognostic utility for patients with PDAC. However, it remains unclear whether non-census genes can also exhibit metabolic—modulating capabilities and impact PDAC progression. In the matched tissue–serum cohort (n = 41), the mutations identified through the integration of causal inference analysis offer an alternative resource for target discovery (Fig. 4a). In contrast to association-based approaches, our study systematically tests the causal hypothesis that genetic variations influence PDAC patient survival by mediating changes in serum metabolite abundance27–29. Applying this approach, we first evaluated PDAC-enriched somatic alterations, including SNVs, indels, and CNVs. We then identified the serum metabolites associated with these genetic variations and delineated their consequences on patient prognosis (Fig. 4b).Fig. 4Causal inference analysis leads to the identification of pancreatic ductal adenocarcinoma-promoting genes that could be independently validated using the TCGA, CTPAC, and CCLE databases.a Schematic diagram of the 3-layer dataset-based causal relationship algorithm. Using this approach, we attempted to identify the causal link among promoting gene mutations, serum metabolites, and survival outcomes. b Sankey plot showing the regulatory relationship from multiple genetic variants to their respective serum metabolites and finally to survival. c Heatmap showing all serum metabolites affected by potent promoting genes identified by the causal relationship algorithm for each individual, in the column of months, blue color represents the overall survival ≥ 17 months, yellow color represents the overall survival < 17months (upper). Based on these serum metabolites, good and poor survival could be classified, P = 0.00015 (lower). High: the prognostic model score above the cut-off value is designated as high; low: the score below the cut-off value is designated as low. d Independent validation of the panel of potent promoting genes (211 genes) identified by the causal relationship algorithm on pancreatic ductal adenocarcinoma survival prediction based on the transcription of these genes within the TCGA and CTPAC cohort (n = 275 patients). Left: heatmap showing the transcriptional abundances of potent driver genes; Right: prognostic efficacy based on the median value of the transcriptional abundances for each gene in the panel (P = 0.027), supporting the potential role of these genes in pancreatic ductal adenocarcinoma progression. High: the prognostic model score above the cut-off value is designated as high; low: the score below the cut-off value is designated as low. e Distribution of the HR of individual genes for the classification of good and poor survival in the TCGA and CTPAC cohort (n = 275 patients). Genes associated with good survival are highlighted in red, whereas genes associated with poor survival are highlighted in green. f Representative example showing a consistent prognostic trend for CXCR5, a previously reported promoting gene in our matched tissue–serum cohort (n = 41 patients) and TCGA and CTPAC cohort (n = 275 patients). High: the prognostic model score above the cut-off value is designated as high; low: the score below the cut-off value is designated as low. g Dot plot showing dependency values of our causal inference-identified genes in genome-wide Cas9 knockout screening of pancreatic ductal adenocarcinoma cell lines in the CCLE collection. h Fisher enrichment analysis of positive hits in causal inference-identified genes that showed significant dependency for pancreatic ductal adenocarcinoma cell lines (P = 0.0013, one side). Enrichment significance was assessed using one-sided Fisher’s exact test. Nominal P values are reported without correction for multiple comparisons. The HR was calculated using the Cox proportional hazards model, with HR = exp(β) and 95% CI derived from the model’s Wald test. P-values were calculated using the two-tailed log-rank test. The fitted regression line is displayed with its 95% confidence bands (shaded area) (c, d, f). Source data are provided as a Source data file. Causal inference analysis leads to the identification of pancreatic ductal adenocarcinoma-promoting genes that could be independently validated using the TCGA, CTPAC, and CCLE databases. a Schematic diagram of the 3-layer dataset-based causal relationship algorithm. Using this approach, we attempted to identify the causal link among promoting gene mutations, serum metabolites, and survival outcomes. b Sankey plot showing the regulatory relationship from multiple genetic variants to their respective serum metabolites and finally to survival. c Heatmap showing all serum metabolites affected by potent promoting genes identified by the causal relationship algorithm for each individual, in the column of months, blue color represents the overall survival ≥ 17 months, yellow color represents the overall survival < 17months (upper). Based on these serum metabolites, good and poor survival could be classified, P = 0.00015 (lower). High: the prognostic model score above the cut-off value is designated as high; low: the score below the cut-off value is designated as low. d Independent validation of the panel of potent promoting genes (211 genes) identified by the causal relationship algorithm on pancreatic ductal adenocarcinoma survival prediction based on the transcription of these genes within the TCGA and CTPAC cohort (n = 275 patients). Left: heatmap showing the transcriptional abundances of potent driver genes; Right: prognostic efficacy based on the median value of the transcriptional abundances for each gene in the panel (P = 0.027), supporting the potential role of these genes in pancreatic ductal adenocarcinoma progression. High: the prognostic model score above the cut-off value is designated as high; low: the score below the cut-off value is designated as low. e Distribution of the HR of individual genes for the classification of good and poor survival in the TCGA and CTPAC cohort (n = 275 patients). Genes associated with good survival are highlighted in red, whereas genes associated with poor survival are highlighted in green. f Representative example showing a consistent prognostic trend for CXCR5, a previously reported promoting gene in our matched tissue–serum cohort (n = 41 patients) and TCGA and CTPAC cohort (n = 275 patients). High: the prognostic model score above the cut-off value is designated as high; low: the score below the cut-off value is designated as low. g Dot plot showing dependency values of our causal inference-identified genes in genome-wide Cas9 knockout screening of pancreatic ductal adenocarcinoma cell lines in the CCLE collection. h Fisher enrichment analysis of positive hits in causal inference-identified genes that showed significant dependency for pancreatic ductal adenocarcinoma cell lines (P = 0.0013, one side). Enrichment significance was assessed using one-sided Fisher’s exact test. Nominal P values are reported without correction for multiple comparisons. The HR was calculated using the Cox proportional hazards model, with HR = exp(β) and 95% CI derived from the model’s Wald test. P-values were calculated using the two-tailed log-rank test. The fitted regression line is displayed with its 95% confidence bands (shaded area) (c, d, f). Source data are provided as a Source data file. A total of 112 CNVs, 88 indels, and 14 SNVs were related to PDAC progression, and these genetic variations were related to 211 independent genes (Supplementary Data 2). Using causal inference analysis, we identified 93 serum metabolites whose abundance was collectively driven by these genetic variations and which could individually predict prognostic outcomes. Of these, 63 were further classified as subtype-specific signature metabolites (Fig. 4c). We further validated the prognostic predictive ability of serum metabolites whose levels were influenced by these causally identified genetic variants (HR 0.2443, 95% CI 0.1128–0.5292; P = 0.00015) (Fig. 4c). This analysis indicated that the effects of these genetic variations on serum metabolites consistently allowed subtyping for patients with PDAC, with significant separation between good and poor survival outcomes. Next, we analyzed the biological roles of the genetic variations identified through causal inference in PDAC development and progression, and tried to seek potential regulators that could be experimentally validated. We first evaluated the predictive efficacy of the identified genetic variations for patient survival outcomes in our tissue–serum matched cohort. A Cox proportional-hazard model built using the exact copy number of the 112 CNV-related genes allowed for a separation of patients in the matched tissue‒serum cohort into good and poor survival groups with median survival lengths of 21 and 15 months, respectively (P = 0.048; Supplementary Fig. 6a, b). Similarly, Cox proportional-hazard models using the mutational status of the 88 indel-related genes or the 14 SNV-related genes could also efficiently separate these patients with PDAC into good and poor survival groups (P = 2.21E-11 or P = 0.009, respectively, Supplementary Fig. 6c–f). We subsequently used a public TCGA and CTPAC PDAC cohort and tested whether the PDAC-promoting genes identified through causal inference analysis could also significantly predict the survival of patients with PDAC in this cohort7. Through directly using the median transcriptional abundance of each gene to generate a combined model, we validated their prognostic efficacy in the TCGA and CTPAC PDAC cohort, yielding a P-value of 0.027 (HR 0.708, 95% CI 0.5213–0.9617) (Fig. 4d). When analysing the prognostic prediction efficiency of individual genetic variations, 37 among the 211 genes showed significant discrimination (Cox model, P < 0.05) between individuals with good and poor survival (Fig. 4e). Among these, we observed a previously reported oncogene CXCR5, whose activation has been reported to promote PDAC metastasis30. Using CXCR5 mutational status in our cohort, transcriptional level in the TCGA and CTPAC cohort, as well as CXCR5 mutation inferred serum metabolites (Supplementary Table 6), consistently led to clear separation between good and poor survival (Fig. 4f). These results support the efficacy of this causal inference-based multi-omics analysis in predicting potential promoting genes that modulate the survival outcomes of patients with PDAC. To further evaluate the biological functions of these causal inference-identified genes on PDAC growth, we analysed the outcomes of their depletion on PDAC cell growth in 23 pancreatic cell lines in the CCLE gene-dependency dataset31. In total, 17,670 genes were screened in this dataset, and 979 genes with dependency value < − 0.5 or >0.5 were defined as potential PDAC-promoting genes or tumor-suppressor genes. Among the 211 genes identified by causal inference analysis, 167 genes were screened in the CCLE dependent data set, while the rest 47 genes have not been covered. Among these 167 genes, 21 genes showed significant PDAC-promoting effects on PDAC growth (red spots, Fig. 4g and Supplementary Data 3). Using Fisher’s exact test, enrichment analysis of positive hits within the causal inference-identified gene set indicated a 2.27-fold enrichment (P = 0.0013) (Fig. 4h). Collectively, these results showed that causal inference analysis allows for the identification of PDAC-promoting genes that could be independently validated using TCGA, CTPAC, and CCLE databases. In vitro and in vivo analysis identified GRPEL1 as a PDAC-promoting gene To further verify the robustness of this causal inference analysis for identifying potential promoting genes that modulate PDAC progression, the functions of the representative inferred promoting gene was experimentally validated in PDAC cell lines or patient-derived organoid (POD) models. GRPEL1, which encodes a mitochondrial protein, was prioritized for functional validation due to its highest dependency score among the 21 PDAC-promoting genes (Supplementary Data 3), consistent with the central role of mitochondria in tumorigenesis. In the matched tissue–serum cohort, we observed copy number variations (CNVs) of GRPEL1 ranging from 0 to 5 copies (Fig. 5a). Notably, reduced copy number was associated with favorable prognosis, whereas increased copy number correlated with poor outcome (HR: 2.1844, 95% CI: 1.0425–4.5772, P = 0.036) (Fig. 5b). These survival data further support the tumor-promoting role of GRPEL1 in PDAC. Knockout of GRPEL1 in PDAC cell lines (PANC-1 and SW1990) led to a suppression of cell proliferation (Fig. 5c, d). Consistently, trans-well assays revealed a impairment in both the migration and invasion capacities of GRPEL1-depleted cells (Fig. 5e, f). Moreover, PANC-1 cells with overexpression of GRPEL1 (Supplementary Fig. 7a) showed elevated cell proliferation rate (Supplementary Fig. 7b), migration, and invasion (Supplementary Fig. 7c, d). These results indicate that GRPEL1 depletion suppresses the proliferation and metastasis of pancreatic cancer cells.Fig. 5Promoting gene GRPEL1 is essential for pancreatic tumor cell proliferation and migration.a The copy number variations (CNVs) status of GRPEL1 in the matched tissue–serum cohort (n = 41 patients). b The copy number of GRPEL1 was associated with prognosis in the matched tissue–serum cohort (n = 41patients; HR: 2.1844, 95% CI: 1.0425–4.5772, P = 0.036). The HR was calculated using the Cox proportional hazards model, with HR = exp(β) and 95% CI derived from the model’s Wald test. P-values were calculated using the two-tailed log-rank test. The fitted regression line is displayed with its 95% confidence bands (shaded area). c Immunoblot of the GRPEL1 expression levels in pancreatic ductal adenocarcinoma cell lines (PANC-1 and SW1990) after GRPEL1 depletion. GAPDH was used as the loading control. Immunoblot of the GRPEL1 expression levels was repeated three times, and a representative result is shown. d Cell count determination was used to detect the viability of PANC-1 and SW1990 cells after GRPEL1 depletion; n = 5 biologically independent experiments. e, f Representative images (left panel) and quantification (right panel) of the migration and invasion of PANC-1 and SW1990 cell lines after GRPEL1 depletion; images shown are representative of three biologically independent experiments, n = 3 biologically independent experiments. g Immunoblot of KOPC-H026X organoids after GRPEL1 depletion (upper). β-Actin was used as the loading control. GRPEL1 immunoblots were repeated three times with similar results; a representative image is shown. Representative images of KOPC-H026X organoids after GRPEL1 depletion (lower), scale bar, 100 μm. h Proliferation curve (g) of KOPC-H026X organoids after GRPEL1 depletion; n = 10 biologically independent experiments. i Survival analysis of the PANC-1 BALB/c nude model. Mice met the study endpoint if the longest tumor diameter exceeded 150 mm. The significance of the results from the Kaplan–Meier survival analysis was determined using a log-rank Mantel–Cox test, n = 4 mice. j Representative images (left panel) and quantification (right panel) of immunohistochemical staining for GRPEL1 in 79 patients with pancreatic cancer and matched normal tissues. In the box plot, the upper and lower whiskers represent the maximum and minimum values, respectively. The upper and lower boundaries of the box correspond to the upper and lower quartiles, and the line within the box indicates the median. Upper images: scale bar, 100 μm; lower images: scale bar, 25 μm. Data are presented as mean ± s.e.m. Statistical analysis was performed using two-tailed Student’s t tests (e, f), one-way ANOVA with Tukey test (d, h, i), and two-tailed paired Student’s t tests (j). Source data are provided as a Source data file. Promoting gene GRPEL1 is essential for pancreatic tumor cell proliferation and migration. a The copy number variations (CNVs) status of GRPEL1 in the matched tissue–serum cohort (n = 41 patients). b The copy number of GRPEL1 was associated with prognosis in the matched tissue–serum cohort (n = 41patients; HR: 2.1844, 95% CI: 1.0425–4.5772, P = 0.036). The HR was calculated using the Cox proportional hazards model, with HR = exp(β) and 95% CI derived from the model’s Wald test. P-values were calculated using the two-tailed log-rank test. The fitted regression line is displayed with its 95% confidence bands (shaded area). c Immunoblot of the GRPEL1 expression levels in pancreatic ductal adenocarcinoma cell lines (PANC-1 and SW1990) after GRPEL1 depletion. GAPDH was used as the loading control. Immunoblot of the GRPEL1 expression levels was repeated three times, and a representative result is shown. d Cell count determination was used to detect the viability of PANC-1 and SW1990 cells after GRPEL1 depletion; n = 5 biologically independent experiments. e, f Representative images (left panel) and quantification (right panel) of the migration and invasion of PANC-1 and SW1990 cell lines after GRPEL1 depletion; images shown are representative of three biologically independent experiments, n = 3 biologically independent experiments. g Immunoblot of KOPC-H026X organoids after GRPEL1 depletion (upper). β-Actin was used as the loading control. GRPEL1 immunoblots were repeated three times with similar results; a representative image is shown. Representative images of KOPC-H026X organoids after GRPEL1 depletion (lower), scale bar, 100 μm. h Proliferation curve (g) of KOPC-H026X organoids after GRPEL1 depletion; n = 10 biologically independent experiments. i Survival analysis of the PANC-1 BALB/c nude model. Mice met the study endpoint if the longest tumor diameter exceeded 150 mm. The significance of the results from the Kaplan–Meier survival analysis was determined using a log-rank Mantel–Cox test, n = 4 mice. j Representative images (left panel) and quantification (right panel) of immunohistochemical staining for GRPEL1 in 79 patients with pancreatic cancer and matched normal tissues. In the box plot, the upper and lower whiskers represent the maximum and minimum values, respectively. The upper and lower boundaries of the box correspond to the upper and lower quartiles, and the line within the box indicates the median. Upper images: scale bar, 100 μm; lower images: scale bar, 25 μm. Data are presented as mean ± s.e.m. Statistical analysis was performed using two-tailed Student’s t tests (e, f), one-way ANOVA with Tukey test (d, h, i), and two-tailed paired Student’s t tests (j). Source data are provided as a Source data file. We then validated the functions of GRPEL1 using an ex vivo organoid system. The PDAC patient-derived organoid strains KOPC-H026X and KOPC-H027X were generated from two patients with PDAC. Using the Cas9-mediated knockout system, GRPEL1 was successfully depleted in PDAC organoid (Fig. 5g). Likewise, GRPEL1 deletion led to a consistent reduction in cell proliferation, although to different extents, in two PDAC tumor-derived organoids from two different patients (Fig. 5g, h, and Supplementary Fig. 8a, b). Given that GRPEL1 was prioritized for functional investigation based on its mitochondrial localization, we first examined the metabolic consequences of its manipulation. Depletion of GRPEL1 in PANC-1 cells resulted in significantly impaired endogenous mitochondrial respiration and oxidative phosphorylation (OXPHOS) capacity (Supplementary Fig. 9a). Given that OXPHOS functions as a central regulatory machinery for core carbon metabolism32, these defects were accompanied by substantial alterations in glycolysis10 and pentose phosphate pathway33,34 following GRPEL1 deletion (Supplementary Fig. 9b, c), suggesting that GRPEL1-driven metabolic remodeling contributes to PDAC pathogenesis. To further validate the tumor-promoting role of GRPEL1 in vivo, we established cell line-derived xenograft (CDX) models using GRPEL1-knockout, GRPEL1-overexpressing, and control PANC-1 cells. Tumors derived from GRPEL1-depleted cells exhibited suppressed growth, while those from GRPEL1-overexpressing cells showed enhanced tumor formation (Supplementary Fig. 7e–g). Consistently, mice bearing GRPEL1-depleted tumors demonstrated prolonged survival compared to controls (Fig. 5i). Immunohistochemical analysis of a PDAC tissue microarray (79 cases) further revealed that GRPEL1 expression was elevated in tumor tissues relative to adjacent normal tissues (Fig. 5j). Moreover, metabolomic profiling revealed substantial overlap between the serum metabolic signatures of patients with high GRPEL1 copy number (≥2 copies vs. ≤1 copy) and those of GRPEL1-overexpressing PANC-1 cells: 88 of 238 altered metabolites and 8 of 17 enriched metabolic pathways were concordant (Supplementary Fig. 10 and Supplementary Data 4). This convergence provides further evidence for a causal link between GRPEL1 genetic alterations and systemic metabolic features. Collectively, these in vitro and in vivo findings establish GRPEL1 as a critical promoter of pancreatic cancer progression. GRPEL1 showed cross-talk with oncogenic downstream events of the mTOR pathway GRPEL1 has been previously identified to encode a conserved nucleotide-exchange factor that facilitates the exchange of ADP with ATP and plays vital roles in maintaining mitochondrial metabolic functions35,36. In this study, we identified GRPEL1 that promotes PDAC progression. To further investigate the underlying molecular mechanism of its PDAC-promoting functions, we profiled the serum metabolites that were correlated with mutations in this gene within the matched tissue–serum cohort. In total, 13 metabolites were significantly correlated with GRPEL1 mutations, and these metabolites also showed efficient prognostic power, indicating that GRPEL1 may regulate PDAC progression via metabolic regulation (Fig. 6a, b). We further analysed genes showing MUMS with significant enrichment of correlated metabolites within GRPEL1-related metabolic profiles (Fig. 6a and Supplementary Data 5). GO analysis of these genes showing MUMS revealed enrichment of pathways including the inhibition of TSC complex formation by PKB and the PI3K-Akt signaling pathway (Fig. 6a, highlighted in green), indicating that GRPEL1 may share a common downstream metabolic signature with mTOR/PI3K/Akt signaling (Fig. 6c).Fig. 6GRPEL1 showing cross-talk with oncogenic downstream events of the mTOR pathway.a Sankey plot showing GRPEL1 significantly correlated with serum metabolites, and genes showing that MUMS were also associated with these metabolites (genes involved in mTOR/PI3K-ATK signaling pathways are highlighted in green). MUMS mutations upstream of the metabolomic signature. b Cox curve showing GRPEL1 correlated serum metabolites could efficiently discriminate between good and poor survival (HR: 2.0981, 95% CI: 1.0136–4.3433, P = 0.044). The HR was calculated using the Cox proportional hazards model, with HR = exp(β) and 95% CI derived from the model’s Wald test. P-values were calculated using the two-tailed log-rank test. The fitted regression line is displayed with its 95% confidence bands (shaded area). HR hazard ratio. High: the prognostic model score above the cut-off value is designated as high; low: the score below the cut-off value is designated as low. c Gene ontology enrichment analysis of genes showing MUMS related to GRPEL1, showing enrichment of the mTOR and PI3K-AKT pathway genes. Metabolite set enrichment analysis (MSEA) was performed using MetaboAnalyst 6.0. For quantitative data, enrichment was assessed using the global test, while binary query lists were analyzed with two-tailed Fisher’s exact tests. P values were adjusted for multiple testing using the Benjamini–Hochberg false discovery rate (FDR) method, and an adjusted P < 0.05 was considered statistically significant. MUMS mutations upstream of the metabolomic signature. d Workflow of PANC-1 cell line-based analysis of rapamycin inhibitory effect on control and GRPEL1-depleted/overexpressed cells. Created in BioRender. Lou, X. (2026) https://BioRender.com/87lhrfz. e IC50 values of rapamycin on control (black line) and GRPEL1-depleted PANC-1 cell lines (red line). Data are from n = 3 biologically independent experiments. f IC50 values of rapamycin in control (black line) and GRPEL1-overexpressed PANC-1 cell lines (red line). Data are from n = 3 biologically independent experiments. g Dot plot showing significant negative correlation of GRPEL1 expression level and rapamycin responses (IC50, µM) in pancreatic cancer cell lines using the GDSC database. Pearson correlation coefficients are reported with their associated P values, which were derived from two-tailed Student’s t tests. The regression coefficient (β) was estimated using a generalized linear model (GLM). The regression line is shown with its 95% confidence bands (shaded area). GDSC genomics of drug sensitivity in cancer. h Schema showing the process through which MUMS analysis indicated downstream cross-talk of promoting mutations and corresponding clinical application. Data are presented as mean ± s.e.m. Statistical analysis was performed using two-tailed Student’s t tests (e, f). Source data are provided as a Source data file. GRPEL1 showing cross-talk with oncogenic downstream events of the mTOR pathway. a Sankey plot showing GRPEL1 significantly correlated with serum metabolites, and genes showing that MUMS were also associated with these metabolites (genes involved in mTOR/PI3K-ATK signaling pathways are highlighted in green). MUMS mutations upstream of the metabolomic signature. b Cox curve showing GRPEL1 correlated serum metabolites could efficiently discriminate between good and poor survival (HR: 2.0981, 95% CI: 1.0136–4.3433, P = 0.044). The HR was calculated using the Cox proportional hazards model, with HR = exp(β) and 95% CI derived from the model’s Wald test. P-values were calculated using the two-tailed log-rank test. The fitted regression line is displayed with its 95% confidence bands (shaded area). HR hazard ratio. High: the prognostic model score above the cut-off value is designated as high; low: the score below the cut-off value is designated as low. c Gene ontology enrichment analysis of genes showing MUMS related to GRPEL1, showing enrichment of the mTOR and PI3K-AKT pathway genes. Metabolite set enrichment analysis (MSEA) was performed using MetaboAnalyst 6.0. For quantitative data, enrichment was assessed using the global test, while binary query lists were analyzed with two-tailed Fisher’s exact tests. P values were adjusted for multiple testing using the Benjamini–Hochberg false discovery rate (FDR) method, and an adjusted P < 0.05 was considered statistically significant. MUMS mutations upstream of the metabolomic signature. d Workflow of PANC-1 cell line-based analysis of rapamycin inhibitory effect on control and GRPEL1-depleted/overexpressed cells. Created in BioRender. Lou, X. (2026) https://BioRender.com/87lhrfz. e IC50 values of rapamycin on control (black line) and GRPEL1-depleted PANC-1 cell lines (red line). Data are from n = 3 biologically independent experiments. f IC50 values of rapamycin in control (black line) and GRPEL1-overexpressed PANC-1 cell lines (red line). Data are from n = 3 biologically independent experiments. g Dot plot showing significant negative correlation of GRPEL1 expression level and rapamycin responses (IC50, µM) in pancreatic cancer cell lines using the GDSC database. Pearson correlation coefficients are reported with their associated P values, which were derived from two-tailed Student’s t tests. The regression coefficient (β) was estimated using a generalized linear model (GLM). The regression line is shown with its 95% confidence bands (shaded area). GDSC genomics of drug sensitivity in cancer. h Schema showing the process through which MUMS analysis indicated downstream cross-talk of promoting mutations and corresponding clinical application. Data are presented as mean ± s.e.m. Statistical analysis was performed using two-tailed Student’s t tests (e, f). Source data are provided as a Source data file. The mTOR pathway has been identified as a master regulator of cell growth and metabolism that is commonly activated in tumors and controls cancer cell metabolism37–40. Blocking this pathway with the mTOR inhibitor rapamycin arrests the proliferation of PDAC cell lines in genetically engineered PDAC mouse models41–43. To further characterize the biological relevance of GRPEL1 and mTOR signaling, we performed an epistatic analysis by calculating the growth-inhibiting effect of rapamycin on PDAC cell lines with GRPEL1 depletion or overexpression (Fig. 6d). Within a concentration range of 10–640 nM, the IC50 of rapamycin in control PANC-1 cells was 222 nM, while it increased to 566 nM in GRPEL1-depleted PANC-1 cells, indicating that GRPEL1 loss caused a resistance response to mTOR inhibition (Fig. 6e). On the basis of these findings, we hypothesized that GRPEL1 may perform oncogenic-specific functions downstream of the mTOR pathway or be collaboratively required for tumor growth driven by mTOR downstream signaling. In contrast, GRPEL1 overexpression sensitized the PANC-1 cells to rapamycin, causing proliferation arrest and reducing the IC50 to 120 nM (Fig. 6f). This may be explained by increased nutrient and energy demand to meet the needs of GRPEL1 overexpression-induced rapid proliferation, which made these cells more sensitive to mTOR inhibition and caused a decrease in nutrient support. As shown in Fig. 5j, GRPEL1 expression was significantly upregulated in PDAC tumor tissues, which increased the possibility that patients with GRPEL1high would be more sensitive to mTOR inhibition. This hypothesis was preliminary supported, since we observed a significant negative correlation between GRPEL1 expression level and rapamycin responses (via the IC50 (μM) value) for PDAC cell lines within the GDSC database (Fig. 6g). Furthermore, patient-derived organoids lacking GRPEL1 exhibited resistance to mTOR inhibition-induced suppression of clone formation (Supplementary Fig. 8c, d). This finding, coupled with the known synergy between mTOR inhibitors and chemotherapy44–47, highlights GRPEL1 as a potential target to potentiate tumor-specific mTOR inhibition-based therapies (Fig. 6h). Discussion In this study, using our MUMS analysis in a PDAC cohort, we observed that genetic mutational diversity contributes to prognostically distinct metabolic subtypes and, in turn, MUMS for such prognostic-related serum metabolomic signatures could reflect cross-talk of pathways downstream of individuals’ driver mutations. Moreover, an enrichment of the mTOR/PI3K-Akt pathways was identified among MUMS of identified PDAC-promoting gene GRPEL1, and functional validation supported that GRPEL1 sensitized PDAC cells to mTOR inhibition-induced proliferation arrest (Fig. 6h). Our findings indicated that the MUMS-based downstream cross-talk analysis provides an approach for identifying neo-targets for targeted therapy or for providing rational-based combined therapy design, as well as expanding applicable patients for current therapeutic approaches for PDAC treatment. In addition to genomic research, we integrated genomic and metabolic data and carried out a three-dataset layer-based causal inference analysis. In this way, we developed a PDAC prognostic model, and the prediction ability of our model was better than the driver gene-based prediction method. Previous genomic studies showed that mutations in KRAS, TP53, CDKN2A, and SMAD4 conferred bad prognostic outcome and that a greater number of altered driver genes conferred a worse OS8,48, whereas results from another study showed no prognostic prediction ability based on the expression of these driver genes49. A limitation of genomic studies is that they evaluated only genomic alterations and did not account for the downstream effects of gene products, such as metabolites, which could account for the differences between driver gene status and driver gene expression in the association of survival outcomes. In our study, we also tested the role of the driver genes in predicting the outcomes of patients with PDAC and could not obtain a significant prognostic separation based on driver genes. However, a prognostic separation was achieved using our model based on serum metabolism driven by tumor genomic variations. Metabolic changes are the direct reflection of gene expression, which accounts for not only the genomic alterations but also the downstream effects of gene products. Notably, metabolites can also act as ligands to activate or inhibit multiple signaling pathways50. In the nine-metabolite panel established in this study, we found that LysoPCs (18:2, 18:1, and 14:0) are largely positively associated with survival (Supplementary Table 4). The protective role of LysoPC (18:2) is supported by Denise Wolrab and colleagues51 and was further validated as a protective lipid for PDAC occurrence in two independent cohorts52. Similarly, LysoPC (18:1) and LysoPC (14:0) have also been identified as protective factors against PDAC development53. In contrast, the remaining six metabolites in our panel have not been previously linked to PDAC. Although androsterone sulfate, alpha-CEHC, and glycodeoxycholic acid sulfate remain underexplored, lysophosphatidylcholines (LysoPCs) and lysophosphatidylethanolamines (LysoPEs) are among the most extensively studied metabolites in both physiological and pathological contexts. Beyond serving as structural components of cell membranes54, LysoPCs act as ligands for multiple G protein-coupled receptors55, which are known to activate MAPK and PI3K/AKT signaling pathways in PDAC56,57. These observations raise important questions regarding whether, how, and in which structural contexts the LysoPCs and LysoPEs included in our nine-metabolite panel contribute to PDAC pathogenesis. It is worth noting that although multiple clinical variables were adjusted for and two independent validation cohorts were employed to minimize environmental confounders, further validation in geographically distinct populations and mechanistic dissection of this metabolite panel remain warranted. Moreover, additional studies investigating the prognostic specificity of this nine-metabolite panel may help elucidate the underlying biological mechanisms linking these metabolites to PDAC progression. Another important study finding was that metabolites integrate the functions of multiple driving mutations, and MUMS in genes sharing common downstream metabolic signatures may be functionally relevant. This facilitated the identification of MUMS in gene sets relevant to the metabolic signatures of targeted therapy genetic mutations and of testing as to whether these MUMS could independently respond to the relevant therapy. Similarly, analyses of MUMS showed that the newly identified PDAC-promoting gene GRPEL1 exhibits a common regulatory relationship with the mTOR pathway, and experimental validation also showed that depletion of GRPEL1 ended mTOR inhibition-induced growth arrest in PDAC cell lines. As a generally required signaling pathway for cell growth and metabolism, the mTOR pathway is commonly activated in tumors, and non-selective blocking of this pathway to facilitate cancer treatment could contribute to serious side-effects37–40. Our findings showed that GRPEL1 may be a cancer-specific target for mTOR pathway inhibition-dependent cancer treatment. In line with the relationships of genes showing MUMS with targeted therapy responses for the identification of treatment targets, these synergistic genes correlating with serum metabolite profiles may also be the most downstream molecular signatures that reflect favorable responses to such targeted treatments. Further evaluation of the potential of these downstream metabolite signatures as noninvasive biomarkers for predicting targeted therapy responses is worthwhile. It should be noted that our analyses were limited to the exomes of PDAC. Other types of genetic changes, such as chromosomal rearrangements, translocations, large deletions and insertions, inversions, chromothripsis, and intronic alterations, would have been missed by our approach. We therefore cannot exclude the possibility that one of these genetic changes may affect the prognostic efficiency of our model, which was built based on whole‒exome sequencing. In this study, we used whole‒exome sequencing as an example to demonstrate a proof-of-concept approach to identify noninvasive biomarkers based on the establishment of causal inference links between tumor genetic mutations and serum metabolites. In the future, a whole-genome-sequencing-based study may overcome this limitation and supply more exact noninvasive biomarkers based on the alteration of serum metabolism driven by tumor genomic variations. In summary, we introduced the MUMS strategy, which tightly correlates tumor gene mutations with serum metabolomic signatures. The use of MUMS facilitates the prediction of prognostic outcomes in PDAC and the identification of rapamycin for treating PDAC with functional GRPEL1, both of which highlight the clinical applications of MUMS. Methods Our research complies with all relevant ethical regulations. This study was approved by the Independent Ethics Committee of the National Cancer Center/Cancer Hospital, CICAMS (Approval No. 18-218/1796), and the Medical Ethics Committee of Tongji Medical Hospital (Approval No. TJDBPS02). All animal experiments were performed in accordance with protocols approved by the Laboratory Animal Ethics Committee of the National Cancer Center/Cancer Hospital, CICAMS (Approval No. NCC2023A198). Patient recruitment, study cohorts, and sample preparation From 2017 to 2022, consecutive samples were collected from patients with PDAC at two independent centers, namely, Cancer Hospital, Chinese Academy of Medical Sciences (CICAMS), and Peking Union Medical College Hospital. From 2022 to 2024, consecutive samples were collected from patients with PDAC at Affiliated Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology. A final diagnosis of pancreatic cancer was made by a pathologist based on histological examination (either on a resection specimen or using pancreatic fine-needle aspiration), and patients without PDAC were excluded. Those who had received radiation or chemotherapy or had a history of other cancers were also excluded. PDAC was staged according to the American Joint Committee on Cancer and the International Union for Cancer Control tumour size, node, and metastasis (TNM) staging system. All patients with PDAC enrolled in this study were followed for at least 24 months to obtain the OS. Baseline characteristics and clinicopathological features, including the OS, for all study patients with PDAC were recorded (listed in Supplementary Table 1). Blood samples were collected prior to cancer treatment. Briefly, pre-treatment blood samples were obtained after overnight fasting using evacuated blood collection tubes (BD, NJ, America) without anticoagulant. Whole blood was incubated at room temperature for 60 min, followed by centrifugation at 1500 × g for 15 min (4–8 °C) to obtain serum. The serum-containing supernatant was mixed in a new centrifugation tube through inversion, split into 0.5-mL aliquots, and stored at −80 °C in 2 mL screw-cap tubes. The samples were then shipped on dry ice to an analytical laboratory. In total, 350 patients with PDAC were enrolled in our study. Of these, 41 were classified into a matched tissue–serum cohort, with matched tumoural and paratumoural tissue samples (at least 4 cm from the tumoural boundary) collected during surgery. Formalin-fixed paraffin-embedded (FFPE) samples were prepared and sliced into 10-mm-thick sections. Tissue and serum samples derived from this cohort were used for WES and untargeted metabolomic analysis, respectively. Pseudo-targeted metabolomic detection of PDAC prognostic biomarkers was conducted using serum samples from 309 patients with PDAC with available OS data. These patients were divided into a modeling cohort (n = 151 patients), an internal validation cohort (n = 79 patients), and a separate external validation cohort (n = 79 patients). All patients provided written informed consent and were not compensated for their participation. Sex was defined based on biological attributes and aligned with self-reported data. However, sex-based analysis was not conducted due to a lack of evidence suggesting that prognostic outcomes of patients with PDAC are influenced by sex or gender. This multi-center design enhances the generalizability of the findings compared to single-center studies; several potential biases should be considered when interpreting the results. First, as all three centers are tertiary referral hospitals specializing in pancreatic diseases, there is a potential referral bias. The cohort may overrepresent patients with complex conditions or those with better performance status seeking advanced treatments, which could limit the applicability of our findings to patients treated in primary or secondary care settings. Second, the relatively long recruitment period introduces a potential temporal bias. Although we excluded patients with prior treatment, subtle shifts in perioperative management or supportive care over the years could have influenced long-term outcomes. This is an inherent limitation of longitudinal cohort studies. Third, the selection of the surgical sub-cohort (n = 41 patients) for tissue analysis introduces a selection bias towards patients with resectable, typically early-stage disease. Therefore, the genomic features identified by WES are primarily representative of the tumor biology in surgically resectable PDAC and may not reflect the mutational landscape of advanced or metastatic disease. Despite these potential biases, the inclusion of a separate external validation cohort helps to mitigate these limitations and ensure the robustness of the prognostic model. Quality control (QC) sample preparation Equal volumes of serum derived from each patient in the matched tissue–serum cohort were pooled to obtain a QC sample. Each batch contained three QC samples. The accuracy and precision of the semi-quantitative untargeted metabolomic profiling were estimated as described in the “Data analysis” section. Untargeted metabolite profiling Untargeted metabolite profiling was performed on 309 serum samples from the modeling cohort (n = 151 patients), an internal validation cohort (n = 79 patients), and a separate external validation cohort (n = 79 patients). Each sample was analyzed once, with no technical replicates. Metabolites were extracted from 60 μL of previously thawed serum by adding 240 μL of acetonitrile:isopropanol (3:1 v/v, Thermo Fisher), followed by 60 μL of ammonium formate (0.5 g/mL, Thermo Fisher) and 6 μL of an internal standard solution containing 100 µg/mL L-tyrosine-(phenyl-3,5-d2) (Sigma Aldrich), 10 µg/mL 13C-cholic acid (Cambridge Isotope Laboratories), and 60 µg/mL doxercalciferol (MedChemExpress) to precipitate serum proteins. After vortexing for 4 min and centrifugation at 17,949 × g for 5 min, 200 μL of the supernatant was transferred to a new tube and dried using a Centrivap cold trap centrifuge at −60 °C. The dried metabolite extracts were then reconstituted in 75 μL of 55% methanol (Thermo Fisher) containing 0.1% formic acid (Thermo Fisher). The metabolic extracts were analysed using reversed-phase liquid chromatography-mass spectrometry (MS) in both positive- and negative-ionization modes. Five microliters of the sample were injected into the Q Exactive mass spectrometer (Thermo Fisher Scientific) coupled to an UltiMate 3000 UPLC system (Thermo Fisher Scientific) equipped with a CORTECS (Waters) C18 column (2.1 × 100 mm, 1.6 μm). The column temperature was maintained at 35 °C with a flow rate of 0.3 mL/min. The mobile phases consisted of (A) acetonitrile containing 0.1% formic acid (FA) and (B) water containing 0.1% FA (Merck Millipore). The following gradient elution program was applied: 5% A (initial), increased linearly to 45% A from 0.5 to 14 min, to 75% A at 32 min, to 80% A at 42 min, and then to 100% A from 50 to 55 min, followed by a return to 5% A and holding for 5 min for column re-equilibration. Data were acquired within the mass/charge ratio (m/z) range of 130–1200 Da at a resolution of 700,000 in the full MS scan mode. Metabolomic data pre-processing Raw data for untargeted metabolomics were acquired using Thermo Xcalibur (Ver 3.0). Metabolomic feature extraction, alignment, and quantification were performed using the R package XC‒MS software (Ver 3.6.1). To acquire reliable peaks, we retained values up to one decimal place for retention times and three decimal places for m/z data to aggregate the metabolites. To filter out background signals, metabolites with an abundance <50,000 in all patients or zero abundance in >85% of patients were filtered out. To eliminate batch-to-batch differences, an R preprocess Core software package (Ver 1.47.1) was used for robust multiarray average (RMA) normalization, and the abundance ratio of metabolites was calculated. The coefficient of variation (CV) for each metabolic feature was calculated based on its abundance in QC samples from each batch. Metabolites with a CV > 30% were also excluded. Metabolite annotation and inference MS/MS spectra of the QC samples were acquired at different fragmentation energies (25 and 50 NCE) for the highest 10 parent ions. Metabolite annotation was performed as previously described, with some modifications58. Briefly, the QC MS1/MS2 spectra were imported into the MS-DIAL 4.24 program and searched against public databases, including the Human Metabolome Database (HMDB, https://hmdb.ca/), and MassBank (https://massbank.eu/MassBank/). Finally, a manual check of the reference database was performed to confirm and distinguish between similar readouts. For metabolites with MS/MS data that could not be reliably acquired, the exact m/z values were checked in relation to the Metlin, HMDB, and BioML databases to determine their potential identities, and the cutoff was set to <10 ppm59. The confidence levels for metabolite annotation were as follows: the highest confidence identification, a validated identification (Level 1), confirms a structure with a minimum of two orthogonal properties from a pure reference standard acquired under identical analytical conditions. A lack of reference standard acquisition but predictive or externally acquired structure evidence, namely MS/MS data, exhibiting diagnostic fragments or neutral losses consistent with a specific structure, would be considered a putative identification (Level 2). Preliminary identifications (Level 3) arise when accurate mass and isotopic distribution patterns produce tentative structures from database searches. Molecular formula candidates (Level 4) and a deconvoluted experimental m/z (Level 5) complete the less confident annotation classifications60. Pseudo-targeted metabolite profiling Pseudo-targeted metabolite profiling was performed once on each of the 309 serum samples from the matched tissue-serum cohort, with no technical replicates. For metabolite extraction in pseudo-targeted metabolomic detection, 10 μL of internal standard solution (5 μg/mL 13C-cholic acid) was added to 80 μL of serum with 150 μL of acetonitrile:isopropanol (4:1 by volume, Thermo Fisher) and 50 μL of ammonium formate (0.5 g/mL), and the mixture was vortexed and centrifuged at 17,949 × g for 5 min. Then 60 μL of supernatant was diluted with 150 μL of HPLC-grade water before use. We developed a pseudo-targeted method based on pure standards, similar to previously reported studies58,61, and determined the relative levels of all metabolites in the identified panel using the same reference pool sample to normalize the abundances for each patient. Pseudo-targeted metabolomic detection was conducted with an AB SCIEX Triple Quad™ 4500 system and run in separate ion modes (positive and negative). The mobile phase and column used for reversed-phase liquid chromatography were the same as those used for untargeted metabolite profiling. The column temperature was set at 45 °C; mobile phase A: 0.1% formic acid-water; mobile phase B: 0.1% formic acid-acetonitrile. The injection volume was 15 μL for each mode. The metabolites were eluted from the column at a flow rate of 0.3 mL/min with gradually increasing concentrations of mobile phase B, starting from 12% mobile phase B and reaching 60% mobile phase B after 2.5 min. Linear 60–85% and 85–100% phase B gradients were set at 6 min and 8.5 min, respectively. Pooled C samples were used as QC samples for targeted analysis. The delustering potentials and collision energies were optimized using QC samples from the control group. For targeted metabolomics, data acquisition was performed using Sciex Analyst software 1.6.3. The metabolite peaks were integrated using Sciex Analyst 1.6.3 software. WES and mutation analysis Library preparation and sequencing DNA was extracted from FFPE slides of tumoural and paratumoural tissue samples using a QIAamp DNA FFPE Tissue Kit (QIAGEN), in accordance with the manufacturer’s recommendations. All 41 pairs of DNA samples passed QC and were sheared into 180–250-bp fragments for library construction. After end repair, 3′A tailing, adapter ligation, and polymerase chain reaction amplification, the prepared gDNA libraries were subjected to an exome capture process using the SureSelectXT Human All Exon V6 (Agilent) system, in accordance with the manufacturer's recommendations. The final exome libraries were quality-checked and sequenced on Illumina NovaSeq6000 instruments to yield 150-bp paired-end reads and achieve a target coverage of 200×. Sequencing data processing and mutation analysis After cutting adapters and filtering low-quality reads, raw sequencing data were aligned to the hg38 human reference genome using the Burrows–Wheeler aligner, and the mapping rate, duplication rate, capture ratio, and coverage depth of the target region were calculated62. Subsequently, Mutect2 software was used to call somatic SNVs/indels between paired tumoural and paratumoural tissues, and these mutant sites were annotated using the Enliven® system. Somatic CNVs were detected between paired tumoural and paratumoural tissues using control-FREEC, and annotated using ANNOVAR. GATK was used to call germline mutations (SNVs/indels), and ANNOVAR was used for annotation63. Somatic variants (indels and SNVs) located in exons, 5′/3′-UTRs, and splicing sites causing frame-shift deletion/insertion, in-frame deletion/insertion, mis-sense, nonsense, nonstop, or splice site/region mutations were used for subsequent analysis, whereas synonymous mutations were filtered. Furthermore, we required each somatic variant to have a variant allele frequency lower than 2% in patient-matched normal tissue, as well as in all other normal tissues. In addition, all somatic CNVs were included. Correlation analysis between known genetic variants and serum metabolites Genes that were significantly mutated and repeatedly discovered in previous large PDAC sequencing studies were used to determine their effects on the serum metabolome using a Spearman association analysis. The associations of three types of genetic variants (SNVs, indels, and CNVs) with serum metabolites were calculated. The cutoff value for a significant association was set at P < 0.001. Identification of PDAC-promoting genetic variations and genetic variations affecting serum metabolites using causal inference analysis Based on genetic variants, serum metabolites, and survival time within the same patient cohort, we sought to determine the causal links showing that certain genetic variants lead to altered serum metabolites and thus predict the survival time of patients with PDAC. To achieve this, we applied a causal inference analysis in this study, optimized from the previously reported likelihood-based causality model selection (LCMS) test29. This algorism uses a multi-omics dataset (at least 3 different types of datasets) derived from the same cohort of subjects, and uses conditional correlation measures to determine which relationship among different datasets is best supported by the data, which is achieved by selecting the model with the smallest Akaike Information Criterion (AIC) value. Using this approach, we identified potential novel promoting mutations that affect PDAC progression. Three types of genetic variants (SNVs, indels, and CNVs) were used, and the variants that showed causal links were identified. To further validate the roles of these variants and their related genes in PDAC prognostic outcomes, a Cox model was established to evaluate their efficiency in distinguishing between short- and long-surviving patients. Additionally, the prognostic value of these variant-related genes was examined using PDAC variants and transcriptome databases from the TCGA and CTPAC PDAC dataset. Selection of metabolites for the PDAC prognosis prediction model We selected metabolite features that were significantly associated with key driver mutations or were involved in novel promoting gene-induced metabolites. We implemented the LASSO algorithm with 10-fold cross-validation for feature selection from these metabolites. Logistic regression was used to create a prognostic model for patients with PDAC based on an untargeted metabolome. Chemical structure annotation, including MS2 ion pairs, was established using MS/MS spectrum matching, as described previously64. Subsequently, the metabolites involved in the prognostic model were transferred to the targeted metabolic platform (4500) in MRM mode, based on their specific precursor ion and product ion pairs. Their prediction efficiency was first evaluated using targeted LC-MS with the modeling cohort to establish a prognostic model, determine the cutoff value, and provide further validation in terms of independent validation in patients. Validation of novel PDAC-promoting gene function using CRISPR/Cas9 knockout in pancreatic cancer cell lines Cell culture SW1990 (Cat No. 3101HUMSCSP5302) and PANC1 (Cat No. 1101HUM-PUMC000023) cells were purchased from the National Infrastructure of Cell Line Resources in China (Beijing, China). These cells were cultured in Dulbecco’s modified Eagle medium supplemented with 10% foetal bovine serum (FBS; Gibco, USA), penicillin, and streptomycin (Gibco). All cells were incubated at 37 °C with 5% CO2. Plasmids and stable cell line construction The lentiCRISPRv2 plasmid was used to clone guide oligos targeting the GRPEL1 genes65. PANC1 and SW1990 cells were plated and transfected overnight with lentiCRISPRv2 constructs. Two days post-transduction, the cells were subjected to 1‒2 weeks of selection in 1 μg/mL puromycin (Solarbio). Knockout efficiency was validated using immunoblotting. Sequences of the gene-specific sgRNAs are listed in Supplementary Table 7. Antibodies and immunoblotting For immunoblotting analysis, proteins were extracted from cells or tissues using radioimmunoprecipitation assay lysis buffer (Beyotime) supplemented with protease inhibitors (Beyotime). Equal volumes of lysates were loaded and separated on 10% sodium dodecyl sulfate-polyacrylamide gel electrophoresis gels and blotted on polyvinylidene difluoride membranes (Millipore). After being blocked with 5% skim milk for 2 h at room temperature, the membranes were incubated with primary antibodies overnight at 4 °C. The primary antibodies used in the western blot analysis included anti-GRPEL1 (1:1000, Cat No.12720-1-AP, Proteintech), anti-ACTIN (1:20,000, Cat No. 66009-1-Ig, Proteintech), and anti-GAPDH (1:20,000, Cat No. 60004-1-Ig, Proteintech). The membranes were then incubated with anti-rabbit horseradish peroxidase (HRP; 1:5000; Cat No. #7074, Cell Signalling Technology) and anti-mouse HRP (1:5000; Cat No. #7076, Cell Signalling Technology) as secondary antibodies at room temperature for 1 h. The blots were detected using a chemiluminescence kit (34577; Thermo Fisher Scientific, USA) and imaged using the MiniChemi 610 system (Sage Creation Science, Co., Ltd.). Proliferation assay Approximately 1.5 × 104 cells were seeded in triplicate into 12-well plates and cultured for the indicated times until they were ready for counting. The cells were then collected for cell counting using a Neubauer-improved counting chamber (BR717810; Sigma-Aldrich). Migration and invasion assays For the invasion assays, cell‒culture inserts (0.8 μm, Corning, 3422) were coated with 1:20 diluted Matrigel (Corning, 356234) for 1.5 h and rehydrated with Opti-MEM (GIBCO, USA) for 0.5 h. For the migration assays, the inserts were not coated. Cells (4 × 105) in 0.2 mL of serum-free medium were placed in the upper chamber, and the lower chamber was loaded with 0.8 mL of medium containing 20% FBS. After incubating the plates for 20 h at 37 °C, the cells remaining on the upper side of the membrane were gently wiped off, and the cells that migrated to the lower side of the membrane were fixed with 4% methanol for 15 min and stained with 0.1% crystal violet. Images of the stained cells in five random fields were captured using an optical microscope (Olympus, Tokyo, Japan), and the cells were counted. Measurement of mitochondrial respiration Endogenous mitochondrial respiration was measured using the Oxygraph-2k system (Oroboros, Innsbruck, Austria). Briefly, 2 × 10⁶ cells were added to an electrode chamber containing 2.5 mL of 1× TD buffer (25 mM Tris-base, 137 mM NaCl, 10 mM KCl, 0.7 mM NaH₂PO₄, pH 7.4). After the signal stabilized, baseline respiration was recorded. Subsequently, 0.1 μM oligomycin (Sigma-Aldrich), an ATP synthase inhibitor, was injected to inhibit Complex V, thereby reducing oxygen consumption rate by impairing electron flow through the electron transport chain. Next, 0.1 μM carbonyl cyanide 4-(trifluoromethoxy) phenylhydrazone (FCCP, Sigma-Aldrich), a mitochondrial uncoupler, was added to assess non-phosphorylating proton leak and maximal mitochondrial respiration capacity. Untargeted metabolite analysis of the top three contributing genes following shRNA knockdown in pancreatic cancer cell lines Plasmids and stable cell line construction pLKO.1-based shRNA plasmids targeting FAT4, LATS2, MEK2, and MYD88, as well as the pLKO.1-control plasmid were purchased from SYNBIO TECHNOLOGIES (Suzhou, China). For lentivirus production, 293T cells were transfected with the respective pLKO.1-shRNA plasmids, after which the lentivirus-containing supernatant was harvested. PANC-1 cells were seeded and transduced with the lentiviral particles overnight. Two days post-transduction, the cells were selected with 1 μg/mL puromycin (Solarbio) for 1–2 weeks. Knockdown efficiency was validated by quantitative RT-PCR. Primer sequences are listed in Supplementary Table 7. K-means clustering analysis Metabolomic data underwent comprehensive preprocessing to ensure analytical robustness. Raw abundance data were subjected to log2 (x + 1) transformation to reduce skewness and stabilize variance, followed by Z-score normalization (centering and scaling) to eliminate magnitude differences. Low-variance metabolites in the bottom 25th percentile were filtered out to minimize noise interference. Unsupervised clustering was performed using the factoextra R package (v1.0.7) with k = 3 clusters, initialized with 25 random starts to ensure solution stability. For pattern visualization, the 50 most discriminative metabolites were selected based on ANOVA F-statistics (lowest P-values). Heatmaps were generated using the pheatmap R package (v1.0.13). OPLS-DA modeling Orthogonal projections to latent structures-discriminant analysis (OPLS-DA) was employed to maximize class separation and identify metabolite biomarkers. The model was constructed using preprocessed metabolomic data with unit variance scaling. Model quality was assessed through 7-fold cross-validation, yielding R2Y (goodness-of-fit) and Q2 (predictive ability) metrics. Permutation testing (200 iterations) was conducted to evaluate overfitting risk. Metabolites with variable importance in projection (VIP) scores >1.5 were considered significant contributors to class separation. The OPLS-DA model was validated using both cross-validation ANOVA (CV-ANOVA) and permutation testing (P < 0.05 indicating statistical significance). All analyses were performed using the ropls R package (v1.38.0). Differential analysis Metabolites with P < 0.05 were marked as differentially expressed metabolites between sample groups. Pathway enrichment analysis was performed using MetaboAnalyst 5.0 (www.metaboanalyst.ca/MetaboAnalyst/). Rapamycin treatment and IC50 calculation in PDAC cell lines For the CCK-8 assay, cells were seeded in 96-well plates at a density of 8000 cells per well. After cell adhesion, the medium was replaced with fresh complete medium containing 10% CCK-8 reagent (DOJINDO, Japan), and the cells were incubated at 37 °C with 5% CO₂ for 2 h. Absorbance at 450 nm was measured daily for 5 days using a microplate reader (Thermo Fisher Scientific). To evaluate the suppressive effect of rapamycin on cell proliferation, GRPEL1-overexpressing and GRPEL1-knockout PANC-1 cells were treated with rapamycin at concentrations of 10, 20, 40, 80, 160, 320, and 640 nM. IC50 values were subsequently calculated based on the resulting dose–response curves. Validation of PDAC-promoting gene function using a PDAC organoid-based ex vivo system Based on background genetic mutations and transcriptome profiles, organoids derived from patients diagnosed with PDAC were selected for further genetic manipulation to validate the function of candidate genes. The PDAC patient‑derived organoid strains KOPC‑H026X and KOPC‑H027X (K2 ONCOLOGY) were generated from tissue samples obtained from two patients (a 51‑year‑old male and a 59‑year‑old female). Ethics approval was granted under the number Lun Shen 2020 Ke 060 by the Medical Ethics Committee of Yixing People’s Hospital. Written informed consent was obtained from the patient for the use of their samples to generate organoids. Consent to publish personal data (sex and age) was also obtained from the patient. We used the CRISPR/Cas9 system to knock out candidate genes and evaluate their roles in organoid growth. The detailed methodology is as follows. Organoid culture Organoids were cultured with OrganoPro pancreatic cancer organoid culture reagent (K2 ONCOLOGY) in an incubator at 37 °C in CO2. The EHS gel Matrix glue (K2 ONCOLOGY) was placed in an ultralow-adsorption 24-pore-cell Petri dish (Cat# K2OS-24-M, K2 ONCOLOGY). Organoids were dissociated into single cells using passaged enzymatic hydrolysates (Catalog # K2O-TPCSA, K2 ONCOLOGY). sgRNA design and organoid transfection Gene-specific sgRNA oligos were cloned into lentiCRISPR v2 puro (Addgene). The sgRNA sequence was obtained from PrimerBank (https://pga.mgh.harvard.edu/primerbank) and is summarized in Supplementary Table 7. Organoids at a concentration of 2 × 105 cells/well were inoculated into a low-absorption 6-well plate, and 24 h later, lentivirus at a multiplicity of infection of 20 and 8 µg/mL polybrene were added. Two to 3 days after transfection, the culture medium was exchanged with medium containing 2 μg/mL puromycin. Single puromycin-resistant organoids were selected for clonal expansion. Western blotting was used to validate knockout efficiency. Proliferation assay of organoids The cell viability of each group (n = 10) was measured using the CellTiter-Glo® Luminescent Cell Viability Assay (G7570; Promega, USA) on days 0, 1, 3, 5, and 7. Cells were seeded in a 96-well plate at 8000 cells/well the day before treatment. Prior to treatment, the medium was aspirated and replaced with fresh medium. At the end of treatment, 100 μL of CellTiter-Glo® reagent was added to the 100 μL of the medium present in each well. The plate was mixed for 2 min on an orbital shaker to induce cell lysis and incubated at room temperature for 10 min to stabilize the luminescence signal. The absorbance at 490 nm was measured using a plate reader (BMG FLUOstar). Rapamycin treatment Cell proliferation was assessed using bright-field imaging on days 0, 3, 6, and 9, followed by live-dead staining with Calcein-AM at the endpoint (day 9). The resulting images were analyzed via the Xellar biosystems platform to quantify live cell area changes and evaluate drug-induced inhibitory effects on tumor organoids. The cell proliferation suppression efficiency of rapamycin in GRPEL1-knockdown organoid was evaluated with rapamycin concentrations of 240 nM. Validation of promoting gene function via animal studies Male BALB/c nude mice aged 5 weeks (Huafukang Company, Beijing, China) were housed under specific pathogen-free conditions in a controlled environment. The housing room was maintained on a 12-h light/dark cycle, with ambient temperature kept at 22 ± 2 °C and relative humidity maintained at 50 ± 10%. Animals were provided with standard laboratory chow (Cat.# 1021, Huafukang Company, Beijing, China) and sterile water ad libitum. For survival analyses, the animals were then randomly grouped (n = 4 mice per group) and subcutaneously injected with 0.1 mL of a cell suspension containing 2 × 106 PANC-1 cells, and the number of days from inoculation to tumor growth to a longest diameter of 15 mm was recorded. In our experiments, the maximum tumor size at the longest diameter was 15 mm, which was lower than the maximum tumor size permitted by the Institutional Animal Care and Use Committee (IACUC; 20 mm). The tumor volume (mm3) was measured with a Vernier calliper and calculated using the following formula: V = 0.5 (length × width2). After completion of the observation period, the mice were euthanised using cervical dislocation, and the tumors were carefully resected for measurement of tumor weight. Human PDAC tissue array analysis A paraffin-embedded tissue microarray (TMA) that contained 79 PDAC tissue samples and matched adjacent normal pancreatic tissues was purchased from YEPCOME Biotechnology Co., Ltd. (ZL-PanAsur1601; Shanghai, China). The TMA was dewaxed with xylol, ethanol, and demineralized water and then inactivated with 3% H2O2. Antigens were retrieved by boiling the samples in ethylenediaminetetraacetic acid (pH 8.0) for 15–20 min. After two washes with phosphate-buffered saline, the slides were blocked with 5% bovine serum albumin for 35 min and incubated with the primary antibody (anti-GRPEL1, 1:200; 12720-1-AP, Proteintech) at 4 °C overnight. Bound antibodies were detected using a commercial kit (RecordBio, RC0080R; Shanghai, China) according to the manufacturer’s instructions. Briefly, anti-rabbit biotin-conjugated secondary antibody was added, and the slides were incubated at 37 °C for 35 min. The color reaction was completed using a DAB-positive substrate. Cell nuclei were counterstained with haematoxylin. Designated protein expression was quantified using a HALO® image analysis platform based on the percentage of protein-positive cells and the intensity of staining. Statistics and reproducibility Statistical analyses were performed using GraphPad Prism 10.1.2 software. See the corresponding legend for the use of statistical methods. Sample sizes were not predetermined by statistical power calculations. Instead, they were determined by the availability of consecutively collected serum samples from PDAC patients with complete clinical data and at least 24 months of follow-up. The total sample size of 350 patients with OS data was deemed sufficient for this exploratory study as it exceeds those reported in similar metabolomic biomarker studies and allows for robust multivariate analyses with multiple covariates. All available eligible samples during the recruitment period were included to minimize selection bias and maximize the generalizability of the findings. For cell culture experiments, at least three independent biological experiments were performed for each condition, consistent with standard laboratory practice. For animal studies, group sizes of 4 mice per group were chosen based on prior experience and published literature in the field, which demonstrated that this sample size is adequate to detect significant differences in key outcome measures while minimizing the number of animals used in accordance with ethical guidelines. Given the observational nature of this study, no data were excluded from the analyses, the experiments were not randomized, and the investigators were not blinded to allocation during experiments or outcome assessment. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Data 5 Reporting Summary Transparent Peer Review File Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Data 5 Reporting Summary Transparent Peer Review File Source data Source Data Source Data