Senescence-driven molecular subtyping in pancreatic cancer: a multi-omics framework for precision medicine Authors: Meilong Shi, Penghao Li, Bo Li, Huan Wang, Xiaoyi Yin, Xiaohan Shi, Suizhi Gao, Yikai Li, Chuanqi Teng, Shuai Yuan, Xinyu Liu, Zhendong Fu, Xiaochao Kang, Wei Jin, Bin Song, Kailian Zheng, Yijie Zhang, Xiongfei Xu, Shiwei Guo, Gang Jin Published: 2025 ABSTRACT AbstractBackgroundPancreatic ductal adenocarcinoma (PDAC) remains a lethal malignancy with a five-year survival rate below 15%, largely due to tumor heterogeneity and limited therapeutic options. While senescence-related genes (SRGs) are implicated in cancer progression, their pancreas-specific roles in PDAC subtyping and treatment remain unexplored.MethodsWe integrated multi-omics data (RNA-seq, ATAC-seq, and whole-genome sequencing) from 402 pancreas-specific SRGs to classify PDAC subtypes through unsupervised clustering. Independent validation cohorts (TCGA-PAAD, n = 183; patient-derived organoids, n = 40) and drug sensitivity screens were used to define subtype-specific therapeutic vulnerabilities. A machine learning-based random forest model identified key SRG biomarkers for clinical stratification.ResultsThree distinct PDAC subtypes were identified: Cluster 1, characterized by extensive immune infiltration; Cluster 2, mixed features with moderate prognosis; and Cluster 3, defined by significant metabolic reprogramming. Drug screens revealed Cluster 3 as uniquely sensitive to Metformin and Trametinib, suggesting combinatory therapy potential. A 20-gene random forest classifier achieved high accuracy in subtype prediction (AUC = 0.96).ConclusionThis study establishes the first pancreas-specific SRG-driven classification of PDAC, resolving prior inconsistencies in Metformin trial outcomes. Our framework enables risk stratification and subtype-guided therapy, with immediate clinical implications: Metabolic-targeting agents (Metformin) may benefit the high-risk Cluster 3, while immunotherapy could be prioritized for Cluster 1.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12885-025-15341-z.Keywords: Pancreatic ductal adenocarcinoma, Senescence, Multi-omics analysis, Chromatin accessibility, Organoid, Drug sensitivity Introduction Pancreatic ductal adenocarcinoma (PDAC) is one of the most lethal cancers, with a five-year survival rate of less than 15% [1], posing a significant global health concern. In 2022, the United States alone reported over 66,000 new PDAC cases and more than 50,000 deaths, highlighting the urgent need for more effective treatments [1]. A major obstacle in PDAC treatment is its significant tumor heterogeneity, manifesting in distinct molecular subtypes and varying patient responses to standard therapies [2]. This complexity hinders the effectiveness of precision medicine, making it essential to improve the classification of PDAC subtypes for more tailored therapeutic approaches. Aging, a known risk factor for cancer, and cellular senescence (a hallmark of aging), plays a crucial role in tumor initiation and progression [3]. The accumulation of senescent cells leads to tissue dysfunction, and increased cancer susceptibility [4]. Recent multi-omics studies have revealed shared characteristics between aging and tumors, identifying senescence-related pathways as potential therapeutic targets [5]. For instance, combined treatment with Trametinib and Palbociclib has been shown to induce the senescence-associated secretory phenotype (SASP), promoting vascular remodeling, enhancing chemotherapy, and boosting T cell infiltration into PDAC lesions through SASP-mediated endothelial activation [6]. However, most studies focus on age-related senescence, overlooking the heterogeneity of senescent states within tumor tissues and their influence on treatment outcomes. Senescence can be induced by oncogenic signals or external stressors like radiation and chemotherapy, leading to oncogene-induced senescence (OIS) [7] or therapy-induced senescence (TIS) [8], respectively. While senescence initially acts as a defense against tumor growth, senescent cells can paradoxically promote tumorigenesis through the release of cytokines [9]. This dual role underscores the need to better characterize senescent states in PDAC to refine patient classification, guide therapeutic strategies, and improve clinical outcomes. In this study, we systematically investigated pancreas-specific senescence-related genes (SRGs) based on the AgeAnno [10] database to explore their role in PDAC and their association with disease progression. We identified three distinct senescence-related PDAC subtypes, each characterized by unique survival outcomes, genomic alterations, immune infiltration, metabolic profiles, and therapeutic responses. Multi-omics data reinforced the link between senescence and tumor characteristics. Furthermore, a pancreatic cancer-specific senescence signature was established, and comprehensive drug screening was performed to propose targeted therapeutic interventions for the identified senescence states. Of note, patients in Cluster 3 showed potential benefits from combining standard chemotherapy with Metformin. Overall, our findings provide a comprehensive framework for understanding the role of senescence in PDAC and propose novel therapeutic strategies to advance precision medicine approaches in PDAC treatment. Materials and methods Study design The local dataset consisted of two distinct cohorts: a discovery cohort of 286 pancreatic cancer patients and a validation cohort of 40 pancreatic cancer patient-derived organoids (PDOs). In the discovery cohort, whole-exome sequencing (WES) was performed on paired tumor and blood samples (n = 190), while RNA sequencing was completed for all 286 tumors. Tumor samples were obtained from patients undergoing surgical resection at Shanghai Changhai Hospital, with all samples pathologically confirmed. Clinical data were collected prospectively, and informed consent was obtained from all participants for data utilization and publication of potentially identifiable information. For the organoid validation cohort, RNA sequencing and Assay for Transposase-Accessible Chromatin (ATAC) sequencing were performed on all 40 organoids. Additionally, drug screening experiments were performed on a subset of 18 tumor organoids. The public validation dataset comprised of RNA sequencing data and its relevant clinical data from 183 pancreatic cancer samples were obtained from TCGA Data Portal (https://tcga-data.nci.nih.gov/). Ethical approval for this study was granted by the Changhai Hospital Ethics Committee (Approval number: CHEC2017-147). RNA-Seq library preparation and sequencing Total RNA extraction from tumor tissues and PDO samples was performed using the TRIzol reagent (Invitrogen, Carlsbad, CA, USA), followed by purification with the RNAeasy Mini Kit (Qiagen, Hilden, Germany) followed by the manufacturer’s instructions. RNA qualities were evaluated by agarose gel electrophoresis, and purity was measured using NanoPhotometer® spectrophotometer (IMPLEN, CA, USA). RNA integrity was confirmed using the RNA Nano 6000 Assay Kit of the Bioanalyzer 2100 system (Agilent Technologies, CA, USA). A minimum of 1 µg of total RNA per sample was utilized for cDNA library construction. cDNA library quality was evaluated with the Agilent Bioanalyzer 2100, and cluster generation was performed on a cBot System (Illumina, San Diego, California, USA) before sequencing on an Illumina Hiseq XTEN platform, generating 150 bp paired-end reads. RNA-Seq processing Raw sequencing data were processed with fastp v0.23.2 [11] for adapter trimming and quality control. The cleaned reads were aligned to the human reference genome (GRCh38) using HISAT2 v2.2.1 [12]. Read counts were generated with featureCounts v2.0.6 [13], referencing GENCODE v.42. Batch effects in RNA-seq data were corrected using the ComBat function from the sva R package [14]. Differential expressed genes were analyzed using the R package DESeq2 v.1.40.2 [15]. Across all samples, the total reads were 1.89 × 10^8 ± 6.56 × 10^7 per sample, with an average total mapping rate of 0.97 ± 0.01. WES library preparation and sequencing Genomic DNA extraction from fresh frozen tissue and whole blood samples was performed using the DNeasy Blood & Tissue Kit (Qiagen, Germany) following the manufacturer’s instructions. DNA concentration was measured by Qubit® DNA Assay Kit with a Qubit 3.0 fluorometer (Invitrogen, Carlsbad, CA, USA). 0.2–0.4 µg of DNA was used per sample. The SureSelect All Exon v6 Kit (Agilent Technologies, Santa Clara, USA) was used for target capture. DNA was fragmented to 350 bp using a Covaris E220 Focused-ultrasonicator (Covaris, Brighton, UK), followed by end-repair, A-tailing, and ligation of indexed adapters (Illumina, San Diego, California, USA). Library purification was performed using AMPure XP beads (Beckman Coulter, Beverly, USA). The concentration and size distribution of the libraries were evaluated using a Qubit®3.0 fluorometer (Invitrogen, Carlsbad, CA, USA) and NGS3K/Caliper. Cluster generation was executed on a cBot system (Illumina, San Diego, California, USA). High-throughput sequencing was conducted on the Illumina HiSeq XTEN platform, generating 150 bp paired-end reads with a target coverage of 50X. Mutation calling Raw sequence reads were first aligned to the Human Genome Reference (GRCh38) using BWA v.0.7.17 (BWA-mem) [16]. Duplicate reads were eliminated, and base quality scores were recalibrated using GATK v.4.1.98 [17]. Somatic single nucleotide variants (SNVs) were detected with MuTect, and somatic indels with Strelka. Identification of germline mutations was conducted using the GATK HaplotypeCaller. High-confidence somatic mutations were called based on the following criteria: (1) tumor and normal samples had sufficient coverage (≥ 10X); (2) at least 5% of total reads in the tumor and less than 1% in normal tissue supported the variant; and (3) at least three reads in the tumor supported the variant. Annotation was performed using ANNOVAR [18]. To enhance the specificity of our analysis, somatic SNVs and indels with a minor allele frequency exceeding 1% in the 1000 Genomes Project or those situated within segmental duplications were excluded from further consideration. Mutational signature and copy number variant analysis Mutational signatures of tumors from pancreatic cancer patients with different subtypes was conducted using the R package deconstructSigs [19]. The mut.to.sigs.input method was applied to prepare the input data, followed by whichSignatures to identify which COSMIC signatures v2 (https://cancer.sanger.ac.uk/cosmic/signatures) were present in the samples and their respective contributions to the mutational spectra. Copy number variants were analyzed using the GISTIC2 [20] algorithm. Copy number alterations were classified as follows: gains for genes with copy numbers exceeding the median ploidy; amplifications for those more than twice the ploidy; losses for those below the ploidy; and deletions for genes with a copy number of zero. Microenvironment gene expression signatures Gene signatures were analyzed using an established gene set to identify tumor microenvironment (TME) subtypes from transcriptome profiles [21, 22]. These gene signatures were applied to assess immune and stromal infiltration across senescence-related subtypes. Enrichment scores were calculated using single-sample Gene Set Enrichment Analysis (ssGSEA). Consensus clustering for senescence-related subtypes Pancreas-specific SRGs were extracted from the AgeAnno database [10], which provides single-cell annotations of aging in humans. Experimentally validated pancreatic tissue-specific senescence genes were retained. We then applied unsupervised clustering using the R package ConsensusClusterPlus v.1.64.0 [23], exploring solutions from two to ten clusters, with parameters set to “reps = 1000, pItem = 0.8, pFeature = 1”, and using the hc algorithm. Based on silhouette scores, changes between k values, and the proportion of ambiguous clustering, we identified three main clusters. Immune and stomal infiltration analysis Immune cell abundance (immune score), stromal infiltration (stromal score), and tumor purity were estimated using the ESTIMATE algorithm [24], which evaluates the abundance of stromal and immune cells in malignant tissues based on gene expression. Statistical significance of differences between groups was assessed through pairwise comparisons using Student’s t-tests. The xCell tool (v.1.1.0) [25] was utilized to measure the levels of different immune cell subtypes. Gene set enrichment analysis We retrieved hallmark gene sets, representing specific biological states or processes, from the Molecular Signatures Database (MSigDB). Single-sample gene set enrichment analysis (ssGSEA) was then performed using the GSVA [26] package in R to calculate pathway scores for each sample. Characteristics of transcriptomic changes in cluster 2 patients We used the Mfuzz [27] algorithm to identify characteristics of transcriptomic changes in Cluster 2 patients. Gene clustering was performed using DEGs enriched in Cluster 1 and 3, based on their average expression levels across tumor samples from Clusters 1, 2, and 3. We focused on genes with peak expression levels exceeding 0.5 to ensure the selection of biologically relevant transcriptional programs. ATAC-seq processing and data analysis Raw sequence reads were first processed using fastp [11] v0.23.2 to remove adapters and low-quality reads. Clean reads were then aligned to the human reference genome (hg38) using BWA v0.7.17 (BWA-mem) [28], with PCR duplicates removed using Picard v2.22.8. Peaks were identified using MACS2 v2.2.4 [29], with default settings except for the following parameters: -B -q 0.05 --cutoff-analysis --call-summits. Peak annotations were performed using the ChIPSeeker package v1.36.0 in R [30]. Peak differences between senescence-related clusters were compared using DiffBind v3.10.1 in R [31]. Significant differentially accessible peaks were identified with an FDR threshold of < 0.05, and transcription factor (TF) binding motif analysis was conducted using Homer v.4.11. Promoters were defined as regions within ± 1 kb of the transcription start site (TSS). Organoid culture The methods and conditions for organoid culture followed those outlined in a previous publication [32]. The basic medium consisted of advanced DMEM/F12, 10 mM HEPES, 1 × GlutaMAX-I, 100 µg/ml Primocin, and 1 × penicillin/streptomycin. The complete medium additionally contained 500 nM A83-01, 10 µM Y-27,632, 1.56 mM N-acetylcysteine, 10 mM nicotinamide, 10 ng/ml FGF10, 1 × B27 supplement, 10 µM forskolin, 30% Wnt3A-conditioned medium, 2% R-spondin conditioned medium, and 4% Noggin-conditioned medium. For all sequencing and functional assays in this study, PDOs were collected after being stably expanded for 10–15 passages, thereby ensuring biological stability and reproducibility. High-throughput screening of chemical and chemotherapeutic drugs For drug sensitivity testing, PDO samples were selected according to two criteria: (i) organoids had been stably expanded and maintained for 10–15 passages to ensure biological stability and reproducibility; and (ii) representative PDOs were chosen from each SRG-derived cluster. To enhance the robustness of the results, six PDOs from each cluster were ultimately included in the screening. Notably, the PDOs in this local validation cohort were derived from independent PDAC specimens and did not overlap with those used in the discovery cohort. 384-well plates with white, clear bottoms were coated with 10 µl of collagen and incubated at room temperature for at least one hour. Patient-derived organoids were dissociated with Tryp-LE, resuspended in medium, and seeded into the wells at a density of 3,000 cells per well. The next day, 58 compounds from Selleck, along with DMSO controls, were added in duplicate using the Bravo robotic workstation. Each compound was tested in seven different concentrations to assess sensitivity. Compound details are listed in Supplementary Table 10. After five days of incubation, cell viability was assessed using 25 µl of CellTiter-Glo Reagent, and luminescence was measured with an Envision plate reader. Viability values exceeding baseline were set to 100. Drug concentrations (nmol/L) were log10 transformed, and the area under the curve (AUC) was calculated using GraphPad Prism 9, with normalization achieved by dividing each AUC by the maximum AUC for that drug. For drug combination assays, gemcitabine was tested either as a single agent or in combination with metformin at a fixed concentration of 0.8 nM, which was selected based on the dose–response curves to induce cytotoxicity in Cluster 3 PDOs while exerting minimal effects on Clusters 1 and 2. Statistical significance between gemcitabine monotherapy and combination treatment within each cluster was determined using paired t-tests. Random forest model We analyzed RNA-sequencing data from 286 PDAC samples, representing three identified clusters. Gene expression values were quantified as Transcripts Per Million (TPM) and adjusted for batch effects using sva v.3.48.0 R package [14]. Non-protein or non-coding genes were excluded, and the top 10,000 genes with the highest variance were selected for further analysis. A stratified, repeated 5-fold cross-validation (caret::trainControl, method = “repeatedcv”, number = 5, repeats = 3) was performed to assess model performance. Within each resampling iteration, zero- and near-zero-variance genes were removed to prevent data leakage. A random forest classifier (randomForest v4.7-1.1.1.1) was trained with 500 trees (ntree = 500) and mtry set to 100 (square root of the number of predictors). Model performance was evaluated exclusively on held-out folds, yielding an average accuracy of 0.907 (95% bootstrap CI 0.895–0.918) and a Cohen’s κ of 0.858 across all 15 test sets. Gene importance was ranked by the mean decrease in Gini index, and the top 20 genes were identified as potential markers for cluster differentiation. Survival analysis Institutional follow-up for this study was conducted by dedicated follow-up specialists, with third-party professional data provided by LinkDoc Technology Co. Ltd. (Beijing, China). Follow-up methods have been previously described [33]. For patients alive at the study’s cut-off date, follow-up data were censored at the last confirmed date of survival. Patients who did not return for clinical follow-ups or could not be reached were classified as lost to follow-up. Overall survival (OS) was defined as the time from surgery to death, while disease-free survival (DFS) was defined as the time from surgery to cancer recurrence or death. Survival analysis were conducted using Kaplan-Meier curves with pairwise log-rank tests, and hazard ratios (HRs) were estimated from Cox proportional hazards models with 95% confidence intervals (CIs). Post hoc power analyses for key pairwise survival comparisons were performed using the powerSurvEpi package. Statistics All statistical analyses were conducted using R (version 4.3.2; https://cran.r-project.org/). Multiple hypothesis testing was uniformly corrected using the Benjamini–Hochberg false discovery rate (FDR) procedure, with FDR < 0.05 considered statistically significant. Continuous variables were compared using the Student’s t-test or Wilcoxon rank-sum test, as appropriate, and categorical variables were analyzed using the Chi-square test or Fisher’s exact test. All statistical tests were two-sided, and P < 0.05 (or FDR < 0.05, where applicable) was considered statistically significant. All clinical data, including imaging reports and laboratory results during hospitalization, were retrieved from the institutional electronic medical record system. If specific variables were unavailable, they were recorded as “NA” and excluded from analyses requiring those data. Results SRG-based classification of PDAC: implications for prognosis and recurrence To differentiate various senescence states in PDAC, we utilized a comprehensive dataset integrating RNA-seq and WES data obtained from PDAC tumors and matched blood samples. Additionally, we incorporated TCGA-PAAD dataset as a public validation cohort (Fig. 1A). Pancreas-specific SRGs were obtained from the AgeAnno database [10], which offers single-cell annotations of age-related genes by integrating transcriptomic and chromatin accessibility data across various human tissue types. Our analysis ultimately resulted in the inclusion of 402 pancreas-specific SRGs for further investigation (Fig. 1A). Utilizing these SRGs, unsupervised clustering was performed, stratifying PDAC patients into three distinct clusters: Cluster 1 (n = 112), Cluster 2 (n = 72), and Cluster 3 (n = 102) in the discovery cohort (Fig. 1B and Supplementary Table 1). Significant differences in overall survival (OS) and disease-free survival (DFS) were observed among these subtypes, with Cluster 3 demonstrating the poorest prognosis, indicating it as a high-risk subgroup (log-rank tests: P < 0.001 for both OS and DFS) (Fig. 1C and D). Our findings were corroborated by analysis of the TCGA-PAAD database, where we identified a cohort of 183 PDAC patients exhibiting three distinct senescence states (public validation cohort; Cluster 1, n = 68; Cluster 2, n = 26; Cluster 3, n = 89) (Fig. 1E). These three clusters demonstrated a notable separation trend in OS and progression-free survival (PFS). Importantly, Cluster 3 exhibited the poorest prognosis, consistent with the results observed in our discovery cohort results (log-rank test; P = 0.002, = 0.004, respectively) (Fig. 1F and G). Identification of three pancreas-specific senescence-related clusters in PDAC. A, Schematic overview of the workflow used to identify pancreas-specific senescence-related clusters, followed by multi-omics analysis. B, Heatmap displaying 286 samples from the local discovery cohort, classified into three distinct senescence-related clusters based on unsupervised clustering of 402 pancreas-specific senescence-related genes. C, Overall survival (OS) analysis (n = 276) of patients from the local discovery cohort, stratified by senescence-related clusters. D, Disease-free survival (DFS) analysis (n = 262) of patients from the local discovery cohort, stratified by senescence-related clusters. E, Same as (B), but for TCGA-PAAD public validation cohort (n = 183). F, OS (n = 123) analysis of public validation cohort patients with PDAC stratified by senescence-related clusters. G, Progression-free survival (PFS) analysis (n = 123) of the public validation cohort, stratified by senescence-related clusters Transcriptional profiling reveals distinct molecular signatures in identified subtypes Expression patterns of pancreas-specific SRGs varied significantly across the three identified PDAC clusters, highlighting their potential to differentiate patient outcomes. To unravel the underlying molecular mechanisms, we conducted a comprehensive analysis of transcriptomic data. This revealed distinct distribution patterns between Cluster 1 and Cluster 3, with Cluster 2 displayed a more dispersed distribution, occupying an intermediate profile (Fig. 2A). Differential expression analysis identified 6,494 differentially expressed genes (DEGs) between Cluster 1 and Cluster 3, including 3,279 up-regulated and 3,215 down-regulated genes (Supplementary Fig. 1A). The most significant DEGs illustrated clear functional distinctions. Notably, Cluster 3 exhibited notable enrichment of genes involved in metabolic processes, such as FABP1 and IRS4, both of which are associated with cancer progression [34–36]. Conversely, Cluster 1 displayed upregulation of immune-related genes, including DEFB132 and MYOC (Fig. 2B). To further elucidate the robust functional attributes of each cluster, we identified DEGs that were consistently upregulated in one cluster compared to the other two. This analysis revealed 734 genes enriched in Cluster 1 and 1,645 genes enriched in Cluster 3 (Fig. 2C and D). KEGG pathway analysis corroborated this distinction, with Cluster 1 was enriched in immune-related pathways, such as cytokine-cytokine receptor interactions, while Cluster 3 was enriched in pathways associated with energy metabolism and drug metabolism, including cytochrome P450 and fat digestion (Fig. 2E and Supplementary Table 2). GO analysis further substantiated these findings, reinforcing the functional disparities between the clusters (Supplementary Fig. 1B). Distinct characterization of three senescence-related clusters. A, t-SNE plot displaying the distribution of samples across the senescence-related clusters. B, Bar and scatter plot showing differentially expressed genes across pairwise comparisons between the clusters. C, Venn diagram depicting upregulated and downregulated genes in Cluster 1 compared with the other two clusters. D, Venn diagram showing upregulated and downregulated genes in Cluster 3 compared with the other two clusters. E, KEGG pathway enrichment analysis of genes from Clusters 1 and 3. F, Boxplots illustrating the immune scores among the three senescence-related clusters, calculated using the ESTIMATE algorithm. G, Heatmap of 286 PDAC samples based on ssGSEA enrichment scores for 29 immune-related Fges. The ssGSEA scores are Z-scored. H, Boxplots comparing CD4 + and CD8 + T cell levels among the three senescence-related clusters, calculated using the Xcell algorithm. I, GSEA enrichment results showing glycolysis in Cluster 3 compared with other clusters. J, Dynamic changes in the expression of Cluster 1 and 3 enriched DEGs divided into two subtypes using Mfuzz soft clustering. Only RNAs with a membership value > 0.5 were retained To explore the immune infiltration profiles of the identified clusters, we employed the ESTIMATE algorithm to analyze transcriptomic data, which deconvolutes immune, stromal, and microenvironmental gene signatures. Remarkably, Cluster 1 exhibited significantly higher immune scores compared to Clusters 2 and 3, indicating enhanced immune cell infiltration in Cluster 1 (Fig. 2F). To further elucidate the microenvironmental features of the three clusters, we evaluated 29 functional gene signatures (Fges) previously reported in PDAC and pan-cancer studies [21, 22]. Cluster 1 displayed significant enrichment across nearly all functional gene signatures, encompassing both anti-tumor immune effectors and immunosuppressive or tumor-promoting elements, indicating a highly inflamed yet immunoregulatory and fibrotic microenvironment (Fig. 2G). Additionally, we employed the xCell algorithm to compare immune cell populations among the clusters, revealing significant enrichment of CD4+ T cells, CD8+ T cells, B cells, and dendritic cells in Cluster 1 (Fig. 2H and Supplementary Table 3). Taken together, these findings suggest that Cluster 1 harbors a higher degree of immune infiltration, which may contribute to its distinct clinical behavior compared to Clusters 2 and 3. To further explore the fundamental characteristics of the identified clusters, we performed GSEA and ssGSEA using hallmark gene sets. GSEA revealed significant enrichment of metabolic pathways, particularly glycolysis, in Cluster 3 (NES = 1.93, FDR = 0.007) (Fig. 2I and Supplementary Table 4). Furthermore, ssGSEA revealed distinct enrichment patterns across the three clusters. Cluster 1 exhibited significant enrichment in immune-related pathways, notably the IL6-JAK-STAT3 signaling pathway, while Cluster 3 showed enrichment in pathways associated with tumor growth and energy metabolism (Supplementary Fig. 1D). Similarly, senescence biology pathway analysis revealed that Cluster 1 was predominantly enriched for stress-induced senescence pathways, while Cluster 3 showed enrichment for DNA damage- and oxidative stress-induced senescence pathways, with Cluster 2 displaying an intermediate profile (Supplementary Fig. 1E). Interestingly, Cluster 2 displayed an intermediate enrichment profile, suggesting its transitional nature between the immune-enriched Cluster 1 and metabolism-enriched Cluster 3. Given the observed enrichment of immune-related and hallmark gene sets, we hypothesized that the tumor biology of Cluster 2 patients represents an intermediate state between Clusters 1 and 3. To further explore the characteristics of Cluster 2, we applied the Mfuzz algorithm to analyze DEGs enriched in Clusters 1 and 3, focusing on those with dynamic expression profiles. This analysis categorized the genes into two distinct subtypes: Subtype 1 (n = 1,643), which showed a gradual increase in expression, and Subtype 2 (n = 736), which displayed a gradual decrease (Fig. 2J and Supplementary Table 5). Notably, samples from Cluster 2 consistently exhibited intermediate expression levels between Clusters 1 and 3 across both subtypes. Further comparison between Subtype 1 and Subtype 2 genes with those enriched in Clusters 1 and 3 revealed differences in only two genes: AGXT and RPS2P45 (Supplementary Fig. 1F). The high expression of AGXT in one Cluster 1 sample accounted for the difference, while the remaining samples exhibited similar expression across all clusters. RPS2P45 displayed similar characteristics in both Subtype 1 (0.48) and Subtype 2 (0.52) (Supplementary Table 5). Overall, the transcriptional profile of Cluster 2 supports its classification as an intermediate state between Clusters 1 and 3, aligning with the observed clinical outcomes. Integrative genomic and transcriptomic profiling of identified pdac clusters: concordance between copy number variations and gene expression profiles in cluster 3 samples To investigate the relationship between the identified PDAC clusters and somatic mutations, we conducted WES on matched tumor-blood samples from 190 of the 286 patients (Fig. 3A, Cluster 1, n = 64; Cluster 2, n = 51; Cluster 3, n = 75). No significant differences were observed in total mutation burden or the frequency of recurrent mutated genes among the three clusters (Fig. 3A). Furthermore, mutation signature analysis did not identify any cluster-specific characteristics (Supplementary Fig. 2A, B and Supplementary Table 6). Subsequently, we utilized GISTIC2 to identify copy number variation (CNV) regions within the clusters. All three clusters displayed extensive genomic amplifications across numerous regions (Fig. 3B and Supplementary Table 7, Cluster 1, n = 30; Cluster 2, n = 29; Cluster 3, n = 29), with 12 regions exhibiting significant amplifications across all clusters. Additionally, each cluster exhibited unique amplified regions, potentially reflecting their distinct biological features (Fig. 3C). Gene enrichment analysis of these amplified regions suggested that Cluster 1 were potentially associated with immune-related processes, such as the Herpes simplex virus 1 infection pathway (Fig. 3D and Supplementary Table 8). Conversely, Cluster 3 demonstrated strong associations with tumor-related pathways, including the PI3K-Akt signaling cascade and growth factor response mechanisms (Fig. 3D and Supplementary Table 8). Cluster 2, however, did not show significant enrichment in any particular pathway. These findings suggest that tumors in Cluster 3 may possess enhanced proliferative capabilities and represent more aggressive malignant phenotypes, consistent with the observed differences in clinical survival outcomes. Lastly, we evaluated the correlation between CNVs and RNA expressions, revealing a high degree of concordance between genomic amplifications and corresponding gene expression patterns (Fig. 3E). Genetic characterization of three identified clusters. A, Landscape of somatic mutation profiles across 190 PDAC patients, stratified by identified clusters (Cluster 1, n = 64; Cluster 2, n = 51; Cluster 3, n = 75). B, Significantly altered CNV regions among the three clusters. Red lines indicate amplification regions, while blue lines indicate deletion regions. C, Venn diagram depicting the distribution of genomic amplification regions across the three identified clusters. D, GO enrichment analysis of genes located within amplification regions specific to Clusters 1 and 3. E, Scatter plot depicting the relationship between RNA expression and CNV alterations. The left panel shows p-values, with green and red lines representing significance thresholds of P = 0.05 and P = 0.001, respectively. The right panel displays correlations, with the red line indicating a correlation coefficient of 0.3 Global increase in chromatin accessibility for metabolism-related genes in cluster 3 ATAC-seq was performed to assess chromatin accessibility changes across the different clusters in organoid cohort (local validation cohort) (Supplementary Fig. 3A, Cluster 1, n = 16; Cluster 2, n = 11; Cluster 3, n = 13). Consistent with the trends observed in both the discovery and public validation cohort, patients in Cluster 3 exhibited the worst OS and DFS (log-rank test; P = 0.032, = 0.095, respectively) (Supplementary Fig. 3B). While no significant differences in the gain of binding sites were observed between Cluster 1 and other Cluster samples, Cluster 1 exhibited a substantial loss of binding sites, whereas Clusters 2 and 3 showed increased chromatin accessibility (Fig. 4A). Chromatin accessibility and transcription factor binding motifs across the three senescence-related clusters. A, Enrichment of ATAC-seq signals around 1.5 kb upstream and downstream of the transcriptional start site (TSS) for samples from the merged clusters. B, Volcano plot showing differentially binding sites between Cluster 1 and Cluster 3. Red dots indicate gain of binding affinity, while blue dots indicate loss of binding affinity. C, Distribution of gain (Cluster 1) and loss (Cluster 3) peaks across the genome. D, GO enrichment analysis of genes associated with gain binding sites in Cluster 1 and Cluster 3. E, Line plot showing ATAC-seq signal for DEGs enriched in Cluster 1 and Cluster 3, based on RNA-seq analysis. Data is shown within ± 1 kb of the transcriptional start site. F, IGV visualization of peaks on PGK1 across samples from the identified clusters. G, The top 5 predicted binding motifs in the gain binding sites of Cluster 3 and the enriched DEGs in Cluster 3 Focusing on chromatin accessibility between Cluster 1 and 3, we identified 1,060 gained and 37,455 lost binding sites (Fig. 4B), indicating enhanced chromatin accessibility in Cluster 3 samples. Mapping these regions to the human genome (hg38) revealed that most open regions in Cluster 3 were in promoter regions, implying a stronger influence on gene expression (Fig. 4C). Clustering analysis of annotated genes indicated that regions with increased accessibility in Cluster 3 were enriched in pathways related to cell growth, Wnt signaling, and fatty acid metabolism (Fig. 4D). In contrast, Cluster 1 showed no enrichment for immune activation-related pathways, likely because its immune phenotype is driven by immune cell infiltration into the stroma, rather than directly by the tumor organoids of Cluster 1 (Fig. 4D). Further, we validated the genes enriched in Cluster 1 and 3 identified in the transcriptome using ATAC-seq data, which aligned with RNA-seq findings. Regions corresponding to genes enriched in Cluster 3 displayed the highest chromatin accessibility in Cluster 3 (Fig. 4E). Given the ATAC-seq peak annotation and the validation of Cluster 3 enriched gene regions, we anticipated a significant increase in chromatin accessibility signals for metabolism-related genes in Cluster 3. As expected, we observed elevated signals in genes such as PGK1, G6PD, GOT2, LDHA, PC and PKM (Fig. 4F and Supplementary Fig. 4). To identify potential transcription factors (TFs) regulating gene expression in Cluster 3, we performed motif analysis on the enriched peak regions. Additionally, we integrated TFs identified from the transcriptome analysis of Cluster 3 enriched DEGs. The top five TFs identified were BACH1, GATA1, HNF1, GATA3, and LHX3 (Fig. 4G and Supplementary Table 9). Previous studies have shown that BACH1 promotes pancreatic cancer metastasis [37], GATA1 contributes to gemcitabine resistance [38], GATA3 is involved in tumor metastasis and differentiation [39], and LHX3 is recognized as a metastatic oncogene [40]. These findings indicate that the increased chromatin accessibility and TF activity in Cluster 3 are associated with more aggressive tumor behavior and a poorer prognosis. Drug screening for identified clusters and therapeutic implications for cluster 3 patients To explore the potential correlation between the identified PDAC clusters and differential drug responses, particularly focus on Cluster 3, which has the worst prognosis, we conducted a comprehensive drug sensitivity screening using patient-derived 3D organoid (PDO) cultures. This screening evaluated variations in drug sensitivity across the identified clusters (Fig. 5A). A custom library of 58 compounds was assembled for the screen (Supplementary Table 10). Drug sensitivity screening was performed on 6 representative samples from each cluster to ensure robust results and consistent results. Drug sensitivity profiles of PDAC organoids across identified clusters. A, Heatmap illustrating drug sensitivity data from 18 PDO samples (Cluster 1, n = 6; Cluster 2, n = 6; Cluster 3, n = 6). Colors range from red to blue, representing Z-scores based on AUC values. Blue indicates high sensitivity, while red indicates low sensitivity. B, Metformin treatment response in organoids: Dose-response curves (left panel) and corresponding statistical analysis (right panel) for 18 PDAC organoids. C, Trametinib treatment response in organoids: Dose-response curves (left panel) and corresponding statistical analysis (right panel) for 18 PDAC organoids. D-F, Comparison of treatment responses to gemcitabine alone and in combination with metformin (0.8 nM) across organoids from different clusters. The top panels show the dose–response curves, and the bottom panels display the corresponding statistical analyses. Six PDOs per cluster were included in the analysis. G, Kaplan-Meier analysis of overall survival (OS) for Cluster 3 patients (n = 95) from the local discovery validation cohort, stratified by Metformin usage. H, Kaplan-Meier analysis of disease-free survival (DFS) for Cluster 3 patients (n = 95) from the local discovery cohort, stratified by Metformin usage. I, Schematic overview of the random forest model used to distinguish between the three identified clusters. J, Conceptual diagram outlining proposed clinical treatment strategies based on the identified PDAC clusters Remarkably, Cluster 3 samples, characterized by elevated metabolic activity across multi-omics datasets, demonstrated marked sensitivity to both Metformin and the senescence-inducing drug Trametinib. In contrast, Clusters 1 and 2 did not exhibit increased sensitivity to any single drug (Fig. 5A). Area under the curve (AUC) analysis confirmed significant differences in sensitivity to Metformin and Trametinib between Cluster 3 and the other clusters (Fig. 5B and C). To further investigate the therapeutic potential of metabolic targeting, we compared responses to gemcitabine monotherapy versus gemcitabine combined with metformin (0.8 nM) across six PDOs per cluster. Combination treatment significantly enhanced gemcitabine sensitivity in Cluster 3 PDOs compared with gemcitabine alone, whereas no significant synergistic effect was observed in Clusters 1 and 2 (Fig. 5D-F). Furthermore, patients in Cluster 3 (discovery cohort) treated with Metformin had better OS compared to those who were not treated (Fig. 5G, log-rank test; P = 0.030), although the trend in DFS was not statistically significant (Fig. 5H, log-rank test; P = 0.092). To enhance patient stratification, a random forest algorithm was applied, dividing the discovery cohort into a training set (n = 202) and a validation set (n = 84). This approach identified 20 highly informative genes that could accurately distinguish between the identified clusters, facilitating rapid and accurate clusters classification for individual patients (Fig. 5I). Based on the integration of our multi-omics data and drug sensitivity results, we propose that clinical management for Cluster 3 patients should include standard chemotherapy in combination with either Metformin or Trametinib (Fig. 5J). This personalized treatment strategy aims to exploit the unique molecular characteristics of Cluster 3 tumors and potentially improve patient outcomes. Discussion In this study, we uncovered significant heterogeneity within the identified PDAC clusters, encompassing variations in immune activation, metabolic reprogramming, and drug response. These variations directly influence treatment strategies and patient survival outcomes. By analyzing pancreas-specific SRGs, we identified three distinct clusters through unsupervised clustering of tumor samples. Cluster 1 exhibited marked immune activation within tumor tissues, indicating potential benefits from immunotherapies [41]. In contrast, Cluster 3 exhibited substantial metabolic reprogramming, particularly in glucose, lipid, and drug metabolism pathways, while Cluster 2 appeared as an intermediate state, potentially transitioning toward the metabolic profile of Cluster 3. Drug sensitivity analysis pointed to a potential therapeutic approach for Cluster 3, which is linked to the worst prognosis. A combination of metabolic agents like Metformin or senescence-inducing drugs such as Trametinib may provide therapeutic advantages for this high-risk group [42, 43]. Emerging evidence highlights the dual role of senescent cells in tumors [5]. Initially, these cells release SASP factors, which interact with the TME [9]. This interaction leads to vascular remodeling, recruitment of CD8+ T cells and dendritic cells, and presentation of senescence-specific antigens to immune cells [44]. These processes can activate immune responses, potentially transforming the immune microenvironment into a suppressive state [45]. Previous pan-cancer studies of the TME have identified two immune-enriched subtypes: fibrotic (IE/F) and non-fibrotic (IE), both associated with better prognoses compared to other subtypes [21]. Our observations in Cluster 1 align with these findings. Furthermore, the IE subtype has been shown to have a higher mutational burden and is more likely to respond to immunotherapy. However, while our study focused on pancreatic tumors, and the mutational burden of Cluster 1 did not significantly differ from the other two clusters, its immune-enriched TME suggests that this subtype might benefit from immunotherapy. Tumor cells rely heavily on metabolism adaptations, especially in pancreatic cancer, which is marked by extensive fibrosis and extracellular matrix deposition. This creates a unique tumor microenvironment characterized by high interstitial pressure, vascular collapse, and reduced nutrient and oxygen diffusion [46]. To survive under these harsh conditions, tumor cells intensify their utilization of alternative energy sources, such as glutamine [47], ornithine [48], carbohydrates [49], and lipids [50], which promotes tumor growth, proliferation, and metastasis. These metabolic adaptations align with the transcriptional characteristics observed in Cluster 3 and correspond to the poor clinical outcomes typically associated with advanced tumors undergoing metabolic reprogramming. Additionally, studies have shown that a metabolically reprogrammed subset of cancer-associated fibroblasts in pancreatic cancer drives tumor progression by supplying lipids for mitochondrial oxidative phosphorylation via the ABCA8a transporter [51]. Our enrichment analysis of bulk tumor samples in the discovery cohort revealed similar metabolic reprogramming patterns. Comparing our findings with previous research on colorectal cancer senescence subtypes, which identified three distinct groups [52], we noted similarities with Cluster 1, particularly in its immune characteristics and clinical outcomes. However, Cluster 3’s highly active metabolic profile appears unique to pancreatic cancer. Notably, drug sensitivity analysis suggests that patients in Cluster 3 could benefit from metabolic-targeted treatments or senescence-inducing drugs. Metformin, in particular, shows promise as both a preventive and therapeutic agent, though evidence has been inconsistent. A large cohort study from the Netherlands reported no link between metformin use and reduced risk of gastrointestinal cancers, including PDAC [53]. Metformin has also been shown to inhibit the initiation and progression of PDAC in genetically engineered mouse models [54], with further studies revealing its ability to suppress PDAC metastasis and epithelial-mesenchymal transition [42]. A 3D co-culture model of PDAC showed that metformin exerts anti-tumor effects by inhibiting pancreatic stellate cell migration through MMP2 downregulation and reducing cancer stemness factors [55]. These findings align with our results and support the potential for metabolic-targeted therapies in PDAC treatment. Previous randomized controlled trials (RCTs) on metformin produced negative results [56], possibly because patients were not stratified by tumor different subtypes. This RCTs suggesting that future research should focus on patients with hyperinsulinemia and tumors exhibiting sensitivity to energy stress markers, such as AMP kinase deficiency-a key regulator of cellular energy homeostasis. Our study enhances this classification by identifying subtypes that could benefit from stratified treatments. The observed sensitivity to Trametinib in Cluster 3 may be due to its ability to induce tumor cell senescence, maintain repressive chromatin states, silence pro-proliferative genes, and, in vivo, potentially release more SASP factors, promoting immune infiltration and inhibiting tumor growth and invasion [57, 58]. This provides further support for its use in patients with specific senescence-related profiles. To facilitate clinical implementation of our classification system, we employed a random forest-based method to efficiently categorize patients into the respective clusters. Through this approach, we identified 20 top senescence-related marker genes capable of accurately indicating cluster membership. Among these, genes highly expressed in Cluster 3 include HTATIP2, DLG3, STAP2, and FUT3. Notably, DLG3 has been reported to enhance protein glycosylation, leading to increased secretion of pro-tumorigenic growth factors that promote PDAC growth and metastasis [59]. Furthermore, FUT3, an N-fucosylation-related gene, is closely associated with pancreatic cancer metastasis [60]. These markers could play a crucial role in shaping patient treatment strategies in clinical practice. Our study has several limitations. First, as a retrospective study, the strength of the evidence is inherently limited, and detailed information on metformin treatment protocols, dosing, and patient selection criteria was not uniformly available. Well-designed prospective studies, including randomized controlled trials, are warranted to validate our findings and strengthen their clinical applicability. Second, although we utilized extensive multi-omics data for both analysis and validation, the lack of single-cell transcriptomic validation and mechanistic studies leaves gaps in understanding the biological drivers behind the identified clusters. Lastly, we did not offer clear therapeutic recommendations for patients in Cluster 1 or 2, and that additional metabolic inhibitors were not tested to further validate the metabolic vulnerability of Cluster 3. While it is reasonable to hypothesize that Cluster 1 patients might benefit from immune checkpoint inhibitors and Cluster 2 could require a more combined treatment strategy, further clinical and experimental validation is needed to support these approaches. Looking ahead, the 20-gene classifier established in this study provides a promising tool for patient stratification in clinical trials and precision therapy of pancreatic cancer, with the potential to guide metformin treatment decisions and deepen mechanistic understanding through integrated clinical and molecular data. Conclusion In conclusion, our study provides a novel classification of PDAC into three distinct clusters based on pancreatic tissue-specific senescence markers. Through comprehensive analysis of the tumor microenvironment, metabolic reprogramming, and epigenomic landscapes, combined with drug sensitivity data, we identified a targeted therapeutic strategy for Cluster 3. These findings offer crucial insights and promising avenues for advancing precision medicine in PDAC, with the potential to enhance treatment outcomes for this highly aggressive cancer. Supplementary Information Supplementary Material 1. Supplementary Figure 1. Functional and gene expression differences among the three senescence-related clusters. A, Workflow for differential expression analysis comparing gene expression between each pair of the three clusters. B, GO enrichment analysis of genes from Clusters 1 and 3. C, Boxplots comparing the levels of B cells and dendritic cells (DCs) among the three senescence-related clusters, as calculated using the Xcell algorithm. D, Heatmap of 286 PDAC samples based on ssGSEA enrichment scores for 50 hallmark gene sets. The ssGSEA scores are Z-scored. E, Senescence-related pathway enrichment analysis reveals distinct biological programs across SRG-based clusters. F, Venn diagram illustrates the distribution of genes from the two subtypes identified by Mfuzz in relation to the Cluster 1 enriched and Cluster 3 enriched genes. Supplementary Material 2. Supplementary Figure 2 Mutation signatures in PDAC samples. A, Stacked bar plot displaying the top 11 mutation signatures and other mutation signatures for each sample, with fractions shown. The clusters are annotated below each bar. B, t-test results comparing the top 11 mutation signatures with the remaining signatures classified as "Others". Supplementary Material 3. Supplementary Figure 3 Senescence-related clusters identified in organoid cohort. A, Heatmap showing gene expression for Cluster 1 and Cluster 3 enriched DEGs in 40 organoid samples. B, Overall survival (OS; top panel, n = 39) and disease-free survival (DFS; bottom panel, n = 34) of patients with PDAC, stratified by senescence-related clusters. Supplementary Material 4. Supplementary Figure 4 ATAC-seq signal profiles across metabolism-related genes in distinct PDAC cluster samples. IGV visualization of peaks on G6PD (A), GOT2 (B), LDHA (C), PC (D) and PKM (E)across samples from the three identified clusters. Supplementary Material 5. Supplementary Material 6. Supplementary Material 7. Supplementary Material 8. Supplementary Material 9. Supplementary Material 10. Supplementary Material 11. Supplementary Material 12. Supplementary Material 13. Supplementary Material 14. Acknowledgements Not applicable. Author contributions Meilong Shi: Methodology, Data curation, Validation, Writing - original draft. Penghao Li: Data curation. Bo Li: Data curation, Visualization. Huan Wang: Data curation. Xiaoyi Yin: Data curation, Visualization. Xiaohan Shi: Resources, Formal analysis. Suizhi Gao: Resources, Formal analysis. Yikai Li: Resources, Formal analysis. Chuanqi Teng: Resources, Formal analysis. Shuai Yuan: Data curation. Xinyu Liu: Data curation. Zhendong Fu: Data curation. Xiaochao Kang: Data curation. Wei Jin: Data curation. Bin Song: Data curation. Kailian Zheng: Data curation. Yijie Zhang: Data curation. Xiongfei Xu: Data curation. Shiwei Guo: Supervision, Writing—review & editing. Gang Jin: Supervision, Conceptualization, Funding acquisition, Writing—review & editing. All authors reviewed the manuscript. Funding This work was supported by the Noncommunicable Chronic Diseases-National Science and Technology Major Project (2025ZD0552300), National Natural Science Foundation of China (grant number 81972913), Shanghai Science and Technology Innovation Action Plan (grant number 23Y41900200), Special Clinical Research Project of Shanghai Municipal Health Commission (grant number 202240305), the Clinical Medicine Research Special Project for Major Disease Multi-center Clinical Research of Naval Medical University First Affiliated Hospital (grant number 2024LYA04), the China Postdoctoral Science Foundation (grant number 2023M744284), and 2025 National Major Science and Technology Project — New Organizational Model Initiative (Project No. 2025ZD0552300). Data availability The data that support the findings of this study are available from the corresponding author upon reasonable request. Declarations Ethics approval and consent to participate The study was conducted in accordance with the national guidelines and was approved by the Ethics Committee of Changhai Hospital (Approval number: CHEC2017-147). Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Penghao Li is the second co–first author. Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Meilong Shi, Penghao Li, Bo Li and Huan Wang are contributed equally. Contributor Information Shiwei Guo, Email: gestwa@163.com. Gang Jin, Email: Jingang@smmu.edu.cn. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Figure 1. Functional and gene expression differences among the three senescence-related clusters. A, Workflow for differential expression analysis comparing gene expression between each pair of the three clusters. B, GO enrichment analysis of genes from Clusters 1 and 3. C, Boxplots comparing the levels of B cells and dendritic cells (DCs) among the three senescence-related clusters, as calculated using the Xcell algorithm. D, Heatmap of 286 PDAC samples based on ssGSEA enrichment scores for 50 hallmark gene sets. The ssGSEA scores are Z-scored. E, Senescence-related pathway enrichment analysis reveals distinct biological programs across SRG-based clusters. F, Venn diagram illustrates the distribution of genes from the two subtypes identified by Mfuzz in relation to the Cluster 1 enriched and Cluster 3 enriched genes. Supplementary Material 2. Supplementary Figure 2 Mutation signatures in PDAC samples. A, Stacked bar plot displaying the top 11 mutation signatures and other mutation signatures for each sample, with fractions shown. The clusters are annotated below each bar. B, t-test results comparing the top 11 mutation signatures with the remaining signatures classified as "Others". Supplementary Material 3. Supplementary Figure 3 Senescence-related clusters identified in organoid cohort. A, Heatmap showing gene expression for Cluster 1 and Cluster 3 enriched DEGs in 40 organoid samples. B, Overall survival (OS; top panel, n = 39) and disease-free survival (DFS; bottom panel, n = 34) of patients with PDAC, stratified by senescence-related clusters. Supplementary Material 4. Supplementary Figure 4 ATAC-seq signal profiles across metabolism-related genes in distinct PDAC cluster samples. IGV visualization of peaks on G6PD (A), GOT2 (B), LDHA (C), PC (D) and PKM (E)across samples from the three identified clusters. Supplementary Material 5. Supplementary Material 6. Supplementary Material 7. Supplementary Material 8. Supplementary Material 9. Supplementary Material 10. Supplementary Material 11. Supplementary Material 12. Supplementary Material 13. Supplementary Material 14. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request.