Single-cell and bulk transcriptomics identify senescence-related EMT transcriptional programs and a prognostic framework in pancreatic ductal adenocarcinoma

Abstract

Background

Senescence-related transcriptional programs and epithelial-mesenchymal transition (EMT) are implicated in pancreatic ductal adenocarcinoma (PDAC) progression and therapy resistance. However, the single-cell distribution, functional heterogeneity, and clinical relevance of EMT-associated senescence-like programs remain incompletely defined.

Methods

We integrated single-cell RNA sequencing (scRNA-seq) data from 10 treatment-naive PDAC tumors with bulk transcriptomic profiles from TCGA-PAAD and an independent validation cohort. Senescence-related transcriptional activity was estimated using ssGSEA based on a curated CellAge gene set. EMT-like tumor cell subclusters were identified by unsupervised clustering after marker-based annotation, and differentiation trajectories were reconstructed using pseudotemporal analysis. A senescence-related gene signature was established for molecular subtyping and prognostic modeling. Drug sensitivity was explored using GDSC-based prediction and connectivity mapping.

Results

scRNA-seq analysis suggested that EMT-like tumor cells carried relatively high senescence-related transcriptional scores within the PDAC tumor microenvironment. Subclustering revealed five distinct EMT-like subsets, among which KRT19⁺ and UBE2C⁺ subsets showed high senescence-related scores, VMP1⁺ and NFE4⁺ subsets showed intermediate scores, and the MZB1⁺ subset showed low scores. Pseudotemporal ordering suggested an association between senescence-related transcriptional activity and EMT-like progression, but did not establish causality. Twelve senescence-related genes stratified patients into two subtypes (SubA and SubB). SubB patients exhibited poorer overall survival (median OS 17.0 vs. 34.8 months, log-rank P = 0.04) and an immune-excluded phenotype characterized by CD8 + T-cell depletion and reduced stromal/immune scores, despite evidence of altered tumor antigenicity. A refined 5-gene risk score (CDK1, FOXM1, NDRG1, PAK4, TPX2) showed moderate prognostic performance. Experimental validation supported up-regulation trends for the five model genes in PDAC, with statistically significant overexpression observed for NDRG1, PAK4, and TPX2 in the small validation cohort. TPX2 expression was positively associated with P16 expression. UMAP-based mapping revealed subcluster-specific enrichment of model genes. Drug-response prediction revealed distinct predicted IC₅₀ profiles between high- and low-risk patients, while CMap analysis nominated negatively connected compounds that may reverse the high-risk transcriptional signature.

Conclusions

EMT-like tumor subclusters with elevated senescence-related transcriptional activity may be associated with PDAC aggressiveness and therapy-related phenotypes. The molecular subtypes and 5-gene signature provide a hypothesis-generating framework for prognosis prediction and therapy selection. Prospective validation and functional experiments are required before clinical application or senescence-targeted treatment recommendations can be made.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12876-026-05043-6.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is a critical oncologic challenge. It is a digestive malignancy with insidious onset and no early specific symptoms [1]. Its aggressive biological behavior includes rapid local invasion, early distant metastasis, and profound therapy resistance. These features contribute to a dismal 5‑year survival rate below 10%. Radical surgery remains the only potentially curative intervention, yet approximately 80% of patients present with unresectable locally advanced or metastatic disease at diagnosis due to occult progression and inadequate screening biomarkers [2]. This clinical crisis stems from its complex tumor ecosystem, where dynamic epithelial-mesenchymal transition (EMT) programs facilitate metastatic dissemination. Fibrotic stroma orchestrates therapeutic resistance through physical and biochemical barriers [3, 4]. Recent advances in single-cell transcriptomics have begun mapping intratumoral heterogeneity, revealing EMT-like subclusters as key mediators of treatment failure [5, 6]. However, the functional integration between EMT plasticity and cellular senescence remains incompletely characterized. Cellular senescence is a potent driver of tumor microenvironment (TME) remodeling. This limits targeted therapeutic development.

Cellular senescence exerts context‑dependent effects in PDAC [7, 8]. On one hand, cellular senescence suppresses tumor initiation through cell‑cycle arrest and immune surveillance, processes mechanistically sustained by cellular senescence-associated secretory phenotype (SASP) factors such as interleukin-6 (IL-6), interleukin-8 (IL-8), CXCR2 agonists, and transforming growth factor-β (TGF-β) signaling [9, 10]. On the other hand, in established PDAC, persistent senescent cells acquire a pro‑tumorigenic cellular senescence‑associated secretory phenotype that remodels the tumor microenvironment—activating cancer‑associated fibroblasts, promoting extracellular matrix deposition, and recruiting immunosuppressive cells—thereby driving therapy resistance and metastatic progression [11, 12]. Importantly, emerging evidence indicates functional crosstalk between cellular senescence and EMT, a key driver of PDAC aggressiveness [13, 14]. However, how these two programs co-occur and interact within the PDAC ecosystem and mechanistically reinforce each other remains largely unexplored. Elucidating this interplay may uncover novel therapeutic vulnerabilities for this lethal malignancy.

To resolve this, we integrated high-resolution single-cell RNA sequencing (scRNA-seq) of treatment-naive PDAC tumors with bulk transcriptomic profiling across multiple cohorts. Our analysis suggests that EMT-like tumor cells may carry a relatively high senescence-related transcriptional burden. We identified five functionally distinct EMT-like subclusters with divergent differentiation trajectories and metabolic states. These subclusters appeared to demonstrate hierarchical organization along a differentiation continuum. KRT19⁺/VMP1⁺ EMT-like populations display stem‑like properties, while MZB1⁺EMT-like cells show features of terminal differentiation. Pseudotemporal ordering suggested, but did not prove, a relationship between senescence-related transcriptional activity and EMT-like progression. Leveraging these findings, we propose a clinically oriented, hypothesis-generating framework that may help stratify patients by senescence-related molecular subtypes and predicted therapeutic vulnerabilities. The flowchart of this study is shown in Fig. 1.

Fig. 1.

The flowchart of this study

Methods and materials

Data acquisition and preprocessing

Transcriptomic RNA-seq data and clinical metadata for pancreatic ductal adenocarcinoma (PDAC) were obtained from TCGA-PAAD. Primary tumor samples with documented overall survival (OS > 0 days) and chemotherapy response were classified as responders/sensitive cases (R, n = 36) or non-responders/resistant cases (NR, n = 43); four normal tissues served as controls. Independent validation cohort GSE62452 (65 PDAC samples with OS data) was acquired from Gene Expression Omnibus (GEO). Single-cell RNA-seq data from 10 treatment-naive primary PDAC tumors (GSE154778) were downloaded. A curated cellular senescence gene set was obtained from CellAgeDB (Build 3, https://genomics.senescence.info/cells/), containing 949 experimentally validated cellular senescence-associated genes (not deduplicated, Table S1) [15, 16].

Single-cell RNA-seq processing and quality control

Raw scRNA-seq data were processed using Seurat (v5.1.0) under R 4.3.2. Low-quality cells (< 200 genes, > 7,000 UMIs, or > 20% mitochondrial genes) were excluded; counts were normalized via SCTransform and integrated across samples using Harmony (v2.0) with parameters (theta = 2, lambda = 1, dims.use = 1:34, group.by.vars = orig.ident) to correct batch effects. Dimensionality reduction was performed via Principal Component Analysis, followed by Uniform Manifold Approximation and Projection (UMAP) visualization with 30 neighbors. The number of principal components for downstream analysis was determined by visual inspection of an ElbowPlot, and the first 34 PCs were retained. Potential doublets were identified and removed using DoubletFinder (v2.0.4) with pK = 0.09 and expected doublet rate nExp = 0.075×total cells.

Cell type annotation and EMT-like subclustering

After automated annotation with SingleR(v2.6.0), we manually refined cell type assignments for lineages that SingleR could not clearly distinguish. To identify EMT-like tumor cells, we first examined the UMAP projection and identified clusters that showed high expression of the EMT‑associated marker genes listed in the supplementary table of reference [17](see additional file 2, 13073_2020_776_MOESM2_ESM.xlsx). These candidate clusters were merged into a single group. To validate this annotation, differential expression analysis was conducted between the merged EMT-like population and all other cells via the Seurat. FindAllMarkers algorithm, with standardized thresholds: min.pct = 0.25, |log₂ fold change| > 1, and adjusted p-value < 0.05. The top differentially expressed genes in this merged group included the EMT markers from the reference list (e.g., SRPX, HMGA2, ANGPTL4, SERPINE2, DCBLD2), supporting its identity as the EMT-like tumor cell population. Finally, we subclustered the validated EMT-like cells using the same differential expression approach, which resolved five distinct subclusters (KRT19⁺, VMP1⁺, NFE4⁺, MZB1⁺, UBE2C⁺).

Cellular senescence scoring

Single-sample Gene Set Enrichment Analysis (ssGSEA) was implemented via the GSVA package (v1.52.2). The CellAge gene set was used to quantify cellular senescence-related enrichment scores for each cell and bulk sample. Scores were z-normalized within each dataset to enable cross-comparison. These scores were interpreted as senescence-related transcriptional activity rather than definitive evidence of irreversible cellular senescence.

Pseudotemporal trajectory reconstruction

Cell differentiation states were inferred using CytoTRACE (v0.3.3) based on gene counts per cell. This generates scores from 0 (terminally differentiated) to 1 (undifferentiated). Pseudotemporal ordering was performed with Monocle2 (v2.32.0) using DDRTree dimensionality reduction, with the root state determined by maximum CytoTRACE score. We identified pseudotime‑associated genes via differentialGeneTest with q‑value < 0.01.

Cell-cell communication analysis

Ligand-receptor interactions were deciphered using CellChat (v1.6.1) with default databases. Communication probabilities were computed via trimean-weighted permutation testing (100 permutations). Significant interactions (false discovery rate-adjusted P < 0.05) were visualized as directed networks and ligand-receptor heatmaps.

Differential gene expression analysis

Bulk RNA-seq count matrices were analyzed using DESeq2 (v1.42.1). Differentially expressed genes (DEGs) were defined by absolute log₂ fold change > 1 and Benjamini-Hochberg adjusted P-value < 0.05 for three comparisons: chemoresistant versus chemosensitive tumors, tumor versus normal tissues, and molecular subtype SubB versus SubA. Gene Ontology (Biological Process, Molecular Function, Cellular Component) and KEGG pathway enrichment analyses were performed using clusterProfiler (v4.10.1) with adjusted P < 0.05 significance threshold.

Prognostic model construction and validation

Univariate Cox proportional hazards regression identified OS-associated genes (P < 0.05). LASSO-penalized Cox regression (glmnet v4.1.3) with 10-fold cross-validation (λ selected by “lambda.1se” criterion) refined the signature. Risk scores were calculated as Σ(LASSO coefficient_i × z-scored expression_i) and patients stratified into high/low-risk groups by median score. Model performance was validated in GSE62452 using Kaplan-Meier survival curves (log-rank test), time-dependent receiver operating characteristic analysis (timeROC v0.4), and risk-stratified heatmaps.

Consensus molecular subtyping

Consensus clustering was performed using ConsensusClusterPlus (v1.54.0) based on prognostic senescence-related genes with parameters: k = 2–6 clusters, 80% resampling rate over 1,000 iterations, Euclidean distance, and Partitioning Around Medoids (PAM) algorithm; optimal k was determined by cumulative distribution function (CDF) stability and relative change in area under the CDF curve.

Tumor microenvironment characterization

Immune cell infiltration was quantified via CIBERSORT (LM22 signature matrix) and ssGSEA (23 immune cell types). Stromal, immune, and microenvironment scores were computed using ESTIMATE (v1.0.13). Expression of 45 immunoinhibitory and 30 immunostimulatory genes (sourced from TISIDB) was analyzed. Immunotherapy response was predicted via Tumor Immune Dysfunction and Exclusion (TIDE) scores and Immunophenoscore (IPS) computed using IOBR (v0.99.8).

Co-expression network analysis

Weighted Gene Co-expression Network Analysis (WGCNA v1.61) was applied to SubA/SubB DEGs. Soft-thresholding power β = 9 was selected to achieve scale-free topology (scale-free fit index R²>0.85). Modules were detected via dynamic tree cutting (minimum module size = 30) and merged at dendrogram cut height = 0.25. Module-trait associations (SubB status) were tested by Pearson correlation.

Therapeutic vulnerability profiling

Half-maximal inhibitory concentrations (IC₅₀) for 138 compounds were predicted using pRRophetic (v0.5) with Genomics of Drug Sensitivity in Cancer (GDSC) training data. Drugs with > 50% missing values were excluded. Connectivity Mapping (CMap; https://clue.io/) identified small molecules reversing high-risk signatures (negative connectivity scores). Enriched drug mechanisms were analyzed for significant compounds (|normalized enrichment score|>2, P < 0.01).

Validation and visualization

Model gene expression was mapped to EMT-like subclusters using Seurat’s FeaturePlot. All statistical analyses used R 4.3.2 with Benjamini-Hochberg FDR correction (adj.P < 0.05 significance threshold) applied where appropriate.

RNA extraction and reverse transcription quantitative PCR

Total RNA from PDAC and normal pancreatic tissues was extracted with TRIzol, reverse-transcribed (PrimeScript RT Kit), and amplified (TB Green Premix) on a Bio-Rad CFX96 system. Cycling conditions: 95 °C for 30 s, 40 cycles of 95 °C for 5s, 60 °C for 30s. GAPDH was the reference gene; relative expression was calculated via the 2^(–ΔΔCq) method with triplicates. Primers used in this study are listed in Supplementary Table S2 .

Multiplex immunofluorescence

Formalin‑fixed, paraffin‑embedded (FFPE) tissue sections (5 μm) were collected from 39 PDAC patients with complete follow‑up data (March 2020 to June 2024) at the Department of Pathology, Yijishan Hospital. This retrospective study was approved by the Ethics Committee of Affiliated Yijishan Hospital of Wannan Medical College, and written informed consent was obtained from all subjects. Sections were deparaffinized, rehydrated, subjected to antigen retrieval, and blocked with BSA. They were then incubated sequentially overnight at 4 °C with primary antibodies: rabbit anti-TPX2 (ProteinTech, Inc; 1:400) and rabbit anti-P16 (ProteinTech, Inc; 1:1000). This was followed by incubation with an HRP-conjugated goat anti-rabbit secondary antibody (Simuwubio, Inc.) and tyramide signal amplification (TSA) using a commercial multiplex fluorescence kit (Simuwubio, Inc.) according to the manufacturer’s instructions. The sequential staining process, including the secondary antibody incubation and TSA development, was repeated for the second primary antibody. After each TSA round, antibody stripping/HRP inactivation was performed according to the multiplex fluorescence kit protocol. Nuclei were counterstained with DAPI, and slides were mounted with an anti-fade mounting medium. Images were captured using an inverted fluorescence microscope (Nikon Eclipse CI-S) and quantified with Image J software.

Statistical analysis

All statistical analyses were conducted in R 4.3.2. Differential expression was determined using DESeq2 (Wald test; Benjamini-Hochberg [BH] adj. P < 0.05, |log₂FC| > 1). Survival analyses employed Cox proportional hazards regression (univariate: P < 0.05; LASSO-penalized Cox regression: 10-fold cross-validation, λ = “lambda.1se”) and Kaplan-Meier curves with log-rank tests. Continuous variable comparisons used Wilcoxon rank-sum tests (FDR-adjusted P < 0.05). Categorical associations were assessed by χ² tests. GO and KEGG enrichment analyses were considered significant at FDR-adjusted P < 0.05. For GSEA, FDR-adjusted P < 0.05 and |normalized enrichment score| > 1.5 were used as significance criteria.Trajectory inference (Monocle2) and cell-cell communication (CellChat) utilized permutation-based significance (FDR-adjusted P < 0.05). Time-dependent ROC curves (1–3 years) were generated via the timeROC package. All P-values were two-sided and BH correction was applied for multiple testing where appropriate.

Results

Single-cell transcriptomics identifies EMT-like tumor cells with high senescence-related transcriptional scores and reveals differentiation trajectories in PDAC

To delineate the PDAC cellular landscape, we analyzed publicly available scRNA-seq data (GSE154778). After rigorous re‑analysis and batch correction, we performed PCA and UMAP clustering. This revealed distinct populations. Annotation using SingleR and markers identified seven major types: epithelial tumor cells (ETC), cancer-associated fibroblasts (CAF), dendritic cells (DC), tumor-associated macrophages (TAM), T cells (TIL), endothelial cells (Endo), and epithelial-mesenchymal transition (EMT) cells (Fig. 2A, Figure S1, Table S3). Then, cell type-specific differentially expressed genes (DEGs) characteristic of each type were identified. The top five DEGs were visualized in Fig. 2B, and one representative DEG per cell type was displayed in UMAP feature plots (Figure S2). ssGSEA using a defined cellular senescence gene set revealed significant heterogeneity in senescence-related transcriptional enrichment across cell types (Kruskal‑Wallis rank‑sum test, P < 0.001). EMT-like tumor cells exhibited the highest average ssGSEA score (Fig. 2C, Table S4), suggesting that they are enriched for senescence-related transcriptional activity within the PDAC TME.

Fig. 2.

Single-cell transcriptomics identifies EMT-like tumor cells with high senescence-related transcriptional scores in PDAC. A UMAP plot of annotated cell types (ETC, CAF, DC, TAM, TIL, Endo, EMT) after batch correction. B Heatmap of top 5 cell type-specific differentially expressed genes per cell type. C Ridge plot of ssGSEA senescence-related scores across cell types (EMT highest). D UMAP plot of EMT-like subclusters (KRT19⁺, VMP1⁺, NFE4⁺, MZB1⁺, UBE2C⁺). E Bubble plot of representative marker expression across EMT-like subclusters. F Violin plots of log₂ fold change for key type-specific genes across EMT-like subclusters. G KEGG pathway enrichment of EMT-like subcluster marker genes

Given this enrichment, we further dissected EMT-like tumor cell heterogeneity through subcluster analysis. We identified five distinct subclusters: KRT19⁺EMT-like, VMP1⁺EMT-like, NFE4⁺EMT-like, MZB1⁺EMT-like, and UBE2C⁺EMT-like. This was based on their top cell type-specific differentially expressed genes identified using the FindAllMarkers function (Fig. 2D, Table S5). A bubble plot visualized expression of representative type-specific genes for each subcluster (Fig. 2E), supporting subcluster specificity. Differential gene expression highlighted distinct transcriptional profiles (Fig. 2F). KEGG pathway enrichment analysis revealed significant functional divergence among subclusters, with each exhibiting distinct pathway enrichment patterns (Fig. 2G, Table S6). Based on senescence-related score comparison, we classified the five EMT-like subclusters into three categories (Figure S3): high-score (KRT19 + EMT-like and UBE2C + EMT-like; mean ssGSEA scores 0.277 and 0.248, respectively), medium-score (VMP1 + EMT-like and NFE4 + EMT-like; 0.191 and 0.186, respectively), and low-score (MZB1 + EMT-like; 0.102).

To investigate cellular differentiation relationships, we computed a differentiation score (predict.order) for each cell (Table S7, Figure S3). CytoTRACE and pseudotime analysis indicated that KRT19⁺EMT-like and VMP1⁺EMT-like subclusters had the highest differentiation potential (Figure S4). To support the pseudotime ordering, we examined the expression dynamics of five EMT marker genes (SRPX, HMGA2, ANGPTL4, SERPINE2, DCBLD2) along the pseudotime trajectory. All five genes showed progressive upregulation with increasing pseudotime, supporting that the inferred trajectory is consistent with EMT-like progression (Figure S4). Cell-cell communication analysis revealed that the strongest collagen-mediated interactions involved the senescence-high subclusters (KRT19⁺ and UBE2C⁺), whereas the senescence-low MZB1⁺ subcluster exhibited minimal outgoing communication (Figure S5).

Cellular senescence-related genes are associated with chemotherapy response and prognosis

To delineate the role of senescence-related genes in therapeutic response and disease progression, we performed differential expression analysis on bulk transcriptomic data from TCGA. Comparison of chemotherapy-sensitive versus chemotherapy-resistant PAAD groups identified 3,467 DEGs (1,825 up-, 1,642 down-regulated in resistant tumors; Fig. 3A-B, Table S8). Intersection with EMT-like subcluster marker genes (from scRNA-seq) and a curated cellular senescence gene set yielded 66 core cellular senescence-associated resistance genes (Fig. 3C, Table S9). Functional enrichment analysis of these genes revealed involvement in cell cycle and senescence-related pathways: GO terms were significantly enriched for mitotic cell cycle processes (BP), cytosol (CC), and enzyme binding (MF) (Fig. 3D), and KEGG analysis highlighted “Cellular senescence” and “p53 signaling pathway” as top enriched pathways ( Fig. 3E).

Fig. 3.

Senescence-related genes associated with chemoresistance and prognosis. A Volcano plot of DEGs between chemosensitive vs. chemoresistant PDAC tumors. B Heatmap of DEGs from (A), clustered by resistance status. C Venn intersection of: (i) chemoresistance DEGs, (ii) EMT-like subcluster markers (scRNA-seq), (iii) cellular senescence genes (CellAgeDB); D-E GO/KEGG enrichment analysis of 66 genes. F Volcano plot of DEGs between tumor vs. normal pancreas. G Heatmap of tumor vs. normal DEGs. H Venn intersection of 66 cellular senescence-resistance genes and tumor vs. normal DEGs. I-J KEGG/GO enrichment of 26 genes.K Forest plot of univariate Cox regression for 12 prognostic genes

Further intersecting these 66 cellular senescence-resistance genes with DEGs from tumor versus normal pancreatic tissue (3,098 genes; 1,711 upregulated, 1,387 downregulated in tumors; Fig. 3F-G, Table S10; note that the heatmap shows some sample intermingling due to small normal sample size and tumor heterogeneity) identified 26 conserved cellular senescence-associated genes (Fig. 3H, Table S11). These genes exhibited enrichment for cell cycle regulation (KEGG: Cell cycle, p53 signaling pathway; GO: DNA damage response, microtubule cytoskeleton,protein/enzyme binding; Fig. 3I-J). Critically, univariate Cox regression analysis confirmed the clinical relevance of these 26 genes. It identified 12 genes significantly associated with poor overall survival (Hazard Ratio [HR] > 1 for all, P < 0.05; Fig. 3K, Table S12). Strikingly, elevated expression of these genes was associated with increased mortality risk, suggesting their potential prognostic relevance in PDAC. This underscores their collective role that is potentially associated with prognosis and therapy-related phenotypes.

Senescence-related molecular subtypes stratify survival and reveal transcriptional features associated with aggressive PDAC

To translate the identified senescence-related gene signature into a clinically oriented framework, we performed consensus molecular subtyping within TCGA-PAAD cohort. Unsupervised consensus clustering based on the expression of the 12 prognostic cellular senescence-associated genes defined two robust molecular subtypes: SubA (n = 43) and SubB (n = 36) (optimal k = 2; Fig. 4A-C, Table S13). SubB patients exhibited significantly worse OS than SubA (median OS: 17.0 vs. 34.8 months, Kaplan-Meier analysis with the log-rank test, P = 0.04; Fig. 4D). This suggests the subtypes have prognostic relevance. Notably, while these senescence-related subtypes showed no significant association with standard clinicopathological features (age, sex, TNM stage, grade, or initial response, Table S14), they were associated with vital status (alive vs. dead), which should be interpreted as supportive of outcome association rather than an independent biological validation (Fig. 4E). Although the heatmap does not show complete separation, the SubA/SubB classification is supported by consensus clustering stability, survival differences, and transcriptional and immune microenvironment alterations. Differential expression analysis between SubA and SubB identified 6,777 genes (Fig. 4F-G, Table S15), including dysregulated senescence-related genes (Fig. 4H, Table S16). WGCNA of these DEGs (soft-threshold power β = 9; Fig. 4I-J) resolved 14 co-expression modules (Fig. 4K). The blue module (858 genes) showed the strongest positive correlation with SubB (|r|=0.66, P < 0.05; Fig. 4L, Table S17). This identifies a core transcriptional network underlying this high-risk, senescence-related phenotype.

Fig. 4.

Senescence-related molecular subtyping and clinical translation. A Cumulative distribution function (CDF) for consensus clustering (k = 2–6). B Consensus matrix heatmap at optimal k = 2 (SubA: green; SubB: yellow). C Relative change in area under CDF curve, supporting k = 2. D Kaplan-Meier survival curves: SubB (yellow) vs. SubA (green). E Heatmap of 12 prognostic genes across clinicopathological groups. F Volcano plot of DEGs between SubB vs. SubA. G Heatmap of DEGs (SubB vs. SubA), clustered by subtype. H Box plots of top 20 senescence-related DEGs (Mann-Whitney U test, ****P < 0.0001). I WGCNA scale-free topology fit: R² vs. soft-thresholding power (β = 9 selected for R²>0.85). J WGCNA mean connectivity vs. soft-thresholding power. K Gene dendrogram with dynamic tree cutting (14 color-coded modules). L Module-trait correlation: Blue module positively correlates with SubB

Different subtypes exhibit distinct immune microenvironments with therapeutic implications

To assess the immunogenicity of the identified senescence-related molecular subtypes (SubA and SubB), we comprehensively profiled the tumor immune microenvironment. CIBERSORT analysis revealed reduced anti-tumor immune populations, including CD8 + T cells and monocytes, together with elevated immunosuppressive components such as Tregs in SubB (Fig. 5A). ssGSEA showed reduced enrichment of 20/28 immune cell types in SubB (Fig. 5B). ESTIMATE analysis also revealed global immune exclusion in SubB, with significantly lower immune, stromal, and microenvironment scores (Fig. 5C-E). Mutation burden analysis showed a non-significant increasing trend in SubB (Figure S6).

Fig. 5.

Distinct immune microenvironments of senescence-related subtypes. A CIBERSORT immune infiltration. B ssGSEA immune enrichment. C ESTIMATE immune scores. D ESTIMATE microenvironment scores. E ESTIMATE stromal scores. F Box plot of differences in Immunoinhibitor genes between molecular subtypes. G Box plot of differences in Immunostimulators genes between molecular subtypes. H TIDE scores. I IFNG scores. J IPS scores. Differences between SubA and SubB were assessed using the two‑sided Mann‑Whitney U test with Benjamini‑Hochberg correction for multiple testing. *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001

Further analysis uncovered widespread dysregulation of immune checkpoint molecules. SubB exhibited altered expression of 14 immunoinhibitors (e.g., PD-L1, CTLA4) and 23 immunostimulators (e.g., CD40, ICOS, Fig. 5F-G), indicating checkpoint imbalance. Assessment of immunotherapy response potential yielded divergent insights: TIDE scores showed no inter-subtype difference (Fig. 5H), whereas SubB demonstrated significantly reduced IFN-γ signaling (P < 0.01, Fig. 5I) but elevated Immunophenoscore (IPS; P < 0.05, Fig. 5J). These internally divergent prediction results suggest a complex immune-excluded phenotype, but immunotherapy response remains uncertain and requires validation in immunotherapy-treated PDAC cohorts.

Construction and validation of a 5-gene senescence-related risk score

To refine the prognostic senescence-related signature, we employed LASSO Cox regression on the 12 candidate genes. This identified five core genes (CDK1, TPX2, FOXM1, NDRG1, PAK4 ) with non-zero coefficients (Fig. 6A-B). A composite risk score was calculated using their LASSO-weighted expression. In the TCGA training cohort (n = 79), patients stratified into high-risk (H) and low-risk (L) groups by median score exhibited significantly divergent survival (log-rank test P = 0.0076; Fig. 6C, Table S18). The model showed consistent predictive accuracy at 1–3 years (AUC > 0.6; Fig. 6D), and elevated gene expression was associated with higher risk (Fig. 6E).

Fig. 6.

Construction and validation of a 5-gene senescence-related prognostic model. A LASSO coefficient profiles for the 12 candidate genes. B LASSO cross-validation curve. C KM survival curves in TCGA training cohort: High-risk vs. low-risk groups. D Time-dependent ROC curves in TCGA training cohort. E Risk stratification heatmap (TCGA). F KM validation in GSE62452 cohort. G ROC curves in GSE62452 cohort. H Risk heatmap in GSE62452

This prognostic power was validated in an independent GEO cohort (GSE62452, n = 65). High-risk patients again showed significantly worse survival (log-rank test P = 0.012; Fig. 6F, Table S19) and consistent time-dependent AUC > 0.6 (Fig. 6G). Gene expression trends mirrored the training set (Fig. 6H). GSEA and protein‑protein interaction analysis suggested that these genes were enriched in cell-cycle-related pathways and formed an interaction network (Figure S7).

Model-guided drug-response prediction and CMap analysis

To delineate the single-cell distribution and therapeutic implications of the 5-gene senescence-related signature, we mapped model genes within EMT-like subclusters using UMAP FeaturePlot visualization. CDK1, FOXM1, and TPX2 exhibited predominant enrichment in UBE2C + EMT-like cells, while NDRG1 localized to VMP1 + EMT-like and PAK4 to KRT19 + EMT-like (Figure S8). This supports subtype-specific expression patterns of prognostic targets and links senescence-related transcriptional status to predicted therapeutic vulnerability. Drug sensitivity profiling across 138 compounds identified eight chemotherapeutics with significantly higher IC₅₀ (P < 0.01) in high-risk patients (Fig. 7A–H, Table S20-21). Complementary connectivity mapping (CMAP) of risk-associated DEGs prioritized small molecules predicted to reverse the high-risk phenotype (negative enrichment scores, Table S22). Representative negatively connected compounds included SGI-1776, NVP-TAE684, trifluridine, BVT-948, regorafenib, dorsomorphin, tivozanib, PF-03758309, CGP-60474, and losartan. Enriched mechanisms (e.g., highlighted HDAC inhibition, kinase inhibition, CDK inhibition; Fig. 7I-J) suggest candidate druggable pathways for future validation.

Fig. 7.

UMAP-based mapping and therapeutic vulnerability profiling guided by the 5-gene model. A-H Box plots comparing predicted IC₅₀ values of 8 compounds between high- vs. low-risk groups; Continuous variable comparisons used the two-sided Mann-Whitney U test (BH adj. P < 0.05). I Representative negatively connected compounds identified by CMap analysis. Only compounds with negative raw connectivity scores were included, and the x-axis represents the absolute value of raw_cs. J Mechanism-of-action profiles of negatively connected CMap compounds. The x-axis represents the absolute value of raw_cs

Experimental validation of senescence-related gene expression and prognostic significance

RT–qPCR of clinical specimens revealed up-regulation trends of CDK1, FOXM1, NDRG1, PAK4, and TPX2 in PDAC tissues (n = 10) vs. adjacent normal tissues (n = 10). However, due to the limited sample size, only NDRG1, PAK4, and TPX2 showed statistically significant over-expression (P < 0.05). FOXM1 exhibited a trend toward up-regulation that did not reach statistical significance (P = 0.06), while CDK1 also showed a trend but with a P value of 0.117 (Fig. 8A-E). These results should be interpreted with caution, and larger independent cohorts are needed to confirm the differential expression of all five model genes. Given TPX2’s marked dysregulation and emerging importance in PDAC pathogenesis reported in previous studies [18, 19], we selected this mitotic regulator for further validation. Interrogation of the GEPIA database revealed a statistically significant positive correlation between TPX2 and the established cellular senescence marker P16 (CDKN2A) at the transcriptomic level (Pearson’s correlation test, R = 0.36, P < 0.001; Fig. 8F). This association was assessed at the protein level using multiplex immunofluorescence (mIF) in an independent cohort of 39 PDAC patients (Table S23). TPX2-high tumors exhibited significantly increased density of P16-positive cells (Mann-Whitney U test, P < 0.05; Fig. 8G-H). Survival analysis based on mIF quantification demonstrated that patients with high TPX2 expression had significantly worse OS (log-rank test, P < 0.05; Fig. 8I). Because P16 alone does not prove classical cellular senescence in PDAC, these findings support a TPX2-P16 association rather than definitive TPX2-driven senescence.

Fig. 8.

Experimental validation of senescence-related gene expression and clinical prognosis. A-E RT–qPCR analysis of the five model genes in PDAC (n = 10) versus adjacent normal tissues (n = 10). Data are presented as mean ± SEM; statistical significance was assessed by two-tailed Student’s t-test. F GEPIA correlation between TPX2 and P16 expression. G Representative multiplex immunofluorescence images of TPX2 (green), P16 (red) and DAPI(blue) staining in PDAC tissues. The full cohort consisted of 39 PDAC patients. H Comparison of P16⁺ cell density between TPX2-high and TPX2-low patient groups (Mann-Whitney U test, P < 0.05). (I) Kaplan-Meier survival analysis for patients stratified by high vs. low TPX2 protein expression (log-rank test, P < 0.05)

Collectively, these experiments partially validate the bioinformatically identified senescence-related genes and specifically support TPX2 as a senescence-associated prognostic biomarker candidate in PDAC.

Discussion

This study addresses key clinical challenges in PDAC management. First, it identifies EMT-like tumor subclusters with elevated senescence-related transcriptional activity that may contribute to therapeutic resistance, offering candidate targets. Second, it develops a 12‑gene molecular subtype and a 5‑gene risk score for patient stratification. Third, it proposes computationally predicted treatment vulnerabilities as hypotheses for future precision-medicine studies.

Cellular senescence and EMT are often co-regulated by shared tumor suppressors and EMT-related transcription factors [13, 20]. Our integrated single-cell and bulk transcriptomic profiling establishes EMT-like tumor subclusters as cellular compartments with high senescence-related transcriptional scores in PDAC. This suggests a functionally coupled EMT-senescence-like axis that may be associated with stromal fibrosis and chemotherapy response. Through re-analysis of scRNA-seq data, we found that EMT-like tumor cells exhibit the highest senescence-related enrichment across the tumor microenvironment. KRT19 + EMT-like subclusters were positioned near the pseudotemporal trajectory root (indicating stem‑like plasticity) and show pronounced senescence-related transcriptional activity. Pseudotemporal trajectory analysis identified select differentiation-dependent genes (e.g., MFAP5) associated with extracellular matrix reorganization. Pathway enrichment reveals involvement of focal adhesion mechanisms in key EMT subclusters. Transcriptomic screening identified cellular senescence-associated resistance genes dysregulating p53 signaling and cell-cycle pathways. SASP factors (e.g., IL6) and cell-cycle regulators (e.g., TPX2) may converge within specific EMT-like niches, suggesting a potential functional link. These findings suggest a hierarchy wherein EMT differentiation states influence senescence-related transcriptional intensity. This may contribute to collagen deposition through extracellular matrix reorganization and may associate with chemotherapy response through cell-cycle dysregulation (e.g., CDK1/TPX2 overexpression). The UBE2C⁺EMT-like subcluster exhibited a relatively high senescence-related score despite intermediate differentiation potential, which may indicate a distinct stress or cell-cycle-associated program, as UBE2C is known to regulate cell cycle progression. Notably, our cell-cell communication analysis did not reveal prominent SASP‑related ligand-receptor signaling directed toward EMT-like cells (Figure S5). This suggests that the elevated senescence-related scores in EMT-like cells may arise from intrinsic cellular programs, such as cell-cycle dysregulation and p53 pathway activation, rather than from paracrine SASP input from other TME components. Previous studies also link EMT and cellular senescence in malignancies [14–22]. For instance, Park et al. showed that p16INK4A-positive senescent tumor cells at the invasive front associate with EMT phenotypes, ECM degradation, and epigenetic reprogramming. This indicates cellular senescence as a driver of malignant evolution and invasive partial EMT states linked to metastasis and poor prognosis. However, our ssGSEA‑based senescence score reflects a composite transcriptional signature. The elevated senescence scores in EMT-like subclusters could partly arise from active cell‑cycle programs or stress responses rather than irreversible senescence. Distinguishing between true senescence, transient stress adaptation, and proliferative EMT intermediates requires functional assays (e.g., SA‑β‑Gal staining, live‑cell tracking, or lineage tracing). We thus interpret our findings as a link between EMT‑associated transcriptional programs and senescence‑related gene expression, not as proof of classical senescence.

Molecular classification based on tumor cell characteristics is a critical foundation for precision medicine [23–25]. Our consensus molecular subtyping delineates two biologically distinct PDAC entities, SubA and SubB, via a validated 12-gene senescence-related signature. SubB patients show median OS of 17.0 months vs. 34.8 months in SubA. This suggests that the classification has prognostic relevance beyond conventional clinicopathological variables, although independence should be confirmed by multivariable Cox analysis. SubB was characterized by hyperactivated cellular senescence-related/EMT transcriptional programs and predicted poorer outcome despite no significant clinicopathological associations. Further refinement via LASSO Cox regression yielded a tractable 5-gene risk score (CDK1, FOXM1, NDRG1, PAK4, TPX2) that stratifies patients across TCGA and GEO cohorts. This molecular taxonomy may inform therapeutic hypotheses. Drug-response prediction revealed differential predicted IC₅₀ profiles between risk groups. Collectively, this framework moves PDAC stratification from purely anatomic descriptors toward functionally adaptive molecular features, but it requires prospective and mechanistic validation before clinical use.

SubB tumors exhibit characteristics of an immune-excluded phenotype. Hyperactivated innate immunity (activated NK cells up‑regulated) paradoxically coexists with adaptive immune failure. This is evidenced by CD8⁺ T‑cell depletion, immunosuppressive networks (Tregs up‑regulated), dismantled checkpoint architecture (CTLA4 down‑regulated), and reduced stromal scores (lower ESTIMATE immune/stromal scores). These changes coincided with suppressed IFN-γ signaling despite heightened immunogenicity (higher IPS). Mechanistically, we speculate that immune exclusion in SubB might stem from multifaceted suppression. Transcriptomic profiling showed widespread gene dysregulation in SubB, including altered expression of key immune checkpoint molecules (e.g., PD-1, CTLA4). This, combined with depleted CD8 + T cell infiltration, is consistent with impaired T-cell activity [26]. Furthermore, this immunosuppressive microenvironment may be exacerbated by SASP-like components, which could impair antigen presentation capacity and hinder T-cell priming. Hypoxic stress within fibrotic niches may also disrupt T-cell infiltration and function [27, 28].

Our integrated drug-response prediction and molecular subtyping framework may help prioritize candidate therapeutic hypotheses for future validation. Conventional pan-senolytic strategies are limited by systemic toxicity and compensatory resistance mechanisms [29, 30]. In contrast, our 5-gene risk score stratified patients, with high-risk subtypes showing both elevated risk and differential predicted drug response. We identify eight compounds with higher predicted IC₅₀ values in high-risk patients. Connectivity mapping prioritizes signature‑reversing small molecules (e.g., SGI-1776, NVP-TAE684, trifluridine, BVT-948, regorafenib, dorsomorphin, tivozanib, PF-03758309, CGP-60474, losartan).UMAP-based mapping of the 5-gene signature within EMT-like subclusters provides a bridge between molecular risk and functional targeting. These findings suggest differential predicted drug response between risk groups and should be validated experimentally before therapeutic interpretation (e.g., CDK inhibitors targeting CDK-high subclusters [31]). Selected agents may not only directly target tumor cells but also disrupt senescence-like cell-stroma crosstalk, reducing stromal barriers and sensitizing tumors to conventional chemotherapy [32, 33]. By tethering drug selection to a biology-driven risk stratification model and single-cell-resolved targets, our approach may help prioritize candidate therapeutic hypotheses for future validation.

While this study highlights senescence-EMT crosstalk as a potential regulator of PDAC aggressiveness, key questions remain. First, our analysis does not provide causal mechanistic evidence; all interpretations based on transcriptomic and proteomic data are correlative. Functional perturbation experiments, such as knockdown or knockout of key senescence regulators in specific EMT-like subclones, are required to establish causality. Future studies should also validate the bioinformatically predicted roles of SASP components through in vitro co-culture systems and in vivo genetic perturbation models. Second, we relied on a single predefined cellular senescence gene set (CellAge Build 3). The composition of any senescence signature may affect results, including pathway enrichment and prognostic gene selection. Future studies employing multiple senescence-related signatures or experimentally derived gene sets would help validate and generalize our findings. Third, while our scRNA-seq captured treatment-naive tumors, longitudinal post-chemotherapy sampling is needed to define therapy-induced senescence evolution. Fourth, prospective trials must assess whether senolytic agents or genetic ablation of senescence-associated genes in EMT cells can resensitize SubB tumors to chemotherapy/immunotherapy and test the functional relevance of these subclusters. Fifth, the experimental validation cohort was relatively small, which may have limited statistical power to detect modest but biologically relevant differences. Sixth, future work should also explore non-cell-autonomous effects of senescence-EMT-like subclusters on neural invasion and metastatic organotropism. Therefore, the findings presented here should be considered hypothesis-generating rather than definitive proof of senescence-dependent mechanisms.

Conclusions

In conclusion, our study defines EMT-like tumor subclusters with elevated senescence-related transcriptional activity as candidate cellular compartments associated with aggressive biology in PDAC. We establish a senescence-related molecular taxonomy and a 5-gene prognostic signature. The high-risk subtype, characterized by an immune-excluded tumor microenvironment, presents both a clinical challenge and a potential opportunity for further therapeutic exploration.

Authors' contributions

LJC, LLZ and XMW conceived and designed the study. LJC, LW and SHC performed the bioinformatic and statistical analyses. SHC and SZ conducted the experimental validation. JW collected clinical samples and performed patient follow-up. LJC and LW wrote the first draft of the manuscript. LLZ and XMW revised the manuscript. All authors contributed to manuscript revision, read, and approved the submitted version.

Funding

This study was funded by Wuhu Health Commission Scientific Research Project (WHWJ2023y010), the Special Research Start-up Funds for Introduced Talents of The First Affiliated Hospital of Wannan Medical College (YR202411), and Wuhu “Huatuo Plan” High-level Talent Introduction and Training Project (HTRCRD202420).

Data availability

Publicly available datasets analyzed in this study were obtained from TCGA-PAAD, GSE154778, GSE62452, CellAgeDB, GDSC, and CMap/CLUE. Additional processed data are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the Declaration of Helsinki. It was approved by the Institutional Review Board of Affiliated Yijishan Hospital of Wannan Medical College (Approval No. 2025066). Written informed consent was obtained from all subjects.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.