Abstract
Background: Pancreatic adenocarcinoma (PAAD) is a highly malignant cancer posing severe clinical challenges. Although the dual role of pyroptosis in tumor progression is increasingly recognized, the prognostic value of its molecular heterogeneity in PAAD remains underexplored. Methods: We integrated multi-omics data and applied interpretable machine learning to construct a predictive framework centered on pyroptosis heterogeneity. Using non-negative matrix factorization (NMF) on pyroptosis-related genes (PRGs), patients were classified into distinct molecular subtypes. Evaluating 117 machine learning combinations, we employed random survival forest (RSF) to build the final model, followed by comprehensive internal and external validation. SHapley Additive exPlanations (SHAP) analysis provided global and local interpretability. Clinical potential was assessed via nomogram, drug sensitivity prediction, single-cell analysis, and immunohistochemical validation. Results: We identified two biologically distinct pyroptosis subtypes and developed a ten-gene pyroptosis subtype-associated gene signature (PSAGS). PSAGS demonstrated robust performance across training, test, and multiple external validation cohorts, outperforming most published models. Multivariate analysis confirmed its independent prognostic value, and a PSAGS-based nomogram exhibited clinical utility. PSAGS-stratified subgroups showed differential responses to immunotherapy, chemotherapy, and targeted agents. Single-cell analysis revealed cell type-specific links between PSAGS scores and pyroptosis activity, indicating that high-PSAGS malignant cells foster an immunosuppressive microenvironment through extracellular matrix (ECM)-mediated signaling. Protein-level validation confirmed upregulation of signature genes in PAAD tissues. Conclusions: This work presents a biologically reliable prognostic model for personalized PAAD management and elucidates how pyroptosis heterogeneity drives tumor progression through cellular interactions.