Explainable multi-omics modeling for risk stratification in pancreatic ductal adenocarcinoma

Pancreatic ductal adenocarcinoma (PDAC) remains one of the most lethal malignancies due to a lack of reliable tools for individualized risk stratification. A comprehensive understanding of the multi-omics landscape may uncover clinically applicable biomarkers and inform precision prognostic assessment. This study aims to establish a prognostic model directly from the complete omics landscape and extract biomarkers.

We developed prognostic models using multi-omics data from a PDAC proteogenomic cohort comprising 75 deceased tumor samples. An independent cohort of 63 deceased PDAC cases from The Cancer Genome Atlas (TCGA)-pancreatic adenocarcinoma (PAAD) was used for external validation. Logistic regression models with least absolute shrinkage and selection operator (LASSO) regularization were constructed, and SHapley Additive exPlanations (SHAP) were applied to evaluate feature importance and identify signature genes. Model selection was based on the average area under the receiver operating characteristic curve (AUROC) across cross-validation folds. Functional validation was performed in PANC-1 cells by knockdown (KD) or overexpression (OE) of representative microRNA-, RNA-, and proteomics-derived signature genes, followed by Cell Counting Kit-8 (CCK-8) proliferation and Transwell migration assays.

Systematic evaluation of 120 multi-omics combinations identified a top-performing prognostic model integrating RNA, microRNA, proteomics, and mutation features. This model achieved a mean AUROC of 0.92±0.11 and accuracy of 0.87±0.01 on internal validation, and 0.99±0.00 and 0.98±0.01 on the TCGA test set. The sensitivity, specificity, precision, recall and F1 scores on the TCGA test set were 0.98±0.01, 0.97±0.02, 0.98±0.02, 0.98±0.01, 0.98±0.01, respectively. SHAP analysis revealed interpretable and clinically relevant prognostic biomarkers, many of which are implicated in immune signaling, metabolic regulation, and cell cycle control. Importantly, modulation of representative signature genes in PANC-1 cells significantly altered proliferation and migration in directions consistent with model-predicted risk associations.

Our findings demonstrate that explainable multi-omics machine learning frameworks can identify robust prognostic biomarkers and achieve highly accurate survival prediction in PDAC. Functional validation further supports the biological relevance of these signatures, underscoring their translational potential for personalized risk assessment.