A Generalizable and Interpretable Framework for Molecular Subtype Classification of Pancreatic Ductal Adenocarcinoma Integrating Conformal Uncertainty Quantification and Consensus-Based Explainable Artificial Intelligence Across Multiple Cohorts.

Pancreatic ductal adenocarcinoma (PDAC) has two principal molecular subtypes-classical and basal-like-with divergent prognosis and chemotherapy response, yet transcriptomic classifiers rarely generalize across cohorts or quantify per-patient uncertainty. We trained a classical-versus-basal-like classifier on CPTAC-PDAC (n = 140) and externally validated it on histology-filtered TCGA-PAAD (n = 150). Twelve algorithms were benchmarked under stratified nested cross-validation with four-method consensus feature selection; domain adaptation (naive transfer, CORAL, and ComBat), four conformal procedures (split, weighted, CV+, and Conformal Risk Control), and a four-method consensus explainable-AI framework (SHAP, LIME, permutation importance, and decision-curve ablation) with pathway enrichment were then evaluated. Top models reached a cross-validated AUROC ≈ 0.96 and external AUROC 0.913-0.938 (top-3 ensemble 0.961); batch correction did not improve transfer, indicating minimal residual batch effect. Consensus explainability recovered keratinization biology and nominated five candidate genes (GSDMC, A2ML1, PIP5K1B, IL20RB, and AKR7L) beyond the Moffitt signature. All four conformal procedures plateaued near 0.85 coverage at α = 0.05 under zero-shot transfer, whereas local recalibration on a small target sample restored nominal coverage. We present a transparent, externally validated, uncertainty-aware and TRIPOD+AI-compliant PDAC subtype classifier, best deployed as a calibrated decision-support tool with site-specific recalibration.