Pancreatic ductal adenocarcinoma (PDAC) remains one of the most lethal human malignancies, in part due to late diagnosis and the lack of robust molecular biomarkers. Although aberrant DNA methylation is a defining feature of PDAC, most studies rely on single cohorts, limiting reproducibility and biological interpretation. Here, we performed an integrative multi-cohort analysis of genome-wide DNA methylation profiles from four independent PDAC datasets generated on Illumina EPIC and HumanMethylation450 platforms, comprising 364 tumors and 99 normal controls. Using a harmonized preprocessing cross-platform normalization strategy, we identified hundreds of CpG sites that were consistently differentially methylated across all cohorts. Integration with pancreatic chromatin-state annotations, hydroxymethylation profiles, and protein–protein interaction networks was used to contextualize recurrent DNA methylation changes. This analysis showed that hypermethylation preferentially targets promoter- and enhancer-associated regulatory elements linked to neuronal and developmental gene networks. To assess predictive relevance, we trained interpretable and non-linear machine-learning models with strict cross-cohort evaluation, and combined SHAP-based feature attribution with deep neural network saliency analysis. Intersection of statistical, biological, and machine-learning evidence identified a compact 18-CpG candidate signature that stratified tumor and normal samples across the analyzed cohorts. Together, this study demonstrates that PDAC methylation remodeling exhibits consistent and reproducible patterns across cohorts that are biologically interpretable. Furthermore, the study shows that integrating chromatin context, network topology, and interpretable machine learning can help identify candidate epigenetic biomarkers with translational potential.
Inayat et al. (Wed,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: