Machine learning-enabled multi-omics discovery of prognostic biomarkers and signaling targets in pancreatic cancer.

Abstract

Pancreatic ductal adenocarcinoma (PDAC) remains difficult to subtype using single omics layers. We conducted an exploratory investigation integrating reverse-phase protein array (RPPA) and DNA methylation data from the cancer genome atlas (TCGA)- pancreatic adenocarcinoma (PAAD) to assess the feasibility of multi-omics subtyping, alongside a supervised machine learning analysis of a small gene expression omnibus (GEO) transcriptomic cohort (n = 26) to identify candidate diagnostic genes. RPPA-based K-means clustering suggested a weak, possible two-subtype structure (silhouette ≈ 0.16) that remained unassociated with overall survival (log-rank p = 0.113) and lacked independent prognostic value. An independently performed similarity network fusion (SNF) analysis integrating RPPA and methylation data showed low concordance with RPPA-derived subtypes (Adjusted Rand Index (ARI) = 0.014), indicating limited convergence between molecular modalities. Supervised machine learning analysis of the GEO cohort using a fully nested leave-one-out cross-validation pipeline achieved a mean (area under the curve) AUC of 0.896 across four classifiers and identified four-fold-stable candidate genes (ESCO2, COL17A1, BCL2L14, and SOWAHB). However, this gene panel demonstrated limited external validity across two independent PDAC cohorts (log-rank p = 0.438 for both GSE62452 and GSE28735), indicating limited generalizability despite robust internal performance. Collectively, these findings provide limited evidence for a robust, prognostically significant multi-omics subtype or a validated diagnostic gene signature; instead, this study serves as a hypothesis-generating resource and highlights the importance of rigorous cross-validation and independent external validation in small-sample transcriptomic biomarker discovery.