Gene-structured histology for deriving and predicting pancreatic cancer molecular subtypes

Abstract

into basal-like and classical states is a critical prognostic determinant, yet clinical implementation remains limited by the cost and turnaround time of transcriptomic sequencing. 1 – 3 Although routine histopathology captures rich morphological features, deep learning models often lack a principled connection to gene-level molecular structure. 4,5 We propose a graph-constrained histology model that maps morphology-derived latent features onto a fi xed, data-driven gene co-expression network for pancreatic cancer molecular subtype prediction. The gene-structured outputs are interpreted as latent features constrained by gene co-expression structure, rather than as direct estimates of patient-level gene expression or empirically recovered gene-network alignment. Our model uses a multi-stage gene sampling work fl ow designed to identify biologically informative genes from transcriptomic data for gene discovery. Using bulk RNA-seq data from 797 patients across TCGA-PAAD and PANCAN cohorts, we established molecular ground truths via single-sample gene enrichment analysis (ssGSEA). A hierarchical Monte Carlo screening process evaluated candidate genes to derive 50-gene modules with high predictive potential. In fi ve-fold cross-validation, the Stage 2 screening of 200-gene modules achieved a mean test AUC of 0.773 ± 0.026. The fi nal selected module contained predominantly protein-coding genes, with a smaller number of non-coding transcripts including snoRNAs and long non-coding RNAs shown in Fig. 1a. Gene Ontology analysis and functional enrichment analysis are shown in Fig. 1b for the derived 50 gene module for the fi rst run. The subsequent Stage 3 optimization of the 50-gene module achieved a mean test AUC of 0.841 ± 0.026, indicating improved predictive performance after focused gene selection. The top genes for all Montecarlo runs for the highest performing gene modules are