Predicting cancer survival by connecting tissue patterns to genes

ProtoPathway combines whole-slide histopathology images and gene expression data to predict cancer survival with built-in biological interpretability. On the imaging side, learnable morphological prototypes compress variable-length tissue patches into fixed representations; on the genomics side, a graph network encodes genes within their known biological pathways (Reactome). Cross-modal attention connects the two, revealing which tissue patterns correlate with which molecular programs. Tested on five TCGA cancer cohorts, it matches or beats existing methods while cutting computation and providing interpretable pathway-level attribution from genes to tissue.