Inferring tumor dependency maps from histopathology through imputed transcriptomics

Gene dependency maps from cancer cell lines define context-specific essential genes, but comparable dependencies in tumors remain largely inaccessible. This gap limits the identification of actionable therapeutic vulnerabilities. Here we present a two-stage deep learning framework that transfers cell-line dependency knowledge to tumors by imputing transcriptomic states from routine whole slide images (WSIs). A weakly supervised WSI model first predicts transcriptomic features, which are then mapped to gene dependency to generate tumor dependency landscapes without additional molecular or functional assays. Through systematic interpretation, we connect inferred vulnerabilities to known molecular drivers and to spatially localized histologic evidence. This framework provides an approach for linking histopathology, transcriptomic states, and cell line dependency to support target prioritization and therapeutic hypothesis generation.