Gene dependency-informed inference of response to targeted cancer therapies.

Abstract

Targeted therapies such as small-molecule inhibitors act by blocking proteins essential for cancer cell survival, yet omics-based modeling of drug sensitivity often lacks mechanistic grounding. We present FORGE (Factorization Of Response and Gene Essentiality), a joint matrix factorization framework that co-models drug response and target gene dependency to enable biologically informed stratification of treatment groups. FORGE derives a Benefit Score from basal gene expression to estimate therapeutic potential. In unseen cell lines treated with erlotinib, FORGE achieves high concordance for dependency (0.69) and IC₅₀ (0.62), with benefit score stratification showing increased dependency and decreased IC₅₀ across quartiles. Joint modeling improves predictive performance over single-task approaches and enhances agreement between gene-level effects (p = 0.039). Validation across independent datasets shows that higher benefit scores associate with tumor regression in patient-derived xenografts and with predicted drug susceptibility in the Tahoe-100M dataset. Mechanistic analyses further identify gene programs underlying drug susceptibility.