Dose-dependent modeling of combinatorial drug responses stratifies patient survival and reveals therapeutic vulnerabilities in precision oncology

Abstract

A substantial proportion of cancer patients fail to benefit from their prescribed combination regimens, yet identifying superior alternatives from the vast pharmacological space prior to treatment failure remains an unsolved clinical challenge. Existing computational approaches either rely on multi-omics profiles unavailable in standard oncological practice or reduce drug efficacy to scalar metrics that discard the dose-dependent resolution essential for therapeutic optimization. Here, we present XACT, a hierarchical deep learning framework that reconstructs full dose-dependent drug responses for both monotherapy and drug combinations using only clinically accessible transcriptomic profiles. By leveraging an asymmetric X-Linear Attention mechanism that models second-order interactions between molecular drug substructures and intracellular signaling pathway activities, XACT captures concentration-dependent pharmacodynamics with state-of-the-art accuracy and generalizability to unseen transcriptomic landscapes. When applied to the TCGA pan-cancer cohort, XACT-derived resistance scores were significantly associated with clinical treatment outcomes and stratified overall survival as the strongest independent prognostic factor after multivariate adjustment for tumor stage and cancer type. Systematic virtual screening revealed therapeutic vulnerabilities and nominated alternative regimens for treatment-refractory sarcoma and pancreatic adenocarcinoma. These results establish XACT as a scalable, interpretable, and clinically translatable framework that advances precision oncology from computational prediction toward data-driven therapeutic prescription.