DeepDrugs: a mechanism-aware tri-linear attention framework for synergistic drug-combination prediction.

Abstract

Accurate prediction of drug synergy is critical for the rational design of effective combination therapies against cancer. However, existing computational approaches usually characterize the effect of an individual drug on a cell line separately and then merge the effect representations of two drugs for synergy prediction, which seriously limits their abilities to capture how two drugs act together within a specific cellular environment. We introduce DeepDrugs, a mechanism-aware deep learning framework that employs a tri-linear attention network to directly characterize how two drugs jointly act within a specific cellular context to produce synergy. Extensive experiments demonstrate that DeepDrugs outperforms state-of-the-art approaches in predictive accuracy, robustness, and generalization. Systematic model interpretation analyses identify key pharmacophores that are consistent with experimental validations. Furthermore, DeepDrugs predicts multiple unseen drug combinations (e.g. the Docetaxel-Bortezomib pair in the MCF7 cell line) that align with empirical findings.