Multi-task learning predicts drug combination synergy in cells and in the clinic

Abstract

Combination therapies for various cancers have been shown to increase efficacy, lower toxicity, and circumvent resistance. However, despite the promise of combinatorial therapies, the biological mechanisms behind drug synergy have not been fully characterized, and the systematic testing of all possible synergistic therapies is experimentally infeasible due to the sheer volume of potential combinations. Here we apply a novel big data approach in the evaluation and prediction of drug synergy by using the recently released NCI-ALMANAC. We found that each traditional drug synergy metric (Bliss, Loewe, ZIP, HSA, ALMANAC Score) identified unique synergistic drug pairs with distinct underlying joint mechanisms of action. Leveraging these findings, we developed a suite of context specific drug synergy predictive models for each distinct synergy type and achieved significant predictive performance (AUC = 0.89-0.953). Furthermore, our models accurately identified clinically tested drug pairs and characterized the clinical relevance of each drug synergy metric, with Bliss Independence capturing clinically tested combinations best. Our findings demonstrate that drug synergy can be obtained from numerous unique joint mechanisms of action, captured by different synergy metrics. Additionally, we show that drug synergy, of all kinds, can be predicted with high degrees of accuracy with significant clinical potential. This breakthrough understanding of joint mechanisms of action will allow for the design of rational combinatorial therapeutics on a large scale, across various cancer types.