An interpretable flux-based machine learning model of drug interactions across metabolic space and time
Drug combinations are a promising strategy to counter antibiotic resistance. However, current experimental and computational approaches do not account for the entire complexity involved in combination therapy design, such as the effect of the growth environment, drug order, and time interval. To address these limitations, we present an approach that uses genome-scale metabolic modeling and machine learning to explain and guide combination therapy design. Our approach (a) accommodates diverse data types, (b) accurately predicts drug interactions in various growth conditions, (c) accounts for time- and order-specific interactions, and (d) identifies mechanistic factors driving drug interactions. The entropy in bacterial stress response, time between treatments, and gluconeogenesis activation were the most predictive features of combination therapy outcomes across time scales and growth conditions. Analysis of the vast landscape of condition-specific drug interactions revealed promising new drug combinations and a tradeoff in the efficacy between simultaneous and sequential combination therapies.