A flux-based machine learning model to simulate the impact of pathogen metabolic heterogeneity on drug interactions

A bstract 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 pathogen metabolic heterogeneity, changes in the growth environment, drug treatment order and time interval. To address these limitations, we present a comprehensive approach that uses genome-scale metabolic modeling and machine learning to guide combination therapy design. Our mechanistic approach (a) accommodates diverse data types, (b) accounts for time- and order-specific interactions, and (c) accurately predicts drug interactions in various growth conditions and their robustness to pathogen metabolic heterogeneity. Our approach achieved high accuracy (AUC = 0.83 for synergy, AUC = 0.98 for antagonism) in predicting drug interactions for E. coli cultured in 57 metabolic conditions based on experimental validation. The entropy in bacterial metabolic response was predictive of combination therapy outcomes across time scales and growth conditions. Simulation of metabolic heterogeneity using population FBA identified two sub-populations of E. coli cells defined by the levels of three proteins (eno, fadB and fabD) in glycolysis and lipid metabolism that influence cell tolerance to a broad range of antibiotic combinations. Analysis of the vast landscape of condition-specific drug interactions revealed a set of 24 robustly synergistic drug combinations with potential for clinical use. S ignificance Worldwide, 700,000 people die each year from drug-resistant infections. Drug combinations have great potential to reduce the spread of drug-resistant bacteria. However, their potency is impacted by both the pathogen growth environment and the heterogeneity in pathogen metabolism. The metabolic heterogeneity in a pathogen population allows them to survive antibiotic treatment. Here we present a flexible machine-learning framework that utilizes diverse data types to effectively search through the large design space of both sequential and simultaneous combination therapies across hundreds of simulated growth conditions and pathogen metabolic states. Our approach can serve as a useful guide for the selection of robustly synergistic drug combinations.