Integration of metabolic modelling with machine learning to identify mechanisms underlying antibiotic killing

Abstract

Microbial pathogens are becoming a pressing global health issue due to the rapid appearance of resistant strains, accompanied by slow development of new antibiotics. In order to improve these treatments and engineer novel therapies, it is crucial that we increase our understanding of how these antibiotics interact with cellular metabolism. Evidence is increasingly building that the efficacy of antibiotics relies critically on downstream metabolic effects, in addition to inhibition of primary targets. Here we present a novel computational pipeline to expedite investigation of these effects: we combine computational modelling of metabolic networks with data from experimental screens on antibiotic susceptibility to identify metabolic vulnerabilities that can enhance antibiotic efficacy. This approach utilizes genome-scale metabolic models of bacterial metabolism to simulate the reaction-level response of cellular metabolism to a metabolite counter screen. The simulated results are then integrated with experimentally determined antibiotic sensitivity measurements using machine learning. Following integration, a mechanistic understanding of the phenotype-level antibiotic sensitivity results can be extracted. These mechanisms further support the role of metabolism in the mechanism of action of antibiotic lethality. Consistent with current understanding, application of the pipeline to M. tuberculosis identified cysteine metabolism, ATP synthase, and the citric acid cycle as key pathways in determining antibiotic efficacy. Additionally, roles for metabolism of aromatic amino acids and biosynthesis of polyprenoids were identified as pathways meriting further investigation.