Machine Learning Guided Prediction of Drug Microbiome Interactions

Abstract

The human gut microbiome plays essential roles in metabolism, immunity, and energy homeostasis. Disruption of this microbial ecosystem has been implicated in conditions ranging from inflammatory bowel disease to metabolic syndrome. While antibiotics are recognized disruptors of the gut microbiome, recent evidence indicates that approximately one-quarter of human-targeted, non-antibiotic drugs also inhibit commensal bacteria in vitro. Importantly, drug-induced microbiome alterations have documented clinical consequences: certain non-antibiotic drugs increase susceptibility to invasive bacterial infections, impair drug efficacy, and contribute to common adverse effects including gastrointestinal symptoms and metabolic dysfunction. Large population cohorts confirm that drugs spanning multiple therapeutic classes such as proton pump inhibitors, statins and beta-blockers significantly alter commensal composition in humans. However, systematically characterizing drug–microbiome interactions through experimental screening alone is impractical given the scale of the pharmaceutical landscape. This dissertation addresses the problem of efficiently identifying drugs with unintended antibacterial effects on gut commensals and on specific bacterial metabolic enzymes, providing a computational framework for early-stage microbiome safety assessment in drug discovery.We first conducted a descriptive statistical analysis comparing the physicochemical properties of bacteria-affecting and non-affecting drugs to establish a rationale for ligand-based prediction. Bacteria-affecting drugs exhibited significantly higher lipophilicity (mean LogP 3.46 (SD=2.39) vs. 2.18 (SD=2.01); Cohen's d = 0.61) and greater aromatic ring count (mean 2.05 (SD=1.14) vs. 1.35 (SD=0.96); Cohen's d = 0.70) compared to non-affecting drugs. Highly lipophilic drugs (LogP > 3) showed a 2.55-fold enrichment in antibacterial activity. These medium-to-large effect sizes indicated that molecular structure contains predictive information about antibacterial potential, motivating the development of machine learning models trained on structural features. To address this challenge, we developed and evaluated machine learning approaches that predict drug–microbiome interactions directly from molecular structure. We trained 81 computational pipelines, combining nine machine learning algorithms with nine molecular fingerprint representations, using published high-throughput screening data on 1200 drugs tested against 40 representative gut bacterial strains. Model performance was assessed through stratified cross-validation using metrics appropriate for imbalanced classification, including balanced accuracy, area under the receiver operating characteristic curve, and Matthew’s correlation coefficient. Prospective predictions were validated experimentally using in vitro monoculture growth inhibition assays, efflux pump knockout studies in Escherichia coli, and synthetic microbial community experiments. Tree-based ensemble methods paired with pharmacophore-aware fingerprints achieved the strongest predictive performance. Prospective application to 1,485 investigational drugs predicted that a substantial fraction may affect commensal growth. Experimental validation confirmed that entrectinib, an approved tyrosine kinase inhibitor, and PSI-697, an investigational P-selectin inhibitor, potently inhibit multiple commensal species including Bacteroides thetaiotaomicron, Akkermansia muciniphila, Prevotella copri, and Ruminococcus gnavus. Both compounds were identified as substrates of bacterial efflux pumps BamB and TolC, suggesting potential contributions to antimicrobial resistance selection. Entrectinib exposure reduced richness in an eight-member synthetic gut community. In a parallel investigation, we applied a knowledge transfer strategy to predict inhibitors of bacterial acetohydroxyacid synthase (AHAS), leveraging structural conservation between enzyme orthologs and validating predictions using a whole-cell colorimetric enzyme activity assay. The knowledge transfer approach identified pazopanib as a novel AHAS inhibitor in Phocaeicola vulgatus, experimentally confirmed through biochemical assay. While false positive predictions were observed in both applications, and the in vitro findings require further validation to establish clinical relevance, this work demonstrates that structure-based machine learning can identify drugs with previously unrecognized effects on gut bacteria and bacterial enzymes, providing a scalable approach for early-stage assessment of a drug like molecule’s microbiome liability. The validated predictions while preliminary suggest that computational screening could complement experimental methods in characterizing drug–microbiome interactions and inform microbiome-aware drug development.