(A-10) Predicting the Effects of Drugs on Commensal Bacteria with Machine Learning

Introduction:: The human microbiome is increasingly recognized as a major determinant of human health, but its composition and metabolism might be impacted by various pharmaceuticals taken by the human hosts [1,2]. For example, a high throughput screen reported 24% of human targeted drugs inhibited the growth of at least one of the representative strains while patient cohort studies estimated up to 50% of non-antibiotic drugs altered the microbiome composition or its metabolic potential [1,2]. For example, drugs like the antipsychotic risperidone negatively impacts the microbiome by shifting the composition to an increased Firmicutes to Bacteroides ratio and is associated with weight gain in treated mice [3]. Conversely, there are positive modulators like diuretics and beta blockers that promote abundance of anti-inflammatory strains like Roseburia in the gut microbiome [4,5]. Taken together, microbiome-drug interactions can adversely impact the composition of the microbiome but can also promote the microbiome’s diversity to improve host health. Therefore, it is important to understand such effects for drug development and microbiome engineering. However, current methods to study drug-microbiome interactions are limited to patient data analysis or high throughput screens which are laborious and resource intensive. There is a significant unmet need for tools that rapidly and accurately predict drug-microbiome interactions. Here, we prototype and validate machine learning algorithms trained on available experimental and biochemical data to predict drug-microbiome interactions rapidly and accurately.


Materials and Methods:: We extracted data from high-throughput experiments that tested 1197 drugs for inhibition of growth of 40 commensal bacteria strains in isolated cultures [1,6]. Molecules were described using molecular fingerprints and annotated as ‘active’ or ‘inactive’ if they significantly lower the OD595 in bacterial monocultures after 24h growth when incubated with the drug. The dataset was balanced by removing molecules with lowest reported activity against bacterial strains. We trained 81 different machine learning models and evaluated them using 5x5-fold cross validation to identify the top performing model. Of the top performing models, Gradient Boosting with Feature Morgan fingerprints, was used to predict the potential effects of 1485 investigational drugs (DrugBank) on 40 gut commensal bacteria. We validated our top predictions by monoculture experiments in B. thetaiotaomicron, A. muciniphila, P. copri, R. gnavus and E. coli. We grew monocultures overnight by inoculating each strain in 5ml of pre-reduced BHI in an anaerobic chamber at 37°C. After overnight inoculation, the cultures were adjusted to a total OD of 0.02 with freshly pre-reduced BHI. Drugs were dissolved in DMSO and added to the bacteria in at least three technical replicates at final concentrations of 20µM in 1% DMSO. We then monitored growth at 37°C for 24h at OD595nm compared to vehicle control, antibiotic treatment, and medium blank.

Results, Conclusions, and Discussions:: We evaluated multiple machine learning models for their ability to predict which molecules would modulate the growth of commensal bacteria from their chemical structure (Figure 1A). Gradient Boosting with Feature Morgan fingerprints was the most accurate model, with mean Cohen’s kappa values reaching 0.440 in cross validation experiments. This model was used to predict the effects of 1485 investigational drugs on 40 gut commensals bacteria (Figure 1B). Overall, 40% of the drugs are predicted to inhibit the growth of multiple commensals. We focused our experimental validation on B. thetaiotaomicron as a model organism given its prevalence in human microbiomes. The top drugs predicted to inhibit B. thetaiotaomicron growth were Entrectinib (P = 0.997) and PSI-697 (P = 0.959). We experimentally validated that Entrectinib (IC50 = 7.6µM) and PSI-697 (IC50 = 6.9µM) potently inhibited B. thetaiotaomicron growth in vitro in monocultures (Figure 1C). Encouraged by these findings, we tested whether the predictions for other strains were also correct for these drugs. As predicted, Entrectinib inhibited the growth of A. muciniphila, P. copri, and R. gnavus while PSI-697 inhibited the growth of P. copri and R. gnavus in monocultures (Figure 1D). Also as predicted, E. coli K12 was resistant to these drugs and its growth was not affected (Figure 1E). To test why E. coli K12 is resistant to these drugs, we tested our drugs in E. coli K12 efflux knockout for BAM B and TOL C efflux proteins, E. coli BW25113 ΔbamBΔtolC and observed that the knockout strain was sensitive to these drugs – indicating that drug efflux might explain the resistance of E. coli against these drugs (Figure 1E).

Our study highlights how machine learning can support the identification of drug-microbiome interactions rapidly and accurately. In the future, we will validate the relevance of the identified effects in complex mixtures of microbes and in vivo. Already, our machine learning algorithms stand to support drug development and microbiome engineering by identifying specific microbes that are affected by small molecular drugs.

Acknowledgements (Optional): : We would like to thank the Duke Biomedical Engineering Department and Duke Microbiome Center for the opportunity to perform this research. We also thank all members of the Reker, Rawls, You, Valdivia, Lawrence and Surana Lab for helpful discussions, feedback, and continued support. We acknowledge the Duke Biomedical Engineering Fellowship (Wanda K. Neu Fellowship) for support.


References (Optional): : 1. Maier L, Pruteanu M, Kuhn M, et al. Extensive impact of non-antibiotic drugs on human gut bacteria Europe PMC Funders Group. Nature. 2018;555(7698):623-628. doi:10.6084/m9.figshare.4813882
2. Vich Vila A, Collij V, Sanna S, et al. Impact of commonly used drugs on the composition and metabolic function of the gut microbiota. Nat Commun. 2020;11(1). doi:10.1038/s41467-019-14177-z
3. Bahr SM, Weidemann BJ, Castro AN, et al. Risperidone-induced weight gain is mediated through shifts in the gut microbiome and suppression of energy expenditure. EBioMedicine. 2015;2(11):1725-1734. doi:10.1016/j.ebiom.2015.10.018
4. Ley RE, Bä Ckhed F, Turnbaugh P, Lozupone CA, Knight RD, Gordon JI. Obesity Alters Gut Microbial Ecology.; 2005. www.arb-home.de
5. Forslund SK, Chakaroun R, Zimmermann-Kogadeeva M, et al. Combinatorial, additive and dose-dependent drug–microbiome associations. Nature. 2021;600(7889):500-505. doi:10.1038/s41586-021-04177-9
6. McCoubrey LE, Elbadawi M, Orlu M, Gaisford S, Basit AW. Machine learning uncovers adverse drug effects on intestinal bacteria. Pharmaceutics. 2021;13(7). doi:10.3390/pharmaceutics13071026