A recent article published in Nature Communications by researchers from the Faculty of Medicine at Tel Aviv University reported development of a machine learning (ML) approach that aims to predict how certain drugs or medications would affect intestinal microbes [1].
Many drugs have been shown to have a negative impact on gut bacteria, depleting beneficial species, and simultaneously promoting potentially harmful taxa. The data-driven approach developed in this study successfully predicts outcomes of drug-microbe interactions in in vitro experiments. The authors mapped a large array of interactions between individual medications and gut bacteria and revealed that adverse effects—including diarrhea, bloating, and nausea—of certain pharmaceuticals are clearly correlated with their antimicrobial properties. This computational approach has the potential to advance the development of personalized microbiome-based medicine, one of the most rapidly evolving biological fields of research.
Computational model: training and predictions
The authors developed a random forest computational model that can predict the influence of any drug on any microbial species. This model takes into account 148 microbial features (genomic information) and 92 drug features (physical-chemical properties). The goal of the model is to predict if a particular drug would inhibit microbial growth. The model was trained on a large dataset containing both genomic and chemical information derived from 40 microorganisms and almost 1200 drugs. After vigorous testing and cross-validation, the model was capable of successfully predicting drug–microbe interactions both in vitro and in vivo. The authors demonstrated that “beyond its ability to predict the impact of specific drugs on specific microbes”, this computational approach was able to uncover links between the “anti-commensal properties of drugs and gastrointestinal side effects”.
The researchers made a substantial effort to experimentally verify obtained predictions by analyzing results from large-scale longitudinal studies. This wasn’t an easy task because the current state of experimental drug-microbiome research provides only very limited data on which computational prediction can be experimentally verified. There are several key reasons for this. Firstly, only a fraction of microbes “available” in a computational model can be actually cultured in the lab. Secondly, the number of drug-microbiome interaction studies remains limited; and those studies, the results of which are fully publicly available, are even more scarce. Despite these limitations, the authors were able to demonstrate “a reasonable predictive power” of the model and captured some “real, non-study- or method-specific aspects of drug–microbiome interactions”, validation of some of which will be a goal of future studies.
Experimental science vs. machine learning
This study is a great example of how modern data-driven computational methods and more classical biological approaches can complement each other. High-throughput laboratory experiments provide definitive insights into highly intricate biological systems. To further deepen these insights, ML algorithms use all the available information from all the conducted studies to predict drug-microbe interactions that either haven’t been discovered yet or simply cannot be discovered in the laboratory. Certain interactions cannot be demonstrated in a lab experiment because some microbes (the vast majority, actually) cannot be cultured in the lab—which is where culture-independent computational approaches can fulfill the gap. This is why many researchers, including the authors of the paper discussed here, believe that computational methods are a crucial component of the fundamental research on the interactions between pharmaceuticals and the human microbiome.
Future of drug-microbiome interactions
A clear understanding of complex interactions between pharmaceuticals and gut microorganisms will undoubtedly help the field of precision medicine to have a more direct evolution trajectory. The authors of the present study conclude that being able to successfully predict drug-microbiome interactions in vivo will increase the prediction accuracy of how a certain drug would impact the intestinal microbiomes of whole populations. Several studies have already demonstrated the pivotal role that the human microbiome plays in pharmaceutical treatments [2-4].
The use of the computational approach discussed in this article and other data-driven computational tools can substantially complement these studies by providing crucial information connecting chemical properties of pharmaceuticals and genomic content of gut microorganisms. The future of personalized medicine and tailored microbiome-based therapeutics will largely rely on the use of these computational methods.
This article was written by Alexey Vorobev and edited by Courtney Thomas.
References:
Algavi YM, Borenstein E. A data-driven approach for predicting the impact of drugs on the human microbiome. Nat Commun. 2023;14(1):3614.
Haiser HJ, Gootenberg DB, Chatman K, et al. Predicting and manipulating cardiac drug inactivation by the human gut bacterium Eggerthella lenta. Science. 2013;341(6143):295-298.
Javdan B, Lopez JG, Chankhamjon P, et al. Personalized mapping of drug metabolism by the human gut microbiome. Cell. 2020;181(7):1661-1679 e1622.
Maini Rekdal V, Bess EN, Bisanz JE, et al. Discovery and inhibition of an interspecies gut bacterial pathway for levodopa metabolism. Science. 2019;364(6445).
Comments