Deploying machine learning and metagenomics for predicting gut microbial drug biotransformation in colorectal cancer.

Abstract

Colorectal cancer (CRC) is one of the leading causes of death globally. Chemotherapy is considered as one of the standard treatment option for CRC. Although effective in many cases, drugs used in chemotherapy are often associated with adverse toxicity. Inter-individual variations in drug response are often attributed to pharmacogenomic associations, and current evidence suggests involvement of gut microbes as a contributor to drug response variations. While several studies have implemented machine learning to predict drug responses mediated by gut microbial enzymes, these predictions mainly rely on substrate information and bacterial genome data. In this study, we have developed CRCGMMeta Predictor, a tool that predicts gut microbial enzymes and species capable of biotransforming drugs. Unlike existing tools, CRCGMMeta Predictor integrates enzyme prediction with species-level abundance in CRC metagenomes using eight geographically diverse metagenomic studies. The tool was trained on 437,079 species-specific enzyme data, incorporating 3457 substrate and protein-encoding features. Multiple machine learning models were evaluated, with the random forest model achieving 99% accuracy, precision, recall, F1 score and ROC-AUC metrics. Validation of the generated model with a blind set achieved 97% accuracy, while 10-fold cross-validation achieved an accuracy of 0.831 ± 0.068. Based on the above results, the random forest model was selected to build the CRCGMMeta Predictor tool. This was further assessed against fluoropyrimidine drugs and its ability to predict enzymes supported by literature evidence and those predicted by existing computational tools. These findings highlight the potential of CRCGMMeta Predictor in identifying microbiome-mediated drug responses.