Explainable Artificial Intelligence Reveals Potential Candidate Mechanism of Strain-Specific Drug Depletion

Oral medications can be bioaccumulated or metabolised by gastrointestinal bacteria in a process collectively termed drug depletion. The precise biological mechanisms governing strain-specific depletion remain poorly understood, and systematic experimental classification of drug-strain interactions via in vitro studies is both costly and time-consuming. In this study, artificial intelligence (AI) methodologies combining machine learning (ML) and natural language processing (NLP) were applied to predict strain-specific drug depletion. The dataset comprised 16,802 drug-strain interaction pairs, with drugs represented by physicochemical descriptors and bacterial strains represented by whole-genome sequences. NLP techniques were used to transform genomic data into feature representations suitable for ML model training. The resulting models achieved strong predictive performance, with a balanced accuracy of 0.90 {+/-} 0.02 and Matthews correlation coefficient of 0.54 {+/-} 0.10. Feature importance analysis revealed that both drug properties and genomic features contributed to model predictions. Among the highest-ranking genomic features, BLASTX annotation identified several enzymes with known or plausible roles in drug metabolism. To further explore the mechanistic relevance of these features, two candidate enzymes were selected for molecular docking against drugs experimentally observed to be depleted. Glycosidase was found to possess binding energies of -8.69 and -7.88 kcal/mol for the two cardiac glycoside drugs digitoxin and digoxin, respectively; whereas acetyl-CoA carboxylase biotin carboxylase presented with binding energies for between -7.09 and -7.74 kcal/mol at one of its druggable sites. Collectively, these findings establish a proof-of-concept AI-driven framework that integrates predictive performance with mechanistic interpretability in the study of drug-microbiome interactions. The broader implications and limitations of applying AI in this context are also discussed. These preliminary findings offer a promising strategy for accelerating drug developments through using AI to rapidly highlight potential drug interactions. Graphical Abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=135 SRC="FIGDIR/small/701358v1_ufig1.gif" ALT="Figure 1"> View larger version (34K): org.highwire.dtl.DTLVardef@16cf66corg.highwire.dtl.DTLVardef@a640acorg.highwire.dtl.DTLVardef@e00ecdorg.highwire.dtl.DTLVardef@1ebbd62_HPS_FORMAT_FIGEXP M_FIG C_FIG