High-throughput identification of bacterial β-glucuronidase inhibitors using machine learning

The human gut microbiome plays a vital role in regulating host physiological functions and influencing the pharmacokinetics of interventions, particularly drug metabolism, which in turn affects pharmacodynamics. Gut microbial β-glucuronidase (GUS) is a key bacterial enzyme that modulates drug therapeutic outcome and gastrointestinal toxicity through deconjugating glucuronidated drug metabolites. Despite this, systematic high-throughput prediction of GUS inhibitors remains limited by sparse experimental data and the translational shortcomings of known compounds. Here, machine learning is applied as a powerful tool to identify potential GUS inhibitors from more than 10,000 FDA-approved drugs, food additives, and excipients. In this study, both unsupervised and supervised machine learning models were trained on literature-derived data describing the inhibitory potency of 122 compounds against Escherichia coli GUS (EcGUS). These models were compared with a newly developed SMILES-based 1D feature-embedded, self-attention classification model (IC-tf) designed for high-throughput screening. To improve interpretability, a dual-level analysis that combines SHAP attribution of handcrafted descriptors with branch-level transformer attention was applied to the IC-tf model. All models demonstrated strong predictive performance, with ROC-AUC values of 85.9%-89.3% under 3-fold cross-validation, with the IC-tf model showing the highest predictive power. In vitro validation with an external set of 20 compounds confirmed strong predictive accuracy for the Random Forest and IC-tf models. This work establishes a scalable computational framework for high-throughput discovery of gut microbial GUS inhibitors, facilitating efficient screening of co-administered drugs and excipients with the translational potential to improve drug bioavailability and reduce gastrointestinal toxicity.