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

Area of Science:

• Pharmacology
• Microbiology
• Computational Chemistry

Background:

• The gut microbiome influences drug metabolism and toxicity.
• Gut microbial β-glucuronidase (GUS) deconjugates drug metabolites, affecting therapeutic outcomes and gastrointestinal toxicity.
• Limited data and translational issues hinder the discovery of GUS inhibitors.

Purpose of the Study:

• To apply machine learning for high-throughput identification of potential GUS inhibitors.
• To screen FDA-approved drugs, food additives, and excipients for GUS inhibitory activity.
• To develop a scalable computational framework for discovering GUS inhibitors.

Main Methods:

• Trained unsupervised and supervised machine learning models on 122 compounds' inhibitory data against *Escherichia coli* GUS (EcGUS).
• Developed and compared a novel SMILES-based 1D feature-embedded, self-attention classification model (IC-tf).
• Employed SHAP attribution and transformer attention for model interpretability.

Main Results:

• Models achieved strong predictive performance (ROC-AUC 85.9%-89.3%) with the IC-tf model showing the highest accuracy.
• *In vitro* validation confirmed high predictive accuracy for Random Forest and IC-tf models.
• Identified potential GUS inhibitors from a large dataset of approved drugs, additives, and excipients.

Conclusions:

• Established a scalable computational framework for discovering gut microbial GUS inhibitors.
• The framework facilitates efficient screening of co-administered drugs and excipients.
• Potential to improve drug bioavailability and reduce gastrointestinal toxicity.