Leveraging Machine Learning and Zinc-Aware Docking to Discover Natural Inhibitors of EptA Resistance Enzymes
Mason Cheng, Mariame Diabate
Affiliation: Saratoga High School
IJSCAR Vol. 3, Issue 1 (2026) · pp. 37–46
Abstract
Antibiotic resistance in multidrug-resistant (MDR) bacteria poses a critical threat to global health contributing to millions of deaths annually. Polymyxins a class of antibiotics have been reintroduced as a last-resort treatment against Gram-negative MDR infections but resistance has spread rendering them ineffective. In light of this problem we employed machine learning and zinc-aware docking to discover natural inhibitors of EptA resistance enzymes a type of enzyme conferring resistance to polymyxins. We curated a training dataset of 2088 compounds derived from known EptA inhibitors and their structural analogs with activity labels generated from docking scores to train a random forest (RF) model. The model achieved 93% accuracy and was applied to screen 5000 compounds from the COCONUT natural product database predicting 166 active compounds of which the top 10 candidates were validated through high-accuracy docking. The top candidates demonstrated strong binding affinity to EptA’s zinc-dependent active site (docking scores: -6.384 to -8.129 kcal/mol) and formed consistent hydrogen bonds with catalytic residues warranting further evaluation through in vitro and in vivo validation.
Keywords: Bacterial resistance, Machine learning, Colistin, EptA, Polymyxin resistance enzymes, Zinc-aware molecular docking, Antibiotic resistance, Polymyxin antibiotics, Zinc-dependent metalloenzymes, Structure-based virtual screening, Gram-negative bacterial infections