Machine learning–guided virtual screening and docking analysis of penicillin-binding protein 2a inhibitors

Background Methicillin-resistant Staphylococcus aureus (MRSA) remains a major global health threat due to its resistance to β-lactam antibiotics, primarily mediated by penicillin-binding protein 2a (PBP2a). Identification of effective PBP2a inhibitors is therefore essential for the development of alternative anti-MRSA therapeutic strategies. Methods In this study, an integrated computational workflow combining quantitative structure–activity relationship (QSAR) modeling and molecular docking was developed to identify computationally prioritized experimentally reported PBP2a inhibitors. Experimentally reported IC₅₀ data were retrieved from the ChEMBL database, curated, standardized, and converted to the pIC₅₀ scale. Molecular representations were generated using 2048-bit Morgan fingerprints. A Random Forest regression model was optimized and evaluated using a chemically aware cluster-based external validation strategy to minimize information leakage from structurally similar analogues. Model robustness was further assessed through cross-validation, residual analysis, applicability domain evaluation, and Y-randomization using 100 permutations. Molecular docking was performed using AutoDock Vina against the allosteric site of PBP2a, and the docking protocol was validated through redocking of the co-crystallized ligand from PDB structure 4CJN. Results The optimized Random Forest model achieved an external test-set R² of 0.5248, with an RMSE of 0.7385 and MAE of 0.5022 under cluster-based validation conditions. Y-randomization analysis produced near-zero or negative R² values for the randomized model, confirming that the observed predictive performance was not due to chance correlations. Applicability domain analysis identified only ten compounds that fell outside the defined applicability domain. Mapping of influential Morgan fingerprint bits to representative substructures revealed that hydroxyl-rich fragments, aromatic systems, and heteroatom-containing linkages were strongly associated with predicted activity. Docking validation yielded a heavy-atom RMSD of 0.924 Å between the crystallographic and redocked ligand poses, indicating good agreement with the experimental binding mode. Several compounds exhibited favorable predicted activity and docking scores within the defined applicability domain. Conclusion This study presents a chemically interpretable and statistically validated computational framework for computationally prioritizing experimentally reported PBP2a inhibitors. The integration of cluster-based QSAR modeling, substructure-level interpretation, applicability-domain analysis, and validated molecular docking provides a robust strategy for guiding future anti-MRSA drug discovery and experimental validation efforts.