Integrated virtual screening, machine learning and molecular dynamics identify novel phytochemical FabI inhibitors...

Methicillin-resistant Staphylococcus aureus (MRSA) infections pose a critical global health threat, necessitating the discovery of novel antibacterial agents targeting essential bacterial pathways. Enoyl-acyl carrier protein reductase (FabI) represents a validated antimicrobial target due to its essential role in bacterial fatty acid biosynthesis and absence in mammalian cells. This study aimed to identify novel polyphenolic and related phenolic FabI inhibitors from natural product databases through an integrated computational approach combining structure-based virtual screening, machine learning (ML) classification, explainable artificial intelligence (XAI) analysis, molecular dynamics simulations, and binding free energy calculations. A library of 41,423 drug-like polyphenolic and related phenolic compounds from the COCONUT database was screened against S. aureus FabI (PDB: 4BNF) using AutoDock Vina, and ten ML algorithms were trained on ECFP4 molecular fingerprints to classify compounds as active or inactive. Top candidates selected through XAI-guided criteria were validated using 100 ns molecular dynamics simulations, followed by MM-GBSA binding free energy calculations, principal component analysis (PCA), and free energy landscape (FEL) analysis. Virtual screening identified 26,097 compounds (63%) with superior binding affinity compared to the reference inhibitor PV4 (-8.21 kcal/mol), with coumarins dominating the top hits (56%) followed by flavonoids (20%). The SVM-RBF classifier achieved optimal performance with ROC-AUC of 0.9894 and Matthews Correlation Coefficient of 0.9017. Feature importance analysis revealed phenolic hydroxyl/ether groups (40% of top features) and alkyl chains (30%) as critical structural determinants for FabI inhibition. N-(1H-benzimidazol-2-yl)-2-(2-oxo-4-phenyl-chromen-7-yl)oxy-acetamide (CNP0408084) emerged as the most promising lead compound, exhibiting high binding affinity (-12.57 kcal/mol), maximum coverage of ML-important features (55%), and superior binding stability (RMSD: 1.79 ± 0.32 Å) compared to PV4 (2.26 ± 0.42 Å). MM-GBSA calculations confirmed the superior binding thermodynamics of CNP0408084 (ΔG_bind = -38.72 ± 3.85 kcal/mol) compared to PV4 (-24.56 ± 2.94 kcal/mol), with van der Waals interactions as the dominant driving force. PCA revealed that CNP0408084 constrains FabI conformational dynamics with the most compact distribution, while FEL analysis demonstrated a single deep energy minimum characteristic of stable, tight-binding inhibitors. This study demonstrates the utility of integrating ML, explainable AI, molecular dynamics, and advanced post-MD analyses for natural product-based drug discovery, providing a generalizable framework for prioritizing virtual screening hits based on mechanistically relevant structural features.