Linking machine learning and biophysical structural features in drug discovery.

Introduction: Machine learning methods were applied to analyze pharmacophore features derived from four protein-binding sites, aiming to identify key features associated with ligand-specific protein conformations. Methods: Using molecular dynamics simulations, we generated an ensemble of protein conformations to capture the dynamic nature of their binding sites. By leveraging pharmacophore descriptors, the AI/ML framework prioritized features uniquely associated with ligand-selected conformations, enabling a mechanism-driven understanding of binding interactions. This novel approach integrates biophysical insights with machine learning, focusing on pharmacophoric properties such as charge, hydrogen bonding, hydrophobicity, and aromaticity. Results: Results showed significant enrichment of true positive ligands-improving database enrichment by up to 54-fold compared to random selection-demonstrating the robustness of this approach across diverse proteins. Conclusion: Unlike conventional structure-based or ligand-based screening methods, this work emphasizes the role of specific protein conformations in driving ligand binding, making the process highly interpretable and actionable for drug discovery. The key innovation lies in identifying pharmacophore features tied to conformations selected by ligands, offering a predictive framework for optimizing drug candidates. This study illustrates the potential of combining ML and pharmacophoric analysis to develop intuitive and mechanism-driven tools for lead optimization and rational drug design.