More Reliable Binding Affinity Prediction of Protein Ligands Combining Molecular Dynamics Simulations and Machine Learning Models

Abstract

Efficient and accurate prediction of protein-ligand binding affinities is essential for advancing drug discovery. This thesis presents a novel approach that combines molecular docking results with features derived from short molecular dynamics (MD) simulations to enhance the prediction accuracy of binding affinities. The thesis aims to bridge the efficiency of traditional docking methods with the more accurate Free Energy perturbation (FEP) techniques. MD simulations were conducted on a dataset of around 5,000 protein-ligand complexes and extracted features related to binding enthalpy and dynamic behavior related to the more complex binding entropy. These features, alongside molecular docking scores, were used to train machine learning models. The CatBoost regressor was identified as the most effective model, achieving better predictive accuracy than Molecular Docking alone. This method facilitates more reliable binding affinity predictions by successfully integrating docking insights with dynamic protein-ligand behavior, thereby accelerating the early stages of drug discovery. Keywords: