Optimizing protein-ligand docking through machine learning: algorithm selection with AutoDock Vina

Understanding protein-ligand interactions is fundamental to drug design, where optimizing docking parameter selection can potentially enhance computational efficiency and resource allocation in virtual screening. While numerous algorithms exist for protein-ligand docking, achieving an optimal balance between accuracy and computational speed remains challenging. AutoDock Vina has emerged as a crucial tool in molecular docking, distinguished by its precision, computational efficiency, and adaptability. However, its application in virtual screening across diverse molecular structures presents challenges, particularly in optimizing search parameters. To address these limitations, we developed a machine learning (ML) framework that automates the selection of optimal docking parameters. We developed a comprehensive algorithm set comprising eighty-one distinct configurations of the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm within AutoDock Vina. These configurations were systematically generated by varying two critical parameters: box size (10 Å to 125 Å) and exhaustiveness (8 to 100). The docking process was automated using molecular descriptors and 199 carefully selected substructure fingerprints for each protein-ligand interaction. Our approach employs advanced machine learning techniques, including multi-output regression models (Ridge, Lasso, Random Forest, Gradient Boosting, and XGBoost) to predict optimal docking configurations for novel ligands. Through computational analysis, our Algorithm Selection framework demonstrated superior performance compared to individual docking algorithms while providing valuable insights into the relationship between molecular features and docking outcomes. The feature selection process was optimized using Gini importance metrics, with model performance evaluated through mean squared errors and mean absolute errors.