Algorithm Selection on Molecular Docking for State-of-the-Art Performance
Molecular docking is a major task for drug discovery and design, with a practical nature. A variety of docking approaches with distinct algorithmic principles have been developed. This study highlights the effectiveness of algorithm selection on molecular docking to outperform the state-of-the-art docking approaches. XGBRegressor, a variant of XGBoost model which is designed for regression problems, is applied to predict and rank the performance of various docking algorithms based on 2D features of proteins and ligands, using the extensive PDBBind 2020 dataset. The model is trained and evaluated with squared error. The predicted results by the our algorithm selection model show promising results that improve the performance of docking by selecting optimal docking algorithms in different cases. The performance is further improved when top 2 and top 3 algorithms predicted by the model are used. The study also finds that building the model with the most important features derived from the original model could simplify the docking process, making it more efficient while still accurate. These all show the effectiveness of using the this algorithm selection model for optimizing molecular docking processes and drug design.
