Does Machine Learning Learn the Physics for Pose Ranking of Fragment-Sized Ligands? A Comparison between Machine Learning and Physics-based Methods

In fragment-based drug discovery using in silico methods, predicting the binding pose is a crucial step to ensure the accurate prediction of binding affinities. Recent studies have focused on the challenges of docking fragments compared to drug-like molecules, with findings suggesting that more sophisticated scoring functions can improve the accuracy of identifying correct binding poses. In this work, we conducted extensive ABFEP (Alchemical Binding Free Energy Perturbation) calculations on a fragment benchmarking dataset to evaluate the accuracy of ABFEP in ranking binding poses of fragments and compared ABFEP rescoring with Vina and two machine learning (ML)-based scoring functions. Indeed, ABFEP, which has a theoretically more rigorous scoring function, significantly outperforms Vina. In-depth comparison between ML-based scoring functions and ABFEP shows that ML-based scoring functions behave similarly to ABFEP on the prediction accuracy and failed cases, indicating that ML is capable of increasing the prediction accuracy over traditional scoring functions through learning the underlying physics rather than memorizing the coordinates in the training data.