RefineScore: Improving Ligand Docking Accuracy and Interpretability by Predicting MDN Corrective Physical Interactions

Precisely evaluating the protein-ligand interaction is crucial in drug screening and optimization. There are significant advances in the application of machine learning approaches to developing protein-ligand scoring functions in recent years. However, traditional docking softwares and existing deep-learning methods remain unsolved limitations in terms of pose quality and binding affinity prediction accuracy. Furthermore, machine learning-based scoring functions are hard to generalize to unseen cases due to the scarcity of structure-affinity data, and mostly lack physical interpretability. In this study, we propose RefineScore, which learns the underlying physical laws by the neural network parameterized physical energy terms. In addition, mixed density network (MDN) is introduced to predict the distance likelihoods of van der Waals (vdW) and hydrogen bonding (Hbond) pairs and further correct the classical energy equations for the absolute binding free energy calculation. RefineScore provides a versatile atomic interaction evaluation framework within protein-ligand complexes and is highly interpretable for the key interaction analysis and energy contribution decomposition. By integrating these carefully tailored techniques, RefineScore maintains the advantages of machine learning models while providing physical interpretability and capturing complex interaction patterns. Our evaluation results show that RefineScore has superior scoring and ranking power.