HierAffinity: Predicting Protein-Ligand Binding Affinity With Hierarchical Modeling

Predicting protein-ligand binding affinity is of paramount importance in drug discovery. The majority of deep-learning methods attempt to predict from the interactions between proteins and ligands, which, however, encounter several critical issues: 1) Assuming that ligand and protein nodes within a specific distance could form interaction pairs, neglecting their intricate spatial relationships or the limited number each node can accommodate. 2) Disregarding the hierarchical interactions between protein and ligand, which vary in nature based on different distances. 3) Ignoring the hierarchical aggregation process from each interaction to ligand atoms and then to the entire ligand. To address these issues, we propose a framework to predict binding affinity with hierarchical modeling (HierAffinity), including three modules: The first module learns the structural embedding of protein and ligand separately; The second module introduces the interact-KNN method to effectively discern probable interaction pairs between a protein and a ligand. These pairs are then classified into distinct types based on their distance for more representative interaction embedding. The third module comprehends the interaction embedding originating from each ligand node into the whole ligand, considering the interaction distribution and the spatial position of individual ligand nodes. We conducted experiments on multiple datasets and baselines, and the results verified the superiority of our framework and the necessity of each individual module.