ABSTRACT Developing robust methods for evaluating protein-ligand interactions has been a long-standing problem. Here, we propose a novel approach called EquiScore, which utilizes an equivariant heterogeneous graph neural network to integrate physical prior knowledge and characterize protein-ligand interactions in equivariant geometric space. To improve generalization performance, we constructed a dataset called PDBscreen and designed multiple data augmentation strategies suitable for training scoring methods. We also analyzed potential risks of data leakage in commonly used data-driven modeling processes and proposed a more stringent redundancy removal scheme to alleviate this problem. On two large external test sets, EquiScore outperformed 21 methods across a range of screening performance metrics, and this performance was insensitive to binding pose generation methods. EquiScore also showed good performance on the activity ranking task of a series of structural analogs, indicating its potential to guide lead compound optimization. Finally, we investigated different levels of interpretability of EquiScore, which may provide more insights into structure-based drug design.

EquiScore: A generic protein-ligand interaction scoring method integrating physical prior knowledge with data augmentation modeling