TransPocket: Structural and Geometric Transformer for Ligand Binding Site Detection

The detection of Ligand binding sites is foundational to drug discovery. Most current methods initially voxelize proteins by transforming them into 3D images, and subsequently using CNN models to extract features for prediction. Nevertheless, these CNN-based methods are fraught with several critical issues: 1) ignore the complexity of protein structures; 2) susceptible to rotation; 3) deficient in handling global and long-range geometric information on a protein. In response to these issues, we introduce a novel Transformer framework, TransPocket, which harnesses both structural and geometric features to predict binding sites. Specifically, our Transformer is composed of two integrated modules: 1) Protein Structure Modeling Module, designed to extract intricate structure information in a protein through our proposed Two-Stage EGNN model while maintaining rotational invariance/equivariance. 2) Protein Geometric Modeling Module, crafted to learn short- and long-range geometric features of a protein utilizing proposed Transformer-Unet model. The experimental results on multiple datasets demonstrate that our model either matches or exceeds the performance of the state-of-the-art, while also validating the efficacy of the individual modules.