A multi-geometric graph fusion network for protein-ligand affinity prediction

Protein-ligand binding affinity prediction plays a crucial role in drug discovery. While recent works use two-dimensional graph neural networks to improve affinity prediction, we find that the three-dimensional geometric information of proteins and ligands can significantly affect their binding mechanisms. However, existing studies have not thoroughly explored the three-dimensional spatial conformations of proteins and ligands. In this paper, we introduce MGGNet, a multi-geometric graph fusion model based on geometric graph neural networks. MGGNet captures the atomic interactions and spatial conformations of protein-ligand complexes by leveraging the three-dimensional structural data. Specifically, we employ a heterogeneous network for the ligand and protein pocket regions to capture non-covalent interactions and binding conformations, while constructing separate homogeneous networks with 3D message passing for each. By incorporating geometric features from multiple coordinate systems, MGGNet effectively learns covalent interactions and 3D spatial conformations, ensuring invariance to spatial transformations. Experimental results on publicly available benchmark datasets demonstrate that MGGNet outperforms state-of-the-art methods. Additionally, visualizations of the binding heatmaps of ligands further highlight the biological relevance of MGGNet's predictions for protein-ligand binding.