Designing a Protein-Protein Interface Scoring Function Using Graph Neural Networks

Machine learning approaches have been applied to numerous biological problems, including the scoring of protein-protein interface (PPI) models. Graph neural networks (GNNs) have been identified as an effective tool for PPI scoring due to their ability to represent three-dimensional protein structures as nodes and edges of a graph, preserving rotational invariance without loss of structural information. However, in a comparison study of leading PPI scoring functions, we observed that GNN-based models consistently underperformed compared to non-GNN-based methods. Using a large dataset of unique heterodimers, we demonstrated that the number of heavy atom contacts at the protein interface showed stronger Spearman correlations with the ground truth score, DockQ, than two recently published GNN-based models. We propose that the poor performance of GNN-based scoring functions can be attributed to several factors, including a nonuniform distribution of ground truth scores in GNN training datasets and overgeneralized or non-local node and edge features. To address these issues, we propose a new GNN-based PPI scoring function trained on a dataset that uniformly samples DockQ, with equal representation of near-native and non-native models. We also incorporate local geometric and physical features into our scoring function using contacts defined through Voronoi tessellations, such as the local packing fraction and interface-specific electrostatic and hydrophobic interactions.