Investigating decoy detection for protein-protein interaction models using state-of-the-art scoring methods and a novel graph neural network

Computational prediction and design of proteins is a difficult task that results in models with a wide variation in quality. Decoy detection algorithms seek to classify computational models as high-quality or low-quality without knowledge of the experimental structures. Recently, dramatic improvements have been made in decoy detection of models for single proteins, but decoy detection of models of protein-protein interfaces (PPI) remains challenging. To assess the current state-of-the-art for PPI decoy detection, we scored computational models generated from RosettaDock, ZDOCK, and HDOCK from a dataset of 32 heterodimeric proteins (with high-resolution x-ray crystal structures) against a standard measure of similarity to the x-ray crystal structure. We found that for some targets, the decoy scores were strongly correlated to the structural similarity scores. However, for other targets nearly all decoy scores were not correlated with the structural similarity scores, which indicates the importance of improving PPI scoring functions. To improve PPI decoy detection, we developed a graph attention neural network model. The model creates a graph using the amino acids as nodes and node features determined using natural language processing on the amino acid sequence. We show results for PPI decoy detection after training the model on the Dockground 1.0 and ZDock decoy datasets, totaling over 170 unique heterodimers.