Quality assessment of computational models of protein-protein interfaces

Protein-protein interactions are essential in many biological processes ranging from cellular signalling to enzymatic activity. Incorrect interactions such as protein aggregation can result in neurological diseases and other disorders. In this work, we describe novel methods to determine whether computational models of protein-protein interfaces are experimentally accurate or not. We first constructed a dataset of high-quality crystal structures of heterodimers and generated computational models of each complex using state-of-the-art docking methods. We focus on several features of the cores of the interfaces to classify the computational models, e.g. the number, atomic overlap energy, hydrophobicity, and packing fraction of the core residues at the interfaces. Using these features, we find important differences between the interfaces of experimentally determined structures and those in poor-quality computational models, or decoys. Based on these results, we developed a machine-learning classifier that can predict the quality of computational models even for interfaces that have not been studied experimentally.