Assessing the current state of computational models for protein-protein interfaces

Computational prediction and design of protein-protein interfaces (PPI) is a difficult task that results in models with a wide range of quality. Thus, PPI scoring algorithms have been designed to classify computational models as high- or low-quality without direct knowledge of the experimentally determined structures. To assess PPI scoring algorithms, we scored models of heterodimeric proteins generated with a rigid-body docking algorithm (ZDOCK) and compared these scores to a measure of structural similarity (i.e. DockQ score) to the associated heterodimeric x-ray crystal structures. We generated hundreds of thousands of computational models, sampled uniformly over the DockQ score, to create a balanced dataset of redocked models. We scored this dataset using seven state-of-the-art PPI model scoring algorithms. We show that only a small fraction of PPI targets possesses strong correlations (ρ > 0.8) between the model scores and DockQ. We use three physical features to differentiate easy- and hard-to-score targets: interface flatness, fractional interface surface area, and number of interfacial contacts. We show that there are strong correlations between model and DockQ scores for PPI targets with large, intertwined interfaces, yet weak correlations for targets with small, flat interfaces. We create a support vector regression score using these physical features that matches or exceeds the performance of current state-of-the-art PPI scoring functions.