Improving the scoring of protein-protein docking models by statistical analysis of residue-residue contacts

Resumen del trabajo presentado en el 43th Annual Meeting of the Spanish Society of Biochemistry & Molecular Biology (SEBBM21), celebrado desde Barcelona (España) de forma virtual, del 19 al 22 de julio de 2021Given that the 3D structure of the majority of protein-protein interactions remains elusive, computational molecular docking algorithms is an alternative to structurally characterize protein interactions at large scale. One of the most successful approaches for rigid-body docking and scoring is pyDock (Cheng et al 2007), with excellent results in CAPRI and CASP blind-test assessments (Lensink et al 2019). Despite the success of pyDock energy-based scoring function in rigid, globular proteins forming strong complexes, we need to improve it for challenging cases with large conformational changes or weak complexes. We previously observed that rigid-body docking models selected by pyDock scoring function contained information about the residues that were important for the complex energy (Fernández-Recio et al. 2004), which was applied to successfully predict interface and hot-spot residues (Grosdidier & Fernández-Recio, 2008). Now we hypothesize that the frequency of residue-residue contacts in such docking subsets contains energetics information that can be extracted by statistical analysis to improve modeling. Indeed, a similar approach called CONSRANK has been used to rank docking ensembles based on the most conserved inter-residue contacts (Chermak et al. 2015). Here we have developed a new scoring protocol, pyDockConStat, based on a statistical analysis of residue-residue contacts at the docking interfaces in different sub-sets of models from pyDock. First, the frequencies of contact pairs were computed at different levels of detail. Then, we have used these values (raw frequencies and normalized propensities) to re-score the docking models. The predictive performance has been tested in selected cases from the protein-protein docking benchmark 5.0 (Vreven et al. 2015). In most of the cases, the new contact-based scoring outperforms pyDock scoring function, often showing an excellent correlation with the quality of the docking models