Combining interface core and whole interface descriptors in postscan processing of protein‐protein docking models

Computational protein‐protein docking scans commonly produce correct and nearly correct interaction models, but sometimes these models are ranked low. We present here postscan processing procedures that dramatically enhance the distinction between nearly correct and false predictions. The procedures employ propensity descriptors calculated for the interface core, an interface‐core clusters count, solvation energy, and a geometric‐electrostatic‐hydrophobic complementarity score. The various descriptors rank high different selections of false models (shuffling effect), and therefore, are used as Boolean yes/no classifiers in “soft intersection” filters, which eliminate large proportions of false models. Furthermore, the standardized descriptors are used in new scoring functions that highlight nearly correct models (NCMs). All the tests are performed on unbound docking models produced with MolFit without use of external data. We find that the discrimination between nearly correct and false models by the various descriptors is class dependent; hence, our postscan processing is class specific. The filters reduce the number of putative models from 10,726, 12,517, and 11,054 to 758, 157, and 1218 for enzyme‐inhibitor, antibody‐antigen, and nonclassified systems. When combined with the new scoring functions, they improve the average rank of the highest ranking NCMs from 673 to 122. Application to 23 CAPRI targets demonstrates the effectiveness of the postscan procedures in cases where external information is used in the production of the putative models. Our new per‐molecule residue propensity descriptors show that interacting interfaces are enriched with high propensity residues except for antigenic sites, which resemble more the noninteracting regions of protein surfaces. Proteins 2009. © 2009 Wiley‐Liss, Inc.