Abstract

The prediction of protein–protein interactions based on independently obtained structural information for each interacting partner remains an important challenge in computational chemistry. Procedures where hypothetical interaction models (or decoys) are generated, then ranked using a biochemically relevant scoring function have been garnering interest as an avenue for addressing such challenges. The program PatchDock has been shown to produce reasonable decoys for modeling the association between pig alpha-amylase and the VH-domains of camelide antibody raised against it. We designed a biochemically relevant method by which PatchDock decoys could be ranked in order to separate near-native structures from false positives. Several thousand steps of energy minimization were used to simulate induced fit within the otherwise rigid decoys and to rank them. We applied a partial free energy function to rank each of the binding modes, improving discrimination between near-native structures and false positives. Sorting decoys according to strain energy increased the proportion of near-native decoys near the bottom of the ranked list. Additionally, we propose a novel method which utilizes regression analysis for the selection of minimization convergence criteria and provides approximation of the partial free energy function as the number of algorithmic steps approaches infinity.