A machine learning approach for ranking clusters of docked protein‐protein complexes by pairwise cluster comparison

Reliable identification of near‐native poses of docked protein–protein complexes is still an unsolved problem. The intrinsic heterogeneity of protein–protein interactions is challenging for traditional biophysical or knowledge based potentials and the identification of many false positive binding sites is not unusual. Often, ranking protocols are based on initial clustering of docked poses followed by the application of an energy function to rank each cluster according to its lowest energy member. Here, we present an approach of cluster ranking based not only on one molecular descriptor (e.g., an energy function) but also employing a large number of descriptors that are integrated in a machine learning model, whereby, an extremely randomized tree classifier based on 109 molecular descriptors is trained. The protocol is based on first locally enriching clusters with additional poses, the clusters are then characterized using features describing the distribution of molecular descriptors within the cluster, which are combined into a pairwise cluster comparison model to discriminate near‐native from incorrect clusters. The results show that our approach is able to identify clusters containing near‐native protein–protein complexes. In addition, we present an analysis of the descriptors with respect to their power to discriminate near native from incorrect clusters and how data transformations and recursive feature elimination can improve the ranking performance. Proteins 2017; 85:528–543. © 2016 Wiley Periodicals, Inc.