A feature-based approach to modeling protein–protein interaction hot spots

Abstract

Identifying features that effectively represent theenergetic contribution of an individual interfaceresidue to the interactions between proteinsremains problematic. Here, we present severalnew features and show that they are more effectivethan conventional features. By combining the proposedfeatures with conventional features, wedevelop a predictive model for interaction hotspots. Initially, 54 multifaceted features, composedof different levels of information including structure,sequence and molecular interaction information,are quantified. Then, to identify the bestsubset of features for predicting hot spots, featureselection is performed using a decision tree. Basedon the selected features, a predictive model for hotspots is created using support vector machine(SVM) and tested on an independent test set. Ourmodel shows better overall predictive accuracythan previous methods such as the alanine scanningmethods Robetta and FOLDEF, and theknowledge-based method KFC. Subsequent analysisyields several findings about hot spots. Asexpected, hot spots have a larger relative surfacearea burial and are more hydrophobic than otherresidues. Unexpectedly, however, residue conservationdisplays a rather complicated tendencydepending on the types of protein complexes, indicatingthat this feature is not good for identifyinghot spots. Of the selected features, the weightedatomic packing density, relative surface area burialand weighted hydrophobicity are the top 3, withthe weighted atomic packing density proving tobe the most effective feature for predictinghot spots. Notably, we find that hot spots areclosely related to n–related interactions, especiallyn n interactions.