Information of Binding Sites Improves Prediction of Protein-Protein Interaction

Protein-protein interaction is essential to cellular functions. In this work, we describe a simple, novel approach to improve the accuracy of predicting protein-protein interaction by incorporating the binding site information. First, we assess the importance of the seven attributes that are used by Bradford et. al (2005) for predicting protein binding sites. The leave-one-out cross validation experiments and principal component analysis indicate that some attributes such as residue propensity and hydrophobicity are more important than other attributes such as curvedness and shape index in differentiating a binding patch from nonbinding patch. Second, we incorporate those attributes to predict protein-protein interaction by simple concatenation of the attribute vectors of candidate interacting partners. A support vector machine is trained to predict the interacting partners. This is combined with using the attributes directly derived from the primary sequence at the binding sites. The results from the leave-one-out cross validation experiments show significant improvement in prediction accuracy by incorporating the structural information at the binding sites