A neural network method to improve prediction of protein-protein interaction sites in heterocomplexes

In this paper we describe an algorithm, based on neural networks that adds to the previously published results (ISPRED, www.biocomp.unibo.it) and increases the predictive performance of protein-protein interaction sites in protein structures. The goal is to reduce the number of spurious assignment and developing knowledge based computational approach to focus on clusters of predicted residues on the protein surface. The algorithm is based on neural networks and can be used to highlight putative interacting patches with high reliability, as indicated when tested on known complexes in the PDB. When a smoothing algorithm correlates the network outputs, the accuracy in identifying the interaction patches increases from 73% up 76%. The reliability of the prediction is also increased by the application the smoothing procedure.