Prediction of protein-protein interaction sites using back propagation neural networks

Identifying protein-protein interaction sites plays a very important role in biological process and protein function. This paper proposes a novel Back Progagtion (BP) neural network improved by Levenberg-Marquardt (LM) algorithm that can predict protein-protein interaction sites. Sequence profile and entropy information of sequential adjacent residues are extracted to construct the different sliding windows as the input feature vectors for the BP network. 10-cross validation method is used to build the training and testing set. The improved BP network is applied to distinguish whether a surface residue in the testing set is an interface residue or not. The results based on a non-redundant data set of heterodimers such as the accuracy is 75.3% and the correlation coefficient is 0.302 that efficiently proves the validity and feasibility of our method. The results also show that the model we proposed is superior to the classical computational methods mentioned in this paper.