Predicting protein-protein interactions based on BP neural network

In this paper, we present a method which only employs protein primary structure to predict protein-protein interactions. The statistical method is used to generate sequence features, which are normalized for satisfying experiments. Six parameters of physicochemical properties are calculated for each protein, including assessable residues, buried residues, hydrophobility, molecular weight, polarity and average area buried. The sequence features are extracted both from interaction proteins and non-interaction proteins. And BP neural network is used to classify two kinds of protein. The statistical evaluation of the BP neural network classifier shows that it performs well above 87% accuracy rate through 10-fold cross-validation. 2000 sequences which come from Scerevisiae yeast dataset are classified in our experimentation. The results demonstrate that 1780 sequences are classified correctly, which show that our proposed method is effective and feasible.