Neural prediction of protein-protein interactions based on physicochemical correlation coefficients and bootstrapping for artificial data generation

Abstract

Although using only protein sequences might be sufficient for predicting, there are major problems in the prediction of protein-protein interactions by classifying technique such as supervised neural network. The first one is extracting the feature of protein pair sequences to form a feature sequence. The second problem is conserving the information when equalizing the lengths of feature sequences before classifying into interacting and non-interacting classes. This dissertation proposed a method to predict protein-protein interactions from amino acid sequences using only artificial boundary data generation and boosting procedures to improve the prediction accuracies. The feature extraction is based on the correlation coefficients of physicochemical properties, the statistical means and standard deviations of secondary structures and protein properties. The important data which lie into the boundary of each subcluster were only used to generate the artificial boundary data by bootstrap resampling technique. Finally, the only artificial boundary data of both positive and negative protein pairs were predicted by boosting method based on neural network classifier. The empirical study has shown that our proposed method yielded better prediction accuracy than the sequence-based methods when performed on Yeast Saccharomyces Cerevisiae data set. Moreover, the number of feature and the number of training data were less than others. The prediction models were also evaluated by cross-species test data sets. The result showed that the proposed method also capable to predict with the good performance on cross-species data.