Sequence-based protein-protein interaction prediction via support vector machine

This paper develops sequence-based methods for identifying novel protein-protein interactions (PPIs) by means of support vector machines (SVMs). The authors encode proteins ont only in the gene level but also in the amino acid level, and design a procedure to select negative training set for dealing with the training dataset imbalance problem, i.e., the number of interacting protein pairs is scarce relative to large scale non-interacting protein pairs. The proposed methods are validated on PPIs data of Plasmodium falciparum and Escherichia coli, and yields the predictive accuracy of 93.8% and 95.3%, respectively. The functional annotation analysis and database search indicate that our novel predictions are worthy of future experimental validation. The new methods will be useful supplementary tools for the future proteomics studies.