A hybrid ranking method for constructing negative datasets of protein-protein interactions

Abstract

Lack of availability of negative examples in the study of computational Protein-Protein Interaction (PPI) prediction is a crucial problem. This leads to computational methods for creating such examples. Most of these methods rely on the fact that proteins not sharing common information tend not to be interacting. While using this fact as the basis for the selection method for non-PPI pairs may yield a negative dataset with high prediction accuracy, it does come with more bias as it is too selective. Other methods simply use random selection as an alternative for fair selection. However, these approaches do not guarantee the prediction accuracy. A method for constructing non-PPI datasets named AIDNIP is proposed. It is a hybrid of the above approaches. Thus, it can reduce biases of selection, while maintaining prediction accuracies. When compared to the existing methods using a Support Vector Machine-based PPI predictor, the proposed method performs better in several metrics investigated in this study. The Perl code and data used in this study are publically available at https://sites.google.com/a/fsksm.utm.my/aidnip.