Combining One-Class Classification Models Based on Diverse Biological Data for Prediction of Protein-Protein Interactions

Abstract

This research addresses the problem of prediction of protein-protein interactions (PPI) when integrating diverse biological data. Gold Standard data sets frequently employed for this task contain a high proportion of instances related to ribosomal proteins. We demonstrate that this situation biases the classification results and additionally that the prediction of non-ribosomal based PPI is a much more difficult task. In order to improve the performance of this subtask we have integrated more biological data into the classification process, including data from mRNA expression experiments and protein secondary structure information. Furthermore we have investigated several strategies for combining diverse one-class classification (OCC) models generated from different subsets of biological data. The weighted average combination approach exhibits the best results, significantly improving the performance attained by any single classification model evaluated.