Biclique extension as an effective approach to predict novel interaction partners in metabolic compound-protein interaction networks.

Abstract

MotivationMetabolic networks are complex systems of connected chemical reactions with physical interactions between metabolites and proteins playing a critical role for both metabolic conversion and regulation. In this study, we aimed to predict previously unknown compound-protein interactions (CPI) by transferring the concept of biclique extension, which was developed in the context of drug-target interaction prediction and that is based on the rationale that interactions that readily extend an existing biclique are real, to metabolic CPI networks. ResultsWe developed and tested a workflow to predict CPIs based on the concept of extending existing bicliques and applied it to E. coli and human using their respective known CPI network as input. Depending on the chosen biclique size, for the E. coli network we reached a sensitivity of 39% with an associated precision of 59%. For the larger human CPI network, a sensitivity of 78% with a false-positive rate of less than 5% and an associated precision of 75% was obtained. At more stringent settings, a precision as high as 95% was attainable at the expense of a lowered recall. Prediction performance significantly exceeded that obtained using randomized networks as input. Predicted novel interactions were tested for biomolecular function involvement, with TCA-cycle and ribosomal processes found associated with particularly pronounced statistical significance. As we demonstrate, our approach holds great potential to increase efficiency of experimental testing of CPIs and can readily be transferred to other species of interest. Availability and implementationThe R code and datasets are available at https://github.com/SandraThieme/BiPredict.