Prediction of kinase–substrate relations based on heterogeneous networks

Abstract

Protein phosphorylation catalyzed by kinases plays essential roles in various intracellular processes. With an increasing number of phosphorylation sites verified experimentally by high-throughput technologies and assigned as substrates of specific kinases, prediction of potential kinase–substrate relations (KSRs) attracts increasing attention. Although a large number of computational methods have been designed, most of them only focus on local protein sequence information. A few KSR prediction approaches integrate protein–protein interaction and protein sequence information into existing machine learning algorithms at the cost of high feature dimensions or reduced sensitivity. In this work, we introduce two novel heterogeneous networks, HetNet-PPI and HetNet-SEQ, by incorporating PPI and similarity of protein sequences into the kinase–substrate heterogeneous networks, respectively. Based on these two heterogeneous networks, we further propose two new KSR prediction methods, HeteSim-PPI and HeteSim-SEQ, by adopting the HeteSim algorithm, which is recently proposed for relevance measure in heterogeneous information networks. Comprehensive evaluation results of the two methods show that similarity of protein sequences is more effective in improving KSR prediction performance as HeteSim-SEQ outperforms HeteSim-PPI in most cases. Further comparison results demonstrate that HeteSim-SEQ is superior to existing methods including BDT, SVM and iGPS, suggesting the effectiveness of the proposed network-based method in predicting potential KSRs.