Exploiting Two-Layered Support Vector Machine to Predict Phosphorylation Sites on Virus Proteins

Abstract

Abstract-Protein phosphorylation in viruses plays crucial regulatory roles in enhancing progression, replication, and inhibition of host cell functions. Due to the difficulty of mass spectrometry-based identification of viral phosphorylation sites, we are motivated to develop a new method to investigate the substrate motifs and identify protein phosphorylation sites on viruses. The experimentally verified phosphorylation data were extracted from a public resource and a recursively statistical method is applied to cluster whole data set of phosphorylated sequences into subgroups containing remarkably sequence motifs around the phosphorylation sites. Two-layered Support Vector Machine (SVM) is then applied to learn a predictive model by integrating the detected sequence motifs. A five-fold cross validation evaluation on the SVM model yields an average accuracy of 0.88 for Serine and 0.83 for Threonine. Furthermore, the independent testing data collected from UniProtKB and Phospho.ELM indicates that the proposed method is comparable with three popular kinase-specific phosphorylation site prediction tools. The cross validation and independent testing demonstrated that the sequence motifs are informative for the prediction of potential kinases for virus protein phosphorylation sites. Furthermore, the proposed method is a practical means of preliminary analysis for virus phosphorylation dynamics.Index Terms-Virus, protein phosphorylation, substrate motif, support vector machine.