PSPhos: PK-Specific Phosphorylation Site Prediction Using Profile SVM

Because of the importance of phosphorylation and the mystery of the corresponding mechanism in vivo, predicting potential phosphorylation sites with protein kinases (PKs) information is a critical and intriguing challenge in bioinformatics. Although some in silico methods have been introduced for this purpose, sophisticated methods are still in urgent demand to further improve prediction performances. In this paper, we present a novel, versatile and comprehensive program, PSPhos (Profile SVM for Phosphorylation site prediction), based on profile Support Vector Machine. We estimated the performance of PSPhos using experimentally verified substrate sites of two important PK families: ABL and DNA-PK, and compared the performance with the GPS and PPSP algorithms. With the Jack-knife tests, PSPhos consistently performed ∼20% better in sensitivity. Then a more comprehensive comparison between PSPhos and most available web servers for phosphorylation prediction was made on two manually-curated testing data of novel substrate sites. With a high specificity of 100%, the sensitivity obtained by PSPhos on the ABL data set was 16% higher than the best result obtained by several existing methods. On DNA-PK data, only PSPhos could successfully identify true phosphorylation sites with high stringency for minimal false positives. All these results indicate that PSPhos is a useful computational resource for the identification of PK-specific phosphorylation sites, and can provide helpful insights for future experimental design and verification.