Improving the performance of protein kinase identification via high dimensional protein–protein interactions and substrate structure data

Abstract

As a crucial post-translational modification, protein phosphorylation regulates almost all basic cellular processes. Recently, thousands of phosphorylation sites have been discovered by large-scale phospho-proteomics studies, but only about 20% of them have information regarding catalytic kinases, which brings a great challenge for correct identification of the protein kinases responsible for experimentally verified phosphorylation sites. In most existing identification tools, only a local sequence was selected to construct predictive models, and information regarding protein-protein interaction (PPI) was adopted for further filtering. However, the limited information utilized by these tools is not sufficient to identify protein kinases responsible for phosphorylated proteins. In this work, a novel computational approach that fully incorporates PPI and substrate structure information is proposed to improve the performance of human protein kinase identification. To handle the issue of high-dimensional PPI and structure data, a two-step feature selection algorithm that incorporates a support vector machine (SVM), is designed to detect information useful in discriminating the corresponding kinase of phosphorylation sites. Benchmark datasets for kinase identification are constructed using human protein phosphorylation data extracted from the latest Phospho.ELM database. With the selected PPI and structure features, the performance of kinase identification is significantly enhanced as compared with that obtained by using only sequence information. To further verify our method, we compared it with the state-of-the-art tools NetworKIN and IGPS at two stringency levels with medium (>90.0%) and high (>99.0%) specificity. The results show that our method outperforms existing tools in identifying protein kinases. Further evaluation demonstrates that our method also has superior performance on different hierarchical levels including kinase, subfamily, family and group.

GPS-PAIL: prediction of lysine acetyltransferase-specific modification sites from protein sequences

“…We also carefully checked the literature and UniProt database34, and all known site-specific HAT-substrate relations (ssHSRs) in the prediction results were pinpointed (Supplementary Table S4). Previously, it was demonstrated that various functional features of proteins, such as gene ontology (GO) annotations and protein-protein interactions (PPIs), were beneficial for the prediction of kinase-specific phosphorylation sites4041. In this work, the GO information was not used, because the functional diversity of HAT-specific acetylated substrates was high and no particularly significant GO terms were detected from the statistical enrichment analysis.…”

Prediction of post-translational modification sites using multiple kernel support vector machine

“…In this study, ten-fold cross-validation as described in existing studies (Gao et al, 2010; Xu et al, 2014a; Xue et al, 2006) was applied to assess the predictive performance of the proposed method. For a given PTM, 9/10 randomly chosen samples were used as the training data while the remaining 1/10 were used as the test data.…”

“…During the past few decades, many efforts including experimental strategies and computational approaches have been undertaken to identify potential PTM sites (Fan et al, 2014; Gao et al, 2016; Xu et al, 2014a), and most of these methods used local sequence information for prediction due to the fact that PTMs generally occur at specific yet conserved motif in the target protein (Blom, Gammeltoft & Brunak, 1999; Eisenhaber & Eisenhaber, 2010; Miller & Blom, 2009). For example, to predict phosphorylation sites, a number of local sequence based tools have been developed, such as GPS 2.0 (Xue et al, 2008), Musite (Gao et al, 2010), PhosphoSVM (Dou, Yao & Zhang, 2014), NetPhos (Blom et al, 2004) and KinasePhos 2.0 (Wong et al, 2007).…”

Prediction of human disease-specific phosphorylation sites with combined feature selection approach and support vector machine

“…We ranked all 315 features with a widely used filtering feature selection algorithm mRMR [19] (1) while the redundancy between the feature s i and all other selected features is donated as RED and calculated by…”

“…However, large amount of features with redundant and noisy data may leave heavy burden on the classifier, which usually leads to decreased performance [19] [20]. To remedy this shortcoming, recently several phosphorylation site prediction approaches incorporating different feature selection methods were proposed.…”