An ensemble method approach to investigate kinase-specific phosphorylation sites

Abstract

Protein phosphorylation is one of the most significant and well-studied post-translational modifications, and it plays an important role in various cellular processes. It has made a considerable impact in understanding the protein functions which are involved in revealing signal transductions and various diseases. The identification of kinase-specific phosphorylation sites has an important role in elucidating the mechanism of phosphorylation; however, experimental techniques for identifying phosphorylation sites are labor intensive and expensive. An exponentially increasing number of protein sequences generated by various laboratories across the globe require computer-aided procedures for reliably and quickly identifying the phosphorylation sites, opening a new horizon for in silico analysis. In this regard, we have introduced a novel ensemble method where we have selected three classifiers (least square support vector machine, multilayer perceptron, and k-Nearest Neighbor) and three different feature encoding parameters (dipeptide composition, physicochemical properties of amino acids, and protein-protein similarity score). Each of these classifiers is trained on each of the three different parameter systems. The final results of the ensemble method are obtained by fusing the results of all the classifiers by a weighted voting algorithm. Extensive experiments reveal that our proposed method can successfully predict phosphorylation sites in a kinase-specific manner and performs significantly better when compared with other existing phosphorylation site prediction methods.

Keywords: cell signaling; phosphate; post-translational modification.

Figure 1

Block diagram of the proposed ensemble method that uses protein sequences as input. Abbreviations: K-NN, K-nearest neighbor; LSSVM, least square support vector machine; MLP, multilayer perceptron.

Figure 2

Average accuracy versus number of features for each of the nine kinases families. Abbreviations: CAM KII, calmodulin-dependent protein kinase II; CDK1, cyclin-dependent kinase 1; CK2, casein kinase 2; GSK-3, glycogen synthase kinase 3; MAPK, mitogen-activated protein kinase; PKA, protein kinase A; PKB, protein kinase B; PKC, protein kinase C; SRC, tyrosin kinase SRC.

Figure 3

Accuracy with various window sizes for all the nine kinases while varying the number of resample datasets. Abbreviations: CAM KII, calmodulin-dependent protein kinase II; CDK1, cyclin-dependent kinase 1; CK2, casein kinase 2; GSK-3, glycogen synthase kinase 3; MAPK, mitogen-activated protein kinase; PKA, protein kinase A; PKB, protein kinase B; PKC, protein kinase C; SRC, tyrosin kinase SRC.

Figure 4

Comparison of various kinase-specific phosphorylation site prediction methods (PPSP, KinasePhos, GPS2.0, Scansite, and NetphosK) with our proposed methods in terms of precision. Abbreviations: CAM KII, calmodulin-dependent protein kinase II; CDK1, cyclin-dependent kinase 1; CK2, casein kinase 2; GSK-3, glycogen synthase kinase 3; MAPK, mitogen-activated protein kinase; PKA, protein kinase A; PKB, protein kinase B; PKC, protein kinase C; SRC, tyrosin kinase SRC.

Figure 5

Comparison of various kinase-specific phosphorylation site prediction methods (PPSP, KinasePhos, GPS2.0, Scansite, and NetphosK) with our proposed methods in terms of recall. Abbreviations: CAM KII, calmodulin-dependent protein kinase II; CDK1, cyclin-dependent kinase 1; CK2, casein kinase 2; GSK-3, glycogen synthase kinase 3; MAPK, mitogen-activated protein kinase; PKA, protein kinase A; PKB, protein kinase B; PKC, protein kinase C; SRC, tyrosin kinase SRC.

Figure 6

Comparison of various kinase-specific phosphorylation site prediction methods (PPSP, KinasePhos, GPS2.0, Scansite, and NetphosK) with our proposed methods in terms of accuracy. Abbreviations: CAM KII, calmodulin-dependent protein kinase II; CDK1, cyclin-dependent kinase 1; CK2, casein kinase 2; GSK-3, glycogen synthase kinase 3; MAPK, mitogen-activated protein kinase; PKA, protein kinase A; PKB, protein kinase B; PKC, protein kinase C; SRC, tyrosin kinase SRC.

Figure 7

Comparison of various kinase-specific phosphorylation site prediction methods (PPSP, KinasePhos, GPS2.0, Scansite, and NetphosK) with our proposed methods in terms of F-measure. Abbreviations: CAM KII, calmodulin-dependent protein kinase II; CDK1, cyclin-dependent kinase 1; CK2, casein kinase 2; GSK-3, glycogen synthase kinase 3; MAPK, mitogen-activated protein kinase; PKA, protein kinase A; PKB, protein kinase B; PKC, protein kinase C; SRC, tyrosin kinase SRC.

Figure 8

Prediction accuracy comparison of our proposed method with various nonkinase-specific phosphorylation site prediction methods (DISPHOS, PPRED, NetPhos) for serine, threonine, and tyrosine phosphoresidues.