Table 3.

Performance comparison between models trained with and without distributed contextual feature vectors for kinase-specific phosphorylation site prediction.

	ws	AGC/PKA	AGC/PKC	CMGC/CDK	CMGC/CK2	TK/Src
Residue-level	0	0.915 +/− 0.013	0.898 +/− 0.022	0.799 +/− 0.040	0.839 +/− 0.040	0.956 +/− 0.011
Context2vec inference	7	0.609 +/− 0.043	0.653 +/− 0.056	0.570 +/− 0.042	0.736 +/− 0.056	0.828 +/− 0.051
	11	0.918 +/− 0.013	0.862 +/− 0.031	0.890 +/− 0.029	0.885 +/− 0.048	0.913 +/− 0.014
	15	0.895 +/− 0.020	0.862 +/− 0.028	0.885 +/− 0.030	0.893 +/− 0.043	0.895 +/− 0.016
	19	0.880 +/− 0.018	0.865 +/− 0.026	0.872 +/− 0.031	0.889 +/− 0.044	0.875 +/− 0.022
Context2vec add	7	0.885 +/− 0.025	0.809 +/− 0.022	0.903 +/− 0.030	0.814 +/− 0.055	0.931 +/− 0.010
	11	0.925 +/− 0.013	0.872 +/− 0.026	0.883 +/− 0.026	0.882 +/− 0.037	0.912 +/− 0.014
	15	0.906 +/− 0.015	0.882 +/− 0.023	0.872 +/− 0.034	0.898 +/− 0.032	0.893 +/− 0.017
	19	0.892 +/− 0.016	0.885 +/− 0.021	0.869 +/− 0.034	0.901 +/− 0.032	0.877 +/− 0.017
Residue-level + Context2vecinference	7	0.928 +/− 0.011	0.897 +/− 0.028	0.925 +/− 0.018	0.866 +/− 0.041	0.960 +/− 0.009
	11	0.938 +/− 0.009	0.909 +/− 0.027	0.908 +/− 0.018	0.907 +/− 0.037	0.964 +/− 0.008
	15	0.937 +/− 0.010	0.907 +/− 0.028	0.902 +/− 0.024	0.919 +/− 0.034	0.966 +/− 0.007
	19	0.937 +/− 0.011	0.909 +/− 0.024	0.899 +/− 0.025	0.917 +/− 0.033	0.965 +/− 0.008
Residue-level + Context2vecadd	7	0.927 +/− 0.012	0.895 +/− 0.028	0.907 +/− 0.020	0.864 +/− 0.040	0.957 +/− 0.009
	11	0.939 +/− 0.010	0.911 +/− 0.023	0.896 +/− 0.020	0.911 +/− 0.034	0.962 +/− 0.009
	15	0.939 +/− 0.010	0.913 +/− 0.023	0.890 +/− 0.022	0.927 +/− 0.026	0.964 +/− 0.008
	19	0.940 +/− 0.010	0.915 +/− 0.020	0.894 +/− 0.024	0.929 +/− 0.024	0.964 +/− 0.008
Residue-level + prot2vecadd(ProtVec)	inf	0.908 +/− 0.013	0.874 +/− 0.022	0.738 +/− 0.042	0.827 +/− 0.036	0.954 +/− 0.010

The prediction performance was evaluated in terms of the average AUC score and the standard deviation.