Deep learning methods for protein torsion angle prediction

BMC Bioinformatics

Table 3 The prediction performance of phi and psi angles of using different local window sizes with four deep learning methods

Torsion Angle	Window size^a	Features^b	DRNN	DReRBM	DNN	DRBM
phi	1	56	21.09	21.81	22.13	21.95
	3	168	20.52	20.77	21.49	21.07
	5	280	20.39	20.92	21.24	20.89
	7	392	20.39	21.03	21.22	20.84
	9	504	20.40	20.95	21.28	21.62
	11	616	20.49	20.88	21.04	21.57
	13	728	20.56	20.98	21.27	21.79
	15	840	20.63	21.12	21.19	21.69
	17	952	20.69	21.04	21.38	21.66
psi	1	56	31.68	32.93	33.55	33.02
	3	168	29.29	29.86	30.14	29.74
	5	280	28.96	29.94	29.25	28.92
	7	392	28.85	30.11	29.25	28.85
	9	504	28.86	29.94	29.38	29.61
	11	616	29.06	29.95	29.06	29.75
	13	728	29.27	30.13	29.38	30.19
	15	840	29.44	30.48	29.24	30.25
	17	952	29.72	30.36	29.54	30.33

^aNumber of window size range from 1 to 17
^bNumber of features as input for the deep learning model. For each residue, we used 7 kinds of features, represented by 56 numbers
The bold fond denotes the best result for each method

ISSN: 1471-2105