Table 1. Predictive performance of Phi and Psi angles based on different local window sizes using the PSI-BLAST profile.

Torsion angles	Local window size	Number of features	Number of support vectors	CC	RMSE	MAE
Phi	3	60	69370	0.455	49.25	31.44
	5	100	69358	0.478	48.57	30.44
	7	140	69299	0.484	48.33	30.05
	9	180	69285	0.486	48.24	29.92
	11	220	69243	0.483	48.27	29.94
	13	260	69343	0.478	48.42	30.04
	15	300	69382	0.472	48.55	30.25
	17	340	69350	0.466	48.73	30.46
	19	380	69369	0.459	48.90	30.71
	21	420	69344	0.451	49.13	30.99
Psi	3	60	69955	0.469	80.79	63.33
	5	100	69923	0.537	76.85	58.84
	7	140	69855	0.563	75.27	56.91
	9	180	69712	0.575	74.55	55.96
	11	220	69738	0.581	74.18	55.43
	13	260	69718	0.581	74.24	55.38
	15	300	69736	0.580	74.44	55.43
	17	340	69724	0.577	74.70	55.68
	19	380	69719	0.573	75.08	56.04
	21	420	69665	0.569	75.43	56.41

The results were obtained using an independent test set of 1,026 proteins from the set of PDB data compiled by Wu and Zhang [11], where the rest 500 proteins were used for training.