Table 3. Prediction performance comparison of TANGLE with ANGLOR and the random amino acid-specific predictor.

		Phi angle (°)			Psi angle (°)
		MAETANGLE	MAEANGLOR	MAErandom	MAETANGLE	MAEANGLOR	MAErandom
AAa	ALA	21.9	22.5	27.4	38.2	42.7	79.7
	CYS	25.5	27.7	32.5	45.0	48.7	85.3
	ASP	29.7	30.8	32.2	48.7	48.9	73.5
	GLU	22.3	23.3	27.4	39.1	43.1	75.2
	PHE	23.6	24.2	32.0	39.4	40.8	85.8
	GLY	84.1	75.1	95.1	76.7	66.9	79.2
	HIS	29.6	31.8	35.7	46.4	48.2	76.4
	ILE	17.5	18.1	26.4	32.1	35.3	84.4
	LYS	24.8	25.6	30.6	41.8	45.6	79.0
	LEU	17.8	18.3	24.4	35.2	38.1	81.4
	MET	22.0	22.4	29.5	36.5	40.9	81.6
	ASN	37.1	37.6	42.3	45.2	45.9	68.2
	PRO	13.6	15.2	19.5	59.3	61.3	86.3
	GLN	23.9	25.1	30.0	39.4	43.0	76.9
	ARG	23.5	25.0	30.4	40.9	44.1	80.5
	SER	30.6	32.3	35.3	53.5	55.4	87.0
	THR	23.9	26.0	29.9	50.4	51.1	88.6
	VAL	19.1	20.1	28.5	34.8	37.6	83.1
	TRP	22.8	23.1	29.8	41.6	43.5	86.4
	TYR	23.7	25.3	32.4	40.1	42.3	85.5
All		27.8	28.2	33.8	44.6	46.4	80.9
SSb	H	9.9	11.0	19.0	18.7	28.2	29.3
	E	26.1	27.9	28.1	38.9	39.9	36.1
	C	40.8	41.8	51.5	66.0	63.9	81.3
SAc	E	30.7	31.2	55.7	47.0	49.9	84.6
	B	24.1	24.1	52.0	40.2	41.5	84.0

Prediction performance is categorized according to twenty amino acid types, three secondary structure types (H, helix; E, beta-strand; and C, coil) and two-state solvent accessibility (E, exposed and B, buried), evaluated by the mean absolute error (MAE). 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.

^aTwenty amino acid types.

^bThree secondary structure types. H: alpha-helix; E: beta-strand; C: coil.

^cTwo-class solvent accessibility: E: exposed; B: buried.