Modular prediction of protein structural classes from sequences of twilight-zone identity with predicting sequences

Table 6 Results of the experimental comparison of the proposed MODAS method and the competing SCEC and SCPRED methods on the ASTRAL_testdataset with the four major structural classes.

Algorithm	Training dataset	Accuracy					MCC				GC²
		α	β	α/β	α+β	Overall	α	β	α/β	α+β
MODAS with 4 classes	ASTRAL_training	94.05	83.48	85.12	71.47	83.05	0.92	0.79	0.78	0.61	0.63
MODAS with 6 classes	ASTRAL_training	93.27	82.73	82.31	71.07	81.84	0.92	0.79	0.77	0.62	0.64
MODAS with 7 classes	ASTRAL_training	91.71	82.27	82.31	70.29	81.17	0.91	0.79	0.77	0.61	0.64
SCPRED	ASTRAL_training	93.13	78.33	83.38	64.27	79.14	0.92	0.77	0.70	0.52	0.57
SCPRED	25PDB	92.81	79.09	80.05	63.74	78.36	0.92	0.78	0.67	0.51	0.56
SCEC	Web server	75.74	72.73	78.42	28.14	62.80	0.65	0.55	0.59	0.22	0.29

The MODAS method was used to make predictions for all 7, 6 (excluding the small proteins class), and the 4 major classes. The SCEC method was trained on the ASTRAL_training with the proteins from the 4 major classes (this method can handle only prediction of the 4 classes) and on the 25PDB dataset based on results in [53]. The SCEC predictions were generated using the web server at http://biomine.ece.ualberta.ca/Structural_Class/SCEC.html. Bold font indicates the best results.

ISSN: 1471-2105