Active machine learning for transmembrane helix prediction

BMC Bioinformatics

Table 3 Comparison of TMpro NN: applying active vs. passive learning algorithms for updating training set from PDBTM (191 proteins, 789 TM segments).

	Methods	# of Proteins in Training-Set	Qok	Qhtm	Qhtm	Qhtm	Q2
				Fscore	%obs	%prd
1	Random	1	20	51	54	49	20
		2	32	69	72	66	32
		5	35	76	78	73	35
		10	35	78	81	75	35
2	Node-Coverage	1	50	91	93	90	79
		2	47	90	91	88	79
		5	49	91	92	89	79
		10	50	91	93	90	79
3	Confusion-Rated	1	20	48	51	46	70
		2	36	81	84	78	75
		5	38	85	90	81	74
		10	46	90	92	87	78
4	Node-Coverage & Confusion-Rated	1	50	91	93	90	79
		2	49	90	92	88	78
		5	48	91	93	89	79
		10	51	91	93	90	79

For description of columns, see caption of Table 1. Qhtm,%obs and Qhtm,%pred have been computed per-protein and averaged over all the proteins.

ISSN: 1471-2105