Active machine learning for transmembrane helix prediction

BMC Bioinformatics

Table 1 Comparison of TMpro NN: applying active vs. passive learning algorithms for updating training set from benchmark analysis.

	Methods	# of Proteins in Training-Set	Qok	Qhtm	Qhtm	Qhtm	Q2
				Fscore	%obs	%prd
1	Random	1	14	27	29	25	55
		2	36	63	67	60	65
		5	51	82	84	80	70
		10	54	91	95	88	73
2	Node-Coverage	1	61	94	97	92	75
		2	61	94	97	91	75
		5	63	94	97	92	75
		10	61	94	97	92	75
3	Confusion-Rated	1	14	27	29	25	55
		2	52	91	95	87	73
		5	55	91	95	88	73
		10	59	93	96	89	74
4	Node-Coverage & Confusion-Rated	1	61	94	97	92	75
		2	59	92	96	88	73
		5	58	92	96	89	73
		10	61	94	96	91	74

It can be seen that TMpro achieves high segment accuracy (F-score) even if the classifier is trained with just one protein that is found by Active Learning algorithms. The columns from left to right show: method being evaluated; Number of proteins in training-set; Protein level accuracies: Qok, which is the percentage of proteins in which all experimentally determined segments are predicted correctly, and no extra segments are predicted; that is, there is a one to one match between predicted and experimentally determined segments; Segment F-score which is the geometric mean of Recall and Precision; Recall (Qhtm,%obs, percentage of experimentally determined segments that are predicted correctly); and Precision (Qhtm,%pred percentage of predicted segments that are correct). Q2 is the residue level accuracy when all residues in a protein are considered together, and the Q2 value for the entire set of proteins is the average of that of individual proteins. See [30]for further details on these metrics.

ISSN: 1471-2105