Predicting substrates of the human breast cancer resistance protein using a support vector machine method

BMC Bioinformatics

Table 2 Performance parameters of 100 runs using various ratios of training/test sets

Training/Test ratio	Category	ACC	SE	SP	MCC
0.5/0.5	Training	83.8	94.6	65.8	0.649
	Test	67.8	82.2	44.0	0.288
	External	69.5	78.3	54.7	0.344
0.6/0.4	Training	85.1	94.7	69.1	0.678
	Test	69.0	83.2	45.5	0.315
	External	70.8	79.0	57.1	0.372
0.7/0.3	Training	83.4	93.3	67.1	0.642
	Test	71.1	84.2	49.1	0.360
	External	70.8	78.4	58.1	0.375
0.75/0.25	Training	84.5	93.8	69.1	0.665
	Test	69.7	83.3	47.2	0.332
	External	70.9	77.7	59.6	0.382
0.8/0.2	Training	83.6	93.5	67.1	0.644
	Test	70.4	83.6	48.7	0.35
	External	70.6	77.5	59.1	0.376
0.85/0.15	Training	82.9	93.5	65.1	0.627
	Test	70.3	85.1	46.4	0.347
	External	70.9	78.5	58.1	0.376

The total number of molecules used in the training and test data sets were 223. The number of molecules in the external validation data set was 40. ACC, accuracy (overall prediction accuracy); SP, specificity (prediction accuracy for the non-substrates); SE, sensitivity (prediction accuracy for the substrates); MCC, the Matthews correlation coefficient (a more balanced prediction parameter than ACC). The external data set was only used to validate the prediction power of the models constructed, and was not used for model selection.

ISSN: 1471-2105