Table 3.

The average performance of A-cell epitope prediction models on training and independent dataset

Features	Threshold	Sensitivity (%)	Specificity (%)	Accuracy (%)	MCC	Parameters
Internal validation: performance on training dataset, evaluated using fivefold cross-validation
DPC	− 0.2	87.49 ± 1.41	98.70 ± 0.16	97.68 ± 0.22	0.86 ± 0.01	g: 0.0005, c: 1, j: 4
DPC + Motif	− 0.2	87.81 ± 1.01	99.30 ± 0.10	98.25 ± 0.17	0.89 ± 0.01	g: 0.0005, c: 1, j: 4
External validation: performance on independent dataset
DPC	− 0.2	87.54 ± 4.31	98.87 ± 0.28	97.84 ± 0.41	0.87 ± 0.02	g: 0.0005, c: 1, j: 4
DPC + Motif	− 0.2	77.86 ± 5.84	99.28 ± 0.30	97.33 ± 0.58	0.83 ± 0.04	g: 0.0005, c: 1, j: 4

These training and independent datasets were created from alternate datasets using bagging. In alternate dataset, negative or non-epitopes were derived from human proteins. The performance values have been reported as mean ± standard deviation for each model

MCC Matthews correlation coefficient, DPC dipeptide composition, DPC + motif dipeptide composition with MERCI motif score, SVM parameters g gamma parameter of the radial basis function, c trade-off between training error and margin, j regularization parameter (cost-factor, by which training errors on positive examples outweigh errors on negative examples)