Table 3. Performance of the classifiers illustrated by confusion matrices of the prediction of IVIG binding on the test set.

		Classifiers trained on:
		Original training seta		Balanced training seta
Classifierb	Prediction	Actual bindersc	Actual non-bindersc	Actual bindersc	Actual non-bindersc
ML-advanced	Binding	2,420	916	2,735	1,539
	Non-binding	1,001	9,303	686	8,680
2-PWM	Binding	2,162	1,260	n.d.d	n.d.d
	Non-binding	1,260	8,958	n.d.d	n.d.d
El-Manzalawy	Binding	2,253	1,033	2,416	1,453
	Non-binding	1,168	9,186	1,005	8,766

^aTraining sets consist of either three times more “non-binding” than “binding” peptides (original) or an equal number of both groups (balanced).

^bOur classifiers (ML-advanced, PWM with treshold 2.45; see Figure 1 for workflow) and the classifier of El-Manzalawy et al. [22] [23].

^cCorrect predictions are underlined.

^dnot determined.