Table 2.

Domain-specific neural network and PSSM results. The application of a domain-specific strategy in the detection of binders reveals the strong effect of the data unbalancing. Class I binding domains have a lower percentage of binders within the datasets and in the corresponding results both PSSM and neural networks display low performances, with no clear benefit in preferring one method to the other. The results of class II binding domains, where a higher percentage of binders (Rvs167, Yfr024, Ysc84) is present, clearly show the prevalence of neural networks. For Boi1 and Boi2 the estimation of PSSM and NN is less significant due to the scarcity of binders.

Class	Domain	Number of Binders	PSSM				NN
			Prec	Sens	Spec	Corr	Prec	Sens	Spec	Corr
I	BOI1	15 (2.2%)	50	25	99	0.34	4	80	47	0.09
	MYO5	35 (5.2%)	57	67	98	0.60	38	53	97	0.41
	RVS167	19 (2.8%)	0	0	99	-0.01	31	68	96	0.43
	SHO1	37 (5.5%)	70	64	98	0.65	64	84	97	0.71
	YFR024	25 (3.7%)	14	14	97	0.11	25	37	94	0.25
	YSC84	12 (1.8%)	100	33	100	0.57	10	80	81	0.24
II	BOI1	16 (2.3%)	17	50	95	0.27	19	38	97	0.25
	RVS167	44 (6.2%)	53	62	96	0.54	59	77	96	0.65
	YFR024	123 (17.4%)	47	56	87	0.40	56	78	87	0.58
	YSC84	67 (9.5%)	61	55	96	0.54	60	83	94	0.67