Integrative approaches to the prediction of protein functions based on the feature selection

BMC Bioinformatics

Table 1 Prediction quality of different prediction approaches

Specificity	# of GO terms	KLR					KL1LR				KLR with Relief
		Pfam	Interpro	All		ES	λ / λ _max = 0.01		λ / λ _max = 0.1
				NOS	S		NOS	S	NOS	S
3-10	952	0.60	0.58	0.55	0.54	0.55	0.73	0.75	0.72	0.74	0.58
11-30	435	0.74	0.76	0.70	0.70	0.82	0.85	0.88	0.79	0.85	0.69
31-100	239	0.79	0.79	0.82	0.82	0.84	0.84	0.88	0.78	0.86	0.64
101-300	100	0.80	0.80	0.84	0.84	0.83	0.82	0.86	0.73	0.84	0.61

GO terms are categorized into four groups based on the number of genes covering the GO term (specificity in the first column). Prediction quality is estimated using AUC values for KLR using Pfam or Interpro data only, KLR using all data sources, KLR using a data source selected by exhaustive search (ES), KL1LR, and KLR using data sources selected by the Relief method. In the case of KL1LR, two different values of the regularization parameter λ are used. NOS stands for non-standardization of features, and S for standardization.

ISSN: 1471-2105