Improving classification in protein structure databases using text mining

BMC Bioinformatics

Table 7 Classifier performance in the enzyme dataset

CONDITIONS				PERFORMANCE
				Superfamily Classification N = 352, 5 classes
	Test Set	Reference Set	Lucene Analyser	AUC	MCC	F
1	Abstract	Dp20 – DX33 -Ann	Stop	0.75	0.51	0.56
2	Annotations	Dp20 – DX33 -Ann	Stop	0.77	0.53	0.58
3	Abstract	Dp20 – Ann	Stop	0.74	0.50	0.53
4	Abstract	Dp20 – Ann	Standard	0.70	0.33	0.44
5	Abstract	Dp20	Stop	0.74	0.49	0.52
6	Abstract	Dp1	Stop	0.64	0.31	0.40

Classifier performance was assessed using AUC, MCC and F-measure under six conditions of the reference set PDB145: Inclusion of additional abstracts from related articles (Dp20); inclusion of annotations (Ann); filtering using the SVM model (DX33). Conditions 1 and 2 : 20 SVM filtered abstracts per enzyme, Stop analyser, inclusion of PDB/UniProt annotations.Condition 3 : 20 abstracts per enzyme, Stop analyser, inclusion of PDB/UniProt annotations. Condition 4 : 20 abstracts per enzyme, Standard analyser, inclusion of PDB/UniProt annotations. Condition 5 : 20 abstracts per enzyme, Stop analyser. Condition 6 : 1 abstract per enzyme, Stop analyser. For the Superfamily classification task all 352 enzymes of REGS352 were classified in 5 superfamilies. Abstracts were used in the test set, except of condition 2 were annotations from UniProt and PDB fields were used for comparison.

ISSN: 1471-2105