Automatic reconstruction of a bacterial regulatory network using Natural Language Processing

BMC Bioinformatics

Table 2 Final network extraction system evaluation metrics

Source	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
RegulonDB (as of June 2006)	3108	-	3397	3397	100%	100%	-	-	-
AS	3148	768	661	1429	45%	45%	0.45	0.77	0.57
AS*	2649	569	535	1104	41%	35.5%	0.41	0.72	0.47
RP	2650	711	605	1316	49%	42%	0.49	0.78	0.55
RP*	2202	522	491	1013	46%	32%	0.46	0.74	0.45
RN	1643	555	471	1026	62%	33%	0.62	0.85	0.47
RN*	1354	426	385	811	59%	26%	0.59	0.81	0.39
EA	627	262	140	402	64%	12%	0.64	0.95	0.22
EA*	554	217	114	331	59%	10%	0.59	0.91	0.19
RS	718	254	146	400	55%	12%	0.55	0.91	0.22
RS*	630	207	121	328	52%	10%	0.52	0.86	0.18
ST	691	199	143	342	49%	11%	0.49	0.90	0.19
ST*	628	170	118	288	45%	9%	0.45	0.86	0.16
RA	414	207	115	322	77%	10%	0.77	1.00	0.18
RA*	370	180	97	277	74%	8%	0.74	0.99	0.16

An asterisk [*] next to source name indicates that no multiple-unit objects (like the individual elements of two-system components, or operons) were added; The RegulonDB "dual" interactions (that is, presenting both activation and repression), are counted here as two distinct interactions. AS represents a file containing all compiled interactions found in all textual sources used in this work, and constitutes the sum of all non-redundant interactions extracted from the full-text and abstract documents.
1. Unique, non-repeated interactions found for each file
2. Interactions that match RegulonDB interaction pairs, but whose function (activation or repression) was not determined by the system.
3. Interactions that match RegulonDB entries, and also match repressor/activator function
4. Overall matches (column 2 + column 3), including both interactions with complete information as well as under specified ones
5. % of total interactions in file which are correct: (1)/(4)
6. Recall: percentage of RegulonDB's 3108 interactions that is represented by correct interactions in file: ((4)*100)/3108
7. Precision 1: Number of overall matches (4)/unique interactions in file (1)
8. Precision 2: Number of overall matches (4)/interactions in file that contain a RegulonDB-annotated regulator
9. F-Measure (F = 2RP/R+P)

ISSN: 1471-2105