Biomedical Relation Extraction: From Binary to Complex

Table 5

Performance of existing PPI extraction methods on the data corpora used.

Category	Result (%)		Corpus	References
	Recall	Precision
Rule-based	86.8	94.3	834 and 752 sentences obtained by a MEDLINE search using these keywords, “protein binding,” “yeast,” “E. coli,” “protein,” and “interaction.”	[21]
	60	87	550 sentences were retained containing at least one of four keywords “interact,” “bind,” “associate,” “complex,” or one of their inflections from 3343 abstracts retrieved from MEDLINE with the following keywords: “Saccharomyces cerevisiae,” “protein,” and “interaction.”	[22]
	80.0	80.5	About 1200 sentences were kept from the top 50 biomedical papers retrieved from the Internet by querying using the keyword “protein-protein interaction.”	[4]
ML methods	57	90	Training set consists of 500 abstracts from MEDLINE. Evaluation set consists of 56 abstracts collected using search strings “protein” and “inhibit.”	[48]
	21	91	3.4 million sentences from approximately 3.5 million MEDLINE abstracts dated after 1988 containing at least one notation of a human protein.	[49]
	71.9	60	AIMed	[38]
	87.2	72.5	LLL	[39]
	76	70	The test corpus consists of 300 randomly selected sentences.	[24]
	70.7	70.3	LLL	[10]
	71.9	60	AIMed	[30]
	59.26	63.37	LLL	[9]
	89	73	LLL	[11]