Extracting Regulatory Gene Expression Networks From Pubmed

Extracting Regulatory Gene Expression Networks from PubMed Jasmin ˇSari´c EML Research gGmbH Heidelberg, Germany saric@eml-r.org Lars J. Jensen EMBL Heidelberg, Germany jensen@embl.de Rossitza Ouzounova EMBL Heidelberg, Germany ouzounov@embl.de Isabel Rojas EML Research gGmbH Heidelberg, Germany rojas@eml-r.org Peer Bork EMBL Heidelberg, Germany bork@embl.de Abstract We present an approach using syntacto- semantic rules for the extraction of rela- tional information from biomedical ab- stracts. The results show that by over- coming the hurdle of technical termi- nology, high precision results can be achieved. From abstracts related to baker’s yeast, we manage to extract a regulatory network comprised of 441 pairwise relations from 58,664 abstracts with an accuracy of 83–90%. To achieve this, we made use of a resource of gene/protein names considerably larger than those used in most other biology re- lated information extraction approaches. This list of names was included in the lexicon of our retrained part-of-speech tagger for use on molecular biology ab- stracts. For the domain in question an accuracy of 93.6–97.7% was attained on POS-tags. The method is easily adapted to other organisms than yeast, allowing us to extract many more biologically rel- evant relations. 1 Introduction and related work A massive amount of information is buried in scientific publications (more than 500,000 pub- lications per year). Therefore, the need for in- formation extraction (IE) and text mining in the life sciences is drastically increasing. Most of the ongoing work is being dedicated to deal with PubMed1 abstracts. The technical terminology of biomedicine presents the main challenge of apply- ing IE to such a corpus (Hobbs, 2003). The goal of our work is to extract from bio- logical abstracts which proteins are responsible for regulating the expression (i.e. transcription or translation) of which genes. This means to extract a specific type of pairwise relations between bio- logical entities. This differs from the BioCreAtIvE competition tasks2 that aimed at classifying en- tities (gene products) into classes based on Gene Ontology (Ashburner et al., 2000). A task closely related to ours, which has re- ceived some attention over the past five years, is the extraction of protein–protein interactions from abstracts. This problem has mainly been ad- dressed by statistical “bag of words” approaches (Marcotte et al., 2001), with the notable exception of Blaschke et al. (1999). All of the approaches differ significantly from ours by only attempting to extract the type of interaction and the partici- pating proteins, disregarding agens and patiens. Most NLP based studies tend to have been fo- cused on extraction of events involving one par- ticular verb, e.g. bind (Thomas et al., 2000) or in- hibit (Pustejovsky et al., 2002). From a biological point of view, there are two problems with such approaches: 1) the meaning of the extracted events 1PubMed is a bibliographic database covering life sci- ences with a focus on biomedicine, comprising around 12 × 106 articles, roughly half of them including abstract (http: //www.ncbi.nlm.nih.gov/PubMed/). 2Critical Assessment of Information Extraction sys- tems in Biology, http://www.mitre.org/public/ biocreative/ will depend strongly on the selectional restrictions …

Content truncated to 500 words