Extracting Molecular Binding Relationships from Biomedical Text

Extracting Molecular Binding Relationships from Biomedical Text Thomas C. RINDFLESCH National Library of Medicine 8600 Rockville Pike Bethesda, MD, 20894 tcr@nlm.nih.gov Jayant V. RAJAN University of Rochester Rochester, NY, 14620 Jayant.Rajan@ mc.rochester.edu Lawrence HUNTER National Cancer Institute 7550 Wisconsin Avenue Bethesda, MD, 20894 lhunter@nih.gov Abstract ARBITER is a Prolog program that extracts assertions about macromolecular binding relationships from biomedical text. We de- scribe the domain knowledge and the under- specified linguistic analyses that support the identification of these predications. After discussing a formal evaluation of ARBITER, we report on its application to 491,000 MEDLINE ~ abstracts, during which almost 25,000 binding relationships suitable for entry into a database of macro- molecular function were extracted. Introduction Far more scientific information exists in the lit- erature than in any structured database. Con- venient access to this information could signifi- cantly benefit research activities in various fields. The emerging technology of information extraction (Appelt and Israel 1997, Hearst 1999) provides a means of gaining access to this in- formation. In this paper we report on a project to extract biomolecular data from biomedical text. We concentrate on molecular binding affinity, which provides a strong indication of macro- molecular function and is a core phenomenon in molecular biology. Our ultimate goal is to auto- matically construct a database of binding rela- tionships asserted in MEDLINE citations. The National Library of Medicine's MEDLINE textual database is an online reposi- tory of more than 10 million citations from the biomedical literature. All citations contain the title of the corresponding article along with other bibliographic information. In addition, a large number of citations contain author-supplied ab- stracts. Initial studies indicate that there are ap- proximately 500,000 MEDLINE citations rele- vant to molecular binding affinity. Our decision to apply information extraction technology to binding relationships was guided not only by the biological importance of this phenomenon but also by the relatively straight- forward syntactic cuing of binding predications in text. The inflectional forms of a single verb, bind, indicate this relationship in the vast major- ity of cases, and our initial work is limited to these instances. For example, our goal in this project is to extract the binding predications in (2) from the text in (1). (1) CC chemokine receptor 1 (CCR1) is ex- pressed in neutrophils, monocytes, lympho- cytes, and eosinophils, and binds the leuko- cyte chemoattractant and hematopoiesis regulator macrophage inflammatory protein (MIP)- 1 alpha, as well as several related CC chemokines. (2) <CC chemokine receptor 1> BINDS <leukocyte chemoattractant> <CC chemokine receptor 1> BINDS <hematopoiesis regulator macrophage inflammatory protein- 1 alpha> <CC chemokine receptor 1> BINDS <related CC chemokine> Considerable interest in information extrac- tion has concentrated on identifying named enti- ties in text pertaining to current events (for ex- ample, Wacholder et al. 1997, Voorhees and Harman 1998, and MUC-7); however, several recent efforts have been directed at biomolecular data (Blaschke et al. 1999, Craven and Kumlien 1999, and Rindflesch et al. 2000, for example). The overall goal is to transform the information …

Content truncated to 500 words