Answering Gene Ontology terms to proteomics questions by supervised macro reading in Medline

Abstract

Motivation and Objectives Biomedical professionals have at their disposal a huge amount of literature. But when they have a precise question, they often have to deal with too many documents to efficiently find the appropriate answers in a reasonable time. Faced to this literature overload, the need for automatic assistance has been largely pointed out, and PubMed is argued to be only the beginning on how scientists use the biomedical literature (Hunter and Cohen, 2006). Ontology-based search engines began to introduce semantics in search results. These systems still display documents, but the user visualizes clusters of PubMed results according to concepts which were extracted from the abstracts. GoPubMed (Doms and Schroeder, 2005) and EBIMed (Rebholz-Schuhmann et al, 2007) are popular examples of such ontology-based search engines in the biomedical domain. Question Answering (QA) systems are argued to be the next generation of semantic search engines (Wren, 2011). QA systems no more display documents but directly concepts which were extracted from the search results; these concepts are supposed to answer the user’s question formulated in natural language. EAGLi (Gobeill et al, 2009), our locally developed system, is an example of such QA search engines. Thus, both ontology-based and QA search engines, share the crucial task of efficiently extracting concepts from the result set, i.e. a set of documents. This task is sometimes called macro reading, in contrast with micro reading – or classification, categorization – which is a traditional Natural Language Processing task that aims at extracting concepts from a single document (Mitchell et al, 2009). This paper focuses on macro reading of MEDLINE abstracts. Several experiments have been reported to find the best way to extract ontology terms out of a single MEDLINE abstract, i.e. micro reading. In particular, (Trieschnigg et al, 2009) compared the performances of six classification systems for reproducing the manual Medical Subject Headings (MeSH) annotation of a MEDLINE abstract. The evaluated systems included two morphosyntactic classifiers (sometimes also called thesaurus-based), which aim at literally finding ontology terms in the abstract by alignment of words, and a machine learning (or supervised) classifier, which aims at inferring the annotation from a knowledge base containing already annotated abstracts. The authors concluded that the machine learning approach outperformed the morphosyntactic ones. But the macro reading task is fundamentally different, as we look for the best way to extract then combine ontology terms from a set of MEDLINE abstracts. The issue investigated in this paper is: to what extent the differences observed between two classifiers for a micro reading task are observed for a macro reading one? In particular, the redundancy hypothesis claims that the redundancy in large textual collections such as the Web or MEDLINE tends to smoothe performance differences across classifiers (Lin, 2007). To address this question, we compared a morphosyntactic and a machine learning classifiers for both tasks, focusing on the extraction of Gene Ontology (GO) terms, a controlled vocabulary for the characterization of proteins functions. The micro reading task consisted in extracting GO terms from a single MEDLINE abstract, as in the Trieschnigg et al’s work; the macro reading task consisted in extracting GO terms from a set of MEDLINE abstracts in order to answer to proteomics questions asked to the EAGLi QA system.