GeneView – Gene-Centric Ranking of Biomedical Text

Background: Life scientists spend a great amount of time searching for gene-specific information. It is widely acknowledged that research results are primarily published in scientific literature and current curation efforts can not keep up with the fast increase of such literature. It can therefore be estimated that the plethora of gene-specific knowledge is still hidden in large text repositories like MEDLINE. Searching text data sources is difficult, as user queries are usually ambiguous and lead to hundreds of results. Faced with such a number of relevant publications, an appropriate article ranking is important. PubMed, for example, ranks articles per default by indexing date, making it difficult to find seminal papers about a specific topic. In this paper, we introduce GeneView, a genecentric text mining application capable of searching, ranking, and visualizing biomedical publications. Results: Our ranking algorithm relies on the assumption that the relevance of a gene for a specific article depends on the frequency with which it is mentioned and on the sections it appears in. For ranking we introduce a simple evaluation strategy by using the NCBI Gene2Pubmed mapping as gold-standard. This strategy is used to evaluate different section specific rankers, where the best one achieves on average a precision of 75.5 %.The evaluation further confirms our expectations, that sections like title, abstract and result are more relevant for gene specific ranking than others. Surprisingly, incorporation of figureand table-captions decreased the quality of ranking results.