Identifying Relevant Full-Text Articles for GO Annotation Without MeSH Terms

Gene Ontology (GO) is a controlled vocabulary. Given a gene product, GO enables scientists to clearly and unambiguously describe specific molecular functions of the gene product, specific biological processes in which it is involved, and specific cellular components to which it is localized. In this paper, we present our approach to identifying which papers have experimental evidence warranting annotation with GO codes. The training data set contains 375 relevant full-text articles and 5,462 irrelevant ones, and the test data set contains 420 positive full-text articles and 5,623 negative ones. We regarded this problem as a binary classification problem, and employed Support Vector Machines (SVMs) to distinguish positive articles from negative ones. Title, abstract, figure/table captions, and three standard sections – Results, Discussion, and Conclusion were the targets of feature extraction. Without incorporating MeSH (Medical Subject Headings) terms as part of the features, our system achieved 0.381 in Normalized Utility measure.