Extraction of gene regulatory networks from biological literature

A gene regulatory network (GRN) is a network of interacting cellular components. The components are genes and their products, and the interactions represent regulatory relationships among genes, specifically activation and inhibition of gene expression, under certain conditions. Many regulatory relationships are known in the literature. However, assembling isolated relationships into networks is a challenging, albeit important task. We have developed a system for automatically extracting GRNs from the literature. As a first filtering step, our system makes use of Textpresso (an existing ontology-based system for information retrieval from scientific literature) to identify document sentences that contain genes of interest and regulatory relationships involving those genes. The ontology-annotated Textpresso search results, provided in XML format, are examined for regular expressions that specify regulatory relationships. We use a set of positively labeled relations between genes (equivalently, pairs of genes that exhibit a regulatory relationship) to infer scores for patterns present in such relations. The pattern scores are further used to compute total scores for putative relationships between pairs of genes present in a sentence. Pairs with total score greater than a threshold are assumed to be positive. Availability of very little amounts of labeled relationships along with large amounts of unlabeled data motivated us to study the performance of our approach also in a semi-supervised setting. We show that unlabeled data can, in some cases, improve the performance of our approach.