The BioCreative VI Precision Medicine Track corpus Selection , annotation and curation of protein-protein interactions affected by mutations in scientific literature

The Precision Medicine Track in BioCreative VI aims to bring together the biomedical text mining community for a novel challenge: mining the biomedical literature in search of information of value to precision medicine initiatives such as mutations disrupting/affecting protein-protein interactions (PPI). The Precision Medicine track is organized into two tasks: 1) the triage task – focusing on selection of relevant PubMed articles describing PPI affected by mutations, and 2) the relation extraction task – focusing on extracting the interacting gene pairs for the interactions that are affected by the presence of a mutation. To support this track with an effective training dataset and limited curator time, the track organizers used a two-staged approach. First, for the creation of the training dataset, the organizers and curators worked on leveraging the information from expertly curated and publicly available PPI databases, augmenting it with a set of articles selected via publicly available state-of-the-art text mining tools. 4,082 PubMed articles were thus carefully reviewed, annotated and released for system development. They contained 1,729 articles labelled positive for curation, out of which, 597 contained 752 curated relations. The second stage pertained to the creation of the testing dataset, which consisted of 1,464 PubMed articles, previously not curated in any of the known PPI databases. These articles were highly likely to describe PPI and sequence variants according to several text mining tests. Each article in the testing dataset was annotated by at least two curators, for relevance relation extraction. Five BioGRID annotators participated and reviewed more than 600 articles each. The testing set contained 730 articles labelled positive for curation, out of which, 688 articles contained 930 curated relations. We detail here the data collection, manual review and annotation process. We give a report on the precision medicine track corpus characteristics. This analysis will provide useful information to developers and researchers for comparing and developing innovative text mining approaches for the https://thebiogrid.org/ BioCreative VI challenge and other Precision Medicine related applications. Keywords—corpus creation, manual annotation, protein-protein interaction, mutation, relation extraction, information extraction.