Assisting document triage for human kinome curation via machine learning

In the era of data explosion, the increasing frequency of published articles presents unorthodox challenges to fulfill specific curation requirements for bio-literature databases. Recognizing these demands, we designed a document triage system with automatic methods that can improve efficiency to retrieve the most relevant articles in curation workflows and reduce workloads for biocurators. Since the BioCreative VI (2017), we have implemented texting mining processing in our system in hopes of providing higher effectiveness for curating articles related to human kinase proteins. We tested several machine learning methods together with state-of-the-art concept extraction tools. For features, we extracted rich co-occurrence and linguistic information to model the curation process of human kinome articles by the neXtProt database. As shown in the official evaluation on the human kinome curation task in BioCreative VI, our system can effectively retrieve 5.2 and 6.5 kinase articles with the relevant disease (DIS) and biological process (BP) information, respectively, among the top 100 returned results. Comparing to neXtA5, our system demonstrates significant improvements in prioritizing kinome-related articles as follows: our system achieves 0.458 and 0.109 for the DIS axis whereas the neXtA5’s best-reported mean average precision (MAP) and maximum precision observed are 0.41 and 0.04. Our system also outperforms the neXtA5 in retrieving BP axis with 0.195 for MAP and the neXtA5’s reported value was 0.11. These results suggest that our system may be able to assist neXtProt biocurators in practice.