Using Biomedical Literature Mining to Consolidate the Set of KnownHuman Protein–Protein Interactions

摘要

This paper presents the results of a largescaleeffort to construct a comprehensivedatabase of known human protein interactionsby combining and linking knowninteractions from existing databases andthen adding to them by automatically miningadditional interactions from 750,000Medline abstracts. The end result is anetwork of 31,609 interactions amongst7,748 proteins. The text mining systemfirst identifies protein names in thetext using a trained Conditional RandomField (CRF) and then identifies interactionsthrough a filtered co-citation analysis.We also report two new strategiesfor mining interactions, either by findingexplicit statements of interactions in thetext using learned pattern-based rules ora Support-Vector Machine using a stringkernel. Using information in existing ontologies,the automatically extracted datais shown to be of equivalent accuracy tomanually curated data sets.