The Cinderella of Biological Data Integration:Addressing Some of the Challenges of Entity and Relationship Mining from Patent Sources

摘要

Most of the global corpus of medicinal chemistry data is only published in patents. However, extracting this from patent documents and subsequent integration with literature and database sources poses unique challenges. This work presents the investigation of an extensive full-text patent resource, including automated name-to-chemical structure conversion, licensed by AstraZeneca via a consortium arrangement with IBM. Our initial focus was identifying protein targets in patent titles linked to extracted bioactive compounds. We benchmarked target recognition strategies against target-assay-compound relationships manually curated from patents by GVKBIO. By analysis of word frequencies and protein names we assessed the false-negative problem of targets not specified in titles and false-positives from non-target proteins in titles. We also examined the time-signals for selected target and non-target names by year of patent publication. Our results exemplify problems and some solutions for extracting data from this source.