PubChem knowledge panels summarize important relationships between chemicals, genes, proteins and diseases uncovered by analysis of term co-occurrences in biomedical literature [1]. A part of PubChem compound, target and disease pages [2], they are built using automatic annotations of PubMed records by natural language processing (NLP) software LeadMine [3], and term matching to PubChem and other public databases. The most relevant information is identified using statistical analysis, summarized, and visualized using relevance-based sampling, allowing a compact, highly informative and easy-to-comprehend knowledge representation. Our ongoing project includes quality enhancements in both intelligent data analysis and interpretation of annotations provided by NLP software. In statistical analysis, we introduced an adjusted entity co-occurrence score, with a correction for random co-occurrence of frequently-annotated entities applied to the record count. To improve the annotation quality, we are working on more accurate type identification and matching to database records through more sophisticated postprocessing that includes processing of annotations of different types together, taking context into account, using curated data, and iteratively improving annotation of the whole database by discovering term-use patterns. Besides improving quality of knowledge summarization in PubChem papers, our efforts allow to better understand the typical use, rank synonyms, correct submitter-provided names, and verify submitter-provided links in our databases, helping to automate biocuration efforts.

Enhancing data-driven summarization of relations between chemicals, genes, proteins, and diseases based on text mining of biomedical literature

2019 Fall National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.