NCBI public repositories provide extensive resources for biomedical discovery. These resources are growing rapidly and have already a significant size (90M compounds and 200M substances in PubChem, 25M records in PubMed, etc.). There is significant variation in the amount of data linked to these records, with information being abundant in some cases and very limited or completely unavailable in others. Even when available, contributor-provided information is not always accurate, consistent, and complete. It is desirable to develop automated methods for cross-validating PubChem contributed information, rating its relevance and quality, and improving PubChem public presentation. A promising approach is to match the data from PubChem to entities found in publicly available documents (such as PubMed records, patents, records from various government agencies, etc.) to help build evidence and support for associations between entities. As a pilot project, we located chemical names in PubMed records, matched them to PubChem depositor-provided synonyms, and then used them to crosscheck the tens of millions of links between PubChem and PubMed records. We used text mining programs, which included: NextMove Software’s dictionary/grammar-based tool LeadMine and NCBI’s machine learning-based tool PubTator to find chemical entities (among others) in PubMed titles and abstracts. Then we matched these detected entities from text-mining to PubChem synonyms (chemical names) using very simple rules. While a straightforward approach allowed us to match chemical entities in more than 90% of PubMed records, only 0.25% of PubChem filtered synonyms had any match, covering 0.2% of compound records (320K PubChem names for more than 200K compounds matched to terms in more than 11M PubMed records). These results were encouraging but made us wonder: can we improve the matching procedure? There are several known reasons for name mismatches: variations in correct spelling; ambiguities in text transformations from one encoding to another (such as UTF8 to ASCII, etc.); spelling errors; and so on. We are currently investigating and will present results on combination of two methodologies: 1) Applying LeadMine name correction capabilities (applied to entities in the documents on one hand and PubChem Compound synonyms on the other hand); 2) Adaptive matching with some level of mismatch depending on the complexity of the name and presence of names of other compounds that are closely spelled.

Improving chemical names matching for verification, rating, and validation of PubChem compound records

253rd National Meeting (2017)

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