Extracting chemical reactions from text using Snorkel.

Area of Science:

• Biochemistry
• Bioinformatics
• Computational Biology

Background:

• Understanding biological processes relies on enzymatic and chemical reactions.
• Existing chemical reaction databases struggle to keep pace with the rapid growth of biomedical literature.
• Conventional text mining requires substantial labeled data, limiting scalability for new relationship types and large document sets.

Purpose of the Study:

• To develop and evaluate a weakly supervised learning framework for extracting chemical reaction relationships from biomedical literature.
• To define a chemical reaction as the transformation of chemical A to chemical B for extraction purposes.
• To assess the system's performance on both curated and general biomedical text corpora.

Main Methods:

• Applied Snorkel, a weakly supervised learning framework, to identify chemical reaction relationships.
• Evaluated the system on two corpora: MetaCyc_Corpus (curated bacteria-related abstracts) and Bacteria_Corpus (general abstracts related to bacteria).
• Defined chemical reaction extraction as identifying the transformation of chemical A to chemical B.

Main Results:

• Achieved 84% precision and 41% recall (55% F1 score) on the MetaCyc_Corpus.
• On the larger Bacteria_Corpus, precision was 62% and recall was 37% (46% F1 score).
• Extracted 6,871 chemical reaction relationships from nine million candidates in the Bacteria_Corpus, demonstrating scalability.

Conclusions:

• Successfully built a database of chemical reaction relationships from nearly 900,000 abstracts using a weakly supervised approach.
• Demonstrated the generalizability of the chemical reaction extraction application from a specialized corpus to a broader set of scientific articles.
• The developed method bypasses the need for large labeled training datasets, offering a scalable solution for biomedical literature analysis.