EnzChemRED, a rich enzyme chemistry relation extraction dataset

Area of Science:

• Biochemistry
• Bioinformatics
• Computational Biology

Background:

• Expert curation of enzyme functions from scientific literature is crucial for FAIR open knowledgebases.
• Current manual curation cannot match the pace of new discoveries and publications.
• Automated methods are needed to assist and scale enzyme curation efforts.

Purpose of the Study:

• To introduce EnzChemRED (Enzyme Chemistry Relation Extraction Dataset), a novel dataset for training and benchmarking Natural Language Processing (NLP) models.
• To facilitate the development of NLP methods, including large language models, for assisting enzyme curation.
• To enable scalable extraction of enzyme-catalyzed reaction information from scientific literature.

Main Methods:

• EnzChemRED comprises 1,210 expert-curated PubMed abstracts.
• Enzymes and catalyzed chemical reactions are annotated using UniProtKB and ChEBI identifiers.
• Language models were fine-tuned using EnzChemRED for entity recognition and relation extraction.

Main Results:

• Fine-tuned language models achieved high F1 scores: 86.30% for protein and chemical identification.
• Extraction of chemical conversions reached an 86.66% F1 score.
• Extraction of catalyzing enzymes achieved an 83.79% F1 score.

Conclusions:

• EnzChemRED significantly enhances NLP model performance for enzyme function extraction.
• The developed methods can process PubMed-scale abstracts to create a draft map of enzyme functions.
• This resource will guide curation efforts in UniProtKB and Rhea, improving biological knowledgebases.