FuncFetch: an LLM-assisted workflow enables mining thousands of enzyme-substrate interactions from published...

Area of Science:

• Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• Thousands of publicly available genomes contain genes with poorly defined functions.
• A gap exists between experimentally characterized protein activities and database entries, largely due to slow biocuration processes.

Purpose of the Study:

• To develop and validate an automated workflow, FuncFetch, for accelerated text-mining of protein and enzyme activities from scientific literature.
• To leverage large language models to overcome the bottleneck in biocuration and improve the deposition of functional information.

Main Methods:

• FuncFetch integrates NCBI E-Utilities, OpenAI's GPT-4, and Zotero for manuscript screening and data extraction.
• The workflow extracts species, enzyme names, sequence identifiers, substrates, and products.
• Extensive validation and quality analyses were performed, including comparison against manually curated datasets.

Main Results:

• GPT-4 demonstrated high precision and recall in identifying characterized enzyme activities within abstracts.
• FuncFetch successfully screened 26,543 papers for nine plant enzyme families, retrieving 32,605 entries from 5,459 papers.
• Analysis revealed approximately 70% of experimentally characterized enzymes remain uncurated, with identified errors highlighting the need for continued manual curation.

Conclusions:

• FuncFetch significantly advances biocuration by automating the extraction of enzyme functional data.
• The workflow facilitates the creation of comprehensive functional fingerprints for enzyme families.
• FuncFetch lays the groundwork for predicting functions of uncharacterized enzymes and addressing the vast uncurated data in public domains.