Comprehensive Characterization of Oxidative Stress-Modulating Chemicals Using GPT-Based Text Mining

Area of Science:

• Toxicology and Environmental Health
• Computational Chemistry
• Bioinformatics

Background:

• Limited toxicological databases hinder environmental pollutant screening.
• Oxidative stress characterization is challenging yet crucial for disease research.
• Large language models (LLMs) offer automated information extraction from scientific texts.

Purpose of the Study:

• To develop and evaluate an LLM-based workflow for extracting oxidative stress information.
• To build a comprehensive dataset of prooxidants and antioxidants from literature.
• To identify potential toxicological substructures associated with oxidative stress.

Main Methods:

• Utilized GPT-4 for text mining of scientific articles on oxidative stress.
• Implemented a workflow including data collection, text preprocessing, and prompt engineering.
• Performed structural alert analysis on identified compounds.

Main Results:

• Extracted 17,780 records from 7,166 articles, covering 2,558 compounds.
• Identified 1,416 prooxidants (pharmaceuticals, pesticides, metals) and 1,102 antioxidants (pharmaceuticals, flavonoids).
• Discovered key prooxidant (e.g., chlorobenzene) and antioxidant (e.g., flavonoid) substructures.

Conclusions:

• LLM-based text mining is a feasible and cost-efficient method for building toxicological databases.
• The extracted information can significantly advance environmental and health research.
• This approach facilitates the identification of hazardous compounds and their mechanisms.