Systematic analysis of hepatotoxicity: combining literature mining and AI language models

Area of Science:

• Toxicology
• Computational Chemistry
• Bioinformatics

Background:

• The volume of toxicological literature is rapidly increasing, posing challenges for researchers.
• Synthesizing vast amounts of published information is crucial for staying current.

Purpose of the Study:

• To automatically identify potential hepatotoxicants from over 50,000 compounds.
• To leverage scientific publications and knowledge for toxicity assessment.

Main Methods:

• Comparison of three automatic information extraction methods: text mining, word embeddings, and large language models.
• Calculation of hepatotoxicity scores for over 50,000 compounds.
• Assessment of method performance using Drug-Induced Liver Injury (DILI) as a use case.

Main Results:

• Text mining achieved an Area Under the Curve (AUC) of 0.8 for DILI validation.
• Large language models demonstrated superior performance with an AUC of 0.85.
• Combining methods yielded the highest AUC of 0.87 for DILI validation.

Conclusions:

• Automated text mining successfully assesses compound toxicity.
• Large language models, particularly with prompt engineering, offer the best performance.
• Findings are available for download to advance toxicity assessment research.