Automated Continuous LLM-driven Biomedical Literature Mining for a Large-Scale Hepatotoxicity Knowledge Base

An automated LLM pipeline built a hepatotoxicity knowledge base with 189,212 findings across 17,002 chemicals from 53,575 PubMed/PMC articles (1995–2026), covering 235 biomarkers.

• Why it matters: Reliable liver and mitochondrial disturbance evidence is needed for drug safety and environmental risk assessment, but training data are limited and often hand-curated.
• What they did: The pipeline queries PubMed/PMC, retrieves full texts or abstracts, prompts an LLM to extract effects into a fixed schema with controlled vocabularies, and de-duplicates into a continuously updated database.
• The result: In a manual review of 217 findings from 35 articles, 208 findings from 33 articles were fully correct, enabling biomarker-specific structure-based QSAR models where classic baselines remain competitive.

AI summary of the abstract · may contain errors · not medical advice · read the paper.