Transforming animal study toxicology reports into structured, harmonized data using large language models.

Area of Science:

• Computational toxicology
• Data science in drug safety
• Pharmacology and toxicology

Background:

• Preclinical toxicology reports contain valuable expert interpretations but are often unstructured.
• Unstructured data limits systematic reuse and integration with computational safety approaches.

Purpose of the Study:

• To develop a Large Language Model (LLM)-supported pipeline for converting unstructured toxicology reports into structured, machine-readable datasets.
• To harmonize extracted data with Standard for Exchange of Nonclinical Data (SEND) terminology.

Main Methods:

• The pipeline integrates automated document preprocessing, section identification, schema-constrained information extraction, and semantic harmonization.
• Targeted human curation complements the automated processes.
• System performance was evaluated on 200 Roche toxicology study reports.

Main Results:

• The pipeline demonstrated strong extraction performance across domains, with high sensitivity and precision for most parameters.
• Histopathology, organ weight, and No Observed Adverse Effect Level (NOAEL)-related endpoints showed high robustness (sensitivity >95%, precision >97%).
• Lower performance for certain parameters was attributed to heterogeneous reporting, not LLM limitations.

Conclusions:

• LLM-assisted extraction reliably captures expert toxicological interpretations at scale.
• Structured datasets enable cross-study querying, identification of toxicological liabilities, and development of predictive toxicology models.
• This approach provides a foundation for data-centric safety assessment and translational toxicology research.