Preclinical toxicology study reports contain the expert interpretations required to distinguish test article-related effects from incidental findings, yet these conclusions often remain embedded in unstructured text form that limit systematic reuse and integration with computational safety approaches. To address this gap, we developed a Large Language Model (LLM) - supported pipeline that converts toxicology reports into structured, machine-readable datasets harmonized with SEND terminology. The pipeline combines automated document preprocessing, section identification, schema-constrained information extraction, and semantic harmonization, complemented by targeted human curation. We evaluated the system performance using 200 Roche toxicology study reports, encompassing clinical pathology, histopathology, organ weights, exposure data, and study-level conclusions. Across domains, extraction performance was strong, characterized by consistently high sensitivity and precision for most parameters. Histopathology, organ weight, and NOAEL-related endpoints demonstrated the greatest robustness, with sensitivity typically above 95% and precision frequently exceeding 97%. Lower performance for parameters such as route of administration and substance identifiers reflected heterogeneous reporting practices rather than LLM-based method limitations. The structured datasets generated by this pipeline enable cross-study querying, identification of compounds with defined toxicological liabilities, integration with raw SEND data, and development of high-quality labels for predictive toxicology models. Practical utility has been demonstrated through representative use cases. These results demonstrate that LLM-assisted extraction can reliably capture expert toxicological interpretations at scale and provide a foundation for data-centric safety assessment, strategic decisions, reverse and forward translational toxicology research.
Doktorova et al. (Thu,) studied this question.