ALURES (Animal Use Reporting System) contains Non-Technical Summaries (NTS) that describe the objectives, study types, and procedures used in animal experiments across Europe. To analyze the methods reported in these documents, we developed a four-stage pipeline as part of WP3 – Extension to studies other than systemic toxicity within the IHI VICT3R Project (Developing and implementing Virtual Control Groups to reduce animal use in Toxicology Research).

VICT3R is a 42-month initiative that began in September 2024, funded by the Innovative Health Initiative Joint Undertaking (IHI JU). The project aims to develop, validate, and support the regulatory acceptance of Virtual Control Groups (VCGs) — a data-driven approach to reducing animal use in toxicology research.

We selected a dataset from the ALURES NTS web portal using the following filters:

• Species: Mice (Mus musculus)
• EU Submission: Yes
• Purpose of the project: Basic Research – Nervous System

Stage 1 – Extraction: LLMs are applied to extract descriptions of experimental procedures from NTS records. A multi-agent debate framework enhances accuracy across multiple languages and standardizes all outputs into English.

Stage 2 – Categorization: The extracted procedures are classified into five categories by LLMs:

• Experimental Frameworks
• Manipulations / Procedures
• Tests/Apparatus
• Data Acquisition Methods

Stage 3 – Clusterization: Conceptually similar procedures are merged to reduce redundancy arising from lexical variation. This step uses hierarchical agglomerative clustering based on semantic embeddings.

Stage 4 – Expert Curation: The resulting taxonomy was presented in an interactive browser-based tree. Domain experts refined it by adding, removing, renaming, or reorganizing nodes, ensuring both accuracy and conceptual clarity.