About ToxTempAssistant

ToxTempAssistant drafts ToxTemp descriptions of cell-based toxicological test methods from the documents you already have: protocols, SOPs, publications and lab notes. For each of the 77 ToxTemp questions, a large language model drafts an answer where your documents contain the information and cites its sources; you fill the gaps, review, edit and export the result.

Why ToxTemp, and why it takes so long

New Approach Methodologies (NAMs), such as human cell-based assays, can only support chemical safety assessment when they are documented transparently and completely. Most NAMs developed in academic labs are not OECD Test Guidelines, and inconsistent reporting makes them hard to evaluate and reuse.

The ToxTemp template (Krebs et al., ALTEX 2019) addresses this by turning the requirements of OECD Guidance Document 211 into 77 guided questions, from the cell source and culture conditions to the exposure scheme, endpoints, prediction model and transferability. The ToxTemp template is used by projects such as VHP4Safety, RISK-HUNT3R and ONTOX. Filling it in by hand can take toxicologists up to a week, often re-typing information that already exists in publications and protocols.

How it works

1. 1Upload your documents

   Give the assay a name and a short description of its purpose, test system and endpoint, then upload your context documents. Images inside PDF and Word files can be extracted as extra context.

2. 2Get draft answers

   The model drafts answers from the information in your documents and cites the document behind each statement. Where your documents don’t cover a question, it replies “Answer not found in documents.” instead of guessing, so expect some questions to be left for you.

3. 3Review and refine

   For most people, correcting a draft is quicker than writing from scratch. Edit and accept answers one by one, regenerate a single answer with extra instructions, and share the assay with colleagues in a workspace.

4. 4Export

   Download the finished ToxTemp as PDF, Word, HTML, Markdown, JSON, XML or LaTeX. The model and its parameters are recorded in the export metadata.

See it in action

This demo shows version 2.16.2, so the current version may look slightly different. Watch on PeerTube.

The 77 ToxTemp questions

The questions are kept word for word from the original template and grouped into 11 sections:

1. Overview
2. General information
3. Description of general features of the test system source
4. Definition of the test system as used in the method
5. Test method exposure scheme and endpoints
6. Handling details of the test method
7. Data management
8. Prediction model and toxicological application
9. Publication/validation status
10. Test method transferability
11. Safety, ethics and specific requirements

Read all 77 questions, with the template’s notes and examples

How accurate is it?

We evaluated ToxTempAssistant (v1.3.1) in a peer-reviewed study, published in Evidence-Based Toxicology (2026), under two controlled conditions: eight publicly available, expert-completed ToxTemps as input (does it recover the experts’ answers?) and eight out-of-scope documents (does it correctly decline to answer?). An answer counted as correct when its meaning closely matched the expert’s answer (cosine similarity above 0.6).

Completeness: share of answerable questions the model answered. Precision: share of those answers that were correct. Specificity: share of questions correctly declined for out-of-scope documents. Accuracy: correct answers and correct refusals combined.

Overall accuracy was similar across models, but they trade off differently: cautious models decline more often, while high-coverage models are more likely to answer from unrelated text. That is why every answer is a draft for expert review. The evaluation data and scripts are open, so the results can be reproduced.

Your data

• EU data residency: answers are generated by gpt-4o-mini, hosted by Microsoft in the EU through Azure AI Foundry.
• Your investigations, studies and assays are visible only to you, unless you share an investigation with colleagues through a workspace.
• Your documents are used to generate your draft. Letting the development team keep the uploaded files to benchmark and improve the tool is optional: you decide when you create an assay, and you can stop sharing them at any time under Privacy in the user menu.
• ToxTempAssistant is a documentation aid. It doesn’t make safety, medical or legal decisions, and you review every generated statement.

Getting access

ToxTempAssistant is in public beta. Sign up with your email address or ORCID iD; the maintainers admit new accounts. Every new account comes with a read-only demo assay, so you can explore a finished draft before uploading your own documents.

The free version uses gpt-4o-mini, hosted by Microsoft in the EU, which had the highest overall accuracy in our evaluation. Need a different model? Contact us and we can upgrade your account at our discretion.

Frequently asked questions

Team and funding

ToxTempAssistant was conceived at the 4th VHP4Safety Hackathon and is developed by Jente Houweling (RIVM; Maastricht University) and Matthias Arras, with Egon Willighagen, Danyel Jennen and Chris Evelo (Maastricht University) and Anne Kienhuis (RIVM; Utrecht University).

The work is part of the Virtual Human Platform for Safety Assessment (VHP4Safety) project, funded by the Dutch Research Council (NWO) through the Dutch Research Agenda programme “Research on Routes by Consortia” (NWA-ORC 1292.19.272).

How to cite

If ToxTempAssistant helps your work, please cite:

Houweling, J. M., Arras, M. M. L., Willighagen, E. L., Jennen, D. G. J., Evelo, C. T., & Kienhuis, A. S. (2026). ToxTempAssistant: using large language models to standardise cell-based toxicological test method descriptions. Evidence-Based Toxicology, 4(1).

Please also consider citing ToxTemp (Krebs et al., 2019).

The software is open source under the GNU Affero General Public License v3.0: source code on GitHub, with archived releases on Zenodo. Questions or ideas? Open an issue on GitHub.