![\"\"](\"/assets/images/ebt_preprint_tta.png\")
Jente Houweling published her first PhD thesis chapter earlier this year:\n“ToxTempAssistant: using large language models to standardise cell-based toxicological test method descriptions”\n(doi:10.1080/2833373X.2026.2638036). So far, I have been blogging about\nmany of the articles on which I am (co-)author. To put it in context. To reflect on the work. I have been postponing\nwriting about this paper because there is a lot to reflect on. I will pick out two things. First, I look at the use\nof AI. The second is the unique, innovative publishing model of the journal where the article was published.\nIf you want to just see it in action, ToxTempAssistant is running\non the Virtual Human Platform for safety assessment.
\n\nBefore we go there, just a quick note on what ToxTemps are and ToxTempAssistent actually is:
\n\n\n\n\nThe ToxTemp template, based on OECD Guidance Document 211, standardises reporting for cell-based NAMs. However,\ncompleting its 77 questions constitutes a substantial bottleneck. The aim of this study is to introduce ToxTempAssistant,\na Large Language Model (LLM)-assisted web tool that supports toxicologists in drafting ToxTemp documents based on\nuser-supplied context documents. This study quantifies the tool’s baseline performance under controlled conditions.\nToxTempAssistant uses grounded, per-question prompting with mandatory source attribution.
\n
AI is very old. Arguably, my PhD thesis\nhad this as key topic, though I preferred to use to term chemometric or machine learning.\nCritical thinking has been essential to this field for a long time, and much of the PhD thesis is actually\nabout critically assessing the performance of the methods used in the thesis. There is decades of research\nhow you do this. Sadly, when it comes to Large Language Models (LLMs), these are not routinely used.
\n\nLLMs are indeed something new. I have seen natural language processing\nresearch when I was Cambridge with Peter Murray-Rust. The current LLMs are different, less deterministic, more probabilistic.\nFrom a chemometrics perspective, that makes sense. Language is complex, and not so deterministic in itself. Moreover,\nwhen representing words, sentences as numbers, you can integrate any resource (think data tables, images, etc). Even more,\ndigital representation was even more the central theme of my PhD thesis.
\n\nBut because of the nature of the method, the nature of how the commercial, better known models are trained (and the\nimpact on the notion of copyright), the impact on the climate emergency,\nthe black box that these LLMs often are, there are so many technical, scientific, and ethical reasons to stay away from them.
\n\nYou can write books about that. Literally (I did not read them yet):
\n\n(I have the feeling I am missing one title I wanted to highlight. If I remember, I will add it.)
\n\nAnd more here and here.\nAnd I am looking forward to reading Deep Unlearning: The Rise of AI and the Radicalization of a Tech Idealist.\nAlso, I recommend at least following Timnit Gebru and\nEmily Bender.
\n\nA year ago, I signed the Open Letter: Stop the Uncritical Adoption of AI Technologies in Academia,\nnow signed by more than 2,000 people (it is not too late).
\n\nSo, with all these things mind, I am happy that Jente did critically adopt LLMs in her work. The paper includes\nvarious experiments to explore the impact of various model parameters on generating nonsense. She also explored\nto use of alternative LLMs platforms, opening the option that some day it runs on other, more ethical platforms.\nToxTempAssistant, moreover, limits the material it takes information from, from a limited set of sources, provided\nby the users. Of course, the model itself is still training on resources with questionable provenance.
\n\nJente’s paper describes the use of positive and negative controls to put the performance in perspective.\nIn doing so, the paper formalizes how to evaluate the use of LLMs in situations where the LLM is used\nto summarize other reports, a common thing to do. This allows us to monitor the impact of, for example,\nnew LLM releases.
\n\nThe results are promising and various independent projects have shown interest in adoption. The ToxTemp\nreports are important for safety assessment: they provide essential context to experimental results and\nas such essential to the FAIR-ness of toxicology data.
\n\nThe ToxTempAssistant paper is published in the relatively new journal Evidence-Based Toxicology (EBT):
\n\n\n\n\nEvidence-Based Toxicology is a broad-focus, gold open-access journal, created to support the use of open science practices and\nevidence-based methods in toxicology and environmental health.
\n
So, CC-BY license (gold open access) and support for Open Science. And Open Peer Review, as we will see. Also,\nI understand it is not a diamond open access journal, so expect APCs.
\n\nThe journal has a community on Zenodo for preprints and peer-reviews.\nI think this is a really nice choice. Of course, a journal specific preprint server has downsides too. For example,\nif it gets rejected from EBT, do you use the EBT preprint server when submitting to another journal? Do you upload\na new version (after all, you should address some of the comments why it was rejected) to another preprint server?
\n\nThe EBT preprint gets a record and revisions are uploaded as new versions to the same Zenodo record\n(doi10.5281/zenodo.17192970):
\n\n
But because EBT uses open peer review, there is a parallel Zenodo entry with the reviews\n(doi10.5281/zenodo.17278785):
\n\n![\"\"](\"/assets/images/ebt_preprint_tta_reviews.png\")
The last version here is the acceptance notice:
\n\n![\"\"](\"/assets/images/ebt_preprint_tta_acceptance.png\")
One of the reviewer suggestions was to use the TRIPOD-LLM template\n(doi:10.1038/s41591-024-03425-5), which was included in a\nlater revision. That made a lot of sense and is a nice example how the journal actively works\non applying open science ideas. The journal webpage writes:
\n\n\n\n\nWe define “open science” as the set of practices aimed at improving the transparency, validity,\nreproducibility, and accessibility of scientific research, while promoting equality of\nopportunity to participate in and benefit from the products of said research.
\n
One comment here is that the template asks on what page that point of the checklist is discussed\nin the article. That is nice, but links it directly to the revision of the manuscript that\nform applies too, but that is not reported in the template. Not ideal. Then again, the point\nis the checking, so maybe not a big deal.
\n\nAnother comment is that it seems the journal website’s page for the article does not seem to actually\nlink to the preprints nor the peer-review reports (or acceptance note). So, in time, these\nopen science aspects of this article will likely get lost in time. Who will find those reports\nif the article does not cite them? This will require Taylor&Francis to modernize the publishing\nmodel, and I sincerely doubt that that will ever happen.
\n\nProf. Anne Kienhuis, one of the co-authors, suggested\nthis journal lead by Prof. Paul Whaley.\nAnd I am happy to have seen this new approach in action and like to thank Paul for pushing for\nthese innovations. I recommend trying it yourself for your next toxicology work.
\n\n