Elamur Research⌕
☰
home / papers / 2026 / ToxReason: A Benchmark for Mechanistic Chemical Toxicity Reasoning via Adverse Outcome Pathway

ToxReason: A Benchmark for Mechanistic Chemical Toxicity Reasoning via Adverse Outcome Pathway

Paper · 2026 · Jueon Park, WonJune Jang, Chanhwi Kim et al.

Summary

ToxReason is a benchmark that tests whether LLMs can explain chemical toxicity through biologically plausible causal mechanisms, not only predict whether a compound is toxic. Existing datasets such as Tox21 and ClinTox only check outcome labels, so a model can write a fluent but mechanistically wrong rationale and go undetected. The authors build the benchmark on the Adverse Outcome Pathway (AOP) framework, which traces toxicity from a molecular initiating event (MIE) through key events to an organ-level adverse outcome (AO). They combine three sources:

  • AOPs for liver, heart and kidney from AOP-Wiki
  • disease–chemical links from CTD
  • target activity data from ChEMBL

Each chemical's MIE is inferred from structurally similar compounds with measured activity. Several LLMs are scored on prediction accuracy and on reasoning quality, judged by an LLM evaluator. The main finding is that high accuracy often comes without sound mechanistic reasoning. Training that explicitly rewards reasoning, via supervised fine-tuning plus reinforcement learning, lets a compact 4B model beat larger state-of-the-art models on prediction while also reasoning better. This matters most in early drug discovery, where clinical evidence is not yet available.

Key points

  • Scale and sources: 23 AOPs from AOP-Wiki (Release 2.7) covering 25 MIE protein targets; the test benchmark has 193 chemicals. Data come from CTD (Nov. 2025 release) and ChEMBL (v36).
  • MIE labelling rule: a chemical counts as activating or inhibiting a target if its EC50 or IC50 is below 10,000 nM. For query chemicals, the direction of action is set by majority vote over structurally similar compounds.
  • Reasoning evaluation: an LLM judge scores reasoning on four complementary metrics. Its overall score agrees well with a Needleman–Wunsch alignment against reference AOP steps (Pearson r = 0.703, Spearman ρ = 0.837).
  • Training results: three learning paradigms are compared. The GRPO-trained ToxReason-4B-GRPO produces step-by-step explanations that follow the reference causal chain, for example glucocorticoid receptor activation leading to hepatic steatosis.
  • Limitations: the benchmark covers only three organs, its evaluation set is small, its MIE space is limited to the selected AOPs, and judge-based scores should be read as relative rather than absolute.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Korea UniversitySeoul · South Korea26 works · 9,490 citations🎓The University of Texas Health Science Center at HoustonHouston · United States5 works · 36 citations🎓Myongji UniversitySeoul · South Korea2 works

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksComputational Theory and Mathematics · 252 worksMaterials Science · 26 worksMaterials Chemistry · 26 worksSmall Animals · 1 workVeterinary · 1 work

🧭 Topics 4Research subjects and languages the work is about.Reasoning · 1,426 worksEvaluation & benchmarks · 1,220 worksBias, fairness & safety · 517 worksInterpretability · 130 works

Cite

Cite as Jueon Park, WonJune Jang, Chanhwi Kim, Yein Park and Jaewoo Kang. 2026. ToxReason: A Benchmark for Mechanistic Chemical Toxicity Reasoning via Adverse Outcome Pathway. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.977

@inproceedings{park2026toxreason,
  title = {{ToxReason: A Benchmark for Mechanistic Chemical Toxicity Reasoning via Adverse Outcome Pathway}},
  author = {Jueon Park, WonJune Jang, Chanhwi Kim, Yein Park and Jaewoo Kang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.977},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.977}
}

References

48 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Gerald T Ankley, Richard S Bennett, Russell J Erickson, Dale J Hoff, Michael W Hornung, Rodney D Johnson, David R Mount, John W Nichols, Christine L Russom, Patricia K Schmieder, and 1 others. 2010. Adverse outcome pathways: a conceptual framework to support ecotoxicology research and risk assessmen… find
  2. Anthropic. 2025a. Introducing Claude Haiku 4.5. https://www.anthropic.com/news/ claude-haiku-4-5. Released: 2025-10-16. ↗
  3. Anthropic. 2025b. Introducing Claude Sonnet 4.5. https://www.anthropic.com/news/ claude-sonnet-4-5. Released: 2025-09-30. ↗
  4. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901.
  5. Yi-Qi Chen, Tao Yu, Zheng-Qi Song, Chen-Yu Wang, Jiang-Tao Luo, Yong Xiao, Heng Qiu, Qing-Qing Wang, and Hai-Ming Jin. 2025. Application of large language models in drug-induced osteotoxicity prediction. Journal of Chemical Information and Modeling, 65(7):3370–3379. find
  6. Allan P. Davis, Thomas C. Wiegers, Daniela Sciaky, Frances Barkalow, Matthew Strong, Benjamin Wyatt, Jill Wiegers, Rebecca McMorran, Shakil Abrar, and Carolyn J. Mattingly. 2025. Comparative toxicogenomics database’s 20th anniversary: update 2025. Nucleic Acids Research. Advance access, October 10,… find
  7. Lee R Dice. 1945. Measures of the amount of ecologic association between species. Ecology, 26(3):297– 302. find
  8. Domenico Gadaleta, Marina Garcia de Lomana, Eva Serrano-Candelas, Rita Ortega-Vallbona, Rafael Gozalbes, Alessandra Roncaglioni, and Emilio Benfenati. 2024. Quantitative structure–activity relationships of chemical bioactivity toward proteins associated with molecular initiating events of organ-spec… find
  9. Kaitlyn M Gayvert, Neel S Madhukar, and Olivier Elemento. 2016. A data-driven approach to predicting successes and failures of clinical trials. Cell chemical biology, 23(10):1294–1301. find
  10. Google DeepMind. 2025. Gemini 3 flash model card. find
  11. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  12. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  13. Ruili Huang, Menghang Xia, Dac-Trung Nguyen, Tongan Zhao, Srilatha Sakamuru, Jinghua Zhao, Sampada A Shahane, Anna Rossoshek, and Anton Simeonov. 2016. Tox21challenge to build predictive models of nuclear receptor and stress response pathways as mediated by exposure to environmental chemicals and dr… find
  14. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276.
  15. Hyeon Hwang, Yewon Cho, Chanwoong Yoon, Yein Park, Minju Song, Kyungjae Lee, Gangwoo Kim, and Jaewoo Kang. 2025. Assessing llm reasoning steps via principal knowledge grounding. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 19925–19948. find
  16. Yunhui Jang, Jaehyung Kim, and Sungsoo Ahn. 2025. Structural reasoning improves molecular understanding of llm. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 21016–21036. find
  17. Dongki Kim, Wonbin Lee, and Sung Ju Hwang. 2025. Mol-llama: Towards general understanding of molecules in large molecular language model. arXiv preprint arXiv:2502.13449. — [arXiv:2502.13449](https://arxiv.org/abs/2502.13449) find
  18. Marcel Leist, Ahmed Ghallab, Rabea Graepel, Rosemarie Marchan, Reham Hassan, Susanne Hougaard Bennekou, Alice Limonciel, Mathieu Vinken, Stefan Schildknecht, Tanja Waldmann, and 1 others. 2017. Adverse outcome pathways: opportunities, limitations and open questions. Archives of Toxicology, 91(11):34… find
  19. Hao Li, He Cao, Bin Feng, Yanjun Shao, Xiangru Tang, Zhiyuan Yan, Li Yuan, Yonghong Tian, and Yu Li. 2025. Beyond chemical qa: Evaluating llm’s chemical reasoning with modular chemical operations. arXiv preprint arXiv:2505.21318. — [arXiv:2505.21318](https://arxiv.org/abs/2505.21318) find
  20. Xuan Liu, Siru Ouyang, Xianrui Zhong, Jiawei Han, and Huimin Zhao. 2025. Fgbench: A dataset and benchmark for molecular property reasoning at functional group-level in large language models. arXiv preprint arXiv:2508.01055. find
  21. Saul B Needleman and Christian D Wunsch. 1970. A general method applicable to the search for similarities in the amino acid sequence of two proteins. Journal of molecular biology, 48(3):443–453. find
  22. Minh-Vuong Nguyen, Linhao Luo, Fatemeh Shiri, Dinh Phung, Yuan-Fang Li, Thuy Vu, and Gholamreza Haffari. 2024. Direct evaluation of chain-of-thought in multi-hop reasoning with knowledge graphs. In Findings of the Association for Computational Linguistics: ACL 2024, pages 2862–2883. find
  23. OpenAI. 2025a. Gpt-5 system card. find
  24. OpenAI. 2025b. Gpt-5.1 instant and gpt-5.1 thinking system card addendum. find
  25. OpenAI. 2025c. Openai o3 and o4-mini system card. find
  26. Jueon Park, Yein Park, Minju Song, Soyon Park, Donghyeon Lee, Seungheun Baek, and Jaewoo Kang. 2025. Cotox: Chain-of-thought-based molecular toxicity reasoning and prediction. arXiv preprint arXiv:2508.03159. — [arXiv:2508.03159](https://arxiv.org/abs/2508.03159) find
  27. Payal Rana, Stephen Kogut, Xuerong Wen, Fatemeh Akhlaghi, and Michael D Aleo. 2020. Most influential physicochemical and in vitro assay descriptors for hepatotoxicity and nephrotoxicity prediction. Chemical Research in Toxicology, 33(7):1780–1790. find
  28. Laura Aliisa Saarimäki, Michele Fratello, Alisa Pavel, Seela Korpilähde, Jenni Leppänen, Angela Serra, and Dario Greco. 2023. A curated gene and biological system annotation of adverse outcome pathways related to human health. Scientific data, 10(1):409. find
  29. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, and 1 others. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.
  30. Jacob Silberg, Kyle Swanson, Elana Simon, Angela Zhang, Zaniar Ghazizadeh, Scott Ogden, Hisham Hamadeh, and James Y Zou. 2024. Unitox: leveraging llms to curate a unified dataset of drug-induced toxicity from fda labels. Advances in Neural Information Processing Systems, 37:12078–12093. find
  31. Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, and 1 others. 2025. Gemma 3 technical report. arXiv preprint arXiv:2503.19786. find
  32. Yoshihiro Uesawa. 2024. Efficiency of pharmaceutical toxicity prediction in computational toxicology. Toxicological Research, 40(1):1–9. find
  33. Mathieu Vinken, Dries Knapen, Lucia Vergauwen, Jan G Hengstler, Michelle Angrish, and Maurice Whelan. 2017. Adverse outcome pathways: a concise introduction for toxicologists. Archives of toxicology, 91(11):3697–3707. find
  34. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171.
  35. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824– 24837.
  36. David Weininger. 1988. Smiles, a chemical language and information system. 1. introduction to methodology and encoding rules. Journal of chemical information and computer sciences, 28(1):31–36. find
  37. Ziting Xian, Jiawei Gu, Lingbo Li, and Shangsong Liang. 2025. Molrag: unlocking the power of large language models for molecular property prediction. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15513–15531. find
  38. Tuan Xu, Leihong Wu, Menghang Xia, Anton Simeonov, and Ruili Huang. 2020. Systematic identification of molecular targets and pathways related to human organ level toxicity. Chemical research in toxicology, 34(2):412–421. find
  39. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025a. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  40. An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, and 1 others. 2024. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115. find
  41. Hengzheng Yang, Jian Xiu, Weiqi Yan, Kaifeng Liu, Huizi Cui, Zhibang Wang, Qizheng He, Yilin Gao, and Weiwei Han. 2025b. Large language models as tools for molecular toxicity prediction: Ai insights into cardiotoxicity. Journal of Chemical Information and Modeling, 65(5):2268–2282. find
  42. Barbara Zdrazil, Eloy Felix, Fiona Hunter, Emma J Manners, James Blackshaw, Sybilla Corbett, Marleen De Veij, Harris Ioannidis, David Mendez Lopez, Juan F Mosquera, and 1 others. 2024. The chembl database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods. N… find
  43. Jiangyan Zhang, Haolin Li, Yuncong Zhang, Junyang Huang, Liping Ren, Chuantao Zhang, Quan Zou, and Yang Zhang. 2025. Computational toxicology in drug discovery: applications of artificial intelligence in admet and toxicity prediction. Briefings in Bioinformatics, 26(5):bbaf533. find
  44. Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Guoyin Wang, and 1 others. 2023. Instruction tuning for large language models: A survey. ACM Computing Surveys.
  45. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, and 1 others. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in neural information processing systems, 36:46595–46623.
  46. Yizhen Zheng, Huan Yee Koh, Jiaxin Ju, Madeleine Yang, Lauren T May, Geoffrey I Webb, Li Li, Shirui Pan, and George Church. 2025. Large language models for drug discovery and development. Patterns, 6(10). find
  47. Jiaxi Zhuang, Yaorui Shi, Jue Hou, Yunong He, Mingwei Ye, Mingjun Xu, Yuming Su, Linfeng Zhang, Ying Qian, Guolin Ke, and 1 others. 2025. Reasoning-enhanced large language models for molecular property prediction. arXiv preprint arXiv:2510.10248. — [arXiv:2510.10248](https://arxiv.org/abs/2510.10248) find
  48. Johanna Zilliacus, Monica K Draskau, Hanna KL Johansson, Terje Svingen, and Anna Beronius. 2024. Building an adverse outcome pathway network for estrogen-, androgen-and steroidogenesis-mediated reproductive toxicity. Frontiers in Toxicology, 6:1357717. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/park-2026-toxreason-benchmark-mechanistic-chemical-toxicity/