---
license: cc-by-4.0
task_categories:
- text-generation
- text-classification
language:
- en
tags:
- chemistry
- toxicity
- drug-discovery
- llm
- reasoning
- aop
- benchmark
pretty_name: ToxReason
size_categories:
- 10K<n<100K

configs:
- config_name: default
  data_files:
  - split: mie_matched
    path: MIE_matched.jsonl
  - split: mie_ao_matched
    path: MIE_AO_matched.jsonl
  - split: evaluation
    path: evaluation.jsonl
---
# ToxReason

🚀 Accepted at ACL 2026 Findings

ToxReason is a benchmark dataset for mechanistic chemical toxicity reasoning based on Adverse Outcome Pathways (AOPs).

The dataset is designed to evaluate whether large language models can generate biologically interpretable toxicity reasoning trajectories that connect molecular structures, Molecular Initiating Events (MIEs), pathway perturbations, and organ-level adverse outcomes.

## Dataset Overview

ToxReason consists of three splits:

| Split | Description |
|---|---|
| `MIE_matched` | Large-scale silver training data constructed by matching chemicals with experimentally supported MIE bioactivity data from ChEMBL. |
| `MIE_AO_matched` | Higher-quality training data linking MIEs and curated adverse outcomes through biologically related AOP trajectories. |
| `evaluation` | Evaluation data for assessing mechanistic toxicity reasoning performance. |

The benchmark focuses on three major organ toxicity categories:

- Liver Toxicity
- Cardiotoxicity
- Kidney Toxicity

## Task Description

Given a molecular structure represented as a SMILES string, models are expected to generate mechanistic toxicity reasoning trajectories.

The expected reasoning may include:

- Molecular Initiating Events (MIEs)
- Biological pathway perturbations
- Adverse Outcome Pathway information
- Organ-level adverse outcomes
- Stepwise mechanistic toxicity explanations

## Intended Use

ToxReason is intended for:

- Toxicity reasoning evaluation
- Mechanistic toxicity generation
- LLM reasoning benchmark research
- AI-assisted drug safety assessment
- Explainable AI for toxicology

## Limitations

- The dataset includes reasoning trajectories derived from curated biological knowledge and rule-based construction.
- Biological mechanisms may not fully represent all real-world causal toxicity pathways.
- The current benchmark focuses on liver, heart, and kidney toxicities.
- This dataset should be used for research purposes and should not be considered a substitute for expert toxicological evaluation.

## License

This dataset is released under the CC BY 4.0 license.

## Citation

If you use this dataset, please cite:

```bibtex
@article{park2026toxreason,
  title={ToxReason: A Benchmark for Mechanistic Chemical Toxicity Reasoning via Adverse Outcome Pathway},
  author={Park, Jueon and Jang, Wonjune and Kim, Chanhwi and Park, Yein and Kang, Jaewoo},
  journal={arXiv preprint arXiv:2604.06264},
  year={2026}
}
```