MolAct is an Agentic RL framework that trains LLMs to design molecules through a multi-turn "Think-Tool-Observation" loop. By leveraging GRPO and a two-stage training paradigm—mastering basic editing before tackling complex property optimization—MolAct learns to autonomously invoke chemical tools to ensure every modification is both physically valid and property-optimized.
- 2025-12-24: 📄 Our paper is now available on arXiv!
- 2025-12-23: 🎉 We released the inference code, training datasets, and pre-trained models for MolAct! You can now run inference with our pre-trained models on molecular editing and optimization tasks.
git clone https://github.com/little1d/MolAct.git
cd MolActPlease refer to the official AgentFly installation guide for setup instructions.
For running evaluations on ChemCoTBench, install additional dependencies:
pip install python-Levenshtein nltk rouge-score selfies scikit-learnDownload NLTK data (required for evaluation metrics):
python -c "import nltk; nltk.download('wordnet', quiet=True); nltk.download('punkt', quiet=True)"We have open-sourced our training datasets and pre-trained models on HuggingFace. Please download the models before running inference.
Run inference on molecular editing tasks using the provided script:
bash scripts/1\ run_mol_edit_inference.sh \
[MODEL_DIR] [BENCH_DIR] [OUT_DIR]Recommended approach: Instead of passing parameters via command line, you can directly edit the script file scripts/1 run_mol_edit_inference.sh to modify the default paths (MODEL_DIR, BENCH_DIR, OUT_DIR) and other parameters (MAX_NEW_TOKENS, TEMP, TOP_P, BACKEND). This is more convenient and recommended.
The script will process all subtasks (add, delete, sub) automatically and generate JSON result files in the outputs/ directory. Each result file contains the model's reasoning trajectory (thought process, tool calls, and observations) along with the final predictions, which can be used for subsequent evaluation.
Run inference on molecular optimization tasks using the provided script:
bash scripts/3\ run_mol_opt_inference.sh \
[MODEL_DIR] [BENCH_DIR] [OUT_DIR]Recommended approach: Instead of passing parameters via command line, you can directly edit the script file scripts/3 run_mol_opt_inference.sh to modify the default paths (MODEL_DIR, BENCH_DIR, OUT_DIR) and other parameters (MAX_NEW_TOKENS, TEMP, TOP_P). This is more convenient and recommended.
The script will process all subtasks (logp, drd, jnk, gsk, qed, solubility) automatically and generate JSON result files in the outputs/ directory. Each result file contains the model's reasoning trajectory (thought process, tool calls, and observations) along with the final predictions, which can be used for subsequent evaluation.
Note: When running inference scripts from the root directory, the TDC library will automatically download oracle files to the oracle/ directory.
Evaluate results on ChemCoTBench using the provided scripts. The evaluation uses the ChemCoTBench evaluation framework.
Recommended approach: Instead of passing parameters via command line, directly edit the script files to modify the default paths:
scripts/2 eval_mol_edit_chemcotbench.shfor molecular editing tasksscripts/4 eval_mol_opt_chemcotbench.shfor molecular optimization tasks
The scripts accept three parameters:
BENCH_DIR: Path to ChemCoTBenchbaseline_and_evaldirectory (contains evaluation scripts)PRED_DIR: Directory containing prediction JSON files generated from the inference step aboveOUT_DIR: Output directory for evaluation results
# For molecular editing tasks
bash scripts/2\ eval_mol_edit_chemcotbench.sh \
[BENCH_DIR] [PRED_DIR] [OUT_DIR]
# For molecular optimization tasks
bash scripts/4\ eval_mol_opt_chemcotbench.sh \
[BENCH_DIR] [PRED_DIR] [OUT_DIR]The scripts will automatically evaluate all subtasks (add/delete/sub for editing, logp/drd/jnk/gsk/qed/solubility for optimization) and save results in the output directory.
Example results: We provide example inference and evaluation results in the outputs/ directory for reference. You can check these files to understand the expected output format.
The agents use a comprehensive set of chemistry tools for molecular manipulation and property calculation:
- Molecular Validation: Validate SMILES strings
- Property Calculation: Calculate molecular properties (logP, QED, solubility, etc.)
- Molecular Editing: Functional group transformations (addition, deletion, substitution) implemented via SMARTS-based pattern matching and modification
- Scaffold Analysis: Murcko scaffold extraction and similarity metrics
- Oracle Scoring: Protein activation scoring for optimization tasks (DRD-2, JNK-3, GSK-3β)
We evaluate on ChemCoTBench for both molecular editing and optimization tasks. ChemCoTBench is a comprehensive benchmark for step-wise reasoning on complex chemical problems.
- Molecular Editing: Add, delete, and substitute functional groups
- Molecular Optimization: Optimize for physicochemical properties (QED, LogP, Solubility) and protein activation (DRD-2, JNK-3, GSK-3β)
The following shows an example of model training on molecular editing task using Qwen-2.5-7B as the base model on 4 H200 hardware.
demo.mp4
If you use MolAct in your research, please cite:
@article{molact2025,
title={MolAct: An Agentic RL Framework for Molecular Editing and Property Optimization},
author={Zhuo Yang and Yeyun Chen and Jiaqing Xie and Ben Gao and Shuaike Shen and Wanhao Liu and Liujia Yang and Beilun Wang and Tianfan Fu and Yuqiang Li},
year={2025},
eprint={2512.20135},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2512.20135}
}We would like to thank the following projects and frameworks that made this work possible:
- SwanLab for experiment tracking and visualization
- ChemCoTBench for providing comprehensive benchmarks for molecular editing and optimization tasks
- AgentFly for the agentic RL framework infrastructure