This repository contains the implementation of a novel framework that leverages LLMs as chemical reasoning engines to guide traditional search algorithms in chemistry. Our approach demonstrates how LLMs can be effectively used for:
- Strategy-aware Retrosynthetic Planning: Enable chemists to specify desired synthetic strategies in natural language and find routes that satisfy these constraints.
- Mechanism Elucidation: Guide the search for plausible reaction mechanisms by combining chemical principles with systematic exploration.
- 🧪 Natural language specification of synthetic strategies
- 🔍 LLM-guided search through chemical space
- 📊 Benchmark datasets for both synthesis planning and mechanism elucidation
- 🤖 Support for multiple LLM providers (Claude, GPT-4, DeepSeek)
# Install from PyPI (TBD)
pip install steer
# Install from source
pip install git+https://github.com/schwallergroup/steer.git# Run the complete synthesis benchmark
steer synth --model=claude-3-5-sonnet bench
# Run a single task
steer synth --model=claude-3-5-sonnet bench --task=ea8df340d54596eda93e23f04dff3a9b# Run mechanism elucidation benchmark
steer mech --model=claude-3-5-sonnet benchThe repository includes two main benchmarks:
- Multiple target molecules of varying complexity
- Strategic constraints specified in natural language
- Evaluation metrics for route-to-prompt alignment
- 12 diverse organic reactions
- Ground truth mechanisms with elementary steps
- Performance metrics for mechanism prediction
If you use this work in your research, please cite:
@misc{bran2025chemicalreasoningllmsunlocks,
title={Chemical reasoning in LLMs unlocks steerable synthesis planning and reaction mechanism elucidation},
author={Andres M Bran and Theo A Neukomm and Daniel P Armstrong and Zlatko Jončev and Philippe Schwaller},
year={2025},
eprint={2503.08537},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2503.08537},
}Click to expand development instructions
git clone https://github.com/schwallergroup/steer.git
cd steer
pip install -e .pip install tox
tox- Andres M Bran
- Théo A. Neukomm
- Daniel Armstrong
- Zlatko Jončev
- Philippe Schwaller
For questions and feedback: