This repository contains code to train a symmetry-constrained monte-carlo tree search (MCTS) algorithm which uses fragments obtained from patent-mined data. Details about the method and results can be found at Symmetry-Constrained Generation of Diverse Low-Bandgap Molecules with Monte Carlo Tree Search. Details about how the patent data was obtained can be found at Automated patent extraction powers generative modeling in focused chemical spaces.
All packages are provided in environment.yml and can be installed with:
conda env create --name mcts --file=environment.yml
Place a config file config.json inside a folder (say test_folder) containing training parameters. We provide a sample version in test_folder.
git clone [email protected]:chemprop/chemprop.git
cd chemprop && pip install -e .
Code to preprocess, cluster and decompose fragments from patents can be found in fragments/. We provide the output files from these steps so that training can be done without having to run these first.
Example code to train chemprop can be found in train_chemprop/. We already provide checkpoints, so re-training is not necessary to reproduce MCTS training.
The following command will run one training of the MCTS model.
python train_mcts.py --output_dir test_folder --environment patent --iter 0
If you would like to run multiple iterations of training to generate diverse candidates (see more details in the paper), you can run training repetitions. A simple example of how this can be done in a SLURM-based cluster can be found in exammple_scripts/run_repetitions.py.
If you use code from this repository, please cite:
@article{subramanian2024symmetry,
title={Symmetry-Constrained Generation of Diverse Low-Bandgap Molecules with Monte Carlo Tree Search},
author={Subramanian, Akshay and Damewood, James and Nam, Juno and Greenman, Kevin P and Singhal, Avni P and G{\'o}mez-Bombarelli, Rafael},
journal={arXiv preprint arXiv:2410.08833},
year={2024}
}
If you use the patent-mined datasets, please additionally cite:
@article{subramanian2023automated,
title={Automated patent extraction powers generative modeling in focused chemical spaces},
author={Subramanian, Akshay and Greenman, Kevin P and Gervaix, Alexis and Yang, Tzuhsiung and G{\'o}mez-Bombarelli, Rafael},
journal={Digital Discovery},
volume={2},
number={4},
pages={1006--1015},
year={2023},
publisher={Royal Society of Chemistry}
}