This repository implements machine learning force field method on transition state search. We use this method to efficiently search the transition state in catalysis reaction. Based on this method, we have achieved end-to-end automatic catalyst screening, which can be widely applied to homogeneous catalysis systems and provides significant assistance for high-throughput catalyst screening.
.
├── assets
| ├── Screening.png
│ └── TSFF_framework.png
├── checkpoints
│ └── placeholder.txt
├── configs
│ ├── neb_config
│ └── training_configs
├── data
│ ├── cata_data
│ ├── gen_cata_data
│ └── t1x_data
├── data_processing
│ ├── dataloader.py
│ ├── reaction_data_creation.py
│ ├── transition1x_data_creation_opt_neb.py
│ └── transition1x_data_creation.py
├── evaluating
│ ├── evaluate_direct.py
│ ├── evaluate_gen.py
│ └── evaluate.py
├── experiment_logs
│ └── rmsd_t1x_neb.log
├── final_checkpoints
│ ├── cata_ckpt
│ ├── gen_cata_ckpt
│ └── t1x_ckpt
├── logs
├── ocpmodels
│ ├── common
│ ├── datasets
│ ├── __init__.py
│ ├── models
│ ├── modules
│ ├── preprocessing
│ ├── __pycache__
│ ├── tasks
│ └── trainers
├── README.md
├── results
├── save
│ ├── cata_neb
│ └── t1x_neb
├── scripts
│ ├── evaluate_cata.sh
│ ├── evaluate_t1x.sh
│ ├── run_diff_arch_t1x.sh
│ ├── run_screening.sh
│ ├── training_cata.sh
│ ├── training_gen_cata.sh
│ ├── training_t1x.sh
│ ├── ts_search_cata.sh
│ └── ts_search_t1x.sh
├── searching
│ ├── interpolate_gen.py
│ ├── interpolate.py
│ ├── neb_direct.py
│ ├── neb_predictor.py
│ └── ts_search.py
└── training
└── main.py
.
├── cata_data
│ ├── new_test_fold1.h5
│ ├── new_train_fold1.h5
│ ├── reaction
│ │ ├── reaction.lmdb
│ │ └── reaction.lmdb-lock
│ ├── test
│ │ ├── Rh.lmdb
│ │ └── Rh.lmdb-lock
│ └── train
│ ├── Rh.lmdb
│ └── Rh.lmdb-lock
├── gen_cata_data
│ ├── all_gen_opt_test.h5
│ ├── all_gen_opt_train.h5
│ ├── final_gen
│ │ ├── barriers.csv
│ │ ├── Cat1174
│ │ ├── Cat1176
│ │ ├── Cat1306
│ │ ├── Cat1355
│ │ ├── Cat1402
│ │ ├── Cat1440
│ │ ├── Cat1452
│ │ ├── Cat1460
│ │ ├── Cat1476
│ │ ├── Cat1494
│ │ ├── Cat1496
│ │ ├── Cat1594
│ │ ├── Cat1643
│ │ ├── Cat1821
│ │ ├── Cat1868
│ │ ├── Cat2097
│ │ ├── Cat2128
│ │ ├── Cat2174
│ │ ├── Cat2189
│ │ ├── Cat229
│ │ ├── Cat2303
│ │ ├── Cat2347
│ │ ├── Cat2380
│ │ ├── Cat2391
│ │ ├── Cat2459
│ │ ├── Cat246
│ │ ├── Cat2507
│ │ ├── Cat2574
│ │ ├── Cat2621
│ │ ├── Cat2666
│ │ ├── Cat275
│ │ ├── Cat2766
│ │ ├── Cat2834
│ │ ├── Cat2908
│ │ ├── Cat292
│ │ ├── Cat326
│ │ ├── Cat353
│ │ ├── Cat51
│ │ ├── Cat520
│ │ ├── Cat529
│ │ ├── Cat548
│ │ ├── Cat628
│ │ ├── Cat632
│ │ ├── Cat682
│ │ ├── Cat715
│ │ ├── Cat788
│ │ ├── Cat805
│ │ ├── Cat814
│ │ ├── Cat872
│ │ ├── Cat887
│ │ └── read_db.py
│ ├── new_test_fold1.h5
│ ├── new_train_fold1.h5
│ ├── reaction
│ │ ├── all
│ │ ├── gt_barrier_50.csv
│ │ ├── pred_barrier_50.csv
│ │ └── sort_barrier_all.csv
│ ├── test
│ │ ├── Rh.lmdb
│ │ └── Rh.lmdb-lock
│ └── train
│ ├── Rh.lmdb
│ └── Rh.lmdb-lock
└── t1x_data
├── reaction
│ ├── reaction.lmdb
│ └── reaction.lmdb-lock
├── test
│ ├── transition1x.lmdb
│ └── transition1x.lmdb-lock
├── train
│ ├── transition1x.lmdb
│ └── transition1x.lmdb-lock
└── val
├── transition1x.lmdb
└── transition1x.lmdb-lock
You can download the all dataset from here, unzip and put the data folder in a directory structure similar to the previous code.
You can download the all model checkpoints from here, unzip and put the model folder in a directory structure similar to the previous code.
unzip data.zip
unzip final_checkpoints.zip
Run
pip install -r requirements.txt
to install the required packages.
Run
pip install torch_cluster-1.6.1+pt20cu118-cp38-cp38-linux_x86_64.whl
pip install torch_scatter-2.1.1+pt20cu118-cp38-cp38-linux_x86_64.whl
pip install torch_sparse-0.6.18+pt20cu118-cp38-cp38-linux_x86_64.whl
pip install torch_spline_conv-1.2.2+pt20cu118-cp38-cp38-linux_x86_64.whl
to install other torch dependencies
We use the scripts/training_t1x.sh script to train the model.
bash scripts/training_t1x.shWe use the scripts/ts_search_t1x.sh script to search the transition state.
bash scripts/ts_search_t1x.shWe use the scripts/evaluate_t1x.sh script to evaluate the transition state search.
bash scripts/evaluate_t1x.shWe use the scripts/training_cata.sh script to train the model.
bash scripts/training_cata.shWe use the scripts/ts_search_cata.sh script to search the transition state.
bash scripts/ts_search_cata.shWe use the scripts/evaluate_cata.sh script to evaluate the transition state search.
bash scripts/evaluate_cata.shOur overall framework is shown in the following figure:
In the data/gen_cata_data/reaction folder, catalyst 3d structure with ComplexGen generated directly from the ligand is present.
We use the scripts/training_gen.sh script to train the model.
bash scripts/training_gen.shWe use the scripts/ts_search_gen.sh script to search the transition state.
bash scripts/ts_search_gen.shWe use the scripts/evaluate_gen.sh script to evaluate the transition state search.
bash scripts/evaluate_gen.shIf you find this code useful, please cite:
@article{yin2025cats,
title={CaTS: Toward Scalable and Efficient Transition State Screening for Catalyst Discovery},
author={Yin, Jun and Li, Wentao and Chen, Honghao and Qiu, Jiangjie and Feng, Huasheng and Xu, Xiangya and Jin, Qiuyan and Wang, Xiaonan},
journal={ACS Catalysis},
volume={15},
pages={15754--15764},
year={2025},
publisher={ACS Publications}
}