Full Atom Protein Pocket Design via Iterative Refinement (FAIR) NeurIPS 2023 Spotlight [pdf]

Install conda environment via conda yaml file

conda env create -f fair_env.yaml
conda activate fair_env

Datasets

Please refer to README.md in the data folder.

The data used for training / evaluating the model are organized in the data Google Drive folder.

For a quick reproduction, you can download the preprocessed lmdb file and name2id file:

• crossdocked_pocket10_processed_final.lmdb
• crossdocked_pocket10_name2id.pt

Then place these files in the data folder.

Training

The model hyperparameters can be adjusted in config.

python train.py

Checkpoint and Testing

A checkpoint of our model is provided in the checkpoint folder.

python test.py

Expected results on the CrossDocked dataset:

AAR	RMSD
40.8% ± 10.9 %	1.44 ± 0.06

Latest Version

We'd like to thank Yifei for the suggestions and discussions of experimental settings. In our latest version, we do not use the original backbone for reference and obtain comparable results after retraining. The code is released in the latest folder.

Citation

@article{zhang2023full,
  title={Full-Atom Protein Pocket Design via Iterative Refinement},
  author={Zhang, Zaixi and Lu, Zepu and Hao, Zhongkai and Zitnik, Marinka and Liu, Qi},
  journal={NeurIPS},
  year={2023}
}