chengtan9907 / Global-context-aware-generative-protein-design

The official implementation of the ICASSP'2023 paper Global-context aware generative protein design.

Global-context aware generative protein design

This repository contains the Pytorch implementation code for paper:

Global-context aware generative protein design

Dependencies

• torch == 1.8.0 (with suitable CUDA and CuDNN version)
• scikit-learn == 0.24.2
• numpy
• argparse
• tqdm

Overview

• dataset/ contains a script for downloading datasets.
• experiments/ contains a pretrained model and its hyperparameter configure file.
• model/ contains code for building models including structGNN, structTrans, GCA.

Install

This project has provided an environment setting file of conda, users can easily reproduce the environment by the following commands:

  conda env create -f environment.yml
  conda activate gca_protein_design

Usage

To download the dataset in need, simply run download_dataset.sh file in ./dataset folder:

bash download_dataset.sh

To validate the reported results, please open reproduce.ipynb and execute the blocks inside.

model_param.json is generated for every experiment, users can reproduce the experiment with the same setting as the trained one by the following python code:

import json
import argparse

config = json.load(open(svpath + 'model_param.json','r'))
exp = Exp(argparse.Namespace(**config))

To train a model from scratch by yourself, users can run main.py with optional arguments. For example, we would like to run model GCA with 10 epochs:

python main.py --model-type gca --epochs 10

More optional arguments are available in parser.py.

Citation

If you are interested in our repository and our paper, please cite the following paper:

@article{tan2022generative,
  title={Generative De Novo Protein Design with Global Context},
  author={Tan, Cheng and Gao, Zhangyang and Xia, Jun and Hu, Bozhen and Li, Stan Z},
  journal={arXiv preprint arXiv:2204.10673},
  year={2022}
}

Contact

If you have any questions, feel free to contact us through email ([e-mail]). Enjoy!