ProteinMAE

Official PyTorch implementation of "ProteinMAE: Masked Autoencoder for Protein Surface Self-supervised Learning".

Dataset

We use Baidu Cloud Disk to share the datasets we use: https://pan.baidu.com/s/1lkq4g5TlRz3tja9_LsQGfQ?pwd=data Password: data

Pre-Training

python main.py --config cfgs/pretrain_protein.yaml --num_workers 8

Downstream tasks

Train

Binding site identification (init with pre-trained weight):

python train_site.py --ckpt ./checkpoints/ckpt-last.pth

Protein-protein interaction prediction (init with pre-trained weight):

python train_search.py --ckpt ./checkpoints/ckpt-last.pth

Ligand-binding pocket classification (init with pre-trained weight):

python train_ligand.py --ckpt ./checkpoints/ckpt-last.pth

Inference

Binding site identification (scratch):

python test_site.py --checkpoint ./checkpoint/Transformer_site_batch32_yuanshi_epoch107

Binding site identification:

python test_site.py --checkpoint ./checkpoint/Transformer_site_batch32_yuanshi_pre6.11_epoch27.pth

Protein-protein interaction prediction (scratch):

python test_search.py --checkpoint ./checkpoint/Transformer_search_batch32_group512_size16_downsample512_6.15_epoch493.pth

Protein-protein interaction prediction:

python test_search.py --checkpoint ./checkpoint/Transformer_search_batch32_pre_group512_size16_downsample512_6.16_epoch382.pth

Ligand-binding pocket classification (scratch):

python test_ligand.py --checkpoint ./checkpoints/Transformer_ligand_downsample512_group768size16_new_epoch395.pth

Ligand-binding pocket classification:

python test_ligand.py --checkpoint ./checkpoints/Transformer_ligand_pre_downsample512_group768size16_new_epoch295.pth

Citation

If you find this code useful for your work or use it in your project, please consider citing:

@article{yuan2023proteinmae,
  title={ProteinMAE: masked autoencoder for protein surface self-supervised learning},
  author={Yuan, Mingzhi and Shen, Ao and Fu, Kexue and Guan, Jiaming and Ma, Yingfan and Qiao, Qin and Wang, Manning},
  journal={Bioinformatics},
  volume={39},
  number={12},
  pages={btad724},
  year={2023},
  publisher={Oxford University Press}
}

Acknowledgments

In this project we use (parts of) the official implementations of the followin works:

• dMaSIF
• Point-MAE

• MaSIF

  We thank the respective authors for open sourcing their methods.