Self-Supervised Pre-training for Protein Embeddings Using Tertiary Structures

requirment

# See cuda version:
cat /usr/local/cuda/version.txt or nvcc -V
conda install -c dglteam dgl-cudaXX.X  (replace the XX.X with cuda version)
pip install torch==1.7.1
pip install dgl
pip install lmdb
pip install lie-learn
pip install boto3
pip install h5py
pip install Bio
pip install pyyaml
pip install matplotlib

inputs/outputs examples

pdb/task01/hdf5.files

sample feature

export PYTHONPATH="/path/to/tFold-SE3-loc"
cd ../..
cd experiments/ebm_struct/
modify config_sample_feat.yaml

for config_sample_feat.yaml: pid_fpath_rcsb_tst: .txt fpath including all pdbid. (same with bc-0.out file) save_feat_dpath_rcsb: save dpath for the generated features mdl_dpath: model dpath

other hyper-paremeters are not using in this sampling version

The model provided here (dpath: tFold-SE3-loc/pdb/task01/models/sample) is not one of the models introduced in the paper. But it may achieve similar results.

CUDA_VISIBLE_DEVICES=0 python main.py --config_fname config_sample_feat.yaml

load the generated features

import torch
feat = torch.load('fpath/to/the/feature/file.pt')

the feat is a python dict including 5 keys: (['pdbid', 'seq', 'inter_residue_feat', 'residue_level_feat', 'protein_level_feat']) where pdbid is the pdbid and the seq is the protein sequence. inter_residue_feat is the structrue embedding shape = L * L * C. residue_level_feat is the embedding after 1D pooling shape = L * C. inter_residue_feat is the embedding after 2D pooling shape = C. where C = 64 in this model.

Notes: After severial experiments, we found that the use of inter_residue_feat (L * L * C) has the best effect, and the performance is gradually weakened after 1D and 2D pooling. Therefore, we suggest users to use our structrue embedding in 2D downstream networks, such as 2D cnn or graph neural network.

Citation

Guo, Yuzhi, Jiaxiang Wu, Hehuan Ma, and Junzhou Huang. "Self-supervised pre-training for protein embeddings using tertiary structures." In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 6, pp. 6801-6809. 2022.