/Protein-Peptide-Interactions

Primary LanguageJupyter Notebook

De Novo Design of Target-Specific Ligands Using BERT-Pretrained Transformer

Requirements

We think the requirements for the environment are so loose that any version can be used

cuda 10.2
conda install gxx_linux-64 7.3
conda install pytorch==1.8.1 torchvision==0.9.1 torchaudio==0.8.1 cudatoolkit=10.2 -c pytorch
conda install pyg -c pyg -c conda-forge
conda install tqdm
conda install -c jrybka sklearn
conda install -c conda-forge tensorboard
conda install -c conda-forge jsonlines

Prepare Data

pepbdb data:http://huanglab.phys.hust.edu.cn/pepbdb/ propedia data:http://bioinfo.dcc.ufmg.br/propedia/download

Detailed data processing steps are given in the readme of the dataprocess

Train ESM-IF1 of our model

Go to the ESM-IF1 folder and run the train_mySE3.py file

python train_esm.py

Empty logs/, models/, and save/ folders is needed to store procedures, breakpoints, and final results.

Train SE(3)-Tansformer of our model

Go to the SE(3)-Tansformer folder and run the train_mySE3.py file

python train_mySE3.py

Empty logs/, models/, and save/ folders is needed to store procedures, breakpoints, and final results.

Train ProtGVP of our model

Go to the ProtGVP folder and run the train_ProtGVP.py file

python train_ProtGVP.py

Empty logs/, models/, and save/ folders is needed to store procedures, breakpoints, and final results.

Attention

Fig. 4 Visualization and Attention matrix on 4HFZ. L54 forms a hydrogen bond with W23 (yellow line segment), where L57 is physically close to the W53 amino acid (blue) and P27 is physically close to L54 (red).

Fig. 5 Visualization and Attention matrix on 5NTN. L504 forms a hydrogen bond with K336 (yellow line segment), where L505 is physically close to K336 amino acid (pink), G506 is physically close to K336 (Burgundy), and Q502 is physically close to Q349 (blue).

Fig. 6 Visualization and Attention matrix on 5EIV. G4 creates a hydrogen bond with Y137 (yellow line segment), where A6 is physically close to the amino acid Y137 (red)

model performance evaluation

Fig.2 Changes in AUC under Different Similarities of Each Model.

Fig.4 The relationship between precision and the amount of training data and pocket size.

tode

More details will be given in the future.