Context-Aware Generative Models for Multi-Domain Proteins using Transformers

Code Organization

src folder

The folder src contains the source code of the paper:

• ProteinsDataset.py contains the function linked with the handling of the data.
• ProteinsTransformer.py contains the code for the Transformer model
• MatchingLoss.py contains our losses
• ardca.py functions related to the arDCA model (wrapping the julia library ArDCA.jl)
• DCA.py functions related with the contact prediction
• utils.py the other functions

files.config.json

shallow.config.json, large.config.json and large_renyi.config.json contains the hyperparameter of the shallow model, the large Transformer and the large Transformer using the entropic regularization. You can easily use a new set of hyperparameter by modifying one of this file or creating your own json file.

models

models are saved in the models folder.

data

We provide two datasets in the forlder data to test the code: PF00207_PF07677 & PF03171_PF14226.

training file

The training is controlled from the train.py and the train.sh. The arguments are: --trainset : path to train dataset
--valset : path to testset --save : path for saving the model --load : path to load to a model and continue training it --modelconfig : path to the json file with the hyperparameters --outputfile : output file where scores are written during training

Run the code

You can either use the python command:

python -m train --trainset "data/pMSA_PF00207_PF07677_train.csv" --valset "data/pMSA_PF00207_PF07677_val.csv" --save "models/saved_PF00207_PF07677.pth.tar" --load "" --modelconfig "shallow.config.json" --outputfile "output.txt"

or use the shell script train.sh.