Skip to content

Latest commit

 

History

History
39 lines (25 loc) · 1.22 KB

README.md

File metadata and controls

39 lines (25 loc) · 1.22 KB

BertFold

BertFold is a 30 layers BERT model to predict the distance map of protein 3D structure. ProtBert is used as a pre-trained model.

More details are in my article.

Getting Started

Download pre processed dataset.

mkdir -p data/ProteinNet/casp12
mv *.pqt data/ProteinNet/casp12/

Install requirements. You may have to install apex and torch-scatter manually.

Run train script.

cd src
python run_train.py params/001.yaml

Results

Model Val Test
ProtBert (seq only) 4.855 7.027
ProtBert-BFD (seq and evolutionary) 4.480 6.127

Once finishing train the model, predicted distance map is available by using the visualization script

1ZU4

Credits