Intrinsic-Extrinsic Convolution and Pooling for Learning on 3D Protein Structures

This is the official code of the ICRL 2021 paper Intrinsic-Extrinsic Convolution and Pooling for Learning on 3D Protein Structures.

Citation

If you find this code useful please consider citing us:

    @article{hermosilla2021ieconv,
      title={Intrinsic-Extrinsic Convolution and Pooling for Learning on 3D Protein Structures},
      author={Hermosilla, Pedro and Schäfer, Marco and Lang, Matěj and Fackelmann, Gloria and Vázquez, Pere Pau and Kozlíková, Barbora and Krone, Michael and Ritschel, Tobias and Ropinski, Timo},
      journal={International Conference on Learning Representations},
      year={2021}
    }

Instalation

Open a docker container with the following command:

sudo docker run --gpus all --privileged -it -v ${PWD}:/working_dir -w /working_dir tensorflow/tensorflow:1.12.0-devel-gpu-py3

Execute the following command to compile the custom ops of tensorflow:

cd IEProtLib/tf_ops
python genCompileScript.py --cudaFolder /usr/local/cuda
sh compile.sh

We already provide a compiled version of the library for the docker container, so if you are using the docker container indicated above you can skip the compilation.

Download the preprocessed datasets.

In the following links the different datasets can be downloaded:

• Enzymes vs Non-Enzymes:

https://drive.google.com/uc?export=download&id=1KTs5cUYhG60C6WagFp4Pg8xeMgvbLfhB

Extract content in: Datasets/data/ProteinsDD/

• Scope 1.75:

https://drive.google.com/uc?export=download&id=1chZAkaZlEBaOcjHQ3OUOdiKZqIn36qar

Extract content in: Datasets/data/HomologyTAPE/

• Protein function:

https://drive.google.com/uc?export=download&id=1udP6_90WYkwkvL1LwqIAzf9ibegBJ8rI

Extract content in: Datasets/data/ProtFunct

Train Ennzymes vs Non-Enzymes

Execute the following commands to train a network on the task:

cd Tasks/ProteinsDD
python Train.py --configFile confs/train_fold0.ini
python Train.py --configFile confs/train_fold1.ini
python Train.py --configFile confs/train_fold2.ini
python Train.py --configFile confs/train_fold3.ini
python Train.py --configFile confs/train_fold4.ini
python Train.py --configFile confs/train_fold5.ini
python Train.py --configFile confs/train_fold6.ini
python Train.py --configFile confs/train_fold7.ini
python Train.py --configFile confs/train_fold8.ini
python Train.py --configFile confs/train_fold9.ini

Train SCOPe 1.75

Execute the following commands to train a network on the task:

cd Tasks/ProtHomology
python Train.py --configFile confs/train.ini

To evalute the trained model on the different test set use the following commands:

python Test.py --configFile confs/test_fold.ini
python Test.py --configFile confs/test_superfamily.ini
python Test.py --configFile confs/test_family.ini

Train Protein function prediction

Execute the following commands to train a network on the task:

cd Tasks/ProtFunct
python Train.py --configFile confs/train.ini

To evaluate the trained model execute:

python Test.py --configFile confs/test.ini

Trained models comming soon