1. PyTorch 1.1
2. TorchProteinLibrary (https://github.com/lupoglaz/TorchProteinLibrary, commit 16166ce4847ad3ba95cd5afdb8bb4503cad32caa)
3. SE3CNN (https://github.com/mariogeiger/se3cnn)

Training dataset (13Gb)

Training logs (39Mb)

Preprocessed Docking Benchmark V4 (50Mb)

Preprocessed Docking Benchmark V5 (50Mb)

1. Chose some directories where you will store your data. In the file src/init.py change the storage_dir for the local machine.

2. Download the dataset and unpack it in the storage_dir. The directories configuration should look like storage_dir/LocalDockingDataset/...

3. Download logs and models and unpack it in the storage_dir, move the folders DPD_experiments and DPD_models to the storage_dir.

4. Download preprocessed benchmarks and store them in your storage_dir/LocalDockingDataset. Main point of interest are directories Matched in these benchmarks. They contain matched chains between bound and unbound structures. For more details see here.

Your storage_dir should look like:

storage_dir/LocalDockingDataset/..

storage_dir/LocalDockingDataset/DockingBenchmarkV4/..

storage_dir/LocalDockingDataset/DockingBenchmarkV5/..

storage_dir/DPD_experiments/..

storage_dir/DPD_models/..

If you want to make your own dataset, follow intruction in the scipts/Dataset directory. The whole process takes around 2-3 days, also there might be some inconsistencies with the directories. If you find, that it does not work, please submit an issue.

1. Dataset loaders can be found in src/Dataset directory. We use two loaders: one for the generated dataset and one for the DockingBenchmark.

2. Models are defined in src/Models. They are split into Protein Representation Models and Docking Models. Additionally, here one can find the ranking loss function and volume multiplication modules

3. Training iteration is defined in src/Training/LocalTrainer.py

4. Docking iteration is defined in src/Docker/Docker.py

Main scripts for training and testing are: src/train_local.py and src/test_local.py. Training and testing parameters we used in the paper can be found in src/server_train_local.sh and src/server_test_local.sh

5. Rotations loading and circular volume plotting can be found in src/Utils

We used rotation generated by https://mitchell-lab.biochem.wisc.edu/SOI/index.php that can be found in data directory. However, they have different license than this repository. Be sure to obtain the license from MitchellLab.

Trainig a model from scratch takes around 4-5 days on TitanX Maxwell. To launch the scripts change the directory to /src and run local_train.py with the appropriate parameters.

An example of the training parameters can be found in server_train_local.sh.

To evaluate a model on the benchmark head to /src and run local_test.py with the appropriate parameters. The evaluation takes 3 days on 5 nodes using K40 GPU.

We did not evaluate docking of the bound structures. In case you would like to do this experiment change lines 62 and 63 in the file src/local_test.py to pass bound structures to the docker.

An example of the testing parameters can be found in server_test_local.sh.

The instruction on how to plot the results are here

@article {Derevyanko738690,
	author = {Derevyanko, Georgy and Lamoureux, Guillaume},
	title = {Protein-protein docking using learned three-dimensional representations},
	elocation-id = {738690},
	year = {2019},
	doi = {10.1101/738690},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2019/08/19/738690},
	eprint = {https://www.biorxiv.org/content/early/2019/08/19/738690.full.pdf},
	journal = {bioRxiv}
}