DiffDock modified for protein-protein

NOTE: This was the prototype for DiffDock-PP. Please refer to that repository for usage.

---

Note: Add --debug flag to run on subset of DIPS (or switch to DB5) to debug.

ESM computation is batched and cached, but for all of DIPS, might be better to load on the fly (asynchronously) rather than all at once.

Model training:

./src/train.sh

To modify dataset/model/training parameters, change config_file.

Model inference:

./src/predict.sh

Currently the model only works for >1 GPU :') But trust me it'll take much too long if you try with only 1.

DIPS and DB5 work. SabDab has not been updated for a long time but it's an interesting dataset that can also incorporate the receptor flexibility aspects, to-be-developed.

Data

My data splits

ln -s /data/rsg/chemistry/rmwu/src/sandbox/glue/data data

Note: you can look at the above directory for clues as to formatting (mainly, instead of Octavian's 3 files, 1 per split, I have a CSV with the split = train val test appended as a column.

You can specify the data's location via data_path (you are welcome to use the existing path to my directory)

Conda environment

pip install numpy

pip install dill
pip install tqdm
pip install pyyaml
pip install pandas

pip install scikit-learn
pip install biopython

# install PyTorch

pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116

pip install --upgrade e3nn

pip install tensorboard
pip install tensorboardX

# install compatible pytorch geometric in this order WITH versions

pip install --no-cache-dir  torch-scatter==2.0.6 torch-sparse==0.6.9 torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-1.13.0+cu116.html