This package implements EPoCS - an ESM-based Pocket Cross-Similarity metric for the comparison and contextualisation of protein binding sites, as well as systematic debiasing of train-test splits for pocket-centric machine-learning models. EPoCS combines protein language models (specifically, ESM-2) with real-space tesselation to generate vector embeddings for protein binding sites. The embeddings are the basis of the EPoCS similarity metric that gives rise to the pocket atlas. See the bioRxiv preprint for details.

For python3.9:

conda create --prefix ./venv python=3.9
conda activate ./venv
conda install -c tmap tmap  # requires python>=3.9,<3.10.0a0

For python>3.9 we need to build tmap from source. Also, note that ESM is not officially supported for python>3.9.

conda create --prefix ./venv python=3.12 gcc_linux-64=13.2 gxx_linux-64=13.2
conda activate ./venv
conda install ogdf -c tmap
conda install cmake pillow numpy scipy matplotlib  # some tmap dependencies
pip install git+https://github.com/reymond-group/tmap.git

Then, for any python3:

conda install conda-forge::pymol-open-source
pip3 install torch torchvision torchaudio
conda install biopandas pytest scipy tqdm -c conda-forge
pip install fair-esm
pip install faerun

To verify your installation, run

pytest tests/ --esm_parameters_path /path/to/esm2_t36_3B_UR50D.pt

The input file should include the paths for protein and ligand files in .cif or .pdb formats. Instead of an explicit ligand input, you can provide a 3D point (x,y,z) as reference point for the tesselation. You can set custom paths for file output, see run_epocs.py --help. If you run EPoCS for thousands of pockets, make sure to consider disk space requirements in advance.

If downloading the ESM model weights from server is not possible or desirable, you can point to an existing parameter file using --esm_parameters_path /path/to/esm/weights/esm2_t36_3B_UR50D.pt.

Example:

python run_epocs.py -f ./example/pocket_list -pp /path/to/esm2_t36_3B_UR50D.pt -np 8

The manuscript is available on bioRxiv, please cite if you found the method and/or code useful:

@article{oruc_epocs_2024,
    doi = {...},
    url = {...},
    author = {...},
}