Implementation of Protein Invariant Point Packer (PIPPack)
PIPPack is a graph neural network (GNN) that utilizes geometry-aware invariant point message passing (IPMP) updates and recycling to rapidly generate accurate protein side chains.
PIPPack has now been published! Please check out the manuscript in Proteins!
To get started right in your browser, click this button to open the PIPPack notebook in Google Colab:
Please let us know if you have any issues or suggestions!
To build the environment from scratch:
# Create and activate the pippack environment
conda create -n pippack
conda activate pippack
# Install PyTorch (see https://pytorch.org/get-started/locally/)
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
# Install Lightning (see https://lightning.ai/docs/pytorch/stable/starter/installation.html)
conda install lightning=2.0.1 -c conda-forge
# Pip installs:
# - PyTorch Geometric (see https://pytorch-geometric.readthedocs.io/en/latest/install/installation.html)
# - BioPython (see https://biopython.org/wiki/Download)
# - Hydra (see https://hydra.cc/docs/intro/#installation)
python -m pip install torch-geometric biopython hydra-core -U
Alternatively, you can use the environment file env/pippack_env.yaml to build the environment:
# Build pippack environment from yaml file
conda env create -f env/pippack_env.yaml
All test datasets (CASP13, CASP14, CASP15, CASP15+context, and Top2018) and prediction results are publicly available at https://zenodo.org/records/11236817.
This project is licensed under the MIT License - see the LICENSE file for details.
If you find PIPPack useful in your research or project, please cite our paper:
@article{randolph2024pippack,
title={Invariant point message passing for protein side chain packing},
author={Randolph, Nicholas Z. and Kuhlman, Brian},
journal={Proteins: Structure, Function, and Bioinformatics},
year={2024},
pages={1-14},
doi={10.1002/prot.26705},
}