SHAPES

Structural and Hierarchical Assessment of Proteins with Embedding Similarity.

Preprint Paper

Install

Clone the repo, then

cd protein_shapes
conda env create -n protein-shapes --file environment.yaml
pip install -e .

The environment was created with

conda install pytorch torchvision pytorch-cuda=12.4 -c pytorch -c nvidia
conda install numpy scipy scikit-learn pandas
conda install -c conda-forge -c bioconda foldseek
pip install ProDy hydra-core ml-collections dm-tree tqdm matplotlib seaborn typer-cli umap-learn
pip install --upgrade huggingface_hub
pip install esm
pip install -e .

which can be adapted for different PyTorch / CUDA versions.

Usage

Specify the embedding type to use and paths to reference/sampled structures in configs/default.yaml, then run with

python run.py

The provided embeddings are "prot_domain_classifier", "esm3", "proteinmpnn", "foldseek",, spanning different protein structure hierarchies, from global features by the domain classifier to residue environments with varying sizes by ESM3 and ProteinMPNN encoder layers and strictly local features with Foldseek tokens. A Huggingface token is needed to use ESM3. Set the environment variable

export HUGGINGFACE_TOKEN=<your_token>

or leave it unset to have an interactive prompt for token entry.

Save time by setting embed_reference: False to reuse precomputed reference structure embeddings or embed_samples: False to reuse precomputed sampled structure embeddings.

Precomputed statistics for CATH are stored in data/ and can be loaded and used as the reference statistics by setting reference_structures: cath.

Outputs are logged in the specified output_dir.

Plots

For continuous embeddings, visualize sampled structures overlayed on reference set structures with UMAP with

python examples/plot_embeddings.py

E.g. to visualize ProteinMPNN embeddings at the first layer, use

python examples/plot_embeddings.py --embed-type proteinmpnn --layer 0

Plots are by default saved in plots/.

Acknowledgements

• ProtDomainSegmentor

  @article{eguchi2020multi,
  title={Multi-scale structural analysis of proteins by deep semantic segmentation},
  author={Eguchi, Raphael R and Huang, Po-Ssu},
  journal={Bioinformatics},
  volume={36},
  number={6},
  pages={1740--1749},
  year={2020},
  publisher={Oxford University Press}
  }
  
  

• ProteinMPNN

  @article{dauparas2022robust,
  title={Robust deep learning--based protein sequence design using ProteinMPNN},
  author={Dauparas, Justas and Anishchenko, Ivan and Bennett, Nathaniel and Bai, Hua and Ragotte, Robert J and Milles, Lukas F and Wicky, Basile IM and Courbet, Alexis and de Haas, Rob J and Bethel, Neville and others},
  journal={Science},
  volume={378},
  number={6615},
  pages={49--56},
  year={2022},
  publisher={American Association for the Advancement of Science}
  }
  

• Foldseek

  @article{van2024fast,
  title={Fast and accurate protein structure search with Foldseek},
  author={Van Kempen, Michel and Kim, Stephanie S and Tumescheit, Charlotte and Mirdita, Milot and Lee, Jeongjae and Gilchrist, Cameron LM and S{\"o}ding, Johannes and Steinegger, Martin},
  journal={Nature biotechnology},
  volume={42},
  number={2},
  pages={243--246},
  year={2024},
  publisher={Nature Publishing Group US New York}
  }