Protein Structure Landscape

This is a companion repository to the paper:

Szczerbiak P, Szydlowski LM, Wydmański W, Renfrew PD, Koehler Leman J, Kosciolek T. Large protein databases reveal structural complementarity and functional locality. Nat Commun 16, 7925 (2025). https://doi.org/10.1038/s41467-025-63250-3

The repository contains data, scripts and notebooks to reproduce main-text Figures and most Supplementary Figures.

Content

• data

  • Contains Parquet files used to reproduce the results presented in the paper. These can be accessed using the pandas.read_parquet('{filename}.parquet') function from the Pandas Python library.
  • Also includes intermediate files (CSV, Pickle) used by the notebooks (see below).

• notebooks

  • Figure_*.ipynb: Reproduces the figures from the paper.
  • data-loader.ipynb: Describes the contents of each Parquet file and demonstrates how to load them.
  • commands.md: A collection of commands used for clustering, generating embeddings, and performing dimensionality reduction.
  • embeddings.ipynb: Generates Geometricus embeddings for the ProtGPT and BFVD datasets. The notebook can be easily adapted for other datasets.
  • dimensionality_reduction.ipynb: Performs dimensionality reduction on the ProtGPT and BFVD embeddings. Also adaptable for other datasets.

• plots

  • Contains an interactive Sankey diagram (sankey_diagram.html), referenced in the Cluster heterogeneity section of the supplementary materials as well as interactive version of Figure 6B (esm_biomes.html).

• zconda_environment_setup.info: Provides instructions for setting up the Python Conda environment required to run the notebooks.

Note: To run embeddings.ipynb and dimensionality_reduction.ipynb you must first download the embeddings.zip file from the Figshare data repository (https://dx.doi.org/10.6084/m9.figshare.27203073), unzip it, and place the contents in the data/ folder.

Novel methods

The source code for new methods and tools developed during the research is available in dedicated data repositories:

• go2cog
• deepFRI v1.1
• web server

Data sources

We analyzed datasets coming from:

• AFDB: Based on the lists of light and dark proteins stored in https://afdb-cluster.steineggerlab.workers.dev/, we downloaded AlphaFold models using gsutil API directly from the AFDB (see details here: https://github.com/google-deepmind/alphafold/tree/main/afdb) (last access: Feb 14 2024).
• ESMAtlas: We downloaded foldcomp files (regular expression: highquality_clust30.*) directly from https://foldcomp.steineggerlab.workers.dev/ (last access: Jan 23 2024).
• MIP: We downloaded Zenodo archive from https://zenodo.org/records/6611431 (last access: Jan 26 2024).

Web server

The web server is available at: https://protein-structure-landscape.sano.science.

It currently supports the following features:

• Structure search by:
  • Protein IDs e.g. A0A1F9R7C9, MGYP002780493870, MIP_00292964
  • Gene Ontology (GO) function names e.g. organic acid metabolic process
  • Origin (biome / taxon) e.g. Marine, Wastewater, Escherichia coli
• Interactive structure visualization using PyMOL
• Download of structures in PDB format
• Export selected entries as a .tsv file
• Filtering options, including:
  • taxonomy
  • source database (AFDB, ESMAtlas, MIP)
  • superCOG functional category
  • protein length
  • pLDDT (for AFDB models only)

When a protein is selected or searched, the server displays:

• Detailed information about the representative structure
• Location of the representative structure in the structural landscape
• A 3D visualization of the structure
• DeepFRI v1.0 function predictions

The web server will be continuously improved.