ESM Atlas: Protein Features and Structures
Biohub
Predicted structures and interpretable features for about 6.8 billion proteins.
Computational outputs from the ESMC and ESMFold2 models over a deduplicated set of more than 6.8 billion public protein sequences, including viral and metagenomic sequences. Includes sparse autoencoder features and structure predictions. Intended for protein modeling and interpretability research.
Access
- Cost
- FreeNo charge to download; hosted in an AWS S3 bucket.
- License
- Provider's own termsChan Zuckerberg Biohub terms, with third-party data under original terms (see record)
- Sign-up
- No sign-up
- How you get it
- API and Bulk downloadParquet, PDB/mmCIF
What's in it
- One row is
- One protein sequence with its features and predicted structure
- Covers
- Not geographic
- Years
- v1.0 release
- Updated
- One-time release
- Size
- About 6.8 billion proteins
Quality
Enormous resource with documentation and a versioned release. Predictions, not experimental structures.
- Predicted, not experimental
- Third-party data carries its own terms
Connects to
- Protein Data Bank 3D Structural Biology Data: Experimental structures to check predictions against
- OpenProteinSet: Multiple sequence alignments for related training and research
Questions it could answer
- How many predicted structures are confidently novel versus known folds?
- Which metagenomic proteins lack any annotation?