Get Started
ESMFold2
ESMFold2 is a state-of-the-art model for biomolecular complex structure prediction. Built on the representations of the ESMC language model, it produces the fastest and most accurate all-atom predictions of protein complexes and protein-protein interactions, including antibody-antigen structures.
Available in two variants: ESMFold2 is optimized for accuracy on difficult targets, and ESMFold2 Fast is optimized for speed. Both are available as open weights on Hugging Face under an MIT license, and can be run locally via the esm Python package or accessed through the Biohub Platform.
Predict a structure
Open NotebookBinder design protocol
ESMFold2 extends state-of-the-art structure prediction to the design of protein-protein interactions. Validated in laboratory experiments across five therapeutic targets, the approach generates mini-protein binders and single-chain antibodies (scFvs) with high success rates, nanomolar affinities, target specificity, and functional activity. The notebook walks through the full design loop from target sequence to ranked binder candidates. We are releasing the full protocol along with the paper to enable independent reproduction of these results.
app = DesignApp()
app.load(use_scaling_critics=False)
sequence, trajectory, scores = app.design(
# PD-L1. https://www.uniprot.org/uniprotkb/Q9NZQ7. 17-132.
target_sequence="AFTVTVPKDLYVVEYGSNMTIECKFPVEKQLDLAALIVYWEMEDKNIIQFVHGEEDLKVQHSSYRQRARLLKDQLSLGNAALQITDVKLQDAGVYRCMISYGGADYKRITVKVNA",
# VHVL framework from Trastuzumab. '#' denotes a mutable position.
binder_sequence="EVQLVESGGGLVQPGGSLRLSCAAS#######YIHWVRQAPGKGLEWVARI#####TRYADSVKGRFTISADTSKNTAYLQMNSLRAEDTAVYYCSR###########WGQGTLVTVSSGGGSGGGSGGGSGGGSDIQMTQSPSSLSASVGDRVTITC###########WYQQKPGKAPKLLIY#######GVPSRFSGSRSGTDFTLTISSLQPEDFATYYC#########FGQGTKVEIK",
is_antibody=True,
)ESMC
ESMC is a state-of-the-art protein language model. Available in three sizes (300M, 600M, and 6B parameters), ESMC learns a general world model of protein biology whose representations improve predictably with scale. The 6B parameter model powers both ESM Atlas and ESMFold2, enabling applications from large-scale functional mapping to state-of-the-art structure prediction, and protein design.
Each model size delivers state-of-the-art performance for its class, and all three are available as open weights on Hugging Face under an MIT license. They can be run locally via the esm Python package or accessed through the Biohub Platform.
ESM Atlas
ESM Atlas is an open database of the predicted structures and functional annotations for over 6.8 billion proteins, with over 1.1 billion paired with high-resolution structures predicted by ESMFold2. Sparse coding of ESMC's representation space computed for all proteins in the Atlas surface structural and functional relationships.
Atlas API and Documentation
Learn about SAE features and the other key concepts behind the Atlas. Discover how to search for proteins of interest, visualize 3D structures and activated features, and investigate cluster relationships. The Atlas API provides programmatic access to the same predicted structures, biological features, and protein clusters available in the Atlas web app. Full API reference documentation is available here.
Download Atlas Data
The ESM Atlas dataset is available for download from AWS S3 at no cost. The table below lists the available datasets with their approximate sizes and the AWS CLI command to download each one.
SAE Clusters is the most manageable entry point for exploring cluster-level organization. For structure-based work, download Structures Only. SAE Features (Pooled) can be used for per-protein feature vectors. Download All Data for a complete set of sequences, structures, and features across all 6.8 billion proteins.
Dataset | Size | CLI Command |
|---|---|---|
SequencesProtein sequences (6.8B proteins) | 2.20 TB | |
StructuresProtein structures (1B proteins) | 68.9 TB | |
SAE featuresPer protein and per-residue feature vectors (6.8B proteins) | 306 TB | |
SAE Clusters708M non-singleton clusters, 7.7M with ≥50 members | 26.0 GB | |
HMM ResultsPredicted pfam and taxonomy (6.8B proteins) | 653 MB | |
Protein_to_accessionMapping of protein IDs to accession numbers (6.8B proteins) | 162 GB | |
NormalizationSAE feature normalization | 192 KB | |
All DataComplete set of sequences, structures, features, and clusters | 377 TB |
All data is made available under CC-BY-4.0 (see link). The AWS CLI must be installed to use the commands above. See the AWS documentation for setup instructions.
Provide Feedback
If you encounter issues or have questions, please post in the ESM Slack community or submit a Feedback Form.