ESMFold2 is a state-of-the-art model for protein structure prediction and design that defines a new frontier for speed and accuracy. The model predicts high-resolution, all-atom 3D protein structures directly from amino acid sequences, with optional multiple sequence alignment (MSA) input for enhanced accuracy on challenging targets. The model outputs comprehensive structural information including all-atom coordinates (backbone and side chains), confidence metrics (pLDDT, pAE, pTM, iPTM), and optional distogram predictions for detailed analysis of predicted structures. Unlike ESMFold, ESMFold2 is able to predict structures for all biomolecules, including small molecules, DNA, RNA, and modified amino acids.
ESMFold2 is capable of either single-sequence or MSA conditioned structure prediction for improved accuracy on difficult targets. The ESMFold2-Fast variant is an inference optimized single-sequence structure prediction model and is not MSA conditioned.
To run this model with the Biohub Platform API, visit the Biohub Platform.
Inference optimized single-sequence structure prediction model
Sept 2021
Performance Metrics
ESMfold2 was evaluated against state-of-the-art single-sequence and MSA-based structure prediction models on the FoldBench benchmark. ESMFold2 meets or exceeds performance by AlphaFold3 on antibody-antigen complex prediction, protein-protein complex prediction and Runs N' Poses benchmarks. Inference-time compute can dramatically improve performance of ESMFold2, especially across antibody-antigen complexes.
Refer to the paper for details on additional performance metrics.
You can also fold complex biomolecules — proteins, DNA/RNA (with modified residues), and small-molecule ligands all at once. Here's an example folding the HhaI DNA methyltransferase + its cognate DNA (with a trapped 5-fluoro-2′-deoxycytidine, CCD C36) + the SAH cofactor (PDB 1MHT):
For the Biohub API, first generate an API key and add it to your Biohub account. The code below assumes the environment variable ESM_API_KEY=$YOUR_API_KEY.
For multi-chain complexes (e.g. antibody–antigen), you can supply paired MSAs to take advantage of inter-chain co-evolution. See the ESMFold2 tutorial for details.
Training Data
ESMfold2 was trained on sequences from the Protein Data Bank (PDB) and the AlphaFold DB (AFDB).
Frontier Safety
Biohub has established a safety team to assess the benefits and potential risks of our models and tools prior to release, and develop mitigations where necessary. Risk assessment was conducted for ESMFold2 prior to release. Further details are available in our corresponding paper appendix.
Informed by our risk assessments, we are releasing the source code and model weights for ESMFold2.
Biohub.ai Platform: We implement guardrails that detect and restrict the use of keywords and sequences corresponding to controlled pathogens and toxins on our freely accessible platform. For further details regarding these guardrails, please refer to our Biohub platform Resources page.
Biases and Limitations
Dataset biases: The model may reflect biases present in the training data (PDB, AFDB), including over-representation of certain protein families, experimental conditions, or structural classes. Performance may vary for underrepresented protein types.
Dataset limitations: PDB historically lacks comprehensive data on protein conformations, post-translational modifications, disordered regions, etc. Like all other structure prediction models trained on the PDB, performance may degrade on other biomolecules.
Computational demand: Highest accuracy structure predictions require scaling inference time compute. Predictions made with reduced inference parameters may lead to suboptimal performance.
Experimental validation required: All predictions should be considered hypotheses requiring experimental validation. The model cannot replace experimental structure determination methods (X-ray crystallography, cryo-EM, NMR) for definitive structural characterization.
Always review and validate outputs generated by the model.
Treat model outputs as machine-generated hypotheses that require further experimental validation, not as established biological facts.
We are committed to advancing the responsible development and use of artificial intelligence.
Should you have any security or privacy issues or questions related to this model, please reach out to our team at support@biohub.org.
Citation
@misc{candido2026language,
title = {Language Modeling Materializes a World Model of Protein Biology},
author = {Candido, Salvatore and Hayes, Thomas and Derry, Alexander and Rao, Roshan
and Lin, Zeming and Verkuil, Robert and Wu, Bryan and Lee, Jin Sub
and Bruguera, Elise S. and Keval, Jehan A. and Kopylov, Mykhailo
and Pak, John E. and Wu, Wesley and Thomas, Neil and Mataraso, Samson
and Hsu, Alvin and Trotman-Grant, Ashton C. and Fatras, Kilian
and dos Santos Costa, Allan and Badkundri, Rohil and Ak{\i}n, Halil
and Oktay, Deniz and Deaton, Jonathan and Montabana, Elizabeth
and Sitwala, Hrishita and Yu, Yue and Wiggert, Marius
and Carlin, Dylan Alexander and Goering, Anthony W. and Blazejewski, Tomasz
and Sandora, McCullen and Hla, Michael and Jia, Tina Z.
and Kloker, Leon H. and Sofroniew, Nicholas J. and Uehara, Masatoshi
and Pannu, Jassi and Bachas, Sharrol and Liu, Daniel S.
and Sercu, Tom and Rives, Alexander},
year = {2026},
url = {https://www.biorxiv.org/content/10.64898/2026.06.03.729735},
note = {Preprint}
}
Acknowledgements
Many people on the Biohub AI Research team and prior EvolutionaryScale team contributed to the development of this model. It would not have been possible without them.
