Protein Structure · biohub
ESMFold2
Predicts all-atom 3D structures for proteins, nucleic acids, and ligands from sequence or sequence alignments.
What ESMFold2 is
ESMFold2 is an all-atom biomolecular structure prediction and design model developed by Biohub. It generates high-resolution three-dimensional coordinates for proteins directly from amino acid sequences, with optional conditioning on multiple sequence alignments (MSAs) to improve accuracy on difficult targets. Unlike its predecessor, ESMFold2 predicts structures across diverse biomolecular classes, including DNA, RNA, small-molecule ligands, and modified amino acids.
The model addresses the challenge of modelling complex biological assemblies and generating functional biomolecules, including protein binders and antibodies. On benchmarks such as FoldBench and Runs N' Poses, ESMFold2 meets or exceeds the performance of AlphaFold3 on protein-protein and antibody-antigen complex predictions.
How it works
ESMFold2 accepts single sequences or multi-chain inputs, with the option to incorporate paired MSAs to exploit inter-chain co-evolutionary signals. Users can also specify nucleic acid sequences with modified bases and small-molecule chemical component dictionary (CCD) identifiers. Structure generation uses iterative sampling, where parameters such as the number of recycling loops, sampling steps, and diffusion samples can be scaled at inference time to enhance structural fidelity.
The model was trained on structural data from the Protein Data Bank (PDB) and the AlphaFold DB (AFDB) using a data cutoff of September 2021. It outputs full all-atom Cartesian coordinates along with confidence metrics, including the predicted Local Distance Difference Test (pLDDT), predicted Aligned Error (pAE), predicted Template Modelling score (pTM), interface pTM (ipTM), and optional distograms.
Key facts
- Tasks
- Protein structure prediction and macromolecular complex modelling
- Input modalities
- Protein sequences, DNA, RNA, small-molecule ligands, multiple sequence alignments
- Output modalities
- All-atom 3D coordinates (PDB/mmCIF), confidence metrics (pLDDT, pAE, pTM, ipTM)
- Training data
- Protein Data Bank (PDB) and AlphaFold DB (AFDB)
- Data cutoff
- September 2021
- GPU memory
- 16 GB
What you can use it for
- Antibody-Antigen Complex Modelling
- Predicting the quaternary arrangement and binding interfaces of antibody-antigen complexes using paired sequence alignments.
- Multi-Biomolecule Complex Prediction
- Simultaneously folding systems comprising proteins, DNA, RNA, post-translationally modified residues, and small-molecule cofactors.
- Protein Binder Generation and Design
- Guiding de novo protein design campaigns by evaluating structural hypotheses and interface confidence metrics such as ipTM.
- Single-Sequence Protein Screening
- Predicting tertiary structures rapidly from single protein sequences without generating multiple sequence alignments.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
ESMFold2
Strengths and limits
Good at
- Predicts full atomistic structures across proteins, nucleic acids, modified residues, and small molecules within a single workflow.
- Supports both single-sequence prediction and MSA-conditioned inference to balance speed and accuracy.
- Exhibits improved prediction accuracy when scaling inference-time compute parameters like loops and sampling steps.
- Provides residue-level and complex-level confidence scores, including pLDDT, pAE, pTM, and ipTM.
Watch out for
- Reflects historical biases in the PDB and AFDB, leading to potential performance drops on underrepresented protein families or non-protein biomolecules.
- Lacks comprehensive performance guarantees on disordered regions and conformations that are poorly captured in training databases.
- Requires increased inference-time compute to achieve its highest prediction accuracy, especially on challenging complexes.
- Produces structural hypotheses that require confirmation through experimental methods such as X-ray crystallography, cryo-EM, or NMR.
Inputs and outputs
You provide
Protein sequence
One-letter amino-acid codes.
You get back
All-atom coordinates in PDB format, viewable and downloadable.
Residue count, mean pLDDT (0–100) and pTM.
Licence
Check the licence before commercial use · See model card
The See model card licence carries terms worth reading before you depend on it.
Questions
Can I use ESMFold2 commercially?
Check the licence before commercial use. The See model card licence carries terms worth reading before you depend on it.
How do I run ESMFold2 without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on CPU only, typically in 12s, and a run costs about 28 credits. Visitors without an account get 1,000 credits a month.
What biomolecules can ESMFold2 fold?
ESMFold2 predicts structures for proteins, DNA, RNA, small-molecule ligands, and modified amino acids, including multi-component complexes containing all of these elements.
Is a multiple sequence alignment required to use ESMFold2?
No. ESMFold2 can perform single-sequence predictions, although providing single or paired MSAs is supported to improve accuracy on difficult targets and multi-chain complexes.
How does ESMFold2 differ from ESMFold2-Fast?
ESMFold2 is a larger model that supports both single-sequence and MSA-conditioned prediction, whereas ESMFold2-Fast is an inference-optimised variant restricted to single-sequence inputs.
What confidence estimates does the model output?
The model provides residue-level pLDDT confidence scores, predicted Aligned Error (pAE), global pTM scores, interface ipTM scores, and optional distograms.
How to cite it
Candido, S., Hayes, T., Derry, A., Rao, R., Lin, Z., Verkuil, R., Wu, B., Lee, J. S., Bruguera, E. S., Keval, J. A., Kopylov, M., Pak, J. E., Wu, W., Thomas, N., Mataraso, S., Hsu, A., Trotman-Grant, A. C., Fatras, K., dos Santos Costa, A., Badkundri, R., Akın, H., Oktay, D., Deaton, J., Montabana, E., Sitwala, H., Yu, Y., Wiggert, M., Carlin, D. A., Goering, A. W., Blazejewski, T., Sandora, M., Hla, M., Jia, T. Z., Kloker, L. H., Sofroniew, N. J., Uehara, M., Pannu, J., Bachas, S., Liu, D. S., Sercu, T., & Rives, A. (2026). Language Modeling Materializes a World Model of Protein Biology. bioRxiv. https://www.biorxiv.org/content/10.64898/2026.06.03.729735
Written from ESMFold2's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.