ESMFold2 - Biohub Platform

ESMFold2 is the successor to ESMFold and sets a new state of the art for single-sequence structure prediction. It also enables the generation of new functional proteins through searching the ESMC model’s latent space. The model predicts high-resolution, all-atom 3D structures of biomolecular complexes directly from sequence, with optional multiple sequence alignment (MSA) input for enhanced accuracy on challenging targets.

Hugging Face

GitHub

Paper

API reference

Get started

Get started with ESMFold2 with our Quickstart and Tutorials.

Quickstart guide

Install the esm Python package

pip install esm

Create an API key

Generate an API key from your Biohub account. This API key manages your access to credits and tokens, and the term API key/token is often used interchangeably within documentation.

Connect to the Biohub Platform API

Call the inference client with your selected model and replace <your API token> with your token name.

from esm.sdk.forge import SequenceStructureForgeInferenceClient

client = SequenceStructureForgeInferenceClient(model="esmfold2-fast-2026-05", url="https://biohub.ai", token="<your API token>")

Run your inference

Now you are ready to use your model. For examples of specific use cases, check out our Tutorials.

Model tutorials

Folding with ESMFold2

Fold proteins in combination with DNA, RNA, and small-molecule ligands.

Binder design

Design antibodies and minibinders with high hit rates. Implements the protocol featured in our paper, which produced binders exhibiting nanomolar affinity, target specificity, and functional activity in laboratory assays.

Model details

For additional information, see the Hugging Face link.

Model card

Cite this model

Language Modeling Materializes a World Model of Protein Biology

@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}
}

Overview
Details
Variants
Usage


Version	2026-05
Architecture	ESM representations that power a series of looped folding layers. A diffusion model projects pairwise representations to atomic resolution predictions.
Supported Modalities	Sequence and structure
Training Data	ESMFold2 was trained on sequences from the Protein Data Bank (PDB) and the AlphaFold DB (AFDB).

Intended use

ESMFold2 predicts high-resolution, all-atom 3D protein structures directly from amino acid sequences, with optional multiple sequence alignment (MSA) input. The outputs include comprehensive structural information including all-atom coordinates (backbone and side chains), confidence metrics, and optional distogram predictions for detailed analysis of predicted structures.

Limitations & risks

The model predicts single static conformations and is not designed for modeling protein dynamics, conformational flexibility, or multiple conformations of the same protein. Validate outputs experimentally. Not intended for clinical or therapeutic applications without further validation.This model is released under the MIT License.


Type of model architecture	Hybrid architecture that combines the representational power of large-scale protein language models with an efficient structure prediction module. ESM representations that power a series of looped folding layers. A diffusion model projects pairwise representations to atomic resolution predictions. The architecture stages include a language model embedding stage which uses a frozen ESMC 6B parameter language model to generate rich single-sequence embeddings, a trunk of looped folding layers that refine pairwise and single representations, and a diffusion-based structure module that employs a diffusion process to generate all-atom 3D coordinates (backbone and sidechains) and uses an ODE (Ordinary Differential Equation) solver for efficient sampling.
Training datasets	ESMFold2 is trained on all biomolecules in the Protein Data Bank (PDB) and 8.8M synthetic structures from the AlphaFold DB (AFDB).

Model	MSA Conditioning	Description	Release Date
esmfold2-fast-2026-05	No	Inference optimized single-sequence structure prediction model	2026-05
esmfold2-2026-05	Yes	Flagship model, capable of either single-sequence or MSA conditioned structure prediction for improved accuracy on difficult targets	2026-05
esmfold2-2026-05-cutoff-2025	Yes	Flagship model trained with a data cutoff of 2025. Not available on Hugging Face	2026-05


Primary Use Case	ESMFold2 produces high-resolution, all-atom 3D protein folding structures directly from input 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.
Supported Input Modalities	Protein sequences
Access	Through Biohub or Hugging Face
License	This model is released under the MIT License.
Not recommended for	Clinical diagnosis or treatment recommendations. Computational metrics do not replace wet-lab validation. Treat model outputs as machine-generated hypotheses that require further experimental validation, not as established biological facts.

ESM Atlas data

The ESM Atlas is a map of 6.8 billion proteins covering the full breadth of life’s biodiversity and more than one billion predicted structures. ESM Atlas organizes protein space by the shared representation space learned by ESMC, with high-resolution structures predicted by ESMFold2. Users can download the dataset used to generate the Atlas from AWS S3 at no cost. See the Atlas data files and download commands in the ESM Atlas guide.

Models