oteam/esmfold2-fast

ESMFold2

Pre-trained models for all-atom structure prediction and gradient-based protein binder design.

Disclaimer

This is an UNOFFICIAL implementation of the Language Modeling Materializes a World Model of Protein Biology by Salvatore Candido, et al.

The OFFICIAL repository of ESMFold2 is at Biohub/esm.

The team releasing ESMFold2 did not write this model card for this model so this model card has been written by the OFoldX team.

Model Details

ESMFold2 uses a frozen ESMC language model to supply sequence representations to a recurrent folding trunk. A diffusion head generates atomic coordinates, and distogram and confidence heads score the prediction. Supported checkpoints also accept protein MSA evidence. Experimental variants expose differentiable inversion for protein binder sequence design.

Please refer to the Training Details section for more information on the training process.

Variants

• oteam/esmfold2: The full 48-block folding model with optional MSA conditioning.
• oteam/esmfold2-fast: The 24-block single-sequence folding model.
• oteam/esmfold2-experimental: The binder-design workflow.
• oteam/esmfold2-experimental-fast: The experimental fast model used for binder sequence optimization.

Model Specification

Variant	Folding Blocks	MSA	ESMC	Single Size	Pair Size	Atom Size	Token Size	Recurrent Updates	Sampling Steps	Num Parameters (M)
ESMFold2	48	yes	6B	384	256	128	768	4	14	6579.742
ESMFold2 Fast	24	no	6B	384	256	128	768	4	14	6533.723
Experimental	48	yes	6B	384	256	128	768	4	14	6577.328
Experimental Fast	24	no	6B	384	256	128	768	4	14	6530.863

Parameter counts include the frozen ESMC encoder.

Links

• Code: ofoldx.models.esmfold2
• Paper: Language Modeling Materializes a World Model of Protein Biology
• Developed by: Salvatore Candido, Thomas Hayes, Alexander Derry, Roshan Rao, Zeming Lin, Robert Verkuil, Bryan Wu, Jin Sub Lee, Elise S. Bruguera, Jehan A. Keval, Mykhailo Kopylov, John E. Pak, Wesley Wu, Neil Thomas, Samson Mataraso, Alvin Hsu, Ashton C. Trotman-Grant, Kilian Fatras, Allan dos Santos Costa, Rohil Badkundri, Halil Akın, Deniz Oktay, Jonathan Deaton, Elizabeth Montabana, Hrishita Sitwala, Yue Yu, Marius Wiggert, Dylan Alexander Carlin, Anthony W. Goering, Tomasz Blazejewski, McCullen Sandora, Michael Hla, Tina Z. Jia, Leon H. Kloker, Nicholas J. Sofroniew, Masatoshi Uehara, Jassi Pannu, Sharrol Bachas, Daniel S. Liu, Tom Sercu, and Alexander Rives
• Model type: ESMC-conditioned recurrent all-atom folding model with diffusion coordinate generation
• Original Repository: Biohub/esm

Usage

The model file depends on the ofoldx library. You can install it using pip:

pip install ofoldx

Direct Use

Structure Prediction

You can use this model directly with a pipeline for all-atom structure prediction:

import ofoldx
from ofoldx.io import read_structure

structure, report = read_structure("input.cif")
pipeline = ofoldx.pipeline("fold", model="oteam/esmfold2")
result = pipeline(structure, seed=7)

The result contains predicted structures and confidence values. The standard and Fast models share the frozen MultiMolecule ESMC-6B checkpoint; its exact revision is recorded in the model configuration and downloaded through the Hugging Face cache. Protein sequences are also accepted directly.

Binder Design

You can use this model directly with a pipeline for experimental binder sequence optimization:

import ofoldx

pipeline = ofoldx.pipeline("design", model="oteam/esmfold2-experimental")
result = pipeline("MKTAYIAKQRQISFVKSHFSRQ", num_binder_residues=80, seed=7)

The workflow optimizes binder sequences with gradients and ranks candidates using folding predictions. It requires the full experimental repository, including when scaling critics are disabled. Antibody designs are detected automatically unless is_antibody is specified. Antibody detection and CDR scoring use ANARCI and require HMMER's hmmscan executable on PATH.

Training Details

ESMFold2 combines pretrained folding models with learned confidence and scaling components for structure prediction.

Training Data

Biohub reports training on Protein Data Bank (PDB) and AlphaFold DB (AFDB) data, with a September 2021 cutoff for ESMFold2 and ESMFold2-Fast. ESMC was pretrained separately; consult its model card for the sequence corpus.

Training Procedure

ESMFold2 is trained with diffusion, distogram, and confidence objectives for all-atom structure prediction.

Citation

@misc{candido2026language,
  title = {Language Modeling Materializes a World Model of Protein Biology},
  author = {Candido, Salvatore and Hayes, Thomas and Derry, Alexander and others},
  year = {2026},
  url = {https://biohub.ai/papers/esm_protein.pdf},
  note = {Preprint}
}

The artifacts distributed in this repository are part of the OFoldX project. If OFoldX supports your research, please cite the OFoldX project as follows:

@software{chen2026ofoldx,
  author = {Chen, Zhiyuan},
  title  = {OFoldX},
  year   = {2026},
  url    = {https://github.com/OTeam-AI4S/OFoldX},
}

Contact

Please use GitHub issues of OFoldX for any questions or comments on the model implementation or model card.

Please contact the authors of the ESMFold2 paper for questions or comments on the paper/model.

License

The model weights, implementation, configuration, and documentation distributed by OFoldX are licensed under the GNU Affero General Public License.

For additional terms and clarifications, please refer to our License FAQ.

SPDX-License-Identifier: AGPL-3.0-or-later