BoltzGen

Pre-trained model for protein and peptide binder design through joint sequence and structure generation, optional sequence redesign, and refolding-based evaluation.

Disclaimer

This is an UNOFFICIAL implementation of BoltzGen: Toward Universal Binder Design by Hannes Stark, Felix Faltings, et al.

The OFFICIAL repository of BoltzGen is at HannesStark/boltzgen.

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

Model Details

BoltzGen is an all-atom diffusion model for biomolecular binder design. A recycled Pairformer combines atom, token, MSA, template, contact, binding-site and secondary-structure features before an EDM denoiser generates sequence and atomic coordinates together. A separate inverse-folding model can redesign the sequence on the generated backbone.

The complete workflow contains seven child pipelines: diverse and adherence structure generators, BoltzIF and SolubleMPNN inverse folders, Boltz-2 refolding and affinity scoring, and ESMFold2 refolding. Please refer to the Training Details section for more information on the training process.

Variants

oteam/boltzgen: The complete design workflow with full Pairformer structure generators and all evaluation children.
oteam/boltzgen-diverse and oteam/boltzgen-adherence: Independently callable structure generators.
oteam/boltzgen-structure-small: A Miniformer structure generator.
oteam/boltzgen-inversefolding: The BoltzIF model for sequence assignment on fixed backbones.

Model Specification

Variant	Representation Size	Trunk	Denoiser / Decoder Blocks	Num Parameters (M)
Diverse / Adherence	384 / 128	64 Pairformer + 4 MSA Blocks	24 token Blocks	482.213
Structure-trained small	384 / 128	12 Miniformer + 3 MSA Blocks	8 token Blocks	136.886
Inverse folding	128 / 128	6 graph encoder Blocks, 30 neighbors	3 graph decoder Blocks	3.069

Links

Code: ofoldx.models.boltzgen
Paper: BoltzGen: Toward Universal Binder Design
Developed by: Hannes Stark, Felix Faltings, MinGyu Choi, Yuxin Xie, Eunsu Hur, Timothy O'Donnell, Anton Bushuiev, Talip Uçar, Saro Passaro, Weian Mao, Mateo Reveiz, Roman Bushuiev, Tally Portnoi, Tomáš Pluskal, Josef Sivic, Karsten Kreis, Arash Vahdat, Shamayeeta Ray, Jonathan T. Goldstein, Andrew Savinov, Jacob A. Hambalek, Anshika Gupta, Diego A. Taquiri-Diaz, Yaotian Zhang, Samuel J. Snyder, A. Katherine Hatstat, Angelika Arada, Nam Hyeong Kim, Haoyu Fan, Ethel Tackie-Yarboi, Dylan Boselli, Lee Schnaider, Chang C. Liu, Gene-Wei Li, Denes Hnisz, David M. Sabatini, William F. DeGrado, Jeremy Wohlwend, Gabriele Corso, Regina Barzilay, and Tommi Jaakkola
Model type: Constraint-guided all-atom biomolecular diffusion and inverse folding
Original Repository: HannesStark/boltzgen

Usage

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

pip install ofoldx

Direct Use

Biomolecular Design

You can use this model directly with a pipeline for binder design with the protein-anything protocol:

import ofoldx
from ofoldx.io import read_structure

structure, _ = read_structure("target.cif")
pipeline = ofoldx.pipeline("design", model="oteam/boltzgen")
result = pipeline(structure, protocol="protein-anything", total_len=120, num_candidates=20, seed=7)
best = result.ranked_candidates[0]

The result retains every analyzed candidate and stable ranked, top-budget and diversity-selected views. Each candidate contains its original codesign structure, the active design structure, refolding predictions and named analysis values. Screening uses the available atoms, as in the released workflow; each analysis retains its atom coverage and reports complete-chemistry metrics separately when available. After inverse folding, screening uses the refolded structure; without inverse folding, it uses the original codesign. For single-chain redesign, the cross-chain iPTM scores use the upstream empty-interface convention and do not describe an actual interface; this can leave candidates tied in ranking.

The other released protocols are peptide-anything, protein-small_molecule, nanobody-anything, antibody-anything and protein-redesign. The default inverse folder is SolubleMPNN, or BoltzIF for protein-small_molecule. Use inverse_fold_model="boltzif" to select BoltzIF or inverse_fold_model=None to retain the generated sequence and coordinates without inverse folding.

RNA and DNA can be fixed context but cannot be generated segments. When inverse folding changes a residue's identity, the new side-chain coordinates remain unavailable until a later structure prediction.

Downstream Use

Partial Workflows

Use design_backbones() for joint sequence and structure generation or inverse_fold() for sequence redesign without evaluation. These operations return candidates in generation order; only the complete design() call provides analysis and ranking. The full and Miniformer structure-generation pipelines and BoltzIF are also independently callable.

The pipeline guide describes staged execution and composition. The BoltzGenPipeline API describes protocol controls, selection budgets and child components; the analysis guide covers candidate and diversity analysis.

Checkpoint Conversion

Convert the original design and evaluation checkpoints together with a native ESMFold2 repository:

python -m ofoldx.models.boltzgen.convert \
  boltzgen1_diverse.ckpt \
  boltzgen1_adherence.ckpt \
  boltzgen1_ifold.ckpt \
  solublempnn_v_48_020.pt \
  boltz2_conf_final.ckpt \
  boltz2_aff.ckpt \
  oteam/esmfold2 \
  ./boltzgen

The complete design workflow requires all seven converted child pipelines.

Training Details

BoltzGen trains coordinate generation with EDM denoising and inverse folding with a separate structure-conditioned sequence objective.

Training Data

BoltzGen builds on the Boltz-2 data pipeline with biomolecular structures, MSA evidence, molecular dictionaries and optional distillation data. The inverse-folding model uses structure-conditioned sequence targets.

Training Procedure

The structure designer learns from cropped structures with sampled design and conditioning regions. Training noise follows sigma = sigma_data * exp(-1.2 + 1.5 * z), z ~ N(0, 1), with EDM loss weight (sigma^2 + sigma_data^2) / (sigma * sigma_data)^2.

OFoldX's built-in Runner uses the shared aligned EDM coordinate error for structure generation. The separate diffusion_loss function exposes the fuller family coordinate objective; see the Runner guide for training and export.

Coordinate Loss

diffusion_loss combines weighted, rigid-aligned EDM MSE, smooth lDDT and optional inter-residue bond-length error. Smooth lDDT is enabled by default; enable bond loss with add_bond_loss=True. Its API documents masking, sample and pair-memory budgets, and the unreduced (B, S) component values.

Citation

@article{stark2025boltzgen,
  author = {Stark, Hannes and Faltings, Felix and Choi, MinGyu and others},
  title = {BoltzGen: Toward Universal Binder Design},
  year = {2025},
  doi = {10.1101/2025.11.20.689494},
  journal = {bioRxiv}
}

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 BoltzGen 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

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

Other

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support