boltz2 / README.md

Sync model card with OFoldX README

a3c946d verified about 1 hour ago

6.85 kB

tags:
  - Biology
  - Protein
  - RNA
  - DNA
  - Structure Prediction
  - Binding Affinity
license: agpl-3.0
library_name: ofoldx
pipeline_tag: other

Boltz-2

Pre-trained model for all-atom structure prediction and protein–ligand binding affinity prediction.

Disclaimer

This is an UNOFFICIAL implementation of the Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction by Saro Passaro, et al.

The OFFICIAL repository of Boltz-2 is at jwohlwend/boltz.

The team releasing Boltz-2 did not write this model card for this model so this model card has been written by the OFoldX team.

Model Details

Boltz-2 uses atom attention, template and MSA encoders, a recycled Pairformer trunk, and EDM coordinate diffusion for structure prediction. A separate affinity model adds pairwise affinity heads. Its two-stage affinity workflow first predicts and ranks structures, then resamples and scores the selected complex.

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

Variants

oteam/boltz2: Biomolecular structure and confidence prediction.
oteam/boltz2-affinity: A composite structure-and-affinity workflow for protein-ligand complexes.

Model Specification

Variant	Pairformer Blocks	Template Blocks	Confidence Blocks	Affinity Blocks	Single Size	Pair Size	Trunk Passes	Num Parameters (M)
Structure	64	2	8	none	384	128	4	506.762
Affinity scorer	64	0	8	8 + 4	384	128	6	515.088

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("complex.cif")
pipeline = ofoldx.pipeline("fold", model="oteam/boltz2")
result = pipeline(structure, seed=7)

The result contains predicted structures and confidence values. designed_positions selects already-designed component occurrences for design-focused confidence; the exact selection and the molecular scope expanded through covalent connections remain distinct.

Binding Affinity Prediction

You can use this model directly with a pipeline for binding affinity prediction on a folded complex:

import ofoldx
from ofoldx.io import read_structure

structure, report = read_structure("complex.cif")
pipeline = ofoldx.pipeline("score", model="oteam/boltz2-affinity")
result = pipeline(structure, seed=7)

The pipeline selects the highest-confidence folded structure, then resamples and scores it with its affinity model. The result retains the structure and affinity predictions.

Downstream Use

Physical Guidance

Boltz-2 steering uses physical potentials during sampling. Active geometry and van der Waals pairs require known atomic numbers and supported tabulated radii; unsupported elements raise errors. A soft element distribution does not determine one radius. These empirical potentials guide sampling and do not establish chemical identity or validity.

See the pipeline guide for staged execution and composition.

Training Details

Boltz-2 is trained for biomolecular structure prediction and protein–ligand binding affinity prediction.

Training Data

The Boltz-2 report describes PDB structures before 2023-06-01, NMR data, molecular-dynamics trajectories, AlphaFold 2 monomer distillation, and Boltz-1 complex distillation. Affinity data comes from PubChem, ChEMBL, BindingDB, and filtered assay sources with decoy generation and structural-confidence filtering. OFoldX does not redistribute these datasets.

Training Procedure

Boltz-2 stages structure, confidence, and affinity training. The affinity objectives include detached-trunk absolute and difference Huber losses, together with a focal binary objective for binder classification.

OFoldX exposes the structure coordinate-denoising objective. It does not reproduce staged confidence and affinity training or the upstream data and decoy-generation pipelines.

Citation

@article{passaro2025boltz2,
  author = {Passaro, Saro and Corso, Gabriele and Wohlwend, Jeremy and others},
  title = {Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction},
  year = {2025},
  doi = {10.1101/2025.06.14.659707},
  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 Boltz-2 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

oteam
/

boltz2