oteam/boltz1

Boltz-1

Boltz-1 is OFoldX's all-atom structure-prediction implementation for biomolecular complexes, with a recycled representation trunk, EDM coordinate sampling, confidence scoring, pocket conditioning, and optional Boltz-1x sampling-time steering.

Disclaimer

This is an unofficial OFoldX implementation of Boltz-1. The upstream authors did not write, review, or endorse this implementation or model card.

Model Details

Boltz-1 predicts all-atom structures of biomolecular complexes. Atom and token features are refined by an MSA module and recycled Pairformer trunk, an EDM diffusion model generates coordinates, and a dedicated confidence trunk predicts pLDDT, PAE, PDE, pTM, ipTM, and ranking scores. The same model supports pocket conditioning and optional sampling-time steering.

Variants

• oteam/boltz1: Boltz-1 structure and confidence model.

Boltz-1x physical-potential and Feynman-Kac steering are inference controls on the same model rather than separate variants.

Model Specification

Pairformer Blocks	MSA Blocks	Denoiser Blocks	Confidence Blocks	Single Size	Pair Size	Trunk Passes	Sampling Steps	Num Parameters (M)
48	4	24	48	384	128	4	200	592.008

Links

• Code: ofoldx.models.boltz1
• Developed by: Jeremy Wohlwend, Gabriele Corso, and the Boltz team
• Model type: Pairformer-based all-atom biomolecular structure predictor with EDM coordinate diffusion
• Original Repository: jwohlwend/boltz

Feature and geometry contract

All public model coordinate boundaries consume an explicit finite numerical projection. Unavailable values are selected out under the declared mask before arithmetic or dtype conversion and have zero coordinate derivatives. Active nonfinite values are invalid requests; finite values outside the model dtype range are an unsupported numeric range. Normalization returns the same input carrier type without changing canonical membership, source masks, or source precision. Decode uses prepared atom correspondences. APIs without an availability mask require complete finite model fields.

Biological classes, token grouping and actual centers are prepared once. The token-relation bit is the symmetric presence of any represented positive chemical edge; it discards kind, role, order and direction and cannot certify covalence or known absence. Reference conformers are complete model features, so their augmentation availability is independent of experimental source observation masks. Generated reference geometry does not create canonical observations. Directed dative geometry receives an explicit capability error at the conformer operation that cannot represent it.

Usage

The model depends on the ofoldx library. Install it with pip:

pip install ofoldx

Structure Prediction

Load the concrete folding Pipeline:

import ofoldx
from ofoldx.io import read_structure

structure, report = read_structure("input.cif")
pipeline = ofoldx.pipeline("fold", model="oteam/boltz1")
output = pipeline.fold(structure)

The Pipeline is callable with the same arguments as fold. Staged execution uses Boltz1Input, a NumPy generator for preparation and a PyTorch generator for sampling:

import numpy as np
import torch

generator = torch.Generator(device=pipeline.processor.device).manual_seed(42)
prepared = pipeline.prepare(structure, generator=np.random.default_rng(42))
raw = pipeline.forward(prepared.input, num_samples=1, generator=generator)
result = pipeline.postprocess(prepared, raw)

Training Details

Training Data

Refer to the official Boltz project for the training data and data-processing procedure used by Boltz-1.

Training Procedure

The released model uses an EDM coordinate-denoising objective together with a trunk distogram objective and a confidence/ranking objective. Training noise is drawn as sigma = sigma_data * exp(-1.2 + 1.5 * z), where z ~ N(0, 1), and the coordinate loss weight is (sigma^2 + sigma_data^2) / (sigma * sigma_data)^2. OFoldX exposes the EDM coordinate-denoising term; it does not reproduce the upstream joint schedule for distogram, confidence, and ranking-head training, or the upstream data pipeline.

Physical guidance reads the declared exact atomic-number category only for endpoints of the final available geometry and van der Waals pairs. These endpoints need known elements with supported radius-table entries; unknown elements or unsupported radii raise typed errors. A soft element distribution does not define a physical radius. Unavailable endpoints and atoms outside all such pairs impose no radius requirement. This table is an empirical guidance convention, not a chemical identity or validity test.

Citation

Please cite the Boltz project described by the official repository. If OFoldX supports your research, also cite the OFoldX software:

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

Contact

Please use OFoldX GitHub issues for questions about this implementation or model card.

License

The original model parameters are licensed under MIT. OFoldX conversion changes parameter names, packing and vocabulary axes.

This model implementation is licensed under the GNU Affero General Public License.

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