Protenix
Pre-trained model for all-atom structure prediction of proteins, nucleic acids, small molecules, and their complexes.
Disclaimer
This is an UNOFFICIAL implementation of the Protenix - Advancing Structure Prediction Through a Comprehensive AlphaFold3 Reproduction by Xinshi Chen, Yuxuan Zhang, et al.
The OFFICIAL repository of Protenix is at bytedance/Protenix.
The team releasing Protenix did not write this model card for this model so this model card has been written by the OFoldX team.
Model Details
Protenix combines atom and token embeddings with an MSA-conditioned recycled Pairformer trunk, EDM coordinate diffusion, and distogram and confidence heads. Template, structure-conditioned, and constraint variants provide additional structural evidence; the constraint model supports atom contacts, token contacts, pockets, and substructures.
Please refer to the Training Details section for more information on the training process.
Variants
- oteam/protenix: The default v1.0 all-atom structure predictor.
- oteam/protenix-20250630: The v1.0 model trained on a later data cutoff.
- oteam/protenix-v0.5: The v0.5 base model, also available as oteam/protenix-base.
- oteam/protenix-constraint: The v0.5 model with constraint conditioning.
- oteam/protenix-mini-tmpl: The v0.5 mini model with structure conditioning.
- Protenix-v2: The larger template-conditioned model, using locally converted parameters.
The converter supports ten checkpoint identifiers, grouped below by their registered model and preparation rules. Select the identifier explicitly; checkpoint filenames and tensor shapes do not select it.
| Graph | Supported checkpoint identifiers |
|---|---|
| Default | protenix_base_default_v0.5.0, protenix_mini_default_v0.5.0, protenix_tiny_default_v0.5.0, protenix_mini_esm_v0.5.0, protenix_mini_ism_v0.5.0 |
| Template | protenix_base_default_v1.0.0, protenix_base_20250630_v1.0.0, protenix-v2 |
| Structure condition | protenix_mini_tmpl_v0.5.0 |
| Constraint | protenix_base_constraint_v0.5.0 |
The two v1 checkpoints share a graph and differ in weights. Mini ESM and ISM share a graph and package the selected ESM2 companion weights with the model. The mini template checkpoint instead uses noisy structure conditioning.
Model Specification
| Variant | Pairformer Blocks | MSA Stages | Single Size | Pair Size | Recycling Steps | Num Parameters (M) |
|---|---|---|---|---|---|---|
| Base v0.5 | 48 | 4 | 384 | 128 | 10 | 368.123 |
| Constraint v0.5 | 48 | 4 | 384 | 128 | 10 | 368.339 |
| Template v1 | 48 | 4 | 384 | 128 | 10 | 368.549 |
| Mini v0.5 | 16 | 1 | 384 | 128 | 4 | 134.102 |
| Mini ESM / ISM v0.5 | 16 | 1 | 384 | 128 | 4 | 2967.899 |
| Structure condition v0.5 | 16 | 1 | 384 | 128 | 4 | 134.212 |
| Tiny v0.5 | 8 | 1 | 384 | 128 | 4 | 109.535 |
| Protenix-v2 | 48 | 4 | 384 | 256 | 10 | 464.535 |
The Mini ESM and ISM counts include the frozen ESM2 encoder.
Links
- Code: ofoldx.models.protenix
- Paper: Protenix - Advancing Structure Prediction Through a Comprehensive AlphaFold3 Reproduction
- Developed by: Xinshi Chen, Yuxuan Zhang, Chan Lu, Wenzhi Ma, Jiaqi Guan, Chengyue Gong, Jincai Yang, Hanyu Zhang, Ke Zhang, Shenghao Wu, Kuangqi Zhou, Yanping Yang, Zhenyu Liu, Lan Wang, Bo Shi, Shaochen Shi, and Wenzhi Xiao
- Model type: Pairformer-based all-atom biomolecular structure predictor with optional learned constraints
- Original Repository: bytedance/Protenix
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/protenix")
result = pipeline(structure, seed=7)
Represented atoms require known elements and formal charges. Predictions use a new coordinate frame and rank by
0.8 * ipTM + 0.2 * pTM - 100 * has_clash; this implementation reports no disorder estimate.
Downstream Use
Chemical Guidance
You can enable training-free chemical guidance during sampling:
import ofoldx
from ofoldx.io import read_structure
from ofoldx.models.protenix import ProtenixTfgGuidance
structure, report = read_structure("input.cif")
pipeline = ofoldx.pipeline("fold", model="oteam/protenix")
result = pipeline(structure, tfg_guidance=ProtenixTfgGuidance(), seed=7)
Guidance is disabled by default. Its configuration controls eight chemical terms, schedules, Monte Carlo smoothing,
and iteration counts; setting a term to None removes it.
Template repositories accept templates=. Structure-conditioned repositories accept structure_condition= with an
in-memory structure. Constraint repositories accept atom_contacts, token_contacts, pocket, and substructure=.
The mini template checkpoint uses the structure-conditioned graph, rather than the ordinary template graph.
See the pipeline guide for staged execution and composition.
Training Details
Protenix is trained with coordinate, geometric, distogram, and confidence objectives in a staged training program.
Training Data
The Protenix report describes the datasets and data policies used for its four-stage upstream training program. Reproducing that program requires the upstream dataset construction and condition-dropout curriculum.
Training Procedure
Protenix combines coordinate, smooth lDDT, distogram, and confidence objectives with stage-dependent weights. Full confidence supervision uses an independent mini-rollout, prediction-dependent truth matching, and detached trunk representations.
OFoldX exposes a shared EDM coordinate objective with compute_auxiliary_heads=False. Its rigid-aligned MSE weights
resolved atoms equally and omits Protenix's molecule weights, 1/3 coordinate normalization, auxiliary losses, and
stage-dependent composition. The differentiable raw model operations do not by themselves reproduce the upstream
mini-rollout, optimizer, or orchestration. See processor data preparation
for the shared input/target interface.
Citation
@article{protenix2025advancing,
author = {ByteDance AML AI4Science Team and Chen, Xinshi and Zhang, Yuxuan and others},
title = {Protenix - Advancing Structure Prediction Through a Comprehensive AlphaFold3 Reproduction},
year = {2025},
doi = {10.1101/2025.01.08.631967},
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 Protenix 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