📣📣📣 We're hiring!
Positions in Beijing, China and Seattle, US !
Interested in machine learning, computational chemistry/biology, structural biology, or drug discovery?
👉 Join us »
We’re excited to introduce Protenix — a trainable, open-source PyTorch reproduction of AlphaFold 3.
Protenix is built for high-accuracy structure prediction. It serves as an initial step in our journey toward advancing accessible and extensible research tools for the computational biology community.
- PXMeter is an open-source toolkit designed for reproducible evaluation of structure prediction models, released with high-quality benchmark dataset that has been manually reviewed to remove experimental artifacts and non-biological interactions. The associated study presents an in-depth comparative analysis of state-of-the-art models, drawing insights from extensive metric data and detailed case studies. The evaluation of Protenix is based on PXMeter.
- Protenix-Dock: Our implementation of a classical protein-ligand docking framework that leverages empirical scoring functions. Without using deep neural networks, Protenix-Dock delivers competitive performance in rigid docking tasks.
- 2025-07-17: Protenix-Mini released!: Lightweight model variants with significantly reduced inference cost are now available. Users can choose from multiple configurations to balance speed and accuracy based on deployment needs. See our paper and model configs for more information.
- 2025-07-17: New constraint feature is released! Now supports atom-level contact and pocket constraints, significantly improving performance in our evaluations.
- 2025-05-30: Protenix-v0.5.0 is now available! You may try Protenix-v0.5.0 by accessing the server, or upgrade to the latest version using pip.
- 2025-01-16: The preview version of constraint feature is released to branch
constraint_esm. - 2025-01-16: The training data pipeline is released.
- 2025-01-16: The MSA pipeline is released.
- 2025-01-16: Use local colabfold_search to generate protenix-compatible MSA.
We benchmarked the performance of Protenix-v0.5.0 against Boltz-1 and Chai-1 across multiple datasets, including PoseBusters v2, AF3 Nucleic Acid Complexes, AF3 Antibody Set, and our curated Recent PDB set.
Protenix-v0.5.0 was trained using a PDB cut-off date of September 30, 2021. For the comparative analysis, we adhered to AF3’s inference protocol, generating 25 predictions by employing 5 model seeds, with each seed yielding 5 diffusion samples. The predictions were subsequently ranked based on their respective ranking scores.
We will soon release the benchmarking toolkit, including the evaluation datasets, data curation pipeline, and metric calculators, to support transparent and reproducible benchmarking.
pip3 install protenixFor development on a CPU-only machine, it is convenient to install with the --cpu flag in editable mode:
python3 setup.py develop --cpu
Check the detailed guide: Docker Installation.
For details on the input JSON format and expected outputs, please refer to the Input/Output Documentation.
If your input is a .pdb or .cif file, you can convert it into a JSON file for inference.
# ensure `release_data/ccd_cache/components.cif` or run:
python scripts/gen_ccd_cache.py -c release_data/ccd_cache/ -n [num_cpu]
# for PDB
# download pdb file
wget https://files.rcsb.org/download/7pzb.pdb
# run with pdb/cif file, and convert it to json file for inference.
protenix tojson --input examples/7pzb.pdb --out_dir ./output
# for CIF (same process)
# download cif file
wget https://files.rcsb.org/download/7pzb.cif
# run with pdb/cif file, and convert it to json file for inference.
protenix tojson --input examples/7pzb.cif --out_dir ./outputWe provide an independent MSA search utility. You can run it using either a JSON file or a protein FASTA file.
# run msa search with json file, it will write precomputed msa dir info to a new json file.
protenix msa --input examples/example_without_msa.json --out_dir ./output
# run msa search with fasta file which only contains protein.
protenix msa --input examples/prot.fasta --out_dir ./outputIf you installed Protenix via pip, you can run the following command to perform model inference:
# 1. The default model_name is protenix_base_default_v0.5.0, you can modify it by passing --model_name xxxx
# 2. We provide recommended default configuration parameters for each model. To customize cycle/step/use_msa settings, you must set --use_default_params false
# 3. You can modify cycle/step/use_msa by passing --cycle x1 --step x2 --use_msa false
# run with example.json, which contains precomputed msa dir.
protenix predict --input examples/example.json --out_dir ./output --seeds 101 --model_name "protenix_base_default_v0.5.0"
# run with example.json, we use only esm feature.
protenix predict --input examples/example.json --out_dir ./output --seeds 101 --model_name "protenix_mini_esm_v0.5.0" --use_msa false
# run with multiple json files, the default seed is 101.
protenix predict --input ./jsons_dir/ --out_dir ./output
# if the json do not contain precomputed msa dir,
# add --use_msa (default: true) to search msa and then predict.
# if mutiple seeds are provided, split them by comma.
protenix predict --input examples/example_without_msa.json --out_dir ./output --seeds 101,102 --use_msa trueAlternatively you can run inference by: Alternatively, run inference via script:
bash inference_demo.shThe script accepts the following arguments:
model_name: Name of the model to use for inference.input_json_path: Path to a JSON file that fully specifies the input structure.dump_dir: Directory where inference results will be saved.dtype: Data type used during inference. Supported options:bf16andfp32.use_msa: Whether to enable MSA features (default: true).
Note: By default, layernorm and EvoformerAttention kernels are disabled for simplicity. To enable them and speed up inference, see the Kernels Setup Guide.
Refer to the Training Documentation for setup and details.
Protenix supports specifying contacts (at both residue and atom levels) and pocket constraints as extra guidance. Our benchmark results demonstrate that constraint-guided predictions are significantly more accurate.See our doc for input format details.
We introduce Protenix-Mini, a lightweight variant of Protenix that uses reduced network blocks and few ODE steps (even as few as one or two steps) to enable efficient prediction of biomolecular complex structures. Experimental results show that Protenix-Mini achieves a favorable balance between efficiency and accuracy, with only a marginal 1–5% drop in evaluation metrics such as interface LDDT, complex LDDT, and ligand RMSD success rate. Protenix-Mini enables accurate structure prediction in high-throughput and resource-limited scenarios, making it well-suited for practical applications at scale. The following comparisons were performed on a subset of the RecentPDB dataset comprising sequences with fewer than 768 tokens.
For details on memory usage and runtime during training and inference, refer to the Training & Inference Cost Documentation.
If you use Protenix in your research, please cite the following:
@article{bytedance2025protenix,
title={Protenix - Advancing Structure Prediction Through a Comprehensive AlphaFold3 Reproduction},
author={ByteDance AML AI4Science Team and Chen, Xinshi and Zhang, Yuxuan and Lu, Chan and Ma, Wenzhi and Guan, Jiaqi and Gong, Chengyue and Yang, Jincai and Zhang, Hanyu and Zhang, Ke and Wu, Shenghao and Zhou, Kuangqi and Yang, Yanping and Liu, Zhenyu and Wang, Lan and Shi, Bo and Shi, Shaochen and Xiao, Wenzhi},
year={2025},
journal={bioRxiv},
publisher={Cold Spring Harbor Laboratory},
doi={10.1101/2025.01.08.631967},
URL={https://www.biorxiv.org/content/early/2025/01/11/2025.01.08.631967},
elocation-id={2025.01.08.631967},
eprint={https://www.biorxiv.org/content/early/2025/01/11/2025.01.08.631967.full.pdf},
}
Protenix is built upon and inspired by several influential projects. If you use Protenix in your research, we also encourage citing the following foundational works where appropriate:
@article{abramson2024accurate,
title={Accurate structure prediction of biomolecular interactions with AlphaFold 3},
author={Abramson, Josh and Adler, Jonas and Dunger, Jack and Evans, Richard and Green, Tim and Pritzel, Alexander and Ronneberger, Olaf and Willmore, Lindsay and Ballard, Andrew J and Bambrick, Joshua and others},
journal={Nature},
volume={630},
number={8016},
pages={493--500},
year={2024},
publisher={Nature Publishing Group UK London}
}
@article{ahdritz2024openfold,
title={OpenFold: Retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization},
author={Ahdritz, Gustaf and Bouatta, Nazim and Floristean, Christina and Kadyan, Sachin and Xia, Qinghui and Gerecke, William and O’Donnell, Timothy J and Berenberg, Daniel and Fisk, Ian and Zanichelli, Niccol{\`o} and others},
journal={Nature Methods},
volume={21},
number={8},
pages={1514--1524},
year={2024},
publisher={Nature Publishing Group US New York}
}
@article{mirdita2022colabfold,
title={ColabFold: making protein folding accessible to all},
author={Mirdita, Milot and Sch{\"u}tze, Konstantin and Moriwaki, Yoshitaka and Heo, Lim and Ovchinnikov, Sergey and Steinegger, Martin},
journal={Nature methods},
volume={19},
number={6},
pages={679--682},
year={2022},
publisher={Nature Publishing Group US New York}
}
We welcome contributions from the community to help improve Protenix!
📄 Check out the Contributing Guide to get started.
✅ Code Quality:
We use pre-commit hooks to ensure consistency and code quality. Please install them before making commits:
pip install pre-commit
pre-commit install🐞 Found a bug or have a feature request? Open an issue.
The implementation of LayerNorm operators refers to both OneFlow and FastFold.
We also adopted several module implementations from OpenFold, except for LayerNorm, which is implemented independently.
We are committed to fostering a welcoming and inclusive environment. Please review our Code of Conduct for guidelines on how to participate respectfully.
If you discover a potential security issue in this project, or think you may have discovered a security issue, we ask that you notify Bytedance Security via our security center or vulnerability reporting email.
Please do not create a public GitHub issue.
The Protenix project including both code and model parameters is released under the Apache 2.0 License. It is free for both academic research and commercial use.
We welcome inquiries and collaboration opportunities for advanced applications of our model, such as developing new features, fine-tuning for specific use cases, and more. Please feel free to contact us at [email protected].
We're expanding the Protenix team at ByteDance Seed-AI for Science!
We’re looking for talented individuals in machine learning and computational biology/chemistry. Opportunities are available in both Beijing and Seattle, across internships, new grad roles, and experienced full-time positions.
“Computational Biology/Chemistry” covers structural biology, computational biology, computational chemistry, drug discovery, and more.
| Type | Expertise | Apply Link |
|---|---|---|
| Full-Time | Computational Biology / Chemistry | Experienced & New Grad |
| Full-Time | Machine Learning | Experienced & New Grad |
| Internship | Computational Biology / Chemistry | Internship |
| Internship | Machine Learning | Internship |
| Type | Expertise | Apply Link |
|---|---|---|
| Full-Time | Computational Biology / Chemistry | Experienced, New Grad |
| Full-Time | Machine Learning | Experienced, New Grad |
| Internship | Computational Biology / Chemistry | Internship (opening ~August) |
| Internship | Machine Learning | Internship (opening ~August) |
