PXDesign

Pre-trained model for protein backbone generation and binder evaluation through Preview and Extended workflows.

Disclaimer

This is an UNOFFICIAL implementation of PXDesign by ByteDance Research.

The OFFICIAL repository of PXDesign is at bytedance/PXDesign.

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

Model Details

PXDesign-d is an all-atom diffusion model for de novo and target-conditioned protein backbone generation. The complete workflow assigns a ProteinMPNN sequence to each backbone, evaluates the complex and unbound binder with AlphaFold 2, and filters and ranks the candidates.

Extended evaluation additionally uses Protenix base or the structure-conditioned Protenix mini model, following the released PXDesignBench composition. Please refer to the Training Details section for more information on the training process.

Variants

PXDesign backbone: Independently callable PXDesign-d coordinate generation through PxDesignBackbonePipeline.
PXDesign Preview: The complete workflow with 100 backbones by default, ProteinMPNN assignment, AlphaFold 2 evaluation and released selection rules.
PXDesign Extended: The complete workflow with 500 backbones by default, full-target Protenix routing, two-step screening and 20-step reruns for candidates passing basic screening.

Model Specification

Single Size	Pair Size	Token Denoiser Blocks	Atom Encoder Blocks	Atom Decoder Blocks	Num Parameters (M)
384	128	16	4	4	139.028

The parameter count is for PXDesign-d.

Usage

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

pip install ofoldx

Direct Use

Preview Design

You can use this model directly with a pipeline for 100-residue binder design with Preview evaluation:

import ofoldx
from ofoldx.io import read_structure
from ofoldx.models.pxdesign import from_binder

structure, _ = read_structure("target.cif")
pipeline = ofoldx.pipeline("design", model="oteam/pxdesign")
result = pipeline(structure, placement=from_binder(100, [("target", [(0, 150)])]), seed=7)

The result retains all candidates and ranked selection views. from_binder names the target instances and zero-based component spans. The complete repository contains six child pipelines; Preview uses the backbone generator, ProteinMPNN and two AlphaFold 2 evaluations.

Extended Design

You can use this model directly with a pipeline for binder design with Extended evaluation and full-target MSA evidence:

import ofoldx
from ofoldx.io import read_structure
from ofoldx.models.pxdesign import from_binder
from ofoldx.msa import read_msa

structure, _ = read_structure("target.cif")
msa = {"target": read_msa("target.a3m", entity_id="1")}
pipeline = ofoldx.pipeline("design", model="oteam/pxdesign")
result = pipeline(
    structure,
    placement=from_binder(100, [("target", [(0, 150)])]),
    msa=msa,
    mode="extended",
    seed=7,
)
best = result.ranked_candidates[0]

Here target is the target instance ID and 1 is its entity ID; use the IDs in your input structure.

Extended additionally uses both Protenix children and reruns every candidate passing basic screening, including strict passes, for 20 diffusion transitions. Returned structures come from those reruns.

Both modes retain admitted target heavy atoms, including Se/SE, and preserve MSE rather than substituting MET as in the upstream input convention. This can change the checkpoint's feature distribution; observation masks control usable geometry without changing chemical membership.

Downstream Use

Backbone Generation

The backbone component is independently callable through PxDesignBackbonePipeline and returns structures in generation order without sequence assignment or evaluation. The pipeline guide describes component loading and staged calls; PxDesignPipeline documents the six component roles and the chemical matching used in evaluation.

Training Details

PXDesign-d learns protein backbone generation with an EDM coordinate-denoising objective.

Training Data

PXDesign-d is trained on PDB structures up to 1 May 2021 and predicted structures from AlphaFold DB and MGnify.

Training Procedure

PXDesign-d is trained in two stages, first emphasizing monomer structures and then target-conditioned complexes.

OFoldX exposes the differentiable one-step denoiser for training and fine-tuning; it does not reproduce PXDesign-d's full training pipeline.

Citation

Please cite the PXDesign work and repository described by the upstream project.

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 PXDesign project 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

oteam
/

pxdesign