PXDesign shows up in searches as the newest thing, an AlphaFold3-class binder designer from ByteDance, which invites an obvious question: is it the one I should be using? Having run it on real campaigns, the honest answer is that the question is malformed. No single generator wins every target. PXDesign is a strong model, and its real value is not that it replaces the others, it is what a different structural prior adds to a pool you are going to screen anyway.
This post explains what PXDesign actually is, then where it sits in a de novo binder toolkit.
What PXDesign actually is
PXDesign is ByteDance’s open-source de novo binder design tool, released under an Apache 2.0 license. It is built on Protenix, ByteDance’s AlphaFold3-class structure model, and that is where the name comes from: PX for Protenix. The pipeline generates backbones with the Protenix model, designs sequences with ProteinMPNN, and ranks the candidates with AlphaFold2 Initial Guess (Bennett et al., 2023), a confidence filter that re-predicts each binder against the target.
The thing that distinguishes it from the RFdiffusion lineage is the prior. RFdiffusion is a diffusion model trained directly on protein structures. PXDesign generates from an AlphaFold3-class model. Those are two different learned views of what a plausible protein complex looks like, so they reach for different backbones on the same target.
We do not single it out
The temptation with any new AF3-class model is to treat it as the upgrade, the one tool that makes the others obsolete. That is not how it behaves in practice.
It is the same principle that governs running multiple parameter strategies within one model, only applied across models. Each generator’s prior biases it toward certain folds. Run one and your screening pool is all variations of that model’s preferences. Add a generator with a genuinely different prior and you cover more of the design space that is actually accessible, at the cost of a few extra runs.
The generator toolkit
For general de novo binders against a structurally defined target, four open generators cover most of the ground. They are complementary, not ranked.
- RFdiffusion. Broad backbone diversity by denoising, sequence added by ProteinMPNN. The workhorse for discovery and volume. See BindCraft vs RFdiffusion for the detailed comparison and when each wins.
- BindCraft. Narrower sampling with scoring built into generation, so each output is a pre-filtered backbone and sequence pair. Higher quality per design, fewer designs per GPU-hour.
- BoltzGen. An all-atom diffusion generator that chooses backbone and sequence together, reasoning over side-chain atoms during design rather than deferring sequence to a later step.
- PXDesign. An AlphaFold3-class prior via Protenix, ranked with AF2 Initial Guess. The different-prior addition to the pool.
Antibody formats sit on a separate axis. If the deliverable has to be a nanobody or an scFv, the generator is different, see scFv vs nanobody: which fragment, which tool.
Why a different prior is the point
The argument for adding PXDesign is not that it is more accurate than RFdiffusion. It is that its errors and its preferences are different. Two models with different priors fail on different targets and succeed on different targets. When you screen the union of their outputs, you get binders neither would have surfaced alone.
This is why the decision is not “which one generator.” On a genuinely hard target, the highest-value move is to run generators that see the problem differently, then let the experiment adjudicate. PXDesign is a good addition to that set precisely because its AlphaFold3-class view is not a variation on the diffusion theme, it is a different route into the same design space.
The trap: chasing the newest single model
Two mistakes follow from the “which one” framing.
The first is treating the newest AF3-class generator as the answer and dropping the others. That narrows your pool to one model’s preferences and throws away the diversity that actually drives hit rate. The second is over-trusting the ranking. AF2 Initial Guess is a confidence filter, and a useful one, but a high score is a prediction that the binder folds and docks as intended, not a measurement that it binds. Read it as a filter that removes obvious failures, not as a guarantee. Either way, the binder earns the name only when a display screen measures it.
Quick reference
| Generator | Structural prior | Sequence step | Ranking | What it adds to the pool |
|---|---|---|---|---|
| RFdiffusion | structure-trained diffusion | ProteinMPNN | AF2 confidence | broad backbone diversity, volume |
| BindCraft | hallucination with built-in scoring | integrated | during generation | pre-filtered, high quality per design |
| BoltzGen | all-atom diffusion | joint with backbone | refolding self-consistency | side-chain-aware generation |
| PXDesign | AlphaFold3-class (Protenix) | ProteinMPNN | AF2 Initial Guess | a different learned prior |
Practical recommendations
- Do not pick one generator, pick a set. The hit rate comes from screening the union of models that see the target differently.
- Add PXDesign for its prior, not as a replacement. Its AlphaFold3-class view reaches backbones the diffusion lineage does not.
- Read AF2 Initial Guess as a filter, not a verdict. It removes weak designs; it does not promise binding.
- Let the screen decide. Every generator proposes. Yeast display measures which proposals are real.
Run PXDesign self-serve on tools.ranomics.com against a target structure, or run it alongside RFdiffusion and BindCraft and screen the union. For the detailed head-to-head on the two most-used generators, see the BindCraft vs RFdiffusion guide.
Summary
PXDesign is ByteDance’s open AlphaFold3-class binder generator, built on Protenix and ranked with AF2 Initial Guess. It is a strong model, but the useful question is not whether it beats RFdiffusion or BindCraft. It is what a different structural prior adds to a screening pool. Run several generators that see the target differently, add PXDesign for its AlphaFold3-class view, treat the ranking as a filter rather than a verdict, and let the wet-lab screen pick the winners.
Related Ranomics services
- BindCraft vs RFdiffusion: the detailed head-to-head on the two most-used generators, and how we chain them.
- PXDesign technology page: how the AlphaFold3-class Protenix generator runs on the Ranomics platform.
- Binder Pilot: take the union of your generator outputs to a ranked, wet-lab-validated hit list on a fixed-scope campaign.