PXDesign: Fast, Modular, and Accurate
De Novo Design of Protein Binders

PXDesign is a model suite for de novo protein-binder design — a diffusion generator (PXDesign-d) paired with Protenix and AF2-IG confidence models for selection. Across six targets, PXDesign delivers 20-73% nanomolar hits on five and 2-6× gains over strong baselines such as AlphaProteo.

Protenix Team, ByteDance Seed

Try PXDesign Web Server Technical Report biorxiv PXDesignBench on GitHub

Highlights at a glance

High wet-lab success: 20-73% hit rates (K_D < 1000 nM) on diverse targets including IL-7RA, PD-L1, VEGF-A, SC2RBD, TrkA; TNF-α remains challenging.

Model-driven generation & selection: PXDesign-d boosts pass rate & diversity; Protenix-based filters strongly enrich true binders and complement AF2-IG; combining predictors further improves enrichment.

Ready to use: Open PXDesignBench and a public web server for the community.

PXDesign Server Walkthrough

How PXDesign works

1) Generation

PXDesign-d (Diffusion) is a DiT-style model for backbone generation — proposes many target-conditioned binder backbones; sequences are then assigned with ProteinMPNN.

2) Prediction & filtering

Complex structures are predicted with Protenix. Candidates are filtered by confidence scores (e.g., ipTM) and structure consistency using Protenix and AF2-IG.

3) Selection for wet-lab

Passing designs are clustered by structural similarity (Foldseek) to preserve diversity; within each cluster, higher-confidence designs (Protenix ipTM) are prioritized for expression and BLI affinity assays.

Why PXDesign works

1) Filtering & ranking power

Protenix-based filtering strongly enriches and prioritizes true binders; together with AF2-IG, it captures complementary true positives, and using both is likely to yield stronger enrichment.

2) In-silico success rate

PXDesign-d attains higher success rates and broader fold diversity than RFDiffusion on 10 targets; diffusion is also more throughput-efficient than hallucination for large campaigns.

3) Wet-lab validation

PXDesign achieves high nanomolar hit rates, leading or matching the best on multiple targets.

Download our designs (.zip)

Method	IL7RA	PD-L1	VEGF-A	SC2RBD	TrkA	TNF-α
PXDesign	–	72.7	47.1	50.0	20.0	0.0
PXDesign (v0)	40.0	–	–	22.2	–	–
RFDiffusion	17.0	13.0	–	–	0.0	–
AlphaProteo	25.0	11.4	21.3	9.3	4.5	0.0
Chai-2	80.0	70.0	–	–	–	21.0
Chai-1d	10.0	5.0	–	–	–	–
Latent-X	16.0	33.0	–	44.0	–	–

Experimental hit rates (%) for designed binders. Bold = per-target best; underline = per-target second best. “–” = not tested.

Per-target experimental success rates across methods.

Representative PXDesign-designed nanomolar binders.

Beyond protein binders

While our in-silico and wet-lab validation focuses on protein binders, PXDesign-d is naturally extensible to diverse molecular targets (e.g., nucleic acids, small molecules, post-translationally modified proteins). We provide cross-modality demonstrations and a preliminary benchmark on cyclic-peptide binders.
Cyclic-peptide generation is available on our public server as an experimental feature.

Ship Notes & Key Takeaways

What’s released

Benchmark & tooling:

PXDesignBench Github repository (filtering/benchmark framework).

Hosted access:

PXDesign web server for binder and cyclic-peptide design.

Paper assets:

Full technical report with protocols, thresholds, and methodology details.

Key Takeaways

Strong hits at scale:

Diffusion-based generation + orthogonal filtering yields high hit rates and diverse binders.

Understand structure predictors:

AF2-IG and Protenix cover different regions of the binder space, and combining them strengthens robustness.

Ready to try:

Open benchmarks + a public server make it easy to reproduce and build upon the pipeline.

PXDesign: Fast, Modular, and AccurateDe Novo Design of Protein Binders