PXDesign

PXDesign is an open-source software suite for the de novo design of protein binders, released by the Protenix Team within ByteDance's Seed organization and posted to bioRxiv in December 2025. The system addresses one of the central challenges in protein engineering: generating novel mini-protein binders that bind a chosen target with high affinity, starting only from the target structure and with no natural binder to use as a template. Rather than relying on a single generative trick, PXDesign packages a full, modular pipeline — backbone generation, sequence design, and confidence-based filtering — into a reproducible framework.

PXDesign is closely related to, but distinct from, ByteDance's Protenix model. Protenix is an open-source reproduction of AlphaFold 3 used for biomolecular structure prediction. PXDesign is a design suite that consumes structure-prediction models — including Protenix and AlphaFold2 in interface-guided mode (AF2-IG) — as scoring and filtering engines to rank candidate designs. In other words, Protenix is one of the components PXDesign relies on, not a substitute for it.

The work is notable for pairing strong in-silico benchmarks with extensive wet-lab validation across multiple diverse targets, and for shipping under a permissive Apache 2.0 license alongside a free public web server, lowering the barrier for academic and industrial groups to run binder-design campaigns.

#Key Features

• Two complementary generators: PXDesign-d uses a diffusion-based backbone generator, while PXDesign-h uses a hallucination-based approach; the diffusion path offers higher throughput and broader fold diversity for large campaigns, while hallucination provides an alternative strategy for difficult targets.
• Confidence-based filtering and ranking: Candidate designs are scored and filtered using confidence metrics from multiple structure predictors, including Protenix and AF2-IG, which the authors systematically analyze to improve hit rates.
• Modular pipeline: Backbone generation, ProteinMPNN-based sequence design, and structure-prediction filtering are composed into a single reproducible workflow rather than a monolithic black box.
• Strong experimental hit rates: Reported nanomolar hit rates (dissociation constant below 1000 nM) of 17–82% across six of seven diverse protein targets, surpassing prior methods such as AlphaProteo.
• Open access: Released under Apache 2.0 for both academic and commercial use, with a free public web server and an accompanying benchmark suite, PXDesignBench.

#Technical Details

PXDesign-d is a diffusion model that generates protein-binder backbones conditioned on the target structure; PXDesign-h instead hallucinates binders by optimizing inputs against a structure predictor's confidence signal. Generated backbones are passed to ProteinMPNN for sequence design, and the resulting designs are filtered and ranked using confidence estimates from Protenix and AF2-IG. The authors report that PXDesign-d achieves higher success rates and broader fold diversity than RFdiffusion across ten targets in silico, and is more throughput-efficient than hallucination for large-scale campaigns. Experimentally, PXDesign produced nanomolar binders with hit rates of 17–82% on six of seven targets drawn from a diverse panel including IL-7RA, PD-L1, VEGF-A, the SARS-CoV-2 receptor-binding domain, TrkA, and EGFR. Model checkpoints (the diffusion generator plus Protenix base and mini variants) download automatically on first run, and external tool weights such as AF2 and ProteinMPNN are fetched via a provided script.

#Applications

PXDesign targets researchers in therapeutic antibody alternatives, biologics discovery, and synthetic biology who need to generate high-affinity binders against a structurally defined target without an existing template. Typical use cases include designing mini-protein binders to receptors and viral antigens for diagnostics, as research reagents, or as starting points for therapeutic development. Because the suite is modular and open-source, computational groups can swap in alternative generators or filters, while experimentalists with no machine-learning infrastructure can use the free web server to produce ranked candidate sequences for direct wet-lab testing.

#Impact

By coupling competitive in-silico performance with experimentally validated nanomolar hit rates and an unrestricted Apache 2.0 license, PXDesign contributes a practical, reproducible alternative to closed or restrictively licensed binder-design pipelines. Its public web server, launched in September 2025, has supported numerous researchers in identifying binders for wet-lab validation, and the companion PXDesignBench framework offers a shared yardstick for comparing future methods. Key limitations follow from its dependence on structure-prediction confidence as a proxy for binding: hit rates vary substantially by target, success is not guaranteed for the most difficult interfaces, and as a preprint-stage release some benchmark claims await peer review and broader independent reproduction.