ByteDance’s PXDesign suite achieves nanomolar hit rates of 17–82% across six drug-relevant targets, outperforming leading methods through smarter generation and multi-predictor filtering

The design of protein binders — proteins engineered from first principles to bind a target with high affinity and selectivity — is one of the defining challenges of modern structural biology and drug discovery. For decades, the field depended on immunization, directed evolution, or display-based selection to identify candidates from large libraries, with limited control over which epitope was engaged. Deep learning changed that calculus substantially. With the arrival of diffusion-based generative models like RFdiffusion in 2023 and hallucination-based pipelines such as BindCraft shortly after, it became possible to propose binder sequences computationally, short-cutting rounds of experimental optimization that previously took months. Yet despite these advances, head-to-head comparisons under consistent evaluation conditions remained scarce, and critical questions around the accuracy of confidence-based filtering — the computational step that converts thousands of candidate structures into a workable experimental shortlist — were largely unresolved.

That gap is addressed directly by the team behind Protenix, ByteDance Seed’s open-source AlphaFold3 reproduction. In August 2025, the same group released PXDesign, a modular suite integrating two complementary generative strategies — diffusion and hallucination — with a rigorous multi-predictor filtering framework. The system was validated in vitro across seven pharmacologically relevant targets, including PD-L1, VEGF-A, SARS-CoV-2 RBD, IL-7RA, TrkA, EGFR, and TNF-α, benchmarked directly against AlphaProteo and RFdiffusion under matched conditions.

“PXDesign achieves nanomolar binder hit rates of 17–82% across six of seven diverse protein targets, surpassing prior methods such as AlphaProteo by 2–6×.”

The PXDesign pipeline operates in three stages. First, PXDesign-d, a Diffusion Transformer backbone trained to generate Cartesian atom coordinates directly rather than SE(3)-equivariant backbone frames, proposes diverse target-conditioned binder structures ranging from 60 to 160 amino acids, with sequences then assigned by ProteinMPNN. Critically, PXDesign-d samples a broader region of fold space than its predecessors — producing binders that are not overwhelmingly α-helical, a well-documented compositional bias in earlier diffusion models. Filtering then applies interface confidence metrics, specifically ipTM and pTM, from both Protenix and AF2-IG predictors in combination, since a key experimental finding is that the two predictors capture largely non-overlapping populations of true binders. Ensemble filtering, using both models, is materially better than either alone.

The experimental results make a strong case for this integrated approach. Against PD-L1, 14 of 17 tested designs bound with KD below 1 μM, and most came in well below 100 nM — compared to the strongest BindCraft binder against that same target, which reached approximately 1 μM. On the SARS-CoV-2 RBD, the hit rate rose from 20% in an earlier version of the pipeline to 50% following generation improvements, with multiple binders confirmed at sub-10 nM affinity by biolayer interferometry. TNF-α remains an outstanding challenge — a notoriously flat and symmetric interface that stymied AlphaProteo as well — which honestly makes the performance elsewhere more impressive rather than less. The full pipeline is freely accessible via a public web server and an open-source GitHub repository, removing the compute and engineering barrier for groups wanting to run their own binder campaigns. As structure prediction and generative design continue to converge, PXDesign offers a clear signal that the computational-to-experimental pipeline for protein binder discovery is compressing from months to days.