AI-designed peptide binders: where de novo design stands in 2026

Designing a peptide that binds a chosen protein used to mean screening huge libraries (phage or mRNA display) and optimising the hits. Generative models are changing that. Structure-aware diffusion models now propose binders that often work in the first round of testing. Here is where the field stands.

RFpeptides: cyclic peptides from a predicted structure

In Nature Chemical Biology (2025), Rettie, Baker, Bhardwaj and colleagues at the University of Washington's Institute for Protein Design described RFpeptides. It is a diffusion pipeline that builds macrocyclic peptide backbones against a target surface, then designs their sequences with ProteinMPNN. They tested 20 or fewer designs per target against four different proteins and obtained medium- to high-affinity binders for all four. For one target, the best macrocycle bound with a K_d below 10 nM, even though the design started from a predicted target structure rather than an experimental one. Crystal structures of the complexes matched the design models to within 1.5 Å (Cα RMSD).

Macrocycles matter because they combine some of the stability and target-surface coverage of larger proteins with the synthetic accessibility of peptides.

Scale and speed: Proteina-Complexa

NVIDIA's Proteina-Complexa, posted in March 2026 with Rosetta Commons co-authors, pairs a generative model with search at inference time. It outputs full-atom designs, including side chains. The developers report that it runs 30–60 times faster than RFdiffusion. In a benchmark of more than a million designs against 127 targets, run with Manifold Bio, it found binders for 86 targets (about 68%), with gains on flexible peptide targets. These are developer-reported figures from a preprint.

Automating the design loop

Preprints this year also describe AI agents running whole design campaigns: choosing tools, filtering designs and iterating. In one August 2026 preprint, an agent built on open-source design tools had 28 of 90 designs bind the target RBX1, against 9 of 245 in an open design competition. This is early and not yet peer-reviewed, but it points to design work becoming cheap and parallel.

The bottleneck moves to synthesis and validation

When computation can suggest dozens of credible binders in an afternoon, the slow and expensive steps are making and testing them:

• Synthesis: designed macrocycles often need non-canonical residues, specific cyclisation chemistry (head-to-tail, thioether, disulfide) and careful purification.
• Identity and purity: confirm every design by MS and HPLC before you test binding. A failed binder made from the wrong material teaches you nothing about the model.
• Binding: use orthogonal biophysics (SPR or BLI for kinetics, ITC for thermodynamics) and include controls such as scrambled sequences.
• Structure: crystallography or cryo-EM where it matters. Model confidence scores are not binding data.

The labs best placed to use these tools are those with a fast, well-characterised synthesis and assay pipeline behind them.

Sources

• Rettie S.A. et al. Accurate de novo design of high-affinity protein-binding macrocycles using deep learning. Nature Chemical Biology (2025). doi:10.1038/s41589-025-01929-w (open access)
• Rosetta Commons: Next generation generative model unlocks de novo designs at scale (April 2026). rosettacommons.org
• Agent-driven de novo protein binder design (preprint, August 2026). alphaXiv
• Peptide-based drug design using generative AI. Chemical Communications (2026). RSC Publishing

Photo: Google DeepMind / Pexels