AI-Designed Peptides: How RFpeptides Rewrote the First Step of Drug Discovery

A neural network from a Nobel laureate’s lab now designs ring-shaped peptides that grip their targets like finished drugs. The lottery of discovery just got shorter.

For most of pharmaceutical history, finding a peptide that binds a chosen protein meant screening enormous libraries and hoping something stuck. The odds were long, the timelines were measured in years, and the bill ran into the millions. A June 2025 paper in Nature Chemical Biology offers a route that looks nothing like that. It describes RFpeptides, a deep-learning system that designs AI-designed peptides from a target’s structure alone — no library, no luck.

The work comes from the Institute for Protein Design in Seattle, the lab of David Baker, who shared the 2024 Nobel Prize in Chemistry for computational protein design. That prize was the field’s coming-of-age. RFpeptides is what the technology does next.

What RFpeptides actually does

RFpeptides specializes in macrocycles — peptides bent into a closed ring. The architecture matters. A ring resists the enzymes that shred ordinary peptides, and it settles into shallow protein pockets that defeat conventional small-molecule drugs. Chemists have prized macrocycles for decades and struggled to design them on purpose.

The Seattle team fed their model a target protein and asked it to invent macrocyclic binders to fit. Then they tested the output. Against each of four different proteins, fewer than twenty designs yielded working binders. Against one target, a protein called RbtA, they built a molecule with sub-10-nanomolar affinity — pharmaceutical-grade grip — starting only from a predicted structure. No crystallography. No screen. A shape guessed by software, refined by math, made real at the bench.

Why a plate of twenty beats a library of millions

Sit with the arithmetic. The old approach spent months and fortunes on the front end of discovery, sifting vast collections for a single hit. The new approach spends an afternoon of computation and a plate of twenty candidates. A hit rate that high, from that few designs, changes what counts as a viable target.

The practical payoff is reached when the search for a binder stops being a gamble; molecules that were never worth chasing become worth a week’s work. Difficult targets — the flat, featureless protein surfaces that small molecules slide off — come back into play.

RFpeptides did not appear from nowhere. It extends RFdiffusion, the open-source design engine the Baker lab released in 2023, and the new macrocycle capability has been folded into that toolkit. Open release is the multiplier. Labs from Boston to Basel can run this on their own targets today, which means the pace of follow-on work will outrun any single group.

The honest limits

None of this is an approved therapy, and the paper does not pretend otherwise. A high-affinity binder is a starting point, not a medicine; it still needs to survive the body, reach its target, and prove safe and useful in people. That road remains long.

What RFpeptides delivers is a faster, cheaper, more reliable beginning — and beginnings are where most drug programs die. Compress the odds at the front, and more programs live long enough to be tested where it counts.

What it means for peptide medicine

The obesity and diabetes drugs dominating headlines were all discovered the slow way, then engineered by hand over years. Tools like RFpeptides suggest the next generation may be conceived in silico and validated in weeks. Design collapsing from years to days is not a marketing line here; it is the measured result in a peer-reviewed journal, reproduced across four targets.

For the enthusiast tracking where peptide science goes next, this is the story under the story. The blockbuster readouts get the coverage. The design tools decide what gets discovered at all.

Frequently asked questions

What are RFpeptides? A deep-learning system from the Institute for Protein Design that designs macrocyclic (ring-shaped) peptide binders directly from a target protein’s structure — no screening library required. It was described in Nature Chemical Biology in June 2025 and folded into the open-source RFdiffusion toolkit.

Who is David Baker? A biochemist at the University of Washington whose lab pioneered computational protein design. He shared the 2024 Nobel Prize in Chemistry for the field, alongside DeepMind’s Demis Hassabis and John Jumper.

Are AI-designed peptides used in real drugs yet? Not yet as approved medicines. RFpeptides produces high-affinity binders — strong starting points — but a binder still has to survive the body and prove safe and effective in people. The advance is a faster, cheaper beginning, not a finished drug.

Sources

1. Rettie, S.A., et al. “Accurate de novo design of high-affinity protein-binding macrocycles using deep learning.” Nature Chemical Biology, June 20, 2025. DOI: 10.1038/s41589-025-01929-w.
2. Institute for Protein Design, University of Washington — research announcements on RFpeptides and RFdiffusion.
3. Watson, J.L., et al. “De novo design of protein structure and function with RFdiffusion.” Nature, 2023.
4. The Royal Swedish Academy of Sciences — Nobel Prize in Chemistry 2024 (David Baker, Demis Hassabis, John Jumper), nobelprize.org.
5. Baker Lab, University of Washington — bakerlab.org.