Deep learning designs ring-shaped peptides that bind their targets as drawn

Why rings

Cyclic peptides, or macrocycles, sit between small-molecule pills and antibodies. They are big enough to grip flat protein surfaces that small molecules cannot hold, and the ring shape makes them stiffer and harder to digest than straight chains. Several are approved drugs. Until recently, finding one meant screening huge libraries.

What the team did

The group adapted RFdiffusion, the tool the Institute for Protein Design uses to invent proteins, so that it could draw ring-shaped backbones around a chosen spot on a target protein. A second tool picked amino acid sequences to fill those backbones.

They designed binders for several targets, including proteins involved in cancer and cell recycling, and one bacterial protein for which they had no prior structure of a bound partner. For each target they made a small number of designs, about 20 or fewer, rather than millions. Some bound strongly, and X-ray structures of the bound pairs were very close to the computer models.

What it changes

The striking part is the hit rate and the accuracy. Getting a binder from a few dozen designs, and seeing it sit where the model said it would, makes rational design of ring peptides look practical.

The work stops at binding. A medicine also has to get into the right tissue, last in the body and avoid hitting other proteins. Those problems are not solved by better design software alone.

Source Rettie S. A. et al., Nature Chemical Biology, 2025 (RFpeptides)

Amino Weekly reports research. It is not medical advice and never gives doses.