RFdiffusion

What it is

Proteins are long chains of building blocks called amino acids, and the chain folds into a particular three-dimensional shape. That shape, or backbone, decides what the protein does. RFdiffusion builds new backbones using a method called diffusion. The model is trained by watching known protein structures get blurred into random noise, one small step at a time, and learning to undo each step. Once trained, it can be handed pure noise and asked to run the process backwards, nudging the jumble towards something that looks like a real protein. The result is a shape, not yet a sequence, so a second tool is normally used to work out which amino acids would fold that way.

The designer can also constrain the process. Part of a structure can be held fixed while the rest is generated around it, so the model fills in a scaffold that holds a given loop in place, or grows a new protein against the surface of an existing one. Because each run starts from fresh noise, the same request yields many different candidates. These are predictions, so they are usually passed through separate structure prediction software, scored, and only the most convincing few are made in the laboratory.

Why scientists use it

Natural proteins are the ones evolution happened to find, and they are a thin slice of what is chemically possible. Searching for new ones by hand, or by mutating existing ones, is slow and tends to stay close to where it started. A generative model can propose shapes nobody has seen, which matters when the aim is a binder that sticks to a chosen target, an enzyme pocket of a particular geometry, or a small protein that fits a receptor on a cell surface.

It also changes the economics. Laboratory work is the expensive part, so anything that narrows thousands of ideas down to a handful worth making saves time and materials. The same approach appears across quite different fields: medicine, where the target is a receptor or a virus; structural biology, where the question is how two molecules fit together; and industrial chemistry, where the goal is a catalyst that survives heat. In each case the model stands in for an expensive search.

In the record

Limits

The model proposes shapes; it does not know whether they will fold, stay stable, or do anything useful. Success rates in the laboratory can be low, and failures are rarely explained. It works best on compact, regular structures and struggles with floppy regions, with proteins that change shape as they work, and with anything needing a precise chemical environment rather than a precise shape. Designs are also judged mostly by other software, so two models can agree confidently and both be wrong. Experimental testing remains the only real check.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Model family

• Diffusion model

Every article