RFdiffusion designs proteins that do not exist in nature

Watson, Juergens and colleagues at the Baker Lab published RFdiffusion in Nature on 11 July 2023. By fine-tuning the RoseTTAFold prediction network on denoising tasks, they turned a model that reads protein structures into one that invents them. Hundreds of the designs were then made in the laboratory.

Background

AlphaFold answers one question: given this sequence of amino acids, what shape does it fold into. That was the fifty-year problem and it is now largely solved.

The reverse question is harder and more useful. Given a job you want done, what protein would do it, and no protein like that having evolved. Deep learning had been applied to this before, and diffusion models in particular had worked poorly on proteins, which the authors attribute to the awkwardness of backbone geometry.

What happened

RFdiffusion was published in Nature on 11 July 2023 by Watson, Juergens and colleagues at the Baker Lab.

They took RoseTTAFold, a network trained to predict structures, and fine-tuned it on denoising tasks until it became a generative model instead. Starting from random coordinates, it removes noise over fifty to two hundred steps until a plausible protein backbone remains, and it can be steered: hold a functional site fixed and design a scaffold around it, or aim a binder at a chosen target.

The same approach handled several problems that had needed separate methods: designing monomers to a specified topology, designing binders, building symmetric assemblies, and scaffolding enzyme active sites.

What followed

Hundreds of the designs were made and measured, which is the part that matters and the part most computational biology skips. A cryo-electron microscopy structure of one designed binder attached to influenza haemagglutinin came out nearly identical to the model that proposed it.

David Baker shared the 2024 Nobel Prize in Chemistry for computational protein design, the line of work this belongs to, awarded alongside the prize for structure prediction. The two halves of that prize are the two directions of the same question.

Why it mattered

AlphaFold answers what shape a given sequence folds into. This asks the opposite question and produces a protein for a job that evolution never needed doing. The designs were then built and measured, which is the part that separates this from most computational biology: the claim was tested in glassware rather than only in a benchmark.

Sources

• De novo design of protein structure and function with RFdiffusion. nature.com. Primary source
• RFdiffusion: A generative model for protein design. bakerlab.org. Official

Cite this page

AI Achievements. (2023). RFdiffusion designs proteins that do not exist in nature. Retrieved 2026-08-29, from https://achievements.ai/milestone/rfdiffusion-de-novo-protein-design

@misc{achievements_rfdiffusion_de_novo_protein_design,
  title  = {RFdiffusion designs proteins that do not exist in nature},
  author = {{AI Achievements}},
  year   = {2023},
  url    = {https://achievements.ai/milestone/rfdiffusion-de-novo-protein-design}
}