RFdiffusion3 Designs Proteins Atom by Atom — and the Results Work in the Lab

October 19 — The detailed design of RFdiffusion3, a generative AI system that can design proteins and nucleic acids at the level of individual atoms, has now been published. Proteins newly designed with RFdiffusion3 can perform their intended functions in real chemical reactions.

An international team from the University of Washington, the University of Oxford and Nvidia built the system, which models individual atoms and designs proteins and nucleic acids, and published the work in the journal Science on the 16th local time.

The team open-sourced RFdiffusion3 on 3 December 2025, and the paper in Science explains in detail the design principles behind the system.

AI-based protein research has moved on from predicting the three-dimensional structures of existing proteins to designing entirely new proteins with a target structure and function.

David Baker, a professor at the University of Washington, is the corresponding author of the RFdiffusion3 study. He won the 2024 Nobel Prize in Chemistry for his work on using computers to predict protein structures. One of Baker’s signature achievements is RosettaFold, an AI system for predicting protein structures.

RosettaFold learns the spatial relationships between the amino acids that make up proteins and uses that information to predict a protein’s three-dimensional structure. RFdiffusion1, released in 2023, and RFdiffusion2, released in 2025, were both built on the structure-prediction network that RosettaFold uses.

RFdiffusion uses diffusion models to determine where amino acids go, and in that way completes a protein design. Diffusion models are a class of AI model that learn the distribution of existing data and then generate new data.

RFdiffusion3 was developed with reference to the network architecture of another protein-structure prediction program, AlphaFold3. RFdiffusion3 can describe a single amino acid with up to 14 atoms, and designs entirely new protein structures by directly adjusting each atom’s three-dimensional coordinates.

With the system, the team achieved precise design of enzyme active sites; enzymes catalyse all kinds of chemical reactions. Enzymes are proteins — biological catalysts that speed up the rates of chemical reactions in living organisms.

When the researchers specified not only the positions of the atoms a reaction needs but also the conditions for hydrogen bonding and the range within which particular atoms must sit inside the protein, RFdiffusion3 could generate protein structures that satisfy all the constraints.

The researchers compared designs against 41 active sites from existing enzymes, and RFdiffusion3 outperformed RFdiffusion2 in 37 of the tests. It also generated protein structures about ten times faster than RFdiffusion2.

A paraoxon-degrading enzyme and a peptide-cleaving enzyme designed by RFdiffusion3 both completed the expected chemical reactions in real experiments. “In future, RFdiffusion3 will be widely used in all kinds of biomolecular design tasks and related applications,” the team said.