De novo design of protein structure and function with RFdiffusion

“De novo design of protein structure and function with RFdiffusion,” by Joseph Watson, David Baker, and colleagues, was published in Nature on July 11, 2023. Where AlphaFold predicts the shape of a protein that already exists, RFdiffusion runs the problem backward: it designs entirely new protein structures to meet a specification.

RFdiffusion was built by fine-tuning RoseTTAFold, the Baker lab’s structure-prediction network, on a denoising task, turning it into a generative diffusion model of the same kind that powers AI image generation. Starting from noise, it iteratively refines a protein backbone toward a valid, designable structure.

The paper demonstrates strong results across many design tasks: unconditional generation of new monomers up to 600 residues, symmetric oligomers, scaffolding of functional motifs and enzyme active sites, and design of proteins that bind a chosen target. The authors validated designs experimentally, including a binder whose cryo-EM structure matched the design model to sub-Angstrom accuracy. Earlier methods often required testing tens of thousands of candidates to find one that worked; RFdiffusion produced far higher success rates.

David Baker shared the 2024 Nobel Prize in Chemistry for computational protein design, the field that RFdiffusion advanced; the other half went to DeepMind’s Demis Hassabis and John Jumper for structure prediction.