RFdiffusion is a landmark open-source diffusion model for de novo protein design, developed by the Baker Lab and the RosettaCommons community. Hosted on GitHub with roughly 3,100 stars and still maintained in 2026, it brought image-diffusion techniques to protein backbone generation and has been used to design protein binders, symmetric assemblies and enzyme scaffolds that were validated experimentally. Few AI tools have had comparable impact in structural biology.
RFdiffusion fine-tunes the RoseTTAFold structure-prediction network into a generative denoising model: starting from noise, it iteratively sculpts a protein backbone that satisfies the design task, whether that is an unconditional monomer, a binder docked against a target protein, a symmetric oligomer, or a scaffold around a functional motif. The repository contains the inference code, model weights and worked examples for each design mode, along with guidance on noise scales and sampling settings for practical design campaigns.
RFdiffusion was developed at the Institute for Protein Design at the University of Washington and described in a 2023 Nature paper that reported experimentally validated designs, including high-affinity binders and symmetric assemblies confirmed by cryo-EM. In practice, design campaigns pair RFdiffusion backbones with a sequence-design step (commonly ProteinMPNN) and structure-prediction screening, a pipeline that has become a standard template in computational protein engineering groups worldwide. The repository’s examples walk through each major design mode with recommended settings.
RFdiffusion is free, open-source code published for the research community. Running it requires your own GPU compute; there is no hosted or paid version.
Computational biologists, protein engineers and drug-discovery researchers designing novel proteins, binders and assemblies in silico.