RFdiffusion is an open-source machine learning tool for generating protein backbone structures from minimal input. The designed backbones – or the regions of the output structures designed by RFdiffusion – are entirely composed of glycine (backbone only) residues. Tools, such as ProteinMPNN, can then be used to generate a sequence based on the fold.

The potential uses of RFdiffusion range from protein monomer backbone design to scaffolding enzyme active sites. The success of the tool across these disparate use cases hinges upon its use of denoising diffusion probabilistic models (DDPMs), often referred to as diffusion models. Diffusion models are used in image and music generation tools due to their ability to generate highly diverse outputs. In the case of RFdiffusion, this means that the tool generates a variety of possible protein backbone structures that still resemble the training data. Diffusion models can also incorporate conditioning information, which is used here to ensure the backbone structure incorporates a specific motif, can act as scaffold for another macromolecule, design potential binders, and design around another target protein. For an example of how RFdiffusion has been used to advance science, see how its conditional denoising process was used to design proteins that neutralize lethal snake venom toxins. 

The process of generating a backbone structure using RFdiffusion is referred to as an inference run. The inference run uses a pre-trained model to predict the protein backbone structure based on the set of inputs and constraints from the user without having to further train the model. For a list of possible inputs and constraints see <insert link to documentation here>. RFdiffusion has several different models to choose from depending on the inference task being performed, see <insert link to different use cases> to learn more. 

See the Ready to Get Started? section below for the different ways to run, download, and/or install RFdiffusion. When you're ready to try running something, we recommend beginning with our Unconditional Monomer Tutorial.