Abstract
De novo protein design has the potential to expand the molecular toolkit available to engineers and scientists. Here we describe RFdiffusion, a generative model for protein backbone design that leverages the strengths of diffusion models while building in knowledge of protein structure through fine-tuning of the RoseTTAFold structure prediction network. RFdiffusion can generate diverse protein structures conditioned on a wide range of design specifications, including functional site scaffolding, symmetric oligomer design, and protein binder design.