How to Use RFdiffusion

RFdiffusion is the original broadly applicable protein diffusion model for generating protein backbones with or without structural conditioning. The method changed protein design by making backbone generation flexible enough to cover unconditional folds, motif scaffolding, binder design, partial diffusion around an existing structure, and symmetry-aware oligomer design inside one general framework.

On Neurosnap, the same service can be used for several design modes. An Input Structure becomes important for binder design, scaffolding, or partial diffusion, while Binder Input Chain, hotspot residues, motif ranges, symmetry settings, and cyclic-chain options determine which design regime is being explored.

RFdiffusion uses denoising diffusion over protein backbone coordinates to sample structures consistent with a chosen set of constraints. The core paper showed that this strategy works across many protein-design tasks because the model can interpolate between purely generative sampling and tightly conditioned design around motifs, targets, or symmetry. That breadth is what distinguishes RFdiffusion from narrower single-purpose backbone generators.

On Neurosnap, Timesteps and Use Beta Model control sampling behavior, while Backbone Only decides whether the run stops at structural generation or continues into downstream sequencing and folding checks. Binder-design controls such as Binding Pocket Residue Start, Binding Pocket Residue End, Binder Length Minimum, Binder Length Maximum, and Hotspots are useful when the target is large and a more focused site is needed. The platform also exposes motif scaffolding, partial diffusion, unconditional generation, symmetry, custom contigs, and optional guiding potentials for more advanced campaigns.

Researchers typically interpret RFdiffusion as a design-campaign engine. The useful question is not whether one structure looks good, but whether the generated family contains backbones that preserve the intended motif, place interfaces plausibly, and justify the next stage of ProteinMPNN sequencing, AlphaFold validation, or experimental testing.

• One framework for many design problems: RFdiffusion covers unconditional generation, motif scaffolding, binder design, symmetry, and partial diffusion in one service.
• Scientifically meaningful controls: Target cropping, hotspots, motif ranges, symmetry, cyclic chains, and custom contigs let researchers tailor the design regime instead of accepting one generic protocol.
• Backbone-first campaign design: The model is especially useful for exploring structural hypotheses before committing to sequence and validation steps.
• No-code access to an advanced design stack: Neurosnap exposes a difficult diffusion workflow without requiring local hydra configs, model weights, or custom post-processing.