How to Use RFdiffusion3

RFdiffusion3 is an all-atom diffusion model for protein design that generates explicit side chains together with the backbone while reasoning about non-protein context such as ligands and nucleic acids. This makes it a strong fit for binder design, enzyme design, and other problems where atom-level interactions matter too much to be treated as a later refinement step.

On Neurosnap, researchers can start from an optional Input Structure, specify generated and fixed regions through a Contig, constrain chemistry with Fixed Atoms, and include a Ligand when the design task is pocket- or active-site-centered. The workflow is especially useful when a design must satisfy precise geometric or chemical constraints rather than only a broad fold objective.

The main methodological shift in RFdiffusion3 is that diffusion happens at the all-atom level. Side chains are generated together with the rest of the structure, allowing the model to reason directly about hydrogen-bonding geometry, solvent accessibility, atom placement, and nonprotein context. The paper also emphasizes a much leaner architecture than RFdiffusion2, with comparable or better design performance at substantially lower compute cost.

On Neurosnap, Contig and Fixed Atoms are the key design-specification tools. They let researchers decide which residues or atoms are immutable motifs, which regions should be inpainted, and whether the task should be interpreted as indexed scaffolding or Unindexed Motif guidepost conditioning. Length Min, Length Max, Timesteps, Step Scale, and Number Designs then determine how much sequence-structure space is explored.

Researchers usually interpret RFdiffusion3 as a campaign generator. Candidate sets are compared for motif preservation, atom-level plausibility, and target-context compatibility before any downstream sequencing, affinity screening, or experimental build decisions are made.

• Atomically detailed design: RFdiffusion3 reasons about side chains and nonprotein context during generation instead of postponing that chemistry to a later stage.
• Flexible structural specification: Contig, Fixed Atoms, ligand identity, and unindexed motif settings make the workflow adaptable to binders, enzymes, and scaffolding tasks.
• Efficient search: The model is designed to deliver all-atom design performance without the heavier cost profile of earlier atom-level methods.
• Research-pipeline fit: Neurosnap exposes the design controls that matter scientifically without requiring local diffusion-model setup.