How to Use RFdiffusion2

RFdiffusion2 is an enzyme-design model built for scaffolding catalytic sites from atom-level functional descriptions rather than from sequence motifs alone. The method is especially relevant for enzyme engineering because it can reason about active-site atoms, ligand chemistry, and backbone generation together instead of asking users to enumerate catalytic residue placements separately.

On Neurosnap, researchers supply an Input Structure, define designable and fixed segments with Contigs, optionally refine atom-level constraints with Contig Atoms, and specify a Ligand Code when catalytic chemistry depends on a bound substrate or cofactor. This makes the workflow a good fit for active-site transplantation, theozyme scaffolding, and chemistry-aware enzyme ideation.

The paper emphasizes atom-level active-site scaffolding. RFdiffusion2 simultaneously infers scaffold geometry and catalytic side-chain placement, which removes one of the major bottlenecks in earlier enzyme-design pipelines that had to search rotamers and backbone placements in separate steps. It also supports unindexed, sequence-agnostic motif conditioning, which is important when catalytic atoms matter more than a fixed residue numbering scheme.

On Neurosnap, Contigs define which structural spans are preserved and which are generated, while Contig Atoms lets researchers constrain only the chemically essential atoms of a catalytic motif. Length Min, Length Max, and Contigs as Guidepost (unindexed mode) determine how much scaffold diversity the model is allowed to explore around the active site.

Researchers typically use RFdiffusion2 to produce a candidate set of catalytic scaffolds, then evaluate those designs with downstream sequence design, folding, or activity-focused screening. The output is most valuable when read as a structured search over enzyme architectures rather than as one final answer.

• Catalytic-design-native inputs: RFdiffusion2 accepts structural motifs, atom-level constraints, and ligand context, which matches the way many enzyme-design problems are actually specified.
• Simultaneous scaffold and side-chain inference: The method is built to place catalytic chemistry and backbone geometry together instead of in separate search stages.
• Guidepost flexibility: Indexed and unindexed contig modes support both motif-preserving and more exploratory active-site scaffolding workflows.
• Enzyme-campaign utility: Neurosnap turns a complex active-site design protocol into a reproducible no-code workflow for generating catalytic hypotheses.