How to Use RFdiffusion3
Commercially Available Online Web Server
RFdiffusion3 guide to all-atom protein design, binder contigs, chain breaks, fixed atoms, hotspots, origin placement, and candidate validation.
RFdiffusion3 is an all-atom protein design model that generates side-chain geometry together with the protein backbone and can condition on ligands, nucleic acids, and selected atoms. It is useful when a design must satisfy a specific contact geometry or preserve a functional motif rather than merely adopt a broad fold.
On Neurosnap, Input Structure, Contig, Fixed Atoms, Hotspots, and Unindexed Motif define the structural problem. Sampling and origin controls govern how generation explores that problem. Applications include protein binder design, ligand-centered pocket design, and scaffolding around atom-level functional constraints.
Choose RFdiffusion3 for explicit structural conditioning
Compared with the original RFdiffusion workflow, RFdiffusion3 exposes more direct control over atoms and non-protein context during design. RFD4-Proteina is a separate joint-generation service with different task presets, selection conventions, and output interpretation.
Treat generated designs as candidates. Preserving selected atoms or contacting a hotspot does not establish folding, binding affinity, enzyme activity, or selectivity. The useful question is whether a proposed sequence and structure satisfy the intended geometry and remain plausible under independent structural and experimental assessment.
How RFdiffusion3 Works
Prepare the structure and define retained context
Upload Input Structure for tasks involving residue-referencing contigs, fixed atoms, hotspots, unindexed motifs, or ligands. Inspect chain identifiers, actual residue numbers, missing atoms, and the intended biological assembly. The input can retain nucleic acids and heteroatoms, while waters are removed.
A cropped target should preserve sufficient structural context around the intended interface. Record any residue remapping and compare generated complexes with the complete protein. Artificially exposed core residues, omitted domains, or missing glycans can create misleading binder surfaces.
For ligand-centered design, Ligand is a three-letter CCD code matching a ligand already present in the uploaded structure. Entering a code alone does not supply a molecule or place it in a pocket. Verify both the chemical identity and its coordinates before conditioning a design around it.
Write a contig with explicit chain breaks
A Contig combines retained input ranges and generated lengths, separated by commas. A token such as A1-100 retains those input residues; 40-60 specifies a generated segment between 40 and 60 residues. Use /0 explicitly to make independent chains.
For an illustrative target chain A numbered continuously from 1 through 100 and a separate 40–60-residue binder:
A1-100,/0,40-60
Without the chain break, the segments describe one continuous chain rather than the intended separate binder. Verify the actual output chain separation before proceeding to sequence selection or further modeling.
When a contig is supplied, Length Min and Length Max are ignored. When no contig is provided, those fields control generated length. Do not confuse a target residue range with the desired binder length or assume a length field overrides an explicit contig.
Distinguish fixed atoms, hotspots, and unindexed motifs
Fixed Atoms specifies the atoms to preserve at selected residues. Each form row has Selection and Atoms fields. Atom choices include explicit comma-separated names and the supported selections ALL, BKBN, and TIP. For example, selecting A10-12 with BKBN is different from preserving all side-chain atoms in those residues.
Hotspots specifies target atoms the generated protein should contact. Use the same two-column selection style, such as A64 with TIP or A88-90 with ALL, only if those residues exist in the retained target. A hotspot is an interaction objective, not a fixed amino-acid position in the binder.
Unindexed Motif retains motif residues without fixing their exact sequence placement in the final design. Do not list the same residues as both indexed contig motifs and unindexed motifs. Indexed placement, atomic preservation, and desired contact are separate decisions; choose each to express the experiment precisely.
The official RFdiffusion3 input specification explains these concepts. Match examples to the fields exposed by Neurosnap rather than copying unrelated command-line options.
Place generation near the intended epitope
Infer Origin Strategy can use default behavior, the input center of mass (com), or selected hotspots. For a localized epitope, hotspot-based placement can be more appropriate than the center of a large target. It does not guarantee the final binder will be compact or correctly placed.
Origin Override supplies explicit x,y,z coordinates and supersedes the inferred strategy. Leave it empty unless those coordinates are justified in the uploaded structure's frame. An old override can silently defeat a carefully chosen hotspot strategy.
Sampling controls and their tradeoffs
| Setting | Interpretation |
|---|---|
| Number Designs | Candidate count, not an accepted-binder or experimental-hit count. |
| Timesteps | Diffusion steps per generation; the default is 200. |
| Step Scale | Changes sampling behavior; the form supports 1–2, with default 1.5. |
| Gamma 0 | Added-noise control; lower values reduce this noise and can reduce diversity. |
| Non Loopy | Biases generation toward more structured folds and fewer loops. |
Begin with a small candidate set and an explicit hypothesis. Upstream examples can use settings outside this form's supported range; do not assume every upstream parameter value is available here. Change one factor at a time and compare the resulting topology, interface placement, and candidate diversity.
Example: a localized epitope binder pilot
For a verified target A1–100 with accessible intended contacts at A64 and A88, use the example contig above. Add hotspot rows A64 / TIP and A88 / ALL, select the hotspots origin strategy, and leave Origin Override empty. These positions are illustrative; replace them with inspected residues for the actual target.
Generate a small pilot, then inspect both the designed chain and retained target. Check compactness, chain separation, intended contact location, clashes, and whether the interface remains accessible in the complete assembly. If designs are extended or misplaced, revisit the crop, contig, hotspot map, and origin before increasing candidate count.
Evaluate RFdiffusion3 outputs and common problems
Compare generated structures for motif preservation, plausible side-chain geometry, and diversity. Do not reinterpret displayed B-factors as a calibrated binding-affinity score or transfer confidence thresholds from another predictor. Independently refold shortlisted sequences and compare them with the intended design geometry before experimental selection.
Why was a binder fused to the target? Check for a missing /0 chain break in Contig.
Why was generation centered away from the epitope? Check hotspot mapping, retained target residues, Infer Origin Strategy, and any explicit override.
Do nonconserved hotspots guarantee selectivity? No. Map the interface onto relevant homologs and assess shortlisted candidates against those off-targets. Favorable on-target geometry is not proof of specificity.
What is Neurosnap?
Neurosnap is the leading platform for bioinformatics and computational science focused on expanding access to powerful modeling and simulation tools. Because many state-of-the-art machine learning systems remain complex to install, configure, and scale, Neurosnap offers a clean, browser-based workspace that removes the burden of infrastructure management, dependency conflicts, and command-line tooling.
Built for biologists, chemists, and cross-disciplinary scientists, the platform enables advanced computational workflows without requiring expertise in software engineering or cloud architecture. Researchers can launch analyses through an intuitive interface, connect programmatically through a comprehensive API, and rely on automated resource management to scale workloads efficiently. By taking care of the underlying compute and operational complexity, Neurosnap allows teams to devote their energy to scientific progress and faster iteration. Security and data protection remain foundational principles, with clear safeguards outlined in our Terms of Use and Privacy Policy to ensure your work stays protected.
Advancing Discovery with RFdiffusion3 on Neurosnap
Using RFdiffusion3 on Neurosnap could drastically accelerate all-atom protein design with explicit control over motifs, ligands, and atom-level interactions.
- 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.
How to Use RFdiffusion3 on Neurosnap
To harness the capabilities of RFdiffusion3, researchers can follow this streamlined workflow within Neurosnap:
- Access Neurosnap: Start by logging in to the Neurosnap website.
- Select Tool: From the list of available tools, choose RFdiffusion3.
- Provide Inputs: Provide all the inputs specified within the submission panel and optionally configure the tool as desired.
- Run Tool: Submit the RFdiffusion3 job and Neurosnap will execute it in the cloud, automatically notifying you as soon as your results are ready.
- Review Output: Explore your results through rich visualizations, including figures, plots, and interactive views designed to help you analyze findings with clarity and confidence.
Citations
Please cite the original work when using RFdiffusion3 in publications or research outputs.
|
Butcher, Jasper, Krishna, Rohith, Mitra, Raktim, Brent, Rafael I. Li, Yanjing, Corley, Nathaniel, Kim, Paul T. Funk, Jonathan, Mathis, Simon, Salike, Saman, Muraishi, Aiko, Eisenach, Helen, Thompson, Tuscan Rock, Chen, Jie, Politanska, Yuliya, Sehgal, Enisha, Coventry, Brian, Zhang, Odin, Qiang, Bo, Didi, Kieran, Kazman, Max, DiMaio, Frank, and Baker, David. De novo Design of All-atom Biomolecular Interactions with RFdiffusion3. bioRxiv. 2025. |
|
Neurosnap Inc. (2022). Neurosnap: An online platform for computational biology and chemistry. Available at: https://neurosnap.ai/ |
Similar Services
Explore related tools that support similar research workflows:
Proudly supporting 50,000+ scientists worldwide, including 7,000+ leading biotech and global biopharma organizations.
Making Scientific Research
Faster & Easier
Register for free — upgrade anytime.
Interested in getting a license? Contact Sales.
Try Free