iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
RFdiffusion3 (RFD3) is an all-atom diffusion model for designing new biomolecules around precisely specified catalytic, binding, ligand, or nucleic-acid interactions. Instead of starting with only a protein backbone or residue frames, it models polymer atoms directly, so a researcher can constrain the actual atoms that must contact a substrate, metal, DNA base, RNA, or another protein. The output is a set of structural hypotheses—not guaranteed functional proteins—that still require sequence design, computational filtering, and laboratory testing.
What RFdiffusion3 does
RFdiffusion3 extends the RFdiffusion family from backbone-centered generation to direct atom-level conditioning. The model diffuses and predicts atomic coordinates while preserving user-specified geometric constraints, then builds a surrounding protein scaffold that could support those constraints. The approach is intended for de novo design rather than reconstruction of a known natural sequence.
The Institute for Protein Design announced the open-source release on December 3, 2025. Its release describes proteins designed to interact with molecules commonly found inside living cells and says that training code and model weights were released through the project’s GitHub and Rosetta Commons ecosystem. The associated bioRxiv preprint is dated November 19, 2025; its publication or version status should be checked before using it as a final peer-reviewed reference.
Free tools Windows power users keep installed
One-click scans. No signup required.
How atomic functional-site scaffolding works
Specify chemistry, not just residues
A functional site can be represented as a set of catalytic or binding atoms whose three-dimensional arrangement must remain intact. Those atoms may belong to disconnected parts of a motif, a ligand, a nucleic acid, or a combination of partners. Because the constraint is expressed at the atom level, residue numbers do not necessarily need to be known in advance.
#1 Best Overall
Generate the surrounding fold
RFD3 uses the constrained atoms as anchors while generating coordinates for the rest of the polymer. It can infer where residues should be placed and which side-chain rotamers can present the required chemical groups. This is particularly useful when a catalytic geometry spans several residue islands or when contacts involve a small molecule or nucleic acid rather than a protein-only interface.
Why this differs from residue-indexed motif design
Earlier motif-scaffolding workflows commonly required a specified backbone motif or residue-indexed representation. The RFdiffusion2 paper described a key shift: an active site can be supplied as functional-group atoms without fixing residue indices, allowing generation to determine residue placement and rotamer conformations. RFdiffusion3 generalizes that all-atom idea to a wider set of protein–small-molecule, protein–DNA, protein–protein, and enzyme-design tasks.
What the reported evidence shows
The evidence combines computational benchmarks with selected experimental demonstrations. These categories should not be treated as interchangeable: a passing in-silico design satisfies geometric or structural filters, while an experimentally characterized protein has been produced and tested under particular laboratory conditions.
Recommended Free Tools
- Atomic Motif Enzyme benchmark: The 2025 bioRxiv preprint reports that RFdiffusion3 outperformed RFdiffusion2 on 37 of 41 cases, or 90% of the benchmark problems.
- Complex, multi-island motifs: For cases containing more than four residue islands, the reported passing-design rate was 15% for RFdiffusion3 versus 4% for RFdiffusion2.
- Earlier RFdiffusion2 result: A related Nature Methods report states that RFdiffusion2 scaffolded all 41 active sites in its benchmark, compared with 16 for previous methods. That figure concerns RFdiffusion2 and should not be read as an RFD3 result.
- Experimental demonstrations: The RFdiffusion3 preprint reports designing and experimentally characterizing DNA-binding proteins and cysteine hydrolases.
- Efficiency: The preprint reports one-tenth the computational cost of prior RFdiffusion versions on its benchmarks. The Institute for Protein Design release describes a ten-fold speed improvement over RFdiffusion2 in its computational benchmarks. The exact hardware, sampling settings, and accounting method belong to those respective benchmark contexts.
These numbers show progress on selected tasks; they do not establish that every ligand, catalytic mechanism, polymer context, or expression system will work.
RFdiffusion3 compared with earlier RFdiffusion models
| Capability | Original RFdiffusion | RFdiffusion2 | RFdiffusion3 |
|---|---|---|---|
| Primary representation | Protein backbone or residue-frame generation | Expanded all-atom motif handling | All-atom diffusion with explicit polymer atoms |
| Residue indices and rotamers | Backbone motifs generally define residue positions | Can use unindexed atomic motifs and infer placement and rotamers | Uses atom-level constraints so placement and side-chain conformations can be inferred during generation |
| Ligand and nucleic-acid context | Not the central capability in the original formulation | Supports atomic-motif extensions | Designed to cover protein–small-molecule, protein–DNA, protein–protein, and related interaction tasks |
| Reported efficiency | Baseline for later comparisons | RFD3 release reports RFD3 as ten times faster in its computational comparison | Preprint reports one-tenth the computational cost of prior versions on its benchmarks |
| Evidence level | Included a 25-problem in-silico motif-scaffolding benchmark | Reported 41 benchmark active sites scaffolded versus 16 by previous methods | Reports benchmark gains plus experimental characterization of DNA-binding proteins and cysteine hydrolases |
The table compares the research reports’ stated scopes, not a universal performance ranking. Hardware, sampling budgets, target classes, and filtering criteria can change practical results.
What researchers can design
Enzymes and catalytic sites
For an enzyme, the input can describe catalytic atoms and their required geometry rather than merely asking for a fold that resembles a known enzyme. The model can then search for a protein environment that presents those groups. RFD3’s reported cysteine-hydrolase demonstrations indicate that this workflow has reached experimental testing for at least one catalytic class.
DNA- and RNA-binding proteins
Nucleic-acid atoms can be included in the conditioning context, enabling designs around contacts to bases, sugars, or phosphates. The reported DNA-binding experiments are evidence that this is more than a geometric feature in the model interface, although activity and specificity remain properties to measure for each design.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsSmall-molecule and protein interfaces
Because ligands and other non-protein atom constellations can be represented explicitly, a design can be optimized around contacts to a small molecule or another macromolecule. This is useful when the interaction depends on exact hydrogen-bond, charge, coordination, or steric relationships that a coarse backbone representation cannot express directly.
A practical RFdiffusion3 design workflow
- Define the design objective. Specify the ligand, nucleic acid, protein partner, or catalytic chemistry, and identify the atoms whose geometry is essential.
- Prepare the atomic motif. Encode the required coordinates and any molecular partners. Check atom names, protonation or charge assumptions, bond definitions, and coordinate units before generation.
- Generate candidates. Run RFdiffusion3 to sample multiple backbones and side-chain arrangements under the atomic constraints. A single generated structure is not a reliable design set.
- Design sequences. Use a compatible sequence-design method to assign amino acids to the generated backbones. The related RFdiffusion2 workflow explicitly uses LigandMPNN for this stage; an RFD3 project should verify compatibility with the current release.
- Predict structures independently. Apply an all-atom structure predictor to test whether the designed sequence is likely to recover the intended fold, motif geometry, and partner contacts.
- Filter candidates. Remove structures with motif drift, steric clashes, exposed or buried groups inconsistent with the chemistry, poor confidence, implausible packing, or incompatible oligomeric states.
- Plan experiments. Select a diverse set rather than only the highest-scoring design. Order genes, express and purify the proteins, and assay folding, binding, catalytic turnover, specificity, and stability under defined conditions.
Computational agreement between the generator and a structure predictor is useful triage, not proof of function. Experimental controls should include an appropriate negative design and, where possible, a positive reference or catalytically altered variant.
Rank #4
Limitations and failure modes
Geometry is not mechanism
A generated arrangement can place atoms near one another without reproducing the proton transfers, electrostatics, dynamics, solvent effects, or conformational changes required for catalysis. A plausible active-site snapshot therefore needs biochemical testing.
Benchmarks cover selected problems
The 41-case enzyme benchmark and the reported DNA-binding and cysteine-hydrolase experiments are informative but finite. Performance can differ for a new ligand, unusual metal coordination, flexible RNA, membrane proteins, post-translational chemistry, or a target requiring large conformational rearrangements.
Sequence design and expression remain bottlenecks
RFD3 generates structural hypotheses; it does not guarantee that a sequence will fold, express solubly, avoid aggregation, or remain stable in the intended assay buffer. Sequence-design choices, synthesis quality, purification conditions, and assay design can dominate the final outcome.
Compute claims are benchmark-dependent
“Ten-fold faster” and “one-tenth the computational cost” describe the cited comparisons, not a promise for every workstation or sampling campaign. Generation time depends on model version, number of samples, atom count, hardware, and downstream prediction and filtering.
Access, reproducibility, and responsible use
The Institute for Protein Design release states that training code and model weights are available through GitHub and the Rosetta Commons ecosystem. Use the official release and project repository for the current installation instructions, model files, software dependencies, and citation metadata; these details can change as versions are updated.
For a reproducible record, retain the model version, checkpoint, input atomic coordinates, random seeds, sampling parameters, sequence-design software and version, structure-prediction settings, and every filtering threshold. Report which structures were only computationally evaluated and which proteins were experimentally characterized.
As with any generative protein-design system, proposed sequences should be screened under the relevant institutional biosafety, dual-use, and chemical-safety procedures. The model’s ability to describe an interaction does not by itself establish that a design is safe, active, or suitable for deployment.
Bottom line
RFdiffusion3’s defining advance is precision control at the atom level: researchers can describe the chemical contacts they need and ask the model to construct a protein around them, even when residues are disconnected or not indexed beforehand. Reported benchmark improvements, efficiency gains, and experiments with DNA-binding proteins and cysteine hydrolases make it a significant design tool, but every generated structure remains a hypothesis until independent structural checks and laboratory assays confirm folding and function.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
