RFdiffusion3 (RFD3) is an all-atom diffusion model for designing new proteins around chemically defined functional sites. Instead of treating a protein mainly as a backbone or set of residue frames, it models polymer atoms explicitly and can condition generation on catalytic atoms, ligand contacts, nucleic-acid interfaces and other non-protein geometry. The result is a way to ask for a fold that makes a specified chemical interaction possible, rather than starting with a fold and hoping the right chemistry appears.
What RFdiffusion3 does
RFD3 generates three-dimensional protein structures de novo while preserving atom-level constraints supplied by the designer. Those constraints can describe a catalytic arrangement, a binding pocket, a protein–DNA interface or contacts to a small molecule. The model then diffuses and predicts the surrounding atomic coordinates, building a protein scaffold around the prescribed geometry.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
Molecular Biology of the Cell | $203.99 | Buy on Amazon |
| 2 |
|
Molecular Biology: Principles and Practice | $169.53 | Buy on Amazon |
| 3 |
|
Molecular Biology of the Cell | $160.49 | Buy on Amazon |
| 4 |
|
BRS Biochemistry, Molecular Biology, and Genetics (Board Review Series) | $64.99 | Buy on Amazon |
| 5 |
|
Molecular Cell Biology (842581) | $329.81 | Buy on Amazon |
As an Amazon Associate I earn from qualifying purchases.
The Institute for Protein Design announced the open-source release on December 3, 2025. Its release describes interactions with molecules commonly found inside living cells and says that training code and model weights are available through GitHub and the Rosetta Commons ecosystem. A bioRxiv preprint dated November 19, 2025 presents the method and its benchmarks; publication or version status should be checked before relying on a later software release.
How atomic functional-site scaffolding works
1. Describe the chemistry, not just a residue pattern
A functional site is represented as a set of atoms whose relative geometry matters. These may include side-chain functional-group atoms, ligand atoms or atoms from a nucleic-acid partner. The description can preserve disconnected catalytic or binding groups and does not necessarily require residue numbers to be known in advance.
#1 Best Overall
2. Condition generation on those coordinates
During diffusion, RFD3 predicts atomic coordinates while being conditioned on the constrained atoms. It must place residues, side chains and backbone segments so the requested distances, orientations and contacts remain geometrically plausible. Because the model works at atom level, it can infer where residues should occur and which rotamers can support the motif instead of receiving a complete residue-indexed template.
3. Build a complete protein around the motif
The output is a candidate protein structure containing the preserved functional site and a newly generated surrounding scaffold. This is especially useful when the motif consists of several residue “islands,” when sequence indices are unknown, or when the design must account simultaneously for a ligand or nucleic acid.
Rank #2
What researchers can design
- Enzymes: catalytic atom arrangements can be scaffolded into new folds, including the cysteine hydrolases experimentally characterized in the preprint.
- Small-molecule binders: ligand atoms can be included in the conditioning geometry so the pocket is designed around the intended contacts.
- DNA-binding proteins: RFD3 has been used to design and experimentally characterize proteins that bind DNA.
- Protein–protein interfaces: interaction geometry can be specified for binder design and other complex-formation tasks.
- General functional-motif scaffolding: disconnected or atom-defined motifs can be embedded in a new protein framework without first fixing every residue index.
These capabilities describe reported tasks, not a guarantee that every ligand, catalytic mechanism, nucleic-acid sequence or cellular condition will work. The model generates structural hypotheses that still require computational filtering and laboratory testing.
RFdiffusion3 compared with earlier RFdiffusion models
| Aspect | Original RFdiffusion | RFdiffusion2 | RFdiffusion3 |
|---|---|---|---|
| Representation | Primarily backbone or residue-frame representations for motif scaffolding | Introduced atom-level descriptions for active-site motifs | All-atom diffusion model with explicit polymer atoms and broader interaction conditioning |
| Residue indices and rotamers | Motif residues generally supplied as a defined structural template | Can use unindexed atomic motifs and infer residue placement and rotamers | Extends atom-level conditioning to complex ligand, nucleic-acid and interaction contexts |
| Partners supported | Original work centered on protein motifs and a 25-problem in-silico benchmark | Active-site scaffolding, including atomic motifs | Reported protein–protein, protein–DNA, protein–small-molecule, enzyme and motif-scaffolding tasks |
| Efficiency | Baseline for the later comparisons | Reference point for RFD3 speed claims | The official 2025 release reports a ten-fold computational speed improvement over RFdiffusion2; the preprint reports one-tenth the computational cost of prior RFdiffusion versions on its benchmarks |
| Evidence level | In-silico motif-scaffolding evaluation | Benchmark evidence, including all 41 AME active sites scaffolded in the related paper | In-silico benchmark gains plus reported experimental characterization of DNA-binding proteins and cysteine hydrolases |
The representation change is the important distinction. RFdiffusion2 showed why atom-level, unindexed motifs matter: a catalytic description can specify the chemically relevant atoms without dictating which residue numbers must occupy them. RFD3 presents that idea as a broader all-atom foundation model for multiple interaction types.
Rank #3
What the reported results show
- On the Atomic Motif Enzyme benchmark, the 2025 bioRxiv preprint reports that RFD3 outperformed RFdiffusion2 on 37 of 41 cases, or 90% of the benchmark cases.
- For cases containing more than four residue islands, the preprint reports passing designs for 15% of RFD3 outputs versus 4% for RFdiffusion2.
- The related RFdiffusion2 paper reports 41 active sites scaffolded, compared with 16 by previous methods, in its benchmark.
- The Institute for Protein Design release reports a 10× computational speed improvement over RFdiffusion2. The preprint’s “one-tenth the computational cost” statement refers to its own benchmark settings, so neither figure should be treated as a universal runtime for every hardware configuration or design length.
Benchmark percentages measure selected computational tests. Experimental characterization of DNA-binding proteins and cysteine hydrolases is stronger evidence than a geometry-only score, but it still covers particular designs and assays rather than all possible targets.
A practical RFdiffusion3 design workflow
- Define the target interaction. Specify the catalytic atoms, ligand, DNA or other partner and identify which distances, orientations and contacts must be retained.
- Generate structural candidates. Run RFD3 with the atom-level constraints, producing multiple backbones and side-chain arrangements rather than relying on a single sample.
- Design sequences. Use a compatible sequence-design method to assign amino-acid sequences to the generated backbones. The related RFdiffusion2 workflow explicitly uses LigandMPNN for this stage.
- Predict structures with an all-atom method. Re-predict each sequence–structure pair, including the ligand or nucleic-acid context when relevant.
- Filter candidates. Check motif RMSD or equivalent geometry, clashes, unsatisfied contacts, pocket shape, interface burial, overall foldability and consistency between the generated and re-predicted structures.
- Prioritize a small experimental set. Select designs with robust motif preservation and plausible expression, purification and assay plans.
- Test experimentally. Measure folding and stability first, then binding or catalytic activity under defined laboratory conditions. Iterate from measured failures rather than treating the first computational batch as finished.
Why validation is still essential
A generated structure is a hypothesis. Diffusion can satisfy geometric constraints while missing protonation states, water networks, conformational dynamics, transition-state stabilization, cellular competition or expression problems. A sequence that looks compatible in an all-atom prediction may misfold, aggregate or bind the wrong species. For that reason, computational validation should separate three claims: the model generated a candidate, an independent predictor supports its structure, and experiments demonstrated the intended function. Only the third establishes activity under the tested conditions.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Access, openness and reproducibility
The December 3, 2025 Institute for Protein Design release states that RFD3 training code and model weights are available through GitHub and the Rosetta Commons ecosystem. Use the project’s current repository and release notes for installation requirements, weight files, inference instructions and citation metadata, because those details can change between versions. Record the exact model checkpoint, sampling settings, hardware, sequence-design tool, structure-prediction model and filtering thresholds for every published result.
Recommended Free Tools
Why the atom-level shift matters
Backbone-only generation is powerful when the desired motif is already expressed as a stable structural fragment. Enzyme and interaction design often starts with chemistry instead: a few catalytic groups, a metal-coordination arrangement, a ligand pose or a nucleic-acid contact surface. RFD3 moves the design input closer to that chemical description. Its reported benchmark gains, broader interaction scope and lower computational cost suggest a more practical route from functional requirements to candidate proteins, while the limited set of experimentally characterized examples makes laboratory validation indispensable.
Quick Recap
Best Value
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.
