Protein Design · Baker Lab / UW IPD × NVIDIA
RFD4-Proteína
Designs a protein's sequence and all-atom structure together, conditioned on a target, motif, ligand or length.
What RFD4-Proteína is
RFD4-Proteína is a flow-matching model for biomolecular design from the Baker Lab at the Institute for Protein Design, University of Washington, and NVIDIA GenAIR. It generates protein sequences and all-atom structures together, conditioned on the molecular context you want to preserve: a binding target, an enzyme motif, a ligand or selected structural features.
The release is a matched pair of a flow model and the autoencoder it was trained against, and the two are not interchangeable across releases. The current autoencoder is unified: one autoencoder covers proteins, nucleic acids and small molecules, with a 147-class residue vocabulary and a 32-dimensional latent.
A design is specified as an annotated CIF structure, which the project's own tooling builds from a YAML condition specification. The annotations state what to hold fixed and what to design, so an ordinary, unannotated structure is not a complete specification. An unconditional monomer needs only a length, with no input coordinates. The hosted tool here takes the finished annotated CIF or a length; it does not read a YAML specification.
How it works
The released flow model is the pretrained base with a rank-32 LoRA fine-tune merged in. The paired bundle stores 435,622,465 parameters in FP32, and the weights file is 1.74 GB.
Generation takes 60 sampling steps. The three compute presets differ only in how often the trunk is recycled at each step: low uses 120 trunk passes, medium, the default, uses 180, and high uses 300.
Key facts
- Current release
- 2026-10-19
- Model
- Joint sequence and structure generation with a matched flow model and autoencoder
- Inputs
- Annotated CIF structures, built with the project's YAML condition specifications; here, an annotated CIF or a length
- Outputs
- Designed molecular structures in CIF format, with generated residue identities and coordinates
- Weights
- 435,622,465 stored parameters in FP32; 1.74 GB
- Residue vocabulary
- 147 classes
- Latent dimension
- 32
- Sampling
- 60 steps; low, medium and high presets of 120, 180 and 300 trunk passes
- Runtime
- Linux x86_64, Python 3.12, NVIDIA GPU with a CUDA 13-compatible driver
What you can use it for
- Protein binder design
- Design a partner for a target such as PD-L1 from an annotated input that holds the target fixed and leaves the binder to be designed.
- Enzyme motif scaffolding
- Scaffold an enzyme motif: the annotations say which atoms to hold fixed, and the rest of the protein is designed around them.
- Antibody design
- The project ships an antibody design example against HER2 alongside the binder and enzyme examples.
- Unconditional monomers
- Generate a small protein from a length alone, with no input coordinates, as a starting point for further screening.
- Beyond protein-only design
- The project's own examples include a nanobody complex, a protein–DNA binder, an RNA pseudoknot, a ligand pocket, a glycan complex and a protein assembly.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
RFD4-Proteína
Strengths and limits
Good at
- Designs the sequence and the all-atom structure in one pass, so the result is a protein rather than a backbone that still needs a sequence.
- A single unified autoencoder covers proteins, nucleic acids and small molecules, so a design can be conditioned on a ligand or a nucleic acid.
- The three recycling presets trade compute for sampling depth without changing the weights.
- The weights are small at 1.74 GB, and peak GPU memory depends on token count, batch size and recycle depth rather than on the weights.
Watch out for
- An ordinary structure file is not a design specification: the CIF has to be annotated with what to hold fixed and what to design.
- Generation does not refold or score its own designs. The authors' pipeline refolds with AlphaFold2 or OpenFold3 and computes metrics, which need their own environments and weights.
- The published efficiency benchmarks were measured on the previous release pair and have not been re-measured on the current one.
- Computational passers do not establish experimental activity, so designs need structural and experimental evaluation.
- The authors tested it on Blackwell, Hopper, Ada Lovelace and Ampere A100 data-centre and workstation GPUs, and require a CUDA 13-compatible driver.
Inputs and outputs
You provide
What to designoptional
Length (residues)optional
Monomer mode only. The whole specification: there is no starting structure.
Annotated CIFoptional
Conditioned mode only. An ordinary PDB will not do — it has coordinates but says nothing about what to design. The model repository ships worked binder, enzyme and antibody examples.
Designsoptional
Independent samples. They run one after another, so time scales with this.
Sampling effortoptional
The model's own recycling schedules. All three take 60 sampling steps and differ in how often the trunk is recycled.
Seedoptional
The same seed and settings usually give the same design. The GPU kernels are not run in deterministic mode, so it is not guaranteed.
You get back
All-atom coordinates for the first design.
Every design and chain produced, with the sequence the model assigned it.
Licence
Commercial use permitted · BSD-3-Clause
Read again on 2026-11-04 at the pinned Hub commit and on GitHub, not from the tags. BSD 3-Clause, copyright the Institute for Protein Design, University of Washington: the LICENSE file is the same blob (db22701) in the code repository and the Hub repository, the Hub card states that it covers the model, the weights and the code, the GitHub licence API and the PyPI metadata both report BSD-3-Clause, and the repository is public and not gated. The card separates what it does not cover: the optional AlphaFold2 and OpenFold3 refolding runtimes in the same repository are third-party images under their own upstream terms, some of which restrict commercial use. They are neither downloaded nor run here (the loader fetches config.json and model.safetensors and nothing else), so generation stays inside the BSD grant. The training data's own terms, on the companion dataset card, are CC0 for PDB entries and CC BY 4.0 for AlphaFold DB structures and CATH annotations. The Python environment carries the usual GPU libraries under NVIDIA's SDK licence, and three copyleft libraries that are imported by a hosted service and not distributed to anyone: Open Babel (GPL-2.0, through AtomWorks and tmol), frozendict (LGPL-3.0, through tmol) and biotraj (LGPL-2.0 or later, through biotite). One thing a reader should know rather than discover: the authors' own hosted Studio screens every input and every design with SecureDNA before returning it. That screening belongs to their deployment, not to these weights, and running the model here does not reproduce it.
Questions
Can I use RFD4-Proteína commercially?
Commercial use permitted. Read again on 2026-11-04 at the pinned Hub commit and on GitHub, not from the tags. BSD 3-Clause, copyright the Institute for Protein Design, University of Washington: the LICENSE file is the same blob (db22701) in the code repository and the Hub repository, the Hub card states that it covers the model, the weights and the code, the GitHub licence API and the PyPI metadata both report BSD-3-Clause, and the repository is public and not gated. The card separates what it does not cover: the optional AlphaFold2 and OpenFold3 refolding runtimes in the same repository are third-party images under their own upstream terms, some of which restrict commercial use. They are neither downloaded nor run here (the loader fetches config.json and model.safetensors and nothing else), so generation stays inside the BSD grant. The training data's own terms, on the companion dataset card, are CC0 for PDB entries and CC BY 4.0 for AlphaFold DB structures and CATH annotations. The Python environment carries the usual GPU libraries under NVIDIA's SDK licence, and three copyleft libraries that are imported by a hosted service and not distributed to anyone: Open Babel (GPL-2.0, through AtomWorks and tmol), frozendict (LGPL-3.0, through tmol) and biotraj (LGPL-2.0 or later, through biotite). One thing a reader should know rather than discover: the authors' own hosted Studio screens every input and every design with SecureDNA before returning it. That screening belongs to their deployment, not to these weights, and running the model here does not reproduce it.
How do I run RFD4-Proteína without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 40GB, typically in 20s, and a run costs about 70 credits. Visitors without an account get 1,000 credits a month.
What input does RFD4-Proteína need?
An annotated CIF, which the project's own tooling builds from a YAML condition specification. The annotations state what to hold fixed and what to design. For an unconditional monomer, a length with no input coordinates is enough. The hosted tool here takes the finished annotated CIF or a length, not the YAML specification itself.
Does it design the sequence as well as the structure?
Yes. It generates residue identities and all-atom coordinates together, rather than a backbone that needs a separate sequence-design step.
How do the low, medium and high effort settings differ?
They are the model's own recycling schedules over 60 sampling steps, with 120, 180 and 300 trunk passes. Medium is the default.
Does it check its own designs?
No. Generation-only runs omit external refolding and scoring. The authors use AlphaFold2 or OpenFold3 refolding and metrics, which are a separate download.
What hardware does it need?
Linux x86_64, Python 3.12 and a CUDA 13-compatible driver. The weights are 1.74 GB in FP32; peak GPU memory depends on token count, batch size and recycle depth.
How to cite it
Baker Lab, Institute for Protein Design, University of Washington, and NVIDIA GenAIR. RFD4-Proteína: joint sequence and all-atom structure design. Software, 2026. https://huggingface.co/nvidia/RFD4-Proteina
Written from RFD4-Proteína's own documentation (model card, code). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.