NVIDIA and the University of Washington’s Institute for Protein Design have made RFD4-Proteina available through Hugging Face Spaces, where users can generate protein designs in a browser. NVIDIA is sponsoring free access to the dedicated GPU Space from Oct. 21 through Nov. 4, 2026; the release gives researchers and other users a way to try the model without first installing it, though its generated binders remain computational predictions until tested experimentally.

RFD4-Proteina is a BioNeMo open model designed to generate protein sequence and all-atom structure together. Hugging Face’s description says it is intended to reduce the handoffs among separate models for backbone generation, sequence design and all-atom structure modeling. Its design settings can specify a target, binding region and hotspot residues or atoms the generated protein should engage.

The model can work in molecular contexts involving proteins, small molecules, nucleic acids, glycans, metal ions and modified amino acids. It combines a flow-matching model with an autoencoder: the model generates anchor-atom coordinates and learned representations of chemical identity and local geometry, then decodes these into sequences and all-atom structures. This is the system’s stated approach; the description does not establish that any individual design will bind or function.

Users can start with supplied examples in either the dedicated RFD4-Proteina Space, which runs on dedicated GPUs, or the ZeroGPU Space, which uses shared capacity. The browser interface displays the target and generated design, and lets users download annotated target, design and complex files in CIF format. The shared guide recommends keeping track of the random seed when comparing runs. It says the dedicated Space stops being free after Nov. 4, while ZeroGPU remains free subject to Hugging Face account quotas.

Protein Design Moves Into the Browser

Browser access lowers the setup barrier for exploring a model that combines sequence and structure generation. A user can inspect an example, set binding hotspots and generate a candidate without assembling a local pipeline of separate tools. The downloadable annotated CIF files also give users a route to continue analysis with their own software after a session.

The practical value depends on what happens after generation. The supplied material describes computational filtering and follow-up analysis, including checks for predicted structure and developability properties such as stability and aggregation potential. Those steps can help prioritize candidates, but they do not show that a design binds in a laboratory or has the desired biological effect. Experimental characterization remains necessary.

From Separate Models to One Workflow

Protein design often involves several stages: specifying a target and desired interaction, generating a structure, selecting a sequence and modeling the resulting molecule. Hugging Face says RFD4-Proteina brings sequence and all-atom structure generation into one model, aiming to reduce information loss between those stages. It was co-developed by NVIDIA and the University of Washington’s Institute for Protein Design.

The release also describes a “thinking” setting of low, medium or high, with medium as the default. According to the model material, higher settings use more compute to refine designs and may increase the likelihood that they pass computational filters. The source reports a benchmark across 24 protein-binding targets: at high thinking, RFD4-Proteina produced 76.40 structurally distinct designs per GPU-hour meeting the stated computational criteria, compared with 13.24 for Proteina-Complexa. The stated criteria included predicted binder pTM of at least 0.80, minimum interface PAE of no more than 1.5 Å and target-aligned binder backbone RMSD of no more than 2.5 Å. These are benchmark results under specified computational measures, not evidence of experimentally confirmed binding.

Binding Remains to Be Tested

The supplied material does not report laboratory validation of the example PD-L1 binder or establish how often generated candidates bind successfully in experiments. It also does not give enough information here to assess the benchmark’s full methodology, statistical uncertainty or performance across targets outside the 24-target set. The reported computational thresholds indicate how candidates were counted in that benchmark, not whether they work in biological settings.

Access conditions after the sponsored period are partly specified: the dedicated Space will no longer be free after Nov. 4, 2026, while ZeroGPU remains free subject to account quotas. The material does not state whether dedicated access will then be available at a paid rate or provide details about future capacity.

Try the Space, Then Validate

During the sponsored window, users can run the dedicated Space on an example or their own design problem, vary the seed and thinking setting, and save the annotated CIF outputs for further computational analysis. After Nov. 4, users can continue through the ZeroGPU Space under its account quotas. The material also points users to a protein design competition sponsored by Anthropic and Adaptyv.

The next scientific step for any promising candidate is independent analysis followed by experimental testing. No outcome or schedule for such validation is specified in the supplied announcement.

Where I land

I see the browser release as a useful way to make a complex protein-design workflow easier to try and inspect. Combining sequence and structure generation could reduce friction for researchers exploring candidates, and the benchmark reported by the developers gives a concrete computational comparison under defined filters.

The strongest counterargument is that computational throughput and structural filters do not establish biological performance. A large number of candidates that pass prediction thresholds may still fail to bind or behave as intended in experiments. I would put more weight on the model’s practical impact if independent groups reported reproducible experimental results across multiple targets, alongside enough methods detail to evaluate how the candidates were selected. Until then, I regard the release as accessible research tooling with promising but unproven experimental utility.

Key Questions

What is RFD4-Proteina?

It is a BioNeMo open model co-developed by NVIDIA and the University of Washington’s Institute for Protein Design that generates protein sequences and all-atom structures together.

How long is the dedicated Hugging Face Space free?

NVIDIA is sponsoring free access from Oct. 21 through Nov. 4, 2026. The ZeroGPU Space is described as remaining free afterward, subject to Hugging Face account quotas.

Does a generated binder mean the target has been experimentally confirmed?

No. The generated structure is a computational result. The source says experimental validation is needed to confirm binding.

What does the thinking setting change?

The low, medium and high settings control how much the model refines a design, according to the product material. Higher settings use more compute per design and are described as increasing the likelihood of passing computational filters.

Source: Hugging Face