NVIDIA and UW Open RFD4-Proteina Protein Design Model on Hugging Face Spaces

The blog post says protein design usually begins with a goal—a target to bind, a molecule to recognize, or an arrangement of atoms to preserve—and turning that goal into a design has often meant splitting the workflow across models that each handle one modality or stage, then reassembling the pieces. Information is lost at every handoff, the post says. RFD4-Proteina reduces those handoffs between separate backbone, sequence, and all-atom structure design models by bringing sequence and all-atom structure generation into one model. It can generate proteins in molecular contexts that include other proteins, small molecules, nucleic acids, glycans, metal ions, and modified amino acids, such as a phosphorylated residue on a target protein.

The model pairs a flow-matching model with an autoencoder. The autoencoder represents each standard or modified amino acid, sugar, DNA or RNA nucleotide as an anchor-atom coordinate—for amino acids, the alpha carbon—and a learned latent vector encoding its chemical identity and local geometry. The flow model generates anchor coordinates and latent vectors together, and the decoder expands them into all-atom 3D structures and sequences. According to the post, this lets the model generate sequences and structures together around a supplied constraint: the target a binder must engage, or the catalytic motif a protein must contain.

Users can run the model in a browser before installing anything in either of two Hugging Face Spaces: the dedicated RFD4-Proteina Space, which runs on dedicated GPUs, and the ZeroGPU Space, which runs on shared capacity. Both begin with supplied examples that let users inspect a target and its generated designs and save input and output for further work. The post says the dedicated Space runs on NVIDIA RTX PRO 6000 GPUs during the sponsored free-access window. After Nov. 4, 2026, the dedicated Space stops being free, but the ZeroGPU Space remains free to use on shared capacity, subject to Hugging Face account quotas.

Both Spaces run AtomWorks Studio, a browser-based viewer and annotation editor for molecular inputs that also runs standalone, according to the post. The built-in PD-L1 example shows a target and its design annotations before generation; an expandable binder segment is part of the input specification. Every design starts from a design context: at minimum, a protein structure to design binders against, a region where the binder should bind, and the specific residues or atoms the design process should engage. Those engagement points are called hotspots and are marked on the target atoms. With the context defined, a user can generate a single design using the example’s default settings, locate the designed protein and supplied target in the interactive viewer, and check that the designed protein contacts the marked hotspots.

Generated designs, annotated targets, and their complexes can be downloaded in CIF file format, the post says. It recommends noting the seed number as a baseline, then changing generation settings to see how each affects results. Holding input and other settings fixed, a new seed generates a different sample from the same design space, and those binders will usually look quite different from the first set, according to the post. Users can also try other settings, including a thinking setting that the post says is similar to how large language models can refine their answers. Downloaded CIFs carry the design context and can be reloaded in a later session. From there, users can predict developability properties from generated sequence or structure, such as thermal stability or aggregation potential, to decide which designs are worth experimental characterization.

The post notes that the PD-L1 example is a computational result and that experimental validation would be needed to confirm binding. It also points users to their own design problems, including a protein design competition sponsored by Anthropic and Adaptyv. The free dedicated Space is available from Oct. 21 to Nov. 4, 2026; after that, users can continue with the ZeroGPU Space under Hugging Face quotas.

Editor's Summary

RFD4-Proteina combines protein sequence and all-atom structure generation in a single BioNeMo open model from NVIDIA and the University of Washington, aimed at reducing handoffs in design workflows. Hugging Face Spaces offer browser access, with a dedicated GPU Space free through Nov. 4, 2026, and a ZeroGPU option remaining free under quotas. The blog post presents the model as a computational design tool whose outputs still require experimental validation.