Skip to content
Neural Daily

ModelsNews

NVIDIA RFD4-Proteina: open model for protein design

NVIDIA and the University of Washington released an open model that generates protein sequence and all-atom structure together. You can try it in the browser, free until November 4.

By: Mikola Máté Patrik · · 2 min read

Two ribbon views of the PD-L1 protein target in teal with green hotspots, the right one with an orange generated binder
Image: NVIDIA · Press kit · Official Hugging Face blog post social image (og:image) · Source

What changed?

On October 21 NVIDIA introduced (Hugging Face blog) RFD4-Proteina, an open BioNeMo model for programmable protein design, co-developed with the University of Washington's Institute for Protein Design.

Protein design has usually been split across several models: one for the backbone, one for the sequence and one for the all-atom structure. Information is lost at every handoff. RFD4-Proteina generates sequence and full atomic structure in a single model. It can design in contexts that include other proteins, small molecules, nucleic acids, glycans, metal ions and modified amino acids.

What's new

  • Thinking setting: like LLMs, the model can refine each design at three levels (low, medium, high). More refinement raises the chance of a good design but costs more compute.
  • Efficiency: in a benchmark across 24 binding targets, high thinking produced 76.40 structurally distinct designs per GPU-hour that met the computational criteria. Proteina-Complexa, the second-best method, produced 13.24.
  • Reward-guided search: external evaluators such as a structure prediction model or a physics-based energy calculation score the designs and steer exploration toward the properties you define.
  • Architecture: a flow-matching model paired with an autoencoder, about 0.4B parameters according to the Hugging Face model page.

How to try it

It runs in the browser in two Hugging Face Spaces. NVIDIA is sponsoring free access to the dedicated Space (NVIDIA RTX PRO 6000 GPUs) from October 21 to November 4. After that, the ZeroGPU Space stays free within Hugging Face quotas. Both Spaces run SecureDNA biosecurity screening on the input and on every generated design.

Locally it installs as a pip package (rfd4-proteina), from the GitHub repo or through the Diffusers library. It needs Linux x86_64, Python 3.12, an NVIDIA GPU and a CUDA 13-compatible driver, though a CUDA 12 build also exists.

Why it matters

An all-atom design model is now within reach of smaller teams, even without a local install. The designs are computational results, so binding still has to be confirmed in the lab. The Anthropic and Adaptyv protein design competition is an obvious place to test it.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles