Skip to content
AI.info

The Pulse

NVIDIA and UW Open RFD4-Proteina for Protein Design

The University of Washington’s Institute for Protein Design and NVIDIA introduced RFD4-Proteina, an open generative model that creates protein sequences and structures together. UW reports faster results on computational benchmarks and says lab tests validated designs across biomolecular tasks.

NVIDIA and UW Open RFD4-Proteina for Protein Design

“RFD4-Proteina will open the door to new solutions in medicine, sustainability, materials and biotechnology,”

David Baker, IPD director and professor of biochemistry at the University of Washington School of Medicine

The University of Washington’s Institute for Protein Design and NVIDIA introduced RFD4-Proteina on October 21, releasing a generative model designed to help researchers create proteins for specific tasks. Unlike a structure-prediction system that estimates a shape from a protein sequence, RFD4-Proteina generates a sequence and an all-atom structure together, using a target molecule or other design constraints as context.

The release joins the Institute for Protein Design’s RosettaFold Diffusion work with NVIDIA’s Proteina-Complexa line. The model is open under the BSD 3-Clause license, so researchers can use and modify its code and weights. NVIDIA also describes access through its BioNeMo tools and partner platforms; the model repository lists hardware and software requirements for running it locally.

Sequence and structure change together

Protein design requires choosing from an enormous range of possible amino-acid sequences and structures. RFD4-Proteina’s approach lets users specify what parts of a molecular structure should stay fixed, what should be designed, and what function a candidate should pursue. During generation, the model refines a protein’s sequence and shape alongside the structures of molecules it may bind or interact with.

NVIDIA’s technical explanation describes an optional search process that scores partial designs and continues developing promising candidates. Researchers can set objectives based on factors such as geometry, molecular interactions or prediction confidence. The method aims to concentrate computing on candidates that meet those objectives rather than treating every generated design as equally promising.

Faster design and lab tests, UW reports

RFD4-Proteina achieved five to ten times higher design success rates across computational benchmarks and generated candidates five to ten times faster than earlier models, according to UW’s announcement. Those are the institution’s reported benchmark comparisons, not evidence that the designs work as medicines in people. NVIDIA’s announcement likewise separates computational results from experimental performance, noting that laboratory experiments establish how a design behaves under tested conditions.

The UW team also tested generated proteins in the lab across biomolecular tasks with different sizes and functions. Seth Woodbury, a UW bioengineering predoctoral candidate and National Science Foundation fellow, said his team designed a pesticide-degrading enzyme in days that performed many times better than its previous design, which had taken months. Meg Lunn-Halbert, a UW Ph.D. candidate, said the researchers controlled which glycans—sugars involved in immune recognition—attached to proteins in less than a week, a task she said had taken months with earlier models.

Those accounts point to practical experiments, but the release does not report clinical tests or establish that any candidate is ready as a treatment or commercial enzyme. Laboratory testing remains necessary to determine whether a computational design performs its intended function.

An open model, with boundaries

The released model and weights are listed at 1.74 gigabytes in FP32, with a tested setup that includes Linux, Python 3.12 and a supported NVIDIA GPU. The repository warns that some optional third-party software bundled for structure evaluation has separate licensing terms; the BSD license applies to RFD4-Proteina’s own code, model and weights. That distinction matters for groups planning commercial use or redistribution.

The project was led at the Institute for Protein Design by doctoral students Nathaniel Corley, Maxwell Kazman and Rafi Brent, alongside visiting student Kieran Didi from the University of Oxford. Its next test is not another benchmark alone: it is whether teams outside the collaboration can reproduce useful designs, validate their functions in the lab and show where the model’s speed translates into better experimental results.

Sources

  1. Original source · Press release UW Medicine Newsroom newsroom.uw.edu

Other pages read

  1. NVIDIA Newsroom nvidianews.nvidia.com
  2. NVIDIA Technical Blog developer.nvidia.com
  3. Hugging Face model repository huggingface.co

Companies in this story

Related stories

Explore

More articles