The UW Medicine Institute for Protein Design has released its first AI model co-developed with Nvidia, an open-source tool called RFD4-Proteina that researchers say designs new proteins with far greater speed and success than previous methods. The Seattle institute, directed by 2024 Nobel Chemistry laureate David Baker, says the model can generate everything from antibodies to enzymes that break down environmental toxins.
A New Model Built on Two Research Lineages
RFD4-Proteina co-designs protein sequences and their full atomic structures at once, using what developers call reward-guided inference-time search, according to a technical blog post from NVIDIA. The model merges two separate computational lineages: the Institute for Protein Design's RosettaFold Diffusion line and Nvidia's Proteina-Complexa, as detailed in a release from the UW Medicine Newsroom. The institute, known as IPD, previously built RoseTTAFold for predicting protein structures and RFdiffusion for designing new ones, and RFD4-Proteina draws on both systems along with Nvidia's architecture, as reported by GeekWire.
UW and Nvidia reported fivefold to tenfold higher design success rates in computational benchmarks, with the model generating protein candidates five to 10 times faster than prior models, per the UW Medicine Newsroom's announcement. Lab bench tests confirmed that the model's protein outputs behaved as anticipated in experimental settings, the same release states. The model is immediately available as an open-source resource under the BSD-3 license, hosted through Berkeley Software Distribution.
From Years of Research to Days of Design
Baker said RFD4-Proteina will enable solutions in medicine, sustainability, materials and biotechnology, per GeekWire's reporting. The practical impact is already visible in the institute's own halls. UW bioengineering predoctoral candidate Seth Woodbury used RFD4-Proteina for enzyme work and designed a pesticide-degrading enzyme in just a few days that performed many-fold better than his previous best, GeekWire reports. Woodbury had previously published a paper on related enzyme research in Nature in 2025.
The model supports a wide range of design tasks from a single system, including monomers, protein and DNA binders, antibodies and nanobodies, enzymes, motif scaffolding, and small-molecule pockets, according to the RosettaCommons GitHub repository. Researchers can access it through a Hugging Face release, a Python API, and command-line tools for generation, evaluation and training, the repository notes. Nvidia's technical blog adds that the model cuts diffusion sampling from roughly 400 steps in Proteina-Complexa down to 60. The project was led by UW Ph.D. students Nathaniel Corley, Maxwell Kazman and Rafi Brent, working alongside visiting University of Oxford student Kieran Didi, according to the UW Medicine Newsroom.
A Massive Computing Grant Powers the Work
The breakthrough comes on the heels of a major infusion of computing resources. IPD received a large grant of computing power from Jensen Huang and Lori Huang through their foundation, GeekWire reported, with the institute beginning to access Nvidia GPUs through CoreWeave in June 2026. The UW Medicine Newsroom describes the Jen-Hsun and Lori Huang Foundation's in-kind grant as totaling nearly 7 million hours of compute time, delivered through a partnership with CoreWeave using Nvidia GPUs. Baker described the foundation's grant as transformational for the institute's work, according to GeekWire.
The computing grant is meant to support research into precision cancer therapeutics, durable vaccines, currently untreatable diseases, and new enzymes that degrade environmental toxins, the UW Medicine Newsroom states. IPD plans to make the resulting technology, including AI models, software tools and research data, broadly available to scientists worldwide, and initial releases of additional AI models are anticipated to begin in late 2026 and continue into early 2027, pending research progress and experimental validation, per the same release.
Open Questions and a Track Record of Spinouts
Not every detail about the new model has been made public. Nvidia's model card for RFD4-Proteina states that the training data's composition, count and content have not been disclosed. The model card lists approximately 434 million parameters, including a 178-million-parameter trunk and 256-million-parameter autoencoder, while the Hugging Face release lists 435.6 million parameters across the paired bundle.
The institute's track record lends weight to the launch. IPD lists 11 spinout companies, according to GeekWire, including vaccine developer Icosavax and protein-design startup Archon Biosciences. The Institute for Protein Design separately says it has launched more than ten spinout companies that together have raised over $2 billion. Nvidia, in a separate announcement, framed the broader goal plainly: the model is meant to help researchers explore and prioritize candidates for laboratory testing, with potential applications including proteins that interrupt disease processes or detect signs of disease in diagnostic tests, rather than replacing experimental validation itself.
