Nvidia and UW release open-source AI model for protein design

A new era of protein design

The Institute for Protein Design (IPD) at the University of Washington School of Medicine has teamed up with Nvidia to launch RFD4-Proteina. This open-source generative model utilizes biomolecular reasoning to assist scientists in crafting new proteins that perform specific tasks. It is built upon the BioNeMo platform, which Nvidia created for AI applications in the fields of biology and chemistry.

The model is intended to help researchers move beyond simply discovering proteins that already exist in nature. Instead, it allows them to design entirely new ones from scratch. These custom proteins could be used to interrupt disease processes or detect signs of illness in new diagnostic tests.

According to the developers, the model allows scientists to search and refine designs using specific molecular inputs and scoring objectives. This means researchers can set specific goals for what a protein should do and then use the AI to find the best structure to achieve it. The project was led by three UW PhD students and a visiting student from the University of Oxford.

How the technology works

RFD4-Proteina combines two major technologies: the IPD’s RosettaFold Diffusion and Nvidia’s Proteina-Complexa. RosettaFold Diffusion is a guided diffusion model used for generating new proteins, while Proteina-Complexa is a generative model for designing protein binders. The new model integrates these to create a more powerful tool for atomistic protein design.

One of the model's key technical features is its ability to refine a protein’s sequence and structure at the same time. It also accounts for the structures of interacting molecules simultaneously. This means that if a researcher makes a change to one part of the system, the model can effortlessly update the rest of the structure to match.

In lab bench tests, the model outperformed previous baselines across four different benchmark suites. These included protein binding, small-molecule binding, enzyme scaffolding, and DNA binding. The results showed that RFD4-Proteina provides a significant leap in both the speed of generation and the overall success rate of the designs.

Speeding up the lab process

The model has demonstrated significant improvements in efficiency for researchers. Developers reported that it achieves five-fold to 10-fold higher design success rates while generating candidates five to 10 times faster than previous models. This allows scientists to explore and prioritize many more candidates for laboratory testing in a shorter amount of time.

For example, a UW PhD candidate reported using the model to precisely control the types of glycans, which are sugars important for immune-system recognition, attached to proteins. She noted that she was able to achieve this in under one week, whereas previous models would have taken months to complete the same task.

Another researcher used the tool to design a pesticide-degrading enzyme that performed many times better than his previous best. That previous design had taken him months to complete, but the new AI-assisted design was finished in just a few days. These examples highlight how the model can compress months of research into a matter of days.

Broad applications and access

The developers believe RFD4-Proteina will have applications far beyond medicine. Nobel laureate David Baker, the director of the IPD, suggested the model could create enzymes that break down pesticides or help in making bioplastics. It could also help find new ways to break down polymers, which are large molecules made of repeating subunits.

To ensure wide adoption, the model is being released as an open-source resource under the BSD-3 license. It is accessible through a variety of partners, including Amazon Bio Discovery, Anthropic’s Claude Science, Benchling, Databricks, Google Cloud, Hugging Face, Nebius, OpenAI, and Tamarind Bio. These partnerships allow researchers to integrate the model into various cloud and scientific computing workflows.

The project received significant funding from various organizations, including the Howard Hughes Medical Institute and the Gates Foundation. The Gates Foundation provided a grant of $25,748,619 with a date of November 2025. Additionally, a computational grant from the Jen-Hsun and Lori Huang Foundation was essential to the development of the model.