New AI model designs proteins 10x faster for medicine and sustainability

A leap in protein engineering

The University of Washington (UW) Medicine Institute for Protein Design (IPD) has introduced a new AI model called RFD4-Proteina. Developed in partnership with Nvidia, this tool is designed to create new proteins with significantly higher efficiency than older systems.

Proteins are the workhorses of biology, performing everything from fighting infections to breaking down waste. While scientists have long tried to engineer custom proteins, the process has traditionally been slow and often resulted in designs that failed to function in real-world tests. RFD4-Proteina aims to change that by offering a more reliable path from a digital concept to a physical molecule.

As reported by Hoodline, the model is capable of generating a wide variety of biological structures. These include antibodies for medicine, enzymes for industrial use, and even proteins designed to break down environmental toxins.

How the technology works

The model achieves its speed by merging two distinct computational lineages. It combines the IPD's RosettaFold Diffusion line with Nvidia's Proteina-Complexa. This allows the AI to co-design protein sequences and their full atomic structures at the same time, rather than treating them as separate steps.

To do this, the system uses a method called reward-guided inference-time search. This technique helps the AI navigate the massive number of possible protein shapes to find the ones most likely to work. In technical terms, Nvidia noted that the model cuts diffusion sampling—the process of refining a design—from roughly 400 steps down to just 60.

The result is a significant jump in performance. In computational benchmarks, the researchers reported that the model produced protein candidates five to 10 times faster than previous models, while also showing fivefold to tenfold higher success rates. Lab tests have already confirmed that the proteins produced by the AI behave as expected when tested in a physical environment.

From theory to practical use

The practical benefits of this technology are already being seen in the lab. For example, a predoctoral candidate at the university used RFD4-Proteina to design a pesticide-degrading enzyme. The new design performed many-fold better than his previous best and was completed in just a few days.

The institute, led by 2024 Nobel Chemistry laureate David Baker, intends for this tool to be a foundation for several industries. The research is focused on creating precision cancer therapeutics, durable vaccines, and new materials. Because the model is open-source, scientists worldwide can access it to prioritize which candidates they should move into expensive laboratory testing.

The project was fueled by a major infusion of computing power. The Jen-Hsun and Lori Huang Foundation provided an in-kind grant totaling nearly 7 million hours of compute time. This massive amount of processing power allows the team to run the complex simulations required to train and refine these high-performance models.

The road ahead for protein AI

While RFD4-Proteina is a major milestone, it is not intended to replace physical experiments. Instead, it serves as a filter to help researchers identify the most promising designs before they spend time and money in a lab. Nvidia clarified that the goal is to help scientists explore and prioritize candidates rather than replacing experimental validation entirely.

The Institute for Protein Design has a long history of turning these types of models into real-world products. They have already launched more than ten spinout companies that have raised over $2 billion. These companies include vaccine developers and protein-design startups, showing a clear path from academic AI research to commercial biotechnology.

The team plans to release more AI models and research data through late 2026 and into 2027. As the technology matures, it will likely become a standard tool for anyone looking to engineer new biological solutions for medicine, sustainability, and materials science.

Technical specifications

For those interested in the underlying architecture, the model is quite large. It features approximately 435.6 million parameters across a paired bundle. This includes a 178-million-parameter trunk and a 256-million-parameter autoencoder.

The model is currently hosted through Berkeley Software Distribution and is available under the BSD-3 license. Researchers can interact with it through several different interfaces, including a Hugging Face release, a Python API, and command-line tools designed for generating, evaluating, and training new protein structures.

Data and transparency

While the model's capabilities are well-documented, some specifics remain private. Nvidia's model card notes that the exact composition, count, and content of the training data have not been disclosed. This is common in many AI releases where the specific datasets used to train a model are kept confidential by the developers.

Despite this, the track record of the UW Medicine Institute for Protein Design provides significant weight to the launch. By making the model open-source, they are providing the global scientific community with a high-powered engine for biological discovery.