← Back to Home

Designing Sequence and Structure Together: NVIDIA and University of Washington Open-Source Protein AI Model

RFD4-Proteina seeks to shorten the path from molecular concept to experimental validation, with applications spanning drug discovery and pesticide-degrading enzymes. The computational performance gains reported by the team must still be interpreted separately from how the proteins perform in the laboratory and in humans.

By SURL BioNews

Creating proteins that can bind disease targets or break down environmental pollutants takes more than drawing an elegant molecular structure; the amino acid sequence, folded shape, and interaction partners must all work together. On October 21, the Institute for Protein Design at the University of Washington School of Medicine and NVIDIA released the open-source model RFD4-Proteina, incorporating these requirements into a single design process in the hope of helping researchers find candidate molecules worth taking into the laboratory more quickly.

Announced at the GTC conference in Berlin, the model combines technologies from the institute’s RFdiffusion series and NVIDIA’s Proteina. It can jointly adjust a protein’s sequence, three-dimensional structure, and the structures of molecules that interact with it, allowing a change in one part to prompt corresponding adjustments elsewhere. Researchers can also specify molecular constraints and scoring objectives to guide the model’s search and devote more computation to harder problems.

Specific applications include designing molecules that bind particular proteins or small molecules, structures that support enzyme active sites, and proteins that bind DNA. NVIDIA said the research draft achieved higher computational design success rates and candidate generation efficiency than the comparison models tested across all four benchmark categories. The University of Washington reported success rates five to ten times those of previous models, with generation speeds also five to ten times as high. These are computational testing metrics and cannot be directly translated into the probability of success in drug development.

There are also preliminary experimental results. According to researchers’ accounts published by the University of Washington, the team designed pesticide-degrading enzymes within days that outperformed the best previous designs, which had taken months to obtain. Another effort controlled the types of sugar chains attached to proteins in less than a week. Sugar chains play a role in immune recognition, and precisely arranging these modifications is an important challenge in protein engineering. However, these timelines come from the team’s experience with individual applications and do not yet indicate that all design tasks can be accelerated to the same extent.

The open-source release also extends beyond a single model. Rosetta Commons explained on October 23 that the model had been made publicly available through its GitHub organization and integrates sequence and structure generation using a latent-space approach. The accompanying tool TMol provides GPU-accelerated design scoring and structural relaxation, while AtomWorks supports training and inference, giving other teams an opportunity to further develop, test, and modify the entire workflow.

The progress supported by the evidence so far is more efficient generation of candidate designs, with some designs demonstrating their intended functions in experiments. Public news materials do not provide complete sample sizes, success rates, or testing conditions for each experiment, and the results cited by NVIDIA are still presented in a research draft. For medical applications, binding a target is only the starting point; protein stability in the body, immune responses, safety, and actual efficacy remain questions for subsequent research to address individually.

References

  1. Genetic Engineering and Biotechnology News
  2. UW Medicine
  3. NVIDIA Newsroom
  4. Rosetta Commons