← Back to Home

NVIDIA and University of Washington Open Protein Design Model to Bring Molecular Ideas into the Lab Faster

RFD4-Proteina incorporates protein sequences, shapes and interactions into its design process. The research team reports improved computational performance and some experimental validation; whether this can translate into new drugs remains subject to much longer biological and clinical testing.

By SURL BioNews

Designing a protein that appears capable of binding to a disease target and actually producing a molecule with the intended function often requires repeated revisions and experiments. NVIDIA and the Institute for Protein Design at the University of Washington School of Medicine jointly introduced RFD4-Proteina on October 21, aiming to shorten this back-and-forth between computers and lab benches so researchers can more quickly identify designs worth testing.

Announced at the GTC conference in Berlin, the model combines the two parties’ existing protein generation technologies. Researchers can specify structural, sequence and chemical constraints, then have the model generate, score and refine candidate molecules. The University of Washington explained that it can simultaneously adjust a protein’s amino acid sequence, three-dimensional structure and the structures of molecules that interact with it; for more difficult problems, it can also devote more computation to searching for solutions.

The reported performance gains currently come mainly from the research team’s tests. According to the University of Washington, RFD4-Proteina achieved five to ten times the design success rate and candidate protein generation speed of previous models on computational benchmarks. NVIDIA said the draft research paper evaluated four types of tasks: protein binding, small-molecule binding, enzyme scaffold design and DNA binding. These figures reflect performance under specific test conditions and cannot be directly translated into success rates for new drug development.

Laboratory results offer evidence beyond computational scores. The University of Washington said some of the designed proteins have demonstrated their intended functions in experiments, including enzymes that break down pesticides and designs that control the types of sugar chains attached to proteins. However, the announcement did not provide complete sample sizes or experimental success rates for each task, and the researchers’ individual results still need to be reproduced by more teams.

The model is open source under the BSD 3-Clause License, allowing other research teams to use and extend it. NVIDIA is also working to connect design with automated experimentation: according to its announcement, the AI research assistant from Genedata, a Danaher company, will coordinate protein design and testing in automated laboratories and use experimental results to guide the next round of design. The value of such workflows lies in feeding evidence from each test back into the design process, rather than merely increasing the number of candidate molecules.

A collaboration with Siemens Healthineers announced the same day shifts the focus to cancer radiation therapy. The two companies are exploring the use of Nemotron models to assist treatment planning discussions, alongside CUDA-accelerated dose calculations and the Holoscan platform connected to real-time sensor data, to support radiation therapy that adjusts to a patient’s condition. This remains a development direction; the announcement provided no clinical results showing that this integrated approach improves patient outcomes.

From protein design to radiation therapy planning, these collaborations share a common goal: shortening the distance between data, computation and practical implementation. But a protein that functions in experiments still faces hurdles involving its activity in the body, safety and manufacturing before it can become a drug; treatment planning tools must also demonstrate clinical reliability. The more concrete advances for now are open access to research tools and experimental support for some molecular designs. How much patients will benefit remains unanswered.

References

  1. marketscreener.com
  2. NVIDIA Newsroom
  3. UW Medicine Newsroom