NVIDIA Debuts Open Protein Design Model RFD4-Proteina with UW Medicine

NVIDIA on October 21, 2026, announced a set of healthcare and life sciences collaborations at GTC Berlin, led by the debut of RFD4-Proteina, an open protein-design model developed with the UW Medicine Institute for Protein Design, and a new adaptive radiotherapy collaboration with Siemens Healthineers.

NVIDIA framed the announcements around two challenges facing the field: life sciences researchers must search vast biological design spaces and shorten the cycle between computational design and laboratory testing, while cancer care teams must translate complex imaging and treatment data into precise plans for each patient. Kimberly Powell, vice president of healthcare at NVIDIA, said agentic and physical AI are opening new possibilities across medicine, spanning protein design, autonomous laboratories and more personalized cancer care.

RFD4-Proteina is an NVIDIA BioNeMo open model that combines programmable protein design with search capabilities. It generates targeted protein designs, enabling scientists and agents to search and refine candidates using molecular inputs, constraints and scoring objectives. NVIDIA listed potential applications including proteins designed to interrupt disease processes or to detect signs of disease in diagnostic tests.

In testing described in the research draft, NVIDIA reported, RFD4-Proteina achieved higher computational design success rates and generation throughput than the evaluated baselines across four benchmark suites: protein binding, small-molecule binding, enzyme scaffolding and DNA binding. The company said the model helps researchers explore and prioritize candidates for laboratory testing, where experiments establish how a design performs under tested conditions.

David Baker, 2024 Nobel Laureate in chemistry and director of the Institute for Protein Design at the University of Washington, described the work as a shift from discovering the proteins nature has made to designing entirely new ones. “RFD4-Proteina expands what scientists can create, from proteins that could advance drug discovery to enzymes that break down pesticides,” he said, adding that making the model open will allow researchers around the world to build on it.

NVIDIA said the model is available through partners including Amazon Bio Discovery, Anthropic’s Claude Science, Benchling, Databricks, Google Cloud, Hugging Face, Nebius, OpenAI and Tamarind Bio, enabling researchers and developers to integrate it into agent, cloud and scientific computing workflows.

Model Architecture, Licensing and Availability

A model card published on Hugging Face by the Baker Lab at the Institute for Protein Design and NVIDIA GenAIR describes RFD4-Proteina as a flow-matching model for biomolecular design that generates protein sequences and all-atom structures together, conditioned on the molecular context a user wants to preserve: a binding target, an enzyme motif, a ligand or selected structural features. The released weights carry 435.6 million parameters, are stored in FP32 at 1.74 GB and run in BF16 under a BSD 3-Clause license.

The card dates the weights to an October 19, 2026 release and states that support for them was merged into the Diffusers library on October 20, 2026. Its worked examples — including a PD-L1 protein binder, a HER2 antibody redesign, an enzyme catalytic motif and a small-molecule binding pocket — demonstrate the conditioning workflow, and the card states they are not experimental binding or catalytic-activity results. On a single NVIDIA H200, the card reports, a 250-token design takes 5.0 seconds in BF16, or 2.8 seconds after regional compilation.

A model card for the NVIDIA NIM deployment describes RFD4-Proteina as a unified generative model produced by NVIDIA’s Generative AI team and the David Baker Lab at the Institute for Protein Design, serving seven design tasks: generation, motif scaffolding, enzyme design, small-molecule binder design, antibody design, partial diffusion and inverse folding. The deployment supports reward-guided beam search over the diffusion trajectory, so a request can optimize against locally computable metrics rather than returning a single unranked sample. The card lists an October 20 release on build.nvidia.com and NGC at GTC Berlin, states that the weights were trained by the Institute for Protein Design with the training-data composition undisclosed, and marks the model ready for commercial or non-commercial use.

Siemens Healthineers Adaptive Radiotherapy Collaboration

Siemens Healthineers and NVIDIA are working to develop the next generation of adaptive radiotherapy with agentic and robotic intelligence, combining imaging and oncology expertise with accelerated computing and simulation. NVIDIA Nemotron open foundation models are being explored to power agentic co-planning, with the aim of helping oncologists evaluate and refine personalized treatment strategies. According to the release, the models could then be combined with CUDA-accelerated dose calculation for rapid patient-specific simulation and the NVIDIA Holoscan platform, which connects live sensor data with AI and simulation to support adaptive treatment.

Approximately 35 million people worldwide are projected to be diagnosed with cancer each year by 2050, the release states, and care teams face growing pressure to deliver precise, personalized treatment while managing limited specialist capacity. “Adaptive radiotherapy is built on a single principle: treatment should adapt to the patient, not the other way around,” said Arthur Kaindl, head of Varian, the Siemens Healthineers cancer-care business. He pointed to the launch of the Accela system as the company’s most recent advance in precision and adaptation.

Scientific Agents and Hospital Operations

The release puts global scientific R&D at $3.8 trillion, with annual pharmaceutical budgets approaching $300 billion, and presents scientific agents and autonomous laboratories as a way to connect design, experimentation and learning in a continuous cycle.

Research and laboratory technology companies are using NVIDIA BioNeMo, an open development platform for AI-driven biology and drug discovery, alongside Nemotron and other NVIDIA technologies to build scientific agents and the specialized models and tools connecting drug design, scientific data and laboratory workflows. Benchling is grounding agents in experimental data, research context and connected scientific workflows, while Biolevate is connecting biological data, specialized tools and agentic reasoning across drug discovery and development. Boltz is developing specialized models that agents can use to predict molecular structures and design and evaluate candidates, and is previewing a scientific agent that plans workflows and interprets results. Automata is connecting scientific agents with automated laboratory infrastructure so they can initiate experiments and capture structured, contextualized results.

Genedata’s AI co-scientist will coordinate protein design and testing in the protein design foundry within Danaher’s autonomous lab, in a workflow that will incorporate RFD4-Proteina so experimental results guide the next round of designs. Hologen is combining its large medicine models with NVIDIA Nemotron to build AI clinical scientists that turn imaging and clinical data into insights about patient outcomes to inform clinical trial design, and Therna Biosciences is using large language model reasoning to connect experimental evidence and RNA medicine design.

On the clinical operations side, the release states that healthcare systems across Europe are working to expand access to specialized care while reducing administrative demands on care teams. Avelios, Corti, Dedalus, Dawn Health and Sword are building on NVIDIA Nemotron, with applications ranging from AI scribes and AI-assisted clinical care to personalized patient support services. PolyAI is building on the NVIDIA Agent Toolkit in a multilingual dialog deployment with medical software provider Corilus, whose network spans more than 70,000 healthcare professionals across Belgium and France. Apian and VitVio are using NVIDIA Cosmos world models and NVIDIA Omniverse libraries to build digital twins that monitor and improve hospital operations.