Loading page…
RFD4-Proteina directs extra computing toward promising designs. Its computational scores help select candidates, but do not by themselves prove that a protein works.
Listen to this story
Released on October 21, 2026, RFD4-Proteina lets researchers define a target, fixed molecular regions and success criteria, then use additional inference-time search to pursue promising designs. NVIDIA and the University of Washington report stronger results across four benchmark suites, while UW also cites laboratory validation and a researcher’s improved enzyme design. For research teams, the practical change is a more integrated design-and-search workflow; scores remain proxies, and experiments are needed to establish whether a candidate works.
NVIDIA says sampling takes 60 steps, down from about 400 in Proteina-Complexa, and reports faster 200-residue generation with lower peak memory than RFD3.
NVIDIA reports higher computational design success rates and throughput across four benchmark suites, attributing the results to testing described in a research draft.
The tutorial’s PD-L1 example rewards structural confidence, not proof of binding; interaction metrics and laboratory tests remain necessary.
Scientists have a new tool for designing proteins and refining promising candidates before making them in a laboratory. NVIDIA and the University of Washington’s Institute for Protein Design released RFD4-Proteina on October 21, 2026. The open model generates a protein’s sequence and full atomic structure together, then can spend extra computing power searching for better designs.
The starting point is a scientific objective, not an unrestricted request for a new molecule. Researchers specify molecular inputs, the parts that must remain fixed, the regions that can change and the criteria used to judge success. Those constraints steer generation toward a particular task.
RFD4-Proteina combines the institute’s RosettaFold Diffusion lineage with NVIDIA’s Proteina-Complexa approach. It produces the sequence and structure jointly, removing a separate step in which another model would assign a sequence to a generated structure. UW says it also refines the structures of interacting molecules alongside the protein.
The optional search loop adds another layer. During generation, the system scores designs, keeps promising partial results and branches from them. External scoring tools can evaluate geometry, molecular interactions or confidence in a predicted structure. This is reward-guided inference-time search: extra computation spent pursuing a chosen objective while the model runs.
NVIDIA’s tutorial illustrates the process with candidates designed to bind PD-L1, an oncology target. Researchers fix the target’s sequence and backbone, allow some atomic arrangements to change and identify contact points for the new protein. An agent can coordinate the tools and record results, while scientists define objectives and use experimental evidence to guide subsequent rounds.
NVIDIA reports 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. Its announcement attributes those results to testing described in a research draft.
The technical post gives a narrower generation comparison against RFD3. NVIDIA attributes the efficiency gains partly to representing proteins at the residue level rather than atom by atom internally. It also says diffusion sampling—the iterative process used to generate a design—falls from about 400 steps in Proteina-Complexa to 60.
Those computational results are distinct from experimental performance. UW says laboratory tests validated model outputs across tasks of varying size, complexity and function. One researcher, Seth Woodbury, said he designed a pesticide-degrading enzyme in days that performed many times better than his previous best, which had taken months to design.
A high search score still needs careful interpretation. NVIDIA’s example rewards confidence in the designed protein’s structure, not evidence that it binds the target. The tutorial explicitly directs users to inspect interaction and structure-checking metrics separately. Laboratory experiments establish how a candidate actually performs under the conditions tested.
NVIDIA describes two main ways to use RFD4-Proteina, with different levels of control and setup:
NVIDIA also lists availability through partners including Amazon Bio Discovery, Anthropic’s Claude Science, Benchling, Databricks, Google Cloud, Hugging Face, Nebius, OpenAI and Tamarind Bio. Databricks describes day-zero support through Genesis Workbench, an open blueprint connecting biological models with input preparation, candidate evaluation and downstream analysis.
The partner post also calls the AI Runtime integration planned, even as it describes example workflows. Its proposed setup puts GPU execution beside research data, experiment tracking and governance. Researchers still review generated code, select scientific tools and interpret results; the workflow does not hand those decisions to the coding agent.
Loading discussion...
Join the conversation
Explain why deeper search would or wouldn’t beat exploring more options.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.