One model where there used to be two
Nvidia and the University of Washington’s Institute for Protein Design have released RFD4-Proteina, an open model for designing proteins that generates the sequence and the all-atom structure in one pass.
That is the structural change. Protein design has usually run as a relay: one model proposes a backbone, a second model such as ProteinMPNN picks amino acids to fit it, and a third checks the result. RFD4-Proteina pairs a flow-matching model with an autoencoder that represents each residue, sugar or nucleotide as an anchor-atom coordinate plus a latent vector, and generates both together. The handoff disappears. The model can also design around molecular context the relay handled awkwardly: target proteins, small molecules, nucleic acids, glycans, metal ions and modified residues.
What the headline number counts
Across 24 protein-binding targets, Nvidia says RFD4-Proteina at its high “thinking” setting produced 76.40 structurally distinct designs per GPU-hour that met its pass criteria. The second-best method it tested, Proteina-Complexa, produced 13.24.
The criteria are worth reading, because they are what “passing” means here: a binder pTM of at least 0.80, a minimum interface PAE of 1.5 ångström or better, and a target-aligned binder backbone RMSD of 2.5 ångström or better. Those are scores a computer assigns to a computer’s prediction. Nothing in the figure says a designed protein was synthesised, folded as predicted, or bound anything.
Nvidia says so itself. The company’s press release notes that the broader results — higher computational success rates and throughput than baselines across protein binding, small-molecule binding, enzyme scaffolding and DNA binding — come from a research draft, and that lab testing is still needed to confirm them. The Hugging Face post repeats it: experimental validation is needed to confirm binding.
The “thinking” setting has three levels, low, medium and high, and controls how much the model refines each design as it generates. Higher settings raise the odds of passing the in-silico filters and cost more compute per design, which is the trade the headline figure is measuring.
Shipped through everyone at once
The distribution list is unusual. Nvidia says the model is available through Amazon Bio Discovery, Anthropic’s Claude Science, Benchling, Databricks, Google Cloud, Hugging Face, Nebius, OpenAI and Tamarind Bio — three competing frontier labs’ platforms on the same day.
A dedicated Hugging Face Space runs it on RTX PRO 6000 GPUs, with Nvidia sponsoring free access from 21 October to 4 November. After that the dedicated Space stops being free, and a ZeroGPU Space on shared capacity remains free subject to account quotas. Both Spaces run SecureDNA screening on inputs and outputs, with a local client that hashes and encrypts sequences before anything leaves the container.
What to watch
David Baker, who directs the Institute for Protein Design and shared the 2024 Nobel Prize in Chemistry, framed the release as a move “from an era of discovering the proteins nature has made”. Kimberly Powell, Nvidia’s vice-president of healthcare, tied it to the company’s wider agentic-AI push in medicine.
The number that will settle this is not 76.40. It is how many of those designs bind when someone makes them. Nvidia named Genedata, working inside Danaher’s autonomous lab, as a partner that will run an AI co-scientist coordinating protein design and testing with the model. That is where a wet-lab hit rate would come from, and until one does, the throughput figure describes how fast the model produces candidates, not how many of them work.
