On October 21, NVIDIA and the University of Washington's Institute for Protein Design released RFD4-Proteina, an open-weight biomolecular design model published with a preprint, inference code, weights and a benchmark explorer. Filed under another model release, it would be easy to skip. The technically consequential claims sit one level down: the model is built to be steered at inference time by an external scoring function, and it is evaluated on a frontier of design quality against GPU-hours rather than on quality alone. Both choices change what kind of system this is.
The architecture starts with representation. An autoencoder maps standard and modified amino acids, sugars, DNA and RNA into a shared 32-dimensional chemical and conformational latent space, trained against force-field energies so that decoded conformations are physically plausible. Generation happens in that space and is decoded back to atoms. One space for several chemistries is what allows a single set of weights to address protein binding, small-molecule binding, DNA binding and glycan recognition without task-specific heads. The benchmark spread below is a test of that claim as much as of raw performance.
The authors describe two modes of reasoning. Intrinsic reasoning is recurrent refinement: the model revisits and refines its own latent state, spending more computation where it helps, exposed to users as a low, medium or high thinking budget. Extrinsic reasoning is inference-time search steered by an arbitrary score, whether energetics, geometric criteria or a custom filter. The second is the one that changes the interface. Generation stops being a single draw and becomes the proposal step of an optimizer: propose, score, refine, repeat. A score is a number and a loop is a loop, which is precisely why an agent can drive it.
That also defines the failure mode. A search that maximizes a geometric score will find whatever maximizes that score, whether or not the resulting protein does anything useful. The evaluation design is aware of the problem. Designs are filtered against the intended structure and task-specific criteria, then re-folded with AlphaFold3, and only a refold that matches the intended structure counts as a pass. Pass rates are clustered, so reproducing the same successful topology repeatedly earns nothing. That is stricter than counting samples, and it removes the most common way design papers inflate numbers. It remains a structural proxy: a protein that folds as intended has not been shown to bind, catalyze, or avoid an off-target.
On the four in-silico suites, the reported numbers are a 10.6% clustered pass rate for protein binding against 5.9% for BindCraft, at 42-fold higher throughput; 14.3% for small-molecule binding against 2.4%; 12.1% for enzyme active-site scaffolding against 5.7%; and 3.5% for DNA binding against 0.8% for BoltzGen and 0.5% for fixed-DNA RFD3. Target coverage is arguably the more informative figure: 22 of 24 protein targets, all 77 small-molecule targets, 35 of 37 enzyme tasks and all three DNA targets. An average pass rate can hide a set of targets that never succeed; coverage cannot.
The compute trade-off is explicit. Moving from the lowest to the highest thinking setting lifts the protein-binding pass rate from 7.7% to 10.6% while throughput falls from 2,388 to 986 designs per GPU-hour. Roughly 2.4 times the compute for about 1.4 times the pass rate, before counting the cost of the downstream filter, which the figure captions include in a separate throughput measure. Whether that trade pays depends on the price of validation, not on the model: if every design must be refolded and filtered anyway, a cheaper sampler with more rejects is often the better pipeline. The load-bearing claim is therefore not the high setting but that the lowest setting already beats every baseline.
A separate result deserves attention. Without additional training, the same weights are reported to match or beat specialist methods on side-chain packing, fixed-backbone sequence design, conformational sampling and design refinement. If that holds up, it argues the latent space is general rather than tuned per task, and that the usual division of labor between a generator and a set of post-hoc fixers is partly redundant. The comparison is the authors' own, against named specialist tools, so it is a published result rather than an independent one.
Wet-lab validation is where the paper spends its credibility. Calcium-dependent calmodulin binders: 12 of 18 designs below 1 micromolar with no detectable binding without calcium. Peptide-MHC binders at 340 picomolar, 710 picomolar and 1.60 nanomolar, with over 100-fold preference for the on-target peptide where off-targets differ by one or two residues. Binders to intrinsically disordered peptides, including what the authors describe as the first de novo binders to the tested YAP1 and Tau epitopes, at 112 and 527 picomolar. For unconstrained human TrkA, 158 of 188 designs were sub-micromolar, alongside designs that discriminate mouse TrkA and the TrkB and TrkC paralogues. Glycan shielding: 8 of 9 designs above 90% oligomannose. An eight-residue phosphotriesterase active site reached a catalytic efficiency of 3,600 per molar per second, a 1.5x10^7-fold rate enhancement over uncatalyzed hydrolysis. And a designed ligase converted 79.0% of the (S)-amine in a lenacapavir synthesis step, against 8.8% of the (R)-amine.
The caveats are stated in the paper and worth repeating. The headline 69 picomolar calmodulin affinity comes from a 1:1 fit at the off-rate fitting floor, and the authors say the precise affinity and ranking remain unresolved. The phosphopeptide figures are response midpoints, not equilibrium affinities. And the distance between computational and experimental success is the number to watch: the lenacapavir ligase campaign found 4 of 67 designs forming product above background, about 6%, while in-silico filters passed 111 of 1,000 designs for the phosphotriesterase scaffold. Both are strong results by the standards of designed enzymes, and both are an order of magnitude below what a pass rate suggests if read carelessly. Hit rate, affinity and catalytic efficiency are three different measurements, and a single percentage cannot stand in for all of them.
For agent-driven workflows, the practical reading is that a model with a scoreable interface is a tool an agent can iterate against, but three things have to be fixed before the loop means anything: the scoring function, including what it rewards and what it ignores; the budget, now denominated in thinking level and GPU-hours; and the stop rule, which for anything that matters is still an assay.
On reproducibility, the release is more complete than most: preprint, code covering inference, reward-guided search, training and evaluation, hosted weights, and a benchmark explorer over the underlying data. That means the numbers can be re-run rather than trusted. One shared dependency is worth flagging: AlphaFold3 serves both as a filter during design and as the arbiter in evaluation, so any systematic bias in that predictor would appear on both sides of the comparison and would not show up as a discrepancy.
The honest summary is narrower than the headline. The shift is from generation as a sample to generation as a search step with an explicit budget and a pluggable objective. That framing says the binding constraint has moved from the sampler to the scoring function and the assay, and it is a more useful thing to know than another pass-rate record.
Sources
RFD4-Proteina preprint and project site: https://rfd4-proteina.org/ (paper: https://rfd4-proteina.org/assets/paper/RFD4-Proteina.pdf)
NVIDIA Newsroom, "Healthcare and Life Sciences Leaders Build on NVIDIA to Advance Drug Discovery and Expand Patient Care," October 21, 2026: https://nvidianews.nvidia.com/news/healthcare-and-life-sciences-leaders-build-on-nvidia-to-advance-drug-discovery-and-expand-patient-care
Code: https://github.com/RosettaCommons/RFD4-Proteina | Weights: https://huggingface.co/nvidia/RFD4-Proteina