A joint sequence-and-structure biomolecular design model from the Baker Lab and NVIDIA, with the test-time-compute vocabulary of language models applied to protein design. The model card describes RFD4-Proteína as "a flow-matching model for biomolecular design" that generates protein sequences and all-atom structures together, conditioned on context to preserve — a binding target, an enzyme motif, a ligand, or prescribed structural features. It is distributed as a paired bundle of a flow model (with a rank-32 LoRA fine-tune merged in) and a unified autoencoder covering proteins, nucleic acids and small molecules, 435,622,465 stored parameters in total, under BSD 3-Clause. The autoencoder encodes each standard or modified amino acid, sugar or nucleotide as an anchor-atom coordinate plus a 32-dimensional latent vector over a 147-class residue vocabulary (up from 16 dimensions and 33 classes in the previous release), with a Rosetta-energy loss during fine-tuning that "lowered mean Rosetta energy by 1.07 kcal/mol per residue and reduced severe bond-angle deviations by 62%."

The framing contribution is "reasoning" at two levels. Intrinsic reasoning is recurrent refinement inside each denoising step — recycling moved from an upstream feature stage into every diffusion timestep, with a variable number of passes so compute concentrates where the trajectory needs it. Extrinsic reasoning is guided search across trajectories against arbitrary user metrics, exploiting the paper's observation that design objectives resolve at different times during generation: interface area becomes predictable at roughly 20% of the trajectory, hydrogen bonding at 50%, AlphaFold 3 interface confidence at 80%. Pruning trajectories against each criterion as it resolves "yielded three times as many unique passing designs per GPU-hour" than naive sampling on the paper's bond-cleaving enzyme task. The latent formulation extends La-Proteina and Proteina-Complexa; RFdiffusion3 is the all-atom predecessor that, as the paper puts it, "does not design high-quality sequences for its generated structures."

In silico, the gains are in pass rate and throughput at the same time. On clustered pass rate (distinct passing structural clusters per 200 attempts), high-thinking RFD4-Proteína reports 10.6% on protein binding versus 5.9% for BindCraft, at 42-fold higher throughput; 12.1% versus 5.7% for AlphaProtein-novo codesign on the Atomic Motif Enzyme benchmark (5.9× throughput); 14.3% versus 2.4% for RFD3 on small-molecule binding, with passing designs for all 77 targets versus 64; and 3.5% versus 0.8% for BoltzGen on the three-target DNA benchmark. Counting AlphaFold 3 evaluation time, that is 5.0–10.4× more passing clusters per GPU-hour than the next-best baseline per task. The card cautions that these numbers were measured on the superseded 2026-10-15 pair against an October 9 benchmark snapshot and "have not been re-measured on the current release," and that computational passers do not establish experimental activity.

The wet-lab results are what separate this from a benchmark paper. Of 18 conformation-selective calmodulin designs, 12 bound the Ca2+-bound form with apparent KD < 1 µM and no detectable binding without calcium, affinities spanning 69 pM to 37 nM. Binders to phosphorylated epitopes reached KD = 112 pM (YAP1) and 527 pM (Tau). Against human TrkA, 158 of 188 designs were sub-micromolar (84%), versus 18/60 (30%) for the prior campaign. Four of 67 designed ligases showed amide-bond formation where an RFD3 campaign got none from 72, the best converting 79% of the (S)-amine with 8.8% for the (R)-enantiomer — which the authors call the "first proof-of-concept demonstration of a de novo designed enzyme catalyzing a step in active pharmaceutical ingredient (API) synthesis."

The preprint — "Biomolecular design through latent reasoning with RFD4-Proteina," CC BY 4.0, not peer reviewed — is corresponded by Kieran Didi (NVIDIA) and David Baker, with title-page affiliations led by the Institute for Protein Design and Department of Biochemistry at the University of Washington and NVIDIA, plus HHMI, Oxford, Manchester, Southampton, ETH Zurich, Tsinghua and Duke. Training compute for the released model is not disclosed; the paper states only that "all ablation runs—and baselines—are trained on 16 H200 NVIDIA GPUs." A Diffusers pipeline was merged on 2026-10-20 and ships in the next Diffusers release.

Model Details

Parameters 436M
License BSD 3-Clause

Benchmark Scores

Benchmark Score Mode
Protein binder clustered pass rate (high thinking) 10.6% (BindCraft 5.9%) —
Atomic Motif Enzyme clustered pass rate 12.1% (AlphaProtein-novo 5.7%) —
Small-molecule binder clustered pass rate 14.3% (RFD3 2.4%) —
DNA binder clustered pass rate (3 targets) 3.5% (BoltzGen 0.8%) —
Human TrkA experimental hit rate (sub-micromolar) 158/188 (84%) —
sciencebiologyproteinopen-weightdiffusionreasoning

Related