NVIDIA and UW's Institute for Protein Design open RFD4-Proteina

The open stake this week is a shared chemical-conformational latent that co-designs sequence and all-atom structure across amino acids, sugars, DNA, and RNA. Baker's lab keeps saying design; the artefact labs can install is RFD4-Proteina.

Between 20 and 21 October 2026, the RFD4-Proteina project site, UW Medicine's newsroom, and NVIDIA's healthcare newsroom pointed at the same release: an open BioNeMo-line model from the Institute for Protein Design with NVIDIA co-authors, installable from PyPI and Hugging Face Diffusers.

Shared latent reasoning joins chemistry, conformation, and search

RFD4-Proteina's novelty claim is a unified 32-dimensional chemical and conformational latent space supervised with force-field energies, plus intrinsic reasoning (recurrent refinement that allocates compute) and extrinsic reasoning (inference-time search steered by geometric or energy scores). The landscape covers standard and modified amino acids, sugars, DNA, and RNA in one generator. Correspondence is listed to Kieran Didi (NVIDIA) and David Baker (UW).

The open-release stake for agents and labs is practical: pip install "rfd4-proteina[all-cuda13]", Diffusers RFD4ProteinaPipeline, and annotated CIF conditioning, without waiting for a closed API.

In silico clustered pass rates and a wet-lab binder panel

In silico, the site reports clustered AlphaFold3 pass rates against generation throughput on protein binding (24 targets), small-molecule binding (77), enzyme scaffolding (37), and DNA binding (3). High-thinking single-shot protein binding is listed at 10.6% clustered pass rate versus 5.9% for the next-best named baseline (BindCraft) at far higher designs-per-GPU-hour; small-molecule, enzyme, and DNA rows likewise claim higher pass rate and throughput than named comparators. Thinking presets (low/medium/high) raise pass rate as throughput falls.

On the wet-lab side, binder panels include calcium-dependent calmodulin designs (12 of 18 below 1 µM with no binding without calcium), peptide-MHC binders with over 100-fold on-target specificity, and disordered-peptide binders that include picomolar-class YAP1 and Tau figures, plus TrkA, glycan-shielding, and phospho-specific histone H3 designs.

Enzyme work on the same page covers improved phosphotriesterases and an enantioselective ligase for a lenacapavir synthesis step, with 79% conversion on the best design. Those figures are author-reported experimental panels. Multi-lab replications are still outstanding.

AF3 clustered passes reward diversity; BindCraft is the speed foil

Clustered pass rates count distinct folds, so repeating one success does not inflate the score. Throughput includes generation and, in some reported bars, AF3 evaluation cost. BindCraft and related hallucination methods are the accuracy-at-low-throughput foil; diffusion baselines are the throughput foil. Specialist tasks (side-chain packing, fixed-backbone sequence design, conformational sampling, design refinement) are claimed competitive without extra training. Fairness still depends on matching task definitions and AF3 thresholds to each baseline's original protocol. The project site is the primary table; no external meta-analysis is cited here.

Therapeutic efficacy, GMP process chemistry, and multi-lab replication stay open

The release does not establish clinical efficacy, regulatory-grade manufacturing routes, or independent wet-lab replication of every panel. Some affinity figures carry fitting-floor caveats on the site itself (for example the calmodulin 69 pM discussion). Agent toolchains that call the Diffusers pipeline still need wet-lab gates before any therapeutic story.

PyPI, GitHub, Hugging Face Diffusers, and BSD-style packaging

Public artefacts named on the site include rfd4-proteina on PyPI (CUDA 12 and 13 extras), RosettaCommons/RFD4-Proteina on GitHub, nvidia/RFD4-Proteina-Diffusers on Hugging Face, and prep utilities for annotated CIF. Licence framing on the open package is BSD-3-class per the discovery brief and project packaging; confirm the repository LICENSE file before redistribution. NIM card references appear in the 20 October NVIDIA ecosystem posts.

Open design loops get a co-design checkpoint; brochure success rates wait

A computational protein lab or science agent stack can install the Diffusers pipeline on a CUDA 12/13 GPU, run conditioned CIF generation, and keep its own AF3 and wet-lab filters. The newsroom success mosaic shows what the release enables in author labs; it still needs local validation before anyone pitches patient-facing binders from the launch figures alone.

RFD4 meets Proteina after RFD3 and closed binder campaigns

RFD4-Proteina sits in the Baker lab diffusion-design line (including prior RFD3 DNA-binding comparisons on the site) and NVIDIA BioNeMo co-design. It is adjacent to Anthropic's published Claude scorpion-binder wet-lab loop without sharing that toxin spine: here the open model and multi-chemistry latent are the story, and the wet-lab panel is supporting evidence for what the release enables.