RFD4-Proteina: NVIDIA and Baker Lab Open-Source Protein Design Model

Episode 1 · NVIDIA and Baker Lab Release Open-Source Protein Design Model RFD4-Proteina (2026-10-21, 65 posts)

After roughly a year of collaboration, NVIDIA and the University of Washington's David Baker lab (Institute for Protein Design, a Nobel laureate team) jointly released the open-source de novo protein design model RFD4-Proteina at GTC Berlin on October 21, open-sourced under a BSD license (m9 mentioned BSD-3). The model combines UW's RFdiffusion with the lineage of NVIDIA Health's Proteina models and is regarded as a major advance in protein design.

Confirmed

• The model was co-developed by NVIDIA's BioNeMo / GenAIR / Health teams and UW's Institute for Protein Design; the paper is titled "Biomolecular design through latent reasoning…"
• A single model starts from a shared latent space and jointly generates protein sequences and all-atom structures, unifying design tasks spanning proteins, enzymes, binders, antibodies, peptides, and DNA binding
• The preprint reports surpassing evaluated baselines in both computational design success rate and generation throughput across four benchmark suites: protein binding, small-molecule binding, enzyme scaffolds, and DNA binding
• Official numbers: on 24 binding targets, 76.40 designs passed per GPU hour, versus only about 13 for the next-best method
• Available day-0 via the Tamarind platform through the NVIDIA Health team, and also accessible via Amazon Bedrock (m5)
• The official demo includes generating binding proteins around hotspot residues of the PD-L1 target

Why it matters

• Merging the RFdiffusion and Proteina lineages into a single model covers the full spectrum of protein design tasks, simplifying workflows that previously required multiple specialized models
• The high throughput (dozens of approved designs per GPU hour) can significantly shorten the path from candidate molecules to lab validation, accelerating drug and therapeutic development
• Fully open-sourced under a BSD license, lowering barriers for both academia and industry; vendors like Databricks also covered the release on launch day

• RFD4-Proteina combines RFdiffusion and Proteina lineages for frontier protein design — AllThingsApx · 2026-10-21
• NVIDIA and UW release open protein-design model RFD4-Proteina at GTC Berlin — AllThingsApx · 2026-10-21
• NVIDIA and David Baker's lab open-source RFD4-Proteina protein design model — AllThingsApx · 2026-10-21
• Baker Lab and NVIDIA launch de novo protein design model RFD4-Proteina — AllThingsApx · 2026-10-21
• NVIDIA and UW's open-source RFD4-Proteina designs antibodies, enzymes and peptides in one model — AllThingsApx · 2026-10-21
• RFD4-Proteina: one open model for protein sequence and all-atom design, 76 passing designs per GPU-hour — AllThingsApx · 2026-10-21
• Baker Lab and NVIDIA Unveil RFD4-Proteina, a Faster Protein Design Model That Beats Prior Methods — jueseph · 2026-10-21
• RFD4-Proteina beats baselines on design success rate across four benchmarks — AllThingsApx · 2026-10-21
• NVIDIA and David Baker's lab debut open protein-design model RFD4-Proteina — AllThingsApx · 2026-10-21
• NVIDIA and UW's Institute for Protein Design release open protein design model RFD4-Proteina — AllThingsApx · 2026-10-21
• IPD and NVIDIA release RFD4-Proteina, a generative model spanning proteins, DNA and RNA — AllThingsApx · 2026-10-21
• NVIDIA and Nobel laureate David Baker's lab release open protein design model RFD4-Proteina — AllThingsApx · 2026-10-21
• NVIDIA and UW release RFD4-Proteina, unifying sequence and all-atom protein design — AllThingsApx · 2026-10-21
• Diffusers merges RFD4-Proteína, a protein design flow-matching model from Baker Lab and NVIDIA — RisingSayak · 2026-10-21
• NVIDIA details RFD4-Proteina: reward-guided inference-time search speeds protein design 3.8x — AllThingsApx · 2026-10-21
• First protein design model RFD4-Proteina merges into Diffusers — RisingSayak · 2026-10-21
• NVIDIA's RFD4-Proteina lands on HuggingFace, outdesigning rivals 5.8x per GPU-hour — AllThingsApx · 2026-10-21
• Baker Lab and NVIDIA unveil RFD4-Proteina, a latent-reasoning protein design model — AllThingsApx · 2026-10-21
• Baker Lab and NVIDIA open-source RFD4-Proteina for generative protein design — AllThingsApx · 2026-10-21
• RFD4-Proteina unifies proteins, DNA, RNA and sugars in one AI design space — AllThingsApx · 2026-10-21

45 more related posts →

Episode 2 · RFD4-Proteina Open-Sourced: Big Gains in Protein Design Pass Rates and Throughput (2026-10-21, 9 posts)

AllThingsApx released and open-sourced a new protein design model, RFD4-Proteina, on October 21-22. Its high-thinking mode substantially outperforms existing methods across multiple benchmarks, with several hands-on testers calling it a "generational leap."

Confirmed

• On the DNA-binding protein design benchmark, RFD4-Proteina (high-thinking) achieved a 3.5% computational clustering pass rate versus 0.8% for the fixed-DNA approaches BoltGen/BoltzGen, covering three targets—pointing to improved gene-targeting tool design.
• On enzyme active-site scaffold design, it reached a 12.1% pass rate versus 5.7% for AlphaProtein-novo codesign, nearly double; the authors say it can produce thousands of designs per GPU hour.
• For protein-protein binder design, it delivers 42x the generation throughput of BindCraft, a single-backbone method in the benchmark, with higher computational clustering pass rates; the authors frame this as a jump from "folding-based design" to "molecular recognition" design.
• The model is live on Hugging Face, accompanied by NVIDIA-related resources from AllThingsApx (per post m5).
• Community adoption was rapid: a protein designer said they will use RFD4-Proteina in the Protein Design Competition co-hosted by Adaptyv and Anthropic; CotetTudor also recommended benchmarking it in that competition right away.
• Tester RafiBrent said he was initially skeptical, but after testing considers RFD4-Proteina a "generational leap" in computational biomolecular design, with his posted data showing binder design pass rates far exceeding existing methods.

Why it matters

• Simultaneous large gains in pass rate and throughput mean protein design pipelines previously bottlenecked by compute and success rates could speed up dramatically, with direct implications for gene-targeting tools, enzyme engineering, and therapeutic binder design.
• The model is open-sourced and entered a live competition on release day, showing the AI protein design community's move from model release to real-world application is shrinking dramatically.

All figures come from the releasers' and reposting testers' posts; no independent third-party reproduction has been seen yet.

• Protein designers adopt RFD4-Proteina for Adaptyv x Anthropic competition — AllThingsApx · 2026-10-21
• RFD4-Proteina benchmarks show step-change gains in protein binder design pass rates — AllThingsApx · 2026-10-21
• RFD4-Proteina designs binders for phospho-amino acids beyond the canonical 20 — AllThingsApx · 2026-10-21
• RFD4-Proteina designs phosphorylation binders, distinguishing adjacent histone sites — AllThingsApx · 2026-10-22
• RFD4-Proteina pitched as a benchmark tool for the Anthropic x Adaptyv protein contest — AllThingsApx · 2026-10-22
• RFD4-Proteina delivers 42x higher binder-design throughput than BindCraft — AllThingsApx · 2026-10-22
• RFD4-Proteina doubles enzyme scaffolding pass rate to 12.1% vs 5.7% — AllThingsApx · 2026-10-22
• RFD4-Proteina hits 3.5% DNA-binder pass rate vs 0.8% for BoltzGen — AllThingsApx · 2026-10-22
• RFD4-Proteina Goes Beyond Canonical Amino Acids, Distinguishing Adjacent Histone Sites — AllThingsApx · 2026-10-22

Episode 3 · RFD4 Replaces Self-Conditioning with Recurrent Updates, Echoing Looped-Flow 'Thinking' (2026-10-22, 2 posts)

The RFD4-Proteina team dropped long-used self-conditioning in favor of recurrent updates: the model keeps xt fixed while iteratively refining its hidden state, performing more compute per step with fewer total steps.

• Researchers break down RFD4-Proteina's CFM trick: recurrent hidden-state refinement, fewer ODE steps — AllThingsApx · 2026-10-22
• RFD4 team swaps self-conditioning for recurrent updates, echoing looped flows — AllThingsApx · 2026-10-22