Baker Lab and NVIDIA Release RFD4-Proteina with 84% Binder Success Rate

The Baker Lab at UW's Institute for Protein Design, together with NVIDIA's GenAIR team after a year of collaboration, released RFD4-Proteina (RFdiffusion 4), a protein design model built around latent reasoning for biomolecular design. Authors include David Baker, Frank DiMaio, and Hannes Staerk. The model also demonstrates agent-style tool calling for a variety of design tasks.

Confirmed

• The model was developed over one year by the Baker Lab/IPD and NVIDIA's GenAIR team; the paper is released with authors including David Baker, Frank DiMaio, and Hannes Staerk.
• The core method performs molecular design via latent-space reasoning.
• Per data disclosed in Hannes Staerk's reading-group preview: 158 of 188 designed binders for the TrkA target bound below 1 µM, versus 18 of 60 for the next-best method—roughly an 84% hit rate.
• Per examples shared by JacobMolBio, the model supports tool calling, completing tasks such as inserting an ALFA tag into GFP and iterative design.

Why it matters

• The results substantially exceed the success rates of existing binder-design methods at the experimental-validation level, marking a major step toward practical AI protein design.
• The introduction of latent reasoning and agent tool calling enables conversational, multi-step completion of complex design tasks, expanding the interaction paradigm of protein design models.

2026-11-06 ~ 2026-11-07 · 5 related posts

• Baker Lab + NVIDIA unveil RFD4-Proteina: agents run protein design tool calls on GFP — AllThingsApx · 2026-11-06
• [source] RFD4-Proteina hits 158/188 binder success rate, crushing prior protein design methods — HannesStaerk · 2026-11-06
• RFD4-Proteina binders: 158 of 188 designs bound TrkA under 1 µM, vs 18 of 60 for best alternative — AllThingsApx · 2026-11-06
• [source] RFdiffusion protein binders: 158 of 188 designed binders bind TrkA below 1 µM — HannesStaerk · 2026-11-07
• Baker Lab and NVIDIA unveil RFD4-Proteina for biomolecular design via latent reasoning — AllThingsApx · 2026-11-07