An open protein-design model produced 67 candidate enzymes for one step in the synthesis of an HIV drug; four showed activity and the best converted 79% of the substrate.
What Was Done
According to a University of Manchester release (October 22, 2026) and Phys.org coverage, the University of Washington's Institute for Protein Design (IPD), led by David Baker, and Nvidia developed RFD4-Proteina, a unified generative model that combines the IPD's RosettaFold Diffusion line with Nvidia's Proteina-Complexa. Researchers at the Manchester Institute of Biotechnology (Anthony Green, Henry Jones and Zefei Xu) tested it on a concrete challenge: designing enzymes that catalyze an amide bond formation step in the synthesis of lenacapavir, an HIV prevention drug.
The Lab Numbers
The 4-of-67 rate is what makes the result honest: AI design doesn't deliver a finished enzyme, it delivers a few candidates to validate. Even so, the best enzyme also suppressed a competing reaction with water, a typical problem in this kind of catalysis. Green called the initial results "spectacular"; Baker said the model "will open the door to new solutions in medicine, sustainability, materials and biotechnology."
How to Read the Result
The 5-to-10x comparisons come from the authors' own computational tests, not an independent evaluation. The work is a preprint ("Biomolecular design through latent reasoning with RFD4-Proteina") and hasn't yet been peer reviewed. The model is open source under a BSD-3 license, so third parties can reproduce it.
What It Means for a Company
Sources: University of Manchester (October 22, 2026), Phys.org (October 23, 2026). Preprint data not yet peer reviewed.