Best AI-Designed Enzyme Converted 79 Percent of Key Molecule in HIV Drug
The University of Manchester has built and tested enzymes designed with the newly released open-source model RFD4-Proteina from the University of Washington and NVIDIA. Four of 67 designs were active, and the best converted 79 percent of a bulky chiral amine — a key step in the synthesis of the HIV-prevention drug lenacapavir. But so far the results exist only as a university announcement, not as peer-reviewed research.
New tool, new result — same week
The story consists of two events that converged in the same week in October 2026. On October 21, NVIDIA and the Institute for Protein Design (IPD) at University of Washington Medicine announced the release of RFD4-Proteina as an open-source model, presented as a BioNeMo open model at NVIDIA's GTC conference in Berlin. It is the first AI model for protein design that the University of Washington has developed in collaboration with NVIDIA, GeekWire reports (GeekWire).
The same week, the University of Manchester reported that researchers at the Manchester Institute of Biotechnology have used the model to design entirely new enzymes — and actually built and tested them in the laboratory. That makes the Manchester example the first externally reported, experimental use of the model in the source material.
What RFD4-Proteina is
The model's name reveals its construction, according to Genetic Engineering & Biotechnology News (GEN): It combines IPD's RosettaFold Diffusion, a guided diffusion model for generating new proteins, with NVIDIA's Proteina-Complexa, a generative model for designing protein binders against protein and small-molecule targets. The model is released as open source under the BSD-3 license.
The release builds on a computational fellowship from the Jen-Hsun and Lori Huang Foundation, with GPU access through CoreWeave since June 2026, GeekWire reports.
The developers themselves present the model as a generational shift: In computational benchmark tests, RFD4-Proteina is said to have achieved design success rates five to ten times higher than previous models, and to have generated protein candidates five to ten times faster. These, however, are claims from IPD and NVIDIA, based on a "research draft" that is not publicly available in peer-reviewed form. NVIDIA's own press release states that the model is said to have outperformed the evaluated reference models in four test areas: protein binding, small-molecule binding, enzyme scaffolding, and DNA binding (NVIDIA via GlobeNewswire).
What Manchester actually showed
The Manchester experiment is therefore interesting precisely because it does not rest on benchmark numbers, but on molecules in a test tube. Professor Anthony Green, Dr. Henry Jones, and Dr. Zefei Xu worked in collaboration with IPD to design enzymes for the affordable and sustainable production of global health medicines, the university writes (University of Manchester).
The target was a challenging step in the synthesis of lenacapavir, an HIV-prevention drug: an amide bond-forming reaction. Of 67 enzymes selected and experimentally tested, four showed clear activity. The best design converted 79 percent of a bulky chiral amine into the desired amide product, and according to the announcement the enzyme showed high selectivity for the S-enantiomer of the amine — exactly the form the lenacapavir synthesis requires.
The hit rate may seem modest — four of 67 — but in de novo design of enzymes for catalyzing real organic reactions, any active design is significant. The university describes the work as "a critical proof-of-concept demonstrating that new enzymes can be designed computationally to accelerate challenging steps in the synthesis of active pharmaceutical ingredients."
What is verified — and what is not
The distinctions here matter:
- The Manchester numbers (4 of 67 active, 79 percent conversion, S-selectivity) come from the university's own news release. No peer-reviewed article or preprint exists in the source material, so the results cannot yet be independently verified.
- The five- to tenfold improvements are developer claims from IPD and NVIDIA, tied to an unpublished draft. Benchmark details and public reference figures for the comparison are not available.
- The model itself, on the other hand, is genuinely available: open source under BSD-3, launched as a BioNeMo model on October 21, 2026.
Why it matters
Two reasons make this week worth noting. First, the Manchester example shows that computational enzyme design can attack hard steps in the synthesis of active pharmaceutical ingredients — here a bulky chiral amine, a classically difficult task for both chemical synthesis and biocatalysis. Since lenacapavir is an HIV-prevention drug, cheaper and more sustainable production of the active ingredient has direct relevance for global health.
Second, the case illustrates how open release changes the logic of testing: The model was made public, and within the same week a university on the other side of the Atlantic could apply it to a problem the developers had not themselves chosen. At UW, PhD candidate Seth Woodbury describes the same effect from the inside: "In just a few days, we designed an enzyme that breaks down pesticides and performs many times better than my previous best, which took months to design," he told GeekWire.
Open questions
Several things remain before this can be counted as an established result:
- Peer review: Neither the Manchester enzyme results nor NVIDIA's benchmark figures have been published in peer-reviewed form. Both could change under review.
- Reproducibility: The details of the comparison basis for the "five to ten times" claims are not publicly available.
- Scaling: One active candidate for one reaction in one synthesis is a proof of concept. How general the model is for industrial enzyme design, no one yet knows.
The model is now out in the world. The coming months — with publication of the Manchester work and independent uses from other laboratories — will show whether the first week's demonstration was an outlier or the beginning of a pattern.