A new artificial-intelligence approach has produced entirely novel enzymes that catalysed a difficult step in the synthesis of an HIV-prevention drug, researchers report. The work demonstrates a proof of concept for using next‑generation generative models to design proteins that perform specific chemical reactions, potentially lowering the cost and environmental footprint of manufacturing essential medicines.
AI that designs proteins for chemistry
The system, called RFD4‑Proteina, was developed by the Institute for Protein Design in collaboration with NVIDIA and applied by scientists at the Manchester Institute of Biotechnology. According to the team, the model can design amino‑acid sequences and three‑dimensional folds together while taking into account molecules those proteins must interact with — a departure from many earlier approaches that treated structure and function in separate stages.
In head‑to‑head comparisons reported by the developers, RFD4‑Proteina produced candidate proteins five to 10 times faster than previous models and achieved computational design success rates that were five to 10 times higher. Crucially for laboratory use, the model can reallocate computing effort to the most promising leads, narrowing the experimental search space.
Application to lenacapavir synthesis
Researchers at the University of Manchester, including Professor Anthony Green, Dr. Henry Jones and Dr. Zefei Xu, applied the model to design enzymes to catalyse a key amide‑bond formation used in the manufacture of lenacapavir, a drug employed for HIV prevention. The target reaction is chemically demanding: the engineered catalyst must favour formation of the desired amide while suppressing a competing reaction with water, which is present in large excess in biocatalytic conditions. It also needs to favour the correct stereoisomer — in this case the S‑enantiomer — to produce the active drug precursor.
From 67 computational designs selected for laboratory testing, four showed clear catalytic activity. The top performer converted 79 per cent of a bulky chiral amine into the target amide product and displayed high selectivity for the required S‑enantiomer.
| Metric | Value |
|---|---|
| Designs tested experimentally | 67 |
| Designs showing activity | 4 |
| Best conversion to desired product | 79% |
What this means and limits
The results are being shared as a preprint and represent a proof of concept rather than a finished industrial process. Nevertheless, the study indicates that generative protein models can be steered to perform very specific chemical tasks, potentially enabling:
- more sustainable routes to active pharmaceutical ingredients by replacing multi‑step chemical syntheses with single biocatalytic steps;
- faster lead generation for enzymes tailored to proprietary processes or scarce feedstocks;
- reduced experimental screening effort through more accurate computational triage.
Experts caution that moving from a laboratory demonstration to scalable manufacturing requires additional work: enzymes must be optimised for stability, productivity and compatibility with industrial conditions, and regulatory pathways for biologically produced intermediates must be navigated. The preprint status also means results have not yet undergone peer review.
Still, the approach addresses two bottlenecks in protein design: the need to consider sequence and structure in an integrated way, and the practical challenge of discovering rare functional solutions among vast possible sequences. By focusing compute power on promising candidates and modelling interactions with substrates directly, the system reduced the number of inventions that need to be tested experimentally.
The Manchester team emphasised the potential public‑health benefits. Designing low‑cost, enzyme‑enabled syntheses could make essential medicines more affordable and easier to produce in a decentralised manner — a particularly important consideration for global health products.
As generative models for proteins continue to improve, the field will need transparent benchmarks, robust experimental validation and careful assessment of biosecurity and intellectual‑property implications. For now, RFD4‑Proteina adds to a growing body of evidence that artificial intelligence can accelerate the translation of protein science into practical chemical tools.
Key collaborators: Institute for Protein Design, NVIDIA and the Manchester Institute of Biotechnology.
