Manchester researchers tested AI-designed enzymes for one reaction relevant to lenacapavir synthesis. The preprint reports laboratory activity, not a finished manufacturing process or a clinical advance.
Laboratory evidence for a single synthesis step
Manchester researchers have tested an AI-designed enzyme for one chemical reaction relevant to lenacapavir synthesis, with four of 67 designs showing activity above background. The strongest converted 79.0% of the targeted starting amine. These enzyme tests provide laboratory evidence for a single step in drug synthesis, not a completed manufacturing process or a clinical result.
The work brings together computational protein design and experimental chemistry. Its significance lies in testing whether newly generated proteins can perform a demanding reaction, rather than relying only on predictions about their structures.
Manchester's role in the AI-designed enzyme collaboration
Professor Anthony Green, Dr Henry Jones and Dr Zefei Xu, from the Manchester Institute of Biotechnology, worked with the Institute for Protein Design on the experimental challenge. The University of Manchester announcement was published on 22 October 2026 at 20:23 Europe/London.
That is the announcement’s publication time. It should not be read as the date or time when the laboratory experiments were performed.
The researchers tested enzymes generated using RFD4-Proteina. David Baker’s team at the University of Washington’s Institute for Protein Design developed the system with NVIDIA, bringing together the RosettaFold Diffusion line and Proteina-Complexa.
Manchester’s contribution gives the story its local research connection: laboratory testing of newly designed proteins against a specific chemical task. The wider collaboration supplied the computational approach used to produce candidate enzymes.
The work is presented as a preprint, rather than a peer-reviewed publication. Its experimental findings and computational comparisons are therefore reported here as the researchers’ results. The team describes the enzyme work as a first proof-of-concept demonstration of a newly designed enzyme for a step in active pharmaceutical ingredient synthesis.
That description concerns the chemical ingredient used in a medicine, not a demonstration of an entire production route. The distinction matters because the reported success addresses one reaction within a wider synthesis problem.
Why this amide-forming reaction is difficult
Enzymes accelerate chemical reactions. Traditionally, researchers have started with naturally occurring enzymes and engineered them to perform a desired task. Computational design offers another approach: creating new proteins by finding compatible amino-acid sequences, three-dimensional structures and molecular interactions.
Here, the target was an amide-bond-forming reaction relevant to lenacapavir synthesis. It involved joining an activated ester to a bulky chiral amine. An amine is one of the starting chemical components; the desired outcome is the formation of the amide product.
Chirality adds a further requirement. The amine has two mirror-image forms, called enantiomers and labelled S and R. The enzyme needed to favour the S form required for the synthesis rather than treating the two forms alike.
The reaction also faces competition from water. Water is present in substantial excess in enzyme-mediated transformations and can react with the activated ester. This competing process, hydrolysis, diverts it away from the desired amide-forming reaction.
A useful AI-designed enzyme therefore had to do more than accommodate the starting molecules. It needed to favour both the required chemical reaction and the required form of the amine.
These requirements explain why a plausible protein structure alone would not settle the question. The enzyme tests had to establish whether the designed proteins actually promoted the intended chemistry under the reported experimental conditions.
What the laboratory tests actually found
The RFD4-Proteina project results describe designs built around a cysteine-protease catalytic triad: a group of three amino-acid residues involved in catalysis. The team used multi-state AlphaFold3 filtering to favour aminolysis, the desired amide-forming chemistry, over ester hydrolysis.
Four of the 67 RFD4-Proteina designs tested formed amide above background. An earlier campaign using RFD3 found none among 72 tested designs. These are results from the reported campaigns, not a universal success rate for either system.
Design 52 produced the strongest reported conversion of the desired starting material. It converted 79.0% of the S-amine, compared with 8.8% of the R-amine. Design 58 converted 61.7% of the S-amine and 14.6% of the R-amine.
The paired figures matter. They show why the result is not simply a question of whether a reaction occurred: the researchers also examined how strongly each design favoured one mirror-image starting material over the other.
The project reports acid formation of 0.29 millimolar for design 52, compared with 0.26 millimolar without enzyme. This is a separate measurement from starting-amine conversion and relates to the competing reaction. It should not be combined with the conversion percentages as though they measured the same outcome.
The 79.0% figure is consequently a specific experimental result, not the overall yield of lenacapavir manufacture. It does not measure medicine purity or patient outcomes.
The evidence establishes activity in several designed enzymes for this reaction. It does not establish commercial deployment, production-scale performance, cost savings or carbon savings. Those outcomes cannot be calculated from the reported conversion figure.
AI enzyme design: generating sequence and structure together
AI enzyme design must address both what a protein is made from and how its atoms are arranged. A sequence identifies the amino-acid residues; a structure describes their positions in three dimensions. For this project, the model generates those elements together.
NVIDIA describes RFD4-Proteina as a flow-matching model that jointly generates protein sequences and all-atom structures. Its output therefore contains both residue identities and atomic coordinates, rather than only a proposed sequence.
Generation can be conditioned on molecular context that needs to remain in place. The model documentation gives examples including a binding target, an enzyme motif, a ligand or selected structural features. These constraints connect the generated protein to a specified molecular task.
An enzyme motif is a relevant arrangement within the protein; a ligand is a molecule included in its molecular context. Preserving such features allows the design process to work around requirements rather than generate an unconstrained protein.
The resulting identities and coordinates are supplied in CIF format, a structural data format. That file records a computational design. It is not, by itself, evidence that the protein carries out a reaction.
This is the central distinction between the design system and the Manchester experiment. AI enzyme design produces candidates; laboratory measurements establish which candidates show the intended activity.
How the model refines and selects candidates
The project distinguishes two forms of computational reasoning. Its term “intrinsic reasoning” describes repeated refinement of a design and the allocation of computation. “Extrinsic reasoning” describes steering the search using external energy or geometric scores.
These are technical descriptions of how the search proceeds, not evidence of human-like understanding. They explain how candidate generation can be adjusted and how candidates can be selected for further attention.
In the project’s illustrated branching search, filtering considers interface area, then hydrogen bonding, and finally AlphaFold3 confidence as progressively finer molecular properties emerge. The checks move from broader features towards more detailed assessments of the proposed molecular arrangement.
This process helps distinguish generating a candidate from retaining it after computational evaluation. A design can be produced without passing the later checks, and passing those checks remains different from demonstrating chemical activity.
The confidence scores belong to the computational evidence. The starting-amine conversion measurements belong to the experimental evidence. Keeping those categories separate makes it possible to understand what each part of the collaboration actually established.
What computational benchmark success measures
The project also reports numerical comparisons with other design approaches. These provide context for the model, but they are separate from the Manchester laboratory results.
In a protein-binding benchmark, high-thinking single-shot generation achieved a clustered pass rate of 10.6%, compared with BindCraft’s 5.9%. The project reports 42-fold higher generation throughput for that comparison.
The wording is important: this is a protein-binding benchmark and a measure of candidate generation. It is not a claim that the Manchester reaction ran 42 times faster, or that a medicine could be manufactured at that rate.
A separate enzyme-scaffolding comparison reported clustered pass rates of 12.1% for RFD4-Proteina and 5.7% for the next-best model. These are computational benchmark results, not the proportion of experimentally active enzymes in the Manchester campaign.
There are therefore three distinct stages to keep in view: generating candidate designs, passing computational checks and demonstrating activity in laboratory experiments. A favourable result at one stage does not supply the measurements needed to establish success at another.
For readers interested in drug synthesis, the benchmark figures describe research-tool performance. They do not establish lower medicine prices, faster industrial manufacture or clinical benefit. The laboratory conversion results remain the relevant evidence for what happened in the reported chemical reaction.
The code and model resources researchers can access
As checked on 24 October 2026, the project provides public research information, while NVIDIA’s RFD4-Proteina model card provides documentation and downloadable weights. The model card identifies a release date of 19 October 2026, separate from Manchester’s announcement three days later.
NVIDIA publishes a paired flow model and autoencoder. The model and weights are available under BSD 3-Clause terms, but optional third-party assets have separate terms. The licence statement should not be extended to every dependency or additional asset.
The documentation also separates generation downloads from training data. It warns that mismatching flow-model and autoencoder releases produces invalid results. Access to the components therefore comes with a specific compatibility requirement, not simply a choice of interchangeable downloads.
These resources allow researchers to examine the method and its reported results. Their availability is distinct from independent confirmation of every finding or readiness for industrial use.
The supported conclusion is narrower and useful: Manchester researchers found activity in several newly designed enzymes for one demanding reaction relevant to drug synthesis. Readers seeking the technical detail can follow the project results alongside the model documentation, keeping experimental findings separate from computational performance claims.
Frequently asked questions
Who carried out the Manchester enzyme tests, and when were they announced?
Professor Anthony Green, Dr Henry Jones and Dr Zefei Xu at the Manchester Institute of Biotechnology collaborated with the Institute for Protein Design. The university announced the work on 22 October 2026 at 20:23 Europe/London; that timestamp does not establish when the experiments took place.
Has the research been peer reviewed?
The work is presented as a preprint, not a peer-reviewed publication. The experimental and computational findings should therefore be understood as results reported by the research team.
What reaction was the enzyme designed to support?
It was designed for one amide-forming step relevant to lenacapavir synthesis. The reaction joins an activated ester to the S-enantiomer of a bulky chiral amine while limiting competing hydrolysis.
What does the 79% conversion result mean?
Design 52 converted 79.0% of the starting S-amine in the reported reaction. This is not the overall yield of lenacapavir manufacture, a measure of medicine purity or evidence of clinical effectiveness.
Does this research establish a new treatment?
No. It concerns laboratory testing of one reaction relevant to drug synthesis, rather than a clinical trial, a newly approved medicine or a completed manufacturing process.
What does RFD4-Proteina generate?
It generates protein sequences and all-atom structures together while preserving specified molecular context. Its outputs include residue identities and three-dimensional coordinates; experimental testing is needed to establish enzyme activity.
Does the reported computational speed improvement mean cheaper medicines?
No such conclusion is established. The reported 42-fold improvement concerns generation throughput in a particular protein-binding benchmark, not medicine production costs or prices.
Can researchers access the model?
As checked on 24 October 2026, NVIDIA provides downloadable paired model components and weights under BSD 3-Clause terms. Optional third-party assets have separate terms, and the flow-model and autoencoder releases must match.