Live·Open questions in longevity research
All news
Scientific ComputingScience ResearchTherapeutics

AI designed molecule-joining enzymes from scratch: one labels a specific lysine (an amino acid) among 36 others, and another performs a step in a potential synthesis of lenacapavir, a drug for HIV prevention (preprint)

17 November 2026· 261117001

AI designed molecule-joining enzymes from scratch: one labels a specific lysine (an amino acid) among 36 others, and another performs a step in a potential synthesis of lenacapavir, a drug for HIV prevention (preprint)

On November 11, David Baker's laboratory at the University of Washington, together with chemists Donald Hilvert and Anthony Green, posted a preprint. These enzymes, called ligases, form the amide bonds found in proteins and many drugs, without directed evolution (selection of mutants in the laboratory).

To form an amide bond, chemists activate the carboxyl group of one molecule and bring it together with the amino group of another. They block additional amino groups with protecting groups. An enzyme selects the intended amino group itself, but adapting it to a new pair of molecules may require directed evolution.

The new enzymes were based on cysteine proteases, which cleave proteins. Baker's laboratory described their design from scratch in Nature in October. The active site of such a protease is built around cysteine, a sulfur-containing amino acid: the sulfur accepts an acyl group from the donor molecule and transfers it to water.

For the enzyme to act as a ligase, the amino group of a second molecule, the acceptor, must receive the acyl group instead of water, forming an amide. Cysteine was chosen for its sulfur: an acyl group bonded to sulfur reacts more readily with amines than one bonded to oxygen. Water competes for that group, however. When water receives it, the donor is broken down by hydrolysis. The RFdiffusion neural network built the protein around this active site, and the AlphaFold 3 neural network checked whether the active site's geometry was maintained across several reaction states.

The first target was FSLTK, a five-amino-acid peptide. The task was to attach a dye to the start of the chain, even though the lysine at the other end also has an amino group. Of 96 proteins, 51 passed computational screening across six reaction states, while 45 passed screening for only one state. Ten proved to be ligases, all from the group of 51. The best, NL6, attached the dye to 90% of the peptide within 16 hours.

The second target was a specific lysine in a peptide attached to a carrier protein. The ligase KL37 attached a label to this lysine. When it was replaced with arginine, which lacks that type of amino group, no product formed within 20 hours, despite the presence of 36 other lysines. In separate reactions with a peptide containing both a free chain start and the target lysine, NL6 labeled only the chain start, while KL37 labeled only the lysine.

The authors had expected hydrolysis to be the main obstacle:

“Yet we obtained several ligases that catalyzed amide formation without appreciably accelerating donor hydrolysis.”

In the authors' model, amide formation outpaces hydrolysis because the acceptor peptide itself arranges the active site for the reaction. AlphaFold 3 calculations suggest that, without the peptide, the histidine in KL37's active site adopts different positions and may activate water less effectively. With the peptide present, it settles into nearly the same position each time. Crystal structures of KL37 show that peptide binding shifts a helix that partially blocks the peptide's binding site.

The third target was a step in a potential synthesis of lenacapavir: joining the donor to an (S)-amine, one of two mirror-image forms of a molecule, like right and left hands. Here, RFdiffusion3 yielded no ligases, while the new version, RFD4-Proteina, yielded four among 67 proteins. The best converted 79% of the (S)-amine to an amide without accelerating hydrolysis. In separate reactions, it coupled the (S)-form 28 times more readily than the mirror-image (R)-form, according to the authors' calculation after correcting for background activity. The authors propose that a helix forming a lid over the active site keeps water out: this helix is present in nine of the 19 initial designs, including all four ligases, and absent from the enzyme that only cleaved the donor. They suggest that this selectivity could enable the separation of mixtures of mirror-image forms and alternative routes to a key lenacapavir intermediate, potentially reducing production costs.

Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#protein-design#ligases#amide-bonds#lysine-labeling#lenacapavir#rfdiffusion