Baker’s group used AI to design an enzyme from scratch that attaches the protein tag ubiquitin to a chosen amino acid (lysine) in PCNA, which holds DNA polymerase in place: after a single substitution in the enzyme, about 90% of PCNA was tagged in vitro
Baker’s group used AI to design an enzyme from scratch that attaches the protein tag ubiquitin to a chosen amino acid (lysine) in PCNA, which holds DNA polymerase in place: after a single substitution in the enzyme, about 90% of PCNA was tagged in vitro
On November 11, David Baker’s group at the University of Washington and researchers at ETH Zurich posted a preprint on enzymes designed using neural networks. Their best enzyme, PL4.1, selectively targets lysine 164 in PCNA among 45 other lysines.
Ubiquitin is a protein that acts as a tag. Cells attach it to a lysine in another protein, and the tag can direct that protein for degradation or recruit it to respond to DNA damage. Attachment involves a sequence of enzymes: E1 uses ATP to activate ubiquitin, E2 carries it, and E3 ligases (of which humans have more than 600) select the target protein and its lysine. The function of most of the roughly 100 thousand identified attachment sites remains unknown. Studying that function requires a protein with a single tag on a chosen lysine. Existing methods for making such proteins are often complex or require a particular sequence around the lysine.
Baker’s group set out to combine the roles of E2 and E3 in a single enzyme and specify its target through the design itself. The group borrowed the reaction chemistry from cysteine proteases, which cleave protein chains: in October, it reported in Nature that it had built such enzymes from scratch. In a protease, cysteine accepts part of the chain being cleaved, and water then releases it. In the new enzyme, cysteine accepts ubiquitin, which is then transferred to the target lysine instead of water, attaching the tag. The neural network RFdiffusion3 built proteins around the active site and the target, while AlphaFold 3 screened out designs whose predicted structures differed from the intended structures.
The researchers first tested the idea on a short peptide containing a single lysine. Ubiquitin was activated in advance by attaching it to a molecule of MESNa, and 35 of 48 designs attached the tag. The authors then chose PCNA, a ring of three identical proteins that slides along DNA and holds the enzyme that copies it in place. Following DNA damage, cells attach ubiquitin to lysine 164, and the ring recruits polymerases that can copy DNA past the damage.
PCNA is a difficult target: the ring contains three target lysines and 45 other lysines, and the region around the target lysines lacks a convenient site for the enzyme to bind. The researchers therefore added a short segment to the enzyme, designed to align with the PCNA chain beneath lysine 164, anchor the enzyme, and bring the lysine into the active site. Four of the 48 PCNA designs attached the tag. Three also tagged other lysines, while the fourth, PL4, was selective: no tag was detected on PCNA when lysine 164 was replaced with arginine.
Replacing asparagine 148 in PL4 with alanine produced PL4.1, which tagged 66% of PCNA in one hour, compared with 2,6% for PL4. AlphaFold 3 predicts that this asparagine occupies a different position in PL4 than intended, blocking a group of backbone atoms needed to stabilize a reaction intermediate. The researchers then dispensed with advance activation of ubiquitin: E1, ATP, and MESNa activated it directly in the reaction mixture, and PL4.1 tagged about 90% of PCNA within 24 hours without E2 or a natural E3. The E2 enzyme UbcH5c carrying the S22R substitution, used to tag PCNA in a 2012 method, tagged 60% under similar conditions and produced several different products rather than one.
The authors suggest that the design method could, in principle, target “any lysine in any protein for which a structure or a confident model is available.” For use in cells, they propose:
“Delivered into cells, designed ligases could allow individual ubiquitination events to be induced on demand, turning proteome-wide maps of ubiquitination sites from a catalog into a set of testable hypotheses.”