AI designed zinc enzymes from scratch to cut protein chains at specified sites: 36% of 135 designs worked
AI designed zinc enzymes from scratch to cut protein chains at specified sites: 36% of 135 designs worked
On October 14, Nature published a paper by David Baker (University of Washington, Seattle) and colleagues. Neural networks in the RFdiffusion family designed proteins in which a zinc ion activates water to break a bond between amino acids. The best design accelerated bond cleavage by more than 10⁸ times relative to the reaction without an enzyme. After four substitutions, the acceleration exceeded 10¹⁰ times.
Without an enzyme, it takes hundreds of years for half the peptide bonds linking amino acids in a protein to break in water. In cells, enzymes called proteases cleave these bonds. Previous computationally designed enzymes containing a metal ion had targeted esters, whose bonds are easier to break. The researchers based the active site on that of a natural zinc protease, aminopeptidase N: three amino acids (two histidines and a glutamate) hold the zinc ion, a water molecule binds to it, and a second glutamate activates the water to attack the bond. The neural network built a protein unlike known proteins around this site and fitted it to a chain of 8–13 amino acids, the substrate it would cleave. Each enzyme had its own substrate, positioned so that the target bond aligned precisely with the active site.
According to the preprint of the same study, the researchers selected designs using structures predicted by AlphaFold 3. A design counted as active if it cleaved more than 10% of its substrate chains within 12 hours. Of 95 designs selected using a model containing a zinc ion alone, 14 (15%) were active. Of 40 selected using a model that included water alongside zinc, 35 (88%) were active. Together, that was 49 of 135, or 36%. One inactive design illustrates the reason for the difference: adding water to the model moves the target bond away from the active site, so the model with water would have excluded this design. Replacing any amino acid that holds the zinc reduced activity by at least 7 times in all five enzymes examined, supporting the conclusion that the designed active site drives the reaction.
Zn45 was the most active enzyme and cleaved every substrate tested. On a substrate designed for another enzyme, it made 0,025 cuts per second per enzyme molecule, which is 1000–10 000 times slower than the best natural proteases. Screening 378 substitutions around the active site identified ten positions where a substitution accelerated cleavage by at least 30%. Combining four substitutions produced Zn45_v2, which made 0,75 cuts per second, 30 times faster than Zn45 and approximately 100 times slower than the best natural proteases. In simulations, the water-activating glutamate in Zn45 adopted several positions, whereas in Zn45_v2 it stayed in place, as it does in the natural enzyme.
The targets were TDP-43 (a protein that aggregates in amyotrophic lateral sclerosis), amyloid-β (which forms plaques in Alzheimer's disease), and serum amyloid A (which forms deposits in amyloidosis). For the latter two, the researchers cleaved ten-amino-acid fragments. For TDP-43, they chose a cleavage site in the protein's flexible tail, which controls aggregation, with the aim of breaking up aggregates while leaving the rest of the protein functional. TDPn3, an enzyme designed from scratch for this site, cleaved it more efficiently than other substrates and cut full-length TDP-43 in a test tube into fragments of the expected sizes. Lead author Anqi Chen tells GeekWire that after cutting one target molecule, the enzyme “can move on to a second, a third, and thousands of other targets.”
Cleavage can also act as a switch. The researchers fitted Neo2, an artificial analogue of interleukin-2 (an immune signalling protein), with a mask that blocks its receptor-binding site. The mask is attached through a sequence that Zn45 cleaves. In the cell assay, masked Neo2 required a concentration of 0,5 nM to produce a half-maximal response. After treatment with Zn45, that concentration fell to 0,01 nM, matching Neo2 without the mask. MMP2, a tumour-associated enzyme that removed the mask in the Baker laboratory's Prokine system, did not activate masked Neo2.