David Liu’s team used AI to stabilize proteases so laboratory evolution could find new functions
David Liu’s team used AI to stabilize proteases so laboratory evolution could find new functions
On July 22, David Liu’s team published a study in Nature. ProteinMPNN redesigned three proteases, and automated laboratory evolution trained them to cleave new protein targets. The best variant targeting ataxin-2 was more than 79 times as selective as the variant evolved from the natural enzyme.
For an enzyme to cleave a new target, mutations must alter target recognition without disrupting the enzyme’s three-dimensional structure. A useful new mutation often makes the protein less stable. The protein then folds less efficiently, and selection eliminates it along with its new function.
Liu’s team stabilized each protease before selection. ProteinMPNN received the three-dimensional structures of botulinum neurotoxin proteases and modified regions far from the catalytic site. For one protease, 58 of 74 variants retained activity, and 22 produced more soluble protein. AI provided a reserve of stability, while the experiments tested which new functions the protein could acquire without exhausting that reserve.
The authors then used PACE, a continuous evolution system based on bacteriophages. A phage could reproduce only when the protease cleaved the specified peptide. Within one day, the system completed dozens of generations of mutation and selection. For the most difficult of the three new targets, a functional enzyme emerged in all four lines started from the redesigned D3 variant and in two of four lines started from the natural protease.
When the authors transferred the identified mutations between proteins, mutations that had evolved in D3 often lost their function in the natural protease. Mutations that evolved from the natural starting enzyme, by contrast, remained functional in D3. The same amino acid substitution can behave differently in different protein backgrounds. This result is consistent with D3 tolerating some mutations that improve function but reduce stability.
In the final campaign, the researchers evolved the protease to target ataxin-2, a protein associated with the risk of amyotrophic lateral sclerosis, while selecting against cleavage of SNAP25, the protease’s natural target. For the best D3-derived variant, the ratio of ataxin-2 cleavage to SNAP25 cleavage was more than 79 times higher than for the best variant evolved from the natural starting enzyme. In cultured human HEK293T cells, it produced more of the intended product and fewer off-target fragments.
In a May study from the same laboratory, ProteinMPNN stabilized PE8, a DNA editor that had already undergone evolution. In the new article, the model is used earlier. It prepares the starting enzyme, which selection then evolves into variants with a new function.
The authors tested three related proteases using bacterial selection and HEK293T cells. Therapeutic use will require delivery of the enzyme to motor neurons, followed by studies of the immune response, safety, and efficacy in animals.