AI protein design improves enzymes made through lab evolution

AI protein design enzymes performed better when paired with laboratory evolution than enzymes evolved from natural proteins, according to a Broad Institute of MIT and Harvard team. The result matters because engineered proteins are being explored for medicine and other uses, where both activity and stability can shape whether a design works in cells.

In a proof-of-concept study published in Nature, the researchers redesigned botulinum neurotoxin proteases with artificial intelligence before evolving them in the lab. The team then retooled the enzymes to cut several protein targets, including ataxin-2, which the Broad Institute said is linked to neurodegeneration.

David Liu, the study’s senior author and director of the Merkin Institute of Transformative Technologies in Healthcare at the Broad, said the work shows that AI-stabilized versions of natural proteins can be stronger starting material for laboratory protein evolution than the natural proteins researchers have long used.

What is laboratory evolution?

Laboratory evolution is a method for generating many protein variants and repeatedly selecting the versions that best perform a chosen task. Liu’s lab introduced a rapid version called PACE in 2011 and has used it to evolve proteins with activities that include gene activation and precise genome editing, according to the Broad Institute.

The problem, the researchers said, is that proteins that gain a new activity through lab evolution often lose some stability. A protein may do the selected job yet fold less reliably or express less well inside cells.

To test whether AI could improve the starting point, graduate student Nick Krasnow led work using ProteinMPNN, an AI model developed by David Baker’s laboratory. The model proposed amino acid sequences for botulinum neurotoxin proteases that kept the natural enzyme’s three-dimensional shape while increasing stability, according to the study description.

Botulinum neurotoxin proteases are best known as the active component in Botox, where they paralyze muscles by cutting specific protein targets. In the Broad experiments, the researchers used PACE to evolve the AI-redesigned protease so it would cut ataxin-2 and remove a sticky region that contributes to protein clumping in the brain during neurodegeneration.

The resulting protein cut ataxin-2 79 times better than versions evolved from the natural botulinum protease, the Broad Institute said. The researchers also tested the strategy with multiple botulinum proteases and different target substrates, and reported that AI-designed starting proteins outperformed natural ones in each case.

Further experiments suggested why the combination worked. Starting with a more stable protein gave the enzyme more room to tolerate mutations needed for new activity. The team also found that mutations that improved activity in the AI-redesigned proteins could not be moved into the natural proteins without fully destabilizing them.

Krasnow said the findings fit a common trade-off in protein engineering: proteins that gain new functions often give up stability. A more stable starting protein, he said, has more stability available to spend as it changes.

The researchers said the work points to a broader strategy: use AI to build sturdier protein scaffolds, then use laboratory evolution to give them new functions. They said the same approach may be useful for other protein classes, including reverse transcriptases whose stability can limit prime editors.

The study, “AI-redesigned starting points and outcomes enhance protein evolution,” was published in Nature by Nicholas A. Krasnow and colleagues.

This story draws on original reporting from Phys.org.