Combining AI with laboratory evolution yields improved proteins

When scientists want to create a protein that can potentially treat disease, one approach is to begin with a different protein and evolve it in the lab, selecting over many generations for versions that have a new, desired function.

Broad Institute researchers have now discovered that starting this laboratory evolution process with more stable, AI-designed proteins leads to better results than starting with natural proteins. Their finding could change the way scientists engineer proteins for medicine and other applications.

In a proof-of-concept study published this week in Nature, the scientists combined AI protein design and laboratory evolution to reprogram botulinum neurotoxin proteases to cleave a variety of different protein targets, including ataxin-2, a protein involved in neurodegeneration. First, they used AI to make a more stable version of the botulinum protease, then they evolved its function so that it could cut ataxin-2. They showed that the resulting proteins were more stable and specific for this function than proteins evolved from natural botulinum protease.

“The most important finding is that using AI to stabilize natural proteins can provide much better starting points for laboratory protein evolution than what we and other researchers have been using for decades,” said David Liu, senior author of the study, Richard Merkin Professor, and director of the Merkin Institute of Transformative Technologies in Healthcare at the Broad. “This insight could change the way researchers conduct protein evolution.”

AI + evolution

In 2011, Liu’s lab pioneered a method called PACE that rapidly evolves new proteins in the lab with desired functions. His lab has used PACE to evolve improved versions of dozens of types of proteins with a wide range of activities including activating genes and precisely editing genomes in patients. However, laboratory-evolved proteins are often unstable, because in many cases they carry out the selected function but partially lose the ability to fold reliably and express well in cells.

Recently, AI tools have emerged that can design more stable versions of existing proteins. Liu’s group hypothesized that one of these AI-designed proteins might offer a better starting place for PACE to evolve a new and still stable protein.

“Laboratory evolution requires the commitment of time and resources. So what you start with is incredibly important as a major determinant of what you end up with,” said Liu, who is also a Howard Hughes Medical Institute investigator and a professor at Harvard University.

To test whether AI-designed starting points work better, Liu’s team—led by graduate student Nick Krasnow—used an AI model called ProteinMPNN developed by David Baker’s laboratory to redesign natural botulinum neurotoxin proteases. These enzymes are best known as the active component in Botox, where they paralyze muscles by cutting specific protein targets. The AI model suggested new amino acid sequences that folded into the same 3D structure as the natural enzyme, but with more stability.

Then, Liu, Krasnow, and their colleagues used PACE to guide the evolution of the AI-redesigned protein so that it could snip ataxin-2, removing the sticky region that causes proteins in the brain to clump in neurodegeneration. The new protein was 79 times better at cutting ataxin-2 than versions evolved from natural botulinum protease. 

The researchers tested the approach across multiple botulinum proteases and different target substrates, and in every case, the AI-designed starting points outperformed their natural counterparts. Further experiments showed that starting with the more stable, AI-designed protein gave the protease more flexibility in acquiring new mutations to gain function. 

Other experiments showed the difference between the natural and AI-designed proteins: the activity-enhancing mutations that evolved in the AI-redesigned ones couldn't be transplanted into the natural proteins without completely destabilizing them.   

“When proteins evolve new functions, they typically sacrifice stability in the process,” said Krasnow. “That limits how much they can change during evolution. But if you start with a more stable protein, it has more stability to spare, so it can afford larger changes in pursuit of new functions.”

The results suggest that neither AI design alone, nor laboratory evolution alone achieves what their combination can accomplish. The researchers say this new approach could be useful for many other types of proteins, including reverse transcriptases whose stability can bottleneck prime editors, and are now broadly applying the strategy.