AI-designed starting points made evolved enzymes 80 times sharper

Liu's lab used ProteinMPNN to redesign a protease before evolving it, and the descendants cut an ALS-linked target 80x more efficiently.

David Liu's group at the Broad Institute of Harvard and MIT has changed where directed evolution begins. Rather than starting from a natural enzyme, the team used ProteinMPNN — the protein-design model from Nobel laureate David Baker's lab at the University of Washington — to generate 58 variants predicted to be more stable. The top three, expressed in E. coli, stayed soluble instead of clumping, and some were more active than the natural versions they were based on.

Those redesigned enzymes were then fed into PACE, the continuous-evolution system Liu's lab reported in 2011, which runs dozens of rounds a day by growing bacteriophages under selection pressure. The target was a protein tied to neuron health: in ALS, a repetitive stretch within it expands, causing aggregates that damage neurons. Natural enzymes have had limited success cutting the mutant form before it clumps. Enzymes descended from the AI-redesigned starting points were nearly 80 times more efficient at cutting it and over 56 times more selective for the intended site than those evolved from natural proteases. The pattern held across three neurotoxin types and multiple substrates.

The bottleneck this addresses is time, and time is most of the cost of a biologic. Directed evolution is slow largely because mutations destabilize the protein faster than they improve it, so runs collapse or need chaperones and extra stabilization rounds. A more mutation-tolerant starting point means more of each run converts into a usable enzyme — and it makes targets reachable that no natural enzyme fits.

This is a proof of concept measured in the lab, not a therapy. No clinical work is claimed here, and the ALS protein is a target, not a treated disease.

Source: Singularity Hub