Redesign Then Evolve: AI-Redesigned Starting Points Enhance Directed Protein Evolution
A paper from David Liu's lab (Broad Institute/Harvard) asks a simple but underexplored question: are computationally redesigned enzymes better starting points for directed evolution than their natural counterparts? The authors combine ProteinMPNN (David Baker's lab, structure-conditioned sequence redesign) with PACE/PANCE (phage-assisted continuous/non-continuous evolution, David Liu's own platform) to test it directly, and the answer is yes, consistently, across three related enzymes and a therapeutically relevant specificity-reprogramming task.
The stability-function trade-off
Directed evolution is good at finding new activity and specificity, but the mutations that grant it often destabilize the protein, reducing folding, expression, or aggregation resistance. Computational redesign is close to the opposite: methods like ProteinMPNN and PROSS reliably improve stability and expression, but designing a highly active new catalytic function directly is still hard. The paper proposes a division of labor instead of asking one method to do both jobs: use computational design to build a more stable, mutation-tolerant scaffold, then let directed evolution discover the new function on top of it.
The redesign workflow
The redesign target was the zinc-dependent protease domain of three botulinum neurotoxin proteases, BoNT/E, BoNT/F, and BoNT/X, whose activity can be coupled to phage replication in a PACE circuit. ProteinMPNN worked on a fixed backbone rather than generating anything de novo: residues within a set distance of the substrate or catalytic zinc, and residues strongly conserved across natural homologs, were fixed, and everything else was redesigned. Distance cutoffs of 14 and 18 Å and conservation thresholds of 30% and 60% were tested, with designs filtered by AlphaFold2 self-consistency (pLDDT and RMSD to the input structure) before experimental screening for expression, activity, and thermal stability.
How well did it work?
For BoNT/E, 78% of 74 tested ProteinMPNN designs retained detectable protease function, 45% matched or exceeded wild-type activity, and 30% combined retained activity with higher soluble yield; the top three variants (D1–D3) showed 1.7–2.8-fold higher catalytic efficiency than wild type. The approach generalized: for BoNT/F, 20% of 81 designs were active and all six that matched wild-type activity also expressed better; for BoNT/X, a striking 99% of 69 designs were active, 26% matched or exceeded wild type, and two variants raised the melting temperature by 22°C and 14.5°C. The redesign direction also runs both ways: grafting the mutations from a previously PACE-evolved but poorly expressed BoNT/E variant onto the redesigned backgrounds rescued soluble expression by 2.2- to 5.2-fold without losing its evolved activity.
Redesign reshapes the fitness landscape
The central experiment evolved wild-type BoNT/E and the redesigned D3 variant side by side against three SNAP25-derived substrates of increasing difficulty. On the easiest substrate, D3-derived proteins reached 6.1–7.7-fold circuit activation versus 2.6–3.0-fold for wild type; on the hardest, two of four wild-type PACE replicates failed outright while all four D3 replicates succeeded, reaching 16- and 20-fold activation versus about 10-fold for the one successful wild-type lineage. Sequencing evolved populations showed a telling asymmetry: mutations that worked in wild-type populations usually also worked in D3, but D3-specific solutions were rarely functional when transferred into wild type. One mutation, K225E, was strongly destabilizing and produced high activity in D3 but nothing useful in wild-type BoNT/E, a clean example of epistasis, where a mutation's effect depends on the sequence background it lands in. This wasn't a quirk of one lucky redesign either: three additional starting points (a second ProteinMPNN design and a PROSS design) across 44 independent evolution experiments all outperformed wild type, though each supported a different set of evolved mutations.
A therapeutically relevant test: ataxin-2
As a more demanding demonstration, the authors evolved redesigned BoNT/E to cleave ataxin-2, a protein implicated in neurodegenerative disease, while selecting against cleavage of BoNT/E's native substrate, SNAP25. The best D3-derived proteases cleaved ataxin-2 about 1.4- to 1.7-fold faster than the best wild-type-evolved comparator, with up to roughly 79-fold greater selected specificity; one leading D3-derived variant showed no detectable SNAP25 cleavage at all, while wild-type-evolved variants retained low but measurable off-target activity. In HEK293T cells, the redesigned-derived proteases also accumulated to higher levels and produced more of the intended cleavage product. Transferring the D3-evolved mutations into wild-type BoNT/E killed ataxin-2 cleavage entirely, while the reverse transfer stayed active, reinforcing that the redesigned background is what made the more complete specificity switch possible.
Worth knowing before you get too excited
This is three redesigned backgrounds and one enzyme family (BoNT proteases), so generalization to unrelated enzymes is still an open question the authors flag explicitly. Stability alone doesn't explain everything either: one redesign, D4, was highly stable but had 20-fold lower initial activity than wild type, yet still evolved into some of the best-performing variants, so the detailed epistatic background matters as much as raw melting temperature. The distance and conservation thresholds used to constrain ProteinMPNN are heuristic choices, not universally optimized rules, and AlphaFold2/3 filtering is a prioritization tool, not proof of correct folding or catalytic geometry; AlphaFold3 didn't even correctly predict the experimentally determined ataxin-2 cleavage site. PACE circuit activation is also a composite signal (activity, expression, folding, and circuit compatibility bundled together), not a pure measurement of catalytic efficiency. And the ataxin-2 proteases were only tested in biochemical assays and transfected HEK293T cells, not in neurons or animal models, so this is a research lead, not a validated therapeutic.
The core lesson still holds up well: before evolving a marginally stable enzyme toward a demanding new function, it may be worth first building several stable, sequence-diverse, computationally redesigned starting points and evolving those in parallel, rather than evolving the natural protein alone.
For similar tools, see the BioMoDes Protein Engineering Applications page.
References
- Krasnow, N.A., Xu, J.A., Zhang, E., et al. AI-redesigned starting points and outcomes enhance protein evolution. Nature (2026).
- Sequence design guide (code) (GitHub).
This post is an AI-reworded, expanded version, in my own voice, of a summary I originally posted on LinkedIn.
I try my best to make the information on this website as accurate as possible.
If you find any errors in the contents of this page or any other page on this website,
I would greatly appreciate that you kindly get in touch with me at
contact[at]abeebyekeen[dot]com.
If you are interested in joining my free weekly “BioModes and Top Reads” newsletter, please subscribe below.