AI-redesigned enzymes beat wild-type starting points in directed evolution, 79-fold specificity gain

2026-07-28

Redesigning proteases with ProteinMPNN makes them better starting points for directed evolution than wild type, with a 79-fold specificity gain for the therapeutic target ataxin-2.

What problem this solves

Directed evolution is the standard way to engineer an enzyme: mutate and screen repeatedly in the lab until a natural protein learns a new job. The trouble is that the natural protein is a poor starting point. Evolution did optimize it, but for its original survival task, not for stability or engineerability. Many key mutations simply break the enzyme, so the reachable sequence space gets stuck in a small pocket around the wild type.

The team tests an intuition: first use AI to redesign the protein sequence into a more stable, more mutation-tolerant version, then use that as the starting point for directed evolution. Can it reach sequences wild type cannot?

Method

The pipeline stitches two stages. The first is ProteinMPNN, an inverse-folding sequence-design model that takes a protein backbone and proposes more stable amino-acid sequences. The team redesigned three botulinum neurotoxin (BoNT) proteases this way, getting variants with higher stability and no loss of catalytic efficiency.

The second stage is directed evolution with PACE (phage-assisted continuous evolution), a David Liu lab technique that compresses mutation and selection into a continuously flowing reactor, evolving proteins orders of magnitude faster than standard methods. The key design is a side-by-side control: under the same selection pressure, one line starts from an AI-redesigned enzyme, the other from the matching wild type, and they simply race.

This control runs across four independent evolution campaigns, ending in a real therapeutic case: evolving the BoNT/E protease to cleave ataxin-2, a therapeutic target linked to neurodegenerative disease.

Results

The four campaigns agree: enzymes starting from AI-redesigned points show systematically higher activity than those starting from wild type.

The strongest evidence is the ataxin-2 campaign:

Starting pointSelectivity for ataxin-2
Best variant evolved from AI-redesigned pointbaseline
Best variant evolved from wild type79-fold lower

Mutations from the redesign-evolved line stop working when placed back into a wild-type background, which shows those sequences are genuinely out of wild type's reach. Catalytic efficiency and stability are both higher, with less off-target cleavage of the native substrate.

Why it matters

AI protein design and directed evolution used to be separate communities on separate tracks: design models invent new proteins, evolution fine-tunes old ones. This paper chains them into one workflow, where AI is not just a final product but a more stable starting line for evolution. For enzyme-engineering and protein-therapy groups, this is a copyable workflow, not a proof of concept.

Limitations

The full text sits behind the Nature paywall; this reading is based on the official abstract, so specific numbers (the stability gain per variant, PACE rounds, absolute kinetic constants for ataxin-2 cleavage) were not available. The abstract gives only a qualitative "consistently better" conclusion plus the single 79-fold anchor, without boundary conditions such as how many rounds wild type needs to hit its ceiling, or whether the AI starting point wins on every substrate. Conclusions like this usually rest on extensive single-mutation backcrossing, and judging their robustness needs the full text.

Terms

Source

What people are saying

All paper explainers