When AI Gives Evolution a Head Start

How AI-designed proteins are helping laboratory evolution create enzymes with new functions.

Rebioin AI | November 2026

In nature, evolution unfolds over countless generations. In a laboratory, scientists can accelerate this process dramatically, sometimes generating dozens of rounds of mutation and selection in a single day. But there is a catch: when a protein evolves a new skill in the lab, it often becomes shaky and unstable. It learns a new trick but forgets how to fold properly.

In July 2026, researchers at the Broad Institute and Harvard published a clever solution in Nature. Instead of evolving a natural protein directly, they first asked an AI model to redesign it into a more stable version and then allowed evolution to do its work. The result was better-performing enzymes with improved functional properties. Let's unpack how this works, in plain language, whether you come from the wet lab or the world of code.

Summary

1. The problem: Laboratory evolution can create proteins with new functions, but the evolved proteins often lose stability along the way [1,2].
2. The idea: Start evolution with an AI-redesigned, more stable version of the protein instead of the natural one [1,2].
3. The result: Across the evolution campaigns tested, AI-redesigned starting points consistently enabled better evolutionary outcomes than their natural counterparts, including improved enzyme activity and access to new functional variants. One evolved enzyme achieved more than 79-fold greater selected specificity for a disease-linked protein target [1,2].

First, the basics

Proteins and enzymes

Proteins are long chains of amino acids that fold into three-dimensional shapes, and their shapes help determine their functions. Enzymes are proteins that speed up chemical reactions. A protease is a special type of enzyme that works like molecular scissors: it cuts other proteins at specific locations.

Directed evolution: breeding proteins in a dish

Think of breeding dogs, but for molecules. Scientists:

1. Introduce many random mutations into a protein's gene.
2. Select the versions that perform the desired function best.
3. Repeat the process for multiple rounds.

This powerful approach earned Frances Arnold a share of the 2018 Nobel Prize in Chemistry for her work on the directed evolution of enzymes [5].

PACE: evolution on fast-forward

In 2011, David Liu's lab introduced PACE (phage-assisted continuous evolution), a method that evolves proteins rapidly and continuously in the laboratory, with little hands-on work between rounds [3]. Since then, it has been used to improve many types of proteins, including genome-editing tools [2].

ProteinMPNN: the AI sequence designer

ProteinMPNN, developed in David Baker's lab, is a deep-learning model. You provide it with a three-dimensional protein backbone, and it suggests amino acid sequences that are likely to fold into that structure [4]. In this study, it proposed new sequences designed to maintain the natural enzyme's overall structure while improving stability [1].

The hidden problem: evolution spends stability

Here is a simple way to picture it. Stability is like money in a bank account. Every new mutation that gives a protein a new function can come at the cost of stability. If the protein starts with a small balance, it can only tolerate a limited number of changes before it becomes unstable and stops folding properly. That limits how much it can evolve [2].

As first author Nick Krasnow explained, a more stable protein has stability to spare, allowing it to tolerate bigger changes in pursuit of new functions [2].

The Liu lab's hypothesis was simple: What if we give evolution a richer starting account?

The experiment

The team worked with botulinum neurotoxin (BoNT) proteases, the protein-cutting components of botulinum neurotoxins. These toxins are used in carefully controlled medical applications, including Botox treatments [2].

Here is how the experiment worked:

1. Redesign with AI. The researchers used ProteinMPNN to redesign three distinct BoNT proteases, obtaining variants with improved stability while maintaining catalytic efficiency [1]. The paper also describes PROSS-redesigned enzymes as alternative starting points [1].
2. Evolve side by side. They ran PACE campaigns starting from the AI-redesigned enzymes and, in parallel, from the original natural (wild-type) enzymes [1].
3. A disease-related target. They then evolved a BoNT/E protease to selectively cleave ataxin-2, a protein associated with neurodegenerative diseases. This could help researchers investigate the potential of targeted protein cleavage in disease-related processes [1,2].

What they found

1. Better starting points consistently won. Evolving from three different redesigned enzymes produced proteases with higher activity than those evolved from their natural counterparts under the same selection conditions [1]. This advantage was observed across multiple proteases and different target substrates [2].
2. A big gain in specificity. The protease evolved from the redesigned starting point achieved higher catalytic efficiency and stability while reducing unwanted cleavage of its native substrate. It demonstrated more than 79-fold greater selected specificity for ataxin-2 compared with the best variant evolved from the natural enzyme [1].
3. Access to "locked" solutions. Mutations that evolved in the redesigned proteins did not work when transplanted back into the natural enzyme because they destabilized it [1,2]. In other words, the natural protein could not easily access those evolutionary solutions.

Why does this work?

A good analogy is a mountain landscape. Evolution is a hiker looking for the highest peak (the best function). A fragile protein stands on a narrow ridge: one wrong step and it falls off (unfolds). A stable, AI-redesigned protein stands on a wider plateau, allowing it to explore more directions without losing its structure.

The authors describe this advantage as increased mutational robustness. By making the starting protein more stable, AI redesign allows it to tolerate more mutations without losing its structure or function, expanding the range of sequences that evolution can explore [1].

The key takeaway from the Broad Institute research is that neither AI design alone nor laboratory evolution alone achieves what the combination can accomplish [2].

Why it matters

1. More efficient protein engineering. Laboratory evolution requires time and resources, so choosing a better starting point can improve the chances of success and reduce experimental effort [2].
2. Beyond natural limits. This approach could help scientists evolve protein properties that may be difficult or impossible to reach using natural starting proteins [2].
3. Many more proteins to explore. The researchers suggest that this strategy could also benefit other proteins, including reverse transcriptases whose stability can limit the performance of prime editors. They are exploring broader applications of this approach [2].

Important caveats

Good science communication means being honest about limitations:

1. This is a proof-of-concept demonstrated in a specific enzyme family using botulinum protease models [2]. It is a protein-engineering demonstration, not a treatment for neurodegenerative diseases.
2. It has not yet been established that the same improvements will occur across every protein family. Further research is needed to test how broadly the strategy can be applied.
3. AI redesign was particularly useful here because protein stability was an important limitation. Results may differ when stability is not the main bottleneck.

What students can take from this

1. If you are a wet-lab biologist: AI is not replacing your experiments. It can help make experimental starting points more promising. Selection and laboratory validation remain essential parts of the work.
2. If you are a bioinformatician: The value of a computational model is ultimately demonstrated through experimental validation. This study is a great example of how a computational tool can change the outcome of an experimental pipeline.
3. For everyone: The best scientific results often come from combining approaches instead of choosing between them.

Mini glossary

1. Wild-type (WT): The natural, unmodified version of a protein.
2. Protease: An enzyme that cuts other proteins.
3. Directed evolution: Repeated rounds of mutation and selection in the laboratory to improve a protein.
4. PACE: A rapid, continuous directed-evolution method that uses bacteriophages.
5. ProteinMPNN: An AI model that proposes amino acid sequences for a given three-dimensional protein structure.
6. Specificity: How selectively an enzyme acts on its intended target rather than other possible targets.
7. Ataxin-2: A protein associated with neurodegeneration and used as the target in this study.

References

1. Krasnow NA, Xu JA, Zhang E, Mahadeshwar GK, Tao YA, McCreary J, Hemez CF, Brown LE, Jiang W, Liu DR. AI-redesigned starting points and outcomes enhance protein evolution. Nature. Published 22 July 2026. doi:10.1038/s41586-026-10820-0
2. Williams SCP. Combining AI with laboratory evolution yields improved proteins. Broad Institute News, 23 July 2026. https://www.broadinstitute.org/news/combining-ai-laboratory-evolution-yields-improved-proteins
3. Esvelt KM, Carlson JC, Liu DR. A system for the continuous directed evolution of biomolecules. Nature. 2011;472:499–503. doi:10.1038/nature09929
4. Dauparas J, et al. Robust deep learning-based protein sequence design using ProteinMPNN. Science. 2022;378:49–56. doi:10.1126/science.add2187
5. NobelPrize.org. The Nobel Prize in Chemistry 2018 (Arnold, Smith, Winter). https://www.nobelprize.org/prizes/chemistry/2018/press-release/