Enzymes Redesigned by Artificial Intelligence Mark New Milestone in Precision Molecular Medicine

The field of protein design has made a qualitative leap in recent years. What once required years of laboratory experimentation — random mutation, selection, repetition — can now be done in days, with AI models predicting amino acid sequences that fold into functional shapes. The current frontier is no longer just designing proteins that bind to a target: it is designing proteins that execute specific chemical reactions on other proteins, with surgical precision.

A set of recent studies published between July and October 2026 demonstrates this advance in impressive form. Researchers from the Broad Institute, MIT, and Scripps Research combined AI tools with laboratory evolution to create proteases — enzymes that cut other proteins — capable of recognizing and degrading targets associated with neurodegenerative diseases such as Parkinson's and Huntington's.

From botulinum toxin to disease treatment

The most emblematic case involves botulinum toxin proteases, the same ones that, in Botox form, paralyze muscles by cutting the SNAP-25 protein in nerve endings. Researchers at the Broad Institute, led by David Liu, used the ProteinMPNN model to redesign these proteases, creating versions more stable than the natural ones. They then applied PACE — an assisted evolution technique developed at the Broad — to reprogram the specificity of these enzymes to cut ataxin-2, the protein linked to Huntington's disease.

The result was 79 times more efficient than versions evolved from natural proteases. And, most importantly, the enzymes evolved from AI-designed starting points were intrinsically more stable and specific than those evolved from wild-type versions.

"When proteins evolve new functions, they typically sacrifice stability," explained Nick Krasnow, first author of the study. "This 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 same group published a separate study in Nature showing that this paradigm — AI as starting point, evolution as refinement — works consistently across multiple different proteases (BoNT/E, BoNT/F, BoNT/X), each with distinct therapeutic applications.

Proteases that hunt cancer

Meanwhile, researchers at MIT and Microsoft Research developed CleaveNet, an end-to-end AI pipeline for designing peptide sequences that are cleaved by specific proteases. Unlike the Broad studies, which focused on reprogramming the enzyme to attack a new target, CleaveNet focused on designing the target to be cleaved by an existing enzyme.

The demonstrated application: early cancer detection. Nanoparticles coated with the designed peptides act as sensors that emit a signal when they encounter MMP13 — a protease overexpressed in cancer cells that helps metastasis by cutting collagen. The system was able to generate peptides cleaved by MMP13 with remarkable efficiency and selectivity, something that traditional trial-and-error methods would take much longer to achieve.

The human genome encodes about 600 proteases. CleaveNet offers, for the first time, a computational path to map and explore this diversity systematically, not just for diagnosis, but potentially for therapy.

Reclaiming Botox for Parkinson's

In a third study presented in Nature, researchers at Scripps Research, led by Hartmut Sondermann, used directed evolution to reprogram the botulinum toxin protease against α-synuclein — the protein that abnormally accumulates in neurons of Parkinson's disease patients.

The mechanism is particularly elegant because it exploits an inherent characteristic of botulinum toxin: it naturally self-delivers to neurons. By modifying the enzyme to attack α-synuclein, the researchers created a tool that already knows how to find its destination. In human cells, "Protease 5" eliminated nearly all α-synuclein without affecting other cellular proteins.

"Instead of needing a specific binding site, they can be engineered to recognize and cut α-synuclein directly, preventing it from dangerously accumulating in the brain," said Sondermann. "Botox has been used for decades with minimal immune reaction, which makes a promising starting point for a therapeutic enzyme."

What this means for the future of protein design

These three studies, published by different teams in different journals, converge in one clear direction: the combination of generative AI models with experimental evolution is transforming proteases from research tools into programmable therapeutic platforms.

The most important point is stability. A fundamental problem of enzyme design has always been that, when acquiring new functions, enzymes lose structural stability. Natural versions are already at the limit of their evolutionary robustness — there is no margin for major changes. AI-redesigned versions, on the other hand, possess excess stability. This "thermodynamic slack" allows enzymes to absorb mutations without destabilizing, exploring a much larger sequence space.

This is not just an incremental improvement. It is a paradigm shift in how we think about enzymatic engineering: instead of "starting with nature and improving what works," we begin with an AI-optimized version of nature and explore possibilities that natural evolution never produced.

The question now is not whether this approach will be adopted by the pharmaceutical industry — it already is — but which disease it will be the first to treat effectively, and how long it will take to reach patients.

Sources: Broad Institute, EurekAlert / Scripps Research, MIT News

✓ Independent sources cross-checked and verified before publishing