When the Starting Point Is the Limit

Here's a problem that's been haunting protein engineers for decades: enzymes that gain new functions tend to lose stability. It's a classic trade-off — and it's been a real bottleneck for anyone trying to design therapeutic enzymes from scratch.

Most mutations that give an enzyme a new ability — say, to cleave a disease-related protein it doesn't normally touch — are thermodynamically destabilizing. They make the protein structure less robust. And when the starting enzyme is already marginally stable (as wild-type enzymes often are), it simply can't survive the journey. It collapses before it gets there.

Now, a study published in Nature suggests that AI can fundamentally change this equation. By redesigning enzyme starting points before evolution begins, researchers were able to access mutational landscapes that were unreachable from natural wild-type sequences — and evolve proteases with specificity improvements of up to 79-fold compared to what wild-type starting points achieved.

The implications extend far beyond the specific enzyme family studied here. If this approach generalizes, it could become a blueprint for engineering therapeutic enzymes that target disease-linked proteins with unprecedented precision.

---

The Approach: Stabilize First, Evolve Later

The study's logic is elegant: instead of accepting the enzyme nature gives you and trying to evolve it under duress, use AI to redesign the starting sequence so it's structurally robust. Then, let evolution do what evolution does — but from a position of strength.

Two computational tools carry the heavy lifting here:

• ProteinMPNN: a deep-learning model that generates new amino acid sequences predicted to fold into a specified structure. Think of it as reverse-engineering the protein — you give it the shape, it suggests sequences that should maintain that shape while changing much of the underlying code.
• PROSS (Protein Repair One-Stop Shop): a computational stabilization method that predicts mutations to increase a protein's thermodynamic stability. It was developed at the Weizmann Institute of Science — so yes, there's an Israeli angle to this story.

The researchers applied these tools to redesign the catalytic domains of botulinum neurotoxin (BoNT) proteases — specifically BoNT/E, BoNT/F, and BoNT/X. These are the enzymes behind botulinum toxin, one of the most potent biological toxins known. Why use them as a model? Because their structure and biochemistry are extremely well characterized, making them an ideal testbed for whether redesigned starting points actually outperform wild-type in evolution.

The redesign process imposed specific constraints: residues within a certain distance of the substrate binding site and catalytic zinc ion were held fixed, while residues that were highly conserved across homologous sequences were also protected from change. Everything else was fair game for AI to rewire.

---

The Numbers: What AI Redesign Actually Achieved

The researchers generated 74 ProteinMPNN redesigns of the BoNT/E protease. Here's what happened:

• 78% retained catalytic activity — a strong hit rate given how radically the sequences were altered.
• Top redesigns (designated D1–D3) showed catalytic efficiencies 1.7 to 2.8 times higher than the wild-type enzyme. The best performer, D2, achieved a catalytic efficiency (kcat/KM) of 310 mM⁻¹s⁻¹ compared to 110 mM⁻¹s⁻¹ for the wild-type. That's nearly triple.
• Thermal stability jumped significantly, with melting temperatures (Tm) reaching up to 59.5°C. For context, wild-type BoNT/E has a meaningfully lower Tm, and this stability improvement is what enables the protease to tolerate mutations that would destroy the original.

The redesigned proteases also expressed far better in bacterial cells — practical evidence that the structural stabilization translates to real-world yield. When combined with mutations from a previously evolved protease, the D2 and D3 designs showed over 24-fold improvement in HEK293T cell expression.

---

44 Parallel Evolution Campaigns: The Real Test

Here's where it gets interesting. The redesigned starting points weren't just better enzymes in a test tube — they proved to be superior launching pads for directed evolution.

The researchers ran 44 parallel continuous evolution campaigns on an automated eVOLVER platform, comparing wild-type BoNT/E proteases against their redesigned counterparts. The challenge: evolve proteases that could cleave increasingly difficult variants of the natural substrate SNAP25 (designated substrates 415, 413, and 412).

Against the most challenging substrate (412), wild-type evolution campaigns failed 50% of the time. Redesigned starting points? All succeeded.

The reason comes back to that stability-activity trade-off. Wild-type proteases hit a stability wall — the mutations needed for new activity destabilized them too much. Redesigned proteases had enough structural margin to tolerate those destabilizing mutations.

One particularly striking finding: a mutation called K225E, which conferred high catalytic function, was non-functional when introduced into the wild-type background — the protein simply couldn't handle it. In the redesigned background, it worked.

Even more counterintuitive: a redesigned variant (D4) that started out 20-fold slower than wild-type still evolved to a higher final activity than wild-type campaigns achieved. Starting slower but more stable beat starting fast but fragile.

---

The Therapeutic Breakthrough: Targeting Ataxin-2

The study's most clinically significant demonstration involved repurposing BoNT/E to cleave human ataxin-2 — a protein implicated in neurodegenerative diseases including ALS.

Starting from the D3 redesigned protease, the researchers evolved variants that could recognize and cut this entirely new substrate. The top result, designated D3(428)2, showed remarkable specificity:

• 79-fold greater specificity for ataxin-2 compared to the best protease evolved from the wild-type starting point.
• No detectable cleavage of the original substrate (SNAP25) even at the highest concentration tested (50 μM) — meaning the evolved protease had effectively lost all activity against its natural target while gaining precision for the new one.
• 16% sequence divergence from the natural BoNT/E framework — the evolved protease is substantially different from the toxin it was derived from.

That 16% figure deserves a moment. It means this isn't just a slightly modified toxin — it's a genuinely novel enzyme, guided by AI and shaped by evolution to do something the original could never do.

---

Why Stability Opens the Evolutionary Floodgate

The core insight of this study is conceptual, not just empirical: stability is a prerequisite for evolvability.

Think of it this way. A marginally stable protein is like a car with a worn-out engine — it runs, but any modification risks breaking it entirely. A redesigned, stabilized protein is like a car with a reinforced engine and chassis. You can modify it aggressively, push it hard, and it keeps going.

The researchers showed this through multiple lines of evidence:

1. Redesigned starting points accessed mutational combinations that were non-functional in the wild-type background.
2. Even kinetically slower redesigned proteases outperformed wild-type in evolution campaigns.
3. Destabilizing but functionally beneficial mutations (like K225E) only worked in the redesigned context.

This suggests that wild-type enzymes aren't just suboptimal starting points — they're constrained starting points, trapped in a limited region of mutational space by their own marginal stability. AI redesign breaks them out of that trap.

---

The Israeli Connection

For anyone in the local biotech or computational biology scene, there's a relevant detail here. The PROSS computational tool — one of the two key AI methods used in this study — was developed at the Weizmann Institute of Science. Its webserver is hosted at pross.weizmann.ac.il.

PROSS works by analyzing a multiple sequence alignment of homologous proteins and sampling mutations predicted to be stabilizing. In this study, it was applied to BoNT/E using the same structural constraints as ProteinMPNN, serving as a complementary approach to the deep-learning model.

The Weizmann Institute has a strong track record in computational protein design, and PROSS has been widely adopted across the protein engineering community. Its role in a high-profile Nature paper is another data point in a broader trend: Israeli computational biology infrastructure is increasingly integral to cutting-edge international research — even when the primary labs are based elsewhere.

---

What This Means for Protein Engineering

The implications of this work extend across several domains:

Therapeutic enzyme development. If you can engineer a protease that specifically targets ataxin-2 (a neurodegeneration-linked protein) with 79-fold improved specificity, the same approach could potentially be applied to other disease targets. The bottleneck has always been whether engineered enzymes can achieve sufficient specificity and stability for clinical use. This study suggests AI redesign can substantially raise that ceiling.

Directed evolution as a platform. Phage-assisted continuous evolution (PACE) is already a powerful tool, but it's been limited by the quality of starting points. AI redesign could dramatically expand what's achievable through PACE and similar directed evolution methods.

The stability-activity paradigm. For decades, protein engineers have accepted the stability-activity trade-off as a fundamental constraint. This study suggests it's not fundamental — it's a limitation of starting from wild-type sequences. Redesign the starting point, and the trade-off largely dissolves.

Scalability. The computational tools used here (ProteinMPNN and PROSS) are publicly available and relatively fast to run. The eVOLVER platform for automated evolution is also becoming more accessible. The workflow described in this study could, in principle, be applied by any well-equipped protein engineering lab.

---

What Comes Next

The authors are careful to note that their work was demonstrated primarily on BoNT protease families. Whether the advantages of AI-redesigned starting points extend to unrelated enzyme families remains to be verified. Different protein architectures may respond differently to the redesign process.

There are also practical questions about delivery, efficacy, and safety in therapeutic contexts. Engineering a protease that can cleave ataxin-2 in a test tube is one thing; delivering it to the right cells in a patient, at the right concentration, without off-target effects, is another.

But the proof-of-concept is strong. And the underlying principle — that AI can create starting points that evolution then optimizes — is likely to be a recurring theme in protein engineering for years to come.

If you're working on computational biology, therapeutic protein design, or enzyme engineering, this paper is worth a close read. The methods are detailed enough to be reproducible, and the results are clear enough to be convincing.

As one reviewer might put it: sometimes the best way to go fast is to start by going steady.