Ivan Kairatov is a distinguished figure in the biopharmaceutical sector, recognized for his profound expertise in merging computational deep learning with practical protein engineering. With a career rooted in high-stakes research and development, Kairatov has spent years exploring how technological innovations can overcome the biological hurdles that often stall drug discovery. His insights into the intersection of artificial intelligence and enzyme evolution provide a roadmap for the future of therapeutic design, particularly in the creation of highly specific proteases for neurodegenerative diseases. By bridging the gap between theoretical AI models and the messy reality of the laboratory, Kairatov offers a unique perspective on how we can now push enzymes far beyond the limits imposed by natural selection.
How does utilizing deep-learning models like ProteinMPNN to stabilize an enzyme’s starting structure fundamentally change the trajectory of an evolutionary campaign compared to traditional methods?
In traditional enzyme engineering, we often feel like we are walking a tightrope where every step toward a new function risks the total collapse of the protein’s structure. Most natural enzymes, like the wild-type botulinum neurotoxin BoNT/E, exist with only marginal biophysical stability, meaning they barely hold themselves together under physiological conditions. When we introduce mutations to change what the enzyme does, those mutations are frequently destabilizing, causing the protein to lose its structural integrity before it ever reaches the desired level of activity. By using ProteinMPNN and the PROSS tool, we can essentially rebuild the protein’s internal scaffolding before the evolution even begins, creating a much more robust starting point. In this study, we saw that 78% of the 74 redesigned variants retained their catalytic activity, which is an incredible success rate for AI-based design. This approach allows us to bypass the common “stability-activity trade-off,” providing a resilient framework that can survive the rigorous mutations required to develop entirely new therapeutic functions.
Could you walk us through the specific performance gains observed when these redesigned BoNT/E proteases were put to the test against their wild-type counterparts?
The performance gap between the redesigned enzymes and the wild-type versions was nothing short of transformative. For instance, the variant we labeled D2 achieved a catalytic efficiency of 310 mM⁻¹s⁻¹, which is nearly triple the 110 mM⁻¹s⁻¹ efficiency we recorded for the wild-type BoNT/E. This isn’t just a theoretical improvement; it translated into a massive increase in thermal stability, with some redesigned enzymes reaching melting temperatures as high as 59.5°C. When we moved these designs into human HEK293T cells, the results were even more striking, as we saw a 24-fold increase in protein expression for the D2 and D3 variants. Even when we tested a kinetically impaired redesign like D4—which started out 20 times slower than the wild-type—it eventually evolved to a higher final activity because its superior stability allowed it to explore a much wider range of beneficial mutations.
The study mentions using 44 parallel continuous evolution campaigns on an automated eVOLVER platform; what does this high-throughput approach reveal about the “evolvability” of AI-redesigned proteins?
Using the automated eVOLVER platform allowed us to watch evolution happen in real-time across dozens of different scenarios, providing a level of statistical depth that manual experiments simply cannot match. We challenged the enzymes with increasingly difficult substrates, specifically the SNAP25 variants 415, 413, and 412, to see where they would break. On the most challenging substrate, 412, the wild-type evolution campaigns failed in 50% of the lagoons, literally hitting an evolutionary dead end where they could no longer adapt. In contrast, every single lagoon containing a redesigned starting point succeeded, proving that AI-redesigned enzymes have a significantly expanded fitness landscape. This success is partly due to their ability to tolerate “forbidden” mutations, like K225E, which provide high function but would completely destroy the activity of a standard wild-type enzyme.
What are the therapeutic implications of successfully reprogramming a protease to target human ataxin-2, and how does the AI-driven approach ensure such high specificity?
Reprogramming a protease to target ataxin-2 is a major milestone because this protein is a primary driver in neurodegenerative conditions like ALS. The challenge is ensuring the enzyme only cuts the target protein and ignores the native SNAP25 substrates that are vital for normal cellular function. Through AI-assisted redesign and subsequent evolution, we produced a variant that was 79-fold more specific for ataxin-2 than the best enzyme evolved from a wild-type starting point. This specific variant showcased a 16% sequence divergence from the natural protein framework, a level of change that is rarely achievable through standard methods without losing all function. Most impressively, even at high concentrations of 50 µM, our top-evolved redesign showed no detectable cleavage of the original native substrate, suggesting we can create “surgical” enzymes that minimize off-target risks in the human body.
How does this integration of AI and continuous evolution address the long-standing problem of proteins losing their structural integrity as they gain new functions?
The core of the problem has always been that nature evolves proteins for survival, not for our industrial or medical needs, leaving them with very little “buffer” to handle new mutations. By setting structural distance constraints—specifically between 10 and 18 Å from the substrate and catalytic zinc ions—we used AI to preserve the essential machinery of the enzyme while optimizing the rest of the sequence. We also utilized multiple-sequence alignment conservation thresholds between 30% and 60% to ensure we weren’t straying too far from viable biological patterns. This creates a “super-stabilized” enzyme that can absorb the shock of destabilizing but functional mutations that would otherwise cause a protein to unfold. Essentially, we are giving the evolution process a much larger “budget” of stability to spend on acquiring complex, non-native catalytic activities that were previously out of reach.
What is your forecast for the scalability of this AI-redesign framework across other enzyme families beyond botulinum neurotoxins?
I believe we are standing at the threshold of a new era where we can treat enzyme engineering as a predictable modular process rather than a game of chance. While this study focused on BoNT proteases, the principles of using ProteinMPNN to establish high-stability starting points are theoretically applicable to almost any enzyme family, from lipases used in green chemistry to polymerases for advanced diagnostics. The next major hurdle will be validating these AI-enhanced enzymes in complex disease models to ensure their delivery, efficacy, and safety are as robust as their catalytic rates. If we can replicate the 79-fold specificity gains we saw with ataxin-2 in other targets, we could rapidly build a library of customized therapeutic enzymes that are safer and more potent than anything currently found in nature. We are moving toward a future where we don’t just find enzymes; we design the exact evolutionary path they need to take to solve a specific human problem.
