AI Protein "Scissors": Baker Lab Creates Proteases with a 10-Billion-Fold Reaction Acceleration

Scientists at David Baker's lab at the University of Washington have announced the creation of artificial proteases—enzymes capable of cleaving proteins at precisely targeted sites with unprecedented speed. In one of the developed design variants, catalytic activity reached a level where the rate of peptide bond cleavage is accelerated by 10 billion times (10¹⁰) relative to the background reaction level. The team also demonstrated that such enzymes can target proteins associated with neurodegenerative diseases, including Alzheimer's disease and frontotemporal dementia. The work, published in October 2026, represents one of the most striking examples of generative artificial intelligence applied to protein engineering.

The Peptide Bond and a Challenge Nature Left Unanswered

Proteins are chains of amino acids linked by peptide bonds. In an aqueous environment, these bonds can persist for hundreds of years without enzymatic assistance, making them extremely stable. Proteases—natural catalysts that accelerate the cleavage of such bonds—recognize specific regions of a protein sequence. The challenge set by Baker's team was to use AI to design catalysts with pre-determined three-dimensional structures and the ability to recognize specific amino acid sequences, rather than being limited to modifying existing natural enzymes.

Three Neural Network Tools Combined

To design the enzymes, the researchers employed a pipeline of three AI tools. RFdiffusion2 generated the backbone of the protein structure, taking into account the arrangement of amino acids required for the chemical reaction and the designated cleavage site. Next, ProteinMPNN selected specific amino acids to ensure proper protein folding, active site function, and interaction with the target molecule. The resulting structures were evaluated using AlphaFold3. Two types of natural proteases served as templates: a zinc-based protease found in the digestive system of cancer cells and a cysteine-based protease derived from yeast. Based on these, the team developed dozens of design variants.

Not Just One Structure: The Path to 10¹⁰

Initial results showed that a single correct three-dimensional structure is not enough for an enzyme to function effectively. Among 69 cysteine protease variants, only 13—about 19%—proved functional. In the series of zinc proteases, initial designs often showed weak catalytic activity. This problem was partially solved by defining the position of the water molecule involved in the reaction: 35 out of 38 variants designed this way showed activity. Further optimization—altering amino acids near the active site that are not directly involved in catalysis—achieved record metrics. The best cysteine protease variant accelerated peptide bond cleavage by a factor of 10⁹ while maintaining high selectivity. The improved zinc protease became 30 times more active than the original design and accelerated the reaction by 10¹⁰ times. However, high speed alone did not guarantee accuracy: one of the most active zinc protease variants cleaved proteins with insufficient selectivity, highlighting the complexity of simultaneously optimizing catalytic efficiency and specificity.

Target Targets: Alzheimer's and Frontotemporal Dementia

As a separate test, the team developed zinc proteases targeted against proteins associated with Alzheimer's disease and frontotemporal dementia. As a result, they successfully obtained an enzyme capable of cleaving the target protein directly inside cells. This demonstrates that the approach is applicable to real therapeutic targets, although a significant path remains before clinical application.

Barriers on the Road to Medicine

Several significant hurdles remain before practical medical application. Proteins are difficult to deliver into the correct tissues and cells, and artificial enzymes themselves can trigger an immune response. If the target is located inside a cell, delivering genetic instructions so the cell produces the required protease itself may be necessary. Furthermore, current AI tools excel at structural design but cannot directly specify the desired catalytic activity—it remains the result of iterative optimization. Nevertheless, the work proves the fundamental feasibility of creating artificial enzymes for specific molecular targets, while their efficacy and suitability for treatment, according to the authors, still need to be improved and verified in further experiments.