AI can now help researchers imagine proteins that do not closely resemble anything found in nature. A recent enzyme-design effort shows why that matters for plastic digestion, and why the hardest part is not drawing a promising protein shape but making the whole chemical mechanism run again and again.
Why ester bonds matter for plastic digestion
The research focused on breaking an ester bond. In plain terms, that bond links two carbon chains through an oxygen atom, while one of the neighboring carbons is also attached to another oxygen.
Breaking the bond requires adding a water molecule. The result is two separated chemical pieces: one carbon chain ends up attached to an alcohol group, written as COH, and the other becomes an organic acid group, written as COOH.
Ester bonds are common in biology, so living systems already provide many examples of enzymes that can handle them. They also appear in large-scale plastic polymers. Polyester, for example, gets its name from the many instances of this bond in the material.
That makes ester-bond chemistry an attractive target for enzyme design. If researchers can build enzymes that break these bonds efficiently, they may open new ways to digest some plastics or work with reactions that biology has not naturally optimized for human uses.
The simple reaction with a complicated mechanism
The challenge is that the chemistry only looks simple from a distance. During the reaction, one piece of the ester can become chemically attached to an amino acid inside the enzyme. If that attachment is not broken in a later step, the enzyme stops working.
A true catalyst cannot merely participate once. It must return to a usable state and repeat the reaction. That is where the design problem becomes difficult.
According to the source, the process has at least four distinct stages. It also depends on several amino acids being placed inside the active site with atomic precision. One key amino acid must operate near the typical pH of living things, sometimes taking a proton from surrounding water and passing it to another amino acid, and at other moments removing a proton from an amino acid and transferring it to part of the ester.
For AI protein design, that distinction is crucial. Designing a protein that appears suited for one step is much easier than designing one that can move through every stage of the catalytic cycle.
How RFDiffusion and PLACER changed the search
The team began with protein-design tools that included RFDiffusion. That AI system uses a random seed to generate different protein backgrounds. Here, it was asked to match the average positions of amino acids found in a family of enzymes that break esters.
Those candidate designs were then passed to another neural network. Its job was to choose amino acids that could form a pocket for an ester. The chosen ester would break down into a fluorescent molecule, allowing the researchers to track activity by watching for glow.
The first results were narrow. Of the 129 proteins designed by this software, only two of them resulted in any fluorescence.
That led the team to add another AI system, PLACER. The software was trained using known structures of proteins attached to small molecules. Some of those structures were randomized, forcing the system to learn how to move them back toward a functional arrangement.
The aim was to capture more than one rigid protein state. Enzymes often need to support shifting configurations as a reaction proceeds, so the researchers hoped PLACER could screen for designs that were better suited to those structural changes.
Adding PLACER helped. Repeating the process with the extra screening step increased the number of enzymes with catalytic activity by over three-fold.
Activity was not enough
Even with that improvement, the first active designs had a serious limitation. They could cleave the ester, but then they stalled after a single reaction. One piece of the ester remained chemically bonded to the enzyme.
That meant the proteins were not yet acting as useful catalysts. They were effectively becoming part of the reaction and then getting stuck.
The researchers adjusted the screening strategy. Instead of only looking for structures suited to the starting arrangement, they used PLACER to evaluate whether designs could adopt a key intermediate state in the reaction.
This change produced a stronger result. In that round, 18 percent of the designs cleaved the ester bond. Two of them, named “super” and “win,” could cycle through multiple rounds of reactions. At that point, the team had produced enzymes rather than one-use reactive proteins.
The process did not stop there. By alternating additional rounds of structural suggestions from RFDiffusion with PLACER screening, the team increased the frequency of functional enzymes. Eventually, they designed one with activity similar to some enzymes made by living things.
They also showed that the same process could be used to design an esterase capable of digesting bonds in PET, a common plastic.
What this says about the future of enzyme design
The result is encouraging, but not because AI made enzyme design effortless. The opposite lesson is just as important: even when researchers know related enzymes from living systems, making a new working enzyme remains a serious challenge.
The useful shift is where more of the trial-and-error can happen. A larger share of the search can now be done on computers, before researchers need to order DNA, get bacteria to produce a protein, and test for activity.
The designed enzymes also did not share many sequences with the known enzymes used as references. That matters because it suggests flexibility. AI-guided protein design may be able to reach enzyme solutions that differ substantially from natural ones, including for esters that living things have never encountered.
The broader implication is measured rather than dramatic. AI protein design can generate new possibilities, and the plastic-related PET result shows a practical direction. But enzymes are not just static shapes. They are molecular machines that must guide several precise steps, recover from intermediate states, and keep working across multiple reaction cycles.
For plastic digestion and other difficult chemistry, the future may depend on combining protein generation, structural screening, and the kind of iterative testing shown here. AI has made the search more powerful. It has not made the chemistry simple.