DeepMind has moved decisively beyond reading the language of proteins to writing it. In a paper published in Nature this week, the Google-owned lab reports that its updated protein-design system, AlphaProteo 2, has generated entirely novel enzymes that, once synthesised and tested in the laboratory, broke down common plastics faster and more durably than the natural enzymes biotechnologists have relied on for years. It is a result that marks a clear departure from the AlphaFold era of structure prediction and edges artificial intelligence into the business of designing functional biology to order.
The headline figure is striking. Several of the model’s top candidate enzymes degraded polyethylene terephthalate (PET) — the plastic used in drinks bottles and food packaging — at rates the authors describe as substantially higher than PETase and its engineered descendants, while remaining stable at the higher temperatures that make industrial recycling economically viable. Crucially, these were not tweaks to existing proteins but sequences with no close natural analogue.
From prediction to creation
AlphaFold, the program that won DeepMind’s John Jumper and Demis Hassabis a share of the 2024 Nobel Prize in Chemistry, answered a decades-old question: given a protein’s amino-acid sequence, what shape will it fold into? That was a read-only achievement, illuminating the natural world rather than extending it.
AlphaProteo 2 inverts the problem. Rather than predicting structure from sequence, it generates new sequences engineered to perform a specific job — in this case, gripping and cleaving the molecular bonds that hold PET together. The first version of AlphaProteo, released in 2024, focused on designing proteins that bind tightly to a target. The new model adds catalytic function, the far harder trick of designing a protein that actively transforms another molecule.
“Binding is a handshake; catalysis is a dance,” said Dr Priya Venkataraman, a computational biochemist at Imperial College London who was not involved in the work. “Designing a stable active site that holds a substrate in exactly the right geometry and then does chemistry to it — and doing that from scratch — is the problem the field has been circling for twenty years. These results suggest the circling is over.”
Why plastic-eating enzymes matter
The choice of target is no accident. Enzymatic recycling has become one of the most watched corners of green chemistry, promising a way to break used plastic back down into its chemical building blocks for reuse, rather than downcycling it into lower-grade material or burning it. Natural and engineered PET-degrading enzymes already exist, but they tend to be slow, fragile, or quickly inactivated under real-world conditions.
DeepMind says its designed enzymes retained activity through repeated cycles and elevated temperatures, the two factors that most often sink a process before it reaches commercial scale. The lab partnered with academic groups to express the proteins in bacteria and validate their performance — an important detail, given the long history of computational predictions that fail to survive contact with a test tube.
- Speed: top candidates degraded PET measurably faster than benchmark engineered enzymes.
- Stability: activity maintained at temperatures relevant to industrial recycling.
- Novelty: sequences with little resemblance to known natural enzymes, suggesting the model is exploring genuinely new design space.
A generative biology toolkit emerges
The broader significance lies less in plastic than in the precedent. If a model can design a working enzyme for one reaction, the same approach could in principle be aimed at others: enzymes to manufacture pharmaceuticals more cleanly, to capture carbon, or to break down so-called forever chemicals.
That prospect is already reshaping how analysts view DeepMind’s spin-out, Isomorphic Labs, and the wider commercial landscape.
“AlphaFold gave away a map of the natural protein universe for free,” said Tom Eldridge, a biotech analyst at the London advisory firm Kestrel Partners. “A generative design engine is a different proposition commercially. The intellectual property doesn’t sit in nature — it sits in the sequences the model invents. That is where the value, and the licensing fights, will concentrate.”
Experts urge some caution. Designing a handful of high-performing enzymes against a well-studied target is not the same as reliably producing them for arbitrary, harder reactions. Independent replication, and tests on the messy mixed-plastic waste streams found in real recycling plants, will be the true measure.
What this means
AlphaProteo 2 signals that AI-driven biology is shifting from describing life to engineering it. For the recycling industry, validated enzymes that beat nature could accelerate a credible route to closing the loop on plastic waste — though scale-up and cost remain unproven. For the wider field, the message is larger still: the bottleneck in molecular design is moving from imagination to manufacturing and validation. If DeepMind’s results hold up under independent scrutiny, the question for biotechnology will no longer be whether we can design proteins to do useful new things, but how quickly, how cheaply, and under whose control.
Photo by Google DeepMind on Pexels
