Google DeepMind says its new system designed an enzyme that makes a key drug ingredient 99 times faster than the industrial process chemists currently rely on. That's not an improvement on nature. Nature never built this enzyme at all.
Pushmeet Kohli, DeepMind's vice president of science and strategic initiatives, announced a preprint for a system called AlphaProtein Novo on X on October 5. He described it as a generative model for designing enzymes from scratch, rather than tweaking ones that already exist in biology. The team built two custom enzymes to prove it works. One synthesizes piperidine, a molecular ring structure found in a huge share of pharmaceutical compounds. The other breaks down DEHP, a phthalate plasticizer regulators have flagged for links to reproductive harm and liver damage, and that the EPA now considers a probable carcinogen.
Neither enzyme exists anywhere in nature. That's the point. Enzyme engineers have spent two decades mutating natural proteins, nudging an existing fold a few amino acids at a time until it does something slightly different or slightly faster. AlphaProtein Novo instead generates an entirely new protein sequence and structure built around a target chemical reaction, with no natural template to start from. DeepMind calls designing a functional enzyme this way, from a blank sheet, one of the field's long-standing grand challenges, and says the new system hit state-of-the-art activity on both benchmark reactions it was tested against.
Piperidine isn't an exotic molecule. It's a six-membered ring that shows up as a building block in a long list of drugs, from antihistamines to antipsychotics, which is exactly why DeepMind picked it as a proving ground. An enzyme that produces it at 99 times the rate of the existing industrial method would matter to any manufacturer currently running that synthesis through conventional chemistry. That typically means higher temperatures, metal catalysts, and more purification steps than a clean enzymatic reaction needs. Cut the steps and you cut the cost and the time to get a drug ingredient out the far end of a plant.
Isomorphic Labs is DeepMind's drug-discovery spinout, built on the AlphaFold lineage. It raised $2.1 billion in a round that closed earlier this year: backers included Thrive Capital, Alphabet, Abu Dhabi's MGX and Singapore's Temasek, pushing its total raised to roughly $2.6 billion. That company is aiming to get its first AI-designed drug candidates into clinical trials by the end of 2026, working with partners including Novartis, Eli Lilly and Johnson & Johnson. AlphaProtein Novo is a separate research effort. But it lands in the same world: AI systems that don't just predict how a molecule folds, but design new ones that do a specific job better than anything chemistry or biology has produced so far.
US and 15 Nations Sign Kyoto Pledge to Put AI at the Center of Science Michael Kratsios of the White House's OSTP wrote the Kyoto statement, timed to the STS forum opening October 4 in Kyoto, Japan, with Japan, South Korea, the UK and the UAE among the signatories naming AI-run labs as the goal. - AI and science research international agreement signed today - governments commit to AI powered scientific experiments globally
The second test case points somewhere different entirely. DEHP is the most common phthalate used to make rigid PVC flexible, and it's everywhere: medical tubing, vinyl flooring, food packaging. It leaches out of that plastic into air, dust and soil, and the CDC's toxicology profile on the chemical lists neurological effects, oxidative stress and reproductive harm among the documented risks. DeepMind's enzyme was designed to break it down. Natural phthalate-degrading enzymes exist, but they tend to lose activity at anything above room temperature, which makes them close to useless in an actual industrial recycling or remediation line that runs hot.
That's the detail worth sitting with. A plastics recycler doesn't operate at room temperature. If DeepMind's enzyme keeps working in the heat where natural enzymes quit, it closes a gap that has kept enzymatic plastic breakdown mostly confined to lab benches rather than factory floors.
The same research lineage already has a track record here. DeepMind previously worked with the University of Portsmouth's Centre for Enzyme Innovation, using AlphaFold to help researchers screen around 100 candidate enzymes. It generated 3D structures for PET-degrading enzymes in days rather than months, feeding directly into the kind of engineered PET hydrolases that labs like UT Austin's have since pushed toward commercial plastic recycling.
AlphaProtein Novo's enzyme work was done with external collaborators, including a group at the Francis Crick Institute. That's the same pattern DeepMind used to validate its AlphaProteo protein-binder system last year, when outside labs confirmed binders designed by the AI could block SARS-CoV-2 variants from infecting cells. Independent validation matters more here than in most AI announcements, because a sequence nobody has ever synthesized before carries no track record. The only way to know it works is to make it in a lab and watch the reaction happen. That's exactly what both the piperidine and DEHP results represent.
None of this means AI-designed enzymes are heading into a factory next quarter. A preprint is not a peer-reviewed paper, and a successful benchmark reaction is not the same as a validated, scaled-up industrial process running for months without degrading. But the direction is now hard to miss. Two years ago, AlphaFold's achievement was predicting what already existed. This one is manufacturing what never did, and pointing it at two problems, drug costs and plastic pollution, that investors in AI-driven biotech have been circling for a while.
Also read: Reflection AI unveils Beam, a 501-billion-parameter open model to rival China • Reflection AI launches Beam, an open model it says beats the West on efficiency • ReleasePad Wants Your AI Coding Assistant To Write Your Release Notes Too
Ilya Sutskever's SSI Has Raised $8 Billion and Shipped Nothing At All Ilya Sutskever's Safe Superintelligence has raised about $8 billion, including a fresh $5 billion from Nvidia, without shipping a single product. Sutskever says the industry's scaling era is over and SSI is betting on something else entirely, but two years in, the company still hasn't shown the world what that something is. - how to price a SaaS product for enterprise - cold email template that gets replies from investors
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.