Google's AI Lab Designs Brand-New Enzymes From Scratch, Including Ones for Chemistry Nature Never Evolved
Google DeepMind and chemists at Caltech posted a study today describing an AI program that invents enzymes, the tiny natural machines that speed up chemical reactions. Its designs made a common building block of medicines far more selectively than any natural enzyme tested, and others broke down a pollutant in conditions that ruin natural ones. The results haven't been reviewed by outside scientists yet, and the new enzymes are still much weaker than the best natural ones.
Google DeepMind, Google's AI research lab, posted a research paper online today, working with chemists at Caltech, including Nobel Prize winner Frances Arnold, and at the University of Pittsburgh. It describes AlphaProtein Novo, an AI program that designs enzymes from scratch.
Enzymes are proteins that speed up chemical reactions. Every living thing depends on them, for jobs like digesting food, and industry uses them to make medicines, detergents and other products. Scientists usually start with an enzyme found in nature and adjust it. That's hard when nature never evolved an enzyme for the job.
Here is how the new approach works. A chemist describes the small working core an enzyme would need: the few pieces that grip the chemicals and carry out the reaction. The AI program builds a whole protein around that core. Another DeepMind program, AlphaFold, which predicts the 3D shape a protein will fold into, checks which designs are likely to hold together. The most promising ones are then made and tested in a real lab. The team tested more than 5,600 designs across five different reactions.
Two results stand out. The first is piperidine, a small ring of atoms found in many medicines. The natural enzymes that can do this kind of reaction tend to make a slightly smaller ring instead. Of 188 natural and lab-improved enzymes the team checked, none made piperidine more than 30% of the time. The best AI design made it 99 times out of 100. The second is DEHP, a chemical added to plastics to make them soft that also pollutes the environment. The AI produced several new families of enzymes that break it down, and one kept working at 90°C (194°F) and in a harsh solvent that ruins natural enzymes.
The researchers are open about the limits. This is a preprint, a paper shared before other scientists have reviewed it. The new enzymes are still far less active than the best enzymes found in nature or improved over years in the lab. And an expert still has to design the working core by hand. "Often, multiple catalytic motifs must be explored before finding one capable of supporting the desired chemistry," the authors write, meaning several cores often have to be tried before one works. The authors call designing enzymes from scratch "a powerful complement to natural enzyme diversity," not a replacement for it. DeepMind has made the program free for non-commercial use, and Google-affiliated companies have filed for a patent on related methods.
Sources
- bioRxiv — Designing enzymes for new-to-nature chemistry and non-natural substrates with AlphaProtein Novo (preprint)biorxiv.org · primary
- Google DeepMind — AlphaProtein Novo code release (GitHub)raw.githubusercontent.com · primary
- MPost — AlphaProtein Novo designs enzymes for reactions nature never evolvedmpost.io
- Aliquant (Daniel Buchan) — Engineering enzymes for everythingaliquant.substack.com