DeepMind and Frances Arnold designed enzymes from scratch that outperformed natural enzymes in reactions they struggled to catalyze
DeepMind and Frances Arnold designed enzymes from scratch that outperformed natural enzymes in reactions they struggled to catalyze
On 5 October 2026, Google DeepMind and Frances Arnold’s laboratory introduced AlphaProtein Novo (AP Novo) in a paper on bioRxiv. The system builds an enzyme from scratch to catalyze a specific reaction. In two practical applications, piperidine synthesis and the breakdown of the toxic plastic additive DEHP, these designs outperformed screens of natural enzymes for the first time.
An enzyme is a protein catalyst that accelerates a specific reaction. Conventional enzyme engineering starts with a natural protein and improves it through random mutations using directed evolution, the method for which Arnold received the 2018 Nobel Prize in Chemistry. First, however, researchers must find a suitable natural protein, and for many reactions, none exists.
AlphaProtein Novo bypasses this search. A diffusion model, using the same technology that turns noise into images in AI image generators, builds a protein scaffold around the geometry of the active site and the target molecule. The scaffold is then assigned an amino acid sequence, and candidates are screened using predictions from AlphaFold 3, a program that predicts a protein’s shape from its sequence. The system was built by some members of the original AlphaFold team, whom DeepMind reassigned to enzyme design in July 2026.
The first application was the synthesis of piperidine, a ring structure found in many drugs. Of 188 natural and engineered proteins tested, none produced more than 30% piperidine alongside 70% byproduct: the available enzymes favored a different molecular structure. The best AP Novo design shifted that ratio to 99% to 1% and increased piperidine yield by two orders of magnitude relative to the control protein. In addition, 94% of the product consisted of one of the two mirror-image forms of the piperidine molecule. In drugs, usually only one form is active and safe; the other may be ineffective or harmful.
The second application was the breakdown of DEHP, one of the most widely used plastic additives, which has been linked to developmental and reproductive disorders. Very few natural enzymes can break it down. The molecule has bulky side chains and is insoluble in water, properties that make it particularly difficult for these enzymes to process. At room temperature, one AP Novo design underperformed two natural enzymes. For another design, heating to 90°C increased activity by 14 times, while a 75% acetonitrile solution doubled it. Both natural enzymes lost all activity under these conditions.
For years, previous attempts at de novo enzyme design encountered the same problem: an excellent scaffold could be paired with an unsuitable amino acid sequence, and testing could not distinguish between failure of the scaffold and failure of the sequence. The authors had the system evaluate several sequences for each scaffold and retained only scaffolds that passed the filter with every sequence. This separated scaffold quality from sequence quality. In two benchmark laboratory reactions, the proportion of functional designs increased 30 times to 44% and doubled to 23%, respectively. The team tested the hypothesis on three model reactions before selecting piperidine and DEHP as applications in which the available natural enzymes were predictably weak.
The authors state the limits of the result: these enzymes have activities orders of magnitude below those of natural catalysts and enzymes improved through directed evolution, and the model uses statistical selection to identify successful candidates without reproducing the physics of catalysis. Arnold herself is a coauthor, having developed the method that this work only partly bypasses.