DeepMind's AI designs brand-new enzymes from scratch

AlphaProtein Novo is an open-source machine learning pipeline from Google DeepMind that designs entirely new enzymes from scratch.

Instead of searching natural databases and tweaking existing proteins through endless rounds of trial and error, this pipeline lets researchers start from first principles to generate custom biocatalysts for reactions nature never invented.

What it can do

AlphaProtein Novo designs both the 3D backbone structure and the corresponding amino acid sequences simultaneously.

In wet lab validations, researchers used it to create new-to-nature nitrene transferases that synthesize piperidine, an essential pharmacophore in pharmaceutical manufacturing, with unprecedented product selectivity.

In another real-world test, it designed enzymes capable of degrading DEHP, a toxic plasticizer and persistent environmental pollutant. The engineered enzyme successfully broke down DEHP under harsh conditions that normally denature natural enzymes. It also achieved state-of-the-art catalytic efficiency across two widely benchmarked model reactions.

How it works and availability

Previous de novo enzyme designs struggled with low catalytic activity. AlphaProtein Novo solves this by integrating AlphaFold 3 predictions directly into its iterative design loop.

The system evaluates backbones using mechanism-inspired metrics derived from AlphaFold 3 alongside sequence-ensemble scoring. It continually predicts how the designed protein will fold and how active sites bind to non-natural substrates before any wet-lab synthesis takes place.

The underlying study is currently hosted as a preprint on bioRxiv and has not yet completed peer review. Google DeepMind has already open-sourced the repository on GitHub, complete with download scripts and setup instructions for running the pipeline.

Why it matters

For computational biologists, synthetic biology startups, and drug developers, AlphaProtein Novo is one of the first de novo design toolsets to prove itself with meaningful catalytic activity in wet labs. If you are struggling with fragile or nonexistent natural enzymes for a non-natural substrate, this repo is worth pulling down today.

We think this shifts generative biology from passive structure prediction to active functional engineering. The main caveat to keep in mind is that the research is still an unreviewed preprint, meaning researchers will still need substantial wet-lab validation and expression tuning before taking these custom enzymes to pilot production.