Enzymes are proteins that accelerate chemical reactions. Designing a new one requires finding both a three-dimensional structure and an amino-acid sequence that can arrange the components needed for a target reaction. AlphaProtein Novo’s core diffusion model tackles those pieces together, generating protein structures and sequences around a catalytic motif and the molecule it needs to interact with.
A pipeline, not a single model
The project repository describes an end-to-end workflow. After the diffusion model generates candidates, researchers can optionally redesign their sequences with LigandMPNN, predict their structures with AlphaFold 3 and run evaluation metrics. The repository’s pipeline script runs those stages in sequence.
The public repository includes code and example design campaigns. Its pretrained weights must be downloaded separately and are governed by their own terms.
Google DeepMind researcher Zach Wu said the team developed the system on several standard model reactions and validated designs in its own wet lab before asking biocatalysis scientists to propose harder, less established reactions.
What the team says it achieved
Researcher Jue Wang said AlphaProtein Novo reached state-of-the-art “novel-scaffold” activity on two benchmark reactions.
Wang also said the team designed enzymes aimed at two practical reactions: synthesizing the pharmaceutical motif piperidine and degrading the plasticizer DEHP. These are research results presented by the project team and linked preprints, not evidence of a deployed manufacturing or cleanup system.
Arnold described the work as a collaboration to make “new-to-nature” enzymes. The release gives researchers a documented codebase for exploring the approach, while reproducing the full workflow requires separate model weights and the supporting tools named by the repository.
