HeyHomees Lab Notes · No. 3 · Google DeepMind · Published October 16, 2026
By HeyHomees Newsroom ·
In 30 seconds
- Runs a diffusion model to co-generate protein structures and sequences conditioned on a motif and ligand context
- Optional stages include LigandMPNN sequence redesign and AlphaFold 3 structure prediction
- Pretrained model weights must be downloaded separately from Google Cloud Storage
What is it?
AlphaProtein Novo (AP Novo) is a generative diffusion pipeline for de novo enzyme design via structural motif scaffolding. It co-generates protein structures and sequences conditioned on a catalytic motif and ligand context, and can be run with the provided run_generator.py script.
How does it work?
The diffusion model generates structures and sequences, then optional LigandMPNN can redesign sequences, and AlphaFold 3 can predict folded structures. The full workflow is orchestrated by run_pipeline.py, which executes all stages in sequence from a JSON manifest.
What does it need to run?
Two Python environments are recommended: a primary environment with JAX for diffusion, folding, and evaluation, and an optional LigandMPNN environment with PyTorch for sequence redesign. Pretrained weights (generator.bin.zst and af3_leaving_atom.bin.zst) must be downloaded and placed in ./models directories.
By the numbers
- Recommended Python environments: Two
- Pretrained generator weight file: generator.bin.zst
- AlphaFold 3 leaving-atom weight file: af3_leaving_atom.bin.zst
- Default generator model directory: ./models/apnovo_generator
The honest read
The release provides counts of environments and file names for required pretrained weights, but it does not include any performance benchmarks, accuracy scores, or runtime measurements. All reported figures are self-described by the repository and the pretrained weights must be obtained separately under the terms of use, so the repository itself does not demonstrate the quality of generated enzymes.
What we don't know yet
- No benchmark results or quantitative evaluation of generated designs are provided
- Hardware requirements beyond needing a GPU are not specified
- Exact download size of the pretrained weight files is not stated
Sources
- README, GitHub (primary)
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.