Google DeepMind Publishes AlphaProtein Novo Enzyme-Design Pipeline
The public code connects motif-conditioned generation to redesign, folding and evaluation, but its weights and research claims require separate evidence.
Edited by Tyronne Panaino
Google DeepMind has published the AlphaProtein Novo repository, documenting a generative diffusion pipeline that co-generates protein structures and sequences around a catalytic motif and ligand context. As verified on October 12, the public package is aimed at computational protein-design work and connects generation with optional sequence redesign, structure prediction and problem-specific evaluation.
The release matters because it exposes the stages and configuration of the design workflow rather than presenting enzyme generation as a single opaque model call. The source also sets important limits: pretrained parameters are not bundled with the package, their use is governed separately, and the repository is software documentation rather than independent validation of research performance.
What the pipeline combines
AlphaProtein Novo begins with a three-dimensional arrangement of side-chain and ligand atoms representing a transition state or intermediate for a target reaction. Its design objective is to generate a protein that holds that catalytic motif in the intended conformation while producing a compatible surrounding structure and amino-acid sequence.
Google DeepMind's documented end-to-end path then adds optional sequence redesign through LigandMPNN, structure prediction through AlphaFold 3 and evaluation metrics selected for the design problem. A pipeline runner coordinates design generation, sequence design, folding and metric calculation from a single manifest, giving researchers a staged record of how a campaign is configured and processed.
How a design campaign is represented
The manifest separates run-level settings from default sampling values and individual design jobs. Additional sections control whether sequences are redesigned, which generated or redesigned structures are folded, which ligand-bound or unbound states are evaluated, and which problem-specific metric suite is applied.
Within a design job, researchers can describe fixed catalytic residues and ligand atoms, designable regions, chain layout, sequence-length constraints and whether the motif position is fixed in the sequence or chosen during generation. This makes the reaction geometry the central constraint while leaving selected parts of the surrounding protein available for generation or redesign.
What the examples show
The repository supplies example campaigns for Kemp eliminase and serine esterase designs, two DEHPase routes, carbene transfer and piperidine or nitrene transfer. Those examples cover both new scaffolds and partial diffusion from a parent design, and their manifests connect the requested chemistry to folding states and evaluation calculations.
The examples are useful for understanding the interface, but they are not exact replicas of every internal research run. Google DeepMind describes the repository as a port of its original internal pipeline and says some settings, including resequencing and folding-seed counts, were reduced so the examples can complete in a more practical time.
Generated material is organized by stage: initial structures and sequences, redesigned sequences, folded predictions and evaluation outputs, alongside campaign metadata and logs. That organization can help researchers separate what the generator produced from what later redesign, folding and scoring stages changed or assessed.
Availability and evidence limits
The repository lists an Apache 2.0 licence for the code, while model parameters, generated outputs and prohibited uses have separate terms or policy files. Pretrained AlphaProtein Novo and AlphaFold-related parameter files must be obtained separately, so a public code checkout alone is not the complete execution package.
This article does not report enzyme-performance results. The official repository links to associated research, but the two bioRxiv pages did not return readable evidence during this run; without those records, experimental results, comparative claims and biological effectiveness would exceed what the verified source establishes. The next useful checkpoint is accessible primary research evidence or independent reproduction of the documented workflow.
Status
Learning, with medium internal confidence. The official Google DeepMind repository establishes the pipeline architecture, examples and packaging boundaries, but it is a first-party software record and does not independently validate scientific performance.
Sources
Update note: Last reviewed 2026-10-12. We will revise this post if the associated primary research becomes accessible or independent reproductions clarify performance and limitations.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.