AlphaProtein Novo Designs Enzymes That Beat Natural Mining

Designed From Scratch: AlphaProtein Novo Outperforms Nature Mining on New Chemistry

TL;DR: Google DeepMind, Caltech and collaborators reported on Oct. 5, 2026 that the AlphaProtein Novo pipeline computationally designed enzymes from first principles that beat natural sequence mining at making piperidines and degrading DEHP, plus state-of-the-art efficiencies on two model reactions.

Google DeepMind and the California Institute of Technology introduced a machine-learning pipeline for de novo enzyme design called AlphaProtein Novo, or AP Novo, in a bioRxiv preprint published Oct. 5, 2026. The work is described in the preprint titled "Designing enzymes for new-to-nature chemistry and non-natural substrates with AlphaProtein Novo" and was led by Zachary Wu and Zi-Qi Li with corresponding author Jue Wang and collaborators including the group of Frances H. Arnold, according to posts by Jue Wang on Oct. 5, 2026.

According to that preprint's abstract, AP Novo demonstrates, for the first time, that de novo enzyme design can outperform natural sequence mining in addressing challenging chemistry. The authors report new-to-nature nitrene transferases to synthesise the pharmacophore piperidine with unprecedented product selectivity, enzymes that degrade the environmental toxin DEHP under conditions that denature natural enzymes, and state-of-the-art catalytic efficiencies on two well-studied model reactions.

A second, related bioRxiv preprint published Oct. 2, 2026, titled "Design-Evolution Synergy Enables New-to-Nature Enzyme Chemistry," describes the companion design-evolution strategy. That preprint lists Zi-Qi Li, Ruijie K. Zhang, Pei-Pei Xie, Jacob Kelly, Yu Zhang, Ziyang Qin and colleagues, with Jue Wang and Frances H. Arnold among the authors and Arnold as corresponding author, and Google DeepMind, California Institute of Technology and University of Pittsburgh among affiliations.

A pipeline built to start where evolution cannot

According to the abstract of "Designing enzymes for new-to-nature chemistry and non-natural substrates with AlphaProtein Novo," creating highly active enzymes for arbitrary reactions is a transformative goal for the molecular sciences. The same abstract states that traditional protein engineering is constrained by a reliance on existing natural starting points, which are often challenging to discover.

The abstract contrasts that constraint with de novo design, which it describes as offering the potential to overcome this by creating enzymes from first principles, but which has yet to achieve practically relevant catalytic properties prior to this work. The reported conclusion is that de novo design has become a powerful complement to natural enzyme diversity for discovering catalysts for a range of applications.

The companion preprint "Design-Evolution Synergy Enables New-to-Nature Enzyme Chemistry" frames the same problem from the evolution side. Its abstract states that directed evolution and machine-learning-assisted evolution have transformed biocatalysis by enabling rapid optimisation of catalytic performance within compatible protein scaffolds. It adds that advances in deep-learning-driven de novo protein design have expanded accessible protein architectural space, enabling the generation of functional catalysts for increasingly challenging transformations.

That second abstract reports a synergistic design-evolution strategy to encode new-to-nature regio- and enantioselective C–H amination reactions using engineered de novo haemoproteins to synthesise chiral piperidines, described there as a privileged pharmacophore in drug discovery. The stated integration is motif scaffolding, computational re-design, and directed evolution, with reported outcomes including programmable regio- and enantioselectivity, high catalytic efficiency, excellent enzyme stability, and scaffold diversity.

What AP Novo actually is

According to the repository titled "AlphaProtein Novo - google-deepmind," AlphaProtein Novo is a generative diffusion pipeline for de novo enzyme design via structural motif scaffolding. The repository is distributed under the Apache License 2.0 and implemented in Python.

The same repository states that the package includes a diffusion model which co-generates protein structures and sequences conditioned on a catalytic motif and ligand context. That generation step is run using run_generator.py, according to the documentation.

To form an end-to-end pipeline, the documentation states that the diffusion model is combined with optional sequence redesign using LigandMPNN, structure prediction using AlphaFold 3, and evaluation metrics. The file run_pipeline.py runs all stages in sequence from a single manifest file containing configuration settings.

The documentation specifies two Python environments. The primary environment, named alphaprotein_novo, contains JAX and runs diffusion generation, AlphaFold 3 folding, evaluation metrics, and the pipeline orchestrator. The optional LigandMPNN environment contains PyTorch and is used to run LigandMPNN for sequence design.

How a design campaign is launched

According to "AlphaProtein Novo - google-deepmind," the standard way to run AP Novo is via run_pipeline.py. The documented launch command executes design generation, sequence design, folding, and metrics calculation end-to-end from a manifest file.

The provided example uses examples/kemp_eliminase/kemp_manifest.json with an output directory such as ./kemp_campaign, as documented in the Quickstart section. The same Quickstart documents options for pointing to model directories and to a LigandMPNN installation, and documents a test manifest, examples/kemp_eliminase/kemp_test_manifest.json, which runs the same Kemp eliminase benchmark with reduced sampling steps to verify installation.

The repository notes that pretrained model weights are not part of the package and must be downloaded from Google Cloud Storage. Use is subject to terms of use, and scripts look for weights under ./models/apnovo_generator by default, or in another directory pointed to with flags or manifest settings, according to the documentation.

The documentation identifies two weight packages. AP Novo generator weights are downloaded as generator.bin.zst into ./models/apnovo_generator. Structure prediction uses AlphaFold 3 Leaving Atom weights, downloaded as af3_leaving_atom.bin.zst into ./models/af3_la, because, according to the documentation, AF3-LA is fine-tuned for leaving atom handling and therefore models the covalent intermediates common in AP Novo designs.

The piperidine test: selectivity nature did not provide

Piperidines are a 6-atom ring common in medicines, according to the Oct. 5 explanation posted by Jue Wang. The same explanation states that enzymes could help produce them more cheaply, and that screening 188 natural enzymes found 0 with high specificity to piperidines.

According to that same account, in one round of design with AP Novo the team achieved higher piperidine specificity than any natural enzyme screened. The main preprint abstract describes this outcome as new-to-nature nitrene transferases to synthesise the pharmacophore piperidine with unprecedented product selectivity.

The companion preprint "Design-Evolution Synergy Enables New-to-Nature Enzyme Chemistry" describes the same chemical objective as encoding new-to-nature regio- and enantioselective C–H amination reactions using engineered de novo haemoproteins to synthesise chiral piperidines. Its abstract reports programmable regio- and enantioselectivity, alongside high catalytic efficiency, excellent enzyme stability, and scaffold diversity, through motif scaffolding, computational re-design, and directed evolution.

Together, the two preprints present piperidine synthesis as the head-to-head case where de novo design outperformed natural sequence mining. The main preprint states explicitly that the pipeline demonstrates, for the first time, that de novo enzyme design can outperform natural sequence mining in addressing challenging chemistry.

The DEHP test: activity where natural enzymes denature

DEHP is a plastic additive and toxic pollutant, according to the Oct. 5 explanation by Jue Wang. The same explanation states that some natural enzymes can degrade it but are very rare.

According to the main preprint abstract, AP Novo was used to create enzymes that degrade the environmental toxin DEHP under conditions that denature natural enzymes. The Oct. 5 explanation specifies those conditions as 75% acetonitrile or 90C, which the author states destroys most enzymes.

The same explanation states that the designed DEHPases were still active under those conditions, and that this could be useful in harsh industrial processes. That harsh-tolerance claim is the second case in which the team reports an advantage over mined natural diversity, complementing the piperidine-selectivity result.

The main preprint frames both the nitrene-transferase result and the DEHP result as evidence that de novo design has become a powerful complement to natural enzyme diversity for discovering catalysts for a range of applications.

Benchmark reactions: fastest novel-scaffold rates reported by the team

In addition to piperidine and DEHP, the main preprint abstract reports state-of-the-art catalytic efficiencies on two well-studied model reactions. The Oct. 5 posts identify those as benchmark reactions and describe the results as SoTA novel-scaffold activity.

According to the detailed Oct. 5 post by Jue Wang, the team made a Kemp eliminase with kcat/Km of 14,600 M-1s-1 and a serine esterase on 4MU-Ac with kcat/Km of 360,000 M-1s-1 and kcat of 5/s. The same post states that both were obtained without redesigning previous designs.

The same post qualifies the result as SoTA novel-scaffold activity, adding in parentheses that it is not strictly zero-shot because the team iterated on filters. That qualification distinguishes the scaffold novelty claim from a claim of a single untouched computational shot including filter development.

The repository's Quickstart and example design campaigns are consistent with that benchmark focus. The documented example manifest is for Kemp eliminase, and the documentation points users to example design campaigns for more detailed examples beyond the basic launch command.

Why filtering mattered more than generation alone

According to the main preprint abstract, iterative design and analysis found that key enablers of success were mechanism-inspired metrics based on AlphaFold 3 predictions and sequence-ensemble-based scoring of designed backbones.

The Oct. 5 technical explanation elaborates on the architecture as a generator-MPNN-AF3-filter pipeline. It states that the generator does sequence-structure co-diffusion via a simple transformer and no protein-specific inductive biases, and that it is SoTA on in silico benchmarks but that for wetlab success, filters do the heavy lifting.

According to that same explanation, the key filter scores a backbone by redesigning multiple sequences on it and aggregating over the ensemble. The author describes this as measuring preorganization of catalytic geometry due to a backbone rather than a single design sequence.

The documentation in "AlphaProtein Novo - google-deepmind" reflects the same multi-stage logic. After generation, candidates go through LigandMPNN sequence design, AlphaFold 3 folding via run_alphafold.py, and evaluation metrics via evaluate_design.py, with outputs organized into stage-specific subdirectories under the output directory.

Credits, code and what the authors say is still missing

The Oct. 5 announcement by Jue Wang states that the work was developed in collaboration with the group of Frances Arnold, and that the work was led by Zachary Wu and Zi-Qi Li with teammates including Alexander E. Chu, Ruijie K. Zhang, Luca Schulz, Jacob Kelly and others, plus Josh Abramson, Thomas Frerix, Amy Danson, Tristan Kwan and Harshnira Patani among contributors. The main preprint author list names Zachary Wu, Josh Abramson, Thomas Frerix, Alexander E. Chu, Ruijie K. Zhang, Luca Schulz, Amy E. Danson, Tristan O. C. Kwan and Jacob Kelly among Google DeepMind authors, Zi-Qi Li, Masy Domecillo, Ariane Mora, Julia C. Reisenbauer and Yu Zhang among California Institute of Technology authors, and Jue Wang as corresponding author.

The same announcement points to preprints and code, consistent with the repository "AlphaProtein Novo - google-deepmind," which asks that any publication disclosing findings arising from using the source code, model parameters, or outputs cite the paper titled "Designing enzymes for new-to-nature chemistry and non-natural substrates with AlphaProtein Novo."

The authors also listed limitations in the Oct. 5 technical post. The author wrote that the best esterase borrows a natural structural scaffold, that backbone-level preorganization does not get all the way to high natural-like activities, and that input motif can matter more than generator and filters. The post concludes that more research is needed and invites users to try the model.

Taken together, the sources published by Oct. 5, 2026 describe AP Novo as a diffusion-plus-filter system that co-generates structure and sequence around a catalytic motif and ligand context, validates candidates with AlphaFold 3-based mechanistic checks and ensemble scoring, and delivered laboratory enzymes for piperidine synthesis, DEHP degradation at 75% acetonitrile or 90C, a Kemp eliminase at 14,600 M-1s-1, and a 4MU-Ac serine esterase at 360,000 M-1s-1 with kcat of 5/s.