Google DeepMind has introduced AlphaProtein Novo (AP Novo), a machine-learning pipeline for de novo enzyme design. According to a preprint published this week on bioRxiv, the system shows for the first time that computationally designed enzymes can outperform natural sequence mining on challenging chemistry. Built jointly with Caltech and the University of Pittsburgh, AP Novo constructs enzymes from first principles rather than optimizing existing proteins, addressing a long-standing bottleneck in biocatalyst discovery. Because natural enzyme discovery depends on sequences that evolution has already produced, reactions never adopted by living organisms leave little to mine — precisely the gap a from-scratch approach is built to close.

At the core of the pipeline is motif scaffolding. A diffusion model co-generates protein structures and amino acid sequences around a catalytic motif — the arrangement of side chains and ligands required for a hypothesized reaction mechanism. Candidate designs are then screened using metrics derived from AlphaFold 3 predictions, which evaluate mechanistically relevant atomic details, such as the geometry between a catalytic base and its substrate. In total, the team tested more than 5,600 designs across five reactions, reporting hit rates of up to 80% in the best design facets and state-of-the-art catalytic efficiencies on benchmark reactions without iterative experimental optimization.

Two headline applications show the breadth of the approach. The first targets nitrene transferases that synthesize piperidines, a heterocycle found in numerous FDA-approved drugs. The competitive cyclization of a substrate can yield either a five-membered pyrrolidine or a six-membered piperidine ring, and natural heme enzymes strongly favor the former. Screening 188 natural and engineered variants surfaced no enzyme exceeding a 30:70 piperidine-to-pyrrolidine ratio. The lead de novo design, GDM_NT_270, achieved 99:1 regioselectivity for piperidine with 94% enantiomeric excess — a measure of how exclusively a molecule is produced in one mirror-image form, a purity variable that matters in drug synthesis — and 22 turnovers, inverting the natural bias through direct control over active-site geometry.

The second application targeted di(2-ethylhexyl)phthalate (DEHP), a pervasive plasticizer and endocrine-disrupting environmental contaminant whose bulky side chains and water insolubility defeat most natural hydrolases. A recent screen of 65 natural esterases identified only one active DEHPase; AP Novo, by contrast, produced seven new structural families capable of the reaction, with a novel-scaffold hit rate reaching 11% (39% for recycled scaffolds). While the designs remain less active than natural enzymes in aqueous conditions, they display properties rarely seen in nature: one 191-residue enzyme was 14-fold more active at 90°C than at room temperature and stayed functional in 75% acetonitrile — conditions that fully denature natural esterases. Such robustness carries practical weight, since elevated temperatures accelerate reactions and organic solvents help dissolve water-insoluble substrates like DEHP. Its small size and high expression in E. coli further reduce production costs.

Sequence Ensembles as the Key Signal

The study's principal methodological insight concerns filtering. The researchers found that evaluating ensembles of LigandMPNN-derived sequences against the same backbone — requiring every resequenced variant to pass mechanism-inspired filters — dramatically amplified predictive power, lifting serine esterase hit rates up to 30-fold. Combining this with partial-diffusion re-sampling of the backbone pushed retrospective hit rates to 60% for Kemp eliminases and 80% for serine esterases. The authors argue this explains why iterative redesign pipelines work: they implicitly select for backbones whose local sequence-structure space broadly supports the intended catalytic geometry.

Limitations persist. Catalytic activities remain orders of magnitude below those of natural or directed-evolution-optimized enzymes, motif construction requires reaction-specific expertise, and the models do not yet meaningfully capture the physics of catalysis. Even so, with code and weights released for non-commercial use — access that lets outside research groups apply the pipeline to target reactions of their own — AP Novo signals that generative protein design is maturing into a practical complement — and, in select cases, an alternative — to mining natural biodiversity.

Source: Metaverse