Artios: Designing Functional De Novo Enzymes with All-Atom Diffusion and Steerable Potts Models

Abstract

Designing enzymes for any desired reaction is a long-standing goal of protein engineering, with applications across chemistry, medicine and industry. Directed evolution can optimise existing enzymes, but it depends on first finding a natural starting point with measurable activity. De novo design could remove this dependence by designing new enzymes directly from a description of the active site, yet it remains extremely challenging. Here we present Artios, a two-stage enzyme design system. ArtiosBackbone, an all-atom diffusion model, generates a protein backbone together with its ligand around an atom-level active-site motif and chooses where the catalytic residues sit in the sequence. ArtiosPotts, a ligand-aware Potts-model sequence designer, samples whole sequences from an explicit energy, to which terms can be added at sampling time without retraining. On the Atomic Motif Enzyme (AME) benchmark, Artios achieves state-of-the-art performance, and when the designed pocket must also keep its ligand it produces five times as many passing designs as the strongest previous method. We also show that such terms steer ArtiosPotts away from low-complexity sequences and towards a natural amino-acid composition, and that its designed sequences are more expressible than those designed with LigandMPNN. In the laboratory, Artios produced multiple functional new-to-nature luciferases, esterases and cutinases, testing no more than a single 96-well plate of designs in each campaign.

Cite

@article{ligo2026artios,
  title  = {Artios: Designing Functional De Novo Enzymes with
            All-Atom Diffusion and Steerable Potts Models},
  author = {{Ligo Biosciences Team}},
  year   = {2026},
  month  = oct,
  url    = {https://papers.ligo.bio/artios}
}