Designing new proteins de novo is a vital field of research for engineering enzymes capable of carrying out specific chemical reactions. However, designing a protein capable of binding to a target molecule does not inherently guarantee that it will exhibit catalytic activity. Developed by researchers at Ligo Biosciences, Artios is an artificial intelligence framework aimed at overcoming this hurdle by co-designing molecular structure and amino acid sequence in protein engineering.
Artios consists of two core components. ArtiosBackbone is a diffusion model that designs the protein’s atomic-level backbone structure, accounting for the target molecule and catalytically critical amino acid residues. ArtiosPotts, on the other hand, is a method based on the Potts model that leverages target-molecule-associated information to guide the selection of the amino acid sequence during design. These components aim to generate both protein structure and amino acid sequence in alignment with the targeted function.
The researchers evaluated Artios across comparative benchmarks assessing protein design and target molecule preservation. In the more stringent AME (Atomic Motif Enzyme) benchmark—which requires the ligand (the molecule that binds to the protein) to be preserved within the designed structure—6.1% of the designs met the success criteria. By comparison, this rate was reported as 1.2% for the combination of Proteina-Complexa with LigandMPNN, and 0.9% for RFdiffusion3 with LigandMPNN. In this evaluation, Artios satisfied 34 out of a total of 38 design requirements. In a comparison using the same RFdiffusion3 backbones, ArtiosPotts increased the proportion of designs meeting the ligand-preservation criterion from 0.9% with LigandMPNN to 5.3%. Similarly, in the SeqVal benchmark comprising native protein–ligand complexes, ArtiosPotts succeeded across 23 out of 114 complexes, whereas the baseline LigandMPNN method achieved success in only 7 complexes.
The researchers also investigated the experimental activity of the designed proteins. In cell-free protein expression assays, ArtiosPotts designs yielded a median luminescence signal of 2.53 × 10⁶ relative light units (RLU), compared to 1.53 × 10⁶ RLU for LigandMPNN designs. These results demonstrate that higher median luminescence signals were achieved with the ArtiosPotts designs. In enzymatic activity assays, 7 out of 85 designs tested in the initial luciferase design campaign and 9 out of 45 designs tested in a subsequent campaign were found to be active. For esterase designs evaluated against a butyrate derivative substrate, 17 out of 88 designs demonstrated activity. Furthermore, for the design called ArCut1, catalytic efficiency against the 4MU-octanoate substrate was measured at (kcat/KM) = (1.84 ± 0.20) × 10³ M⁻¹ s⁻¹.
The study indicates that uniting atomic-level structural design with steerable amino acid sequence design can facilitate the development of proteins bearing novel functions. Nonetheless, the activities of the engineered enzymes remained below the levels of natural or optimized enzymes. The researchers also note that structural prediction methods used, such as AlphaFold2 and Protenix, present limitations in resolving the precise geometries required for catalysis. Computational structural predictions do not constitute experimental validation of catalytic interactions, and the experimental structures of the proteins designed in this study have not yet been determined.
The Artios models and design framework have been released as open-source by the researchers. Consequently, while the study presents an approach for the computational design of novel functional proteins, the findings must be interpreted in light of the current limitations of the methodology.
Translated by: Nur Tanem Altundaş
Editor: Hazel Birgül
Reference:Cort, O., Egerton-Warburton, E., Goreci, A., & Harris, E. *Artios: Designing Functional De Novo Enzymes with All-Atom Diffusion and Steerable Potts Models*. Ligo Biosciences, Research Preview (October 2026). https://papers.ligo.bio/artios/paper.pdf