Co-designing sequence and structure of functional de novo enzymes with EnzyGen2

Abstract

Proteins underpin essential biological functions across all kingdoms of life. The capacity to design novel proteins with tailored activities holds transformative potential for biotechnology, medicine, and sustainability. However, since protein functions, particularly enzymatic activities, depend on precise interactions with small-molecule ligands, accurately modeling these interactions remains a formidable challenge in de novo protein design. Here, we introduce EnzyGen2, a protein foundation model designed for the simultaneous co-design of sequence and structure under ligand-guided functional targeting. Comprising 730 million parameters, EnzyGen2 is trained on 720,993 protein-ligand pairs using multi-task learning objectives that encompass joint prediction of sequence, structure, and protein-ligand interactions. In rigorous in silico benchmarks, EnzyGen2 consistently outperforms state-of-the-art baselines, including Inpainting, RFdiffusion/ProteinMPNN, RFdiffusion2/LigandMPNN, and RFdiffusion3/LigandMPNN, as measured by the enzyme-substrate prediction score, AlphaFold2 confidence metric, and structural fidelity, while it generates samples 400x faster than prior methods. We further experimentally validated EnzyGen2 across multiple enzyme families, including chloramphenicol acetyltransferase, aminoglycoside adenylyltransferase, and thiopurine S-methyltransferase. De novo enzymes generated by our family-specific EnzyGen2 exhibited catalytic activities comparable to or exceeding those of natural enzymes, while retaining substantial novelty with sequence identities as low as 51.6%. These results establish EnzyGen2 as a robust Artificial Intelligence-based tool for functional enzyme design, demonstrating the power of large protein foundation models to create high-performance, novel biocatalysts.