Expanding the galactose oxidase substrate repertoire via a multi-modal computational strategy

Biocatalysis offers a greener alternative to chemical processes, yet industrial adoption is bottlenecked by the prohibitive cost of enzyme engineering. Here, we report a multi-modal computational framework that integrates synergistic descriptors of substrate binding, protein stability, and aggregation traits. Using galactose oxidase as a model system, this approach achieves 77% precision in identifying active variants across 13 structurally diverse secondary alcohols despite being trained exclusively on a single substrate. By implementing sequence conservation-based filtering, we demonstrate that training data can be reduced by 41% without compromising predictive fidelity. Furthermore, the multi-modal model generalizes to combinatorial variants (2-5 mutations) with 100% precision on a limited library. Experimental validation yields a curated panel of starting points for the synthesis of high-value functional ingredients, including optical brighteners and fragrance stabilizers. Our multi-modal approach facilitates more cost-effective enzyme engineering, supporting the development of biocatalytic routes for sustainable chemical manufacturing.