The rapid maturation of artificial intelligence (AI) has catalyzed a fundamental transition in biocatalysis, moving from structural analysis toward the prescriptive design of bespoke enzymes. This review synthesizes this AI-driven revolution, evaluating breakthroughs like AlphaFold2 and CLEAN that now bridge sequences with catalytic properties, including kinetic parameters and substrate specificity. We critically compare rational design strategies, contrasting evolutionary-guided redesign with the emerging generative de novo paradigm, where diffusion models and protein language models (PLMs) explore uncharacterized sequence space. By dissecting algorithms such as Graph Neural Networks and Transformers, we illustrate their role in deciphering protein chemistry's linguistic "grammar". Grounded in industrial cases, we demonstrate how AI overcomes bottlenecks like the stability-activity trade-off. Finally, we delineate the trajectory toward autonomous biofoundries and virtual cell modeling, envisioning engineered biocatalysts systematically integrated into complex metabolic networks─providing a roadmap for next-generation computational enzymology.
Shi et al. (2026) studied this question.