From machine learning to multimodal models: The AI revolution in enzyme engineering

Protein engineering is a powerful tool for applications spanning synthetic biology, biocatalysis, and drug discovery. Recent advances in artificial intelligence (AI), from conventional machine learning (ML) algorithms to large-scale pre-trained protein models, have greatly accelerated enzyme engineering field entering a data-driven era. This review provides a guidance map of current enzyme engineering tasks and builds an integrative perspective on AI methods, model types, landmark tasks, and data resources. We begin by delineating the core modeling tasks in enzyme engineering, which include encompassing function annotation, structural modeling, and property prediction and by reviewing recent advances alongside dominant algorithmic frameworks. Next, we outlined the evolution of AI into enzyme engineering, tracing its progression through four stages: classical machine learning approaches, deep neural networks, protein language models (pLMs), and emerging multimodal architectures. Finally, we highlight four trends that are redefining the landscape of AI-driven enzyme design: (i) the replacement of handcrafted features with unified, token-level embeddings; (ii) a shift from single-modal models toward multimodal, multitask systems; (iii) the emergence of intelligent agents capable of reasoning; and (iv) a movement beyond static structure prediction toward dynamic simulation of enzyme function. Together, these developments are paving the way for intelligent, generalizable, and mechanistically interpretable AI platforms poised to synthetic biology.