Enzyme design is a transformative approach for addressing a wide array of scientific, industrial, and medical challenges by engineering enzymes to perform specific functions, through which the limitations of natural enzymes, such as the lack of stability, specificity, or activity required for novel applications or non-native conditions, can be addressed. However, the vast space for enzyme design, comprising countless possible amino acid combinations, renders exhaustive exploration, even making it infeasible (Cobb et al., 2013).

Traditional enzyme engineering methods predominantly rely on top-down strategies, including rational design, semi-rational design, and directed evolution (Chen and Arnold, 1993; Lerner et al., 1964; Tian et al., 2024). These approaches have achieved significant milestones, such as improving the efficiency of enzymes like P450 and alcohol dehydrogenase and engineering enzymes to catalyze new reactions (Brandenberg et al., 2019; Jensen et al., 2021; Kim et al., 2019; Li et al., 2024; Liu et al., 2019; Xu et al., 2024).

Theoretically, every specific reaction has an optimal enzyme sequence that yields maximal activity. While natural evolution could eventually reach this sequence given infinite time, directed evolution accelerates the process by artificially enhancing the rate of sequence variation and selection. Whether there is a fundamentally different strategy that can achieve comparable or superior results within a much shorter timescale by efficiently sampling a broad design space from scratch rather than gradually modifying existing sequences is worth exploring.

De novo enzyme design is the computational creation of novel protein sequences and structures from first principles or learned models, rather than modifying natural enzymes. Unlike traditional rational design or directed evolution, which explore sequence space locally around existing scaffolds, de novo design enables access to novel folds and functions absent in nature. By starting from a desired function or structure and leveraging generative models and physics-based simulations, it can efficiently search vast regions of sequence–structure space. This global exploration allows it to bypass local optima and more directly identify high-performance solutions, accelerating the discovery of enzymes with enhanced activity, specificity, and stability.

Early de novo designs relied on physicochemical principles (Hodges et al., 1981). For instance, DeGrado et al. (1987) laid the foundation for de novo design of enzymes by using geometric parameters of their tertiary and quaternary structures, and Kuhlman et al. (2003) advanced the field with fragment-based methods that enabled backbone generation and sequence optimization. Despite some successes in physicochemical principle-based methods, including high thermodynamically stable helical bundles and water-soluble α-helical barrels (Huang et al., 2014; Thomson et al., 2014), they require extensive knowledge of target enzyme structures and face challenges in parameterizing diverse enzyme families.

Recent advances in machine learning (ML) have revolutionized the landscape of de novo enzyme design, providing powerful tools to address the limitations of those traditional approaches. ML-driven strategies can design enzymes by targeting known functional sites or directly addressing specific design requirements (Ingraham et al., 2023; Munsamy et al., 2024; Watson et al., 2023), which often serve as starting points, enabling rapid in silico design and functional validation thereafter. Moreover, ML plays a pivotal role in subsequent refining, experimental validation, and optimization, which are essential for achieving enhanced enzyme performance. Fig. 1 highlights the chronological milestones for this transformative approach.

In this review, we highlight recent advances in AI-driven bottom-up de novo enzyme design, focusing on three key areas: 1) the application of ML techniques for de novo enzyme design, 2) the use of ML for rapid in silico functional and interaction validation, and 3) leveraging ML for efficient enzyme modification and optimization. We conclude by discussing the challenges and future directions of AI-driven de novo enzyme design, emphasizing its transformative potential in advancing the field.