Rank-guided learning accelerates automated enzyme engineering

Enzyme catalysis has become a cornerstone of modern synthesis, offering sustainable routes to structurally complex molecules. However, the rugged fitness landscape and limited model capacity to learn from sparse data constrain the development of biocatalysts. Here, we present Rank-guided Exploration for Automated enzyme reProgramming (REAP), a closed-loop platform that combines artificial intelligence with robotic experimentation to accelerate enzyme engineering. At its core, REAP is powered by RankReg, a hybrid loss function that jointly optimizes ranking fidelity and quantitative accuracy, enabling adaptive in-loop learning directly from experimental feedback. This ranking-regression framework allows REAP to identify functional hotspots across both catalytic centers and previously underexplored distal regions, and to progressively extend model-guided exploration from single mutations to productive combinatorial variants. Applied to cytochrome P450 BM3, REAP achieved a 57-fold improvement in desired activity within five cycles, and further produced up to 104-fold enhancement in Staphylococcus aureus Sortase A (SaSrtA). By coupling algorithmic innovation with automation, REAP establishes a scalable and generalizable framework for accelerating biocatalyst discovery and expanding the scope of enzyme-enabled synthesis.