From Promiscuity to Specificity: Machine Learning-Driven Multidimensional Evolution of Tropinone Reductase for Challenging Substrates

Abstract A fundamental challenge in harnessing substrate promiscuity for industrial biocatalysis lies in reconciling broad substrate acceptance with high catalytic specificity. Although promiscuous enzymes provide attractive starting points for accessing non-natural transformations, their native active-site architectures and conformational dynamics often lack the precision required for efficient and selective conversion of desired unnatural substrates. Thus, systematically decoding and reprogramming the interplay between promiscuity and specificity remains a key obstacle toward versatile biocatalysts. Here, we address this challenge by integrating virtual screening, machine learning, and molecular dynamics simulations into a unified engineering framework. A tropinone reductase from Cochlearia officinalis (CoTR) was identified as a broadly promiscuous scaffold. By converting complex enzyme–substrate interaction dynamics into quantifiable and actionable design principles, we predict and validate key residues that govern substrate promiscuity. Guided by these insights, distinct CoTR variants were engineered for three representative substrates. The optimized variants—CoTR‑8M, CoTR‑5M‑1, and CoTR‑5M‑2—exhibited 7.0‑ to 16.3‑fold higher catalytic efficiencies than wild-type CoTR toward their respective target substrates and supported substrate loadings of up to 50 or 100 g L–1 while maintaining >99% ee. This work not only delivers scalable routes to three industrially valuable chiral alcohols but also provides mechanistic insights into the synergistic modulation of promiscuity and specificity. This study demonstrates a general strategy for translating intricate molecular-recognition features into rational design principles, thereby facilitating the development of tailored biocatalysts for non-native substrates.