Terpenoids are the largest and most structurally diverse class of natural products, synthesized by terpene synthases (TPSs) through complex cyclization and hydroxylation cascades. Although product specificity can be tuned by modulating the enzymatic microenvironment around transient carbocation intermediates, efficient design of enzyme variants for specific products remains highly challenging, as current approaches largely rely on sequences and static structural information. Here, guided by the mechanism understanding obtained from molecular simulations, we validated the binding environment of a sesquiterpene synthase Agr5 from Agrocybe aegeriid by reprogramming it to generate non-native products. Building on V314G variant, designed based on substrate binding environment including a previously unrecognized water channel, we developed an engineering protocol that combines dynamic cross-correlation matrix analysis with a machine learning-based score DeEnzymeScore, to screen a designed variant library. This approach effectively explored the sequence-fitness landscape and identified both remote and active-site mutations, Notably, the most active variant enhanced the Agr5’s catalytic efficiency for viridiflorol by 11-fold. The observed activity improvement stems from opening a gate between E246 and Y261 at the base of the catalytic site, which stabilizes an adjacent flexible loop favorable and facilitates carbocation conversion. This work establishes a mechanism-guided strategy for fine-tuning closely related sesquiterpene synthases and demonstrates an efficient machine learning-driven workflow that improved TPS activity while minimizing mutagenesis and screening efforts.
Zhou et al. (Thu,) studied this question.