Discovery and Characterization of Terpene Synthases Powered by Machine Learning
The exponential growth of uncharacterized enzyme sequences in genomic repositories demands novel tools for functional annotation. Here, we combined alignment-driven structural domain analysis with protein language models to create EnzymeExplorer, a machine-learning pipeline for enzyme function prediction. We applied this approach to terpene synthases (TPSs), which present an ideal model case because they catalyze complex carbocationic rearrangements with unpredictable product outcomes. We detected new structural domains and achieved significantly higher average precision than existing methods for function prediction. By analyzing the UniRef90 database, we identified TPSs overlooked by existing computational methods. Remarkably, we discovered and experimentally confirmed three archaeal TPSs, expanding the known taxonomic distribution of TPS catalysis to a new domain of life. Further in silico screening of archaeal proteomes revealed that terpene biosynthesis is widespread across Archaea. Our approach offers a powerful framework for characterizing enzyme "dark matter" in the rapidly expanding genomic and metagenomic datasets.