Microbial-derived secondary metabolites (SMs) hold great therapeutic potential but are predominantly discovered from cultured species, representing only a fraction of microbial biodiversity. Advances in metagenomics have unveiled reservoirs of biosynthetic gene clusters (BGCs), but translating genomic sequences into precise chemical structures remains challenging owing to the structural complexity of cryptic BGCs and the context-dependent substrate tolerance and cross-reactivity of modular biosynthetic domains. Here we present DeepSeMS, a transformer-based large language model that accurately predicts secondary metabolite chemical structures from BGC sequences. By encoding biosynthetic genes as functional domains and leveraging a feature-aligned data augmentation, DeepSeMS outperformed existing methods and successfully generated chemically valid predictions for 96.38% of cryptic BGCs. Applying DeepSeMS to a global ocean metagenome, we characterized over 60,000 secondary metabolites, revealing chemical diversity, ecological specificity and considerable biomedical potential, especially as antibiotics. This study underscores the capability of deep learning-driven approaches in revealing hidden biosynthetic potential of Earth's largest, yet largely unexplored, microbial ecosystem.
Xu et al. (Thu,) studied this question.