Advancing Natural Product Discovery Through Integrated Machine Learning, Bioactivity-Guided Genome Mining, and Quantitative Benchmarking of Biosynthetic Gene Cluster Similarity

Abstract

Natural products remain a critical source of therapeutic agents, yet effectively translating biosynthetic potential into bioactive compounds requires improved integration of genomic, chemical, and computational strategies. This dissertation advances natural product discovery through the combined application of machine learning, bioactivity-guided genome mining, and quantitative benchmarking of biosynthetic gene cluster (BGC) similarity. By systematically linking genomic features to chemical structure and biological activity, this work establishes a framework for more efficient prioritization of bioactive natural products. A machine learning-guided genome mining workflow was developed and applied to nocardicin-like BGCs to prioritize clusters with predicted antibacterial activity. Comparative genomics and predictive modeling enabled the targeted evaluation of candidate strains, successfully enriching for bioactive producers. Although no new nocardicin analogs were identified, this strategy led to the unexpected discovery of siderophore-derived antibacterial metabolites from Actinoplanes teichomyceticus. Metabolomic analysis and bioactivity-guided fractionation revealed ferrioxamine and desferrioxamine-type siderophores with selective Gram-positive activity, including enhanced potency upon metal chelation, suggesting a Trojan metal-like mechanism. To strengthen genome-to-structure prediction, this work also quantitatively benchmarked BGC similarity metrics against corresponding product structural similarity. While all evaluated methods showed moderate correlation, performance varied across biosynthetic classes, with BiG-SCAPE demonstrating the most robust clustering after dataset curation. Collectively, this dissertation demonstrates how integrating predictive machine learning, experimental validation, and quantitative benchmarking can improve the accuracy, interpretability, and efficiency of modern natural product discovery.