NPOmix: a machine learning classifier to connect mass spectrometry fragmentation data to biosynthetic gene clusters

Microbial specialized metabolites are an important source of and inspiration for many pharmaceutical, biotechnological products and play key roles in ecological processes. However, most bioactivity-guided isolation and identification methods widely employed in metabolite discovery programs do not explore the full biosynthetic potential of an organism. Untargeted metabolomics using liquid chromatography coupled with tandem mass spectrometry is an efficient technique to access metabolites from fractions and even environmental crude extracts. Nevertheless, metabolomics is limited in predicting structures or bioactivities for cryptic metabolites. Linking the biosynthetic potential inferred from (meta)genomics to the specialized metabolome would accelerate drug discovery programs. Here, we present a k-nearest neighbor classifier to systematically connect mass spectrometry fragmentation spectra to their corresponding biosynthetic gene clusters (independent of their chemical compound class). Our pipeline offers an efficient method to link biosynthetic genes to known, analogous, or cryptic metabolites that they encode for, as detected via mass spectrometry from bacterial cultures or environmental microbiomes. Using paired data sets that include validated genes-mass spectral links from the Paired Omics Data Platform, we demonstrate this approach by automatically linking 18 previously known mass spectra to their corresponding previously experimentally validated biosynthetic genes (i.e., via NMR or genetic engineering). Finally, we demonstrated that this new approach is a substantial step towards making in silico (and even de novo) structure predictions for peptidic metabolites and a glycosylated terpene. Altogether, we conclude that NPOmix minimizes the need for culturing and facilitates specialized metabolite isolation and structure elucidation based on integrative omics mining. SignificanceThe pace of natural product discovery has remained relatively constant over the last two decades. At the same time, there is an urgent need to find new therapeutics to fight antibiotic-resistant bacteria, cancer, tropical parasites, pathogenic viruses, and other severe diseases. Here, we introduce a new machine learning algorithm that can efficiently connect metabolites to their biosynthetic genes. Our Natural Products Mixed Omics (NPOmix) tool provides access to genomic information for bioactivity, class, (partial) structure, and stereochemistry predictions to prioritize relevant metabolite products and facilitate their structural elucidation. Our approach can be applied to biosynthetic genes from bacteria (used in this study), fungi, algae, and plants where (meta)genomes are paired with corresponding mass fragmentation data.