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

Abstract

Microbial natural products, in particular secondary or specialized metabolites, are an important source and inspiration for many pharmaceutical and biotechnological products. However, bioactivity-guided methods widely employed in natural product discovery programs do not explore the full biosynthetic potential of microorganisms, and they usually miss metabolites that are produced at low titer. As a complementary method, the use of genome-based mining in natural products research has facilitated the charting of many novel natural products in the form of predicted biosynthetic gene clusters that encode for their production. Linking the biosynthetic potential inferred from genomics to the specialized metabolome measured by metabolomics would accelerate natural product discovery programs. Here, we applied a supervised machine learning approach, the K-Nearest Neighbor (KNN) classifier, for systematically connecting metabolite mass spectrometry data to their biosynthetic gene clusters. This pipeline offers a method for annotating the biosynthetic genes for known, analogous to known and cryptic metabolites that are detected via mass spectrometry. We demonstrate this approach by automated linking of six different natural product mass spectra, and their analogs, to their corresponding biosynthetic genes. Our approach can be applied to bacterial, fungal, algal and plant systems where genomes are paired with corresponding MS/MS spectra. Additionally, an approach that connects known metabolites to their biosynthetic genes potentially allows for bulk production via heterologous expression and it is especially useful for cases where the metabolites are produced at low amounts in the original producer.

Enhanced correlation-based linking of biosynthetic gene clusters to their metabolic products through chemical class matching

“…Integrative genome-metabolome mining is a complex problem that will require many different solutions and smart ways to integrate those. Combining and streamlining NPLinker with other novel linking methods, such as NPOmix, will be key for advancing this field to understand complex microbial communities and prioritise NP discovery [24]. Other possible routes are to implement additional feature-based scores such as a score based on shared substructures as inferred to be present from the genomic and metabolomic data, the first for example through iPRESTO that finds sub-clusters in BGCs that likely encode for biosynthetic scaffolds or substructures [25], and the latter for example through the use of data-driven approaches that find (MS2LDA) or contain (MotifDB) mass spectral patterns that can be connected to chemical substructures [26,27].…”

FERMO: a Dashboard for Streamlined Rationalized Prioritization of Molecular Features from Mass Spectrometry Data

“…Also, the integration of LC-MS/MS data with biological metadata is only the first step in integrative data analysis. The integration of LC-MS/MS metabolomics with other types of omics data, such as metagenomics or transcriptomics, would further strengthen prioritization workflows, as demonstrated with the recently published tools NPOmix [47] or NPLinker [48]. While challenging, such methods promise to further facilitate hypothesis formulation and potentially automate molecular feature prioritization.…”

Enhanced correlation-based linking of biosynthetic gene clusters to their metabolic products through chemical class matching

“…Integrative genome-metabolome mining is a complex problem that will require many different solutions and smart ways to integrate those. Combining and streamlining NPLinker with other novel linking methods, such as NPOmix, will be key for advancing this eld to understand complex microbial communities and prioritise NP discovery [24]. Other possible routes are to implement additional feature-based scores such as on the basis of shared substructures as inferred to be present from the genomic and metabolomic data, the latter for example through the use of data-driven approaches like MS2LDA and MotifDB [25,26].…”

“…The molecular networks listed in the PoDP were used, while antiSMASH 6 was run on the listed 24 and 11 genomes, respectively. CANOPUS was run for the three datasets within NPLinker with the aforementioned default settings, which took around 24 S4). We decided to do this as it did not hamper our effort of nding the validated links in the three datasets.…”