A Metabolome- and Metagenome-Wide Association Network Reveals Microbial Natural Products and Microbial Biotransformation Products from the Human Microbiota

Abstract

The human microbiome consists of thousands of different microbial species, and tens of thousands of bioactive small molecules are associated with them. These associated molecules include the biosynthetic products of microbiota and the products of microbial transformation of host molecules, dietary components, and pharmaceuticals. The existing methods for characterization of these small molecules are currently time consuming and expensive, and they are limited to the cultivable bacteria. Here, we propose a method for detecting microbiota-associated small molecules based on the patterns of cooccurrence of molecular and microbial features across multiple microbiomes. We further map each molecule to the clade in a phylogenetic tree that is responsible for its production/transformation. We applied our proposed method to the tandem mass spectrometry and metagenomics data sets collected by the American Gut Project and to microbiome isolates from cystic fibrosis patients and discovered the genes in the human microbiome responsible for the production of corynomycolenic acid, which serves as a ligand for human T cells and induces a specific immune response against infection. Moreover, our method correctly associated pseudomonas quinolone signals, tyrvalin, and phevalin with their known biosynthetic gene clusters.IMPORTANCE Experimental advances have enabled the acquisition of tandem mass spectrometry and metagenomics sequencing data from tens of thousands of environmental/host-oriented microbial communities. Each of these communities contains hundreds of microbial features (corresponding to microbial species) and thousands of molecular features (corresponding to microbial natural products). However, with the current technology, it is very difficult to identify the microbial species responsible for the production/biotransformation of each molecular feature. Here, we develop association networks, a new approach for identifying the microbial producer/biotransformer of natural products through cooccurrence analysis of metagenomics and mass spectrometry data collected on multiple microbiomes.

Keywords: association network; biotransformation; mass spectrometry; metagenomics; microbiome; natural products; xenobiotic.

FIG 1

The pipeline includes the following steps: extracting microbial (a) and molecular (b) features from the raw data, searching for pairs of associated features and computing false discovery rates (c), constructing the association network (d), and assigning molecular features to phylogenetic clades (e).

FIG 2

(a) Microbial natural products can be detected as positive correlations between the occurrences of the microbial species and the molecules in the association network. (b) Microbial biotransformation products can be detected as negative correlations between the microbial species and the precursor molecules, along with positive correlations between the microbial species and the product molecules. The feature tables are mock-up data.

FIG 3

(a) Association network of AGP. (b) Pseudomonas bacteria are positively associated with phenazine-1-carboxylic acid, rhamnolipids, and PQS. (c) The correlation between Desulfovibrio and cholic acid is noncausal. (d) Clostridiales biotransform bile acids. Here, we combined the nodes that represent the same molecules or taxa in the same family.

FIG 4

(a) Chemical structure of corynomycolenic acid. (b) Metabolite graph of corynomycolenic acid. (c) Fragmentation graph of corynomycolenic acid. (d) Annotation of the mass spectra of corynomycolenic acid (only explained peaks are shown).

FIG 5

BGC of phevalin. (a) Putative nonribosomal peptide synthetase (NRP) BGC discovered by antiSMASH. This BGC contains two adenylation domains (A), two thiolation domains (T), one condensation domain (C), and one NAD binding domain (NAD). Under each adenylation domain are the associated amino acids and scores predicted by NRPSPredictor. The greater the score, the greater the likelihood that the amino acid will be recognized by the adenylation domain. The surrounding structures are the putative molecules that can be produced by the BGC. (b) Fragmentation tree of Val-Phe (phevalin) and Val-Tyr (tyrvalin) given by Dereplicator+. (c) Mass spectral annotations given by Dereplicator+.

FIG 6

(a) Putative BGC of corynomycolenic acid in Corynebacterium kuscheri strain DSM 20755. (b) Known BGC of corynomycolic acid in Corynebacterium diphtheria strain NCTC 13129. Genes annotated with the same function in the two BGCs are in the same color. Genes in gray are unique genes of the two BGCs. (c) Chemical structure of corynomycolenic acid. (d) Chemical structure of corynomycolic acid. The structural difference between the two molecules is highlighted in black boxes.

FIG 7

Assigning the molecular features that are positively associated with the microbial features at a P value threshold of 10⁻²⁰ to the phylogenetic tree. The tree is trimmed to the taxonomic-order level. Numbers in boldface show the counts of molecules assigned to the corresponding clades. Heatmap shows −log₁₀(P), where P is the minimal P value between the molecule and an OTU within the clade. Dereplicator+ molecular annotations for the known molecules are shown. The molecular features were extracted by MSClustering based on tandem mass spectral data and annotated by spectral library search and Dereplicator+ (level 2 and 4 metabolite identification) (46).

FIG 8

Benchmarking various feature extraction methods and association tests. Different methods are compared based on the number of associations discovered (a) and the number of unique metagenomic features associated with a molecular feature (b) at different false discovery rate thresholds. Here, we benchmark MS-Clustering and Optimus (binarized abundance with thresholds 10, 10², …, 10⁶) with Fisher’s exact test association and Optimus (continuous abundance) with Pearson’s correlation test association, Spearman’s rank correlation test association, and mutual information criterion. In the case of Pearson’s correlation, no association was discovered at a false discovery rate of 0.01.

FIG 9

Benchmarking MS-Clustering and Optimus (binarized abundance with thresholds 10, 10², …, 10⁶) with Fisher’s exact test association. Data on the x axis represent false discovery rates estimated by the Benjamini-Hochberg procedure. Data on the y axis represent the numbers of metabolite-microbe associations discovered.