With the advent of next-generation sequencing platforms, a major influx of studies have sought to identify microbial composition differences in various habitats. However, such studies rarely consider environmental variables, such as metabolites or proteins, resulting in incomplete systemic clarity and potentially erroneous assumptions. This study represents one of the first successful attempts to integrate components of the adult gut mucosal ecosystem. We chose to perform analysis on two distinct colonic regions to ensure reproducibility of findings. Notably, all mucosal samples were collected from subjects who had undergone bowel preparation. While standard for both clinical and research endoscopy, bowel preparation is known to alter microbial alpha and beta diversity
. Accordingly, such depletion of mucosal microbiota is likely to reduce the scope of detectable inter-omic relationships. However, we reason that the observed relationships are representative of the native mucosa. Also, bowel preparation should result in less dietary and enteric secretion input from the proximal intestine, thereby increasing biogeographic resolution and decreasing noise from dietary metabolites. Nonetheless, it is possible that bowel preparation introduces metabolic changes in the microbial community that elicits non-physiologic inter-omic relationships. Therefore, the scope and quality of these inter-omic relationships merit additional assessment in undisturbed mucosal sites.
This study revealed significant inter-omic structure in both the cecum and sigmoid colon that was independent of age or disease status. While the relationship between the microbiome and metabolome appeared strongest in the cecum by Procrustes and coinertia analysis, a larger number of significant correlations were observed in the sigmoid compared to the cecum, possibly reflecting the known biogeographic differences of both microbes and metabolites
[11, 45, 51]. Despite this biogeographic dissimilarity, we observed significant overlap in findings between the cecum and sigmoid data. While only 342 metabolites were measured in both colonic regions, several observations remained consistent between datasets. As highlighted in Figures
7, the cecum and sigmoid microbial clusters were very similar in composition and function. Furthermore, as shown in Table
2, the relative ratios of inter-omic correlations were nearly identical at the phylum level. However, some biogeographic differences were observed. For example, the sigmoid data had nearly double the number of significant inter-omic correlations involving Proteobacteria and Actinobacteria, which might suggest differing biogeographic functional roles. Furthermore, using prediction strength, six microbial clusters were predicted for the cecum data while twenty were predicted for the sigmoid, suggesting that microbial function in the sigmoid is much more distinct than in the cecum.
This study also identified microbial clusters that were both metabolically and metagenomically concordant. Using observed metabolic correlations to govern cluster assignment of microbes revealed similarities among diverse groups of bacterial phyla that might not otherwise have phylogenetic or genomic associations. The metabolic relationships defined by these clusters may provide a new avenue to consider in vivo microbial function and host response. As noted above, species of Firmicutes, including Faecalibacterium, Phascolarctobacterium, and Roseburia tend to be depleted in IBD while Proteobacteria species tend to be enriched
. All three IBD-depleted genera were substantially confined to cluster 4 of both colonic regions, which was enriched for fatty acid biosynthetic genes and depleted of fatty acid metabolism genes. Fatty acids with the most relevance are SCFA, which are produced as fermentative byproducts. Multiple dietary inputs can be used by bacteria to produce SCFA, including microbial or dietary-derived starch, acetate, lactate, linoleic acid, and fiber
. Fatty acid biosynthesis is particularly relevant to the host because SCFA, like butyrate, can 1) act as energy sources for colonocytes; 2) inhibit Nuclear Factor-κB activation in human colonic epithelial cells, resulting in decreased levels of inflammatory cytokines; and 3) stimulate mucin production, which could result in increased barrier protection
. Accordingly, reduction in SCFA availability to the host, which could occur if cluster 4 bacteria were depleted, could increase mucosal propensity for and susceptibility to inflammation. However, cluster 4 is also highly enriched for various amino acid metabolic and biosynthetic pathways that could also influence the environmental availability of such molecules and thereby contribute to mucosal homeostasis in ways that have not yet been examined (Figures
6C). Furthermore, cluster 4 was depleted of genes from the KEGG glutathione metabolism pathway, which includes both biosynthetic and metabolic genes. Glutathione is important for mitigating oxidative stress and acts as a powerful redox buffer
. Therefore, depletion of genes involved in production and metabolism of glutathione could indicate that bacteria in cluster 4 were more susceptible to oxidative stress and would therefore be at a selective disadvantage in oxidative inflammatory conditions like those found in IBD. Conversely, cluster 1, which contained many Proteobacteria, was significantly enriched with genes from the glutathione metabolism pathway, which could explain why such bacteria tend to more abundant in IBD
. Therefore, grouping metabolically similar bacteria into clusters aids functional analyses by 1) reducing the dimensionality of data; 2) allowing assignment of potential functions to bacteria that might not be culturable in vitro; and 3) defining collective relationships between bacteria that might not be overtly related by phylogenetic sequences.
Another central finding of this study was the rich network of significant correlations between the microbiome and metabolome. Such correlation structure likely arises from a combination of two general processes: 1) catabolism and anabolism of metabolites by microbes, and 2) stimulation and inhibition of microbial growth by metabolites. Indeed, it is widely accepted that dietary alteration is accompanied by shifts in gut microbiome composition and that microbial composition influences the intestinal metabolome
[16, 25, 26]. However, the metabolites and metabolic pathways involved in such processes are unknown. Therefore, while it is difficult to conclusively assign cause and effect to correlation data, a central goal of this study was to determine whether bioinformatic signatures of either process could be detected.
To observe whether catabolic or anabolic reactions contributed to the inter-omic correlation structure, KEGG pathway representation of imputed OTU metagenomes was correlated with the correlation coefficients from the pair-wise microbe-metabolite comparisons. While roughly half of these comparisons had significant positive correlation coefficients, indicating concordance between observed metabolite abundance and metagenomic abundance, a large proportion of correlations were insignificant or significantly negative (Additional file
6). While the mechanisms remain unclear, multiple possibilities exist for the differing directionality of observed correlations. For example, a metabolic end product might have a positive correlation with the OTU (and thus the originating pathway) that produces and exports it, but a negative correlation with the OTU (and thus the originating pathway) that imports and processes it in a downstream pathway. Unfortunately, this also means that some metabolites might have less significant correlation curves due to organisms that encode enzymes producing such metabolites, but are not correlated with the metabolites because they are not exported, which was required for us to measure the association. Given the immature understanding of gut metabolic pathways and the imperfect nature of putative metabolite ID picking, we were not able to resolve this issue. Regardless, these data suggest that our predicted metagenomic and putative metabolite ID data were concordant. This was an important finding because defining metabolite IDs using biochemical methods is subject to numerous limitations that could be simplified with metagenomic data. To try and detect signatures of microbial inhibition or stimulation, we attempted to quantify the significance of community structure between microbes with shared metabolite correlations. While significant community structure was observed between microbes, we were unable to deconvolute the transitive effects of correlation and thus, could not conclude that metabolite-mediated microbial stimulation/inhibition occurred. However, we provided two examples of communities that appeared to have exceptional structure, even compared against other metabolite-associated microbial communities. Additional file
10 shows the structure of a microbial community that was defined by common correlation with a cecal metabolite (mass = 230.1845, retention time = 0.4078 minutes). This metabolite negatively correlated with 17 bacteria. When correlated with each other, these bacteria formed two tight clusters with extremely well-defined co-exclusion structure. An appealing explanation for this type of behavior is that the two communities actively compete with each other for consumption of the metabolite. The metabolite-associated community in Additional file
11 involved a sigmoid metabolite (mass = 434.1867, retention time = 0.6621 minutes). This metabolite was negatively correlated with one Tenericute and eighteen Proteobacteria OTUs and positively correlated with seven Firmicutes OTUs. When correlated with each other, these bacteria also formed two distinct communities that were extremely co-exclusive; one community contained all the Firmicutes, and the other community contained most of the Proteobacteria. This behavior could be indicative of a Firmicutes-produced metabolite that is inhibitory to Proteobacteria. While it is possible that the observed structure was due to the transitive nature of the correlations, these observations would seem to suggest that metabolites might be driving microbial community structure and may directly or indirectly modulate inter-species competition.
Regardless of the mechanism of correlation, the strong correlation between individual metabolites and microbes has numerous potential implications for future innovations. For example, strong correlative relationships may have value as biomarkers for individual or groups of analytes. An obvious application would be the use of specific metabolites as indicators of the presence of certain bacteria, which would presumably be faster than culture- or sequence-based approaches. Furthermore, knowledge of the relationships between metabolites and bacteria may prove useful in either direct (therapeutic) or indirect (dietary) interventions for chronic disorders with microbial compositional shifts, such as IBD.