Figure 1.
Methodology for characterizing microbial interactions using a compendium of similarity measures.
16S data from the Human Microbiome Project (HMP) were collected from 18 body sites in a cohort of 239 healthy subjects and assessed using 16S rRNA gene sequencing. We analyzed microbial co-occurrence and co-exclusion patterns in these data by developing two complementary approaches: a compendium of Generalized Boosted Linear Model (GBLMs) and an ensemble of similarity and dissimilarity measures. Each approach produced a network in which each node represented a microbial taxon within one body site, and each edge represented a significant association between microbial or whole clade abundances within or across body sites. The resulting association networks produced by each individual method were merged as p-values using Simes method, after which FDR correction was performed. Associations with FDR q-values>0.05, inconclusive directionality, or fewer than two supporting pieces of evidence were removed. This provided a single global microbial association network for taxa throughout the healthy commensal microbiota.
Figure 2.
Significant co-occurrence and co-exclusion relationships among the abundances of clades in the human microbiome.
A global microbial interaction network capturing 1,949 associations among 452 clades at or above the order level in the human microbiome, reduced for visualization from the complete network in Figure S1. Each node represents a bacterial order, summarizing one or more genus-level phylotypes and family-level taxonomic groups. These are colored by body site, and each edge represents a significant co-occurrence/co-exclusion relationship. Edge width is proportional to the significance of supporting evidence, and color indicates the sign of the association (red negative, green positive). Self-loops indicate associations among phylotypes within an order; for a full network of all phylotypes and clades, see Figure S1. A high degree of modularity is apparent within body areas (skin, urogenital tract, oral cavity, gut, and airways) and within individual body sites, with most communities forming distinct niches across which few microbial associations occur.
Table 1.
16S rRNA gene sequencing data from the Human Microbiome Project used to assess microbial co-occurrence relationships in the human microbiome.
Figure 3.
Global network properties summarizing key microbial hubs and interaction patterns.
A) Node degree distributions of overall, co-occurrence, and co-exclusion associations in the human microbiome. This is well-fit by a power law with slope −1,7 (dotted red regression line, adjusted R2 = 0.9). Node degree indicates the number of links that connect a node to others in the network. Power law degree distributions, referred to as scale-free, mean that most nodes have only a few edges and are often connected by a few high-degree hub nodes. The top five most connected hubs as indicated in callouts, mainly signature oral taxa including Porphyromonas in the tongue dorsum. B) and C) Node proportions after division of the network into body sites (B) or classes (C). Both pie charts show that the composition of the network (in agreement with underlying data) is skewed towards the oral cavity (B) and its constituent Firmicutes (including Bacilli and Clostridia) (C). (B) further agrees with published measures of body sites' alpha diversity [84]. D) and E) Composition of relationships among microbes grouped according to body site (D) and taxonomic class (E). In E), the first two bars (green and red) include the fraction of all possible edges incident to at least one node representing a class or one of its members (root scaled for visualization). The second two bars (lime and orange) only include pairs of microbes that are members of the same class, again normalized as a fraction of total possible interactions and root scaled. The Bacilli, Bacteroidia, and Fusobacteria contain significantly more negatively associated microbes than expected by permutation testing (see Table S2), and classes overall are depleted for negative associations, indicating that members of the same class tend not to compete strongly with each other in these communities.
Table 2.
Summary statistics of microbial associations in the normal human microbiota.
Figure 4.
Co-occurrence of microbial clades within and among body areas.
Nodes represent microbial classes colored by phylum, with edges summarizing aspects of their interactions over all body sites. Classes are linked when the number of edges between them is significantly larger than expected (randomization p<0.05, see Methods). Edge type (solid or dashed) indicates the body area contributing the most edges to the total interactions between two classes, with the label specifying the percentage contributed by this dominant body area. For instance, 80% of the edges between Bacilli and Actinobacteria come from skin sites. Green indicates co-occurrence, red exclusion. Most inter-class interactions occur in the mouth, with the Actinobacteria and Bacilli forming negative hubs.
Figure 5.
Related microbial niches as determined by associations spanning habitats at multiple human body sites.
Each node represents a body site, with edge width indicating significant cross-site correlations (randomization p<0.05, see Methods). Green edges show co-occurrence, red co-exclusion. Skin, vaginal, oral soft tissue, and tooth plaque moieties are apparent, with the gut and airways notably lacking significant interactions with other available body site niches. However, most relationships between microbial relative abundances occur specifically within, rather than between, individual body sites.
Figure 6.
Functional and phylogenetic similarities between co-occurring microbes.
Evolutionary (phylogenetic) distances among microbial clades were compared to the clades' functional potentials as defined by the Jaccard index of orthologous gene (COG) families shared between genomes (see Methods). Each point represents a pair of significantly associated microbes colored by direction of the association (green positive, red negative) and shaped by the type of relationship (triangle: between body sites, square: within site). Phylogenetic distances were inferred by FastTree [82] using species-level 16S sequences. Most interactions lie along the diagonal, reflecting the baseline correlation between these functional and evolutionary distances, with highly related clades co-occurring among related habitats (e.g. bilateral skin sites, proximal oral sites) in the lower left. Off-diagonal examples include potential competition among dominant gut signature taxa (e.g. Prevotellaceae/Bacteroides) and functional complementarity between distinct oral pathogens (e.g. Treponema/Prevotella).