Fig 1.
Time course genus abundance information was acquired from metagenomic sequencing of mouse gastrointestinal tracts under varying experimental conditions. Missing time points from experimental data were estimated such that genus abundances existed at the same time points across all treatment groups. Next, genus abundances were binarized such that Boolean regulatory relationships could be inferred. A dynamic Boolean model was constructed to explore gut microbial dynamics, therapeutic interventions, and metabolic mediators of bacterial regulatory relationships.
Fig 2.
Construction of a network model of the gut microbiome from time course metagenomic genus abundance information.
Principal component analysis coefficients associated with each sample in the metagenomic genus abundance dataset was completed for A) interpolated genus abundances and B) binarized interpolated genus abundances. ‘*’ = Healthy; ‘^’ = clindamycin treated; ‘#’ = clindamycin+ C. difficile treated. C) Consensus binarization of genus abundance information. Each heatmap represents the consensus binarization for each treatment group. The horizontal axis represents the day of the experiment that the sample came from. The vertical axis represents the specific genera being modeled. Each genus was binarized to a 1 (ON; above activity threshold) or 0 (OFF; below activity threshold). D) Interaction rules were inferred from the binarized data. The interaction rules were simplified for visualization (compound rules were broken into simple one-to-one edges).
Fig 3.
Steady states and node perturbations in the gut microbiome model.
A) Heatmap of the three steady states in the gut microbiome model. These steady states are identical to steady states identified in the three experimental groups. B) The effect of node perturbations represented by four heatmaps. On the Y-axis of each of the four heatmaps are nodes (genera) in each steady state. On the x-axis of each of the four heatmaps are the steady states found under normal model conditions (i.e. no node perturbations) and also the specific perturbation of a single network node. The two heatmaps in the left column of the figure demonstrate the effect of addition (forced overabundance) of individual genera, and the two heatmaps in the right column of the figure demonstrate the effect of removal (knockout) of individual genera. The top row heatmaps show the effect of node perturbations on the clindamycin treated group and the bottom row heatmaps show the effect of node perturbations on the clindamycin+ C. difficile treatment group. *Genus abundance of 0 means present in 0% of asynchronous simulations and is indicated in blue; Genus abundance of 1 means present in all (100%) of asynchronous simulations, shown in yellow. n = 1000 simulations were applied for all Boolean model simulations.
Fig 4.
Subsystem enrichment analysis highlights metabolic differences between taxa.
The p-values from the enrichment analysis are log-transformed and negated, such that darker regions indicate greater enrichment. The enrichment analysis quantifies the likelihood that a given subsystem (row) would be as highly abundant as observed within a given metabolic reconstruction (column) by chance alone. A subset of 22 interesting subsystems is shown here. Subsystems of interest include those for which all taxa are enriched, such as glycolysis, and nucleotide sugars metabolism, highlighting the fact that all taxa contain relatively full compliments of reactions within those subsystems. Similarly, subsystems for which a single genus differs from the remaining genera are interesting, such as cyanoamino acid metabolism, where C. difficile is highly enriched for reactions in that subsystem. Some subsystems are differentially enriched between Barnesiella and Lachnospiraceae, and C. difficile such as lipopolysaccharide biosynthesis and cyanoamino acid metabolism.
Fig 5.
Metabolic competition scores and in vitro data indicate a non-metabolic interaction mechanism.
A) Competition scores for all pairs of genera with C. difficile. Notice that Barnesiella has nearly the lowest competition score. B) Maximum growth rates for all growth conditions. C. difficile grew more slowly in B. intestinihominis spent media (n = 16, p-value < 0.005, by one-sided Wilcoxon rank sum test). The co-culture with both B. intestinihominis and C. difficile grew more slowly than C. difficile alone (n = 16, p-value < 0.05, by one-sided Wilcoxon rank sum test). C) Area under the curve (AUC) was not significantly different for C. difficile in fresh media or B. intestinihominis spent media (n = 16, p-value = 0.22 by one-sided Wilcoxon rank sum test). D) The experimental (red, solid line) and simulated (blue, dashed line) co-culture growth curves. “Binte” indicates B. intestinihominis, while “Cdiff” stands for C. difficile. On average, the experimental co-culture growth curves maintained a lower density than the simply additive null model. Error bars represent the standard error of the mean from 16 independent replicates.
Fig 6.
Computational models can bring us closer to true interaction networks.
A) Potential inhibitory mechanisms include direct inhibition of C. difficile by Barnesiella (e.g. via competition for scarce resources, or toxin production), or indirect inhibition (e.g. through a host antimicrobial response). B) A great deal has been published on the topic of network inference from complex data sets, and more can be done to improve inference methods. Particularly for microbial interaction networks, it is essential to identify, not only the nature of the interactions, but also the underlying mechanisms. Metagenomic genus abundance information can be used to infer causal relationships between bacteria; however, other information sources are required to determine the exact nature of these interactions. Each individual network edge may have very different underlying causes (metabolic, physical interaction, toxin-based, etc.). Including more tools in the pipeline, such as metabolic network reconstructions, bioinformatics tools, etc., will help elucidate these mechanisms, allowing far more rapid hypothesis generation, leading to a more focused effort in the wet lab.