A minimal model for microbial biodiversity can reproduce experimentally observed ecological patterns

Abstract

Surveys of microbial biodiversity such as the Earth Microbiome Project (EMP) and the Human Microbiome Project (HMP) have revealed robust ecological patterns across different environments. A major goal in ecology is to leverage these patterns to identify the ecological processes shaping microbial ecosystems. One promising approach is to use minimal models that can relate mechanistic assumptions at the microbe scale to community-level patterns. Here, we demonstrate the utility of this approach by showing that the Microbial Consumer Resource Model (MiCRM) - a minimal model for microbial communities with resource competition, metabolic crossfeeding and stochastic colonization - can qualitatively reproduce patterns found in survey data including compositional gradients, dissimilarity/overlap correlations, richness/harshness correlations, and nestedness of community composition. By using the MiCRM to generate synthetic data with different environmental and taxonomical structure, we show that large scale patterns in the EMP can be reproduced by considering the energetic cost of surviving in harsh environments and HMP patterns may reflect the importance of environmental filtering in shaping competition. We also show that recently discovered dissimilarity-overlap correlations in the HMP likely arise from communities that share similar environments rather than reflecting universal dynamics. We identify ecologically meaningful changes in parameters that alter or destroy each one of these patterns, suggesting new mechanistic hypotheses for further investigation. These findings highlight the promise of minimal models for microbial ecology.

Figure 1

A minimal model for investigating microbial biodiversity. (a) The Microbial Consumer Resource Model extends the classic Consumer Resource Model of MacArthur and Levins by incorporating the generic exchange of secondary metabolites observed in microbial communities, as described in the Methods. Each consumed resource type α (stars, squares, circles) with abundance R_α is taken up by species i at a rate c_iαR_α, and transformed into other resource types through metabolic reactions inside each cell with normalized stoichiometry matrix D_αβ. A fraction l of the resulting chemical flux returns to the environment, where it can be consumed by other microbes, while the rest is retained and used for growth. (b) Communities are initialized by randomly sampling subsets of species from a given regional pool, simulating the effect of stochastic colonization. The importance of dispersal limitation for community assembly can be tuned by adjusting the number of species in each the subset. (c) Each community is supplied with a constant influx of specified resource types, and all resources are diluted at a fixed rate. We assume that each community is well-mixed, so that its state is fully defined by the set of resource abundances R_α and microbial population sizes N_i. (d) Heat map of randomly sampled matrix of consumer preferences c_iα with S = 200 species and M = 100 resource types. (e) Heat map of randomly sampled metabolic matrix D_αβ, which encodes the allowed metabolic transformations and their relative rates, shown here with M = 100 resource types.

Figure 2

Relationship between diversity and environmental harshness is modulated by environmental complexity. Left: Gray dots are the number of distinguishable strains observed in each sample of the EMP, plotted vs. pH and temperature. Black dots represent the 99th percentile of all communities at a given pH or temperature. Colored lines are fits of a Laplacian and a Gaussian distribution to the 99the percentile points. Reproduced from Figure 2 of the initial open-access report on the results of the EMP. Right: The number of species surviving to steady state in simulated communities, plotted vs. environmental harshness. Harsher environments at extreme pH or temperature were simulated by increasing the total amount of resource consumption m_i required for growth (by the same amount for all species). Blue squares are simulation results when all the energy was supplied via a single resource type, while orange circles are simulations where the incoming energy was evenly divided over all 90 possible resource types. See main text and Methods for simulation details.

Figure 3

Nestedness of community composition indicates selection-dominated community assembly. Top: Presence (colored) or absence (white) of each microbial phylum in a representative set of 2,000 samples from the EMP. Reproduced from Figure 3 of the EMP report. Different colors represent different biomes. Bottom: Presence (black) or absence (white) of species in simulated communities. Two different regimes of community assembly were simulated. The first is the selection-dominated scenario of Fig. 2, where variability in diversity is produced by variations in environmental harshness, and all samples are initialized with the vast majority (150/180) of the species in the regional pool. The second is a dispersal-dominated scenario, where environmental conditions are identical for all samples, but each sample is initialized with a different number of species, varying from 1 to 180. See main text and Methods for simulation details.

Figure 4

Incorporating metabolic and taxonomic structure. (a) Three-tiered secretion model used for simulating human and marine microbiomes. M = 300 resource types are grouped into T = 6 classes of equal size, labeled A through F. These groups represent different kinds of metabolites, e.g. lipids, sugars, amino acids, etc. Group F is the “waste” class, containing common byproducts generated by many metabolic pathways, e.g., carboxylic acids. A fraction f_s of the byproduct flux from metabolism of a given resource is partitioned among resources of the same class. A fraction f_w of the flux is partitioned among “waste” resources (class F). The rest of the flux is nonspecifically partitioned among all the other classes. In all simulations shown here, f_s = f_w = 0.45. (b) Heatmap of a metabolic matrix D_αβ encoding the three-tiered secretion model. (c) Taxonomic structure used for human and marine microbiome simulations. Microbial species are grouped into “families,” with each family specializing in a different resource class. Specialist families allocate a fraction q of their consumption capacity to their favored resource class. In all the simulations shown here, q = 0.9. There is also a generalist family whose preferences are uniformly sampled across all resource types.

Figure 5

Metabolic and taxonomic structure give rise to Fisher log series Left: Tag sequence count distribution for a representative sea surface sample from the Tara Oceans Project. Data was subsampled 300 times at a depth of 10,000 reads (out of 129,135 in the original sample), and species with 5 reads or less in the raw data were treated as extinct for the purpose of computing the Fisher log series parameters (see Methods). Right: Abundance distributions for simulated communities. 1,000 individuals were sampled from each of 900 simulated communities, with environments and colonization as described for the “Simple Environments” panel of Fig. 6 below. Each point is an average over all 900 communities of the number of species with a given number of individuals. All simulations were performed with the metabolic structure described in Fig. 4 above. The left-hand panel also incorporated taxonomic structure, with different families specializing in different resource classes, with specialization level q = 0.9. The right-hand panel did not have taxonomic structure (q = 0), and consumption preferences for all species were sampled from the same Bernoulli distribution. Green curve (“Truncated Gaussian”) comes from assuming that species’ invasion fitness are sampled from a Gaussian distribution, and that population sizes for surviving species are proportional to the invasion fitness, while species with negative invasion fitness go extinct. See Methods for details.

Figure 6

Low-dimensional nutrient supply variation reproduces patterns in human microbiome survey data. Top: Each column represents one sample from the Human Microbiome Project (HMP). Colored segments represent relative abundances of different phyla in each community. Reproduced from Figure 2 of the initial open-access report on the results of the HMP. Bottom: Each column represents one of 900 simulated samples, each stochastically colonized with 2,500 species from a regional pool of 5,000 species, comprising seven metabolically distinct families. Colored segments represent relative abundances of the seven families defined in Fig. 4. Each of the three “body sites” was supplied with resources from a different pair of resource classes, with total nutrient supply fixed. In the first set of simulations (left), one resource from each class was supplied, and the ratio of the two supply rates was randomly varied from sample to sample. In the second set (right), all resources from each class were supplied, with randomly chosen supply rates for each sample, normalized to keep the total supply fixed. The brown family present in all three environments specializes in the typical byproducts (e.g., carboxylic acids) generated from all the other resource classes. Within each body site, samples are sorted by relative abundance of this family. See main text and Methods for simulation details.

Figure 7

Correlations between inter-site nutrient variation and metabolic structure affect distinguishability of body sites. Left: Principal coordinate analysis (PCoA) of MetaHIT OTU-level community compositions, using the Jensen-Shannon distance metric. Data points are colored by the body site from which the sample was taken. Reproduced with permission from Figure 1 of Ref. . Right: Jensen-Shannon PCoA of species-level compositions of the simulated communities. In the first set of simulations (left), the nutrients supplied to different body sites come from different resource classes. This is the same set of simulations used for the left-hand panel of Fig. 6, but similar results are obtained if the simulations of the other panel are used instead, or if consumption preferences are uniformly random with no taxonomic structure (See Supplementary Fig. S2). In the second set of simulations (right), each environment is supplied with a randomly chosen set of resource types, with each site being supplied with about one third of the 300 possible resources. See main text and Methods for simulation details.

Figure 8

Pattern in ordination of compositions from single body site admits of multiple explanations. Left: Jensen-Shannon PCoA of MetaHIT stool samples, showing a characteristic ‘U’ shape that has been observed in many independent studies. Colors indicate three hypothesized enterotypes, which we do not discuss here. Reproduced with permission from Figure 3 of Ref. . Right: Jensen-Shannon PCoA of simulated samples from Body Site 1 under two different levels of dispersal limitation. In the first (top), each community was initialized with 2,500 randomly chosen species out of the regional pool of 5,000. The communities display a continuous gradient in the population size of the most abundant species (over all samples) along the ‘u’ shape from one end to the other. In the second (bottom), each community started with 4,900 species. These communities display a continuous gradient of environmental conditions along the ‘U’ shape from one end to another.

Figure 9

Host-specific dynamics are compatible with dissimilarity-overlap correlation. Top left: The composition of pairs of samples can be compared in two independent ways: “overlap” measures the fraction of each sample comprised by species common to both, and “dissimilarity” measures how different the relative abundance profiles are within this shared pool. The four pairs shown here have increasing overlap and decreasing dissimilarity from left to right, corresponding to the four points indicated in the scatter plot. Dissimilarity and overlap are plotted for 17,955 pairs of stool samples from the HMP, analyzed at the genus level. Solid line is a Lowess smoothing of the data, and red points correspond to the sample pairs illustrated in the first panel. Reproduced with permission from. Top right: Dissimilarity and overlap for 10,000 pairs of simulated samples from the metabolically distinct simple environments of Fig. 6, with one resource supplied from class A and one from class B. Solid line is a Lowess smoothing of the data. Blue and green points correspond to two representative pairs of communities selected for further analysis in the bottom panel. Bottom: For each sample, the population dynamics near the steady state was approximated with a generalized Lotka-Volterra model. Effective carrying capacities and interaction coefficients computed from the mechanistic model parameters together with the population sizes and resource abundances, as described in the Methods. We have plotted the carrying capacity of each species for two representative pairs of communities with low (left) and high (center) overlap. These pairs are indicated in the scatter plots by blue and green points, respectively. We also show the normalized root-mean-square variability in carrying capacity for all 10,000 sample pairs (right).

Figure 10

Modularity of community assembly. Left: In the experiments reported in, synthetic beads composed of different kinds of polysaccharides, including agarose, alginate and carrageenan, were incubated with coastal seawater and colonized by the marine bacteria resident in the seawater sample. 16S rRNA amplicon profiling was performed for communities grown on beads composed of a single kind of polysaccharide, as well as mixtures of two kinds of polysaccharides. Relative abundances of amplicon sequences variants for two different mixtures (Agarose/Alginate and Agaraose/Carrageenan) are plotted versus a weighted average of the relative abundances on the pure beads. Solid lines are fits to a linear mixture model, with R² of 0.84 and 0.74, respectively. Right: Abundance of each species in simulated communities supplied with mixtures of two resource types, plotted against the average of the abundances for communities supplied with just one of the resource types, with the total energy supply held constant. For the first two panels, all other parameters are the same as for the human microbiome simulations of Fig. 6, except that each sample is initialized with all 5,000 species from the regional pool. Titles indicate the class labels of the two supplied resources for each scenario, and species are colored by metabolic family following Fig. 4. In the third panel, simulations were run with the same number of resources and species, but with all resources assigned to the same resource class, eliminating all metabolic and taxonomic structure. Solid lines are predictions of the additive model where the abundance in the mixture equals the average of the abundances in the single-resource condition. The R² score of this model is also shown in each panel.