A macroecological perspective on genetic diversity in the human gut microbiome

Abstract

While the human gut microbiome has been intensely studied, we have yet to obtain a sufficient understanding of the genetic diversity that it harbors. Research efforts have demonstrated that a considerable fraction of within-host genetic variation in the human gut is driven by the ecological dynamics of co-occurring strains belonging to the same species, suggesting that an ecological lens may provide insight into empirical patterns of genetic diversity. Indeed, an ecological model of self-limiting growth and environmental noise known as the Stochastic Logistic Model (SLM) was recently shown to successfully predict the temporal dynamics of strains within a single human host. However, its ability to predict patterns of genetic diversity across human hosts has yet to be tested. In this manuscript I determine whether the predictions of the SLM explain patterns of genetic diversity across unrelated human hosts for 22 common microbial species. Specifically, the stationary distribution of the SLM explains the distribution of allele frequencies across hosts and predicts the fraction of hosts harboring a given allele (i.e., prevalence) for a considerable fraction of sites. The accuracy of the SLM was correlated with independent estimates of strain structure, suggesting that patterns of genetic diversity in the gut microbiome follow statistically similar forms across human hosts due to the existence of strain-level ecology.

Fig 1. Distributions of genetic diversity exhibit similar statistical forms across phylogenetically distant species in the human gut.

a,b) Similarity in patterns of genetic diversity was evaluated for sites obtained from the 22 most prevalent bacterial species. c) The distribution of within-host allele frequencies across all hosts as well as d) the distribution of mean allele frequencies were rescaled to determine whether they exhibited similar forms, specifically by rescaling their logarithm using the standard score (i.e., z-score). In c, statistical fits of a gamma (the distribution predicted by the SLM) and a lognormal (a point of comparison) are illustrated as black lines. To limit the effect of the bounded nature of allele frequencies on the distribution, mean frequencies containing observations of f = 1 were excluded from subplot d. e) The relationship between statistical moments of within-host allele frequencies was consistent across species, as there was a strong linear relationship between the mean frequency of an allele and its variance on a log-log scale (i.e., Taylor’s Law). To reduce the contribution of an excess number of zeros towards estimates of $\overline{f}$ and ${\sigma }_f^2$, alleles with non-zero values of f in <35% of hosts were excluded. f) Finally, the fraction of sites harboring alleles present in a given number of hosts decreased in a similar manner across species. All sites in this analysis are synonymous, identical analyses were performed on alleles at nonsynonymous sites (S1 Fig). Species within the same genus were assigned the same primary or secondary color with different degrees of saturation.

Fig 2. The SLM successfully predicts genetic patterns for prevalent alleles.

a) Predicted values of prevalence were obtained (Eq 6) and matched with their observed values, where the data generally fell on the one-to-one line (dashed black line) for sites with a prevalence ≳ 0.01. b) The pattern of the SLM performing better for higher values of prevalence was illustrated by quantifying the relative error of the prevalence predictions. The mean of the logarithm of the relative error of the SLM over all sites ($\overline{{\text{log}}_{10}}\epsilon$, dashed black line) was ∼0.1. c) The contingency of the SLM’s success was illustrated by examining the relationship between the mean frequency of an allele across hosts ($\overline{f}$) and its prevalence. The predictions of the SLM (not a statistical fit) succeed for high mean frequency alleles (dashed black line). All analyses here were performed on alleles at synonymous sites using the common commensal gut species B. vulgatus. The color of each datapoint is proportionate to the number of sites. Visualizations of the predictions in this plot for all species for nonsynonymous and synonymous sites can be found in S1 Text.

Fig 3. The SLM succeeds and fails in a consistent manner across phylogenetically distant species.

a) Distributions of logarithmic relative errors of prevalence obtained from Eqs 6 and 21 were rescaled by their mean and standard deviation to illustrate their similarity across species. b) Binning observed prevalences reveals how the predicted values tend to follow a one-to-one relationship (dashed black line) across species, with variation among species. c) By calculating the correlation between log₁₀-transformed observed prevalence and the relative error of our predictions, one finds a negative correlation for the majority of species. A permutation test confirms that these negative correlations are significant (95% CIs as black lines) for the majority of species (16/22). All analyses in this plot were performed using synonymous sites. Identical analyses and equivalent results for nonsynonymous sites can be found in S11 Fig.

Fig 4. The presence of strain structure is correlated with the accuracy of the SLM.

a) The presence or absence of strain structure was inferred from the distribution of allele frequencies for each species within each host using StrainFinder, providing an estimate of the fraction of hosts that harbor strain structure. b) This per-species estimate of strain structure can be compared to the mean relative error of allelic prevalence predictions obtained using the SLM (Eq 6) to determine whether the success of the SLM is correlated with the existence of strains. By examining this relationship when sites with rare alleles (i.e., low prevalence) are included ($\widehat{\varrho }\ge 0.05$: dashed line) and excluded ($\widehat{\varrho }\ge 0.15$: solid line), one sees a stronger correlation ffor alleles with a high prevalence threshold. c) This trend can be systematically evaluated by calculating the correlation across a range of prevalence values (black dots). A permutation test establishes 95% confidence intervals of the null (grey window).