Gut microbiome heritability is near-universal but environmentally contingent

Abstract

Relatives have more similar gut microbiomes than non-relatives, but the degree to which this similarity results from shared genotypes versus shared environments has been controversial. Here, we leverage 16,234 gut microbiome profiles, collected over 14 years from 585 wild baboons, to reveal that host genetic effects on the gut microbiome are near-universal. Controlling for diet, age, and socioecological variation, 97% of microbiome phenotypes are significantly heritable, including several reported as heritable in humans. Heritability is typically low (mean=0.068), but is systematically greater in the dry season, with low diet diversity, and in older hosts. We show that longitudinal profiles and large sample sizes are crucial to quantify microbiome heritability and indicate scope for selection on microbiome characteristics as a host phenotype.

One Sentence Summary:

The presence and abundance of most taxa in the baboon gut microbiome are heritable, but heritability estimates change over time and depend on host environment and age.

An important goal of microbiome research is to determine the heritability of gut microbiome traits (1–8). Linking microbiome variation to host genetic variation can reveal which aspects of the microbiome are capable of responding to selection on the host, suggest which microbiome traits are under host control, and connect microbial abundance to host pathways and disease states (1, 7). However, current research suggests that heritable gut microbiome taxa are uncommon. In humans, only 3 to 13% of gut microbes have non-zero heritability, and one study estimated that overall microbiome heritability may be as low as 0.019 (1, 2, 4, 6, 7). Furthermore, the few heritable microbiome phenotypes in humans, such as the abundance of the family Christensenellaceae, exhibit widely varying heritability estimates across studies (h²=0.31 – 0.64) (1, 2, 4, 6, 7, 9).

There are challenges in accurately estimating narrow-sense heritability (h², the proportion of phenotypic variance explained by additive genetic variance) for the human microbiome. First, relatives—especially twins and other first-degree relatives, which are the basis for most microbiome heritability studies—often share diets, behaviors, and built environments, which can lead heritability to be overestimated (10). Controlling for gene-environment correlations requires fine-grained, individual-based environmental and behavioral data, which have not been available in previous studies (1, 2, 4–6). Second, all current estimates of microbiome heritability in humans rely on cross-sectional microbiome sampling, even though microbial abundances are dynamic and difficult to accurately phenotype from one-time measures (1, 2, 4, 6, 7, 11). Further, h² can change over a host’s lifetime due to shifting environmental conditions and host attributes (e.g., h² for body mass index decreases with age, as dietary and behavioral effects increase relative to the effects of genotype: (12, 13)). To date, no studies of gut microbiome heritability have fully accounted for this temporal variability or its dependence on the environment.

Estimating microbiome heritability in a natural primate population

To overcome these challenges, we estimated h² for gut microbiome traits in 585 wild baboons (Papio cynocephalus, the yellow baboon, with some admixture from anubis baboons, Papio anubis; Fig. 1A). To do so, we used 16,234 16S rRNA gene sequencing-based microbiome profiles from fecal samples collected longitudinally over 14 years. These samples were collected from the Amboseli baboon population (Fig. 1B), which has been the subject of individual-based research since 1971 (14). Each study subject had on average 28 samples collected across 4.5 years (range=1 to 177 samples per baboon; median days between samples=28; Fig. 1A).

Fig. 1. Time-series data used to estimate microbiome heritability.

(A) The data set consisted of 16,234 microbiome samples collected from 585 individually-recognized baboons. Each point represents a sample; the y-axis is ordered by baboon age at first sample collection. (B) Map of the 90% kernel density estimate (KDE) home ranges and active dates for the 10 baboon social groups sampled over the study period, based on 71,645 GPS points collected during group monitoring. (C) For each microbiome sample, we had data on the diet consumed by members of the corresponding social group in the 30 days prior to sample collection. Each vertical bar represents one sample, ordered by collection date. Colors represent diet components (see table S1). (D) The relative abundances of microbial phyla in the current study are similar to prior primate studies (18–21). (E) The relative abundance of microbial phyla in all 16,234 samples ordered by collection date. ‘Rare’ taxa: <0.5% mean relative abundance per sample. In (C) and (E), the x-axis starts in the year 2000. The same legend applies for (D) and (E).

Baboons lead shorter lives than humans, so these time series often span a substantial fraction of the baboon lifespan (female life expectancy at birth is 10 years; females and males achieve sexual maturity at 4.5 and 5.7 years respectively; (15)). Each microbiome sample is accompanied by detailed information on the pedigree relationships of its donor (fig. S1), as well as fine-grained data on environmental conditions, social behavior, demography, and group-level diet composition at the time of sampling (Fig. 1C; table S1, S2). These complementary data allowed us to achieve precise estimates of heritability and quantify the impact of shifting environmental and social conditions on heritability. They also break apart gene-environment correlations: baboon social groups contain a wide range of maternal, paternal, and non-relatives (median within-group relatedness in a given year=0.055, SD=0.11), yet all group members travel in a coordinated fashion across the landscape and feed on the same seasonally available foods (14). Additionally, groups exhibit substantial home range overlap (Fig. 1B; (16)). Here, we studied ten social groups that varied in size from 17 to 118 members (mean=58).

Each 16S gut microbiome profile was generated from a fecal sample collected from an individually recognized baboon and processed following (17) (figs. S2–S4; table S3). Similar to other primates (Fig. 1D; (18–21)), the most common gut microbial phyla were Firmicutes, Bacteroidetes, and Actinobacteria (Fig. 1E). Both the abundances of these phyla and the composition of baboon diets show cyclic fluctuations (Fig. 1C, 1E), which reflect Amboseli’s wet-dry seasonal dynamics (14).

Using these microbiome profiles, we estimated the narrow-sense heritability (h²) of 1,034 gut microbiome phenotypes. These included 7 community phenotypes, or measures of microbiome community composition (amplicon sequence variant (ASV) richness, ASV Shannon’s H index, and the first five principal coordinates (PCs) of a Bray-Curtis dissimilarity matrix), and 283 single-taxon phenotypes representing the relative abundance of individual microbiome taxa, from ASVs through phyla, found in >50% of samples (figs. S5–S6; (1–6, 17, 22)). We also estimated h² for 744 presence/absence phenotypes, which reflect whether a taxon is present or absent in a sample (limited to taxa found in 10% to 90% of samples).

We estimated h² separately for each phenotype using the animal model, implemented in ASReml-R v3 (tables S4–S5; (23)). This mixed-effects model estimates each individual’s additive genetic value as a random effect, based on the expected covariance in additive genetic effects between relatives in a pedigree (24–26). It also partitions phenotypic variance across additive genetic variance and other random effects, after conditioning on fixed effects. Following the typical approach in human genetics and plant and animal breeding, we estimated total phenotypic variance (the denominator of h²) after correcting for fixed effects (12, 24, 27). This allowed us to exclude the effects of environmentally variable traits like diet and rainfall, technical effects, and demographic variables like sex and age (figs. S7, S8).

Genetic effects on the gut microbiome are near-universal

We found that 97% of single-taxon and community phenotypes were significantly heritable, including all 7 community phenotypes and 93% (273/283) of single-taxon phenotypes (likelihood ratio test; FDR threshold=0.1; Figs. 2A, 2B, S9, S10; table S6). Heritability was not limited to prevalent taxa, as 95% of the 744 presence/absence phenotypes were also significantly heritable, some of which were found in only 10% of samples (Fig. 2B; fig. S9A, S9B, S9C; table S7). However, more prevalent taxa tended to have higher h² (Pearson’s R=0.28, p=2.3×10⁻¹⁵; fig. S9D). Importantly, the proportion of significantly heritable single-taxon phenotypes was robust across phylogenetically- and compositionally-aware data transformations (PhILR transformation=96% heritable; CLR transformation=99% heritable; FDR threshold = 0.1; Fig. 2B; tables S8–S9). Heritability estimates were correlated between single-taxon phenotypes and CLR transformed single-taxon phenotypes (Pearson’s R=0.82, p=2.3×10⁻⁶⁹), and between single-taxon phenotypes and presence/absence phenotypes (Pearson’s R=0.68, p=3.2×10⁻²⁹; fig. S9E, S9F).

Fig. 2. Most microbiome phenotypes are heritable.

(A) Heritability estimates for the 40 most heritable single-taxon phenotypes and all 7 community phenotypes. Red text indicates taxa that are also heritable in humans (1, 2, 4–6). (B) Heritability estimates were robust across data transformations. Dark purple bars show significantly heritable phenotypes; thin yellow bars indicate mean heritability. (C) Additive genetic variance explained significantly more variance in microbiome phenotypes than host identity or maternal effects. The y-axis is ordered by taxonomic level and h², as given in table S7. (D) For the 32 microbial taxa heritable in our study (x-axis) and at least one human study (y-axis; (1, 2, 4–6, 32, 33)), h² was correlated between baboons and humans (Pearson’s R = 0.52, p = 0.002).

The most heritable phenotype among the single-taxon and community phenotypes was the first principal coordinate of a PCoA of Bray-Curtis dissimilarities, which captures a global summary of variation in the baboon gut microbiome (h²=0.21; p=5.7×10⁻¹⁵; Fig. 2A; Bray-Curtis PC1 explained 19% of the variance in microbiome composition overall (17)). More closely related ASVs tended to have similar h² (Moran’s I=0.0996, p=0.001; and Pagel’s lambda=0.73, p=0.001), especially ASVs belonging to the families Prevotellaceae, Lachnospiraceae, and Ruminococcaceae (local Moran’s I p<0.05; fig. S11), suggesting a phylogenetic signal in microbe heritability.

While h² for single-taxon and community phenotypes tended to be low to modest (mean h² among the 280 significant phenotypes=0.068; range=0.008 – 0.21; Fig. 2C), heritability values for presence/absence traits were significantly higher (paired t-test p=2.2×10⁻²⁷; mean h²=0.077, maximum h²=0.26; Fig. 2B; fig. S9A, S9B, S9C, S9F), as are heritability estimates from compositionally-aware abundance transformations (paired t-test p=2.7×10⁻²⁷; mean h²=0.084, maximum h²=0.20; Fig. 2B; fig. S9E). Overall, these values are similar to the heritability of social behavioral traits in nonhuman primates (fig. S12; table S10) and traits with strong social components in humans (28, 29), but exceed most available estimates for fitness in animal populations (30).

Across traits, host genotype explained more variance than host identity (paired t-test p=2.4×10⁻²⁷) or maternal effects (paired t-test p=5.5×10⁻⁸⁶). These results suggest that host genotype is more important in creating familial similarity in baboon microbiome composition than matrilines, even though matrilines form the core kinship units in baboon societies (Fig. 2C; fig. S9B, S9C). Further, we found no evidence that microbial transmission between relatives or assortative mating inflates h². Parent pairs did not have more similar microbiome composition than non-parent female-male pairs, as would be expected under assortative mating by microbiome composition (Mantel test r=0.004, p=0.22; fig. S7B). In addition, accounting for grooming-based social interaction networks (in the subset of models where such networks could be robustly estimated; n=500) decreased h² by only 0.0051 on average, and did not significantly improve any models (fig. S7C; table S11). We note that the weak effects of social networks on microbiome similarity are likely due to the longitudinal nature of this data set. In our population, social effects on microbiome composition are strongest between samples collected in the same month, and samples from social partners separated by long time periods are not especially similar (18, 31).

Humans and baboons share heritable taxa

We next investigated whether similar gut microbiome taxa are influenced by host genotype across baboons and humans, which would suggest that trait heritability in the microbiome is conserved. Heritability estimates were correlated for the 32 microbiome taxa found to be heritable in both our study and in at least one of 7 human data sets from 5 studies (n=3,511 aggregate sample size in total; (1, 2, 4–6, 17, 32, 33)), despite substantial methodological variation in data collection and methods for h² estimation (Pearson’s R=0.52, p=0.002; results are consistent using a linear mixed model that controls for study: b=0.91, p=0.014; Fig. 2D; table S12). Shared, heritable taxa include the family Christensenellaceae, one of the most consistently heritable phenotypes in humans (baboon: single-taxon h²=0.12; presence/absence h²= 0.20; humans: 0.31–0.64); Fig. 2A; fig. S9A; (1, 2, 4, 6, 7, 9, 34)). In contrast to a previous study in humans, heritable taxa did not co-occur more frequently than expected within hosts (1). However, more heritable taxa did exhibit higher connectivity in taxon co-occurrence networks (Pearson’s R = 0.58, p = 0.006; fig. S13).

Year, season, and host age modify heritability estimates

To understand why microbiome h² estimates often vary across studies (1, 2, 4–6), we then investigated social and environmental factors that systematically influence trait heritability. Here, we focused on a refined set of 100 collapsed phenotypes, including the 7 community phenotypes and 93 single-taxon phenotypes in which we collapsed phylogenetically nested taxa to the lowest taxonomic level (following (1, 2, 6); fig. S14; table S13). We found that host traits and environmental conditions had substantial effects on h². Across years, h² calculated for a single year can differ by up to 0.24 compared to h² calculated using all years (n=15 most heritable collapsed phenotypes, evaluated in years with at least 150 individuals and 1000 samples; table S11; Fig. 3A). For example, although h² for Christensenellaceae R-7 group (the collapsed phenotype for Christensenellaceae)=0.12 across all years, its annual h² estimates range from 0.06 (in 2002) to 0.18 (in 2007).

Fig. 3. Heritability estimates are affected by year, season, and host age.

(A) Heritability estimates varied across years. Panels show h² for the 15 most heritable collapsed phenotypes in years with sufficient sample size (>150 baboons and >1000 total samples). (B) Heritability estimates for all 100 collapsed phenotypes were highly correlated between seasons (black line; R = 0.83, p = 4.7×10⁻²⁷). Dashed line is x = y. (C) Heritability estimates for collapsed phenotypes were higher in the dry season than in the wet season (n = 89 taxa heritable in both seasons; paired t-test p = 4.4×10⁻¹²). (D) Dietary diversity was highest in the wet season (paired t-test; p = 4.2×10⁻⁵). (E) Heritability increased with age for 29/100 collapsed phenotypes). Each density plot represents the observed h² for these 29 collapsed phenotypes across 3-year sliding age classes. The yellow line indicates mean h² across all age classes. (F) Heritability estimates per age window for the 10 collapsed phenotypes with the steepest increase in h² with host age.

Within years, we also observed systematic effects of wet/dry seasonal dynamics on microbiome heritability. On the basis of the 89 collapsed phenotypes that were heritable in both dry and wet season samples (estimated separately; red points in Fig. 3B), we found that h² was, on average, 48% higher in the dry season than in the wet season (paired t-test p=4.4×10⁻¹²; Fig. 3C), even though h² estimates were strongly correlated between seasons (Pearson’s R=0.81, p=3.5×10⁻²²; Fig. 3B). These seasonal differences in h² may be explained by seasonal changes in phenotypic variance (V_p): weather in Amboseli is highly variable during the 7-month wet season, with periods of intense rain followed by several weeks with little or no rain, compared to the near invariant dry season. In support, V_p for microbiome phenotypes was higher in wet versus dry seasons (paired t-test p=4.2×10⁻⁵; fig. S15A). Baboons also consume a greater diversity and evenness of food types in the wet season compared to the dry season (linear mixed model; b=0.15, p=5.9×10⁻¹¹⁴; Fig. 3D). Although diet composition and rainfall per se are included in our models, individuals who eat diverse diets may also experience season-dependent environmental variation that our model does not capture. To test this hypothesis, we stratified the data by dietary diversity, and found that heritability estimates were higher in the low diet diversity data set (paired t-test p=1.0×10⁻¹¹; fig. S15B, S15C; 72% of samples in the high diet diversity data set were collected in the wet season).

Host characteristics such as age can also modify trait heritability (13, 35). Indeed, we found that, for the majority of microbiome phenotypes, h² increased with host age. When we stratified the 100 collapsed phenotypes into overlapping 3-year age classes of similar sample size (table S14), we found that h² changed significantly with age for 32% of phenotypes, and 91% of these phenotypes (29 of 32) resulted in higher h² in older animals (linear models p<0.05; Fig. 3E), with a total increase in h² of up to 0.24 (Fig. 3F). This observation is driven by both increasing genetic contributions to gut microbiome variation with host age (i.e., increased V_A; linear mixed model, b=1.7×10⁻⁵, p=0.0085) and decreasing contributions from residual environmental variance (i.e., decreased V_R; linear mixed model, b=−2.9 ×10⁻⁵, p=1.5 ×10⁻⁴). Older baboons ate less diverse diets than younger baboons, regardless of season (linear mixed model; effect of age on diet diversity in the wet season: b=−1.6 ×10⁻², p=1.6×10⁻²⁴; effect of age on diet diversity in the dry season: b=−1.2 ×10⁻², p=2.0×10⁻¹¹; fig. S16A, S16B). In addition, females exhibited reduced social partner diversity with age (linear mixed model; b=−0.35, p=1.4×10⁻¹⁹; fig. S16C, S16D). Moreover, microbiome diversity (Shannon’s H) also decreased slightly with age (linear mixed model; b=−0.0063, p=0.024; fig. S16E) and its h² exhibited the sixth strongest increase with age (linear model; b=0.013, p=2.5×10⁻⁵; fig. 3F). A possible explanation for this pattern is behavioral canalization that is not fully captured by the diet composition effects in our models, whereby older baboons increase in behavioral conservatism with age.

Longitudinal sampling affects heritability estimation

Together, our results qualitatively differ from similar research on humans: instead of a very small number of heritable microbiome phenotypes, we find near-universal heritability (1, 2, 4, 6, 7, 9). Further, we explain systematic variation in h² on the basis of temporal, environmental, and individual characteristics. These findings suggest that deep, longitudinal sampling is required to accurately characterize microbiome heritability and account for potentially extensive temporal variation (Fig. 4A; note that trait heritability was not correlated with its variability across the life course [coefficient of variation in abundance]: R=−0.14, p=0.15 for n=357 individuals with >10 samples; fig. S17).

Fig. 4. Microbiome phenotypes are dynamic and sampling design affects heritability estimates.

(A) Highly heritable microbiome phenotypes fluctuate in abundance (y-axis) in individual hosts over time (x-axis), as shown by Bray-Curtis PC1 and Christensenellaceae. Each row represents a baboon with >100 samples. (B) Longitudinal sampling improves the detection of heritable phenotypes. Purple circles indicate the percent of significantly heritable taxa in our data set when subset from 1 to 20 samples per individual. Yellow circles are the percent of significantly heritable microbiome phenotypes in 7 human data sets from 5 studies ((1, 2, 4–6, 32, 33); note that the plotted points from (33) and (32) show near-perfect overlap). (C) Heritability varies widely at lower sampling depths, even for highly heritable phenotypes (x-axis). The range of h² from 100 random subsets at each sampling depth is shown on the y-axis. (D) The percentage of significantly heritable traits rises with increasing sample size. Plot shows the percentage of models (out of 100 subsamples) that were improved by adding pedigree information. Each line represents one of the 100 collapsed phenotypes.

In support, we found that sample size and longitudinal sampling affected both our ability to detect heritable microbiome phenotypes and the heritability estimates themselves. Specifically, if we simulated cross-sectional data by randomly subsetting our collapsed phenotype data set to one sample per individual (n=585 samples; repeated 100 times), we found that <5% of phenotypes were significantly heritable, on average (mean = 4.6%; 95% CI = 3.1–6.1%; Fig. 4B; table S12). This proportion is comparable to that described in most human studies, but increases with more longitudinal samples per individual (Fig. 4B). Further, when we randomly subset our collapsed phenotype data set to 1,000 samples (included repeated samples for some individuals), h² estimates fell outside their standard error in the full data set in an average of 74% of cases (across 100 random subsamples; Figs. 4C, S18; table S15). Increasing the subset size to 10,000 samples dropped this percentage to 11% (Figs. 4C, S18) and increased the number of significantly heritable phenotypes. With 1,000 samples, heritable microbiome phenotypes detected in the full data set were significantly heritable in only 38% of 100 subsamples, on average (Fig. 4D), but at 10,000 samples, this concordance rose to 98%.

Conclusions

Nearly all gut microbiome taxa are heritable in baboons, including both prevalent and rare taxa. Although the magnitude of these heritability estimates is typically small, some traits exhibit h² greater than 0.15 (n = 59/744 presence/absence phenotypes; 6/283 single-taxon phenotypes; 1/7 community phenotypes). The universal role played by host genetic variation in our data set contrasts with previous work in humans, which finds few heritable taxa (1, 2, 4, 6, 7). These data sets may have had limited power because all human studies to date have been cross-sectional, and may have lacked data on key environmental variables that mask or modify heritability levels (1, 2, 4, 6, 7). Further, h² for traits detected in both baboons and humans are correlated (Fig. 2D), suggesting that traits with low h² in baboons may also be heritable—but have gone undetected—in humans.

Our findings do, however, agree with the observation that environmental effects on gut microbiome variation are larger than additive genetic effects (7). Future work will help refine our understanding of these environmental influences, including whether they mediate and/or interact with the effects of host genotype. Additionally, as 16S rRNA sequencing data have limited resolution, large-scale metagenomic data will be important for understanding whether individual microbial strains or gene content are also heritable, and perhaps more interestingly, whether microbial genotype affects host heritability. Together, our work argues for a qualitative change in perspective: from a microbial landscape largely unaffected by host genotype to one in which host genetics plays a consistent and sometimes appreciable role. These qualities imply that microbiome traits are therefore visible to natural selection on the host genome.

Data and materials availability:

Our data and code are publicly available, but the original biological and DNA samples cannot be shared due to restrictions on third-party sharing for CITES-regulated samples exported from Kenya. 16S rRNA gene sequences are deposited on EBI-ENA (project ERP119849) and Qiita (study 12949; (36)). Analyzed data and code are available on Zenodo (DOI: 10.5281/zenodo.4662081; (37)).