Interhost dispersal alters microbiome assembly and can overwhelm host innate immunity in an experimental zebrafish model

Significance

Manipulating the microbial communities associated with animals to improve host health requires a comprehensive understanding of the mechanisms driving microbiome variation, which a strict focus on host-specific factors has been insufficient in providing. We performed an experiment to test whether the movement, or dispersal, of microorganisms among zebrafish hosts could alter the effects of important host factors, using a dispersal-based model to guide the interpretation of results. We observed that interhost dispersal can alter the diversity and composition of microbial communities and overwhelm the effects of the host’s innate immune system. These findings suggest that dispersal is an important mechanism driving microbiome variation and should be considered in future microbiome research.

Abstract

The diverse collections of microorganisms associated with humans and other animals, collectively referred to as their “microbiome,” are critical for host health, but the mechanisms that govern their assembly are poorly understood. This has made it difficult to identify consistent host factors that explain variation in microbiomes across hosts, despite large-scale sampling efforts. While ecological theory predicts that the movement, or dispersal, of individuals can have profound and predictable consequences on community assembly, its role in the assembly of animal-associated microbiomes remains underexplored. Here, we show that dispersal of microorganisms among hosts can contribute substantially to microbiome variation, and is able to overwhelm the effects of individual host factors, in an experimental test of ecological theory. We manipulated dispersal among wild-type and immune-deficient myd88 knockout zebrafish and observed that interhost dispersal had a large effect on the diversity and composition of intestinal microbiomes. Interhost dispersal was strong enough to overwhelm the effects of host factors, largely eliminating differences between wild-type and immune-deficient hosts, regardless of whether dispersal occurred within or between genotypes, suggesting dispersal can independently alter the ecology of microbiomes. Our observations are consistent with a predictive model that assumes metacommunity dynamics and are likely mediated by dispersal-related microbial traits. These results illustrate the importance of microbial dispersal to animal microbiomes and motivate its integration into the study of host–microbe systems.

The communities of microorganisms associated with animals, referred to as the “microbiome,” are highly diverse and have the potential to strongly influence host health. Understanding how microbiomes contribute to host physiology, and how to manipulate this relationship to promote host health, requires a comprehensive understanding of the mechanistic drivers of microbiome variation across hosts. Unfortunately, it has been difficult to identify consistent host factors that can explain the large amounts of the variation in microbiome composition across individual hosts, despite large-scale sampling efforts (1). At best, only a small fraction of variation across hosts can be explained by individual host factors, leading to the perception that the rules governing microbiome assembly are idiosyncratic. However, unlike many other attributes of an animal’s biology that impact its health and fitness, an animal’s microbiome is subject to dispersal of microorganisms from other hosts. If the influence of microbial dispersal among hosts is substantial, then a comprehensive model of microbiome dynamics must include consideration of not just the factors associated with individual hosts but also the population of hosts with which they exchange microbiome members.

Dispersal is increasingly recognized as an important determinant of the structure and function of both experimentally assembled (2, 3) and naturally occurring bacterial communities (4, 5), and there is mounting evidence that dispersal is also important to the assembly of nonpathogenic, animal microbiomes. Biogeographic patterns have been observed for microbiomes associated with natural populations of animals (6–8), consistent with predicted effects of dispersal. Social interactions among hosts, a possible facilitator of microbial dispersal, have been shown to correlate with the composition of animal microbiomes, with hosts tending to share more members of their microbiome with the microbiomes of individuals with whom they interact frequently (9–11). Dispersal has also been hypothesized to explain differences in the microbiomes of humans in economically developed and developing regions (12). Studies of laboratory animals often report that the microbiomes of animals housed together are more similar than those in different housing units. These so called “cage effects” routinely explain significant amounts of microbiome variation, as well as variation in phenotypes known or suspected to be mediated by the microbiome (13–15). Interestingly, experiments studying the innate immune system have often shown that cohousing of healthy and immune-deficient animals can transfer phenotypes associated with immune pathway mutants, including increased inflammation and colitis (16, 17). Similar investigations of the link between innate immunity and microbiomes have led to conflicting or inconclusive results, with some finding little to no effect of innate immune pathways on microbiome composition or diversity, especially in cases where both wild-type (WT) and immune-deficient animals were housed together or from the same litter (18–20). These examples are particularly interesting given the role the immune system plays in direct interactions between animals and their resident microorganisms, suggesting dispersal of nonpathogenic microorganisms may have important consequences to animal hosts.

Research on host-associated microbiomes has increasingly utilized frameworks from general ecological theory to guide experiments and interpret patterns such as those described above. Metacommunity theory in particular focuses on dispersal among multiple discrete “local” communities (21) and is thus potentially well suited to describing host–microbe systems, where hosts act as environments that are home to local communities of microorganisms linked by interhost dispersal (22, 23). Dispersal, as well as metacommunity theory specifically, has been invoked to explain many patterns in microbiome diversity and composition (e.g., ref. 12). While the results of these studies are often consistent with the predicted effects of dispersal, these studies are not designed to directly test the importance of interhost dispersal on the assembly of host-associated microbiomes and often struggle to disentangle the effects of dispersal from other confounding factors. Furthermore, they provide limited insight into the mechanisms by which dispersal processes result in such patterns.

Here, we describe an experiment, guided by a quantitative predictive framework, that explicitly manipulates interhost dispersal to test its role in microbiome assembly and gain insights into the underlying mechanistic processes. Specifically, we tested whether dispersal of microorganisms among hosts can influence or overwhelm the effects of individual host factors, namely the innate immune system, on the composition and diversity of zebrafish (Danio rerio) intestinal microbiomes. The relatively simple husbandry and large clutch sizes of zebrafish allowed us to manipulate the transmission of microorganisms among a large number of replicate individuals through cohousing and isolation at a scale not feasible in humans or other vertebrate models, while their genetic malleability made it possible to focus on the effects of host innate immunity through the generation of immune-deficient mutants. To generate specific predictions and to guide the interpretation of our results, we created a computational model assuming metacommunity dynamics across hosts. We observed that the effects of dispersal among zebrafish on microbiome composition and diversity are largely consistent with our model and can overwhelm the effects of host innate-immune activity.

Results

We generated an immune-deficient myd88⁻ mutant zebrafish line and raised homozygous mutant with WT zebrafish under three housing conditions designed to either allow or restrict dispersal among hosts: “solitary” conditions in which each individual zebrafish was in isolation with no exposure to other individuals (i.e., no interhost dispersal), or cohoused, either with members of the same genotype only (“separated”) or with members of both genotypes (“mixed”; Fig. 1). The myeloid differentiation primary response gene 88 (MyD88) encodes a universal adapter protein in the Toll-like receptor (TLR) pathway and is responsible for activating several immune responses in response to signaling from the microbiota, including the production of proinflammatory cytokines and antimicrobial peptides and the detoxification of the bacterial product lipopolysaccharide (24–26). The germline mutation in MyD88 was generated using CRISPR/Cas9, and the resulting myd88⁻ mutant was confirmed to have the expected phenotype of low neutrophil abundance in the intestines of conventionally reared larvae, with abundances similar to those of WT larvae raised germ-free, as previously described for zebrafish injected with a MyD88 morpholino (24) (Generation and Verification of a myd88 Mutant Zebrafish and Fig. S1). To isolate the effects of innate vs. adaptive immunity, we raised fish to 21 d postfertilization (dpf), as adaptive immunity does not become active in zebrafish until ∼28 dpf (27). At 21 dpf, we killed the fish and characterized the microbial communities associated with their intestines and those associated with their food and flask water by 16S amplicon sequencing. As one might expect given the importance of host immunity to defense against pathogens, myd88⁻ fish had higher mortality rates and, notably, their mortality rates were higher in the cohoused treatments compared with the solitarily treatment (Fig. S2). Interestingly, mortality rates for WT fish were also higher when cohoused, especially when cohoused with myd88⁻ fish. Because of this, by the end of the experiment, the number of fish in each flask were no longer equal. However, we did not observe a significant effect of the ultimate number of fish per flask on microbiome composition within treatments [permutational multivariate ANOVA (PerMANOVA): P > 0.05].

Fig. 1.

Experimental design. WT and myd88⁻ zebrafish were raised in one of three housing conditions to manipulate the degree of interhost dispersal: housed alone with no interhost dispersal (solitary), cohoused with only individuals of the same genotype (separated), or cohoused with individuals of both genotypes (mixed).

Fig. S1.

Sequence and functional confirmation of myd88 mutant. (A) Schematic of the myd88 gene exons and introns. The blue box indicates the region targeted by the CRISPR guide RNA. The red arrow indicates the 4-bp deletion. The red amino acids indicate a frameshift mutation followed by a truncation of the MyD88 protein before the Death Domain. (B) Functional MyD88 protein is required for intestinal neutrophil recruitment. We quantified intestinal neutrophil number in wild type and our myd88 mutant, which confirmed a loss of MyD88 function in the mutant. Student t test, ****P < 0.0001.

Fig. S2.

Proportion of surviving individuals in each genotype by housing treatment over time (days post fertilization). For the mixed-housing treatment, the genotype of individuals could not be determined until the end of the experiment, so the survivorship curve represents both genotypes until the last time point.

To determine whether the observed effects of housing treatments on microbiome diversity and composition were consistent with interhost dispersal being the primary driving mechanism, we compared our experimental results with predictions from a computational model assuming metacommunity dynamics (Table 1). In this model, individual hosts are home to local communities of microorganisms that are connected by dispersal to form a metacommunity comprised of all of the hosts within a population/flask (Fig. S3). Additional details regarding model construction and generation of predictions is available in Computational Metacommunity Model.

Table 1.

Predictions from a metacommunity model of the effects of interhost dispersal treatments on the diversity and composition of host-associated microbiomes

Microbiome structure	Solitary	Separated	Mixed
Correlation with host environment	Strong	Strong between/weak within genotypes	Weak
α-Diversity (within host)	Low	Intermediate	High
β-Diversity (interhost)	High	Intermediate	Low
γ-Diversity (across host)	Low	Intermediate	High
Abundance of dispersal specialist	Low	High	High

Fig. S3.

A conceptual description of the metacommunity model used to describe the dynamics of a host-associated metacommunity.

Interhost Dispersal Overwhelms Host Factors.

Overall, there was a significant difference in the composition of microbiomes associated with WT and myd88⁻ zebrafish, but the effect of host genotype was weaker than the effect of housing treatment across the entire dataset (Fig. 2A and Table 2). We predicted that the effect of host factors such as innate immunity would depend on the degree of interhost dispersal, due to the homogenizing effects of exchanging microbial taxa among host types (Table 1). In agreement with our predictions, we observed a strong interaction between housing and genotype. Specifically, there was a much greater difference in microbiome composition between genotypes when hosts were raised in solitary compared with cohoused treatments, either within or across genotypes (Table 2).

Fig. 2.

Relationship between host factors and microbiome composition. (A) A nonmetric multidimensional scaling ordination of Canberra distances among individual microbiomes. Ellipses represent 95% CIs around the group centroid. (B) Variance in microbiome composition explained by individual host factors following a redundancy analysis. Shown are the adjusted R² values for the unique and shared contribution of multiple host factors: il-1β expression (IL1B), c3 expression (C3), and standard length (SL). Negative adjusted R² values are treated and displayed as zeroes (considered as null).

Table 2.

PerMANOVA analysis on the effects of genotype and housing on microbiome composition

Factor	MS	F	R2	P*
Across-housing treatments
Genotype	0.007	7.2	0.05	<0.001
Housing	0.009	9.6	0.14	<0.001
Genotype × Housing	0.008	8.6	0.13	<0.001
Within-housing treatments
Genotype–solitary	0.067	34.1	0.58	<0.001
Genotype–separated	0.014	3.4	0.15	0.010
Genotype–mixed	0.004	2.8	0.05	0.001

^*P values calculated from a distribution of 1,000 random permutations.

To further investigate whether the effects of host factors were being weakened by dispersal at a finer, individual host level, we measured the relationship between attributes of each host and the composition of their microbiome within each housing treatment. We hypothesized that dispersal among heterogeneous hosts would dilute the effects of local host factors, and therefore the relationship would be strongest for solitary hosts and weakest for separated and mixed cohoused hosts. We first measured the standard length of each zebrafish (Fig. S4A), which is known to be an overall indicator of fish development and health (28), and which we had previously shown was a strong predictor of intestinal microbiomes across zebrafish development (29). We also characterized the level of innate immune activity of each host by measuring the transcriptional levels of two immune genes: one, c3, in the MyD88-independent complement pathway, and another, il-1β, in the MyD88-dependent pathway. As expected given their genotype, we found expression of il-1β to be lower in myd88⁻ compared with WT hosts, while expression of c3 was similar between the two genotypes (ANOVA on effects of genotype: F statistic = 13.9, P < 0.001 for il-1β and F statistic = 0.11, P = 0.74 for c3; Fig. S4 B and C). Despite having a strong effect on the microbiome diversity and composition, housing conditions had no clear effect on host innate immune response (ANOVA on effects of housing: P > 0.05 for both il-1β and c3). This reaffirmed our assumption that interhost dispersal primarily altered the degree of filtering of the microbiome by the local host environment rather than changing the host environment itself. To test this, we performed a redundancy analysis to determine the unique and shared contribution of each host factor to explaining the variance in community composition. Consistent with our hypothesis, a greater amount of variance in microbiome composition was explained by host factors in solitary than in separated or mixed hosts (Fig. 2B).

Fig. S4.

Measurements of individual host factors: Standard length (A), and the relative expression of host il-1β (B) and c3 (C) genes.

Interhost Dispersal Increases Diversity.

The strong independent effect of housing treatment on microbiome composition suggests dispersal fundamentally alters the structure of host-associated microbiomes. Our model predictions of the effects of dispersal on diversity were qualitatively in agreement with the predictions of general ecological theory (30, 31): dispersal among hosts increases α-diversity (i.e., within-host diversity) through the maintenance of taxa in hosts where they would otherwise go extinct, decreases β-diversity (i.e., variation among hosts) through the homogenizing effects of sharing individuals, and increases γ-diversity (i.e., across-host diversity) by allowing dispersal-specialized taxa to evade competitive exclusion at the metacommunity scale (Table 1).

Our experimental results were largely consistent with these theoretical predictions. Both within-host α-diversity (Fig. 3A; ANOVA on effects of housing: F statistic = 36.6, P < 0.001) was lower in the solitary treatment, relative to mixed and separated, with the mixed and separated treatments indistinguishable statistically. Variation among hosts, or β-diversity, was significantly greater in solitary hosts than mixed hosts, with separated hosts overall more similar to solitary than mixed hosts (Fig. 3B; ANOVA on effects of housing: F statistic = 26.5, P < 0.001). Across-host γ-diversity (Fig. 3C) was lower in the solitary treatment, relative to mixed and separated. While housing treatment had a large impact on microbiome diversity at multiple scales, there was no detectable difference between genotypes, with the exception of β-diversity within separated hosts (Fig. 3; ANOVA: P > 0.05). The most notable deviation from our predicted effects of dispersal on diversity was the effect of the separated housing treatment, which we anticipated would show a response intermediate to the solitary and mixed treatments. In contrast, we observed no difference in α- and γ-diversity between the separated and mixed cohousing treatments. The simplest explanation for this discrepancy is that innate immunity has a smaller effect on diversity than expected relative to the independent effects of dispersal.

Fig. 3.

Effects of host genotype and housing conditions on the composition and diversity of intestinal microbiomes. (A) Within-host α-diversity, measured by the Shannon diversity index. (B) Interhost β-diversity, measured by the Canberra distance from the group centroid (β-dispersion). (C) Across-host γ-diversity, measured by comparing the average gain in total OTU richness with increased sampling within each metacommunity treatment.

Interhost Dispersal Promotes Dispersal-Related Traits.

We unexpectedly observed that cohousing had a similar effect on microbiome composition and diversity regardless of whether hosts were cohoused with only members of the same genotype (separated) or with members of both genotypes (mixed; Fig. 2A; PerMANOVA: P > 0.05 for both WT and myd88⁻ comparisons). If dispersal acted only to homogenize microbiomes through the simple exchange of microorganisms among hosts, then we would expect that the two genotypes would maintain their distinctiveness in the separated treatment (since dispersal between genotypes is not possible). This suggested that dispersal among hosts altered the nature of selection for microbiome members, potentially by increasing the success of dispersal-adapted microorganisms. In further support of this hypothesis, both separated and mixed microbiomes were overall more diverse than solitary microbiomes (Fig. 3A) and were not limited to being a simple mix of taxa from solitary microbiomes. While the majority of taxa that occurred in solitary microbiomes were also detected in their cohoused counterparts (86% and 88% for solitary WT and myd88⁻ hosts, respectively), a much smaller proportion of taxa that occurred in cohoused microbiomes were also detected in their solitary counterparts (59% and 64% for cohoused WT and myd88⁻ hosts, respectively). Our computational model also predicted a similar increase in diversity in cohoused hosts. In particular, the model predicted that interhost dispersal would favor the persistence of species with greater dispersal rates (Fig. S5A), especially when a trade-off was imposed between dispersal rates and host specificity, such that species with high dispersal rates would be found in a wider range of host types, thereby homogenizing communities (Fig. S5B). This suggests that interhost dispersal allows for the success of taxa in the metacommunity as a whole that would not otherwise persist.

Fig. S5.

The distribution of species dispersal rates, β_d, and host specialization traits, β_n, in metacommunities simulated using a computational model of microbiome assembly. Density plots showing the distribution of dispersal rates (β_d) across species in simulated metacommunities either without (A) or with (B) a trade-off between dispersal rates and host specialization. Solid yellow lines represent microbiomes of solitary hosts (no interhost dispersal), dashed green lines represent microbiomes of separated hosts (interhost dispersal only among similar host genotypes), and dotted blue lines represent microbiomes of mixed hosts (interhost dispersal among all host genotypes).

If it is indeed the case that trade-offs in the life history strategies of microbial taxa underlie the effects of dispersal in this system, then we should see this reflected in the traits of microbial taxa associated with cohoused and solitary microbiomes. To test this prediction, we asked whether taxa associated with cohoused microbiomes were enriched for traits related to dispersal and colonization ability compared with taxa associated with solitary microbiomes. The taxa most differentially abundant in cohoused compared with solitary microbiomes primarily belonged to the genera Vibrio (log₂-fold change = 1.1, P < 0.0001) and Shewanella (log₂-fold change = 0.62, P < 0.01). To infer the traits of individual operational taxonomic units (OTUs) in our study, we used the ancestral state reconstructions implemented in PICRUSt to estimate the gene content of our observed OTUs by matching them to a reference database (32). We then asked which gene pathways were predicted to be enriched in those taxa that were differentially abundant in cohoused microbiomes compared with those associated with solitary microbiomes. Notably, the most strongly enriched pathways in cohoused associated taxa included ones related to bacterial motility (such as flagellar assembly and bacterial chemotaxis), quorum sensing (which is often used to regulate biofilm formation and virulence), two-component regulatory systems (which facilitate responses to changes in the environment), and Vibrio pathogenicity (Table 3). Although these traits are merely predictions based on similarity to a known reference database rather than direct measurements, they independently support our modeling predictions that dispersal among hosts selects for taxa with life history strategies favoring dispersal and repeated colonization of multiple host types. In the case of the Vibrio pathogenicity-related pathways, they may also explain the observed decreased survivorship of cohoused zebrafish of both genotypes (Fig. S2).

Table 3.

The top KEGG orthology pathways predicted to be enriched in taxa that are differentially abundant in cohoused microbiomes

KEGG orthology pathway	χ2 statistic	Adjusted P value
Vibrio cholerae pathogenic cycle	263	1.29 × 10−56
Bacterial chemotaxis	131	7.27 × 10−28
Quorum sensing	130	1.06 × 10−27
Two-component system	130	1.14 × 10−27
Phosphotransferase system (PTS)	100	4.93 × 10−21
Flagellar assembly	79	2.14 × 10−16
Vibrio cholerae infection	70	1.92 × 10−14
Aminobenzoate degradation	70	2.42 × 10−14
Phenylalanine metabolism	49	9.19 × 10−10
Benzoate degradation	47	2.32 × 10−9

Generation and Verification of a myd88 Mutant Zebrafish

Guide Design.

A guide was designed to introduce a lesion in the N terminus of MyD88. It targets a 5′-aaagaaactgggtctgttcc-3′ site that is 66 bp from the initiating ATG (Fig. S1A).

RNA Synthesis.

For making nls-cas9-nls (pT3TS-nCas9n; Addgene plasmid number 46757) (36), the template DNA was linearized with Xba1 and purified (Qiagen). RNA was synthesized with the T3 mMESSAGE kit (Invitrogen) and cleaned using the RNeasy kit (Qiagen). For making guide RNA (gRNA), we ordered a two partially overlapping long oligos and filled in the overhang with T4 RNA polymerase (Invitrogen). The oligos were a gRNA scaffold oligo (PAGE purified) with the following sequence: 5′-gatccgcaccgactcggtgccactttttcaagttgataacggactagccttattttaacttgctatttctagctctaaaac-3′ and a gene-specific oligo with the following sequence: 5′-aattaatacgactcactata aaagaaactaggtctgttcgttttagagctagaaatagc-3′, which includes a T7 promoter, the gene-specific target sequence, and the scaffold RNA overlapping sequence. After filling in by T4 polymerase, the DNA was used as a template for in vitro transcription with the MEGAscript T7 kit (Invitrogen), following manufacturer’s instructions. Resulting RNA was cleaned with the RNeasy kit (Qiagen).

Microinjection.

The mix of myd88 gRNA (50 ng/μL) and nlsCas9nls mRNA (150 ng/μL) was injected at a volume of ∼1 nL into the cell of one-cell stage ABC × Tu embryos. Sequencing confirmed polymorphisms at the target locus. The lesion chosen was a 4-bp deletion that introduces a frameshift and a stop codon that truncates the protein in the first half of the first exon (Fig. S1A) before the death domain.

Neutrophil Quantification.

To verify the phenotype of the mutant, intestinal neutrophils were quantified (23). The 6-dpf zebrafish larvae were fixed in 4% paraformaldehyde overnight. Whole larvae were stained with myeloperoxidase kit (Sigma) following the manufacturer’s protocol. Stained neutrophils were quantified by dissecting the intestine and counting positively stained cells in the intestine under a dissecting microscope (Fig. S1B).

Computational Metacommunity Model

Overview.

In this supplement, we go into greater detail about the computational metacommunity model used in our study. The interested user may also utilize the provided Mathematica code to generate similar predictions (Dataset S1). We begin by walking through each step of the model and noting the control parameters in each step (compiled in Table S1) and then explore the results of the specific choices for parameter values used in the current paper. All simulations were done in Mathematica software (Wolfram).

Table S1.

Parameters used in the computational metacommunity model

Parameter	Usage	Value
βn	Niche trait {0,1}
βd	Dispersal trait {0,1}, 1 = best disperser, 0 = never disperses
ρ	Density of species entering pool	0.001 or 0.1
ntrickle	The number of species entering the system each iteration	100
ndisp	The number of species chosen from the metacommunity to join the host community each step	5
ni	The density of species i
αι,φ	The competition coefficient between species i and j	Between 0 and 1
σ	The width of the competition kernel	0.1
niches	The number of peaks in the carrying-capacity function	3
δ	Offset of the carrying-capacity function	Mean of 0.0 or 0.5
τ	The mean number of timesteps for the Lotka–Volterra competition dynamics to run	10
pf, pd, p0	The probability of feeding, dispersal, or no event	0.49, 0.49, 0.02
f	Feeding factor	0.1
b	Steepness of the remaining density curve	6

Base Model.

We created a computational model tailored to simulate the dynamics of bacterial species as they disperse among, and compete within, animal hosts. Individual microbial species are differentiated by two traits: β_n, which sets the niche (i.e., host type) preference of the species, and β_d, which sets their dispersal ability. The model consists of iterating four steps: (i) immigration to the media from outside the system (i.e., from nonsterile food and media); (ii) colonization of hosts by the bacteria in the media; (iii) competition of the species within a host according to Lotka–Volterra dynamics; (iv) dispersal of bacteria from hosts to the media (Fig. S3).

The metacommunity model is set up to track the population densities of bacterial species as they move from an external species pool, into the media (for example, the water in the fish tank), into the host, and back into the media (Fig. S3). All of the bacterial species are defined by two traits: β_n, a trait that determines the niche requirements of a species; and β_d, a trait that determines that dispersal ability of a species. Both traits are bounded at 0 and 1. A given species in a given place (a host, the media, or the pool) has three attributes, its two traits and its density. Only the density is subject to change. The pool of available species is generated evenly along a grid, with species every 0.01. Species enter the system at a density, ρ, between 0.01 and 0.1, the number of different species entering the system during a single iteration is defined by ntrickle.

The immigrants from the pool join the community of bacteria in the media. It is from this community that colonists to the host are chosen. Species are selected randomly from the media community according to their density in that community. High-density species are more likely to be chosen than low-density species. Species are chosen from this pool with replacement so if more than once species is chosen it is possible for the same species to be chosen more than once. The number of species chosen is ndisp. The chosen species are added to the existing host community at the same density as they were in the media. If the same species is chosen more than once, or if that species is already present in the host, the densities are added together.

The community in a given host competes according to Lotka–Volterra dynamics:

$$\frac{d{n}_i}{dt}=n_i\left(1-\frac{{{\displaystyle \sum }}_j{\alpha }_{i,j}{n}_j}{K\left({\beta }_{n,i}\right)}\right),$$ [S1]

where n_i is the density of species i, α_i,j is the competition coefficient between species i and j, and K is the carrying-capacity function, which depends on the β_n of species i. The value of α_i,j is determined by the difference in β_n of the two species in question. β_n is defined on a ring, so β_n = 1 = 0.

$${\alpha }_{i,j}=e^{\left(\frac{-\mathrm{\Delta }{\left({\beta }_{n,i},{\beta }_{n,j}\right)}^2}{{\sigma }^2}\right)},$$ [S2]

where

$$\mathrm{\Delta }\left({\beta }_{n,i},{\beta }_{n,j}\right)=\min \left({\beta }_{n,i}-{\beta }_{n,j},{\beta }_{n,j}-{\beta }_{n,i}\right)$$ [S3]

and σ is a parameter that controls the width of the function. A higher σ leads to more intense competition for species with more distant β_n, a very low σ means that the species only really compete with those with nearly identical β_n. For all simulations presented in this paper, σ = 0.1. All α_i,j are symmetrical.

$$K\left({\beta }_{n,i}\right)=\frac{\sin \left(2\pi \ast niches\left({\beta }_{n,i}+\delta \right)\right)}{2}+0.5,$$ [S4]

where niches is a parameter that controls the number of peaks between 0 and 1, set to 3 for this simulation, and δ is the parameter that controls the offset of the peaks. This is different in every host, either only slightly between two hosts of the same type or completely out of phase for two hosts of different types. In this simulation, the mean offsets are either 0 or 0.5, to give out-of-phase carrying-capacity functions. Fig. S6 shows the effect of altering δ.

Fig. S6.

The metacommunity model’s carrying-capacity functions with two different offsets each corresponding to a different host environments. A species with a trait value β_n, which is optimal in the host with δ = 0 (A) will be poorly adapted to the second where δ = 0.5 (B). Thus, different hosts with different offsets select for different bacterial species.

The competitive dynamics are numerically simulated for a set number of time steps, the exact amount of time is chosen from a Poisson distribution around a mean time, τ (set to 10 for all simulations). After the dynamics are run, a disturbance event occurs, either a feeding event or a dispersal event. Which one of these events occurs is chosen randomly with the following probabilities: p_f, the probability of a feeding event; p_d, the probability of a dispersal event; or p₀, the probability of no event. After all of these events, the community is amended with new colonists according to ndisp, new immigrants are added to the metacommunity according to ntrickle, and the competitive dynamics are run again. If a feeding event occurs, the entire community is multiplied by the same proportion, according to the feeding factor, f (set to 0.1), to simulate an increase in the carrying capacity caused by additional resources. If a dispersal event occurs, then species either stay in the host or are dispersed into the media depending on their β_d. The density of the species that stays within the host is given by the following:

$$n_i\left(after\hspace{0.17em}dispersal\hspace{0.17em}in\hspace{0.17em}host\right)=n_i{\left(1-{\beta }_{d,i}\right)}^b,$$ [S5]

where b is a parameter that controls the steepness of the curve, set to 6 to induce a strong trade-off. The density of the species that disperse to the media community is given by the following:

$$n_i\left(after\hspace{0.17em}dispersal\hspace{0.17em}in\hspace{0.17em}media\right)=n_i{\beta }_{d,i}.$$ [S6]

Once the event takes place and the new colonists are added to the host community, any species that are at densities lower than the threshold, 0.001, are removed from the community. In the media community, all species are decreased by a factor of 10 (to simulate a death rate in the media) and then any species that do not meet the threshold are eliminated from the community before the new species are added.

We replicate the experimental treatments in the following way: 12 solitary hosts have only a single host per media community. Although all solitary hosts share the same species pool, there is no communication between the simulations. Separated host simulations consist of 12 hosts with the similar δ selected from a Poisson distribution centered on either 0.0 or 0.5 with variance of 0.02. Each host is slightly different to replicate biological background variation, but the difference between the two mean types is much greater than the differences within types. These hosts share the same media community from which they draw their colonists and to which they add their dispersers. Mixed hosts are similar to the separated hosts except that, instead of 12 hosts with offsets all drawn around the same mean, there are 6 hosts of each mean. Again, these 12 hosts all share the same media community. Each simulation was run for 100 loops, and the experiment was repeated 50 times.

Generalist–Specialist Trade-off.

To implement a trade-off between generalists and specialists (depending on the β_d), we used the following modification of Eq. S4:

$$K_j\left({\beta }_{n,i},{\beta }_{d,i}\right)=\frac{\sin \left(2\pi \ast niches\left({\beta }_{n,i}+{\delta }_j\right)\right)\left(1-{\beta }_{d,i}\right)}{2}+\mathrm{0.5.}$$ [S7]

As shown in Fig. S7, this modification leads to high-dispersal species experiencing a “flatter” selective environment within the host; the peaks are lower and the troughs are higher. Good disperser, high β_d species are more generalist because they see less difference between the host types. Bad disperser, low β_d species are able to achieve higher densities within a host that they are well adapted to in their β_n trait.

Fig. S7.

The metacommunity model’s carrying-capacity functions corresponding with two different offsets with the generalist–specialist trade-off implemented (A and B). Solid lines correspond to β_d = 0, dashed lines to β_d = 0.5, and the dotted line to β_d = 0.9.

Parameter Choices.

We simulated the system under a range of parameters for most of the parameter values in the system. During this process, we were looking for parameter values that reproduced the effects seen in the experiment; it is not intended to be a thorough examination of the model’s robustness to parameter choice. Different choices lead to different outcomes.

Immigration from the pool.

There was very little effect of changing ntrickle when ρ = 0.001 because at such low densities they are very unlikely to get chosen to colonize the host and are quickly purged from the community due to their low abundance. When ρ = 0.1, there was a strong homogenizing effect on the β-diversity between the solitary host and the mixed and separated treatments. At very high ρ, the community effects begin to be blurred and the distinct clusters of β_n around the peaks in the carrying-capacity function are lost as the community becomes less structured.

Timing and probability of feeding and dispersal.

When τ was very short, <7, there was strong selection for dispersers probably because the dynamics never get past the initial transients. Dispersers that come in at high densities are still at high densities at the time of the disturbance because they have not yet decreased to their carrying capacities, even if they are poorly suited for that host environment given their β_n. At much larger τ, >20, the dynamics reach their steady state and selection for dispersers is much lower. We chose an intermediate value, τ = 10, to impose high disturbance but allow some competitive dynamics to play out.

We also investigated the relative frequencies of p_f, p_d, and p₀. Increased p_f relative to p_d led to a convergence in α-diversity between the treatments. Greater p_d led to greater α-diversity for the separated and mixed cohoused treatments and decreased α-diversity for the solitary hosts. When p₀ was high, disturbance rarely occurred and α-diversity also converged. We chose equal and high p_f, p_d = 0.49, leading to high rates of disturbance and to a difference in α-diversity qualitatively similar to that seen in the experiment.

Number of new colonists to the hosts.

We set the number of colonists to a level that reproduced the community results, ndisp = 5. When there were many more colonists, ndisp = 15–50, the distribution of β_n within the hosts lost its niche structure and the community started to look like just a random collection of species. In addition, with higher ndisp there was strong selection for high β_d in all hosts, even the ones that were solitary.

Discussion

Dispersal among hosts had a substantial impact on the diversity and composition of zebrafish gut microbiomes in our study and was sufficient to overwhelm the effects of host-specific factors such as innate immune activity. Notably, the strong effects of dispersal in the experiment were specifically the result of “interhost” dispersal, as the overall migration of microorganisms from all other sources, such as from the microbial communities found in the flask water or food, was not directly altered. Furthermore, we only manipulated the potential for dispersal to occur by exposing hosts to one another, while the actual movement of microorganisms among hosts occurred naturally without direct manipulation (such as by gavage, injection, forced feeding, etc.). As a consequence, we do not know the actual rates of dispersal that occurred in this experiment nor how they compare with those found in natural populations of fish or other animals. It is probable that both the strength and nature of the effects of interhost dispersal differ with different rates of dispersal, and deeper investigations of this relationship will help reveal the specific circumstances under which we expect interhost dispersal to be more or less important. It is worth noting, however, that given that the rates of dispersal in this experiment were allowed to occur passively, they are unlikely to be overly unrealistic, and it is therefore reasonable to expect similar effects in many natural systems.

In addition to demonstrating the overall effects of dispersal on microbiome composition and diversity, this experiment has generated unexpected insight into the specific mechanisms by which interhost dispersal influences microbiomes. Contrary to our initial predictions, we observed a homogenization of host microbiomes regardless of whether dispersal was limited to within host genotypes (separated) or allowed to occur between host genotypes (mixed). This homogenization likely occurs because the existence of dispersal among hosts alters the viable trait space for host-associated microorganisms and selects for life history strategies favoring motility and transmission. This interpretation is supported by our genomic predictions that dispersal-related genetic pathways were enriched in the microbiomes of cohoused compared with solitary hosts, and by the ability to reproduce our results in our computational model by imposing a trade-off between dispersal ability and host type specificity. Similar phenomena are predicted in metacommunity models incorporating colonization–competition trade-offs (33) and have been investigated in communities of pathogens (34). This behavior in commensal microbiomes suggests that the consequences of interhost dispersal are likely more complex than simply the homogenization of microbiomes through the sharing of microbial taxa.

Combining experimental studies such as this with surveys of natural systems will undoubtedly help inform both how generalizable experimental results are as well as strengthen inferences about the importance of dispersal based on observed patterns. The importance of dispersal in natural populations of hosts has often been inferred by observed increases in within-host α-diversity and decreases in interhost β-diversity (e.g., ref. 12), or by correlations between microbiome similarity and proxy measurements for dispersal, such as frequency of social interactions (e.g., ref. 10) or geographic distance (e.g., ref. 8). These inferences are frequently based on general predictions from conceptual models of community assembly or even simple intuition. By experimentally testing predictions from a mechanistic model, our study creates a stronger link between such patterns and interhost dispersal as the driving mechanism, justifying such inferences. It also provides further evidence that theories and models developed in other ecosystems may be applied to better understand the assembly and dynamics of host–microbe systems. In particular, we have demonstrated that metacommunity theory provides an appropriate framework for the study of host-associated microbial communities, as has previously been suggested (22, 23). Utilizing and combining these tools will help provide better understanding of how individual microbiomes are influenced by processes occurring at the scale of populations and communities of multiple hosts. Such a holistic understanding will improve our ability to both manipulate microbiomes and predict their responses to changes in host behavior and ecology.

Methods

Zebrafish Husbandry.

All zebrafish experiments were performed using protocols approved by the University of Oregon Institutional Care and Use Committee and followed standard protocols. We generated an immunocompromised myd88 zebrafish mutant using the CRISPR-Cas9 system (35) (Generation and Verification of a myd88 Mutant Zebrafish). WT AB/Tübingen fish with a fully functional immune system and isogenic mutant myd88⁻ zebrafish were then raised in glass Erlenmeyer flasks, such that microorganisms could disperse among hosts in the same flasks, but not between hosts in different flasks. Because there was no practical or noninvasive method to reliably distinguish myd88⁻ from WT zebrafish embryos, the embryos for each genotype were generated from two crosses of homozygous parents. To eliminate potential maternity effects, standard gnotobiotic zebrafish protocols were used to make the embryos “germ-free” (free of microorganisms) before being exposed to a shared inoculum at the beginning of the experiment.

Beginning as germ-free embryos, zebrafish were raised in flasks alone (solitary), or cohoused with 10 total zebrafish of the same genotype (separated), or 5 of each genotype (mixed), per flask (Fig. 1). Initially, 20 fish of each genotype were raised alone in solitary flasks, 40 fish of each genotypes were raised across four replicate separated flasks, and 20 fish of each genotype were raised across four replicate mixed flasks, resulting in a total of 160 zebrafish at the beginning of the study. The volume of embryo media and size of flask was scaled to the number of fish, that is, 50 mL of embryo media in a 125-mL Erlenmeyer flask for solitary conditions and 500 mL of embryo media in a 1-L Erlenmeyer flask for cohoused conditions. Doing so allowed the density of fish to be equivalent across housing conditions. Every day, ∼75–90% of the embryo media in each flask was removed and replaced with fresh, but not sterilized, embryo media. During this time, the majority of food debris and zebrafish feces, as well as any dead fish carcasses, were removed as well. Once the zebrafish fully hatched from their chorions (by 4 dpf), fish were fed live rotifers to a concentration of 20 individuals per mL, followed by the addition of live brine shrimp beginning at 10 dpf once per day.

Sampling and DNA/RNA Extractions.

At 21 dpf, the juvenile zebrafish were killed and dissected to sample their intestinal microbiomes by 16S rRNA gene sequencing, as well as to characterize their innate immune response by qPCR of two genes encoding innate immune cytokines: il-1β and c3. Each individual intestine was aseptically removed and placed in a sterile 2-mL screw cap tube with 200 mL of nuclease-free water while the remainder of the zebrafish carcass was placed in a 2-mL screw cap tube with 1 mL of TRIzol (Life Technologies). Both sample types were then immediately frozen in liquid nitrogen and stored at −80 °C until DNA/RNA extractions were performed. To identify the genotype of mixed cohoused zebrafish, mixed cohoused samples were genotyped by PCR of the MyD88 gene using forward primer, 5′-GTAACGCGGAGATATACAACAAC-3′, and reverse primer, 5′-GAAGCGAACAAAGAAAAGCAA-3′.

DNA was extracted from intestinal samples using the MoBio PowerMicrobiome RNA Isolation kit (product number 26000-50) with the addition of β-mercaptoethanol (product number M3148-25ML; Sigma) using the manufacturer’s suggestions. RNA was extracted from the remaining zebrafish carcasses using a standard laboratory TRIzol extraction protocol (29).

cDNA Conversion and qPCR of Host Innate Immune Genes.

Extracted zebrafish RNA was converted into cDNA using the SuperScript IV reverse transcriptase kit (Invitrogen) following the manufacturer’s instructions. qPCR assays were performed in 20-μL reactions with 20 ng of cDNA, and 400 nM gene-specific or control primers. Gene-specific primers were ordered from Eurofins Genomics with the following sequences: IL-1B forward primer, 5′-CATCAAACCCCAATCCACAG-3′, and reverse primer, 5′-CACCACGTTCACTTCACGCT-3′; C3 forward primer, 5′-CGGACGCTGACATCTACCAA-3′, and reverse primer, 5′-TCCAGGTCTGCTCTCCCAAG-3′. Primers for the housekeeping genes SDHA and ElF-1B (used to normalize the results) were ordered from PrimerDesign. All reactions were performed in triplicate using a Bio-Rad CFX96 Real-Time PCR (qPCR) Thermocycler. Starting concentrations of transcripts were estimated from the resulting amplification curves using the LinRegPCR software (36). Technical replicates were then averaged and divided by the geometric mean of housekeeping genes SDHA and ElF-1B to normalize results.

16S rRNA Gene Sequencing and Processing.

We characterized the microbiomes of individual samples via sequencing of the V4 region of the 16S rRNA gene using 515F and 806R primer sequences (37). We used a single-step PCR to add dual indices and adapter sequences to the V4 region of the bacterial 16S rRNA gene and generate paired-end 150-nt reads on the Illumina HiSeq 2500 platform. The resulting 16S rRNA gene sequences were assembled using FLASH (38) and quality filtered using the FASTX Toolkit (39). Zebrafish host sequences were filtered from the dataset by aligning reads to the zebrafish genome using Bowtie2 (40). OTUs were defined de novo using 97% sequence similarity in the USEARCH pipeline (41). The taxonomy of these OTUs was then assigned using the RDP classifier (42). To infer the genomic content of OTUs, we first matched representative sequences from each de novo OTU to 97% similarity OTUs in the Greengenes 13_8 database (43). We then used the preestimated genomic predictions for the Greengenes OTUs from PICRUSt to infer the genomic content of the OTUs in our study (32). Illumina sequence reads have been deposited under the National Center for Biotechnology Information BioProject accession number PRJNA378677.

Community Analysis and Statistics.

Before analysis, OTU abundance tables were rarefied to 13,700 reads per sample. We measured differences in microbiome composition using the Canberra distance. To assess whether different treatments or host factors had a significant effect on microbiome composition, we performed PerMANOVA with 1,000 random permutations using these distances. To measure the overall variation in microbiome composition within groups, we performed a multivariate homogeneity of group dispersions test. Variance partitioning on microbiome composition by host factors (standard length and il-1β and c3 gene expression) was done by canonical redundancy analysis to measure both the unique and shared contributions of each host factor (44). Identification of differentially abundant taxa was done using a negative binomial distribution model implemented in the DESeq2 package (45), while identification of predicted gene pathways enriched in these taxa was done by Poisson regression followed by a χ² test of significance. Calculation of the Canberra distance, PerMANOVA, Shannon diversity, multivariate dispersions test, and redundancy analysis were all performed in R (46) using the vegan package (47).