Identification of Specialists and Abundance-Occupancy Relationships among Intestinal Bacteria of Aves, Mammalia, and Actinopterygii

Abstract

The coalescence of next-generation DNA sequencing methods, ecological perspectives, and bioinformatics analysis tools is rapidly advancing our understanding of the evolution and function of vertebrate-associated bacterial communities. Delineation of host-microbe associations has applied benefits ranging from clinical treatments to protecting our natural waters. Microbial communities follow some broad-scale patterns observed for macroorganisms, but it remains unclear how the specialization of intestinal vertebrate-associated communities to a particular host environment influences broad-scale patterns in microbial abundance and distribution. We analyzed the V6 region of 16S rRNA genes amplified from 106 fecal samples spanning Aves, Mammalia, and Actinopterygii (ray-finned fish). We investigated the interspecific abundance-occupancy relationship, where widespread taxa tend to be more abundant than narrowly distributed taxa, among operational taxonomic units (OTUs) within and among host species. In a separate analysis, we identified specialist OTUs that were highly abundant in a single host and rare in all other hosts by using a multinomial model without excluding undersampled OTUs a priori. We show that intestinal microbes in humans and other vertebrates display abundance-occupancy relationships, but because intestinal host-associated communities have undergone intense specialization, this trend is violated by a disproportionately large number of specialist taxa. Although it is difficult to distinguish the effects of dispersal limitations, host selection, historical contingency, and stochastic processes on community assembly, results suggest that intestinal bacteria can be shared among diverse hosts in ways that resemble the distribution of “free-living” bacteria in the extraintestinal environment.

INTRODUCTION

Because the structure and composition of intestinal host-associated communities (microbiota) have both beneficial and detrimental effects on the physiology of animals, especially vertebrates (1), the factors that shape these communities have been the subject of intense research (2–6). Microbes help maintain homeostasis by exchanging signals with mammalian immune, circulatory, digestive, and neuroendocrine organ systems (7, 8), and although such specific interactions have been investigated in only a few model species, they are thought to play important roles in the physiologies of most vertebrates. In turn, host type, immune system, diet (9), age (5, 10), associations with cohabitants (11), and other factors shape the microbial community structure in a host-specific fashion (12). The discovery of such intimate associations between host and microbe led to the indirect treatment of infected individuals through direct manipulation of their microbiota (6, 13). Effective management of microbial communities to improve health requires a better understanding of ecological factors that control the community assembly of microbes (14).

Ecological drift and intense selection pressure from the host are thought to drive the divergence and codiversification of some intestinal microbes into host-associated assemblages. Physiological and genomic evidence from some bacteria suggest host specialization, where fine-tuning to the host environment often results in increased fitness and abundance (15, 16). In contrast, it is possible that some bacteria may generalize in a wide range of hosts by utilizing a large range of substrates or by processing them more efficiently. This classic trade-off between two lifestyles has been investigated and discussed thoroughly for “large” organisms and some microbial communities but less so for intestinal host-associated microbial communities. Additionally, identification and targeting of specialists outside the host in environmental samples, such as recreational water bodies or drinking water supplies, can help identify dangerous sources of fecal pollution by serving as host-specific indicators.

The distribution patterns for intestinal microbes have been less studied than those for microbes in other habitats (10). Taxon-area and distance-decay relationships for microbes have been observed in salt marsh sediments (17, 18) and tree holes (19, 20). The interspecific positive relationship between abundance and occupancy (also called the abundance-range or abundance-distribution relationship), whereby more-abundant taxa also tend to be more widespread, has been observed for microbes in aquatic habitats (21–23) and soils (24–26). In such environments where this relationship holds true, there may be a lack of a fitness trade-off between specialist and generalist lifestyles. Given the potential for intestinal microbes to be somewhat dispersal limited and under intense selection pressure from the host, it is unclear if this fitness trade-off exists for microbes in the gut environment.

In this study, we analyzed sequences from the bacterial V6 hypervariable region of the 16S rRNA genes from 106 animal fecal samples to investigate bacterial distributions within microbial communities sampled from members of Aves (birds), Mammalia (mammals), and Actinopterygii (ray-finned fish). We used a multinomial species classification method to identify locally abundant specialist bacterial taxa without discarding operational taxonomic unit (OTU) observations through data normalization. We also characterized the relationship between OTU abundances and distributions between hosts of similar and different species.

MATERIALS AND METHODS

Sample collection.

The overall data set was generated by combining data from newly collected fecal samples and from previously reported data sets obtained using similar methods (Table 1). Common names of host species are used throughout the manuscript for brevity. Fresh fecal samples were collected aseptically using sterile gloves, sterile disposable spatulas, and sterile 50-ml conical tubes. Fish gut contents were removed surgically and stored in 1.7-ml microcentrifuge tubes. All fresh samples were stored on ice immediately after sampling and at −80°C upon arrival to the U.S. Environmental Protection Agency laboratory in Cincinnati, OH.

TABLE 1.

Fecal samples used in this study

Common name of organism	Scientific name of organism	Location(s)	No. of samples	Reference(s)
Chicken	Gallus gallus domesticus	Georgia	5	This study
Cow	Bos taurus	Colorado, Georgia, Ohio, Nebraska	30	4
Deer	Odocoileus virginianus	Wyoming	4	This study
Dog	Canis lupus familiaris	Ohio	4	This study
Duck	Anas platyrhynchos	Ohio	3	This study
Goose	Branta canadensis	Ohio	3	This study
Gull	Larus delawarensis	Wisconsin	4	This study
Horse	Equus ferus caballus	Georgia	5	This study
Human	Homo sapiens	—a	36	2, 35
Perch	Perca flavescens	Wisconsin	4	This study
Swine	Sus scrofa domesticus	Georgia	5	This study
Trout	Oncorhynchus mykiss	Wisconsin	3	This study

^aSample origin was not reported.

DNA extraction, quantification, and sequencing.

All DNA extractions were performed with the FastDNA kit (Q-Biogene, Carlsbad, CA) according to the manufacturer's instructions, as previously described (4). Prior to fecal DNA extraction, GITC buffer (5 M guanidine isothiocyanate, 100 mM EDTA [pH 8.0], 0.5% Sarkosyl) was mixed with ∼1 g (wet weight) of fecal material to create a fecal slurry. Eight hundred microliters of each fecal slurry was bead homogenized at 4.0 m/s for 30 s by using an MP FastPrep-24 instrument (MP Biomedicals, LLC, Solon, OH). DNA was eluted in 100 μl elution buffer and stored at −20°C until further analysis. DNA yield and quality were ascertained using a NanoDrop 2000 instrument (Thermo Scientific, Wilmington, DE). DNA extracts (5 to 25 ng per reaction mixture) were amplified by using previously reported primer sets and conditions (27), which are described in detail elsewhere (http://vamps.mbl.edu/resources/faq.php#tags). Pyrosequencing was performed as described previously (4).

Sequence data preprocessing, preclustering, clustering, and taxonomic assignment.

Quality filtering and trimming of sequences were performed by using the Visualization and Analysis of Microbial Population Structures (VAMPS) interface (http://vamps.mbl.edu) (28), as reported previously (29). Sequence data were binned according to their barcodes, trimmed, and preclustered to minimize the impact of sequencing errors on sample richness. Data were then downloaded from VAMPS. Both chimera removal and read clustering (97% similarity threshold) were performed simultaneously with the USEARCH function cluster_otus (30). Taxonomic assignment was performed by using the RDP Classifier (31), using the Silva v111 RefNR reference alignment (32). Sequence data are stored on VAMPS and can be viewed without special permission by using the generic guest log-in.

Construction of community data matrices and diversity estimates.

OTU distributions were analyzed using the vegan (33) and pvclust (34) packages as well as custom scripts in R and Python. From the original data set, three smaller community data matrices were constructed and formed the basis for all analyses. A human-only data set (859,510 reads) was created by combining data from 33 human samples from the obese/lean data set (35) and from additional samples taken at initial time points from a previous study by Dethlefsen and colleagues (2). The cattle-only data set (629,299 reads) was composed of data from 30 samples (10 from cattle in Colorado fed processed grain [data sets CO1 and CO2], 5 from cattle in Ohio fed unprocessed grain [data set DK], 5 from cattle in Georgia fed forage [USDA], and 10 from cattle in Nebraska [5 fed forage {data set NE1} and 5 fed unprocessed grain {data set NE2}]) (4). The vertebrate data set (75 samples; 1,775,995 reads) was composed of 12 samples from the obese/lean data set, the three initial time points in the study by Dethlefsen and colleagues, the cattle data set (except CO1 and CO2), and all newly sequenced samples listed in Table 1. After removal of OTUs observed only once (“singletons”), each data set was then subsampled randomly without replacement to the minimum sample library size (4,986, 8,017, and 14,844 for vertebrate, human, and cattle data sets, respectively) for 100 iterations to construct community matrices. This normalization procedure resulted in OTU counts of <1 in many cases because some OTUs were not represented in all subsamples taken at each iteration. Normalized counts were used only for abundance-occupancy relationship investigations. Nonnormalized data sets were used to identify specialists with CLAM tests. Pooled diversities and their standard errors were estimated by using the function vegan::estimateR.

Community clustering and analysis of variance.

Square root transformations followed by Wisconsin transformation (see ?vegan::decostand) on the vertebrate data set were used to generate Canberra distances for nonmetric multidimensional scaling (NMDS) using the vegdist function. Canberra distances have performed well on data sets whose OTUs may be arranged in clusters as opposed to gradients (36). Unweighted pair group method with arithmetic mean (UPGMA) clustering was used to group samples and create a dendrogram according to their community similarity by using the function pvclust::pvclust. The dendrogram was pruned to the total number of host groups in the data set (n = 13) before plotting to remove edges linking different host groups. Variation in microbial communities attributable to host taxonomy was estimated by using permutational multivariate analysis of variance using distance matrices (ADONIS).

Abundance-occupancy relationship.

Occupancy was defined as the proportion of all host groups (n = 13) in which the OTU was observed. Additionally, within-species occupancy was calculated for human and cattle data sets by calculating the proportion of samples in which an OTU was observed within like species. All abundance measures were estimated after summing counts for each OTU or phylum within each host species group. Two measures of abundance were used: maximum abundance, the highest count observed within any host species group for each OTU, and local mean, the average OTU counts for host species in which the OTU was observed (i.e., unoccupied species were omitted). Additionally, within-species abundance measures were calculated analogously to within-species occupancy (described above), whereby the maximum abundance and local mean abundance were estimated by using OTU distributions within samples instead of host species. The relationship between abundance and occupancy was investigated with both nonparametric (Loess) and parametric (simple linear regression) methods.

One concern was that the tendency of abundant OTUs to be more easily detected could inflate their observed occupancies relative to rare OTUs, resulting in what would appear to be an abundance-occupancy relationship but would actually be an effect of ascertainment bias or insufficient sampling of rare taxa. To assess the effects of ascertainment bias on the abundance-occupancy relationship in the human data set, the average increase in observed occupancy was estimated over a series of random subsampling depths ranging from 500 to 7,500 observations.

Multinomial species classification (CLAM test).

Nonnormalized community data were used to identify specialists using CLAM tests (37) by successive pairwise group comparisons, resulting in the identification of specialist OTUs for each host species group. As part of the CLAM tests, sample coverage correction based on the number of observed singletons was applied to rare OTUs whose counts were below 10 sequences per group. Relative OTU abundances were used above this threshold. After these corrections, OTUs were classified as specialists if ≥90% of their occurrences were within a specified group. A 0.05 significance cutoff was used for individual tests. OTUs classified as generalists, “too rare” to classify, or specialists outside the group of interest were disregarded in this analysis.

RESULTS

General data set description.

General descriptions of human and cattle data sets were reported previously (2, 4, 35). Normalization and exclusion of singletons for abundance-occupancy analysis resulted in the discarding of 52.9%, 56.6%, and 51.8% of vertebrate, human, and cattle OTUs, respectively. Diversity estimates indicate that we observed roughly half of all OTUs present in samples (Table 2). Percentages of observed OTUs were lowest for gull samples (30.4%) and highest for horse samples (52.5%).

TABLE 2.

Pooled diversities for the vertebrate data set

Organisma	No. of reads	No. of OTUs	Chao	Chao SE	% of OTUs observed
Chicken	44,863	1,829	5,332	311	34.3
Cow_F	208,473	11,358	25,100	508	45.3
Cow_UG	210,593	7,465	18,817	493	39.7
Deer	79,540	5,448	13,066	351	41.7
Dog	117,715	2,615	6,896	302	37.9
Duck	60,795	3,128	9,486	412	33.0
Goose	58,483	3,740	10,298	385	36.3
Gull	161,585	3,859	12,712	532	30.4
Horse	145,594	10,100	19,251	339	52.5
Human	372,632	8,516	24,921	660	34.2
Perch	148,380	2,184	6,249	324	35.0
Swine	106,881	6,443	13,120	307	49.1
Trout	60,461	993	3,101	258	32.0

^a“Cow_F” and “Cow_UG” represent cattle fed forage and unprocessed grain, respectively.

Community taxonomic composition.

The taxonomic compositions of host communities differed greatly between species; however, all species contained a large proportion of bacteria belonging to the Firmicutes phylum (Fig. 1). Firmicutes composed the largest portion of bacterial communities in birds and mammals, while Proteobacteria composed the largest portion in fish. Bacteroidetes composed the sixth, second, and third most abundant groups in birds, mammals, and fish, respectively. Deer, dog, and horse bacterial communities stood out among all mammals, mainly because of their unusually large proportions of Proteobacteria, Fusobacteria, and Lentisphaerae, respectively.

FIG 1.

Phylum-level taxonomic composition of host-associated intestinal microbial communities. The total height of each stacked bar corresponds to all reads from a sample, while shorter, color-coded bars correspond to the proportion of those reads falling within major bacterial phyla. “Cow_UG” and “Cow_F” indicate cattle fed unprocessed grain and forage, respectively.

Microbiota dissimilarity.

NMDS plots arranged animal groups into distinct clusters that agreed well with hierarchical clustering (Fig. 2). Samples of domestic or agricultural origin clustered by host type well, while there was less agreement among bacterial communities sampled from wildlife. Despite cattle being the same species, diet largely determined cattle sample clustering. ADONIS indicated that host taxonomic species and class were able to explain only 29.5% and 5.6% (P < 0.001) of the variation in vertebrate microbial communities, respectively.

FIG 2.

(A) NMDS plot of samples based on microbial community profiles (stress = 0.158). Samples connected by lines were collected from the same population. (B) UPGMA host species group clustering overlaid onto the same NMDS ordination. Ordination was solved using all OTUs in the vertebrate data set. The UPGMA tree was pruned to the number of host species groups. “Cow_UG” and “Cow_F” indicate cattle fed unprocessed grain and forage, respectively. Perch and trout cluster on top of one another.

Abundance-occupancy relationship.

Even after the exclusion of singletons and normalization, the distribution of OTUs was highly skewed toward occupancy in a single host (70.3% of OTUs appeared in a single host type only). Only four OTUs were found in at least one sample from each species: two unclassified Enterobacteriaceae, one Ralstonia OTU, and one Clostridium OTU. One concern was that data normalization obscured the presence of some OTUs by exclusion and artificially decreased the observed occupancy; however, occupancy analysis with nonnormalized counts suggested a similar, highly skewed trend, with the majority of OTUs occupying a single host and only a few widespread OTUs (data not shown). In addition, 100% of the OTUs present in more than one host species group in the nonnormalized data set were also identified in the normalized data set, confirming that normalization did not strongly influence occupancy estimates. Loess curves on plots of abundance versus occupancy suggested that the relationship between maximum OTU abundance and occupancy was positive and linear within a range (Fig. 3). Assuming linearity, regression analysis indicated that the interspecific abundance-occupancy relationship among all OTUs within human, cattle, and other vertebrate data sets in this study was significantly positive (P < 10⁻⁴), with a high degree of error (R² < 0.1). A similar trend was observed when local mean abundance was used as the abundance measure instead of maximum abundance (data not shown).

FIG 3.

Interspecific abundance-occupancy relationship in vertebrate, human, and cattle microbial communities. Within-species occupancy is represented on the x axis for human and cattle data sets, while occupancy is represented on the x axis for the vertebrate data set. Loess curves were estimated by using all data from each data set, while regression lines (“Linear mod”) were estimated after the exclusion of single-occupancy OTUs. Blue shading represents the two-dimensional kernel density of the data. Artificial variance was added after Loess and regression analyses on x axes for plot clarity.

Of 75 OTUs with proportional occupancies of ≥0.5 (observed in >6 host species), 62.7% were classified as Firmicutes, and 34.7% were classified as Proteobacteria. Fifty of these OTUs were widely shared among the three host taxonomic classes, and all 75 were shared between Aves and Mammalia.

The observed relationship between abundance and occupancy could not be due solely to ascertainment bias. Successive subsampling trials with the human data set showed that for each log₁₀ increase in mean local abundance, there was an average increase in proportional occupancy of 0.5 and a similar increase of 0.3 when maximum abundance was used. In contrast, proportional occupancy increased by only ∼0.02 in these same subsampling trials, suggesting that increasing the depth of sequencing would result in only a small increase in observed occupancy—not enough to explain the observed relationship.

Another concern was that the observed abundance-occupancy relationships could be due to overclustering of distinct ecotypes within the same OTU. To investigate this possibility, we performed an entirely separate analysis on OTUs clustered at a 99% similarity threshold in hopes of minimizing clustering of sequences that may have coevolved in distant hosts. Exclusive clustering resulted in 124,563, 27,020, and 54,854 OTUs for the vertebrate, human, and cattle data sets, respectively. There was no discernible difference in the abundance-occupancy relationships between the two clustering thresholds (data not shown), suggesting that OTU clustering threshold parameters, at least those most commonly used, cannot explain our observations.

CLAM tests.

CLAM tests on 54,666 OTUs resulted in the identification of 10,663 (19.4%) specialist OTUs (Table 3). Taking into account abundance, specialist OTUs accounted for 89.4% of all sequence reads. The taxonomic identities of specialists fell roughly in line with the overall community composition, with the dominant taxa making up a large portion of the specialist population within each host group (Fig. 4). Clustering of OTUs at 99% similarity instead of 97% similarity produced similar types and distributions of specialist OTUs (data not shown). A priori exclusion of OTUs via normalization prior to CLAM tests decreased the total number of OTUs identified as specialists (data not shown), presumably because the majority of OTUs veiled in the normalization process were at low abundance and low occupancy. This presumption was supported by both the low occupancy of a large portion of OTUs and the observation of a positive abundance-occupancy relationship.

TABLE 3.

Number of specialist OTUs within each host group belonging to each bacterial phylum^a

Phylum	No. of specialist OTUs in the indicated host species
	Aves				Mammalia							Actinopterygii
	Chicken	Duck	Goose	Gull	Cow_F	Cow_UG	Deer	Dog	Horse	Human	Swine	Perch	Trout
Acidobacteria	0	44	0	7	9	7	6	8	28	0	15	1	1
Actinobacteria	21	49	183	191	169	30	38	9	111	48	64	14	3
Bacteroidetes	10	22	50	32	193	288	120	23	328	67	140	18	11
Chlamydiae	1	0	2	2	10	0	3	0	5	1	2	0	0
Chloroflexi	1	0	1	10	3	1	3	0	21	1	3	0	0
Chrysiogenetes	0	0	4	0	8	1	3	0	4	1	0	3	1
Deferribacteres	0	2	2	0	6	1	0	1	9	1	1	1	0
Deinococcus-Thermus	0	0	0	4	9	2	0	0	4	0	5	0	0
Firmicutes	169	147	302	324	1,211	375	319	287	1,085	611	484	182	46
Fusobacteria	0	4	1	3	4	1	2	28	9	0	2	2	2
Lentisphaerae	0	0	0	0	4	2	4	0	9	1	2	0	0
Planctomycetes	2	5	0	32	6	4	1	0	38	1	3	1	0
Proteobacteria	54	109	82	264	400	78	132	14	474	63	117	142	38
Spirochaetes	0	1	1	2	20	4	5	1	30	1	14	0	0
Synergistetes	1	1	1	3	7	2	1	0	6	1	0	0	0
Tenericutes	0	1	3	2	49	9	19	1	48	4	26	0	0
Verrucomicrobia	3	13	0	17	10	3	12	4	34	1	4	0	0

^aAquificae, Armatimonadetes, BRC1, Caldiserica, Chlorobi, Elusimicrobia, Fibrobacteres, Gemmatimonadetes, Nitrospira, OP11, Thermosulfobacteria, Thermotogae, TM7, and WS3 were represented by 1 to 20 specialist OTUs and were omitted from the table. “Cow_F” and “Cow_UG” represent cattle fed forage and unprocessed grain, respectively.

FIG 4.

CLAM test results for Firmicutes (A) and Bacteroidetes (B) overlaid onto NMDS ordination (same as Fig. 2). OTU bubbles are placed according to their abundance-weighted average ordination scores, which reflect the degree of association with particular host species. The five families containing the most specialist OTUs within each phylum are displayed. Numerical values in the bottom left key represent the total number of specialist OTUs within each family. The diameter of each bubble is proportional to relative abundance as a proportion of all summed OTU counts within each host group.

DISCUSSION

The rules governing the distribution of microbes have long been debated (38), and while there are trends shared not only between bacterial communities from different habitats but also between macro- and microorganisms, the factors structuring these communities likely differ (39). Microbial communities native to the guts of animals are a special case because they may be strongly influenced by the ecology and evolution of their hosts. The degree to which microbes invest in particular host-specific lifestyles can be studied by asking how they fit well-studied macroecological patterns, if at all. Our work shows that intestinal bacteria present both within and among vertebrate host species follow similar abundance-occupancy relationships, which we cannot precisely explain in light of previous explanations, but because the increase in occupancy as a function of abundance far outweighs that attributable to sampling depth, it is unlikely that the relationship is due solely to ascertainment bias. We also found that OTU clustering parameters had little effect on the abundance-occupancy relationships. Similar relationships were observed previously for the human microbiome by comparison of rank-abundance and rank-prevalence (40). Although it is difficult to compare such relationships based on the proportions of host species or individuals occupied to those based on ranges (e.g., latitudes or distances), the observation that similar patterns occur reinforces the idea that host-associated intestinal microbial communities may operate under a set of principles similar to those of free-living communities to a degree (41).

Although some intestinal bacteria have developed mechanisms for survival under oxic and oligotrophic or otherwise harsh conditions, many are not fit for such conditions, limiting survival outside the host to a matter of days (42) and restricting recolonization in distant suitable habitats (i.e., dispersal limitation). Isolation contributes to community dissimilarity through ecological drift (43). Selective pressure, most of which is mediated by the host immune system or other factors, such as diet and outcompetition from highly specialized community members, can restrict successful colonization of the gut from outside members and further contribute to the isolation and divergence of these host-associated communities. The narrow range of abundant specialists suggests that host selection and drift through ecological isolation may have caused a significant portion of intestinal bacteria to deviate from the nearly universal abundance-occupancy relationship. In contrast to previous studies that show clear abundance-occupancy relationships in large organisms, which suggest a lack of a fitness trade-off between generalist and specialist niches, these results confirm that bacteria can benefit significantly by acquiring specialist lifestyles.

Because the CLAM test is relatively robust to biases caused by different sampling depths between samples and stochastic sampling of rare taxa (37), we were able to identify thousands of specialist OTUs without the exclusion of a large amount of data a priori through normalization by subsampling. Typically, host-associated specialist taxa are identified through comparative 16S rRNA gene sequence analysis or enrichment methods (44–47), followed by testing for their presence in other sources with more sensitive methods, such as PCR, which can take years to complete (48–50). Although our methods, like most, cannot confirm the absence of specific taxa, the comparison of OTUs between multiple host-associated communities simultaneously resulted in the identification of both previously identified and potentially new specialist groups in a single step. Independent studies have also identified members of the Enterococcaceae that dominate Larus spp. (gulls) sampled over a wide geographic range but are not found in other species at significant concentrations (50–52). The relative abundance of Lachnospiraceae in mammalian guts was noted previously, and genomic analysis suggests that this group's abilities to form endospores; produce butyrate, a compound thought to be important in host physiology; and carry genes important for protein interactions and signal transduction play prominently in this group's ability to evolve host-specific preferences (53). Similar mechanisms likely exist for other specialist taxa identified in this study. Erysipelotrichaceae, Porphyromonadaceae, and Spirochaetaceae may represent previously unidentified canine, porcine, and equine specialist groups, respectively. In dogs, the abundance of Erysipelotrichaceae drops significantly in diseased states, while no significant change occurs for most other bacterial taxa (54), which suggests that canine-associated OTUs within this environment may have specialized not only to the canine gut environment but also to a healthy-host state within this environment. Such taxonomic groups identified by the CLAM test may represent potential host-associated targets for PCR- or sequencing-based (55) fecal pollution identification methods, and further investigation into their distribution and growth/persistence in the environment is warranted.

There are many considerations and caveats when interpreting CLAM test results. Pretreatment of the input data (e.g., normalization) and alternate user-defined values (e.g., statistical significance threshold) changed the number of specialist bacteria identified within each host. Normalization leads to a larger proportion of taxa classified as too rare to classify (37). The range, type, and physiological states of the host species sampled and their grouping by the analyst (e.g., regarding or disregarding diet regimes) also influence the identification of specialist taxa. Future studies should be directed at describing this variation among hosts to an extent that we could not achieve with such small sample sizes. These methods do not distinguish specialists that have a high abundance within a small proportion of hosts from lower-abundance specialists found in a large proportion of hosts. Such information may be useful when trying to distinguish dispensable from essential community members or to determine the degree of association between two organisms (56).

While such tests help prioritize bacterial groups for future study, a deeper understanding of the ecological and physiological roles that contribute to patterns of abundance and occupancy is needed to fully understand the extent of host-microbe relationships and to test the widespread assumption that the most abundant bacteria also play the most important physiological roles within the host. Functional metagenomic analysis may provide a more accurate picture of the overall community metabolic capability, while single-cell isolation and genome sequencing techniques may be more useful in linking functional capacity to 16S rRNA data such as those produced in this study. A more detailed comparison of host genetics, perhaps through comparison of mitochondrial genomes or whole nuclear genomes, may provide a host phylogenetic “landscape” on which to study the effects of other environmental factors, such as host diet, habitat, or interpopulation social interactions, on microbial communities.

Funding Statement

The U.S. Environmental Protection Agency through its Office of Research and Development funded and managed the research described here.