Similarity and differentiation between bacteria associated with skin of salamanders (Plethodon jordani) and free-living assemblages


  • Benjamin M. Fitzpatrick,

    Corresponding author
    1. Ecology and Evolutionary Biology, University of Tennessee, Knoxville, TN, USA
    • Correspondence: Benjamin M. Fitzpatrick, Ecology and Evolutionary Biology, University of Tennessee, Knoxville, TN 37996, USA.

      Tel.: (865) 974 9734; fax: (865) 974 3067;


    Search for more papers by this author
  • Amanda L. Allison

    1. Ecology and Evolutionary Biology, University of Tennessee, Knoxville, TN, USA
    Search for more papers by this author


All animals and plants have intimate associations with microbes. Opinion has shifted from viewing microbes primarily as pathogens to the idea that healthy animals and plants carry specialized communities of coevolving microorganisms. However, the generality of this proposition is unknown because surveys rarely compare host-associated microbes with samples from relevant microhabitats. Symbiotic communities might be assembled from local environments with little evolutionary specialization. We evaluated the specificity of bacteria associated with salamander skin in comparison with surfaces in their immediate environments using 16S rRNA sequences. Host-associated and free-living samples were significantly different. However, relative abundances were strongly correlated; the most abundant taxa on salamander skin were also most abundant on moist debris on the forest floor. Thus, although bacterial assemblages on salamander skin are statistically differentiated from those on inanimate surfaces, they are not entirely ‘distinct’. Candidate salamander specialists were few in number and occurred at low relative abundances. Within some OTUs, differences in allele frequency suggested genetic specialization at finer levels. Although host-associated and free-living assemblages were similar, a range of more or less specialized symbiotes was evident and bacteria on salamander skin were often specific genotypes of OTUs commonly found on other moist surfaces in the environment.


Microbial symbiosis is widespread, if not universal, among multicellular organisms, suggesting that host–symbiont interactions are among the most important drivers of ecology and evolution (Dale & Moran, 2006; Douglas, 2008; Sachs et al., 2011). Symbiosis encompasses any intimate association between different kinds of organisms from parasitism to mutualism (De Bary, 1879; Walter et al., 2011). Historically, microbiology focused largely on parasite and pathogen interactions, but recent medical, ecological, and evolutionary research emphasizes the importance of mutualisms between bacteria and eukaryotes (Dethlefsen et al., 2007; Fraune & Bosch, 2010). An increasingly popular view is that healthy plants and animals carry intimate, specialized communities of coevolving microorganisms (Rosenberg et al., 2007; Turnbaugh et al., 2007; Gilbert et al., 2010; Bailey et al., 2012; Ogilvie et al., 2012) and might even be reconceived as communal metaorganisms (Dupré & O'Malley, 2007; Rosenberg et al., 2007; Doolittle & Zhaxybayeva, 2010). This idea is reminiscent of Clements' community-as-organism analogy (Clements, 1916), which has been largely rejected in mainstream evolutionary ecology (Callaway, 1997; McIntosh, 1998; Begon et al., 2006). Aside from a few well-studied pairwise interactions, almost nothing is known about the consistency and specificity of microbial symbioses (Dale & Moran, 2006; Bright & Bulgheresi, 2010; Feldhaar, 2011; Mueller, 2012), making it difficult to determine where host-associated microbial communities fall on the continuum between super-organism and casual assemblage.

We examined consistency and specificity of host-associated vs. free-living bacterial communities by comparing samples from salamander skin to environmental samples from the nearest moist surfaces – the rocks and logs under which salamanders were found. Microorganisms associated with skin seem likely to fall on a wider range of levels of specialization in comparison with gut microbes because there is more consistent opportunity for exchange between host skin and environment, and because the gut might be a more distinct environment. On the other hand, gut microbial communities can be rapidly and profoundly altered by diet (Costello et al., 2010; Sharon et al., 2010; Muegge et al., 2011), and anaerobic conditions (typical of the gut) are not uncommon in moist soils (Inglett et al., 2005). Animal skin, as a microbial habitat, might be very consistent relative to both the gut and the external environment. Moreover, amphibian skin in particular might be a chemically and structurally specialized habitat.

Salamanders and other amphibians maintain a mucous layer that helps resist desiccation and often contains chemicals that are distasteful or toxic to predators and pathogens (Brodie & Howard, 1973; Brodie et al., 1979; Clarke, 1997; Petranka, 1998; Dodd, 2004). There is some evidence that bacteria in the mucous layer can help prevent fungal infections (Becker & Harris, 2010), and some researchers are working to develop probiotic treatments against the widespread pathogen Batrachochytrium dendrobatidis, a chytrid fungus implicated in amphibian decline and extinction (Harris et al., 2006). Thus, understanding the evolution and ecology of amphibian skin communities has both basic and applied significance.

Some previous studies of amphibian skin have inferred the existence of ‘distinct’ resident bacterial assemblages (sensu Price, 1938) without directly comparing host-associated and free-living samples. Electron micrographs show bacterial cell division on salamanders and an apparent tendency of bacteria to aggregate around certain glands (Lauer et al., 2007). Density of cultivable bacteria on toad skin grew fourfold between molts in a laboratory study (Meyer et al., 2012), showing endogenous population growth. In addition, samples of putative transient bacteria (rinsed off in preparation for sampling residents) were statistically different from resident bacteria when compared using DGGE of 16S rRNA gene PCR products (Lauer et al., 2007, 2008), and McKenzie et al. (2011) found consistent differences between skin bacterial samples from co-habiting frog and salamander larvae. However, statistically significant differences in mean relative abundance do not necessarily indicate that host-associated bacteria are unique or specialized (Yoccoz, 1991; Anderson et al., 2000).

Loudon et al. (2014) demonstrated statistical differences in relative abundance of bacteria on salamander skin (Plethodon cinereus) and soil samples from their habitat. But the most common bacteria on salamanders were also found in the soil. Moreover, housing salamanders in soil vs. sterile media resulted in major bacteriological differences over a 28-day period (Loudon et al., 2014). Their results indicate that bacterial assemblages on salamander skin are strongly affected by the external environment, including the reservoir of free-living bacteria.

Here, we undertook an assessment of specialization in salamander skin bacteria at community and population levels in comparison with other moist surfaces on the forest floor. We focus on one common species (P. jordani, the Red-cheeked Salamander) and two kinds of inanimate cover objects (rocks and logs) at three localities in Great Smoky Mountains National Park (North Carolina). Woodland salamanders such as P. jordani tend to be sedentary, with small home ranges and typical dispersal distances on the order of tens of meters (Madison & Shoop, 1970; Merchant, 1972; Nishikawa, 1990). Thus, our sampling crosses distances several orders of magnitude greater than typical salamander movement (Fig. 1). We used high-throughput Illumina (Hi-seq) sequencing of the 16S ribosomal RNA gene as a culture-free assay of bacteria in each sample. We used ordination and diversity partitioning to compare variation between host-associated and free-living samples with variation among individual samples, among localities, and between cover types. We searched for candidate host specialists using indicator values and specialization indices. At the population level, we tested whether host-associated populations of particular OTUs were genetically differentiated from free-living populations. These analyses provide a broad assessment of where salamanders and their symbionts fall on the continuum between highly co-adapted, insular metaorganisms and members of broad, diffuse interaction networks.

Figure 1.

Sampling localities in Great Smoky Mountains National Park (USA) consisted of three transects.



Sampling was conducted under North Carolina Wildlife Resources Commission Collection License Permit 11-SC00542, United States National Park Service Scientific Research and Collecting Permit GRSM-2011-SCI-0062, and University of Tennessee Institutional Animal Care and Use Committee Protocol 2017-0611.

Salamander-cover object paired samples were collected in Great Smoky Mountains National Park, North Carolina, during July 2011 (Fig. 1). Twenty-four sampling areas were located as eight evenly spaced sites along each of three 3.2 km transects: Noland Divide (1570–1800 m), Kephart Prong (850–1080 m), and Mt. Sterling Ridge (1585–1775 m). Up to two salamanders were captured per collection site. Total sample size was 75, including 29 salamanders, 39 logs, and 7 rocks (Table S1).

Bacterial DNA collection and preparation

We captured salamanders by hand and rinsed them with 50 mL of sterile, dechlorinated water to remove transient bacteria. We think this is roughly equivalent to the three successive water baths of 20 mL used by Culp et al. (2007). Rinsing is standard practice (Culp et al., 2007; Lauer et al., 2007; McKenzie et al., 2011; Loudon et al., 2014), but raises unresolved questions regarding what is being sampled. Rinsing is a selective procedure (Lauer et al., 2007) that might bias inference if the goal is to assess what microbes are normally in contact with skin. Following earlier studies, we assume rinsing removes bacterial cells that are not truly resident on the skin. Variation in the thoroughness of rinsing might affect how similar a sampled assemblage is to environmental samples.

After rinsing, each salamander was handled only through a fresh, sterile plastic bag. We swabbed the trunk of each salamander with two sterile cotton swabs (Lauer et al., 2007). The cover object, the bottom face of either a rock or log where the salamander was captured, was also swabbed with two sterile cotton swabs for collection of environmental bacteria. Salamanders were then released in the site of capture. At each transect, we swabbed 5–10 cover objects without P. jordani and performed two ‘dummy’ swabs of fresh plastic handling bags as controls for contamination. All swabs were frozen at −20 °C until DNA extraction using Qiagen QIAamp DNA Micro Kits (Germantown, MD). Negative controls never contained detectable DNA after extraction or PCR.

DNA samples were prepared for Illumina sequencing following Caporaso et al. (2011). PCR primers (F515/R806) were used to amplify region V4 of the prokaryote 16S ribosomal RNA. To identify samples after sequencing, 12 bp barcodes (Caporaso et al., 2011) were added to the 5′ end of primer F515 (Table S2). PCRs were carried out in triplicate 24-μL reaction mixtures containing 2.0 μL of each primer, 1.0 μL template DNA, 1.5 mM MgCl2, 0.2 mM dNTPs, 0.25 μL GoTaq DNA Polymerase (5.0 U μL−1), 5.0 μL 5x reaction buffer, and 10 μL water. Thermocycling conditions were as follows (Caporaso et al., 2011): 94 °C for 3 min, 35 cycles of 94 °C (45 s), 50 °C (30 s), 72 °C (90 s), and a final elongation at 72 °C for 10 min. Triplicate PCRs were pooled and cleaned using Qiagen PCR purification kits. A NanoDrop Micro-Volume UV-Vis Spectrophotometer was used to determine DNA concentration. All amplicons were pooled in equalized DNA concentrations and sent to the Yale Center for Genome Analysis (YCGA; West Haven, CT) for paired-end sequencing on a Hi-Seq 2000.


Sequences were processed and run through quality control using qiime (Caporaso et al., 2010). Reads were truncated after two consecutive low-quality base calls and removed from analysis if they contained more than one ambiguous call. After parsing both forward and reverse reads by barcode, reverse reads were reverse-complimented and concatenated with the corresponding forward read with a gap to mark the joint. After removing barcodes, concatenated reads contained up to 90 bp from the 5′ end of the PCR product and 101 bp from the 3′ end (191 bp of data with c. 100 bp unread gap). We eliminated reads with < 75 bp from the dataset. To eliminate PCR chimeras and other artifacts, we first clustered reads using uclust (Edgar, 2010) and kept only those reads within 94% similarity of a sequence in the Greengenes reference set (DeSantis et al., 2006). We performed de novo clustering of this reduced dataset to define operational taxonomic units (OTUs) with 97% sequence similarity. Very rare OTUs have a high probability of being artifacts of sequencing errors (Caporaso et al., 2011; Werner et al., 2012); therefore, we included in our analyses only OTUs detected at least ten times in at least three samples. Each OTU was assigned to a taxonomic group using the RDP classifier (Liu et al., 2008). More restrictive criteria (e.g. eliminating all OTUs with fewer than 1000 reads total) made no difference in community-wide inferences, but necessarily reduced the number of possible indicator species. Sequence data are deposited at NCBI's sequence read archive under Accession PRJNA238561. Sample metadata and read counts used in downstream analyses are available as Supporting Information.

Diversity partitioning

Diversity of bacterial OTUs within samples was represented as Hill numbers (also called ‘numbers equivalents’), which give the ‘effective number of species’ indicated by traditional diversity indices (MacArthur, 1965; Jost, 2006). Hill numbers are indexed by q, the ‘order’ of diversity. Species richness is the Hill number of order q = 0. Order q = 1 gives the Hill number for the Shannon index (i.e., the number of equally abundant species required to give the same Shannon entropy as estimated from the sample). Order = 2 gives the Hill number for the Simpson index. The difference between orders is a function of evenness – assemblages dominated by a few highly abundant taxa will show a steep decline in Hill numbers with increasing q (Jost, 2006). Community ecologists largely agree that Hill numbers, rather than traditional indices, are the appropriate expressions to compare diversity among samples (Ellison, 2010; Chao et al., 2012).

We estimated the contribution of the host-associated vs. free-living dichotomy to the overall diversity of bacteria in the sample using hierarchical partitioning of diversity (Chao et al., 2012). This extends the concepts of alpha, beta, and gamma diversity to multiple levels. We compared two partitioning schemes. One considered transect as the highest level of grouping, with sample type (salamander, rock, log) nested within transect. The alternative considered sample type the highest level, with transect nested within sample type.

When using Hill numbers, multiplicative partitioning of diversity is generally recommended (Jost, 2007; Chao et al., 2012). Under multiplicative partitioning, beta diversity is the effective number of completely distinct assemblages, that is, the number of assemblages sharing no species that would yield the same total diversity (gamma) as observed. We also present additive partitioning of traditional diversity indices (richness, Shannon, and Simpson) owing to the intuitive relationship between additive beta diversity and the fraction of total diversity accounted for at a given level (Crist et al., 2003; Chao et al., 2012). Partitioning was performed using functions multipart and adipart in vegan (Oksanen et al., 2011).

Community similarity between host-associated and free-living samples

Our paired sampling scheme provides a direct and powerful way to determine whether samples of bacteria from salamander skin are different from samples from their immediate environment. For each salamander-cover object pair, we performed a randomization test (Solow, 1993; Manly, 1997) with 10 000 permutations to obtain a P-value for the data given the null hypothesis that the two sets of sequence reads were sampled from the same distribution, accounting for differences in sample size (i.e., assuming any differences in the total number of reads were due to sampling, not biology). This was obtained as a randomization (function chisq.test in r) of 2 × S contingency tables (Hope, 1968), where S is the number of OTUs in at least one of the samples. With 29 tests, the Bonferroni critical value 0.05/29 = 0.00172.

To evaluate how differences between host-associated and free-living samples compare to differences among hosts and among cover objects, we estimated indices of community overlap based on presence–absence (Sorensen index; = 0), Shannon diversity (Horn index; = 1), and probability of identity (Morisita–Horn index; = 2) using the r package Vegetarian (Charney & Record, 2009). These measures of community similarity between samples provide an assessment of the uniqueness of host-associated communities (Jost, 2007; Jost et al., 2011). This is important because the high resolving power of the DNA sequencing assay might expose statistically significant but ecologically subtle patterns. We visualized patterns of community similarity using unconstrained ordination (Borcard et al., 2011). For statistical testing, we used variance partitioning of distance-based redundancy analysis (function varpart), which allows transect and sample type to be treated as crossed factors, and also allows inclusion of spatial and elevational distances as covariates, thus properly using all of the information in our sampling design to partition the among-sample variation (Borcard et al., 1992, 2011).

Identification of specialists

To identify bacterial OTUs with high affinity for salamanders vs. cover objects, we used indicator species analysis (De Cáceres & Legendre, 2009; De Cáceres et al., 2010) with habitat types categorized as logs, rocks, or P. jordani. We estimated group-equalized correlation coefficients, rg, for each category. This point-biserial correlation between relative abundance and habitat reflects both specificity and consistency. For example, an OTU found only on one habitat type (high specificity) but often absent from patches of that habitat type (low consistency) can have a lower rg than an OTU that consistently occurs at high abundance on a particular habitat type, even if it sometimes occurs on other habitat types (lower specificity). Estimates and permutation tests were calculated using the r package indicspecies (De Cáceres & Legendre, 2009).

To provide an estimate of specialization with greater emphasis on specificity, we calculated the paired differences index (PDI) using the r package bipartite (Blüthgen et al., 2006). This index estimates the average difference between the proportion of occurrences of a given OTU on its favored habitat and the proportion of occurrences on other habitats. PDI is unaffected by absences or by variation in overall abundance between species (i.e. PDI is the same if 90 of 100 occurrences of a given species are on one habitat, or if 900 of 1000 occurrences of a given species are on one habitat) and is highly informative in discriminating degrees of specialization and generalization (Blüthgen et al., 2006).

Ecological differentiation within OTUs

To test for genetic differentiation between host-associated and free-living populations within each OTU, we separated the sequences mapping to each OTU and performed de novo clustering at 99% identity threshold. That is, sequences within 99% similarity of each other were considered the same allele to reduce inclusion of false alleles created by PCR and/or sequencing error. The possibility of false alleles was further reduced by eliminating very rare alleles (those not detected at least 10 times in at least three samples). We accept that some true alleles might have been ignored and some false alleles included. We assume these errors were unbiased with respect to habitat type, introducing noise and therefore reducing any signature of systematic differentiation. With reads of 75–191 bp, we certainly will have missed genetic variants that differ only in other parts of the gene (or genome, for that matter). We recognize that our estimates of differentiation represent the region sequenced, and likely underestimate genome-wide differentiation.

We used the allele frequency data to estimate hierarchical F-statistics (Weir, 1996; Goudet, 2005) and test for genetic differentiation between host-associated and free-living populations, accounting for differentiation among populations of the same type and for spatial differentiation (among sampling transects). This analysis evaluates whether different genotypes tend to segregate by habitat (host vs. cover object).


Taxonomic diversity and rarefaction

Our final dataset included 1 965 175 reads assigned to 2316 OTUs (defined by 97% similarity). Reads were classified into 24 classes, predominantly Gammaproteobacteria. Our analyses are based on OTUs clustered at the 97% similarity level, roughly corresponding to ‘species level’ clades (Konstantinidis et al., 2006). We summarize the composition of higher taxa in the supplemental material (Fig. S1) and the most commonly identified genera in Fig. 2 (representing 29% of the total reads in the final dataset). Genus designations were based on RDP classification of the 97% similarity-based OTUs (Liu et al., 2008).

Figure 2.

Heatmap of relative abundance of genera (based on RDP classification of 97% similarity OTUs) in samples from salamander skin and inanimate cover object surfaces. Numbers are average relative abundances by habitat. Shading is roughly inversely proportional to loge (relative abundance), except for the highest bin. Bold values highlight the extremely high abundance of Nevskia on salamanders and rocks. The most abundant 25 (of 74) genera are shown. Groups without a genus name could not be confidently assigned to a known genus using the RDP method. Genera are sorted (top to bottom) from highest to lowest abundance on salamanders.

The number of reads per sample ranged from 1819 to 106 509. The majority of samples had between 11 000 and 39 000 (24% and 76% quantiles), and there was no difference on average between sample types. Variance among samples reflects some combination of imperfect quantification prior to pooling and variation in read quality. To avoid biases arising from variation in sequencing depth among samples, rarefaction to 1819 reads per sample was used in all subsequent diversity analyses (diversity and distance matrices were calculated as averages of 1000 replicate subsamples of 1819 reads per sample).

Diversity partitioning

To evaluate how well 1819 reads represents the total richness of a habitat, we estimated rarefaction curves for every sample (Fig. S2). No sample appeared close to saturation. After rarefaction, diversity per sample was similar across habitat types (Table 1). Average richness per salamander was slightly greater than average richness per cover object, and logs tended to have slightly lower richness. The distribution of relative abundances was quite uneven in all habitat types, as shown by the decline in Hill numbers with order (q). Salamander skins had the most skewed (least even) distribution of bacterial OTU abundances (Table 1).

Table 1. Bacterial diversity by habitat type. All diversity metrics are means after rarefaction to 1819 reads
  1. a

    Salamanders had significantly lower DINV than logs (Tukey's HSD: = 0.0001), but not rocks (= 0.1289). Rocks and logs were statistically similar (= 0.8353). Shannon's diversity (H or eH) was not distinguishable among habitat types (> 0.17).

Reads per sample28 471.151 301.620 010.3
Rarefied richness (S)628.09567.71609.42
Hill no. = 1 (eH)237.38258.55266.74
Hill no. = 2 (DINV)66.26a97.82106.79
Evenness (DINV/S)0.1030.1670.171

Full hierarchical additive partitioning of diversity (Table 2) revealed that the majority of total diversity can be found within each habitat type (salamander, rock, or log). Turnover between habitat types (salamander vs. rock vs. log) accounted for 15–23% of species richness, 3.3–3.5% of Shannon diversity, and only 0.25–0.29% of Simpson diversity. These differences between diversity metrics indicate that most of the differentiation between host-associated and free-living assemblages is accounted for by rare microbes, which are counted in richness but have little effect on higher order diversity indices. Moreover, observed presence of rare OTUs (and therefore richness) is sensitive to sample size: Partitioning the diversity without rarefaction estimated the between-habitat β+ to account for only 2.5% of the richness (not shown), and the average individual sample included over 1500 OTUs (65% of the total richness could be found in a single swab).

Table 2. Additive hierarchical partitioning of bacterial diversity. Habitat types are salamander skin, logs, or rocks
γTotal diversity2313.96.3360.99187
αWithin samples587.0 (25.367%)5.504 (86.870%)0.98514 (99.321%)
inline image Samples within habitat types1092.8 (47.226%)0.553 (8.733%)0.00361 (0.364%)
inline image Habitat types within transects532.6 (23.019%)0.223 (3.526%)0.00286 (0.289%)
inline image Between transects101.5 (4.388%)0.055 (0.871%)0.00026 (0.026%)
Alternative hierarchy:
inline image Transects within habitat types272.6 (11.780%)0.065 (1.028%)0.00057 (0.058%)
inline image Between habitat types361.6 (15.627%)0.213 (3.369%)0.00255 (0.257%)

Multiplicative partitioning of diversity (Hill numbers) gives a complementary perspective (Table 3). The total diversity is only about 1.2–1.4 times the diversity found within a habitat type (host or cover object). That is, there are effectively only about 1.3 completely distinct microbial habitats in the study system.

Table 3. Multiplicative hierarchical partitioning of bacterial diversity. Habitat types are salamander skin, logs, or rocks
γTotal diversity2313.99564.44123.15
αWithin samples586.64254.7090.20
β1Samples within habitat types2.861.701.10
β2Habitat types within transects1.311.231.19
β3Between transects1.051.061.03
Alternative hierarchy:
β2Transects within habitat types1.191.081.19
β3Between habitat types1.191.231.20

Community similarity between host-associated and free-living samples

Unconstrained ordinations show consistent differentiation between host-associated and free-living samples, however, some host-associated samples were more similar to certain free-living samples than to other host-associated samples (Fig. 3). This qualitative pattern held for presence–absence data in addition to relative abundance (Fig. S3). Differences between free-living assemblages on rocks vs. logs were most apparent in presence–absence analyses, suggesting that the occurrence of certain rare species depended on substrate type, but the relative abundances of common species were similar. Indeed, relative abundances were highly correlated across all three habitat types (Fig. 4).

Figure 3.

Principal coordinates plot showing variation between host-associated and free-living bacterial assemblages. Symbols closer together represent samples with more similar bacterial assemblages. Differences in bacterial samples were quantified as 1 – (Horn overlap) following the method of Jost (2007). Other metrics and ordination methods gave similar results (Fig. S3).

Figure 4.

Correlation between relative abundance of bacterial OTUs in host-associated vs. free-living samples. Each point is a ‘species level’ OTU (97% similarity clustering). Dotted lines mark the position of zeros (note log10 scale axes). Light dashed line illustrates the line of identity. Correlation is represented by Kendal's τ (all P-values < 10−15, = 2316). (a) Salamanders vs. logs, (b) Salamanders vs. rocks, (c) rocks vs. logs.

Habitat type (salamander vs. rock vs. log) was a statistically significant predictor of community similarity (Table 4). Differences among transects accounted for only about 3% of the variation in bacterial OTU occurrence or abundance. Adding elevation or geographic distance to the analyses accounted for < 1% of the variation, and these variables are strongly correlated with transect location (not shown). Results were the same for permutational anova and redundancy analyses (not shown). Pairwise randomization tests comparing each salamander-associated sample with a sample from its particular cover object always rejected the null hypothesis of no difference (Table S2), and there was no evidence that a salamander-associated sample tended to be more similar to the sample from its particular cover object than to the average host vs. cover comparison (Table S2).

Table 4. Partitioning of dissimilarity by habitat (salamander skin, log, or rock) and transect (Fig. 1) using constrained analysis of principle coordinates (Legendre & Anderson, 1999; Borcard et al., 2011)
 d.f.Variance (Inertia) F P
(a) q = 0 (Sorenson distance)
Habitat26.37918.85%8.606< 0.001
Transect21.5134.47%2.041< 0.001
(b) q = 1 (Horn distance)
Habitat21.75122.69%10.863< 0.001
Transect20.3254.21%2.017< 0.001
(c) q = 2 (Morisita–Horn distance)
Habitat21.38953.89%44.095< 0.001

Identification of specialists

Indicator species analysis identified 446 bacterial OTUs having statistically significant associations with salamanders. There is a nonlinear relationship between mean relative abundance and rg; OTUs most abundant on salamanders tended to have positive correlations and the least abundant tended to have negative correlations, but for the vast majority of the range of abundances, there appears to be no relationship (Fig. S4).

OTUs having strongest associations with salamanders (as measured by point-biserial correlation rg) tended to occur at moderate abundance on most salamanders (high consistency) and tended to be absent (or too rare to be detected) in most cover object samples (high specificity). On the other hand, OTUs with highest specificity (PDI) tended to be rare and not consistently detected on all salamanders. None of these candidate salamander specialists were found exclusively on salamanders (Table S3).

A statistical cluster of potential salamander specialists is evident in the bimodal distribution of rg, the strength of association with salamanders in particular (Fig. 5). This distribution is better described as bimodal than unimodal (ΔAIC = 367.25, mixture of two normal distributions vs. best single normal using mixtools in r). Many top candidate salamander specialists grouped with the genus Nevskia, but could not be assigned to any described species based on 97% similarity (Table S3). Nevskia was very common across all samples (Fig. 2), but salamanders and cover objects tended to harbor different Nevskia OTUs (Table S3). Nevskia spp. are common in soil and water (Cypionka et al., 2006) and were found previously in samples from aquatic amphibian skin (McKenzie et al., 2011). The most well-known species (N. ramosa) is known for biofilm formation on still water (Cypionka et al., 2006).

Figure 5.

Distribution of the point-biserial correlation, rg, representing strength of association between bacterial OTUs and salamander skin (Plethodon jordani) relative to cover objects. Smooth lines were fitted under the assumption that the data are distributed as a mixture of two normal distributions.

Although not common (low consistency), one of the most specialized OTUs associated with salamanders grouped in the class Chlamydiae (Table S3). This OTU could not be assigned to any finer scale taxonomy, suggesting a heretofore unknown lineage of bacteria. Like the human pathogen Chlamydia, all known Chlamydiae are intracellular parasites of eukaryotes, but most are associated with amoebae rather than animal hosts (Coulon et al., 2012).

Ecological differentiation within OTUs

Most bacterial OTUs lacked sufficient ‘intraspecific’ variation in 16S rRNA gene sequences to test for differentiation among sample types. The OTUs having multiple alleles tended to be those represented by many reads, and we suspect deeper sequencing would reveal multiple alleles within other OTUs. Of 132 OTUs having multiple alleles in the dataset, 131 showed significant genetic differentiation between populations in different habitats (salamanders vs. rocks vs. logs). These included nine of the 10 most common OTUs on salamanders (Table S3).

Genetic differentiation within OTUs was widespread but subtle (Fig. 6). Most of the variation (> 60%) occurred within samples, under 20% of variation was accounted for by turnover between populations on the same habitat type, and turnover between habitat types and between transects each rarely accounted for more than 5% of the variation.

Figure 6.

Distribution of additive partitions of gene diversity (Simpson's index or heterozygosity) within species level OTUs. Only the 132 OTUs with multiple alleles are included. The ‘between habitats’ component represents genetic differentiation between populations on salamanders vs. rocks vs. logs.

OTUs showing strong ecological differentiation were rarely classified below family (Table S3). However, the OTU with greatest genetic differentiation between host-associated and free-living populations was classified as Chlamydophila (another Chlamydiae). The known species of Chlamydophila infect mucosal cells of birds and mammals and can be deadly (Rhode et al., 2010). Other OTUs with substantial host-associated differentiation include two chloroplast OTUs (Table S3), suggesting the possible importance of eukaryotic symbiotes.


Whether host–symbiont systems are highly integrated and exclusive units of organization or parts of more diffuse interaction webs is a key question for understanding how they relate to broader patterns of evolution and ecology. At one extreme, hosts and symbionts might be so interdependent as to operate as collective metaorganisms (Zilber-Rosenberg & Rosenberg, 2008; Doolittle & Zhaxybayeva, 2010; Dupré, 2010). Metaorganismal models of ecology and evolution are still taking shape, but extreme versions conceptualize host-associated microbial assemblages as isolated units with minimal exchange between hosts or with free-living assemblages. At the other extreme, hosts might simply be colonized by free-living microbes from the ambient environment with varying degrees of selectivity (Nyholm & McFall-Ngai, 2004; Mueller, 2012). We evaluated these alternatives by comparing bacterial samples from hosts (Jordan's Salamanders) and their cover objects. The composition and abundance of bacterial OTUs were more consistent with a selective recruitment model of symbiotic community assembly than a metaorganismal model. This conclusion is concordant with the results of Loudon et al. (2014). However, population genetic differentiation within some OTUs suggests restricted dispersal between host-associated and free-living populations on an evolutionarily significant timescale. Overall, our results suggest that bacteria on salamander skin are often salamander-adapted genotypes of bacteria found on other moist surfaces in the region. Interaction between host and symbiotes is ecologically and evolutionarily important, but not so influential as to produce a ‘distinct’ salamander-associated flora, as implied by the metaorganismal hypothesis.

Although our host-associated and free-living samples tended to be different, they did not form distinct multivariate clusters and samples from different habitat types were sometimes quite similar in species composition and relative abundance (Fig. 3). Overall, the most abundant microbes tended to be most abundant in all habitats (Fig. 4). These similarities probably reflect both geographic association and habitat similarity between the moist surface of a salamander and the moist surface of a log or rock on the forest floor. However, relative abundances were not identical and subtle, but consistent differences were detected between salamander-associated and cover object-associated samples in all pairwise comparisons and global multivariate analyses. Thus, the bacteria on salamander skin consist of the same OTUs as those on other moist surfaces on the forest floor, and they occur in similar but not identical relative abundances. This level of similarity is not consistent with long-term evolution of host-associated microbes with minimal exchange between host-associated and free-living assemblages.

Loudon et al. (2014) also found multivariate differences between samples from salamander (P. cinereus) skin and soil from a forest in Virginia (USA). However, they remarked that the most common OTUs on salamanders were also common in environmental samples. The most common taxa they identified included Pseudomonas, which was also common in our samples (Fig. 2), and Verrucomicrobia, which was relatively rare in our data (Fig. S2). They did not count Nevskia among their top candidates for a ‘core microbiota,’ whereas it was by far the most common taxon in our study system (Fig. 2, Table S3). These major differences in taxonomic composition between our studies might reflect differences in geography, habitat, seasonality, host species, etc. However, they are consistent with the inference that salamander skin bacteria are largely recruited from the environment and not maintained in isolation, at least relative to the scales of salamander evolution and biogeography.

Multivariate similarity does not rule out the possibility that a few bacterial taxa might be highly specialized for life on salamander skin. No OTUs were restricted to, or excluded from, salamander skin. However, the distribution of indicator values was bimodal, suggesting that the assemblage of bacteria on salamander skin can be described as a mixture between a specialized and a more casual, peripheral microbiota. However, such a categorization would be arbitrary in the sense that there is no distinct gap in the distribution of specialization indices (Fig. 5). In addition, it glosses over the fact that some of the most specialized microbes might be parasites while others are mutualists. Instead, host-associated bacterial OTUs probably fall on a two-dimensional continuum, from broadly generalized to modestly specialized in one dimension, and from beneficial to antagonistic in the other.

The greatest support for evolutionarily significant association between bacteria and salamander hosts came from genetic differentiation between host-associated and free-living populations of the same ‘species level’ OTUs. This suggests a systematic restriction of exchange between habitat types (host vs. cover object) over and above the restriction owing simply to spatial separation of patches (accounted for in the hierarchical partitioning of genetic distances). That is, population allele frequencies of the 16S rRNA gene sequences tended to be more similar between host-associated samples than between free-living and host-associated samples. Multilocus analysis of particular OTUs would give a more complete picture of how often individual cells might move between the skin and the environment and how often cells from different populations exchange genes, but our analysis is adequate to reject panmixia between free-living and host-associated populations. As with differentiation at the OTU level, there are two non-mutually exclusive interpretations. First, host-associated bacteria might be transferred among hosts with greater frequency than transfer between free-living and host populations – that is individual cells are more likely to disperse between patches of the same kind of habitat rather than between habitats. Second, salamander skin might be a selective environment, such that bacteria with particular genotypes are more likely to survive and reproduce once they arrive on a host, whereas other genotypes are more likely to survive and reproduce on inanimate rocks and logs. Salamander skin is consistently in contact with the environment, and regular sloughing might enhance opportunities for recolonization by free-living bacteria. The relative importance of nonrandom dispersal vs. host-associated selection for population genetic differentiation is an important unanswered question with broad significance in microbial ecology (de Wit & Bouvier, 2006) and the evolution of conflict and cooperation (Bright & Bulgheresi, 2010; Sachs et al., 2011; Mueller, 2012).

Genetic differentiation between host-associated and free-living populations suggests restricted dispersal and/or local adaptation, at least for those particular OTUs showing genetic variation in our dataset. That is, host-symbiote associations are consistent enough, in space and time, to allow modest population genetic differentiation. This inference is not in conflict with the strong similarity observed at the community level. As with the community dissimilarity, the population genetic dissimilarity is statistically significant but subtle, with habitat type accounting for only a small fraction of total genetic diversity. Diversity and dissimilarity at both levels represent the long-term outcome of the same conflicting processes (Vellend, 2010); dispersal and parallel selection tend to homogenize relative abundances of genotypes and species while drift and differential selection tend to produce differences. Horizontal transfer of genes within and between species might allow ecologically important genes to segregate more strongly between habitats, and this represents an interesting direction for future research. Overall, the ecological and evolutionary relationships between salamanders and their skin bacteria appear to be strongly influenced by but not entirely dependent on free-living bacterial communities.

Taken together, the similarities and differences between host-associated and free-living bacterial assemblages are not consistent with long-term isolation of a distinct set of host-associated taxa with a unique community structure on salamander skin. Rather, salamander skin has the same taxa in similar proportions as other moist surfaces on the forest floor. Subtle but consistent differences in prevalence and relative abundance confirm that salamander skin is not ecologically identical to the underside of a rock or log, but the broad similarity of community structure was surprising. Instead of an isolated and exclusive community of host-associated microbes, there is a smooth continuum from modestly specialized to broadly generalized OTUs and genotypes.


Alison Buchan, Brandon Matheny, the HOFF Lab, and two anonymous reviewers provided comments and advice. Funding was provided by the Tennessee Herpetological Society and Department of Ecology and Evolutionary Biology, University of Tennessee Knoxville.