Data Analysis
To evaluate overall patterns of microbiome alpha diversity, t-tests and
Mann-Whitney U tests were performed to compare total observed eukaryotic
or bacterial OTUs across site types (island vs. mainland) or MHC IIB
genotypes (homozygote vs. heterozygote) in SPSS (vrs. 22). Analyses of
microbiome community structure (beta diversity) were conducted using
Mantel tests of community dissimilarity vs. geographic distance or
genetic distance (FST) implemented in the ade4 package
of R (vrs. 1.7-11) (44–47).
To examine associations between microbial communities and geography or
host frog MHC IIB genotype, data were statistically analyzed and
visualized using packages implemented in Python (vrs. 2.7.13) and
Matplotlib (48,49). Associations between microbial communities and
geography or frog MHC IIB genotype were determined by simulating an
expected null distribution of host frog microbiomes. To create the null
distribution, a two-column data table was first created with column 1
being the site type (island or coastal) or MHC IIB genotype
(heterozygous or homozygous) of a host frog and column 2 being one
microbial OTU found on that frog. After the data table was populated for
all frogs and microbes in the dataset, column 2 (microbial OTU) was held
constant while column 1 (site type or frog genotype) was shuffled
randomly. This was repeated 1000 times to create two sets of random
microbial occurrence distributions, one for analysis of microbial
associations with site type and a second for analysis of microbial
associations with host frog genotype.
Co-occurrence between microbial OTUs within and among domains (Bacteria
vs. Eukaryotes) was analyzed with a third null distribution of microbial
communities. Because of potential site effects on microbial presence and
community structure (e.g ., some microbes only co-occur on frogs
because the microbes themselves solely occur at the same subset of
sites) and site-MHC IIB genotype interactions (as homozygotes and
heterozygotes are not evenly distributed across sites or site types;
Table 1), an expected null distribution of microbes accounting for
site-specific presence/absence of each microbe was created. This null
distribution of microbes was achieved through within-site randomization
using MCMC edge swapping, a standard method for network datasets
(50–52). This method allows any configuration to be reached from any
starting point, and allows for even sampling along all allowed states as
forward and backward swaps are equally likely. To achieve this, first,
two microbe-frog pairs were randomly selected (each pair consisting of a
single randomly selected microbial OTU found on a single randomly
selected frog). Microbial OTUs were then swapped between the selected
frogs when three criteria were met: (1) the frogs were different
individuals with the same MHC IIB genotype (either both homozygous or
both heterozygous); (2) the OTUs were different from one another; and
(3) neither frog already hosted the microbe it would receive via the
swap. Microbe swapping was performed with 1000 repetitions for each
frog-microbe pair.
To test whether hypothesized bacterial effects on Bd extend to
diverse microeukaryotic members of the microbiome, bacterial OTUs that
matched the Woodhams et al. (2015) database were binned according to
their hypothesized ecological significance with regard to Bd(Bd inhibitory, Bd enhancing, or no effect on Bd ).
The co-occurrences of bacteria within each category with microbiome
eukaryotes were then compared with the third null distribution of
microbial OTUs.
For all microbial association/co-occurrence analyses, the probability of
non-random microbial association/co-occurrence (p ) was calculated
by comparing observed versus expected counts of microbial
association/co-occurrence. P -values were evaluated at a
significance level of α = 0.05. Using the results of the tests of
co-occurrences within and among all microbial taxa, network analyses
were performed and visualized in SciPy (53).