Bioinformatic analysis.

Sequences were processed using Mothur (53). In brief, sequences were removed from the analysis if they contained ambiguous characters, had homopolymers longer than 8 bp, and did not align to a reference alignment of the correct sequencing region. Unique sequences and their frequency in each sample were identified and then a preclustering algorithm was used to further denoise sequences within each sample (54). Unique sequences were identified and aligned against a SILVA alignment (55) (available at https://www.arb-silva.de/aligner/). Sequences were chimera checked using UCHIME (56), and reads were then clustered into 97% operational taxonomic units (OTUs) using OptiClust (57). OTUs were classified using the SILVA reference taxonomy database (release 132; available at https://mothur.org/wiki/silva_reference_files/#release-132). All data were visualized in R and Excel. For α- and β-diversity measures, all samples were subsampled to the lowest coverage depth and standard indices were calculated in Mothur. Community structure was investigated using the Yue and Clayton similarity estimator, nonmetric multidimensional scaling (NMDS) was used to compare bacterial community structures across all samples, and a multiple-sample analysis of molecular variance (AMOVA) was used to test the significance of differences between microbial communities.