Genetic Data
We sampled 56 individuals from 31 localities (Appendix S1; Fig. 1) from
the known range of the species, including the recently discovered
populations in the Piedmont and Coastal Plain of Georgia (Graham et al.
2012). This includes isolated localities in central and western Alabama,
including near the type locality of Desmognathus chermocki Bishop
and Valentine, 1950, a junior synonym of D. aeneus Brown and
Bishop, 1947 after Chermock (1952). We generated SNP data using
genotype-by-sequencing (GBS; Elshire et al. 2011) with the ApeKI enzyme
at the University of Wisconsin-Madison Biotechnology Center. We
sequenced libraries on a NovaSeq6000 (Illumina Inc.), with coverage
ranging from 2.3–14.6 (mean 10.2) million reads per sample. We
optimized assemblies in ipyrad v.0.9.87 (Eaton and Overcast 2020) using
the criteria proposed by Ilut et al. (2014) and McCartney‐Melstad et al.
(2019). These yielded an optimal threshold of 97% similarity for
clustering on an initial sequencing run of 33 individuals (Fig. S1),
which we retained for the full analysis of 56. We enforced a threshold
of >60% presence per locus, equal to a minimum of 34/56
individual coverage across samples. The raw assembly yielded 13,091 SNPs
with 26,593 alleles from 2,132 loci. After filtering for missing data
(>33%) and singleton alleles (Linck and Battey 2019), we
retained 2,281 SNPs with 4,661 alleles. A group of seven samples had
<50% coverage and were therefore dropped for the conStruct,
IBD, and RDA analyses (see below), which are sensitive to missing data.
Part of these analyses were carried out on the GW HPC Pegasuscluster (MacLachlan et al. 2020).