Biogeographic barriers such as rivers have been shown to shape spatial patterns of biodiversity in the Amazon basin, yet relatively little is known about the distribution of genetic variation across continuous rainforest. Here, we characterize the genetic structure of the brilliant-thighed poison frog (Allobates femoralis) across an 880-km-long transect along the Purus-Madeira interfluve south of the Amazon river, based on 64 individuals genotyped at 7609 single-nucleotide polymorphism loci. A population tree and clustering analyses revealed 4 distinct genetic groups, one of which was strongly divergent. These genetic groups were concomitant with femoral spot coloration differences, which was intermediate within a zone of admixture between two of the groups. The location of these genetic groups did not consistently correspond to current ecological transitions between major forest types. A multimodel approach to quantify the relative influence of isolation-by-geographic distance (IBD) and isolation-by-environmental resistance (IBR) nevertheless revealed that, in addition to a strong signal of IBD, spatial genetic differentiation was explained by IBR primarily linked to dry season intensity (r² = 8.4%) and canopy cover (r² = 6.4%). We show significant phylogenetic divergence in the absence of obvious biogeographical barriers and that finer-scaled measures of genetic structure show patterns that are associated with environmental variables also known to predict the density of A. femoralis.

Extraction of DNA and single-nucleotide polymorphism (SNP) discovery was carried out at Diversity Arrays Technology sequencing Pty. Ltd. (DArTseq) facility (Canberra, Australia; more detail, please see the Supplementary Information in the article). A modified double-digest restriction-site associated DNA sequencing protocol was performed on libraries prepared using a combination of Pstl-Hpall restriction enzymes. The Pstl enzyme adaptor also contained an Illumina adaptor sequence, a primer sequence and a variable-length barcode. The Hpall adaptor contained an Illumina flow cell attachment and overhang sequence. Following enzymatic digestion, fragments were amplified and sequenced on an Illumina HiSeq2500. DNA sequences were aligned via BLAST using the Nanorana parkeri reference genome. To check for contamination, sequences were also blasted to bacterial and fungal genomes (NCBI).

This data set was filtered for missing data using the filter_dart function of the R package RADIATOR v. 0.010. Only individuals and loci with ≥95% SNPs geno- typed were retained. SNPs were also screened for allele coverage, with any SNPs displaying a local and global minor allele frequency threshold of less than 1% removed from the dataset. In cases where multiple SNPs were found within the same read, only one locus was retained (chosen randomly per RAD tag) to avoid statistical bias from physical linkage. Two samples from M14 had <95% of loci genotyped and were removed, which resulted in 64 individuals from 13 populations genotyped at 10 275 SNPs (please see the article Table S2 for summary of filtering steps).