Population Structure
Both DAPC and STRUCTURE indicated that the most likely number of population clusters was two and the second most likely was three, as determined by BIC and the ΔK method, respectively. The ΔKmethod is biased toward K = 2 (Janes et al., 2017; Campana, Hunt, Jones, & White, 2011) and simulation studies have shown that the mean probability (MeanLnP(K)) output from STRUCTURE performs better in scenarios with high gene flow and low F ST (Latch, Dharmarajan, Glaubitz, & Rhodes, 2006). Because K = 3 produced the highest MeanLnP(K) in the STRUCTURE analysis (Table S4a), we consider this a relevant model and show the proportion of individual membership in each cluster as defined by each of the two analyses for both K = 2 and K = 3 (Figure 5). We also ran STRUCTURE separately for individuals caught on MK and MT, with settings described above except we ran simulations for K = 1–5. We found no evidence of structure among MK individuals. MT individuals were divided into two clusters - one with individuals ≥ 2000 masl and one with individuals < 2000 masl, with individuals of mixed ancestry at 2000 masl (Table S4b).
Cluster membership is mostly concordant between the DAPC and STRUCTURE analyses, except STRUCTURE assigned mixed ancestry to many individuals while DAPC did not. This is not unexpected as previous studies have shown that DAPC may underestimate admixture (Frosch et al., 2014) while STRUCTURE is more accurate at assigning mixed ancestry (Bohling, Adams & Waits, 2012). When K = 2, individuals at 2000 and 2400 masl on MT form a separate cluster from low elevation MT + MK (Figures 4 & 5), with mixed ancestry individuals at 2000 masl MT. For K = 3, the divisions are between high elevation MT, low elevation MT, and MK, with individuals assigned mixed ancestry at low elevation MK (900 masl) and 2000 masl MT (Figure 5).
The PCA shows a similar pattern. Eigenvector 1 (7% variation explained) separates the two mountains, with overlap among individuals at 900 masl. Eigenvector 2 (4% variation explained) partially separates individuals by elevation, with lower elevation individuals clustered together (Figure 6). With K = 2, after removing individuals that could not be assigned to a STRUCTURE cluster (cutoff q kvalue < 0.6), F ST is 0.05 (p = 0.0001). The AMOVA showed 95% of variation is partitioned within clusters and 5% between. With K = 3, removing individuals withq k values < 0.6,F ST between MK and low elevation MT was 0.035 (p = 0.001), between MK and high elevation MT 0.092 (p = 0.0001), and between low elevation MT and high elevation MTF ST = 0.065 (p = 0.0005) (Tables 2a & 2b). The AMOVA showed that 94% of variation is distributed within clusters, and 6% among.
Including data for all 80 individuals, the Mantel test revealed a significant, positive correlation between genotypic distance and geographic distance (r = 0.287, p = 0.0001). Including the 58 unrelated individuals, the correlation was weaker but statistically significant (r = 0.05, p < 0.0001). The correlogram showed significant positive autocorrelation between individuals at distances of 200 m and less (r = 0.091, pr -rand ≥ p r -data = 0.0001) and 200 m–1 km (r = 0.036, p r -rand ≥ p r -data = 0.0001); autocorrelation was no longer significant at 2 km (r = -0.001, p r -rand ≥ p r -data =0.598). At subsequent distance classes (5, 10, 15, and 18 km), individuals have greater genetic distance than expected at random; i.e., there is a significant negative correlation between genetic and geographic distance (p r -rand ≤ r- data = 0.009, 0.0001, 0.0001, 0.0001, respectively) (Figure S4). The average geographic distance between pairs of first-order relatives was 162.5 m, second order was 1.2 km, third order was 4.8 km, and between distant or ‘unrelated’ individuals was 12 km (Table 1). Differences between first and third, first and distant, second and distant, and third and distant relatives were significant (p < 0.05).