The Influence of Historical Geneflow, Bathymetry and Distribution Patterns on the Population Genetics of Morphologically Diverse Galápagos’ Opuntia echios

Natural History and Field Samples

The isolated Galápagos archipelago (Ecuador) comprises 13 large islands (>10 km²) and over 130 smaller islands and rocks (Fig. 1) with ages ranging from 0.7 to <5 Myr, following a west-east ageing gradient (Hickman and Lipps 1985; White et al. 1993; Geist 1996). Submerged seamounts located east of the presently emerged island, however, extent timing of first colonisation up to 14 Myr (Christie et al. 1992; Werner et al. 1999). Santa Cruz is amongst the largest and highest island found within Galápagos’ central region and comprises most of the known vegetation zones (Itow 1971). Its 135 km long jagged coastline is characterized by coastal cliffs at the north–north-eastern side (van der Werff and Adsersen 1993). The surrounding islets (Plaza Sur, Baltra, Daphné Major, and Seymour Norte) are characterized by one uniform vegetation zone: the “arid zone”. On Santa Cruz this arid zone ecosystem, which is the most important habitat for Opuntia species on the archipelago, forms a “vegetation” ring around higher altitude regions. Moreover, due to the fact that the southern part of Santa Cruz receives considerably more precipitation than the northern part (Alpert 1946) this arid ring varies in width (Fig. 1).

Opuntia echios samples (n = 444 in total), representing four varieties, were collected during two sampling periods (225 and 219 individuals in 2002 and 2007, respectively) at 22 localities (Fig. 1). Since Santa Cruz is thought to house two geographically isolated varieties and their morphological intermediates, it was sampled most extensively (18 localities). Due to their small size, only one population was sampled on Plaza Sur, Baltra, Seymour Norte, and Daphné Major. Whenever possible, tissues were collected from 30 individuals at any given locality. To avoid clonal sampling, minimal inter-sample distances were set at 100 m (Browne et al. 2003). Using this protocol, many plants were left unsampled for the majority of locations (i.e. distance to other samples <100 m). However, for some locations (especially for the eastern side of Santa Cruz) distances between neighbouring Opuntias exceeded the preconceived 100 m. Here, inter-sample distances were between 150 and 500 m. For each plant a fresh, mature terminal cladode (a modified stem segment resembling a leaf) was separated at the joint with a machete, a process comparable to natural damage by wind, and transported in an individual paper bag to the Charles Darwin Research Station (CDRS) where they were stored at room temperature. Before shipping, pads collected in 2002 were cut into pieces of approximately 10 × 10 cm, and transported in paper bags at ambient temperature to both the Royal Belgian Institute of Natural Sciences (RBINS) and the University of Antwerp (UA) where they were placed in plastic bags and stored at −20°C. New Ecuadorian export regulations, which went into effect shortly after the start of our second collection trip, resulted in an extended preservation period at Galápagos, where moist conditions resulted in suboptimal conservation of fresh material. After the loss of 30 samples (due to rotting), the remaining samples were kept from decay by isolating 25 cm² of tissue (excluding both the cuticle and the parenchyma layer containing a mucilaginous secretion). These samples were stored in four individual plastic vials at −20°C at CDRS. Two of these vials were sent in a cooled container (dry ice) to the UA where they are stored at −20°C whilst the two remaining vials were kept at CDRS (−30°C), where they remain as reserve specimens.

DNA Extraction and Microsatellite Genotyping

Genomic DNA was isolated using DNAeasy Plant Mini Kits (Qiagen). In addition to standard procedures, an extra centrifugation step was executed to the lysate (5 min at 13,000 RPM) to remove most precipitates. To maximize overall DNA yield, extraction products were dissolved in 100 μl dH₂O. All individuals were genotyped using the eight most variable microsatellite markers specifically developed for Galápagos’ Opuntias (Helsen et al. 2006, 2009a). Except for changing cy5 labels in FAM (Opuntia1, Opuntia2, Opuntia5), VIC (Opuntia10, Opuntia11, Opuntia15), and NED (Opuntia9, Opuntia13), amplifications were performed as described by Helsen et al. (2006). Resulting PCR products were analysed on ABI 3730 instruments (Applied Biosystem), genotyped using size calling in GeneMapper Software v3.7 (Applied Biosystems), and visually checked afterwards. Thirty individuals collected during the 2002 trip were reprocessed in 2007 to check for potential scoring differences between years. Due to poor quality of some of the material, a limited number of individuals could not be genotyped (even after repeated amplifications) for all loci. For those specimens, these loci were treated as missing data. We evaluated whether these non-amplifications resulted from the impact of amplicon size on PCR success [reviewed by Broquet et al. (2007)].

Analysis of Genetic Variation

Opuntia echios’ hexaploid nature (Helsen et al. 2009a), hampers genetic analysis since individuals may posses up to six different alleles for each locus. Partial heterozygosity (Bruvo et al. 2004) makes it impossible to score genotypes exactly. For example, individuals bearing two alleles for one specific locus (i.e. A and B) may have any of the following genotypes: AAAAAB, AAAABB, AAABBB, AABBBB and ABBBBB.

We used two methods to counter such scoring problems in polyploids: (i) score alleles as presence/absence data (Rodzen et al. 2004) and (ii) a relatively novel approach termed Microsatellite DNA Allele Counting using Peak Ratios (MAC-PR) (Nybom 2004; Esselink et al. 2004; Nybom et al. 2006), hereafter referred to as the P/A- and MAC-PR datasets, respectively. Since P/A methods score co-dominant microsatellites as being dominant, some information is lost and as such the power of the analysis is reduced (Bjorklund 2005). In the latter method, individual allele dosage at a specific locus is estimated based on the assumption that abundant alleles amplify more often than rare alleles do. Relative allele peak heights are subsequently used to evaluate individual allele dosage, thereby correcting for allele-specific amplification success by comparing peak heights of all combinations of alleles (see Esselink et al. 2004 for further details). Due to the high genetic variability and hexaploid nature of some samples, we occasionally were unable to define unambiguous correction factors to neutralize this effect. In such cases, we used a more robust method (assigning <6 alleles) comparable to the method described by Jenneckens et al. (2001).

Genetic diversity of the P/A dataset was estimated for each locality and putative variety using Nei’s unbiased diversity estimator on the individual marker level (Nei and Roychoudhury 1974). Before calculating total genetic diversity for both the locality and island levels, we evaluated the effect of differences in sampling effort on estimates of genetic diversity. Making use of the theoretical logarithmic relationship between sampling effort and genetic diversity, we used the most thoroughly sampled locations (n ≥ 20) to estimate the number of individuals needed to capture most (i.e. 95%) of the total genetic diversity. For each of these more extensively sampled locations, we calculated genetic diversity of repeatedly (n = 100) composed subpopulations containing one to the total number of sampled individuals, thereby randomly picking unique individuals from the original dataset (resampling procedure in R). The result of this simulation was used in later analysis of genetic diversity by either omitting locations containing less individuals than this threshold value and/or adjusting diversity estimates.

Differences in genetic diversity between varieties were evaluated using a mixed-model ANOVA with variety as fixed effect and locality (nested within group) and allele as random effects. We also tested for inter-locality differences within SAS release 9.1 (SAS Institute Inc). Numbers of private alleles were calculated on both the island and locality level.

Genetic differentiation at different hierarchical levels (within locations, between locations and between islands) was calculated on the P/A dataset using analysis of molecular variance (AMOVA) in ARLEQUIN 3.01 (Excoffier et al. 2005). Since AMOVA analysis assumes Hardy–Weinberg equilibrium, the P/A dataset was tested for deviations from this equilibrium using a Bayesian-based deviance information criterion (DIC) selection statistics model implemented in Hickory v1.0.4 (Holsinger et al. 2002; Holsinger and Wallace 2004). The P/A data of all individuals and those of Santa Cruz exclusively were fitted to four models: (i) a “full model” which allows for inbreeding and includes priors for f, π_i (the mean of the beta distribution (Holsinger et al. 2002)), and θ; (ii) “f = 0” assuming no inbreeding; (iii) “θ _B = 0” implying a zero-valued Bayesian F _ST estimate; and (iv) “f free” model allowing the incorporation of uncertainty about f into the analysis. Default settings were used for burn-in (50000), sampling (250000), and thinning (50), running five independent MCMC runs to test consistency of the outcome. The best fit of these models was evaluated using measures of DIC. Resulting best fitted models were subsequently used to estimate parameters related to genetic structure (θ_B: an estimate of F _ST under random-effects model population sampling).

The relationship between geographical and pair wise genetic distances on the locality level (calculated as F _ST/(1 − F _ST)) was evaluated to determine whether isolation by distance (IBD, Wright 1943) explains the observed distributions of phenotypic differences. Since Opuntias are restricted to the arid ecosystem, which roughly follows the coastline within high altitude islands, we used shortest coastline distances as an alternative of straight in line pairwise geographical distances. These coastline distances were calculated based on a map displaying sample location as revealed by their coordinates using ImageJ software (Abramoff et al. 2004). Whilst it remains to be tested how Opuntias genes spread within Galápagos, both the effect of geographical distances and their log transformation (one vs. two dimensional, Rousset 1997) were implemented in a partial Mantel test (with 3.10⁵ bootstrap randomizations) within IBDWS (Jensen et al. 2005).

We evaluated abrupt changes in genetic variation using both BARRIER software (Manni et al. 2004) and a method described by Koizumi et al. (2006). This latter method makes use of biased residual values of the regression between pairwise genetic and geographic distances to define divergent populations (Koizumi et al. 2006). We, therefore, verified whether populations with the highest mean residuals contained zero within their 95% CI, as described by Koizumi et al. (2006). If the relevant CI did not include zero, the population was excluded and the IBD procedure was rerun.

Since genetic differentiation between islands (and locations within them) is potentially small and as a result of differences in allele frequencies rather than the presence or absence of alleles, we used a model-based Bayesian clustering approach on the MAC-PR dataset to define the most likely number of clusters or subpopulations (K) (STRUCTURE v. 1.0: Pritchard et al. 2000). This model accounts for the presence of Hardy–Weinberg or linkage disequilibrium by introducing population structure and attempts to find population groupings that (as far as possible) are not in disequilibrium. Based on preliminary analysis, we evaluated the likelihood for K = 1–12 for the whole dataset. For each analysis, ten independent runs of the Gibbs sampler were evaluated for each K to check potential mixing. Because of the potential for being closely related, due to migration and possibly shared recent ancestry, a combination of the “admixture model with no prior population (geographical) information” and the “correlated allele frequency model” was used to analyse our data (Pritchard et al. 2000). Based on stationary in several statistics, burn-in length was set to 5.10⁴ after which 2.10⁶ MCMC replicates were run. The most likely number of clusters was evaluated using ΔK values (Evanno et al. 2005). In order to unravel substructuring, analyses were rerun on partial datasets which represent population genetic clusters as defined by the outcome of the first Structure analysis (Coulon et al. 2008).