Keeping up with the neighbours: using a genetic measurement of dispersal and species distribution modelling to assess the impact of climate change on an Australian arid zone gecko (Gehyra variegata)

Abstract

Aim

During this century rapid climate change will have a profound effect on global biodiversity, and species survival will be contingent on their ability to adapt or disperse. Species distribution models are a popular tool for gauging how the distribution of favourable climate may change over space and time. Evaluating the level of dispersal between the current distribution and potential future distribution of a species is a key to predicting their survival, but rarely estimated. Here we applied species distribution models and a genetic estimate of dispersal to quantitatively assess dispersal to new distributions in the timescale imposed by future climate change.

Location/Methods

We sampled 635 adult Gehyra variegata (2n = 40a/38b) throughout central and eastern Australia, encompassing much of the recorded distribution for this gecko. We genotyped all individuals at 16 microsatellite loci, from which we estimated mean annual dispersal distance using Wright's neighbourhood size. Species distribution modelling predicted the current and future distribution of the species, and we used annual dispersal distances to evaluate whether the species could keep up with shifts in the range of their favourable climatic conditions.

Results

Our estimates of mean dispersal showed that 17–41% of the current G. variegata (2n = 40a/38b) distribution was unlikely to contribute to their future distribution given the timescale imposed by future global climate change.

Main Conclusions

Our approach can make further use of molecular and occurrence record datasets to answer whether a species has the capacity to reach future areas of favourable climate and the extent to which the current distribution will contribute to this process.

Introduction

It is common knowledge that global climates have influenced the natural distribution of biodiversity. Evidence from both contemporary observations (Hughes, 2000; Walther et al., 2002; Thuiller et al., 2005) and the fossil record (Davis & Shaw, 2001) demonstrate the influence changing climates exert on species' distributions. Atmospheric greenhouse gases are certain to increase in the future and General Circulation Models (GCMs) predict global warming in the range of 1.1–6.4 °C by the year 2100 relative to 1990 (IPCC, 2007). This unprecedented rate of change in the world's climate is expected to result in numerous extinctions (Thomas et al., 2004). For example, across lizard taxa it is predicted that an appropriate thermal niche will not be retained in situ, and that climate change has already resulted in 12% of local populations becoming extinct across 48 species of Mexican lizard since 1975 (Sinervo et al., 2010). To survive, species may adapt to climate change either through plasticity or evolutionary adaptation, but this might not be possible within the timescales imposed by global climate change (Gienapp et al., 2008; Visser, 2008) or when the conditions have not been experienced in their evolutionary history (Ghalambor et al., 2007). The alternative to adaptation is relocation, the success of which is contingent on the dispersal ability of the organism in question and the scales at which new favourable areas are located (Gaston & Blackburn, 2002; Thomas et al., 2004).

Species distribution modelling (SDM) is a well-established technique used to predict species' distributions under various climate scenarios. The technique correlates species' occurrence records with climate variables to model the environmental requirements, which are then used to predict species' distributional patterns. SDMs have shown how species' ranges may shift in response to climatic changes, thus revealing any loss or gain in the areas with favourable conditions (Cordellier & Pfenninger, 2009; Fouquet et al., 2010). Predictions from SDMs have been coupled with genetic techniques to infer the process of past divergences and the locality of refugia (Knowles et al., 2007; Waltari et al., 2007). A problem with SDM is that this correlative approach predicts species' distributions without explicitly incorporating processes that potentially limit its range (Kearney & Porter, 2004; Guisan et al., 2006; Heikkinen et al., 2007; Morin & Thuiller, 2009). To more accurately predict species' future distributions, consideration needs to be given to mechanistic variables including physiological traits, biotic interactions and dispersal characteristics (Davis et al., 2005; Kearney & Porter, 2009).

Dispersal is likely to be amongst the most important mechanistic factors influencing species' ranges, and may therefore be a key variable for predicting future distributions following climate change. The need to consider dispersal when predicting changes in species' distributions has been recognized many times throughout the last two decades (Pitelka et al., 1997; Cain et al., 1998, 2000; Nathan & Muller-Landau, 2000; Araújo & Guisan, 2006). However, the few SDM studies that have incorporated dispersal to gauge the effect of climate change are plant studies (Ostendorf et al., 2001; Iverson et al., 2004; Engler & Guisan, 2009). Additionally, to our knowledge none have used a genetic estimate of dispersal, which is important when considering dispersal over multiple generations. A major impediment to predict future distributions is obtaining measures of dispersal at the landscape scale.

In the last decade genetic techniques have been utilized across many taxa to indirectly measure dispersal (Stow et al., 2001; Sumner et al., 2001; Watts et al., 2007; Pinsky et al., 2010; Duckett & Stow, 2012). Genetic estimates of dispersal could be combined with SDMs to better assess the likelihood of species reaching potential future distributions within the required timescale. SDM first predicts the current and future distribution of species using a correlative approach with present data and environmental variables. Then, using a dispersal estimate inferred from the genetic data we can evaluate whether a species is likely to keep pace with their predicted range shift. This interdisciplinary approach incorporates both correlative and mechanistic variables that limit species' distributions to quantitatively assess the proportion of a current species distribution that can feasibly reach their future distribution. Increasing the capacity to predict species' responses to climate change will help improve the uncertainty in this field (Heller & Zavaleta, 2009).

In this study we investigated the impact of climate change on an Australian arid zone gecko. The Australian arid zone represents the country's largest biome which houses exceptional levels of biodiversity including a diverse and endemic lizard fauna (Cogger, 2000; Byrne et al., 2008). The arid zone spans c. 70% of continental Australia and is characterized by annual climatic variation, some topographical heterogeneity, ephemeral river systems and a sparse mosaic of vegetation (James & Shine, 2000; Martin, 2006). We used microsatellite markers to estimate levels of dispersal and annual dispersal distances for a small tree dwelling gecko (Gehyra variegata; 2n = 40a/38b) across the major landscape features in arid Australia. We then coupled these estimates with SDMs to determine if the species was capable of reaching new distributions imposed by future climate change. Because of the speed at which the climate was expected to change, we predicted that a proportion of this species current distribution would be too far away from its predicted future distribution, which would result in isolation from favourable climatic conditions. We also predicted that a proportion of the area that was suitable for G. variegata (2n = 40a/38b) in the future was unlikely to be colonized over the next 70 years. If either of the predictions is confirmed we will identify and quantify those areas.

Methods

Study species and sampling strategy

Gehyra variegata is a gecko that is widely distributed throughout the arid regions of Australia and has four distinct chromosomal races (2n = 40a, 38b, 40b, and 44; King, 1979). The rock dwelling 2n = 44 race was recently described as a distinct species (Gehyra lazelli; Sistrom et al., 2009). The 2n = 40b variant may also utilize rocky outcrops as retreat sites (Kitchener et al., 1988; Sarre, 1998) and the 2n = 40a/38b variants predominately retreats beneath exfoliating bark and tree debris (Bustard, 1968; Henle, 1990; Moritz, 1992; Duckett & Stow, 2012). We sampled the most widely distributed and predominately tree dwelling 2n = 40a/38b variant which was found throughout central and eastern Australia. Because the 2n = 44 was sympatric with the 2n = 40a variant in some areas, we used morphological characteristics to avoid misidentification (Sistrom et al., 2009). Here, sequence data from the ND4 region of the mitochondrial genome were collected for each individual and confirmed the presence of a single species (P.E. Duckett, unpublished data). Our study sites were located throughout the arid and semi-arid zones of central and eastern Australia, which closely matched the complete distribution of G. variegata (2n = 40a/38b; King, 1979).

The arid zone was broadly categorized by habitat type and landscape features: arid woodlands, Mitchell grasslands, the channel country, gibber plains, sandy deserts and MacDonnell Ranges which we then referred as ‘landscapes’. Acacia aneura (mulga) and Acacia cambagei (stinking gidgee or wattle) dominates the woody vegetation in most of these landscapes, with the exception of the eastern woodlands which commonly comprised the genera Eucalyptus and Callitris. G. variegata was collected from each of these landscapes to evaluate variation in dispersal, assuming that dispersal characteristics would differ amongst landscapes. We could have evaluated the relationship of dispersal with environmental gradients, such as that represented by the climatic data or soil type. However, in arid Australia, distinctive and adjacent landscapes often have very similar climatic conditions and a single soil classification can be found in contrasting landscapes (e.g. sandy plains versus sand-dune formations). On a broader scale, understanding how landscape features influenced genetic structure had been a major goal of landscape genetics in the last decade (Storfer et al., 2010), and dispersal resistance was often categorized by broad landscape features and characteristics (Spear et al., 2010). Our sampling covered a large proportion of this species distribution and spanned much of the rainfall and temperature gradients experienced in arid Australia along with most of the soil types (ANRA, 2001). We therefore expected that our sampling captured much of the variability in dispersal that might be present throughout the distribution of G. variegata (2n = 40a/38b).

Estimates of dispersal were calculated from individuals collected from two large-scale sampling transects (c. 1700 km) designed to include the majority of landscapes inhabited by G. variegata (2n = 40a/38b; Fig. 1). Additionally, smaller replicate areas were sampled within each landscape allowing dispersal to be assessed at both larger and smaller scales (Fig. 1). Sampling effort within each replicate landscape was similar, yielded a varying number of individuals and encompassed a mean area of (mean 4900 km² ± 3500). The wide variation in the area from where individuals were sampled at different locations was largely due to the harsh environmental conditions within the Simpson Desert, where only a single transect was possible. Here, individuals were only located in a small geographical range c. 275 km² (Fig. 1).

Figure 1. The sampling design: landscape type can be identified by matching the colours and/or the numbers shown in the key. Within each landscape the number of samples (N), samples per km² (D), number of replicate sampling areas (R), individuals within each replicate (solid line) and individuals within each replicate for across catchment comparisons (broken line) are given. Localities where Gehyra variegata (2n = 40a/38b) was captured are marked by an open circle.

Between 2009 and 2010 we captured 635 individual adults by searching potential retreat sites across the predominately tree dwelling G. variegata chromosomal races (2n = 40a/38b; King, 1979), with very few exceptions, individuals were found in retreat sites associated with woody vegetation. The time period of collection coincided with the end of an El Niño-southern oscillation event that had created a long dry period; therefore, dispersal estimates are likely to be associated with these conditions. Adult geckos were classified using the minimum snout-vent length of a gravid female (Kitchener et al., 1988), or if male, by visual recognition of the obvious bulges, just distal to the cloaca, which contain their hemipenes. The location of each tree on which individuals were captured was recorded using a GPS. The distances between capture locations within each landscape ranged from 0 to 300 km, and prior to release a tail-tip tissue biopsy was taken and stored in 95% ethanol for analysis back in the laboratory.

Laboratory procedures

We extracted total DNA using a modified salting-out protocol (Sunnucks & Hales, 1996) and genotyped all 635 individuals at 16 microsatellite loci by Polymerase Chain Reaction (PCR; Table 1). The final reagent concentrations and the thermocycling conditions for PCR are outlined elsewhere (Hoehn & Sarre, 2006; Duckett & Stow, 2010). To test our data quality we reanalysed 5% of individuals at all loci and identical genotypes were obtained for each of these individuals in both runs.

Table 1. Observed (H_O) and expected (H_E) heterozygosity for 16 microsatellites across the 635 individual Gehyra variegata (2n = 40a/38b) samples

Locus	N	N A	Range	H E	H O	F IS
FIS values marked * showed significant deviation from Hardy–Weinberg Equilibrium after adjustment for multiple tests (α < 0.05).
GVVSN	620	45	168–400	0.877	0.742	0.155*
GV4B6	600	44	304–516	0.954	0.767	0.197*
GV92I	622	16	96–168	0.847	0.619	0.270*
GV1C10	620	38	292–480	0.943	0.827	0.123*
GVMJA	621	15	224–296	0.817	0.763	0.067*
GVYR7	612	25	272–488	0.901	0.779	0.136*
GV3C6	618	10	164–232	0.229	0.189	0.173*
GVVVF	607	29	316–532	0.888	0.713	0.197*
GVWD8	614	29	156–268	0.905	0.832	0.081*
GV4C9	608	38	204–504	0.911	0.775	0.150*
GVGVB	621	25	128–240	0.906	0.847	0.066*
GV3E10	615	36	208–364	0.950	0.777	0.183*
GVN6E	624	13	228–276	0.861	0.796	0.076*
GV17E	616	26	324–464	0.897	0.851	0.052*
GV56J	626	25	128–268	0.882	0.832	0.057*
GVT64	610	52	336–608	0.942	0.854	0.094*
Mean		29.15		0.857	0.748	0.128*

Summary genetic data

We calculated summary data including number of alleles (N_A), observed heterozygosity (H_O), and expected heterozygosity (H_E) using genalex v6.0 (Peakall & Smouse, 2006). Measurements of F_IS, linkage disequilibrium and the significance of any deviation from Hardy–Weinberg equilibrium (HWE) were calculated with fstat v2.9.3 with corrections for multiple tests (Goudet, 2001).

Characterizing dispersal by analysis of relatedness

We assessed the geographical distribution of pairwise genetic similarity for adult individuals within the replicate areas of each landscape by spatial autocorrelation using genalex v6.0 (Peakall & Smouse, 2006). The genalex software calculated a spatial autocorrelation coefficient (r) which we referred to as relatedness. Significantly higher relatedness at close geographical scales inferred that dispersal was limited at the spatial scales that were examined. Because ephemeral river systems were common features in some regions within the Australian arid zone, we assessed if rivers were associated with higher or lower levels of dispersal as these might potentially influence the precision of our estimates of average dispersal distance. We analysed the distribution of relatedness across- and along-river systems in the channel country and Mitchell grassland landscape. For all analysis we used the single population option with distance categories defined to maximize the number of pairwise comparisons, and where possible to allow for comparisons across landscapes at similar spatial scales. The 95% confidence intervals around mean relatedness (r) within each distance category were estimated by bootstrapping 9999 times, and the 95% confidence intervals around a random distribution (mean r = 0) was determined by 9999 permutations.

Calculating annual dispersal distance

Following Wright, the ‘neighbourhood’ was a unit of genetic structure that reflected the average distance between the natal and breeding site of the organism in question (Wright, 1943; Slatkin & Barton, 1989). Using a method based on the neighbourhood, we indirectly estimated annual dispersal distance by approximating the neighbourhood size (NS). NS can be estimated from the inverse of a regression slope between a multilocus estimator of individual pairwise genetic distances [F_ST/(1−F_ST)] regressed on geographical distance (Rousset, 1997, 2000). This relationship can be biased by very small spatial scales, thus we exclude comparisons where individuals occur at the same GPS point (same tree), and regress the logarithm of geographical distance against the ê estimator of genetic distance using genepop3.2a (Raymond & Rousset, 1995). The regression method were demonstrated to provide good levels of congruency between direct and indirect estimates of NS (Sumner et al., 2001), and remained robust to low rates of long distance dispersal, when IBD was measured in nonlinear habitats, when IBD was weak, or from higher sampling variance (Watts et al., 2007).

Neighbourhood size was equal to 4πDσ², where D was the population density and σ² was the variance of per-generation dispersal. This facilitated either the evaluation of density if dispersal is known (D = NS/4πσ²) or dispersal if density is known (σ² = NS/4πD). From this two dimensional estimate of dispersal area between parent and offspring we calculated a linear estimate of annual dispersal per individual by considering the species' generation length of 2 years (Bustard, 1969). In each replicate area within landscapes we estimated the effective density (D_E) by dividing the effective population (N_E) by the area occupied by those samples (length multiplied by the width of the furthest points). Various methods existed for calculating N_E from multilocus datasets which varied from short term methods to those that accounted for long-term variation (Leberg, 2005; Wang, 2005). Although coalescent-based approaches, which might consider longer term variation in allele frequencies were used with some success, incomplete sampling could confound results, and so these methods remained more applicable to systems where sampling was known to be complete (Slatkin, 2005). The benefit of using a contemporary estimate of N_E in this study ensured that dispersal estimates would be conservative and not confounded by historical gene-flow, which seemed appropriate when rare events over evolutionary timescales might be irrelevant to current ecological processes in arid Australia (Pinsky et al., 2010). Therefore, we adopted the sib-ship assignment method using a single sample of multilocus genotypes as implemented by colony v2.0.1.1 (Jones & Wang, 2010). In simulation studies this method had outperformed the heterozygote excess, linkage disequilibrium and temporal methods and did not require the assumption of random mating (Wang, 2009). For our analysis we used a full-likelihood approach with a long-run and medium-likelihood precision, no sib-ship prior and a standard genotyping error rate of 2%.

These estimates of dispersal can be potentially biased by the scale of analysis as genetic structure may change with scale (Ruckelshaus, 1998). We evaluated the effect of scale by applying this method to each replicate within each landscape, and also to each of the large-scale sampling transects. For the method to be applied there needed to be no significant difference in average dispersal estimates calculated at the different geographical scales.

Inferring the current and future species' distributions from SDM

Using available occurrence data for G. variegata (2n = 40a/38b), we modelled environmental suitability using maxent 3.3.1. The resulting models were utilized to predict areas of climatic suitability to 2030 and 2070 using four General Circulation Models with the A2 scenario. ‘See Appendix S1 in Supporting Information’ for a full explanation and justification of model building and modelling procedures.

Quantitatively assessing the future species' distributions

We designed R scripts (Appendices S2 and S3) to identify the proportion of the current distribution which could feasibly contribute to areas within the future distribution and the proportion of the future distribution reachable within the measured timescale. We estimated these proportions based on mean annual dispersal distance (± 1 SD) calculated across all landscapes, and using the minimum and maximum annual dispersal distance identified when comparing all landscape types. The three key components of the script involve dilation (which modelled dispersion), boundary detection and intersection of the presence-absence maps (Fig. 2). More information on these image processing and analysis techniques for binary images was available from Russ (1995) and James (1987) whose BASIC functions were adapted for the first two components.

Figure 2. The approach used to model dispersal processes begins with the presence-absence forms of current (M1) and future (M2) distributions of favourable climate conditions produced by the species distribution model. Dilating M1 with a dilation radius equal to (mean annual dispersal distance * (years between M1 and M2)) produces a map (D1) of maximum reachable area. The intersection between D1 and M2 identifies grid cells of the future distribution reachable from the current distribution (I2). Dilating I2 with the same dilation radius D2. The intersection between D2 and M1 identifies grid cells from the current distribution that can supply organisms to grid cells within the reachable future distribution.

Finally, we assessed how much of the predicted distribution populated by G. variegata (2n = 40a/38b) would overlap with agricultural and urbanized environments using land-use GIS data (Australian-Government, 2006). Future projections of land use in Australia are currently unavailable at sufficient quality for the present study, thus we assumed land use remains static until 2070.

Results

Summary statistics for the genetic data

The 16 primer pairs successfully amplified polymorphic loci with unambiguous alleles (Table 1) and revealed high levels of polymorphism, with N_A = 10–52, H_O = 0.189–0.854, H_E = 0.229–0.954. All loci showed a significant homozygote excess across the complete dataset. However, analysis at finer spatial scales (within 4 km²) rarely showed significant deviation from HWE. This indicated that the homozygote excess apparent when all samples are pooled is a Wahlund effect owing to genetic structure (Wahlund, 1928). There was no evidence for linkage disequilibrium, and our high rate of amplification success indicated that null alleles were absent or at a very low frequency.

Inferring dispersal patterns from genotypic structure within each landscape

From the spatial distribution of relatedness we inferred that the level of dispersal differed considerably among the landscapes (Fig. 3a–i). Dispersal in landscapes A–D was generally at shorter distances in comparison to those in landscapes E–I. We showed significantly higher levels of relatedness amongst individuals at close distances to each other (r = 0.02–0.09) within the sand dune (A), mulga woodlands (B), gibber plains (C), and eastern woodlands (D) landscapes. Relatedness did not differ from random expectations at any distance within the channel country (E–F), Mitchell grasslands (G–H), MacDonnell Ranges (I), or within and across catchments, thus suggesting higher levels of dispersal in these landscapes.

Figure 3. The spatial distribution of relatedness varies considerably among landscapes. Data were pooled across two replicate sampling areas (Fig. 1) to generate the spatial correlogram for each landscape. An exception was the sand dune habitat which consisted of samples from a single area. The solid line tracks relatedness, dashed lines represent the upper (U) and lower (L) 95% confidence interval around random expectations whereas bars around relatedness values show the 95% confidence determined by bootstrapping.

Annual dispersal distances

The annual dispersal distance estimated from NS in each landscape largely supported the levels of dispersal inferred from the spatial autocorrelation of relatedness. Annual dispersal distance was lower within the sand dune, mulga woodlands, and gibber plains in comparison to channel country, grasslands and MacDonnell Ranges (Table 2). Annual dispersal distance for the eastern woodlands was also estimated to be high, despite the significant genotypic structure at small spatial scales. The minimum and maximum annual dispersal distances were calculated from individuals sampled within the Simpson Desert (1.69 km year⁻¹) and the channel country (7.59 km year⁻¹) landscapes. The mean annual dispersal distance calculated from our large-scale transects fell within the range of landscape dispersal estimates and was not significantly different to the mean annual dispersal distance calculated across all landscapes (Table 2; t = −0.59, d.f. = 9, P (two-tailed) = 0.57). We therefore used the mean annual dispersal distance pooled across landscapes (4.85 km year⁻¹ ± 1.73) to predict the connectivity from the current distribution to the future distribution. This was the lower of the two estimates and was selected for a conservative assessment.

Table 2. The mean (± SD) of our neighbourhood size (NS), effective density (D_E) and dispersal rate estimates for each landscape and for each broad scale transect

Landscape	NS	D E	Dispersal rate (km year^−1)
Sand dune (Single sample area)	139	1.09	1.69
Mulga woodland	136 ± 75	0.05 ± 0.02	2.96 ± 0.68
Channel country across catchment	614 ± 145	0.24 ±  < 0.01	3.21 ± 0.91
Grassland across catchment	586 ± 493	0.04 ±  < 0.01	4.54 ± 1.20
Gibber plains	413 ± 155	0.03 ±  < 0.01	4.62 ± 0.74
Channel country within catchment	807 ± 752	0.03 ± 0.02	5.70 ± 2.67
Grasslands within catchment	1126 ± 89	0.03 ±  < 0.01	5.86 ± 0.80
Eastern woodland	251 ± 173	< 0.01 ±  < 0.01	6.23 ± 1.05
MacDonnell ranges	715 ± 318	0.01 ±  < 0.01	6.26 ± 0.18
Mean			4.85 ± 1.73
All samples
Transect 1	52	< 0.01	5.05
Transect 2	56	< 0.01	5.50
Mean			5.28 ± 0.32

Predicted distributions from species distribution models

The G. variegata (2n = 40a/38b) SDM AUC scores were high (0.873 ± 0.005), thus the presence records contained in the test sample of the data were predicted reliably between runs which provided strong support for the maxent model. This suggested sampling bias was not a concern and the variables selected were a good fit. The predicted current range (1.39 ± 0.15 million km⁻²) closely resembled the known current range for G. variegata (2n = 40a/38b; Fig. 4a; King, 1979; Cogger, 2000). The current range of G. variegata (2n = 40a/38b) is predicted to decline by 2030 (0.87 ± 0.25 million km⁻²), and further by 2070, (0.62 ± 0.24 million km⁻²; Fig. 4b,c), with a general shift in distribution towards the southeast and eastern areas of Australia. The area of suitability in the south-west region is predicted to increase, and most of this is currently not occupied by the species. Highly localized refuges are predicted to persist in central Australia.

Figure 4. Presence-absence maps for the predicted distributions of Gehyra variegata (2n = 40a/38b). Suitable areas are shown in dark grey and unsuitable areas in light grey. The following distributions are presented: a = the present; b = the year 2030; c = the year 2070; d = the areas that are capable of supplying individuals to the 2070 distribution; e = the area that can be reached by 2070 based on our genetic estimates of dispersal.

Quantitative assessment of the future distribution

Using mean annual dispersal ± 1 SD, we showed that G. variegata (2n = 40a/38b) could populate between 86% and 99% of its predicted future distribution that had favourable conditions, and this was achieved by dispersal of individuals originating from 59% to 83% of the current distribution. From this we inferred that 17–41% of the species' current distribution will not be able to keep up with its predicted range shift (Fig. 4d,e). Furthermore, the percentage of the predicted G. variegata (2n = 40a/38b) distribution that overlapped with areas occupied by agricultural and urbanized areas increased from 4.7% (current) to 11.2% (2070). Using maximum annual dispersal estimates resulted in complete connectivity between the current and future distributions. Using minimum annual dispersal resulted in the species populating 82% of the predicted future distribution and this was achieved from individuals occupying 48% of the predicted current distribution.

Discussion

Knowledge of the dispersal capacity of a species will help refine predictions made of the impact of future climate change. Dispersal is also amongst the most difficult traits to measure, with genetic approaches showing most promise at the landscape scale (Segelbacher et al., 2010). The levels of dispersal that we inferred using two different analytical approaches (spatial patterns of relatedness and Wright's NS), were largely consistent with each other. The annual dispersal distances we revealed seemed reasonable given that dispersal characteristics were known to vary between habitats in this species (Bustard, 1969; Moritz, 1992). Also the highest levels of dispersal we discovered in the MacDonnell Ranges were congruent with previous studies where little population structure across distances up to 100 km were measured (Moritz, 1992). Although recapture studies suggest much lower dispersal distances than characterized here, Moritz (1987) found rapid recolonization events occurred at larger spatial scales, and recapture methods are likely to strongly underestimate true dispersal distances (Shreeve, 1995; Pike et al., 2008).

By estimating average dispersal distances and the areas favourable for G. variegata (2n = 40a/38b) following climate change, we calculated that up to 41% of the current distribution would not contribute towards the colonization of areas with favourable climate by 2070. Furthermore up to 14% of those favourable areas cannot be populated because the extent of the range shift exceeds the distance the species are expected to cover within the timescales imposed by climate change. These estimates changed considerably when either maximum or minimum annual dispersal distances were applied to the model. This highlighted the value in obtaining realistic dispersal estimates across a species geographical range, and to consider how these contemporary rates are likely to reflect future dispersal rates under realistic climate change scenarios. Nonetheless, this approach can potentially take advantage of molecular and occurrence record datasets to provide dispersal estimates for use with assessment of the impact of future climate change.

Our genetic estimates of annual dispersal rates have to be based on past conditions and it is conceivable that dispersal characteristics might change according to the prevailing weather conditions. Therefore, when estimating our dispersal distance from NS we carefully considered how N_E was calculated. Methods were available that estimated contemporary or historical N_E (Leberg, 2005; Wang, 2005). We adopted a contemporary method because the G. variegata (2n = 40a/38b) we sampled were captured towards the end of an El Niño-southern oscillation event. Hence dispersal estimates would be representative of drier rather than wetter conditions. This is preferable because the Australian climate is largely predicted to experience increasing aridity in the near future (IPCC, 2007; Suppiah et al., 2007; Timbal & Jones, 2008).

Potentially, during future changes in distribution highly dispersive individuals may be forced to the leading edge of the range shift in each area. This new concept of spatial sorting has been demonstrated within the rapidly expanding populations of the invasive cane toad in Australia (Shine et al., 2011). We have no information on whether dispersal characteristics would be selected for in this study. However, the approach we adopted (Roussett's, 1997) measure the ‘variance’ of dispersal not ‘mean’ dispersal. Therefore, information from individuals dispersing greater distances than the mean value was incorporated. Furthermore, modelling changes in distribution using minimum, mean and maximum dispersal rates can help assess the influences of highly dispersive individuals. Under favourable conditions the colonization of new habitat may be dominated by the highest dispersers, justifying the use of the maximum dispersal rates in our model. However, favourable conditions seem unlikely in the near future due to the negative impacts of rising temperatures, increasing aridity and vegetation declines within inland Australia (Hughes, 2003; IPCC, 2007). Lower levels of dispersal under these conditions are supported by our lowest dispersal estimates occurring in our driest location, the sand dune landscape (Fig. 3). Thus, the mean annual dispersal distance provided us with a conservative and perhaps more realistic estimate than applying the maximum dispersal rate.

How global climate change will influence biotic interactions, which may subsequently increase or decrease dispersal rates, is unknown for most species (Callaway et al., 2004). Recent studies suggest when species are temporarily released from pathogenic or predation pressure dispersal rates can increase (Van Grunsven et al., 2007; Murrell & Barraquand, 2012). Further, the type and magnitude of competition for resources can also influence dispersal (McCarthy, 1997). Experiments manipulating competition for resources have shown that a reduction in nutrition for gravid lizards decreases the dispersal of their offspring (Massot & Clobert, 1995). Nonetheless, our mean estimate of dispersal was derived from observations collected across a large area, consisting of several distinct landscapes and presumably, varying biotic interactions. Therefore, our sampling strategy is likely to account for contemporary levels of variation in dispersal. Finally, behavioural plasticity or evolutionary in situ adaptation to climatic changes may negate the need to disperse altogether, for at least some individuals. In Australia, nesting lizards (Bassiana duperreyi) have adjusted nest depth and the timing of oviposition in response to rising temperatures, allowing them to temporarily remain in situ (Telemeco et al., 2009). However, long term persistence becomes less probable with increasing rates of climatic change, exemplified by the localized extinction of several lizard species in the last decade (Sinervo et al., 2010).

Our SDM projections may have been influenced by our assumption that the distribution of G. variegata (2n = 40a/38b) was in equilibrium with current climate. The influence of assuming equilibrium is shown by Elith et al. (2010), where they adjust models to better predict the distribution of cane toads in Australia. The cane toad is a vagile invasive species lacking any predator control, and has been spreading since its introduction. Correlative based predictions assuming equilibrium, combined with presence records that are unlikely to have adequately sampled future environmental conditions, were shown to provide inaccurate predictions. These same issues are less likely for the widely distributed and native G. variegata (2n = 40a/38b; Bustard, 1968; Arnold & Poinar, 2008). Currently, G. variegata (2n = 40a/38b) is almost certainly much closer to its distribution equilibrium, rather than a state of spread. Additionally, we sampled over an extensive geographical and climatic range, including some of the driest and hottest locations within Australia; therefore, a number of our presence records are likely to reflect future conditions. This was demonstrated with our extremely low ‘clamping’ estimates due to the overwhelming majority of projected regions falling within the range covered in the SDM training data (Appendix S1).

If our estimates of mean annual dispersal distance are representative of future dispersal rates, then our quantitative modelling approach predicts that rapid climate change will relocate and reduce the distribution of G. variegata (2n = 40a/38b). For those stranded in small areas of suitable climate, or areas with less than optimal climatic conditions, there may be detrimental consequences of isolation and small population sizes (Cushman, 2006). Although some may persist through environmental stochastic challenges for some time, population viability may be compromised in the long term by inbreeding and the random accumulation of deleterious alleles or loss of beneficial ones (Lande, 1993; Frankham, 1998; Frankham et al., 2004). These processes have been demonstrated to reduce fitness and increase extinction risk in a range of organisms including reptiles, birds, mammals and fish (Gilpin & Soulé, 1986; Hanski, 2011). Alternatively, the impact of climatic change in these regions may be buffered to some degree by the persistence of microhabitat, which may have enabled persistence during the climatic fluctuations of the Pleistocene (Duckett and Stow: In review Diversity and Distributions DDI-2012-0265). In addition, human land use may potentially fragment distributions and exacerbate the problems of climate change for G. variegata (2n = 40a/38b). In 2070 the distribution of G. variegata (2n = 40a/38b) is predicted to have a substantial portion overlapping with agricultural and urbanized areas (Australian Government, 2006). In Western Australia both G. variegata (2n = 40b) and the similar-sized Oedura reticulata suffered reduced genetic diversity and increased genetic structure where their natural habitat was fragmented by agriculture (Hoehn et al., 2007). On the other hand, G. variegata (2n = 40a) are often located in human impacted environments (Cogger, 2000), therefore it remains unclear the degree to which land use practices may fragment the distribution of G. variegata (2n = 40a/38b).

Coupling both SDMs and genetic estimates of dispersal will assist with predicting a species' response to climatic change. As we have shown, this approach incorporates both correlative and mechanistic variables to quantitatively identify areas of the current distribution of G. variegata (2n = 40a/38b) that are likely to become isolated and areas of the predicted future distributions that cannot be colonized. Further advances may come from simultaneously incorporating other variables into models such as multispecies interactions or the changes in vegetative cover through time (Hampe, 2004). Nonetheless, we show that available molecular and occurrence record datasets may be amenable to assess and help prioritize those species which will be most vulnerable to the impacts of climate change.

Acknowledgements

This work was supported by Macquarie University and the Wildlife Preservation Society of Australia Limited. Lizards were handled in accordance with Macquarie University and South Australian Animal Ethics committee recommendations. Tissue collection was licensed by the New South Wales National Parks and Wildlife Service (S12353), Northern Territory Parks and Wildlife Commission (33155), Queensland Environment Protection Agency (WITK04616407) and South Australia Department of the Environment and Natural Resources (Z25499-4). We thank Macquarie University volunteers for their assistance during field collections and Cabrelli for assistance in basic SDM techniques.

Biosketches

Paul Edward Duckett has research interests in marine ecology and biogeography with reference to the potential impacts of climate change. He is primarily interested in historical refugia and whether they will be beneficial for the future of global biodiversity.

Peter Wilson has research interests in macroecology and biogeography including the application of species distribution modelling and potential impacts of climate change. He is primarily interested in the ecology and functional morphology of microchiropteran bats, but has also contributed to papers on invertebrates, reptiles, sharks and plants.

Adam Stow is a A/Prof at Macquarie University, Sydney. With research interests in molecular ecology and conservation biology his primary focus is assessing gene flow, dispersal and genetic variability in human impacted environments.

Author contributions: P.E.D and A.S. conceived the ideas; P.E.D. collected the data; P.E.D and P.W. analysed the data; and P.E.D and A.S. led the writing.