1 INTRODUCTION

Investigating the effects of historical climatic changes on species distributions, and genetic diversity and structure, can improve our understanding of species demography and help us to predict species' responses to climate changes (Yannic, Pellissier, Ortego, et al., 2014). Because adaptation of large mammals to rapid environmental change is limited and slow (Hetem et al., 2014), their immediate response to climate change is generally a range shift (Chen et al., 2011). The reorganization of species' distributions reshapes the connectivity of populations and, as a consequence, their population genetic structure (hereafter genetic structure), especially in the context of major glacial cycles (Hewitt, 1999; Pellissier et al., 2014). For example, the range of cold-adapted species from higher latitudes, such as the caribou (Rangifer tarandus), shifted southward during glaciations, where they were isolated in multiple refugia, resulting in distinctive signals in their genetic structure (Taylor et al., 2021; Yannic, Pellissier, Le Corre, et al., 2014). Moreover, range shifts depend on the ability of species to disperse to suitable areas over time (Williams & Blois, 2018); consequently, the study of genetic structure can indicate a species' ability to shift its range in interaction with landscape structure (He et al., 2013). Therefore, by combining species habitat preferences over dynamic landscapes with model simulations and genetic data, we can better understand species' past dynamics and identify their capacity to respond to climate change.

As a result of climate or productivity, animal species generally use only a fraction of the landscape (Peterson et al., 1999); when combined with direct barriers to dispersal, such as topography, ice or rivers, these factors determine the connectivity among populations and their genetic differentiation (Frankham et al., 2002). Typically, species in more complex landscapes, such as mountain ranges, show high levels of genetic differentiation between populations (Badgley, 2010). Genetic differentiation can be the result of recent or more ancient habitat fragmentation, such as that associated with distinct glacial refugia (Hewitt, 2004) or the formation of river valleys during geomorphological processes (Hazzi et al., 2018). The combination of landscape dynamics and species dispersal abilities is expected to have modulated the response of species to past environmental change (Hewitt, 1999). For example, genetic analysis has shown that the mountain hare (Lepus timidus) probably recolonized Ireland over a land bridge from southern France (and not from Scotland) following the last glaciation (Hamill et al., 2006), whereas the pygmy shrew (Sorex minutus) was only able to colonize Ireland with anthropogenic translocations from Britain (McDevitt et al., 2011). While historical species dynamics have been inferred from genetic data since the emergence of the field of phylogeography (Hewitt, 1999, 2004), the development of palaeo-environmental reconstructions and species distribution models (Elith & Leathwick, 2009) and process-based models (Landguth & Cushman, 2010) offers new possibilities for insights into species range dynamics (Yannic et al., 2020).

Combining habitat suitability models (HSMs) with process-based simulations enables the reconstruction of species dynamics through geological time and helps us to infer species demographic parameters (Arenas et al., 2012; Yannic, Pellissier, Ortego, et al., 2014), thus representing a central tool for geogenomics. HSMs have been widely used to predict species geographic distributions (Guisan et al., 2017) and to estimate the location of refugia during the glaciations, often in combination with genetic data (Svenning et al., 2011). To be compared with genetic data, however, HSMs must be combined with models capturing actual processes of dispersal and genetic differentiation (Epperson et al., 2010). These process-based models use a bottom-up approach to simulate populations, following rules defined a priori (e.g. for dispersal) to scrutinize emerging patterns (e.g. in genetic structure; Grimm et al., 2005). HSMs have been used as input in process-based models to explicitly simulate the processes of individual dispersal and genetic drift that shape population isolation (Brown & Knowles, 2012). Since individual-based models are computationally demanding (Currat et al., 2019), another class of more parsimonious spatially explicit simulations has been proposed to investigate and validate multiple hypotheses regarding the role of dynamic landscapes in shaping inter- and intraspecific diversity (Leprieur et al., 2016; Yannic et al., 2020). By combining process-based models with species HSMs and exploring biological parameters, simulations can help us to investigate the effects of species dispersal ability (Knowles & Alvarado-Serrano, 2010) and landscape features on genetic structure (Spear et al., 2010). Models can potentially simulate the connectivity of populations over time based on climate reconstructions, and the comparison of these simulations with empirical genetic data can provide insight into past and future species distribution range dynamics (Yannic, Pellissier, Ortego, et al., 2014).

The Northern chamois (Rupicapra rupicapra) is an alpine ungulate native to Europe and the Near East (Corlatti et al., 2011). It has previously been used to study the impacts of habitat connectivity on population genetics (Crestanello et al., 2009; Rodríguez et al., 2008), as its preference for steep, rugged areas (Corlatti et al., 2022) means that its preferred habitat is naturally fragmented (Buzan et al., 2013; Šprem & Buzan, 2016). Also, gene flow is expected to be limited among some populations (Buzan et al., 2013) because wide valley floors may act as natural barriers to dispersal (Soglia et al., 2010). Therefore, simulating population connectivity based on habitat suitability and comparing these models with observed genetic structure could be informative regarding the effect of landscape features on chamois dispersal (Shirk et al., 2012).

In this study, we aimed to provide insight into the drivers of the contemporary distribution and genetic structure of Northern chamois in the Alps through a combination of HSMs, process-based models to simulate cumulative functional connectivity through time, and a comprehensive newly generated genomic data set (>20 k SNPs). We analysed the contemporary genetic structure to determine the geographic distribution of genetic clusters. Next, we simulated the habitat suitability and connectivity from the last glaciation (20,000 years before present, bp) to the present to track functional connectivity. Cumulative distances between clusters that integrate the evolution of connectivity through time were used to generate a divergence matrix, which was compared with the observed genetic structure. We estimated dispersal abilities of the Northern chamois for different landscape feature scenarios by comparing simulations with the genomic data. We investigated the impact of landscape features on population structure in comparison to different dispersal abilities. Based on the expectation that cumulative distances serve as an approximation for population differentiation, we made the following predictions: (1) life history traits such as dispersal can be estimated realistically from simulations; (2) simulations can estimate the relative impacts of landscape features and dispersal on population genetic structure, that is, whether limited dispersal alone is sufficient to explain the observed genetic structure.

2 MATERIALS AND METHODS

2.3 Modelling habitat suitability

2.3.1 Occurrences and pseudo-absences

The entire study area encompasses the Alps, the Massif Central, the Jura mountains, the Vosges, the Black Forest and the northernmost part of the Dinaric Alps, spanning from −6° to 19° longitude and from 41° to 51° latitude (see Figure 1 for a general workflow). We collected data on the current species distribution from various citizen science projects throughout the study area, and from hunting records from Slovenia, at a 1 km resolution from the last 15 years (see Table S1). Most of the records of the citizen science data have been validated by experts. We used four different occurrence/pseudo-absence sampling strategies (see also Supporting Information) after experimental testing for more combinations and a wider range of parameters, repeating each of them 10 times (Barbet-Massin et al., 2012). We sampled 10,000 occurrences randomly and 10,000 pseudo-absences either with a slight, middle or strong density dependence with the ‘spatstat’ R-package v1.61–0 (Baddeley & Turner, 2005). Additionally, we applied proportional occurrence sampling combined with random pseudo-absence sampling to account for regions with higher densities of occurrences (Phillips et al., 2009). Details are provided on p. 12 of the Supporting Information.

2.3.2 Environmental variables

After preliminarily testing a set of seven climatic variables that potentially restrict habitat suitability for chamois (cf. Corlatti et al., 2022) for multicollinearity (Guisan et al., 2017), we chose two climatic variables, that is, mean annual precipitation and mean annual 2-m air temperature at 0.008333° resolution from CHELSA-TraCE21k (Karger et al., 2017, 2018) to model habitat suitability through time to the glaciation 20,000 years bp (see section below). To test the extent to which chamois rely on steep terrain, we first modelled the distribution without any terrain variables (only the two climatic variables) and then with four topographic variables derived from the Copernicus Land Monitoring Service (2016): mean slope, slope heterogeneity, a binary classification of suitable slopes for the Northern chamois (20–40°; Bačkor, 2010; Papaioannou et al., 2015) and the number of suitable cells at 100 m resolution within ~900 m grid cells. The environmental predictors showed a correlation of r < 0.7 and a variance inflation factor of <3, values considered acceptable for including all parameters in the same model (Guisan et al., 2017; Zuur et al., 2010).

We used data from Seguinot et al. (2018) for the predicted glaciated areas in the Alps, Jura, Vosges and Black Forest until the glaciation. Additionally, we modelled the glaciers from the Massif Central, estimating the equilibrium line elevation based on the mean annual temperature (Linsbauer et al., 2013) from CHELSA-TraCE21k, and we calculated the area proportional to the glacial extent during the glaciation (Van Vliet-Lanoë et al., 1991).

2.3.3 Habitat suitability

To model habitat suitability, we used four statistical modelling approaches in R v.3.6.1 (R Core Team, 2019), following Yannic et al. (2020) for model creation, validation and prediction at a resolution of 0.0083° back in time until the glaciation in 100-year time steps. We validated the hindcasted habitat suitability with a comprehensive fossil collection using the Boyce index (Boyce et al., 2002). The fossil collection included data provided by C. Callou, personal communications, and the IANUS data portal (Heinrich et al., 2016), and also from literature research (Table S2). A detailed description of the habitat suitability modelling is provided on p. 13 of the Supporting Information.

3 RESULTS

3.1 Geographic genomic structure

Our SNP analyses showed that the European Alps host two divergent genetic clades of Northern chamois. The eastern part of the Alps including the Dinaric Alps contain one large genetic cluster, whereas multiple clusters occur in the south-western Alps (Figure 2a). The first two axes of the DAPC exhibit a v-shape (Figure 2c), with the chamois from the central and eastern Alps (clusters 1 and 2 in Figure 3) being close to each other and showing some overlap. Consistently, the phylogenetic tree (Figure 3c) showed two main clades, dividing the western and eastern Alps, each with two clusters. We found signals of significant isolation by distance across the Alps according to the Mantel test correlation between pairwise genetic distances and connectivity distance among individuals (Mantel's r: 0.338; p < 0.01). The Admixture plot indicates admixture at the contact zone between different clusters, for example, around the upper Rhone or Adige valleys, but the overall genetic structure was strong in both the Admixture (Figure 2a,b) and multivariate analyses (Figure 3 and Figure S7).

3.4 Simulated connectivity over time

Our simulation predicted that chamois recolonized the Alps mainly from 15,000 years bp onward (Figure 4). The predicted suitable habitat at the south-western and south-eastern part of the Alps during the glaciation was substantially larger than the small and scattered areas at the northern edge. Some fossil records lay outside the focus area (European Alps) and outside of other mountain ranges, but most of them were located in proximity to predicted suitable habitat (Figure 4 and Figure S10). The largest uncolonized area at 10,000 years bp was predicted in the region of the Adige valley, Italy; otherwise, most of the Alps was colonized by then. This predicted long-lasting separation of the populations resulted in the longer separation of clusters east (1) and west (2) of the Adige valley (Figure 3d). The HSM used for the best simulation of the divergence performed well for the present distribution (AUC = 0.90; TSS = 0.67 and Kappa = 0.67; see also Table S3) and less well for the validation of the hindcasted distribution using fossils (Boyce index including buffer: 0.34; Table S4).

4 DISCUSSION

Modelling species' responses to palaeo-environmental changes can help identify the drivers of contemporary distribution and genetic structure (He et al., 2013), and the important features determining spatial and temporal dynamics. The genomic data of 449 Northern chamois samples across the European Alps suggested two main clades: one located in the south-western Alps and one in the eastern Alps. Generally, the model fit based on simulations decreased with increasing dispersal distances, indicating a low dispersal ability of Northern chamois. Additionally, medium to large rivers or valley bottoms acted as important dispersal barriers. Low dispersal ability and habitat suitability alone were not sufficient to depict the observed genetic structure. We showed that the past distribution of species can be reconstructed using dynamic HSMs based on geological and palaeo-environmental data and process-based models. Our approach suggests that combining genomics and palaeo-environmental reconstructions, including data from geology, glaciology or climatology, can elucidate the evolution of intraspecific diversity in complex environments.

Modelling overall population connectivity across dynamic landscapes and comparing the models with genomic data has been shown to reveal the effects of landscape features on genetic structure (Jenkins et al., 2016; Landguth & Cushman, 2010). Our simulation detailed the importance of the complex topographic terrain of the Alps, but also past climate dynamics, in shaping the current genetic differentiation of the chamois. The simulations support previous findings by Soglia et al. (2010) and Zemanová et al. (2015) reporting that large rivers or valley bottoms act as dispersal barriers, although individual chamois can cross rivers (Safner et al., 2019) and on some occasions can even swim in the sea (Kavčić et al., 2020). The barrier to dispersal likely increases with the size and the flow of the river (Safner et al., 2019; Soglia et al., 2010). Rivers can act as natural dispersal barriers for other ungulates such as the bighorn sheep (Ovis canadensis; Deakin et al., 2020) and roe deer (Capreolus capreolus; Coulon et al., 2006). The comparison between the fit of the simulated divergence with and without the terrain slope variable to the observed genetic structure and the validation with fossil records of the different HSM scenarios both confirm that chamois prefer steep terrain, as is commonly accepted (Aellen et al., 1995). It apparently allows them to escape from predators (Bačkor, 2010; Papaioannou et al., 2015) and to avoid competition with other ungulates (Anderwald et al., 2016). Even small steep areas may have been sufficient for the chamois to thrive, since fossil records have been found outside the main mountain ranges, for example, from Belgium (Couturier, 1938) or south-west of the Massif Central (C. Callou, personal communications). Taken together, these results indicate that the comparison of different landscape scenarios in dynamic simulation models can demonstrate the use of complex mountain landscapes by animals and how they shape differentiation among populations.

Dispersal is known to be a key trait influencing species responses to climate change (Schloss et al., 2012). In line with this, we observed that dispersal was the most important parameter determining the fit of the simulation to the genomic data. The best-performing simulations for the Northern chamois suggested a median effective dispersal of 5.5 km/100 years, which is at the lower end of the range observed for chamois in the Alps today (Loison et al., 2008). The better fit of the simulations with a Weibull shape 1 for the dispersal ability distribution reflects the occurrence of many individuals with short dispersal distances and a few with long dispersal, which could reflect philopatric females and male-biased dispersal (Loison et al., 2008), with very few long-distance migrants (Buzan et al., 2013). In fact, in exceptional cases, chamois may disperse up to 60 km (Clarke, 1986), which may reflect the need of individuals to find suitable habitats. Long-dispersing red fox (Vulpes vulpes) individuals, for example, quickly track newly available habitat towards the tundra in response to climate change (Colson et al., 2017; Norén et al., 2015). The simulated dispersal ability was also below that reported for bighorn sheep (Deakin et al., 2020). Since the model only accounted for dispersal resulting from measures in gene flow, it simulated the effective dispersal rate, which is almost always smaller than the real dispersal rate (Kobayashi et al., 2008). Thus, the actual dispersal ability is probably underestimated by our analyses, reducing the predicted habitat that a species facing climate change could use, and affecting management and conservation strategies (Luo et al., 2015).

Our genetic analysis revealed that Northern chamois in the Alps fall into two main groups in the south-west and south-east (Figures 2 and 3; see also Rodríguez et al., 2008). These two lineages probably represent two sources of recolonization of the Alps after the last glaciation, originating from populations isolated from each other during that time. A similar pattern has been shown for red deer (Cervus elaphus; Doan et al., 2022) and the Valais shrew (Sorex antinorii; Yannic et al., 2012). Here, the contact zone of the two chamois lineages lies around the Rhone valley in Switzerland and corresponds to the contact zone found in some alpine plant species that also underwent a range expansion after the last glaciation (Jay et al., 2012), although this appears to be located further west than that of red deer (Doan et al., 2022). However, in contrast to the red deer range expansion following the last glaciation, which was dominated by the north-eastward expansion of populations from south-western refugia (Doan et al., 2022), the recolonization by chamois was likely dominated by the eastern lineage, which expanded west over large parts of the Alps. Another possible contact zone is the Adige Valley; in fact, this valley has been shown to be a barrier during the last glaciation for several alpine mammals, including chamois, but also mountain hare and red deer (Vernesi et al., 2016), separating populations into genetically differentiated clusters east and west of it. The divergence found in the simulation is less pronounced in the genomic data but still visible, for example, as a contact zone with admixed individuals (Figure 2a). The legacy of a long-lasting (>2 million year) separation between the two lineages (Rodríguez et al., 2008) probably overrides the patterns in the genetic structure since the last glaciation to some extent. Additional analysis, involving ancient DNA for example, could provide further insights in the post-glacial history of the chamois. By coupling genomics and simulations, our analyses recovered the likely colonization dynamics of the chamois in the Alps since the last glaciation.

Computational demands limit many modelling approaches, for example, with respect to their resolution (Brown & Knowles, 2012). Although we used a model of simplified evolutionary processes (our counter of divergence was defined entirely by functional connectivity through time), we had to aggregate habitat suitability layers to reduce storage volume and speed up the calculations. The resulting grid cells were about 31–43 times larger than the average individual home range size of the Northern chamois (Nesti et al., 2010). Thus, information on landscape features and connectivity was lost during aggregation, including small suitable habitat patches, resulting in somewhat similar HSM scenarios. This could explain why the topographical variables only marginally affected the model fit. Increasing the spatio-temporal resolution of the gen3sis simulations while simulating longer time spans, for example, until the separation of the chamois lineages, could help to disentangle the drivers of the contemporary population structure. Furthermore, explicitly simulating the action of neutral evolutionary forces (mutation, drift and gene flow) could improve the fit between simulations and genomic data, but would require a reduction in landscape complexity. Other models, such as Splatche3 (Currat et al., 2019) or Nemo (Guillaume & Rougemont, 2006), make fewer simplifying assumptions about evolutionary processes, but are less capable of exploring complex landscape features and dispersal abilities. General limitations of the underlying HSMs of the simulations have already been discussed extensively in the literature (e.g. Elith & Leathwick, 2009), including the issues that arise when hindcasting over millennia (e.g. Svenning et al., 2011). Species interactions may have changed or evolved during the modelled time period, as human overexploitation after the Neolithic at lower elevations likely contributed to the range contraction of the Northern chamois toward alpine areas (Baumann et al., 2005). Thus, part of the ecological niche occupied during or immediately after the glaciation may now be abandoned as unsuitable and not captured when building HSMs based on the contemporary distribution.

5 CONCLUSIONS

Coupling dynamic habitat suitability modelling through space and time, in geologically complex terrain and considering past climate changes, with empirical population genomics improves our understanding of the formation of diversity in complex mountain landscapes. In our case study, a short dispersal distance and specific landscape connectivity dynamics shaped the intraspecific diversity of Northern chamois within its mountain range in the European Alps. Our study demonstrates that such an interdisciplinary and integrative biogeographic approach, linking palaeo-environmental modelling, biological processes (e.g. dispersal) and genomic data, can provide insight into the dynamics of intraspecific diversity. Beyond our case study of one ungulate's response to Quaternary climate change, our approach could be used to study the role of tectonic, geomorphological and palaeoclimatic dynamics in forming intra- and interspecific diversity in mountainous areas (Salles et al., 2019). In the context of far-reaching environmental changes, this framework could be extended to predictions about the future of biodiversity not only in order to inform conservation and population management programs, but also into the past to understand the origin of biodiversity hotspots in mountains.

CONFLICT OF INTEREST

None declared.

BIOSKETCH

Flurin Leugger is interested in how landscape history shapes today's biodiversity and genetic patterns. He and the other authors collaborate in monitoring biodiversity, developing models to understand the drivers of biodiversity patterns, and forecasting ecosystem responses to global changes (see https://ele.ethz.ch).

Author contributions: GY and LP conceived the ideas and supervised FL. EB, LC, BC, BR, HCH, NS and RV contributed to the sample collection. DR, NCGG, NT and ST performed laboratory work. TB and GY performed the bioinformatic analysis. DNK provided the climatic data. FL compiled the occurrence data, performed the models and simulations, and ran the analysis. FL, GY and LP wrote the first version of the manuscript with the support of all co-authors.