Introduction

Population dispersal and vicariance has become the subject of an increasing number of phylogeographic studies (Chan et al. 2011; Chavez et al. 2014; Valente et al. 2014). Integrating the molecular phylogenies of extant taxa with topographic factors and known climatic oscillations contributes to our understanding of the evolutionary processes that shape genetic patterns, historic colonization routes, and species evolution (Avise 2009; Chavez et al. 2014). Environmental change and climatic oscillations drive species range shifts along geographic gradients, in terms of both latitude and altitude, and have been linked to evolutionary events (Epps et al. 2006; Gorelick 2008; Angert et al. 2011; Liang et al. 2017).

The succession of warm and cold periods during the Pleistocene (starting 2.5 million years ago) has greatly contributed to intra-specific diversification, especially arising from populations persisting in allopatric refugia (Avise 2007; Husemann et al. 2014). Refugia represent areas in which a population can survive during unfavorable environmental conditions (in this case glacial periods). During the cold episodes of the Pleistocene, populations became isolated in refugia as their habitats fragmented (Bennett and Provan 2008; Sommer and Zachos 2009; Stewart et al. 2010). This scenario is particularly relevant to mountainous terrain, where habitat fragmentation gives rise to multiple refugia in valleys, greatly facilitating species diversification (Potts 2013).

Taiwan features a tropical to sub-tropical climate. The island has six steep mountain ranges, encompassing 268 peaks > 3000 m, deep gorges, and wide valleys, together producing diverse altitudinal climate zones and vegetation types at a very fine topographical scale. These features render Taiwan an ideal locale to study refugia and biodiversification. Significant climatic oscillations have occurred in Taiwan throughout its geological history. Alpine cirque glaciers have been identified as major features of the high mountains in northern Taiwan (Böse 2004; Böse and Hebenstreit 2011). Moreover, its low- and mid-altitude areas display evidence of a shift of highland forests to lower elevations (Böse 2004). Vertical distributions of vegetation zones, ranging from cryoflora (alpine tundra and subalpine coniferous woodland) to sub-mountainous evergreen broadleaf forest (tropical rain forest), shape faunal distribution patterns (Böse 2000; Li et al. 2013). Dramatic climate change during the transition of interglacial periods to glacial periods can result in niche fragmentation and produce barriers potentially restricting species dispersal. The effects of climatic oscillations on the genetic structure and demographic history of taxa such as amphibians, freshwater fishes, and small mammals during Pleistocene glaciations have been described for Taiwan (Yuan et al. 2006; Lin et al. 2008, 2012; Oshida et al. 2011). Large herbivores like ruminants are particularly suitable for studying how mammals respond to changes in terms of adaptive strategies because they are highly sensitive to vegetation modification during climate changes (DeMiguel et al. 2010; Strani et al. 2018). However, the ramifications of climate fluctuations for large herbivores in Taiwan have remained unexplored.

Formosan sambar deer (Rusa unicolor swinhoii) is a large herbivore endemic to Taiwan that is found in mountainous areas (1500 to 3500 m). It occurs in both forested and grassland habitats, and areas of rugged terrain (less than 60° slope). It exhibits altitudinal migratory behavior, preferring lower elevations during colder weather (Yen et al. 2013, 2019). Based on its island-wide distribution and migratory behavior, the dispersal history of the Formosan sambar deer represents an ideal model for examining the impact of Pleistocene climatic oscillations on a large herbivore.

Here, we use mitochondrial control region sequences to examine the genetic patterns, lineage divergence, and demographic history of the Formosan sambar deer. We link our findings to climate change during the Pleistocene and hypothesize a possible scenario for how topology and climatic oscillations have influenced sambar dispersal in Taiwan.

Materials and methods

Sample collection

We collected 877 fecal samples from 25 sampling locations in mountainous areas throughout Taiwan. Samples from these sampling sites were divided into nine subpopulations according to geographic location and connectivity, i.e., sampling locations located in adjacent mountain ranges without obvious geographic barriers to dispersal were classified as a subpopulation. Fecal samples were preserved in 95% EtOH and stored at − 20 °C. All sampling protocols were approved by the administration of Taroko National Park, the Forestry Bureau of Taiwan, and the Institutional Animal Care and Use Committee of National Taiwan Normal University (approval no. 101004), and were performed in accordance with relevant guidelines and regulations.

Fecal DNA extraction, PCR amplification, and mitochondrial control region sequencing

Genomic DNA was extracted from fecal samples using the FavorPrep Stool DNA Isolation Mini Kit (Favorgen Biotech Corp., Pingtung, Taiwan). The full-length mitochondrial control region (D-loop) was amplified and sequenced using polymerase chain reaction (PCR) primers designed according to the complete mitochondrial DNA sequence of Formosan sambar deer (GenBank accession number NC_008414): forward primer D-loop F1, 5′-GAAAAGGAGAGCAACCAACC-3′, and reverse primer D-loop R1, 5′-GGCTGGGACCAAACCTGTATG-3′. The highly variable tandem repeats (5′-TATAATAGTACATAAAATTAATGTATTAGGACATATCATG-3′) and cytosine-thymine repeats (5′-ATATTTCCCCCCCCTCCATTAATTTTTCCCCC-3′) within the control region may affect sequencing accuracy. To avoid sequencing errors arising from these regions, we used a nested primer pair to acquire internal sequences: D-loop F2, 5′-TGCCCCATGCTTATAAGC-3′; D-loop R2, 5′-CATTGATGAGTCCAAGCATA-3′. PCR was performed using the Advantage 2 PCR enzyme system (TAKARA Bio, Inc., Shiga, Japan). Thermal cycling conditions were 94 °C for 5 min, 39 cycles at 94° C for 45 s, 60 °C for 30 s, and 68 °C for 80 s, with a final extension at 68 °C for 10 min. The PCR products were purified with the GenepHlow Gel/PCR Kit (Geneaid Biotech Ltd., New Taipei City, Taiwan), and sequenced with an Applied Biosystems 3730 DNA Analyzer (Applied Biosystems, Inc., Foster City, CA). Partial control region sequences were obtained by merging sequence fragments in the EditSeq software (DNASTAR, Inc., Madison, WI).

Haplotype network analysis, Bayesian phylogenetic tree, and divergence time estimation

The highly variable tandem repeats and cytosine-thymine repeats within the control region were excluded from phylogenetic analysis, leaving 889 base pairs (bp) for phylogenetic analysis. Out of 877 fecal samples, a total of 454 bp sequences were successfully acquired, which have been deposited into the National Center for Biotechnology Information (NCBI) GenBank with accession numbers KU531437–KU531440, KU531442, KU531443, and KU531445–KU531456. Multiple sequence alignment was executed in the MegAlign software (DNASTAR Inc.) to identify haplotypes. A haplotype network with 95% probability of parsimony connections was constructed in TCS v1.21 software (Clement et al. 2000) to clarify the relationship among haplotypes. In addition, control region sequences from Rusa unicolor hainana (GQ304777), Muntiacus reevesi (NC_004069), and Muntiacus crinifrons (NC_004577) were obtained from NCBI GenBank. A phylogenetic tree of haplotypes using Bayesian inference was constructed in the BEAST software package v2.3.0 (Bouckaert et al. 2014) using the HKY + I model, which was selected as the best-fitting model in jModelTest v2.1.5 (Bouckaert et al. 2014), and four times Markov chain Monte Carlo (MCMC) sampling was executed for 100 million generations and sampled every 1000 generations. A maximum clade credibility tree was generated after the first 10% of tree samples had been discarded as burn-in using TreeAnnotator v2.3.0 (Bouckaert et al. 2014). Divergence time was estimating by applying a Bayesian strict clock model with a clock rate assuming 0.04 substitutions/million years (Randi et al. 1998) and a calibrated Yule tree prior (Drummond and Suchard 2010; Heled and Drummond 2011). To calibrate divergence times on our phylogenetic tree against the fossil record, we specified a normal prior distribution of 8 ± 1 million years ago (MYA) for the node of the Cervinae and Muntiacinae subfamily (Dong et al. 2004; Gilbert et al. 2006). The tree was visualized and edited in FigTree v1.4.2 (Rambaut 2014). Values for effective sample size (> 200) for all parameter estimates were verified in the Tracer v1.6 software (Rambaut et al. 2013).

Analyses of population structure

A spatial analysis of molecular variance (SAMOVA) was conducted using SAMOVA 1.0 (Dupanloup et al. 2002) to define groups of subpopulations that are genetically homogeneous and maximally differentiated. SAMOVA was conducted with the number of groups (K) ranging from 2 to 8. To verify output consistency, we ran the analysis five times for each K value with 1000 independent iterations, starting from 100 random initial conditions. We determined optimal K as that having the greatest percentage of variation and being most significant. To investigate genetic diversity, the DnaSP v 6.12 software (Librado and Rozas 2009) was used to calculate haplotype diversity (h) and nucleotide diversity (π) in Formosan sambar deer. In addition, pairwise genetic differentiation (FST) and genetic distance (Dxy) among our a priori nine subpopulations were assessed in the Arlequin v3.5.1.2 (Excoffier and Lischer 2010) and MEGAXI (Tamura et al. 2021) software, respectively.

Phylogeographic and historical demographic analyses

To investigate the history of divergence among SAMOVA-identified regions (NSP, SSP, NTR, STR, NDD and CSR; see “Results” and Table 1 for details), we tested 10 scenarios by means of approximate Bayesian computation (ABC) in the DIYABC v2.1 software (Cornuet et al. 2014). We generated CSR-origin (scenarios 1–5) and SSP-origin (scenarios 6–10) scenarios to test the direction of migration and the origin of Formosan sambar. Since the NDD, NTR, and STR regions possessed haplotypes from both genetic clades, our tests also explored different divergence orders for these subpopulations and the existence of secondary contact events. Scenarios 1 and 2 assume that the NDD subpopulation was the first to split from the CSR subpopulation. A secondary contact then occurred between the CSR and NDD subpopulations to give rise to the NTR + STR subpopulation. Scenarios 3 and 4 assume that the earliest divergence occurred between the CSR and STR subpopulations, followed by secondary contact between them. The timing of the split of the SSP and NSP or STR and NTR subpopulations differs within these two pairs of scenarios (scenario 1 versus 2 or scenario 3 versus 4, respectively). Scenarios 6 and 7 predict different origins of the CSR, i.e., from the SSP (scenario 6) or NDD (scenario 7) subpopulations. Both scenarios 8 and 9 allowed us to test if NTR initially split from NSP, but with the order of SSP diverging from NSP and STR diverging from NTR differing. Scenarios 5 and 10 tested the hypothesis that no secondary contact had occurred. Neutrality tests, including Tajima’s D and Fu’s Fs tests, were performed in the Arlequin v3.5.1.2 software (Excoffier and Lischer 2010). Mismatch distributions were calculated and depicted using DnaSP v 6.12. A Bayesian skyline plot (BSP) was generated in BEAST v2.3.0 (Bouckaert et al. 2014) to date changes in population size over time. For that analysis, we applied estimated clock rates from our Bayesian phylogenetic tree using fossil calibrations. Under a coalescent Bayesian skyline model, four MCMC iterations were executed for 300 million generations and then sampled every 1000 generations, followed by removal of 10% of samples as burn-in. BSP reconstructions were visualized in the Tracer v1.6 software (Rambaut et al. 2013).

Table 1 SAMOVA on nine Formosan sambar deer subpopulations
Full size table

Results

Population structure and genetic differentiation

We generated 454 control region sequences from 25 sampling sites and divided them into nine subpopulations according to the geographic distance between sites and the mountain ranges they were located on (Fig. 1; Supplementary Fig. S1). We identified 18 haplotypes arising from 19 variable sites (Supplementary Table S2). We conducted a SAMOVA to examine best-fit grouping patterns by maximizing the proportion of genetic variance among regions, which revealed an optimal grouping of six regions, with the highest variance (68.78%) attributed to variatiofn among regions (P < 0.01) (Table 1). Of the nine a priori subpopulations, five (NSP, SSP, NTR, STR, NDD) were assigned by SAMOVA as being distinct regions, with one further region comprising the remaining SDD, NYS, SYS, and SGL subpopulations (Table 1). Hereafter, these six regions are denoted Southern Sheipa National Park (SSP), Northern Sheipa National Park (NSP), Southern Taroko National Park (STR), Northern Taroko National Park (NTR), Northern Danda Major Wildlife Habitat (NDD), and Central to Southern Range (CSR, formed by the grouping of the four remaining subpopulations).

Fig. 1
figure 1

Geographic distribution of two clades identified from phylogenetic analysis of nine Formosan sambar deer subpopulations. The red and blue sections of the nine subpopulation pie charts represent the frequency of occurrence of Taroko-Sheipa (TS) and Central Mountain Range (CMR) clade haplotypes, respectively. The yellow dots denote mean geographical centers of the sampling sites. Details on sampling sites and subpopulations are presented in Supplementary Table S1

Full size image

Haplotype diversity (h) and nucleotide diversity (π) for the entire Formosan sambar deer population were determined to be 0.814 and 0.00371, respectively. Only one haplotype was found in each of the NSP and SSP regions. Apart from those two regions, the value of h for the remaining four regions ranged from 0.255 (STR) to 0.693 (NTR), and values of π ranged from 0.00056 (CSR) to 0.00399 (NTR). All regions and subpopulations presented low values for π (< 0.005). These results indicate that deer from NTR possess greater genetic diversity than those from other regions (Table 2).

Table 2 Sample size (n), number of haplotypes (nh), haplotype diversity (h), nucleotide diversity (π), and neutrality test results for each region, subpopulation, and clade of Formosan sambar deer
Full size table

We measured genetic distances (Dxy) and fixation indices (FST) to elucidate the genetic differentiation among subpopulations (Table 3). Regions NSP and SSP presented the closest genetic distance (0.0011). The largest genetic distance was identified between the SSP and STR subpopulations (0.0081), rather than between the NSP and SGL subpopulations (0.0069) that are the furthest apart. Similarly, we observed lower genetic distance but greater geographic distance between the SDD and SGL subpopulations (0.0002) than between adjacent STR and NDD subpopulations (0.0043). The lowest FST values were detected between the SDD and SGL subpopulations (0.0038), with the highest respective values being between NSP and SSP (1.0000), representing two adjacent subpopulations (Table 3). These results indicate that geographical distance may not be the only factor contributing to genetic barriers. Consequently, we explored historical demography to explain genetic differentiation across regions.

Table 3 Pairwise genetic differentiation index (FST, below the diagonal) and genetic distance (Dxy, above the diagonal) based on mitochondrial control-region sequences for nine Formosan sambar deer subpopulations
Full size table

Phylogenetic relationships and divergence

To understand the genetic characteristics of the Formosan sambar deer population, we performed a network analysis to depict the genealogy of the 18 identified haplotypes (Fig. 2). We uncovered two primary genetic clades comprising RUS01-04 and RUS05-18, with four diagnostic nucleotide substitutions between them at sites 151 (C/T), 180 (G/A), 352 (A/G), and 371 (C/T) (Supplementary Table S2). To further clarify the phylogenetic relationships among our control region sequences, we constructed a phylogenetic tree based on Bayesian inference. Sequences of Rusa unicolor hainana, Muntiacus reevesi, and Muntiacus crinifrons obtained from NCBI were used as outgroups and for fossil calibration of molecular clock dating. Our Bayesian phylogenetic tree clearly separated the 18 Formosan sambar deer haplotypes into two clades, with a significant Bayesian posterior probability (> 0.98; Fig. 3). The two clades reflect spatial divergence between north and south Taiwan. Hereafter, we name the clade comprising haplotypes RUS01-04 as Taroko-Sheipa (TS), which encompasses deer distributed in regions SSP, NSP, STR, NTR, and NDD in northern Taiwan. Within the TS clade, haplotype RUS01 is restricted to the SSP region; haplotype RUS02 occurs in NSP, NTR, and STR; haplotype RUS03 in NTR; and haplotype RUS04 in STR and NDD (Supplementary Table S3). The second clade (hereafter, the Central Mountain Range (CMR) clade) comprises haplotypes RUS05–18. Haplotypes RUS05 and RUS06 of the CMR clade are only found in regions STR, NTR, and NDD, apart from one individual that was detected in the SDD subpopulation of the CSR region. All other haplotypes (RUS07–18) were solely distributed in the CSR region. Notably, 147 individual deer possessed haplotype RUS07, representing the majority of our samples, and it is more widely distributed than any of the other haplotypes (Supplementary Table S3). Values of h and π for the TS clade were 0.571 and 0.00104, respectively, whereas for the CMR clade, they were 0.701 and 0.00115 (Table 2), indicating greater genetic diversity in the CMR clade relative to the TS clade.

Fig. 2
figure 2

Haplotype network displaying 95% parsimonious connections of Formosan sambar deer haplotypes. The open circles show the 18 detected haplotypes, and the solid circles show hypothetical haplotypes. The lines connecting haplotypes indicate single-step mutations between them. The circle size is proportional to the sample size presented in the lower right corner. CMR, Central Mountain Range clade; TS, Taroko-Sheipa clade

Full size image
Fig. 3
figure 3

Fossil-calibrated Bayesian phylogenetic tree constructed from mitochondrial control region haplotypes. Bayesian posterior probabilities are shown above the nodes. Numbers below the nodes indicate the divergence time (and 95% highest posterior density represented by a blue bar) to the most recent common ancestor of each clade. TS, Taroko-Sheipa clade; CMR, Central Mountain Range clade

Full size image

The divergence times to the most recent common ancestor of each clade were estimated using a calibrated molecular clock analysis on our Bayesian phylogenetic tree (Fig. 3), which revealed coalescence of a common ancestor for Rusa unicolor hainana and Rusa unicolor swinhoii (Formosan sambar deer) approximately 0.97 MYA (95% highest posterior density (HPD) = 0.55–1.41 MYA), i.e., during the late Pleistocene. Coalescence of our two Formosan sambar deer clades occurred 0.41 MYA (95% HPD = 0.21–0.63 MYA), also in the late Pleistocene. The most recent common ancestors for the TS and CMR clades arose 0.16 MYA (95% HPD = 0.04–0.30 MYA) and 0.24 MYA (95% HPD = 0.12–0.37 MYA), respectively.

We also generated a Bayesian skyline plot (BSP) to reconstruct the relationship between the timing of past events and the magnitude of change in the effective maternal population size of Formosan sambar deer since the appearance of its most recent common ancestor (Fig. 4). This analysis revealed a slow and steady increase in overall population size (Fig. 4a). Notably, the effective maternal population size of the CMR clade was relatively constant up to 10,000 years ago (YA) and then rapidly expanded thereafter (Fig. 4b). In contrast, the effective maternal population size of the TS clade has increased only gradually since 2000 YA, expanding to a much lesser extent than observed for the CMR clade (Fig. 4c).

Fig. 4
figure 4

Bayesian skyline plots representing historical demographic trends for a the entire Formosan sambar deer population, b the Central Mountain Range clade, and c the Taroko-Sheipa clade. X-axis shows time before present in million years ago (MYA). Y-axis indicates effective population size on a logarithmic scale. The thick solid lines represent mean estimates for effective population size, and the differentially colored areas reflect the 95% highest posterior density intervals

Full size image

Historic demographic analyses

To better understand the demographic history of the Formosan sambar deer population, we applied neutrality tests for subpopulations, regions, and clades to detect population expansion (Table 2). Region CSR displayed significantly negative values for both Tajima’s D (− 1.84610, P < 0.01) and Fu’s Fs (− 12.08819, P < 0.001). Clade CMR showed a significantly negative value for Fu’s Fs (− 6.11141, P < 0.05) and a unimodal pattern in mismatch distribution (Fig. S3). Together, these results indicate that deer from regions CSR and STR, as well as deer in the CMR clade, have undergone historical population expansion.

We observed that the RUS05 and RUS06 haplotypes from the CMR clade occurred in regions NDD, NTR, and STR, i.e., where haplotypes from the TS clade are also found. Moreover, our neutrality tests and mismatch analysis revealed population expansion for the CMR clade and the CSR region, indicating the possibility of secondary contact between CSR and more northerly regions. To explore the possible dispersal history of Formosan sambar deer in Taiwan, we used coalescent simulations to compare the probabilities of 10 potential scenarios (Fig. 5), which identified Scenario 4 as the most likely demographic history with a posterior probability of 0.5389 (a 95% credibility interval [CI] of 0.4544–0.6234). This scenario supports region CSR as being the likely origin of all others and experiencing secondary contact with STR to give rise to the NDD subpopulation. Other events encompassed by this scenario include splitting of CSR from other regions (t5), divergence of the STR and SSP + NSP regions (t4), secondary contact between the STR and CSR regions giving rise to the common ancestor of NDD (t3), divergence of the SSP and NSP regions (t2) and, finally, splitting of the STR and NTR regions (t1), all occurring in chronological order from most ancient to most recent.

Fig. 5
figure 5

Scenarios of lineage diversification and secondary contact for Formosan sambar deer based on mitochondrial control region sequences. Scenario 4 displayed the best fit to our dataset based on approximate Bayesian computation (ABC). Posterior probabilities (PP) are shown for each scenario. Values in square brackets indicate the 95% highest posterior probability confidence interval

Full size image

Discussion

Our SAMOVA revealed significant genetic differentiation among Formosan sambar deer from six defined regions. This result implies potential barrier effects between those six regions. However, minimal and maximal pairwise genetic distances and FST values did not correspond to geographic distances. This discordance between geographic and genetic distances could be due to non-mutually exclusive factors, such as (1) the dispersal capability of Formosan sambar deer could be higher than expected or (2) topological or climatic factors could restrict movement irrespective of dispersal capacity. To further explore potential contributory factors to the observed genetic patterning, we inferred the dispersal history of Formosan sambar deer.

We noted that haplotype RUS07, which was assigned to the CMR clade, lay at the center of the star-like network and that it is distributed widely throughout the CSR region (Fig. 2). Moreover, we identified greater genetic diversity in the CMR clade relative to the TS clade based on h and π values (Table 2). Accordingly, we postulate that haplotype RUS07 represents the ancestral lineage for the two clades. Although scenarios linked to SP being the origin are also plausible, our approximate Bayesian computation (ABC) supports that CSR acted as the origin of the current subpopulations. Only two haplotypes, RUS01–02, occur in Sheipa National Park (the NSP and SSP regions), and haplotypes of the CMR clade are not found there, evidencing a barrier to gene flow. Thus, from our haplotype phylogeny, we infer that the TS clade was derived from the CMR clade. The two clades are separated by three intermediate haplotypes (Fig. 2, indicated by three black dots), which were not detected among our samples. Neither RUS07 nor the three missing haplotypes were detected at NTR, STR, and NDD, indicating that deer hosting these haplotypes may be rare or have become extinct. This outcome implies that the trajectory of haplotype evolution RUS07 → RUS05 → RUS06 did not occur in TS, but arose from secondary contact due to sambar harboring RUS05 and RUS06 migrating into a region hosting the TS clade.

Haplotypes of both the TS and CMR clades (RUS02–RUS06) were detected in the NDD, NTR, and STR regions (Fig. 1; Supplementary Table S3), with secondary contacts potentially giving rise to this genetic admixture. Geographically, deer of the TS clade are primarily restricted to habitats north of the Chilai ridgeline (indicated as Chilai-Nengkao in Fig. 1). In contrast, deer of the CMR clade occur south of it, except for deer with the RUS05 and RUS06 haplotypes that occur in the NTR, STR, NDD, and CSR regions. Our approximate Bayesian computation (ABC) analysis supports the hypothesis of the CSR being the origin of Formosan sambar subpopulations and that secondary contact between the STR and CSR regions gave rise to the NDD subpopulation (Fig. 5). Contact between subpopulations may have arisen from expansion of the CMR or TS clades, or both. The CMR clade presented a statistically significantly negative value for both Tajima’s D and Fu’s Fs, indicating that it had undergone population expansion (Table 2). In contrast, the same neutrality test results for the TS clade and the entire Formosan sambar deer population were not significant. These outcomes were supported by our BSP analysis, together revealing that the CMR and TS clades have experienced differing demographic trajectories. Sampling biases due to the rough terrain of the Taiwanese mountain ranges that limited sampling effort may have contributed to the fact that only one haplotype for each subpopulation was detected in SSP and NSP and, consequently, why we did not detect significant expansion for SP regions. If such biases exist, it is plausible that sambar in the STR, NTR, and NDD regions originally featured haplotypes of the CMR clade (RUS05 and RUS06) and came into contact with the TS clade (RUS03 and RUS04) during expansion of the TS. However, we assert that the frequency of missed haplotypes is not high, and if SP regions were the source of RUS03 and RUS04, the frequency of these two haplotypes should be high there. Yet, nowhere displays a sufficiently high frequency of haplotypes RUS03 and RUS04 that could have acted as the source of the TS clade, which could have expanded into NTR and STR. Instead, it is more likely that haplotypes RUS03 and RUS04 detected in NTR and STR represent the origin of the TS clade where they evolved into RUS02 and RUS01, with these latter two haplotypes then dispersing into the SP region. This assumption is supported by the DIYABC result that depicts the CMR-origin scenario, with dispersal occurring from both south to north and from east to west. Hence, we suspect that dispersal of deer harboring the RUS05 and RUS06 haplotypes is responsible for the secondary contact between northern and southern Formosan sambar populations.

Based on all of our analyses, we present a likely scenario explaining the genetic patterning of Formosan sambar deer. We hypothesize that Formosan sambar deer first became established in the central or southern mountain ranges of Taiwan upon arrival from continental Asia and then dispersed northward. This northward dispersal resulted in the formation of two clades, i.e., CMR and TS. Based on molecular clock dating and fossil calibration, divergence between these two clades occurred 0.41 MYA, i.e., during the Nebraskan glacial stage (0.33–0.47 MYA) of the mid-Pleistocene, indicating that divergence of the TS clade coincided with a glacial period (Fig. 3). More recently, the ancestor of the TS clade migrated into the northern parts of the Central Mountain Range and into the region of Taroko National Park (from where we sampled the NTR and STR subpopulations) that are not separated by apparent geographic barriers. However, during the last glaciation, equilibrium-line altitudes (ELAs) of the Nanhu Mountain glacial cirque in Taroko National Park expanded, lowering snowlines (1500 m or less, Böse 2000, 2004; Hebenstreit et al. 2006). Thus, glaciers covering alpine hilltops formed closed-circle ELAs in the Taroko area, severely restricting southward migration of deer from the TS clade. Moreover, deer of the TS clade were forced to seek refuge in low-altitude canyons within Taroko where vegetation could still be found. The walls of those canyons are surrounded by steep mountains and cliffs (Supplementary Fig. S3), which would have been capped by glaciers during glacial periods, creating topographic and climatic barriers to gene flow.

Radiotelemetry data and camera-trapping has revealed that Formosan sambar deer display altitudinal migration. In cold months, they migrate to lower elevations from the upland habitats they occupy during warmer months (Yen et al. 2013, 2019). This behavior implies that Formosan sambar may be sensitive to temperature, especially to cold weather. Temperatures during the Last Glacial Maximum (30,000–22,000 YA) were 8–11 °C lower than present (Tsukada 1967), so shifting of vegetation zones into lower elevations unsuitable for sambar deer likely induced genetic barriers during glacial times. Stress-induced selection and allotropic differentiation may have also resulted in the loss of intermediate haplotypes between the CMR and TS clades (represented as black dots in Fig. 2), facilitating their differentiation. However, during subsequent interglacial periods, the deer underwent post-glacial population expansion, allowing secondary contact between both clades. Hence, deer populations of the CMR and TS clades now occur sympatrically in the northern part of the Central Mountain Range (i.e., in the NTR, STR and NDD regions).

In conclusion, we have detected 18 mitochondrial control region haplotypes among Formosan sambar deer and assessed genetic differentiation across subpopulations. We have identified two primary haplotype lineages. Moreover, we have postulated a secondary contact event and predicted a likely evolutionary scenario for the demographic history of Formosan sambar. Finally, we have inferred dispersal histories in the context of Pleistocene climatic oscillations and speculated that the area of Taroko National Park acted as a refugium for Formosan sambar deer during glacial periods. Our study represents an example of how climate change plays an important role in patterns of genetic variation for large herbivorous mammals.