2.1 Sampling and DNA extraction

A total of 488 individual specimens from 39 geographical localities were used in this study, out of those 145 individuals from 15 populations were retrieved from our previous study (Yuan et al., 2015). Our sampling included all the eight Gynaephora species inhabiting the QTP and covered most of the distribution of grassland caterpillars in the QTP meadows. Detailed sampling information is provided in Figure 1 and Table 1. All Gynaephora specimens were hand-netted in alpine meadows of the QTP during July–August between 2013 and 2016. All specimens were stored in absolute ethyl alcohol in the field and transferred to −20°C until they were used for DNA extraction. Total genomic DNA was extracted from each specimen using the DNA Extraction Kit (Tiangen, Beijing, China) following the manufacturer's protocol.

2.2 PCR amplification and sequencing

Two mitochondrial genes (COI and ND5) were amplified using the same primers and procedures described in our previous study (Yuan et al., 2015). All PCR products were purified using a DNA Gel Purification Kit (Omega, Norwalk, CT, USA) and sequenced in both directions by Sanger sequencing using the same PCR primers.

Sequences of the two mitochondrial genes (COI and ND5) were initially aligned using ClustalW in MEGA 5.10 (Tamura et al., 2011) with default parameters to check for stop codons or indels, which could reveal mitochondrial pseudogenes. All sequences newly obtained in this study have been deposited in GenBank under accession numbers OP574353–OP574695 and OP577499–OP577841. Then, sequences of COI and ND5 were concatenated using DAMBE 7.2.6 (Xia, 2018). Mitochondrial haplotypes were identified using DNASP 5.10 (Librado & Rozas, 2009).

2.3 Phylogenetic analysis

A total of 11 Gynaephora species were included in phylogenetic analyses, i.e. eight QTP Gynaephora species, G. groenlandica, G. rossii, and G. selenitica. We selected Lymantria dispar (NC_012893) as the outgroup, according to our previous study (Yuan et al., 2015). For each gene, substitution saturation was evaluated using DAMBE 7.2.6 (Xia, 2018). A lack of evidence for saturation indicated that all codon positions can be used for the phylogenetic analysis. The partitioning schemes and corresponding nucleotide substitution models for the dataset (COI + ND5) were selected by PartitionFinder 1.1.1 (Lanfear et al., 2012), partitioned by genes and codon positions as in our previous study (Yuan et al., 2015). Single partitioning scheme with the HKY + I + G model was used for the following phylogenetic analyses. The phylogenetic tree was constructed based on the mitochondrial dataset (COI + ND5) using maximum likelihood (ML) and Bayesian inference (BI). Both BI and ML analyses were conducted using the CIPRES Science Gateway 3.3 (Miller et al., 2010). The ML analysis was carried out using RAxML-HPC2 on XSEDE 8.0.24 (Stamatakis, 2014) with the GTRGAMMA model, and 1000 bootstrap (BS) replicates were used to estimate node reliability. The BI analysis was conducted using MrBayes 3.2.2 (Ronquist et al., 2012) on XSEDE. Four Markov chains (three hot and one cold chain) were independently run two times for 1 × 10⁶ generations, with sampling every 1000 generations. Chain congruence was assessed by the effective sample size (ESS) (>100) and the potential scale reduction factor (PSRF) (approximately 1.0), as recommended in MrBayes 3.2.2 documentation (Ronquist et al., 2012). The first 25% of samples were discarded as burn-in, and the remaining samples were used to construct a 50% majority-rule consensus tree. Bayesian posterior probabilities (PP) were calculated. Furthermore, a median-joining network was generated for all mitochondrial haplotypes using Network 4.6.0.0 (Bandelt et al., 1999).

2.4 Genetic diversity and differentiation

Haplotype diversity (h) and nucleotide diversity (π) for all 39 geographic populations and each of the four clades (based on the population structure analysis) were calculated using DnaSP 5.10 (Librado & Rozas, 2009). Genetic differentiation (F_ST) between the 39 geographic populations was calculated using Arlequin 3.5 (Excoffier & Lischer, 2010). The significance of F_ST between populations was tested by 10,000 permutations in Arlequin 3.5 (Excoffier & Lischer, 2010). To investigate genetic structuring between sampling locations, an analysis of molecular variance (AMOVA) (Excoffier et al., 1992) was performed with Arlequin 3.5 (Excoffier & Lischer, 2010). Mantel tests were performed using Isolation By Distance Web Service (IBDWS) (Jensen et al., 2005) to test the correlation between genetic distance and geographic distance.

2.5 Population demography

The historical population demographics of QTP Gynaephora species was investigated by neutrality tests (Fu's Fs, Fu and Li's F* and D*, and Tajima's D) and pairwise mismatch distributions for all 39 sampling locations combined and each of four clades separately using DnaSP 5.10 (Librado & Rozas, 2009). Fu's Fs, Fu and Li's F* and D*, and Tajima's D were calculated to detected population demography (Fu, 1997; Tajima, 1989). Significant negative F_S values can be taken as evidence for expansions, while positive values might result from population subdivision or a recent population bottleneck (Fu, 1997). For the mismatch distribution analysis, a unimodal distribution indicates a recent demographic expansion, while multimodal distributions indicate population stability (Harpending et al., 1998; Rogers & Harpending, 1992).

2.6 Divergence time estimation

Divergence times of the QTP Gynaephora were estimated by a molecular clock approach implemented in BEAST 1.8.1 (Drummond et al., 2012) using the phylogenetic tree of haplotypes as a constraint tree. A lognormal prior was used for COI and the mean substitution rate was set to 0.0115 per site per million years (Ma). We scaled the ND5 substitution rate to the mean COI rate (according to the K2P genetic distances between COI and ND5), following the method described in our previous study (Yuan et al., 2018). The following settings were used: linked tree model, HKY + I substitution model for each mitochondrial gene, strict clock, Yule model of speciation, and default values for the remaining parameters. Chain convergence was checked using the ESS implemented in Tracer 1.5 (Rambaut et al., 2018). Posterior estimates of Gynaephora ages and 95% highest posterior densities (HPD) were summarized using TreeAnnotator v1.8.1 (Drummond et al., 2012). All BEAST analyses were conducted on the CIPRES Science Gateway 3.3 (Miller et al., 2010).

2.7 Biogeographical analysis

Ancestral area reconstruction was conducted to investigate the biogeographical history of Gynaephora by using a Bayesian binary MCMC (BBM) method, as implemented in RASP 3.1 (Yu et al., 2015). Seven biogeographical areas were defined based on the species distributions and results of the phylogenetic analysis (Figure 1): the distributions of (a) G. groenlandica; (b) G. rossii; (c) G. selenitica; and (d–g) the four clades A–D in Figure 1. The BBM analysis was run with the following settings: 10.001 credible trees generating from the BEAST analysis were used, the maximum number of areas for all nodes was set to seven, and the remaining parameters were set to default values.

2.8 Modeling the gene flow

Gene flow between the four clades (A–D in Figure 1) identified by phylogenetic analyses was modeled by using a coalescent-based approach implemented in Migrate-n (Beerli, 2006; Beerli & Palczewski, 2010). This software estimates long-term average values of gene flow and can be used to test predictions about refugial structure and the direction of migration from putative refugia. The Sanjiangyuan region, locating in the central and eastern QTP, might represent a refugium of the QTP Gynaephora during glacial periods, as has been inferred in previous studies of other taxa (Qiu et al., 2011; Yang et al., 2009). Therefore, to reduce the number of models, we first divided the 39 populations into two groups based on the phylogenetic results, i.e., group A (Clade A in Figure 1) and group BCD (Clades B–D in Figure 1), and constructed three migration models (M1–M3 Figure S2). Then, we tested 14 migration models for three groups (Clade A, Clade B, and Clades C + D) and nine migration models for four groups (Clades A–D in Figure 1; Figure S2).

Migrate-n was used to estimate the migration rates (M) and effective population size (θ) parameters as well as the marginal likelihood of each gene flow model (Beerli, 2006). The gene flow models were ranked according to log Bayes factors (LBFs) calculated from the Bezier corrected marginal likelihoods of the data given the model. The marginal likelihoods for each model were approximated by thermodynamic integration of the MCMC over four heated chains (Metropolis coupled MCMC), as described previously (Beerli & Palczewski, 2010). We used a uniform prior for θ between 0 and 0.1 and a sampling window of 0.01 where we generate new proposals; for M, we used a uniform prior between 0 and 1000, with a window of 100 model. A four-chain static heating scheme was used, and 100,000 trees were discarded as burn-in before recording 1,000,000 steps with an increment of 100, resulting in 40 million samples. To evaluate convergence, we evaluated ESS (>10,000) and the posterior distributions of the parameters to determine whether they were unimodal smooth curves. Bayes factors (BFs) were calculated as a ratio of the marginal likelihoods to calculate model probabilities.

3.1 Population structure

A total of 40 haplotypes were detected for the combined mitochondrial gene dataset (COI and ND5) (Table 1). Among these haplotypes, each of 29 haplotypes was found only in one population, and the remaining 11 haplotypes were shared by multiple geographic populations (Table 1; Figure 2). Haplotypes H3 and H4 showed the highest frequencies, followed by H2, H5, H7, and H11 (Figure 2). Only one haplotype was detected in all 18 populations, and the DM population (six haplotypes) had the largest number of haplotypes (Table 1).

Phylogenetic analyses based on 40 haplotypes using two methods (BI and ML) yielded nearly identical topologies (Figure 3). The QTP Gynaephora species formed a strongly supported monophyletic group (PP = 1.0, BS = 100) and consisted of four clades (A–D in Figures 1 and 3). Clade A (16 haplotypes in 12 populations) formed a sister group relationship to a clade containing the remaining three clades (Clades B–D), whereas Clade B (four haplotypes in six populations) was sister to Clade C (five haplotypes in 10 populations) and Clade D (15 haplotypes in 11 populations) (Figure 3). These four clades were also obtained by a haplotype network analysis (Figure 2).

3.2 Genetic diversity and differentiation

For QTP Gynaephora species overall, high levels of haplotype diversity (h = 0.913 ± 0.007) and nucleotide diversity (π = 2.288 ± 0.033) were observed (Table 2). Among the four clades (Figures 1-3), Clade A showed the highest genetic diversity (H = 16, h = 0.888 ± 0.012, and π = 0.641 ± 0.019%), followed by Clade D (H = 15, h = 0.837 ± 0.020, π = 0.201 ± 0.006%) and Clade C (H = 5, h = 0.567 ± 0.040, and π = 0.061 ± 0.006%), while Clade B (H = 4, h = 0.112 ± 0.042, and π = 0.073 ± 0.30%) showed the lowest genetic diversity (Table 2).

Almost all pairwise F_ST values between the 39 sampling sites were significantly greater than 0 (Table S1), indicating substantial genetic differentiation between QTP Gynaephora populations. Mantel tests of the 39 QTP Gynaephora populations indicated that there was a significant positive correlation between genetic distances and geographic distances (r = .169–.517, p < .01; Table S2). Among the four clades, a significant correlation between genetic and geographic distances was found only for Clade A (r = .191–.457, p < .05; Table S2).

To further investigate genetic differentiation and genetic structuring between the 39 geographic populations we performed AMOVA for the four clades obtained from the phylogenetic and network analyses (Figures 1 and 2). A significantly positive F_CT (F_CT = 0.8701, p < .001; Table 3) suggested that geographic structuring existed among the examined populations. Genetic differentiation among the four clades accounted for 87.01% of the total genetic variance, whereas differentiation within clades only accounted for 1.35% of genetic variation, indicating a low gene flow among the four clades.

3.3 Population demography

The mismatch distribution for the 40 haplotypes of the 39 geographic populations was multimodal (Figure S1), indicating that no obvious expansion occurred in the QTP Gynaephora species. Similar results were obtained for three clades (Clades A, B, and D), whereas Clade C showed a unimodal distribution, which could be due to either demographic expansion (population expansion) or to positive selection. In neutrality tests, significant negative values were not obtained for QTP Gynaephora as a whole or for Clades A, B, and C (Table 2). Significant positive values for each of four neutrality tests were found in all populations combined, whereas Fu's F_S value was significantly negative in Clade D (Table 2).

3.4 Divergence time estimation

Divergence time estimates for the genus Gynaephora based on a strict mitochondrial molecular clock are shown in Figure 4. Divergence within the genus was dated to the late Miocene, approximately 6.05 Ma (95% HPD: 2.9–10.83 Ma). The divergence of QTP Gynaephora species from non-QTP species likely occurred during the early Pliocene, at around 4.65 Ma (95% HPD: 2.22–8.12 Ma). The earliest split within QTP Gynaephora was dated to the early Pleistocene at about 2.08 Ma (95% HPD: 0.90–3.68 Ma), followed by further splitting at around 1.31 Ma (95% HPD: 0.58–2.39 Ma). Clades C and D diverged about 0.53 Ma (95% HPD: 0.2–1.03 Ma).

3.5 Biogeographic reconstruction

The BBM biogeographic analysis revealed that the genus Gynaephora likely originated in central Europe (see Figure 5c for the distribution of the type species G. selenitica) with high probability; 12 dispersal and six vicariance events led to the colonization of current ranges with two expansion routes: northward dispersal to form several arctic Gynaephora species (e.g., G. rossii and G. groenlandica) and southward dispersal to form the QTP Gynaephora species (Figure 5). The common ancestor of the QTP Gynaephora species likely first dispersed to the Sanjiangyuan Region of the QTP (Clade A in Figures 1-3) and then to the Naqu Region, leading to the emergence of G. alpherakii by two westward dispersal events and one vicariance event, followed by dispersal to the North and Northeast QTP (Figure 5) and the split into two clades (Clades C and D in Figures 1-3).

3.6 Gene flow

A total of 26 models were evaluated using Migrate-n (Table 4; Figure S2). The best-supported models were M3 for two groups (A, BCD), M5 for three groups (A, B, CD), and M20 for four groups (Clades A–D) (Table 4; Figure S2). These best models consistently supported Clade A as the refugium with unidirectional gene flow from Clade A to Clade B (mean Nm = 91.2) and from Clade B to Clade C (mean Nm = 142.9) (Figure 1; Table S3). Bidirectional and asymmetrical gene flow were found between Clades C and D (Figure 1; Table S3). Theta estimates suggested that Ne was highest in Clade A and lowest in Clade C (Table S3).

4.1 The number of QTP Gynaephora species may be overestimated

The unique geographical and environmental conditions of the QTP led to a large number of endemic species. Eight Gynaephora species endemic to the QTP have been reported according to morphological characteristics (Yan, 2006; Zhang & Yuan, 2013), but were not well supported by molecular data in our previous studies (Yuan et al., 2015, 2018). In this study, by extensive sampling covering almost the whole geographical distribution and phylogenetic analyses, the number of valid species was further reduced to four monophyletic groups (Clades A–D in Figure 1). Clades B and C corresponded to G. alpherakii and G. menyuanensis, respectively. Clades A and D each contained multiple morphologically described species (Table 1). Among three species included in Clade A, G. qinghaiensis had a wide geographical distribution, while the two other Gynaephora (G. qumalaiensis and G. jiuzhiensis) are locally distributed (Yuan et al., 2015). G. qinghaiensis has long been considered the only species in alpine meadows of the QTP, i.e., grassland caterpillars destroying alpine meadows are all considered G. qinghaiensis. Although the Latin name G. alpherakii is frequently used to describe the QTP grassland caterpillars in the literature, Chou and Ying (1979) found that no specimens of grassland caterpillars sampled conformed to the morphological characteristics of G. alpherakii, questioning the existence of this species in the QTP. However, both our results and previous results consistently indicated that grassland caterpillars collected from multiple geographical localities in the Naqu Region of Tibet are an independent species, including those in Amdo County, the locality of the type specimen of G. alpherakii. Further studies including morphological characteristics are necessary to confirm whether this recovered monophyletic group is indeed G. alpherakii, as originally described.

Three Gynaephora species (G. aureata, G. ruoergensis, and G. minora) clustered together to form Clade D, as found in our previous study (Yuan et al., 2015). G. aureata is widely distributed, while the two other species are only distributed in Ruoergai County (Chou & Ying, 1979; Zhang & Yuan, 2013). The two sympatric species are morphologically similar, with obvious differences only in wing size (Chou & Ying, 1979), likely due to phenotypic plasticity. In addition, a sympatric distribution may facilitate relatively strong mitochondrial introgression due to interspecific hybridization. Among eight QTP Gynaephora species, G. menyuanensis is the most recently described and is mainly distributed in the northeastern QTP (Zhang & Yuan, 2013). This species can be effectively differentiated based on several morphological characteristics from other Gynaephora species. The validity of this species was also supported by molecular data in our present and previous study (Yuan et al., 2015), suggesting the independent species status of G. menyuanensis.

4.2 High genetic diversity in QTP Gynaephora species

Genetic diversity is a fundamental component of biodiversity. Generally, the higher the genetic diversity within a species, the greater the potential for adaptation in response to changing environments. Genetic diversity can be measured by two indices, i.e., haplotype diversity (h) and nucleotide diversity (π). Nucleotide diversity represents the cumulative degree of genetic variation during evolution, while haplotype diversity reflects the probability that two randomly sampled alleles are different and is more relevant to population dynamics at short time scales (Leitwein et al., 2020; Pauls et al., 2013). Patterns of genetic diversity can be assigned to four categories (Grant & Bowen, 1998): (1) low h and low π, (2) high h and low π, (3) low h and high π, and (4) high h and high π. The results for all 39 Gynaephora populations as a whole were consistent with the fourth category, i.e., high h (>0.5) and high π (>0.5%), suggesting that QTP Gynaephora species was a large stable population with a long evolutionary history or secondary contact between differentiated lineages.

Among the four clades (A–D in Figures 1-3), Clade A showed the highest haplotype and nucleotide diversities, consistent with the hypothesis that source populations generally have high genetic diversity. Given that Clade A consisted of three morphologically described species, we proposed that the high level of genetic diversity found in Clade A may be related to secondary contact between these differentiated species/populations during glacial–interglacial cycle periods. Furthermore, it is likely that there are cryptic species within Clade A that were not resolved with the present mitochondrial data. Clade B (G. alpherakii) fell into the first category (low h and low π), suggesting a contribution of founder events in which new populations were established by a small number of individuals drawn from the large ancestral Clade A. In particular, only one exclusive haplotype was found in geographical populations AD, NQ, RW, SS, and XQ from Clade B, further confirming that genetic drift might eliminate many haplotypes, leaving only one haplotype in these populations. It is possible that Clade B, found in higher-altitude areas, faced stronger environmental pressures, resulting in lineage-specific haplotypes. Both Clades C and D were consistent with the second category, i.e., high haplotype diversity (h > 0.5) and low nucleotide diversity (π < 0.5%). This might be attributed to population expansion after population bottleneck followed by rapid population growth and the accumulation of new mutations (Grant & Bowen, 1998). High haplotype diversity was likely related to the high mitochondrial DNA mutation rate, as indicated by a haplotype network analysis (Figure 2) with 1–2 mutation steps before the generation of a new haplotype.

4.3 Isolation and subsequent divergence is the key speciation mechanism of the QTP Gynaephora species

The divergence between QTP Gynaephora species and non-QTP species was estimated to occur in the early Pliocene (around 4.65 Ma), which was slightly younger than our previous estimate (Yuan et al., 2015). The genus Gynaephora likely originated in central Europe, as indicated by the biogeographic analysis (Figure 5). Therefore, ancestors of QTP Gynaephora arrived at the QTP during a period of uplift. The QTP has experienced uplifts many times; an intensive uplift probably began during the Miocene and the most intensive uplifts during the Pliocene and Pleistocene are known as the Qingzang Movement (3.6–1.7 Ma) and the Kun-Huang Movement (1.2–0.6 Ma) (Shi et al., 1998; Zheng et al., 2000; Zhisheng et al., 2001). These intensive uplifts of the QTP were likely causes of interspecific divergence in QTP Gynaephora date to the early Pleistocene to Pliocene (2.08–0.53 Ma). In addition, changes in vegetation from forests to grasslands due to climate change during the Pleistocene (Wu et al., 2001) not only provided abundant food resources, an important basis for population expansion, but also likely provided new habitats, contributing to population differentiation and speciation. Therefore, we proposed that grassland caterpillars arrived at the QTP after the QTP uplifts and then gradually diverged, resulting in speciation influenced by plateau uplifts and associated climatic fluctuations.

Biogeographic analysis, divergence time estimation, and gene flow analysis showed that the region where Gynaephora first arrived was the interior of the QTP (the Sanjiangyuan region, Clade A in Figure 1) after its migration from Eurasia. This was also supported by the higher genetic diversity in Clade A than in the three other clades. Given the unidirectional gene flow from Clade A to Clade B, the QTP Gynaephora species may have remained in the interior of the QTP during the glacial period, rather than dispersing to low-altitude environments. Although it is not clear whether the QTP formed a large ice sheet during the glacial period, some animals did remain in the QTP interior during these periods. Therefore, the Sanjiangyuan region was likely a refuge for QTP Gynaephora during glacial periods of the Pleistocene. Generally, postglacial differentiation could induce a decrease in genetic diversity along the expansion route, with an increase in the distance from refugia (Comes & Kadereit, 1998; Provan & Bennett, 2008). This hypothesis can explain the decrease in genetic diversity from Clade A to Clade B; however, it was difficult to explain the increasing trend from Clade B to Clade C. Both Clades C and D inhabited relatively low altitudes and showed relatively high genetic diversity, likely due to a high level of bidirectional gene flow between these two clades during the interglacial period. Significant genetic differentiation among the four clades and unidirectional gene flow (Clade A to Clade B and then Clade C), together with the observed association between genetic and geographic distances, indicated that geographic isolation and subsequent divergence was the dominant mode of speciation in the QTP Gynaephora, as has been proposed in some animal taxa endemic to the QTP (Favre et al., 2015; Lei et al., 2014; Wen et al., 2014).