Landscape Heterogeneity Explains the Genetic Differentiation of a Forest Bird across the Sino-Himalayan Mountains

Abstract

Mountains are the world's most important centers of biodiversity. The Sino-Himalayan Mountains are global biodiversity hotspot due to their extremely high species richness and endemicity. Ample research investigated the impact of the Qinghai–Tibet Plateau uplift and Quaternary glaciations in driving species diversification in plants and animals across the Sino-Himalayan Mountains. However, little is known about the role of landscape heterogeneity and other environmental features in driving diversification in this region. We utilized whole genomes and phenotypic data in combination with landscape genetic approaches to investigate population structure, demography, and genetic diversity in a forest songbird species native to the Sino-Himalayan Mountains, the red-billed leiothrix (Leiothrix lutea). We identified 5 phylogeographic clades, including 1 in the East of China, 1 in Yunnan, and 3 in Tibet, roughly consistent with differences in song and plumage coloration but incongruent with traditional subspecies boundaries. Isolation-by-resistance model best explained population differentiation within L. lutea, with extensive secondary contact after allopatric isolation leading to admixture among clades. Ecological niche modeling indicated relative stability in the extent of suitable distribution areas of the species across Quaternary glacial cycles. Our results underscore the importance of mountains in the diversification of this species, given that most of the distinct genetic clades are concentrated in a relatively small area in the Sino-Himalayan Mountain region, while a single shallow clade populates vast lower-lying areas to the east. This study highlights the crucial role of landscape heterogeneity in promoting differentiation and provides a deep genomic perspective on the mechanisms through which diversity hotspots form.

Introduction

Montane ecosystems act as cradles, barriers, bridges, and reservoirs for species and support a remarkable proportion of global terrestrial biodiversity (Rahbek et al. 2019; Perrigo et al. 2020). The geographical and environmental complexity of mountains is likely fundamental to the evolution of species and tightly associated with high biodiversity (Fjeldså et al. 2012; Price et al. 2014; Perrigo et al. 2020). Much research has revolved around speciation, genetic differentiation, and evolutionary diversification in montane systems, including such complex ecosystems as the Alps, the Andes, and the Himalayas (Schmitt 2009; White 2016; Winger 2017; Jardim de Queiroz et al. 2022; Múñoz-Valencia et al. 2023). Climatic fluctuations, geographical barriers, and the landscape heterogeneity of montane ecosystems are known to have profound effects on population genetic differentiation within a species (Rheindt et al. 2009; García-Llamas et al. 2018; Ryser et al. 2021; Manthey et al. 2022; Pujolar et al. 2022).

As one of the most complex montane ecosystems, the Sino-Himalayan Mountains (SHM) harbor exceptionally high species richness and endemicity, high topographic heterogeneity, and a range of climatic characteristics (Myers et al. 2000; Ding et al. 2020; Wambulwa et al. 2021). SHM extends along the southern and eastern margins of the Qinghai–Tibet Plateau (QTP), including the Mountains of Southwest China and Indochina (Cai et al. 2018). They have attracted considerable attention from evolutionary biologists to investigate the diversification process and species differentiation in this region (Li et al. 2013; Lei et al. 2014; Price et al. 2014; Sun et al. 2014; Schumm et al. 2020). The uplift of the Himalayas and QTP has led to a sharp increase in altitude from east to west within China, forming a series of parallel alpine ridges reaching elevations over 5,000 m above sea level (m.a.s.l) in southwest China (Zhang et al. 2009). The uplift of mountains promotes the formation of geographical barriers, which significantly accelerates the vicariance of species (Favre et al. 2015; Cai et al. 2018). Additionally, it substantially increases the available ecological niches, thereby promoting the gradual accumulation of alien species into SHM (Favre et al. 2015), which are also considered to be the main mechanisms for the formation of the diversity pattern of passerine birds in SHM (Johansson et al. 2007; Päckert et al. 2011; Cai et al. 2018). Such a topographical complexity also leads to a dramatic ecological stratification and heterogeneity of environments within the SHM (Qu et al. 2014). Climatic oscillations have been another driver of differentiation: glacial and interglacial cycles throughout the Pleistocene epoch have been considered an important factor in promoting floral and faunal diversification in SHM (Liu et al. 2012; Wang et al. 2013, 2018, 2019; Wang and Pierce 2023).

At the intraspecific level, ample research in this region has focused on exploring the role of glacial climatic oscillations in population formation or identifying geographic barriers related to differentiation across plant and animal taxa (Chen et al. 2011; Wang et al. 2013, 2018; Zuo et al. 2015; Liu et al. 2016a). Great strides have been made in DNA sequencing technology during the past 2 decades, helping to provide more comprehensive insights into species differentiation and population demography by utilizing whole-genome data (Luikart et al. 2003; Davey et al. 2011; Ellegren et al. 2012; Lamichhaney et al. 2015). Although some studies have explored the impact of landscape heterogeneity and environmental factors on population differentiation within the SHM regions, these investigations are few in number and limited in geographic scope (Chen et al. 2019; Hu et al. 2020; Yang et al. 2022; Wang and Pierce 2023; Yan et al. 2023).

Landscape genetics facilitates an understanding of how geographical and environmental features shape population structure (Manel et al. 2003). When genetic differences between populations are positively correlated with their geographic distance, populations exhibit a pattern of isolation by distance (IBD), a concept central to studies of genetic differentiation among populations (Wright 1946; Slatkin 1993). However, IBD does not consider the effects of landscape heterogeneity on differentiation. The concept of landscape heterogeneity includes the variety and spatial arrangement of land cover types, spanning different environmental conditions and levels of resource availability, and influences ecological processes (Tonetti et al. 2023). As an extension to IBD, McRae (2006) proposed the isolation-by-resistance (IBR) model based on circuit theory, hypothesizing that properties of gene flow in populations are analogous to conductance in linear electronic circuits. IBR is a modification of IBD that quantifies the resistance of individuals across the landscape to improve our understanding of how landscape characteristics affect genetic structuring (McRae 2006; McRae et al. 2008; Zhang et al. 2020; Wang and Pierce 2023). A central tenet of IBR is the relationship between genetic differentiation and resistance distances, which can be understood as the probability that an individual disperses from one site to another, weighted by friction to dispersal across unsuitable habitats and/or physical barriers (McRae 2006; Wang and Bradburd 2014). While the concept of isolation by environment (IBE) implies that genetic variation among populations is mostly explained by local environmental features rather than by spatial factors, and genetic differentiation between populations increases with environmental differences (Wang and Summers 2010; Wang and Bradburd 2014). IBE can be generated by multiple unique processes, such as natural selection and sexual selection against immigrants, reduced hybrid fitness, and biased dispersal (Wang and Bradburd 2014).

The red-billed leiothrix (Leiothrix lutea) is a medium-sized passerine belonging to the family Leiothrichidae, native to the SHM, southeast China, and adjacent areas. The species is sedentary and inhabits an altitudinal range up to 3,400 m (Male et al. 2020). There are 5 subspecies: Leiothrix lutea kumaiensis, Leiothrix lutea calipyga, and Leiothrix lutea yunnanensis range across the western, central, and eastern Himalayas, respectively, while Leiothrix lutea kwangtungensis is distributed from eastern Yunnan Province to northern Vietnam and Guangdong Province, and the nominate subspecies Leiothrix lutea lutea lives to the north of the latter, ranging from Sichuan Province to Fujian Province (Gill et al. 2023). Phenotypically, subspecies mainly differ in the depth of plumage colors, especially the color of the breast, abdomen, and base of primary flight feathers (Zheng et al. 1987). This species has been listed in the Convention on International Trade in Endangered Species (CITES) Appendix II (IUCN 2023) and key protected species list in China since 2021 due to high illegal trade risk.

In this study, we investigated the genetic structure and differentiation of L. lutea using genome-wide sequencing data across the species' major distribution range, with a particular emphasis on the SHM. We compared populations in terms of genetic diversity, demographic history, and shifts in suitable habitats across time among different subclades. The effects of geographical distance, environmental factors (including climate, altitude, and land cover types), and landscape heterogeneity on the genetic structure were assessed as well. We further compared phenotypic differences among the genetic clades based on morphometric parameters, plumage colors, and songs. Based on all these analyses from multiple angles, we attempt to explore the influence of complex topographic heterogeneity and various climatic characteristics on genetic differentiation patterns within the SHM system.

Results

Genome Sequencing and Variant Calling

We obtained whole-genome resequencing data from 72 individuals across the distribution of L. lutea (Fig. 1a; supplementary table S1, Supplementary Material online). The total data were 1,425.9 Gb of paired-end reads with an average of 19.8 Gb for each individual. The trimmed reads were mapped to the L. lutea reference genome (GCA_013400445.1, genome size = 1.1 Gb, scaffold N50 = ∼298 kb, contig N50 = 27.8 kb) and resulted in a mean coverage depth of 11.7×, a mean breadth of coverage of 94.0%, and a mean mapping rate of 86.7% (supplementary table S1, Supplementary Material online). We removed an individual (gz01) with a mapping rate below 60%, resulting in a data set of 11,076,689 single nucleotide polymorphisms (SNPs), reduced to 10,683,342 SNPs when only covering autosomes. Based on the results of kinship inference (see Supplementary Material), an individual in the related pairs sampled from Yadong County (xzyd01, Tibet) was excluded. Therefore, a total of 70 samples were retained in downstream analyses.

Fig. 1.

Distribution of sampling sites and population genetic structure of L. lutea. a) The distribution range of L. lutea in the wild is indicated by shadow. The solid triangles in the diagram represent each sampling site. The map is the elevation data downloaded from WorldClim (http://www.worldclim.org). b) PCA using genome-wide SNPs is performed in Plink to project 70 individuals into PC axis 1 and PC axis 2. c) Genetic structure is estimated using the clustering algorithm ADMIXTURE from K = 2 to K = 5.

Genetic Structure and Phylogeny of L. lutea

The phylogenetic tree inferred by 13 mitochondrial protein-coding genes showed no obvious population structure (supplementary fig. S1a, Supplementary Material online). In contrast, 5 separate genetic clusters were inferred by principal component analysis (PCA) using 10,683,342 SNPs of the autosomal data set (Fig. 1b). Along the first principal component (PC1), all individuals were divided into 3 clusters: East, Yunnan, and Tibet. The individuals from Tibet were further separated into 3 clusters along PC2, including Chayu, Motuo, and Southern Tibet (TibetS, including individuals from Cuona, Yadong, and Zhangmu). We also carried out population structure analysis using ADMIXTURE (Alexander et al. 2009) and set the number of postulated ancestry groups (K) from 2 to 6. When K = 2, the results indicated that individuals from Tibet and the East clustered into separate groups (Fig. 1c), while individuals from Yunnan exhibited admixture between Tibetan and Eastern components. Individuals from Cuona, Yadong, and Zhangmu in Tibet emerged as a separate cluster at K = 3, while individuals from Yunnan and Motuo showed signs of mixing. At K = 4 and K = 5, the individuals of Yunnan and Motuo formed their respective clusters, while mixed individuals emerged in Yadong and Gongshan (Yunnan) at K = 5. Two individuals from Motuo and 1 from Yunnan formed a new cluster when K = 6 (supplementary fig. S2, Supplementary Material online). The optimal K from cross-validation was 1, but based on the pattern in which clusters separated in PCA and other analyses, we divided populations into 5 genetic clades, including East, Yunnan, Chayu, Motuo, and TibetS.

Considering the mixed signal in ADMIXTURE analysis, we used HyDe (Blischak et al. 2018) for hybridization detection. A significant hybridization signal was detected in multiple populations (P < 0.05), including East, Yunnan, and Chayu (supplementary table S2, Supplementary Material online). FastTree2 (Price et al. 2010) was used to infer the maximum likelihood (ML) phylogenetic tree of all individuals based on concatenated SNPs of the autosomal data set. ML tree revealed the same lineage divergence pattern as confirmed by PCA (supplementary fig. S1b, Supplementary Material online). Except for the Yunnan population, the other 4 populations formed a monophyletic group, respectively. However, 3 individuals from Yunnan formed a paraphyletic group with other individuals, and this could be explained by that these individuals mixed with more components of other populations, according to the ADMIXTURE results.

Genetic Diversity and Population Divergence

We calculated nucleotide diversity and heterozygosity to compare the genetic diversity of different genetic clades. Heterozygosity was calculated for every individual, and the Wilcoxon rank-sum test was used for comparisons between groups (Fig. 2; supplementary table S1, Supplementary Material online). The individuals in the Yunnan group had the highest nucleotide diversity (π = 0.00201) and heterozygosity (H = 0.012), while TibetS had the lowest (π = 0.00186, H = 0.0087; supplementary table S3 and fig. S4, Supplementary Material online). The Wilcoxon rank-sum test showed that the heterozygosity of the Yunnan group was significantly higher (P < 0.05) than the other populations while there was no significant difference in heterozygosity among the 3 populations of Tibet (Fig. 2). The heterozygosity of the East population was significantly higher (P < 0.05) than that of Motuo and TibetS but significantly lower (P < 0.05) than that of the Yunnan population (Fig. 2). We compared the heterozygosity with other avian species: the results showed that L. lutea had the highest heterozygosity across species from different endangerment categories (supplementary fig. S3, Supplementary Material online). The overall Tajima’s D values of all L. lutea populations were positive, with the highest in the East (Tajima’s D= 0.816; supplementary table S3, Supplementary Material online). To quantify differentiation between different genetic clades, we calculated pairwise genetic differentiation (fixation index, F_ST) and absolute genomic divergence (d_XY). Genome-wide pairwise genetic differentiation F_ST between groups ranged from 0.012 to 0.036, but the F_ST value between the East and Chayu was obviously higher than between other group pairings (supplementary table S4 and fig. S4, Supplementary Material online). All group pairings had a similar magnitude of absolute genomic divergence d_XY (supplementary table S4 and fig. S4, Supplementary Material online).

Fig. 2.

Heterozygosity was estimated for all individuals within the 5 populations and compared using the Wilcoxon rank-sum test (***P < 0.001, **P < 0.01, and *P < 0.05).

Demographic History Reconstruction

Pairwise sequentially Markovian coalescent (PSMC; Li and Durbin 2011) was used to infer the long-term historical demography (Fig. 3). The effective population size (N_e) of the 3 populations from Tibet (Chayu, Motuo, and TibetS) began to decline slowly after reaching a peak around 0.4 Ma (Fig. 3). The Yunnan population maintained a relatively stable N_e after the same peak and declined from 0.1 Ma, while the East population experienced a brief increase and subsequent decline in N_e between 0.1 and 0.2 Ma, with the N_e of the latter 2 populations higher than that of the 3 Tibetan populations (Fig. 3). The N_e of Motuo and Chayu remained relatively stable and was consistently lower than that of the other 3 populations. The N_e of the East declined sharply after the Last Interglacial (LIG; ∼140 to 120 ka).

Fig. 3.

Demographic history reconstruction for 3 individuals from each population with a generation time of 2.5 yr and mutation rate of 0.046 per million years (Li et al. 2010; Smeds et al. 2016). PSMC reconstruction of the past demographic history with 100 bootstraps. LGM, Last Glacial Maximum; LIG, Last Interglacial.

We also used SMC++ (Terhorst et al. 2017) to infer the more recent demographic history covering a time range going back to ∼1 Ma (supplementary fig. S5, Supplementary Material online): in this analysis, the N_e of Chayu and TibetS reached a peak at 0.4∼0.5 Ma, consistent with the results of PSMC. The N_e of Motuo, Yunnan, and East reached a peak at about 80 ka, with the East and Yunnan populations declining after 80 ka to lower levels of N_e than those of the 3 populations in Tibet. SMC++ showed that after 10 ka, the N_e of the 3 Tibetan populations was higher than that of the Yunnan and East populations, which was contrary to the results of PSMC.

Joint Demographic History

To infer the patterns of isolation and gene flow among populations, we used the diffusion approximation method of δaδi (Gutenkunst et al. 2009) to analyze the 2D site frequency spectrum (2D-SFS) and the 3D site frequency spectrum (3D-SFS). The tested models were based on assumptions related to periods of isolation, migration rates, and population size changes. All models tested are shown in supplementary figs. S6 and S7, Supplementary Material online, with optimal models shown in Fig. 4. The detailed results of all demographic inference models are in supplementary table S5, Supplementary Material online. A total of 18 models were tested for 9 population pairs (2D) and compared using the Akaike information criterion (AIC). A suite of 3 secondary contact models was supported as the best fit within the 2D model paradigm, involving symmetrical or asymmetrical gene flow with or without changes in effective population size (Fig. 4a to c). For instance, the best model between Chayu and the other 4 populations involved symmetrical gene flow after secondary contact. For population pairs including the East population, optimal models involved changes in effective population size. The best-fit models for Yunnan in combination with TibetS and Motuo also involved changes in effective population size, whereas asymmetrical gene flow only occurred between East and Motuo (supplementary table S5, Supplementary Material online).

Fig. 4.

Simulation results of optimal population divergence models using 2D-SFS and 3D-SFS in δaδi. Models referenced to dadi_pipeline (Portik et al. 2017). a) Secondary contact after isolation divergence with continuous symmetrical size change. b) Secondary contact after divergence in isolation with continuous symmetrical and instantaneous size change. c) Secondary contact after divergence in isolation with continuous asymmetrical and instantaneous size change. d) The optimal model among 3D models means adjacent secondary contact after divergence in long isolation divergence. T1 and T2 in a) to c) indicate the split time and the secondary contact time, respectively. T1 and T2 in d) indicate the split time, and T3 in d) indicates the secondary contact time. nu1 and nu2 indicate the effective population sizes of population 1 and population 2. nu1a and nu2a indicate the effective population size before the instantaneous size change. nu1b and nu2b indicate the effective population size after the instantaneous size change. m indicates the migration rate between population 1 and population 2. m12 indicates the migration rate from population 1 to population 2. m21 indicates the migration rate from population 2 to population 1. m1 and m2 indicate the migration rate from population 1 to population 2 and the migration rate from population 2 to population 1.

We tested a total of 18 3D models for TibetS, Motuo, and Chayu from Tibet. The results of the simulations corroborated the results of the 2D model analysis and demonstrated secondary contact after isolation (Fig. 4d; supplementary table S5, Supplementary Material online). Comparisons of best-fit models and resulting residuals of the 2D-SFS and 3D-SFS are shown in supplementary fig. S8, Supplementary Material online. In conclusion, results of both 2D and 3D simulations indicated that gene flow has occurred between populations after isolation.

Historical Distribution Range Changes Projected by Ecological Niche Modeling

Ecological niche modeling (ENM; Phillips et al. 2006) was used to reconstruct suitable habitat for L. lutea for 4 evolutionary periods, including the Present, the Middle Holocene (MIH, ∼6 ka), the Last Glacial Maximum (LGM; ∼21 to 18 ka), and the LIG (∼140 to 120 ka). Model fit and accuracy were evaluated via area under the curve (AUC). The average training AUC was 0.929 (Present: 0.930, MIH: 0.929, LGM: 0.931, and LIG: 0.927). The suitable distribution of L. lutea shrank from the LIG to the LGM, especially in eastern China (Fig. 5a and b). The suitable habitat for LIG was mainly distributed in the western and southern Sichuan Basin, while the suitable distribution in the LGM was mainly located around the Sichuan Basin, the western Hengduan Mountains, and Southeast Tibet. The suitable distribution of the Yunnan population increased toward the LGM and then decreased toward the MIH, but the overall distribution range did not change significantly (Fig. 5b and c). The suitable range of the 3 populations in Tibet was consistently small, showing a linear longitudinal distribution along the Himalayan Mountains, and the change in the distribution range of the 3 populations mirrored that of the Yunnan population. The Present suitable distribution of the species was similar to that in the LIG. Overall, the effects of glacial cycles have been more pronounced on the East population than on other populations (Fig. 5e).

Fig. 5.

ENM predicting the suitable distribution of L. lutea. a) to d) represent the suitable distribution areas of L. lutea in the 4 historical periods: the LIG (∼140 to 120 ka), the LGM (∼21 to 18 ka), the MIH (∼6 ka), and the Present. Warmer colors represent higher distribution likelihoods. e) Superposition results of ENM of L. lutea for 4 historical periods. The deepest color denotes regions that emerged as most favorable for L. lutea during all 4 historical periods. In contrast, the lightest color denotes areas suitable for L. lutea during only 1 of these 4 periods.

Testing Isolation Models and Spatial Connection

We tested IBD, IBE, and IBR to explore the main factors affecting the genetic differentiation within L. lutea using ML population effects (MLPE) models (Clarke et al. 2002) and multiple regression on distance matrices (MRDM; Legendre et al. 1994). Euclidean geographic distance between sampling sites (supplementary table S6, Supplementary Material online) was used to test IBD. We performed PCA on a suite of environmental factors extracted from each sampling point and obtained 2 environmental distances (based on the PC1 and PC2) to test IBE. Lastly, a pairwise resistance matrix was calculated in Circuitscape v4.0 (McRae et al. 2016) to test IBR. A genetic distance matrix was calculated as a dependent variable using F_ST/(1 − F_ST). In our MLPE analysis, we used the AIC and Bayesian information criterion (BIC) to evaluate the fit of different models. According to the AIC and BIC ranking (Table 1), IBR was the optimal model to explain genetic structure (AIC = −1193.66, ΔAIC = 2.85; BIC = −1180.2, ΔBIC = 2.85). MRDM was used to further explore the relative contribution of models (Table 1; supplementary table S7, Supplementary Material online). The multivariable linear regression model that included all distance matrices explained up to 71.84% of the genetic distance (P < 0.0001), with resistance distance and geographical distance being the most important variables explaining genetic variation (partial regression coefficient β = 0.907 and −0.425, respectively; P < 0.01; Table 1). Both MLPE and MRDM indicated that resistance distance best explained genetic differentiation within L. lutea. The results of linear regression with genetic distance as the dependent variable and other variables as independent variables showed that resistance distance explained 64.66% of the genetic distance (supplementary table S6, Supplementary Material online).

Table 1.

Summary results of the MLPE mixed model and MRDM for model selection

Model	MLPE						MRDM
	AIC	ΔAIC	BIC	ΔBIC	Log-likelihood	Deviance	β	P(var)	R 2	P(model)
IBR	−1,193.6585	0	−1,180.2701	0	600.8293	−1,201.6585	0.9074	0.0001	0.7184752	0.0001
IBD	−1,190.8093	2.8492	−1,177.4208	2.8493	599.4046	−1,198.8093	−0.4253	0.0021	…	…
IBEPC1	−1,189.8030	3.8555	−1,176.4145	3.8556	598.9015	−1,197.8030	0.3163	0.0018	…	…
IBEPC2	−1,189.710	3.9485	−1,176.322	3.9481	598.855	−1,197.710	−0.0196	0.5998	…	…

The order of the models corresponds to their ranking from best (smallest AIC and BIC) to poorest. The first-ranked model is marked in bold. β is the partial regression coefficients and P(var) is the corresponding P value. R² represents the proportion of genetic distance explained by all models, and P(model) is the corresponding P value.

We used estimated effective migration surfaces (EEMS; Petkova et al. 2016) to visualize the spatial population structure and effective migration across the range of L. lutea (Fig. 6). Effective migration was generally high across the East, barring the effects of some barriers including the Qinling Mountains and the Wuyi Mountains (Fig. 6), which also showed the barrier effect of the montane system. Effective migration in other populations was lower than the global mean, with the Hengduan Mountains and the Himalayas constituting a formidable barrier to gene flow (Fig. 6). Spatial analysis of genetic diversity highlighted 2 main regions of exceptionally low diversity: one in the East and the other in TibetS. Populations in both these areas were also characterized by low effective migration. High spatial genetic diversity was detected only in a small part of Yunnan (supplementary fig. S9a, Supplementary Material online).

Fig. 6.

EEMS to visualize the spatial population structure and effective migration of L. lutea. Intensifying cool-toned shadows indicate higher effective migration while intensifying warm-toned shadows indicate lower effective migration. Solid dots represent the sampling sites corresponding to Fig. 1. The color of the dots represents different genetic clades, the size of the dots represents the number of sample points, and the larger the dots, the more samples.

Phenotypic Divergence

We used PCA and linear discriminant analysis (LDA) to analyze differentiation in morphometry, plumage colors, and songs (supplementary fig. S10, Supplementary Material online; Fig. 7). PCA results showed a certain degree of differentiation in the plumage coloration of wing spots and songs among populations but no differences in morphometry (supplementary fig. S10, Supplementary Material online). We further used LDA to maximize differences between populations. The results of LDA were unable to distinguish populations based on morphometric measurements across 161 specimens (Fig. 7a; supplementary table S8, Supplementary Material online).

Fig. 7.

LDA of phenotypic divergence among populations and the reflectance of wing spot. a) LDA is unable to distinguish populations based on morphometric measurements. b) The reflectance of wing spot feathers at 300 to 700 nm across 5 populations. c) LDA of plumage coloration of the wing spot shows that the East, Yunnan, and Tibet formed the 3 separate clusters without overlapping, while the 3 Tibetan populations show considerable overlap. d) LDA of song parameters across the 3 taxa can be distinguishable from one another.

Measurements of plumage reflectance were made for a total of 69 specimens from all populations (supplementary table S9, Supplementary Material online). We converted the spectral data into the amount of photons captured by the cone cell based on the tetrachromatic visual system of the bird and then performed LDA (Maia et al. 2013). Populations from the East, Yunnan, and Tibet formed 3 separate clusters without overlap, while the 3 Tibetan populations showed considerable overlap (Fig. 7c). The photon capture of ultraviolet wavelength was more related to linear discriminant 1. Plumage color and genetic distance showed a positive correlation trend, but the correlation was low and not significant (r = 0.02143, P = 0.39583; supplementary fig. S12, Supplementary Material online).

The songs of 15 individuals were analyzed (supplementary table S10, Supplementary Material online). The 3 populations (East, Motuo, and Yunnan) were distinguishable from one another in bioacoustics (Fig. 7d). Minimum frequency (F_min) had the highest correlation with linear discriminant 1, followed by average bandwidth. Similarly, there may be additional differences in song parameters collected from different provinces across the East (supplementary fig. S11e, Supplementary Material online). Based on this limited song data set, we detected vocal differentiation in L. lutea, but more material is needed for range-wide conclusions.

Discussion

The SHM is known as one of the global biodiversity hotspots and a complex mountain ecosystem, and numerous researchers have explored the mechanisms and factors that have contributed to the generation of its immense biodiversity (Wambulwa et al. 2021; Wu et al. 2022). In this study, we conducted extensive genomic and phenotypic sampling to investigate the population structure of L. lutea. Our findings highlighted the significant role of landscape heterogeneity, rather than geographic distance and local environmental factors, in shaping the current genetic structure of the species.

Using whole-genome analyses, we identified 5 genetic clades across the range of L. lutea within our sampling regime, including East, Yunnan, Chayu, Motuo, and TibetS. All clades except for the East have a large part of their range in the SHM. We did not find substantial genomic or phenotypic differentiation between L. l. lutea and L. l. kwangtungensis, the 2 traditional subspecies that make up almost the entire range of our East cluster. In contrast, L. l. calipyga, which was mainly distributed in the eastern Himalayas of southeast Tibet, was sorted into 3 genetic clades across 5 sampling locations. Interestingly, 3 of these Tibetan sampling locations (Cuona, Yadong, and Zhangmu) emerged as relatively undifferentiated, whereas Motuo and Chayu exhibited significant genetic differentiation (Fig. 1), suggesting that some geographic features in the eastern Himalayas were more conducive to genetic differentiation than others. For example, Motuo has a relatively lower altitude and more humid climate and has been considered an independent biodiversity hotspot in some taxa (Dufresnes and Litvinchuk 2022).

Previous studies have revealed patterns of longitudinal population subdivision along an East to West axis in many organisms native to the SHM region, particularly in forest passerines of the upper elevational belt (Liu et al. 2012, 2016b; Päckert et al. 2015; Zuo et al. 2015; Jiang et al. 2023). We detected a similar pattern of East to West differentiation in SHM populations of L. lutea, and the series of north to south mountains and rivers in the Hengduan Mountains may have played a major role in hindering the gene flow between East and the other populations. Our data also underscore the notion that the SHM region is a repository of deep genetic diversity: geographic distances among the 4 clusters in Tibet and Yunnan are much shorter than those between some of the sampling locations in the East, yet they are divided by deep genomic divergences. Our results show that the IBR model can best explain the genetic differentiation, and some potential landscape barriers may play an important role. The Yunnan is located in the southwestern Hengduan Mountains and may be separated from other populations by the Nushan and Mekong Rivers in the east and the Gaoligong Mountains, Salween River, and Irrawaddy in the west. Chayu is surrounded by mountains such as Gangrigabu Mountain and Bossula Ridge. The influence of warm and humid air currents from the Indian Ocean makes the climate humid. Motuo is located at the southern foot of the Himalayas and has a warm and humid climate. The Tibetan population is located in the Himalayas, at a higher altitude and with a colder and drier climate. The dense mountains in the SHM may also have hindered gene flow between populations. These geographical barriers, as well as differences in climatic conditions, lead to varying habitat conditions and may have promoted genetic differentiation between populations.

Phenotypic characteristics such as song and plumage color are known to be crucial for sexual selection in birds, while morphometric traits often reflect ecological factors (Byers and Kroodsma 2009; Dunn et al. 2015). We found no significant differences in morphometrics among populations (Fig. 7a). Yet there was slight diversification in plumage color and song (Fig. 7b to d), which to some extent was consistent with genomic data. The relationship between plumage color and genetic distance showed a positive correlation based on the limited amount of morphological data. Our vocal sampling was limited to 3 populations (supplementary fig. S11e, Supplementary Material online). Microgeographic song variation is widespread in oscine songbirds (Xing et al. 2013; Rodríguez-Fuentes et al. 2022), and more song data may help uncover patterns of song differentiation in L. lutea.

Multiple previous studies have explored genetic differentiation in this region, identifying primarily Pleistocene climate change, but also topographic heterogeneity and other environmental aspects, as important drivers of diversification (Li et al. 2013; Lei et al. 2014, 2015; Wang et al. 2018; Dong et al. 2020). Most European and North American species showed suitable distribution area contraction during the LGM and expansion afterwards (Milá et al. 2006). Our study demonstrated that L. lutea had the same pattern and underwent its most significant Pleistocene shifts in suitable distribution across the East (Fig. 5). Although the suitable distribution area of the other 4 populations—in the SHM—has also fluctuated, their total range size has remained much more stable across Pleistocene glacial cycles (Fig. 5). When compared to other avian species in the SHM region (Liu et al. 2012; Wang et al. 2018), L. lutea has been characterized by relatively smaller shifts in distribution.

The PSMC results showed that the effective population size of L. lutea declined after the LIG (Fig. 3), with N_e values generally higher in the East and Yunnan as compared to Tibet, corresponding to a higher genetic diversity of these 2 populations (Fig. 2; supplementary fig. S3, Supplementary Material online). The relatively small distribution range of the populations in Motuo and TibetS may have contributed to their lower genetic diversity and their consistently low N_e. The SMC++ results also revealed a decrease in N_e after the LIG but did not show an increase in N_e after the LGM except for the population in TibetS (supplementary fig. S5, Supplementary Material online). The prediction of the suitable distribution area in the 4 historical periods by the ENM did not completely correspond to the results of population dynamics history. It may be due to the possible dramatic shifts in elevation for this species in the Himalayas and the eastern edge of the QTP during the glacial cycles. On the other hand, gene flow between populations also affects the inference of population dynamic history.

Mitonuclear discordance is a common phenomenon in nature (Toews and Brelsford 2012; Zhang et al. 2019; Kato et al. 2021). We found mitonuclear discordance in L. lutea, manifested by significant genetic structure in autosomal DNA, while mitochondrial DNA showed no structure (supplementary fig. S1, Supplementary Material online). Such patterns can often be attributed to mitochondrial introgression (Rheindt and Edwards 2011; Toews and Brelsford 2012), as mitochondria have a smaller effective population size and are not affected by recombination and linkage (Funk and Omland 2003; Chan and Levin 2005). We used a variety of models to simulate the divergence history of L. lutea in δaδi. All models supported extensive gene flow between population pairs due to secondary contact, sometimes accompanied by changes in effective population size (Fig. 4; supplementary table S5, Supplementary Material online). Mitonuclear discordance in L. lutea is probably due to recent mitochondrial admixture.

Based on the population genomic structure of L. lutea, we used the landscape genetic approach to test which isolation pattern best explained genetic differentiation. We examined 3 separate models: IBD, IBE, and IBR. The least–cost pathway (LCP) is the other commonly used isolation model (Adriaensen et al. 2003), which accounts for characteristics of the landscape that may facilitate or impede movement along a single, optimal pathway, but it cannot incorporate multiple dispersal pathways connecting samples (McRae 2006) and was hence not further considered. Instead, we used EEMS to visualize the spatial population structure and highlight regions of higher-than-average and lower-than-average historical gene flow (Portik et al. 2017). Resistance distance calculations in IBR often utilize data related to species traits or habitats (Thomas et al. 2015; Weber et al. 2017). One common method is to calculate resistance distance based on niche model predictions, since gene flow between populations can be limited by suitable habitats (Myers et al. 2019; Zhang et al. 2020; Shu et al. 2022; Moreno-Contreras et al. 2023). In IBE and IBR analyses, based on L. lutea's elevational distribution range and preference for evergreen broadleaved forest and woodland, we considered 2 other variables in addition to the bioclimatic factors, namely altitude and land cover. Mantel test is one of the most popular approaches to evaluate spatial processes driving population structure (Diniz-Filho et al. 2013; Zhang et al. 2020), but it has been frequently critiqued because results were inconsistent with expectations and there were issues with type I (false-positive) and type II (false-negative) error rates (Legendre et al. 2015; Somers and Jackson 2022). Therefore, we used MLPE and MRDM to determine which model better explains genetic variation in L. lutea (Moreno-Contreras et al. 2023).

The results showed IBR best explained the genetic structure of L. lutea, which suggests that landscape heterogeneity impedes gene flow and promotes the formation of population genetic structure. Resistance to gene flow based on landscape features seemed to be more reflective of genetic variation than the distribution of these features themselves. The EEMS results identified individual geographic features that have impeded gene flow among populations of L. lutea, mainly in the SHM (Fig. 6). Some montane system ranges in the eastern have also shown an impedimentary effect on gene flow but not enough to produce large genetic differentiation in populations. These results demonstrate the important role of SHM heterogeneous landscapes, including climatic and geological characteristics, in promoting species differentiation and diversification.

As a global biodiversity hotspot and a complex mountain system, SHM has extremely high biodiversity (Qu et al. 2014; Lei et al. 2015), but the diversification mechanisms of different taxa in SHM are different. For pheasants, Phylloscopus, and Seicercus warbler, their high species richness in the SHM area is mainly due to the colonization by alien species (Johansson et al. 2007; Cai et al. 2018). But Paridae and babblers originated in the SHM and had a long time for diversification (Johansson et al. 2017; Cai et al. 2020). The Leiothrichidae that L. lutea belonged to also likely originated in the SHM and underwent diversification. Our population genomic results on L. lutea support the theory that these mountains constitute a cradle of diversity and underscore the importance of the SHM as a center for lineage diversification (Liu et al. 2016b; Cai et al. 2018; Päckert et al. 2020) of the early stage during speciation.

In conclusion, our study reveals at least 5 geographic lineages within L. lutea, which are not congruent with traditional subspecies boundaries. Importantly, the SHM region harbors more genetic lineages within a comparatively smaller area than more expansive areas in the East. The resistance to gene flow created by certain landscape features promotes genetic differentiation within L. lutea. The complex topography and climatic heterogeneity of the SHM region within the montane system may play an important role in lineage formation and diversification. The SHM region is likely to act as a differentiation pump in L. lutea and many other organisms in the region. Whole-genome–based phylogeography studies of multiple species in this region are eagerly anticipated to shed further light on the mechanisms that may have led to its remarkable biodiversity.

Materials and Methods

Sampling and Whole-Genome Sequencing

We sampled 63 individuals of L. lutea from 2004 to 2018, deposited in the Institute of Zoology, Chinese Academy of Sciences (supplementary table S1, Supplementary Material online). Whole-genomic DNA was extracted from muscle tissue and blood using the Tissue/Cell Genomic DNA Extraction kit (Aidlab Biotechnologies Co., Ltd., Beijing, China) following the manufacturer’s protocol. DNA libraries were constructed with ∼350 bp insertions and sequenced with Biomarker Technologies (Beijing, China) using an Illumina NovaSeq 6000 platform with a paired-end read length of 150 bp. In addition, we added the sequencing data of 9 Yunnan individuals from F.D.'s unpublished data. Raw read data used in this study are publicly available in the NCBI Sequence Read Archive (SRA) under BioProject PRJNA977805.

Variant Calling

We used FastQC v0.11.8 (Andrews 2010) to assess read quality and Fastp v0.21 (Chen et al. 2018) to trim poor-quality reads and Illumina adapters with default settings. Quality-controlled reads were mapped to a L. lutea reference genome (GCA_013400445.1; Feng et al. 2020) using BWA v0.7.17 (Li and Durbin 2009). BAM files were sorted, duplicates were removed, and files were indexed using SAMtools v1.3.1 (Li et al. 2009). We called variants using HaplotypeCaller in GATK v4.0.9 (McKenna et al. 2010). The resulting gvcf files were merged using CombineGVCFs and genotyped with GenotypeGVCFs in GATK v4.0.9 (McKenna et al. 2010). We used VCFtools v0.1.13 (Danecek et al. 2011) to remove indels and filter VCF according to the following criteria: (i) keep only biallelic SNPs; (ii) minimum quality value > 30; (iii) minimum coverage depth > 5 and maximum coverage depth < 100; (iv) missing ratio across all individuals < 5%; and (v) minor allele frequency ≥ 5%.

We only retained autosomal SNPs to avoid biases associated with sex chromosomes. We inferred scaffolds located in chromosomes by assuming conserved synteny with Parus major and applied BLAST+ 2.2.26 to align our reference genome against the P. major genome (GCA_001522545.3) with an e < 10⁻⁴⁰. We kept the longest alignment if the same scaffold was repeatedly aligned to different chromosomes. We retained scaffolds located in autosomes and obtained an autosomal data set.

Population Structure and Phylogenetic Analysis

For population structure reconstruction, we pruned autosomal SNPs at high linkage disequilibrium (LD) across all individuals using Plink v1.9 (Purcell et al. 2007). We set the independent pair-wise filter at a correlation threshold of 0.1 for a window size of 50 kb and a step size of 1 kb. We performed PCA using Plink v1.9 (Purcell et al. 2007). ADMIXTURE v1.3.0 (Alexander et al. 2009) was run to analyze population structure with cross-validation and 100 bootstraps, using K values ranging from 2 to 6.

We inferred the phylogenetic tree using 13 mitochondrial protein-coding genes and the nuclear whole-genome autosomal data set. Beast v2.5.1 (Bouckaert et al. 2014) was used to infer phylogenetic trees with a clock rate of 2.1% per million years for CYTB (Weir and Schluter 2008) using 13 partitioned mtDNA protein-coding genes (see details in Supplementary Material). FastTree2 (Price et al. 2010) was applied to reconstruct an ML tree based on the general time reversible (GTR) model of nucleotide substitution using concatenated SNPs of the autosomal data set. Leiothrix argentauris was used as the outgroup (SRR15193423). The script Vcf2phylip.py was used to convert the vcf format to fasta format (Ortiz 2019).

Genetic Diversity and Population Divergence

The final BAM files were used to estimate the heterozygosity of every individual. SAMtools v1.3.1 (Li et al. 2009) and BCFtools v1.3.1 (Li 2011) were employed to generate consensus sequences with a range of sequencing depth from 5 to 100. The tools vcfutils.pl (Li et al. 2009) and seqtk (https://github.com/lh3/seqtk) were used to transfer the vcf files into fasta format. We counted the number of heterozygous sites in each genome and calculated heterozygosity using the total number of heterozygous sites divided by the sum of homozygous and heterozygous sites. We used the Wilcoxon rank-sum test to compare the heterozygosity among groups. Tajima’s D, nucleotide diversity (π), and population genetic differentiation (F_ST) between each pairwise combination of populations were calculated using VCFtools v0.1.13 (Danecek et al. 2011) based on a 50 kb nonoverlapping sliding window approach. The Python script egglib_sliding_window.py developed by Simon Martin (https://github.com/simonhmartin/genomics_general) was used to estimate the absolute genomic divergence (d_XY) for each population pair using the sliding window approach, with a nonoverlapping window size of 50 kb and at least 5 sites within each window.

Demographic History Reconstruction

To reconstruct the historical demography of L. lutea populations, we applied 2 coalescent simulation methods. The PSMC model employs allele distribution information based on recombination events (Li and Durbin 2011). Due to the limited number of recombination events, the power of PSMC in inferring recent demographic events is poor. In contrast, SMC++ combines the sampling frequency spectrum data with LD and can analyze hundreds of unphased whole genomes (Terhorst et al. 2017), which is better suited to simulating demographic events of the more recent past. For both methods, we used a mutation rate per site per generation of 4.6 × 10⁻⁹ (Smeds et al. 2016), and 2.5 yr was specified as the generation time (Li et al. 2010; see details in Supplementary Material).

Joint Demographic History

To infer patterns of isolation and gene flow among populations, we used the diffusion approximation method of δaδi (Gutenkunst et al. 2009) to analyze the 2D and 3D folded joint site frequency spectra (2D-JSFS, 3D-JSFS). Although δaδi considered the role of natural selection, it did not consider the effects of linkage effects (Gutenkunst et al. 2009). Therefore, we used the LD filtered data in δaδi. We examined 18 alternative 2D models between all 9 possible pairings of populations (supplementary fig. S6, Supplementary Material online) to test hypotheses including a simple divergence model, ancient migration, and secondary contact. We also examined 16 alternative 3D models for the 3 populations in Tibet (including Chayu, Motuo, and TibetS) and tested hypotheses including simultaneous splitting, secondary contact, ancient migration, and nonsimultaneous splitting (supplementary fig. S7, Supplementary Material online). Both the 2D and 3D models differed in their assumptions as related to migration rates, periods of isolation, and population size changes (supplementary figs. S6 and S7, Supplementary Material online) following the δaδi pipeline (Portik et al. 2017; Barratt et al. 2018). All models were set to 4 rounds and generated a final set of 40 replicate optimizations to estimate the log likelihood. The AIC of the replicate with the highest likelihood for each model was used to compare the models.

ENM

Maximum entropy species distribution modeling in MAXENT v3.3.3 (Phillips et al. 2006) was used to predict range shifts of suitable distribution using bioclimatic variables across 4 periods: the Present, the MIH (∼6 ka), the LGM (∼21 to 18 ka), and the LIG (∼140 to 120 ka). We downloaded 19 bioclimatic variables from WorldClim (Hijmans et al. 2005; http://www.worldclim.org/). All bioclimatic rasters were clipped into the area set in 15 to 40°N to 70 to 125°E. To prevent overfitting, we retained 8 bioclimatic factors that were not highly correlated with each other (correlation coefficient < 0.8). The modern occurrence of L. lutea was defined by sampling sites and museum locality records from the National Animal Collection Resource Center as well as GBIF data (https://doi.org/10.15468/dl.y5yr96). A total of 785 localities were kept after filtering. A total of 80% of the occurrence data were randomly selected to construct models, and the remaining 20% were used to test the models with the default convergence threshold (10⁻⁵) and 5,000 maximum iterations. We executed 10 replicates with the replicated run type as bootstrap and left other parameters at default. Model fit and accuracy were evaluated by AUC (see details in Supplementary Material).

Testing Isolation Models and Spatial Connectivity

To investigate whether population structure and patterns of gene flow are related to geographical barriers and landscape heterogeneity, we assessed IBD, IBE, and IBR. The pairwise genetic distance matrix [F_ST/(1 − F_ST)] of sampling sites was calculated in VCFtools v0.1.13 (Danecek et al. 2011). Euclidean geographic distance between sampling sites was obtained from coordinate data using the R package “geosphere.”

For environmental variables, in addition to the 19 bioclimatic factors of the Present used in the above ENM analysis, we downloaded elevation data from WorldClim (Hijmans et al. 2005; http://www.worldclim.org/) with a resolution of 30 arc-seconds as well as land cover data from the GlobCover 2009 Project (Arino et al. 2012). Finally, 10 factors (correlation coefficient < 0.8) were used for each sampling site to obtain environmental distances by PCA. We extracted the first 2 principal components (PC1 and PC2) to further calculate the environmental distance matrix (ENV_PC1 and ENV_PC2).

For the resistance distance in IBR calculations, we inferred environmental niches to generate resistance surfaces because potential routes of dispersal and gene flow among populations are likely restricted by suitable habitat (Myers et al. 2019; Zhang et al. 2020). To include as many landscape characteristics as possible, we used bioclimatic factors, elevation, and land cover to calculate the niche model. The pairwise resistance matrix was calculated in Circuitscape v4.0 (McRae et al. 2016) by taking the resulting raster from ENM as conductance, with the assumption that high habitat suitability leads to low resistance (Zhang et al. 2020).

We used MLPE models (Clarke et al. 2002) in the R package “ResistanceGA” (Peterman and Jarman 2018) to test which type of distance best explained genetic differentiation. We further performed MRDM in the R package “ecodist” (Goslee and Urban 2007) to infer the relative contribution of different types of distances to genetic distance (Legendre et al. 1994). We applied EEMS (Petkova et al. 2016) to visualize spatial patterns in effective migration and genetic diversity. The surfaces of effective migration rates and effective diversity were generated by the R package “rEEMSplots” (Petkova et al. 2016; see details in Supplementary Material).

Phenotypic Analysis

We analyzed morphometric traits, plumage colors, and songs to determine the phenotypic divergence among populations. The specimens used for measurement were from the National Animal Collection Resource Center (National Animal Collection Resource Center 2023). A single person (X.J.) completed all measurements to reduce observer bias. We performed PCA and LDA for 3 phenotypic characteristics. We used the “princomp” function in R to perform PCA and the “lda” function in the R package “MASS” (Venables and Ripley 2002) for LDA. We plotted the spectrometric distances of wing spot coloration against genetic distance to explore the correlation between phenotype and genetic distance. Details on phenotypic measurements can be found in the Supplementary Material.

Supplementary Material

Supplementary material is available at Molecular Biology and Evolution online.

Author Contributions

The study program was conceived by G.S. and F.L. G.S., F.L., D.Z., F.D., and C.J. led the fieldwork and sample collection. X.J. carried out the lab work, morphological scaling, plumage color, vocal measurements, and data analysis. L.W., D.Z., H.W., and S.W. contributed to data analysis. G.S. and X.J. led the early drafting. F.L., F.E.R., L.Y., H.E.A., and W.Z. contributed to manuscript writing and improvement. All authors assisted in the interpretation of the results and commented on the manuscript.

Funding

This work was supported by the National Key R&D Program of China (2022YFC2602204, 2022YFC2601601), National Science Foundation of China (32130013, 31630069), and the Central Guidance on Local Science and Technology Development Fund of Tibet Autonomous Region (XZ202201YD0015C). The scientific field expeditions to obtain samples were funded by the State Basic Research Program (2019QZKK05010112).

Data Availability

Whole-genome sequencing data generated in this study have been deposited into the NCBI Sequence Read Archive under a BioProject portal of PRJNA977805. Sample accessions are included in supplementary table S1, Supplementary Material online. The code and scripts used in this analysis have been uploaded to GitHub (https://github.com/xljiao/leiothrix_lutea).