Genome-wide patterns of divergence and introgression after secondary contact between Pungitius sticklebacks

Abstract

Speciation is a continuous process. Although it is known that differential adaptation can initiate divergence even in the face of gene flow, we know relatively little about the mechanisms driving complete reproductive isolation and the genomic patterns of divergence and introgression at the later stages of speciation. Sticklebacks contain many pairs of sympatric species differing in levels of reproductive isolation and divergence history. Nevertheless, most previous studies have focused on young species pairs. Here, we investigated two sympatric stickleback species, Pungitius pungitius and P. sinensis, whose habitats overlap in eastern Hokkaido; these species show hybrid male sterility, suggesting that they may be at a late stage of speciation. Our demographic analysis using whole-genome sequence data showed that these species split 1.73 Ma and came into secondary contact 37 200 years ago after a period of allopatry. This long period of allopatry might have promoted the evolution of intrinsic incompatibility. Although we detected on-going gene flow and signatures of introgression, overall genomic divergence was high, with considerable heterogeneity across the genome. The heterogeneity was significantly associated with variation in recombination rate. This sympatric pair provides new avenues to investigate the late stages of the stickleback speciation continuum.

This article is part of the theme issue ‘Towards the completion of speciation: the evolution of reproductive isolation beyond the first barriers’.

1. Introduction

In the majority of speciation events, reproductive isolation is a continuous process rather than an all-or-nothing phenomenon, except for a few cases such as polyploid speciation [1]. Therefore, a given species pair can be placed along an axis, the ‘speciation continuum', given the strength of reproductive isolation or the rate of on-going gene flow between them [1,2]. Recent studies of the speciation continuum have suggested that the transition from weak to strong reproductive isolation may occur suddenly [3,4]. However, whether these so-called ‘tipping points' occur in the majority of speciation events is unknown [5]. Furthermore, although it is well known that ecological adaptation to different environments can initiate divergence even in the presence of gene flow [6–8], the factors that complete speciation remain largely elusive [9,10].

A period of geographical isolation is one of the factors that can promote the completion of speciation [10,11]. First, Bateson–Dobzhansky–Muller (BDM) incompatibility loci will evolve between allopatric populations in a snow ball fashion as a result of genetic drift [12]. Theoretical studies have shown that even a small amount of gene flow can substantially slow down the establishment of genetic incompatibilities [13,14]. Second, the lack of gene flow may increase the fixation probability of selfish elements and suppressor genes within each allopatric population, thereby further driving the establishment of intrinsic genetic incompatibilities [15]. Third, divergently adapted alleles will be more easily fixed in the absence of gene flow [16], which can also drive genomic and ecological divergence. Finally, once postzygotic isolation has been established during allopatry, secondary contact may reinforce prezygotic barriers to prevent inter-specific mating that can produce maladapted progeny [17,18].

A few empirical studies have tested these verbal and theoretical arguments. For instance, allopatric species pairs exhibit stronger hybrid abnormalities than sympatric species pairs in marine reef fishes [19]. By contrast, allopatric and sympatric species of Drosophila species pairs do not differ in the strength of hybrid abnormalities, although pre-zygotic isolation is stronger in sympatric pairs as suggested by theory [20]. Although such comparative studies between currently sympatric and allopatric species pairs are helpful to infer the roles of allopatric phases in speciation, it is usually unclear whether currently sympatric species experienced a period of allopatry in the past or for how long currently allopatric species have been isolated [21,22]. Determining the period of geographical isolation is even more difficult in species pairs that diverged at more ancient times, due to the fact that ancient geological events are generally more difficult to reconstruct than recent events. Recent advances in genomic technologies, however, provide us opportunities to test whether primary divergence or secondary contact is more plausible and to estimate divergence times using genome-wide sequence data [4,23–25]. For instance, the sympatric Japanese stickleback species pair of Gasterosteus aculeatus and G. nipponicus was originally considered to have arisen via allopatric speciation [26,27]. However, a more recent demographic study using whole-genome sequences has shown that some gene flow has occurred between them for the majority of time since their divergence [28]. This illustrates the importance of conducting a detailed demographic analysis using genome-wide sequence data [29–32].

Genome-wide sequence analysis further enables us to understand how genomic divergence proceeds during speciation [29,33,34]. Because speciation generally proceeds with genomic divergence, it is essential to understand how genome-wide differentiation and divergence arises during speciation [29,33,34]. The genic view of speciation predicts that the initial stage of divergence with gene flow is characterized by genetic differentiation only at barrier loci and the nearby linked loci, while other regions show low levels of differentiation due to gene flow [8,35]. Theoretical studies further predict that when the effective migration rate is reduced to some level, genome-wide congealing is expected to occur [5,36]: after this tipping point, a positive feedback between divergent selection and linkage disequilibrium will promote further overall genomic divergence [5,36,37]. Empirical genomic data has demonstrated heterogeneous patterns of genomic differentiation in many young species pairs with gene flow, consistent with the genic view of speciation [5,29,33,38–40]. By contrast, we know less about genomic patterns of divergence at late stages of speciation, except in a few cases [28,41].

Stickleback fishes contain many sympatric species pairs differing in the levels of reproductive isolation and divergence history [2,6,42]. Therefore, they are a great model to investigate how speciation progresses towards completion. However, the majority of previous studies have focused on species pairs with young divergence times that lack irreversible intrinsic incompatibility [27,43]. The genomic studies of such young stickleback species pairs mostly find restricted divergence at adaptive loci [39,44]. One exception is a sympatric pair of two Gasterosteus stickleback species in Japan, the Japan Sea stickleback (Gasterosteus nipponicus) and a Pacific Ocean lineage of the three-spined stickleback (G. aculeatus) [27,28,43]. There is strong reproductive isolation between these species, including hybrid male sterility [27,43], with a very low migration rate (1.05–1.30 × 10⁻⁶) [28]. These two species were estimated to have diverged approximately 0.68 Ma [28]. This is much older than the majority of the three-spined stickleback species pairs, which mostly diverged after the end of the last glacial period (approx. 12 000 years ago) [45]. Although the two Japanese species have exchanged migrants for the majority of time since their divergence, there are high levels of genome-wide divergence and limited regions of introgression between them [28]. It is unknown whether this pattern of genomic divergence is representative of the late stage of stickleback speciation.

Although the sympatric Japanese Gasterosteus species pair is the oldest species pair with on-going gene flow reported thus far among sticklebacks, the closely related genus Pungitius may contain other old species pairs, offering unique opportunities to investigate late stages of speciation [46,47]. In the present study, we focused on a Japanese sympatric pair, P. sinensis and P. pungitius. P. pungitius, the nine-spined stickleback, is widely distributed in both freshwater and brackish-water habitats across the Northern Hemisphere, whereas P. sinensis is endemic to East Asia and found mainly in freshwater habitats (figure 1a). These two species overlap in eastern Hokkaido, Japan [47,48]. Previous studies uncovered hybrid male sterility in both directions of hybridization [49,50], suggesting the possibility that they are at a late stage of speciation. However, a detailed demographic history and genomic patterns of divergence between them are unknown.

Figure 1.

(a) Present distribution and sampling sites of Pungitius pungitius (red) and P. sinensis (blue). (b) Individual strength and relative contribution of each isolating barrier. (Online version in colour.)

The aim of the present study was to characterize the past demographic history of sympatric P. sinensis and P. pungitius and analyse the genomic patterns of divergence as a first step towards a better understanding of the late stages of speciation. To this end, we first quantified the strength of the isolating barriers in this species pair. Next, we constructed a de novo genome assembly of P. sinensis for use as a reference genome sequence for population genomics. Finally, we inferred their demographic history and analysed their patterns of genomic divergence and introgression.

2. Material and methods

(a). Components of reproductive isolation

The contributions of each isolating barrier were calculated using the methods described previously [51,52]. To calculate the strength of eco-geographical and temporal isolation, data of spawning sites and seasons previously obtained by underwater observation with snorkelling at a sympatric site (Shiomi River, Hokkaido, Japan in 2003 May–July) were used [53] (electronic supplementary material, table S1). To calculate the strength of sexual isolation, the mating successes of the pairs with the same species and different species were obtained from previous no-choice mating trials [54]. Genetic and gametic inviability and hybrid sterility data were obtained from previous artificial insemination experiments [49]. It should be noted that P. pungitius and P. sinensis were referred to as the brackish-water and freshwater types in these studies, respectively.

We performed calculations following Lackey & Boughman [52]. Briefly, the strength of reproductive isolation of each barrier was calculated as

$$1-2\left(\frac{H}{C+H}\right),$$

where H is the frequency of heterospecific events (e.g. fertilization, survival and progeny production) and C is the frequency of conspecific events. We next calculated sequential strength of each barrier by ordering isolating barriers sequentially during the life cycle. Total isolation was then calculated as the cumulative effect of all sequential isolating barriers. Finally, the relative contribution of each barrier was calculated by dividing the sequential strength of each barrier by total isolation.

(b). De novo genome assembly

To construct a reference genome sequence of the genus Pungitius, we used one male of a fourth-generation inbred line of P. sinensis collected from the Oboro River, Hokkaido, Japan. Details of the methods used are available in the electronic supplementary methods. Briefly, genomic DNA was sequenced using PacBio Sequel (about 48.6 Gb) (Pacific Biosciences, Menlo Park, CA, USA) and Illumina HiSeq (Illumina, San Diego, CA, USA) in the 300 bp-paired end mode (about 55.4 Gb) (electronic supplementary material, table S2). The HGAP4 assembler was used for obtaining contigs [55,56]. For correcting the sequences, Illumina short reads were mapped to the contigs by bwa-mem [57] and then polished once by Pilon v. 1.22 [58]. Polished contigs were scaffolded with Irys (Bionano Genomics, San Diego, CA, USA). For detecting contamination, a homology search of the scaffolds against bacterial sequences was conducted using BLAST v. 2.2.9. We removed the scaffolds with > 1% of sequences matched to the bacterial genome and mitochondrial sequences from the RefSeq database. Cleaned scaffolds were aligned and oriented using a linkage map constructed with double digest RAD-seq data (electronic supplementary material, methods). The reference genome was annotated with the coding sequences of G. aculeatus using GenomeThreader Gene Prediction Software v. 1.7.1 [59]. Completeness of the assembled genome was calculated using BUSCO implemented in the gVolante v. 1.2.1 web server with 4584 actinopterygian core genes [60,61]. The whole-genome assembly is available from Dryad (https://doi.org/10.5061/dryad.pk0p2ngkd).

(c). Fish sampling and whole-genome re-sequencing

The sampling of the sticklebacks used for whole-genome re-sequencing was described previously (figure 1a; [47,48,62]). Briefly, sympatric P. pungitius and P. sinensis were collected in the Shiomi River, Hokkaido, Japan (N = 12 for each species) [48]. For an allopatric P. sinensis population, we used the previously reported whole-genome sequence data from the Yuza population in the Yatsume River, Yamagata Prefecture, Japan (N = 12) [63]. Additionally, we collected an allopatric P. sinensis population (Musashi) in the Motoarakawa River, Saitama, Japan (N = 4) [62]. To estimate recombination rates, we used P. sinensis collected in the Bekanbeushi River, Hokkaido, Japan (N = 16). For the topology analysis of a phylogenetic tree (see below), we used laboratory strains of allopatric P. pungitius populations collected at Isefjord (N = 2) and Copenhagen (N = 2) in Denmark [47]. Because these allopatric P. pungitius fish are progeny of laboratory crosses, we used these fish only for the f₄ statistic and topology analyses.

Genomic DNA was isolated using DNeasy Blood & Tissue Kits (Qiagen, Hilden, Germany). Sequence libraries were constructed using NEB Next Ultra DNA Library Preparation Kit with each fish indexed with a barcode (NEB, Ipswich, MA, USA). Libraries were run on Ilumina HiSeqX in 150 bp × 2 paired-end mode (Illumina, San Diego, CA, USA) at Macrogen Japan (Kyoto, Japan). The coverage and accession numbers are provided in electronic supplementary material, tables S3 and S4. The resulting short reads were trimmed and mapped to the P. sinensis reference sequence, using CLC Genomics Workbench 10.1.1 (Qiagen, Hilden, Germany) using the same parameters as described previously [62]. Mapping rates (mean ± sd. = 94.49% ± 0.009) and read depths (19.12 ± 2.41) of P. pungitius sequences were slightly lower than those of P. sinensis (97.38% ± 0.007 for mapping rate and 26.88 ± 7.85 for read depth electronic supplementary material, table S4). Next, the BAM files were exported and used for subsequent analyses.

Polymerase chain reaction duplicates were detected using GATK v. 4.1.2 [64]. Three types of single-nucleotide polymorphism (SNP) calls were conducted. First, for analyses requiring only variant sites (SNP-only dataset), we called SNPs using the GATK HaplotypeCaller and GenotypeGVCF following the GATK Best Practices documentation [65]. Then, filtering was performed using vcftools v. 0.1.17 [66] with the following parameters: minQ = 30, max meanDP = 36.6 (twice the mean depth in GATK calling), remove-indels, maf 0.05, max-alleles 2, minDP 8, max-missing 0.8. This resulted in 8 316 192 SNPs. Second, we used bcftools v. 1.9 mpileup to call both SNPs and invariant sites (all-site dataset) when the analysis required both variant and invariant sites [67]. For the allele frequency spectrum analysis (AFS), we then conducted the following filtering procedures using vcftools: remove-indels, max-meanDP 33.826 (twice the mean depth in bcftools calling), minDP 8, minQ 30, max-missing 1 and max-alleles 2. This resulted in 135 722 934 sites with 5 497 550 SNPs. For genomic divergence analysis, we used the following filtering: remove-indels, max-meanDP 33.826, minDP 8, minQ 30, max-missing 0.8, max alleles 2, maf 0.05 and mac 1, resulting in 212 533 216 sites with 13 229 682 SNPs. Third, we used bcftools v. 1.9 and bamcaller.py in MSMC Tools (https://github.com/stschiff/msmc-tools) for the pairwise sequentially Markovian coalescent (PSMC) analysis. We filtered out any sites with a minimum mapping quality below 20, a minimum base quality below 20, and a depth larger than twice the chromosome-wide average per individual. In all analyses, we removed the sex chromosomes (Chromosome 12) because the inter-population genetic differentiation can be confounded by genetic differentiation between the X and Y (P. pungitius) and/or Z and W chromosomes (P. sinensis) [50,63].

(d). Population structure and demographic analysis

Past population size change was estimated from each fish using the PSMC' implemented in MSMC2 [68] with default settings. We assumed 1 year for the generation time and 7.1 × 10⁻⁹ for mutation rate per site per generation [28,69]. When we calculated the mean and standard deviation of N_e, we did not consider the two most recent and three oldest time segments due to the low accuracy at recent and old ages [70].

We conducted AFS analysis to estimate the demographic histories during speciation between sympatric P. sinensis and P. pungitius in Shiomi River. We first calculated folded AFS of the all-site dataset using easysfs.py (https://github.com/isaacovercast/easySFS) and obtained the multi-dimensional data of the observed AFS. We used fastsimcoal2 v. 2.6 for the demographic analysis [23]. We examined six demographic models (electronic supplementary material, figure S1): no gene flow; unidirectional gene flow from P. sinensis to P. pungitius; unidirectional gene flow from P. pungitius to P. sinensis; continuous bidirectional gene flow; bidirectional gene flow after secondary contact; no gene flow after ancestral bidirectional gene flow. Because the PSMC' analysis showed that P. pungitius has a substantially changed effective population size (N_e) (see below), we added a parameter of population size change for P. pungitius. For divergence time estimation, a uniform distribution was applied: because divergence time is likely to be over 1 Ma, it was set to range between 1 × 10⁶ and 5 × 10⁶. For other parameters, a log-uniform distribution was applied. We used 7.1 × 10⁻⁹ for the mutation rate per site per year as described above. It should be noted that the migration rate in models of fastsimcoal2 are coded as backwards in time under the coalescent. However, we described the migration rate as forward in time throughout this paper. Further details are available from Dryad (https://doi.org/10.5061/dryad.pk0p2ngkd).

Following previous work [30,71], we performed 100 independent runs of 100 000 coalescent simulations for each model. A simulation with a maximum-likelihood value among 100 runs was used for the Akaike information criteria (AIC) model selection, where a model with the lowest AIC value was selected as the best. Using approximated likelihood estimation in fastsimcoal2, we obtained a likelihood distribution for each model by calculating the likelihoods of 100 expected AFS under the parameters that maximize the likelihood. If the distributions do not overlap among the models, we judged them as significantly distinguishable. We obtained 95% confidence intervals (95%CI) of the respective parameters for the best model using non-parametric block bootstrapping. Briefly, we split the genome into 1 Mb windows, randomly resampled windows with replacement allowed, and obtained 100 bootstrapped observed AFS. For the sampled AFS, we conducted 10 independent runs for the maximum-likelihood estimation of the parameters. Using the 100 estimates, we calculated the 95%CI.

(e). Genome-wide analysis of divergence and introgression

ADMIXTURE analysis was conducted to analyse the genetic structure. Linked SNPs were removed from the SNP-only dataset using Plink v. 1.9 with the parameter --indep-pairwise 50 10 0.1 [72]. The ADMIXTURE program was run with K values ranging from 1 to 10, and cross-validation error was calculated to select the best K [73].

To investigate genome-wide differentiation, we calculated Weir and Cockerham's F_ST [74] in non-overlapping 50 kb windows with vcftools using the SNP-only dataset. We also calculated the genetic diversity within a population (π) and genetic divergence between populations (d_XY and d_A) in non-overlapping 50 kb windows using the all-site dataset with popgenWindow.py script (https://github.com/simonhmartin/genomics_general) [38]. To examine whether variation in recombination rate could account for the variation in these statistics across the genome, we estimated the population recombination rate (ρ) using LDhelmet v. 1.10 [75], as described previously [76]. These analyses followed read-aware phasing, which was conducted for the SNP-only dataset using shapeit2 with default settings for each chromosome [77,78]. The estimated ρ for every SNP was averaged for each respective non-overlapping 50 kb window and used for analysis. The peak location of the density distribution of F_ST was calculated using the density function in R.

To test for introgression between the sympatric species, we calculated f₄ statistics of the ABBA-BABA test using the SNP-only datasets [79]. We first chose two targeted sympatric populations (P2 and P3) and two allopatric populations (P1 and P4). Then, the numbers of ABBA and BABA sites were counted, respectively, and used to calculate f₄ according to the following formula: f₄ = (number of BABA sites−number of ABBA sites)/ number of total sites. The calculation was conducted using Admixtools v. 5.1 and the admixr package in R [79,80]. The obtained Z-score was used to calculate the corresponding p-values. Next, to determine the regions that show signatures of introgression between sympatric species, we applied a topology weighting method called TWISST [81]. First, we performed read-aware phasing for the SNP-only dataset using shapeit2, as described above. Next, we selected individuals from four populations: sympatric P. sinensis in Shiomi River, allopatric P. sinensis in Yuza, sympatric P. pungitius in Shiomi River and allopatric P. pungitius in Denmark. We then performed maximum-likelihood tree estimation for each non-overlapping window containing 100 SNPs using RAxML v. 8.2.4 [82] and the raxml_sliding_window.py script (https://github.com/simonhmartin/genomics_general). The ASC_GTRCAT substitution model was used to correct the ascertainment bias. The estimated phylogenetic trees were used for the TWISST analysis. The weightings of a tree showing the topology where the sympatric populations clustered together were calculated for each window. To investigate the types of genes that are enriched in these putatively introgressed regions, we conducted gene ontology analysis using the g:Profiler web server [83].

3. Results

(a). Components of reproductive isolation

Among the six components of reproductive isolation investigated previously, four isolating mechanisms substantially contributed to the reproductive isolation between sympatric P. sinensis and P. pungitius (figure 1b). Geographical isolation was the strongest, and its relative contribution was also the highest among the isolating barriers examined (72.4%). Other substantial barriers were temporal isolation (4.62%), sexual isolation (14.7%) and hybrid male sterility (8.31%). The cumulative effect of total reproductive isolation was 90.5%.

(b). Whole-genome assembly and past demography

In our whole-genome assembly of P. sinensis, 62 scaffolds covering 461 Mb could be anchored to 21 stickleback chromosomes, although the contigs belonging to chromosome 4 were split into two. The genome size of a closely related species, the three-spined stickleback, is approximately 463 Mb [84], indicating that this assembly is likely to cover the majority of the genome. Obtained BUSCO scores were as follows: 91.1% for single copy genes, 2.6% for duplicated genes and 2.2% for fragmented genes, with 4.1% being missing. This reference assembly allowed us to conduct subsequent population genomic analysis.

To investigate the demographic histories of P. pungitius and P. sinensis, we first conducted a PSMC analysis, which showed a clear difference in the history of changes in N_e among the populations. Both the P. sinensis and P. pungitius populations at the Shiomi River site consistently declined until approximately 300 000 years ago (figure 2a). Since then, the N_e of P. sinensis has been relatively stable in the Shiomi River (mean N_e ± s.d. = 200 892 ± 32,809 since 300 000 years ago). By contrast, P. pungitius in Shiomi River showed a fluctuation of N_e: N_e gradually increased (N_e ± s.d. = 488 131 ± 36 885 at the peak) until approximately 80 000 years ago and then decreased again (N_e ± s.d. = 156 424 ± 78 237 at the lowest). N_e continued to decrease in both the Yuza and Musashi populations of P. sinensis.

Figure 2.

(a) Past effective population sizes estimated by PSMC. (b) The best demographic model inferred by fastsimcoal2: N_ANC, the effective population size (N_e) of the ancestral species; N_SINE, N_e of P. sinensis; N_ANC_PUNG, N_e of P. pungitius at the onset of divergence; N_PUNG, contemporary N_e of P. pungitius; T_ROOT, time of divergence; T_MIG, time of secondary contact; m_S->P, migration rate from P. sinensis to P. pungitius; m_P->S, migration rate from P. pungitius to P. sinensis. (c) Boxplot of log likelihood of different demographic models. Model 5 is shown in figure 2b. Model 1, no gene flow; model 2, unidirectional gene flow from P. sinensis to P. pungitius; model 3, unidirectional gene flow from P. pungitius to P. sinensis; model 4, continuous bidirectional gene flow; model 5, bidirectional gene flow after secondary contact; model 6, ancestral bidirectional gene flow (also see electronic supplementary material, figure S1). (d–f) Estimated parameter values with 95% confidence intervals. (Online version in colour.)

Next, to investigate the geographical mode of speciation, we conducted AFS analysis. The AFS analysis showed that the model of bidirectional gene flow after secondary contact was the best among the six models tested (figure 2b): the log likelihood distribution of this model did not overlap with the other models, suggesting that this model is strongly supported (figure 2c). In this best model, the species split at 1.73 Ma (95%CI = 1.11 × 10⁶–1.73 × 10⁶) and came into secondary contact at 37 200 years ago (95%CI = 3.07 × 10⁴–4.32 × 10⁴) (figure 2d). The migration rate from P. pungitius to P. sinensis (fraction of migrants within P. sinensis per generation m_P−>S = 5.03 × 10⁻⁷, 95%CI = 5.00 × 10⁻⁷–6.42 × 10⁻⁷; number of migrants from P. pungitius to P. sinensis per generation = 0.134, 95%CI = 0.133–0.171) was higher than that from P. sinensis to P. pungitius (fraction of migrants within P. pungitius per generation, m_S−>P = 1.88 × 10⁻⁷, 95%CI = 1.62 × 10⁻⁷–2.25 × 10⁻⁷; number of migrants from P. sinensis to P. pungitius per generation = 0.0304, 95%CI = 0.0270–0.0365) (figure 2d).

(c). Genomic landscape of differentiation and divergence

ADMIXTURE analysis showed that the cross validation error was the lowest at K = 4. At K = 4, all four populations were distinguished with no apparent signals of admixture between sympatric P. sinensis and P. pungitius species (electronic supplementary material, figure S2). At K = 2, an allopatric P. sinensis population from Musashi revealed a hybridization signature. This was likely due to the ancestral polymorphism shared with P. pungitius, since a previous phylogenetic analysis showed that this P. sinensis population diverged from P. pungitius earlier than other P. sinensis populations [47]. Thus, ADMIXTURE did not detect any clear signatures of introgression in sympatry.

Both sympatric and allopatric pairs of P. sinensis and P. pungitius were highly differentiated in their genome (figure 3 and table 1; electronic supplementary material, figure S3). However, genome-wide F_ST was significantly lower in the sympatric pairs than in the allopatric pairs, suggesting gene flow in sympatry (table 1; permutation test for comparing between allopatric and sympatric pairs, p < 2.2 × 10⁻¹⁶ for both comparisons). Interestingly, the density distribution of F_ST showed that both the sympatric and allopatric pairs had two peaks; however, these two peaks were clearer in the sympatric pairs (figure 3). This is due to the fact that the lower peak is shifted downward in the sympatric pair: the position of the higher peak was 0.916, 0.915 and 0.919 for F_ST of the sympatric pair, F_ST against Musashi and F_ST against Yuza, respectively; the lower peak position was 0.867, 0.896 and 0.878 for F_ST of the sympatric pair, F_ST against Musashi and F_ST against Yuza, respectively. Both d_XY and d_A also showed lower values in the sympatric pairs than in the allopatric pairs (table 1; permutation test, p < 2.2 × 10⁻¹⁶ for both comparisons), again consistent with gene flow. The genetic divergence (d_XY) in the sympatric pair was positively correlated with genetic diversity within each population (π) (Spearman's correlation; P. sinensis, r = 0.282, p < 2.2 × 10⁻¹⁶; P. pungitius, r = 0.596, p < 2.2 × 10⁻¹⁶).

Figure 3.

Sliding window analysis of genetic differentiation (F_ST) and divergence (d_XY) between sympatric pairs and allopatric pairs. Weighting of the tree topology consistent with introgression (i.e. the percentage of trees showing close relationship between the sympatric P. pungitius and P. sinensis) and the estimated recombination rate (ρ) are shown below. Sex chromosomes (chromosome 12) were not analysed. (Online version in colour.)

Table 1.

Genome-wide differentiation and divergence between populations.

comparison	species	geography	FST	dXY	dA
P. sinensis (Shiomi) versus P. pungitius (Shiomi)	different	sympatric	0.873 ± 0.061	0.0251 ± 0.0079	0.0207 ± 0.0071
P. sinensis (Musashi) versus P. pungitius (Shiomi)	different	allopatric	0.893 ± 0.052	0.0281 ± 0.0080	0.0247 ± 0.0072
P. sinensis (Yuza) versus P. pungitius (Shiomi)	different	allopatric	0.880 ± 0.057	0.0267 ± 0.0080	0.0220 ± 0.0072
P. sinensis (Shiomi) versus P. sinensis (Musashi)	same	allopatric	0.807 ± 0.067	0.0082 ± 0.0020	0.0064 ± 0.0014
P. sinensis (Shiomi) versus P. sinensis (Yuza)	same	allopatric	0.629 ± 0.096	0.0062 ± 0.0020	0.0030 ± 0.0011
P. sinensis (Musashi) versus P. sinensis (Yuza)	same	allopatric	0.883 ± 0.058	0.0098 ± 0.0023	0.0076 ± 0.0017

The heterogeneity in F_ST and d_XY was significantly associated with variation in recombination rate estimated as ρ_. F_ST was negatively correlated with ρ in both the sympatric and allopatric species pairs (sympatric: Spearman's correlation r = −0.214, p < 2.2 × 10⁻¹⁶; Yuza versus P. pungitius: r = −0.222, p < 2.2 × 10⁻¹⁶; Musashi versus P. pungitius: r = −0.216, p < 2.2 × 10⁻¹⁶). By contrast, d_XY was positively correlated with ρ in both the sympatric and allopatric pairs (sympatric: Spearman's correlation r = 0.0750, p = 3.90 × 10⁻¹¹; Yuza versus P. pungitius: r = 0.0838, p = 1.49 × 10⁻¹³; Musashi versus P. pungitius: r = 0.0722, p = 1.96 × 10⁻¹⁰). Genetic diversity within each population (π) was also positively correlated with ρ (P. pungitius: Spearman's correlation r = 0.172, p < 2.2 × 10⁻¹⁶; P. sinensis in Shiomi: r = 0.221, p < 2.2 × 10⁻¹⁶; Yuza; r = 0.190, p < 2.2 × 10⁻¹⁶; Musashi: r = 0.133, p < 2.2 × 10⁻¹⁶).

To confirm the presence of gene flow, we calculated the f₄ statistic. Significantly more SNPs were shared by sympatric species than allopatric species (P1 = Musashi, P2 = P. pungitius in Shiomi, P3 = P. sinensis in Shiomi, P4 = Yuza, f₄ = −6.41 × 10⁻⁴, p = 5.91 × 10⁻¹⁴; P1 = Denmark, P2 = P. pungitius in Shiomi, P3 = P. sinensis in Shiomi, P4 = Yuza, f₄ = −1.01 × 10⁻⁴, p = 4.18 × 10⁻³⁰). To identify the loci with signatures of introgression, we next conducted a TWISST analysis. TWISST showed that the majority of the genome showed the species tree topology with the second most common topology being the introgression topology (electronic supplementary material, figure S4). Regions with a topology consistent with introgression were interspersed and broadly distributed across the genome (figure 4). These regions were more often found in the genomic regions with high recombination rates: weighting of the introgression topology was positively correlated with ρ (Spearman's correlation, r = 0.0524, p < 2.2 × 10⁻¹⁶). Gene ontology analysis indicated that genes related to ion channels were enriched at the sites of putative introgression (electronic supplementary material, table S5).

Figure 4.

Comparisons of migration rate and F_ST among stickleback species pairs. Different colours indicate the different systems of sympatric species pairs. For references, see electronic supplementary material, tables S6 and S7. (Online version in colour.)

4. Discussion

In the present study, we have identified a stickleback species pair in the late stage of speciation. Although previous genetic studies have identified F₁ and backcross progeny between these two species [48,53], the long-term migration rates estimated here were very low (m = 1.88 − 5.03 × 10⁻⁷), suggesting that the total reproductive isolation is strong. This migration rate is the smallest compared to other stickleback systems studied so far (figure 4a). Furthermore, they exhibit hybrid male sterility [49,50], which is more likely to evolve between more genetically divergent species [12,85]. Consistent with this prediction, genome-wide differentiation and divergence between these two species is very high compared to other stickleback species pairs examined previously (figure 4b).

In our calculation of reproductive isolation (figure 1), total isolation was 90.5%, suggesting that 10% hybridization may occur. This is inconsistent with the estimated migration rates. Therefore, there are likely to be additional isolating mechanisms, such as selection against immigrants [86] and ecological selection against hybrids [6]. Previous studies have shown that P. pungitius mainly inhabits brackish-water habitats, while P. sinensis inhabits freshwater habitats [53]. Furthermore, P. pungitius has been to shown to be slightly but significantly more tolerant of seawater challenge experiments than P. sinensis [48], suggesting the possibility of physiological selection against immigrants. Although the relative contribution of hybrid sterility to total isolation was not very high compared with other isolating barriers, this is simply because we calculated the contributions of each barrier based on when it occurs during the life cycle and hybrid sterility acts at a later stage. However, hybrid sterility would be important for completing speciation because it is an irreversible barrier. The order of the appearance of isolating barriers during evolution is important for understanding how speciation proceeds, but we lack information about when each isolating barrier has evolved.

The Japanese nine-spined stickleback system can provide us opportunities to investigate the patterns of genomic divergence at late stages of speciation. We have identified heterogeneous patterns of genomic differentiation and divergence, which are highly correlated with recombination rates. Higher F_ST in genomic regions of lower recombination rates is widely observed in many cases of speciation [39,40,87,88]. This can be explained by the fact that regions with low recombination rates reduce intra-population genetic diversity due to the hitchhiking effects of negative selection against deleterious mutations and/or positive selection for adaptive alleles [87–89]. Consistent with this idea, intra-population genetic diversity showed a positive correlation with the recombination rate (electronic supplementary material, figure S3). By contrast, low recombination regions had a lower d_XY. Although this may appear to be counterintuitive, it can be explained if the ancestral populations had a lower genetic diversity in regions of low recombination [87–89]. Moderate positive correlations of d_XY with current genetic diversity π support the idea that d_XY reflects not only divergence between species but also genetic diversity in the ancestor. We also detected signals of gene flow in sympatry. Introgression occurred more often in regions with higher recombination rates, consistent with the theoretical predictions [90] and other empirical data [91]. This is because introgressed neutral loci can recombine away from weakly deleterious alleles that are purged by selection in high recombination regions. Interestingly, the genome-wide density distribution of F_ST had two peaks with the trend for the lower peak being moved downward in sympatry, suggesting the possibility that introgression may contribute to this heterogeneous pattern of F_ST. Because ion channels play important roles in many physiological functions [92–94], introgression of these genes may have some fitness effects. In this study, we used population genetic-based recombination rate. Although ρ is known to be highly correlated with recombination rates measured by linkage analysis [76], the recombination rates will need to be measured using linkage analysis to confirm our results.

Our data are consistent with the hypothesis that allopatry can promote speciation. Our demographic analysis clearly indicated that these sticklebacks experienced an allopatric phase. Secondary contact then occurred during the last glacial period, when P. pungitius may have expanded its distribution to the south where it came into contact with P. sinensis. This demographic history contrasts with that of the Japanese three-spined stickleback species pair, where a model of consistent gene flow was supported [28], although we cannot rule out the possibility of repeated instances of contact and isolation. This difference may account for the higher genomic divergence between the nine-spined stickleback than the three-spined stickleback species pairs. Currently, we do not know whether any prezygotic barriers are enhanced in sympatry compared to in allopatry. Further ecological and behavioural comparisons between sympatric and allopatric pairs are required to test the presence of reinforcement after secondary contact.

Whether speciation has a tipping point remains elusive [5]. Our plot of F_ST and migration rate in multiple stickleback species pairs may be consistent with the hypothesis of the tipping point (figure 4). However, there are too few data points to draw solid conclusions. We need more data on gene flow rates and genomic divergence of species pairs differing in the strength of reproductive isolation and demographic history. Although several theoretical studies have indicated that an allopatric phase promotes the establishment of postzygotic isolation [95], few empirical studies have examined the necessity as well as the duration of the allopatric phase for the establishment of intrinsic incompatibility. Whole-genome sequencing together with sophisticated demographic analysis will enable us to infer whether and, if so, for how long, pairs of species experienced allopatry in the past. Investigating the association between the duration of allopatry and the strength of each isolating barrier, including intrinsic incompatibility, in multiple species pairs will reveal how strongly allopatry in the past influences the evolution of reproductive isolation. In summary, further integration of past demographic history, current gene flow rate, and components of reproductive isolation across many systems will enable us to understand how past divergence history, such as allopatric phases, influence the evolutionary trajectories of speciation.

Ethics

Fish were euthanized before fixation according to the protocols approved by the Institutional Animal Care and Use Committee (31-16, 30-12, 29-11). Specimens were collected under the permits of Hokunaisuimen No. 6.

Data accessibility

The assembled reference genome sequence and files used for the analysis will be deposited in Dryad (https://doi.org/10.5061/dryad.pk0p2ngkd). Short read sequences are available from DDBJ (DRA009343–DRA009347).

Competing interests

We declare we have no competing interests.

Funding

This research was supported by JSPS KAKENHI grant nos 19H01003 and 17KT0028 to J.K., 22687006 to H.T., and 16H06279 to A.T. Y.Y.Y. was supported by NIG Postdoctoral Fellow Program.