Introduction

Reproductive isolation and hybridization are antinomic but related processes that are tuned by the same evolutionary forces. Natural hybridization between two divergent lineages occurs as long as reproductive isolation is not totally established (Dobzhansky 1937; Mayr 1942; Coyne and Orr 2004). The resulting exchange of genetic material beyond the first hybrid generation occurs through the production of backcross and recombinant genotypes that are exposed to selection, allowing neutral and advantageous alleles to spread among divergent lineages but preventing gene flow at the loci involved in reproductive isolation (Barton 1979). Hybrid zones, the geographic areas where hybridization occurs, therefore act as selective filters which allow us to unravel the strength of reproductive barriers and their impact on the patterns and dynamics of gene exchanges at various stage of speciation (Barton and Hewitt 1985; Barton and Bengtsson 1986; Hewitt 1988; Martinsen et al. 2001).

The semipermeable nature of species boundaries was first evidenced from differential introgression patterns in hybrid zones (Harrison 1990), that is, variation among loci in the level of incorporation of alleles from one lineage into the other (Payseur 2010; Harrison and Larson 2016). Introgression at a given neutral marker depends on the antagonistic effects between counter-selection on a nearby selected locus and recombination with that barrier locus (Barton and Bengtsson 1986). Therefore, differential gene flow across the genome reflects both recombinational distance to a barrier locus and selection acting on that locus. Different methods have been developed to detect individual loci with introgression behaviors departing from the genome-wide average. This includes the analysis of spatial allele frequency patterns in a transect spanning the hybrid zone, or variation in ancestry proportion at individual loci compared to genome-wide expectations in admixed genotypes (see Payseur 2010 for a review).

The geographic cline method is a powerful approach to analyze the relationship between allele frequency and geographic distance to the center of a hybrid zone (Endler 1977; Barton and Hewitt 1985; Barton and Gale 1993; Porter et al. 1997; Teeter et al. 2008). Fitting a cline model to observed allele frequency patterns allows inferring fundamental parameters such as cline center, dispersal, and the strength of selection. Therefore, important information concerning the balance between migration and selection can be obtained using the geographic cline method, such as the identification of which loci tend to introgress neutrally and which do not. The genomic cline method is a related approach developed to deal with hybrid zones that are not clinally structured, such as mosaic hybrid zones (Harrison and Rand 1989; Bierne et al. 2003; Larson et al. 2013). The change in allele frequency at individual loci is analyzed along a gradient of genomic admixture instead of a spatial gradient, enabling a comparison of individual locus introgression patterns with regard to the genome-wide average pattern of admixture (Gompert and Buerkle 2009, 2011; Fitzpatrick 2013). As for geographic clines, this genomic cline method also provides a quantitative assessment of the excess of ancestry to a given parental species, as well as the rate of change in allele frequency from one species to the other along the admixture gradient. Although both approaches are useful to document the effect of selection and recombination on differential introgression in admixed genotypes, they do not specifically account for the demographic history of the parental species (Payseur and Rieseberg 2016).

An alternative approach that partly overcomes this limitation is the use of demo-genetic models that account for varying rates of introgression among loci during the divergence history. Such methods for reconstructing the history of gene flow between semi-isolated lineages have been developed within Bayesian (Sousa et al. 2013), approximate Bayesian computation (Roux et al. 2013, 2016), or approximate likelihood frameworks (Tine et al. 2014; Le Moan et al. 2016; Rougeux et al. 2017). Using summary statistics like the joint allelic frequency spectrum, which depicts correlations in allele frequencies between lineages outside the hybrid zone, they capture variable behaviors among loci and allow quantifying the degree of semi-permeability reflecting the overall balance between gene flow and selection. These methods have the advantage of specifically accounting for the history of gene flow during divergence, using contrasted speciation scenarios such as primary differentiation, ancient migration (AM), or secondary contact (SC). However, they do not provide information about individual locus behavior as the cline methods do. Here, we pushed these inference methods one step further in order to assess the probability for a given locus to belong to one of two categories: (i) loci with a reduced effective migration rate due to selection and linkage, and (ii) loci which can readily introgress.

The flatfish species Solea senegalensis and Solea aegyptiaca are two economically important species in the Mediterranean basin that hybridize along the northern Tunisian coasts, where they form a hybrid zone (She et al. 1987). Mitochondrial divergence between them is high (~2%) relative to other vertebrate species pairs that also hybridize in nature, such as the house mice Mus m. musculus and M. m. domesticus (0.3%, based on sequences retrieved from Genebank) or Atlantic and Mediterranean lineages of Dicentrarchus labrax (0.7%, Tine et al. 2014). Based on an analysis of spatial allele frequency patterns at a dozen of allozymic and intronic loci, Ouanes et al. (2011) proposed that the hybrid zone between S. senegalensis and S. aegyptiaca was centered in Bizerte lagoon, acting as a non-stable unimodal tension zone stemming from a SC. They also suggested that the zone could have undergone a recent expansion. Recently, Souissi et al. (2017) showed the existence of morphological transgressions within the contact zone, possibly indicating a reduced fitness of recombined compared to parental genotypes. However, the low number of markers in these latter studies could not provide a clear description of the genetic exchanges across the genome of these two species, nor quantifying the strength of the forces maintaining the incomplete reproductive isolation between them.

In the present work, high-throughput genotyping using restriction-associated DNA (RAD) sequencing (Baird et al. 2008) was carried out in individuals sampled at both sides of the hybrid zone, with a particular effort in the contact zone itself. By combining three different methods (genomic and geographic cline analysis and historical demographic inference) exploiting different aspects of the data, we provide new insights on the history of divergence between S. senegalensis and S. aegyptiaca, and a genome-wide description of varied patterns of introgression attributed to recent or ongoing movement of the hybrid zone.

Materials and methods

Sampling

A total of 161 samples were collected from 10 locations spanning the distribution range of Solea senegalensis and S. aegyptiaca from Senegal to Egypt (Fig. 1). This sampling strategy aimed at covering the geographical distribution of both species, including a detailed transect of their natural hybrid zone in Tunisia. Five sampling locations were collected throughout the S. senegalensis parental zone, two in the Atlantic Ocean (10 individuals from Dakar in Senegal and 16 from the Gulf of Cadiz in Spain), and three in the Western Mediterranean Sea (15 individuals from Annaba and 8 from Mellah Lagoon in Algeria, and 10 from Tabarka in Tunisia). Three locations were sampled across the geographical range of S. aegyptiaca in the Eastern Mediterranean Sea (13 individuals from Kerkennah Island and 15 from El Biban lagoon in Tunisia, and 20 samples from Bardawil lagoon in Egypt). Sampling size was locally increased in the Tunisian region where both species coexist and hybridization has been reported (Bizerte lagoon n = 29 and Gulf of Tunis n = 25). Finally, a total of seven individuals belonging to the closely related species S. solea were sampled in Tunisia to provide an outgroup species for the orientation of ancestral and derived alleles in S. senegalensis and S. aegyptiaca.

Fig. 1
figure 1

Map of the sampling locations used in the present study. Solea senegalensis side: Dakar (Dk), Cadiz (Cx), Annaba (An), Mellah lagoon (Ml), Tabarka (Tb). Hybrid zone: Bizerte lagoon (Bz). Solea aegyptiaca side: Gulf of Tunis (Gt), Kerkennah Islands (Kr), El Biban lagoon (Bb), Bardawil lagoon (Bw)

Full size image

RAD library preparation and sequencing

Whole-genomic DNA was extracted from fin clips using the DNeasy Blood & Tissue kit (Qiagen). The presence of high molecular weight DNA was checked on a 1% agarose gel, and double-stranded DNA concentration was quantified using Qubit 2.0 and standardized to 25 ng per μl. RAD library construction followed a modified version of the original single-end RAD-Seq protocol (Baird et al. 2008). Briefly, 1 μg of genomic DNA from each individual was digested using the restriction enzyme SbfI-HF (NEB), and ligated to one of 32 unique molecular barcodes of 5–6 bp. Ligated products were then combined in equimolar proportions into six RAD libraries, each made of a multiplex of 32 individuals originating from various localities. Each library was finally sequenced in 101 bp single read mode on a separate lane of an Illumina HiSeq2500 sequencer, at the sequencing platform “Génomique Intégrative et Modélisation des Maladies Métaboliques” (UMR 8199, Lille, France).

Bioinformatic analyses

Raw reads were de-multiplexed based on individual barcode information and subsequently end-clipped to 95 bp to homogenize read length after removing barcodes of different lengths. Read quality filtering was performed using the sliding window approach implemented in the module process_radtags from the stacks pipeline (Catchen et al. 2013). This allowed us to exclude reads in which the average quality of 15 adjacent bases fell below a raw phred score of 10. Retained read were then aligned to a draft assembly of the S. senegalensis genome (98,590 scaffolds, total length 740 Mb, N50 contig length 10,767 bp, Manchado et al. 2016) using bowtie 2v.2.1.0 (Langmead and Salzberg 2012) with the --very-sensitive option in --end-to-end mode. In order to take into account both the level of divergence among species and the possibility of introgression while accounting for hidden paralogy, we used a subset of individuals to empirically determine the optimal maximum number of mismatches allowed between aligned reads and the reference genome, following a procedure described previously (Le Moan et al. 2016; Rougemont et al. 2017; Rougeux et al. 2017). We found that a maximum of seven mismatches for a 95 bp RAD tag offered the best compromise to correctly align both species and the outgroup to the S. senegalensis genome. This was set to bowtie 2 options using a minimum alignment score of --score-min = −42 needed for an alignment to be considered as valid, fixing the penalty of mismatch bases to MX = 6 and using default gap penalties. We then used pstacks to call variable positions under the bounded SNP model, setting the upper sequencing error rate to 2.5% and a minimum sequencing depth to 5× per stack. Homologous loci across samples were merged based on their genomic position within scaffolds using cstacks to construct a catalog of loci. Individual stacks were then matched against the catalog of loci with sstacks to determine genotypes. The module rxstacks was ran with the --prune_haplo option to exclude poor-quality loci with multiple sequencing errors using a log-likelihood threshold of −300 (determined empirically, Fig. S1). Individual genotypes were finally exported in the VCF format using the module populations.

We then applied population-specific filters with VCFtools (Danecek et al. 2011) to remove SNPs showing significant deviation to Hardy–Weinberg equilibrium within at least one of the S. senegalensis or S. aegyptiaca locality samples located outside the hybrid zone, using a P-value threshold of 0.01. Next, we excluded loci displaying more than 20% of missing genotypes in at least one locality. Over the 116,385 remaining SNPs, we randomly selected one single SNP for each pair of RAD loci associated with the same restriction site (i.e., within a distance less than 200 bp), in order to limit the impact of linkage disequilibrium. Finally, we only retained loci with available sequence data in the outgroup species S. solea, which had to be fixed for the polymorphic site found in S. senegalensis and S. aegyptiaca. This resulted in a final dataset containing 10,758 independent SNPs that were specifically filtered to comply with the δaδi analysis requirements. The same SNPs were used in all subsequent analyses unless specified otherwise.

Genetic structure and hybridization

Genetic variation within and among S. senegalensis and S. aegyptiaca was characterized with a principal component analysis (PCA) performed with the R package Adegenet (Jombart 2008; Jombart and Ahmed 2011). In order to document the possible existence of population structure within each species and distinguish parental and admixed genotypes, we estimated individual admixture proportions from two to four differentiated genetic clusters that were inferred from the data with no a priori information on individual membership, using fastStructure (Raj et al. 2014) with 108 iterations. In addition, we estimated the proportion of alleles with S. aegyptiaca ancestry for each individual, using the R package Introgress (Gompert and Buerkle 2010) to estimate a Hybrid Index. We then searched specifically the presence of early-generation hybrids using NewHybrids (Anderson and Thompson 2002) with a subset of 296 diagnostic SNPs showing fixed differences between species. For each individual, we estimated the membership probability to each of six categories of pedigree including pure parental species (Psen and Paeg), first and second-generation hybrids (F1 and F2), and first-generation backcrosses in each direction (BCsen and BCaeg), using a total of 10,000 burn-in steps and 50,000 iterations.

Demo-genetic inference of the divergence history

We used a modified version of the δaδi program (Gutenkunst et al. 2009) that improves likelihood optimization using a simulated annealing (SA) procedure (Tine et al. 2014) in order to infer the divergence history between S. senegalensis and S. aegyptiaca. This method uses the joint allele frequency spectrum (JAFS) between two populations as summary statistics to characterize divergence. The program δaδi fits demographic divergence models to the observed data using a diffusion approximation of the JAFS, enabling a comparison of different alternative divergence models in a composite likelihood framework. We used seven divergence models developed in a previous study (Tine et al. 2014) to determine whether and how gene flow has shaped genome divergence between S. senegalensis and S. aegyptiaca. The simplest model of strict isolation (SI) corresponds to an allopatric divergence scenario in which the ancestral population of effective size NA splits into two derived populations of size N1 (S. senegalensis) and N2 (S. aegyptiaca) that evolve without exchanging genes for Ts generations. We then considered three models of divergence including gene flow, either during the entire divergence period (isolation with migration, IM), the beginning of divergence (AM), or the most recent part of divergence (SC). In these models, migration occurs with potentially asymmetric rates (m12 and m21) that are shared across all loci in the genome. We also considered simple extensions of the IM, AM, and SC models that capture the effect of selection by accounting for heterogeneous migration rates among loci (IM2m, AM2m, and SC2m). These semi-permeability models consider that two categories of loci, experiencing different effective migration rates, occur in proportions P and 1−P in the genome.

The JAFS was obtained by pooling the least introgressed populations from S. senegalensis (Dakar, Cadiz, Annaba, and Mellah) for species 1 and S. aegyptiaca (Kerkennah, El Biban, and Bardawil) for species 2 (Fig. 2), in order to avoid including admixed genotypes causing departures to the underlying population model. We also tested the robustness of our inferences with respect to the existence of a subtle genetic structure within each species (Fig. 2a, b) by considering only Annaba and Mellah for S. senegalensis (species 1) and Kerkennah and El Biban for S. aegyptiaca (species 2).

Fig. 2
figure 2

Genetic structure of S. senegalensis and S. aegyptiaca analyzed using 10,758 SNPs, showing introgressive hybridization between species. Principal component analysis (PCA) based on 161 individuals representing a individual coordinates along PC1 and PC2 axes and b along PC1 and PC3 axes. Legend: Dk; Cx; An; Ml; Tb; X Bz; Gt; Kr; Bb; Bw. c Triangle plot showing individual interspecific heterozygosity against hybrid index. d Result of fastStructure analysis performed for K = 2 genetic clusters

Full size image

We used S. solea samples as an outgroup species to determine the most parsimonious ancestral allelic state for each SNP in order to generate an unfolded JAFS oriented with the derived allele (Fig. 3a). The size of the JAFS was projected down to 40 sampled chromosomes per species to account for missing data. For each model, we estimated the parameter values that maximize likelihood using two successive SA procedures before quasi-Newton (BFGS) optimization (Tine et al. 2014). Comparisons among models was made using the Akaike information criterion (AIC) to account for variation in the number of parameters among models. A total of 20 independent runs were used for the optimization of each model. Parameter uncertainties were estimated from non-parametric bootstrapped data using the Godambe information matrix as implemented in δaδi. We used 1000 bootstrapped datasets to estimate confidence intervals as the maximum likelihood parameter value ±1.96 × SE. Estimated parameter values were converted into biologically meaningful quantities by taking into account our SNP filtering procedures following the method in Rougeux et al. (2017).

Fig. 3
figure 3

Detection of introgression based on the inferred demographic divergence history of S. senegalensis and S. aegyptiaca. a The observed joint allele frequency spectrum (JAFS) for S. aegyptiaca (AEG, y-axis) and S. senegalensis (SEN, x-axis) showing the number of SNPs (colored scale on the right side of the spectrum) per bin of derived allele counts using 20 individuals per species. b Schematic representation of the secondary contact model with heterogeneous introgression rates among loci (SC2m model). c The maximum-likelihood JAFS obtained under the best-fit SC2m model, which can be decomposed into a reduced introgression model (Φ) explaining 95% of the genome and a free introgression model (Y) accounting for 5% of the genome. The JAFS of the two models Φ and Y obtained from msms simulations, illustrating which bins are likely to be attained preferentially under the free vs. reduced introgression model

Full size image

Inferring the probability of locus introgression under the best divergence model

Because the JAFS-based inference of the divergence history does not provide an assessment of introgression probability for each locus separately, we developed an approach to estimate the relative probability of each individual locus to be assigned to one of these two categories: (i) loci which can readily introgress between species, and (ii) loci experiencing a highly reduced introgression rate due to selection against foreign alleles and linkage. This probability was estimated using the best-fit model identified in the previous section, which was a SC model with variable introgression rates among loci (SC2m) (Fig. 3b). The SC2m model can be decomposed as a linear combination of two simple models describing gene flow in two different compartments of the genome (Fig. 3c). The best-fit SC2m model estimated that only 5% of the loci can still introgress between species with effective migration rates m1−2 and m2−1, whereas the remaining 95% of loci experience a highly reduced introgression rate with effective migration rates m′1−2 and m′2−1. We used estimated model parameters to perform coalescent simulations with msms (Ewing and Hermisson 2010) under the SC model (assuming theta = 747.202; N1 = 0.818; N2 = 1.136; TS = 4.803; TSC = 0.081) using two different conditions for gene flow (assuming either m1−2 = 4.607 and m2−1 = 0.381, or m′1−2 = 0.057 and m′2−1 = 0.153) to generate a free introgression and a reduced introgression rate dataset. One thousand JAFS with 40 sampled chromosomes per species were produced under each model. We then averaged the number of derived alleles within entries across replicates to obtain a single JAFS for both the free introgression (Y) and the reduced introgression (Φ) model (Fig. 3c). Finally, we used the average number of SNPs per entry (i,j) in each JAFS to estimate the probability Pi,j that a SNP with a derived allele count i in species 1 and j in species 2 can be obtained under the free introgression model:

$$P_{i,j}{\mathrm{ = }}0.05{\mathrm{ \times \Upsilon }}_{i,j}/\left( {0.05 \times {\mathrm{\Upsilon }}_{i,j} + 0.95 \times {\mathrm{\Phi }}_{i,j}} \right),$$
(1)

where Yi,j and Φi,j are the average number of SNPs predicted in the JAFS entry (i,j) under the free introgression and reduced introgression model, respectively. Every SNP from the real dataset was finally associated to the introgression probability Pi,j given by Eq. (1) based on its corresponding entry in the JAFS.

Genomic clines

The Bayesian Genomic Cline program BGC (Gompert and Buerkle 2012) was used to quantify individual locus introgression relative to genome-wide introgression. BGC describes the probability of locus-specific ancestry from one parental species given the genome-wide hybrid index. The BGC model considers two principal parameters, called α and β, which describe locus-specific introgression based on ancestry. Parameter α quantifies the change in probability of ancestry relative to a null expectation based on genome-wide hybrid index, that is, the direction of introgression. A positive value of α reflects an increase in the probability of ancestry from species 1 (introgression into S. aegyptiaca), whereas a negative value of α reflects an increase in the probability of ancestry from species 2 (introgression into S. senegalensis). Parameter β describes the rate of transition in the probability of ancestry from parental population 1 to parental population 2 as a function of hybrid index, that is, the amount of introgression. Positive β values thus denote a restricted amount of introgression, whereas negative β values indicate a greater introgression rate compared to the genome-wide average.

We ran BGC under the genotype-uncertainty model, assuming a sequence error probability of 0.0001 and using information from the data to initialize ancestry and hybrid index. Two MCMC chains each made of 150,000 steps were ran, recording every 20th value. Cline parameter quantiles were calculated to designate outlier loci with respect to parameters α and β based on the assumption that the genome-wide distributions of locus-specific cline parameters are both centered on zero. Therefore, outlier loci are markers with extreme patterns of introgression relative to the remainder of the genome. A locus was considered as an α outlier if its posterior estimate of α was not contained in the interval bounded by the 0.025 and 0.975 quantiles of \({\cal N}(0,\tau _\alpha )\) (likewise for β) (Gompert and Buerkle 2011; Gompert et al. 2013).

Geographic clines

The geographic cline analysis was carried out in order to link allele frequencies at individual loci with geographic position along a transect spanning the hybrid zone. We used the R package HZAR (Derryberry et al. 2014) that fits allele frequency data to classic equilibrium models of geographic clines (Szymura and Barton 1986) using the MCMC algorithm. Instead of searching for the best model separately for each cline, we used the full model that fits cline center, width, and independent introgression tails using estimated values for minimum (pmin) and maximum (pmax) allele frequencies. In this way, we avoided the potentially confounding effects of fitting different models for the comparison of cline center and slope parameters across loci.

Results

A total of 833.3 million raw reads were obtained, 89.1% of which were retained after demultiplexing and quality filtering for reference mapping against the S. senegalensis draft genome (average number of reads per individual: 3,906,809, s.d.=1,878,685). Individual genotype calling in Stacks produced a raw VCF file containing 174,490 SNPs, from which we retained 10,758 oriented SNPs (using S. solea as an outgroup to identify ancestral states) after controlling for linkage and filtering for quality. The mean sequencing depth per locus per individual was superior to 100× (Fig. S2), and the mean genotype missing rate was 1.3% (s.d. = 2.6%) per individual and 1.4% (s.d. = 1.8%) per locus. The final VCF containing 10,758 SNPs was used for all downstream analyses except NewHybrids.

Genetic structure, hybridization, and introgression

The PCA clearly separated S. senegalensis from S. aegyptiaca samples along the first PC axis (PC1), which explained 74.2% of the total genotypic variance (Fig. 2a). S. senegalensis samples were organized along that axis according to their geographical proximity from the contact zone, following a gradient of genetic similarity to S. aegyptiaca increasing from Dakar to Bizerte lagoon. The two species were found to coexist only in Bizerte lagoon, with four S. aegyptiaca individuals being found among a majority of S. senegalensis genotypes. The second principal component revealed a weak signal of differentiation (PC2, 0.73% of associated variance) separating Bardawil lagoon from other samples within S. aegyptiaca. Likewise, the third principal component (PC3, 0.66% of genotypic variance) captured a subtle differentiation signal between the Atlantic samples from Dakar and Cadiz and the Mediterranean S. senegalensis samples (Fig. 2b). The fastStructure analysis (Fig. 2c) confirmed the separation of the two species into different genetic clusters and their coexistence in Bizerte lagoon, as detected from the PCA. However, it failed to detect any further genetic subdivision (Fig. S3), confirming that the signal of genetic structure within each species is at most very small.

The finding of intermediate admixture proportions in some individuals revealed signs of introgressive hybridization between S. senegalensis and S. aegyptiaca around the contact zone.

The relationship between individual hybrid index and interspecific heterozygosity represented in a triangle plot illustrated well the absence of F1 and F2 hybrids in our dataset, and the presence of a few backcrosses together with several introgressed genotypes (i.e., late-generation backcrosses) (Fig. 2d). This was confirmed by the detection of three first generation backcrosses by NewHybrids (detected with an assignment probability of 1), one in the direction of S. senegalensis (in Bizerte lagoon) and two in the opposite direction (in Gulf of Tunis), plus two unassigned individuals likely representing later-generation backcrosses (Fig. S4). Therefore, our results provide evidence for contemporary introgressive hybridization between the two sole species.

Demo-genetic history of divergence

The δaδi analysis showed that the SC model with varying introgression rates along the genome best explained the observed JAFS (SC2m, delta AIC with the second ranked model >25). In comparison, the six other models had a significantly lower performance for different reasons. While the SI model could explain SNPs occupying the outer frame of the JAFS, it could not predict at the same time the presence of loci in the more inner part of the spectrum. The IM, AM, and SC models assuming genome-wide homogeneous migration rates better predicted loci occupying the central part of the JASF. However, they underestimated the density of private and highly differentiated SNPs. Finally, the three models including heterogeneous migration rates along the genome (IM2m, AM2m, and SC2m) yielded significantly improved fits. Among them, only the SC2m model provided a good prediction for both a high density of highly differentiated SNPs between the two species and the presence of SNPs toward the central part of the spectrum. Model selection was not sensible to the potential effect of unaccounted structure within each species (Table S1), consistent with the very weak signals of within-species differentiation detected in PCA analysis.

Using the best-fit obtained for the SC2m model over 20 independent runs to get estimates of model parameters and confidence intervals, we found that the duration of the isolation period between S. senegalensis and S. aegyptiaca was about 60 times longer than the duration of the SC. A large proportion of the genome (~95%) was associated with relatively small effective migration rates and limited gene flow corresponding to less than one effective migrant per generation in both directions (N1m′12 = 0.047, N2m′21 = 0.175). By contrast, the effective migration rate was more elevated in the remaining small fraction of the genome (~5%), especially in the direction from S. senegalensis to S. aegyptiaca where introgression was found 80 times higher than elsewhere in the genome. Introgression was found 12 times lower in direction of S. senegalensis, although still more than twice as high than in the remainder (95%) of the genome (Table 1).

Table 1 The demographic divergence history of S. senegalensis and S. aegyptiaca is best explained by a secondary contact model with heterogeneous introgression rates among loci
Full size table

Genomic clines

The genomic cline analysis performed with the BGC program revealed different behaviors among the 10,758 SNPs with respect to the cline parameters α and β. Considering the locus-wise ancestry shift parameter α, we observed that 48% of the loci have an excess of S. senegalensis ancestry (negative α), whereas only 29% of them have an excess of S. aegyptiaca ancestry (positive α). The quantile method for outlier detection confirmed that a higher proportion of loci was characterized by an excess of S. senegalensis ancestry (43% of negative α outliers) compared to S. aegyptiaca ancestry (17% of positive α outliers). By contrast, the locus-wise slope parameter β depicting the rate of transition between species was symmetrically distributed between negative and positive values, with equal proportions of loci showing a decreased (48% of positive β) and increased (51% of negative β) introgression rates. The proportion of loci showing significant deviation compared to the genomic average amounted to 52%, and were equally distributed among negative (26%) and positive (26%) β outliers.

Geographic clines

The mean position of cline center (C) calculated over all fitted clines was located 12 km to the east of Bizerte lagoon (Fig. 4). Most individual locus clines tended to co-localize at this position, especially for the steepest clines whose centers localized essentially 10 km or less around the center of the hybrid zone. In general, loci whose individual centers did not coincide with the central part of the hybrid zone harbored less steep clines, either because they correspond to rare variants with low information content (i.e., that do not form a cline) or due to softer selection at these loci which are uncoupled from the hybrid zone (Fig. 4).

Fig. 4
figure 4

Distribution of geographic cline slope (S) for the 10,758 loci as a function of their cline center (C). The box represents a zoom in the central part of the contact zone. Cline centers falling outside this central region mostly correspond to rare variants with low information content

Full size image

Links between different approaches

Linear regression models were used to investigate whether different analyses of individual locus introgression that use different aspects of the data tend to produce similar results. First, we evaluated the extent to which an excess of ancestry from a given species relates to a spatial shift of cline center into the other species range. The genomic cline parameter α showed a significant positive correlation with the geographic cline center C on both sides of the hybrid zone (Fig. 5). The correlation was, however, much stronger in the S. aegyptiaca side (R²c−α = 0.274, P < 10−10) compared to the S. senegalensis side where it was barely detectable (R²c−α = 0.023, P < 10−10).

Fig. 5
figure 5

Correlation between genomic cline parameter (α) and geographic position of the cline center (C) on both sides of the hybrid zone: a in the Solea senegalensis geographic range, with the Atlantic/Mediterranean transition zone indicated by the dotted arrow, and b in the Solea aegyptiaca geographic range. Negative α values indicate an excess of S. aegyptiaca ancestry, whereas positive α values indicate an excess of S. senegalensis ancestry

Full size image

We then tested whether the loci exhibiting the most abrupt transition in ancestry between the two species (i.e., a decreased introgression rate) also tended to display the steepest geographic clines. To validate this prediction, we focused on the 26% of positive β outliers detected with BGC, and compared β values with the slope parameter S of the geographic clines. We found a significantly positive correlation (R²S−β = 0.27, P < 10−10) between the two parameters, confirming that loci with a low introgression rate also tend to display steeper geographic clines (Fig. 6).

Fig. 6
figure 6

Correlation between the geographic cline slope parameter (S) and the genomic cline parameter β for positive β outlier loci, showing significantly restricted introgression rates compared to the genome-wide average

Full size image

Moreover, we tried to connect genomic and geographic cline parameters with our estimated probability that individual loci belong to the small fraction of the genome showing the highest introgression rate. As expected, geographic cline slope was significantly negatively correlated with the inferred probability of introgression (R²S−Proba = 0.04, P < 10−10, Fig. S5). Therefore, private SNPs located on the outer frame of the JAFS, for which we inferred a small probability of introgression, were usually associated with steep cline slopes. By contrast, loci occupying the most central part of the JAFS, which were assigned the highest introgression probabilities, were also characterized by shallower geographic clines. Finally, when restricting the analysis to the 5% of loci with the highest introgression probability, we show that a majority of their geographic cline centers are located outside the contact zone, with a spatial shift more pronounced in the S. aegyptiaca direction (Fig. 7a). Similarly, the distribution of the genomic cline parameter α also showed a deficit of values around zero, and a majority of loci with an excess of S. senegalensis ancestry, in keeping with the general asymmetry of the exchanges between the two species (Fig. 7b).

Fig. 7
figure 7

Distribution of geographic and genomic cline centers for the 5% of loci showing a probability of introgression greater or equal to 95%. a Geographic distribution of the geographic cline center (C) with the Atlantic/Mediterranean transition zone indicated by the dotted arrow. b Distribution of the genomic cline parameter α. Negative values for both cline center and α indicate an excess of S. aegyptiaca ancestry, whereas positive values indicate an excess of S. senegalensis ancestry

Full size image

Discussion

Divergence history and semi-permeability to gene flow

Our results using high-throughput genotyping confirmed previous observations, based on a limited number of nuclear markers, that the flatfish S. senegalensis and S. aegyptiaca are genetically divergent sibling species, which are still exchanging genes across their natural hybrid zone near Bizerte lagoon (She et al. 1987; Ouanes et al. 2011; Souissi et al. 2017). However, the proportion of hybrid genotypes detected here in the contact zone (i.e., 3 early generation hybrids among 54 individuals) is much lower than established in previous studies. Although this could be due to large temporal variations in the rate of hybridization, the most plausible hypothesis is that real hybrids (e.g., F1, F2, and first generations backcrosses) are more easily distinguished from introgressed genotypes using a high number of diagnostic markers (Pujolar et al. 2014; Jeffery et al. 2017), whereas earlier studies have used principally non-diagnostic markers. Our results thus support the existence of an abrupt geographic transition between the two species across their contact zone, where predominantly parental genotypes occur in sympatry with an overall low frequency of hybrids. These observations are consistent with the tension zone model (Barton and Hewitt 1985), in which the hybrid zone is maintained by a balance between the influx of parental genotypes from outside the zone and the counter-selection of hybrid genotypes inside the zone. This interpretation is also in keeping with the finding of transgressive body shape variation attributed to a reduced condition index of admixed genotypes from the contact zone (Souissi et al. 2017).

The regions located outside the contact zone, which contain only parental genotypes, provided relevant information to infer the demo-genetic history of divergence between these two species. Gene exchange between S. senegalensis and S. aegyptiaca is best explained by a SC model with heterogeneous rates of introgression along the genome. Similar divergence histories have already been found between geographical lineages or ecotypes of the same species, like in the European sea bass and anchovy (Tine et al. 2014; Le Moan et al. 2016). Using the same approach as implemented here, these studies could determine that a minor but significant fraction of the genome (i.e., 20–35%) does not neutrally introgress between hybridizing lineages/ecotypes, thus providing a quantitative assessment of the degree of semi-permeability to gene flow. In the present case, by contrast, we found that the great majority of the genome (approximately 95%) experiences a highly reduced effective migration rate (i.e., <1 migrant per generation) between species. Moreover, the inferred divergence period estimated using a mutation rate of 10−8 per bp per year and a generation time of 3–5 years corresponds to a separation time of ca. 1.1 to 1.8 Myrs, which together with the mitochondrial sequence divergence of 2%, indicates a relatively ancient speciation event. This timing of divergence is much older than the approximately 300 Kyrs that were inferred between glacial lineages in anchovy and sea bass (Tine et al. 2014; Le Moan et al. 2016). Therefore, our results are consistent with the prediction that more anciently diverged species should display stronger reproductive isolation due to the establishment of more genetic barriers (Orr 1995; Moyle and Nakazato 2010; Roux et al. 2016), which is reflected here by a lower level of permeability to gene flow compared to sea bass and anchovy.

Another aspect of this low miscibility of the two genomes was captured by individual locus cline parameters. Consistent with the view that stepped clines generate a strong barrier to gene flow (Szymura and Barton 1986), we found a high proportion of loci combining extremely low introgression rates (i.e., positive β outliers) and steep geographic clines (i.e., cline width below the average) with centers colocalized in the central part of the hybrid zone. Because this type of situation generates strong linkage disequilibrium, each locus is expected to receive, in addition to its own selective coefficient, the effect of selection on other loci (Barton 1983; Kruuk et al. 1999). This makes the genome acting almost as a single underdominant locus causing a global reduction in gene flow, which sometimes is referred to as a “congealed genome” (Turner 1967; Bierne et al. 2011; Gompert et al. 2012b) relatively to its very low capacity to incorporate foreign genes. In contrast to this genome-wide reduction in gene flow, we also found support for a small proportion of genomic regions with higher-than-average introgression rates between species.

Multiple approaches to detect increased introgression rates

Differential introgression among loci was evidenced by combining several approaches that capture different but complementary aspects of the data. For instance, genomic and geographic cline methods both rely on the sampling of admixed genotypes from the hybrid zone, whereas the modeling approach based on the JAFS relies of the joint distribution of allele frequencies between parental populations located outside of the hybrid zone. By combining these methods, we found that the 5% of loci with higher-than-average introgression rates identified with the JASF approach generally display shifts in both geographic and genomic cline centers, as illustrated by their tendency to have non-zero cline center and α parameter values (Fig. 7). Therefore, our method to detect increased introgression while accounting for the demographic divergence history identifies genomic regions that also depart from the genome average in those places where the two species meet and admix.

Discordant clines were not symmetrically distributed between the S. senegalensis and S. aegyptiaca side of the hybrid zone. The majority of the highly introgressed loci were shifted into the S. aegyptiaca geographic range. For these loci, the excess of S. senegalensis ancestry (measured by the genomic cline parameter α) was positively related to the extent of the spatial shift of the cline center into the S. aegyptiaca territory. This correlation was much stronger than previously observed in a bird hybrid zone study (Grossen et al. 2016), possibly due to a stronger variance in cline shift in the Solea system because of hybrid zone movement (see below). By contrast, much fewer loci were found with increased introgression rates in the opposite direction. Such loci generally displayed cline centers shifted away from the contact zone, probably because the Atlantic samples of S. senegalensis are introgressed to a lesser extent, if any, than those inside the Mediterranean as evidenced by their position along axis one of the PCA (Fig. 2b). The spatial shift in cline center was in this case only weakly positively correlated with the excess of S. aegyptiaca ancestry. This could be expected, however, since the genomic cline method has limited power to infer cline parameters outside the range of observed hybrid indexes in admixed genotypes (Gompert et al. 2012a), which in our case were biased in favor of S. aegyptiaca ancestry.

Differential introgression patterns and the dynamics of the hybrid zone

The inferred history of divergence between S. senegalensis and S. aegyptiaca places the beginning of gene flow ca. 18 to 30 Kyrs ago, supporting a scenario of SC at the end of the last glacial period. Therefore, the hybrid zone is probably sufficiently old for the dynamics of introgression to be well established. This supports the idea that reproductive isolation between S. senegalensis and S. aegyptiaca was strong enough to prevent genetic homogenization throughout most of the genome during this period. At the same time, the small fraction of introgressing loci that were still able to cross the species barrier indicate the existence of genomic regions unlinked to reproductive isolation loci, either due to a local lack of barrier loci and/or to a high local recombination rate (Roux et al. 2013).

Whether prezygotic or postzygotic effects could explain the observed asymmetric introgression under the hypothesis of a stable hybrid zone remains to be fully addressed. Alternatively, the geographic pattern of asymmetrical introgression evidenced here could reveal a movement of the hybrid zone (Buggs 2007), especially if the markers concerned are randomly spread across the genome (Wielstra et al. 2017). In the absence of a well-assembled reference genome, the genomic distribution of introgressing loci could not be addressed. However, a northward movement of the hybrid zone has already been proposed based on a temporal trend of increasing abundance of S. aegyptiaca in Bizerte lagoon (Ouanes et al. 2011). Therefore, our results probably reflect a recent or ongoing shift of the hybrid zone, with S. aegyptiaca incorporating the small fraction of compatible genes from S. senegalensis as it moves northward in its territory, leaving behind a tail of neutral or quasi-neutral introgressed alleles. Similar signatures have already been reported in other sister species such as in house mice (Wang et al. 2011), rabbits (Carneiro et al. 2013), salamanders (Visser et al. 2017), toads (Arntzen et al. 2017), newts (Wielstra et al. 2017), and chickadees (Taylor et al. 2014). The possible causes of hybrid zone movement include the tracking of environmental changes, fitness differences among individuals from the two species, or gradients of population density (Barton and Hewitt 1985; Buggs 2007; Taylor et al. 2015; Gompert et al. 2017). Future studies will have to establish which of these hypotheses accounts for the asymmetric introgression pattern in soles.

Opposite to the preferential direction of introgression, we also found a few loci within the S. senegalensis genetic background whose geographic center was apparently shifted far away from the contact zone. This pattern was also captured by the first and third PCA axes, along which we observed a slight differentiation between the S. senegalensis samples from Dakar and Cadiz on one side and Annaba on the other, which seems to be congruent with increased introgression in Mediterranean compared to Atlantic S. senegalensis samples. Two non-mutually exclusive explanations can be put forward to account for this difference. Strong barriers to gene flow only have a delaying effect on the dynamics of introgression (Barton 1979). If the barrier is strong enough, many generations may be required for advantageous or simply neutral alleles to extricate themselves from the hybrid zone, depending on the balance between the selective coefficient of the focal alleles and that of the deleterious genes contained in their chromosomal vicinity. Hence, spatial allele frequency patterns may appear even well after SC, and since the entrance of the Mediterranean Sea is a well-known barrier to dispersal for many marine organisms (Patarnello et al. 2007), the wave of advance of introgression clines may be delayed as they travel toward Dakar to the south of S. senegalensis’ range. Nevertheless, advantageous alleles are expected to race ahead neutral ones once they are freed from their background, a possibility would thus be that some of these loci exhibiting a shifted center could be related to an adaptive introgression of S. aegyptiaca alleles into the S. senegalensis background. Evidence for adaptive alleles spreading between species has been found in recent hybrid zones induced by human introductions (Fitzpatrick et al. 2010), but should be difficult to observe in historical hybrid zones where advantageous alleles have had ample time to spread (Hewitt 1988). Alternatively, these clines may correspond to alleles providing a local advantage to the S. senegalensis populations in the Mediterranean environment but not in the Atlantic. The hypothesis of adaptive introgression needs to be scrutinized in more details by focusing on the chromosomal signature of differentiation around the putative selected locus (Bierne 2010).

Conclusion

To conclude, our combination of analytical approaches provides new insights into the genomic architecture and the dynamics of gene flow between two divergent but still interacting parapatric species. Despite a relatively modest geographic coverage and the scarcity of available admixed genotypes in the tension zone, the genome-wide analyses taking into account the inferred history of divergence provided an efficient way to detect loci with deviant introgression behaviors. This is complementary to classical approaches based on genomic and geographic cline analyses. Our results bring new support for the tension zone model against a simple coexistence in sympatry. We show that differential gene flow has shaped genetic divergence across the tension zone, although most loci behave as if they were sitting on a congealed genome, rendering very unlikely the future remixing of the two genes pools via the creation of a hybrid swarm. Nevertheless, a few genes seem able to escape counter-selection, possibly due to different underlying processes. The first involves a possible movement of the zone, in which a shift of species range boundary leaves a tail of neutral introgressed alleles behind the front of the invading species. The second is possibly related to the existence of a physical barrier to gene flow near Gibraltar and/or the spread of globally or locally adaptive alleles into the range of the invaded species. Our results thus provide a snapshot on the genetic outcome of evolutionary processes potentially involved when divergent gene pools come back into contact after a long period of geographical isolation.

Data archiving

Demultiplexed RAD sequencing read data (fastq files) are available on NCBI under BioProject Accession: PRJNA443143. VCF files and input files for dadi are available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.h508g38.