Gene-Flow in a Mosaic Hybrid Zone: Is Local Introgression Adaptive?

Abstract

Genome-wide scans of genetic differentiation between hybridizing taxa can identify genome regions with unusual rates of introgression. Regions of high differentiation might represent barriers to gene flow, while regions of low differentiation might indicate adaptive introgression—the spread of selectively beneficial alleles between reproductively isolated genetic backgrounds. Here we conduct a scan for unusual patterns of differentiation in a mosaic hybrid zone between two mussel species, Mytilus edulis and M. galloprovincialis. One outlying locus, mac-1, showed a characteristic footprint of local introgression, with abnormally high frequency of edulis-derived alleles in a patch of M. galloprovincialis enclosed within the mosaic zone, but low frequencies outside of the zone. Further analysis of DNA sequences showed that almost all of the edulis allelic diversity had introgressed into the M. galloprovincialis background in this patch. We then used a variety of approaches to test the hypothesis that there had been adaptive introgression at mac-1. Simulations and model fitting with maximum-likelihood and approximate Bayesian computation approaches suggested that adaptive introgression could generate a “soft sweep,” which was qualitatively consistent with our data. Although the migration rate required was high, it was compatible with the functioning of an effective barrier to gene flow as revealed by demographic inferences. As such, adaptive introgression could explain both the reduced intraspecific differentiation around mac-1 and the high diversity of introgressed alleles, although a localized change in barrier strength may also be invoked. Together, our results emphasize the need to account for the complex history of secondary contacts in interpreting outlier loci.

GENETIC barriers between related lineages are often semipermeable, varying in strength across the genome (Harrison 1986). As such, hybridization can lead to adaptive introgression, the transfer of selectively beneficial alleles between species (Parsons et al. 1993; Arnold 2004; Heliconius Genome Consortium 2012; Hedrick 2013). However, the frequency and importance of such hybridization events remain hotly debated (Mallet 2007; Abbott et al. 2013; Barton 2013). It is now common to address such questions by scanning genome sequences for regions of enhanced or reduced differentiation. Universally advantageous alleles can usually cross barriers to gene flow without much delay (e.g., Piálek and Barton 1997), and neutral loci linked to such alleles are expected to display unusually low levels of differentiation. In contrast, barrier loci will delay the introgression of neutral alleles in proportion to their linkage (Barton 1979; Barton and Bengtsson 1986); that is the basis for the investigation of genomic islands of differentiation (Turner et al. 2005; Hohenlohe et al. 2010; Nadeau et al. 2012).

Interpreting regions of unusual differentiation can be difficult, not only because of variation in recombination across the genome (Nachman and Payseur 2012; Roesti et al. 2013), but also because reduced differentiation might be caused by processes other than adaptive introgression. In particular, population history may promote multiple contacts between the same lineages in different locations, and so the resulting barriers to gene flow might vary from place to place, according to local ecological gradients and population connectivity. In accordance with this hypothesis, reversed or modified associations between genetic differentiation and habitat variation (Bierne et al. 2011; Jackson et al. 2012) and partial parallelism in genomic divergence (Gagnaire et al. 2013) have both been observed in replicated contact zones of the same species pair. Differentiating between these hypotheses is challenging, because adaptive introgression leaves a genomic signature that can be substantially different from the classic “hard sweep” (Maynard Smith and Haigh 1974) where adaptive substitution is associated with a single new mutation. Adaptive introgression may involve multiple migrant copies of the same beneficial allele (Pennings and Hermisson 2006) leading to a “soft-sweep” signature, which can be difficult to discriminate from a local increase in the neutral introgression rate (see Table 1).

Table 1. Reduction in diversity after introgression in studies of adaptive introgression.

Genus	Species pair		Locus	Her/Hed
	Donor	Recipient
Canisa	latrans	lupus	Mitochondria	0.752
Eucalyptusb	cordata	globulus	Chloroplast	0.751
Euremac	hecabe b	hecabe y	Wolbachia	0.388
Littorinad	obtusata	fabalis SS	ARK	0.143
Mytiluse	edulis	trossulus	Mitochondria	0.608
Mytilusf	edulis	galloprovincialis	mac-1	0.957

d denotes the donor species; r denotes the recipient species. Only donor-derived alleles were considered. He_r/He_d, proportional diversity after introgression.

^aLehman et al. (1991).

^bMcKinnon et al. (2004).

^cNarita et al. (2006).

^dKemppainen et al. (2011).

^eQuesada et al. (1999); Rawson and Hilbish (1998).

^fPresent study.

Here we present a multilocus scan for local introgression between the marine mussels Mytilus edulis and M. galloprovincialis. These species appear to have had a complex history of fragmentation and colonization during the Quaternary, as is the case for almost all temperate organisms (Hewitt 2011). Recent analyses suggest that M. edulis and M. galloprovincialis diverged ∼2.5 million years (MY), followed by a secondary contact beginning about 0.7 MY (Roux et al. 2014). Today, they are isolated by multifarious pre- and postzygotic mechanisms (Bierne et al. 2002, 2003a, 2006). Due to natural replications of contact zones, the Mytilus species complex is particularly suitable for studying the different possible outcomes of secondary contact. Here we study an original mosaic hybrid zone in Europe, extending from the Mediterranean Sea to the North Sea (Bierne et al. 2003b; Hilbish et al. 2012) and to the British Isles (Skibinski et al. 1983). We focus on the French section of the mosaic zone that includes two patches per species, one peripheral and one enclosed within the zone (Figure 1).

Figure 1.

Localities of Mytilus spp. samples along the European coast. Panmictic patches of M. edulis are indicated by thick lines; their corresponding samples are (1) Wadden Sea (solid square) in the peripheral patch of the North Sea and (2) Lupin (solid circle) in the enclosed patch of the Bay of Biscay. Panmictic patches of M. galloprovincialis are indicated by double lines; their corresponding samples are (1) Faro (open square) in the peripheral patch of the Iberian Coast and (2) Roscoff (open circle) in the enclosed patch of Brittany. The three hybrid zones described by Bierne et al. (2003b) are indicated by dotted lines. HZ1, Iberian Coast/Bay of Biscay; HZ2, Bay of Biscay/Brittany; HZ3, Brittany/North Sea.

To study introgression between these populations, we used previously published data at 440 loci (422 amplified fragment length polymorphism, AFLP, and 18 codominant nuclear markers; Gosset and Bierne 2012). Initial analyses identified a nuclear marker, mac-1, with high interspecific differentiation in the peripheral patch but an anomalously low level of differentiation in the enclosed patch in Brittany. This outlier locus was therefore a candidate for adaptive introgression in a geographically localized region.

To test the plausibility of this hypothesis, we first sequenced a 3.1-kb region around mac-1 and analyzed how allele frequency varies along the chromosomal region. Unlike another M. edulis locus analyzed previously (Bierne 2010), we failed to identify variation in differentiation on such a small chromosomal scale. We then determined how well our data might be explained by a model of adaptive introgression (Pennings and Hermisson 2006). Inferences indicated that the level of gene flow implied by the adaptive introgression hypothesis was high, but still compatible with the strength of the genetic barrier between the two species. We further propose that the barrier to gene flow in the vicinity of mac-1 could be simply more permeable in certain hybrid zones of the mosaic. Overall, we show that although the molecular signature of adaptive introgression is difficult to identify, dedicated methods that account for the history of speciation can help in interpreting outlying levels of introgression.

Materials and Methods

Study sites and sampling

The zone of study, along the European coast, is characterized by three successive transitions between panmictic patches of each species (Figure 1; Bierne et al. 2003c; Hilbish et al. 2012). We used two geographical samples of M. edulis: the peripheral patch of the North Sea (Wadden Sea, Holland); and the enclosed patch of the Bay of Biscay (Lupin, France). We also used two geographical samples of M. galloprovincialis: the peripheral patch of the Iberian Coast (Faro, Portugal); and the enclosed patch of Brittany (Roscoff, France). These four samples have been described in Faure et al. (2008) and have been established to be representative of monospecific panmictic patches. In addition, we used a sample of M. trossulus (Tadoussac, Canada) to serve as an outgroup. Forty-eight individuals per sample were examined, except for the Brittany sample, which comprised 87 individuals. Genomic DNA was extracted from adults using the DNeasy Blood and Tissue Kit (Qiagen) following the manufacturers protocol.

Multilocus scan

For the multilocus scan, we combined existing data, comprising 422 AFLP markers, 13 codominant nuclear markers (Gosset and Bierne 2012), and 5 allozyme markers (Faure et al. 2008). To combine loci with various level of diversity, nuclear codominant markers were transformed into biallelic loci by pooling alleles according to their frequencies in the M. galloprovincialis and M. edulis reference samples (see McDonald 1994). AFLP markers are subject to several well-known caveats, notably fragment-size homoplasy (Caballero et al. 2008; Whitlock et al. 2008). To reduce this problem, we excluded AFLP bands smaller than 50 bp, which are most prone to homoplasy (Vekemans et al. 2002). Allele frequencies of AFLP markers were estimated with the Bayesian method of Zhivotovsky (1999) in AFLPsurv v. 1.0 (Vekemans et al. 2002). All data were combined for the outlier tests.

To ensure robustness, and following the recommendation of Pérez-Figueroa et al. (2010), we used six distinct methods to identify loci with unusual levels of genetic differentiation. These are the methods of Lewontin and Krakauer (1973), Beaumont and Nichols (1996), Vitalis et al. (2003), Beaumont and Balding (2004), Bonhomme et al. (2010), and a custom simulation method in which the neutral envelope of F_ST is obtained from simulations with the parameter estimates of the best-supported demographic model. Full details are given in Supporting Information, File S1.

Data analyses at the outlier locus mac-1

Chromosomal walking along mac-1:

We sequenced a region of 3.1 kb around mac-1 taking advantage of published sequences (Daguin et al. 2001), an additional upstream sequence (M. Ohresser, personal communication) and looking for homology between coding regions and expressed sequence tags of Mytilus (Tanguy et al. 2008). To describe the local variation of the edulis allele frequency around mac-1, we walked along the sequence and looked for evenly spaced genetic polymorphisms (see Figure S2 for a diagram). Four polymorphisms, two PCR product-length polymorphisms and two SNPs, were chosen, based on the allelic genealogies, to be discriminant between the two species. We extensively genotyped our four samples (North Sea, Brittany, Bay of Biscay, and Iberian Coast) for these four polymorphisms. See File S1 for technical details.

Genetic analysis:

For the PCR product-length polymorphisms described above, we computed classical genetic statistics on the subset of edulis alleles (i.e., alleles that group within the diversity of M. edulis within the mac-1 genealogy; see Figure 3) using Genetix 4.05.2 (Belkhir et al. 1996–2004): the number of allelic classes (k), the heterozygosity (He, Nei 1978), departure from Hardy–Weinberg equilibrium, and genetic differentiation were estimated by f and Θ (Weir and Cockerham 1984), respectively, and significance was tested with 1000 permutations. Finally, we evaluated the departure from the neutral expectation at mutation–drift equilibrium with the Ewens test (ε, Slatkin 1994; Slatkin 1996).

Figure 3.

Allele genealogy reconstructed by the neighbor-joining algorithm apply to the number of nucleotide differences (region 2). Sequences sampled in M. edulis patches (solid squares, North Sea; solid circles, Bay of Biscay) cluster within the “edulis” clade, while those sampled in the Iberian Coast patch (open squares) of M. galloprovincialis cluster within the “galloprovincialis” clade. In Brittany (open circles), sequences are found in both clades.

Sequence alignment was performed with Multalin (Corpet 1988) and verified by eye in BioEdit 7.5.3 (Hall 1999). Alignment gaps were excluded from further analyses. We inferred allelic genealogies for the three regions sequenced using the neighbor-joining algorithm implemented in Mega 5.0 (Kumar et al. 2004). We rooted the allele genealogy with the outgroup M. trossulus. We computed classical genetic statistics on DNA sequences using DNAsp (Rozas et al. 2003): the number of polymorphic sites (S), the number of synonymous (S_s), nonsynonymous (S_ns) and noncoding (S_nc) mutations, levels of nucleotide diversity estimated from the number of polymorphic sites (θ_S, Watterson 1975) and from pairwise differences (θ_π, Tajima 1983) and the minimum number of recombination events (R_m, Hudson and Kaplan 1985). We also computed two indicators of the distortion of the allele-frequency spectrum from the neutral mutation–drift equilibrium expectation: (1) Tajima’s D (Tajima 1989a; 1989b) and (2) Fay and Wu’s H (Fay and Wu 2000); significant departure from standard neutral model were assessed using standard-coalescent simulations.

Model of adaptive introgression:

To assess the hypothesis of adaptive introgression at mac-1, we considered a model of adaptive introgression via recurrent migration, introduced by Pennings and Hermisson (2006).

This model considers a universally beneficial allele that introgresses from the donor population(s) to the recipient population via recurrent migration and then sweeps to global fixation. Formally, the model considers two Wright-Fisher populations, each of N individuals. Each individual is characterized by two completely linked haploid loci. The first locus is biallelic, with alleles B and b. The B allele has a selective advantage s over the b allele. The second, linked locus, is neutral with an infinite number of possible alleles, and new mutations arising at rate μ (the per-gene mutation rate). The donor population is initially fixed for the beneficial B allele, while the recipient population is initially fixed for the alternative b allele, and both populations are assumed to have reached mutation–drift equilibrium at the neutral locus. After equilibration follows a period of secondary contact in which the populations exchange migrants at rate m. During this secondary contact, the beneficial B allele is introduced into the recipient population, where it will eventually reach fixation. The three most important parameters of this model are the scaled quantities θ = 2Nμ, M = 2Nm, and α = 2Ns (Pennings and Hermisson 2006).

To fit this model to data, we used the mac-1 alleles sampled from the donor populations (the two M. edulis patches) and the recipient population (the Brittany patch of M. galloprovincialis). Analytical results assume complete linkage and recent completion of the sweep (see below), and so, to make our inferences robust to the violation of these assumptions, we take as data only those M. galloprovincialis alleles that group within the diversity of M. edulis within the genealogy (i.e., only the alleles denoted with open circles in the edulis clade, Figure 3).

Maximum-likelihood approach:

We first fit the data using a maximum-likelihood (ML) approach. Our data in this case comprised the number of alleles sampled in the donor and recipient populations (n_d and n_r) and the number of distinct allelic classes in these samples (k_d and k_r). Under the assumptions described above, the probability of observing the data from the donor population is given by the Ewens’ sampling formula (Ewens 1972),

$$P\left(k_{\text{d}}|n_{\text{d}},\theta \right)=\frac{{\theta }^{k_{\text{d}}}{S}_{n_{\text{d}}}^{k_{\text{d}}}}{\theta \times \left(\theta +1\right)\dots \left(\theta +n_{\text{d}}-1\right)}=\frac{\mathrm{\Gamma }\left(\theta \right){\theta }^{k_{\text{d}}}{S}_{n_{\text{d}}}^{k_{\text{d}}}}{\mathrm{\Gamma }\left(\theta +n_{\text{d}}\right)},$$ (1)

where $S_{\text{n}}^k$ is an unsigned Stirling number of the first kind, i.e., the number of permutations of n elements in k disjoint cycles (Charalambides and Singh 1988). Pennings and Hermisson (2006) showed that adaptive introgression has some formal similarities to Equation 1. In particular, if α = 2Ns ≫ 1, then after the beneficial allele has reached fixation in the donor population, the number of distinct immigrant alleles that contribute to the sweep also approximate Ewens’ distribution, but with the scaled migration rate, M, playing the role of the scaled mutation rate, θ (Pennings and Hermisson 2006). We cannot observe the number of immigrant alleles directly, but we do know that these migrants were drawn at random from the donor population and that the diversity in the donor population is described by Ewens’ distribution with parameter θ. As such, the likelihood of observing our data are well approximated by

$$\begin{array}{c}\mathcal{L}\left(\theta ,M|k_{\text{d}},k_{\text{r}},n_{\text{d}},n_{\text{r}}\right)\approx P\left(k_{\text{d}}|n_{\text{d}},\theta \right){\displaystyle \sum _{i=k_{\text{r}}}^{i=n_{\text{r}}}P\left(i|n_{\text{r}},M\right)P\left(k_{\text{r}}|i,\theta \right)}\\ =\frac{\mathrm{\Gamma }\left(M\right)\mathrm{\Gamma }\left(\theta \right){\theta }^{k_{\text{d}}+k_{\text{r}}}{S}_{n_{\text{d}}}^{k_{\text{d}}}}{\mathrm{\Gamma }\left(M+n_{\text{r}}\right)\mathrm{\Gamma }\left(\theta +n_{\text{d}}\right)}{\displaystyle \sum _{i=k_r}^{n_{\text{r}}}\frac{\mathrm{\Gamma }\left(\theta \right)M^i{S}_{n_{\text{r}}}^i{S}_i^{k_{\text{r}}}}{\mathrm{\Gamma }\left(\theta +i\right)}}\end{array}$$ (2)

(where the summation is over the number of possible migrants from the donor population, which appear in the sample from the recipient population). Note that the approximation of Equation 2 does not depend explicitly on the strength of selection, α, but this new result does allow us to estimate the parameters θ and M from our mac-1 data. The parameters that maximize Equation 2, $\widehat{\theta }$ and $\widehat{M}$, are the maximum-likelihood estimates. Confidence intervals on a given parameter were the values that reduced the log likelihood by two units—obtained by maximizing the likelihood, conditional on the parameter of interest taking a suboptimal value (Edwards 1992). An R script that implements Equation 2 is included as File S2.

Approximate Bayesian computation approach:

Equation 2 is an approximation (Pennings and Hermisson 2006); furthermore, it considers only the number of allelic classes (k_d and k_r) and, therefore, ignores some of the information in the data, such as the allele-frequency spectrum. Accordingly, we also estimated model parameters using an approximate Bayesian computation (ABC) approach (Beaumont et al. 2002). This approach uses forward simulation of the adaptive introgression model and then compares summary statistics of the simulated and observed data.

To simulate the model, mutation–drift equilibrium was first achieved by sampling directly from the full Ewens (1972) distribution, as implemented in the program MONTECARLO (Slatkin 1994). Then, we ran forward-in-time simulations of the secondary contact, ending each simulation when the beneficial allele had reached fixation in the recipient population.

To compare the fit of the simulations to our data, we used several summary statistics. First, for each of the two species where d denotes the donor species and r denotes the recipient species, we calculated (1) the number of allelic classes (k_d and k_r), (2) the expected heterozygosity at the locus (He_d and He_r; Nei 1978), (3) the standard deviation in frequencies over all allelic classes (Sd_d and Sd_r), and (4) the proportion of private alleles (not found in the other species, %P_d and %P_r). We also used (5) the between-species diversity F_ST (Nei 1973) and (6) the proportional diversity after introgression described by He_r/He_d. All of these summary statistics for the true data are given in Table S1.

To estimate the parameters, we performed 950,000 forward simulations in total. For each simulation, parameters were log-tangent transformed (Hamilton et al. 2005), and we then calculated the Euclidean distance between the observed and simulated transformed parameters. The 1000 simulations with the smallest associated Euclidean distance were then retained, and the posterior distribution of the scaled parameters was estimated by means of weighted nonlinear multivariate regressions of the parameters on the summary statistics (Blum and Francois 2010). For each regression, 50 feed-forward neural networks and 25 hidden networks were trained using the R package abc (Csilléry et al. 2012) and results were averaged over the replicate networks. Priors on the model parameters were all uniform (flat), and used the following ranges: θ ∼ [0.2, 40], M ∼ [0.2, 200], α ∼ [2, 2000].

We evaluated the performance of the ABC method in two ways. First, we performed a goodness-of-fit test (Cornuet et al. 2010). This involved simulating 50,000 data sets with parameters drawn from their estimated posterior distributions and then comparing the summary statistics of these simulated data sets to the statistics of the observed data. Second, we computed the mean bias statistic, $b=1/n{\displaystyle {\sum }_i\left[e_i/v_i\right]}$, where n is the number of summary statistics, e_i is the median estimation, and v_i is the true median value of the ith data set (Excoffier et al. 2005). We generated 100 pseudo-observed data sets with known parameter values drawn from the prior distributions and computed b for each scaled parameter.

Forward simulations of adaptive introgression written in C⁺⁺ are available (see File S2) together with R scripts implementing a simple version of the ABC approach.

Models of speciation:

To compare our estimate of M at mac-1 to the extent of gene flow between the Brittany patch of M. galloprovincialis and M. edulis, we took advantage of the ABC approach described in Roux et al. (2013) that accounts for a putative genome-wide heterogeneity in introgression rates. We obtained DNA sequence data at the eight nuclear loci described in Roux et al. (2014) for the North Sea patch of M. edulis and the Brittany patch of M. galloprovincialis. Only silent positions (i.e., synonymous polymorphisms in coding regions and noncoding polymorphisms in introns or intergenic regions) were retained in the analysis.

We investigated four models of speciation differentiated by their temporal patterns of introgression (see Figure S8): (1) strict isolation (SI) between the two daughter species, (2) isolation with migration (IM) assuming continuous gene flow since the two species started to diverge, (3) ancient migration (AM), where migration is restricted to the early period of speciation, and (4) secondary contact (SC), where the two daughter populations first evolve in strict isolation and then experience gene flow in a secondary contact. For models with gene flow (IM, AM, and SC), we compared two alternative scenarios in which the effective migration rate was either homogeneous or heterogeneous among loci.

Five million multilocus simulations were performed for each model using the coalescent simulator Msnsam (Hudson 2002; Ross-Ibarra et al. 2008). We then compared the simulated and observed data sets by using an array of summary statistics widely used in the literature (Wakeley and Hey 1997; Fagundes et al. 2007; Ross-Ibarra et al. 2008): (1) nucleotide diversity (π, Tajima 1983); (2) Watterson’s θ_w (Watterson 1975); (3) net interspecific divergence (netdiv_edu-gal); and (4) between-species differentiation computed as 1 − π_S/π_T, where π_S is the average pairwise nucleotide diversity within species and π_T is the total pairwise nucleotide diversity of the pooled sample across species. We assessed departure from mutation–drift equilibrium using Tajima’s D (Tajima 1989a,b). We also classified the variable genomic positions according to whether they are exclusively polymorphic in M. edulis (S_x.edu), exclusively polymorphic in M. galloprovincialis (S_x.gal), with different alleles fixed in both species (S_f), or sharing the same polymorphism (S_S). We computed the average and standard deviation of each of these statistics across the surveyed loci by using MScalc (available from http://www.abcgwh.sitew.ch/; see Roux et al. 2011). We provide the observed values computed from the sequenced data in Table S2. Full detail of priors, model choice and model checking procedures, as well as parameter estimation, are described in File S1.

Results

Detection of interspecific and intraspecific outliers

Figure 2A compares interspecific and intraspecific differentiation at screened loci. The corresponding figure in M. edulis is provided (see Figure S1). Differentiation between patches of the same species was generally low (in M.galloprovincialis F_ST = 0.021; in M. edulis F_ST = 0.024), but highly variable between loci, and a few loci were identified as outliers using several methods of detection. In particular, all methods agreed that three loci were outliers within M. galloprovincialis (ACG.CGA.206 F_ST = 0.189; mac-1 F_ST = 0.192 and CAG.CTC.317 F_ST = 0.224; Figure 2A), and two loci within M. edulis (ACG.CGA.152 F_ST = 0. 294 and CAG.CGA.363 F_ST = 0.374; Figure S1). In the comparison between species, differentiation was also low (F_ST = 0.068) but the distribution of F_ST values was overdispersed, as one would expect when a semipermeable barrier to gene flow isolates two species. The locus mac-1, which was highly differentiated within M. galloprovincialis, was also highly differentiated between species (F_ST = 0.507) and was the only locus to be detected as an outlier in both kinds of comparison.

Figure 2.

Genome scan of inter- and intraspecies differentiation. (A) F_ST values between species against values within M. galloprovincialis (Iberian Coast vs. Brittany). Values for the F_ST between species were based on the mean allele frequency of the two patches for each species. Dotted lines represent the threshold at 99% of the Lewontin and Krakauer test for the two comparisons. Dashed lines represent the 95% and 99% quantiles of the neutral envelope of F_ST obtained from our custom simulation method. Loci detected by all methods as outliers within M. galloprovincialis are depicted by open circles while those outliers in four of six methods are shaded. The locus mac-1 is the only outlier with all methods in the interspecific and intraspecific comparisons (top right square). On the margins are represented the corresponding F_ST distributions. Genome scan was based on 440 loci (422 AFLPs and 18 codominant nuclear markers). All methods are described in File S1. (B) Frequency of the galloprovincialis allele at the three outlier loci in the four panmictic patches of the studied area. M. edulis patches are solid (North Sea and Bay of Biscay), while M. galloprovincialis patches are open (Iberian Coast and Brittany). We defined the galloprovincialis allele as the less frequent in the two patches of M. edulis.

To see this in more detail, Figure 2B shows allele frequencies in all four patches at the three loci identified as outliers within M. galloprovincialis. While all three loci show allele-frequency differences within M. galloprovincialis, only mac-1 shows allele-frequency differences between the Brittany patch of M. galloprovincialis and the neighboring patches of M. edulis.

Detailed analysis of mac-1

Because the pattern of differentiation at mac-1 was unique, this locus was investigated in more detail. Specifically, we sequenced three regions, comprising a total of 3.1 kb in the vicinity of mac-1 (Figure S2). All three regions contained several SNPs, and the first region also contained indels that were used to define PCR primers that generate length polymorphism of the PCR product, indel-in0. Summary statistics of the genetic diversity within each sequenced region are found in Table 2, while Table 3 describes the genetic diversity of the two PCR product-length polymorphisms.

Table 2. Genetic statistics of the sequenced regions.

Region	nedu	ngal	Sedu	Sgal	${\text{S}}_{{\text{s}}_{\mathit{edu}}}$	${\text{S}}_{{\text{s}}_{\mathit{gal}}}$	${\text{S}}_{{\text{ns}}_{\mathit{edu}}}$	${\text{S}}_{{\text{ns}}_{\mathit{gal}}}$	${\text{S}}_{{\text{nc}}_{\mathit{edu}}}$	${\text{S}}_{{\text{nc}}_{\mathit{gal}}}$
1	17	19	14	34	1	3	2	4	11	27
2	19	11	53	51	0	0	0	0	53	51
3	27	27	25	19	16	11	9	8	0	0

Region	${\theta }_{{\text{s}}_{edu}}$	${\theta }_{{\text{s}}_{gal}}$	${\theta }_{{\pi }_{edu}}$	${\theta }_{{\pi }_{gal}}$	${\text{R}}_{{\text{m}}_{edu}}$	${\text{R}}_{{\text{m}}_{gal}}$	Dedu	Dgal	Hedu	Hgal
1	0.004	0.006	0.005	0.011	2	2	−0.44	−1.78*	0.82	2.91
2	0.103	0.013	0.138	0.016	7	4	−1.03	−0.77	−0.25	6.4
3	0.005	0.004	0.009	0.007	2	2	−1.4	−1.37	2.58	−3.26

edu denotes the patches of M. edulis (North Sea and Bay of Biscay); gal denotes the Brittany patch of M. galloprovincialis. Only edulis-derived alleles were considered. Significance of D and H were tested with standard-coalescent simulations: *P-value <0.05. n: number of sequences sampled; S, total number of polymorphic sites; S_s, number of synonymous mutations; S_ns, number of nonsynonymous mutations; S_nc, number of noncoding mutations; θ_S, diversity estimated from the number of polymorphic sites (Watterson 1975); θ_π, diversity estimated from pairwise differences (Tajima 1983); R_m, minimum number of recombination events (Hudson and Kaplan 1985); D, Tajima’s D (Tajima 1989a; 1989b); H, Fay and Wu’s H (Fay and Wu 2000).

Table 3. Genetic statistics of the PCR product length polymorphisms.

Locus	nedu	ngal	kedu	kgal	Heedu	Hegal	fedu	fgal	Θ
indel-in0	92	28	9	5	0.364	0.244	0.049	−0.071	−0.006
mac-1	177	60	15	6	0.761	0.728	0.0962	−0.053	0.0097

edu denotes the patches of M. edulis (North Sea and Bay of Biscay); gal denotes the Brittany patch of M. galloprovincialis. Only edulis-derived alleles were considered. Significance of f and Θ were tested with 1000 permutations: *P-value <0.05. n, number of alleles sampled; k, number of allelic classes; He, heterozygosity (Nei 1978); f, departure from HWE (Weir and Cockerham 1984); Θ, genetic differentiation (Weir and Cockerham 1984).

We next constructed genealogies of the three regions. Figure 3 shows the genealogy of region 2 where alleles from the Brittany patch of M. galloprovincialis are found in two distinct clades, grouping with (1) alleles from both M. edulis patches (“edulis clade”; Figure 3) and (2) with the other galloprovincialis alleles (“galloprovincialis clade”; Figure 3). Similar patterns are also found in regions 1 and 3 (Figure S3) and were robust to the exclusion of possible recombinants (Table 2).

Together, these patterns are consistent with the results of F_ST outlier tests and suggest that there has been an introgression of mac-1 alleles from M. edulis into the Brittany patch of M. galloprovincialis, which lies between them. Furthermore, comparison of the Brittany alleles that group in the edulis clade (open circles within the “edulis clade”; Figure 3), with the related alleles found in the two M. edulis patches (solid squares and circles, Figure 3), suggests that very similar levels of genetic diversity are found in all three patches (see Table 2 and Table 3). For example, the proportion of diversity introgressed at the locus indel-in0 is He_gal/He_edu = 0.67 and at mac-1 is He_gal/He_edu = 0.96 (see Table 3). Additionally, there is no evidence of departure from neutrality in any of the three sets of alleles (see the results of Tajima’s D and Fay and Wu’s H in Table 2).

One possible interpretation of these results is that mac-1 has been subject to adaptive introgression from M. edulis into Brittany patch of M. galloprovincialis. The following sections use several methods to test the validity of this hypothesis.

Testing for the chromosomal signature of adaptive introgression

One genomic signature of local adaptation is a single sharp peak in the frequencies of linked neutral alleles, centered on the adaptive polymorphism (Charlesworth et al. 1997). To test for this signature we identified four polymorphisms, two indels and two SNPs, that mapped to an internal branch of their corresponding genealogy (Figure 3 for region 2 and Figure S3 for regions 1 and 3). These polymorphisms were evenly spaced along the 3.1-kb sequence, and their locations are shown in Figure S2 (triangles for indels and stars for SNPs). Figure 4 shows the variation of the edulis allele along the sequence in the four patches. Whereas the frequency of the edulis allele remains consistently low in the Iberian Coast patch (Fq_edu < 0.05) and consistently high in the North Sea and Bay of Biscay patches (Fq_edu > 0.95), we noted a slight increase of the edulis allele toward the 5′ side of mac-1 in the Brittany patch. However, there was a surprising lack of variation along the whole sequence: the edulis allele frequency remains ∼0.4 over several kilobase pairs.

Figure 4.

Frequency of the edulis allele at the four polymorphisms along the studied genomic region in the four patches. Position (in kilobase) is indicated by asterisks for SNPs and by inverted triangles for PCR product-length polymorphisms. M. edulis patches are solid (squares, North Sea; circles, Bay of Biscay), while M. galloprovincialis patches are open (squares, Iberian Coast; circles, Brittany). Sample size was n = 48, except in Brittany (n = 87).

Model fitting of adaptive introgression

The limited scale of our chromosomal walking does not allow us to reject any hypotheses with high confidence. Accordingly, we next tested whether adaptive introgression might plausibly explain the pattern of diversity observed at mac-1.

First, to illustrate the effects of adaptive introgression on diversity, Figure 5 plots the expected reduction in diversity against the rate of gene flow (see also Pennings and Hermisson 2006). When M < 0.01 we see a clear “hard sweep” with a single haplotype reaching high frequency in the recipient deme, while when M > 1, most of the heterospecific diversity introgresses into the recipient background (a classic soft sweep with no detectable reduction in diversity). For intermediate values of M, the effect is also intermediate, with both a soft sweep and a reduction in diversity.

Figure 5.

Expected sweep effect in the adaptive introgression model. The proportional diversity after introgression from d (donor species) to r (recipient species), He_r/He_d, is shown as a function of the scaled migration rate, M. Errors bars represent standard deviation calculated over 5000 replicates. Median values are depicted by circles. Parameter values were N = 10000, θ = 2, α = 1000, r = 0 (complete linkage), and n_d = n_r = 50 (sample size). Simulations were run until the beneficial allele reached fixation in the recipient background.

Building on the results of Pennings and Hermisson (2006), we fitted a model of adaptive introgression to our mac-1 data. We first developed an ML approach to estimate the population mutation rate, θ = 2Nμ, in the M. edulis populations (North Sea and Bay of Biscay), and the effective number of migrants, M = 2Nm, from these populations to the Brittany patch of M. galloprovincialis (see Equation 2). Considering only edulis-derived alleles, the total number of alleles sampled, and the number of distinct allelic classes observed were n_d = 177 and k_d = 15 for the North Sea and Bay of Biscay patches and n_r = 60 and k_r = 6 for the Brittany patch (Table 3). From these data, we obtained a relatively precise estimate of the population mutation rate: $\widehat{\theta }=3.79$ [1.95 – 6.66] (see solid lines, Figure 6A). However, the effective number of migrants was imprecisely estimated, with low migration rates rejected but no upper bound: $\widehat{M}=4.63$ [0.95–] (see solid lines, Figure 6B). The lack of an upper bound on the estimate reflects the fact that nearly all of the allelic diversity from the edulis patches had been introgressed into the Brittany patch of M. galloprovincialis.

Figure 6.

Parameter estimation in the adaptive introgression model. Maximum-likelihood estimation of (A) θ and (B) M are depicted by black curves. Maximum-likelihood values are denoted by solid vertical lines and confidence intervals by dotted lines. Posterior distributions of (A) θ, (B) M, and (C) α are depicted by shaded circles. Medians are denoted by solid vertical lines and 95% highest posterior density values by dotted lines. Values are corrected for parameter bias in the ABC inference (see text for more details).

In an attempt to improve our inference, we next used an ABC approach, which allowed us to relax the assumption of strong selection and to consider more of the information in the data. Parameter posterior distributions, corrected for mean bias, are shown in Figure 6 as shaded circles. In the following, we give their median and 95% credible intervals. We obtained an even more precise estimate of the scaled mutation rate (Figure 6A): θ_median = 2.82 [1.42 – 4.53] whose bounds are similar to the ML estimates. Estimation of the scaled migration rate (Figure 6B): M_median = 99 [24 – 154] showed a clear discrepancy between the two methods with respect to the lower bound. Compared to ML, the lower bound estimation was M ≫ 1 while its upper bound was poorly estimated (value approximates the prior upper bound). Finally, the estimation of selective strength, α_median = 402 [203 – 594] (Figure 6C) was imprecise but compatible with the assumption of strong selection used in the ML inference.

To better understand the discrepancy between the two methods, we reran the ABC estimation, using only the summary statistics that appear in the ML equations (namely, n_d, k_d, n_r, and k_r). In this case, estimates of θ and M were very similar to the MLEs (Figure S4), and there was no information on α as expected. The results of the ABC approach therefore come from the additional use of summary statistics that contain information that is not used in the ML approach. We therefore conducted a goodness-of-fit test to evaluate how well the model fit the data at mac-1 when estimating the parameters by each method (ML and ABC using all statistics). Figure S5 shows the goodness-of-fit of ML estimation (blue dots) and ABC estimation (red dots) for each statistic. In the main, both methods fit the data well (see Table S1 for quantiles and P-values) but looking at the median values, ML performs better. In particular, ABC tends to overestimate k_r and He_r, which basically explains the discrepancy in the estimation of M. The ML method proved to be the most understandable approach with a simpler model that retrieved the best fit to the data and consequently our interpretations rely on the ML estimates.

Finally, we used our simulations to assess the effects of violating some assumptions of our model: (1) the assumption of complete linkage between the selected and neutral loci and (2) the assumption that migration rates remained constant. Figure S6 shows that the sweep pattern was indeed robust to including recombination. Figure S7 suggests that our results will also be fairly robust to temporal variation in migration rates, particularly in mussels where dispersal rates are very high.

Insights into secondary contact history

Both estimation methods agree that M > 1 best explains our data if adaptive introgression has taken place. To assess the plausibility of our estimation, we analyzed several models of speciation (Figure S8) between M. edulis and the Brittany patch of M. galloprovincialis.

We first applied a model choice procedure for each of the three speciation models with gene flow to test whether heterogeneity in introgression rates was supported. We observed clear support for heterogeneous migration (Table S3, “within model”). We then applied the model-choice procedure to the speciation models, assuming heterogeneous migration (Table S3, “between models”). Secondary contact was the best-supported model with posterior probability, P_post = 0.61, confirming a previous study (Roux et al. 2014). We tested the robustness of this result by simulating 500 pseudo-observed data sets for each model. We empirically estimated that given a posterior probability of 0.61, the probability that secondary contact with heterogeneous introgression rates is the correct model was 0.987 (Table S4). It is worth mentioning that when migration rate was assumed to be homogeneous among loci, the method failed to infer secondary contact; instead ancient migration was the best-supported model (SC, P_post = 0.26 and AM, P_post = 0.65).

We then estimated the multilocus distributions of introgression rates in the best-supported SC model with heterogeneous migration (Figure 7). We found an asymmetry of introgression in favor of M. edulis: M_{gallo to edulis} = 0.15 [0.002 – 1.06] (see open bars, Figure 7) and M_{edulis to gallo} = 0.45 [0.013 – 2.84] (see shaded bars, Figure 7). However, compared to the multilocus inference, introgression at mac-1 was strongly unidirectional from M. edulis to M. galloprovincialis as suggested by the comparison of different scenarios of introgression (SC with introgression into M. edulis: P_post = 0.0013; SC with introgression into M. galloprovincialis: P_post = 0.9997; see Table S5). Furthermore introgression rate into M. galloprovincialis at mac-1 was very high compared to the extent of gene flow across loci (M_median = 2.50, solid line, Figure 7) that is consistent with the mac-1 genealogy (Figure 3). In comparison, our ML estimation of M at mac-1 in the adaptive introgression model ($\widehat{M}=4.63$, Figure 6B) was outside the 95% quantile of the multilocus distribution, but its lower bound (M_{ML lower bound} = 0.95, dashed line, Figure 7) remained consistent with the secondary contact inference.

Figure 7.

Multilocus distributions of introgression rates estimated in the secondary contact model between M. edulis (North Sea) and M. galloprovincialis (Brittany). Introgression from M. edulis into M. galloprovincialis is shown as shaded, while introgression in the opposite direction is shown as open. Distributions were obtained after randomly sampling 1000 values from each of the 2000 beta distributions retained by the ABC analysis. The solid line shows the median introgression rate into M. galloprovincialis estimated at mac-1 in the secondary contact model (M_{SC median} = 2.50). For comparison, the lower bound of the maximum-likelihood estimation in the adaptive introgression model is reported as a dotted line (M_{ML lower bound} = 0.95).

Discussion

Adaptive introgression between hybridizing species has long been proposed as a potentially important source of new adaptations in plants and, more recently, in animals too (Arnold 2004; Hedrick 2013). Most of all, compared to adaptation from new mutations or standing variation, adaptive introgression might allow big evolutionary jumps by the acquisition of complex variants (Wright 1949; Mallet 2007; Abbott et al. 2013).

Genome scan surveys of differential introgression among loci can help to localize genomic regions that introgress adaptively (Payseur 2010). Here, we took advantage of the mosaic structure of the hybrid zone of mussels along the western European coastline, to perform a multilocus scan of local introgression. As pointed out by Harrison and Rand (1989), an exciting feature of mosaic hybrid zones is that they involve independent contacts between species, each with a unique evolutionary trajectory. In addition to random fluctuations, the outcomes will be contingent to the relative abundance of the two species, local environmental conditions, and the genetic architecture of their reproductive isolation.

Our multilocus scan, based on AFLP and codominant markers, showed some genetic differences between the peripheral patch of M. galloprovincialis in the Iberian Coast and the patch enclosed in Brittany. In our analysis, AFLP markers were used mainly to obtain an estimation of the genomic distribution of F_ST values. We also have fitted a speciation model to an independent DNA sequence data set. We therefore checked whether the parameters inferred could reproduce the F_ST distribution observed at AFLP markers. The two distributions—AFLPs vs. simulations under the best demographic model inferred—were very similar (Figure S9), suggesting that we have a good estimate of the genomic distribution of F_ST. To mitigate the problem of false positives, due to an underestimation of the neutral variance of F_ST (Robertson 1975), we cross checked the results of six methods, all of which assume different population structures. In this way, we identified three loci with a robust deviation from the neutral expectations (open circles, Figure 2A). When compared to other methods, simulations of the inferred secondary contact model with heterogeneous gene flow produced an elevated variance of F_ST in both the between-species comparison, and—more surprisingly—in the within-species comparison. Nevertheless, even under this scenario, the combination of a high differentiation within and between species observed at mac-1 was not observed under neutrality (see dashed lines, Figure 2A). The genealogy of mac-1 confirmed that this was due to local introgression of M. edulis alleles into the Brittany patch of M. galloprovincialis, but not into the patch on the Iberian Coast (Figure 3). This pattern contrasts strongly with results at the other loci analyzed.

The locus-specific nature of the introgression at mac-1 suggested that selective processes were acting rather than purely demographic ones. The highly asymmetrical introgression rates from M. edulis (Table S5) contrasts with the multilocus pattern (Figure 7), and so we hypothesized that mac-1 might have been subject to adaptive introgression, with edulis alleles sweeping to fixation in the Brittany patch of M. galloprovincialis.

Demonstrating adaptive introgression requires a combination of evidence that is hard to obtain in most species (Vekemans 2010). Few studies have undertaken experiments to confirm the fitness advantage of the introgressed traits/alleles in the recipient backgrounds. Exceptions include the adaptive introgression of drought tolerance in Helianthus annuus (Whitney et al. 2010) and the adaptive introgression of rodent poison resistance in the house mouse, Mus musculus domesticus (Song et al. 2011). In contrast, most studies rely on indirect evidence involving geographical variation in allele frequencies, analysis of DNA polymorphism, and the reconstruction of gene genealogies. For example, Heliconius Genome Consortium (2012), Pardo-Diaz et al. (2012), and Smith and Kronforst (2013) provided strong evidence for the adaptive introgression of wing color pattern genes involved in Müllerian mimicry between hybridizing species of Heliconius. In humans, Mendez et al. (2012a,b, 2013) found a high frequency of divergent haplotypes at two immune genes, STAT2 and OAS1, in modern Melanesians. These haplotypes share recent common ancestry with archaic hominin sequences (STAT2, Neanderthal; OAS1, Denisovan). This result is suggestive of ancient introgressions, but their adaptive nature remains hypothetical. Similarly Roux et al. (2013) identified genomic hotspots of introgression between two highly divergent sea squirt species but could draw no conclusions about their adaptive nature, illustrating that specific tests need to be developed.

The patterns that we observe at mac-1 differ in one respect from previously studied cases of outlying introgression. In particular, in most previous cases, there has been a reduction in the genetic diversity in the recipient population, when compared to the donor population (see Table 1). In contrast, at the mac-1 locus, 96% of the M. edulis diversity was introgressed to M. galloprovincialis, and there were no other signs of a classical selective sweep (Table 1, Table 2, and Table 3).

However, it remains possible that the mac-1 alleles have introgressed via a soft sweep, with little loss of diversity (Pennings and Hermisson 2006). We tried two approaches to test for a soft sweep at this locus. First, we analyzed four polymorphisms along the sequence (Figure 4) with the hope of observing a gradient of introgression. Unfortunately, the chromosomal scale proved to be too small. Second, we used two new inference methods to test whether the rate of migration required to generate the soft sweep effect was consistent with the existence of an efficient genetic barrier at mac-1 elsewhere in the range. Our inference suggested that M = 2Nm > 1 was required to explain our data under the adaptive introgression model, and the ML estimate ($\widehat{M}=4.63$; Figure 6B) was outside of the 95% quantile of the multilocus distribution of introgression rates between M. edulis and the Brittany patch of M. galloprovincialis (Figure 7). However, its lower bound remained consistent with the secondary contact inference (M_{ML lower bound} = 0.95, dashed line, Figure 7), and so we cannot reject the hypothesis of adaptive introgression.

While we cannot reject adaptive introgression, we must also recognize alternative hypotheses. Migration pressure can easily spread gene combinations, even if these are universally selected against (e.g., Barton 1992). It is therefore possible that our results could be explained by a geographically and genomically localized change in the barrier strength between M. edulis and M. galloprovincialis. The genomic region around mac-1 may contain barrier loci that were maintained in some patches, but not in the Brittany patch—perhaps due to differences in local ecological selection or because an asymmetric intrinsic incompatibility managed to uncouple from the tension zone.

Together, our results emphasize the difficulties in interpreting outlier loci in populations that have undergone secondary contact. This is unsuprising given that patterns of introgression after such a contact can vary dramatically, depending on the way divergence occurred in allopatry, the resulting genetic architecture of the barrier, and the local landscape where the contact takes place—including its population density, connectivity, and environmental variation (Bierne et al. 2011, 2013a; Domingues et al. 2012; Strasburg et al. 2012; Abbott et al. 2013; Gagnaire et al. 2013). Nevertheless, our best hope of understanding these issues is to apply dedicated methods, such as those developed here, to much larger genome-wide data sets.