Introduction

Introgressive hybridization can promote or impede the progress of speciation and thus the emergence and maintenance of species diversity (Abbott et al., 2013). On one hand, hybridization may lead to the collapse of distinct species upon secondary contact or when selection regimes change (Seehausen et al., 1997; Taylor et al., 2006; Gilman & Behm, 2011; Vonlanthen et al., 2012; Rudman & Schluter, 2016). Conversely, hybridization may release lineages from constraining genetic correlations (Grant & Grant, 1994) increasing their evolvability (Parsons et al., 2011; Renaud et al., 2012; Seehausen et al., 2014; Selz et al., 2014; Stelkens et al., 2014, Fig. 1), which may lead to the emergence of distinctively adapted hybrid populations that occupy a niche space different from either parental species (Rieseberg et al., 2003; Nolte et al., 2005; Mallet, 2007; Stelkens & Seehausen, 2009; Abbott et al., 2013).

Fig. 1
figure 1

Schematic representation of different hybridization scenarios between species (solid outlined) and their respective potential outcomes (dashed lines): The outlined ellipses represent the 95% confidence ellipses of the phenotypic (P) variance/covariance matrices of two species. The lines indicate the direction of highest covariance (p max) and perpendicular to it the direction with lowest covariance (p min). The more constrained the matrix in phenotype space, the higher is p max and the lower is p min, which results in a more eccentric ellipse. In a the P matrices of the two species are aligned along a common line of least resistance, with p max having the same slope and the same intercept. Gene flow between these two species should lead to an elongation of the P matrix because variance is increased along p max. In b the two P matrices have still the same slope but their intercepts are different, in c the opposite is the case, where the P matrices do not have the same slope but their intercepts are the same and in d both intercept as well as slope are different between the matrices. The matrices in b, c and d all have diverged lines of least resistance; therefore, gene flow leads to P matrices which have a lower eccentricity because variance is increased along p min

Full size image

Adaptation can be characterized as the movement of a population in phenotype space towards a local adaptive optimum, where the mean phenotype expressed in the population fits a given environment (Wright, 1932; Schluter, 2000; Orr, 2005; Calsbeek et al., 2011). Phenotypic evolution towards adaptive peaks is thought to be constrained along the so-called genetic “lines of least resistance” (LLR), which can be quantified as the leading eigenvector of the genetic variance–covariance matrix G (Lande, 1979; Schluter, 1996; Steppan et al., 2002; Klingenberg, 2010; Blows et al., 2015). The LLR is assumed to account for the largest proportion of heritable phenotypic variation and phenotypic evolution is predicted to be biased towards the direction of the LLR (Lande & Arnold, 1983; Schluter, 1996). The LLR is influenced by mutation, gene flow, drift and selection (Lande, 1979; Steppan et al., 2002; Guillaume & Whitlock, 2007; Chapuis et al., 2008; Bailey et al., 2013). Selection may reorient the LLR towards the direction of the most prevalent selection regime (Lande, 1979; Schluter, 1996; 2000; Arnold et al., 2008) and gene flow may reorient the LLR towards that in the source population of gene flow (Guillaume & Whitlock, 2007).

In the absence of quantitative genetic data, the G matrix can be approximated by the P matrix (Cheverud, 1988), which is based on phenotypic data from wild populations (Arnold et al., 2008). P is thus defined as the combination of the genetic and environmental covariance matrices, that is, G + E (Lande, 1979; Arnold & Phillips, 1999), where both effects could also interact (G × E; Falconer, 1989). Consequently, P matrices also include phenotypically plastic effects (Lande, 2009; Draghi & Whitlock, 2012; Wood & Brodie, 2015). Whereas many studies investigated the stability of the P matrix through time, and investigated whether evolution occurs along common LLRs in the P matrix (e.g. Schluter, 1996; Arnold et al., 2008; Eroukhmanoff & Svensson, 2008; Hine et al., 2009; Lucek et al., 2014a, b), few studies have addressed to which degree gene flow may affect the P matrix and hence influence the evolutionary trajectory or potential of a population (Guillaume & Whitlock, 2007; Selz et al., 2014; Roseman, 2016). Theory suggests that gene flow between two diverged populations can change the shape of a P matrix, which is estimated by the eccentricity of the P matrix, i.e. the ratio of its two leading eigenvectors. Biologically, the degree of eccentricity is related to the extent of genetic constraints, where increased eccentricity reflects stronger covariation among traits, while low eccentricities imply reduced covariance among traits and thus fewer genetic constraints (Steppan et al., 2002; Jones et al., 2003; Eroukhmanoff & Svensson, 2011; Bailey et al., 2013). Gene flow may increase the phenotypic variance, leading to an increase (stronger covariation) or decrease (relaxed covariation) in eccentricity depending on whether or not gene flow increases variance along LLR or along other axes of the P matrix (Guillaume & Whitlock, 2007; Fig. 1).

Here we use empirical data to test these predictions for the potential effects of gene flow on the P matrix. We study two sister species of African cichlid fish, Pundamilia pundamilia (Seehausen & Bouton 1998) and Pundamilia nyererei (Witte-Maas & Witte 1985), which co-occur at several islands in the southern part of Lake Victoria (Fig. 2). Variation in water turbidity between islands is associated with different divergent selection regimes affecting Pundamilia species differentiation, with more gene flow between the sympatric species at islands with less clear water (Seehausen et al., 1997, 2008). Theory suggests that if the LLR have diverged between the Pundamilia species, hybridization should reduce the eccentricity of species-specific P matrices (Guillaume & Whitlock, 2007). In contrast, if the Pundamilia species have not diverged in their LLR but rather differ in their mean position along a shared LLR, then gene flow between the two species should lead to an increase in species-specific eccentricity with increasing gene flow (Fig. 1). When the species share the LLR and their mean trait values on the LLR, eccentricity may not change at all (Guillaume & Whitlock, 2007). We therefore first estimate if the Pundamilia species pair has diverged along the LLR or not at Makobe Island, being the island that is the least affected by recent changes in water turbidity and where there is no evidence for ongoing interspecific gene flow between the two species (Seehausen et al., 2008). We then test the hypothesis that introgressive hybridization can relax genetic constraints in those parts of trait space where the two species are diverged in their LLR based on species-specific trait-by-trait P matrices.

Fig. 2
figure 2

Overview of the studied populations: a Map of Lake Victoria (Greg, 2015). b Detailed view of the Speke and Mwanza Gulf where the four islands Makobe, Kissenda, Python and Luanso are located. c Male specimens of Pundamilia nyererei (left) and P. pundamilia (right) for each island. For Luanso, an exemplary specimen of the existing hybrid swarm is depicted. The pairwise F ST values between sympatric Pundamilia species are given for each island (taken from Seehausen et al., 2008)

Full size image

We further illustrate the effects of gene flow in trait-by-trait P matrices on the multivariate phenotype in the context of the morphospace occupied by the most abundant species in the Makobe Island cichlid community (Seehausen et al., 1997, 2008). By including three additional sympatrically occurring species that represent other ecological guilds of the extant adaptive radiation of cichlids from Lake Victoria and together with the two Pundamilia species amount to ~80% of the local fish community in abundance (Seehausen & Bouton, 1997), we aim to put the potential effects of gene flow on Pundamilia phenotypes and P matrices into the community context of this adaptive radiation.

Materials and methods

Sampling

Cichlid fish were collected between 1993 and 1998 from four islands (Makobe, Python, Kissenda and Luanso) within the Mwanza Gulf in southern Lake Victoria (Fig. 2). Each specimen was fixed in formaldehyde-solution immediately after capture and then transferred into an alcohol solution with increasing concentration (30, 50 and finally 70%). Individuals were identified to species level based on their morphology by OS. Only males of each species were used in this study as females cannot be unambiguously assigned to a species. At Luanso Island, the two species cannot be distinguished genetically and only a single phenotypically highly variable population of Pundamilia exists with some males resembling either of the species but most males being intermediate in phenotype (Seehausen et al., 2008). They are consequently treated as a single, phenotypically variable population (P. “hybrid”).

Morphological analysis

For each individual, 13 linear morphological distances were measured to the nearest 0.01 mm using a digital caliper. Measurements were especially taken on the head to capture taxon-specific ecological relevant trophic morphology (Barel et al., 1977): head length (HL), lower jaw length (LJL), lower jaw width (LJW), snout length (SnL), preorbital depth (POD), cheek depth (ChD), eye length (EyL), eye depth (EyD), interorbital width (IOW), preorbital width (POW), snout width (SnW), body depth (BD) and standard length (SL). The latter was measured to size correct all other linear traits (see below). Many of these traits were found to have a heritable component in a common garden experiment using Pundamilia species from Kissenda Island (Magalhaes et al., 2009). Each fish was measured twice to estimate repeatability, which was generally >95% for all measurements (results not shown). For all further analyses, the average of both independent measurements was used. Measurements, which were not reliable due to unusual body positions, were omitted. To account for potential size-related effects and allometry, each measurement was first normalized by mean-scaling (Kirkpatrick, 2009), and subsequently regressed against standard length, retaining the residuals (Reist, 1986). To retain potential differences in trait means among populations, both the mean-scaling as well as the size correction was performed combining all individuals from all populations. Missing data were replaced by the average trait value of a population after mean-scaling.

Comparing P matrices

For each Pundamilia population on each island, the phenotypic variance–covariance (P) matrices of trait-by-trait morphospaces were calculated based on the size-corrected residuals of the two traits. Using the 12 size-corrected phenotypic traits, this resulted in 66 different trait-based pairwise P matrices per population. The eccentricity of each P matrix was then calculated as the ratio between the length of the orthogonal axes of the 95% confidence ellipsoid that accounts for the highest (p max) and lowest (p min) variance, respectively (see Fig. 1). Ellipses, p max and p min, were calculated using a custom made script based on an implementation in the CAR package (Fox & Weisberg, 2011) in R 3.1.2 (R Development Core Team 2014).

To further test if species divergence may commonly involve a shared or a diverged LLR, trait-by-trait-based P matrices were compared between P. pundamilia and P. nyererei from Makobe Island by calculating the pairwise differences in the intercept and the angle between the two leading eigenvectors of two P matrices. The latter is given as the inversed cosine of the dot product that is divided by the summed length of both eigenvectors (Schluter, 1996). The significances of these pairwise measurements were further established using a bootstrapping procedure with 1000 permutations (Berner, 2009). The two species were either considered to share a common LLR for a given trait-by-trait P matrix if neither the intercept nor the slope was significantly different (Fig. 1a) or to deviate from a common LLR if the slope and/or the intercept differed significantly (Fig. 1b–d). The eccentricity of trait-by-trait P matrices was furthermore calculated for each species at each island separately.

The degree of pairwise genetic divergence (F ST) between sympatric Pundamilia species based on microsatellite loci was obtained for each pair from Seehausen et al. (2008) and subsequently used as a measure of gene flow between the two focal species at the different islands (Fig. 2c). A linear mixed effect model was employed to test if eccentricity differs between the P matrices that have a diverged LLR and those that have a common LLR and if eccentricity is affected by gene flow (F ST) and lastly if there is an interaction between either effect using island as a random factor. This analysis was separately performed for P. pundamilia and P. nyererei where in both cases the same set of hybrid individuals were included for Luanso Island.

Estimating the effects on the multivariate morphospace

In order to reconstruct the occupied multivariate morphospace of both Pundamilia species in relation to other cichlid species, the populations from Makobe Island showing the highest genetic divergence among our studied Pundamilia species pairs (Seehausen et al., 2008) were used together with individuals from three other cichlid species present at Makobe Island, including the two most abundant ones [(Seehausen et al., 1998): Mbipia mbipi (Seehausen, Lippitsch & Bouton 1998), Neochromis omnicaeruleus (Seehausen & Bouton 1998) and Paralabidochromis cyaneus (Seehausen, Bouton & Zwennes 1998); see Table 1 for details]. These species differ furthermore in their ecology: P. pundamilia feeds on benthic invertebrates, P. nyererei feeds predominantly on zooplankton, whereas M. mbipi and N. omnicaeruleus feed predominantly on epilithic algae and P. cyaneus on insects and epilithic algae (Seehausen et al., 1998).

Table 1 Sampling summary
Full size table

The size-corrected dataset was used to conduct a principal component analysis (PCA) using all individuals from all species from Makobe Island. The scores of the two leading PC axes were then compared among the five cichlid species using an ANOVA with a TukeyHSD post hoc test. To further infer where individuals from other islands would fall in the Makobe Island morphospace, they were subsequently projected into the Makobe morphospace following an approach described in Selz et al. (2014) and Lucek et al. (2014c). In short, this method takes the PC axes that were calculated using the Makobe Island individuals and subsequently infers where a given individual from another island would be placed in the mutli-species morphospace of Makobe Island. The PC scores of the projected Pundamilia species and the Pundamilia species at Makobe Island were subsequently compared using an ANOVA with species (P. pundamilia, P. nyererei or P. “hybrid” for Luanso Island) and island (Makobe and either Python, Kissenda or Luanso respectively) as factors. All analyses were performed in R.

Results

Trait-by-trait P matrices

Out of 66 pairwise trait comparisons between the two sympatric Pundamilia species at Makobe Island, 18 trait combinations show a significantly diverged LLR, especially involving lower jaw length (6 significant comparisons; Table S1), interorbital width and snout length (5 significant comparisons each; Figure S1; Table S1). Out of these 18 cases, the angle between P matrices was significantly larger than zero on seven occasions, whereas the remaining eleven cases showed a significant difference in the intercept (Figure S1; Table S1).

For P. nyererei, we found that trait-by-trait P matrices with an LLR that both Pundamilia species at Makobe Island have in common were at all islands more eccentric than matrices with diverged LLR (F 1,258 = 13.6, P < 0.001; Fig. 3). Eccentricity in P. nyererei was, however, not affected by the extent of gene flow (F 1,258 = 0.1, P = 0.788), independent of whether LLR were shared with the sister species or diverged from it (a non-significant interaction between gene flow and divergence in the LLR; F 1,258 = 0.8, P = 0.385). For P. pundamilia, on the other hand, P matrices with an LLR that both species at Makobe have in common did not differ in their eccentricity from matrices with a diverged LLR at islands where gene flow is little, but in populations that have more gene exchange with the sister species, P matrices with diverged LLR lose their eccentricity as indicated by the significant interaction between F ST and divergence in the LLR (F 1,258 = 6.0, P = 0.015).

Fig. 3
figure 3

Average eccentricity (±1 SD) of all pairwise trait matrix comparisons that either showed a significantly diverged line of least resistance (LLR; dashed line) or share a common LLR (solid line) for a Pundamilia nyererei and b P. pundamilia separately for each island. In addition, the results of a linear mixed effects model are given, testing for a statistical association of eccentricity with either the degree of interspecific gene flow (F ST), the number of diverged/undiverged LLR and its interaction. For Luanso Islands the same Pundamilia hybrids were used

Full size image

Multivariate phenotypic changes

The two leading principal component (PC) axes for the overall morphospace, comprising five species of the Makobe Island community explained 73.2% of the total variation (63.6 and 9.6% on PC 1 and 2, respectively; Fig. 4). Traits that accounted for most of the variation were LJL and LJW on the first and LJW and ChD on the second PC axis, respectively (Table S2). Individual scores differed significantly among species along the first PC axis (F 4,154 = 198.1, P < 0.001), where all pairwise post hoc comparisons were significant except for P. pundamilia and M. mbipi (Table S3). On the second PC axis, species were significantly different too (F 4,154 = 24.7, P < 0.001), but only the comparisons involving N. omnicaeruleus yielded significant post hoc tests (Table S3).

Fig. 4
figure 4

Individual PC scores for the studied species assemblage from Makobe Island in a common morphospace (top). The ellipses indicate the 95% confidence boundaries for each species and represent the underlying P matrix (red P. nyererei, blue P. pundamilia, orange M. mbipi, green N. omnicaeruleus, light blue Pa. cyaneus). The bottom panels give the position of Pundamilia specimens projected into the common morphospace of the Makobe species community (only P. pundamilia and P. nyererei are shown) for Kissenda and Python as well as the Pundamilia hybrid individuals in Luanso. Dashed lines represent the 95% confidence boundaries of the projected individuals for each island

Full size image

When projecting Pundamilia from the other islands into the Makobe Island morphospace, the populations from Kissenda Island showed a slightly increased morphospace relative to those from Makobe (Fig. 4), where their PC scores differed both between species (F 1,138 = 121.8, P < 0.001) and island (F 1,138 = 4.1, P = 0.044) along the first axis. In contrast, PC scores along the second axis differed only for the factor island (F 1,138 = 17.3, P < 0.001) but not species (F 1,138 = 1.6, P = 0.209). The Pundamilia populations from Python Island are more divergent from the Makobe populations, differing in their PC scores along the first and second axis for the factor species (PC1: F 1,129 = 56.1, P < 0.001, PC2: F 1,129 = 10.5, P = 0.002) and island (PC1: F 1,129 = 9.7, P = 0.002, PC2: F 1,129 = 49.5, P < 0.001). Lastly, the hybrid Pundamilia population from Luanso Island occupies a distinct part of the Makobe morphospace that matches the morphospace occupied by the ecologically distinct species M. mbipi in the Makobe community (Fig. 4), where M. mbipi and P. “hybrid” specimens differ in their PC scores along the first (F 1,62 = 7.3, P = 0.009) but not the second (F 1,62 = 0.1, P = 0.870) axis. Lastly, the PC scores of the Pundamilia species at Makobe Island and the projected P. “hybrid” from Luanso differ along the first PC axis for the factor species (F 1,103 = 38.2, P < 0.001) and island (F 1,128 = 5.3, P = 0.023) but not on the second PC axis (species: F 1,103 = 0.7, P = 0.412; island: F 1,103 = 0.9, P = 0.352).

Discussion

Theory predicts the evolution of the genetic (G) and its related phenotypic (P) variance–covariance matrix to be constrained along the so-called lines of least resistances (LLRs; Schluter, 1996; Steppan et al., 2002; Eroukhmanoff, 2009; Klingenberg, 2010). Both G and P may change rapidly through drift, selection and gene flow (Lande, 1979; Steppan et al., 2002; Chapuis et al., 2008; Bailey et al., 2013). The role of gene flow between populations or introgressive hybridization between species has, however, been mainly studied in theory (Steppan et al., 2002; Guillaume & Whitlock, 2007; Seehausen et al., 2014). Here, we provide empirical data consistent with theoretical predictions for the effect of gene flow or introgressive hybridization on the P matrices in natural populations of cichlid sister species.

Our finding that more than two-thirds of all trait-specific P matrices showed a shared LLR in the sympatric sister species pair of Pundamilia pundamilia and P. nyererei at Makobe Island is consistent with the idea that evolution is constrained by LLRs. This is expectation applies particularly in the early stage of species divergence (Schluter, 1996, 2000) and may thus be prominent in the evolutionary young Lake Victoria radiation of haplochromine cichlids (Seehausen, 2006). Over longer time, selection may overcome such constraints leading to an increased number of trait combinations that evolve diverged LLRs (Schluter, 2000; Eroukhmanoff, 2009). Introgressive hybridization can, however, rapidly release lineages from G matrix constraints and potentially increase the evolvability of the resulting hybrid population (Parsons et al., 2011; Renaud et al., 2012; Seehausen et al., 2014; Selz et al., 2014; Fig. 1).

Interspecific gene flow may particularly affect the eccentricity of P matrices that have a diverged interspecific LLR, where a decreased level of eccentricity may reflect reduced trait covariation and hence a relaxation of former genetic constraints (Guillaume & Whitlock, 2007). Consistent with this, we observe in P. nyererei at all islands weaker eccentricity in P matrices with LLRs that have diverged in the sympatric sister species than those with common LLRs. This may be an effect of low levels of recent gene flow between the species or of gene flow in the past. Recent demographic analyses using whole-genome sequence data have revealed that interspecific gene flow was predominantly from P. nyererei into P. pundamilia with much less gene flow into P. nyererei (Meier et al. 2016). It is possible that gene flow changes the eccentricity of the P matrix of an adapted population only when the migration rates are high (Guillaume & Whitlock, 2007). At the islands, we studied P. pundamilia is less abundant than its sister species P. nyererei (Seehausen & Bouton, 1997), which may explain the asymmetry of gene flow between the species (Meier et al. 2016).

Also consistent with theoretical predictions, we observe among P. pundamilia populations differences in P matrix eccentricity between populations that have received different amounts of interspecific gene flow from P. nyererei (Fig. 3). Specifically, P matrices that have LLRs that are diverged from those in the sister species are less eccentric at islands where P. pundamilia received more gene flow from P. nyererei, whereas LLRs that are shared between the species are more eccentric where gene flow was more common. Overall, P. nyererei shows an increased degree of phenotypic integration in comparison to P. pundamilia.

Because P matrices are based on phenotypic data from wild populations, changes in P may occur due to the combined effects of the genetic and environmental covariance matrices, and their potential interactions (Lande, 1979; Falconer, 1989; Arnold & Phillips, 1999). Although many of the traits that we used to construct the trait-by-trait P matrices have been found to be rather heritable in a common garden experiment involving the same Pundamilia species as studied here (Magalhaes et al., 2009), it is possible that some part of the observed changes in the P matrix occurred due to phenotypic plasticity as a response to differences in the environment (Lande, 2009; Draghi & Whitlock, 2012; Wood & Brodie, 2015). Further laboratory experiments would thus be needed to estimate the actual differences in the genetic G matrix of our studied populations.

The differences in trait-by-trait P matrices that we found between the hybrid population at Luanso Island and the sister species at Makobe Island have implications for the multivariate phenotypes, which became apparent when projecting populations into the multivariate morphospace of the cichlid community at Makobe Island (Fig. 4). Both Pundamilia species from Kissenda and Python Island still showed significant phenotypic divergence and occupied similar parts of the morphospace as their counterparts from Makobe Island, despite some degree of gene flow between them. By contrast, individuals from the hybrid population at Luanso Island do not fall into the full morphological space of the two putative original species. Instead, its morphospace occupation resembled that of another species at Makobe Island, Mbipia mbipi. Interestingly, M. mbipi, the most abundant species in shallow waters of Makobe Island, is a shallow water omnivore that lives close to the rocks where it makes opportunistically use of gaps between the rocks and also feeds on the rock surfaces (Seehausen & Bouton, 1997). The Pundamilia hybrids at Luanso occupy quite exactly the same spatial habitat (Seehausen et al., 1997, 2008) and are also omnivorous. Lastly, at all three islands where the two Pundamilia species experience some degree of gene flow, some individuals lie outside the combined morphospace of the Makobe Island Pundamilia species. This may suggest that hybridization could lead to a relaxation of some genetic constraints and thus the expression of some novel phenotypes (Parsons et al., 2011; Renaud et al., 2012; Selz et al., 2014).

The quantitative genetic framework of G and its related P matrix have become a powerful tool to assess the genetic constraints to selection (Lande, 1979; Lande & Arnold, 1983; Schluter, 1996; Blows et al., 2015) and to study how selection, drift and gene flow may affect the evolution of a population (Chapuis et al., 2008; Bailey et al., 2013). Our findings are consistent with theoretical predictions (Chapuis et al., 2008; Bailey et al., 2013, Fig. 1), suggesting that gene flow and introgressive hybridization can alter the P matrix of wild populations where the respective effect may depend on the relative abundance of hybridizing species and the level of interspecific gene flow. Such changes in the P matrix may potentially redirect a population towards a novel part of the morphospace and shift the P matrix towards a novel distinct adaptive peak (Seehausen et al., 2014).