The power of hybridization between species to generate variation and fuel adaptation is poorly understood despite long-standing interest. There is, however, increasing evidence that hybridization often generates biodiversity, including via hybrid speciation. We tested the hypothesis of hybrid speciation in butterflies occupying extreme, high-altitude habitats in four mountain ranges in western North America with an explicit, probabilistic model, and genome-wide DNA sequence data. Using this approach, in concert with ecological experiments and observations and morphological data, we document three lineages of hybrid origin. These lineages have different genome admixture proportions and distinctive trait combinations that suggest unique and independent evolutionary histories.

Hybridization is often detrimental and maladaptive, such as when hybrid individuals suffer reduced fitness. Alternatively, hybridization can facilitate gene and trait exchange, and hybrid individuals sometimes experience greater fitness than parental forms, especially in novel, and often extreme, environments (Grant and Grant 1992; Rieseberg et al. 2003a; Arnold 2006; Mallet 2007; Fitzpatrick and Shaffer 2007). One of the most dramatic outcomes of hybridization is homoploid hybrid speciation (HHS), the formation of an isolated lineage of hybrid origin without a change in base chromosome number (Buerkle et al. 2000; Mallet 2007; Abbott et al. 2010). Cases of HHS constitute windows on the mechanisms of genomic evolution during speciation because the extant parental species provide a basis for comparison. Cases of HHS also offer excellent opportunities to examine divergence in the face of gene flow and to investigate the contribution of ecological processes to speciation.

The propensity for hybridization to generate genetic and evolutionary novelty, including hybrid species, and the fate of that novelty, is often determined largely by the ecological context in which hybridization occurs (Buerkle et al. 2000; Rieseberg et al. 2003b; Martin et al. 2006; Taylor et al. 2009). Although there has long been debate about the relative fitness of hybrid genotypes in parental and novel habitats, it is clear that the ecological context is of fundamental importance (Arnold and Martin 2010). For example, substantial ecological differences between the parent species’ habitat and the habitat of the hybrids are predicted to facilitate the formation of species of hybrid origin (Buerkle et al. 2000, 2003). Empirical data support this prediction: hybrid species are commonly ecologically differentiated and associated with novel or even extreme environments relative to the parent species (Rieseberg et al. 2003b; Gompert et al. 2006; Abbott et al. 2010; Kunte et al. 2011).

When considering the contribution of ecological processes to hybrid speciation, there is the additional possibility of repeated origins of hybrid species leading to multiple isolated lineages of hybrid origin (multiple origins) (Schwarzbach and Rieseberg 2002; Gross et al. 2007; Abbott et al. 2010). The evolution of multiple hybrid lineages from the same parental species, driven by ecological factors, has been investigated in several plant systems (Wang et al. 2001; Schwarzbach and Rieseberg 2002; Gross et al. 2003; Rieseberg et al. 2003b; Abbott et al. 2010). For example, in Helianthus deserticola, there is evidence of as many as three origins, presumably involving three admixture events (Gross et al. 2007). It is also possible that multiple disjunct populations of hybrid origin could become isolated following a single hybridization event. This might occur if segments of an admixed lineage became isolated in separate populations, perhaps associated with islands, prior to genome stabilization (Buerkle and Rieseberg 2008). These scenarios (single or multiple origins) present opportunities to study adaptation and differentiation among admixed populations. Hybrid populations can evolve either independently in differing selective environments, in concert with selection fixing adaptive alleles in each population, or in parallel in similar selective environments (Gross et al. 2007). Investigations of multiple hybrid populations provide a framework for understanding how populations or species maintain cohesion or undergo differentiation. Here, we investigate four lineages of butterflies, each putatively arising from hybridization between the same parental species, and each occupying high-altitude habitats in different mountain ranges in western North America.

Hybridization between the butterfly species Lycaeides melissa and L. anna (formerly L. idas anna) resulted in an isolated hybrid lineage in the Sierra Nevada of California and Nevada that was documented with multiple molecular marker systems and investigations of morphological and ecological variation (Gompert et al. 2006). Lycaeides melissa populations generally occur in Great Basin habitats on the east side of the Sierra Nevada and the species range extends to southern Canada and the central USA, whereas L. anna populations occupy wet meadows at mid-elevation on the west slope of these mountains and in isolated populations in the central Coast Range of California (Scott 1986; Nice and Shapiro 1999; Guppy and Shepard 2001; Nice et al. 2002, 2005; Gompert et al. 2006). The hybrid lineage in the Sierra Nevada occupies a novel, alpine (above treeline, oreal) habitat, utilizes an alpine-endemic host plant, and exhibits both intermediate and transgressive traits (Nice et al. 2002; Fordyce and Nice 2003; Gompert et al. 2006). Some of these traits, such as a unique absence of egg adhesion to the larval host plant and strong female oviposition preference for their natal host plant, appear to be adaptive in the hybrid lineage’s environment and might contribute to reproductive isolation (Nice et al. 2002; Fordyce and Nice 2003; Gompert et al. 2006). These observations suggest an important role for ecological factors in the evolution of the Sierran hybrid lineage, as predicted by theory (Buerkle et al. 2000). Additional investigations have detected possible evidence of genetic admixture and unique morphological and ecological variation in high-altitude populations of Lycaeides in three other nearby mountain ranges (Nice et al. 2005; Gompert et al. 2008a,b, 2010b) (Fig. 1, Table 1).

Figure 1.

Left: Map depicting the location of putative hybrid taxa (colored circles) in western North America. Each circle represents a sampled population. The approximate geographic ranges of Lycaeides anna (yellow) and L. melissa (blue) are shown, the area of range overlap is shown as intermediate. Right: L. anna (top), the hybrid lineage from the Sierra Nevada (middle), and L. melissa (bottom).

Table 1.  Summary of ecological and morphological differentiation among four putative hybrid lineages and the parental species, L. anna and L. melissa.
Lineage/character L. anna Sierra NevadaSiskiyou Mts.Warner Mts.White Mts. L. melissa
  1. (A) Larval host plant is the genus of host plant used by females of each lineage for oviposition in the wild (CCN, MLF, JAF, ZG, LKL, CAB, pers. obs.) and on which larvae feed. (B) Host fidelity was assessed by oviposition preference experiments (Nice et al. 2002; Gompert et al. 2006; Fordyce et al. 2011). “Low” indicates little to no significant female preference for their natal host (i.e., the host listed as “Larval host plant”); “moderate” indicates significant preference exhibited by the females, but a substantial proportion of ovipositions observed on alternative hosts; “high” host fidelity indicates results of experiments where females laid >80% of their total eggs on the natal host plant. (C) Egg morphology was examined using electron microscopy (Forister et al. 2008). (D) Egg adhesion refers to whether eggs are strongly adhered to the larval host plant when they are laid. These data are a summary of the results of experiments that measured the force required to dislodge eggs from plants (Fordyce and Nice 2003) and from direct observations of egg adhesion during oviposition preference experiments (Nice et al. 2002; Gompert et al. 2006; Fordyce et al. 2011) and direct field observations. (E) Male genitalic morphology was assessed through morphometric analyses (Nice et al. 2005; Lucas et al. 2008) and reanalyzed here. (F) Wing morphology was assessed through principle component analysis of wing pattern character areas (Fig. S2). Wing patterns and the form of the male genitalia were originally used to delineate the parental species (Nabokov 1949) and were classified here based on similarity to the parental species. For details on ecological and morphological variation, see Materials and Methods.

(A) Larval host plant Lupinus, Lotus Astragalus Lupinus Astragalus Astragalus Medicago
       and low-elevation
(B) Host fidelity Moderate High ? High High Low-moderate
(C) Egg morphology L. anna intermediate?similar to L. annasimilar to L. anna L. melissa
(D) Egg adhesion Yes No Yes Yes No Yes
(E) Male morphology L. anna Intermediate L. anna IntermediateIntermediate L. melissa
(F) Wing pattern L. anna L. melissa Reduced L. melissa L. melissa L. melissa

In the present study, we first used genome-wide DNA sequence variation to test the hypothesis of hybrid speciation in each of these lineages. In doing this, we use an explicit, model-based test of the hypothesis of HHS, which contrasts with previous studies (e.g., Wang et al. 2001; Gompert et al. 2006; Mavárez et al. 2006) that have used marker or phenotypic additivity to detect patterns that are consistent with HHS. We used this model-based approach to estimate historical, demographic parameters for each lineage, including initial genomic admixture proportions and time of the admixture (speciation) event. We also measured contemporary genetic differentiation among lineages and estimated contemporary admixture proportions in each lineage. We then quantified and summarized ecological and morphological variation (including reanalysis of data from previous studies) to address whether these lineages exhibit similar traits that are potentially adaptive in alpine habitats or, alternatively, exhibit measurable variation reflective of ecological differentiation and possibly local adaptation.

Our large, multilocus DNA sequence dataset demonstrates that three of the four putative hybrid lineages in Lycaeides have admixed genomes composed of alleles from L. anna and L. melissa. Tests of alternative models of bifurcating and reticulating (hybrid) speciation events provide support for the hypothesis of hybrid speciation in these three admixed lineages. Parameter estimates under the model of hybrid speciation in combination with ecological and morphological measurements suggest that the hybrid lineages differ in their genomic structure and are morphologically and ecologically differentiated. We discuss the possibility of independent evolution in each of the lineages that might be consistent with the hypothesis of local adaptation in each mountain range. This work contributes to our understanding of hybrid speciation and constitutes the first, explicit test of the hypothesis of HHS based on a modeling approach and using a large, multilocus dataset.



Sampling and molecular methods used to generate genome-wide sequence data, and analytical details for next-generation data analysis, are described in more detail elsewhere (Forister et al. 2011; Gompert et al. 2010a; Gompert and Buerkle 2011). Sequence data in this manuscript are for DNA samples from seven Lycaeides populations (Table S1, Fig. S1), each generated by pooling genomic DNA from 15 individuals. We used a restriction fragment-based procedure to amplify a subset of the genome for high-throughput sequencing (the approach we used is similar to the CRoPS method [van Orsouw et al. 2007]). We used the restriction enzymes EcoRI and MseI to create fragments to which unique 10 bp barcodes or multiplex identifier sequences (MIDs) were ligated along with sequencing adaptors. At the time of library preparation, the number of available MID sequences was limited, thus pooling individual DNA samples from populations was necessary because individual indexing was not possible. (Current methods include individual indexing; see Gompert et al. 2012 and Parchman et al. 2012.) The fragments were then amplified with PCR using iProof high-fidelity DNA polymerase (BioRad Laboratories Life Science, Hercules, CA) to reduce PCR error and primers complimentary to the adaptors. PCR products were separated on a 2% agarose gel and fragments between 400 and 550 bp were excised from the gel. Fragments were purified using the GENECLEAN Turbo DNA purification kit (MP Biomedicals, LLC, Santa Ana, CA). (For more details, see Gompert et al. 2010a.) This template was sequenced by SeqWright DNA Technology Services (Houston, TX) using the 454 GS XLR70 Titanium platform. Assembly of the resulting sequence reads was performed using the SeqMan Ngen sequence assembler v2.0.0 (DNASTAR) software.

All sites with two or more observed states were treated as variable. This approach is clearly overly simplistic (see Li et al. 2009), however, in the absence of systematic errors, this assumption is unlikely to bias results. We estimated genome-wide genetic differentiation between each pair of populations (measured as the proportion of genetic variation partitioned between populations, inline image [Excoffier et al. 1992]), using a hierarchical Bayesian model (Gompert et al. 2010a; Gompert and Buerkle 2011). This model assumes that the population-level haplotype counts for each locus are distributed as a multivariate Pólya distribution (i.e., a Dirichlet compound multinomial distribution) with a vector of parameters comprised of the population haplotype (allele) frequencies. The use of this distribution, while integrating over all possible values of allele frequencies, allows estimation of allele frequencies while also accounting for the uncertainty associated with sampling sequences from individuals and sampling individuals from populations. Locus-specific inline image values are assumed to be drawn from a normal distribution with a mean and variance equal to the genome-level inline image (see eqs. 2–4 in Gompert et al. 2010a). (Software to implement the hierarchical, Bayesian analysis of molecular variance, BAMOVA, as described by other studies [Gompert et al. 2010a; Gompert and Buerkle 2011] is available at http://www.uwyo.edu/buerkle/software/.) This analysis was based on 1570 sequence loci, each with greater than inline image coverage across all populations and at minimum inline image coverage per population per locus (see Results). We include low coverage loci, rather than restricting analyses to only loci with some arbitrarily higher coverage, to improve the efficiency and accuracy of inferences. Higher coverage loci clearly contribute relatively more information to estimates of population parameters, but there is still important information from loci with low coverage and discarding the low coverage loci reduces the overall sample size and, therefore, the accuracy and precision of these estimates (see Buerkle and Gompert 2012 for a full discussion of this issue).

We used an ordination method, nonmetric multidimensional scaling (NMDS), to visualize population differentiation. Unlike tree-based clustering methods (e.g., neighbor-joining), NMDS does not force a bifurcating relationship among populations. We employed the matrix of pairwise inline image values to conduct NMDS with two dimensions using the MASS package in R (Venables and Ripley 2002; R Development Core Team 2010).


We tested three competing hypotheses for the origin of each of the alpine Lycaeides populations and each alpine population was modeled independently. The hybrid speciation model (HS) posits the divergence of L. anna (P1) and L. melissa (P2) from a common ancestor (A) at some time in the past (inline image; Fig. 2a). Then, at a more recent time (inline image), L. anna and L. melissa contribute to a single admixture event in which some proportion (inline image) of the hybrid genome is derived from P1, and P2 contributes inline image. The hybrid lineage (PH, an alpine population) and the parental species are isolated following the admixture event. We define population mutation rates (inline image) for the ancestral population (inline image), parental population 1 (inline image), parental population 2 (inline image), and the hybrid lineage (inline image). Alternatively, the L. anna—alpine sister species model (AS) (Fig. 2b) assumes that, following the initial divergence between (P1) and (P2), population P1 experienced a bifurcating speciation event that gave rise to two extant lineages inline image (i.e., L. anna) and an alpine population (PH). This model replaces the admixture time (inline image) with the time of the more recent speciation event (inline image) and does not include the admixture proportion, but is otherwise identical to the HS model. Finally, the L. melissa—alpine sister species model (MS) (Fig. 2c) is a mirror image of the AS model with a speciation event at time inline image that gives rise to inline image (i.e., L. melissa) and the alpine population (PH).

Figure 2.

Graphical representation of (A) the hybrid speciation model, (B) the L. anna—alpine sister species model, and (C) the L. melissa—alpine sister species model. Parameters are described in the text.

As is generally true with all modeling approaches, the models examined here are a simplification of what is more likely a complex evolutionary process, and this modeling approach could be extended to incorporate more parameters. However, given limits on computational time and the low-coverage sequence data, these speciation models capture the basic dimensions of the hypotheses being tested. One specific simplification of the models examined here that requires comment is the exclusion of gene flow following the admixture event from the hybrid speciation model. Although gene exchange among lineages of hybrid origin, or between the hybrids and parental species, is a possibility, we do not consider cases with continuing gene flow to represent hybrid speciation. In other words, hybrid speciation is considered complete when the hybrid lineage is no longer exchanging genes with the parental taxa or any other incipient lineage (Buerkle and Rieseberg 2008). Thus, with respect to gene flow parameters, our simplified model comports with a rigorous definition of hybrid speciation (Buerkle and Rieseberg 2008). Another consideration is that continuing gene flow following admixture would simply have the effect of pushing the admixture time (inline image) forward in the model (i.e., more recent admixture). Furthermore, the effect of postadmixture gene flow would be to reduce the amount of differentiation observed between lineages. In this sense, the exclusion of postadmixture gene flow is a conservative approach because models accounting for such gene flow would presumably also be more likely to favor hybrid speciation compared to a model lacking this parameter if postadmixture gene flow had actually occurred. Consequently, although these models are relatively simple, we believe they adequately capture the dynamics of the hypotheses being tested.

We used an Approximate Bayesian Computation (ABC) approach (Beaumont et al. 2002; Beaumont 2008; Csillery et al. 2010) to estimate the posterior probability of each of these models for each alpine population. These estimates were based on sequence data from both of the putative parental species and each hybrid lineage. Each analysis included a subset of loci from the sequence data described in the previous section that had at least one sequence for each population in the analysis and were compatible with an infinite sites mutation model (660 loci for the Sierra Nevada, 490 loci for the Siskiyou Mts., 715 loci for the Warner Mts., or 691 loci for the White Mts.). We briefly describe the procedure used to estimate model posterior probabilities. We specified a prior probability for each of the models, inline image, and priors for the following parameters conditional on the model (i.e., only a subset of parameters are relevant for a given model): effective population size (Ne) for PA, Ne for P1 or inline image, Ne for P2 or inline image, Ne for PH, the time (number of generations in the past) of admixture tadmix, or bifurcating speciation tP1H or tP2H, the time of the ancestral speciation event tP12, the mean per locus mutation rate inline image, the standard deviation of the mutation rate inline image, and the admixture proportion inline image (Table S2). Note, we placed priors on individual parameters, rather than the composite parameters inline image and inline image used to describe the models. We generated 106 simulated datasets per alpine population by first sampling a model (HS, AS, or MS) and then sampling model parameters from the appropriate priors. We used the software ms (Hudson 2002) to simulate sequence data with the sampled model and parameters following an infinite sites mutation model. The data were simulated to match the sampling and coverage of the observed data. Specifically, we simulated a sample of 30 gene copies per locus per population (i.e., 15 individuals) and then re-sampled the simulated sequences with replacement to match the sequence coverage for the locus.

We calculated 26 summary statistics for the observed genetic data and each of the simulated datasets: the number of alleles at a locus in each population (mean and variance among loci), nucleotide diversity in each population (inline image; mean, variance, and skew among loci), the net number of nucleotide differences between each population pair (inline image; mean, variance, and skew among loci), the mY admixture coefficient estimator (Excoffier et al. 2005), and the admixture coefficient estimator given by equation (6) from Chakrabory et al. (1992). We scaled and rotated the summary statistics using a principal component analysis (PCA) using the prcomp function in R. This is similar to the approach of Bazin et al. (2010) and this orthogonal transformation of the summary statistics is performed to reduce their dimensionality. We then calculated the Euclidean distance between the transformed summary statistics for the observed data and each simulated dataset. We calculated the correlation between each of the transformed summary statistics and the model used to simulate the data (e.g., HS, AS, or MS) for the 50,000 (5%) simulated datasets with the smallest Euclidean distance. We used the two transformed summary statistics (inline image) most correlated with the speciation model to estimate model posterior probabilities. As proposed by Beaumont (2008), we treated the evolutionary model as a categorical variable and used a multinomial logit model to estimate inline image. We performed model parameter estimation using a weighted local regression (i.e., using the 10,000, 1%, simulations closest to inline image), which was conducted using the postpr function in the R package abc (Csillery et al. 2010). Bayes factors were used to compare among the three models of speciation.


We estimated model parameters for the HS evolutionary model to quantify initial admixture proportions and determine whether hybrid, alpine lineages varied in their initial admixture proportions, times of admixture, and speciation times. We used an ABC framework to estimate model parameters using a similar approach as we described in the previous section. However, for parameter estimation, all datasets were simulated under the HS evolutionary model, and we conducted 106 simulations per alpine population. As described previously, we scaled and rotated the summary statistics using a PCA, but for parameter estimation, we calculated the correlation between each rotated summary statistic and seven model parameters: inline image, inline image, inline image, inline image, inline imageinline image, and inline image. The rotated summary statistic most correlated with each parameter was used to estimate the posterior probability distribution for all model parameters, except for inline image, for which the two most closely correlated summary statistics were used. If the same rotated summary statistic was most correlated with multiple parameters, it was included a single time. We computed the Euclidean distance between the retained and rotated summary statistic for each simulated dataset and the observed data. We obtained posterior estimates of parameters using weighted local multivariate regression (Beaumont et al. 2002) using 1% of the simulated datasets with the lowest Euclidean distances. We logit-transformed inline image and log-transformed all other model parameters prior to posterior density estimation. We assessed differences in initial admixture proportion among the hybrid lineages by estimating the posterior probability distribution for the difference in initial admixture proportion between pairs of populations and examining the proportion of times that the difference in initial admixture proportion was greater than zero in the postburnin Markov chain Monte Carlo (MCMC) for a given pairwise comparison. These proportions provide an estimate of the probability that the initial admixture proportions are different in each pairwise comparison.


We evaluated the efficacy of the ABC approach we used to estimate speciation model posterior probabilities by applying the same procedure to 1000 of the simulated datasets, as the evolutionary model for each is known. We then calculated the mean posterior probability of each evolutionary model as a function of the true model.

We also assessed the precision and efficacy of the specific ABC approach we used to estimate demographic parameters by applying the same approach to 500 simulated datasets with known parameter values. We used root mean square error (RMSE) and relative RMSE (rRMSE; RMSE divided by the true parameter value) to measure the difference between each true parameter value and the estimated parameter value (median of the posterior probability distribution). We also quantified the proportion of the time that the true parameter value was contained within the 95% credible interval (CI) of the estimated parameter value (95% COV).


We used a simple Bayesian model to estimate contemporary admixture proportions. For each putative hybrid lineage, we assumed that allele frequencies are a mixture of allele frequencies in the parental species. Under this assumption, the contemporary admixture proportion, a, describes the proportion of hybrid genome contributed by L. anna (parental population 1). 1 a describes the contribution of L. melissa (parental population 2). Thus, we must estimate the allele frequencies in the two parental populations and a for each hybrid population. Following Gompert and Buerkle (Gompert et al. 2010a; Gompert and Buerkle 2011), we model the likelihood of the observed allele counts given the population frequencies using a multivariate Pólya (multinomial-Dirichlet) distribution, which incorporates uncertainty that arose by sampling individuals from a population and sampling sequences (reads) from individuals:


Here inline image is the gamma function, inline image is the number of gene copies (i.e., twice the number of diploid individuals) sampled from population inline image, pijk and xijk are the frequency and observed count of allele k in population j for locus i, and nij is the number of reads from population j for locus i. The likelihood of the data for the admixed population is computed likewise, but with pijk=a pi1k+ (1 −a) pi0k. We placed uniform priors on the allele frequencies and a. To avoid loss of information in ancestry associated with derived haplotypes in the current model framework, we restricted this analysis to a single variable site within each of the contigs (the first variable nucleotide was selected). We obtained the marginal posterior probability distribution for a and parental population allele frequencies using MCMC. We implemented the MCMC algorithm in C++ using the Gnu Scientific Library (Galassi et al. 2009). For each admixed population, we sampled the MCMC for 47,500 steps following a 2500-step burnin. We summarized posterior estimates of a based on the median and 95% CIs of the postburnin samples. As with intial admixture (inline image), we assessed differences in contemporary admixture proportion (a) among the hybrid lineages by estimating the posterior probability distribution for the difference in contemporary admixture proportion between pairs of populations and examining the proportion of times that the difference in initial admixture proportion was greater than zero in the postburnin MCMC for each pairwise comparison.


The larval hosts for each of the lineages has been determined from field observations of ovipositing Lycaeides females. At several localities, including the high-altitude sites in the Sierra Nevada, White Mts., and Warner Mts., potential alternative hosts are available. However, in such cases, females in the field consistently oviposit on a single host, referred to as the “natal host” (Nice and Shapiro 1999; Nice et al. 2002; Fordyce and Nice 2003). Nevertheless, the presence of potential alternative larval host plants at some Lycaeides sampling localities prompted a series of cafeteria style oviposition preference experiments (choice trials) to estimate host fidelity. Wild females were captured at sampling localities, transported to a greenhouse and placed in choice arenas that included fresh-cut stems of two to four potential host plants and including the natal host. The results of these experiments (Nice et al. 2002; Gompert et al. 2006; Fordyce et al. 2011) were reanalyzed to estimate preference for the natal host in each of the lineages using a hierarchical Bayesian model specifically designed for count data (Fordyce et al. 2011).


The external morphology of Lycaeides eggs was examined using scanning electron microscopy. Morphometric analyses were used to describe the variation among populations and lineages (Forister et al. 2006). The morphology of eggs from the Siskiyou Mts. populations has not yet been examined. We reanalyzed these data for six of the seven populations considered here. We contrasted the first principal component between lineages using analysis of variance (ANOVA). Post hoc tests were employed to test for variation among populations.


While conducting oviposition preference experiments with Lycaeides females (Nice et al. 2002; Gompert et al. 2006; Fordyce et al. 2011), we discovered an unusual trait in the hybrid lineages in the Sierra Nevada (Fordyce and Nice 2003). Females from this lineage do not firmly adhere their eggs to the natal host plant. As a consequence, in this population, eggs laid by females on the natal host plant readily fall off of the plant and come to rest at the base of the plant. We recorded observations of egg adhesion in the other lineages during the course of conducting host preference experiments (see above) and from several years of field observations.


Measurements of male genitalic morphology and wing pigment patterns were used originally by Nabokov (1943, 1949) to delineate the parental species L. anna and L. melissa. We used Nabokov’s measurements to examine variation in male morphology across North America, including the populations sampled here (Nice and Shapiro 1999; Nice et al. 2005; Lucas et al. 2008). We reanalyzed these data for the seven focal populations (314 males, Table S4) in this study. PCA and an ANOVA on PC1 scores were used to compare population means.


We measured eight wing pattern characters from 108 male and 61 female Lycaeides (Table S5). We removed the wings from each butterfly and photographed them using a digital camera. We focused on the underside of the hindwings because this is the surface of the wing that is visible when Lyaceides are at rest and is therefore important during courtship, and because this is what Nabokov used in his original descriptions of Lyaceides in North America (Nabokov 1943, 1949). Two features of the wing pigment patterns are obvious: black wing pattern spots on the interior of the wing, and aurorae, which are orange and iridescent elements along the wing margins (Fig. S2). We identified eight wing pattern elements of both the black spots and aurorae that were used in these analyses: M, Sc, M2, Cu1, a2, b2 (the orange portion) of a2, a6, and b2 of a6 (Fig. S2). We drew outlines around each wing and the eight characters using the IMAGEJ software to calculate areas of each element. We standardized each measurement by dividing it by total wing area. We used PCA to explore variation in the wing pattern characters. First, to examine and illustrate the patterns of variation among lineages within the sexes, separate PCAs were performed for females and males. We then used ANOVA on PC1 and PC2 scores to test for differences among lineages within the sexes. ANOVA was also used on the total dataset, with sex was included as a factor the analysis, to test whether the hybrid entities were morphologically distinct from each other and the parental species by contrasting the first principal component for wing pattern (which explained 68.9% of the variation) for all samples (i.e., females and males combined).



We assembled the 341,045 sequences into 15,262 contigs or loci. Files including sequences and quality scores have been deposited in the NCBI Short Read Archive (accession SRA010351). Coverage for individual loci varied substantially and we confined our analyses to loci with greater than inline image coverage. Population genetic analyses of these 1570 DNA sequence loci confirm the intermediate nature of the four lineages, consistent with the hypothesis that these high-altitude populations arose via hybridization. Ordination of pairwise genetic distances between all populations confirmed previous findings from other data (Gompert et al. 2008b) that the populations in the four mountain ranges are differentiated from L. anna and L. melissa, but have intermediate genomic compositions, relative to the putative parental species (Fig. 3, Table S3). The four hybrid lineages also appear to be genomically differentiated from each other, particularly in the second dimension of the ordination.

Figure 3.

Scatter plot depicting ordination-based distances from NMDS of molecular variation among populations.


We used an ABC approach to estimate the posterior probabilities of three speciation models (Fig. 2) for each putative hybrid lineage. The HS model had the highest posterior probability for all alpine populations, which ranged from 0.447 to 0.570 (Table 2). Similarly, the HS model was approximately one to four times as likely as the other models (Table 2). For the Lycaeides lineage from the Siskiyou Mts., our ability to discriminate among the three models was limited and the posterior probabilities for the HS and AS models were similar.

Table 2.  Evolutionary model posterior probabilities and Bayes factors. Probabilities are given for the hybrid speciation (HS), L. anna—alpine sister species (AS) and L. melissa—alpine sister species (MS) models for each of the four hybrid lineages. Bayes factor for the HS model (numerator) relative to each model (denominator).
PopulationPosterior probabilitiesBayes factors
Sierra Nevada0.5420.1470.3111.003.691.74
Siskiyou Mts. 0.447 0.442 0.111 1.00 1.01 4.02
Warner Mts.0.5700.1920.2381.002.972.40
White Mts. 0.526 0.339 0.135 1.00 1.55 3.99


We used ABC methods for parameter estimation under the HS model (Fig. 2a). The posterior probability distributions for each parameter are summarized in Table 3. Estimates of initial admixture proportions, inline image (the proportion of the admixed genome derived from L. anna), for the Sierra Nevada, Warner, and White mountains are consistent with the hypothesis of hybrid speciation. Each of these lineages has an estimate of initial admixture proportion with a 95% CI that does not include zero or one, indicating genomes consisting of elements of both of the putative parental species, L. anna and L. melissa (Fig. 4). The estimate of initial admixture for the population in the Siskiyou Mts. (Siskiyou Mts., inline image = 0.85, 95% CI 0.10−1.00) did not provide strong evidence of admixture and is consistent with either the hybrid speciation model or the bifurcating speciation model involving L. anna (see above). Among the three other lineages, there was variation in the admixture proportion (Sierra Nevada, inline image = 0.55, 95% CI 0.24−0.82; Warner Mts., inline image = 0.40, 95% CI 0.05−0.91; White Mts., inline image = 0.28, 95% CI 0.02−0.91). We made pairwise comparisons of estimates of initial admixture proportions by estimating the posterior probability of the difference in estimates and calculating the proportion of times that the difference was greater than zero in the postburnin MCMC for each pair of lineages. Although the CIs on these estimates were relatively wide, there were clear differences in the point estimates of initial admixture proportions (Fig. S3 and Table 4).

Table 3.  Median and 95% credible intervals for model parameters.
Population inline image inline image inline image inline image inline image inline image inline image
Sierra Nevada1.
  (0.20-3.14) (0.28-2.02) (0.21-3.51) (0.04-10.4) (0.06-2.37) (0.00-0.49) (0.24-0.82)
Siskiyou Mts.0.780.791.
  (0.09-2.57) (0.08-5.58) (0.08-9.13) (0.01-3.09) (0.00-0.26) (0.00-0.08) (0.10-1.00)
Warner Mts.1.630.911.991.
  (0.47-4.49) (0.12-4.36) (0.30-8.30) (0.12-6.55) (0.01-0.50) (0.00-0.30) (0.05-0.91)
White Mts.0.600.812.491.760.150.060.28
  (0.15-1.39) (0.23-2.74) (0.55-9.19) (0.22-9.28) (0.02-0.45) (0.00-0.36) (0.02-0.91)
Figure 4.

Posterior probability distribution for historical model parameters (black lines). Prior distributions (gray lines) are also shown. Inferences were based on 490–715 loci, as indicated below population labels.

Estimates of speciation time of the parental species (L. anna and L. melissa) were similar across the four analyses as expected (Fig. 4). The speciation event for the parental species appears to be a relatively recent divergence and, although we do not have a genome-level mutation rate estimate for Lycaeides and cannot translate our speciation time estimates into absolute time, this result comports with previous investigations with other marker systems (Nice et al. 2005; Gompert et al. 2006, 2008a). However, the recency of the divergence between L. anna and L. melissa obviously constrains the time to admixture, which are also short (i.e., recent events) for all four hybrid lineages (Fig. 4).


To evaluate the efficacy of the ABC approach for modeling speciation, we calculated the mean posterior probability of each evolutionary model as a function of the true model for 1000 of the simulated datasets. The results are given in Table S6. Our ability to discriminate among models was modest. However, the most common incorrect inference was classification of HS as either AS or MS, thus our results are likely conservative and might underestimate the probability of HS for the observed data. An inherent limitation is that as admixture proportions tend toward 0 or 1, the HS model converges on either the AS or MS model.

We also assessed our ability to use ABC methods for parameter inference. The results of this assessment are reported in Table S7. rRMSE was on the order of 0.25 for most parameters (i.e., the difference between the estimated and true parameter value was about 25% of the true value), but higher for inline image and only 0.11 for inline image. Moreover, the mean RMSE for inline image was only 0.05, indicating that on average the estimated value of inline image was within inline image0.05% of the true value. This is rather encouraging, as inline image is the parameter, we were most interested in because we consider this parameter to be most informative about the hypothesis of admixture.


Posterior estimates of a based on the median and 95% CIs of the postburnin samples are summarized in Table 5 and Figure 5. Examination of differences in estimates of contemporary admixture proportions between lineages (using the same approach used to examine pairwise differences in intial admixture proportions) demonstrate substantial differences in all but two of the pairwise comparisons between hybrid lineages (Sierra Nevada vs. Siskiyou Mts. and Sierra Nevada vs. Warner Mts.) (Fig. S4 and Table 4).

Table 5.  Estimates of contemporary admixture proportion (a) and 95% credible intervals for each hybrid lineage.
Population a 95% credible interval
Sierra Nevada0.672(0.592–0.737)
Siskiyou Mts. 0.732 (0.623–0.824)
Warner Mts.0.602(0.530–0.674)
White Mts. 0.513 (0.438–0.589)
Figure 5.

Posterior probability distributions for contemporary admixture proportions in each of the four hybrid lineages: White Mts., Warner Mts., Sierra Nevada, and Siskiyou Mts. For comparison, the posterior distributions for the parental taxa are included: L. anna and L. melissa.

Table 4.  Comparisons of estimates of admixture proportions among hybrid lineages. Values are the probability that the admixture proportion of lineage (population) in that row is greater than the admixture proportion of the column lineage for (A) initial admixture proportions (inline image) and (B) contemporary admixture proportions (inline image).
Lineage:Siskiyou Mts.Warner Mts.Sierra NevadaWhite Mts.
(A) Initial admixture
 Siskiyou Mts.0.000.830.790.87
 Warner Mts. 0.17 0.00 0.31 0.63
 Sierra Nevada0.210.690.000.79
 White Mts. 0.13 0.37 0.21 0.00
(B) Contemporary admixture
 Siskiyou Mts.0.000.970.850.99
 Warner Mts. 0.03 0.00 0.09 0.95
 Sierra Nevada0.150.910.000.99
 White Mts. 0.01 0.05 0.01 0.00


Female host preference was assessed in cafeteria-style choice experiments (Nice et al. 2002; Gompert et al. 2006; Fordyce et al. 2011). There is considerable variation among lineages in the level of female preference for their natal host (Fig. S5). Females from populations of the hybrid lineages had relatively strong preference for their natal hosts compared to populations of L. anna and L. melissa. This was especially true of the hybrid populations from the Sierra Nevada. Female oviposition preference, considered here as a population-level trait (see Fordyce et al. 2011), is coupled in Lycaeides with the potential for extrinsic or ecological reproductive isolation because mating also occurs on, or very near, the natal host plant (Nice and Shapiro 1999; Nice et al. 2002).


Morphological measurements of egg characteristics were made from scanning electron microscopy images (Forister et al. 2006). We contrasted the first principal component (which explained 47.8% of the variation for egg morphology) among lineages using ANOVA. Post hoc tests revealed variation between populations (Fig. S6) with the hybrid populations having egg morphology somewhat intermediate between the parental species, though more similar to L. anna compared to L. melissa.


Lack of egg adhesion was observed in two of the putative hybrid lineages: in the Sierra Nevada and White Mts. To the best of our knowledge, these are the only Lycaeides populations that exhibit this trait. Experiments used to quantify the strength of egg adhesion demonstrated that this is not a consequence of a plant trait. That is, eggs from the Sierra Nevada and White Mts. populations showed significantly less egg adhesion compared to other populations, regardless of which plant the egg was laid on (Fordyce and Nice 2003). We interpreted this to be adaptive, and potentially a mechanism that contributes to isolation of these populations, because the above-ground biomass of the host plants senesces near the end of the growing season and is subsequently removed by winter weather conditions. Overwintering eggs that fall from the host land near the site of new growth the following spring (Fordyce and Nice 2003).

Strong egg adhesion was observed in all other populations, including the populations of the parental species as well as the populations in the Siskiyou and Warner mountains. The egg adhesion in the Warner Mts. is somewhat puzzling given our adaptive scenario for the lack of adhesion observed in the Sierra Nevada and White Mts. because the high-altitude Warner Mts. populations also use an alpine Astragalus as a larval host. However, there may be important ecological differences between these populations or their respective host plants. Alternatively, these different character states may represent underlying differences in the standing variation created by the hybridization event(s) giving rise to these lineages.


Measurements of Nabokov’s (1943, 1949) male genitalic characters for 314 males from the sampled populations (Lucas et al. 2008) (Table S4) were reanalyzed using PCA and an ANOVA on PC1 scores (which explained 90.8% of the variation). Comparison of population means form the basis of the data reported in Table 1 and Figure S7. Males from the Sierra Nevada, White mountains, and Warner mountains showed various degrees of intermediacy. Males from the Siskiyou mountains were not different from L. anna.


Eight wing pigment pattern elements (Fig. S2) were measured from digital photos of wings of both males and females (169 total individuals) from the focal populations (Table S5). Female and male wing patterns exhibited similar patterns of variation across all lineages (Fig. S8). Most variation in wing pattern (80%) was explained by the first principal component. All eight measurements of wing pattern had positive loadings on PC1 suggesting that PC1 corresponded to wing pattern character size. This is consistent with previous analyses of wing pattern variation in Lycaeides (Fordyce et al. 2002; Gompert et al. 2010b). Wing pattern PC2 had positive loadings for black spots, but negative loadings for wing aurorae and accounted for 12% of wing pattern variation. Based on PC1, we detected significant differences in wing patterns between males and females (F= 46.6, P < 0.0001) and among populations (F = 253.4, P < 0.0001). All pairwise comparisons between hybrid entities were significant except between the White Mts. and the Sierra Nevada (Fig. S9). Based on PC2, we detected significant differences in wing pattern between the Warner Mts. and all other lineages (F = 23.1, P < 0.0001) and between males and females (F = 5.9, P < 0.016). Combined ANOVA on PC1 scores with population (i.e., lineage), sex, and the interaction as effects was performed to further test for significant among-lineage differences (Table S8). Post hoc tests revealed substantial variation among populations (Fig. S9), including a wide range of variation among the hybrid lineages compared to each other and the parental species.


The combination of molecular, ecological, and morphological data presents a picture of a complex mosaic of differentiated lineages in western North American Lycaeides. Data from 1570 DNA sequence loci confirm the admixed and intermediate genomic composition of lineages in the Sierra Nevada, Warner, and White mountain ranges. There is less evidence of admixture for the lineage in the Siskiyou mountains. Evaluation of three alternative speciation models using ABC methods supported the hybrid speciation model over bifurcating speciation models in three of the lineages. Again, the lineage in the Siskiyou mountains was the exception with the hybrid speciation (HS) model and the L. anna-alpine sister species (AS) model having nearly equal probabilities and this lineage does not appear to be of hybrid origin. The Siskiyou population is perhaps best described as a sister taxon to L. anna. Although the strength of evidence supporting the HS model in each of the other three lineages was modest, the molecular data are clearly consistent with the hypothesis of hybrid speciation. Despite the evidence of genomic admixture and support for hybrid speciation, these lineages are not homogenous and exhibit distinct ecological, morphological, and genetic characters.

The lineages are similar in that they all occupy high-altitude sites in the four mountain ranges, consistent with the hypothesis that hybrid speciation is facilitated by escape from the influence of the parental species by colonization of novel habitats (Buerkle et al. 2000; Gompert et al. 2006). Morphometrics of the external morphology of Lycaeides eggs showed that the hybrid lineages have intermediate egg morphology relative to L. anna and L. melissa, although the morphology of eggs from the Siskiyou mountains has not yet been examined (Fig. S6). However, beyond these similarities, there are unique combinations of traits that distinguish these lineages from each other and the putative parental species, with both intermediate and extreme trait values exhibited in the hybrids, relative to the parental species (Table 1). Three of the four lineages use a high altitude, endemic Astragalus as the larval host plant. The high-altitude lineages tend to have much higher fidelity to their natal larval host plant, as measured by female oviposition preference, than either of the parental species (Fig. S5). Two of the lineages (in the Sierra Nevada and White Mts.) exhibit an unusual lack of egg adhesion that causes eggs to fall off of the host plant following oviposition. This trait appears to be adaptive in the alpine habitat where the above-ground biomass of the host plant senesces at the end of the high-altitude growing season (Fordyce and Nice 2003). Populations in the Sierra Nevada, Warner, and White mountain ranges are also intermediate in terms of the form of the male genitalia (Lucas et al. 2008), whereas males from the Siskiyou mountains are not significantly different from L. anna male morphology (Fig. S7). Wing patterns for both males and females also vary among the four lineages and the parental species, but while the Sierra Nevada, Warner, and White lineages have wing patterns that tend to be intermediate, they are not significantly different from L. melissa. The Siskiyou lineage has unusual, reduced wing patterns that appear to be transgressive with respect to both of the parental species (Fig. S9). However, as discussed above, the lineage in the Siskiyou mountains might be a sister taxon of L. anna and their unusual wing patterns might not, therefore, be the product of admixture.

In accord with ecological and morphological differences, the three putatively hybrid lineages are distinct genetically. Ordination of pairwise inline image values illustrates the differentiation among populations in the Sierra Nevada, Warner, and White ranges (Fig. 3). Estimation of contemporary admixture proportions also revealed substantial differences in the contribution of each of the parental species to the genomes of the admixed lineages (Fig. 5, Table 4). These comparisons of the admixed lineages strongly suggest that there is little gene flow among them, either as a function of geographic isolation or strong independent selective regimes in each mountain range, or both. Thus, the contemporary patterns of ecological, morphological, and genetic differentiation among the admixed Lycaeides lineages are consistent with a history of independent evolution. There is little indication of cohesive or coordinated evolution among the lineages.

Estimates of initial admixture proportions (at the time of hybridization) also varied among the lineages (Figs. 4, S3; Table 4). Although estimates of these parameters clearly exhibit substantial uncertainty, the bulk of the posterior densities suggests that these populations differed in their initial admixture proportions. Estimates of initial admixture for three of the lineages show considerable genomic contribution from both parental species; however, the 95% CI for the initial admixture estimate for the Siskiyou Mts. lineage includes 1.0 (i.e., no contribution from L. melissa). If the Siskiyou Mts. lineage arose by hybridization, then the evidence suggests that the initial admixture event was asymmetrical with a majority of the genomic contribution derived from L. anna; however, as noted, the L. anna—alpine sister species model cannot be rejected for the Siskiyou lineage.

The variation among the estimates of initial admixture proportion might indicate that the lineages experienced independent evolutionary histories following a common hybridization event. That is, subsequent to hybridization, recombination, selection and drift acted upon the admixture of parental genomes independently resulting in the observed variation in estimates of admixture proportions. Alternatively, the variation in initial admixture might be evidence of repeated hybridization events leading to multiple hybrid lineages.

Distinguishing between these competing hypotheses is difficult. If the admixed lineages are the product of multiple hybridization events, one could expect variation in the estimates of the time to admixture. However, estimated admixture times were very similar and obviously constrained by the relatively recent divergence time for the parental species, L. anna and L. melissa (Fig. 4, Table 3). Thus, even with 1570 sequence loci, we lack the resolution to identify differences in the timing of multiple potential admixture events. However, there is clear evidence of independent evolutionary histories for each lineage. As noted above, pairwise inline image values and estimates of contemporary admixture proportions revealed substantial differences among the four lineages. Differences between initial and contemporary admixture proportions in each lineage (Figs. 4, 5, Tables 3, 4) also indicate that the hybrid lineages experienced further independent evolution following the initial admixture. Consequently, our analyses are consistent with the hypothesis of independent evolution in these lineages following hybridization, but the current data cannot distinguish between one hybrid origin for all admixed lineages, or multiple independent hybrid origins.

The combination of genome-level genetic analyses and quantification of ecological and morphological variation reveals a mosaic of differentiation among the four lineages of Lycaeides butterflies in high-altitude habitats of mountain ranges in western North America. There is little evidence for hybrid speciation in the Siskiyou Mts. However, in the other three ranges (in the Sierra Nevada, Warner, and White Mountains), our analyses support the hypothesis that these lineages constitute species of hybrid origin that have experienced independent evolutionary histories. Although the available data do not allow us to distinguish between the hypotheses of a single or multiple origins, these admixed lineages represent well-documented cases of hybrid speciation and provide the foundation for comparative analyses of the genetic architecture of hybrid species and an opportunity to study in detail the role of stochastic and selective factors in shaping the genomes of hybrid lineages.

Associate Editor: T. Streelman


We thank A. M. Shapiro, B. Fitzpatrick, and E. Baack for discussion and comments on an earlier version of this manuscript. We thank the associate editor and reviewers for constructive criticism. This research was funded by the National Science Foundation (DDIG-1011173 to ZG, NSF EPSCoR WySTEP summer fellowship to LKL, IOS-1021873 and DEB-1050355 to CCN, DEB-0614223 and DEB-1050947 to JAF, DEB-1020509 and DEB-1050726 to MLF, and DBI-0701757 and DEB-1050149 to CAB).