The evolutionary history of Antirrhinum suggests that ancestral phenotype combinations survived repeated hybridizations


For correspondence (fax +44 131 650 5392; e-mail


The model species Antirrhinum majus (the garden snapdragon) has over 20 close wild relatives that are morphologically diverse and adapted to different Mediterranean environments. Hybrids between Antirrhinum species have been used successfully to identify genes underlying their phenotypic differences, and to infer how selection acts on them. To better understand the genetic basis for this diversity, we have examined the evolutionary relationships between Antirrhinum species and how these relate to geography and patterns of phenotypic variation in the genus as a whole. Large population samples and both plastid and multilocus nuclear genotypes resolved the relationships between many species and provided some support for the traditional taxonomic division of the genus into morphological subsections. Morphometric analysis of plants grown in controlled conditions supported the phenotypic distinction of the two largest subsections, and the involvement of multiple underlying genes. Incongruence between nuclear and plastid genotypes further suggested that several species have arisen after hybridization between subsections, and that some species continue to hybridize. However, all potential hybrids appear to have retained a phenotype similar to one of their ancestors, suggesting that ancestral combinations of characters are maintained by selection at many different loci.


Much of our understanding of the genes involved in morphological evolution and speciation has come from taxa that are sufficiently different to be regarded as separate species, but that retain the ability to form fertile hybrids. Both natural and artificial hybrids have been used to detect loci underlying differences between the parental species. In some cases, the genes and mutations have themselves been identified. This has been particularly successful when the research infrastructure developed in a closely related model species is available, for example in Drosophila (McGregor et al., 2007; Jeong et al., 2008).

The garden snapdragon, Antirrhinum majus (Plantaginaceae) has been used as a model to study inheritance, and the genetic control of development and flower colour (Schwarz-Sommer et al., 2003). Its close relatives are native to the western Mediterranean region, mostly the Iberian peninsular, and comprise a monophyletic group – the traditional genus Antirrhinum– in which between 17 and 27 distinct species and subspecies have been recognized in different taxonomic accounts (Rothmaler, 1956; Webb, 1971; Sutton, 1988). Although morphologically diverse and adapted to different, often extreme, environments, all Antirrhinum species can form fertile hybrids with each other and with A. majus when artificially cross-pollinated. Such hybrids have identified genes underlying differences in morphology and flower colour between their parents (Hackbarth et al., 1942; Langlade et al., 2005; Schwinn et al., 2006; Feng et al., 2009). Natural Antirrhinum hybrids have also identified genes involved in flower colour variation, and have suggested how selection acts on them (Whibley et al., 2006). The genus Antirrhinum can therefore provide a model for understanding the genetic basis for patterns of phenotypic diversity and adaptation around the species level, which may be typical of many recently evolved Mediterranean taxa (Thompson, 2005). However, it has been difficult to relate the genetic differences between pairs of parental species to variation in the genus as a whole because the relationships between Antirrhinum species have not been resolved. For instance, it is currently not possible to infer the ancestral state of a character and whether similar phenotypes might have evolved multiple times within the genus.

One obstacle to resolving evolutionary relationships within Antirrhinum is reflected in the unclear taxonomy of its species, many of which have not been reported to have a unique, fixed character (a synapomorphy; Webb, 1971). Therefore, species might not represent discrete genetic entities because they have been delimited artificially. Nevertheless, support for the genetic distinction of some recognized species has been provided by allozymes and DNA sequence variation (Mateu-Andres and Segarra-Moragues, 2000, 2003; Jimenez et al., 2005a; Mateu-Andres and de Paco, 2005). Relationships above the species level are also unclear. The genus has been divided into three morphological subsections –Antirrhinum, Streptosepalum and Kickxiella– but no subsection has been defined by a synapomorphy, and all have been suggested to overlap in phenotype (Rothmaler, 1956; Webb, 1971). The subsections also correlate with ecology: most members of subsection Kickxiella are small prostrate alpines or xerophytes that grow on rock faces, whereas subsections Antirrhinum and Streptosepalum comprise larger, more upright plants that are able to grow in competition (Figure 1). This raises the possibility that each subsection represents an ecotype, and that its species have evolved similar characters independently as adaptations to a particular environment, rather than sharing characters through common descent.

Figure 1.

 The three morphological subsections of Antirrhinum.
A representative of each subsection is shown in situ and in cultivation. Subsection Antirrhinum is represented by A. pseudomajus, subsection Streptosepalum is represented by A. braun-blanquetii and subsection Kickxiella is represented by A. pulverulentum. Scale bars: 150 mm for cultivated plants, and the ruler shown with plants in the field is 35 mm wide.

Attempts to resolve a species phylogeny for Antirrhinum from DNA sequences have been unsuccessful. Relatively little sequence variation has been found in the genus, consistent with its recent origin, and the variation is not distributed consistently between taxa, so that different genes support different relationships between species (Jimenez et al., 2005b; Vargas et al., 2009). Sparse sampling of the taxa used for DNA sequence analysis might also have contributed to a lack of phylogenetic resolution, given that young species are likely to share many unfixed alleles that might not be represented in small taxon samples.

Here, we examine evolutionary relationships within the genus Antirrhinum by comparing populations sampled from across the geographic range of each species. Plastid and multilocus nuclear genotypes resolve the relationships between many species, and further suggest that the traditional morphological subsections largely correspond to separate evolutionary lineages. They also suggest that hybridization has occurred repeatedly between the two major ancestral lineages where they overlap in range. Morphometric analysis of plants grown in common garden conditions supported the phenotypic distinction of the two major subsections, and the involvement of many underlying genes. However, all putative hybrid species appeared to resemble one of their parents, suggesting that ancestral suites of phenotypes have survived hybridization as a result of selection at multiple loci.


Antirrhinum taxonomy reflects discontinuous phenotypic variation

Antirrhinum populations were sampled from across the geographic range of each recognized species and subspecies, so that the depth of sampling in each taxon broadly corresponded to its abundance (Figure 2; Tables 1 and S1). The only recognized species that remained un-sampled was Antirrhinum martenii, which could not be found at its original collection sites in the Moroccan Rif. For convenience, we treated all taxa at the rank of species, including those that are often regarded as subspecies of A. majus.

Figure 2.

 The distribution of recognized Antirrhinum species.
Species ranges were estimated from collection sites of this and previous studies. The range of A. tortuosum extends further eastwards than is shown here. The names of members of subsection Antirrhinum are shown in red, members of Kickxiella are shown in blue and members of Streptosepalum are shown in orange. Each species range is coloured to represent its geographic location, so that species with similar locations are shown in colours with similar hues. For A. tortuosum and A. linkianum, subpopulations that occur in distinct geographic regions are shown with different colours.

Table 1. Antirrhinum species names and abbreviations
SpeciesAbbreviationNumbers used in morphmetricsa
  1. aThe number of populations of each species used in morphometric analysis is shown, followed by the total number of individuals sampled.

  2. bThese taxa have also been considered subspecies of A. majus, e.g. A. majus ssp. striatum or A. majus ssp. majus.

Subsection Antirrhinum
 A. australe Rothm.au8/48
 A. barrelieri Boreauba3/30
 A. boissieri Rothm.bo1/12
 A. cirrhigerum Filcahobci4/33
 A. graniticum Rothm.gr7/33
 A. grosii Font Quergo 
 A. latifolium Millerla5/10
 A. linkianum Boiss. & Reuterbli6/25
 A. litigiosum Paublt7/29
 A. majus L.bma 
 A. pseudomajus Fernández-Casasbps3/15
 A. siculum Millarsi7/16
 A. striatum Fernández-Casasbst6/17
 A. tortuosum Vent.bto9/99
Subsection Kickxiella
 A. charidemi Langech1/4
 A. hispanicum Chav.hi3/16
 A. hispanicum (Moroccan accessions)Mh3/16
 A. lopesianum Rothm.lo1/3
 A. microphyllum Rothm.mi1/3
 A. molle Langemo6/47
 A. mollisimum Rothm.ms3/30
 A. pertegasii Rothm.pe1/6
 A. pulverulentum Lazaropu2/24
 A. rupestre Rothm.ru5/20
 A. sempervirens Lapeyr.se2/11
 A. subbaeticum Guemessu1/3
 A. valentinum Font Querva1/3
Subsection Streptosepalum
 A. braun-blanquetii Rothm.bb1/9
 A. meonanthum Hoffmans. & Linkme1/6

To assess phenotypic variation within Antirrhinum we grew plants from a representative subset of 98 populations together in a glasshouse, and recorded phenotypes for an average of 5.8 plants from each population (Table 2). We chose phenotypes that differed between populations, without regard to their differences between species or subsections, to avoid any bias towards characters that might automatically support traditional taxonomic divisions. Characters were then selected for further analysis on two genetic criteria. The first attempted to reduce the use of characters that were strongly influenced by environment. It assumed that members of the same species are genetically most similar to each other. Therefore, a comparison of the variation within species and between species gives an estimate of the extent to which a phenotype is genetically determined (approximating its broad-sense heritability). Phenotypes for which more than 35% of the total variance occurred within species (stomatal indices in leaves, stems or petals, and leaf epidermal cell size) were excluded from further analysis.

Table 2.   Phenotypes used in de-trended correspondence analysis (DCA)
  1. aPercentage of the total variance in each character that is contributed by differences between species, which approximates to the heritability (H2) of the character in the broad-sense.

  2. bCharacters were either measured on a continuous scale (C) or scored in discrete categories (D), in which case the number of character states is given (e.g. D, 2 refers to a binary character). Units are not given for continuous characters because they were subject to linear rescaling, which made them dimensionless.

PC181First PC from the allometry model of leaves and flowersC
PC283Second PC from the allometry model of leaves and flowersC
PC375Third PC from the allometry model of leaves and flowersC
PC473Fourth PC from the allometry model of leaves and flowersC
Anth91Intensity of red anthocyanin pigmentation in the corollaD, 5
Branch67Ratio of the length of the longest axillary branch to Ht (below)C
CarpL88Carpel length (from ovary base to stigma tip)C
El73Presence of anthocyanin outside the corolla face and dorsal petal sinusD, 2
FlNode67Number of the node at which flowers were first producedD, 29
Ht73Length of the stem from cotyledons to first flowerC
InfNode84Maximum number of nodes bearing open flowers at any one timeD
IntHairs88Density of stem trichomes between nodes 1 and 2C
LfHair90Density of adaxial epidermal trichomes in node-4 leavesC
PeCell65Density of adaxial epidermal cells in dorsal petalsC
PedL75Pedicel lengthC
StemW60Diameter of the stem midway between nodes 1 and 2C
StRatio73Ratio of ventral to dorsal stamen lengthsC
Sulf87Presence of widespread yellow aurone pigment in the corollaD, 2
Ve68Presence of darker anthocyanin pigmentation over corolla veinsD, 2
YelFace78Extent of yellow aurone pigmentation in the corolla faceD, 4
YelHair75Extent of yellow trichomes within the ventral corolla tubeD, 5
YelTube68Extent of yellow pigmentation within the ventral corolla tubeD, 4

The second criterion for screening phenotypes attempted to minimize repeated sampling of the same genetic differences. Differences in leaf and flower size and shape (allometry), for example, which have been used as taxonomic characters in Antirrhinum, do not vary independently of each other because they are affected by the same genes: i.e. the characters are developmentally constrained (Langlade et al., 2005; Feng et al., 2009). Instead of treating them separately, we therefore quantified variation in the allometries of leaves and flowers together using a computer shape model of flattened node-4 leaves, flattened dorsal petals and side views of intact flowers, in which other aspects of variation are apparent (e.g. the angles at which petal lobes are presented). The shape model captures co-variation in the positions of points placed around the organ outlines as orthogonal principal components (PCs; Figure 3). The first four PCs accounted for 93% of the variance within the genus, allowing the leaf and flower allometry of each plant to be described accurately with these four PC values. The types of variation captured by the PCs are shown in Figure 3(b). PC1 describes mainly size variation, showing that organ size is the major difference between plants, although this is correlated with other aspects of shape variation. For instance, an increase along PC1, which involves an increase in leaf and flower size, is accompanied by a narrower leaf shape and more reflexed dorsal petals. PC2 and PC3 capture other aspects of shape and size variation in both leaves and flowers, whereas PC4 mainly describes the extent to which flowers vary independently of leaves (Figure 3b). Because the PCs are uncorrelated to each other, each was treated as a separate phenotypic character.

Figure 3.

 Models describing organ shape and size variation in the genus Antirrhinum.
(a) Points were placed around node-4 leaves, flattened dorsal petals and images of intact corollas. Green points were placed manually and red points were spaced automatically between them. Variation in the positions of all the points within the genus were described by principal components (PCs). The effects of variation of ±2SDs in the first four PCs are shown relative to the mean organ outlines in black. Var shows the percentage of the variation in the genus that is captured by each PC.

We also avoided using other phenotypes that high correlations suggested might be developmentally constrained (e.g. the lengths of the style and stamens), and chose flower colour phenotypes to represent the effects of genes known to be involved in interspecies differences (Schwinn et al., 2006; Martin et al., 1987; see Experimental procedures for details).

For the 22 remaining characters, average values were calculated for each of the 98 populations, and the range of the mean values were adjusted to a common linear scale of 0–22. These rescaled values were used in de-trended correspondence analysis (DCA) to position each population in phenotypic space, defined by the two major axes of co-variation in the data set (Parnell and Waldren, 1996).

A notable feature of the distribution of populations in phenotypic space was that members of subsection Kickxiella formed a cluster distinct from subsections Antirrhinum and Streptosepalum, with Kickxiella populations mainly sharing higher values along DCA axis 1 (Figure 4a). Kickxiella therefore appears to be phenotypically distinct from the other two subsections. The only exception involved populations from Morocco that had previously been included in Antirrhinum hispanicum (subsection Kickxiella; Rothmaler, 1956). These mapped in the same phenotypic space as subsections Antirrhinum and Streptosepalum, whereas Spanish populations of A. hispanicum, from which the species was originally described, fell within the Kickxiella cluster. Because Moroccan populations could not be classified as A. hispanicum on grounds of morphology, they were subsequently treated as a separate taxon within subsection Antirrhinum.

Figure 4.

 Distribution of Antirrhinum species in phenotype space.
The two major axes of co-variation in 22 different phenotypic characters were identified by de-trended correspondence analysis. The mean values for 98 populations were plotted in the space define by these axes (a).Members of subsection Antirrhinum are shown as triangles, members of subsection Streptosepalum are shown as squares and members of subsection Kickxiella are shown as circles. Populations are coloured as in Figure 2, to reflect their geographic origins. The label mh denotes Moroccan accessions previously assigned to A. hispanicum. The contribution of each character to the phenotypic space is shown in (b).
Characters and their abbreviations are detailed in Table 2, and accession numbers for populations are presented in Figure S2.

Although DCA supported the phenotypic distinctiveness of Kickxiella, subsections Streptosepalum and Antirrhinum were not separated from each other. At the species level, populations of the same species tended to cluster together, although many overlapped. The overlap between species in a space involving 22 phenotypes does not rule out the possibility that individual species might be defined by fewer phenotypes or by characters that were not considered here.

The DCA allows co-variation between characters to be represented in the same phenotype space as the populations (Figure 4b), illustrating their effects on the separation of populations. The spread of characters along axis 1 suggested that the phenotypic distinctiveness of subsection Kickxiella, which has higher axis-1 values, does not depend solely on one type of character, for example size or flower colour. This was further supported by the finding that subsection Kickxiella differed significantly from the other two subsections for all phenotypes, except PC2, PC4 and Ve (P ≤ 0.05 in Student’s t-tests). Therefore, subsection Kickxiella can be defined by a combination of phenotypes that are likely to reflect variation in a number of different genes.

In contrast, all subsections overlapped along DCA axis 2, which also reflects variation in a number of different characters, although members of the same species tended to cluster with each other along this axis.

Plastid haplotypes support distinct evolutionary lineages

One explanation for the partial overlap of Antirrhinum species in phenotypic space is that the genus behaves as a single interbreeding population, in which the genetic differences between individuals reflect their geographic separation (isolation by distance). In this case, the similar phenotypes of Kickxiella species should either reflect their closer genetic relatedness and geographic proximity to each other, or may be independent of both relatedness and geography if similar phenotypes have evolved independently. We took this scenario of a single population as the null hypothesis against which to test alternatives, including the possibility that members of subsection Kickxiella share similar phenotypes because they descend from a separate evolutionary lineage.

We first examined the relatedness of individuals from their plastid haplotypes. Because previous studies had detected relatively little sequence variation in Antirrhinum plastids (Jimenez et al., 2005b), we compared 19 non-coding loci across 10 Antirrhinum species, and identified the three most variable (trnD-trnT, trn-S-trnR and trnS-trnFM; Shaw et al., 2005, 2007). Their sequences were then obtained from 90 populations that represented the geographic range of each species, and were found to contain 54 polymorphisms arranged into 34 different haplotypes. The relationships between haplotypes were inferred by parsimony. To allow for the resolution of a purely branching haplotype network (Figure 5), 12 single-nucleotide polymorphisms that were present in rare haplotypes and not fixed in any species were assumed to be homoplasious to mutations supporting deeper branches. Sequences from Misopates and Chaenorrhinum, each of which has been suggested to be sister to Antirrhinum (Ghebrehiwet et al., 2000; Oyama and Baum, 2004; Vargas et al., 2004), both rooted the Antirrhinum haplotype network in the same position.

Figure 5.

 Relationships between plastid haplotypes. Sampled haplotypes are shown in boxes containing the species names and accession numbers of the plants carrying them.
The most parsimonious relationships between haplotypes are shown by lines connecting boxes with cross lines that represent the number of mutations by which haplotypes differ. Diagonal cross lines show mutations that were assumed to be homoplasious with earlier mutations. Branches in which mutations introduced an MseI or TfiI restriction site are also shown. The tree is rooted with sequences from Misopates orontium, which differs from the in-group by 19 mutations.

The network contained four main clades (Figure 5). Clade I was restricted to Antirrhinum siculum. Clade II occurred mainly in subsections Antirrhinum and Streptosepalum, and Clades III and IV occurred mainly in subsection Kickxiella. This distribution of haplotypes cannot be explained solely by geographic separation, and therefore is inconsistent with the null hypothesis that the genus behaves as a single, unstructured population. For instance, members of subsection Antirrhinum carrying a clade-II haplotype occur throughout the geographic range of the genus, and overlap with Kickxiella species carrying clade-IV haplotypes. A more plausible explanation is that the distribution of plastid haplotypes reflects the evolutionary relationships between species: i.e. clade-III and -IV haplotypes were present in the Kickxiella ancestor; clade-II haplotypes were present in the separate lineage leading to subsections Antirrhinum and Streptosepalum; and clade-I haplotypes were present in the lineage represented by A. siculum only. However, several taxa have haplotypes that are incongruent with their phenotypes, in one of two respects. First, A. latifolium (subsection Antirrhinum) has only clade-IV Kickxiella haplotypes, and A. sempervirens, A. lopesianum, A. pertegasii and A. molle in subsection Kickxiella appear fixed for clade-II haplotypes usually found in subsections Antirrhinum and Streptosepalum. Secondly, A. tortuosum, A. barrelieri and A. australe from subsection Antirrhinum, and A. pulverulentum from subsection Kickxiella, carry both clade-II and clade-IV haplotypes.

The depth of taxon sampling was increased by genotyping another 273 individuals from an additional 156 populations (Table S1). Clade-IV haplotypes were identified by the presence of an MseI site in trnDtrnT, and clade-II haplotypes were identified by a TfiI site in trnStrnR. Individuals lacking both these sites were assumed to carry clade-III haplotypes, unless they were accessions of A. siculum (clade I). The additional samples confirmed the distribution of haplotypes identified by DNA sequencing, and also revealed that populations of A. litigiosum (subsection Antirrhinum) from the Valencia region carried clade-IV (Kickxiella) haplotypes, whereas those from central Spain carried clade-II haplotypes only.

One explanation for the cases of incongruence between plastid haplotype and morphology is that ancestral polymorphisms have persisted in the lineages, leading to different subsections (lineage sorting). However, this cannot easily account for the presence of highly diverged haplotypes within the same species. For example, the clade-II and clade-IV haplotypes found in A. tortuosum (subsection Antirrhinum) differ by up to 21 mutations, and those in Spanish A. hispanicum (subsection Kickxiella) differ by up to 17 mutations, although only 24 mutations distinguish the most dissimilar haplotypes within the genus as a whole (Figure 5). Such cases of incongruence are more consistent with transfer of haplotypes by hybridization between lineages. Further evidence for hybridization is provided by the location of populations with incongruent haplotypes. They occur where Kickxiella species overlap with members of the other subsections: for example, with subsection Antirrhinum in the Sierra Nevada of south-east Spain (Figure 6). In contrast, haplotypes are congruent with morphology where subsections grow in isolation (e.g. subsection Antirrhinum in Morocco, western Portugal and northern Spain, and Kickxiella in central Spain). The only notable exception is that of A. latifolium, which carries a clade-IV haplotype, is restricted to the Alps and does not currently co-occur with a Kickxiella species, although this does not rule out the possibility of hybridization with a previously sympatric Kickxiella.

Figure 6.

 Geographic distribution of plastid haplotypes.
Members of subsection Antirrhinum are shown as circles, members of subsection Kickxiella are shown as triangles and members of subsection Streptosepalum are shown as squares. Filled shapes represent haplotypes that are congruent with morphology; unfilled shapes show incongruent haplotypes.

Nuclear genotypes support species delimitations and evolutionary lineages

Because maternally inherited plastid genomes provide only limited information about the relationships between individuals, particularly for hybrids, we also examined variation in nuclear genes. Previous studies had found relatively little nuclear sequence variation and inconsistent distribution between Antirrhinum species (Gübitz et al., 2003; Vargas et al., 2009). We therefore sampled multiple nuclear loci as amplified fragment length polymorphisms (AFLPs) from the set of 293 accessions that had been used for plastid haplotype analysis. Removal of potentially homoplasious markers left 381 informative AFLPs.

The AFLPs were used to infer the population structure within Antirrhinum using a Bayesian model, structure, which assigns individuals to a specified number of populations by minimizing genetic disequilibria within each population (Falush et al., 2003). This method is not dependent on prior taxonomic assumptions, and is able to represent hybrids as genetic admixtures of other populations.

When structure was used to assign each accession to one of two populations (K = 2 in Figure 7), one population comprised subsection Streptosepalum and most of subsection Kickxiella, and the second consisted of subsection Antirrhinum and the Kickxiella species from south-east Spain (A. hispanicum, A. mollissimum, A. rupestre and A. charidemi). This division therefore supported the genetic distinction of Streptosepalum and northern Kickxiella species from subsection Antirrhinum and southern Kickxiella. A. latifolium and A. molle, which showed incongruence of plastid haplotypes and morphology consistent with hybrid origins, were suggested to have admixed nuclear genomes, as was A. siculum.

Figure 7.

 Assignment of individuals to populations.
Plants were assigned to different ancestral populations, each shown in a different colour, by their nuclear genotypes. Representative assignments are shown for models in which the maximum number of populations (K) was set to either two or 17. Black bars denote members of subsection Antirrhinum, dark-grey bars denote members of subsection Streptosepalum and light-grey bars denote members of subsection Kickxiella.

A maximum of 14 different populations could be identified by structure (Figures 7 and S3). Only two of these –A. graniticum and A. latifolium– consisted almost entirely of unique haplotypes, suggesting a high level of coalescence within these species. The two Streptosepalums remained within the same population as A. grosii and A. lopesianum from subsection Kickxiella. The central Spanish Kickxiella species, which are all local endemics and do not overlap in range, were assigned to another population, which we subsequently refer to as core Kickxiella. Many of the remaining species fell into geographic groups with similar genetic compositions. In south-east Spain, for example, A. mollissimum was designated an admixture of genomes from neighbouring A. charidemi and A. rupestre, and the A. rupestre genome was also found admixed in A. hispanicum and A. barrelieri nearby.

Cultivars of A. majus were assigned to the same population as A. pseudomajus and A. striatum from around the eastern Pyrenees, supporting the domestication of A. majus from this population.

The ability to progressively subdivide the genus into populations with boundaries that corresponded to classical species delimitations provided support for traditional taxonomy, and suggested that AFLP genotypes might resolve relationships between species, even though extensive admixture was inferred for all but A. graniticum and A. latifolium. Therefore, genetic distances were calculated between all pairwise combinations of individuals, and were then used in neighbour-joining analysis with Misopates orontium as the out-group. When all individuals were included in the analysis, members of the same species usually clustered with each other, as they did in the structure analysis, but the relationships between species were poorly resolved (Figure S4). One explanation for poor resolution is that some of the taxa have hybrid origins, and so carry alleles from multiple lineages. These shared alleles could reduce resolution both within and between the parental lineages. We therefore excluded potential hybrids – i.e. plants with incongruent nuclear and plastid genotypes, or apparently admixed nuclear genomes – if they reduced the overall support. The remaining accessions were resolved into three major clades (Figure 8). One comprised the core Kickxiella, within which A. pulverulentum formed a supported group with its central Spanish neighbour A. microphyllum, and A. pertegasii grouped with the geographically distant A. subbaeticum. The second major clade consisted of species from subsection Antirrhinum, with A. siculum at its base and A. graniticum sister to the remaining taxa, which included two geographic groupings: A. pseudomajus, A. striatum and A. litigiosum from around the eastern Pyrenees, and A. linkianum and A. cirrhigerum from the west coast of Iberia. All accessions of Moroccan ‘A. hispanicum’ clustered together within subsection Antirrhinum, supporting the phenotypic evidence for their membership of this subsection. The third major clade consisted of A. meonanthum and A. braun-blanquetii (subsection Streptosepalum) from north-west Iberia, together with two of their neighbouring Kickxiella species, A. grosii and A. lopesianum.

Figure 8.

 A neighbour-joining tree of Antirrhinum species.
The tree was made from pairwise nuclear genetic distances between inferred non-hybrid accessions. Support values are shown for nodes that were recovered in more than 50% of 1000 bootstrap replicates.


By comparing the genotypes of Antirrhinum populations at multiple nuclear loci we have identified three well-supported clades of species. The two largest clades do not correspond to individuals from the same geographic locations, although they contain geographic subgroups within them, and so cannot be explained by a genetic structure that relates solely to geographic isolation. The clades are therefore likely to represent different evolutionary lineages.

One clade consists of species from subsection Kickxiella, which was originally defined by the shared phenotypes of its members. Morphometric analysis under common garden conditions confirmed the phenotypic distinctiveness of this subsection, with common characters that include small stature, small, ovate and hairy leaves, and small pale flowers with reflexed petals. The ancestral Kickxiella is therefore likely to have shared these characters, and, like its descendants, to have lived on rock faces. The second major clade comprised mainly members of subsection Antirrhinum. These are dark-pink- or yellow-flowered species that form large, upright plants, with larger, more elongated leaves, and are able to grow in competition. The third major clade contains both of the species from subsection Streptosepalum and two species with Kickxiella phenotypes. The classical division of Antirrhinum into three morphological subsections therefore appears partly natural.

The overlap of the three subsections along DCA axis 2 implies that all subsections vary to a similar extent for the different morphological and flower-colour characters that contribute to this axis. This is consistent with ancestral variation that has not been fixed during divergence of subsections, although fixation may have occurred during speciation because members of the same species tend to cluster along axis 2.

Distinct plastid lineages were also identified, and each was found predominantly in subsections Antirrhinum and Streptosepalum, or in Kickxiella. Many species carry congruent nuclear and plastid genomes, for instance A. microphyllum, A. subbaeticum and A. valentinum have genotypes expected of direct descendants of the core Kickxiella lineage, whereas A. siculum, A. graniticum, A. pseudomajus, A. striatum, A. cirrhigerum and A. linkianum have both plastid and nuclear genomes representative of subsection Antirrhinum.

However, several species appear to have nuclear and plastid genomes from different lineages. These include A. sempervirens, A. pertegasii and A. pulverulentum, which have subsection Antirrhinum plastids but core Kickxiella nuclei. Such cytonuclear incongruence can be explained if hybridization between subsections was followed by introgression involving Kickxiella, leading to the capture of an Antirrhinum plastid by a pre-dominantly Kickxiella nuclear genome. Similarly, A. tortuosum could also have captured its diverse incongruent plastids from multiple Kickxiella donors.

The remaining species share their nuclear genotypes with both Antirrhinum and Kickxiella lineages, and consequently cannot be assigned to one of the three major clades. However, they share plastids and nuclear genes with their geographic neighbours, suggesting that they are the result of hybridizations between subsections Antirrhinum and Kickxiella without significant introgression. This is particularly apparent in the species from south-east Spain, which have similar nuclear genotypes and carry both Kickxiella and Antirrhinum plastids. Hybridization could also have given rise to the nuclear clade consisting of the two Streptosepalum species together with A. grosii and A. lopesianum from subsection Kickxiella, because these species are found together in north-west Iberia and carry diverged plastids. Alternatively, the clade might represent a third ancestral lineage in which some members have captured incongruent plastids from subsection Kickxiella.

It is not possible to resolve the order in which the three major lineages diverged from each other. However, the distribution of the core Kickxiella species between the mountains of central Spain suggests that this lineage was once more widespread, and became fragmented on contraction to its current mountain refugia. This fragmentation could have contributed to the formation of distinct species. A similar range contraction has been proposed to account for the genetic relationships in several alpine species that are currently restricted to mountains around the western Mediterranean (Kropf et al., 2003; Dixon et al., 2007). Contraction to alpine refugia can also explain why all high mountain ranges in Iberia and southern France either have their own endemic core Kickxiella species or, in the case of the Alps (A. latifolium), the Pyrenees (A. molle and A. sempervirens) and the Baetic Cordillera (the south-east Spanish species), retain vestiges of the Kickxiella lineage in their plastid and nuclear genotypes. Populations from the Rif Mountains of Morocco have been recognized as a species within subsection Kickxiella (A. martenii; Rothmaler, 1956), suggesting that Kickxiella might once have extended into North Africa. As Kickxiella species are found on dry rock faces, mostly at higher elevations, contraction in their ranges could have occurred during periods of increasing rainfall and warming. Several major events of this kind have occurred in the Mediterranean region since the end of the last ice age (Fletcher et al., 2010). The same environmental changes that reduced the range of Kickxiella could have allowed subsection Antirrhinum to spread through lowland regions, perhaps from coastal refugia, and therefore to hybridize with Kickxiella in regions of contact.

Character evolution

The genetic relationships between Antirrhinum species suggest that several hybridization events occurred between subsections Antirrhinum and Kickxiella, yet the phenotypes of the two subsections remain distinct from each other. This is most apparent in two species that are sympatric in the Sierra Nevada of south-east Spain: A. rupestre (Kickxiella phenotype) and A. barrelieri (Antirrhinum phenotype), which have indistinguishable AFLP and plastid genotypes, consistent with their continuing hybridization. Therefore, although hybridization has the potential to create new adaptive combinations of phenotypes (e.g., Rieseberg et al., 2003), evolution in Antirrhinum appears to be constrained.

Two factors suggest that the evolutionary constraints within Antirrhinum are not developmental. Firstly, the phenotypic characters that distinguish subsections Kickxiella and Antirrhinum were chosen to represent variation in multiple genes with independent effects. Secondly, hybrids between A. majus and A. charidemi have shown that multiple genes underlie the differences between their Antirrhinum and Kickxiella phenotypes (Langlade et al., 2005; Feng et al., 2009). For instance, 10 genes with additive effects were found to affect leaf size, so that F1 hybrids and almost all F2 progeny had leaf phenotypes that were intermediate between the two parents (i.e. they occupied the gap in phenotypic space between subsections Antirrhinum and Kickxiella).

A more likely explanation is therefore that Antirrhinum phenotypes are constrained by selection. For some characters this might reflect contrasting adaptations to life on bare rock faces or in competition: for instance, small organs and dense hairs might be advantageous in limiting water loss on dry rock faces, but a disadvantage in competing with other species (Parkhurst and Loucks, 1972; Ehleringer et al., 1976). Hybrids with intermediate phenotypes could therefore be maladapted in both parental habitats. Because the differences between Antirrhinum and Kickxiella phenotypes appear to involve multiple unlinked genes, selection would have to act co-ordinately on many loci.

The proposed evolutionary history of Antirrhinum, which involves hybridization between emerging lineages, influenced by environmentally induced range changes, might be typical of the early stages of species formation in many taxa. It suggests that adaptation can exert a major constraint on diversity, even when environments change rapidly and novel genotypes are produced frequently in hybrids.

One testable prediction of our hypothesis that ancestral combinations of genes are reselected following hybridization in Antirrhinum is that similar phenotypes will have a similar genetic basis, even in species that share little of their nuclear genomes. The genus Antirrhinum, and particularly its recent hybrid species, might therefore allow adaptive genes to be identified through genotype–phenotype association.

Experimental procedures

Taxon sampling

The populations used in this study are detailed in Table S1. Populations of most species were sampled in 2006–2007. Sampling sites included locations used in previous taxonomic accounts to aid identification. Seeds of these accessions are available on request. Other accessions were kindly provided by Isabel Mateu-Andrés, Christophe Thébaud, Thomas Gübitz, Enrico Coen and their colleagues.

Phenotype analysis

Plants for phenotype analysis were grown from field-collected seeds or from seeds produced by intercrossing members of the same population. Germination was synchronized in March 2007 by imbibing seeds in 10 μm gibberellin (GA3), and plants were grown in natural light in a glasshouse. Phenotypes were recorded as the fifth flower opened. Phenotypes and their abbreviations are listed in Table 2. Cell and hair densities were measured from impressions made in cyanoacrylate glue on microscope slides by counting the number of cells or hairs within a 2 mm2 area. Other phenotypes were measured directly from plants or from digital images. Flower images were used to score the presence of yellow aurone pigments outside the corolla face, which reflects variation in the sulfurea gene (sulf; Whibley et al., 2006), darker anthocyanin in petal cells overlying major veins (Ve, conditioned by the Venosa locus; Schwinn et al., 2006), and restriction of anthocyanin pigmentation to the corolla face and between dorsal petal lobes, which is characteristic of mutations in the Eluta gene (El; Martin et al., 1987). Although these three genes show epistatic interactions, they have independent effects that allow all combinations of genotypes to be inferred (Martin et al., 1987).

Shape models were initially made for individual organs (node-4 leaves, flattened dorsal petals or intact flowers) using the method of Langlade et al. (2005), in which points were positioned around the outlines of each organ. Matrices of point co-ordinates were then translated and rotated to set their centroids at the origin, and to minimize variance in the positions of points (a Procrustes alignment without scaling). Principal component analysis (PCA) was used to partition the variance between plants into orthogonal PCs. For each type of organ, PC1 identified size as the major source of variation between plants (Figure S1). To minimize re-sampling this size variation (which could have a common genetic basis in the different organs), we therefore made a single shape model for all three organs together. The leaf matrices were first rescaled so that the total variance between plants in the leaf data set was the same as for petals and flowers together. The three point matrices (leaf, petal and flower) representing each individual were combined and subject to PCA.

Plastid haplotype analysis

To compare plastid haplotypes, 19 non-coding loci (Shaw et al., 2005, 2007) were amplified from 10 Antirrhinum species, and sequences of the three loci with the highest proportion of nucleotide substitutions relative to indels were obtained from all populations. Further accessions were genotyped by incubating the product from trnDtrnT with MseI and from trnStrnR with TfiI. Where lack of cleavage suggested a novel haploptype for the species, PCR products were sequenced to confirm the absence of the restriction site.

Amplified fragment length polymorphisms (AFLPs) were produced from DNA digested with PstI and MseI, and were amplified with four primer combinations (P11–M49, P11–M41, P12–M37 and P14–M35, in Keygene nomenclature). AFLPs that did not represent genotypes consistently were identified by carrying out 10 AFLP reactions on each of four independent DNA extractions from the same set of genetically diverse plants. Fragments that were not detected in all replicates of the same plant were removed from the larger data set. Significant negative correlations between AFLP band size and frequency were detected for two primer combinations, suggesting size homoplasy among smaller bands. Fragments smaller than 100 nucleotides were therefore excluded from the analysis.

Individuals were clustered into populations using structure (Falush et al., 2003, 2007). Priors and parameters relating to correlated allele frequencies and the extent of population subdivision were found to have only minimal effects on population assignments and likelihoods of models fitting the data (Jakobsson and Rosenberg, 2007). At least eight simulations were carried out for each value of K (number of populations), without varying other parameters. Each simulation comprised 20 000 burn-in and 120 000 experimental replications. Results were visualized with distruct (Rosenberg, 2004).

Pairwise genetic distances (Jaccard) between individuals were calculated from AFLP data in past (Hammer et al., 2004), and neighbour-joining analyses of distance matrices in phylip (Felsenstein, 1989).


We are grateful to Maureen Erasmus, Kim Coulson, Mary Coulson, Niall Wilson, Monique Burrus and Christophe Thébaud for their considerable help with fieldwork. This work was supported by BBSRC, through grant BB/D552089/1, and a postgraduate studentship to YW, and by a Small Project Grant from the University of Edinburgh Fund.

Accession numbers of sequence data: FR690148-FR160233, FR690277-FR690458 and FR690868-FR690870.