The floral morphospace – a modern comparative approach to study angiosperm evolution

What morphospaces can and cannot do

The structure of a morphospace is determined by the choice of the characters to include and by the way these characters are coded. This selection of morphological properties, which is based on the research question, the systematic group and the material available for study, can affect the assessment of the occupied vs empty parts of the morphospace (Maclaurin, 2003; Huttegger & Mitteroecker, 2011). The actual coding of characters further influences the way morphological (dis)similarity is quantified (the metric imposed on the space) and what the possible transitions between character states are (Mitteroecker & Huttegger, 2009). The choice of a particular type of variable and metric that properly reflect the studied developmental, functional or evolutionary properties of the organisms is often not trivial. For example, in flowers, the evolution toward an increase in petal (or other organ) number may be achieved by the addition of one petal (e.g. from 4 to 5 petals) in a clade or by its multiplication (e.g. from 3 to 6) in another clade. The numbers originate from counts, but these types of characters might rather be treated as nominal (categories) or ordinal data (ranked categories), depending on the evolutionary model employed.

For characters that can be represented by a real number with an interval or ratio scale, a morphospace can be constructed by taking the measurements or parameters as Cartesian coordinates of an Euclidean space (Fig. 1). This is the common approach in morphometrics and for many theoretical morphospaces (Herrera, 1993; Medel et al., 2003; Shipunov & Bateman, 2005; Yoshioka et al., 2006; Gómez et al., 2009; van der Niet et al., 2010). In such a morphospace, the distance between points can be interpreted as morphological dissimilarity. Several points on a straight line in morphospace (e.g. a developmental or evolutionary trajectory) differ in the same morphological pattern, and the angle between two such linear directions can be interpreted as the magnitude of divergence between the corresponding morphological patterns or trajectories (Mitteroecker & Huttegger, 2009). However, Euclidean morphospaces with their wealth of interpretable geometric relationships are only possible for relatively small degrees of morphological diversity and for geometrically independent variables with comparable units. For example, in a space constituted by variables of different units, distances and angles cannot be interpreted – only linearity, intermediacy and patterns such as clusters of organisms (so-called affine geometry; Huttegger & Mitteroecker, 2011). For organisms differing in qualitative properties, such as the presence or absence of organs, or organ arrangement (phyllotaxis), no unique quantitative notion of similarity is possible because the variables possess a nominal scale only. Spaces with weak geometric relations (topological or pre-topological spaces) may still be possible to construct (Stadler et al., 2001). In general, the wider the studied morphological diversity, the weaker the mathematical and geometric properties of the morphospace (Mitteroecker & Huttegger, 2009).

Finally, assessing organ homology can be challenging when dealing with characters of distantly related clades. As a typical example in flowers, petaloid structures may be derived either from bracts, sepals or stamens, and the distinction between sepals and petals is often unclear (Kramer & Jaramillo, 2005; Ronse de Craene & Brockington, 2013).

How to visualize morphospaces

Depending on the type of data (nominal, ordinal or scale), morphospace representations are built using different methods. When dealing with two or three continuous scale characters (e.g. petal length and petal width), a morphospace can be simply represented as a two- (2D) or three-dimensional (3D) diagram in which each axis represents one of the variables (Allen et al., 2008; Fig. 1). When dealing with more than three characters (e.g. the length and width of three different kinds of floral organs), ordination methods are used to reduce the number of variables in such a way that relationships between individuals can be adequately represented in a 2D or 3D diagram. Widely used ordination methods are Principal Component Analysis (PCA), which creates a low-dimensional ordination that maximizes the variation between the measured specimens, and between-group PCA, which maximizes the variation between group means (Dommergues et al., 1996; Drake & Klingenberg, 2010; Córdoba & Cocucci, 2011; Mitteroecker & Bookstein, 2011; Mitteroecker et al., 2013).

While PCA can only be used for variables at an interval scale, metric and nonmetric multidimensional scaling (MDS and NMDS) are more general techniques that are also suited for ordinal and nominal data as long as a notion of (dis)similarity can be expressed numerically (Ollerton et al., 2009; Xiao & Laflamme, 2009; Bröderbauer et al., 2013). MDS, which is also referred to as principal coordinate analysis (PCoA), yields an ordination in which the Euclidean distances between the specimens in the low-dimensional representation approximate the original measures of (dis)similarity in a least-squares sense. If these original measures are Euclidean distances, then MDS leads to the same ordination as PCA. By contrast, NMDS ordinates the original distances or dissimilarity by maximizing the correlation between the ranked distances (Rabinovitz, 1986).

Ordination methods are powerful tools for the exploration of high-dimensional data, but they can also lead to misinterpretation, especially when single components are interpreted separately or in the presence of polynomial relationships between variables. For more details see Swan (1970), Mardia et al. (1979), Mitteroecker et al. (2005) and Bookstein (2013, 2014).

The statistical significance of group differences in mean or variance in the morphospace are usually tested by multivariate nonparametric or randomization tests, such as nonparametric analysis of variance (Anderson, 2001), analysis of similarity (Clarke, 1993), or permutation tests (Good, 2000).

Phylomorphospaces, in which (time-calibrated) phylogenetic frameworks are plotted within morphospaces, provide temporal landmarks for transitions identified in a morphospace. In addition, they allow us to highlight trends in the morphological diversification of taxa and to compare the magnitude and the direction of their morphological evolution (Macholán, 2006; Clabaut et al., 2007; Sidlauskas, 2008). Phylogenetic relationships are usually added to the morphospace by reconstructing possible ancestral states for all nodes of the phylogeny and including them in the morphospace (Sidlauskas, 2008; Klingenberg & Gidaszewski, 2010), or by linking extant taxa and their related fossils in the morphospace (Macholán, 2006; Benson & Choiniere, 2013). This method remains limited by phylogenetic and ancestral state reconstruction uncertainties.

How to measure morphological diversity (disparity)?

Disparity indices quantify the morphological diversity of a clade, that is the distribution of its constitutive taxa into vast or limited regions of the morphospace (Wills et al., 1994; Foote, 1997; Erwin, 2007). Examples of such indices are the sum of the variances of all characters (usually referred to as total variance), the mean pairwise distance between individuals (Foote, 1992, 1995; Wills et al., 1994), or the determinant of the sample variance-covariance matrix (generalized variance; Huttegger & Mitteroecker, 2011). All these indices are scalar summary statistics of multivariate distributions. They crucially depend on assumptions about common scales and the geometric independence of the variables (Huttegger & Mitteroecker, 2011). Reliable scientific inference (i.e. unaffected by factors such as sample size and the number and type of variables) may require the comparison of different disparity indices and ordinations (Foote, 1997; Ciampaglio et al., 2001).

Temporal changes in disparity values can be related to variation in taxonomic diversity. The evolution of taxonomic diversity in time (‘clade shape’) has, for example, been quantified by Gould et al. (1987) for marine invertebrates. They investigated whether the extant diversity of a clade evolved from a larger ancestral diversity (bottom heavy distribution) or, instead, increased steadily over time during evolution (top heavy distribution). Similar approaches have later been applied to quantify clade shape in the context of morphologic diversity (Wagner, 1995; Foote, 1997; Lupia, 1999; Hughes et al., 2013). The evolution of disparity through time and patterns of diversification can also be inferred on phylogenies by calculating the disparity through time (DTT) index, a measure of subclades’ disparity relative to their age (Harmon et al., 2003).

The evolution of disparity is central to understanding the evolutionary history of taxa. For example, it has been suggested that a high disparity might allow taxa to overcome extinction events more easily, as they have a broader spectrum of possibilities for adaptation to changing environments (Foote, 1997; Chevin et al., 2010).

Pollination syndromes

Pollination syndromes, or floral syndromes, are sets of floral traits, like shape, color, scent or rewards, which flowers have evolved as adaptations to different pollen vectors and which, theoretically, may appear repeatedly in unrelated lineages as a result of convergent evolution. A classification of plants according to pollination syndromes was first proposed by Delpino (Vogel, 1954, 2012), and later elaborated by Vogel (1954, 2012), van der Pijl (1961), Faegri & Van Der Pijl (1966), Endress (1994) and Proctor et al. (1996).

Morphospaces have been used in studies on pollination syndromes in three main ways. (1) As tools for describing pollination systems, that is, identifying the floral traits associated with various pollen vectors. Plant species under the same pollination regime are expected to cluster together in a space constructed with these traits (Wilson et al., 2004; Córdoba & Cocucci, 2011; Ortega-Olivencia et al., 2012; Bröderbauer et al., 2013; Maia et al., 2013). For example, Ortega-Olivencia et al. (2012) used a morphospace made of 35 floral traits to show that the bird pollination syndrome in Scrophularia was related to the showy colors of the corolla and staminodes, the lateral narrowing of the corolla tube and the absence of nectar guides. (2) Morphospaces were used to test the relevance of previously described pollination syndromes by investigating whether a given plant species’ distribution in the morphospace corresponds to the definition of the pollination syndrome. For example, Ollerton et al. (2009) ordinated 537 floral trait combinations corresponding to 11 theoretical pollination syndromes into a 3D space (a theoretical morphospace). They added plant species from six plant communities to this space, in which they had made observations of the most effective pollinators. Few of these species fell into the area corresponding to the predicted theoretical pollination syndrome, because these syndromes were not relevant, or because some of the plotted species were not specialized enough to their pollinators to exhibit the corresponding floral syndromes (see also Hingston & McQuillan, 2000; Ollerton & Watts, 2000). (3) Morphospaces were used to infer the pollen vectors of nondocumented pollination systems by including such species in the morphospace and evaluating whether they fall in an area of the space already occupied by closely related species belonging to documented pollination systems (Gibernau et al., 2010; Bröderbauer et al., 2013).

Instead of focusing on trait convergences, morphospace approaches and the calculation of disparity indices have also been used to demonstrate floral morphological divergence. It has been shown that in certain plant communities floral divergence was positively selected for, so that pollinators discriminate between different plant species and do not transfer heterospecific pollen by indiscriminate foraging (Eaton et al., 2012).

Finally, pollination syndromes have also been investigated on the basis of nonmorphological phenotypic characters, for instance scent, nectar composition and time of anthesis, by constructing phenotype spaces (Ollerton & Watts, 2000; Ollerton & Raguso, 2006; Cortis et al., 2009; Jürgens et al., 2013).

The sensory color space

Sensory color spaces are theoretical spaces used to evaluate the pollinators’ ability to discriminate between flower colors, an important prerequisite for studying the evolution of visual communication within small (e.g. co-occurring Oxalis species; de Jager et al., 2011) or large taxonomic groups (e.g. 876 plant species from all over the world; Stournaras et al., 2013).

Sensory color spaces were first proposed by Burkhardt (1989) and Goldsmith (1990) for bird vision, and have been used in studies investigating the ability of birds to discriminate between congeners’ plumages (Endler et al., 2005; Stoddard & Prum, 2008) or egg colors (Stoddard & Stevens, 2011). Each color is described as a point in a polyhedron or a polygon, whose position is determined by the relative stimulation of the different receiver photoreceptors. For example, in birds the sensory color space is a tetrahedron in which positions from each tip correspond to the relative stimulation of one of the four retinal cone types (UV-, small-, medium- and long-wavelength sensitive; Fig. 3).

Sensory spaces have been used to investigate plant color variation patterns in (bumble) bee- (Chittka, 1992, 1997; Tastard et al., 2008; Arnold et al., 2009), fly- (Troje, 1993) and bird-pollinated flowers, and bird-dispersed fruits (Burns et al., 2009; Stournaras et al., 2013). In most of the studies on bee-pollinated species, no pollinator-driven divergence (or convergence) in color characters was found (see also Gumbert et al., 1999; Arnold et al., 2009; de Jager et al., 2011), suggesting that there is no bee-driven selection of flower color. Alternatively, color variation may be constrained and thus hidden by other effects, such as pleiotropy, selection by herbivores or responses to environmental stress (Rausher, 2008; Schaefer & Ruxton, 2011). However, Stournaras et al. (2013) have shown that the volume of the avian color space occupied by fruits is much smaller than the one occupied by flowers, suggesting that bird pollinators and dispersers might exert different selective pressures on flower and fruit colors, or that the constraints on color pigment production are higher on fruits than on flowers (Fig. 3).

The latter study (see also Chittka et al. (1994) for a comparison of leaves and flowers) illustrates that sensory spaces are ‘standardized, generalizable and transportable’, meaning that different biological models can be ordinated in the same space during successive independent analyses (Stoddard & Prum, 2008).

Fitness landscapes

Fitness landscapes, first introduced by Sewall Wright in 1932 are spaces in which allele combinations (of individuals) or gene frequencies (of populations) are associated with fitness values. Simpson's (1944) adaptive landscapes are an extension of Wright's concept to phenotypic traits; adaptive landscapes hence are phenotype spaces or morphospaces, augmented with a fitness dimension (cited in Mitteroecker & Huttegger, 2009). Linking fitness values to floral phenotypes has been used to understand the mechanisms underlying speciation and the maintenance of hybrid zones (Whibley et al., 2006). It has also been used to quantify and represent correlative selection on floral traits (Schluter & Nychka, 1994; Campbell, 2009). For a 2D phenotype space, the adaptive landscape is represented as a surface with ‘peaks’ and ‘valleys’ that correspond to phenotypes of high and low fitness, respectively. Fitness landscapes and adaptive landscapes are central concepts in evolutionary ecology, both as metaphors and as actual mathematical models (Arnold et al., 2001). In Prusinkiewicz et al. (2007), for instance, an axis corresponding to a life history or climatic parameter was added to a 2D theoretical morphospace of inflorescence architecture. Fitness variation was modeled according to this parameter and highlighted by colors so that trajectories through paths of high adaptive value could explain evolution from a type of inflorescence to another in a changing environment. Such high fitness paths are called evolutionary wormholes by the authors (Prusinkiewicz et al., 2007).

Note that inferring fitness only from the distribution of groups in morphospace, without any independent fitness estimate, can be highly misleading. For example, even in the absence of any selection (i.e. equal fitness for all morphologies), the groups are unlikely to be equally distributed across morphospace. Clustering in morphospace is possible even in the absence of fitness peaks and valleys (Pie & Weitz, 2005).

Fitness is the sum of survival and reproductive success (Schluter, 1988). For plants, a good approximation of fitness is the sum of female (seed set) and male (seeds sired) reproductive success (Strauss, 1997; Dafni et al., 2005). These measures can be rather difficult to obtain and the fitness associated with a phenotype is thus often estimated by one or several of its components, such as pollen/pollinium removal (male success), pollinium reception and fruit/seed set (female success; O'Connell & Johnston, 1998; Maad, 2000; Conner et al., 2009; Benitez-Vieyra et al., 2010), pollinator visitation rates (Campbell, 2009), or simulations of population growth from one year to the next (Prusinkiewicz et al., 2007).

Stebbins’ topological space

The first floral morphospace, and to our knowledge the only one assessing the proportion of floral shapes among possible ones that actually evolved in nature, was created by Stebbins in 1951. Stebbins’ aim was to find evidence of the role of natural selection on floral trait evolution by demonstrating that some forms repeatedly appeared among angiosperms. He assumed that the adaptive values of floral traits were interdependent and that selection should therefore be studied using character combinations rather than single characters. Stebbins constructed a floral morphospace with eight binary characters related to the perianth, the androecium and the gynoecium (Table 1), and included data for 288 angiosperm families (as circumscribed in 1951). To visualize this morphospace, he organized the 256 possible combinations (2⁸) into a table of 16 rows and 16 columns (Fig. 4a,b). The number of angiosperm families in which these combinations occurred was then recorded in the corresponding element of the table to highlight the contrast among character state combinations with high and low frequency of occurrence in angiosperms.

Among the 256 possible character state combinations, only 86 were realized in angiosperm families (Fig. 4a,b). The six most common combinations were recorded in 17 to 36 different families, while 37 combinations were found in only one or two families (Fig. 4c). Stebbins interpreted most of the empty squares of the morphospace as ‘structurally impossible’ combinations (n = 96; for instance, the association of an inferior ovary and free carpels, and/or a parietal placentation and free carpels) or ‘apparently unadaptive combinations’ (n = 128; for instance, the union of petals and high stamen number, the presence of zygomorphy and high stamen number, and/or the absence of petals in zygomorphic flowers; Fig. 4a). His explanation for the success of highly represented combinations was that they might evolve easily and confer high ecological advantages. Among the realized combinations, five were represented by only a few but species-rich groups. These groups are Caesalpiniaceae (paraphyletic, now part of Fabaceae, see APG III (2009) for an up-to-date classification of angiosperms), Papilionaceae (= Fabaceae), Orchidaceae, Gramineae (= Poaceae) and Compositae (= Asteraceae). Stebbins hypothesized that direct competition might have prevented other groups from acquiring these trait combinations.

Stebbins’ representation can be easily interpreted in terms of state combination success. However, as Stebbins himself noted, this type of representation has a few caveats. The most important one may be the lack of meaning of the displacements in the chart, as similar morphologies are not grouped together in the representation. In fact, Stebbins’ representation is not even a space in the sense introduced above, as it does not offer any relations or measures of similarity between the different morphologies (see also Mitteroecker & Huttegger, 2009). In addition, this chart does not display any systematic information.

A new analysis of Stebbins’ dataset

Since Stebbins’ work, a range of new methods have been developed to construct morphospaces (see section 3). We created an NMDS ordination from Stebbins’ original dataset (see Supporting Information Methods S1), which allowed us to apply a simple distance metric for binary data (the number of variables in which two families differ) to Stebbins’ variables. We then measured disparity among angiosperm subgroups with three different indexes. Our goal was to construct an empirical morphospace in which we could investigate the spatial distribution of the 288 angiosperm families as sampled by Stebbins and to demonstrate the potential of the morphospace approach to study the evolution of angiosperms as a whole.

Our analyses of Stebbins’ dataset suggest that, when grouped into subgroups (basal angiosperms, monocots, basal eudicots, basal superrosids, fabids, malvids, Caryophyllales, basal superasterids, lamiids and campanulids; adapted from Soltis et al., 2011), angiosperms show significant differences in floral morphology (positions of the occupied areas in morphospace, npMANOVA overall test: df = 9, F = 11.3, r² = 0.19, P = 0.001, Fig. 5) and in disparity (spread of the occupied areas in morphospace, Fig. 6a). The NMDS ordination produces a good representation of the data when three dimensions were retained (stress value = 0.092; see Clarke (1993) for stress value significance; Fig. 5). Lamiids were significantly different from all other groups with a tendency to have dichlamydous flowers (with both petals and sepals), united petals, few stamens and united carpels. Monocots, basal superasterids and campanulids tended to group together, sharing a majority of dichlamydous flowers with united carpels. No apparent morphological grouping could be found among the remaining groups (basal angiosperms, basal eudicots, basal superrosids, fabids, malvids and Caryophyllales; see Fig. S1 and Methods S1 for character state frequencies and post hoc test results).

According to the mean pairwise distances and the sum of variances, malvids, lamiids and campanulids showed the lowest disparity, while basal superrosids, basal superasterids and fabids showed an intermediate level of disparity, and monocots, basal eudicots and basal angiosperms showed the highest disparity (Fig. 6a). According to the range (Fig. 6a,b), the highest disparity was reached by basal eudicots, then basal angiosperms, monocots, fabids, malvids and basal superasterids, whereas Caryophyllales and lamiids, then basal superrosids and campanulids showed less disparity.

Stebbins’ representation of his data is a theoretical morphospace, in which a number of extant angiosperm families were added to each character combination. It allows for the identification of character combinations that currently exist as compared to the ones that do not. Remember, however, that Stebbins’ approach is primarily a tabulation of data, not a space in the mathematical sense, because it lacks any relations or measures of (dis)similarity between morphologies. By constrast, our ordination analysis of the same data (using a simple metric) in order to construct an empirical morphospace, allowed us to visualize the relative positions of 10 angiosperm subgroups and their differences in disparity. These two approaches together provide a comprehensive picture of the proportion of realized trait combinations and of the distribution of floral disparity within angiosperms. The families sampled by Stebbins represent angiosperm diversity to an acceptable degree (compared to the number of families per category in APG III, 2009; Stevens, 2001; Fig. 4d); thus our analysis gives an idea of what will be possible with a more detailed and more comprehensive sampling strategy, more adequate character definitions and an incorporation of fossil data.