Author for correspondence: Lawrence D. Harder Tel:+1 403 2206489 Email: firstname.lastname@example.org
I. Introduction 530
II. The process of floral and inflorescence adaptation 532
III. Experimental studies of flowers as adaptations 538
IV. Floral diversification: microevolution writ large? 539
V. Concluding comments 541
Although not ‘a professed botanist’, Charles Darwin made seminal contributions to understanding of floral and inflorescence function while seeking evidence of adaptation by natural selection. This review considers the legacy of Darwin's ideas from three perspectives. First, we examine the process of floral and inflorescence adaptation by surveying studies of phenotypic selection, heritability and selection responses. Despite widespread phenotypic and genetic capacity for natural selection, only one-third of estimates indicate phenotypic selection. Second, we evaluate experimental studies of floral and inflorescence function and find that they usually demonstrate that reproductive traits represent adaptations. Finally, we consider the role of adaptation in floral diversification. Despite different diversification modes (coevolution, divergent use of the same pollen vector, pollinator shifts), evidence of pollination ecotypes and phylogenetic patterns suggests that adaptation commonly contributes to floral diversity. Thus, this review reveals a contrast between the inconsistent occurrence of phenotypic selection and convincing experimental and comparative evidence that floral traits are adaptations. Rather than rejecting Darwin's hypotheses about floral evolution, this contrast suggests that the tempo of creative selection varies, with strong, consistent selection during episodes of diversification, but relatively weak and inconsistent selection during longer, ‘normal’ periods of relative phenotypic stasis.
‘During the summer of 1839 ... I was led to attend to the cross-fertilisation of flowers by the aid of insects, from having come to the conclusion in my speculations on the origin of species, that crossing played an important part in keeping specific forms constant. I attended to the subject more or less during every subsequent summer ...’
A celebration of Darwin's contributions would be incomplete without consideration of pollination and floral adaptation. The preceding quotation illustrates that Darwin actively examined this subject after age 30 yr and it highlights two of his central themes: adaptation by natural selection and the benefits of outcrossing. Darwin introduced these themes, as illustrated by plant reproduction, in The Origin of Species (Darwin, 1859) and elaborated on them in books on orchid pollination (Darwin, 1862, 1877a: Fig. 1), cross-fertilization and self-fertilization (Darwin, 1876), and heterostyly and dioecy (Darwin, 1877b). His detailed interpretations of orchid flowers as outcrossing adaptations specifically inspired the golden age of pollination biology during the late nineteenth century and, together with his general evolutionary insights, laid the foundation for current understanding of floral function. Indeed, Darwin's perspectives on floral function continue to motivate studies of plant reproduction and some of his hypotheses have been tested only recently (Table 1).
Table 1. Examples of recent tests of Darwinian hypotheses concerning pollination, mating and sexual systems
Darwin reference year (pp.)
Examples of contemporary studies
Coevolution of floral and pollinator traits; especially flower depth and proboscis length
Darwin considered floral traits as ‘beautiful contrivances’ for outcrossing and their remarkable diversity as the accumulation of persistent, gradual selection. As stated repeatedly in The Origin of Species, Darwin's (1859) view of diversification by natural selection was greatly influenced by Lyell's interpretation of the incessant role of erosion in structuring physical landscapes. According to this uniformitarian view, the process that created floral diversity is ongoing at a relatively constant rate and so is amenable to scientific inquiry. Indeed, the inconsistency of this perspective with the Cretaceous radiation of angiosperms created for Darwin an ‘abominable mystery’ (reviewed by Friedman, 2009).
In this review we consider whether contemporary evidence for natural selection on floral and inflorescence traits is consistent with Darwin's view of gradual diversification. We begin with a detailed summary of studies of natural selection, especially the within-generation component of phenotypic selection, to assess the evidence for the adaptive process. We then briefly consider experimental studies for evidence that floral and inflorescence traits are adaptations. Finally, we describe the possible roles of adaptation in floral diversification and review phylogenetic evidence of these roles. In general, current evidence is inconsistent with Darwin's uniformitarian perspective, so we propose a modified explanation of the association between the nature of selection and floral diversity.
II. The process of floral and inflorescence adaptation
As a process, adaptation is synonymous with natural selection, which occurs when fitness varies predictably with a genetically determined phenotypic trait, changing the trait distribution between consecutive generations. Because of selection, each generation tends to perform better than the previous generation, on average, and so is more adapted. Accordingly, complete analysis of the adaptive process must assess: (1) phenotypic selection, or the association between trait and fitness variation within generations; (2) the extent to which genetic inheritance creates resemblance between parents and their offspring; and (3) the change in trait distributions between generations. Few studies consider all components in natural populations, especially for long-lived species, so we review studies of the components of selection separately, focusing on quantitative traits, before considering complete examples. Other evolutionary processes influence floral and inflorescence traits, including genetic drift (Husband & Barrett, 1992), gene flow (Morjan & Rieseberg, 2004), and hybrid introgression (Rieseberg & Wendel, 1993); however, they do not produce adaptation and so are beyond the scope of this review.
1. Phenotypic selection
Selection requires phenotypic variation among individuals. Flowers and inflorescences commonly exhibit coefficients of variation of c. 15% for linear traits and numbers of structures, 25% for flower mass and > 50% for nectar characteristics (Cresswell, 1998). Floral morphology does not vary equivalently among angiosperms. For example, flower size varies more in species with radially symmetrical (actinomorphic) flowers than in those with bilaterally symmetrical (zygomorphic) flowers (Wolfe & Krstolic, 1999; Herrera et al., 2008). These summaries of phenotypic variation overestimate the opportunity for selection, because they combine variation within individuals (see Diggle, 2003) with variation among individuals, which is most relevant to selection. Nevertheless, this summary suggests ample phenotypic opportunity for selection.
Phenotypic selection also requires that fitness varies systematically with trait variation among individuals. For quantitative traits, such associations can be assessed conveniently by regression, as directional selection requires monotonically increasing or decreasing relations, whereas stabilizing and disruptive selection require relatively high or low performance, respectively, by individuals with intermediate phenotypes. To assess the incidence and intensity of phenotypic selection we surveyed studies that considered floral and/or inflorescence traits of open-pollinated plants. We included only studies that used Lande & Arnold's (1983) method, despite its various limitations (Mitchell-Olds & Shaw, 1987; see later), because the associated statistics facilitate comparison among studies. In particular, each observation, i, of absolute performance (‘fitness component’) is divided by the sample mean, transforming it into relative performance, wi, and the analysis considers standardized regression coefficients, which represent the change in relative performance associated with a one standard-deviation change in trait zj. The most complete regression model,
considers the partial regression coefficient of each trait, βj, as the main indicator of directional selection, the coefficient of a trait squared as a measure of stabilizing (γj < 0) or disruptive selection (γj > 0), and coefficients of the products of pairs of traits (γjk) as measures of selection for particular combinations of these traits (correlational selection). If only one trait is considered (i.e. p = 1) standardized estimates of βj and 2γj represent linear and quadratic selection differentials, or the combined direct and indirect effects of the trait on relative performance. By contrast, for analyses of several traits (i.e. p > 1), βj and 2γj approximate selection gradients, or a trait's direct effect, given the other traits included in the analysis. (Stinchcombe et al. (2008) noted that the doubling of γj is commonly overlooked, so reported estimates of stabilizing and disruptive selection are often too small by half. This error does not affect significance testing.)
The 56 studies that satisfied the inclusion criteria considered 44 species from 38 genera in 27 families (see the Supporting Information, Table S1). Ashman & Morgan (2004), and Delph & Ashman (2006) surveyed subsets of this literature, addressing specific aspects of selection (summarized below), rather than general features of floral and inflorescence adaptation. Except for one grass, all studies considered animal-pollinated species, and all but three studies considered herbaceous plants, revealing biases against abiotic pollination, autogamy and woody species. Based on information provided by the authors: 29 of the animal-pollinated species have actinomorphic, rather than zygomorphic, flowers; 21 of 36 species were self-compatible (not reported for remaining species); and 37 species have cosexual, monomorphic populations (the remainder included three dioecious, three gynodioecious and one tristylous species). Easily measured or counted aspects of floral morphology and production dominated these studies compared with aspects of color, scent, nectar, pollen and ovule production. To facilitate analysis, the measured traits were grouped into functional categories, such as attraction, contact of sexual organs with pollinators, etc. (see Table S2). Some studies considered the same population during several flowering seasons (nine studies) and/or examined more than one population (16 studies), and 20 studies considered multiple performance (‘fitness’) measures, resulting in 1073 records (cases) of selection differentials and/or gradients. All but four studies considered female performance (pollen import, fruit and/or seed production) and only four of the 13 studies that assessed male performance measured siring success. In general, phenotypic selection studies have treated seeds, rather than genetic contributions, as the relevant fitness measure, thereby ignoring the potential two-fold genetic advantage of viable selfed seeds over outcrossed seeds (Fisher, 1941). We used repeated-measures general and generalized linear models to analyse these selection estimates (see Figs 2 and 4), rather than formal meta-analysis techniques, because the latter require the complete phenotypic variance-covariance matrix and do not accommodate repeated measurement by individual studies (Kingsolver et al., 2001).
Selection acts on floral and inflorescence traits moderately often – a common result for surveys of phenotypic selection (Kingsolver et al., 2001; Geber & Griffen, 2003). All but three of the 48 study–species combinations (6.3%) that estimated selection gradients detected selection, but only 32.5% of the 692 estimated gradients differed significantly from zero (Table 2). Some of the nonsignificant estimates arose from one or more of three aspects of study design: consideration of traits that may not have been selection targets, but were conveniently measured or analysed traits (including principal components); small samples with limited statistical power to detect weak selection (Fig. 2a); and consideration of only linear regression coefficients (βj), precluding detection of stabilizing or disruptive selection (Fig. 2b: this explanation does not fully explain the disparity between studies that estimated γj or not (compare ‘β only’ column in Table 2 with parenthetical values in the ‘β and γ’ column)). Only 18 of the 145 correlational regression estimates (γjk) identified significant selection for pairs of traits; all but two of which involved flower production or display, or flowering date.
Table 2. Percentage of selection-gradient estimates for 48 study–species combinations that revealed different modes of phenotypic selection on floral or inflorescence traits for studies that estimated only linear selection gradients (β only), or both linear and quadratic gradients (β and γ)
β only (n = 394)
β and γ (n = 298)
Stabilizing and disruptive selection were identified if the maximum or minimum of the regression surface, respectively, was within three standard deviations of the trait mean (i.e. −/2 < 3 SD), which should include 99% of the observations for a normally distributed sample. Directional selection occurred if either differed significantly from 0 and was not estimated or not significant, or −/2 > 3 SD. Selection involved both directional selection and either stabilizing or disruptive selection if 1 SD < −/2 < 3 SD. The percentages in parentheses for cases in which both β and γ were estimated indicate the inferred distribution of selection modes considering β alone. Note that cases of pure stabilizing or disruptive selection may be less common (and cases of mixed selection more common) than indicated, because of a common error of not multiplying estimates by 2 (see Stinchcombe et al., 2008).
Stabilizing and positive directional
Stabilizing and negative directional
Disruptive and positive directional
Disruptive and negative directional
Directional selection occurred about four times more often than either stabilizing or disruptive selection, which were equally frequent (Table 2, ‘β and γ’ column). Positive directional selection occurred more than twice as often as negative directional selection, indicating generally better performance of plants with more and/or larger flowers (see below). Roughly half of cases of disruptive selection also involved (positive) directional selection. If the latter pattern persisted and was coupled with significant heritability, directional selection would eventually dominate as individuals in the most favored tail of the phenotypic distribution increased in frequency. By contrast, stabilizing selection tended to occur in isolation from directional selection, so the affected populations were centered under the prevailing fitness peak. The relatively low incidence of stabilizing selection (Table 2) suggests that floral traits are typically not fully adapted to the prevailing biotic and abiotic environment; however, this interpretation must be considered cautiously for several reasons. First, some nonsignificant cases may involve traits subject to stabilizing selection with a broad fitness peak, which would be difficult to detect without very large samples. Second, directional selection for more and perhaps larger flowers during a single reproductive season in a perennial species could be offset by reduced survival or reproductive success by the same plants during the next season (Obeso, 2002), resulting in stabilized selection based on lifetime fitness. Nevertheless, the available evidence suggests that phenotypic selection typically promotes a shift in mean phenotype, suggesting a mismatch with the prevailing environment, perhaps because the selective environment is dynamic over generations. However, common directional selection should promote diversification when populations differ in selection intensity or the traits involved.
In general, the incidence of phenotypic selection seems largely unaffected by species characteristics, including the existence of self-incompatibility and whether the pollinators had short- or long-proboscides. A possible exception involves floral symmetry, as actinomorphic species tended to experience selection less often (mean = 0.30, LSE = 0.037, USE = 0.040) than zygomorphic species (mean = 0.42, LSE = 0.048, USE = 0.050), although this difference was not quite statistically significant (T1 = 3.29, P > 0.05). Other recorded characteristics (growth habit, dichogamy, sexual system) were not analysed, because one category dominated the survey.
The incidence of phenotypic selection generally varied between years within populations and among populations during the same year, suggesting that Darwin (1859) was somewhat overzealous in claiming that ‘natural selection is daily and hourly scrutinising, throughout the world, every variation ... ; silently and insensibly working, whenever and wherever opportunity offers, at the improvement of each organic being in relation to its organic and inorganic conditions of life’ (p. 84). Of the 86 pairs of selection gradients during consecutive years, 41 involved no change (including 17 cases of no selection during both years), 35 involved selection during one year only and 10 involved transitions between selection modes. Identified causes of inter-year variation in selection included variation in the pollinator fauna (Conner et al., 2003), the presence of herbivores (Sandring et al., 2007), and abiotic factors (Maad, 2000; Caruso et al., 2003; Maad & Alexandersson, 2004). Of the 88 cases estimating selection on a trait simultaneously in multiple populations, 42 found contrasting results among populations and all but five of the cases of consistent results among populations involved no evidence of phenotypic selection. Inferred causes of inter-population differences in phenotypic selection include contrasts in: the pollinator fauna (Totland, 2001; Herrera et al., 2006 (not included in survey); Gómez et al., 2008), perhaps owing to differences in population size and isolation (Moeller & Geber, 2005) or interspecific competition for pollinator visits (Caruso, 2000); flowering-season duration (Hall & Willis, 2006); herbivory (Eckhart, 1993; Sandring et al., 2007); and abiotic conditions (Petit & Thompson, 1998; Caruso et al., 2003). Similarly, Ashman & Morgan (2004) found that the intensity of phenotypic selection on female function (i.e. absolute value of the selection gradient) increased with the severity of pollen limitation for 12 studies (see also Moeller & Geber, 2005). Such spatial and temporal variation in the incidence, mode and intensity of phenotypic selection underscores the relevance of the biotic and abiotic environment in determining a trait's influence on fitness. Biotic components of the environment are themselves subject to selection, so that selection responses by interacting species can reciprocally shape each others’ subsequent evolution (Anderson & Johnson, 2008; Pauw et al., 2009).
Overall, the incidence of selection differed among trait classes, with flower production experiencing selection about three times more often than other traits (Fig. 3). The high incidence of selection on flower production is unsurprising, because it determines a plant's capacity for mating opportunities through both ovules and pollen. Selection on flower production may also commonly induce indirect selection on other traits. For example, Schemske & Bierzychudek (2001) demonstrated that contrasting selection on white- and blue-flowered Linanthus parryi between years with low or high rainfall resulted primarily from opposing flower-production responses between morphs rather than from effects of color on pollination attraction and seed production per flower. Although not included in Fig. 3, because it was measured by only two studies, nectar production (volume and rate) deserves mention as the only trait category for which direct selection was not detected (22 estimates).
The magnitude of selection gradients, or selection intensity, also varied among trait classes (Fig. 4), although in a more complicated manner than the incidence of selection. The mean gradient differed significantly from 0 for three trait classes – flower color, flower production and phenology – but only when fitness was measured during seed production rather than during pollination (Fig. 4a). For a smaller sample of gradients for species with known self-incompatibility systems, significant mean gradients were detected for flower production, flower size, inflorescence display and phenology (Fig. 4b). However, six of the seven significant means for combinations of trait class, fitness stage and incompatibility involved self-compatible species. The interpretation of this imbalance is not obvious as three of the seven significant means were based on pollination, which should not vary directly with incompatibility.
For about three-quarters of cases in which both the overall (differential) and direct impact (gradient) of phenotypic selection were measured for the same trait in a population the outcomes were coincident (Table 3), indicating either consistent evidence of no selection (neither significant), or that direct selection was not strongly opposed by indirect selection on correlated traits (differential and gradient both significant). In the remaining cases, indirect selection counteracted direct selection, in a few cases causing significant selection opposite to that expected from the selection gradient.
Table 3. Comparison of estimates of overall phenotypic selection (differentials) and direct selection (gradients) on floral and inflorescence traits, based on percentages
Direct selection (gradients)
Overall selection (differentials)
When overall selection is significant, but direct selection is not, a trait is subject to only indirect selection via its association with other traits. When direct selection is significant, but overall selection is not, opposing indirect selection on a trait counteracts the effect of direct selection. The number in parentheses for cases in which both direct and overall selection are significant indicate the percentage of those cases in which the corresponding selection differentials and gradients have opposite signs, indicating that indirect selection overwhelmed direct selection.
n = 127
n = 126
Not significant, n = 161
Significant, n = 92
n = 97
n = 38
Not significant, n = 115
Significant, n = 20
The importance of direct versus indirect phenotypic selection is illustrated more clearly by studies that used path analysis or structural equation modeling (Kingsolver & Schemske, 1991; Mitchell, 1992) to test explicit biological hypotheses concerning causal linkages of traits and specific agents of selection (pollinators, herbivores, abiotic factors) to fitness. For example, Mitchell (1994), Gómez (2000) and Irwin (2006) reported that the direct effect of flower production on reproductive capacity can influence selection more strongly than indirect effects mediated by pollinator attraction. Path-analysis studies have specifically clarified four aspects of phenotypic selection on floral and inflorescence traits. First, the relative contributions of different effective pollinators (positive fitness effects) and of pollinators versus cheaters (neutral or negative fitness effects) to selection on floral and inflorescence traits have been distinguished (Schemske & Horvitz, 1988; Stanton et al., 1991; Irwin, 2006). For example, Stanton et al. (1991) found that the benefits of pollen number for increased visitation to Raphanus sativus flowers by small native bees affected selection on pollen production more than the direct effect of pollen number on siring success. By contrast, honey-bee visits varied positively with petal size, but they reduced siring success, causing negative indirect selection on this trait. Curiously, this approach has apparently been overlooked during recent debate about the extent to which pollination systems involve specialized interactions between a plant species and a restricted set of animals, or more generalized interactions with a diverse fauna (reviewed by Fenster et al., 2004). By allowing assessment of the relative efficacy of different animals as pollinators, path analysis would help distinguish between a species’ key pollinators and animals that either play a limited role in overall pollen dispersal, or actually reduce pollination success. Second, path analysis has helped reveal that responses of floral herbivores (Schemske & Horvitz, 1988; Ashman & Penet, 2007) and seed predators (Cariveau et al., 2004; Parachnowitsch & Caruso, 2008) to floral and inflorescence traits can affect phenotypic selection more strongly than pollinator responses (for a review see Strauss & Whittall, 2006). Third, path analysis effectively dissects the effects of trade-offs on phenotypic selection (Stanton et al., 1991; Hansen & Totland, 2006). For example, Hansen & Totland (2006) found a trade-off between flower size and number and weaker selection on flower size in a population where selection on flower number was strongest. Finally, path analysis has demonstrated stronger selection through seed production on flower number than on ovule number (Herrera, 1993; Conner et al., 1996; Gómez, 2008), again demonstrating the primary contribution of flower production to fitness. Together, these insights illustrate the great utility of path analysis for decomposing direct and indirect components of phenotypic selection. Unfortunately, these techniques are poorly developed for nonlinear relations, and so are most appropriate for examining directional selection. Some studies have implemented both Lande & Arnold's (1983) regression approach and path analysis to take advantage of their respective virtues (Herrera, 1993; Gómez, 2000; Parachnowitsch & Caruso, 2008).
Darwin (1859, 1871) distinguished natural selection, which involves survival and/or fecundity differences, from sexual selection, which involves differences in mating success, and he doubted whether that latter acted on immobile hermaphrodites with imperfect senses, which he judged incapable of mate choice. Acceptance of the possibility that sexual selection might act on plants required the replacement of the Darwin's (1876) perspective ‘... that flowers are adapted for the production of seed’ (p. 3), with recognition three decades ago that a hermaphrodite's fitness depends on its maternal and paternal genetic contributions. Sexual selection is now widely appreciated to influence floral and inflorescence evolution (Willson, 1994; Delph & Ashman, 2006), but expectations of its nature have shifted somewhat. Darwin proposed that sexual selection acts asymmetrically, favoring traits that promote many matings for males and mate quality for females. Bateman (1948) subsequently proposed that this asymmetry results because resource availability limits female reproduction. For plants, this asymmetry is expected when pollen import does not limit seed production, which seems to occur with only moderate frequency (Knight et al., 2005). Rather than negating the opportunity for sexual selection, pollen limitation reduces the asymmetry of its action and promotes traits that favor mating frequency for both sex roles. For example, Ashman & Morgan (2004) interpreted the similarity in phenotypic selection on flower size through female and male function as consistent with selection in hermaphrodites acting to balance contributions through the sex roles, on average, and inconsistent with perspectives based on Bateman's principle that attraction primarily serves male function. Similarly, Delph & Ashman (2006) reviewed six studies of phenotypic selection in seven species that jointly considered pollen removal and import and found that 13 species–trait combinations involved unilateral selection through one sex role or the other, eight involved parallel selection through both sex roles, and only 12 involved opposing selection through female and male function. Nevertheless, the possibility that sexual selection acts differently on female and male traits is clearly illustrated by the common sexual dimorphism for floral and inflorescence traits of dioecious species (Eckhart, 1999; see also Lankinen & Larsson, 2009). In addition, sexual selection probably strongly influences both female traits (e.g. stigma and style characteristics, self-incompatibility systems) and male traits (e.g. pollen size, aggregation and aperture number) that influence pollen–pistil interactions and ovule fertilization (Bernasconi, 2003), although they have not been the subject of phenotypic selection studies.
The studies summarized thus far considered selection as though it acts on plants independently of the characteristics of other individuals in the population; however, the fitness of many traits depends on the frequencies of other individuals with the same or alternative traits. In the context of mating, the long-term outcome of frequency-dependent selection depends on whether an individual benefits most from mating with individuals with the same phenotype (positive frequency dependence), which promotes dominance of the majority phenotype, or an alternate phenotype (negative frequency dependence), which favors polymorphism. The few empirical studies of relevant aspects of frequency-dependent selection have considered species polymorphic for floral color (Waser & Price, 1983; Gigord et al., 2001; Eckhart et al., 2006), morphology (Thompson et al., 2003; Brys et al., 2007) or sex phenotype (McCauley & Brock, 1998) and generally demonstrated negative frequency dependence. However, in some cases this selection occurred somewhat heterogeneously: Eckhart et al. (2006) found that some pollinators preferentially visited the majority morph, whereas others preferred the minority morph, and McCauley & Brock (1998) found negative frequency dependence for a gynodioecious species through male function of hermaphrodites, but through seed production of females. Frequency dependence can also affect selection on quantitative traits, but despite being an essential feature of models of evolutionarily stable reproductive strategies (reviewed by Morgan, 2006), it has received little empirical attention (but see also Aragón & Ackerman, 2004).
2. Quantitative genetics of flowers and inflorescences
The capacity for phenotypic selection to alter a trait's distribution between generations depends on both the additive genetic variation for the trait and its genetic associations with other traits (Conner, 2006). Ashman & Majetic (2006) reviewed heritability (h2) and genetic correlations (rg) of floral and inflorescence traits for 41 species, most of which were estimated under controlled conditions (glasshouse, growth chamber), which generally enhances detection of significant variation. On average, 39% of phenotypic variation could be attributed to genetic causes (includes broad- and narrow-sense heritabilities, which did not differ significantly), indicating considerable capacity for selection to modify trait distributions. Corolla dimensions (h̄2 = 0.46), aspects of pollen production (h̄2 = 0.43) and floral traits that can affect the relative incidence of self-pollination versus cross-pollination (e.g. stigma-anther separation, proportion of male flowers: h̄2 = 0.41) exhibited the most genetic variation, whereas aspects of nectar production were the most susceptible to environmental variation (h̄2 = 0.21). Flower number and display (h̄2 = 0.34) and traits associated with ovule production (h̄2 = 0.33) lay between these extremes. The moderate heritability for flower number, the trait most frequently subject to phenotypic selection (Fig. 3) is perhaps surprising, as it is often affected by a plant's resource status (Campbell & Halama, 1993). In general, self-compatible species exhibited lower heritability than self-incompatible species. These results indicate considerable standing genetic variation for floral and inflorescence traits, including flower production, so that genetic capacity should not generally limit the incidence of natural selection, although it may temper its magnitude.
Genetic correlations can constrain or accelerate genetic change, depending on a correlation's alignment with the direction of selection. The genetic correlations summarized by Ashman & Majetic (2006) were mostly positive (r̄g ≈ 0.4), which would tend to enhance the generally positive phenotypic selection reported above. This conclusion applied to correlations between pollen and ovule production, which are generally expected to vary negatively if they draw on the same resource pool. Indeed, the only consistent, if weak, evidence for a trade-off involved flower size and number. Overall, the positive genetic associations of floral traits were twice as strong for species with zygomorphic flowers (r̄g = 0.51) as for those with actinomorphic flowers (r̄g = 0.23), suggesting greater floral integration. In addition, a recent survey of the predicted evolutionary effects of genetic correlations by Agrawal & Stinchcombe (2009), which included some of studies surveyed above, concluded that these correlations impede evolutionary change slightly more often than they enhance it, but they rarely eliminate change in the direction of selection. For example, six generations of artificial selection on corolla-tube length and filament length in Raphanus raphanistrum caused significant evolution in the direction of greatest genetic resistance (Conner, 2006). Thus, genetic correlations may seldom seriously impede, and often enhance, floral adaptation.
3. Inter-generation evidence for natural selection
The few assessments of the combined effects of phenotypic selection and heritability that generate change between generations provide useful confirmation of the separate conclusions from studies of phenotypic selection and quantitative genetics and additional insights. Based on joint measures of phenotypic selection and quantitative genetics, five studies have predicted the impact of directional natural selection according to Δz̄ = Gβ, where Δz̄ is a vector of changes in mean phenotype for one or more traits, G is the genetic variance-covariance matrix, and β is a vector of phenotypic selection gradients for each trait (Lande & Arnold, 1983). Both studies that considered single traits predicted increased flower size. Galen (1996) predicted a 4–17% increase in the flare of Polemonium viscosum corollas between generations, which compared well with the observed 9% increase based on 6-yr survival of seedlings from her phenotypic selection study. This intensity of selection could produce the differences that Galen observed between low- and high-elevation sites within three generations. Morgan & Ashman (2003) predicted increases in flower size of 0.21 and 0.42 standard deviations (SD) for female Fragaria virginiana in gynodioecious populations with high and low hermaphrodite frequencies, respectively, compared with only 0.05 SD for hermaphrodites in both populations, indicating both that selection can differ between sex-morphs within a population and that the mating environment influences selection intensity. The three additional studies that predicted selection responses considered multiple traits and found evidence of indirect selection mediated by genetic correlations. Campbell (1996) found that only one (corolla width) of five measured traits of Ipomopsis aggregata was not subject to conflicting selection through correlated traits. Similarly, Caruso (2004) predicted that a size–number trade-off constrained short-term flower-size evolution of Lobelia siphilitica, even reversing it for some traits, and that this trade-off would slow evolutionary responses of flower number to positive directional selection. By contrast, Mitchell et al. (1998) predicted several cases of enhanced selection responses owing to genetic correlations for Penstemon centranthifolius.
Irwin & Strauss (2005) used the discrete floral-color polymorphism of R. sativus to assess inter-generation effects of natural selection. Based on measurements of the frequency, seed production, and offspring diversity of each morph, they predicted a significant shift in morph frequency between generations. By contrast, morph frequencies did not change, apparently because of differential survival, perhaps related to herbivore preferences. This finding emphasizes that evolutionary responses depend on the action of selection throughout the lifecycle, which can modify relative success during pollination and so complicate selection on floral and inflorescence traits.
III. Experimental studies of flowers as adaptations
‘I have found the study of Orchids eminently useful in showing me how nearly all parts of the flower are co-adapted for fertilisation by insects, and therefore the results of natural selection, – even the most trifling details of structure.’
Letter to Sir Joseph Hooker, 1862
The preceding statement illustrates that Darwin was the first ardent adaptationist, seeking evidence of natural selection in most characteristics of organisms. Interestingly, most of Darwin's (1862) adaptive interpretations of functions of orchid flowers represented hypotheses, rather than conclusions based on evidence, as he admitted that during ‘twenty years of watching Orchids [I] have never seen an insect visit a flower, excepting butterflies twice sucking O[rchis] pyramidalis [now Anacamptis pyramidalis] and Gymnadenia conopsea’ (pp. 34–35: to some extent this was rectified in the second edition (Darwin, 1877a), which included many descriptions of pollinator observations). By contrast, convincing demonstration that a trait, or suite or traits, is an adaptation, such as that provided by recent tests of Darwin's hypotheses (Table 1), requires evidence both of the trait's function and that it confers higher fitness than alternate traits that are reasonably possible given functional, historical, genetic and developmental constraints.
To assess the functional evidence that flowers and inflorescences are adapted to prevailing conditions, we surveyed 37 experimental studies (48 species in 24 families: see Table S3). Of the 48 experiments with different combinations of species and pollen vectors, 17 altered some aspect of attractive floral signals (flower size, number or color, petal number, nectar guides, scent), whereas the remainder involved traits that affect pollen exchange between flowers and pollen vectors (e.g. corolla-tube width, nectar-spur length, nectar volume, protandry). Overall, 73% of experiments detected significant effects of trait variation on performance, with no significant difference between studies of attractive versus interactive traits (Fisher's exact test, P = 0.50). In all but two of the 35 significant cases the response to manipulation indicated that the natural phenotype was more adapted than the altered phenotype(s). Thus, manipulative studies of floral and inflorescence function revealed more consistent evidence of adaptation than the phenotypic selection studies reviewed earlier.
Conclusions of adaptedness from functional studies must be treated cautiously for at least three reasons. First, most of the studies surveyed manipulated phenotype in one direction only, so the effect of change in the opposite direction was untested. In particular, because tissue can be removed more easily than added, most studies reduced floral or inflorescence traits, whereas many phenotypic selection studies with significant results indicated selection for larger traits (Table 2), so that the untested alternative manipulation might have led to contrasting conclusions about adaptation. Second, most studies imposed a more drastic phenotypic alteration than would typically occur during one generation of natural selection (e.g. reduction of flower area by one-half, or elimination of nectar guides). Indeed, many studies manipulated traits beyond the known variation in the study taxon (or its close relatives) and thus may be biologically unrealistic with respect to the ancestral variation during the trait's evolution. Finally, almost half (15/31) of studies for which a true control treatment would have been appropriate did not include this safeguard, raising the possibility that unintended manipulation effects were overlooked. Nevertheless, most well-designed studies that considered a range of alternate phenotypes detected significant negative effects of manipulation, providing support for the general conclusion of adaptedness.
IV. Floral diversification: microevolution writ large?
Darwin (1862) proposed that adaptation for outcrossing ‘carried on during many thousands of generations in various ways, with the several parts of the flower, would create an endless diversity of coadapted structures for the same general purpose’ (p. 351). Accordingly, floral diversification and associated phylogenetic patterns of floral traits should be consistent with the evidence of selection on floral traits reviewed earlier in this review. Detailed consideration of floral diversification is beyond the scope of this review, especially as it has been addressed recently by Armbruster & Muchhala (2009). However, the following brief overview of how pollination-mediated selection can create floral diversity clarifies the macroevolutionary relevance of the adaptive process. With this background we then briefly consider the phylogenetic evidence of diversification as an alternative perspective on floral adaptation.
1. Modes of pollination-mediated diversification
The adaptation of pollination systems could influence reproductive diversification in at least three ways, either independently or in concert. As Darwin (1859, 1862) stated in the quotation in the preceding paragraph and illustrated by his famous prediction that an unknown hawk moth with an 11-inch proboscis pollinated the Malagasy star orchid, he expected that coevolution produced reciprocal floral and pollinator adaptation, driving trait evolution in both mutualistic partners. Pairwise coevolution undoubtedly contributes to diversification in obligate mutualisms, especially brood-site mutualisms such as those between figs and fig-wasps (Thompson, 2005). However, most specialized pollination mutualisms involve greater specialization in the plant than in the animal partners (Ashworth et al., 2004) and this asymmetry leads to diffuse ‘guild coevolution’ between one or more pollinators and the suite of plants they pollinate (Anderson & Johnson, 2009; Pauw et al., 2009). Furthermore, coevolution cannot occur for species that are pollinated abiotically, or that rely on opportunistic pollinators or a varied pollinator fauna, and so they must diversify by some other process(es).
Divergent use of the same pollen vector, a second mode of diversification, is suggested by examples such as the remarkable diversity of inflorescence architecture among wind-pollinated grasses. The variety of processes involved in cross-pollination, including pollinator attraction, pollen removal, transport and deposition on stigmas, creates opportunities for different plant species to use the same pollen vector in contrasting ways with limited interspecific pollination (Harder, 2000). For example, the putative sister species Platanthera bifolia and Platanthera chlorantha (see Fig. 1c), place pollinaria on the tongues and eyes, respectively, of their shared moth pollinators (Maad & Nilsson, 2004), although the nature of selection that produced this transition is not well understood.
Divergence requires that lineages within a species follow contrasting evolutionary trajectories. Geographic separation facilitates divergence, regardless of the cause of differentiation, by both interrupting gene flow and allowing local adaptation, which can produce ecotypes (Turesson, 1922). Like the evolution of edaphic ecotypes in response to a geographic soil mosaic (Linhart & Grant, 1996), pollination ecotypes can evolve in response to a geographical mosaic of pollinators, creating extensive geographic variation in floral traits in association with contrasting pollination environments (Herrera et al., 2006; Johnson, 2006), including selfing races (Lloyd, 1965). Demonstration of pollination ecotypes caused by pollinator shifts requires evidence of a geographical pollinator mosaic (i.e. that pollinator differences between sites exist independently of the plant traits being studied) and that floral differences arose through pollinator-mediated selection and are genetically based, as established by reciprocal transplantation or common-garden experiments. Few published studies meet all of these criteria (reviewed by Herrera et al., 2006; Johnson, 2006).
Evidence for pollination ecotypes is available for cases involving coevolution and pollinator shifts, but has not apparently been examined for divergent use of the same pollen vector. In reciprocally specialized pollination systems, coevolution coupled with limited gene flow can generate remarkable fine-scale geographical variation in both flower and pollinator traits (Thompson, 2005; Anderson & Johnson, 2008). For example, in South-African guilds of long-tubed plants pollinated by long-proboscid flies coevolution primarily involves the flies and common, rewarding guild members, whereas rarer, nonrewarding guild members simply track the resulting changes in their shared pollinator (Anderson & Johnson, 2009; Pauw et al., 2009). Reciprocal translocation of plants within these geographical mosaics has confirmed the adaptedness of local ecotypes (Anderson & Johnson, 2008, 2009). Pollination ecotypes probably evolve more commonly through adaptive shifts between pollinator assemblages (Johnson, 2006). For example, spur-length ecotypes in the Disa draconis complex clearly arose from a pollinator shift between short- and long-tongued flies, rather than coevolution, because such nectarless orchids cannot influence tongue-length evolution in their pollinators (Johnson & Steiner, 1997).
The transition from ecotype (race) to species caused by divergent local adaptation must often occur imperceptibly. This distinction was not an issue for Darwin, because for him races and species differed only in the degree of divergence (Mallet, 2008). However, this uniformitarian view of species was largely supplanted by Dobzhansky's (1937) proposal that species have special properties, especially isolating mechanisms, that are absent in geographical races. Grant & Grant (1965) shared this neo-Darwinian perspective and proposed that pollination ecotypes be considered incipient species if complete isolation has yet to evolve. In practice, distinguishing whether allopatric forms are incipient or biological species is almost impossible and, in most cases, arbitrary from an evolutionary perspective; however, the coexistence of closely related species in sympatry requires reproductive isolation.
Contrasting floral traits of closely related sympatric species may act as isolating mechanisms whenever they promote intraspecific mating, whether selfing or outcrossing (Campbell & Aldridge, 2006). Well-documented cases include ethological isolation of Mimulus cardinalis (hummingbird pollinated) and Mimulus lewisii (bee pollinated) based on flower color (Bradshaw & Schemske, 2003) and mechanical and ethological isolation of Aquilegia formosa (hummingbird pollinated) and Aquilegia pubescens (hawk-moth pollinated) based on flower orientation and color (Fulton & Hodges, 1999). However, sister species often occupy different habitats, which can isolate species more effectively than floral traits (Rieseberg & Willis, 2007), even for species that have undergone pollinator shifts (Ramsey et al., 2003). Thus, pollinator shifts during speciation may often contribute more to phenotypic diversification than to reproductive isolation.
Although allopatry and local adaptation greatly facilitate floral divergence, divergence may also be initiated or accentuated in sympatry. Sympatric origin could be possible if genome duplication both abruptly enlarges floral traits and limits mating between cytotypes (Kennedy et al., 2006; Thompson & Merg, 2008). Contrasting floral dimensions between cytotypes would provide the opportunity for both divergent use of the same pollinator and pollinator shifts. Divergence could also be boosted by character displacement when the ranges of lineages that had previously diverged in allopatry begin to overlap. Indeed, in their survey Armbruster & Muchhala (2009) found character displacement to explain the positive association between floral specialization and species diversity more often than competing explanations, especially in clades with relatively imprecise pollination systems.
3. Macroevolution of angiosperm flowers
Pollination systems and floral traits are often highly labile and associated with cladogenesis, suggesting that reproductive shifts can drive speciation (Armbruster, 1993; Johnson et al., 1998; Weller & Sakai, 1999; Kay et al., 2005). During radiations, floral traits that promote outcrossing can evolve to extreme dimensions, such as the very long floral spurs of angraecoid orchids and the giant flowers of Rafflesia, the evolution of which accelerated, rather than diminished, over time (Barkman et al., 2008). By contrast, stochastic consequences of obligate selfing generally reduce genetic diversity, so selfing lineages tend to be short-lived and not prone to radiation (reviewed by Barrett et al., 2009), providing macroevolutionary confirmation of Darwin's (1859) assertion that ‘no organic being self-fertilises itself for an eternity of generations’ (p. 97).
Although lability is the hallmark of floral evolution, diversification is never unconstrained. Cases of directional, even apparently irreversible, shifts in pollination systems clearly illustrate historical constraints on the course of diversification. In New World lineages, hummingbird pollination evolves much more often from bee pollination than the reverse (Kay et al., 2005; Wilson et al., 2007). Similarly, Whittall & Hodges (2007) concluded that pollinator shifts, rather than coevolution, drove spur elongation in Aquilegia and that pollinator transitions from bees to hummingbirds to hawkmoths occurred without reversals. Functional constraints can also arise from processes other than pollination (Strauss & Whittall, 2006; Rausher, 2008). For example, the evolution of flower color in Dalechampia and Acer is apparently constrained because the same pigments confer protection against leaf herbivores (Armbruster, 2002). Similarly, Hakea species protected by spiny foliage tend to be insect-pollinated, because the spines limit access to flowers by birds, whereas evolutionary loss of these defense traits facilitates shifts to bird-pollination (Hanley et al., 2009). Finally, the pollination environment necessarily determines the opportunity for diversification, so that pollinator transitions are less likely and divergent use of the same pollen vector is more likely in regions with limited functional diversity of pollinators.
Despite constraints, pollinator-driven speciation apparently causes particularly rapid diversification in some lineages (Kay et al., 2006). Such adaptive radiation can arise after the evolution of a key innovation, such as floral spurs (Kay et al., 2006), or because lineages occupy geographical regions with particularly heterogeneous pollinator mosaics (i.e. a key environment: Johnson et al., 1998; Kay et al., 2005). Diversification may be accelerated by coevolution (Thompson, 2005); however, because both pollinators and flowers must evolve, this process may often occur more slowly than by divergent use of the same pollen vector or pollinator shifts. For example, as Yucca coevolved with its specialist moth pollinators it did not diversify faster than its sister clade, Agave, which has more generalist pollination systems (C. Smith et al., 2008). Thus, all modes of diversification probably contribute significantly to the variety of angiosperm flowers and inflorescences, although their relative impact depends on the nature (e.g. abiotic versus biotic, specialist versus generalist) and diversity of pollen vectors.
V. Concluding comments
This review has revealed somewhat contrasting views of floral and inflorescence adaptation. On one hand, the survey of phenotypic selection studies indicates that selection on individual traits acts sporadically within populations and inconsistently among populations. On the other hand, experimental studies of pollination function usually demonstrate the adaptive nature of floral and inflorescence traits in prevailing pollination environments and phylogenetic studies reveal patterns consistent with a key role for adaptation of pollination systems in floral diversification. This contrast probably occurs because many studies of contemporary natural selection do not represent the adaptive process that creates most functional diversity of angiosperm flowers. Specifically, this mismatch likely arises because clades generally diversify rapidly as phenotypic innovation or access to new environments allows evolutionary exploration of new opportunities, but diversification then slows as these opportunities are exhausted (Linder, 2008). Given that rapid diversification happens briefly during a clade's history, most short-term studies of phenotypic selection likely occur during periods of relative stasis when the pollination environment imposes weak and fluctuating selection of floral traits. By contrast, studies of floral function consider the products of preceding periods of intense selection and phylogenetic studies examine this historical process explicitly.
If floral diversification involves cycles of an initial ‘creative’ phase, when either a novel trait or a new environment accentuates the relation of fitness to phenotypic variation, followed by a longer ‘normal’ phase, when existing traits are relatively adapted, how can the creative component of selection on floral traits be studied? Our review of selection studies suggests that simply conducting more selection studies will be inefficient, largely confirming ‘normal’ patterns of selection that are already apparent. Instead, studies of selection should focus on populations with novel traits (either natural or induced; e.g., Cresswell, 2000) or in new environments (owing to natural or assisted colonization or habitat disruption). For example, invasive species may provide an opportunity to examine adaptation during the type of environmental shift responsible for diversification, although such studies are subject of various caveats (Colautti et al., 2009). Thus, just as knowledge of meteorology can assist in timing studies of the infrequent torrential rains that paradoxically shape many desert landscapes, understanding of the circumstances required for creative selection should increase the chance of studying the adaptive process in diversification mode. If studies of selection on novel traits or in new environments find a similar, relatively low, incidence of selection as found in the existing body of selection studies, then Darwin's hypothesis that selection creates the endless diversity of floral contrivances would have to be re-evaluated.
New approaches would also benefit the direct analysis of diversification, whether it considers the performance of pollination ecotypes or the phylogeny of floral and inflorescence traits. The often qualitative changes in floral traits associated with pollinator shifts make such transitions attractive and amenable subjects for diversification studies. For example, tests of association between pollinator type and floral traits in a phylogeny can provide clear evidence of pollinator shifts (Kay et al., 2005; Wilson et al., 2007; Whittall & Hodges, 2007; S. Smith et al., 2008). By contrast, divergent use of the same pollen vector cannot be studied as straightforwardly and has received little attention, even though it clearly contributes to the floral diversity of some clades. Phylogenetic associations between cytotype and floral traits may reveal some cases of such changes; however, divergent use of the same pollen vector need not involve genome duplication. In addition, the timing of phenotypic divergence relative to geographic or reproductive isolation deserves more attention to clarify the extent to which isolation precedes or follows trait change. Such information would also contribute usefully to discussions of the relative importance of the uniqueness of the phenotype or genotype in species identity (Mallet, 2008; Lexer & Widmer, 2008).
The perspective that emerges from this review is essentially Darwinian, acknowledging the conceptual and factual foundation that he established for interpreting the function and evolution of flowers and inflorescences. Experimental evidence of floral function typically demonstrates adaptedness and phylogenetic evidence reveals many instances of coincident cladogenesis and floral evolution. The main modification involves a shift from the expectation, implicit in modern studies of phenotypic selection, that the incidence and intensity of selection observable in many contemporary populations is equivalent to the nature of selection that produced their current adaptations. This hypothesis recognizes selection as the dominant creative evolutionary process but proposes that the tempo of floral adaptation varies. Thus, this view affects how the process of floral adaptation should be studied but does not challenge Darwin's emphasis that this process is responsible for organic diversity.
We thank S.C.H. Barrett, K. Duffy and an anonymous reviewer for comments on the manuscript, and the Natural Sciences and Engineering Research Council of Canada and the National Research Foundation of South Africa for funding.