Direct and indirect responses to selection on pollen size in Brassica rapa L.


Lawrence D. Harder, Department of Biological Sciences, University of Calgary, Calgary, Alberta, Canada T2N 1N4. Tel.: +1 403 220 6489; fax: +1 403 289 9311; e-mail:


Pollen size varies little within angiosperm species, but differs extensively between species, suggesting the action of strong selection. Nevertheless, the potential for genetic responses of pollen size to selection, as determined by additive genetic variance and genetic correlations with other floral traits, has received little attention. To assess this potential, we subjected Brassica rapa to artificial selection for large and small pollen during three generations. This selection caused significant divergence in pollen diameter, with additive genetic effects accounting for over 30% of the observed phenotypic variation in pollen size. Such heritable genetic variation suggests that natural selection could effect evolutionary change in this trait. Selection on pollen size also elicited correlated responses in pollen number (–), flower size (+), style length (+), and ovule number (+), suggesting that pollen size cannot evolve independently. The correlated responses of pollen number, flower size and ovule number probably reflect the genetically determined and physically constrained pattern of resource allocation in B. rapa. In contrast, the positive correlation between pollen size and style length may represent a widespread gametic-phase disequilibrium in angiosperms that arises from nonrandom fertilization success of large pollen in pistils with long styles.


Pollen size varies extensively among angiosperm species, from grain diameters of 5–250 μm (Wodehouse, 1935; Whitehead, 1969; Muller, 1979), which corresponds to differences in volume approaching five orders of magnitude. Several intra- and inter-specific associations suggest that this variation partially reflects evolutionary adaptation to each species’ pollination and fertilization environment. In general, the evolution of pollen size seems to reflect the consequences of size for postpollination processes, rather than pollen transport (at least for animal-pollinated species: Harder, 1998). In particular, pollen size positively affects both pollen-tube growth rate (Van Breukelen, 1982; Lord & Eckard, 1984; Perez & Moore, 1985; Gore et al., 1990; Manicacci & Barrett, 1995), and seed-siring success (Cruzan, 1990; Lau & Stephenson, 1993, 1994). Furthermore, pollen-size differences between species commonly relate positively to variation in pistil characteristics that influence the environment for pollen-tube growth, such as stigma depth and style length (reviewed by Torres, 2000). Such associations imply that sexual selection for fertilizing ability may commonly favour larger pollen. However, this directional selection could be counteracted by a corresponding decline in mating opportunities if increases in pollen size reduce the number of pollen grains produced per flower (see Vonhof & Harder, 1995). Hence a species’ characteristic pollen size may balance the competitive advantages of large pollen against the numerical advantages of small pollen, given its specific reproductive environment.

Except for one study involving a heterostylous species (Manicacci & Barrett, 1995), intraspecific studies of the role of pollen size in postpollination success have relied on phenotypic, rather than genetic variation (van Breukelen, 1982; Lord & Eckard, 1984; Perez & Moore, 1985; Cruzan, 1990; Lau & Stephenson, 1993, 1994). As a result, differences in pollen performance need not indicate the effects of size itself, but could instead represent correlated effects of the environmental conditions responsible for variation in pollen size, such as the availability of minerals to the producing plant (Bell, 1959; Lau & Stephenson, 1993, 1994). To overcome this problem, we initiated a study of the effects of genetic differences in pollen size on postpollination performance. The first stage of this study, which we report here, involved an artificial selection experiment to produce plants that differ genetically with respect to pollen size. The second stage, to be reported elsewhere, involved a series of experiments in which small- and large-pollen competed for siring opportunities. In addition to producing plants with contrasting pollen sizes, the selection experiment revealed several features of the genetic architecture of pollen size that are relevant to pollen-size evolution.

The ability of sexual selection to modify pollen size depends fundamentally on the extent of genetic variation for size within species. Previous genetic studies of pollen size identified a range of control by additive genes. Studies of Spergularia marina (Delesalle & Mazer, 1995), Mimulus guttatus (Fenster & Carr, 1997) and Campanula rapunculoides (Vogler et al., 1999) found little additive genetic variation for pollen size, despite considerable variation for other floral traits. Such limited additive genetic variation is perhaps not surprising for a trait that affects fitness so directly (see Snow & Mazer, 1988; Walsh & Charlesworth, 1992). Nevertheless, genetic variation in pollen size has been maintained in some species. For example, Raphanus sativus exhibits significant additive genetic variance for pollen grain volume (Mazer & Schick, 1991; Mazer, 1992; but see Young et al., 1994) and artificial selection on the diameter of Phaseolus vulgaris pollen caused divergence over three generations (Montes-R & White, 1996). Hence, some species retain the genetic potential to respond to phenotypic selection on pollen size imposed by changes in pollination or postpollination conditions.

The potential for rapid evolutionary modification of pollen size also depends on the genetic associations of pollen size with other pollen and floral characteristics, which could either facilitate or impede adaptation (see Antonovics, 1976; Cheverud, 1984; Shore & Barrett, 1990; Conner & Via, 1993). Given the nature of pollen production and the reproductive role of male gametophytes, three processes could govern genetic correlations involving pollen-grain size. First, individual genes simultaneously affect the phenotype of several floral organs (reviewed in Coen & Meyerowitz, 1991; Weigel & Meyerowitz, 1994; also see Xie et al., 1999). Such pleiotropic genes would create positive genetic associations among floral organs when genotypes differ in the quality and quantity of gene products that affect the sizes of several organs (see Hill & Lord, 1989 and references therein). Secondly, the control of resource allocation could implicate pollen size in genetic correlations. In general, genes that control the relative allocation of resources among competing entities have negative pleiotropic effects on the sizes of those entities. This genetically controlled allocation could lead to two types of negative genetic correlations: tradeoffs in the sizes of competing entities, or functions, such as male and female investment (see Mazer et al., 1999; Koelewijn & Hunscheid, 2000); or tradeoffs between the size and number of competing entities, such as pollen grains (see Mazer & Hultgård, 1993; Stanton & Young, 1994). However, such negative effects are offset by the effects of genes that control the acquisition of resources (Houle, 1991), which can eliminate or reverse genetic correlations between competing entities (e.g. Young et al., 1994; Fenster & Carr, 1997). Finally, pollen size could become involved in genetic correlations if it interacts with other floral characters to cause nonrandom mating within a population. In this situation, certain combinations of alleles at loci controlling pollen size and the associated floral traits will become more common than expected from random mating. Although each of the preceding three mechanisms lead to specific expectations for the associations between floral traits, realized genetic correlations will depend less predictably on the combined influences of these mechanisms.

To date, the potential for pollen-size evolution, as governed by both genetic variation and covariation with pollen number and other floral characters has not been examined in a single study. To assess this capacity, we applied artificial selection for three generations on pollen size in Brassica rapa L. (syn. B. campestris, Brassicaceae). The resulting direct and correlated responses identify several aspects of floral integration that may be common features of pollen production by animal-pollinated angiosperms.

Materials and methods

Study species

Brassica rapa is a self-incompatible, annual or biennial mustard native to Eurasia and now widely naturalized in North America. In this study, we used a rapid-cycling stock of B. rapa developed previously by several cycles of selection for early flowering, rapid seed maturation, absence of seed dormancy, small plant size and high female fertility (Williams & Hill, 1986). Flowering of rapid-cycling B. rapa begins approximately 13 days after planting, and continues for 7 days. During this period, three or four flower buds open daily on inflorescences which produce 20–25 flowers. The plants used in this experiment were derived from two morphologically distinct base populations developed by Wisconsin Fast Plants (University of Wisconsin, Madison, WI, USA): wild-type (W) and anthocyaninless (A). The differences in anthocyanin production were not relevant for this study, but proved useful in a subsequent study of the effects of pollen size on gametophytic competition. Plants were grown singly in 4 × 4 × 5.5 cm potting units in standard greenhouse soil with 0.2 g of fertilizer (Osmocote 14:14:14 controlled release fertilizer) and maintained in a laboratory (23–26.5 °C) under constant illumination from cool-white fluorescent bulbs (230 μmol m–2 s–1 PAR).

Before describing our experimental protocol, we acknowledge two features of our study that could affect the generality of our results for plants in natural conditions (see Mitchell-Olds & Rutledge, 1986; Mazer & LeBuhn, 1999; Harshman & Hoffmann, 2000; Sgrò & Partridge, 2000) and outline our approach to assessing these effects. The first concern is whether the history of rapid-cycling B. rapa, which modified many aspects of life-history and plant stature, also altered the genetic architecture of pollen size. Two features of this history are relevant: the initiation of rapid-cycling lines by crossing early flowering plants of diverse provenance (Williams & Hill, 1986), and maintenance of developing lines under laboratory conditions for many generations. The initial genetic mixing is of the greatest concern, because it probably altered allelic frequencies at many loci, increasing additive genetic variation and altering genetic covariances unpredictably (see Bohren et al., 1966). We cannot assess these effects directly. Instead, we consider the generality of our results by comparing them with the findings of other genetic studies of pollen size. The effects of a history of laboratory conditions are probably less concerning, because the mating protocol used during the development of rapid-cycling B. rapa provided a typical environment for pollen size function. Pollen size in animal-pollinated species seems primarily to affect postpollination success, rather than pollen transport between plants (Harder, 1998), so that the relevant environment involves competition with other male gametophytes in the styles of compatible pistils. Rapid-cycling lines of B. rapa have been perpetuated by hand cross pollination, in part because B. rapa is self-incompatible. This hand pollination involved pollen transfer among many plants with the same brush (P.H. Williams, personal communication 2000), so that each stigma probably received pollen from a variety of plants, creating a typical competitive environment for male gametophytes. Consequently, the history of the plants considered in our experiment has probably maintained most environmental features that affect pollen evolution, including the nonrandom mating associated with competitive ability that could create gametic phase disequilibrium.

The second concern involves the extent to which our laboratory conditions altered the expression of heritable variation and correlated responses. Laboratory conditions per se probably do not affect additive genetic variation, however, they may facilitate detection of that variation by minimizing the environmental component of phenotype variation (although see Young et al., 1994; Weigensberg & Roff, 1996; Mazer & LeBuhn, 1999). We account for this possible effect both by presenting an absolute measure of additive genetic variation in pollen size (coefficient of additive genetic variation, CA=100√(VA/Y¯;) where VA is the additive genetic variation and Y¯ is the mean pollen size: Houle, 1992) and by estimating the approximate magnitude of heritability in natural conditions. If environmental variability affects phenotypic variation in pollen size, VP, but not the mean or additive genetic variance, then we can estimate the heritability expected in natural conditions according to

inline image

where hL2 and VP,L are the heritability and phenotypic variance in the laboratory, and CN is the phenotypic coefficient of variation expected under natural conditions. The consequences of laboratory conditions for correlated responses to selection are much less obvious. Genetic correlations that arise primarily from pleiotropy might be more consistent among environments than those associated primarily with gametic phase disequilibrium, especially if disequilibrium arises from nonrandom mating. Because the relative contributions of these genetic mechanisms probably differ, depending on the traits involved, we will discuss the effects of study conditions separately for each genetic correlation that we consider.

Selection protocol, and data collection

The lines considered in this experiment were derived from two base populations (W and A) initiated from different seed lots obtained from the Carolina Biological Supply Company (Burlington, North Carolina, USA). From each base population we established two replicate lines (W1 and W2, A1 and A2), which in turn served as the parental stocks for two selection lines (‘large-pollen’ and ‘small-pollen’) and a control line (see Fig. 1 for design details, including sample sizes). We staggered the planting of large-pollen, small-pollen and control line plants within a replicate by approximately 2 weeks to overcome the time limitation imposed by the short flowering period of individual plants (maximum 7 days) relative to the time required for data collection and pollinations.

Figure 1.

 Pedigree of genetic lines created during three generations of divergent selection on pollen diameter, with a randomly selected control (L – large-pollen line, S – small-pollen line and C – control line). Each population included two replicates. The numbers of individuals measured for pollen size in each line are indicated throughout.

During each generation we measured and counted pollen with a Particle Data, Elzone 180XY electronic particle analyser. From each plant, we collected all anthers from the fourth, fifth or sixth (occasionally seventh) flower after they dehisced (see Fig. 1 for population sizes). We suspended each pollen sample in 1.5 mL of 70% ethanol and then added 25 mL of 0.63% NaCl solution and sonicated for 5 min. Anthers were examined under a dissecting microscope to ensure that all pollen had been removed prior to analysis. The particle analyser then processed each pollen sample by drawing a 1-mL subsample through a 190-μm diameter aperture. As particles pass through the aperture, they change the resistance to an electric current flowing between electrodes on either side of the aperture. The analyser counts these resistance changes and measures their amplitude. The amplitude changes are converted into particle diameters and displayed as a frequency histogram of counted particles in 128 logarithmic diameter classes. We estimated the total volume of pollen produced by a flower, Vi, by Vi=∑j=11284/3πdi,j3, where di,j is the midpoint of the jth diameter class. Two 1-mL subsamples were analysed and the average pollen diameter (μm), pollen volume per flower (μm3), and pollen number were calculated for each pollen sample. Before counting each subsample, we inverted the counting vial gently, without introducing air bubbles, two or three times to ensure that all pollen grains were randomly suspended.

After quantifying pollen size in the four parental populations of 162–180 plants (see Fig. 1), we chose the 20 individuals with the largest and 20 individuals with the smallest mean pollen diameters as the parents for the large-pollen and small-pollen selection lines. For each of these selected parents we chose 10 plants randomly from the remaining 19 to serve as pollen donors, so that each parental plant received 10 pollinations and donated pollen for pollinations on 10 other plants. Pollen from each donor was used to pollinate a different flower on a maternal plant. During all pollinations, newly dehisced anthers were brushed across the stigma to cover it with pollen. We marked the pedicel of each pollinated flower with coloured tape that identified the paternal donor. When siliques were collected, they were placed in separate envelopes which were labelled with the identity of the maternal and paternal plants.

We imposed selection for three generations. During offspring generations 1 and 2, the 20 individuals with the largest and 20 individuals with the smallest mean pollen diameters were chosen to propagate the large-pollen and small-pollen lines, respectively. For offspring generation 1, we randomly assigned parental pairs as long as they did not share a parent from the previous generation (i.e. no half-sib crosses). In offspring generation 2, only plants with a coefficient of relatedness less than or equal to 0.0625 were paired and mated. This breeding design was adopted to minimize accumulation of genetic load. From the fruits produced by these matings we randomly selected seeds to initiate the next generation for each lineage. The first two offspring generations started with 180 seeds per line to enable further selection. In contrast, the third (final) offspring generation started with 40 seeds because we used these plants only to quantify the final response to selection.

The control lines of each replicate were initiated from plants in the parental populations that were not included in either selection line (hence these lines initially experienced stabilizing selection): thereafter, these lines were maintained by completely random mating. During each generation, we randomly selected 50 control-line plants as the mating group and mated each of these plants with one other plant chosen randomly from the group. Two flowers were pollinated for each combination to ensure at least one fully seeded silique.

To determine the genetic associations between pollen diameter and other floral traits, we collected the fifth, sixth or seventh (occasionally eighth) flower from all individuals in offspring generation 2 (large-pollen and small-pollen lines only) and stored them in separate microtubes in 70% ethanol. We subsequently measured petal length, maximum petal width, filament length, and style length at 6× and ovule number at 25× under a dissecting microscope for 40–50 randomly chosen flowers per replicate. In addition, we used counts of pollen number and pollen volume per flower obtained from the automated particle analyser to test for a tradeoff between pollen size and total pollen production.

Data analyses

The realized narrow-sense heritability of pollen diameter was estimated from the regression of selection response on cumulative selection differential (Falconer & Mackay, 1996). We calculated three estimates of heritability, based on three different measures of the response to selection after each generation (R):

inline image

where X¯L, X¯S and X¯C are the mean pollen diameters for the large-pollen, small-pollen, and control-lines, respectively. Measurement of the selection response as the difference between selected and control-line means (R and R) or as the divergence between the large-pollen and small-pollen line means R? eliminates common environmental effects, although the latter method estimates heritability more precisely (Hill, 1972). Selection differentials were calculated for each generation as

inline image

(Falconer & Mackay, 1996), where di is the pollen diameter of the ith individual selected for breeding, μ is the mean pollen diameter in the entire population, and pi is the proportion of individuals measured in the next generation that were the offspring of plant i. To calculate heritability based on the divergence between the large-pollen and small-pollen lines, we summed the selection differentials for each line (absolute values) to obtain the total selection applied. Variances for realized heritabilities were estimated according to formulae given by Ross (1997, pp. 140–142).

The average pollen size of control lines did not remain stable during the three offspring generations, so we assessed whether these changes could have resulted from genetic drift following Lande’s (1976) approach. This analysis considers the test statistic

inline image

where: Ne is the effective sample size; z2 is squared phenotypic divergence of the final offspring generation from the parental generation, measured in SDs; h2 is the heritability; t is the number of generations (4); and σp2 is the overall phenotypic variance. D is normally distributed with μ=0 and σ=1, so that the probability that drift provides a complete explanation for the change in a control line can be obtained from a table of standard normal deviates. In implementing this analysis, we used estimates of realized heritability based on divergence and estimated σp2 as the MSerror from an ANOVA comparing pollen size in the four generations.

To assess the direct response of pollen diameter to selection and the correlated responses of pollen number, total pollen volume, and floral traits we compared trait means in each population (anthocyaninless and wild-type) with analyses of variance. The analyses for pollen diameter considered replicate (1 or 2), selection line (large-pollen, small-pollen, control), generation, and their interactions as independent sources of variation. The analyses for correlated traits used a similar design with the omission of generation and its interactions with other independent factors, as we assessed correlated responses during only one generation (offspring generation 2 for floral traits, generation 3 for pollen traits). Pollen number and pollen volume were square-root transformed prior to analysis to satisfy the ANOVA assumptions. We analysed the two populations (anthocyaninless and wild-type) separately because their different selection histories could have resulted in different magnitudes and patterns of genetic variation.

We further examined the relation between mean pollen size and number during all generations to test the expectation that pollen production (n) varies inversely with investment per pollen grain (v=volume) from division of a fixed expenditure of resources on male gametes (E), such that n ∝ E/v (Vonhof & Harder, 1995). Because we measured a linear dimension (d=pollen diameter), whereas investment involves a cubic dimension (v), the relevant relation becomes

inline image

We tested this expectation by linear regression analysis after log transforming both sides of eqn 5, yielding

inline image

The full general linear model for this analysis also included population (anthocyaninless and wild-type), replicate nested within population, and their interactions. Selection of the final model involved backward elimination (α=0.05) of nonsignificant effects. We used a single-sample t-test to test the hypothesis that the partial regression coefficient for ln(pollen diameter) did not differ significantly from –3.


Direct responses to selection

Mean pollen size increased significantly in the control lines of replicates A1, W1 and W2 (Fig. 2). Although control lines comprised fewer plants than large-pollen and small-pollen lines, we rejected drift as a possible explanation for these increases in pollen diameter using Lande’s (1976) approach (P < 0.001 in all cases). Changes in the means of control lines probably reflect altered environmental conditions or inadvertent selection.

Figure 2.

 Direct responses to three generations of selection on pollen diameter in four replicate populations of Brassica rapa (A1, A2, W1 and W2). The left-hand panel for each replicate illustrates the mean (±95% CI) pollen diameter of plants in large-pollen (L —), small-pollen (S - - - ), and control (C ········) lines. Means with confidence intervals that do not overlap differ significantly. The right-hand panel depicts deviations of each selection line from the control line (L-C and S-C, see eqns 2 and 3) relative to the cumulative selection differential.

Despite the changes in control-line means, directional selection for large and small pollen diameter caused additional divergence between lines. After three generations of selection, the mean pollen diameters of large-pollen, small-pollen, and control lines differed significantly for all replicates (Fig. 2, left-hand panels). By offspring generation 3, plants in small-pollen lines produced pollen grains that averaged 0.34–0.66 μm (0.76–1.38 SDs) smaller than those in the initial parental generation, whereas in large-pollen lines the average pollen size had increased by 1.18–1.47 μm (1.99–3.27 SDs). Compared with control line plants during offspring generation 3, plants in small-pollen lines produced pollen grains that averaged 0.55–1.27 μm (0.99–2.94 SDs) smaller, whereas those from large-pollen lines were 0.40–1.00 μm (0.93–1.81 SDs) larger (Fig. 2, right-hand panels). By the third offspring generation, pollen grains from large-pollen line plants averaged 1.55–2.04 μm (2.98–4.45 SDs) larger in diameter than pollen from small-pollen line plants (Fig. 3).

Figure 3.

 Relation of the average divergence between lines selected for large or small pollen (see eqn 4) during three generations in four replicate populations of B. rapa (A1, A2, W1 and W2) to the cumulative selection differential (absolute value).

Replicates of each population did not respond consistently to opposing selection on pollen diameter (generation × line × replicate interaction, see Table 1). In particular, replicates A1 and W2 responded relatively symmetrically, whereas the large-pollen line of replicate A2 responded more strongly than the small-pollen line, and replicate W1 exhibited the opposite asymmetrical response (Fig. 2). The most asymmetrical response occurred in replicate W1, for which mean pollen diameter (±SE) increased by 0.11 (±0.06) μm per generation in the large-pollen line, whereas it decreased by 0.46 (±0.10) μm per generation in the small-pollen line.

Table 1. ANOVA evaluating the direct response to selection on pollen diameter in anthocyaninless and wild-type populations of B. rapa.Thumbnail image of

Differences among replicates in the magnitude of response were reflected in the estimates of realized heritability (±SE), which ranged from 0.13 ± 0.04–0.61 ± 0.08 (Table 2, also see Fig. 2, right-hand panels, and Fig. 3). Not surprisingly, these extremes in heritability involved the large-pollen and small-pollen lines of replicate W1, which responded asymmetrically to selection. During the three generations of selection, large-pollen and small-pollen lines diverged similarly for all four replicates (Fig. 3), resulting in equivalent heritability estimates based on divergence (mean h2=0.39, see Table 2).

Table 2.   Phenotypic and genetic variation in pollen diameter for two replicates from each of two populations of B. rapa. The phenotypic statistics depict the mean ± SD in the parental generation. Estimates of realized narrow-sense heritability (±SE) are based on regressions of selection response on cumulative selection differential during three generations. Separate heritabilities were calculated for the large-pollen and small-pollen lines (eqns 2 and 3) as well as an estimate based on the divergence between opposing selection lines (eqn 4). The estimates of the coefficient of additive genetic variation were calculated according to CA = 100√(VPh2/), where h2 is the divergence heritability estimate and and VP are, respectively, the phenotypic mean and variance in the parental generation (based on Houle, 1992). Thumbnail image of

Correlated responses to selection

We identified correlated responses to selection on pollen diameter based on consistent divergences between large- and small-pollen lines after two generations (floral traits) or three generations (pollen traits) of selection. According to this approach, pollen number, petal length, petal width, style length and ovule number exhibited correlated responses. The correlated responses for all these traits, except pollen number, imply positive genetic associations with pollen size. In contrast, total pollen volume and filament length did not change systematically in response to selection.

Pollen production

Pollen number changed in response to selection on mean pollen diameter (i.e. significant line effect), except in replicate A2, resulting in a line × replicate interaction for the anthocyaninless population (Table 3). A tradeoff between pollen size and number was evident in replicates A1, W1 and W2 (Fig. 4a). By offspring generation 3, plants in small-pollen lines of these replicates produced significantly more pollen grains than those in large-pollen lines. However, in general, pollen number did not differ significantly between the large-pollen and control lines. The apparent size-number tradeoff in the small-pollen lines is consistent with the somewhat greater response to selection for smaller pollen grains in these replicates (see Fig. 2).

Table 3.   Analyses of variance evaluating pollen production (offspring generation 3) and floral traits (offspring generation 2) after several generations of selection on pollen diameter for anthocyaninless and wild-type B. rapa. The analyses for pollen production considered control line plants in addition to large-pollen and small-pollen lines. Thumbnail image of
Figure 4.

 Correlated responses of (a) pollen number and (b) the total volume of pollen produced per flower to selection on pollen diameter in large-pollen (open circles), small-pollen (black squares), and control lines (grey diamonds). Means with 95% confidence intervals that do not overlap differ significantly. See Table 3 for statistical results.

Over all generations, pollen number varied negatively with mean pollen diameter (Fig. 5: F1,35=10.04, P < 0.01), irrespective of population (F1,35=2.41, P > 0.10) or replicate within population (F2,35=1.73, P > 0.10). The partial regression coefficient for ln(pollen diameter), which equalled –2.36 ± 0.74 (±SE), did not differ significantly from –3 (two-tailed single-sample t-test: t35=0.86, P > 0.2), suggesting that the pollen size-number tradeoff arose simply from the division of a fixed expenditure of resources to male gametes (see Fig. 5). In support of this conclusion, the total volume of pollen produced per flower did not differ significantly between the large-pollen and small-pollen lines in replicates A1, W1 and W2 (Fig. 4b), affirming that artificial selection on pollen diameter did not change total resource allocation to pollen in those replicates. Interestingly, plants in the small-pollen line of the one replicate (A2) that did not exhibit a correlated response between pollen size and number produced a smaller total volume of pollen than large-pollen or control line plants (Fig. 4).

Figure 5.

 The relation between the adjusted mean number of pollen grains per flower (n) and mean pollen diameter (d) for large-pollen (open circles), small-pollen (black squares), and control lines (grey diamonds). The dashed line represents the predicted number of pollen grains based on ln(n)=18.08−2.36ln(d). The influences of population and replicate (nested within population) have been controlled by adding the residuals from the entire general linear model for each genetic line to the predicted values for the observed pollen diameter.

Floral traits

Selection on pollen diameter induced correlated responses in petal size (Table 3, Fig. 6a, b). After two generations of selection, plants in the large-pollen lines of three of the four replicates (A1, A2 and W1) produced longer petals than those in the corresponding small-pollen lines (Fig. 6a). Furthermore, plants in large-pollen lines of two of these replicates (A2 and W1) also produced wider petals than small-pollen plants. These responses indicate the presence of positive genetic correlations between flower size and pollen grain size.

Figure 6.

 Correlated responses of (a) petal length, (b) petal width, (c) style length, (d) filament length, and (e) ovule number selection on pollen diameter in large-pollen (open circles) and small-pollen lines (black squares). Means with confidence intervals that do not overlap differ significantly. See Table 3 for statistical results.

Style length also apparently exhibited a positive genetic correlation with pollen size in the anthocyaninless population, but not the wild-type population. By offspring generation 2, large-pollen plants in the anthocyaninless population had longer styles, on average, than small-pollen plants (Fig. 6c), although the extent of divergence differed between the two replicates (line–replicate interaction, Table 3).

Filament length did not change consistently in response to selection on pollen size. In both populations, average filament length changed positively with pollen size in one replicate (A2 and W1) and negatively in the other (A1 and W2: Fig. 6d), resulting in significant line–replicate interactions (Table 3). It is unclear whether this inconsistent variation between selection lines reflects slightly different environmental conditions during the growth of the different replicates, or stochastic differences in genetic correlations with pollen size.

Ovule number exhibited the clearest indirect response to selection on pollen diameter (significant line effects, nonsignificant line–replicate interactions; Table 3). By offspring generation 2, plants in large-pollen lines produced significantly more ovules, on average, than those in small-pollen lines in both populations (Fig. 6e). This result persisted when either petal length or style length were included as covariates (analyses not presented), indicating that differences in ovule production occurred independently of differences in flower size. Hence, ovule number seems to be positively genetically correlated with pollen size.


Direct responses to selection

The Brassica rapa‘populations’ that we studied possessed the genetic potential to respond to selection on pollen size. Truncation selection that allowed mating between only the ~11% of individuals with either the smallest or largest pollen caused mean pollen size to diverge 3–4.5 SDs during three generations. If pollen size affects mating success, then such heritable genetic variation would allow natural selection to increase the incidence of pollen sizes that promote pollination and/or fertilization. The extensive variation in pollen size among species (Wodehouse, 1935; Muller, 1979) and limited variation within species (e.g. Vonhof & Harder, 1995; Cresswell, 1998) suggests that pollen size commonly experiences selection. Indeed, pollen size is phenotypically less variable than most other floral characteristics (Cresswell, 1998), which in turn are less variable than vegetative traits (Briggs & Walters, 1997).

Our heritability estimates (Table 2) probably indicate the maximal responsiveness to selection on the size of B. rapa pollen grains. The laboratory conditions under which we grew our plants resulted in phenotypic coefficients of variation around 2%. This variation equals the smallest coefficient of variation observed for pollen diameter for 30 species from natural environments and is about 60% smaller than the median CV for these species (3.3%: see Vonhof & Harder, 1995; Cresswell, 1998). Based on our results and an expected coefficient of variation of 3.3%, eqn 1 predicts h2 ≈ 0.16 in natural conditions. Two points are evident from this result. First, although our laboratory experiment provided strong evidence of additive genetic variation for pollen size in B. rapa, it would be more difficult to identify this genetic variation in natural conditions. Secondly, in natural conditions pollen size would respond to selection of the intensity applied in our experiment about one-third as strongly as we observed in the laboratory. These results illustrate both the virtue of our laboratory approach for isolating additive genetic variation for pollen size and the care that must be exercised when considering the implications of laboratory estimates of heritability.

Studies of six species provide contrasting evidence for genetic variation in pollen size. The histories of the plants studied may contribute to this variation. In particular, the three species with demonstrated heritable variation in pollen size have histories of genetic manipulation by plant breeders or are widespread agricultural weeds (R. sativus, Mazer & Schick, 1991; Mazer, 1992; P. vulgaris, Montes-R & White, 1996; B. rapa, this study), both circumstances that facilitate gene flow and could increase genetic diversity. In contrast, genetic variation for pollen size was not detected for three wild species (S. marina, Delesalle & Mazer, 1995; M. guttatus, Fenster & Carr, 1997; C. rapunculoides, Vogler et al., 1999). However, this seeming dichotomy may also reflect the statistical power to isolate genetic variation. Selection experiments in controlled environments provide a powerful means of detecting genetic variation, in part because they involve many plants, and both studies that applied selection on pollen size observed significant responses (Montes-R & White, 1996; this study). The remaining studies evaluated genetic variation based on either the resemblance of offspring to parents or differences between families. The ability of these approaches to detect genetic variation varies inversely with sample size, so it is not surprising that the study that found significant variation (Mazer & Schick, 1991) also considered two to eight times more families (n=60) than those that did not (Delesalle & Mazer, 1995; n=15 families per population; Fenster & Carr, 1997; n=23 and 25; Vogler et al., 1999; n=7). Two studies of R. sativus provide a particularly interesting comparison, as Mazer & Schick (1991) considered 60 families and found additive genetic variation for pollen size, whereas Young et al. (1994) considered only 11 families and found no significant variation. Indeed, Young et al. (1994) estimated coefficients of additive genetic variation (3–4%) about twice as large as our estimates for B. rapa (Table 2), although they assessed this variation to be not statistically significant. Hence, it is possible that all of these species possess heritable variation for pollen size, but it was not detected for three of them because the studies considered limited samples. Clearly, resolution of whether significant additive genetic variation for pollen size persists in most plant species requires studies based on large samples of a greater variety of species, preferably under natural conditions.

Correlated responses to selection

Despite significant heritable variation for pollen diameter, which should facilitate pollen-size evolution, the correlated responses by other floral characters reveal that pollen size cannot evolve independently. Such associations between traits could arise from gametic-phase disequilibrium and/or pleiotropy (Falconer & Mackay, 1996). Regardless of the genetic mechanism, genetic correlations between floral traits may be maintained by selection as adaptive trait combinations if they produce effective reproductive phenotypes (Stanton & Young, 1994).

The positive genetic correlation between pollen size and style length observed in B. rapa (Fig. 6c) may represent a gametic-phase disequilibrium that arises commonly in angiosperms because of nonrandom mating. In particular, a positive genetic correlation should evolve when large pollen has a higher probability of siring seeds in pistils with long styles because of faster germination and/or faster or more prolonged pollen-tube growth (see, Cruzan, 1990; Lau & Stephenson, 1993, 1994; T.S. Sarkissian, L.D. Harder & N.M. Williams, unpublished data). This mating pattern will cause alleles coding for large pollen grains to become associated with alleles that promote long styles, creating gametic-phase disequilibrium. In contrast, for heterostylous species in which plants with contrasting style lengths have a higher probability of exchanging pollen, disassortative mating should create a negative genetic correlation between pollen size and style length. Indeed, negative phenotypic and genetic associations between pollen size and style length typify heterostylous species (Darwin, 1884; Ganders, 1979; Dulberger, 1992). Although the disequilibrium responsible for such associations could arise solely from male–male competition, it also provides the genetic architecture needed for the evolution of female mate choice (see Fisher, 1958; O’Donald, 1980; Lande, 1981; Kirkpatrick, 1982). Both of these aspects of sexual selection should maintain associations between pollen size and style length. Such sexual selection has probably produced the widespread positive association between species with respect to pollen size and style length (reviewed by Torres, 2000) as an incidental outcome of the evolution of mating patterns within species.

The other genetic correlations that we detected, involving pollen number, flower size and ovule number, could all reflect the genetic control of resource allocation within and between flowers. The only negative genetic correlation detected during this study involved pollen size and number (see Figs 4 & 5), which has also been found in other studies (Mazer & Hultgård, 1993; Stanton & Young, 1994; but see Fenster & Carr, 1997). This inverse relation probably arose simply from division of a fixed expenditure of resources among pollen grains within individual flowers, because total allocation to pollen did not change relative to the control lines in seven of the eight selection lines (Fig. 4b). Such a fixed expenditure of resources among pollen grains within individual flowers enables two opposing allocation options –‘more, smaller’ or ‘fewer, larger’ pollen grains (Fig. 5). This negative genetic association is probably responsible for the parallel phenotypic tradeoff between size and number observed within species for pollen production (see Vonhof & Harder, 1995). Furthermore, such a tradeoff probably governs the optimal size-number combinations that benefit reproductive success in specific pollination and fertilization environments (Harder, 1998), resulting in the negative phenotypic association between pollen size and number observed between species (see Vonhof & Harder, 1995).

A genetic size-number tradeoff may help maintain heritable variation in pollen size if selection on mating success varies spatially and/or temporally (see Haldane & Jayakar, 1963; Via & Lande, 1987; Gillespie & Turelli, 1989). Changes in pollinator type and abundance (see Galen, 1989; Schemske & Horvitz, 1989; Kelly, 1992; O’Neil & Schmitt, 1993; Robertson et al., 1994) and/or changes in population structure may mediate changes in the intensity and/or direction of selection on pollen size (but see Harder, 1998). For example, suppose plants in a sparse population attract few pollinators and suffer insufficient pollination, whereas plants in a dense population receive large pollen loads from multiple donors. The first environment places a premium on more (smaller) pollen grains, whereas the second environment may favour larger (fewer) grains, if large pollen has an advantage during postpollination processes. If the optimal pollen size varies with these different pollination conditions, then the direction and magnitude of selection will also differ. Extensive pollen and/or seed dispersal between such populations would maintain genetic variation in pollen size.

In contrast to pollen number, flower size and ovule number exhibited positive genetic associations with pollen size. Such associations are consistent with current models of the genetic control of flower development, with individual genes influencing the phenotypic expression of many floral organs (reviewed in Coen & Meyerowitz, 1991; Weigel & Meyerowitz, 1994; see Xie et al., 1999 for a pollen-specific example). However, such control must obey the physical conservation of matter, so that these positive correlations must reflect one of the two classes of differences between plants with large vs. small pollen. In one class, plants with large pollen are more proficient in their acquisition or conversion of nutrients, so that they have more resources per flower to invest in petals and ovules than plants with small pollen. In the alternate class, all plants invest equally in reproduction, but plants with large pollen (and larger flowers with relatively more ovules) produce fewer flowers and/or smaller ovules. Unfortunately, we did not measure ovule size, flower number, or total reproductive effort, so that we cannot distinguish among these contrasting allocation responses. Nevertheless, the correlated responses of flower size and ovule number despite no contrasting response of total pollen volume remind us that the hierarchical nature of allocation in plants complicates reproductive tradeoffs greatly (also see Mazer, 1992; de Jong, 1993; Venable, 1996; Koelewijn & Hunscheid, 2000; Worley & Barrett, 2000).

The observed genetic characteristics of pollen size indicate that the evolution of pollen size in a particular reproductive environment will involve many floral traits. Suppose that large pollen grains compete more successfully for fertilizations (e.g. T.S. Sarkissian, L.D. Harder & N.M. Williams, unpublished data), particularly in long-styled pistils. Because pollen size is heritable, the resulting sexual selection will tend to increase average pollen size within the population and create a positive correlation between pollen size and style length. However, the increase in pollen size cannot continue unchecked, because it reduces the number of pollen grains produced, which limits a plant’s pollen export opportunities. This resource tradeoff creates an optimal pollen size that balances the advantages of large pollen size for gametophytic competition against the fecundity disadvantages of fewer pollen grains. Whether selection eventually brings the population to this optimum depends on the mating consequences of correlated changes in flower size and ovule number (and perhaps reproductive effort, flower number and ovule size). Hence, a plant’s pollen size is but one component of the integrated floral design that determines mating success in a particular pollination and fertilization environment.


We thank A.S. Gemmel, W.E. Gross, A. Kurji, J. LaMontagne, N. O’Brien and C. Payne for help with data collection, D. Houle, J.S. Shore and A.C. Worley for advice on experimental design, and S.C.H. Barrett, A.C. Worley and two unidentified reviewers for comments on the manuscript. The Natural Sciences and Engineering Research Council of Canada supported this study through a PGS-A Scholarship to TSS and a Research Grant to LDH.