Transgressive phenotypes and generalist pollination in the floral evolution of Nicotiana polyploids

Polyploidy is an important driving force in angiosperm evolution, and much research has focused on genetic, epigenetic and transcriptomic responses to allopolyploidy. Nicotiana is an excellent system in which to study allopolyploidy because half of the species are allotetraploids of different ages, allowing us to examine the trajectory of floral evolution over time. Here, we study the effects of allopolyploidy on floral morphology in Nicotiana, using corolla tube measurements and geometric morphometrics to quantify petal shape. We show that polyploid morphological divergence from the intermediate phenotype expected (based on progenitor morphology) increases with time for floral limb shape and tube length, and that most polyploids are distinct or transgressive in at least one trait. In addition, we show that polyploids tend to evolve shorter and wider corolla tubes, suggesting that allopolyploidy could provide an escape from specialist pollination via reversion to more generalist pollination strategies.

Polyploidy is an important driving force in angiosperm evolution, and much research has focused on genetic, epigenetic and transcriptomic responses to allopolyploidy. Nicotiana is an excellent system in which to study allopolyploidy because half of the species are allotetraploids of different ages, allowing us to examine the trajectory of floral evolution over time.
Here, we study the effects of allopolyploidy on floral morphology in Nicotiana, using corolla tube measurements and geometric morphometrics to quantify petal shape. We show that polyploid morphological divergence from the intermediate phenotype expected (based on progenitor morphology) increases with time for floral limb shape and tube length, and that most polyploids are distinct or transgressive in at least one trait. In addition, we show that polyploids tend to evolve shorter and wider corolla tubes, suggesting that allopolyploidy could provide an escape from specialist pollination via reversion to more generalist pollination strategies. P olyploidy, or whole genome duplication (WGD), is an important driving force in the evolution of angiosperms. Ancient polyploid events are shared by all seed plants, all angiosperms and all core eudicots, and there is evidence of multiple independent ancient WGDs across the angiosperms [1][2][3][4][5] . Polyploidy is associated with many genomic changes, including chromosomal rearrangements, retrotransposon activity, gene silencing and epigenetic changes, and it can also yield novel phenotypes, which may facilitate establishment of polyploids as species distinct from their progenitors 1 . Potentially, such phenotypic changes in floral structures may affect pollination. Flowers attract pollinators with colour, scent, shape, size of floral display, nectar and pollen rewards, and aspects such as corolla tube or nectar spur dimensions can influence pollinator access to any reward offered. Pollinator-mediated selection on floral traits has been shown in many studies [6][7][8][9][10] and can even occur in generalist pollination systems because of different selective pressures from varying pollinator assemblages 11 .
Despite the importance of the changes following allopolyploidy, clear patterns of phenotypic evolution associated with allopolyploidy are hard to discern. In some cases, floral morphology in hybrids is intermediate between that of the parents, whereas in others phenotypes are transgressive (outside the range of the two parental species). For example, F 4 hybrids between Nicotiana alata and N. forgetiana are intermediate in 12 corolla tube and floral limb (the portion of the flower that opens at anthesis) measurements 12 . However, other homoploid (diploid) hybrids display transgressive phenotypes, especially when alleles segregate in F 2 and subsequent generations 13 . F 1 hybrids between two species of Petunia with different corolla tube length and floral limb size have transgressively larger floral limbs, but intermediate corolla tube length; in the segregating F 2 population, floral limb size is transgressive in both directions, and tube length can be transgressively larger 14 . In addition, autopolyploidy alone can yield floral changes even without the diversity generated by hybridization. For example, autotetraploids of Heuchera grossulariifolia have larger, differently shaped flowers than diploids and attract different floral visitors; 6 of 15 common floral visitor species preferred either diploid or autotetraploid flowers 15 .
Here, we study the effects of allopolyploidy on floral morphology in the genus Nicotiana. Nicotiana is an excellent system in which to study allopolyploidy because about half of its species are allotetraploids of different ages, including polyploids formed approximately 0.2, 1, 4.5 and 10 Myr ago; recent (synthetic) polyploids are also available (Table 1) 16,17 . This series of ever older polyploid lineages enables us to examine both the consequences of polyploidy at formation and how polyploids may have diverged over time from an expected intermediate floral morphology based on that of known diploid progenitors 16,[18][19][20][21] . Examples of floral morphology of Nicotiana polyploids and their diploid progenitors are found in Fig. 1. Our hypotheses are (1) polyploid floral divergence, from its predicted form at polyploid origin, will increase with age; (2) polyploid divergence in floral morphology will increase with greater progenitor morphological differentiation because there is the potential for greater floral variation on which selection/drift can act; and (3) polyploid species will diverge independently along lineage-specific trajectories.

Results
Floral morphological variation in Nicotiana. To examine the extent of variation present in floral limb shape, a geometric morphometric analysis of floral limb shape (Fig. 2a,b) was performed. The first two principal components of the resulting morphospace accounted for 58.18% and 20.33% of the variation present in the data, respectively. The morphospace of these two principal components was defined by two characters on diagonal axes; on one axis, the floral outline changed from round to stellate, and, on the other, relative size of the corolla tube opening ranged from small to large. The top right-hand corner of the morphospace consisted of impossible shapes in which the internal landmarks that mark the tube opening cross, creating negative tube width (marked in grey in all figures). Figure 2b shows the extent of the variation present in floral limb shape. Flowers from the same accession clustered in the first two principal components ( Fig. 2b and Supplementary Fig. 1), but no clustering was seen in subsequent principal components (for example, PC3 and PC4, Supplementary Fig. 1).
Corolla tube length and width were measured to determine the range of floral size. The accessions examined here clustered in the scatter plot of tube width versus length (Fig. 2c). Tube length ranged from 1.13 to 9.36 cm, and tube width ranged from 0.16 to 1.65 cm.
Polyploid floral morphology. An intermediate phenotype between those of the diploid progenitors is the simplest null hypothesis for estimating polyploid phenotype immediately following polyploidization. The extent of the evolution of polyploid floral morphology was assessed by comparing each polyploid mean to its progenitor midpoint (the mean of progenitor averages) and progenitor midpoint range (estimated range of the progenitor midpoint based on the range of progenitors; see Methods for details). The oldest polyploid section, section Suaveolentes, was likely to have been formed before diversification of several Nicotiana sections 21 ; therefore, a progenitor midpoint could not  be calculated for N. suaveolens. Instead, this polyploid was compared with the morphology of all the species in its progenitor sections ( Supplementary Fig. 2). Owing to the difference in comparison, this polyploid was left out of further calculations. In floral limb shape, 50% of polyploids overlapped with their progenitor midpoint range, 27% were distinct from the midpoint range but within the combined ranges of their progenitors and 23% had a transgressive phenotype outside the combined ranges of their progenitors (Table 2, Fig. 3 and Supplementary Fig. 2; see Methods for further explanation of 'overlapping', 'distinct' and 'transgressive'). Evidence from N. alata, N. forgetiana and their hybrids showed that control of tube length development is not genetically linked to that of tube width 12,22 . Therefore, we analysed these characters independently. In tube length, 77% of polyploids overlapped with their progenitor midpoint range, 9% were distinct and 14% were transgressive (Table 2, Fig. 3 and Supplementary  Fig. 2). In tube width, 68% of polyploids were overlapping, 9% are distinct and 23% are transgressive (Table 2, Fig. 3 and Supplementary Fig. 2). In examining all three traits together, most polyploids were either distinct or transgressive in at least one trait, whereas only 23% of polyploids overlapped with their progenitor midpoint range in all traits ( Table 2).
Direction and degree of polyploid floral divergence from expectation. Directional distances of each polyploid mean from its progenitor midpoint were plotted with the origin as the progenitor midpoint (Fig. 4). Using this figure, the direction of change in morphology from predicted for each polyploid could be determined (based on the quadrant in which the polyploid falls). In floral limb shape, 70% of polyploids fell along the round to stellate floral axis, whereas 30% of polyploids diverged along the small to large tube opening axis (Fig. 4a). In tube length and width, 25% of polyploids had a longer and wider tube than expected, 50% had a shorter and wider tube and 25% had a shorter and narrower tube. No polyploid had a longer and narrower tube than expected (Fig. 4b). This distribution was significantly different from equal numbers of polyploids in each quadrant (χ 2 = 12, d.f. = 3, P = 0.0074), and polyploids tend to have shorter and wider tubes than expected.
Older polyploids tend to be more divergent from the progenitor midpoint. To test whether younger polyploids had an intermediate floral morphology between those of their diploid progenitors whereas older polyploids had distinct or transgressive phenotypes, we plotted the distance of each polyploid from its progenitor midpoint against estimated polyploid age (Fig. 5).
In floral limb shape, distance from the midpoint increased with polyploid age. An asymptotic curve was fitted to the data ( y = 0.0953451-5.97903 × 10 −16 /x; Fig. 5a). The shape of this curve implies that divergence from the progenitor midpoint occurs early in polyploid evolution. A Mood's median test on polyploid distance from the midpoint revealed that the median distances of the synthetic and young polyploids (0.2 million years old (myo)) were significantly smaller than those of older polyploids (1 and 4.5 myo; χ 2 = 12.27, d.f. = 3, P < 0.007). However, although young and older polyploid groups were significantly different in distance, their ranges overlapped such that the most divergent synthetic and young polyploids were as divergent as the least divergent polyploids of 1 and 4.5 myo (Fig. 5a).
In tube length, regression analysis showed that polyploid distance from the midpoint and polyploid age were positively associated (F = 4.63; d.f. = 1, P = 0.043); however, older polyploids showed high variance in distance from the midpoint (Fig. 5b). This relationship also confirms our hypothesis that older polyploids tend to be more divergent from the progenitor midpoint than younger ones. There was no significant relationship between distance from the progenitor midpoint in tube width and polyploid age (F = 2.57, d.f. = 1, P = 0.123; Fig. 5c).
We also hypothesized that the morphological distance between diploid progenitor means might affect polyploid divergence. However, there was no correlation between progenitor distance and polyploid distance from the midpoint in any of the traits examined (data not shown).
Homoploid hybrid floral morphology. Several species in Nicotiana were shown to be homoploid hybrids 19,23 . However, the progenitors of these species were not represented by a single extant taxon; instead, it seemed that they were formed before the diversification of some of the sections of Nicotiana. Thus, we compared homoploid hybrid morphology with the total range of floral morphology found in the multiple species of their progenitor sections ( Supplementary Fig. 3). In floral limb shape, half of homoploid hybrids were overlapping with their progenitor range and half were transgressive (Table 2 and Supplementary Figs 2 and 3). In tube length, 83% of homoploid hybrids were overlapping and 17% were distinct (Table 2 and Supplementary Figs 2 and 3). In tube width, all homoploid hybrids overlapped with their progenitor range ( Table 2 and Supplementary  Figs 2 and 3). Most homoploid hybrids were transgressive or distinct in at least one trait, whereas only 33% overlapped with their progenitor range in all traits examined.

Discussion
Our results show that divergence in floral morphology increases with polyploid age (Fig. 5) but not with increased morphological divergence between progenitor diploids (data not shown). In addition, we show that most polyploids are distinct or transgressive in at least one floral trait examined (Table 2) and that polyploids tend to have shorter and wider corolla tubes than expected (Fig. 4b).
Floral evolution following polyploidy. Divergence in floral limb shape proceeds rapidly following allopolyploidization and tends to increase over time (Fig. 5a). Divergence in tube length in polyploids also tends to increase over time (Fig. 5b), but without the initial rapid increase. Divergence in tube width, however, is not associated with polyploid age (Fig. 5c). Differences in the overall patterns of floral morphological divergence in polyploids among characters suggest differing constraints on morphological trait evolution. Floral limb shape seems to be the most evolutionarily labile character as only 50% of polyploids overlap with their progenitor midpoint range in floral limb shape compared with 77% and 68% in tube length and width, respectively ( Table 2). Floral limb shape can also be highly variable within some species (for example, N. glutinosa) 24 , providing further evidence that the evolution of floral limb shape is less constrained. Potentially the large variability in shape among and within species arises because the floral limb is the most prominent feature of the flower for attracting pollinators, and it may be beneficial to be less constrained to facilitate rapid response to changing selective pressures.
Tube length is an important factor in pollination in species with nectar rewards, and it has been proposed that a fit between corolla tube or nectar spur length and pollinator tongue length improves both pollen removal and deposition 25,26 . Indeed, coevolution between floral tube/spur length and tongue length, as proposed by Darwin 27 , has been reported in plant interactions for hawkmoths 25 , long-tongued flies 28 and hummingbirds 29 . Within species (represented by a single population), floral traits that are involved in the fit between flower and pollinator are reported to have the smallest amount of phenotypic variation 30 . Pollinator-mediated selection on corolla tube length has been suggested from studies on six populations of Nicotiana glauca, and the strongest directional selection was proposed to occur in populations where there was the greatest mismatch between flowers and pollinators 31 . Taken together, this evidence suggests that tube length can be constrained by pollinator-mediated selection and is most likely to be under tight developmental control. Studies using N. alata, N. forgetiana and their hybrids show that corolla tube length, stamen length and carpel length are correlated 12 . In all natural polyploids examined here, stamens and carpels reach to the mouth of the corolla tube, even though some progenitors have exserted stamens. Synthetic N. tabacum QM is the exception (its paternal progenitor has exserted stamens), and the extent of stamen and carpel exsertion varies among synthetic lines (E.W. McCarthy et al., unpublished data). However, natural N. tabacum has inserted reproductive organs, perhaps suggesting that this trait is under selection in this species.    Figure 5 | Distance from the polyploid mean to the progenitor midpoint against estimated polyploid age for floral limb shape, corolla tube length and corolla tube width. a, Floral limb shape. An asymptotic curve was fitted to the data, which shows that distance from the progenitor midpoint in floral limb shape increases rapidly with polyploid age. b, Corolla tube length. Regression analysis shows that polyploid distance from the progenitor midpoint in tube length is positively associated with polyploid age. c, Corolla tube width. Regression analysis shows no significant relationship between polyploid distance from the progenitor midpoint in tube width and polyploid age.
Tube width is also likely to be under pollinator-mediated selection. Relative to hummingbirds, bumblebees prefer wider corolla tubes 10,32 . Narrower corolla tubes in hummingbird-pollinated Penstemon species are suggested to be due to both 'pro-bird' and 'anti-bee' adaptations 10 . In contrast, relative to hawkmoths, hummingbirds select for wider corolla tubes 6,7 . However, studies on flowers visited by short-billed and long-billed hummingbirds suggest that tube width is under greater selection from pollinators with shorter tongues and that changes in tube width may be a response to increases in corolla tube length 33 . Potentially, the evolution of a wider tube makes it possible for pollinators with shorter tongues to reach the nectar reward.
The divergence distance of the polyploids from the progenitor midpoint in tube width is not associated with polyploid age, largely because the oldest polyploids examined here have little divergence from expectation and the younger polyploids have a large range in divergence distances (Fig. 5c). The four allotetraploids of section Repandae, the oldest polyploids examined here, had a single origin 16,19,20 . One of these, N. nudicaulis, has a transgressively short corolla tube length, suggesting that even short-tongued pollinators can gain access to the nectar. The other three species, N. repanda, N. nesophila and N. stocktonii, are likely to be pollinated by nocturnal hawkmoths because they have long-tubed, white flowers that open at night. However, field pollination studies in Nicotiana are limited to a few diploid species, so the pollinators of these species have not been documented. Nicotiana nesophila and N. stocktonii are endemic to the isolated Revillagigedo Islands off the Pacific coast of Mexico 24 . Hawkmoths are present on these islands 34,35 , but hummingbirds have not been recorded in recent surveys 36 . There was a single potential sighting of a hummingbird in a survey from 1897 37 . The lack of hummingbirds may have resulted in reduced selective pressure for wider tubes 6,7 .
Pollinators can also use shape cues as they forage. Bees can learn to distinguish between different convex shapes 38 and prefer symmetry to asymmetry 39 . Hawkmoths prefer flowers with more dissected outlines (as opposed to convex) 40 . Shape, however, does not seem to be the most important cue in determining pollinator preference. Colour cues are more important than shape for bees 41 , and diurnal hawkmoths respond to both size and colour before shape while foraging 42 . Shape also appears to be less important than size for nocturnal hawkmoths because a preference for more deeply dissected floral outlines in Nicotiana section Alatae is no longer observed when floral shapes are standardized for surface area 43 . These observations that shape cues play a minor role in determining pollinator behaviour compared with other floral cues are consistent with our results that floral limb shape in polyploids is the most evolutionarily labile of the characters examined here ( Table 2).
Polyploids tend to evolve features associated with generalist pollination. Most polyploids have shorter and wider corolla tubes than expected by their progenitor midpoints (Fig. 4b), indicating a trend towards the evolution of features associated with more generalist pollination (although the pattern of floral evolution in polyploids differs somewhat if ancestral reconstructions of the progenitor morphologies are considered, polyploids still diverge towards shorter and wider tubes more often than is observed in diploids; E.W. McCarthy et al., unpublished data). It was previously found 44 that the number of species of floral visitors decreases as corolla tubes get longer but increases as tubes get wider, suggesting that having shorter and wider corolla tubes can in fact lead to more generalist pollination. Generalist pollination may be advantageous to polyploids, especially during establishment in a new or distinct niche from that of their progenitors. Generalism may increase the probability that a species will extend its range into new habitats and may decrease the likelihood of extinction [45][46][47] . The adaptation of specialism to certain pollinator types can hinder subsequent shifts to other pollinators, suggesting that specialism can be an evolutionary dead-end [48][49][50] . However, reversals in specialization do occur 51 , and in some plant systems there is no evidence to link pollinator specialization to decreased diversification 50 . Nevertheless, polyploids tend to evolve shorter and wider tubes, suggesting that allopolyploidy may provide an escape from specialization in pollination (dead-end or otherwise), allowing lineages to diversify by means of hybridization and WGD.

Methods
Plant growth. We grew plants in a greenhouse with a 16:8 h light/dark cycle that was kept between 10 and 30°C year round with heaters, midterms and fans. The Nicotiana accessions examined here are found in Supplementary Table 1.
Flower photographs. We took photographs at anthesis with a Canon PowerShot A520. Nicotiana otophora is characterized by senescence of the floral limb before full anthesis; therefore, it was photographed when the flower opened, but before the stamens dehisced. We took front and dissected view photographs for each flower (five replicates for each type per flower, five flowers per plant and five plants per accession), unless five plants were unavailable, in which case we took five to ten photographs for as many plants as possible (Supplementary Table 1).
We took front photographs by inserting the corolla tube into a hole in a piece of cardboard so that the floral limb laid flat. In some species (N. setchellii, N. langsdorffii and N. glauca), the floral limb forms a broad cup that cannot be flattened without disrupting the shape, so we placed flowers so the opening of the corolla tube lay against the cardboard and photographed directly above the corolla opening to minimize distortion of shape. For dissected photographs, we dissected flowers, removed the calyx and floral limb, and pinned the corolla tube open.
Geometric morphometric analysis. We used 15 landmarks to quantify floral limb shape (Fig. 2a). We created a TPS file from front photographs using TPSUtil version 1. 38 (ref. 52). We manually landmarked this series of photographs using TPSDig2 version 2.10 (ref. 53). We input these landmarked photographs into TPSRelw version 1.45 (ref. 54), which calculated a consensus used as the reference specimen. TPSRelw aligned all photographs to the reference, calculated partial warp scores for each photograph and a principal components analysis was performed on partial warp scores to calculate the relative warps of the dataset, which can be used to visualize the shape variation present in the dataset.
We imported the principal component scores into Excel and calculated flower and accession averages. We used scatterplots of both the first two principal components and the third and fourth to examine the clustering of individual flowers in the morphospace. We drew convex polygons around flower averages for each accession in two-dimensional plots of the morphospace to delimit the portion of the morphospace occupied by that accession. We marked the mean position of each accession in the two-dimensional morphospace with a circle. We exported images showing the relative warps of the landmarks from the extremes of the principal components from TPSRelw to visualize the limits of the morphospace.
Metric measurements and analysis. We measured corolla tube length (from dissected photographs) and width (from landmarks 5 to 14) using ImageJ version 1.42q (ref. 55). We calculated flower averages from five replicate photographs. We made scatterplots of tube width versus length using flower averages and used convex polygons and means to describe each accession.
Calculation of progenitor midpoint values. We used the average of the progenitor means in the morphospace to calculate the progenitor midpoint. We calculated the range expected around the progenitor midpoint based on the ranges of the individual progenitors. We calculated the range of each progenitor for each trait and chose the larger range in each trait for a pair of progenitors to estimate the range around the progenitor midpoint. We delineated these chosen ranges in the morphospace as an ellipse with the progenitor midpoint at its centroid. The chosen range was divided in half and both added to and subtracted from the progenitor midpoint to calculate the ellipse. This 'progenitor midpoint range' is a conservative measurement because it uses the maximum phenotypic variance of the progenitors.
Analyses of polyploid divergence. To analyse the extent of polyploid divergence from expected, we classified polyploids into three categories: overlapping (if a convex polygon enveloping all individuals of the polyploid accession overlapped with the progenitor midpoint range); distinct (if the polyploid polygon did not overlap with the progenitor midpoint range, but was within the combined ranges of the progenitors; or transgressive (if the polyploid polygon fell outside the combined ranges of the progenitors). We analysed tube length and width characters independently because the development of tube length is not genetically linked to that of tube width 12,22 . Grey boxes delimit the progenitor midpoint range in tube width versus length figures. Owing to the nature of geometric morphometric analyses, principal component (PC) 1 and PC2 do not correspond to individual traits; therefore, we analysed floral limb shape using PC1 and PC2 together as two-dimensional coordinates. We calculated the distance of the polyploid mean from the progenitor midpoint for floral limb shape, tube length and tube width.
We plotted the distance between progenitor midpoint and polyploid mean in each floral character against both estimated polyploid age and distance between diploid progenitors in the same character. An asymptotic curve was fitted to the polyploid distance in floral limb shape versus polyploid age plot in Mathematica 5.0 (Wolfram Research Inc.). We performed regressions on all other plots using Minitab version 15.1 (Minitab, Inc.) and used a Mood's median test to determine significant differences between medians of each polyploid age group.
To examine overall trends in the evolution of floral morphology following polyploidy, we plotted the directional distance of the polyploid mean from the midpoint as a vector initiating at the origin, which represents the progenitor midpoint. The figure shows only the direction and magnitude of the change in polyploid morphology because progenitor midpoints from different pairs of progenitors, which are distinct in morphology, were translated to the origin. Because the range in tube length across the dataset is much larger that of tube width, we standardized the divergence measurements for tube length and width to proportional values (calculated by dividing each species mean by the largest species mean in each character). We used a goodness-of-fit chi-squared test to determine whether the distribution of polyploids into the four quadrants of the graph was significantly different from equal numbers in each.