MATING SYSTEM AND THE EVOLUTION OF STAMEN MORPHOLOGY IN THE MUSTARD FAMILY (BRASSICACEAE)

By

Anne Michelle Royer

A DISSERTATION

Submitted to Michigan State University in partial fulfillment of the requirements for the degree of

Plant Biology – Doctor of Philosophy Ecology, Evolutionary Biology and Behavior – Dual Major

2014

ABSTRACT

MATING SYSTEM AND THE EVOLUTION OF STAMEN MORPHOLOGY IN THE MUSTARD FAMILY (BRASSICACEAE)

By

Anne Michelle Royer

Biotic diversity is characterized by patterns of both divergence and similarity. Both natural selection and constraint may operate to conserve a trait across related species, depending on the ecology of the species. My dissertation investigates the roles of these evolutionary forces in maintaining a family-diagnostic stamen morphology. Flowers in the Brassicaceae (mustard plant family) are characterized by four long and two short stamens within a flower (Zomlefer 1994).

Although found in >95% of the genera in the family (Endress 1992), the reason(s) for the widespread conservation of this morphology are not known.

Existing adaptive hypotheses for the evolution and the maintenance of tetradynamy address how the trait could increase fitness for outcrossing species.

While there is some evidence that tetradynamy is adaptive in self-incompatible members of the Brassicaceae (e.g. Kudo 2003; Conner et al. 2009), how it functions is not clear. Additionally, there have been multiple independent losses of self- incompatibility in the family (Lloyd 1965; Mable et al. 2005), followed in some cases by the evolution of high self-pollination rates (Preston 1986). In these highly selfing species, the maintenance of short stamens, which general appear too short to pollinate the stigma within a flower (Müller 1961), is particularly mysterious.

My dissertation unites approaches from evolution, ecology, and genetics to understand the maintenance of short stamens in the Brassicaceae across the full range of mating systems. I investigate the function and morphology of short stamens in three species: the obligately outcrossing wild radish (Raphanus raphanistrum), the highly selfing model plant Arabidopsis thaliana, and A. thaliana’s sister species,

A. lyrata, which includes both outcrossing and selfing populations.

In the outcrossing species Raphanus raphanistrum, I used experimental manipulations in arrays exposed to pollination in the field to show that having more stamens increases male fitness, and female fitness is also affected by stamen treatment. There were some indications that short stamens were more attractive to pollinators at high overall pollinator visitation rates.

In the highly selfing model plant Arabidopsis thaliana, I showed that short stamens do not significantly increase fitness. I found many populations, particularly near the southern end of the geographic range, have partially lost the short stamens.

Genetic mapping revealed three QTL controlling the number of short stamens, with strong epistasis greatly reducing their individual effects. Ongoing evolutionary loss of non-adaptive short stamens in Arabidopsis thaliana may be slowed by gene interactions, low genetic variance in the north, and an inbreeding mating system.

Finally, I investigated the evolution of floral morphology in selfing and outcrossing populations of the mixed-mating Arabidopsis lyrata. I found that while relative investment in female fitness has increased as predicted in selfing populations, predicted changes in size and evolution of short stamens have not yet occurred. This may be consistent with recent evolution of selfing, continued facilitation of pollination by insects, and/or constraint on the evolution of short stamens.

For Toby, who taught me I’m capable of things I wouldn’t have believed possible, and Matt, who supported and loved me unconditionally.

iv ACKNOWLEDGMENTS

I’m grateful first to Jeff Conner for his tireless dedication as an advisor. He has been an exceptional editor, an encouraging mentor, and an inspiration with his continuing enthusiasm for the unfolding questions being pursued in his lab. He always made sure I had what I needed to do the work I wanted to do, and I rarely had to wait to get requested feedback.

My other three committee members have also played essential roles in my development. Doug Schemske hosted me in his lab as a new graduate student on campus, eventually supporting me in ongoing collaborations with his lab that have been some of my most productive projects. I’ve benefitted immeasurably from the time he’s spent pushing me to think harder and follow my dreams. Jen Lau offered me support and mentoring at KBS, especially since our sons were born and in the last year when Jeff was on sabbatical. Ian Dworkin also offered an important unique perspective on my research, with particular contributions to the genetic mapping section and several of the future directions that will hopefully develop in the next few years.

Financial support for the work in this dissertation came from NSF grants to

Jeff and Doug and MSU funds to me, particularly KBS Lauff and Porter funds. The

BEACON and GK-12 programs supported me for several years and helped me achieve my goals of integrating teaching and research. Thanks to Tom Getty for involving me in both programs and being the most encouraging faculty member at

KBS. Louise Mead at BEACON was also a key mentor in my education pursuits, and my students 2008-2014 have inspired and motivated me.

v Jeff’s excellence as an advisor includes his ability to bring together a great group of people. I was lucky to spend much of my time at KBS with my lab-sister

Raffica LaRosa (a comrade in good times and bad, research and personal life) and lab manager extraordinaire Cindy Mills, who served as a sounding board and compatriot in years of counting stamens and running gels. Other characters that have also played important parts in my time in the Conner lab include Sam

Slowinski (2009 REU and coauthor on my radish work), “younger” labmates Sam

Perez and Amanda Charbonneau, 2013 REU Marvin Osborne, who counted many pollen grains, and a long list of other summer undergraduates Jeff supported to help me with my work over the years.

My colleagues at Kellogg Biological Station made it a fantastic place to work through their enthusiasm for scientific discussions, side projects, and all other kinds of fun. In no particular order, I’m particularly indebted to Colin Kremer, Mike Grillo,

Todd Robinson, Lauren Kinsman, Emily Grman, Rachel Prunier, Idelle Cooper, Liz

Schultheis, Kane Keller, Casey terHorst, and Tomomi Suwa. Amanda Posto at

Indiana University was also an influential friend and collaborator.

On a personal level, I want to thank my yoga teacher Karina Mirsky and the community at Sangha Yoga in Kalamazoo for getting me through grad school happier and healthier than I was when I started. Finally, I’m grateful to my family for believing in me and supporting me in so many ways.

vi TABLE OF CONTENTS

LIST OF TABLES ix

LIST OF FIGURES x

CHAPTER 1 1 INTRODUCTION 1 Background 1 Organization of Dissertation 2

CHAPTER 2 5 STAMEN FUNCTION IN WILD RADISH DEPENDS ON POLLINATOR VISITATION RATE 5 Introduction 5 Methods 9 Field experiment 9 Male and female fitness estimates 12 Analysis 14 Results 15 Slow release hypothesis 15 Trait specialization – long stamen attraction hypothesis 21 Discussion 23 Acknowledgements 29

CHAPTER 3 30 ONGOING LOSS OF A CONSERVED TRAIT: LACK OF FUNCTION, LATITUDINAL PATTERNS, AND GENETIC CONSTRAINTS 30 Introduction 30 Methods 32 Results 33 Discussion 39 Acknowledgements 40

CHAPTER 4 42 EARLY EVOLUTION OF SELFING IN ARABIDOPSIS LYRATA INCLUDES CHANGES IN SEX ALLOCATION BUT NOT FLOWER SIZE 42 Introduction 42 Methods 47 Field common garden 48 Greenhouse common garden 49 Floral measurements 51 Results 53 Discussion 63

vii Acknowledgements 65

CHAPTER 5 66 SUMMARY AND FUTURE DIRECTIONS 66 Summary 66 Future Directions 66

APPENDIX 69 Growth conditions for Arabidopsis thaliana 70 Experiment: function of A. thaliana short stamens in selfing 71 Geographic variation in A. thaliana short stamen production 78 QTL mapping 78 A. thaliana candidate gene search 92

LITERATURE CITED 93

viii LIST OF TABLES

Table 1. Insect visitors observed visiting experimental plants. 17

Table 2. Effects of stamen treatment and pollinator visitation rate on male and female fitness. 18

Table 3. Effects of stamen treatment and overall pollinator visitation rate on visits to individual plants. 22

Table 4. Arabidopsis lyrata populations and sampling scheme 45

Table 5. Models of trait differences between populations and mating systems 54

Table 6. Models testing for increased variance in short stamen length with shift to self-pollination 60

Table 7. Accessions included in the study of geographic variation in short stamen production 70

Table 8. Testing function of short stamens 75

Table 9. Plants, lines, and flowers sampled for QTL analysis 78

Table 10. Locations and 95% credible intervals for main-effect QTL peaks 86

Table 11. Significance of main effects and interactions 87

Table 12. Number that fall within QTL for stamen loss in Arabidopsis thaliana 88

Table 13. Details for candidate genes 89

Table 14. Results of three models in R/qtl 90

ix LIST OF FIGURES

Figure 1. Experimental stamen removal treatments applied to flowers. 10

Figure 2. Male fitness of 2 vs. 4-staminate treatments over varying pollinator visitation rates. 16

Figure 3. Male fitness of with different stamen treatments over varying pollinator visitation rates. 19

Figure 4. Female fitness by treatment 20

Figure 5. Square-root transformed pollinator visitation rates of different treatments over varying overall pollinator visitation 24

Figure 6. Female fitness of with different stamen treatments over varying pollinator visitation rates 26

Figure 7. Effect of stamen removal on per-flower seed set 31

Figure 8. Geographic trends in short stamen number 34

Figure 9. Main effects of QTL 36

Figure 10. Epistasis imposes evolutionary constraint on short stamen loss 37

Figure 11. Arabidopsis lyrata flower preserved in alcohol, with linear measurements marked 44

Figure 12. Self-pollinating populations produce more ovules per flower 55

Figure 13. Self-pollinating populations produce less pollen, and short stamen anthers produce more pollen than long stamen anthers 56

Figure 14. Pollen:ovule ratio is lower in selfing populations 57

Figure 15. Greater herkogamy in selfing populations in the field is due to longer pistils 59

Figure 16. Distribution of seed set in stamen removal treatments 74

Figure 17. Main-effect QTL using stepwise analysis on the complete untransformed data with no epistasis in R/qtl 81

x Figure 18. Epistasis results in reduced short stamen loss 82

Figure 19. Epistasis results in fewer effective paths to evolution of stamen loss by natural selection 83

Figure 20. Distribution of mean short stamen production in the recombinant inbred lines 84

xi CHAPTER 1

INTRODUCTION

Background

Biotic diversity is characterized by patterns of both divergence and similarity. In closely related organisms, similarity due to conserved traits can be the result of current adaptation or constrained evolution of an ancestral trait. From compound flowers in the sunflower family to reduced wing number in flies, conserved traits are a ubiquitous feature of life on earth, but it can be challenging to discern whether these traits are adaptive or simply relics of a shared evolutionary past. Both selection and constraint may be operating to conserve a trait across related species, depending on the ecology of the species. My dissertation investigates the roles of selection and constraint in maintaining a family-diagnostic stamen morphology. Flowers in the Brassicaceae (mustard plant family) are characterized by tetradynamy (having four long and two short stamens within a flower) (Zomlefer 1994). Although it is found in >95% of the genera in the family

(Endress 1992), the reason(s) for the widespread conservation of tetradynamy are not known.

Existing adaptive hypotheses for the evolution and the maintenance of tetradynamy address how the trait could increase fitness for outcrossing species.

While there is some evidence that tetradynamy is adaptive in self-incompatible members of the Brassicaceae (e.g. Kudo 2003; Conner et al. 2009), even in these exclusively outcrossing species, the function of tetradynamy is not clear. Although self-incompatibility is widespread in the family, there have been multiple

1 independent losses of self-incompatibility (Lloyd 1965; Mable et al. 2005) followed in some cases by the evolution of high self-pollination rates (Preston 1986). In these highly selfing species, the maintenance of short stamens, which general appear too short to pollinate the stigma within a flower (Müller 1961), is particularly mysterious.

My dissertation unites approaches from evolution, ecology, and genetics to understand the maintenance of short stamens in the Brassicaceae across the full range of mating systems. I investigate the function and morphology of short stamens in three species: the obligately outcrossing wild radish (Raphanus raphanistrum), the highly selfing model plant Arabidopsis thaliana, and A. thaliana’s sister species,

A. lyrata, which includes both outcrossing and selfing populations.

Organization of the dissertation

Chapter 2: In collaboration with Jeffrey Conner and Sam Slowinski, I investigated how the function of tetradynamy changes with frequency of pollinator visitation in the self-incompatible wild radish (Raphanus raphanistrum). We applied stamen removal treatments to plants that were placed in the field on days with a range of pollinator visitation frequencies. This allowed us to separate the effects of short stamens alone, long stamens alone, within-flower dimorphism, and pollinator abundance. We assessed both seeds produced per flower and seeds sired per flower to determine if selection acted differently through male and female fitness. We found no single treatment consistently performed best, but both female and male fitness were significantly affected by interactions between stamen treatment and pollinator visitation rate. Long stamens appear to serve an important role in

2 attracting pollinators when floral visitors are relatively rare. There were trends suggesting a function for short and dimorphic stamens at visitation rates higher than those observed in our experiment.

Chapter 3: In collaboration with Jeffrey Conner and Douglas Schemske, I tested the function of short stamens in the self-pollinating species Arabidopsis thaliana, described the frequency and distribution of short stamen loss in the native range, and investigated the genetic architecture of short stamen loss. We found that short stamens do not significantly increase selfed seed set. Elimination of these apparently nonfunctional organs is possible, as our common garden study showed that many populations, particularly near the southern end of the geographic range, have partially lost the short stamens; this coincides with a previously described cline in genetic variance, with greater variance in the south. Genetic mapping revealed three QTL controlling the number of short stamens, and strong epistasis that reduces the number of effective pathways for stamen loss. These results suggest that evolutionary loss of non-adaptive short stamens in Arabidopsis thaliana is underway, but may be slowed by gene interactions, an inbreeding mating system, and low genetic variance or correlations with traits that vary adaptively with latitude.

Chapter 4: Loss of self-incompatibility in some Great Lakes populations of

Arabidopsis lyrata has led to variation in mating system from self-incompatible to highly selfing. I used this natural variation to test general expectations for floral evolution with mating system shifts as well as specific predictions for changes in investment in short stamens in the Brassicaceae. I explored the evolution of floral

3 morphology with mating system shift by measuring floral morphology and counting pollen grains from two highly selfing and five outcrossing populations. I found that sex allocation shifted as predicted with the evolution of selfing; selfing populations produce less pollen and more ovules per flower, resulting in a decreased pollen:ovule ratio. The decrease in floral size that generally accompanies the evolution of self-pollination is not apparent in A. lyrata. Close inspection of flowers of both selfing and outcrossing populations showed that contact between short anthers and receptive stigmas is extremely rare. This suggests that, as in sister species A. thaliana, short stamens are unlikely to contribute to selfed seed in A. lyrata. However, there were no indications of a reduction in investment in short stamens in selfing populations (no flowers lacking short stamens or greater reduction in pollen production in short stamen anthers relative to long stamen anthers) or relaxed selection on short stamen position (no difference in variance in stamen length between selfing and outcrossing populations).

4 CHAPTER 2

STAMEN FUNCTION IN WILD RADISH DEPENDS ON POLLINATOR VISITATION

RATE

Introduction

The diversity of floral morphology has been of interest to evolutionary biologists since Darwin (Darwin, C.H. 1862, 1876, 1877; Barrett 2010), with many studies aiming to understand the function of these reproductive traits (Harder and

Johnson 2009). Insect pollination may have played a large role in the rapid radiation of angiosperms (Grimaldi 1999; Friedman 2009), with ~87% of the >350,000 extant flowering plant species pollinated by animals (Ollerton et al. 2011). For these species, understanding interactions with pollinators is critical to gaining insight into the evolution of their floral traits.

Though rarely the most conspicuous floral trait, stamens exhibit morphological variation to rival any other floral structure. Two examples with adaptive hypotheses are the precise delivery system offered by pollinia, which evolved independently in orchids and milkweeds, and the reduced pollen waste in poricidal buzz-pollinated anthers, which are found in 65 different angiosperm families (De Luca and Vallejo-Marín 2013). Heteranthery, having different anther forms within a single flower, has also evolved independently multiple times and exists in at least 16 families of flowering plants (Vallejo-Marin et al. 2010). This can take several forms, including feeding versus pollinating anthers that differ in size, color, and position (Vallejo-Marin et al. 2010); stamens of different heights

5 coordinated with styles of different heights (heterostyly)(Barrett 1992); and stamens of different heights with constant style height.

The tetradynamous stamen condition, with a single style height but each flower having four long stamens and two short stamens, is a form of heteranthery diagnostic for the Brassicaceae (the mustard family). The condition is largely conserved across this large and diverse family, and may be maintained by natural selection or constraint. All studies of tetradynamy have been in Raphanus raphanistrum (wild radish) and the closely related Brassica rapa (canola). In these outcrossing members of the mustard family, results indicate that the trait is adaptive (at least in some circumstances) (e.g. Kudo 2003; Conner et al. 2009), but how it functions is unclear. There are three non-mutually exclusive hypotheses for how heteranthery may increase plant fitness: division of labor, slow release of pollen, and trait specialization.

The division of labor hypothesis classifies anthers as “feeding” or

“pollinating” structures, and dates back to Fritz Müller, one of Darwin’s contemporaries (Müller 1883). Under this hypothesis, pollinators forage primarily on short anthers. This positions their bodies to more efficiently pick up pollen from the long anthers and/or deposit it on the stigma. There is solid support for this mechanism in members of the Melastomataceae (Luo et al. 2008)and Solanaceae

(Vallejo-Marin et al. 2009). However, there are several reasons to doubt the applicability of the positioning hypothesis to the Brassicaceae. Poricidal anthers and lack of nectar characterize most heterantherous taxa, but not the Brassicaceae

(Vallejo-Marin et al. 2010). In both species with good evidence of separate feeding

6 and pollinating anthers, the feeding stamens are longer, with anthers that are darker, larger, and produce more pollen (Luo et al. 2008; Vallejo-Marin et al. 2009).

Wild radish differs from this syndrome in many respects. In addition to abundant nectar production and easily accessible pollen with no color difference between anther types, the short rather than long anthers are larger with more pollen, and pollen feeders remove less pollen per visit from short than long anthers (Conner et al. 1995). Because the classic division of labor hypothesis is unlikely to apply to tetradynamy, we focus instead on the slow release and trait specialization hypotheses.

The slow release of pollen hypothesis for the maintenance of tetradynamy posits that both long and short stamens primarily produce pollen for export, but they optimize male fitness at different pollinator visitation rates. Theory shows that quick release of pollen is adaptive when pollinators are rare, whereas slow release of pollen increases the number of seeds sired when pollinator visitation rates are high (Harder and Thomson 1989). Previous work indicates that in R. raphanistrum, fewer pollen grains are removed per visit from short than long stamens (Conner et al. 1995), from experimentally manipulated flowers with both long and short stamens compared to flowers with only long stamens (Conner et al. 2003), and from flowers with naturally greater difference in anther heights. It is clear that tetradynamy slows pollen release; it has not been established whether this mechanism results in increased fitness (although Conner et al. [2003] did find stabilizing selection on dimorphism across three years).

7 Finally, the trait specialization hypothesis posits that traits may be adaptations to a subset of interacting species (Sahli and Conner 2011); in tetradynamy, this could mean that long anthers increase plant fitness when interacting with one pollinator type, and short stamens function best with another pollinator type. Combining selection from these disparate pollinator groups could result in net selection to maintain tetradynamy in the generalist R. raphanistrum.

From a trait specialization perspective, short stamens, which are inserted in the floral tube directly above the nectaries, may be adaptations for nectar-feeding pollinators.

For pollinators that consume pollen, stamens can serve an important role in attracting visitors to the flower (e.g. Lunau 2000; Luo et al. 2008). Syrphid flies, which are common and effective pollinators of R. raphanistrum (Sahli and Conner

2007), have been shown to make more frequent and lengthier visits to Brassica rapa flowers that have more anthers in an anther removal experiment (Golding et al.

1999). In R. raphanistrum, flowers artificially selected to have less dimorphism

(which results in more visible short stamens) receive more and longer visits from bees (Sapir et al, in prep). Pollen-feeding insects make the majority of visits in nearly all studies of R. raphanistrum pollination (Conner et al. 2009), so it is plausible that the role of long stamens in pollinator attraction results in higher plant fitness.

We tested the slow release hypothesis and the role of long stamens in attraction using stamen removal experiments, pollinator observations, and

8 measures of male and female fitness across a range of pollinator visitation rates in wild radish Raphanus raphanistrum, (Brassicaceae).

Methods

Field experiment

In 2009, we grew 104 Raphanus raphanistrum using seed collected from a wild population near Binghamton, NY (Conner and Via 1993). Each plant in our study was grown from a seed collected from a unique field maternal plant. In early

June, we sowed 4-7 seeds in 10cm pots with MetroMix potting soil and ½ tsp

Osmocote Plus 15-9-12 fertilizer (Scott’s Company LLC) in a pollinator-free greenhouse at Kellogg Biological Station (KBS). They were grown with 16-hour,

24°C days and 8-hour, 20°C nights. Seedlings were thinned to one plant per pot at the two-adult-leaf stage.

We randomly assigned 96 plants to one of four arrays of 24 plants each.

Throughout the experiment, the other eight plants (1-3 per array) were used as substitutes for plants without enough flowers. On the morning an array was taken into the field, each plant was assigned one of four stamen treatments in a stratified random design (Figure 1): two short stamens (2S), two long stamens (2L), two long stamens and two short stamens (2L2S), and four long stamens (4L). Plants were reduced to 12 flowers each spaced as evenly as possible across the available display.

There were a few accidental deviations from the intended flower number of 12, which were noted during pollinator observations. All fitness measures were adjusted to a per-flower basis to correct for this small variation in flower number.

Each flower received the plant’s assigned treatment by plucking anthers just below

9

Unmanipulated,R.#raphanistrum,flower, pis2l,

short,stamens, long,stamens,

4L,, 2L2S, 2L, 2S, , , !X! !X! !X! !X! !X! !X! , X! X! X! !X! , !X! !X!

Figure 1. Experimental stamen removal treatments applied to flowers. Long stamens are represented by black circles with white dots, shorts stamens by white circles with black dots. All of the flowers of each plant placed in the field were given one of four stamen removal treatments: Four long (4L; both short stamens removed), 2 long 2 short (2L2S; two long stamens removed), two long (2L; two long and both short stamens removed), or two short (2S; all four long stamens removed).

10 their attachment to the filaments with fine forceps. To control for disruption caused by anther removal, the forceps were also placed briefly on the filament below each anther that was not removed. Flowers were marked with a small piece of colored tape on the pedicel indicating the date and treatment.

Only one array was used on any given field day. Between July 17th and

August 27th, three of the arrays went in the field four times, and one went out three times for a total of 15 field days. Because wild radish flowers last for only two days

(Conner, pers. obs.), and an average of 84% of the pollen was removed from intact flowers in one hour in a previous study (Rush et al. 1995) each flower was only exposed to pollination on a single day. The arrays went out in succession; one array was used per day, after which the plants were returned to the greenhouse for at least seven days to allow any fruits to begin forming undisturbed (mean = 12.92 days, s.d. = 4.56). Arrays were not set up on rainy or windy days, as pollinators are less active under those conditions (Kevan and Baker 1983).

Each array of 24 plants was randomly divided into eight treatment groups of three plants each. Each treatment group cycled through a staggered rotation of the four stamen treatments. Thus, while there were six plants per treatment each day, the plants in each treatment group and the order in which the eight groups experienced treatments varied. This meant that there was no bias among treatment groups due to previous treatments. Each plant in the first three arrays experienced all four stamen removal treatments, except for a few plants that were replaced because they had less that 12 flowers.

11 In the morning, as soon as flower treatments had been applied, the array was placed in an old field with no naturally occurring R. raphanistrum and a variety of abundant pollinator taxa known to visit R. raphanistrum (Sahli and Conner 2007).

Plants were randomly assigned locations 1m apart in a six-by-four-plant grid.

Locations were held constant across all 15 field days; but the placement of plants within these locations was re-randomized on each field day.

We collected data on visits to flowers each day immediately after plants were placed in the field. We aimed to observe each plant every day, but occasionally weather or lack of personnel interfered. We observed 8-24 plants per day (mean

20.47, standard deviation 5.99) for 10 minutes each, giving a total of 3070 minutes of pollinator observations. For each observation, we recorded the number of pollinators that visited the plant and identified them to functional group

(lepidopterans, small bees, large bees, small syrphid flies, large syrphids, and other).

An insect was required to contact stamens and/or stigma on at least one flower to count as a pollinator. Visiting multiple flowers on the same plant consecutively was scored as one visit, but leaving the plant and returning was recorded as two visits.

For a subset of visitors, we also recorded the total number of flowers probed during a plant visig, how long each floral visit lasted, and whether the insect nectared. We collected detailed accounts of visits from 230 individual insects, with 2-43 detailed visits recorded per day (mean 14.33, SD 11.89).

Male and female fitness estimates

Tissue samples were collected from each plant used in the field and frozen at

-80° C for genotyping. All fruits were removed when mature and seeds were

12 counted to estimate female fitness. Seed set was unusually and inexplicably low for

7/17, so that day was excluded from all fitness analyses (pollinator visitation data from 7/17 were included).

To estimate male fitness, we performed paternity analysis on 1256 offspring, with mean 89.7 (SD 5.5) offspring per field day and 3.8 (SD 1.2) offspring per plant per field day. Seeds were sown in 4cm square wells in 72-well trays with MetroMix and two tablespoons of Osmocote per tray. Offspring tissue samples were collected and stored at -80° C. Parent and offspring DNA was extracted using MP Biomedical’s

FastDNA kit and the FastPrep instrument (Carlsbad, CA). They were then genotyped at four microsatellite loci (Bn35d, BRMS005, Na14E08, Ra2E11; radish.plantbiology.msu.edu) using the PCR protocol described in Sahli et al. (2008).

We scored alleles using FMBIO Analysis 8.0 (Hitachi Software Engineering 1991) and binned them with Allelogram 2.2 (Manaster 2010).

Parent genotypes at all four loci were confirmed on at least two independent gels, with one exception. One locus from one parent had persistent errors in gels and was inferred from the offspring using the program GERUD 2.0 (Jones 2005). After excluding one offspring from this maternal plant because of an incompatible genotype, GERUD was able to infer the parental genotype from the remaining 12 offspring.

Paternity analysis was performed using the program Cervus 3.0 (Kalinowski et al. 2007). For the allele frequency analysis, we used a minimum expected frequency of 5 with a Yates correction when df=1 and Bonferroni correction to evaluate significance. We simulated 10,000 offspring genotypes with proportion loci

13 typed set at 0.95 and proportion mistyped at 0.01. Offspring with fewer than two loci genotyped were excluded from the analysis. We were able to determine paternity with 80% confidence for 938 of the 1256 offspring; these 938 were used to estimate male fitness.

Analysis

To test the effects of stamen treatment and pollination visitation rate on fitness, we used ANCOVA including treatment, mean pollination visitation rate for each field day (continuous), and interaction between treatment and pollination visitation rate as fixed effects; plant ID was included as a random effect. Formally, this model is

w ~ β0 + βt + βp + βtxp + ε

2 β0 ~ N(β0i, σ β0)

Where w is fitness (seeds produced or sired per flower, both relativized by the mean across the entire experiment), 0 is the intercept, t is the stamen removal treatment, p is the mean number of pollinator visits per flower per hour, and i is the individual plant. We used Tukey’s Least Square Mean Honestly Significant Difference test to determine differences between treatment means.

The same model was used for male fitness (seeds sired per flower) and female fitness (seeds produced per flower), both of which produced normally distributed residuals. The effect of total stamen number on fitness was tested with

14 the same model, but with treatments categorized as four stamens (2L2S & 4L) or two stamens (2S & 2L). To test the effect of stamen dimorphism, the model was run with only the treatments with four stamens (4L, 2L2S). The effect of long vs. short stamens was isolated by running the model with only the two staminate treatments

(2S, 2L)

To look at whether pollinators visited treatment at different rates, we used

ANCOVA with the same model described above, with w as visits per flower per hour to each plant. Heteroscedastic residuals were ameliorated by square root transforming visitation rate.

All analyses were performed in JMP 10.0.0.

Results

We observed a total of 261 insects visiting flowers. Most of the visits were from flies and bees, with cabbage butterflies making up almost all the rest (Table 1).

Consistent with expectations, treatments with more stamens had higher male fitness (Table 2, Figure 2).

Slow release hypothesis

If the presence of short stamens increases seed siring success by slowing the release of pollen when pollinators are abundant, then the fitness effects of short stamens will depend on pollinator visitation rate. At low rates of visitation, flowers with more long stamens should sire more seeds (2L fathers more seed than 2S; 4L fathers more seed than 2L2S). At high rates of visitation, we expect the opposite, with the dimorphic treatment (2L2S) performing best.

15 resid m no 2 vs. pollinator visitation rate per flower per hour trt stam# 2 4 r2#=)0.02) r2#=)0.008) 1.0 p#=)0.0496# p#=)0.2593#

0.8

0.6

0.4 resid m no 2 0.2

0.0 Male)fitness)(seeds)sired)per)flower))

-0.2

-0.4 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 pollinator visitation rate per flower per hour Pollinator)visita,on)rate)(visits)per)flower)per)hour))

Figure 2. Male fitness of 2 vs. 4-staminate treatments over varying pollinator visitation rates. Fitness relativized by the mean across the entire experiment. Residuals of a model including only Plant ID.

16 Table 1. Insect visitors observed visiting experimental plants. A grand total of 261 insects were observed visiting flowers. They were identified to the lowest taxonomic category possible without disturbing their behavior. Percentages are reported, with raw numbers in parentheses.

Bees Flies Butterflies Other S bee 33 % (86) S syrphid 7% (18) Pieris 11% (28) Wasp 1% (2) Honey bee 5% (13) M syrphid 17% (45) Beetle 2% (6) Megachilid 2% (5) L syrphid 16% (42) Bombus 6% (15) Muscid 0.4% (1) Total 46% (119) Total 40% (106) Total 11% (28) Total 3% (8)

17 Table 2. Effects of stamen treatment and pollinator visitation rate on male and female fitness. Treatments were modeled three different ways for both male and female fitness, designated on the left side of the table in bold: all four treatments (4L, 2L2S, 2L, 2S), treatments with two stamens compared to treatments with four (4L and 2L2S vs. 2L and 2S), dimorphic compared to monomorphic (2L2S vs 4L) and long vs. short (2S vs. 2L). *Effects significant at p < 0.05.

Male fitness Female fitness (seeds sired per (seeds per flower) flower)

All treatments df F p df F p Stamen treatment 3 3.58 0.0145* 3 12.66 <0.0001* Pollinator visitation rate 1 2.46 0.1177 1 0.01 0.9406 Stamen trt*pollinator rate 3 2.71 0.0453* 3 2 0.1136

Two vs. four stamens Stamen treatment 1 4.9 0.0279* 1 35.15 <0.0001* Pollinator visitation rate 1 2.31 0.1291 1 0.003 0.9568 Stamen trt*pollinator rate 1 5.96 0.0152* 1 5.43 0.0204*

2L2S vs. 4L Stamen treatment 1 6.68 0.0119* 1 0.12 0.7329 Pollinator visitation rate 1 0.44 0.5061 1 1.59 0.2091 Stamen trt*pollinator rate 1 0.86 0.355 1 0.03 0.86

Two short vs. 2 long Stamen treatment 1 1.31 0.2557 1 4.81 0.0308* Pollinator visitation rate 1 9.75 0.0027* 1 2.86 0.0926 Stamen trt*pollinator rate 1 4.03 0.0489* 1 0.12 0.7351

18 resid m no 2 vs. pollinator visitation rate per flower per hour Treatment 2L 2L2S 2S 4L 2 2 r2#=)0.02) 1.0 r2#=)0.03) r #=)0.002) r #=)0.02) p#=)0.14# p#=)0.64# p#=)0.21# p#=)0.26#

0.8

0.6

0.4 resid m no 2 0.2

0.0 Male)fitness)(seeds)sired)per)flower))

-0.2

-0.4 0.0 0.4 0.8 1.2 1.6 0.0 0.4 0.8 1.2 1.6 0.0 0.4 0.8 1.2 1.6 0.0 0.4 0.8 1.2 1.6 Pollinator)visita,on)rate)(visits)per)flower)per)hour))pollinator visitation rate per flower per hour

Figure 3. Male fitness of with different stamen treatments over varying pollinator visitation rates. Fitness relativized by the mean across the entire experiment. Residuals of a model including only plant ID.

19 Mean(seeds mothered per flower) vs. Treatment A" 3.5

B" AB" AB" 3.0

2.5

2.0

1.5 seeds mothered per flower

1.0

Female&fitness&(seeds&sired&per&flower)& 0.5

0.0 2L 2L2S 2S 4L Treatment

Figure 4. Female fitness by treatment. Bars indicate 1SEM. Shared letters indicate lack of significant difference according to Tukey HSD test.

20 Contrary to predictions for the slow release hypothesis, we found a significant interaction between treatment and pollinator visitation rate only for the comparison of two vs. four stamens, and there was no difference between male fitness for 2L2S and 4L (Table 2, Figure 3). Stamen treatment did have significant effects on female fitness, suggesting that the slow release hypothesis is not acting alone, if at all (Table 2, Figure 4).

Trait specialization – long stamen attraction hypothesis

If stamens increase fitness by attracting pollinators to the flowers, we expect higher visitation rates to treatments with more stamens (more visits to 2L2S & 4L than 2S and 2L). Because long stamens are more visible and accessible, we also would expect higher visitation rates to treatments with long than short stamens

(more visits to 4L than 2L2S, and more visits to 2L than 2S.)

These predictions should result in matching patterns for male and female fitness at low visitation, that is, higher fitness in treatments with higher visitation. If pollination is sufficient to fertilize all ovules, we expect an asymptote of seeds per fruit at high visitation. Male fitness at high visitation is likely to be dictated by gamete number, so treatments with four stamens should outperform those with two. Because short stamens have more pollen, our detailed prediction for male fitness at high visitation would be

2L < 2S << 4L < 2L2S.

We found no main effect of stamen removal treatment on pollinator visitation rate to flowers of each treatment (Table 3). This suggests a lack of support for the attraction hypothesis. However, there is a significant interaction with overall

21 Table 3. Effects of stamen treatment and overall pollinator visitation rate on visits to individual plants. Overall pollinator visitation rate is the average across all plants on a give field day; visits to individual plants were always strongly influenced by this. Visit number per flower per hour, square root transformed. Model was run with and without the highest pollinator visitation day. *Effects significant at p < 0.05. High visitation day All days included excluded

All treatments df F p df F p Stamen treatment 3 0.45 0.7191 3 0.36 0.7848 Pollinator visitation rate 1 91.94 <0.0001* 1 66.07 <0.0001* Stamen trt*pollinator rate 3 2.00 0.1137 3 0.96 0.4101

Two vs. four stamens Stamen treatment 1 1.08 0.2992 1 1.15 0.2843 Pollinator visitation rate 1 92.55 <0.0001* 1 68.19 <0.0001* Stamen trt*pollinator rate 1 0.12 0.7294 1 0.56 0.4532

2L2S vs. 4L Stamen treatment 1 0.005 0.9448 1 0.02 0.9016 Pollinator visitation rate 1 42.30 <0.0001* 1 32.28 <0.0001* Stamen trt*pollinator rate 1 0.08 0.7741 1 0.67 0.4151

Two short vs. 2 long Stamen treatment 1 0.31 0.5798 1 0.0139 0.9065 Pollinator visitation rate 1 41.35 <0.0001* 1 27.31 <0.0001* Stamen trt*pollinator rate 1 5.03 0.0264* 1 0.80 0.3735

22 visitation across days for the 2L vs 2S treatment comparison (Table 3); as overall visitation rate increases, visits to the 2S treatment increase faster than visits to the

2L treatment (Figure 5). Although it is not significant, the same trend exists in the

2L2S vs. 4L comparison, with higher visitation to the dimorphic than the monomorphic treatment at high overall visitation rates (Figure 5). Both patterns are consistent with long stamens attracting pollinators at low visitation frequencies.

There were considerably more visitors per flower on our highest visitation day than any other, so we also tried eliminating it from the analyses. When we do this, the significant interactions between per-treatment visitation and overall visitation disappear (Table 3).

Discussion

Support for the slow-release hypothesis was weak. This may be because it is predicted to operate at high pollinator visitation, but the visitation rates we observed are low for R. raphanistrum (0.6 – 3 visits per hour, Conner, unpublished data; 2-6 visits per hour, Rush et al [1995]; up to 9.5 visits per hour, Young and

Stanton [1990]). Because it predicts no effect on female fitness but female fitness effects were found, we can conclude that it certainly does not operate alone. It is also possible that the slow-release model does not apply to R. raphanistrum; while

Thomson and Harder (1989) does not include simultaneous vs. staggered flower opening in the model, later work incorporating variation in stigma presentation time shows that when flowers open simultaneously, rapid release of pollen is adaptive at high pollinator visitation (Stanton 1994). Wild radish often has a few flowers open throughout the day, but most open together in the early morning

23 square root # visits vs. day means poll visit per flower per hour Treatment 2L *% 2L2S 2S *% 4L 2.5

2.0

1.5

1.0 square root # visits

0.5 %Square%root%visits%per%plant%(per%flower%per%hour)% 0.0

0.0 0.4 0.8 1.2 1.6 0.0 0.4 0.8 1.2 1.6 0.0 0.4 0.8 1.2 1.6 0.0 0.4 0.8 1.2 1.6 Mean%pollinator%visits%to%en.re%array%(per%flower%per%hour)%day means poll visit per flower per hour

Figure 5. Square-root transformed pollinator visitation rates of different treatments over varying overall pollinator visitation. Slopes of the 2L and 2S treatments are significantly different from each other (Table 3); the only other pairwise comparison made, 2L2S vs 4L, was not significant.

24 (Stanton 1994, personal observation), throwing the relevance of the slow-release hypothesis into question. Regardless, experiments at higher pollinator visitation rates would be necessary to convincingly test this hypothesis.

Rather than playing a consistent role in pollinator attraction, the contribution of stamens to bringing floral visitors in depends on overall visitation rate, with trends of treatments with more long stamens garnering more visits when fewer pollinators were around. Both male and female fitness of the two-staminate treatments increased with more visitation, suggesting that at low visitation pollinators may be more likely to visit the more rewarding four-staminate treatments. Female fitness in the 2L treatment increased more rapidly with higher visitation rates than in the 2S treatment (Figure 6), which would be consistent with the more apparent, and more easily accessed, long stamens attracting pollinators. At these low pollinator visitation rates, attraction may be one of the primary functions of anthers; at higher rates, the slow-release function could also come into play.

Having more flowers increases pollinator visitation in radish (Conner and

Rush 1996), so one likely reason for the low visitation rates we observed is reduced display. The labor-intensive stamen removal treatments meant we could only include 12 flowers per plant, whereas unmanipulated wild radish can easily produce over 50 flowers at a time, and sometimes up to several hundred (Conner and Rush

1996). Having relatively few plants would have exacerbated this. Because the arrays were only exposed to pollinators for a few hours at a time, per flower total visitation was also lower than normal; in the wild, flowers are generally open for two days

(Young and Stanton 1990).

25 resid f no 2 vs. pollinator visitation rate per flower per hour Treatment 2L 2L2S 2S 4L 5 r2#=)0.06) r2#=)0.005) r2#=)0.03) r2#=)0.05) p#=)0.04*# p#=)0.51# p#=)0.10# p#=)0.04*# 4

3

2

1 resid f no 2 0

-1

-2 Female)fitness)(seeds)produced)per)flower))

-3

0.0 0.4 0.8 1.2 1.6 0.0 0.4 0.8 1.2 1.6 0.0 0.4 0.8 1.2 1.6 0.0 0.4 0.8 1.2 1.6 pollinator visitation rate per flower per hour Pollinator)visita,on)rate)(visits)per)flower)per)hour))

Figure 6. Female fitness of with different stamen treatments over varying pollinator visitation rates. Fitness relativized by the mean across the entire experiment. Residuals of a model including only plant ID.

26 Although visitation rates this low would not be expected over an entire season, cold or rainy weather could easily result in low visitation for a subset of the flowers produced in a year. Certainly some flowers will experience a low-visitation environment, and our results give insight into how stamens function on those days.

It makes sense for the plants to evolve strategies that work across a range of visitation rates. Normal, high visitation rates to R. raphanistrum result in rapid depletion of pollen (Young and Stanton 1990; Rush et al. 1995). Under these circumstances, we expect seed set to not be pollen-limited (Burd 1994).

Accordingly, variation in floral morphology has generally been found to have no effect on female fitness (Stanton et al. 1986; Young and Stanton 1990; Conner et al.

1996), although Conner et al (1996) did find selection through female fitness to increase flower size in one year, and Pfennig and Conner (1997) found weak pollen limitation of fruit set. Because variation in visitation rate clearly influences how stamens function, and most studies have studied effects at high visitation rates, this study fills an important gap in our understanding of how variation in interspecific interactions affects adaptation.

Optimal foraging theory, which makes predictions about how animals change their feeding behavior to maximize net energy gain (Schoener 1971; Sih and

Christensen 2001), may offer an explanation for increasing visitation to flowers with fewer stamens at high overall visitation rates. When pollinators are few, pollen will be more abundant; this low-competition environment is thus analogous to high

“prey” abundance. Optimal foraging models predict predators will prioritize the most profitable prey under these circumstances (i.e., pollinators will visit flowers

27 with the most pollen.) As prey becomes less abundant (in this case, because more foraging pollinators are around), pollinators are expected to begin spending more time visiting less profitable flowers (i.e., ones with fewer stamens) (Schoener 1971;

Sih and Christensen 2001). This certainly applies to pollinators; for example, it has been shown that bumblebees expand foraging to less-rewarding plant species when the bees are at high density (Fontaine et al. 2008). When visiting wild radish, honeybees change their floral size preference with variation in visitation rate, shifting to visit smaller (presumably less-visited) flowers when more bees are around (Young and Stanton 1990).

The comparison between the 2L2S and 4L treatments, or the dimorphic and monomorphic treatments, is arguably the most relevant for understanding the evolution of tetradynamy from the ancestral monomorphic condition. Female fitness was the only fitness component that differed significantly between these two treatments, and the result seems nonsensical (relative fitness for the ancestral 4L treatment decreased with increasing visitation). The trends across all three fitness components support 2L2S outperforming 4L at high visitation – slightly more visits, higher female fitness, slightly higher male fitness.

It is clear that pollinator visitation rate matters for understanding the function of tetradynamy. Our data suggest that at the low pollinator visitation rates that occurred throughout this study, stamens function to attract pollinators. Male fitness is the most likely fitness component to show significant effects on fitness at high visitation, and the slow-release hypothesis may operate under those conditions. Most of our observations were made immediately after plants had been

28 placed in the field, so it is also possible that pollinator behavior changed throughout the day as pollen was depleted. This could explain some of the lack of agreement between pollinator visitation and plant fitness data. In the future, experiments with higher visitation rates, longer plant exposure, and observations spread over the exposure time may yield clearer insights into the function of tetradynamy and how it interacts with pollinator visitation rate.

Acknowledgements

Cindy Mills helped with greenhouse and molecular work. Many undergraduates assisted with application of stamen removal treatments and setting up field arrays. This work was supported by grants from the Kellogg Biological

Station’s G.H. Lauff fund and T. Wayne & Katherine Porter fund.

29 CHAPTER 3

ONGOING LOSS OF A CONSERVED TRAIT: LACK OF FUNCTION, LATITUDINAL

PATTERNS, AND GENETIC CONSTRAINTS

Introduction

While evolutionary biologists often focus on adaptive traits, Darwin recognized that traits being diminished or lost under relaxed selection is also a dominant theme in evolution (Darwin, C.H. 1859). This often occurs after a shift in ecology renders formerly adaptive traits nonfunctional (Fong et al. 1995), such as during colonization of a new environment (e.g. loss of eyes in cave dwelling fish

(Yoshizawa et al. 2012) and armor in freshwater sticklebacks (Le Rouzic et al.

2011)). However, retention of nonfunctional traits is also common (Lahti et al.

2009), including feeding structures in larvae with no digestive tract (Pernet 2003), retinal circadian rhythms in blind cave fish (Espinasa and Jeffery 2006), and anti- rattlesnake behaviors in ground squirrel populations isolated from the predator for

>70,000 years (long enough to lose venom resistance) (Coss 1999). Natural selection acts to reduce or eliminate costly traits, but a loss of function alone does not make trait elimination inevitable. Loss of a costly nonfunctional trait could be slowed by low genetic variance or epistasis (Mezey and Houle 2005; Futuyma

2010), or nonfunctional traits that require few resources may degrade only slowly through mutation accumulation (Lahti et al. 2009). The best-understood examples of nonfunctional trait evolution, cavefish eyes and freshwater stickleback armor, both find trait loss is accelerated by selection (Le Rouzic et al. 2011; Yoshizawa et al.

30 A)!

B)! 35# B# 30# B#

25#

20# A# A# 15# median#0# median#0# median#26# median#30.5# 10#

5# Mean'seed'produc0on'per'fruit' 0#

!5#

none################short#only###########long#only################all# Stamen'treatment'

Figure 7. Effect of stamen removal on per-flower seed set. A) A. thaliana flower showing the pollen from the short stamen deposited on the side of the pistil rather than the stigma. (Source: Jürgen Berger/Max Planck Institute for Developmental Biology, Tübingen, Germany). B) Least square means ± 2SEM, from a model including stamen removal treatment, with population and plant nested in population as random effects. Shared letters above the bars indicate lack of significant difference according to a Tukey HSD test. There were 32-33 flowers per treatment.

31

2012). Understanding the mechanisms maintaining nonfunctional traits requires studies of similar depth in systems where evolutionary loss is slowed or stopped.

We present an example of ongoing and constrained loss of a nonfunctional trait in a model organism: the shift to self-pollination in Arabidopsis thaliana has caused some of the plant’s stamens to lose their pollination function. We demonstrate the lack of function of the trait, report a latitudinal cline of partial loss in the native range, and describe how the genetic architecture may constrain loss.

The combination of a trait that is easy to manipulate and a model organism with abundant genetic resources has allowed us an unusually detailed glimpse of trait loss in progress.

Methods

Arabidopsis thaliana is a member of the Brassicaceae, a family in which almost all of the >3700 species have flowers with four long stamens and two short stamens

(Figure 7a). Adaptive hypotheses for the production of short stamens in outcrossing species have some support (Conner et al. 2003), but A. thaliana is highly (97-99%) self-pollinating (e.g. (Abbott and Gomes 1989; Platt et al. 2010)). Because the anthers of short stamens remain below the level of the stigma throughout floral development (Müller 1961; Smyth et al. 1990), they seem unlikely to deposit pollen on the stigma and thus contribute to selfed seed set (Figure 7a).

To test whether short stamens in A. thaliana are functional, we compared seed set across four stamen treatments: all six removed, only long or only short stamens removed, or all stamens intact but buds probed with forceps to control for

32 stamen manipulations (Figure 7b). Each treatment was performed on at least one bud per plant, with all four treatments performed on adjacent buds. Plants from six accessions from Finland, Germany, Kazakhstan, and Sweden were used (Table 7). If long stamens alone are responsible for seed set in A. thaliana, we would expect seed set in intact flowers and those with short stamens removed to be equally high.

Conversely, we would expect seed set in completely emasculated flowers and those with long stamens removed to be equally low. The results supported these predictions (Figure 7b); short stamens do not contribute significantly to seed set.

A trait that loses its primary function may still be adaptive if it increases the organism’s fitness in some other way (Lahti et al. 2009). “Staminodes” (stamens that have evolved loss of pollination function, usually producing no pollen) with alternative functions are widespread in angiosperms, occurring in nearly a third of families, but universally have roles in facilitating outcrossing via insect pollinators

(Walker-Larsen and Harder 2000). For the selfing A. thaliana, we might expect selection to reduce investment in ineffective short stamens, reallocating resources such as lipids and proteins abundant in pollen (Evans et al. 1991) to other functions.

One way to do that is by eliminating short stamens altogether.

Results

To determine if short stamen loss is common in A. thaliana and to test for geographic trends, we sampled flowers from ten plants from each of 45 populations of A. thaliana chosen as evenly as possible from latitudinal and longitudinal bands across the native range (30 populations in each of the bands, with 15 shared between the longitudinal and latitudinal sample). Up to five matrilines per

33 2" A)! 1.8" 1.6"

1.4"

1.2"

1" r"="0.65" 0.8" p"<"0.0001" 0.6" " 0.4"

Popula'on)Mean)Short)Stamen)Produc'on) 0.2"

0" 35" 40" 45" 50" 55" 60" 65" La'tude)(°N))

B)! 2" 1.8"

1.6"

1.4"

1.2"

1"

0.8" r"=")0.33" 0.6" p#="0.08# 0.4" "

0.2" Popula'on)Mean)Short)Stamen)Produc'on) 0" )10" 0" 10" 20" 30" 40" 50" 60" 70" 80" Longitude)(°E))

Figure 8. Geographic trends in short stamen number. Correlations between stamen number and a) latitude, or b) longitude. Each point is a population mean. Removal of the longitudinal outlier at 73.1°E increases the correlation between stamen number and longitude to r = -0.39, p = 0.04.

34 population were grown (mean = 2.67), for a total of 119 matrilines (Table 7). One flower was sampled from each plant every 2-3 days for an average of 6 flowers per plant (2570 total flowers), with 13-117 flowers sampled per population (mean 56.6,

SD 22.0). We examined the correlations between short stamen loss and latitude and longitude, using the subset of 30 populations for each. Population mean short stamen production varied from 0.49 to 1.96 across the native range of Arabidopsis thaliana. Individual plants that exhibited short stamen loss typically produced three different flower morphs, with zero, one, or two short stamens. Interestingly, population mean stamen number is positively correlated with latitude (Figure 8a).

The correlation with longitude, with stamen production decreasing toward the east, is much weaker and not statistically significant (Figure 8b). Short stamen loss is most advanced in the south, but is incomplete even in populations where it is most common.

To understand the genetics of short stamen loss, we mapped quantitative trait loci (QTL) involved in loss using a set of recombinant inbred lines (RILs) produced using parents from the latitudinal extremes of A. thaliana’s native range,

Sweden and Italy (Agren et al. 2013). Each RIL was genotyped at 348 SNP markers spaced every ~1 cM across the genome. Short stamen number was scored on one to seven flowers (mean = 3) from one to seven (mean = 4.78) plants from each of 519

RILs, for a total of 7435 flowers, plus 502 flowers from the southern and 224 from the northern parents (Table 9). The southern parent involved in the mapping population produced about one short stamen on average (range 0 to 2, mean =

0.93), whereas the northern parent nearly always produced two short stamens

35

2$ PVE$=$28.0! PVE$=$7.0! PVE$=$2.7!

1.9$

1.8$

1.7$

1.6$ Short$Stamen$Number$ $ 1.5$ Swedish$allele$(A,B,C)$ Italian$allele$(a,b,c)$

1.4$ 0.5$ QTL$A$1$ 1.5$ QTL$B$2$ 2.5$ QTL$C$3$ 3.5$ (chromosome$5)$ (chromosome$3)$ (chromosome$1)$

Figure 9. Main effects of QTL. Results from stepwise analysis with full, untransformed data in R/qtl. QTL arranged in order of effect size/percent variance explained (PVE). Bars indicate 2 SEM. All three are significant at p < 0.0001.

36 2.0 A B C phenotype additive 1.8 prediction

1.6

1.4 er b

num 1.2

ABC ABc Abc abc ABC ABc aBc abc ABC aBC aBc abc amen t 2.0 s D E F

Short 1.8

1.6

1.4

1.2

Abc abC ABC AbC abc ABC AbC abc ABC aBC abC abc Genotypes - possible sequences of loss allele accumulation

Figure 10. Epistasis imposes evolutionary constraint on short stamen loss. All possible sequences of accumulating loss alleles are shown, illustrating that epistasis slows loss of short stamens relative to the additive prediction. Dashes show the additive prediction, diamonds show the actual phenotype. Lowercase letters indicate the Italian genotype; uppercase the Swedish genotype. Arrows show which QTL is transitioning from Swedish to Italian in each step. Steps that result in a significant difference according to Tukey’s HSD test are indicated by a black arrow; white arrows indicate non-significant steps. In five of the six intermediate genotypes between the all-Swedish (ABC) and the all-Italian (abc), the additive expectation for short stamen number is lower than the actual phenotype. In four of the six scenarios (a, b, d, and e), the first step results in little or no short stamen loss; in only one scenario (c) does each step produce a significant drop in short stamen number. The additive prediction was made by subtracting the main effect sizes (difference between Swedish and Italian in Figure 9) from the mean short stamen production with the Swedish genotype at all 3 QTL (ABC). Bars indicate 2 SEM.

37 (range 0 to 2, mean 1.96). The RILs were grown in growth chambers mimicking conditions in Italy; the parent phenotypes did not change significantly in a similar experiment mimicking Swedish temperatures.

We found three significant main-effect QTL for short stamen loss (one each on chromosomes 1, 3, and 5; Figures 9, 17), one significant epistatic interaction, and a second nearly significant epistatic interaction (Figures 10 and 18, Tables 10 and

11). In both interactions, epistasis decreased short stamen loss; the stamen-loss alleles on chromosomes 1 and 3 only have a significant phenotypic effect in the presence of the stamen-loss allele on chromosome 5 (Figures 10 and 19). This pattern of epistasis would slow stamen loss by decreasing the loss produced relative to the additive prediction and limiting the number of paths with significant stamen loss at each allele replacement (Figures 10, 19).

Temporal differences in development, with long stamens initiating at the beginning of stage 5 of floral development and short stamens initiating later (Smyth et al. 1990), create an opportunity for heterochrony in gene expression to cause loss of the short stamens. We identified 1388 genes expressed in stamens under the

Bayes 95% credible interval for each QTL (Table 10, 12), including 25 candidate genes with published evidence of involvement in stamen development (Table 13).

Of the candidates, 21 are associated with the jasmonic acid (JA) and/or gibberellic acid (GA) pathways (Table 13). These two pathways are known to interact in stamen development (reviewed in Alvarez-Buylla et al. 2010 p. -), and changes in their regulation could be the mechanism of epistasis.

38 Discussion

Given the lack of short stamen function we describe, evolutionary elimination of the trait is predicted. The fact that loss is universally incomplete suggests that the process of elimination either started recently or is constrained. The pattern of epistasis between stamen-loss loci reduces the phenotypic effect of single loci alone

(Figure 10), while high homozygosity resulting from selfing reduces the effective recombination rate, slowing the rate at which stamen-loss alleles at different loci are brought together. Thus the shift to selfing that made short stamens nonfunctional may be slowing their loss.

The latitudinal cline in short stamen production could be created by differential selection or constraint. Direct selection favoring short stamen production in the north seems unlikely, because short stamens are predicted to increase fitness only when outcrossing rates and pollinator visitation are high

(Harder and Thomson 1989; Conner et al. 2003). Small pockets of increased outcrossing have been reported in the species (e.g. 14.5% (Bomblies et al.)), but even these rates are too low to suggest high pollinator visitation; direct observations of pollinators suggest that 7% or fewer of all flowers are visited

(Hoffmann et al. 2003; Lundemo 2010).

Potential constraint creating the latitudinal cline could be due to indirect selection and/or low genetic variance. Many traits vary adaptively over latitude (e.g. flowering time (Stinchcombe et al. 2004) and cold tolerance (Zhen and Ungerer

2008) in A. thaliana). A genetic correlation with such a trait could constrain stamen loss in the north. Postglacial population dynamics are another possible source of

39 constraint; as with many species (Comes and Kadereit 1998), A. thaliana shows reduced genetic variation in the north consistent with founder events during postglacial recolonization (Beck et al. 2008; Lewandowska-Sabat et al. 2010).

Recent results of a reciprocal transplant conducted with the same set of A. thaliana

RILs found southern alleles at several fitness QTL were adaptive in the north, but not vice versa (Agren et al. 2013). Although none of the QTL identified as involved in latitudinal adaptation are near the stamen loss QTL, it confirms that constraint in the north is found in adaptive traits for A. thaliana.

Although neutral loss through the accumulation of mutations is possible, the latitudinal cline is opposite what would be expected under that scenario. Founder effects can accelerate loss of a neutral vestigial trait through drift (e.g.(Daehler and

Strong 1997)) , so we would expect at least some populations in the northern part of the range to lose short stamens more quickly.

We conclude that evolutionary loss of nonfunctional short stamens is currently underway in Arabidopsis thaliana. Loss is widespread but incomplete, and may be constrained by epistatic genetic architecture and high selfing rates. The latitudinal cline in stamen loss is more likely due to a cline in genetic variance and/or genetic correlations with locally adapted traits rather than latitudinal differences in direct selection on short stamens.

Acknowledgements: We thank James Beck for providing seeds, Chris Oakley for assistance with R/qtl, Emily Dittmar for coordinating the RIL growout, Mark

Hammond and Cindy Mills for lab assistance, Mike Grillo for advice on identifying candidate genes, Raffica LaRosa for assistance in the lab and editing, Colin Kremer

40 for assistance with R, and Sam Perez and many undergrads for phenotyping thousands of flowers. Keith Karoly planted the seed of this project by mentioning that A. thaliana should have lost the short stamens. This manuscript was improved by comments from Ian Dworkin. Supported by grants from the Kellogg Biological

Station’s G.H. Lauff fund to A.M. Royer, NSF DEB 1022202 to D. Schemske and NSF

DEB 0919452 to J.K. Conner.

41 CHAPTER 4

EARLY EVOLUTION OF SELFING IN ARABIDOPSIS LYRATA INCLUDES CHANGES IN

SEX ALLOCATION BUT NOT FLOWER SIZE

Introduction

The evolution of self-pollination from outcrossing, one of the most common transitions in flowering plants (Stebbins 1957; Barrett 2002), is associated with a suite of trait changes known the “selfing syndrome” (Darwin, C.H. 1876; Ornduff

1969; Sicard and Lenhard 2011). A switch to self-pollination generally involves shifting resources away from attracting pollinators and increasing investment in female fitness compared to male fitness (Charlesworth and Charlesworth 1981). It includes less nectar, less scent, smaller flowers, reduced herkogamy (distance between anthers and stigma) and a diminished pollen-to-ovule (P/O) ratio relative to outcrossing plants (e.g. Lloyd 1965; Ritland and Ritland 1989; Goodwillie and

Ness 2005). The replicated evolution of this selfing syndrome offers opportunities to study the evolutionary and ecological basis of these trait shifts.

The selfing syndrome does not necessarily evolve in the same way in every case; ovule counts and the evolutionary timing of flower reduction are two examples. While a reduction in pollen production is largely responsible for the lower P/O ratio in selfing plants (Darwin, C.H. 1876; Ornduff 1969), changes in ovule number are variable. Ovule production increases in many selfing taxa along with decreases in pollen count (Mione and Anderson 1992; Delesalle et al. 2008;

Mazer et al. 2009). However, ovule counts may also decrease, reducing the effect of pollen decrease on the P/O ratio (Ritland and Ritland 1989). Similarly, while small

42 flowers are a constant feature of the selfing syndrome, they may evolve together with the shift in mating system (as a reduction in overall flower size decreases herkogamy) (Guerrant 1989; Elle and Carney 2003), or gradually after selfing becomes established because selection favors reduced investment in showy flowers

(Charnov 1982; Bodbyl Roels and Kelly 2011).

There have been many independent transitions to self-pollination in the mustard family (Brassicaceae) (e.g. Lloyd 1965; Mable et al. 2005; Foxe et al. 2009).

These species display the selfing syndrome described above, but possess an additional family-specific trait that may be affected by a mating system shift. The

Brassicaceae are characterized by tetradynamous stamens: four long and two short stamens within a flower (Figure 11)(Endress 1992; Zomlefer 1994). Although the reason for widespread conservation of this trait is not known, all of the current adaptive hypotheses center on a function for outcrossing (Golding et al. 1999;

Conner et al. 2003, 2009; Kudo 2003) . In the model mustard Arabidopsis thaliana, short stamens are too short to reach the stigma (Müller 1961) and do not contribute significantly to selfed seed (Chapter 3). Therefore, we might expect that for mustards, the selfing syndrome would include a greater reduction in investment in short stamens relative to long stamens, and relaxed selection on short stamen length due to loss of function.

Arabidopsis lyrata (Brassicaceae), A. thaliana’s sister species, is an excellent system for testing how the investment in short stamens changes during mating system shifts. Throughout most of its circumpolar distribution, A. lyrata exhibits the sporophytic self-incompatibility typical of the mustard family (Mable et al. 2003). In

43

Figure 11. Arabidopsis lyrata flower preserved in alcohol, with linear measurements marked. 1) Petal width, 2) petal length, 3) sepal height, 4) sepal length, 5) pistil height, 6) long anther length, 7) long stamen height, 8) short anther length, 9) short stamen height, 10). Calipers set at 2mm for scale. The horizontal line was used as a common baseline for height measurements; it was set at the short stamen attachment point, perpendicular to the pistil. Lines were placed and measurements made with ImageJ.

44 Table 4. Arabidopsis lyrata populations and sampling scheme. The Saugatuck population was only grown in the greenhouse. A. OR = outcrossing rate, %SC = percentage of the population that is self-compatible. Mating system estimates for IDU, PIN, RON, and BP are from Mable et al (2005); estimates for BP, PIN, and PP are from Mable et al (2007). Both studies used controlled self-pollinations to estimate self-compatibility and multilocus microsatellite genotypes of individuals grown from field-collected seed to estimate outcrossing rates. MS = binned population mating system, outcrossing or selfing; Field = number of plants sampled from the field common garden; GH = number of plants sampled from the greenhouse common garden. B. Mean (sd) sample sizes in the common garden experiments, not including flowers sampled for pollen counts. C. Number of plants in the greenhouse study that were the result of self-pollinations (self), outcrossing after early emasculation (outcross [early]), and outcrossing after late emasculation (outcross[late]). The plants sampled for populations not included in this table were all the result of outcrossing after late emasculation.

A. Population OR %SC MS Field GH Ludington State Park, MI (LUD) 1.0 .07 Out 56 28 Saugatuck State Park, MI (SAU) .97 .24 Out 21 Indiana Dunes National Lakeshore, IN (IDU) .91 .31 Out 53 66 Bruce Peninsula National Park, ON (BP) .88 .03 Out 40 51 Pinery Provincial Park, ON (PIN) .84 .06 Out 46 24 Rondeau Provincial Park, ON (RON) .21 .97 Self 53 46 Point Pelee National Park, ON (PP) .02 1 Self 52 33

B. Field GH Populations 6 7 Maternal lines per population 7.67 (0.82) 5.43 (1.72) Plants per maternal line 6.5 (1.15) 6.16 (1.87) Plants per population 50 (5.9) 38.4 (16.5) Flowers per plant 1 2.7 (0.6) Plants 299 269 Flowers 299 723 C.

Population Self Outcross (early) Outcross (late) PP 10 6 17 RON 8 17 21 BP NA 17 17 IDU NA 21 72

45 the Great Lakes region of North America, however, some populations have evolved a loss of self-incompatibility; populations vary in frequency of self-compatibility and show roughly corresponding inbreeding rates (Mable et al. 2005; Mable and Adam

2007; Willi and Maattanen 2010) (e.g. Table 4). Populations with high frequency of self-compatibility have evolved autogamous selfing, the ability to pollinate within a flower without the assistance of pollinators (Mable et al. 2005). This range of mating systems in natural populations presents an opportunity to examine the evolution of the selfing syndrome at the earliest possible stage.

This study investigates how A. lyrata floral morphology is evolving in response to shifts in mating system. I first assessed whether short stamens contact the stigma in one self-incompatible and one highly-selfing population of A. lyrata. I then quantified differences in floral morphology between selfing and outcrossing populations using two common garden studies (one in the greenhouse, one in the field). I measured the number of pollen grains and ovules produced, width of petals, and length of petals, sepals, pistils, stamens (long and short), and anthers (long and short). These measures were used to estimate flower size, herkogamy, and allocation to male and female reproduction in seven populations that span the range from self-incompatible to highly selfing. The predictions are divided into general ones, which would apply to any group of flowering plants, and specific ones dealing with tetradynamy.

My general predictions are that selfing populations will exhibit decreased

P/O ratio, decreased flower size, petal reduction that is stronger that the reduction in overall flower size, and decreased anther-stigma herkogamy. The mustard-

46 specific predictions are a greater reduction in short-stamen pollen grain numbers compared to long-stamen, and increased variance in short stamen length due to relaxed selection on short anther position.

Methods

To assess function of short stamens in Arabidopsis lyrata, I sampled one healthy inflorescence from nine individuals of a highly selfing population (Rondeau, which has been observed to produce autogamously selfed seed (Mable et al. 2005)) and nine individuals of a largely self-incompatible population (Saugatuck). Seeds collected from the populations were sown directly into 4” square pots with

MetroMix potting soil and ½ tsp Osmocote Plus 15-9-12 fertilizer (Scott’s Company

LLC). They were grown in the greenhouse at Michigan State University (MSU)’s

Kellogg Biological Station (KBS) in Hickory Corners, MI.

I dissected each flower or bud on the inflorescence from the most mature

(lowest) flower that still retained its anthers to one unopened bud beyond the last one that had a dehisced anther. A total of 58 flowers were sampled (mean 3.2, SD 1.0 per plant). For each one, I assessed whether the stigma appeared to be receptive

(papillae visible and appear damp), whether the long and short anthers were dehisced, and whether they were contacting the stigma (i.e., whether they were capable of autogamous selfing).

To examine changes in floral morphology with shifts in mating system, flowers were collected from two highly selfing and five primarily outcrossing Great

Lakes populations of A. lyrata grown in two randomized common garden experiments (Table 4), allowing me to evaluate evolved differences in outcrossing

47 vs. selfing populations. In 2012 I sampled six populations in a field common garden at KBS. In 2013 I sampled seven populations (the same six as 2012, with one additional population) in a greenhouse common garden experiment on MSU’s main campus in East Lansing, Michigan.

Field common garden

Germination of seeds for the 2012 field common garden was begun on

9/9/11. Field-collected seeds were sterilized in 30% bleach solution with 0.001%

Triton X 100 surfactant for ten minutes, washed twice with sterilized water, suspended in 0.1% agar, and sown on a 100mm X 15 mm petri dish with agar

(Phytoblend, Caisson Laboratories Inc.). Petri dishes were filled with 20mL of 0.8%

Agar with 2.31% Gamborg’s B-5 basal medium and autoclaved after the media set.

Sown plates were wrapped in Parafilm and stratified for four days at 6.5°C. Six of the

38 plates were contaminated, and those seeds were transferred to soil before germination. Stratified seeds were incubated at 22°C with 16 hours days at 100 PAR for 4-8 days. Seedlings were then transplanted into 2 inch square pots filled with soil-less potting media (Suremix Perlite, Michigan Grower Products Inc.) The seedlings were grown in chambers at 22°C with 16 hours days at 100 PAR for 10-14 days. On October 7th, they were transported to KBS, where they were moved daily between the field and greenhouse to harden off for 11 days.

The transplant site was field soil covered with a layer of black weed barrier, with 9 cm diameter circular holes every 30.5 cm in a grid oriented to the cardinal directions. A soil plug corer (Sod Plugger, Hummert International) was used to create a hole for each plant by removing field soil (round plug of 7.5 cm diameter

48 and 7.5 cm depth). The soil was saved, passed through a soil shredder, and mixed to create a homogeneous, well-aerated soil used for backfilling during transplanting.

The seedlings were transplanted into 6 rows, with rows oriented East/West, on

October 18th. Each plant was given a unique, randomly assigned ID number and location. Seedling deaths within a week of transplant were attributed to transplant shock and replaced with extra plants that had been kept outside in pots.

Of the plants that survived the winter and flowered in 2012, I randomly selected two plants per maternal line for full fruit and flower counting and sampled all the remaining plants (Table 4B). A single fully opened flower, the most recently opened on its inflorescence, was collected from each of the remaining plants (299 plants, mean 50 per population [Table 4]) on two consecutive days (5/10-11).

Flowers were placed in 1ml microcentrifuge tubes which were then filled with 75% ethanol to preserve the flowers.

Greenhouse common garden

The 2013 greenhouse common garden study included seven populations – the same six as the 2012 study, with one additional outcrossing population. Plants for this study were grown from seeds resulting from controlled crosses within populations, both self-pollination and crosses between different individuals within a population. Seeds were produced by first emasculating flowers, then pollinating when they became receptive. Most flowers for outcrossed seed were emasculated in the bud just before the flowers opened; in four of the populations, flowers for selfed seed and a subset of the outcrossed seed were emasculated at an earlier bud stage

(no visible petals). For outcrossed seeds, pooled pollen from at least two plants from

49 at least four different, non-maternal lines from the same population was used. For selfed seeds, flowers were pollinated with pooled pollen from at least three flowers on the same plant. For the two selfing populations, plants that were the result of both selfing and outcrossing were used (Table 4C); data for selfed plants of outcrossing populations were excluded from all analyses.

Seeds were sown on 9/21/12 using the same protocol described above for the 2012 field study, then stratified for one week at 4°C. Seeds were germinated and grown at 22°C for 17 days. Low germination limited the replication of some lines.

Healthy seedlings were transplanted into 3 inch plastic pots filled with Sure-mix perlite potting medium (Michigan Grower Products, Inc., Galesburg, MI). They were then moved to a growth chamber (22°C, 16h days, 150 PAR) and sub-irrigated as needed with 0.5X Hoagland’s solution. Plants that died from transplant shock within a few days were replaced with seedlings if available. After two weeks (10/29/12) of growth, the plants were vernalized for 10 weeks (4°C, 10h days, 75 PAR). They were then moved into a greenhouse set at 20°C (although temperatures fluctuated depending on the weather). Supplemental lighting was 16h days under sodium vapor lamps. The plants were sub-irrigated with water with 100 ppm N Plantex 18-

9-18 pH reducer fertilizer (Plant Products Co. Ltd., Brampton, ON).

Once plants flowered, up to three of the most recently opened flowers per plant were collected in 75% alcohol as described above (for full sampling scheme, see Table 4B). Plants in this study had been flowering longer and were more stressed (dry and thrips-infested) than those in 2012 field common garden, so there were not always three healthy flowers. If there was a fourth healthy flower, it was

50 used for pollen collection. I collected two long anthers in one vial and two short anthers from the same flower in a second vial, sampling a total of 177 flowers; mean

25.21 (SD 13.42) per population. Pollen was counted using a Coulter counter according to the protocol in Rush et al. (1995) and reported as number of grains produced per anther.

Floral measurements

Two sepals and two petals were removed from preserved flowers to expose the pistil and stamens before photographing them to measure floral traits, using

ImageJ 1.43u (Schneider et al. 2012) (Figure 11). I measured the lengths of a short stamen, short anther, long stamen, long anther, sepal, pistil, and both length and width of a petal. I also measured the longest axis of a short anther, long anther, and sepal. After measurements on the photos were complete, ovules were counted on the same flowers by removing the pistil and gently squashing the pistil with a cover slip on a glass slide with a drop of blue food coloring for contrast.

Analyses: Flower size was estimated as PC1 from a principal component analysis including length of short stamen, long stamen, pistil, petal, sepal, and both length and width of the petal. The amount of petal displayed to pollinators is determined by how wide the petal is and how much of it is claw (base of petal) vs. limb (the outer, showy part of the petal). Sepal length was used as a proxy for claw length, so petal display size was estimated by subtracting sepal length from petal length and multiplying by petal width ([petal length – sepal length] * petal width).

To estimate pollen:ovule ratios, I first calculated total pollen per flower by multiplying the per-anther pollen counts by the number of each anther type and

51 adding them together (pollen per short anther *2 + pollen per long anther *4). I then divided total pollen per flower by the mean ovule count from the flowers sampled on the same plant.

For the field study, I tested whether PC1, ovule count, herkogamy (measured as pistil length – long stamen length), long stamen length, short stamen length, pistil length, and petal display changed with mating system using linear models including mating system and population nested in mating system as fixed effects. Pistil, long stamen length, and short stamen length were included to understand how those two traits individually might contribute to differences in herkogamy.

For the greenhouse study, I used the same models with the addition of plant nested in population nested in mating system as a random effect because I sampled multiple flowers per plant. Pollen-ovule ratios were natural log transformed to eliminate heteroscedasticity in the model residuals, and analyzed with a linear model including mating system and population nested in mating system as fixed effects.

While long stamens are expected to evolve closer contact between anther and pistil in selfing populations, short stamens are likely to experience relaxed selection on position if they are not contributing to seed set. To assess the likelihood of short stamens contributed to selfed seed, in addition to the two-population bud dissection study, I inspected the distribution of short stamen herkogamy across both common garden studies for evidence of contact between short anthers and stigmas. To test for an increase in variance of short stamen length with a shift to selfing, I used AIC values to compare the performance of a model including mating

52 system as a fixed effect and population to a model with population, mating system, and different variances for selfing and outcrossing populations. For the greenhouse population, I ran additional models that also included individual plant as a random effect. The mixed-effects models were run in R (R Core Team 2012) using the nlme package (Pinheiro, J. et al. 2013).

To test for decreased pollen production overall in selfing populations, and in short relative to long anthers, pollen counts were analyzed with a model including mating system, stamen length, mating system*stamen length interaction, and population nested in mating system as fixed effects.

For all variables, means across populations were compared using Tukey’s

HSD test. Because I had multiple preserved flowers per plant from the greenhouse study but not from the field, the two data sets required different models and were analyzed separately. All models were also run with and without the BP population, which is categorized as outcrossing (Table 4) but appeared to have a selfing phenotype for many traits. With the exception of short-stamen length variance models, all analyses were performed in JMP 10.0.0 (SAS Institute 2012).

Results

In the seven A. lyrata populations investigated, investment in both female and male function changes as predicted with a shift to selfing. Ovule production increases in populations that self at high rates relative to outcrossing populations

(Table 5, Figure 12), while selfing populations have less pollen on the anthers of both long stamens and short stamens (Table 5, Figure 13). The P/O ratio drops accordingly

53 Table 5. Models of trait differences between populations and mating systems. Pop(ms) indicates population nested in mating system, (ms) indicates mating system, (sl) indicates stamen length. The pollen count and P/O ratio sections for the field common garden are blank because pollen was not collected from that experiment. Plant ID was included as a random effect in the greenhouse common garden models, but for simplicity the results are not reported here. Significant results (p <0.05) are in bold, except population. Trait Field common garden Greenhouse common garden r2 N df F p r2 N df F p Ovule number 0.29 295 ms 1 7.21 0.008 0.81 626 ms 1 10.08 0.002 pop(ms) 4 26.6 <0.0001 pop(ms) 5 25.55 <0.0001

Pollen count 0.15 312 ms 1 6.08 0.01 pop(ms) 5 5.37 <0.0001 sl 1 17.18 <0.0001 sl * ms 1 0.008 0.9285

Pollen:ovule ratio 0.18 176 ms 1 3.10 0.08 pop(ms) 5 7.35 <0.0001

Pollen:ovule ratio 0.08 144 ms 1 8.13 0.005 (no BP) NA pop(ms) 4 1.51 0.20

Petal display 0.30 262 ms 1 2.21 0.14 0.60 585 ms 1 0.19 0.67 pop(ms) 4 26.7 <0.0001 pop(ms) 5 13.6 <0.0001

Petal display 0.20 229 ms 1 1.24 0.27 0.56 498 ms 1 5.34 0.02 (no BP) pop(ms) 3 17.88 <0.0001 pop(ms) 4 3.87 0.005

Flower size 0.19 256 ms 1 0.52 0.47 0.60 586 ms 1 0.49 0.48 pop(ms) 4 14.56 <0.0001 pop(ms) 5 10.88 <0.0001

Flower size 0.18 223 ms 1 0.08 0.78 0.58 499 ms 1 4.32 0.039 (no BP) pop(ms) 3 16.24 <0.0001 pop(ms) 4 8.87 <0.0001

54 Table 5 (cont’d)

Trait Field common garden Greenhouse common garden Long stamen - pistil 0.17 297 ms 1 52.11 <0.0001 0.70 629 ms 1 0.67 0.41 pop(ms) 4 1.31 0.27 pop(ms) 5 3.49 0.005

Pistil 0.16 297 ms 1 27.55 <0.0001 0.57 639 ms 1 0.002 0.97 pop(ms) 4 7.90 <0.0001 pop(ms) 5 2.55 0.03

Long stamen 0.15 297 ms 1 0.12 0.73 0.59 629 ms 1 0.84 0.36 pop(ms) 4 13.20 <0.0001 pop(ms) 5 7.56 <0.0001

55 Outcrossing" A" Selfing" 45" B" 40" B" p"="0.008" A" *" 35" A" A" A" 30" 25" 20" 15" 10" 5" 0" LUD" IDU" BP" PIN" RON" PP" OUT" SELF" B" 45" 40" D" CD" p"="0.002" 35" BC" AB" AB" *" 30" A" A"

Ovule"produc;on"per"flower" 25" 20" 15" 10" 5" 0" LUD" SAU" IDU" BP" PIN" RON" PP" OUT" SELF" Popula;on"""""""""""""""""""""""""Ma;ng"system"

Figure 12. Self-pollinating populations produce more ovules per flower. LS means ± 2SEM. A) Field common garden study. B) Greenhouse common garden study. Populations are ordered left to right from highest to lowest outcrossing rate. Letters indicate differences between populations according to Tukey’s HSD test. The last two bars compare the selfing and outcrossing populations, with an asterisk indicating a difference significant at p < 0.05.

56 Outcrossing" 12000" Selfing" 10000" AB" A" ABC" p"<"0.0001" ABC" p"<"0.0001" 8000" BC" *" *" C" C" 6000"

4000"

2000"

0" Pollen"produc.on"(per"anther)"

Popula.on"""""""""""""""Ma.ng"system""""Stamen"type"

Figure 13. Self-pollinating populations produce less pollen, and short stamen anthers produce more pollen than long stamen anthers. LS means ± 2SEM. There was no significant mating system by stamen type interaction. Data from the greenhouse common garden study. Populations are ordered left to right from highest to lowest outcrossing rate. Letters indicate differences between populations according to Tukey’s HSD test. The last two bars compare the selfing and outcrossing populations, with an asterisk indicating a difference significant at p < 0.05.

57 Outcrossing" 2500" A" A" Selfing"

2000" A"

"ra9o" A" p"="0.08" AB" 1500" AB" B" 1000"

500" Pollen:ovule

0" LUD" SAU" IDU" BP" PIN" RON" PP" OUT" SELF" Popula9on"""""""""""""""""""""""""Ma9ng"system"

Figure 14. Pollen:ovule ratio is lower in selfing populations. LS means ± 2SEM; marginal result becomes highly significant when BP population is excluded. Populations are ordered left to right from highest to lowest outcrossing rate. Letters indicate differences between populations according to Tukey’s HSD test. The last two bars compare the selfing and outcrossing populations, with the p-value indicating a lack of statistically significant difference.

58 with a shift to selfing, although the very low ratio of the BP population makes this shift borderline statistically significant (Table 5, Figure 14).

I found no decrease in flower size in the common garden studies with a shift to selfing – either overall size, measured as PC1, or petal display size (Table 5). Principal component one (PC1) was a good estimate of flower size, explaining 71.8% of the variance in flower size in the greenhouse and 63.4% in the field. Loadings for all traits were between

0.75 – 0.91 in the greenhouse 0.64 – 0.89 in the field.

There is no support for decreased herkogamy in selfing populations. There were differences in herkogamy between mating systems opposite the predicted direction in the field, but not in the greenhouse (Table 5, Figure 15). Selfing populations grown in the field had greater herkogamy than outcrossing populations, with long stamen anthers below the stigmas; outcrossing populations had little herkogamy. This was due to selfing populations in the field having longer pistils than outcrossing populations. Long stamen length did not differ between mating systems (Table 5, Figure 15).

In the study looking for contact between short anthers and stigmas in inflorescences from nine individuals from a selfing and an outcrossing population, no flower at any developmental stage in either population population ever had a short stamen that was dehisced and in contact with the stigma. In the common garden studies, my measure of short stamen herkogamy indicated a possibility of overlap between short anthers and stigma in only six of 1018 flowers measured (<0.6%). This makes autogamous selfing via short stamens unlikely. These results strengthen my expectation of seeing reduced investment in short stamens and increased variance in their length, as a lack of function should release any stabilizing selection on their position.

59 " A" B" B" p"<"0.0001" 0.3" *" 0.2" A"

herkogamy 0.1" A" A" A"

0" LUD" IDU" BP" PIN" RON" PP" OUT" SELF"

50.1"

50.2" Outcrossing" Long"stamen" 50.3" Selfing" B" 4.2" A" A" p"<0.0001" A" 4" *" 3.8" B" B" 3.6" B" 3.4"

PisFl"length" 3.2"

3" LUD" IDU" BP" PIN" RON" PP" OUT" SELF"

C" 4.2" A" 4" B" p"="0.73" 3.8" B" B" B" B" 3.6" 3.4" 3.2" 3" Long"stamen"length" LUD" IDU" BP" PIN" RON" PP" OUT" SELF"

PopulaFon"""""""""""""""""""""""""MaFng"system"

Figure 15. Greater herkogamy in selfing populations in the field is due to longer pistils. LS means (mm) ± 2SEM, only showing results from the field common garden study. A) Herkogamy (long stamen length – pistil length), B) pistil length, C) long anther length. Populations are ordered left to right from highest to lowest outcrossing rate. Letters indicate differences between populations according to Tukey’s HSD test. The last two bars compare the selfing and outcrossing populations, with an asterisk indicating a difference significant at p < 0.05.

60 Table 6. Models testing for increased variance in short stamen length with shift to self-pollination. A. Models explaining variation in short stamen length with and without taking variance by mating system into account were compared using AIC; dAIC is (model AIC) – (best model AIC). Lower AIC values indicate a better model, so dAIC of 0 is the best model fit. All models included mating system (ms) as a fixed effect and population (pop) as a random effect. For the greenhouse common garden, which included multiple flowers sampled per plant, I also fit a model with plant ID. Comparing those base models to ones including variance by mating system (ms variance) shows that taking variance by mating system into account does not improve model fit. This indicates no significant difference in short stamen length variance between mating systems. B. Estimates of variance in short stamen length between mating systems. There is no significant difference between variance in outcrossing and selfing populations.

A. field common greenhouse common garden garden model AIC dAIC df AIC dAIC df ms + pop 276.86 0 4 373.41 70.94 4 ms + pop + ms variance 278.79 1.93 5 375.39 72.92 5 ms + pop + plant 305.43 1.61 5 ms + pop +plant + ms variance 303.82 0 5

B. variance selfing outcrossing field common garden 0.121 0.140 greenhouse, individual not in model 0.100 0.102 greenhouse, individual in model 0.069 0.061

61 However, I found no evidence for either decreased pollen production on short anthers or greater variance in length of short stamens. The lack of significant interaction between stamen length and mating system in the pollen count models indicates there was no decrease in pollen production in the short stamen anthers relative to long stamen anthers of selfing populations relative to outcrossing populations (Table 5). Models including a difference in short stamen length variance between mating systems did not fit better than those without the variance effect (Table 6), indicating that variance did not change with a shift to selfing. In fact, for models without individual plant in them, mean variance in short stamen length was nonsignificantly greater in outcrossing populations

(Table 6).

The Bruce Peninsula population (BP), categorized as outcrossing, was unusual in several respects. The floral morphology fit the predictions for a selfing rather than outcrossing population for ovule count, pollen production, P/O ratio, and the size of flowers (including petal display) (Table 5, Figures 12-14), although removing it from the models only changed the results for mating system for flower and petal display size, and only for the greenhouse study (Table 5). However, selfing and outcrossing populations can exist close together, and are in fact known to at this particular site (Foxe et al. 2010), so reassessing outcrossing rate for this sample may be useful.

Variation in floral morphology between populations was substantial in both the field and common garden studies (Table 5). Although not all traits differed with mating system, every trait tested differed significantly between populations (with the exception of long stamen herkogamy in the field common garden and the P/O ratio).

62 Discussion

Populations of Arabidopsis lyrata with high selfing rates exhibit only a subset of the selfing syndrome, indicating which traits are evolving first in this early stage of the mating system shift. As predicted, selfing A. lyrata have sex allocation shifted toward female reproduction; the pollen/ovule ratio is reduced due to both decreased investment in male fitness (less pollen) and increased investment in female fitness (more ovules) (Table 5).

However, they show no reduction in flower size (Table 5), and herkogamy in selfers is either unchanged or shifted to greater separation between anthers and stigma in selfing species, counter to predictions (Table 5, Figure 15).

One possible explanation for increased herkogamy in the selfing populations is early pollination followed by expansion of the pistil as ovules begin to mature. A. lyrata’s sister species, the highly selfing A. thaliana, is known to self in the bud. Because the ovaries comprise most of the pistils of many mustards (including Arabidopsis spp.), the pistils may grow rapidly after fertilization (this is certainly true in Raphanus raphanistrum; JK Conner, personal communication). This may not have been observed in the greenhouse plants because they were sampled when older and more stressed, possibly leading to fertilized ovules not developing. Future studies should measure herkogamy earlier, ideally before anthers dehisce, to eliminate this possibility.

Lack of contact between the short stamens and stigma suggests that, as in sister species A. thaliana (Chapter 1), short stamens are unlikely to contribute to autogamously selfed seed in A. lyrata. In spite of this, there is no evidence that selfing populations are reducing investment in short stamens more than long stamens. Both stamen types have equally diminished pollen production relative to outcrossing populations (Table 5, Figure

63 13). There is also no support for increased variance in short stamen length (Table 6). This may point toward constraint on independent evolution of short stamens, as in A. thaliana

(Chapter 1). Alternatively, as with other traits that do not conform to the selfing syndrome, it could indicate that short stamens maintain adaptive function in relatively rare outcrossing events or for insect-mediated self-pollination within a plant (geitonogamy), or that selection to eliminate the trait is weak because it requires few resources.

The relative frequency of geitonogamous vs. autogamous selfing in A. lyrata populations is currently unknown. Although the reduced P/O ratios in selfing populations suggest that autogamous selfing is shaping the evolution of floral morphology, persistence of low levels of self-incompatibility in these populations means not all plants are capable of even geitonogamous selfing. Among self-compatible plants, not all are capable of autogamous selfing (Mable and Adam 2007), and even those that can self within a flower do not do so consistently (Mable, personal communication; personal observation.)

The geographic distribution of selfing populations supports the idea that selfing may have evolved through geitonogamy. Populations with low outcrossing rates are found on the edges of the species range, where low genetic diversity due to range expansion, exacerbated in some cases by small population size, can cause selection for the loss of self- incompatibility (Griffin and Willi 2014). Thus, the evolution of selfing in A. lyrata is likely not related to a stressful environment or lack of pollinators, both of which would be expected to cause selection for smaller flowers. Geitonogamy contributes significantly to pollination in many other selfing plant species (Goodwillie et al. 2005); if this is true in A lyrata, a reduction in flower size could be maladaptive even in the most highly selfing

64 populations. Future studies quantifying autogamous selfing rates will be critical for understanding the evolution of floral morphology in A. lyrata.

Inspecting the floral morphology of selfing A. lyrata populations shows us that these plants have only traveled a short way down the path to autogamous selfing. They are allocating less to male and more to female fitness, but meet none of the other criteria of the selfing syndrome. As an example of the evolution of selfing in progress, it is clear that in this case reduced flowers would be a possible eventual result, not a cause, of a shift in mating system. Further work on the mechanisms of self-pollination in this species would shed light on why the flowers of selfing A. lyrata populations have not yet evolved the full selfing syndrome.

Acknowledgments

Jeff Conner gave advice from conception through writing. Raffica LaRosa commented on the manuscript. Emily Dittmar, Nick Batora, and Chris Oakley grew the

2012 outdoor common garden plants from seed to transplant stage. Mark Hammond coordinated the field transplant. Chris Oakley and Jon Spoelhof allowed me to sample flowers from the 2013 greenhouse common garden experiment. Doug Schemske included me in collaborative experiments in his lab, including advice and materials. Shawn Szabo and Marvin Osborne assisted in the lab. Colin Kremer helped with statistical analysis in R.

This work was supported by grants from the Kellogg Biological Station’s G.H. Lauff fund and T. Wayne & Kathryn Porter fund.

65 CHAPTER 5

SUMMARY AND FUTURE DIRECTIONS

Summary

My dissertation addresses the evolution of tetradynamy across the range of mating systems, with a focus on the function and evolution of short stamens. I confirmed in

Chapter 2 that tetradynamy is adaptive for outcrossing via pollinators, and discovered that the function of the trait is affected by pollinator visitation frequency. In Chapter 3, I show that tetradynamy is not adaptive for autogamous selfing, and in particular the short stamens appear to be nonfunctional. It appears that constraint, in the form of epistasis and possibly low genetic variance or genetic correlations, exacerbated by inbreeding, is slowing the evolutionary elimination of short stamens in the model plant Arabidopsis thaliana. I explored investment in short stamens in populations that have recently evolved selfing in

Chapter 4. I found that, as I expected from my results in sister species A. thaliana, selfing populations of A. lyrata have not detectably reduced investment in their newly nonfunctional short stamens. By studying the function and evolution of short stamens across a range of mating systems, I have been able to illustrate that while tetradynamy is almost certainly maintained by natural selection in outcrossing species, it is nonfunctional in autogamous selfers and its evolutionary loss is likely constrained.

Future directions

While previous work has established the adaptiveness of tetradynamy, discovering how the trait functions has proven a tremendous challenge. The insight that pollinator visitation rate impacts function is a step forward, but there are several clear next steps to take in exploring how mustard stamens work. Sahli and Conner (2011) showed that

66 different pollinator taxa differ in the selection they exert on dimorphism. I have data on community composition for the field days in Chapter 2; looking at whether selection varies with pollinators present may be a first step. Blending Sahli and Conner’s approach of focusing on individual pollinator taxa with the experimental manipulations in Chapter 2, and perhaps adding variation in overall visitation rate, could help further flesh out the possibilities of the trait specialization hypothesis. In order to understand how different pollinators together could favor tetradynamy, we particularly need data on how nectar vs. pollen feeders interact with short or long stamens in isolation. Less ambitious experiments that would still further our understanding could include detailed observations of pollinator interactions with flowers. These could be combined with experiments using different colors of fluorescent dyes on short and long stamens, observing where the powder lands on pollinators and how much of it is transferred to subsequently visited flowers. There is certainly abundant material for further work in understanding how tetradynamy functions in outcrossing.

Because of the model system status of Arabidopsis thaliana, the future directions suggested by Chapter 3 are particularly expansive. We hope to both deepen and broaden this work. To go deeper, we have explored knockout lines for candidate genes, and are seeking funding to identify the genes or even specific mutations responsible for short stamen loss. To broaden in the short term, we are pursuing QTL mapping in independent sets of RILs to establish whether short stamen loss has evolved through different pathways.

One set of lines will be planted in July 2014, with several more candidate sets available as we have the resources to allocate. Long term, I would like to investigate short stamen function and loss in other highly selfing members of the Brassicaceae. There is at least one

67 other species, Cardamine hirsuta, which shows a remarkably similar pattern of short stamen loss, including interpopulation variation (Matsuhashi et al. 2012). Exploring the function of short stamens and genetic architecture of their production in additional species would help us understand if our results are general.

The conclusions I could draw in Chapter 4 are limited in part by the low sample size of selfing populations, but also by treating mating system as a categorical variable when, in

Great Lakes populations of A. lyrata, it is clearly continuous. Ideally, I would like to set up a common garden sampling all known self-compatible populations of A. lyrata and balance the design with a random sample of self-incompatible populations. Data on pollination biology in natural populations is also much-needed; while several European labs are doing excellent work on the evolution of self-incompatibility in the species, field observations are lacking. For understanding the evolution of the selfing syndrome, we need to know how much of self-pollination is geitonogamous, i.e. whether selfing is a response to lack of pollinators or low diversity in SI alleles.

68

APPENDIX

69 Growth conditions for Arabidopsis thaliana

Except where specified, plants for the geographic variation observational study and stamen removal experiments were grown in growth chambers under the conditions described below either at Michigan State University’s main campus in East Lansing,

Michigan (MSU), or at Michigan State University’s Kellogg Biological Station in Hickory

Corners, Michigan (KBS). Most seeds were obtained from the Arabidopsis Biological

Resource Center (abrc.osu.edu). Collections in Sweden and Italy were made by D.

Schemske, and Ukrainian collections were from J. Beck (Wichita State University).

MSU growth conditions: seeds were sterilized in 30% bleach solution with 0.001%

Triton X 100 surfactant for ten minutes, washed twice with sterilized water, suspended in

0.1% agar, and sown on a 100mm X 15 mm petri dish with agar (Phytoblend, Caisson

Laboratories Inc.). Petri dishes were filled with 20mL of 0.8% Agar with 2.31% Gamborg’s

B-5 basal medium and autoclaved after the media set. Sown plates were wrapped in

Parafilm, stratified in the dark for four days at 5˚C, then moved to 22˚C with 16-hour days with ~100- 150 µEinsteins photosynthetically active radiation (PAR). Once the first true leaves formed, seedlings were transplanted to 2” square pots with Suremix “Perlite” soil- less media. Plants were bottom-watered with ½ strength Hoagland’s Nutrient solution; soil was allowed to dry down between watering. Plants were kept at 22˚C with 16-hour days and ~100- 150 µEinsteins PAR for 3 weeks until rosettes were formed, and then vernalized for ~8 weeks at 6˚C with 10-hour days and ~50 µEinsteins PAR to promote flowering. The plants were then moved to 22˚C with 16-hour days with ~100- 150 µEinsteins PAR for the duration of the study.

70 KBS growth conditions: 3-10 seeds were direct-seeded into 2” square pots with

MetroMix growing mix (Sun Gro). Seeds were then stratified in the dark at 6˚C for four days, followed by five weeks at 22˚C with 16-hour days. After germination, seedlings were thinned to a single plant and pot locations were randomized. Plants were vernalized for 10 weeks at 6˚C with 10-hour days. Conditions were then changed to 22˚C with 14-hour days for flowering. Pots were bottom-watered with deionized water and allowed to dry down between watering.

Experiment: function of A. thaliana short stamens in selfing

The four stamen removal treatments (removing all stamens, removing no stamens but probing inside the bud, removing short only, or removing long only) were performed before anther dehiscence in unopened buds using fine forceps under a dissecting microscope. Pedicels were painted with unique colors indicating the treatment applied.

There was an additional control treatment that included painting the pedicel but no other interference with the flower.

Because we only used flowers with two short stamens, populations with high short stamen production were utilized to guarantee all treatments could be performed. In 2007, the stamen removal treatments described above were performed on 18 plants grown at MSU from two Swedish populations (originally collected by D. Schemske near the towns of

Rödåsen and Skuleberget). In 2011, I repeated the experiment on 2-4 individuals from each of five populations grown at KBS (total 16 plants) sampled from across Europe (Table 7); these plants were placed in the field after stamen removal until the manipulated flowers senesced to see if short stamens functioned in a more natural environment, then returned to a growth chamber to mature any developing fruits. Resulting seeds were counted. Seed

71 Table 7. Accessions included in the study of geographic variation in short stamen production. Includes accessions in studies of short stamen function (indicated with an asterisk next to the code) and parents of the recombinant inbred lines (Rodasen [Sweden] and Belmonte [Italy]). Except where noted, stock numbers are those used by the Arabidopsis Biological Resource Center (abrc.osu.edu). Beck and Schemske lines were collected by James Beck and D. Schemske. Locations were taken from the name of the line and/or the nearest location to the lat/long coordinates provided to identify the lines. In cases where the two disagreed, the coordinates took precedence. When coordinates were unavailable, the named location of collection was used to find them. An “X” to the left of the latitudinal or longitudinal coordinates indicates that accession was included in the corresponding sampling band of the native range.

Codes Location # Lines Stock Numbers Latitude Longitude Stam # Austria Gr Graz 5 1198, 1200, 1202, 1204, X 47.1 N 15.4 E 1.6 1206 Uod Uod 2 22612 - 13 X 48.07 N X 14.53 E 1.27 Belgium An Antwerp 2 946, 22626 51.3 N X 4.3 E 1.14 Czech Republic Br Brno 1 22628 X 49.12 N X 16.37 E 0.49 Lp2 Lipovec 2 22594 - 95 X 49.22 N X 16.39 E 1.29 Sap Slapy 1 6854 X 49.8 N X 14.4 E 1.11 Zdr, Bor Zdarec/Borky 4 22588 - 91 X 49.38 N X 16.27 E 1.14 Pu2 Prudka, Croatia 2 22592 - 93 X 42.38 N 18.07 E 1.21 Finland Tamm* Tammisari 2 22604 - 05 X 59.58 N 23.26 E 1.78

72 Table 7 (cont’d)

Codes Location # Lines Stock Numbers Latitude Longitude Stam # France Cen Caen 1 1066 49.2 N X 0.35 W 1.71 Gy La Miniere 1 22631 49 N X 2 E 1.41 Rennes Rennes 5 22269, 22271 - 72, 22610 - 11 48.5 N X 1.41 W 1.02 Germany Bay Bayreuth 2 954, 22633 X 49 N X 11 E 1.57 Bd, Sp Berlin 2 962, 1530 X 52.5 N X 13.4 E 1.2 Dr Dresden 1 1114 X 51 N X 13.75 E 1.67 Ei* Eifel 4 1126, 1128, 1130, 22616 50.3 N X 6.3 E 1.44 Er Erlangen 1 1142 X 49.59 N X 11.04 E 1.24 Ga Gabelstein 2 1182, 22634 50.3 N X 8 E 1.22 Got Göttingen 5 22608 - 09, 22308 - 10 51.32 N X 9.55 E 1.92 Je* Jena 1 1246 X 50.9 N X 11.6 E 1.68 Krot Krottensee 3 3886 - 88 X 49.6 N X 11.6 E 1.38 Nd Niederzenz 2 1390, 22619 51 N X 10 E 1.31 No Nossen 1 1394 X 51.05 N X 13.3 E 1.03 Ste Stendal 1 1536 X 52.6 N X 11.85 E 1.47 Wt Wietze 5 1604, 1606, 1608, 1610, 22637 52.3 N X 9.3 E 1.53 Italy NA Belmonte 5 Schemske 1, 2, 4, 12, 13 X 42.12 N 12.48 E 1.35 NA Bolsena 5 Schemske 3, 11, 15, 17, 18 X 42.65 N 12 E 1.07 Ct Catania 1 22639 X 37.3 N 15 E 0.68 Pa Palermo 3 1438, 1440, 1442 X 38.1 N 13.4 E 1.32 Kazakhstan KZ* Karagundy 5 22442 - 44, 22606 - 07 49.5 N X 73.1 E 1.15 Netherlands Nok Noordwijk 4 1398 1400 1402 22643 52.3 N X 4 E 1.9 Poland La Landsberg/Gorzów 3 1298, 1302 X 52.5 N X 15.5 E 1.8 Wa Warsaw 1 22644 X 52.23 N X 21 E 1.54

73 Table 7 (cont’d)

Codes Location # Lines Stock Numbers Latitude Longitude Stam # Russia Ws Wassilewskija 1 22623 52.5 N X 30 E 1.12 Sweden X 63.01 N 18.19 E 1.86 Fab Fäberget 2 22576 - 77 X 62.48 N 18.05 E 1.79 Lov Lövvik 2 22574 - 75 X 56.12 N 15.18 E 1.17 Omo2 Östra Möcklö 2 22584 - 85 X 56.12 N 15.18 E 1.17 NA* Rodasen 5 Schemske 1, 5, 9, 47, 51 X 62.8 N 18.2 E 1.96 NA* Skuleberget 5 Schemske 4, 20, 27, 28, 36 X 63.05 N 18.22 E 1.74 Spr1 Sprattleboda 2 22582 - 83 X 56.32 N 14.29 E 1.79 M3385S, Stockholm 3 3111, 3114, 1534 X 59.2 N 18.4 E 1.62 M7943S, St Var2 Vårhallarna 2 22580 - 81 X 55.33 N 14.2 E 1.43 Ukraine En-D Donetsk 1 920 48 N X 37.8 E 0.76 Koch Kocherov 5 22823 - 27 50.36 N X 29.96 E 1.11 NA Veselinovka, Kyiv 5 Beck 794 (1-5) 50.26 N X 31.5 E 1.27 Oblast

74 unmanipulated

30

A)! 25 untouched) 20

15

10

Frequency) 5

0 removed none 0 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 30 B)! 25 probed) 20

15

10

Frequency)) 5 removed short 0 0 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 30 C)! removed)shorts) 25

20

15

10 Frequency)) 5 removed long 0 0 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 D)! 30 25 removed)longs)

20

15

10 Frequency)) 5 removed all 0 0 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 E)! 30 25 removed)all) 20

15

10 Frequency)) 5

0 0 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 Seed)set)per)fruit) Figure 16. Distribution of seed set in stamen removal treatments. Treatments are depicted in cartoons to the right of histograms; d) was a control with flowers probed with forceps, whereas e) was an unmanipulated control. Increasing frequency of flowers with zero seed set from a-b and b-c is consistent with a side effect of greater disruption of the flower from probing and short-stamen-removal. The increase in flowers with zero seed set from c-d is not likely due to damage, because long stamens are easier to remove without disruption to the flower than short ones.

75 Table 8. Testing function of short stamens. Treatment was highly significant (F = 42.37, p <0.0001, N = 157). Flowers in the “no stamens removed” treatment were probed with the forceps, and flowers in the “unmanipulated” treatment were untouched. For Tukey’s HSD, a shared letter indicates no significant difference.

Treatment Least Square 1 SE Tukey’s HSD N Mean All stamens removed 0.25 3.42 A 33 Long stamens removed 2.60 3.43 A 32 Short stamens removed 21.18 3.43 B 32 No stamens removed 25.04 3.42 B/C 33 Unmanipulated 29.69 3.54 C 27

76 set was analyzed in JMP version 10.0.0 (SAS Institute 2012) with a model including treatment as a fixed effect, population as a random effect, and plant ID as a random effect nested in population. Formally, this is:

γ ~ β0

γ ~ β0 + βt +ε 2 β ~ N(β0 p,σ β 0 ) 2 βi ~ N(β0 pi,σ β 0 )

Where γ is the number of seeds produced in a single fruit, 0 is the intercept, t is the treatment, p is the population, and i is the individual plant nested in population. Differences between treatment means were tested using a Tukey HSD test (Table 8).

We initially included year and the year-by-treatment interaction in the model for seed set in the stamen removal experiments, but there was no significant effect of environment (indicating no increase in function in the field vs. growth chamber), so we left them out of the final model. We tried to control for bud damage during stamen removal by including a control in which buds were probed but not manipulated further, and this treatment did show an increase in fruits with seed set of zero (likely due to damage) compared to a completely unmanipulated treatment (Fig. 16). However, the control treatment was likely not aggressive enough to mimic damage levels short stamen removal.

The position of short stamens deeper in the buds makes it likely that short stamen removal will either disrupt long anther position (reducing contact with the stigma) or that the pistil itself will be damaged. Evidence for this comes from the greater number of flowers with zero seed set in the short-stamen-removal treatment compared to the manipulated control

(Fig. 16). Four flowers had unusually high seed set (≥ 24) in the long- and all-stamens-

77 removed treatments (Fig. 16). These were likely the result of accidental pollinations; three of the four (including the highest in both treatments) occurred in growth chambers, where no potential insect pollinators were present. They were left in the analysis as we had no a priori reason to eliminate them.

Geographic variation in A. thaliana short stamen production

Plants were grown at KBS with populations interspersed in trays. Tray positions in growth chambers were rotated weekly to minimize position effects. Replicate plants for each line were divided evenly in two blocks grown under identical conditions two months apart. Flowers were preserved in 70% ethanol for later stamen counts using a dissecting microscope. Short stamen number was used in the analysis because a reduction in total stamen count is nearly always due to loss of short stamens (only 31 of the 2570 flowers examined, or 1.2%, produced fewer than four long stamens, compared to 1496, or 58.2%, producing fewer than two short stamens). Correlations between stamen production and longitude and latitude were performed in JMP 10.0.0 (SAS Institute 2012).

QTL mapping

Seeds were sown in sterilized petri dishes on media consisting of Gambog’s B-5 nutrient mix, Bacto Agar, and ultrapure water. They were stratified at 4°C for five days and then moved to 22°C with 16-hour days for 8-10 days. Seedlings were then transferred to soil in tubes and returned to the growth chambers for 8 days. Of the seedlings that survived transplant, six per RIL and 20 of each parental line were transferred to growth chambers programmed to mimic the temperature fluctuations encountered during the Arabidopsis thaliana growing season in Italy (Dittmar et al, unpublished). Plants were randomized to positions in six trays divided between two growth chambers, with each tray divided into

78 Table 9. Plants, lines, and flowers sampled for QTL analysis. The multiple parent lines are all descended by selfing from the two individuals used to produce the RILs. Fewer Swedish parents were sampled because they flowered late and completed flowering quickly.

Type of lines Total Total # Total # Mean Mean flowers # lines plants flowers plants per per plant line (SD) (SD) RILs 519 2468 7435 4.76 (1.35) 3.01 (0.30) Italian parent 10 168 502 16.8 (1.03) 2.99 (0.15) Swedish parent 10 74 224 7.40 (2.01) 3.03 (0.37)

79 four sections: long side of the perimeter, short side of the perimeter, inner circle, and center. Plants were randomized across the six trays and sections within trays, and trays were rotated randomly within and between chambers every three days until flowering began. Trays were watered as needed with deionized water and ½-strength Hoagland’s solution.

On the day of flower collection, we collected the three most recently opened flowers on plants with at least 10 open flowers. If there were fewer than three healthy open flowers, we took whatever was available; if the plant was near finishing flowering but still healthy, we took the entire flowering end of the inflorescence, occasionally resulting in more than three flowers sampled per plant. There is a small but significant effect of flower rank, with later flowers producing more short stamens, but even for late flowers the large difference in production between the Italian and Swedish RIL parents is preserved, so variation in rank of flowers collected should not bias the results. Flowers were stored in

70% ethanol. In total, we collected and scored 8161 flowers from 519 RILs plus the parents

(Table 9). Stamens were counted under a dissecting microscope. Stamen number averaged across flowers and individual plants for each RIL was used for the QTL analysis.

To assess position effects on stamen production and variation between and within

RILs, we partitioned variance with a random-effects model in JMP version 10.0.0 (SAS

Institute 2012) that included RIL, individual plant nested in RIL, tray, and position nested in tray. Formally this is:

80 γ ~ β0 +ε

2 β0 ~ N(β0r,σ β 0 ) 2 β p ~ N(β0 pr,σ β 0 ) 2 βt ~ N(β0t,σ β 0 ) 2 βu ~ N(β0tu,σ β 0 )

Where γ is the number of short stamens produced in a single flower, r is the RIL, p is the individual plant, t is the tray, and u is the position nested in tray. Variance between RILs was substantial (36.16%), indicating a strong genetic component controlling the trait. The small variance between plants within RILs (0.04%) was consistent with this. Most of the variance in stamen production was among flowers within individual plants (60.49%), reflecting high within-plant plasticity - plants with stamen loss produced all three flower types, with 0, 1 and 2 stamens. Because we sampled consecutive flowers, we may have even underestimated the within-plant variance. Growth chamber position effects on stamen production were minimal (2.18% for tray and 1.14% for position within the tray).

This suggests that while mean short stamen production is strongly genetically controlled, it is poorly canalized, resulting in common production of multiple flower morphs on a single plant. This lack of buffering in response to minor environmental variation suggests a trait in the early stages of loss.

The skewed distribution of stamen loss is the RILs is consistent with epistasis: a disproportionate fraction of the lines had no variation in short stamen production, with all flowers producing both stamens (Fig. 20). In lines with heterozygosity this could indicate

81 stepwise_stamenqtl_full_untransformed_LodProfile

[email protected]

40

30 lod 20

[email protected] 10

[email protected]

0

1 3 5

Chromosome

Figure 17. Main-effect QTL using stepwise analysis on the complete untransformed data with no epistasis in R/qtl. The numbers above the peaks give the chromosome location for peaks that cross the threshold for significance (e.g. [email protected] = peak at 74.4 cM on chromosome 1). The LOD threshold (2.62), designated by a horizontal dotted line, was estimated by permutation for 5% significance.

82 2$ p<0.0001$$ 2$ p$=$0.077$$" $" 1.9$ 1.9$ 1.8$ 1.8$ 1.7$ 1.7$ 1.6$ 1.6$ 1.5$ 1.5$ QTL A allele 1.4$ 1.4$ QTL A allele produc2on$ produc2on$ 1.3$ loss$ 1.3$ loss$ Mean Short StamenMeanShort Number 1.2$ 1.2$ loss production loss production QTL B allele QTL C allele

Figure 18. Epistasis results in reduced short stamen loss. Epistatic interactions between QTL A and QTL C, and between QTL A and QTL C; p-values are from the ANOVA testing for all possible interactions between the two loci. Error bars are ±2 SEM.

83 2.0

b 0 c

bc 1.8 a

1.6 ac

Short Stamen Number Short Stamen ab 1.4

abc

1.2

Figure 19. Epistasis results in fewer effective paths to evolution of stamen loss by natural selection. Starting with the all-Swedish genotype (0), which produces the most short stamens, this illustrates the phenotypes of the 12 possible transitions to accumulating the Italian genotype at all three QTL. Dotted lines indicate transitions that do not result in a significant decrease in short stamen number (determined using Tukey’s HSD test on epistasis ANOVA results). Five of the 12 transitions do not result in significant loss of short stamens, and only one path has a significant reduction at each step (0 -> a -> ac -> abc, corresponding to the upper right panel of Fig. 3b).

84 Swedish"parent" 1.95"(0.02)" 180"

160"

140"

120"

100"

80"

Number'of'RILs' Italian"parent" 60" 0.92"(0.05)" 40"

20"

0" 0.4")"0.59" 0.6")"0.79" 0.8")"0.99" 1.0")"1.19" 1.2")"1.39" 1.4")"1.59" 1.6")"1.79" 1.8")"1.99" 2" RIL'Mean'Short'Stamen'Number'

Figure 20. Distribution of mean short stamen production in the recombinant inbred lines. RIL parent means are depicted as stars, with 1 SEM in parentheses. The left skew of the distribution hints at the epistasis uncovered in the QTL analysis.

85 dominance, but because the RILs are nearly completely homozygous, it can only be caused by epistasis.

The main-effect QTL analyses were carried out in R/qtl (Broman et al. 2003)

(Figure 17). We performed Haley-Knott regression with 10,000 permutations to set a LOD threshold, followed by automated stepwise analyses (Broman et al. 2003) without epistasis, with alpha set at 0.05. We tested for epistasis between the main- effect QTL with an ANOVA in JMP (SAS Institute 2012) based on RIL genotypes at the marker closest to the QTL peaks. The model included all QTL and all possible interactions between them (Table 11). Formally, this model is:

γ ~ β0 + βa + βb + βc + βaxb + βaxc + βbxc + βaxbxc

Where γ is the mean number of short stamens produced in a RIL, a is the effect of the stamen-loss allele at QTL1, b is the effect of the stamen-loss allele at QTL3, and c is the effect of the stamen-loss allele at QTL5.

The distribution of stamen number was highly skewed (stamen loss is relatively rare in the RILs, Figure 20) and could not be normalized through transformation.

Because the automated stepwise procedure in stepwiseqtl is sensitive to non- normal residuals, we analyzed the untransformed data and then made two new, complementary versions of the dataset: one binary (lines coded as loss vs. no loss), and one quantile-normalized, including only lines with stamen loss and analyzed them the same way (Broman et al. 2003). The analyses identify the same three main-effect QTL: one each on chromosomes 1, 3, and 5 (Fig. 3, Tables 12,

86 Table 10. Locations and 95% credible intervals for main-effect QTL peaks. QTL are designated by chromosome number. QTL from R/qtl and QTL Cartographer (Cart) are similar; the peaks from QTL Cartographer fall near or within the confidence intervals for the peaks in the R/qtl analysis. R/qtl peaks, and Bayes 95% credible intervals were calculated using the full, untransformed data; kb locations are based on the AGI map. Percent variance explained for R/qtl was calculated using the LOD scores from the Haley-Knott regression (Broman and Sen, 2009).

QTL R/qtl peaks Bayes 95% Bayes 95% QTL Cart peaks (PVE) credible interval credible interval (PVE) (cM) (kb) 1 74.4 cM (2.7%) 58.6 – 83.8 cM 21,551 – 30,246 80 cM (2.6%) 3 16.5 cM (7.0%) 15.6 – 19.7 cM 7,329 – 8,201 26 cM (7.8%) 5 7.7 cM (28.4%) 6.7 – 7.7 cM 2,408 – 2,986 10 cM (27.9%)

87 Table 11. Significance of main effects and interactions. From analyses in R/qtl using the full untransformed data; main effects are from the R/qtl stepwise model without epistasis, interactions are from ANOVA including main effects and all possible interactions in JMP.

QTL location 1 SE F p Chromosome 1 0.01 19.51 <0.0001 Chromosome 3 0.01 58.51 <0.0001 Chromosome 5 0.01 98.37 <0.0001 Epistasis: 1*3 0.01 0.10 0.75 Epistasis: 1*5 0.01 3.14 0.08 Epistasis: 3*5 0.01 35.29 <0.0001 Epistasis: 1*3*5 0.01 1.41 0.24

88 Table 12. Number that fall within QTL for stamen loss in Arabidopsis thaliana. Candidates are known to be expressed in stamens and are also implicated in stamen production in publications. “Total loci” column includes all known genes within the Bayes 95% credible interval for the QTL, including candidates and non-candidates expressed in stamens.

Chromosome of QTL Total loci Expressed in Candidate stamens s 1 2278 1171 22 3 182 117 2 5 176 100 1

89 Table 13. Details for candidate genes. Locus designations are from TAIR. The four candidates not known to be involved in the GA and/or JA pathways are italicized; the genes on the third and fifth chromosome QTL are bold.

QTL locus Location (kb) function 1 AT1G61680 22,772-22,775 JA-induced 1 AT1G62990 23,337-23,341 JA-induced, downregulated by DELLAs 1 AT1G66140 24,620-24,622 upregulated by DELLAs 1 AT1G66350 24,748-24,750 DELLA protein, opposed by GA 1 AT1G67730 25,391-25,394 transcription upregulated by JA JAG, works with NUB to regulate stamen 1 AT1G68480 25,684-25,686 development 1 AT1G68690 25,789-25,792 JA-induced 1 AT1G69490 26,122-26,123 regulated by BRAHMA 1 AT1G70940 26,743-26,746 PIN3, some mutants have no stamens 1 AT1G73870 27,779-27,781 upregulated by DELLAs 1 AT1G74430 27,975-27,977 JA-induced 1 AT1G74670 28,053-28,054 upregulated by DELLAs 1 AT1G75750 28,441-28,442 upregulated by DELLAs 1 AT1G75900 28,499-28,501 downregulated by DELLAs ASK1, interacts with UFO to affect B 1 AT1G75950 28,517-28,518 genes 1 AT1G76240 28,603-28,604 downregulated by DELLAs 1 AT1G76410 28,669-28,670 JA-induced 1 AT1G76890 28,873-28,875 JA-induced 1 AT1G77450 29,100-29,101 JA-repressed 1 AT1G77590 29,148-29,152 JA-induced 1 AT1G78440 29,512-29,513 downregulated by DELLAs 3 AT3G22060 7,771-7,772 upregulated by DELLAs 3 AT3G22800 8,063-8,065 downregulated by DELLAs initiates /regulates JA response in 5 AT5G08750 2,852-2,855 stamens

90 Table 14. Results of three models in R/qtl. Higher pLOD indicates higher maximum-likelihood estimated model explanatory power. The peaks are in cM. “Effect” is the effects size (2SE) per-allele, so is half the difference between the two homozygotes from different parents. Effect size for the binary model cannot be compared to the others due to the binary phenotype and lack of multiple QTL.

QTL1 QTL3 QTL5 95% 95% 95% Data pLOD peak CI effect LOD peak CI effect LOD peak CI effect LOD

full untrans- 45.25 68.4- 0.05 16.5- 0.08 7.5- 0.17 formed 7 74.4 80.9 (0.02) 4.8 19.7 22.3 (0.02) 12 7.7 8.5 (0.02) 42.3

27.40 68.4- 0.06 15.6- 0.10 6.7- 0.15 no 6 transformed 5 74.4 79.6 (0.02) 4.1 16.5 22.3 (0.02) 9.2 6.6 8.5 (0.02) 26.1 2.6- .-1.0 binary 12.74 NA NA NA NA NA NA NA NA 6 8.6 (0.28) 15.2

91 14). The QTL on chromosome 5 has the largest effect in all three analyses, and is the only QTL in the binary analysis. The analysis in QTL Cartographer found the same three QTL in nearby locations on the same chromosomes with comparable effect sizes (Table 10).

Because there was some variation in the number of plants sampled per line

(Table 9), we excluded the 37 lines with fewer than three plants sampled and reran the analysis as described above on untransformed RIL means. The main results were unchanged.

While there is some segregation distortion in this set of recombinant inbred lines, it is most substantial on chromosome 4 (which does not include a QTL for short stamen loss in our main analysis), and with our large sample size it is not substantial enough to affect the QTL mapping at regardless (Agren et al. 2013).

A. thaliana candidate gene search

To investigate possible biological mechanisms underlying both the main effect

QTL and epistasis, candidate genes were identified using the TAIR Gene Search

(www.arabidopsis.org). Twenty-two of the candidates were in the broad QTL on chromosome 1, two in the chromosome 3 QTL and one in the chromosome 5 QTL.

92

LITERATURE CITED

93 LITERATURE CITED

Abbott, R. J., and M. F. Gomes. 1989. Population genetic structure and outcrossing rate of Arabidopsis thaliana (L) Heyhn. Heredity 62:411–418.

Agren, J., C. G. Oakley, J. K. McKay, J. T. Lovell, and D. W. Schemske. 2013. Genetic mapping of adaptation reveals fitness tradeoffs in Arabidopsis thaliana. Proc. Natl. Acad. Sci. U. S. A. 110:21077–21082.

Alvarez-Buylla, E. R., M. Benítez, A. Corvera-Poiré, Á. C. Cador, S. de Folter, A. G. de Buen, A. Garay-Arroyo, B. Garciá-Ponce, F. JaimesMiranda, R. V. Pérez-Ruiz, A. Piñeyro-Nelson, and S.-C. Y.E. 2010. Flower Development. in The Arabidopsis Book. The American Society of Plant Biologists, Rockville, MD.

Barrett, S. C. H. 2010. Darwin’s legacy: the forms, function and sexual diversity of flowers. Philos. Trans. R. Soc. B Biol. Sci. 365:351–368.

Barrett, S. C. H. 1992. Heterostylous genetic polymorphisms: model systems for evolutionary analysis. Pp. 1–39 in Evolution and Function of Heterostyly. Springer- Verlag, New York.

Barrett, S. C. H. 2002. The evolution of plant sexual diversity. Nat. Rev. Genet. 3:274– 284.

Beck, J. B., H. Schmuths, and B. A. Schaal. 2008. Native range genetic variation in Arabidopsis thaliana is strongly geographically structured and reflects Pleistocene glacial dynamics. Mol. Ecol. 17:902–915.

Bodbyl Roels, S. A., and J. K. Kelly. 2011. Rapid Evolution Caused by Pollinator Loss in Mimulus Guttatus. Evolution 65:2541–2552.

Bomblies, K., L. Yant, R. A. Laitinen, S.-T. Kim, J. D. Hollister, N. Warthmann, J. Fitz, and D. Weigel. Local-Scale Patterns of Genetic Variability, Outcrossing, and Spatial Structure in Natural Stands of Arabidopsis thaliana. Plos Genet. 6.

Broman, K. W., H. Wu, S. Sen, and G. A. Churchill. 2003. R/qtl: QTL mapping in experimental crosses. Bioinformatics 19:889–890.

Burd, M. 1994. Bateman Principle and Plant Reproduction - the Role of Pollen Limitation. Bot. Rev. 60:83–139.

Charlesworth, D., and B. Charlesworth. 1981. Allocation of resources to male and female functions in hermaphrodites. Biol. J. Linn. Soc. 15:57–74.

Charnov, E. L. 1982. The theory of sex allocation. Princeton University Press, Princeton, N.J.

94 Comes, H. P., and J. W. Kadereit. 1998. The effect of quaternary climatic changes on plant distribution and evolution. Trends Plant Sci. 3:432–438.

Conner, J. K., R. Davis, and S. Rush. 1995. The effect of wild radish floral morphology on pollination efficiency by 4 taxa of pollinators. Oecologia 104:234–245.

Conner, J. K., A. M. Rice, C. Stewart, and M. T. Morgan. 2003. Patterns and mechanisms of selection on a family-diagnostic trait: Evidence from experimental manipulation and lifetime fitness selection gradients. Evolution 57:480–486.

Conner, J. K., and S. Rush. 1996. Effects of flower size and number on pollinator visitation to wild radish, Raphanus raphanistrum. Oecologia 105:509–516.

Conner, J. K., S. Rush, and P. Jennetten. 1996. Measurements of natural selection on floral traits in wild radish (Raphanus raphanistrum) .1. Selection through lifetime female fitness. Evolution 50:1127–1136.

Conner, J. K., H. F. Sahli, and K. Karoly. 2009. Tests of adaptation: functional studies of pollen removal and estimates of natural selection on anther position in wild radish. Ann. Bot. 103:1547–1556.

Conner, J., and S. Via. 1993. Patterns of phenotypic and genetic correlations among morphological and life-history traits in wild radish, Raphanus raphanistrum. Evolution 47:704–711.

Coss, R. G. 1999. Effects of relaxed natural selection on the evolution of behavior. Pp. 180–208 in S. A. Foster and J. A. Endler, eds. Geographic variation in behavior: Perspectives on evolutionary mechanisms. Oxford University Press.

Daehler, C. C., and D. R. Strong. 1997. Reduced herbivore resistance in introduced smooth cordgrass (Spartina alterniflora) after a century of herbivore-free growth. Oecologia 110:99–108.

Darwin, C.H. 1859. On the Origin of Species by Means of Natural Selection, or the Preservation of Favoured Races in the Struggle for Life. John Murray, London.

Darwin, C.H. 1862. On the Various Contrivances by which British and Foreign Orchids are Fertilized by Insects. Murray, London.

Darwin, C.H. 1877. The different forms of flowers on plants of the same species. John Murray, London, UK.

Darwin, C.H. 1876. The effects of cross and self fertilization in the vegetable kingdom. John Murray, London.

De Luca, P. A., and M. Vallejo-Marín. 2013. What’s the “buzz” about? The ecology and evolutionary significance of buzz-pollination. Curr. Opin. Plant Biol. 16:429–435.

95 Delesalle, V. A., S. J. Mazer, and H. Paz. 2008. Temporal variation in the pollen:ovule ratios of Clarkia (Onagraceae) taxa with contrasting mating systems: field populations. J. Evol. Biol. 21:310–323.

Elle, E., and R. Carney. 2003. Reproductive assurance varies with flower size in Collinsia parviflora (Scrophulariaceae). Am. J. Bot. 90:888–896.

Endress, P. K. 1992. Evolution and floral diversity - the phylogenetic surroundings of Arabidopsis and Antirrhinum. Int. J. Plant Sci. 153:S106–S122.

Espinasa, L., and W. R. Jeffery. 2006. Conservation of retinal circadian rhythms during cavefish eye degeneration. Evol. Dev. 8:16–22.

Evans, D. E., P. E. Taylor, M. B. Singh, and R. B. Knox. 1991. Quantitative analysis of lipids and protein from the pollen of Brassica napus L. Plant Sci. 73:117–126.

Fong, D. W., T. C. Kane, and D. C. Culver. 1995. Vestigialization and loss of nonfunctional characters. Annu. Rev. Ecol. Syst. 26:249–268.

Fontaine, C., C. L. Collin, and I. Dajoz. 2008. Generalist foraging of pollinators: diet expansion at high density. J. Ecol. 96:1002–1010.

Foxe, J. P., T. Slotte, E. A. Stahl, B. Neuffer, H. Hurka, and S. I. Wright. 2009. Recent speciation associated with the evolution of selfing in Capsella. Proc. Natl. Acad. Sci. 106:5241–5245.

Foxe, J. P., M. Stift, A. Tedder, A. Haudry, S. I. Wright, and B. K. Mable. 2010. Reconstructing origins of loss of self-incompatibility and selfing in North American Arabidopsis lyrata: a population genetic context. Evolution 64:3495–3510.

Friedman, W. E. 2009. The meaning of Darwin’s “abominable mystery.” Am. J. Bot. 96:5–21.

Futuyma, D. J. 2010. Evolutionary constraint and ecological consequences. Evolution 64:1865–1884.

Golding, Y. C., M. S. Sullivan, and J. P. Sutherland. 1999. Visits to manipulated flowers by Episyrphus balteatus (Diptera : Syrphidae): Partitioning the signals of petals and anthers. J. Insect Behav. 12:39–45.

Goodwillie, C., S. Kalisz, and C. G. Eckert. 2005. The evolutionary enigma of mixed mating systems in plants: Occurrence, theoretical explanations, and empirical evidence. Annu. Rev. Ecol. Evol. Syst. 36:47–79.

Goodwillie, C., and J. M. Ness. 2005. Correlated evolution in floral morphology and the timing of self-compatibility in Leptosiphon jepsonii (Polemoniaceae). Int. J. Plant Sci. 166:741–751.

96 Griffin, P. C., and Y. Willi. 2014. Evolutionary shifts to self-fertilisation restricted to geographic range margins in North American Arabidopsis lyrata. Ecol. Lett. 17:484– 490.

Grimaldi, D. 1999. The co-radiations of pollinating insects and angiosperms in the Cretaceous. Ann. Mo. Bot. Gard. 86:373–406.

Guerrant, E. O. 1989. Early maturity, small flowers and autogamy: a developmental connection. Evol. Ecol. Plants 61:84.

Harder, L. D., and S. D. Johnson. 2009. Darwin’s beautiful contrivances: evolutionary and functional evidence for floral adaptation. New Phytol. 183:530–545.

Harder, L. D., and J. D. Thomson. 1989. Evolutionary options for maximizing pollen dispersal of animal-pollinated plants. Am. Nat. 133:323–344.

Hitachi Software Engineering, C., Ltd. 1991. FMBIO Analysis. Hitachi Software Engineering, Co., Ltd.

Hoffmann, M. H., M. Bremer, K. Schneider, F. Burger, E. Stolle, and G. Moritz. 2003. Flower visitors in a natural population of Arabidopsis thaliana. Plant Biol. 5:491– 494.

Jones, A. G. 2005. Gerud 2.0: a computer program for the reconstruction of parental genotypes from half-sib progeny arrays with known or unknown parents. Mol. Ecol. Notes 5:708–711.

Kalinowski, S. T., M. L. Taper, and T. C. Marshall. 2007. Revising how the computer program CERVUS accommodates genotyping error increases success in paternity assignment. Mol. Ecol. 16:1099–1106.

Kevan, P., and H. Baker. 1983. Insects as Flower Visitors and Pollinators. Annu. Rev. Entomol. 28:407–453.

Kudo, G. 2003. Anther arrangement influences pollen deposition and removal in hermaphrodite flowers. Funct. Ecol. 17:349–355.

Lahti, D. C., N. A. Johnson, B. C. Ajie, S. P. Otto, A. P. Hendry, D. T. Blumstein, R. G. Coss, K. Donohue, and S. A. Foster. 2009. Relaxed selection in the wild. Trends Ecol. Evol. 24:487–496.

Le Rouzic, A., K. Østbye, T. O. Klepaker, T. F. Hansen, L. Bernatchez, D. Schluter, and L. A. Vøllestad. 2011. Strong and consistent natural selection associated with armour reduction in sticklebacks. Mol. Ecol. 20:2483–2493.

97 Lewandowska-Sabat, A. M., S. Fjellheim, and O. A. Rognli. 2010. Extremely low genetic variability and highly structured local populations of Arabidopsis thaliana at higher latitudes. Mol. Ecol. 19:4753–4764.

Lloyd, D. G. 1965. Evolution of self-compatibility and racial differentiation in Leavenworthia (Cruciferae). Contrib. Gray Herb. Harv. Univ. 3–134.

Lunau. 2000. The ecology and evolution of visual pollen signals. Plant Syst. Evol. 222:89–111.

Lundemo, S. 2010. Molecular studies of genetic structuring and demography in Arabidopsis from northern Europe. Norwegian University of Science and Technology, Trondheim, Norway.

Luo, Z., D. Zhang, and S. S. Renner. 2008. Why two kinds of stamens in buzz- pollinated flowers? Experimental support for Darwin’s division-of-labour hypothesis. Funct. Ecol. 22:794–800.

Mable, B. K., and A. Adam. 2007. Patterns of genetic diversity in outcrossing and selfing populations of Arabidopsis lyrata. Mol. Ecol. 16:3565–3580.

Mable, B. K., A. V. Robertson, S. Dart, C. Di Berardo, and L. Witham. 2005. Breakdown of self-incompatibility in the perennial Arabidopsis lyrata (Brassicaceae) and its genetic consequences. Evolution 59:1437–1448.

Mable, B. K., M. H. Schierup, and D. Charlesworth. 2003. Estimating the number, frequency, and dominance of S-alleles in a natural population of Arabidopsis lyrata (Brassicaceae) with sporophytic control of self-incompatibility. Heredity 90:422– 431.

Manaster, C. 2010. Allelogram: a program for normalizing and binning microsatellite genotypes.

Matsuhashi, S., S. Sakai, and H. Kudoh. 2012. Temperature-dependent fluctuation of stamen number in Cardamine hirsuta. Int. J. Plant Sci. 173:391–398.

Mazer, S. J., L. S. Dudley, V. A. Delesalle, H. Paz, and P. Galusky. 2009. Stability of pollen-ovule ratios in pollinator-dependent versus autogamous Clarkia sister taxa: testing evolutionary predictions. New Phytol. 183:630–648.

Mezey, J. G., and D. Houle. 2005. The dimensionality of genetic variation for wing shape in Drosophila melanogaster. Evolution 59:1027–1038.

Mione, T., and G. J. Anderson. 1992. Pollen-Ovule Ratios and Breeding System Evolution in Solanum Section Basarthrum (Solanaceae). Am. J. Bot. 79:279–287.

98 Müller, A. 1961. Zur Charakterisierung der Blüten und Infloreszenzen von Arabidopsis thaliana (L.) Heynh. Kulturpflanze 9:264–393.

Müller, F. 1883. Two Kinds of Stamens with Different Functions in the Same Flower. Nature 27:364–365.

Ollerton, J., R. Winfree, and S. Tarrant. 2011. How many flowering plants are pollinated by animals? Oikos 120:321–326.

Ornduff, R. 1969. Reproductive Biology in Relation to Systematics. Taxon 18:121– 133.

Pernet, B. 2003. Persistent Ancestral Feeding Structures in Nonfeeding Annelid Larvae. Biol. Bull. 205:295–307.

Pinheiro, J., Bates, D., DebRoy, S., Sarkar, D., and R Development Core Team. 2013. nlme: Linear and nonlinear mixed effects models.

Platt, A., M. Horton, Y. S. Huang, Y. Li, A. E. Anastasio, N. W. Mulyati, J. Agren, O. Bossdorf, D. Byers, K. Donohue, M. Dunning, E. B. Holub, A. Hudson, V. Le Corre, O. Loudet, F. Roux, N. Warthmann, D. Weigel, L. Rivero, R. Scholl, M. Nordborg, J. Bergelson, and J. O. Borevitz. 2010. The scale of population structure in Arabidopsis thaliana. Plos Genet. 6.

Preston, R. E. 1986. Pollen-ovule ratios in the Cruciferae. Am. J. Bot. 73:1732–1740.

R Core Team. 2012. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria.

Ritland, C., and K. Ritland. 1989. Variation of Sex Allocation Among Eight Taxa of the Mimulus guttatus Species Complex (Scrophulariaceae). Am. J. Bot. 76:1731.

Rush, S., J. K. Conner, and P. Jennetten. 1995. The effects of natural variation in pollinator visitation on rates of pollen removal in wild radish, Raphanus raphanistrum (Brassicaceae). Am. J. Bot. 82:1522–1526.

Sahli, H. F., and J. K. Conner. 2011. Testing for conflicting and nonadditive selection: floral adaptation to multiple pollinators through male and female fitness. Evol. Int. J. Org. Evol. 65:1457–1473.

Sahli, H. F., and J. K. Conner. 2007. Visitation, effectiveness, and efficiency of 15 genera of visitors to wild radish, Raphanus raphanismum (Brassicaceae). Am. J. Bot. 94:203–209.

Sahli, H. F., J. K. Conner, F. H. Shaw, S. Howe, and A. Lale. 2008. Adaptive Differentiation of Quantitative Traits in the Globally Distributed Weed, Wild Radish (Raphanus raphanistrum). Genetics 180:945–955.

99 SAS Institute, I. 2012. JMP, version 10.0.0. SAS Institute, Inc, Cary, NC.

Schneider, C. A., W. S. Rasband, and K. W. Eliceiri. 2012. NIH Image to ImageJ: 25 years of image analysis. Nat. Methods 9:671–675.

Schoener, T. W. 1971. Theory of Feeding Strategies. Annu. Rev. Ecol. Syst. 2:369– 404.

Sicard, A., and M. Lenhard. 2011. The selfing syndrome: a model for studying the genetic and evolutionary basis of morphological adaptation in plants. Ann. Bot. 107:1433–1443.

Sih, A., and B. Christensen. 2001. Optimal diet theory: when does it work, and when and why does it fail? Anim. Behav. 61:379–390.

Smyth, D. R., J. L. Bowman, and E. M. Meyerowitz. 1990. Early flower development in Arabidopsis. Plant Cell 2:755–767.

Stanton, M. L. 1994. Male-male competition during pollination in plant populations. Am. Nat. 144:S40–S68.

Stanton, M. L., A. A. Snow, and S. N. Handel. 1986. Floral Evolution: Attractiveness to Pollinators Increases Male Fitness. Science 232:1625–1627.

Stebbins, G. L. 1957. Self Fertilization and Population Variability in the Higher Plants. Am. Nat. 91:337–354.

Stinchcombe, J. R., C. Weinig, M. Ungerer, K. M. Olsen, C. Mays, S. S. Halldorsdottir, M. D. Purugganan, and J. Schmitt. 2004. A latitudinal cline in flowering time in Arabidopsis thaliana modulated by the flowering time gene FRIGIDA. Proc. Natl. Acad. Sci. U. S. A. 101:4712–4717.

Vallejo-Marin, M., E. M. Da Silva, R. D. Sargent, and S. C. H. Barrett. 2010. Trait correlates and functional significance of heteranthery in flowering plants. New Phytol. 188:418–425.

Vallejo-Marin, M., J. S. Manson, J. D. Thomson, and S. C. H. Barrett. 2009. Division of labour within flowers: heteranthery, a floral strategy to reconcile contrasting pollen fates. J. Evol. Biol. 22:828–839.

Walker-Larsen, J., and L. D. Harder. 2000. The evolution of staminodes in angiosperms: Patterns of stamen reduction, loss, and functional re-invention. Am. J. Bot. 87:1367–1384.

Willi, Y., and K. Maattanen. 2010. Evolutionary dynamics of mating system shifts in Arabidopsis lyrata. J. Evol. Biol. 23:2123–2131.

100 Yoshizawa, M., Y. Yamamoto, K. E. O’Quin, and W. R. Jeffery. 2012. Evolution of an adaptive behavior and its sensory receptors promotes eye regression in blind cavefish. BMC Biol. 10.

Young, H. J., and M. L. Stanton. 1990. Influences of Floral Variation on Pollen Removal and Seed Production in Wild Radish. Ecology 71:536–547.

Zhen, Y., and M. C. Ungerer. 2008. Clinal variation in freezing tolerance among natural accessions of Arabidopsis thaliana. New Phytol. 177:419–427.

Zomlefer, W. B. 1994. Guide to Flowering Plant Families. University of North Carolina Press, Chapel Hill, NC.

101