Phenotypic selection on floral traits in the arctic plant Parrya nudicaulis (Brassicaceae)

Abstract The evolution of floral traits is often attributed to pollinator‐mediated selection; however, the importance of pollinators as selective agents in arctic environments is poorly resolved. In arctic and subarctic regions that are thought to be pollen limited, selection is expected to either favor floral traits that increase pollinator attraction or promote reproductive assurance through selfing. We quantified phenotypic selection on floral traits in two arctic and two subarctic populations of the self‐compatible, but largely pollinator‐dependent, Parrya nudicaulis. Additionally, we measured selection in plants in both open pollination and pollen augmentation treatments to estimate selection imposed by pollinators in one population. Seed production was found to be limited by pollen availability and strong directional selection on flower number was observed. We did not detect consistently greater magnitudes of selection on floral traits in the arctic relative to the subarctic populations. Directional selection for more pigmented flowers in one arctic population was observed, however. In some populations, selection on flower color was found to interact with other traits. We did not detect consistently stronger selection gradients across all traits for plants exposed to pollinator selection relative to those in the pollen augmentation treatment; however, directional selection tended to be higher for some floral traits in open‐pollinated plants.

Furthermore, the non-graminoid arctic vascular flora as a whole contains a relatively high percentage of anthocyanin-pigmented taxa with many capable of producing nectar and scent (Jaakola & Hohtola, 2010;Whittall & Carlson, 2009), which is suggestive of pollinator-mediated selection contributing to the maintenance of those traits.
Here, we estimate the magnitude of phenotypic selection on floral traits (flower number, petal size, corolla depth, anther height, and flower color), using seed set as a proxy for fitness in the arctic and subarctic mustard, Parrya nudicaulis (Brassicaceae). High within-population variation in flower size, petal orientation, and pigmentation is common in P. nudicaulis (Figure 1). This species is largely pollinator dependent and severely pollen limited (Fulkerson et al., 2012). With the use of pollen augmentation and control treatments in P. nudicaulis (see Sandring & Ågren, 2009), we predict that the strength of pollinator-mediated selection is greater than nonpollinator-mediated selection on floral traits associated with increased pollinator attraction. Last, as pollinator service is expected to be of poorer quality in arctic relative to subarctic populations, we predict that phenotypic selection on floral traits is greater in the more northerly populations.

| Study system
Parrya nudicaulis L. Regel (Brassicaceae) is found from northeastern Asia, across Alaska and to the western Canadian Arctic Archipelago (Al-Shehbaz, 2010;Hultén, 1968). Flowering occurs in late May to mid-June in subarctic sites in Alaska and several weeks later on the Arctic Coastal Plain. At reproductive maturity, this perennial herb produces a single raceme of 8-14 flowers, which normally persists between 10-14 days with individual flowers senescing after 3 days.
Flowers are protandrous; the upper anthers dehisce shortly after the flowers open, followed by the lower two anthers within approximately 12 h, and the stigma becomes bilobed and receptive during the second day. Flower color of P. nudicaulis is highly variable among individuals in many populations (Butler et al., 2014). While the hue is quite consistent, the lightness values range dramatically among individuals. Flowers range from pure white, and produce no anthocyanins, to dark violet with substantial anthocyanin production (Dick et al., 2011). Most flowers emit a sweet fragrance comparable to Syringa species. Nectar is secreted at the base of the corolla and less than 4 µl is produced in plants bagged for 24 h (Fulkerson et al., 2012). Floral visits to P. nudicaulis at the studied populations in Alaska are infrequent (mean of 0.14 and 0.58 visits/flower/hour in 2009 and 2010), and although a diversity of visitors drink nectar and collect pollen on P. nudicaulis, muscid and syrphid flies make up the largest proportion of floral visitors (Fulkerson et al., 2012).
This study was conducted at two arctic and two subarctic pop- Provence and are a graminoid tundra habitat dominated by tussock sedge, dwarf shrubs, and moss and lichens (Raynolds et al., 2005).  June. Infructescences were collected at the end of July, prior to seed dehiscence.

| Pollination treatments
To remove the component of phenotypic selection due to pollinator visitation, mixed pollen from at least 10 haphazardly selected individuals that were >10 m distance from the recipient were used to hand-pollinate flowers. Phenotypic selection was not estimated from plants that served as pollen donors. Manipulated flowers were marked with a small amount of "puffy paint" at the base of the pedicel. Every flower was hand-pollinated every day, until there were signs of flower senescence to ensure that stigma receptivity was not missed. Supplemental pollen added to the entire inflorescence reduces the chance of differential resource allocation interfering with the detection of pollen limitation Knight et al., 2006;Zimmerman & Pyke, 1988). The fate of all flowers was followed to estimate probability of seed set and fecundity for each plant.

| Phenotypic measurements
We used the measurements of six floral traits that we expected could be under pollen-mediated selection: flower number, petal width, petal length, corolla depth, anther height, and flower color. Petal length was highly correlated with petal width (r = 0.65 p < .001), and both measurements were reflective of flower size, and therefore to reduce multicollinearity, petal length was not included in the analysis. Pistil height was correlated with corolla depth and the stigma became receptive when it neared the corolla opening; we did not measure pistil position to avoid contact or damage to the stigma.
We counted the total number of flowers produced at the end of the flowering season. All other traits were measured at anthesis when the flowers were fully open and anthers were accessible to pollinators. We measured the width and lengths of the largest petal, corolla depth, and height of the tallest anthers to the nearest 0.01 mm with a digital caliper at Eagle Summit, Galbraith, and the 2009 Ivishak plants. To capture a large enough sample with limited time, at the Twelve-Mile and 2010 Ivishak populations, we measured corolla depth and anther height with digital calipers, but we measured petal F I G U R E 2 Variance-standardized linear (a and c) selection gradients (β σ ) and non-linear (b and d) selection gradients (γ σ ) for plants subjected to pollen-mediated selection (black squares) and pollen-augmented plants (open squares) on probability of seed set and fecundity at Eagle Summit 2010. Bars display the 95% CI length and width using digital photographs of individual flowers with a scale bar; measurements were subsequently made in ImageJ (Rasband, 1997(Rasband, -2018 image analysis software. Means and variance measurements of all traits are summarized in Tables S1-S5. A Royal Horticultural Society Colour Chart (RHS, 2007) was used to quantify the variation in flower color between plants at the time of anthesis. Using this chart, however, limits the factor of "color" to categorical data. To determine lightness values of the color chips, we used the techniques followed by Fulkerson et al. (2012) to create CIE L* values: L* values range from 0 to 100, where "0" is black or "nearblack" and "100" is white or "near-white" (see Stevens et al., 2007;Voss, 1992). Flower color was characterized by a total of 18 color chips in these populations and ranged from L* value 59.5 to 99.5.
Parrya nudicaulis petals fall within a narrow range of purple-violet of the RHS Colour Chart, and L* is highly correlated with anthocyanin concentration (J. B. Whittall, unpublished data).

| Selection analysis
The strength and direction of selection on the floral traits were measured using a multivariate regression analytic framework (Lande & Arnold, 1983). We used variance-standardized partial linear regression coefficients to estimate the strength of directional selection on traits independent of all other measured traits (i.e., selection gradients, β σ ) (Lande & Arnold, 1983). Additionally, we calculated mean-standardized selection coefficients (β μ ), as this metric has been shown to avoid the problem of conflating selection and variation and it is particularly useful for summarizing the strength of selection for diverse traits, and for facilitating a more accurate estimate of response to selection (see Hereford et al., 2004). Trait standardizations were made for the individuals used in each regression model.
Mean-standardized results are presented in the Tables S1-S5. The number of individuals was not sufficient to measure nonlinear selection (convex or concave) for all populations, although the sample size approached recommended levels for Eagle Summit open-pollinated and pollen augmentation treatments (see Walsh & Lynch, 2018). We therefore quantified nonlinear selection and correlational selection for variance-standardized traits at Eagle Summit between pairs of traits using quadratic (γ ii ) and 15 cross-product (γ ij ) terms in the regression model (Sandring & Ågren, 2009). These regression coefficients were multiplied by 2 to derive the nonlinear selection coefficients (Stinchcombe et al., 2008). Fitness was estimated by two separate values: the probability of producing seed and fecundity for those individuals which produced seed. Thus, the first fitness metric separates plants that had reproductive failure to those that reproduced (i.e., either received insect visitation or self-fertilized). The second fitness metric encompasses the quality and quantity of pollination of plants that did reproduce. This approach also facilitates use of different regression models without violating assumptions. These fitness values were relativized by dividing by the population mean.
Multiple logistic regression was used to estimate selection on the probability of seed set due to the dichotomous nature of this fitness measure (Janzen & Stern, 1998). Binomial logistic regression coefficients were transformed into linear regression coefficients using the methods of Janzen and Stern (1998). Secondarily, we measured selection gradients on those individuals that did set seed at the experimental population at Eagle Summit and Ivishak using standard multiple regression methods. Contrasts in the magnitude and direction of selection gradients between open-pollinated and pollenaugmented treatments and among arctic and subarctic populations were compared with means and 95% confidence intervals to avoid the pitfalls of null hypothesis significance testing (Anderson et al., 2000;Fidler et al., 2006;Rinella & James, 2010). All analyses were conducted using R version 2.12 (R Development Core Team, 2011). Twelve-Mile, produced 5.33 ± 0.91 SE seeds/plant. In 2009, the arctic sites at Galbraith and Ivishak produced 2.00 ± 0.59 SE seeds/ plant and 6.26 ± 0.87 SE seeds/plant, respectively. In 2010, seed production was 9.95 ± 0.40 SE seeds/plant at the Ivishak population.  Table 1A).

| RE SULTS
In the pollen-augmented treatment, the probability of seed set was greater for individuals with shorter corolla tubes (Figure 2, Table 1A). For those individuals that set seed in the pollen augmentation treatment, fecundity was also lowest for individuals with intermediate anther position (i.e., disruptive selection, Figure 1, Table 1B).  (Table 4).
In a New Zealand alpine plant, the strength of selection on flower color (whiter flowers had greater fitness) was stronger under a lower pollination limitation treatment than when pollen was more limiting (Campbell & Bischoff, 2013); however, in this case non-pollinator- Selection on floral traits can also occur when pollen limitation is absent (Galen, 1996;Parachnowitsch & Kessler, 2010). In this study of P. nudicaulis, while we did not detect consistently Selection gradients at the arctic Ivishak population were also stronger in the year with less favorable weather and much lower natural seed set. The 2009 flowering season at Ivishak was marked with a wet, windy, and cold climate that would likely limit insect flight time and pollinator availability (Bergman et al., 1996;Totland, 1994). In Additionally, our selection results may be underestimates since they do not include the male component of fitness. Male fitness is also expected to increase with increasing number of flowers; unfortunately, selection on male fitness is rarely studied, despite its importance (Sutherland & Delph, 1984). Plants containing a greater number of ovules than are on average fertilized have been hypothesized to benefit from occasional "jackpot" chance visits in environments with highly stochastic pollinator visits Burd et al., 2009). Pollinator visits to P. nudicaulis in tundra habitats occur at much lower rates and depend on windows of favorable climate compared to plants in temperate habitats (Fulkerson et al., 2012). Indeed, an increase in ovule number would be beneficial for occasional pollinator visits, but an increase in flower number would further enhance the probability of seed set for an individual through geitonogamy, as well as presumably promoting pollen export (male fitness). The greater strength of pollinator-mediated, relative to nonpollinator-mediated, selection on flower number is consistent with our prediction of selection favoring traits associated with enhanced pollinator attraction. Contrary to our prediction, however, we did not detect selection for larger petal size. Pollinators have been shown to prefer flowers with larger petals and corollas in a number of other studies Campbell et al., 1996;Galen, 1996;Gómez, 2003;Parachnowitsch & Kessler, 2010;Sandring & Ågren, 2009;. However, the pollinator guilds of the arctic and subarctic habitats are moderately diverse, generally dominated by flies, and dissimilar from previously studied regions (Fulkerson et al., 2012;Tiusanen et al., 2016;Tiusanen et al., 2019), and phenotypic selection is typically higher in plants pollinated by bees, long-tongued flies, or birds (Caruso et al., 2019). Additionally, directional selection on the size of the corolla or pollination unit (e.g., capitulum in Asteraceae) has not always been detected (Parachnowitsch et al., 2012), even when seed set is significantly pollen limited (Andersson & Widén, 1992;. It is possible that some of the apparent phenotypic selection on flower size observed in these studies could be a product of covariation in ovule number (see Hansen et al., 2003;however, see Stanton & Preston, 1988). In P. nudicaulis, we suspect that flower size has a minor impact on the overall floral display perceived by pollinators TA B L E 1 A Variance-standardized linear (β σ ) gradients (and 95% confidence intervals in parentheses) for open-pollinated and pollen augmentation treatments using logistic regression on probability of seed set, and multiple linear regression on fecundity (seed number) for those individuals that did set seed at Eagle Summit in 2010 Note: Gradients marginally and significantly different from 0 are shown in bold (*p < .10 > .05; **p < .05 > 0.01; ***p < .01). Probability of seed set selection gradients is transformed from logistic regression coefficients using the method of Janzen and Stern (1998). Regression coefficients for γ matrix diagonals were multiplied by 2 to calculate concave and convex gradients. Positive γ values indicate concave (disruptive) selection, and negative values indicate convex (stabilizing) selection. The regression model included all five traits and fifteen cross-product terms. Probability of seed set selection gradients is transformed from logistic regression coefficients using the method of Janzen and Stern (1998).
and unmeasured traits such as scent production may be significantly more important in pollinator perception (Parachnowitsch et al., 2012).
An alternative prediction to selection for enhanced pollinator attraction in pollen-limited environments could be selection for increased capacity for selfing. Parrya nudicaulis is protandrous and largely pollinator dependent, but a small frequency of flowers that had pollinators excluded do set seed (Fulkerson et al., 2012); thus, it suggests that sufficient phenotypic variation exists in traits associated with selfing to respond to selection in these populations.
We did not, however, detect directional selection on reduced petal size, lower anther position, or reduced floral pigmentation in openpollinated plants in either arctic or subarctic populations. It is also possible for herbivores, seed predators, and resource costs to generate selection on reduced floral displays (Caruso et al., 2019;Descamps et al., 2021;Galen, 1999). The arctic population expe- latitudes (and altitudes) as weather and climate appropriate for pollinator service declines (Bergman et al., 1996;Inouye, 2020;Totland, 1994). In a number of cases we detected greater selection in the arctic populations; however, the strength of selection was inconsistent among populations, traits, and years. Often the direction of linear selection was divergent for traits (or trait combinations) between the arctic and subarctic populations. The arctic populations are approximately 400 km to the north of the subarctic populations and have substantially lower mean July temperatures on average (Dick et al., 2011) that would be expected to be associated with reduced pollinator activity; however, the subarctic sites are at a higher elevation and are also often subjected to inclement weather. Year-to-year variation in weather is likely to make detection of regional patterns in selection gradients difficult to detect.
We provide modest evidence of stronger selection gradients for pigmentation at the higher latitudes compared to the lower-latitude sites, where darker violet individuals had higher fecundity. Indeed, at a population level, anthocyanin pigmentation of P. nudicaulis increases in frequency with increasing latitude (Dick et al., 2011).
Flower color did not affect the probability of seed set at the Ivishak population, but selection coefficients for darker flower color were strongly significant with a greater number of flowers and marginally on its own. Flower color did not enhance the probability of seed set unless it interacted with another trait in the other sites. Selection on flower color can not only be a result of herbivores, pathogens, or abiotic factors directly acting on the trait but also be a result of indirect selection through correlated traits (Campbell & Bischoff, 2013;Frey, 2004;Rausher, 2008;Strauss & Whittall, 2006). Note: Gradients marginally and significantly different from 0 are shown in bold (·p < .10 > .05; *p < .05 > .01; **p < .01). Regression coefficients for γ matrix diagonals were multiplied by 2 to calculate concave and convex gradients. The regression model for fitness estimated by probability of seed set included only the five traits, as most individuals set seed in this year and site, limiting confidence in estimates of regression coefficients. The regression model of fitness estimated through fecundity, however, had sufficient sample size to include all five traits and fifteen cross-product terms. Probability of seed set selection gradients is transformed from logistic regression coefficients using the method of Janzen and Stern (1998).

DATA AVA I L A B I L I T Y S TAT E M E N T
Data will be accessible to the public through our Alaska Center for