Plasticity of plant defense and its evolutionary implications in wild populations of Boechera stricta

Abstract

Phenotypic plasticity is thought to impact evolutionary trajectories by shifting trait values in a direction that is either favored by natural selection (“adaptive plasticity”) or disfavored (“nonadaptive” plasticity). However, it is unclear how commonly each of these types of plasticity occurs in natural populations. To answer this question, we measured glucosinolate defensive chemistry and reproductive fitness in over 1,500 individuals of the wild perennial mustard Boechera stricta, planted in four common gardens across central Idaho, USA. Glucosinolate profiles—including total glucosinolate concentration as well as the relative abundances and overall diversity of different compounds—were strongly plastic both among habitats and within habitats. Patterns of glucosinolate plasticity varied greatly among genotypes. Plasticity among sites was predicted to affect fitness in 27.1% of cases; more often than expected by chance, glucosinolate plasticity increased rather than decreased relative fitness. In contrast, we found no evidence for within-habitat selection on glucosinolate reaction norm slopes (i.e., plasticity along a continuous environmental gradient). Together, our results indicate that glucosinolate plasticity may improve the ability of B. stricta populations to persist after migration to new habitats.

Introduction

The role of phenotypic plasticity in adaptive evolution has been a subject of great controversy and research interest for decades (Bradshaw 1965; Via and Lande 1985; Via et al. 1995; Pigliucci 2005; Ghalambor et al. 2015; Hendry 2015). It has long been recognized that both an organism’s genotype and its environment shape its phenotype (Supplementary Figure 1), which then determines its evolutionary fitness. Phenotypic variation caused by environmental stimuli is not heritable and therefore cannot result in evolution through systematic changes in allele frequencies (Falconer and Mackay 1996). Nevertheless, plasticity is predicted to impact evolution by shifting phenotypes that are under natural selection (Bradshaw 1965; Supplementary Figure 2). Furthermore, if patterns of plasticity are genetically variable, then plasticity itself may evolve in response to selection (Gomulkiewicz and Kirkpatrick 1992). It remains unclear how commonly these phenomena occur in natural populations, and whether the adaptive value of plasticity varies for different traits, environments, and spatial scales.

One way that plasticity could impact evolution is by accelerating or hindering adaptation to a novel environment—e.g., upon migration to a new habitat or in response to a relatively sudden ecosystem shift, as might result from climate change (Donohue et al. 2001; Richards et al. 2006; Ghalambor et al. 2007; Anderson et al. 2012). Plasticity that moves a phenotype in a direction favored by selection is often called “adaptive” plasticity because it increases fitness relative to a non-plastic genotype (Supplementary Figure 2); however, whether this type of plasticity actually facilitates genetic adaptation is controversial. Strong adaptive plasticity could profoundly increase relative fitness, removing the selective force that would otherwise drive adaptation through allele frequency change and thereby inhibiting local adaptation. Alternatively, moderate adaptive plasticity may enable a population to survive in the new environment long enough for selection to increase the frequency of beneficial alleles, thus promoting local adaptation (Baldwin 1896; Price et al. 2003; Ghalambor et al. 2007). The opposite pattern—in which “nonadaptive” plasticity moves phenotypes farther from the new optimum—may either increase the risk of extinction or lead to rapid adaptive evolution by intensifying natural selection (Conover and Schultz 1995; Ghalambor et al. 2007; Ghalambor et al. 2015; Huang and Agrawal 2016). In this manuscript, we do not attempt to determine whether plasticity constrains or facilitates long-term genetic adaptation. Rather, our goal is to assess the relative frequency of adaptive versus nonadaptive plasticity in natural populations.

A related but distinct question is whether, and how, plasticity might evolve as an adaptation to environmental heterogeneity within a single habitat. Plasticity in response to fine-scale environmental variation is often imagined as a reaction norm, with the trait value as some function of a continuous environmental predictor (Supplementary Figure 3; Schmalhausen 1949). For traits that exhibit genotype-by-environment interactions, genetic variation exists for reaction norm shape. Natural selection acting on variation for plasticity can be detected using established quantitative genetic methods such as genotypic selection analysis on reaction norm coefficients (Supplementary Figure 3; Weis and Gorman 1990; Rausher 1992; Baythavong and Stanton 2010). Natural selection is predicted to result in increased plasticity if patterns of plasticity are heritable, if the spatial scale of changing selective pressures is similar to the organism’s dispersal distance (Levins 1962; Gomulkiewicz and Kirkpatrick 1992; Baythavong 2011), if reliable environmental cues for the selection pressure are available (Levins 1963; Donohue et al. 2000; Schmitt et al. 2003; Reed et al. 2010), and if costs of plasticity are minimal (Auld et al. 2010).

Despite several excellent empirical studies (Dudley and Schmitt 1996; Schmitt et al. 1999; Donohue et al. 2000; Donohue et al. 2001; Sultan 2001; Baythavong 2011), more examples from natural populations are needed to test theoretical predictions about the fitness impacts of both between-environment and within-environment plasticity (Hendry 2015). Data on the plasticity and evolution of physiological traits (as opposed to morphological or life-history traits) is particularly scarce (Palacio-López et al. 2015). Because variation in phytochemistry may affect not only the evolution of the plant but also entire communities and ecosystems (Wimp et al. 2007; Hopkins et al. 2009), phytochemical plasticity has been identified as a high priority research target (Hendry 2015). Here, we address these needs by studying plasticity and evolution of glucosinolate defensive chemistry in the wild perennial herb Boechera stricta (Graham) Al-Shehbaz, a close relative of Arabidopsis. Goals of this study were (1) to characterize genotype-by-environment interactions underlying glucosinolate variation in B. stricta, (2) to assess whether glucosinolate plasticity alters relative fitness after transition to novel habitats, and (3) to test whether natural selection acts on glucosinolate reaction norms within habitats.

We measured glucosinolate profiles, size, and fecundity of 25 B. stricta genotypes replicated in 80 experimental blocks divided among four common gardens in diverse habitats (Figure 1). Because Boechera has limited dispersal (<0.5 m on average; Bloom et al. 2002), the environmental variation encompassed by the widely separated common gardens is much greater than what individual B. stricta populations normally encounter; thus, plasticity among field sites describes plasticity after a sudden environmental change or migration to a new habitat. To assess whether glucosinolate plasticity among habitats exhibits an “adaptive” or “nonadaptive” pattern, we compared the direction of selection in each site with the direction of diverse genotypes’ plastic responses to that site (Supplementary Figure 2). Then, we quantified within-habitat glucosinolate plasticity and assessed its relationship to fitness in each habitat using genotypic selection analysis on reaction norm coefficients (Supplementary Figure 3). We found that substantial genotype-by-environment interactions underlie glucosinolate variation in B. stricta, and plasticity among sites tended to move trait values in an adaptive direction; however, we did not detect selection on glucosinolate plasticity within habitats.

Figure 1.

Field sites and experimental design. (a) Map of field sites and wild B. stricta populations used in common garden experiment. Common gardens are denoted with white triangles and labeled. Circles mark collection sites of the 25 genotypes included in the experiment. Not shown: one Eastern genotype collected in Colorado. Map data: Google. (b) Each common garden contained 16 to 22 experimental blocks. Each randomized block contained two individuals of each of the 25 genotypes, planted in a 10-cm grid. Panel (c) shows a block placed within the natural vegetation; (d) shows one experimental rosette with its identifying tag.

Methods

All statistical analyses were performed in R version 3.3.2 (R Core Team 2016) with heavy use of the packages ggplot2, lme4, lmerTest, dplyr, tidyr, and stringr (Wickham 2009; Bates et al. 2015; Kuznetsova 2015; Wickham and Francois 2015; Wickham 2016a,b). Throughout, P-values were corrected for multiple comparisons using the sequential Bonferroni correction (Holm 1979). Additional details for all sections are available in Supplementary Methods. All data and R code are available for download in a Dryad repository (doi:10.5061/dryad.f9s4424; Wagner & Mitchell-Olds, 2018).

Study system

The short-lived perennial herb Boechera stricta (Graham) Al-Shehbaz is common in montane meadows and forests throughout its native range in western North America (Rushworth et al. 2011). Natural populations are strongly genetically differentiated (F_ST=0.56; Song et al. 2006) and have adapted to diverse habitats that vary in climate, water availability, elevation, soil composition, plant community diversity and density, and microbial community composition (Supplementary Figure 4; Anderson et al. 2013a, b; Wagner et al. 2016).

B. stricta produces a variety of glucosinolates, which are sulfur-rich, biologically active phytochemicals that protect against generalist insect herbivores and pathogens and may also affect nonpathogenic root-associated microbes (Agrawal 2000; Tierens et al. 2001; Brader et al. 2006; Halkier and Gershenzon 2006; Bednarek et al. 2009; Bressan et al. 2009; Clay et al. 2009; Hopkins et al. 2009; Schranz et al. 2009; Sanchez-Vallet et al. 2010). Glucosinolates are constitutively produced, although attack by natural enemies often induces additional production (Agrawal 1998, 2000; Brader et al. 2001; Agrawal et al. 2002; Textor and Gershenzon 2009; Abdel-Farid et al. 2010; Manzaneda et al. 2010). B. stricta produces four aliphatic glucosinolate compounds with varying biological activity (Figure 2a; Windsor et al. 2005; Schranz et al. 2009; Prasad et al. 2012). Total glucosinolate concentration and relative abundances of these compounds vary extensively among individuals (Figure 2b–c).

Figure 2.

Glucosinolate variation in 1,505 field-grown Boechera stricta rosettes, sorted by increasing BC-ratio. (a) Chemical structures of the four primary glucosinolates in B. stricta (Kanehisa et al. 2002), with their amino acid precursors. 6MSOH is derived from methionine; the other compounds are all derived from branched-chain amino acids. Individuals rarely produce both 1MP and 2OH1ME; the exact reason for this is currently unknown, but appears to be related to population structure. (b) The proportions of four aliphatic glucosinolates are shown for each measured plant. (c) Three emergent properties of each glucosinolate profile are shown for each individual, which were calculated from the directly measured concentrations of the four compounds shown in panel (b). Three genotypes in this study lack branched-chain glucosinolate functionality and only produce 6MSOH; individuals of these genotypes are seen in the lower left-hand corner of panel (c).

Design and installation of field experiment

In October 2013, we planted 4,000 self-full siblings of 25 naturally inbred Boechera stricta genotypes (Supplementary Table 1) in fully randomized blocks (two replicates per genotype per block) in four common gardens in central Idaho (Figure 1). These field sites are all home to wild B. stricta populations, and are distinguished by many biotic and abiotic environmental characteristics (Supplementary Table 2; Supplementary Figure 4). Each genotype was derived from an accession from one wild B. stricta population (Figure 1a; Supplementary Table 1), which we propagated by self-fertilization in standard greenhouse conditions to minimize variation caused by maternal environmental effects. These genotypes represent the breadth of B. stricta genetic diversity, comprising 12 from the West subspecies and 13 from the East subspecies (Lee and Mitchell-Olds 2013). Because B. stricta primarily self-pollinates and is naturally inbred (F_IS=0.89; Song et al. 2006), self-full siblings are essentially genetically identical. Therefore, phenotypic variation among individuals of the same genotype describe that genotype’s plastic response to environmental variation.

Measurement of plant performance in the field

During summer 2014, we returned to each site several times to measure survival, developmental stage, and height. Mid-summer, we measured insect herbivory on each plant by recording the total number of leaves, the number of damaged leaves, and a visual estimation of the percent of leaf area that was damaged on an average damaged leaf. From these records we calculated overall herbivory as:

$$\mathit{\text{Percent Damage}}=\frac{\mathit{\text{Number of damaged leaves}}\ast \mathit{\text{Average }}\%\mathit{\text{ damage}}}{\mathit{\text{Total number of leaves}}}$$

We measured herbivory separately for rosette leaves and cauline (stalk) leaves.

At the end of the growing season, we measured fruit production for each surviving individual. Survival ranged from 44.4% at Jackass Meadow to 71.0% at Silver Creek; a moderate proportion of the survivors successfully set fruit, ranging from 22.0% at Jackass Meadow to 59.7% at Silver Creek. Because B. stricta is a perennial that can reproduce more than once, our fecundity data may underestimate lifetime fitness; however, previous multi-year field experiments have found that most plants produce most of their lifetime fruit set in their first reproductive event (T. Mitchell-Olds, unpublished). Because B. stricta is predominantly self-pollinating (Song et al. 2006), fruit production reflects both male and female fecundity, and thus is a good estimate of reproductive fitness for a given year. Previous work has confirmed that fruit output is a strong predictor of seed production in this species (A. Manzaneda 2008, unpublished).

Although glucosinolates are frequently studied for their protective effects against herbivory, particularly in B. stricta (Schranz et al. 2009; Prasad et al. 2012), these phytochemicals may have additional important functions that are not mediated by insects—for instance, pathogen resistance (Brader et al. 2006; Clay et al. 2009) and responses to drought and other abiotic stressors (Khokon et al. 2011; del Carmen et al. 2013; Martínez-Ballesta et al. 2015). Furthermore, B. stricta genotypes vary in their ability to tolerate and regrow after insect damage, complicating the relationship between herbivory and evolutionary fitness (Manzaneda et al. 2010). For these reasons, we focus on true reproductive fitness in this study, rather than herbivore damage. For phenotypic selection analyses (below), we used fecundity (in mm of fruit produced) as a measurement of reproductive fitness:

$$\mathit{\text{Fecundity}}=\mathit{\text{Number of fruits}}\times \mathit{\text{Length of average fruit}}$$

Surviving individuals that did not set fruit were assigned a fecundity of zero. For estimation of genotypic fitness, we also calculated the probability of survival for each genotype l :

$$\mathit{P}(\mathit{\text{survival}})=\frac{\mathit{\text{number of surviving individuals of genotype l}}}{\mathit{\text{number of individuals of genotype l originally planted}}}$$

We then calculated the total evolutionary fitness for each genotype as:

$$\mathit{w}=\mathit{P}(\mathit{\text{survival}})\times \mathit{\text{mean}}(\mathit{\text{fecundity of survivors}})$$

Measurement of glucosinolate profiles

Because insect attack can induce additional production of glucosinolates (Agrawal 1998), we measured glucosinolate profile as early as possible in the summer, before peak herbivory. Although it is unlikely that herbivore damage was completely absent at the time of glucosinolate measurement, damage levels were too low to justify the effort of scoring them at this early timepoint. Therefore, our glucosinolate measurements mostly reflect constitutive levels of chemical defenses, although we cannot rule out the possibility that induction by early-season natural enemies contributed some of the observed variation in these traits, including plastic variation within and among field sites.

On the earliest census date for each site, we collected ~20–30 mg of rosette leaf tissue from each surviving plant into tubes containing 70% methanol. Samples were shipped to Duke University, then fully randomized onto 96-well plates. Glucosinolates were extracted from the methanol leachates using established protocols (Supplementary Methods). We used high-performance liquid chromatography (HPLC) to measure the abundance of four aliphatic glucosinolates (Figure 2a–b) in each sample. Three of these compounds (1ME, 1MP, 2OH1ME) have branched-chain structures and biological activity that differs from that of the fourth, straight-chain compound (6MSOH; Figure 2a; Schranz et al. 2009; Prasad et al. 2012). We calculated absolute concentrations (µmol per mg dry weight) of each compound by comparing each peak to an internal standard and dividing by the dry weight of each leaf sample (Supplementary Methods).

From the absolute concentrations of all four compounds, we calculated three summary metrics for each sample’s glucosinolate profile:

$${\mathit{\text{Total}}}_{[\mathit{\text{AGS}}]}=[2\mathit{\text{OH}}1\mathit{\text{ME}}]+[1\mathit{\text{ME}}]+[1\mathit{\text{MP}}]+[6\mathit{\text{MSOH}}]$$

$${\mathit{\text{BC}}}_{\mathit{\text{ratio}}}=\frac{[2\mathit{\text{OH}}1\mathit{\text{ME}}]+[1\mathit{\text{ME}}]+[1\mathit{\text{MP}}]}{{\mathit{\text{Total}}}_{[\mathit{\text{AGS}}]}}$$

$${\mathit{\text{BC}}}_{\mathit{\text{diversity}}}=-\sum _{i=1}^k[{\mathit{\text{BC}}}_i]\ast \text{log }([{\mathit{\text{BC}}}_i])$$

where k = the total number of branched-chain compounds present in the sample and [BC_i] = the concentration of the i^th branched-chain glucosinolate. Total [AGS] describes the combined concentration of all aliphatic glucosinolates. BC-ratio describes the proportion of aliphatic glucosinolates that are derived from branched-chain amino acids, which is an ecologically and evolutionarily important trait in B. stricta due at least in part to its strong influences on herbivory resistance as well as drought tolerance (Schranz et al. 2009; Manzaneda et al. 2010; Prasad et al. 2012; T. Mitchell-Olds, unpublished). Finally, BC-diversity describes the balance of the three types of branched-chain glucosinolates, taking low values when glucosinolate profiles are dominated by one compound and high values when multiple compounds are present in similar amounts (Figure 2b–c). We calculated BC-diversity using the Shannon diversity index in the R package vegan (Oksanen et al. 2013).

Partitioning variance in glucosinolate profiles

To assess plasticity of glucosinolate profiles among habitats, we used univariate REML mixed models and ANCOVA to partition variance in each of the three glucosinolate traits among genetic and environmental predictors. We modeled each trait as:

$$\mathit{\text{Trait}}=\mathit{\text{Genotype}}+\mathit{\text{Site}}+\mathit{\text{Genotype}}\ast \mathit{\text{Site}}+\mathit{\text{Block}}(\mathit{\text{site}})+\mathit{\text{Genotype}}\ast \mathit{\text{Block}}(\mathit{\text{Site}})+\mathit{\text{Plant height}}+\mathit{\text{Developmental stage}}+\mathit{\text{Batch}}+\mathit{\text{error}}$$

where Height, Developmental Stage, and Batch were nuisance variables to control for (respectively) the “general vigor problem” of large plants having more resources to invest in defense (Agrawal 2011), ontogenetic changes in rosette glucosinolate profiles, and HPLC batch effects. Block (nested in Site), Genotype*Block, and Batch were random-intercept terms; the rest were modeled as fixed effects. One genotype was omitted from the analysis because no individuals of that genotype survived at one field site. Spearman’s rank correlation tests of least-squares means resulting from this model (for the Genotype fixed effect) revealed that these traits were partially genetically correlated, although correlations between BC-ratio and both other traits were driven by three outlier genotypes that entirely lacked branched-chain glucosinolate functionality (Supplementary Table 3). Least-squares mean trait values for Site, Genotype, and Genotype × Site fixed effects were used to quantify plasticity among habitats.

To better understand how variation in individual glucosinolate compounds underlies variation in these three emergent properties of glucosinolate profiles, we repeated the above analysis for square-root-transformed concentrations of 2OH1ME, 1ME, 1MP, and 6MSOH. Similar to the emergent properties that are the focus of this study, concentrations of individual compounds were strongly genetically controlled but also highly plastic within and among habitats. (Supplementary Tables 4, 9b; Supplementary Figure 5). However, genetic correlations between individual compounds were even stronger than those between emergent glucosinolate profile properties (Supplementary Table 3b). For this reason, and because previous work has confirmed the ecological and evolutionary importance of emergent glucosinolate profile properties (Prasad et al. 2012; Müller et al. 2010), we focus the remainder of the analyses on BC-ratio, Total [AGS], and BC-diversity rather than individual compounds.

Characterizing selection on glucosinolate profiles

We evaluated natural selection on three separate glucosinolate-related traits (Figure 2c) in each of four common gardens (Figure 1) by conducting twelve phenotypic selection analyses (3 glucosinolate traits × 4 sites). The within-site relative fecundity of all measured individuals was regressed onto standardized trait values, while controlling for microsite variation in habitat quality using a random intercept Block term:

$$\mathit{\text{Individual relative fecundity}}=\beta \ast \mathit{\text{Trait}}\_\mathit{\text{value}}+\mathit{\text{Block}}+\mathit{\text{error}}$$

A significant β regression coefficient indicated nonzero linear selection. This regression coefficient is the selection differential on one trait at one site (e.g., the slopes of the lines in Supplementary Figure 2b–c). The signs and magnitudes of the regression coefficients from these models thus indicate the direction and strength of linear selection on each trait at each site.

Next, to test whether selection differentials for each trait varied among sites (i.e., whether selection was spatially variable), we analyzed the relative fecundity and phenotype data from all four sites together, although fecundity and trait values remained relativized and normalized within each site as described above. We did this by fitting a mixed-effects ANCOVA model with an additional Site × Trait interaction term, which described heterogeneity of linear selection on the trait among habitats:

$$\mathit{\text{Individual relative fecundity}}=\mathit{\text{Site}}+\mathit{\text{Block}}(\mathit{\text{Site}})+\mathit{\text{Trait}}+\mathit{\text{Site}}\ast \mathit{\text{Trait}}+\mathit{\text{error}}$$

Because stabilizing or disruptive selection could also affect the fitness consequences of plasticity (i.e., movement of trait values either toward or away from optimum values at each site), we also conducted quadratic selection analysis for each trait at each site:

$$\mathit{\text{Individual relative fecundity}}=\mathit{\text{Block}}+\beta \ast {\mathit{\text{Trait}}}_{\mathit{\text{value}}}+\gamma \ast {(\mathit{\text{Trait\_value}})}^2+\mathit{\text{error}}$$

A significant γ regression coefficient indicated nonzero quadratic selection; negative values were taken as evidence of stabilizing selection, while positive values indicated disruptive selection. However, quadratic regressions were considered statistically significant only if they gave significantly improved fit over a linear selection model, and if we could reject the null hypothesis that the true fitness minimum or maximum fell outside the range of observed trait values using generalized linear hypothesis tests (Supplementary Methods; Mitchell-Olds and Shaw 1987).

Testing for adaptive plasticity among habitats

The question of whether plasticity can aid survival in new environments hinges on whether the plastic response moves trait values in a direction favored by natural selection, thus increasing relative fitness in the new environment. We explored this interaction between plasticity and selection by combining results from the variance-partitioning model and selection analyses described above.

For each site-trait combination, we calculated each genotype’s expected change in fitness due to plasticity (Δω) by evaluating the selection differential at two trait values:

$$\mathrm{\Delta }\omega =f(T_{\mathit{\text{il}}})-f({\overline{T}}_l)$$

where f(x) is the selection differential (i.e., relative fitness as a function of a given trait at a given site; see above), T_il is the least-squares mean trait value of genotype l at site i (depicted by the black and red points in Supplementary Figure 2), and T̄_l is the least-squares mean trait value of genotype l averaged across all sites (depicted by the black and red dashed lines in Supplementary Figure 2). Both T_il and T̄_l were calculated from the REML variance-partitioning model described above. T̄_l represents the trait value of a hypothetical genotype that is identical to genotype l except that it lacks plasticity. Therefore, Δω is an estimate of the fitness change that can be attributed to plasticity among sites of a trait that is under selection. Positive values of Δω constituted evidence of adaptive plasticity; negative values indicated non-adaptive plasticity; when the 95% CI of Δω included zero, plasticity was considered to be “neutral,” neither adaptive nor non-adaptive.

Our estimates of Δω allowed us to estimate how frequently glucosinolate plasticity among sites affected fitness. We also conducted an exact binomial test of the null hypothesis that such “non-neutral” plasticity is equally likely to move glucosinolate trait values in an “adaptive” or a “non-adaptive” direction. Cases of plasticity with no predicted effect on fitness, in which the 95% CI of Δω included zero, were excluded from this analysis. Finally, we conducted Pearson’s chi-squared tests to determine whether the two subspecies differed in the relative frequency of adaptive versus non-adaptive plasticity, or in the relative frequency of neutral versus non-neutral plasticity.

We considered all linear and quadratic selection—for each trait in each site—when performing binomial tests and chi-squared tests. Because selection curvature was usually very weak, however, the linear and quadratic fitness predictions were extremely similar. Therefore, in the interest of caution, we performed these tests separately for the linear and quadratic selection differentials, and found that they led to identical conclusions. For simplicity, we present the results using linear selection in the main text, and provide the quadratic selection results in the Supplementary Information for comparison.

Characterizing within-habitat plasticity using reaction norms

In this study, we focused on phenotypic plasticity induced by spatial environmental variation at a single time-point. Because each plant in this study only experienced a single spatial environment, plasticity of individual plants could not be measured. Instead, spatial plasticity of glucosinolate profiles is a property of a genotype, estimated by comparing the phenotypes of individuals that shared the same genotype but were growing in different experimental blocks. To infer whether natural selection was acting on fine-grained glucosinolate plasticity within B. stricta habitats, we (1) quantified plasticity among blocks for each genotype as a continuous function or reaction norm, and (2) used genotypic selection analysis to test whether reaction norm steepness—a measure of plasticity—predicted evolutionary fitness of each B. stricta genotype.

First, for each of 25 genotypes we fit one reaction norm for each glucosinolate trait in each field site—for a total of 300 reaction norms (3 traits × 4 sites × 25 genotypes). Each reaction norm describes one glucosinolate trait as a continuous linear function of an environmental index (EI, a numerical descriptor of microhabitat conditions within each experimental block). We assumed that most environmental factors causing glucosinolate plasticity are unknown, and so the relevant environmental characteristics are best “measured” using plant phenotype data. Therefore, we assigned the grand mean trait values observed in each block (for all 25 genotypes, pooled) to be the environmental indices (Finlay and Wilkinson 1963). The grand mean BC-ratio in each block was used as the EI for fitting BC-ratio reaction norms; the grand mean Total [AGS] in each block was used as the EI for fitting Total [AGS] reaction norms; and the grand mean BC-diversity was used as the EI for fitting BC-diversity reaction norms. We calculated the reaction norm for each genotype and each trait in each site using linear regression of genotype-specific block mean trait values onto the EIs:

$$\mathit{\text{Genotype trait mean in block}}=\beta \ast \mathit{\text{Block environmental index}}+\mathit{\text{error}}$$

The linear regression coefficient (β) estimated by each model is a reaction norm slope, which describes the magnitude and direction of the plastic response for one genotype. We also calculated the height of each reaction norm by evaluating the linear function at the mean EI value; thus, reaction norm height for a given site describes the genotype’s predicted trait value in an “average” block (Supplementary Figure 3) at that site.

Second, to test whether reaction norm slopes were heterogeneous among genotypes, we analyzed genotype-specific block mean trait values and EIs from all 25 genotypes together by fitting an ANCOVA model with an additional Genotype × EI interaction term that described genetic variation for reaction norm slopes:

$$\mathit{\text{Genotype trait mean in block}}=\mathit{\text{Genotype}}+\mathit{\text{Block EI}}+\mathit{\text{Genotype}}\ast \mathit{\text{Block EI}}+\mathit{\text{error}}$$

If the interaction term was significant, we concluded t hat reaction norm slopes were heterogeneous among genotypes. We performed this analysis separately for each glucosinolate trait in each site.

Testing for selection on reaction norm slope and reaction norm height

The slope and height of the reaction norms are measurable characteristics of plant genotypes, corresponding to plasticity and average trait values, respectively (Supplementary Figure 3). We measured linear selection gradients on these reaction norm parameters to determine whether glucosinolate plasticity in response to fine-grained environmental variation affects fitness. Because we calculated site-specific reaction norms (see above), and to allow for the possibility that plasticity is not equally advantageous in all habitats, we measured selection on reaction norm parameters separately for each site.

To test for linear selection on reaction norm slopes, we conducted genotypic selection analysis separately for each glucosinolate trait at each site. Genotypes’ relative fitness within each site (survival rate × fecundity of survivors for each genotype, divided by the mean fitness for all genotypes at that site) was regressed onto the genotypes’ site-specific reaction norm height and slope, generating two partial regression coefficients (β₁ and β₂) that describe linear selection on average trait values and trait plasticity, respectively (Supplementary Figure 3g–h; Supplementary Methods):

$$\mathit{\text{Genotype relative fitness}}={\beta }_1\ast \mathit{\text{RN height}}+{\beta }_2\ast \mathit{\text{RN slope}}+{\beta }_3\ast \mathit{\text{Diameter}}+\mathit{\text{error}}$$

The mean rosette diameter for each genotype at time of transplanting, which reflects differences in performance in the greenhouse and not in the field, was included as a nuisance variable.

Results

Genotype and environment interact to control glucosinolate profiles

All three glucosinolate profile features—BC-ratio, Total [AGS], and BC-diversity—varied among genotypes and among field sites (Figure 3a; Table 1). All three traits were also affected by developmental stage and plant size (Table 1; Supplementary Figure 6). Genotypic variation for BC-ratio has been previously reported in B. stricta Schranz et al. 2009; Manzaneda et al. 2010); our data show that Total [AGS] and BC-diversity are also genetically controlled. In this experiment the East subspecies harbored more genetic variation for BC-ratio than the West; however, genetic variability for the other traits was comparable between the subspecies (Figure 3a). Genotypes varied less in glucosinolate concentration than in BC-ratio and BC-diversity, except for SAD12, which produced nearly twice the concentration of glucosinolates as the others. SAD12 is also notable as the only genotype in the experiment that originated in Colorado; the others are from Idaho or Montana. Nevertheless, analysis of Total [AGS] yielded the same results regardless of whether the outlier SAD12 was included.

Figure 3.

Genotype and environment influence glucosinolate profiles both independently and via their interaction. (a) Both genotype and habitat influence glucosinolate profiles. Plotted are least-squares mean trait values for each Genotype (circles) and for each Site (triangles) from a REML mixed model that also controlled for developmental stage, plant size, genotype-by-site interactions, and block and batch effects (Table 1). Thus, circles show the mean trait value for each genotype (averaged across all sites); triangles show the mean trait value at each site (average for all genotypes). Note that the horizontal position of the triangles is meaningless—they were placed in order to not obscure the genotype means. Error bars are 95% confidence intervals. (b) Glucosinolate plasticity among sites is genetically variable. Genotypes are delimited by vertical dashed lines. The points are least-squares mean trait values for each genotype in each common garden, with 95% confidence intervals. BC-ratio and BC-diversity are unitless; units for Total [AGS] are µmol/mg dry tissue.

Table 1.

Variance partitioning of glucosinolate traits using REML mixed models reveal strong genotype-by-site interactions. Total [AGS] was square-root transformed before analysis. All effects are fixed except for Block, Genotype × Block, and HPLC batch, which were random-intercept terms. Significance of random effects was assessed using likelihood ratio tests.

	BC-ratio	BC-ratio	BC-diversity
R2	0.99	0.49	0.89
Genotype	F23,1035=343.68	F23,949=9.24	F23,987=125.34
	P<3e−16	P <3e−16	P <3e−16
Site	F3,109=21.28	F3,99=14.07	F3,149=44.62
	P =1.3e−10	P =1.0e−07	P <3e−16
Genotype × Site	F69,1035=3.93	F69,899=3.02	F69,966=11.82
	P <3e−16	P =7.2e−14	P <3e−16
Dev. Stage	F2,785=4.31	F2,1332=23.78	F2,1286=5.56
	P =0.014	P =2.1e−10	P =0.0079
Height	F1,1109=33.05	F1,1344=46.86	F1,1279=10.08
	P =2.3e−08	P =3.5e−11	P =0.0015
Block	χ21=22.74	χ21=59.23	χ21=0.19
	P =3.7e−06	P =4.2e−14	P=0.66
Genotype × Block	χ21=58.90	χ21=0.03	χ21=7.96
	P =5.0e−14	P=0.86	P =0.0096
HPLC batch	χ21=4.09	χ21=9.66	χ21=8.84
	P =0.043	P =0.006	P =0.0059

All three traits showed significant plasticity among sites. For Total [AGS], the magnitude of the site effect was comparable to variation attributed to genotype—in particular, plants growing at Mahogany Valley produced only 50% the concentration of glucosinolates as those growing at the other sites, on average (Figure 3a). In contrast, for BC-diversity and especially BC-ratio, the magnitude of plasticity among sites was minor compared to the variation due to genotype. Additionally, BC-ratio and Total [AGS] varied significantly among blocks within sites (6.3% and 11.2% of the total variance, respectively), indicating that meter-scale environmental heterogeneity affected expression of these traits (Supplementary Figure 7a; Table 1).

Finally, significant genotype-by-site interactions confirm that genotype and environment interacted to shape glucosinolate profiles (Table 1). In general, Eastern genotypes were more sensitive to environment than Western genotypes, especially for BC-ratio and BC-diversity. However, patterns of plasticity among sites varied even within subspecies (Figure 3b). In addition to these genotype × site interactions, genotype × block interactions accounted for 65.3% and 35.9% of the variance in BC-ratio and BC-diversity, respectively. This indicates strong genetic variation for plasticity of glucosinolate composition in response to environmental heterogeneity on the meter scale within habitats. Consistent with the observed subspecies difference in plasticity among habitats (Figure 3b), Eastern genotypes displayed greater plasticity among blocks within habitats, as well (Supplementary Figure 7b).

Evidence for frequent adaptive plasticity among sites

Selection on glucosinolate profiles in all habitats

Phenotypic selection analysis revealed ten cases of statistically significant selection: seven cases of linear selection on glucosinolate traits, and three cases of disruptive (quadratic) selection (Figure 4; Supplementary Figure 8; Supplementary Table 5). High Total [AGS] was associated with decreased fecundity in all four common gardens (Figure 4g–j; Supplementary Table 5). Pooling the data from all sites and testing for a Site*Total [AGS] interaction failed to reject the null hypothesis that these selection differentials were equivalent (Supplementary Table 6), suggesting that high glucosinolate concentrations were equally costly or disadvantageous in all four habitats. In contrast, linear selection on BC-diversity varied among sites (Supplementary Table 6). Specifically, high BC-diversity was associated with higher fecundity at Jackass Meadow but lower fecundity at Alder Creek and Silver Creek; however, we detected no significant selection on BC-diversity at Mahogany Valley (Figure 4k–n; Supplementary Table 5).

Figure 4.

Evidence for adaptive glucosinolate plasticity among field sites. For each trait, we asked whether patterns of plasticity among sites resulted in increased or decreased relative fitness (Δω) due to linear selection at each of four sites (the results for non-linear selection are shown in Supplementary Figure 8). To do this for each genotype, we evaluated the selection differential (plotted as red lines) at two values: the genotype’s LS mean trait value across all sites (which represents a hypothetical “no plasticity” scenario) and the genotype’s true LS mean trait value at that site (a). The averages of these trait values across all genotypes are depicted as dotted and solid vertical lines, respectively, in the left side of panels (c–n). In the right side of panels (c–n), Δω due to plasticity is plotted for each genotype. Positive and negative estimates of Δω are considered evidence for adaptive and non-adaptive plasticity, respectively, if their 95% CIs do not contain 0 (colored in pink and blue respectively). Panel (b) summarizes counts of neutral, adaptive, and non-adaptive plastic responses for each genotype. Overall, plasticity was in an adaptive direction more often than expected by chance (71.3% of all cases of non-neutral plasticity; exact binomial test, P=5.6e⁻⁵).

For thoroughness, we also calculated linear selection gradients to assess direct selection on each trait while controlling for indirect selection on the other glucosinolate traits (Supplementary Methods). The selection gradients generally agreed with the selection differentials, and also indicated that BC-ratio may be under negative selection at Mahogany Valley (Supplementary Table 7).

Finally, we found evidence that disruptive selection is acting on BC-ratio at Silver Creek and on Total [AGS] at Alder Creek and Silver Creek (Supplementary Figure 8; Supplementary Table 8).

Plasticity among habitats was more likely to increase fitness than decrease fitness

To determine whether each trait exhibited adaptive or maladaptive plasticity (or neither) at a given site, we used the selection differentials calculated above to estimate the fitness effects of plastic changes in trait values among sites. For simplicity, here we present results based on directional selection only (Figure 4); however, the use of quadratic selection differentials (either separately or in combination with directional selection) yielded identical conclusions (Supplementary Figure 8).

Because the strength and direction of plasticity varied among genotypes (Figure 3b), we assessed the frequency of adaptive versus non-adaptive plasticity separately for each genotype. We used each genotype’s experiment-wide LS mean trait values as baselines, representing a hypothetical genotype that was identical except that it lacked plasticity among sites (Supplementary Figure 2). Relative to this baseline, plasticity interacted with directional selection to substantially alter relative fitness in 27.1% of all cases.

Of the cases in which a genotype’s plasticity was predicted to alter its fitness, trait values shifted in an adaptive direction 73.1% of the time (95% CI = 61.8% to 82.5%)—considerably more than would be expected by chance (exact binomial test, P=5.6e⁻⁵; Figure 4b). This especially benefitted the East subspecies; although the subspecies did not differ in the relative frequency of adaptive versus non-adaptive plasticity (χ²= 0.04, P=1), Eastern genotypes generally had stronger plastic responses than Western genotypes (Figure 3b), which translated into greater fitness effects. Plasticity among sites affected fitness nearly 40% of the time for Eastern genotypes, but only 15% of the time for Western genotypes (χ²=6.2, P=0.015; Figure 4b).

Glucosinolate reaction norms were genetically variable but not under detectable selection

After correction for multiple testing, we detected genetic variation for reaction norm slopes for BC-ratio and Total [AGS], each in two of four field sites (Figure 5; Supplementary Table 9). However, genotypic selection analysis failed to detect significant selection differentials on reaction norm slopes at any site (Supplementary Table 10), suggesting that glucosinolate plasticity in response to continuous environmental gradients was neither beneficial nor costly within any of these habitats.

Figure 5.

Genotype-specific glucosinolate reaction norms. The mean trait values of each genotype in each block were regressed onto the “environmental index”, or the grand mean trait values for all genotypes in each block. Each resulting regression line is a reaction norm for one genotype, plotted here in red or black for Eastern or Western genotypes, respectively. Genotypes with steeper reaction norm slopes exhibit more plasticity in response to continuous environmental gradients. The “average reaction norm” (equivalent to the line y=1*x, where the expressed trait value equals the block mean trait value) is shown as a blue dashed line. The purple vertical dotted line denotes the value at which reaction norm height was evaluated, which corresponds to the mean environmental index for each trait (i.e., an average micro-environment within that habitat or field site). Stars indicate statistically significant genetic variation for reaction norm slopes (Supplementary Table 9).

Discussion

In Boechera stricta, plant genotype and environment interact to affect the concentration and composition of glucosinolate profiles. All three emergent traits of glucosinolate profiles—Total [AGS], BC-ratio, and BC-diversity—were under both genetic and environmental control, with strong genotype-by-environment interactions. Genotype was the strongest determinant of BC-ratio and BC-diversity, whereas genotype and environment had similar effect sizes for Total [AGS] (Figure 3a). We observed abundant genetic diversity for plasticity both among and within field sites (Figure 3b; Figure 5; Table 1). Particularly striking was the 65% of BC-ratio variation that was explained by genotype-by-block interactions (Supplementary Figure 7b; Table 1).

In general, Eastern genotypes were more plastic than Western genotypes, and there was additional genetic variation in reaction norm shape and inter-site plasticity within the East subspecies (Figure 3b; Figure 5). However, this divergence does not reflect adaptation to within-habitat heterogeneity. With a few exceptions, all of the environmental variables that we measured had similar variances in the two Eastern habitats and the two Western habitats (Supplementary Figure 9), indicating that neither subspecies generally occupies more complex habitats than the other. The lack of variation for glucosinolate plasticity among Western genotypes is consistent with observed patterns of reduced molecular diversity relative to Eastern genotypes (Baosheng Wang, personal communication), and might limit further evolution of reaction norms within the West but not the East subspecies.

Glucosinolate plasticity may aid colonization of new habitats

Because the distances separating our field sites are much greater than the dispersal distance of Boechera (Bloom et al. 2002), it is unlikely that the inter-site plasticity observed in this study evolved as an adaptation to heterogeneity among sites (Via and Lande 1985; Gomulkiewicz and Kirkpatrick 1992). Instead, the differences between pairs of sites are reasonable simulations of transitions to novel environments, or when streams carry seeds to lower-elevation sites. Thus, we could assess whether this coarse-grained glucosinolate plasticity might impact the likelihood that a B. stricta population could survive a major environmental change or colonize a new habitat (Ghalambor et al. 2007).

Whether plasticity of a trait improves the relative fitness of a population upon encountering a new environment depends on the directions and magnitudes of both natural selection and the plastic response. Characterizing plastic responses in this way requires the definition of a “baseline” against which the site-specific trait values can be compared; here we used each genotype’s experiment-wide mean trait value as the baseline. We note that if the four field sites used in this experiment were somehow unrepresentative of the wider range of habitats occupied by wild B. stricta populations, then this baseline might not reflect the “true” average glucosinolate profile, and thus our estimations of plasticity might be incorrect. We have no reason to suspect this is the case, because we chose these sites to reflect the diversity of habitat types in the region (Supplementary Figure 4; Supplementary Table 2): Alder Creek and Silver Creek are riparian meadows—representative habitats of the Western subspecies—while the other two sites are typical Eastern subspecies habitats—dry, high-elevation meadows (Lee and Mitchell-Olds 2011, 2013). Nevertheless, repetition of this experiment in a wider range of B. stricta habitats would be necessary to rule out this possibility.

In this experiment, we detected significant natural selection acting on at least one glucosinolate trait at all four field sites (Figure 4; Supplementary Table 5). Selection consistently favored lower Total [AGS], suggesting a cost of producing these defensive compounds (Mauricio 1998). Alternatively, glucosinolate production could be consistent among plants and this pattern could reflect the dilution of glucosinolate content over a greater amount of tissue in larger, more vigorous individuals. In contrast, the direction of selection on BC-diversity varied among sites. Strikingly, plasticity of BC-diversity in Eastern genotypes mirrored these varying selection pressures. Trait values increased in sites where BC-diversity was under positive selection, and decreased where it was under negative selection, resulting in significant fitness boosts for the most plastic genotypes (Figure 4k–n). Across all traits, sites, and genotypes, plasticity was predicted to impact fitness in 27.1% of cases. Such “non-neutral” plasticity was much more likely to increase fitness than to decrease it (Figure 4b; exact binomial test, P=5.6e⁻⁵).

The whole of the data suggest that glucosinolate plasticity often changes defensive chemistry to better match the local selection pressures, and therefore might aid B. stricta populations in colonizing new habitats (Ghalambor et al. 2007). Whether such adaptive plasticity promotes glucosinolate evolution in the long term, however, will be a more difficult question to answer (Ghalambor et al. 2007; Ghalambor et al. 2015; Huang and Agrawal 2016). Identifying the environmental cues perceived by the plants that induce changes in defensive chemistry, as well as the ecological causes of local selection pressures, should be a priority for future research on adaptive glucosinolate plasticity.

Finally, we highlight the implications of genetic variation for plasticity. Although the relative frequency of adaptive versus nonadaptive plasticity was similar between subspecies, genotypes differed in the strength of their plastic responses. Based on our results, we expect glucosinolate plasticity to aid the colonization of new habitats 19.8% of the time overall. However, considering only Western genotypes, this estimate drops to 10.4%, whereas Eastern genotypes are expected to benefit from glucosinolate plasticity 29.2% of the time. In contrast, non-adaptive glucosinolate plasticity is expected to hinder colonization of new habitats 7.3% of the time species-wide, with similar rates for both subspecies (Figure 4b). These results illustrate that the contribution of glucosinolate plasticity to persistence after environmental change is not uniform across the species.

No evidence for selection on plasticity within habitats

In this experiment, we detected little evidence for selection on glucosinolate reaction norms, or plasticity in response to continuous environmental gradients. Because this study included only 25 genotypes, evidence for selection on reaction norms may become clearer as more genotypes are analyzed. Consistent with this, selection gradients on reaction norm height (i.e., mean trait values across all blocks) lacked statistical support but agreed qualitatively with the patterns detected using the phenotypic selection analysis (compare β values in Supplementary Table 5 with β_H values in Supplementary Table 10). Another possible reason for this negative result is that selection pressures on glucosinolate profiles may not vary on such a fine spatial scale, reducing the opportunity for adaptive plasticity within habitats (Via and Lande 1985; Gomulkiewicz and Kirkpatrick 1992). The lack of observed selection against plasticity suggests that glucosinolate plasticity does not carry a significant cost (Auld et al. 2010).

In addition, inter-annual variation is one potential cause of plasticity and variable selection that we did not address in this study. Other experiments have shown that herbivory pressure on B. stricta varies considerably over a span of a few years within a single site (T. Mitchell-Olds, unpublished); consecutive generations of a B. stricta lineage might therefore experience very different predatory environments, potentially causing temporally heterogeneous selection on glucosinolate profiles. The consistent negative selection on glucosinolate concentration in this study (Figure 4g–j; Supplementary Table 5) suggests that herbivory intensity may have been lower than usual, reducing the usefulness of these expensive phytochemicals (Mauricio and Rausher 1997; Mauricio 1998). Temporal fluctuations in resource availability (e.g., due to limited or abundant rainfall) or pathogen pressure could also affect the relative costs and benefits of glucosinolate profile properties. Because B. stricta is a perennial, even a single individual may need to adjust to several environments over its lifetime; in theory, plasticity could evolve as an adaptation to such fine-scale variability (Gomulkiewicz and Kirkpatrick 1992; Reed et al. 2010; Baythavong 2011). The relative importance of temporal environmental variation compared to spatial variation could be assessed by extending a similar experimental design over multiple growing seasons. Follow-up studies should prioritize temporal variation as a potential driver of adaptive plasticity in glucosinolate profiles and other important B. stricta traits.

Importantly, evidence of natural selection on reaction norms would not be sufficient to demonstrate that plasticity evolved as an adaptation. Reliable environmental cues for the different selection pressures are also required, preventing the evolution of plasticity as an adaptation to truly stochastic environmental fluctuations (Levins 1963; Donohue et al. 2000; Schmitt et al. 2003; Reed et al. 2010). In other words, if an organism’s offspring reliably encounter multiple contrasting selection pressures, and if some perceptible environmental factor predicts how selection will act, plasticity (rather than habitat specialization) will evolve as a long-term adaptation to a heterogeneous environment (Donohue et al. 2000; Sultan and Spencer 2002). Even when these conditions are not met, however, plasticity can still improve short-term relative fitness—that is, it may sometimes be “adaptive” even if it did not evolve as an “adaptation” (Sultan 1987).

Future directions

The molecular basis of genotype-by-environment interactions is a key research goal for understanding the evolution of phenotypic plasticity and the robustness of genetic improvements in crop species (El-Soda et al. 2014). The ample genetic variation for plasticity of three glucosinolate traits provides an opportunity to explore this phenomenon at the molecular level in B. stricta, which offers resources such as near-isogenic lines varying in glucosinolate synthesis genes, a recently developed genome-wide association panel, a fully sequenced genome, and genetic similarity to the model plant Arabidopsis thaliana (Rushworth et al. 2011; Prasad et al. 2012; Lee et al. 2017). The Arabidopsis glucosinolate biosynthetic pathway is well characterized (Halkier and Gershenzon 2006), although genes affecting plasticity are not necessarily part of this core pathway. Indeed, in Arabidopsis, patterns of flux through this pathway were robust to several biotic and abiotic environmental stimuli (Olson-Manning et al. 2015), suggesting that variation in environment-sensing genes or other upstream genes may be more important for glucosinolate plasticity, per se.

The exact ecological causes of glucosinolate plasticity and selection in this experiment are still unclear and cannot be determined from this dataset; additional experiments with environmental manipulations will be necessary to identify the causal stimuli (Anderson et al. 2014). Many environmental features distinguish the field sites used in this study (Supplementary Figure 4). For this reason, we used plant phenotype to define the environmental index of each block, rather than guessing which environmental parameters were most relevant (Finlay and Wilkinson 1963). The disadvantage of this method is that the phenotype is a “black box” obscuring the actual ecological features driving trait plasticity. The extent to which the stimuli causing glucosinolate plasticity overlap with those exerting selection on glucosinolate profiles may determine the consistency with which plasticity can “track” selection and boost local fitness. Previous work indicated that insect herbivores are one important agent of selection on BC-ratio; however, the protective effect of high BC-ratio is not observed in all habitats, indicating that insect communities vary in the selective pressure they exert on this trait (Prasad et al. 2012). Different herbivore species also varied in their ability to induce glucosinolate plasticity in the greenhouse (Manzaneda et al. 2010), but it is unknown whether entire communities of B. stricta herbivores also vary in this way, much less whether that variation corresponds to the differential selection on glucosinolate profiles by the same herbivores. Other environmental properties that could potentially drive glucosinolate plasticity, selection, or both include pathogens (Clay et al. 2009) and a variety of abiotic stressors including drought, soil salinity, and nutrient limitation (Khokon et al. 2011; Martínez-Ballesta et al. 2013). Identifying which ecological factors cause plasticity of glucosinolate profiles—and selection on them—will be an important step towards understanding the apparent link between natural selection and phenotypic plasticity in diverse habitats across a natural landscape.

Finally, it is an open question whether the patterns of adaptive plasticity observed in this experiment are typical of other functional traits. Although glucosinolates are known to be ecologically and evolutionary important, they represent only a fraction of the traits contributing to evolutionary fitness in B. stricta and other Brassicaceae. Additional experiments are needed to clarify whether the evolutionary implications of glucosinolate plasticity are generalizable to other adaptively important traits in B. stricta and other study species. Until more examples from wild populations are available, it will remain challenging to identify patterns and draw general conclusions about the role of plasticity in adaptive evolution.