A change in climate causes rapid evolution of multiple life-history traits and their interactions in an annual plant

Authors


Steven J. Franks, Department of Biological Sciences, Larkin Hall, Fordham University, 441 E. Fordham Road, Bronx, NY 10458, USA.
Tel.: +1 718-817-3609; fax: +1 718-817-3645;
e-mail: franks@fordham.edu

Abstract

Climate change is likely to spur rapid evolution, potentially altering integrated suites of life-history traits. We examined evolutionary change in multiple life-history traits of the annual plant Brassica rapa collected before and after a recent 5-year drought in southern California. We used a direct approach to examining evolutionary change by comparing ancestors and descendants. Collections were made from two populations varying in average soil moisture levels, and lines propagated from the collected seeds were grown in a greenhouse and experimentally subjected to conditions simulating either drought (short growing season) or high precipitation (long growing season) years. Comparing ancestors and descendants, we found that the drought caused many changes in life-history traits, including a shift to earlier flowering, longer duration of flowering, reduced peak flowering and greater skew of the flowering schedule. Descendants had thinner stems and fewer leaf nodes at the time of flowering than ancestors, indicating that the drought selected for plants that flowered at a smaller size and earlier ontogenetic stage rather than selecting for plants to develop more rapidly. Thus, there was not evidence for absolute developmental constraints to flowering time evolution. Common principal component analyses showed substantial differences in the matrix of trait covariances both between short and long growing season treatments and between populations. Although the covariances matrices were generally similar between ancestors and descendants, there was evidence for complex evolutionary changes in the relationships among the traits, and these changes depended on the population and treatment. These results show that a full appreciation of the impacts of global change on phenotypic evolution will entail an understanding of how changes in climatic conditions affect trait values and the structure of relationships among traits.

Introduction

Traits related to the timing of flowering are among the most important life-history traits in plants (Primack, 1985; Rathcke & Lacey, 1985). These phenological traits, which include time to first flowering, duration of flowering, size at flowering and others, influence resource allocation, reproductive availability and individual fitness (Widén, 1991; Sandring et al., 2007). Phenological traits are known to vary among individuals within populations and among populations across the landscape (Primack, 1985), with climatic factors strongly influencing traits related to the timing of flowering (Fox, 1990; Lambrecht et al., 2007).

Given the influence of climatic conditions on reproductive timing, it is likely that global climate change will alter patterns of plant phenology and other life-history traits, potentially leading to sweeping, rapid evolutionary changes across many taxa worldwide. Several reviews and meta-analyses have found clear trends of advancing phenology with climate warming in hundreds of species (Fitter & Fitter, 2002; Peñuelas et al., 2002; Parmesan & Yohe, 2003; Menzel et al., 2006). Although most previous studies that have shown correlations between climate changes and shifts in phenology were not designed to address whether the phenological differences are due to evolution or phenotypic plasticity, there is some evidence for evolution in response to climate change. Genetically based changes in response to climatic warming have been found for parturition date in northern red squirrels (Realéet al., 2003), for photoperiod response in pitcher plant mosquitoes (Bradshaw & Holzapfel, 2001, 2008) and for chromosomal inversions related to heat tolerance in drosophila (Umina et al., 2005).

Although these cases show rapid evolution, there are several types of constraints that could slow the ability of species to adapt to changes in climate. One type of constraint is due to correlations among traits, with negative correlations potentially slowing the responses to selection (Lynch & Walsh, 1998). Because evolution operates not on individual traits in isolation but also on traits through their covariance with other characters (Lande & Arnold, 1983), understanding the evolution of correlated traits is of fundamental concern and prerequisite for predicting evolutionary change (Lynch & Walsh, 1998). Furthermore, agents of selection can alter not only traits individually and through their covariances, but also the matrix of trait covariances itself. Implications of changes in the matrix of trait covariances have been examined theoretically (Turelli, 1988; Arnold, 1992), but empirical demonstrations are rare (Roff & Mosseau, 2005).

Evolutionary change can also be slowed by developmental constraints. If an event such as first flowering can only occur at a given developmental stage, then it is more difficult to evolve changes in the timing of flowering than if development is not fixed. According to models of heterochrony, evolutionary changes in the time to first flowering can occur through one of three basic pathways: plants can flower earlier due to earlier onset of growth, more rapid growth and development, or to flowering at an earlier ontogenetic stage (Diggle, 1999). If flowering can occur at an earlier developmental stage, vegetative and reproductive growth are uncoupled, and developmental morphology is not as much of a constraint to phenological evolution. There is currently mixed evidence for such a disassociation between vegetative and reproductive growth (Diggle, 1999), and it is unclear to what extent developmental constraints may hinder the ability of plants to evolve in response to changes in climate.

Previously, Franks et al. (2007) found rapid evolution of one phenological trait – time to first flowering – in the annual plant Brassica rapa following a natural drought in California. They found this evolutionary change by using a resurrection approach of comparing offspring of individuals collected before and after the drought. Here we use data from the same experiments to examine the effects of drought on evolutionary changes in multiple life-history traits. Specifically, we asked to what extent a natural 5-year drought in southern California altered several phenological and morphological traits that are associated with the timing of flowering. We also examined evolutionary changes in correlations among these traits and investigated whether there was evidence for developmental constraints affecting the rates of evolutionary change. We looked at these evolutionary changes in populations from a dry site and a wet site, which previous work has shown have diverged in flowering time traits (Franke et al., 2006; Franks et al., 2007). We also subjected plants from both populations and both collection generations to either a simulated drought treatment or a long watering season treatment to examine experimentally how the life-history traits and their relationships would be affected by different environmental conditions. We could thus compare the multivariate evolutionary changes that occurred over the course of the 5-year drought with those that have resulted in the divergence of flowering time traits in the two populations growing in sites with different conditions and also to what evolutionary changes could potentially be expected based on experimentally altering conditions.

Materials and methods

Study system

Brassica rapa (syn. campestris) L. (Brassicaceae) is a self-incompatible, weedy winter annual plant naturalized in California. For the natural strains we studied from southern California populations, germination can occur without stratification and usually begins within a few days of seeds imbibing water. The plants produce a series of basal leaves before bolting. Several more widely spaced cauline leaves are then produced along the stem. Growth is indeterminate, and multiple flowering racemes are produced from the bases of stem nodes. The plants continue to produce flowers until resources become limiting or, if resources continue to be supplied, until the onset of senescence. Under nonlimiting resource conditions, plants can flower for as long as approximately 80 days (Franks et al., 2007). Individual flowers last approximately 3 days.

Time to first flowering has been shown to vary widely among individuals and to be influenced by a number of factors including plant genotype, plant size and developmental stage, photoperiod and water and nutrient status. A substantial amount of research has been carried out on elucidating the environmental cues and genetic controls of flowering time in B. rapa (Kim et al., 2007) and the related plant Arabidopsis thaliana (Simpson & Dean, 2002). In southern California, B. rapa plants tend to begin flowering in winter at the start of the rainy season and generally continue to flower until the end of the rainy season in spring. The length of the rainy season, and thus the growing season for B. rapa, tends to vary widely among years (Franke et al., 2006). In southern California, the 5 years preceding 1997 were all above-average precipitation, long growing season years, whereas the 5 years before 2004 were all below-average precipitation, short growing season years (Franks et al., 2007).

We collected seeds from two natural populations in southern California. The Back Bay (BB) site is located in upper Newport Harbor and the Arboretum (Arb) site is near the San Joaquin Marsh and adjacent to the UCI Arboretum. The Back Bay site has sandy soils that drain rapidly, which tends to make this a dry area, whereas the Arboretum has soils higher in clay content and a high water table, making this generally a more wet area (Franke et al., 2006). Seeds were collected from both sites in May 1997 and June 2004, with large numbers (> 10 000) of seeds collected per site to avoid drift and founder effects. Moreover, the seeds were collected from plants that were widely spaced throughout the population, and thus separate individuals are not likely to be very closely related in this gravity dispersed species. We refer to the 1997 seeds as predrought, ancestral genotypes and the 2004 seeds as post-drought, descendant genotypes.

Experimental methods

The plants and experimental design for this study were used previously to examine the effects of natural drought on time to first flowering (Franks et al., 2007). In this study, we include additional data on multiple life-history traits collected from the same individuals as in the previous work (Franks et al., 2007).

The field-collected seeds were grown for a generation in a greenhouse to ameliorate maternal environmental and seed storage effects. These F1 progeny were then assigned to 1997 × 1997, 2004 × 2004 and 1997 × 2004 hybrid crossing groups with each population. The F2 progeny from these crosses were then used in the following experiment.

Plants were grown from seed in the greenhouse with six plants to each 25-cm pot. Each pot contained one of each combination of population (Arb and BB) and cross (1997, 2004 and hybrid). Pots were assigned one of three watering treatments in a randomized block design. The watering treatments were short season (watered daily to saturation for 33 days) and medium season (watered daily to saturation for 51 days) and long season (watered 88 days). The sample size of the experiment was = 1800 plants (three cross types × two populations × three watering treatments × 100 replicates). For all analyses, the hybrids were intermediate between the 1997 and 2004 generations, and the medium season treatment was intermediate between the short and long season treatments. For simplicity, we thus present results only from the 1997 and 2004 generations and for the short and long watering seasons, giving a total sample size of 800 individuals.

We checked all plants daily and recorded germination and date of first flowering. Every 3 days, we hand pollinated all plants and also counted all open flowers on all plants to obtain the complete flowering schedules.

Statistical analyses

We calculated values for the life-history traits as follows. Days to first flowering or flowering time (FT) was the difference between germination date and date of the first full opening of the first flower. Duration of flowering (DS) was the difference between the date of first flowering and the last date that a flower was recorded. Skew (SK) of the flowering schedule was calculated by treating the number of open flowers each day as frequencies and calculating the skew of this flowering schedule distribution (Primack, 1985). Total number of flowers (TT) was calculated as the sum of the flowers occurring on each census date. A term describing the shape of peak of the flowering schedule (MX) was calculated as the ratio of the maximum number of flowers on 1 day to the total number of flowers. Number of stem nodes (ND) was the total number of nodes determined by leaves (cotyledons and true leaves) and leaf scars at the time of first flowering. Diameter of the stem (DM) was the diameter at the level of the cotyledons. Height (HT) was the distance from the base of the plant to the uppermost node. Note that some of the variables we used are not strictly independent, and in some cases residuals may be correlated. For example, the flowering schedule shape parameter contains maximum number of flowers in the numerator and total number of flowers in the denominator, whereas total number of flowers is an additional variable. Although there did not appear to be a problem with multicollinearity, and the data, based on examining residual plots, fit the assumption of multivariate normality, it is important to consider this issue when interpreting the results.

To determine the effects of collection generation, population and watering treatment on life-history traits, we conducted a multivariate analysis of variance (manova) and univariate analyses of variance (anova) using sas (version 9.1; SAS Institute, Cary, NC, USA). Dependent variables were the life-history traits and independent variables included generation, population, watering treatment and all interactions, as well as block. The effect of watering treatment was tested over the treatment by block interaction because the watering treatment was applied at the level of the pot rather than the individual plant. Duration of flowering and skew were transformed with a natural logarithm function to improve model fit and multivariate normality.

We examined the relationships among the life-history traits in several ways. We determined the matrices of trait correlations (using Pearson product moment correlations on untransformed values) and covariances for each level of generation, population and watering treatment. We then examined how the correlations and the structure of the correlation matrix differed among the different generations, populations and watering treatment environments using individual matrix element comparisons and common principal components analyses (CPCA).

Individual correlations were compared using the Fisher’s Z transformation (Snedecor & Cochran, 1980) for each pair of traits at each level of generation, population and watering treatment, for a total of 336 tests. These tests determine if individual correlations differ depending on the generation, population or treatment. We corrected for multiple tests using a false discovery rate adjustment (Benjamini & Hochberg, 1995), and report raw and adjusted P-values.

To compare overall matrix structure, we used CPCA with the program CPC (Phillips, 1998). CPCA allows testing not only if matrices differ but how they differ, including whether they are proportional or share any principal components (Flury, 1988; Phillips & Arnold, 1999). We used both the ‘jump-up’ and the model building approaches (Flury, 1988) to matrix comparisons. In the jump-up approach, each level of the Flury hierarchy is compared with unrelated structure. This method is especially appropriate for hypothesis testing at each level of structure (Baker & Wilkinson, 2003). In the model building approach, the Akaike information criterion (AIC) is used to determine which level of shared structure in the Flury hierarchy best fits the data.

Common principal components analysis was designed based on the covariance matrix, which gives the natural scale of variation for each trait but is also therefore biased toward weighting traits with the largest variances, even if the size of the variances changes simply with measurement scale (Phillips & Arnold, 1999). Using the correlation matrix (which is the equivalent of first standardizing the traits to zero mean and unit variance) avoids this scale bias, but in this case the method does not test if the matrices are proportional and also P-values for other tests in the hierarchy may be suspect (Phillips & Arnold, 1999). We conducted CPCA on both the covariance and correlation matrices. For each pair of two factors from the group of three (generation, population and watering treatment), comparisons of the matrices were made for the third factor, for a total of 12 matrix comparisons at each of eight levels of structure. For example, for each combination of population and watering treatment, matrices were compared across generations.

We used several multivariate analyses to further examine relationships among the life-history traits and how they varied across generations, populations and watering treatments. We conducted both principal components analyses (PCA) and canonical discriminant analyses (CDA) for each combination of generation, population and watering treatment using sas. PCA creates a set of orthogonal vectors (principal components) that explain decreasing amounts of variation in a multivariate data set, such that the first principal component explains the largest amount of variation, the second principal component the second largest and so on. The loadings show how each principal component is related to each original variable. CDA also creates new orthogonal vectors from a multivariate data set. With CDA, the orthogonal vectors are canonical discriminant axes that are ordered in terms of the amount they separate a predetermined set of groups, which in our case were the population, generation and watering treatment groups. Loadings on the canonical discriminant axes also show how they relate to the original variables.

Results

Trait evolution

Evolutionary changes in life-history trait values were assessed by comparing ancestors and descendants. There were significant differences between the generations in all traits except for height (Table 1). For some traits, the effect of generation differed between the populations and/or watering treatments, as shown by the interaction terms (Table 1). The effect of generation differed between the populations for days to first flowering, duration of flowering, ratio of maximum to total flowers and number of stem nodes (Table 1). The effect of generation differed by watering treatments for ratio of maximum to total flowers, and there was a significant three-way interaction between generation, population and watering treatment for the total number of flowers (Table 1). In the multivariate analysis, there were significant effects of generation, population, watering treatment and all two-way interactions (Table 1).

Table 1.   Univariate analysis of variance (anova) and multivariate analyses of variance (manova) table for the effects of source population (Pop: ARB vs. BB), generation (Gen: 1997 vs. 2004) and watering treatment (W: short vs. long season length) on life-history traits.
SourceFTDSSKMXTTNDDMHT
Univariate
Pop262.2***11.0***13.2***15.3***19.4***32.8***51.6***26.5***
Gen64.4***11.1***3.9*8.1**5.4*51.6***26.4***0.3
Pop × Gen27.3***6.7**1.04.9*3.016.4***1.31.3
W21.2***489.4***2.1197.3***540.8***11.4**45.5***22.4***
Pop × W19.8***14.0***0.119.1***50.1***5.6*29.0***9.7**
Gen × W0.30.40.18.8**3.50.51.80.7
Pop × Gen × W0.11.00.61.34.0*1.50.82.4
 NDFDDFWilks’λFP
  1. Traits include days to first flowering (FT), duration of flowering (DS), skew (SK), maximum number of flowers open on 1 day (MX), total number of flowers (TT), number of leaf nodes (ND), stem diameter (DM) and plant height (HT) at first flowering. For the univariate analyses, degrees of freedom are 1, 1589 for Pop; 2, 1589 for Cross; 2, 1589 for Pop × Cross; 2, 48 for W; 2, 1589 for Pop × W; 4, 1589 for Cross × W; and 4, 1589 for Pop × Cross × W. Values given are F-values. NDF, numerator degrees of freedom; DDF, denominator degrees of freedom. Bold indicates < 0.05. *< 0.05; **< 0.01; ***< 0.001.

Multivariate
Pop86810.6938.9< 0.0001
Gen86810.9010.0< 0.0001
Pop × Gen86810.945.2< 0.0001
W86810.36151.9< 0.0001
Pop × W86810.8910.5< 0.0001
Gen × W86810.972.20.0255
Pop × Gen × W86810.981.80.0849

Descendants flowered earlier than ancestors in both the short season (Fig. 1a) and long season (Fig. 1i) watering treatments and for both populations, although the difference between ancestors and descendants was greater in the Arboretum (wet site) than in the Back Bay (dry site). Descendants also flowered for longer than ancestors (Fig. 1b,j). Ancestors had overall negative skew of the flowering schedule, indicating that individuals tended to produce more flowers late in the flowering period. For the descendants, skew was not significantly different from zero (Fig. 1c,k). Descendants had a significantly lower ratio of total to maximum flowers in the short season treatment (Fig. 1d) but not in the long season treatment (Fig. 1l). The significant three-way interaction for total number of flowers indicates that this variable differed depending on the combination of generation, population and watering treatment (Table 1). Plants produced more flowers in the long than in the short season treatments (Fig. 1e,m). In the short season treatment, plants from the dry site (Back Bay) produced relatively more flowers than plants from the wet site (Arboretum) (Fig. 1e), whereas, in the long season treatment, plants from the wet site produced more flowers than plants from the dry site (Fig. 1m). There was a decrease in the total number of flowers produced by the descendants compared with the ancestors only in the wet site and under long season conditions (Fig. 1m).

Figure 1.

 Evolutionary changes in life-history traits. Shown are least-squared means (± SE) for days to first flowering (a,i), duration of flowering (b,j), skew (c,k), maximum number of flowers (d,l), total number of flowers (e,m), number leaf nodes (f,n), stem diameter (g,o) and height (h,p) for plants in the short season treatment (a–h) and long season treatment (i–p) from the wet site: Arboretum (Arb) and dry site: Back Bay (BB) populations. Plants from the 1997 ancestral, predrought generation are shown with open bars, and plants from the 2004 descendant, post-drought generation are shown with hatched bars. Note the differences in the axis scales for the short and long season treatments.

Descendants had fewer nodes at time of first flowering than ancestors in both watering treatments, and the magnitude of the difference was greater for the Arboretum population than for the Back Bay population (Fig. 1f,n). Descendants had thinner stems at time of first flowering for plants from both populations and in both watering treatments (Fig. 1g,o). Plants were taller in the long season treatment (Fig. 1h) than in the short season treatment (Fig. 1p). In the long season treatment, plants from Arboretum were taller than plants from Back Bay (Fig. 1p). Height was not affected by generation (Table 1). Mean-squared errors (MSE) for all univariate analyses are given in Table S1.

Phenotypic correlations

Many of the life-history traits were significantly correlated with each other. In both the short and long season treatments (with generation and population pooled), 21 of the 28 pairwise correlations among the traits were significant at < 0.05 (Table 2). The correlations varied widely in magnitude and direction, with r values ranging from −0.56 to 0.70 (Table 2).

Table 2.   Phenotypic correlations for life-history traits.
 FTDSSKMXTTNDDMHT
  1. Shown are Pearson product moment correlations (r) for plant traits, pooled across collection generation and population. Plants in the short season watering treatment are in the upper right, and plants in the long season watering treatment are in the lower left. Traits include days to first flowering (FT), duration of flowering (DS), skew (SK), ratio of maximum number of flowers open on 1 day to total number of flowers (MX), total number of flowers (TT), number of leaf nodes (ND), stem diameter (DM) and plant height (HT) at first flowering. Sample sizes ranged from = 349 to 397. Bold indicates < 0.05. *< 0.05; **< 0.01; ***< 0.001.

FT 0.43***0.27***0.56***0.34***0.49***0.26***0.14**
DS0.26*** 0.22***0.70***0.69***−0.090.060.19***
SK0.23***0.21*** 0.12*−0.04−0.090.11*0.10
MX0.020.53***−0.04 0.56***0.22***0.010.19***
TT0.44***0.26***0.25***0.22*** 0.14**0.38***0.28***
ND0.64***0.15**0.14**−0.040.53*** 0.43***0.05
DM0.67***0.12*0.16**0.010.65***0.69*** 0.18***
HT0.22***−0.020.05−0.020.23***0.30***0.25*** 

Time to first flowering (FT) was correlated with several of the other life-history traits, but much of the variation in the flowering schedule was not explained by FT. For example, FT explained only 19.4% of the variation in total number of flowers (TT) in the long season treatment and 11.6% of the variation in TT in the short season treatment (Table 2).

Matrix comparisons

We analysed both how individual correlations as well as how the covariance matrix differed among the generations, populations and watering treatments. Many of the individual correlations differed between the watering treatments and between the populations, with relatively fewer individual correlations differing between ancestors and descendants (Fig. 2, Table S2). In some cases, the direction of the correlation changed depending on the watering treatment. For example, the correlation between time to first flowering and total number of flowers was positive in the long season treatment (= 397, r = 0.44, < 0.0001; Fig. 2) and negative in the short season treatment (= 359, r = −0.34, < 0.0001; Fig. 2). This indicates that plants in the long season treatment that flowered later produced more total flowers than plants that flowered early, whereas plants in the short season treatment that flowered early produced more total flowers than plants that flowered late.

Figure 2.

 Matrix of trait correlations. Shown are correlations among the life-history traits of days to first flowering (FT), duration of flowering (DS), skew (SK), ratio of maximum number of flowers open on 1 day to total number of flowers (MX), total number of flowers (TT), number of leaf nodes (ND), stem diameter (DM) and plant height (HT) at first flowering for plants from the wet site (Arboretum) and dry site (Back Bay) populations in the short season (S) and long season (L) watering treatments in the 1997 predrought (97) and 2004 post-drought (04) lines. The position of the bubble indicates the watering treatment and generation, and the size of the bubble indicates the magnitude of the correlation (see the legend). Open bubbles indicate positive correlations, and dark, filled bubbles indicate negative correlations. Lines indicate correlations that are significantly different based on a false discovery rate set at 0.05.

To compare overall differences in the patterns of trait correlations, we used CPCA. The watering treatments caused major changes in the relationships among the life-history traits, as shown by comparing individual correlation coefficients (Table S3) and by CPCA (Fig. 3). The short and long season treatments were unrelated at all levels of CPC structure for both populations and both generations, according to both jump-up (Fig. 3a) and model building (Fig. 3b) approaches. The dry site and wet site populations were also different from each other in trait relationships. The populations differed at the CPC2 level for both treatments and both generations, according to the jump-up approach (Fig. 3a). Evolutionary changes in the relationships among the traits could be evaluated by comparing the ancestral and descendant generations. The generations were either equal, proportional or differed at one of the CPC levels, depending on the population, watering treatment and type of analysis used. The generations were equal or proportional if the soil moisture conditions in the treatment and source population matched. For example, the model building approach showed that the ancestor and descendant trait matrices were proportional for the dry site (Back Bay) with the dry treatment, as well as for the wet site (Arboretum) with the wet treatment (Fig. 3b). By contrast, the generations differed at the CPC2 level for the dry site with the wet treatment and for the wet site with the dry treatment (Fig. 3b). Thus, when the population and watering treatment matched, the generations did not differ or were proportional in covariance structure. When the population and watering treatment did not match, the generations shared structure at the CPC3 or CPC5 levels (jump-up, Fig. 3a) or at the CPC2 level (model building, Fig. 3b).

Figure 3.

 Comparisons of correlation matrices based on common principal components analyses (CPCA). Plants from the wet site (Arboretum) are shown as squares on the left planes, and plants from the dry site (Back Bay) are shown as circles on the right planes. Plants in the long season treatment are in the front planes, and short season plants are in the back planes. Predrought 1997 plants are in the top planes, and post-drought 2004 plants are in the bottom planes. The size of the lines indicates the degree to which the correlation matrices differ, with the thickest lines showing correlation matrices that are not different from equal to each other. The best fit model for each correlation comparison according to the jump-up (a) and model building (b) approach to CPCA is shown alongside each bar, with dotted bars indicating that the best fit model is that the matrices differ in all principal components.

For the matrices that are proportional, it is possible to determine the coefficient of proportionality, which shows how the covariance matrices are related. If the proportionality coefficient is > 1, then the correlations are tighter or the traits are more closely linked in the second compared with that in the first matrix. If the proportionality coefficient is between −1 and 1, then the covariances are weaker in the second matrix compared with that in the first.

We found that the equation describing the relationship between the covariance matrices for the ancestral compared with the descendant lines in the Arboretum (wet site) population in the long season treatment was ancestral = −0.01 + 0.878 × descendant (r2 = 0.81; SE of coefficient = 0.083; t = 10.6; < 0.001). This indicates that the trait covariance structure was weaker, with the traits less tightly correlated, in the descendants compared with that in the ancestors in the Arboretum site.

In addition to conducting CPCA on covariances, we also performed the analyses with the trait correlation matrices, which corrects for unequal variances but has other limitations (Phillips & Arnold, 1999). Generally, the analyses with correlations matched those with covariances. One exception was that the comparison between the ancestral and descendant correlation matrices for the Arboretum population in the long season treatment was not different from equal, whereas the covariance matrices were not equal but were proportional and shared all principal components.

Multivariate analyses

A principal component analysis revealed that traits were grouped by those expressed before and after the onset of reproduction. In the short season treatment, PC1 was dominated by total number of flowers, ratio of maximum number of flowers to total and duration of flowering (Fig. 4), which are all related to flowering effort after the onset of reproduction. PC1 can thus be thought of in this case as representing the amount and temporal pattern of reproduction. PC2 was dominated by days to first flowering, number of stem nodes and stem diameter (Fig. 4), which relate to the timing and developmental pattern of vegetative growth leading up to the onset of reproduction. Thus, PC2 in this case represents the timing of first flowering. In the long season treatment, there was generally a reversal in the loadings compared with that in the short season treatment (Fig. 4). For example, time to first flowering loaded negatively on PC1 and positively on PC2 in the short season treatment but positively on PC1 and negatively on PC2 in the long season treatment. Thus, season length influences the relative importance of preflowering and post-flowering traits.

Figure 4.

 Principal component analysis trait loadings. Shown are the loadings (eigenvectors) for principal component 1 (PC1) and principal component 2 (PC2) for days to first flowering (FT), duration of flowering (DS), skew (SK), ratio of maximum number of flowers open on 1 day to total number of flowers (MX), total number of flowers (TT), number of leaf nodes (ND), stem diameter (DM) and plant height (HT) at first flowering. The loadings indicate the contribution to each trait to both principal components. Loadings are shown for plants from the wet site: Arboretum (Arb) and dry site: Back Bay (BB) populations from the predrought 1997 (97) and post-drought 2004 (04) collection generations in the short season (S) and long season (L) treatments.

Canonical discriminant analyses verified that the life-history traits we measured adequately distinguished the different populations, treatments and generations (Table 3). Like the principal components, the canonical components separately grouped the preflowering (Can2) and post-flowering (Can1) traits. Can1 showed strong discrimination between the season length treatment groups (Fig. 5). This indicates that the watering treatments had a large effect on traits expressed following flowering, such as the duration of flowering and the total number of flowers. Can2 showed strong discrimination between the populations, with positive means for all plants in the Arb population and negative means for all plants in the BB population (Fig. 5). Can2 also discriminated between the collection years, with 1997 values always greater than 2004 within each population and treatment group (Fig. 5). Thus, the differences between the populations and between the generations appears to be more closely linked to preflowering traits such as the timing and developmental stage of first reproduction.

Table 3.   Canonical discriminant analyses.
TraitShort seasonLong season
FCan1Can2FCan1Can2
  1. Shown are F statistics (NDF = 7, DDF = 736) for each life-history trait as well as the loadings for each trait on canonical component 1 (Can1) and canonical component 2 (Can2) for plants in the short season and long season watering treatments. Traits include days to first flowering (FT), duration of flowering (DS), skew (SK), ratio of maximum number of flowers open on 1 day to total number of flowers (MX), total number of flowers (TT), number of leaf nodes (ND), stem diameter (DM) and plant height (HT) at first flowering. Bold indicates < 0.05. *< 0.05; **< 0.01; ***< 0.001.

FT37.9***0.960.1778.3***0.92−0.36
DS4.4***−0.370.014.1**−0.010.62
SK2.1−0.250.014.8**−0.270.17
MX8.5***0.500.120.7−0.02−0.06
TT8.3***−0.390.4216.3***0.50−0.16
ND14.6***0.500.6120.2***0.48−0.66
DM4.3**0.280.2226.7***0.61−0.33
HT2.9*0.10−0.4010.3***0.400.20
Figure 5.

 Canonical discriminant analysis. Shown are class means for Can1 and Can2 for the wet site: Arboretum (Arb) and dry site: Back Bay (BB) populations from the predrought 1997 (97) and post-drought 2004 (04) collection generations in the short season (S) and long season (L) treatments. Note that Can1 separates the short and long season treatments, and Can2 separates the populations and the generations within populations.

The relationships among the traits and their evolutionary changes can also be viewed using bivariate plots, also known as biplots. These plots show the vector direction and magnitude for each trait plotted in multivariate space. The biplots here give further support to the idea that season length appears to have a strong impact on the traits and the structure of their correlations. The short season and long season plots appear quite distinctive for both the Arboretum and Back Bay populations (Fig. 6). By contrast, the two populations seem fairly similar to each other for a given season length treatment (Fig. 6). Thus, population seems to have less of an influence on trait structure than season length.

Figure 6.

 Biplots. Shown are biplots, which give the eigenvectors from a principal components analysis plotted with the axes as principal component 1 (PC1) and principal component 2 (PC2). The traits are days to first flowering (FT), duration of flowering (DS), skew (SK), ratio of maximum number of flowers open on 1 day to total number of flowers (MX), total number of flowers (TT), number of leaf nodes (ND), stem diameter (DM) and plant height (HT) at first flowering. The plots are for the wet site (Arb), dry treatment (a); wet site (Arb), wet treatment (b); dry site (BB), dry treatment (c); and dry site (BB), wet treatment (d). Vectors shown with solid lines are for plants from the predrought 1997 generation, and vectors shown with hatched lines are from the post-drought 2004 generation.

It is also interesting to note that the biplots reveal a distinct pattern in the direction of trait evolution. The evolutionary patterns can be seen by comparing the vectors for 1997–2004. When comparing these vectors for the dry site (Back Bay) in the dry treatment (short season) as well as for the wet site (Arboretum) in the wet treatment (long season), the vectors generally rotate in the anticlockwise direction from 1997 to 2004 (Fig. 6). By contrast, the vectors generally rotate in the clockwise direction from 1997 to 2004 for the dry site in the wet treatment as well as for the wet site in the dry treatment (Fig. 6). This shows roughly orthogonal shifts in the relative importance of PC1 vs. PC2 when the population conditions and treatment was congruent (dry site and dry treatment or wet site and wet treatment) in contrast to when the population conditions and treatment were opposed. In other words, the evolutionary changes appeared to follow one multivariate trajectory when population and treatment soil moisture conditions were similar and the opposite trajectory when they were different. This finding is similar to that shown by the CPC analyses above, and further indicates that the largest disruptions in the relationships among the life-history traits seems to occur when there is the greatest mismatch between past and current conditions.

Discussion

Life-history evolution and climate change

We found that a natural drought caused rapid evolutionary changes in multiple life-history traits in B. rapa. Several life-history traits differed between the ancestral genotypes collected before and descendant genotypes collected after the natural drought. This shows directly that an evolutionary change in the life-history traits has occurred during a 5-year drought.

The evolutionary changes in trait levels following the drought are consistent with predictions from life-history theory. For example, descendant genotypes flowered earlier than ancestors. An earlier flowering time allows plants to escape late season drought and have higher fitness in drought years than late flowering genotypes (Cohen, 1976; Kozlowski, 1992). Descendants had a longer duration of flowering than ancestors due to the shift to earlier flowering and the lack of an equivalent change in date of last flowering, possibly because of greater evolutionary pressure on date of first than of last flowering. The plants tend to continue to flower for longer given sufficient resources, as shown by the fact that plants in the long season treatment flowered for twice as long as plants in the short season treatment. The drought also caused an evolutionary shift to lower skew and lower maximum number of flowers. This resulted in flatter, more evenly spread flowering schedule curves for the descendants compared with that for the ancestors. These results suggest that the drought selected for individuals that flowered earlier, continued to flower for longer given sufficient resources, and produced a more consistent, evenly distributed pattern of flowering over time.

Previous reviews have suggested, based on observational evidence, that changes in climate can alter patterns of phenology (Fitter & Fitter, 2002; Peñuelas et al., 2002; Parmesan & Yohe, 2003; Menzel et al., 2006) as well as plasticity in phenology (Nussey et al., 2005) and correlations among life-history traits (Both & Visser, 2005). By using the direct approach of comparing ancestors and descendants, we showed here that a shift in phenology following a climatic fluctuation was a true genetically based evolutionary change rather than an expression of phenotypic plasticity. This approach of comparing ancestors and descendants to show an evolutionary change in phenology due to climate change has been used only in a few previous studies, including work on photoperiod response in pitcher plant mosquitoes (Bradshaw & Holzapfel, 2001, 2008) parturition date in red squirrels (Realéet al., 2003) and time to first flowering in B. rapa (Franks et al., 2007). Whereas these previous studies have focused on a single phenological trait, we report here that a climatic fluctuation can lead to the rapid evolution of a suite of interconnected life-history traits. Examining a range of interacting life-history traits is important for understanding the full impact of changes in climate on both developmental and evolutionary trajectories and for assessing the degree to which development may constrain the rate of evolution.

Developmental constraints

Descendant, post-drought genotypes flowered earlier than ancestral, predrought genotypes and had fewer stem nodes and thinner stems. This suggests that descendants flowered at an earlier ontogenetic stage rather than developing more rapidly and flowering at the same stage. This interpretation is also supported by the fact that time to first flowering was highly positively correlated with both number of stem nodes and stem diameter at the time of first flowering. A flexible switch from vegetative growth to the onset of reproduction has also been found in several other systems (Jones, 1992; Wiltshire et al., 1994; Jones & Watson, 2001). It thus seems that for at least some species with indeterminate growth, the onset of reproduction is somewhat uncoupled from vegetative growth, and developmental constraints may not impose severe limits on responses of flowering time and other life-history traits to selection.

Multivariate life-history evolution

All of the life-history traits we measured showed a high degree of phenotypic correlation. If they are also genetically correlated (which was not determined in this study), evolution will act not only on traits individually but also through their correlations and potentially on the structure of these correlations itself. Our multivariate analyses suggest that drought can not only change life-history traits in B. rapa, but can also alter some aspects of the structure of the covariance matrix among these traits.

Although the trait covariances matrices were in some cases similar comparing the ancestors and descendants, there were also some differences, indicating evolutionary changes in the relationships among the traits. For example, in the Arboretum site, the trait correlation matrix for the 2004 generation was proportional to the 1997 generation, with the coefficient of proportionality less than one. This means that there was less of a close relationship among the traits following the drought and gives evidence that the drought selected for decoupling such traits as flowering time and size at first flowering or at least disrupted such relationships. This disruption makes sense in the light of the fact that selection during the drought was probably the strongest on the time of first flowering and less strong on size at first flowering. This differential selection pressure should result in a weaker correlation between size at flowering and timing of flowering especially if, as appears to be the case, there are not absolute developmental constraints on flowering time.

In addition to the differences in the trait covariance matrices between generations, there were also substantial differences between the dry site and wet site populations. Because the populations differ in soil moisture levels (Franke et al., 2006), the differences in the relationships among the traits could be due to evolutionary divergence driven by differences in drought intensities at the two sites. However, the sites could also differ in other factors, or the populations could differ in the trait relationships due to drift rather than selection.

The major differences in the trait covariance matrices between the short and long season treatments further indicate that drought can substantially alter the relationships among traits. These differences could be caused either by plastic responses of individuals to different environmental conditions or by drought selecting on different trait combinations and leading to evolutionary changes in the trait covariance matrix. To distinguish between these possibilities, it would be interesting to conduct future studies that directly examined changes in genotypic covariance matrices under different environmental conditions.

Although differences in many of the life-history trait values between the ancestors and descendants indicate rapid evolution of these traits following drought, there was not as much evolutionary change in the trait covariance matrix. This finding bodes well for quantitative genetic models that assume constancy of the P and G matrices in predicting short-term evolutionary changes (Lande & Arnold, 1983; Turelli, 1988; Arnold, 1992). However, the limited differences in the matrices between the ancestors and descendants and the larger differences between watering treatments and between populations do indicate the potential for changes in the covariances among life-history traits, which should certainly be considered in long-term models of evolutionary change.

Multivariate analyses showed that the life-history traits tended to group into those expressed before and those expressed after first flowering. These analyses also showed that these groups of traits were all well separated by season length, population and generation. Furthermore, the watering treatment strongly influenced the relative importance of the preflowering and post-flowering traits. Post-flowering traits such as duration of flowering and total number of flowers are probably plastic and responsive to such environmental conditions as soil moisture. Preflowering traits, by contrast, have been shown previously to be strongly genetically determined, and to evolve following generations of changes in conditions (Franks et al., 2007). Thus, changes in season length led both to plastic changes in patterns of post-flowering reproductive allocation and to evolutionary changes in preflowering traits.

Multivariate evolutionary trends differed depending on whether the soil moisture conditions in the source populations and experimental treatments were similar or different. For example, if the population and treatment conditions were similar (dry site with dry treatment or wet site with wet treatment) CPC analysis showed that the 2004 generation covariance matrices were proportional to those from the 1997 generation. If the population and treatment conditions were opposed, then the covariance matrices were less similar between 1997 and 2004, and thus showed more of an evolutionary change. In addition, the biplots showed that the vectors associated with each trait rotated in the clockwise direction between 1997 and 2004 in the treatment and population conditions were similar and in the anticlockwise direction if the conditions were opposed. This shows that the multivariate evolutionary trajectory depended on the interaction between the population and treatment conditions. If, based on the trait loadings, PC1 is taken to be flowering effort and PC2 flowering timing, then the relative importance of flowering timing vs. effort seems to differ when the populations and environment are similar compared with when they differ. This would make sense if selection to disrupt the relationship between reproductive timing and effort is the greatest when conditions do not match. For example, if the wet site experiences a dry treatment, then there is greater selection for earlier flowering, leading also to a lower correlation between flowering time and other traits, than if the dry site experiences a dry treatment, as the trait of flowering time and the relationships among the traits should already be closer to the optimum for dry conditions in that case.

Previous work with other species has also revealed complex multivariate evolutionary patterns and has shown that directional selection can weaken or disrupt correlations among characters. For example, Blows et al. (2004) examined a set of cuticular hydrocarbons on the fruit flies Drosophila serrata. They found that females exerted directional sexual selection on male cuticular hydrocarbon patterns, and that the direction of this selection was in opposition to the direction of additive genetic variance for these traits. This suggests that the directional selection exerted by females could reduce and disrupt the genetic variances and covariances for the hydrocarbon traits. This finding is similar to our result that directional selection on flowering time can disrupt associations among related life-history traits, which has important implications for predicting multivariate evolutionary change.

The importance of complete phenological information

The vast majority of studies on plant phenology focus exclusively on date of first flowering (Fitter & Fitter, 2002). This approach potentially loses a great deal of important information on the full distribution of reproductive availability over time. This may not be a problem if other parameters of the flowering schedule are highly correlated with time to first flowering. Our results for B. rapa, however, suggest otherwise.

We found that although time to first flowering was significantly correlated with other parameters of the flowering schedule, time to first flowering explained relatively little of the variation in these additional parameters. For example, time to first flowering explained only 9% of the variation in total number of flowers and only 1% of the variation in duration in flowering. These additional parameters are critical for determining many factors of ecological and evolutionary importance such as fitness, degree of assortative mating and reproductive isolation within and among populations. Thus, for questions pertaining to these important issues, it may be necessary to obtain full flowering schedules and to calculate the life-history parameters that can be obtained from this complete distribution of reproductive availability.

The ancestor–descendant approach to quantifying evolution

The fact that multiple life-history traits and the relationships among these traits evolved after only a 5-year drought indicates that global climate change is likely to have rapid and complex effects on the evolutionary trajectories of many species. By using the ancestor–descendant comparison approach, it is possible to take advantage of changing climatic conditions to gain insight into evolutionary processes. This approach is particularly useful for disentangling evolutionary changes in suites of integrated life-history traits.

Traditional quantitative genetic methods (Lande & Arnold, 1983; Lynch & Walsh, 1998) predict changes in multivariate trait means based on selection gradients and on the matrix of genetic variances and covariances (G matrix). This method assumes constancy in the selection gradients and G. In our study, we examined the phenotypic rather than the genotypic covariance matrix; so, our inferences about changes in the G matrix and about potential future multivariate evolution are indirect. But given this caveat, our results show that relationships among several life-history traits in B. rapa changed only a little after a few generations, making the assumption of consistency in the G matrix likely to be valid at least over the short term. However, the watering treatment and population differences showed that large changes in the correlation matrix are possible. Regardless, the ancestor–descendant method circumvents this limitation by examining multivariate evolution directly. The trait values of ancestors and descendants, as well as the covariances among these traits, can be examined to determine how the traits and their relationships have changed. This method thus allows testing of multivariate trait evolution hypotheses such as those related to the extent and importance on developmental constraints, and provides additional information which can be used in developing improved predictions of evolutionary change.

Acknowledgments

We thank the following for help with data collection and plant maintenance: K. Afshar, V. Chandrasekaran, A. Dick, A. Franks, L. Gonzalez, C. Herman, L. Hua, E. Ko, T. Kossler, P. Le, K. Musser, A. Ng, M. Ngugen, A. Ogura, P. Rath, S. Sim, K. Torosian, P. Tran, W. Yang, A. M. Weis, and E. Weiss. D. Franke made the 1997 seed collections. Discussion with Kjell Bolmgren, Elizabeth King and Ellen Simms improved the manuscript. This research was supported by NSF grants DEB-0345030, DEB-0440595 and DEB-0636812 to A. E. Weis.

Ancillary