Phenotypic Plasticity and Integration in Response to Flooded Conditions in Natural Accessions of Arabidopsis thaliana (L.) Heynh (Brassicaceae)

Abstract

Flood response is a crucial component of the life strategy of many plants, but it is seldom studied in non‐flooded tolerant species, even though they may be subjected to stressful environmental conditions. Phenotypic plasticity in reaction to environmental stress affects the whole plant phenotype and can alter the character correlations that constitute the phenotypic architecture of the individual, yet few studies have investigated the lability of phenotypic integration to water regime. Moreover, little has been done to date to quantify the sort of selective pressures that different components of a plant’s phenotype may be experiencing under contrasting water regimes. Genetic differentiation and phenotypic plasticity at the single‐trait and multivariate levels were investigated in 47 accessions of the weedy plant Arabidopsis thaliana, and the relationship of plastic characters to reproductive fitness was quantified. Results indicate that these plants tend to be highly genetically differentiated for all traits, in agreement with predictions made on the basis of environmental variation and mating system. Varied patterns of apparent selection under flooded and non‐flooded conditions were also uncovered, suggesting trade‐offs in allocation between roots and above‐ground biomass, as well as between leaves and reproductive structures. While the major components of the plants’ multivariate phenotypic architecture were not significantly affected by environmental changes, many of the details were different under flooded and non‐flooded conditions.

INTRODUCTION

Being sessile organisms, plants unavoidably experience fluctuations in their external environment which, under certain circumstances, may lead to the evolution of phenotypic plasticity in a variety of traits (Bradshaw, 1972; Pigliucci, 2001). Plasticity is the property of a given genotype to produce different phenotypes depending on the environment (Schmalhausen, 1949). Changing environmental conditions are also known to induce distinct patterns of character correlations (Schlichting and Levin, 1986), although the extent to which this is a result of natural selection for ‘phenotypic integration’ or of genetic constraints in the form of pleiotropy and/or linkage is not clear.

The evolution of patterns of phenotypic covariation is of particular relevance to two major areas of inquiry attempting to link macro‐ and microevolutionary processes (Steppan, 1997). On one hand, quantitative genetics models evolutionary change as the result of natural selection and other forces acting on the genetic architecture of phenotypes. On the other hand, research on evolutionary constraints views phenotypic covariances as the expression of underlying genetic and developmental constraints. As indicated by several authors (e.g. Lande, 1979; Arnold and Wade, 1984; Arnold and Phillips, 1999; Phillips and Arnold, 1999), covariance patterns are fundamental to quantitative models of phenotypic divergence over long periods of time. Thus, in order to apply microevolutionary models to the understanding of macroevolutionary patterns, the extent to which genetic covariation remains constant or proportional across populations or species over macroevolutionary time scales must be verified (Lande, 1979; Steppan, 1997). Therefore, the experimental and comparative study of phenotypic integration becomes a crucial component of evolutionary quantitative genetics.

Water is among the chief environmental factors that induce both plasticity of single traits and of character correlations (Pigliucci et al., 1995b; Conner and Zangori, 1998; Meyer and Allen, 1999; Barrilleaux and Grace, 2000). Water is obviously a crucial and highly variable abiotic factor for every living organism. For plants, both a shortage (drought) and an excess (waterlogging or flooding) of water is stressful and the two conditions elicit distinct coping mechanisms (Blom and Voesenek, 1996; Mauseth and Plemons‐Rodriguez, 1998; Galen, 2000; Zhang et al., 2000). In the case of excess water, an important problem is decreased oxygen concentration in the soil. This primarily affects the root system and, as a consequence, above‐ground growth.

Much attention has been paid to plant responses to flooding (Baruch and Merida, 1995; Rubio et al., 1995; Blom and Voesenek, 1996; Visser and Voesenek, 1996; Youssef and Saenger, 1996; Moog, 1998; Grichko and Glick, 2001), with research focusing primarily on flood‐tolerant species such as rice, Carex and Rumex sp. However, it is equally important to understand how species that are not flood tolerant cope with the occasional excess of water, e.g. due to a heavy rainfall event that can saturate soil pore spaces and cause anoxia.

We have investigated genetic differentiation of reaction norms to flooding as well as flood‐induced changes in patterns of character correlations in a collection of populations of the wild mustard Arabidopsis thaliana (L.) Heynh (Brassicaceae). A. thaliana is a well‐established model system for studies in molecular biology and physiology (Meyerowitz, 1989; Meinke, 1994; Convay and Poethig, 1997; Coupland, 1997; Ellis et al., 1999), and has recently received much attention from an ecological and evolutionary standpoint (Clauss and Aarssen, 1994; Mauricio and Rausher, 1997; Li and Jun‐Ichirou Suzuki, 1998; Pigliucci, 1998).

In this paper we address the following questions: (1) is there genetic variation for trait means among populations when exposed to different water regimes that include flooded conditions? (2) Are there plasticity and variation for plasticity to flooding for the traits of interest? (3) Do plastic traits affect reproductive fitness, suggesting that they may be under natural selection depending on environmental conditions experienced by the plants? (4) Do environmental changes affect the variance–covariance matrix relating different life history and architectural traits, and if so, to what extent?

MATERIALS AND METHODS

Collections of seeds from 47 accessions of Arabidopsis thaliana (L.) Heynh. were obtained from the Arabidopsis Information Management System (AIMS at www.arabidopsis.org) (see Appendix). All accessions selected for the experiment represented early flowering populations of A. thaliana and had been bulk‐propagated at AIMS to maintain genetic variation. To minimize maternal effects and increase seed availability, we grew the material for one generation under controlled laboratory conditions of 16/8 h of light/darkness at 23–25 °C and provided bottom watering every other day. All other conditions were as detailed below.

Second‐generation seeds were placed on a wet filter paper and cold‐treated for 1 week at 5 °C in a refrigerator. Imbibed seeds were then transferred to a mix of top soil/coarse sand/turface (2 : 2 : 1 by volume) and placed under combined fluorescent and incandescent light on a three‐shelf growth rack. Seedlings were randomly thinned after 5 d, when the first set of true leaves had emerged, leaving one plant per 7 cm (diameter) × 5 cm (deep) pot. We applied two treatments: flooding and no flooding. Plants assigned to the non‐flooded regime were bottom‐watered every other day, letting the soil saturate with water for 2 h, and then drained. Pots containing plants assigned to the flooded treatment were kept at saturation, so that plants were waterlogged during the experiment. Water in flooded pots was changed every other day. Both treatments were top fertilized weekly for 5 weeks with 2 ml of an 11 : 11 : 11, N : P : K solution. All other conditions were as described above for the first generation. Plants were harvested near the end of their natural period of senescence (at most 4 months after planting) and measurements were taken as specified below.

Measurements

Three sets of traits were measured: vegetative, architectural and reproductive. Vegetative traits were quantified at the bolting stage, when the rosette begins to produce the flowering stem: (1) bolting time (time from seed planting to the beginning of the elongation of the main stem); (2) rosette leaf number, quantifying meristem allocation to vegetative growth; and (3) rosette diameter, a measure of plant size at bolting. Plant architecture traits were measured at the time of harvest, approx. 1 week after maturation of the first silique: (4) number of lateral branches; (5) above‐ground fresh weight (a measure of plant allocation of resources to above‐ground growth), comprising the rosette plus the main stem and branches bearing fruits; (6) below‐ground fresh weight (plant allocation of resources to below‐ground growth); and (7) total number of basal stems (allocation of resources to secondary meristems). Reproductive traits were also measured when plants set the first fruits: (8) time of first reproduction, when the first seeds matured and siliques started opening, counted as days from bolting (i.e. from the beginning of the reproductive phase); and (9) total fruit production (reproductive fitness). Weights were measured using a balance with a precision of 0·001 g. Metric characters were measured in millimetres using a digital calliper.

Experimental design and statistical analyses

Plants from each accession were randomly assigned to one of the two treatments (flooded or non‐flooded), with every accession represented by six replicates within each treatment. The total size of the experimental population was 564 plants (47 accessions × two treatments × six replicates). Individuals were placed in two three‐level light racks equipped with two fluorescent and two incandescent light tubes/bulbs per level, yielding a photon flux of 240 µmol m^–2 s^–1 and a neutral light spectrum. Each level on each rack housed four trays and contained two replicates of each family randomly assigned to one of the four trays, yielding 94 individual plants per level and 23–24 pots per tray.

Measured variables deviating from normality or homoscedasticity were transformed appropriately (Sokal and Rohlf, 1995). A nested mixed‐model ANOVA (split‐plot design: SYSTAT, 2000) was used to estimate the significance of the following factors: (1) accession, testing for genetic differentiation in character means among accessions independently of the environment; (2) treatment, estimating overall phenotypic plasticity independent of genetic effects; (3) accession × treatment interaction, testing for the existence of genetic differentiation for plasticity among accessions; and (4) tray (nested within treatment), estimating the degree of micro‐environmental variation due to the experimental set‐up. Treatment was considered a fixed effect, while accession was treated as a random effect. Following Sokal and Rohlf (1995), if the tray effect was significant, then the treatment effect was tested over tray. Also, if the accession × treatment interaction showed a significant effect, we tested accession over the interaction term. Otherwise, factors were tested over the error mean square. [This balanced use of conservative statistical tests is advocated by Sokal and Rohlf (1995), and we consider it better than always testing over interaction or lower‐level effects, regardless of their significance. Our approach compromises between the often diverging criteria of statistical and biological significance.] Given the high number of multiple comparisons, we used a sequential Bonferroni correction to adjust the nominal α‐values for the ANOVAs across rows in Table 1 (this correction is moderately conservative, as opposed to a straight Bonferroni correction which tends to overcorrect for type II errors; Rice, 1989). We then plotted mean (accession) reaction norms for all traits with a significant treatment or accession × treatment term.

Table 1.

ANOVA table showing mean squares (top line) and associated P‐values (bottom line)

Trait	Accession (46 d.f.)	Treatment (1 d.f.)	Accession × treatment (46 d.f.)	Tray (treatment) (22 d.f.)	Error (401–435 d.f.)
Bolting day	207·34	11·30	20·14	143·86	13·76
	0·0000	0·3653	0·0297	0·0000
Leaf number at bolting	84·29	32·00	7·20	23·81	5·01
	0·0000	0·2588	0·0364	0·0000
Rosette diameter (log)	0·81	37·10	0·22	0·57	78·05
	0·0000	0·0000	0·0000	0·0000
Set of first fruits (log)	0·09	0·17	0·04	0·09	0·03
	0·0022	0·1910	0·0336	0·0000
Shoot fresh weight (log)	1·68	684·13	0·49	3·31	0·39
	0·0000	0·0000	0·1259	0·0000
Root fresh weight	0·01	0·18	0·01	0·01	0·00
	0·0549	0·0000	0·0000	0·0000
Lateral branches	4·11	309·03	1·62	2·20	1·21
	0·0000	0·0000	0·0733	0·0138
Basal stems	4·37	1072·80	2·67	4·97	1·30
	0·0499	0·0000	0·0001	0·0000
Total fruit production (log)	0·85	442·20	0·33	1·48	0·22
	0·0000	0·0000	0·0158	0·0000

Bold type indicates statistically significant effects after a table‐wide sequential Bonferroni correction.

d.f. Indicates the degrees of freedom of each factor.

Transformations to compensate for lack of normality or heteroscedasticity are also detailed.

To explore the relationship between environmentally variable character expression and reproductive fitness, we used regression analyses investigating the presence of linear and/or quadratic relationships (Lande and Arnold, 1983) between each measured trait and total fruit production. Since our interest in this paper focused on environmental effects, for this analysis we considered only characters that showed either a significant treatment or treatment × accession interaction. Regressions were conducted separately for the two treatments.

We also calculated treatment‐specific correlation matrices and ran standard principal components analyses on them to visualize the sets of co‐varying traits in each environment as well as to determine how differentiated was the multivariate phenotype of the accessions. We ran a series of correlation tests comparing each eigenvector (in order of variance explained) in one environment with the corresponding one in the other environment to formally test the degree of multivariate similarity between the two matrices. We chose not to use common principal components analysis (CPC; Steppan, 1997; Phillips and Arnold, 1999), as advocated by some authors, because on several of our data sets this technique rejected the hypothesis of any similarity of structure when it was evident by visual inspection and by carrying out alternative tests that the matrices were in fact structurally similar. We think that this is most likely a problem with the sensitivity of the CPC method, which may have a tendency to reject the null hypothesis too often when the data sets are of sufficient size.

RESULTS

Genetic variation among accessions

ANOVAs showed that all traits except root fresh weight and basal stems were significantly variable among accessions (Table 1). However, these two traits—together with rosette diameter—were the only ones that showed any significant accession × treatment interaction, which indicates presence of inter‐accession differentiation for plasticity. Bolting day, leaf number at bolting and time to first reproduction were the only traits that did not show a main effect of treatment (i.e. overall plasticity). Not surprisingly, all of the traits measured showed significant effects due to micro‐environmental variation (tray within treatment effect).

Reaction norms were plotted for the traits showing a significant accession × treatment effect (Fig. 1). All three plastic traits had higher values under non‐flooded conditions than under flooded ones, in agreement with the intuitive expectation that flood constitutes a stress. However, some genotypes responded minimally to the change in environment, especially in terms of rosette diameter and root weight.

Fig. 1. Reaction norms of the traits showing genetic variation for plasticity. The LER genotype, a common laboratory line of arabidopsis used in many experiments, is found towards the bottom of each graph (i.e. it produced small rosettes, few roots and few basal stems under both treatments).

Effects of plastic traits on reproductive fitness

To explore the effect of plastic traits on reproductive fitness, we conducted a series of linear and quadratic regression analyses (Table 2). Under flooded conditions, all linear terms were significant, indicating directional selection for increase in rosette size, shoot weight, root weight (marginally significant), number of lateral branches and number of basal stems. As far as the quadratic terms were concerned, only shoot fresh weight was statistically significantly associated with reproductive fitness under flooded conditions. However, a visual inspection of the trait–fitness relationship (not shown) indicated that the quadratic term did not actually add to the explanatory power of the linear one and that its statistical significance was due to the presence of a few outliers.

Table 2.

. Regression analysis conducted on traits showing plasticity or genetic differentiation for plasticity

Treatment	Effect	Standardized coefficient	t	P (2‐tail)
	Rosette diameter	0·30	6·45	0·0000
Linear terms	Shoot fresh weight	0·35	0·55	0·0000
Flooded	Root fresh weight	0·10	2·09	0·0372
	Number of lateral branches	0·14	3·86	0·0001
	Number of basal stems	0·25	6·21	0·0000
Non‐flooded	Rosette diameter	–0·06	–1·15	0·2520
	Shoot fresh weight	0·56	9·89	0·0000
	Root fresh weight	0·09	1·60	0·1089
	Number of lateral branches	0·18	5·01	0·0000
	Number of basal stems	0·31	8·17	0·0000
	Rosette diameter	0·25	1·73	0·0855
Quadratic terms	(Shoot fresh weight)2	–0·94	–8·86	0·0000
Flooded	(Root fresh weight)2	–0·17	–2·05	0·0413
	(Number of lateral branches)2	0·10	0·94	0·3493
	(Number of basal stems)2	0·10	1·43	0·1540
Non‐flooded	(Rosette diameter)2	–0·48	–3·26	0·0013
	(Shoot fresh weight)2	0·00	–0·02	0·9811
	(Root fresh weight)2	–0·33	–2·89	0·0042
	(Number of lateral branches)2	–0·07	–0·55	0·5808
	(Number of basal stems)2	0·03	0·26	0·7939

The table reports the standardized selection coefficients and the associated t‐test and P‐values. Bold type indicates statistical significance. Linear and quadratic terms were included in the model. These are normally interpreted, respectively, as directional and disruptive/stabilizing (depending on the sign) selection.

When we conducted the same analysis for data from the non‐flooded conditions (Table 2) we found that shoot weight, number of lateral branches and number of basal branches showed statistically significant linear terms, while rosette diameter and root fresh weight were characterized by significant quadratic terms. Here, the quadratic terms were actually informative since there was a clear peak of higher fitness for intermediate rosette and root size, on each side of which the plant’s fitness clearly decreased (details not shown).

Multivariate phenotype

We calculated the correlation matrices among all traits under both environmental conditions (Table 3). The matrices appeared visually similar, differing only slightly in the magnitudes of some correlations. Three correlations were statistically significant under flooded conditions but not in the non‐flooded environment (Table 3); all involved fruit set, with bolting time (–0·22), leaf number at bolting (–0·24) and number of lateral branches (0·29), respectively. Conversely, four correlations were significant under non‐flooded but not under flooded conditions: three involved lateral branches and, respectively, root fresh weight (0·24), shoot fresh weight (0·32) and rosette diameter (0·38). The last correlation was between bolting day and rosette diameter (0·66). While the fact that one correlation is significantly different from zero and another one is not does not necessarily imply that the two correlations are in fact distinct, this appears to be the case in many of these instances when one considers the magnitudes of the same correlation under the two treatments (we did not carry out formal tests because of the high number of pairwise comparisons and the lack of a priori hypotheses about which pairs should be significantly heterogeneous).

Table 3.

Correlation matrix among characters for flooded conditions (below diagonal) and non‐flooded conditions (above diagonal)

	Bolting day	Leaf number	Root fresh weight	Number of lateral branches	Number of basal stems	Rosette diameter (log)	Senescence (log)	Shoot fresh weight (log)	Total fruit (log)
Bolting day	1	0·84	0·56	0·11	–0·10	0·66	0·15	0·41	0·20
Leaf number	0·77	1	0·59	0·19	–0·07	0·68	–0·01	0·37	0·19
Root fresh weight	0·39	0·51	1	0·24	0·23	0·71	–0·06	0·62	0·54
Number of lateral branches	–0·17	–0·04	0·15	1	0·18	0·38	–0·03	0·32	0·39
Number of basal stems	–0·01	–0·04	0·33	0·10	1	0·33	0·00	0·48	0·64
Rosette diameter (log)	0·01	0·34	0·47	0·09	0·30	1	–0·14	0·68	0·64
Set of first fruits (log)	–0·22	–0·24	–0·13	0·29	0·18	–0·17	1	0·04	0·07
Shoot fresh weight (log)	0·30	0·39	0·70	0·19	0·44	0·72	–0·14	1	0·79
Total fruit (log)	0·04	0·2	0·60	0·31	0·48	0·66	0·05	0·87	1

Bold type indicates significant correlations after Bonferroni correction.

To determine whether the environment altered the overall structure of the correlation matrix, we performed a principal components analysis to allow visual inspection of the eigenvectors and a vector correlation analysis to carry out a formal test of the degree of matrix similarity between treatments. Visual inspection of the eigenvectors (not shown) revealed that the first two factors (those explaining the largest amount of variance) were very similar, the major difference being a change in the magnitude of the vector associated with the time of first reproduction (which explained little of the variance on either component under non‐flooded conditions). For the rest, the two plots were essentially mirror images of each other along the second factor. Since the actual direction in multivariate space is arbitrary, this means that most of the relationships among variables were unaffected by the environmental change. A more careful inspection of the loadings, however, did reveal some subtle differences between the two treatments (Table 4), including a dissociation of lateral branches from the main group of traits under non‐flooded but not under flooded conditions, and a somewhat opposite behaviour of basal branches.

Table 4.

Principal components analyses detailing the composition of the first two eigenvectors under flooded and non‐flooded conditions

	Non‐flooded		Flooded
Component loadings	PC‐1	PC‐2	PC‐1	PC‐2
Eigenvalues	4·13	1·76	3·66	1·95
Percentage of total variance     explained	45·98	19·57	40·67	21·69
Bolting day	0·43	–0·72	0·69	0·63
Leaf number	0·58	–0·65	0·70	0·62
Root fresh weight	0·83	–0·09	0·82	0·13
Lateral branches	0·20	0·53	0·44	–0·23
Basal stems number	0·48	0·45	0·42	–0·74
Rosette diameter	0·77	0·11	0·92	0·09
Set of first fruits	–0·16	0·57	0·00	0·04
Shoot fresh weight	0·93	0·12	0·84	–0·28
Total fruit	0·83	0·40	0·77	–0·53

Bold type indicates which component (within each treatment) was associated with the highest load for a given variable.

A formal correlation analysis comparing vectors (in order of explained variance) between the two treatments showed a high degree of similarity of the first two eigenvectors (Table 5). However, the environment did have minor effects on the multivariate structure, as evidenced by the low and non‐significant correlations between vectors 3–7 and 9.

Table 5.

Correlations among the principal vectors expressed under the two environmental regimes

	Vector 1	Vector 2	Vector 3	Vector 4	Vector 5	Vector 6	Vector 7	Vector 8	Vector 9
Pearson’s coefficient (flooded vs. non‐flooded)	0·9306	–0·8344	0·4618	–0·6678	–0·0912	–0·3193	–0·1581	0·9701	0·5889
(P‐value)	(0·0003)	(0·0052)	(0·2108)	(0·0494)	(0·8155)	(0·4023)	(0·6846)	(0·0000)	(0·0953)
Total variance explained for flooded (%)	40·67	21·7	11·65	9·42	5·94	4·55	3·30	1·92	0·87
Total variance explained for non‐flooded (%)	45·98	19·57	11·88	9·26	4·66	3·66	2·16	1·58	1·25

The vectors explaining the majority of the variance appeared to be highly correlated with each other. Several of the smaller vectors, however, did show very low similarity between treatments. Bold type indicates significant correlations after Bonferroni’s correction.

DISCUSSION

Phenotypic evolution involves an understanding of the amount of variation for characters, of their lability to environmental conditions, their association with fitness, as well as their relationship with other aspects of the phenotype (Schlichting and Pigliucci, 1998). Here, we have attempted to characterize phenotypic divergence among accessions of a wild weedy species and to study how the correlations among traits are affected by changes in an important component of the environment, water availability. As Armbruster and Schwaegerle (1996) pointed out, studies carried out among accessions or populations give a picture of the outcome of recent evolutionary events leading to differentiation, rather than of potential response to environmental forces acting in the future—typically the target of intra‐population variation studies. Such an intermediate level of analysis is important to bridge population biology with macroevolution at and above the species level.

We investigated individual traits and their relationship with reproductive fitness, as well as the multivariate patterns of phenotypic integration and their lability to environmental change (Schlichting, 1989). There is a considerable interest in the study of phenotypic and genetic correlations because of their relevance to evolutionary theory (Roff and Mouseau, 1999), especially with regard to the validity of the assumptions embedded in quantitative genetic models of evolutionary change (Turelli, 1988; Pigliucci and Schlichting, 1997) and to the understanding of multivariate phenotypic evolution (Schlichting and Pigliucci, 1998).

Genetic variation among accessions

We observed widespread genetic variation for across‐environment trait means among our accessions of Arabidopsis thaliana when grown under contrasting water regimes. We also observed widespread phenotypic plasticity (six out of nine traits), while only three traits showed genetic differentiation for plasticity. These results indicate that there was much more genetic differentiation for trait means irrespective of the environment than differentiation for plasticity to water availability among accessions of A. thaliana. Similar results were reported by Pigliucci et al. (1995b) in their work on different populations of A. thaliana, showing a moderate amount of plasticity in the populations studied, but a lack of genetic differentiation for plasticity. This is consistent with the species’ life history: A. thaliana flowers in the spring and is probably exposed to random fluctuations in flooding regimes which depend on the local geography. Under these conditions, these plants are not expected to evolve adaptive plasticity to water, but rather, genetic specialization for whatever water regime they encounter more often (Pigliucci, 2001). In addition to fluctuating rainfall/water levels, edaphic conditions are important: sandy soils do not cause much water logging since water can percolate through the sediment with ease, whereas soils richer in silts or clay are character ized by slower drainage of water (Podbielkowski and Podbielkowska, 1992). Unfortunately, little is known about the edaphic conditions typical of A. thaliana populations, although personal observations (M.P.) seem to indicate that this species often occupies sandy soils.

Therefore, it is perhaps not surprising that our accessions did not show specialization for root size (one of the few traits that was not heterogeneous among provenances) and that all accessions performed poorly under the relatively novel environment of water‐logged conditions. This implies that the observed plasticity was mostly the result of a passive response to severe stress rather than an adaptive response. Similar results were also obtained by Anderson and Pezeshki (2001) in their work on three bottomland tree species. In their study, all of the species tested showed a decrease in dry root mass when exposed to varying periods of flooding, with the highest decrease elicited by long‐term flooding.

In general, a high degree of differentiation among accessions for trait means has been observed in other studies of this species (Pigliucci et al., 1995a, b) and is consistent with the highly selfing mating system of this taxon (Abbot and Gomes, 1987), which leads to high among‐population and low within‐population genetic variation. It is an open question whether the observed differentiation is chiefly the result of historical (drift) or deterministic (selection) phenomena.

Plasticity and fitness

Using regression analysis performed on the traits showing plasticity or genetic differentiation for plasticity (rosette diameter, number of lateral branches, number of basal stems, and shoot and root fresh weight), we detected directional selection for an increase of all these traits under the flooded treatment, and directional selection for an increase in shoot fresh weight and degree of branching (both lateral and basal stems), as well as stabilizing selection for rosette size and root weight under non‐flooded conditions. This indicates that it is advantageous to invest more in aerial structures (especially shoots and branches) regardless of the environment, and no selection on plasticity per se on these traits would be expected. The observed differentiation for plasticity of basal stems is therefore unlikely to be the result of past selection in response to varying water conditions. Surprisingly, few studies of selection on plant plasticity in response to water can be found in the literature. The only work available (Dudley, 1996a, b) actually deals with adaptation to dry conditions, and is not therefore particularly informative for our system.

Under flooded conditions, the vast majority of plants produced very small roots, again indicating that they were under severe stress caused by the anoxic conditions of the soil (since the water level never dropped below full soil capacity in our experiment). Indeed, we observed that plants grown under flooded conditions not only had much shallower root systems, but produced adventitious roots from their hypocotyls, a well‐known compensation mechanism employed by plants grown in anoxic environments (Armstrong and Brandle, 1994). Under non‐flooded conditions, root size seemed to have an optimal intermediate level, because plants with root systems that were both too small or too large were at a clear disadvantage. This is probably due to the fact that too little root growth does not provide enough nutrients and water for normal plant growth, but too much allocation to roots when water is not particularly scarce may be detrimental to above‐ground biomass. Also, under non‐flooded conditions, we found stabilizing selection on rosette size, implying the existence of an optimal intermediate value for this trait on either side of which reproductive fitness is lowered if water conditions are normal. This could be due to the detriment to reproductive structures that may be caused by either too little or too much allocation to vegetative ones.

Finally, we have observed apparent ‘stabilizing selection’ under flooded conditions for shoot weight, but this was probably an artefact caused by the presence of a few outliers, and the linear term of the model was sufficient to explain the observed pattern.

Environment and the stability of the covariance matrix

While classical studies in evolutionary ecology have tended to focus on the variation in single characters (with some notable exceptions: Berg, 1960; Clausen and Hiesey, 1960), there has been a recent increase in interest in the co‐variation among characters and its consequences for phenotypic evolution (Steppan, 1997; Arnold and Phillips, 1999; Phillips and Arnold, 1999; Waldmann and Anderson, 2000). We were particularly interested in the relationship between phenotypic integration (assessed by the pattern of character correlations) and environmental variation, i.e. in how the environment can alter the patterns of phenotypic correlations among traits, which in this case represent genetic differentiation among accessions.

Both a visual inspection of the correlation matrices obtained under either environment and the principal components analyses showed a fairly high degree of similarity between the major aspects of the character architecture as expressed under the two treatments. As indicated by our vector correlation analyses, the first two vectors—which explain over 60 % of the observed variance—were highly correlated to each other across treatments. This implies that the environment did not significantly affect the mechanisms underlying character correlations. However, it is also noticeable that many of the minor components were in fact not correlated, if one takes the overall degree of variance explained by each of them as a sufficient matching criterion (which it may not be, especially for the components explaining very small portions of variance).

Lack of matrix divergence among populations has been observed before, for example by Arnold and Phillips (1999) during the work on coastal/inland divergence in garter snakes. Roff and Mouseau (1999) reviewed the literature on divergence of genetic correlations across different taxonomic levels and found that the results were mixed, as one would expect considering the heterogeneity of methods, types of characters, and taxa that have been employed and sampled to date.

The field of multivariate phenotypic evolution is also plagued by methodological problems. For example, Roff (2000) compared various methods for examining multivariate genetic/phenotypic divergence and found that no single method yields satisfactory results. Our own experience with the currently popular common principal components (CPC) analyses (Phillips and Arnold, 1999; Waldmann and Andersson, 2000) is actually rather unsatisfying: both in the current study and on other occasions (M. Pigliucci and A. Kolodynska, unpubl. res.), we discovered that CPCs tend to be too sensitive and yield a verdict of no similarity among matrices even when it is obvious by both visual inspection and other methods (Mantel tests, vector correlation analyses) that there is a high degree of overlap between matrices. This problem of finding satisfactory statistical methods to quantify changes in phenotypic integration (see also Smouse et al., 1986; Cowley and Atchley, 1992; Shaw, 1992; Shipley, 2000) is perhaps the major stumbling block against progress in this important field of inquiry into phenotypic evolution.

ACKNOWLEDGEMENTS

The authors would like to thank Mitchell Cruzan, Otto Schwarz and Randy Small for valuable comments on previous drafts of this manuscript. This work was carried out with partial funding from the National Science Foundation to M.P. (grant DEB‐0089493).

Table 1A.

Accession number	Source
CS0911	Estland (Germany)
CS0913	Petergof (Russia)
CS0915	Wassilewskija (Russia)
CS0916	Condara (Tadjikistan)
CS0917	Darmstadt (Czechoslovakia)
CS0920	Enkheim (Ukraine)
CS0922	Hodja‐Obi‐Garm (Tadjikistan)
CS0925	Litvania (Litvania)
CS0932	Aberdeen (UK)
CS1184	Gudow (Germany)
CS1214	Guckingen (Germany)
CS1226	Hilversum (Netherlands)
CS1240	Isenburg (Germany)
CS1252	Vranov (Czechoslovakia)
CS1258	Jamolice (Czechoslovakia)
CS1282	Rodenbach (Germany)
CS1284	Köln (Germany)
CS1504	Seis (Italy)
CS1514	Slavice (Czechoslovakia)
CS1604	Wietze (Germany)
CS1630	Wildbad (Germany)
CS1635	Canterbury (UK)
CS1637	East Malling (UK)
CS1640	Tsu (Japan)
CS3109	Copenhagen (Denmark)
CS3110	Weiningen (Switzerland)
CS3179	Graz (Austria)
CS3180	Coimbra (Portugal)
CS6003	Köln (Germany)
CS6004	Maidstone Kent (UK)
CS6015	West Malling, Kent (UK)
CS6016	Maidstone (UK)
CS6023	Sedmouth (UK)
CS6034	Bretagny (France)
CS6036	Bretagny (France)
CS6038	Kelsterbach (Germany)
CS6041	Kelsterbach (Germany)
CS6046	Köln (Germany)
CS6047	Maidstone (UK)
CS6068	Kent (UK)
CS6105	Kelsterbach (Germany)
CS6187	Washington (USA)
CS6194	Blanes (Spain)
CS6195	Wurzburg (Germany)
CS6682	Dijon (France)
CS6731	Gluckingen, (Germany)
LER, Landsberg erecta

Received: 21 November 2001; Returned for revision: 6 February 2002; Accepted: 15 April 2002