The study population was derived from a wild population of A. thaliana (Brassicaceae), a highly selfing temperature annual. It was originally collected in Kendalville, Michigan, USA. A bulk sample from the Kendalville population has been maintained by Lehle Seed Inc. (Round Rock, TX, USA), and a previous study in our lab had established a total of 68 full-sib maternal families from this bulk collection (Camara et al., 2000). All 68 of these were used as the base population for artificial selection. The 68 random families in the base population are referred to as ‘founding genotypes’. Beyond this base generation, our artificial selection protocol results in maternal families that are not necessarily random, and we therefore refer to them as ‘sublines’. Also, given the almost completely selfing mating system (Abbott & Gomes, 1989), we chose not to impose outcrossing artificially. This departure from some previous studies (e.g. Ward et al., 2000; Fischer et al., 2004) mirrors the choice made by other researchers (e.g. Mauricio, 1998). The realism and potential biases introduced by this choice are discussed later in this report.
All studies were conducted in an air-conditioned, artificially lit (see below) walk-in growth room, with high plant densities. Seeds were imbibed in the dark for 48 h on Petri plates containing an agar-based medium of MS Basal Salts (Sigma, St. Louis, MO, USA). Plates were then randomly assigned to light treatments. After 1 week, seedlings were transplanted into small pots in 96-pot greenhouse flat inserts, with each pot containing about 120 mL of standard greenhouse growth medium (Fafard #2, Agawam, MA, USA) and returned to their assigned light treatment. Flats were bottom watered twice weekly. Temperature was maintained at 22–24 °C.
Characterizing the base population
We grew 68 founding genotypes in two light treatments: control and reduced R : FR, both with an 18L/6D photoperiod. In the control treatment we used a bank of seven standard fluorescent tubes (F20T12 Sylvania, Danvers, MA, USA) interspersed with four 25 W incandescent bulbs. In the reduced R : FR, three of the four standard tubes were replaced with specialized tubes that emit far-red wavelengths (F20S-FR74 Toshiba Ltd., Tokyo, Japan), with standard and specialized tubes alternated and again interspersed with four 25 W incandescent bulbs. The R : FR ratios were approximately 1.0 and 0.5 in the control and reduced treatments, respectively. Quantities of photosynthetically active radiation was approximately the same in both treatments. There were seven replicate flats per treatment, with one individual per genotype in each flat (n = 952). In this and in follow-up experiments, flats were rotated frequently to minimize any effects of patchiness in the light treatments.
All plants were checked at least every other day for bolting, indicated by the appearance of ∼2 mm of an elongating inflorescence. We recorded the number of days between bolting and initial exposure of seeds to light; we also counted the number of rosette leaves. For each founding genotype, the leaf number data was used to calculate a mean plasticity index defined as the proportional change in mean leaf number when plants are exposed to reduced R : FR conditions [=1− (genotype mean in low R : FR/genotype mean in high R : FR)]. This index was used as the basis for selection.
To evaluate whether this base population harboured genetic variation for the plasticity of either bolting time trait (i.e. leaf number at bolting or days to bolting), we analysed data using SAS PROC MIXED and restricted maximum likelihood (REML) (SAS Institute Inc., 1999). We conducted separate analyses for the two traits. Our mixed models included two levels of the R : FR treatment [fixed effect, tested with an F-value and Satterthwaite approximated degrees of freedom(d.f.)], genotype, and the genotype-by-R : FR treatment interaction (random effects, tested with Wald's Z). Results of these tests did not differ from tests involving sequential comparison of models with individual random effects included vs. excluded (Littell et al., 1996). If a population harbours genetic variation for a trait (expressed within environments or averaged across environments), there should be a significant variance component for the genotype effect. A broad-sense heritability (H2) can be estimated by the ratio of this variance component to total variance. If a population harbours genetic variation for plasticity, there should be a significant variance component for the genotype-by-R : FR treatment interaction term; a broad-sense heritability for the plasticity of a trait can be calculated as the ratio of this variance component to total variance (Scheiner, 1993; Pigliucci, 2001; van Kleunen et al., 2002). Broad-sense heritabilities are often considered to be inflated in comparison to narrow-sense heritabilities, but are perhaps more appropriate and relevant here for two reasons: we are working with a naturally selfing plant and our selection protocol targets plasticity, which is an attribute of a group of individuals with similar genotype (e.g. inbred families, clones) rather than individuals.
Establishing the lines
Six lines were initiated from the base population of 68 founding genotypes: two ‘high plasticity’ (HP1, HP2) and two ‘low plasticity’ (LP1, LP2) selection lines and two control lines (C1, C2). We used a line-sorting procedure that narrowed the number of founding genotypes gradually, across three episodes of selection, rather than all at once. This procedure maintained a large number of sublines in each line across several generations, avoided any artificial disruption of this species’ highly selfing mating system, and sampled multiple maternal plants of the same genotype in an effort to represent any potential within-family variation.
Using the plasticity index for each genotype, we identified the top (or bottom) 34 genotypes in the base population. For the 17 genotypes in the top (or bottom) quartile, three randomly selected maternal plants (51 maternal plants total) were chosen to contribute seed for three sublines in the next generation. For 17 genotypes in the next highest (or next lowest) quartile, one randomly selected maternal plant contributed seed for just one subline in the next generation. (To minimize heterogeneity of maternal effects, only maternal plants grown under high R : FR conditions contributed seed from one generation to the next.) To create control lines, we randomly chose 17 founding genotypes and drew seed from three randomly selected maternal plants and another 17 founding genotypes from which we drew seed from just a single randomly selected maternal plant. Selection procedures for the HP, LP and C lines were repeated to produce a second line of each type, drawing from the same 68 founding genotypes in the base population, but not necessarily from the same maternal plants (6 lines × 68 sublines = 408 sublines). Since the same founding genotypes may have contributed seed to both selection lines, we were careful to track the founding genotype of each subline throughout the study. Also, replicate lines are not completely independent with respect to the initial episode of selection.
During the subsequent episodes of selection, however, sublines within all six lines were handled independently. This entailed growing all 408 sublines for a second and a third generation in the high and low R : FR treatments in the growth room, with three individuals per subline per treatment (n = 2448). Individual plants were scored for leaf number and the R : FR plasticity index was calculated for each subline in each generation. Between the second and third generation, this index was used to rank and select sublines for each HP and LP line; sublines were randomly selected for each control line. A final episode of selection occurred after the third generation and prior to the follow-up experiment for evaluating the selection lines. The R : FR plasticity index was used to rank sublines and the top (or bottom) 18 sublines were selected within each line, with one randomly selected maternal plant per subline contributing seed. This shift in the selection protocol was logistically necessary because the experiment to evaluate selection lines involved four rather than just two light treatments and our growth room could not accommodate multiple individuals from all 408 sublines in all four environments.
For all six lines, selection differentials, S were calculated for the base, second and third generations of the selection experiment as the difference between the mean of selected parents and the mean of all individuals in the parental generation before selection was imposed. Means were weighted since sublines differed in the number of progeny contributed (see above). The response to selection, R was calculated as the difference between the current generation's mean as compared to the previous generation's mean. For the four selection lines, S and R were standardized by dividing each line's differential and response by the differential and response for the C1 and C2 lines averaged. Standardized responses to selection were plotted against standardized selection differentials. Finally, cumulative responses and differentials across the three episodes of selection were calculated and used to estimate realized heritabilities, h2 = R/S.
Loss of sublines; expected and actual sorting of sublines
In the base population study, 42 of 952 plants (4.4%) failed to germinate or died as seedlings. No sublines were lost. In the next two generations, lack of germination or early mortality sometimes resulted in loss of sublines. After the second generation, three sublines were lost from the HP2 line, two each from HP1 and LP2 and one each from C1, C2 and LP1. After the third generation, one additional subline went extinct in LP1.
Our study is clearly a departure from experimental evolution studies conducted with out-crossing species, because Arabidopsis has a highly selfing mating system and we did not experimentally impose an out-crossed mating scheme (given its artificiality for this species). Potential limitations and biases introduced by using a selfing species are discussed later in this report. Such an approach also entails selection via line sorting, and here we briefly sketch its potential impact given the protocols described above. At one extreme, the final 18 sublines in any given selection or control line could represent 18 founding genotypes (i.e. of a total of 68 in the base population). At the other extreme, the three episodes of selection described could result in representation by a minimum of two founding genotypes. Also, since replicate lines were derived from a common pool of founding genotypes, ‘sharing’ of founders was possible and is expected to be more prevalent in selection as compared to control lines. To examine the extent to which these potential line-sorting artifacts are actually present in our lines, we tracked each subline's founding genotype in all six lines.
Follow-up experiment: evaluating lines
We conducted a factorial experiment that exposed all sublines to high and low R : FR conditions and to both 12L/12D (‘short’) and 18L/6D (‘long’) treatments. In the short day treatment, 12 h of the high or low R : FR lights described above were used for the short treatment (see Characterizing the base population). In the long day treatment, these conditions were augmented with 6 h of dim incandescent light (i.e. low R : FR at the end of the day). A light-proof curtain shielded plants in the short photoperiod treatment from this extended period of dim light. The ecological realism of these conditions represents a compromise between two other experimental objectives. First, we aimed at equalizing the quantity of PAR not only between contrasting R : FR conditions, but also between short and long photoperiods. Second, we chose ‘short’ and ‘long’ photoperiod that would not unduly prolong the experiment and such that the ‘short’ treatment elicited a significant delay in bolting time.
The experiment involved a total of 7 × 18 = 126 sublines (n = 2106), with the seventh line composed of 18 sublines randomly drawn directly from the base population (saved as seed). If fewer than three plants per treatment were available for estimating a subline's trait mean and plasticities, the subline was omitted from the analysis (four from C2; two each from HP2, LP1 and LP2; one from C1). Plants grown from saved seed from the base population behaved anomalously, bolting with many fewer leaves compared to all other lines and expressing different phenotypes than in the base population study. This line was exposed to unique seed storage, and studies with other Arabidopsis genotypes have demonstrated that seed age, stratification and maternal environment effects can affect life-history traits such as flowering time (Shaw et al., 2000; Munir et al., 2001). Although we included data from the base line in our analyses, this paper compares the HP and LP selection lines to the average of the controls lines (C).
Data were analysed using two approaches. First, with data from individual plants, we performed a mixed model analysis using SAS PROC MIXED with REML. Leaf number data were square root transformed to improve normality. F-values with Satterthwaite approximated d.f. were used to test the fixed effects of selection line, R : FR treatment, photoperiod treatment and two- and three-way interaction terms. Variance components were calculated and tested for significance with Wald's Z statistic for four random effects, all nested within selection line: subline, subline-by-R : FR treatment, subline-by-photoperiod treatment and subline-by-R : FR treatment-by-photoperiod treatment.
Data from the follow-up experiment were also used to calculate each subline's R : FR-mediated plasticity index for leaf number (as described above), a comparable R : FR-mediated plasticity index for days to bolting and a photoperiod-mediated plasticity index. This latter index was defined as the proportional acceleration in mean leaf number when plants were exposed to long photoperiods [=1 − (subline mean for trait in long photoperiod/subline mean for trait in short photoperiod)].
For these three indices and for leaf number in high R : FR conditions, we examined differentiation among selection lines by conducting univariate one-way anovas and two orthogonal contrasts: between the two control lines and the four selection lines and between the two HP and two LP lines. Formal tests for unequal variance among selection lines (Levene's test) found heteroscedasticity for one of the plasticity indices (R : FR plasticity index for leaf number) and for leaf number and we could not successfully homogenize variances using transformations of either variable. In the case of leaf number, this is most likely because there is a bimodal distribution in the L1 and L2 lines. However, one-way anovas are robust to violations of heteroscedasticity and normality assumptions, particularly when samples sizes are well balanced, as in this case (Zar, 1999, p. 185). For simplicity, we present parametric analyses of untransformed data, with the caveat that deviations from normality potentially compromise single d.f. planned contrasts (Sokal & Rohlf, 1995).
To evaluate indirect responses to selection, we examined genetic correlations, estimated as Pearson product-moment correlations based on subline (i.e. family) means within HP, LP and C lines. We focused specifically on correlations between the plasticity subject to direct selection (i.e. the R : FR-mediated plasticity index for leaf number) and (1) the mean of that trait, (2) the R : FR-mediated plasticity index for days to bolting and (3) the photoperiod-mediated plasticity index for leaf number. This method of estimating genetic correlations has been criticized as biased (Lynch & Walsh, 1998), but the approach has been explored and supported by simulation studies (Roff & Preziosi, 1994; Windig, 1997) and employed for studies with Arabidopsis (Mauricio, 1998) and with the clonal plant Ranunculus reptans (van Kleunen et al., 2002). It is an appropriate approach here because plasticity indices cannot be scored on individuals, but require estimation from subline means (i.e. family means).
To supplement our examination of genetic correlations, we used an ancova approach that essentially asks whether direct and indirect responses to selection are linearly interdependent. The ancova models were modifications of the among-line anova. In all models, the plasticity index for leaf number was the dependent variable. Each of the three ancova models added one of the following continuous covariates and its interaction with the main effect of line: (1) mean leaf number, (2) R : FR-mediated plasticity index for days to bolting and (3) photoperiod-mediated plasticity index for leaf number. F-values were used to evaluate the significance of the covariate effect and the covariate-by-line interaction term (Willis, 1996).