Rapid reversal of a potentially constraining genetic covariance between leaf and flower traits in Silene latifolia

Abstract Genetic covariance between two traits generates correlated responses to selection, and may either enhance or constrain adaptation. Silene latifolia exhibits potentially constraining genetic covariance between specific leaf area (SLA) and flower number in males. Flower number is likely to increase via fecundity selection but the correlated increase in SLA increases mortality, and SLA is under selection to decrease in dry habitats. We selected on trait combinations in two selection lines for four generations to test whether genetic covariance could be reduced without significantly altering trait means. In one selection line, the genetic covariance changed sign and eigenstructure changed significantly, while in the other selection line eigenstructure remained similar to the control line. Changes in genetic variance–covariance structure are therefore possible without the introduction of new alleles, and the responses we observed suggest that founder effects and changes in frequency of alleles of major effect may be acting to produce the changes.

thinner leaves than females (Delph & Bell, 2008;Delph, Knapczyk, & Taylor, 2002). Although the covariance is positive, it represents a potential constraint to the evolution of traits toward their optima in male plants. Increased flower number contributes to higher fecundity, but the correlated increase in SLA means plants also make thinner leaves, which are associated with higher transpiration rates and increased mortality . Selection for decreased SLA in males is significant in dry habitats (Delph, Andicoechea, et al., 2011).
Despite the conflict between SLA and flower number in S. latifolia, a positive and significant correlation between them has been observed in several studies. A phenotypic correlation between flower number and SLA was seen across nine populations in the United States and Europe (Delph et al., 2002), and in a different population the estimated genetic correlation between SLA and flower number was positive and significant (Delph, Andicoechea, et al., 2011). A selection experiment on flower size generated correlated change in both flower number and SLA, suggesting genetic integration of these traits (Delph, Gehring, Arntz, Levri, & Frey, 2005). Lastly, a quantitative trait locus (QTL) for flower number was found to overlap with a QTL for SLA, indicating possible linkage or pleiotropy between the traits (Delph, Arntz, Scotti-Saintagne, & Scotti, 2010).
In general, a positive genetic correlation between SLA and flower number appears to be pervasive and stable, and may constrain the evolution of each trait toward its optimum value. In other words, there may be environments in which making relatively many flowers and having thick leaves would be optimal for male plants via both fecundity and viability selection, but this combination of traits opposes the observed correlation.
We anticipate that a change in the covariance between flower number and SLA in male S. latifolia is possible through changes in existing genetic variation and can therefore occur in a small number of generations. Covariances are potentially shaped by linkage disequilibrium, pleiotropy, and differential epistasis and are influenced by the relative contributions of these mechanisms as well as the rate of changes in allele frequency and loss of alleles caused by fixation. Loci in linkage disequilibrium can generate a correlation between two traits that is susceptible to rapid change (Falconer & Mackay, 1996).
While selection favors linkage when covariance is adaptive (Sinervo & Svensson, 2002), covariance generated by linkage disequilibrium is expected to decrease over time in the absence of selection. In addition, fixation resulting from selection or drift at one or both loci eliminates the influence of the linkage on covariance, reducing its magnitude even if linkage is not disrupted. Pleiotropy has the potential to generate stable, lasting genetic covariation (Mitchell-Olds, 1996). However, a small number of pleiotropic alleles and/or alleles with variable or opposing pleiotropic effects can reduce the stability of covariation, such that changes in allele frequency and loss of alleles could significantly alter the magnitude of covariation in a small number of generations (Agrawal, Brodie, & Rieseberg, 2001;Bohren et al., 1966;Conner, 2012;Falconer & Mackay, 1996;Gromko, 1995;Houle, 1991). Differential epistasis occurs when the extent of pleiotropy caused by a locus is modified by alleles at a second locus (Cheverud, 2001) and may also contribute to a rapid change in covariance. Moreover, changes at a locus influencing differential epistasis have the potential to modify genetic covariances without changing trait means (Pavlicev et al., 2008). In sum, several genetic mechanisms have the potential to cause a rapid change in genetic covariance, suggesting that shifts away from a constraining correlation are possible under certain conditions and selection regimes.
The time scale over which genetic covariance can change has implications for the ability of organisms to escape developmental or functional constraints and adapt to new environments. To determine whether a genetic covariance could change rapidly while trait means changed little, we conducted an artificial-selection experiment in S. latifolia in which we selected families with combinations of trait values for flower number and SLA that had the potential to reduce the covariance but keep trait means constant.

| Study species
Silene latifolia (Caryophyllaceae) is a short-lived perennial herb native to Europe and widely common as an introduced species in the United States. It is dioecious and sexually dimorphic for many reproductive and vegetative traits (see review in Delph, 2007). It flowers in its first year, begins flowering in mid-spring or early summer, and produces flowers indeterminately on a cyme (Morton, 2005).
Its white, fragrant flowers open at night and are primarily pollinated by moths (Shykoff & Bucheli, 1995). Fruits are capsules containing many small seeds.

| Trait measurements and selection procedures
The base generation for selection was founded from seeds collected in a wild population in Giles County, Virginia. Seeds from 103 capsules were grown in a greenhouse at Indiana University (IU). These plants were used as parents in a breeding design that used a male and a female originating from each capsule and mated each individual to three other individuals to generate 150 full-sib families. Each family had two other families with the same mother and two other families with the same father, generating additional half-sib relationships. For example, a female from family 1 was mated to males from families 2, 3, and 4, and a female from family 2 was mated to males from families 3, 4, and 5 ( Figure 1). See Steven, Delph, and Brodie (2007) for further details of the crossing design. A randomly selected subset of 120 of the resulting full-sib families was used in this study.
We grew all plants in this study in the same environment in pollinator-free greenhouses at IU, and we used a mix of Metromix We measured specific leaf area (SLA) 10 days after a plant opened its first flower on the leaf two nodes down from that flower. If the leaf at the second node was smaller than 3 cm 2 , we used the third leaf down. Leaf area was measured with a portable leaf-area meter Selection to directly alter the correlation was similar to the selection employed in . From the base generation, we established a control line and two replicate selection lines by selecting eight full-sib families for each. The goal of selection was to decrease the covariance between male SLA and male flower number while keeping the trait means the same. In other words, we performed disruptive correlational selection on the two traits.
To achieve this, we identified families with trait mean combinations most distant from the major axis of the correlation using principal components analysis. We used the second principal component value (PC2) to identify families for selection ( Figure 2). In an attempt to maintain trait means, we selected full-sib families equally from both extremes of the second axis; four families with the most positive PC2 values, and four with the most negative. In the first round of selection from the base generation, we attempted to maintain similar selection strength in the two selection lines (Table 1). Families for the control line were selected randomly, and we allowed a family to be used in both the control line and a selection line to avoid bias.
To select parents from each full-sib family for breeding, we randomly selected a female and chose the male with trait values most similar to the family mean. In some cases, selection of the male was determined by availability of flowers and pollen. We performed crosses using the same general design used to create the base generation ( Figure 1). This design generated 24 full-sib families. When performing selection for the next generation, we selected from these full-sib families.
We selected to reduce the correlation between SLA and flower number in males over a total of four generations. In generations 1-3, we grew eight seeds per family under the same conditions as the base generation and followed the same selection and crossing procedures. Sample sizes ranged from 38 to 69 males measured per line.   (Table 2). We also used this package to calculate average relatedness for the full pedigree in each line.

| Statistical analysis
To determine whether selection changed mean trait values, we used a one-way analysis of variance with line as the factor for each trait. We compared means between lines using Tukey's HSD.
We estimated genetic parameters for the base generation and each of the lines in generation four using the multivariate animal model (Kruuk, 2004;Wilson et al., 2010)  To visualize the differences in overall G matrix structure among lines, we generated an ellipse representing the eigenstructure of the matrix for the base generation and the last generation of each line. Using eigenvalues, we calculated major and minor axes scaled to 95% of the variation; these values determine the x and y dimen- We also used random skewers to compare G matrices among lines (Cheverud & Marroig, 2007). The random skewers test compares matrices within the context of the multivariate breeder's equation. Random selection vectors are multiplied by each variancecovariance matrix, and the similarity of the response vectors is determined by the vector correlation between them. If two matrices result in similar responses to the same selection vectors, the correlation is high. To determine statistical significance, this correlation is compared against a null distribution of vector correlations between vectors selected randomly from a uniform distribution. Therefore, a p-value <.05 indicates greater matrix similarity than expected by chance. We used the R package phytools 0.5-20 (Revell, 2012) to conduct pairwise comparisons between the genetic variance-covariance matrices for the base generation and the control and selection lines.

| RE SULTS
After four generations of selection, the genetic covariance in selection line 1 was negative and significantly <0 (χ 2 = 4.58, p = .032; Table 3). The eigenstructure of the G matrix in selection line 1 also deviated significantly from the base generation and the control line ( Figure 3, Table 4). However, the eigenstructure of the G matrix in selection line 2 was similar to the control line, and the genetic covariance was not significantly different from 0 (χ 2 = 0.819, p = .37). In both selection line 2 and the control line, genetic variation increased somewhat (Table 3), and the genetic covariance in the control line was also not significantly different from 0 (χ 2 = 2.18, p = .140).
The genetic correlations calculated from the variance and covariance estimates reflect the changes in both variance and covariance. In the control line and selection line 2, where variance increased for both traits, the genetic correlation between traits was slightly smaller than the genetic correlation in the base generation (Table 3). In selection line 1, the reduction in genetic variance in both traits and the reversal in sign of the genetic covariance resulted in a highly negative genetic correlation between traits. Phenotypic correlations were generally small and not similar to the underlying genetic correlations (Table 3).

TA B L E 3
Additive genetic variance and covariance for flower number (logtransformed and multiplied by 50) and specific leaf area (cm 2 /g) in males for plants in the base generation, the control line after four generations of random mating, and selection line 1 and selection line 2 after four generations of selection to reduce the correlation between the traits line 2 (Table 2). Selection line 2 also showed an appreciable increase in additive genetic variance for SLA ( Table 2).
The differing responses in the two selection lines were also evident in the eigenstructure of the G matrix after selection. The eigenstructure of selection line 1 diverged significantly from the eigenstructure of the base generation and the control line, reflecting a significant change in genetic variation and covariation in this line (Table 4). The rotation of the matrix increased above 90 degrees, indicating a negative correlation between flower number and specific leaf area. In addition, the magnitude of variation along the minor axis decreased considerably (Figure 3). However, selection line 2 did not deviate significantly in eigenstructure from either the base generation or control line (Figure 3, Table 4).
Matrix comparison by random skewers also revealed that the G matrix of selection line 1 was altered significantly compared with that of the base generation and the control line (Table 5). The G matrix of selection line 2 differed significantly from that of the control, although correlations among all matrices were generally high.

F I G U R E 3
Ellipses representing the eigenstructure of the G matrix in the base generation, the control line after four generations of random mating, and selection line 1 and selection line 2 after four generations of selection to reduce the covariance between the traits. The magnitude of the major and minor axes correspond to eigenvalues, and the angle of the matrix represents matrix rotation, as determined by eigenvectors

Note:
To determine whether the lines deviated from the base generation, and whether the selection lines differed significantly from the control line, we conducted randomization tests that generated a null distribution of 1,000 iterations from a combined dataset, and p-values were calculated from the number of iterations that were more extreme than the observed values. by the covariance that appears in the numerator. This phenomenon has potentially produced a very tight genetic correlation between the two traits but reduced genetic variance for the traits in general.

TA B L E 4 Comparison of G matrix
While both the magnitude and orientation of the G matrix in selection line 1 was altered by our selection regime, the eigenstructure of the G matrix was generally similar across the base generation, the control line, and the second selection line. In addition, the random skewers test provided evidence that the overall effect of the reduced covariance on the outcome of selection is surprisingly modest. When multiplied by random selection vectors, the G matrix for selection line 1 generated a response to selection correlated with the response in the other selection line and the control line, indicating that the matrices retain some similarity after selection. However, the greatest difference in matrices was between selection line 1 and the control, reflecting the significant change in covariance in this line.
The magnitude and rapidity of the change in the genetic covariance in selection line 1 suggest that one or a few loci of major effect may be driving the switch in sign. Other studies have demonstrated that single alleles with pleiotropic effects can significantly alter a correlation over a few generations (Agrawal et al., 2001). Segregating alleles at a single locus in Arabidopsis thaliana had effects on the value of a genetic correlation (Stinchcombe, Weinig, Heath, Brock, & Schmitt, 2009), and an increased mutation rate significantly altered the covariance structure of size and reproductive investment traits in A. thaliana, while trait means did not change (Camara & Pigliucci, 1999). In addition, strong selection for an insecticide resistance allele in a leafroller led to changes in diapause and larval weight, suggesting a constraining pleiotropic effect (Carriere & Roff, 1995). Note: Values are vector correlations between response vectors for the two matrices; higher correlations indicate greater similarity in responses to random selection vectors. The asterisk indicates the correlation that showed significant similarity between the two matrices.

TA B L E 5
Comparison of genetic variance-covariance matrices by random skewers reduces effective population size. Therefore, we chose a population size of 12, which assumes 50% of the matings among the 16 individuals were between first cousins. We generated a total of 1,000 correlations. The average correlation was 0.84, which is larger than the genetic correlation of 0.62 estimated using the animal model and reflects the anticonservative nature of BLUPs. In the simulation, eighteen correlations were <0.5, and two were <0.2 ( Figure 6).
These findings suggest that the initial selection of individuals from the base generation had the potential to establish lines with very small initial genetic covariance between the traits and highlight the importance of stochastic factors in changes in the G matrix.
We  and making fewer flowers may be under relatively strong viability selection, such that any selection via fertility on flower number is overwhelmed. Random skewers analysis found that the G matrix for the base generation and selection line 2 were similar, but the base generation and the control line were not, suggesting that selection may promote the structure of the matrix found in the base generation. Variability in selection pressures across a season or among years may help to maintain alleles that generate a negative or zero covariance at low frequencies by occasionally favoring these alleles (Houle, 1991).
Restructuring a covariance in a relatively small number of generations has implications both for escape from constraint and for the development of adaptive covariance between traits. Although the lability of a covariance is dependent on its underlying genetic structure, the likelihood of change depends upon optimum phenotype combinations (Arnold, 1992;Brodie, 1992). If the multivariate adaptive landscape is relatively static, the structure of the covariance may converge with the structure of the adaptive landscape even when change in covariance is slow (Arnold et al., 2008). If the covariance is capable of change that is contemporaneous with changing trait optima, the interplay between covariation and correlational selection could be a key aspect of the dynamics of adaptation even over short time scales.
Selection could potentially create genetic covariance that enhances trait integration and adaptive evolution, as observed in some studies of multivariate selection (Agrawal & Stinchcombe, 2009).

| CON CLUS ION
Our findings illustrate the potential instability of a genetic covariance when exposed to novel selection regimes and genetic drift and suggest that variation in the underlying genetic architecture may contribute to the lability of a covariance. Changes in covariance caused by changes in allele frequency and fixation of alleles may be especially important for covariances strongly influenced by few loci of major effect. The independent evolution of correlated traits is potentially constrained not by the covariance itself, but by the combination of covariance and patterns of selection exerted by the environment.

ACK N OWLED G EM ENTS
This work was supported by a grant from the U.S. National Science Foundation (DEB-0210971) to LFD and EDBIII. Discussions with T.
Linksvayer were especially helpful in guiding our interpretation of the results. The authors thank J. Conner and A. McAdam for comments that have improved the manuscript.

CO N FLI C T O F I NTE R E S T
The authors declare that they have no conflict of interest.

AUTH O R CO NTR I B UTI O N S
LFD and EDBIII conceived of the research question and designed the experiment; JCS and IAA collected and analyzed data. JCS drafted the manuscript and figures, which were then reviewed and revised by LFD, EDBIII, and JCS.

DATA AVA I L A B I L I T Y S TAT E M E N T
Data files are archived in the Dryad Digital Repository at https ://doi. org/10.5061/dryad.dbrv1 5dx7.