Pleiotropy of FRIGIDA enhances the potential for multivariate adaptation

An evolutionary response to selection requires genetic variation; however, even if it exists, then the genetic details of the variation can constrain adaptation. In the simplest case, unlinked loci and uncorrelated phenotypes respond directly to multivariate selection and permit unrestricted paths to adaptive peaks. By contrast, ‘antagonistic’ pleiotropic loci may constrain adaptation by affecting variation of many traits and limiting the direction of trait correlations to vectors that are not favoured by selection. However, certain pleiotropic configurations may improve the conditions for adaptive evolution. Here, we present evidence that the Arabidopsis thaliana gene FRI (FRIGIDA) exhibits ‘adaptive’ pleiotropy, producing trait correlations along an axis that results in two adaptive strategies. Derived, low expression FRI alleles confer a ‘drought escape’ strategy owing to fast growth, low water use efficiency and early flowering. By contrast, a dehydration avoidance strategy is conferred by the ancestral phenotype of late flowering, slow growth and efficient water use during photosynthesis. The dehydration avoidant phenotype was recovered when genotypes with null FRI alleles were transformed with functional alleles. Our findings indicate that the well-documented effects of FRI on phenology result from differences in physiology, not only a simple developmental switch.


Introduction
Populations of a species are frequently distributed across climatic gradients, where natural selection can lead to adaptation to local conditions. The environmental conditions that cause local adaptation have been well documented through reciprocal transplants and studies of clines [1][2][3][4][5][6]. These experiments show that divergent patterns of selection cause shifts in the mean values of many traits leading to a multivariate response. Such a response to selection improves fitness and promotes successful adaptation to local conditions. Despite a large body of research, it remains a challenge to determine the specific genetic loci that respond to selection and confer local adaptation [7][8][9][10].
Long-term breeding programmes and quantitative genetic studies have demonstrated variation in nearly all traits, and thus a simple lack of additive genetic variation is not expected to constrain adaptation [11]. Instead, a limited amount of genetic variation along vectors of selection has been shown to limit adaptive evolution [12][13][14]. Theoretically, independence of all loci and phenotypes will improve the potential for adaptation by optimizing evolvability [15] and the response to selection (R) [14,16,17]. However, certain genetic correlations can disrupt the optimal genetic architecture by reducing the amount of genetic variation which is available to selection and causing correlated responses of non-adaptive traits [12,18]. Although this maladaptive role for genetic correlations is not universal [19], genetic correlations may affect R by limiting the dimensionality of the genetic (co)variance matrix or restricting genetic variation to vectors which are not aligned with selection [12][13]18,20].
The combined effects of the pleiotropic loci, which cause genetic correlations, may have a profound impact on patterns of local adaptation. As pleiotropy can constrain multivariate adaptation and cause correlated evolution of adaptive and deleterious phenotypic values, these loci are typically considered 'antagonistic' [21][22][23][24][25]. Adaptation is especially constrained when pleiotropic gene action limits phenotypic correlations along a vector orthogonal to that of selection and reduces R [13]. Antagonistic pleiotropy is well documented and has led to the belief that all pleiotropy is maladaptive [26]. However, recent theoretical work has countered this viewpoint by demonstrating that intermediate levels of pleiotropy may actually improve the conditions for adaptation and evolution of complexity [27][28][29].
To study the adaptive value of pleiotropic loci, it is necessary to assess the effects of genetic variation on the structure of many phenotypes which are subject to correlational selection in nature. Adaptation to drought in plants provides an ideal system to achieve this goal [30 -33]. In natural and agricultural systems, annual plants can be adapted to local drought conditions by either growing and reproducing before the onset of drought (drought escape) [31,34 -36] or by delaying reproduction, increasing water use efficiency (WUE) and conserving resources (dehydration avoidance) [37][38][39]. For example, accessions which exhibit early-flowering time (FT) and low WUE were selected for in consistently wet soil and late-season drought conditions [36,40], whereas direct selection on increased WUE favoured a dehydration avoidance strategy in environments with earlyseason drought [41]. Therefore, adaptation to different local soil moisture conditions and seasonal rainfall patterns contributes to the observed strong correlations between FT, growth rate and WUE within and among species [6,32,38,[41][42][43]. Several studies have suggested that pleiotropy may also affect this correlation [38,44,45].
Here, we provide empirical evidence for an adaptive role of pleiotropy. Using genome-wide approaches, allelic variants and transgenic manipulation, we demonstrate that the 'FT' gene, FRIGIDA (FRI) pleiotropically affects phenotypic variation in growth rate, WUE and FT. Derived, null FRI alleles produce a drought escape phenotype (decreased WUE, increased growth rate, decreased FT) relative to the ancestral adaptive strategy. This phenomenon, which we term 'adaptive pleiotropy', enhances the likelihood of adaptation by increasing adaptive responses to selection.
The FRI-NIL was generated by introgressing a functional FRI allele from the Sf-2 line into wild-type (WT) Col-0, the reference A. thaliana accession with a null FRI allele [47 -49]. The tr-FRI transgenic over expressed line was generated by ligating FRI-GFP into the XmaI and XhoI sites of 35SpBARN vector and then transformed a into Col-0 background using the floral dip method. We used only FRI transgenic lines that exhibited a late flowering phenotype. The FRI-NIL and the transgenic line 'FRI-GFP Col T2 #20' are available from S.D.M. and X.Y. We also present WUE data from FRI-NIL and Columbia genotypes with knocked-out FLC alleles. See Michaels & Amasino [50] for details on these lines. Gas exchange physiology was measured with an LI-6400 photosynthesis system (LiCor Inc, Lincoln, NB) equipped with a custom whole-plant gas exchange cuvette. A total of 20 measurements were taken over a 2 min period for each of 20 plants (10 replicates/genotype) at two time points (14 and 21 days post-germination). The photosynthetic parameters (A, c i and g s ) were estimated following von Caemmerer & Farquhar [51]. Gas exchange data were analysed in a mixed-model framework where genotype was fixed, and measurement and date were nested within individual as a random effect in JMP GENOMICS v. 5.0 (SAS Institute, Cary, NC). We also generated A/c i curves by measuring photosynthetic rate across nine levels of external CO 2 concentrations using a different set of plants grown hydroponically. We compared A between the FRI-NIL and Col-0 controlling for variation in c i with a mixed effect ANOVA. The genotype was a fixed effect, and c i was a continuous, random covariate.
We measured WUE, growth rate and FT for each plant (n/genotype ¼ 10). Flowering initiation was recorded when a visible bolting structure first appeared at the apical meristems; FT is calculated as the number of days between germination and initiation of flowering. We analysed carbon isotope composition (d 13 C), a surrogate measure of WUE [38,52], on lyophilized, finely ground rosette leaves at the Stable Isotope Facility at University of California, Davis (UCD; http://stableisotopefacility. rspb.royalsocietypublishing.org Proc R Soc B 280: 20131043 ucdavis.edu/). Leaves were harvested before the onset of flowering of the earliest accession at a single time-point for all lines. Using images taken directly over the rosette, we assessed leaf area in IMAGEJ (http://rsbweb.nih.gov/ij/). These data were used to calculate relative growth rate of leaf area (GR la ¼ [ln(LA t2 )ln(LA t1 )]/(t 2 2 t 1 )), where LA is leaf area at time 1 (t 1 ) and time 2 (t 2 ). We analysed the effect of genotype on WUE, FT and GR la via one-way ANOVA in JMP GENOMICS v. 5.0.

(c) Quantitative trait locus analysis
We analysed quantitative trait loci (QTL) for WUE and FT in R/qtl [53] using the following settings: (i) imputations (256 draws) to generate a complete and even genome-wide pseudomarker grid of 2 cM for mapping, (ii) 10 000 permutations to calculate QTL incorporation thresholds at an experiment wise a ¼ 0.05, (iii) stepwise model selection scanning for epistatic and additive QTL at each step [54], (iv) iterative position refinement analysis by holding all but one QTL constant and varying the position of the focal QTL and re-calculating the penalized logarithm (base 10) of the odds (LOD) score for the model and (v) fitting the refined model via ANOVA to calculate the effect size, per cent variance explained and LOD score for each QTL. The allelic effect at FRI was compared via one-way ANOVA in JMP GENOMICS v. 5.0. To further refine the bi-phenotype QTL position, we standardized the LOD scores by the largest value for each phenotype (stand. LOD) and summed the bivariate scores for each point on the genotype grid, then calculated the bivariate QTL interval as the point where the summed LOD scores decreased to the average single phenotype odds ratio at a given map position.

(d) Quantification of the FRI-NIL Sf-2 introgression
Whole genome sequence was obtained by paired-end Illumina sequencing at the UCD Genome Center (http://www.genome center.ucdavis.edu/). A reference-based assembly of the TAIR 9 Columbia genome was conducted in SHORE (http://1001 genomes. org/software/shore.html) to call single nucleotide polymorphisms (SNPs) [55] and identify the size of the introgression.

(e) Gene expression analysis
Genome-wide gene expression was determined via Affymetrix (Affymetrix Inc. Santa Clara, CA) AthSNPtile arrays for all TK RILs. We screened for all genes within 50 kb of the QTL point estimate (CH 4 237 060-337 060 bp) and compared expression levels between TSU-1 and KAS-1 alleles at each gene, then corrected for multiple comparisons via q-value calculations (R package qvalue http://www.bioconductor.org/packages/release/bioc/ html/qvalue.html).
(f ) Analysis of population structure at FRIGIDA We conducted three separate population genetic analyses using the publicly available genome-wide SNP data (http://cynin. gmi.oeaw.ac.at/home/resources/atpolydb) [56,57]. We imputed the functionality of FRI for all lines by extracting all SNPs within 100 kb of FRI and training a classification model, support vector machines (SVMs) with a radial basis function [58], using data on SNPs and FRI functionality for the 317 genotypes in our panel that had both SNP and FRI data. After a grid search of tuning parameter values, our final SVM model predicted FRI functionality with 95 per cent accuracy in fourfold cross-validation. We tested the accuracy of the SVM model using n-fold crossvalidation: after selecting n accessions at random, we tested the accuracy of SVM models in n-fold cross-validation (i.e. leaveone-out cross-validation) for n ¼ 10, 15, 20, 25, 30, 35 and 40. For each value of n, we cross-validated the SVM predictions for 20 random subsets of accessions. For n ¼ 10, SVM models were on average 87 per cent accurate in cross-validation. By n ¼ 20, models were 93 per cent accurate in cross-validation. This signifies that the SNP associations with FRI functionality are easily observable in even small samples of accessions. Using this model, we then imputed the allelic state (binned into functional or null categories) for all accessions in the SNP database.
We calculated genome-wide F ST in PLINK [59] by classifying the accessions as 'functional' or 'non-functional' FRI and calculating the molecular variance between and within these allele classes. We generated 5000 random divisions at the same frequency as the FRI allele classes. These permutations allow us to assess the significance of the F ST measure compared with random evolution. We conducted two additional analyses with subsets of the available accessions. Ten of the 574 sites sampled by Horton et al [57] showed within-population variation at FRI. Using these populations and geographical clusters at the country level [57], we calculated an average heterozygosity over SNPs sampled at 50 kb intervals (H t ). Then, we split the population based on FRI phenotype, calculated an average heterozygosity within each subpopulation (H s ) in the same way and took the mean of those. We used these H t and H s values to calculate genome-wide F ST based on the FRI phenotype. We bootstrapped to calculate significance by dividing the data at a random subset of 5000 SNPs with similar frequency to FRI and recalculating F ST .
(g) Comparison of climatic variables associated with FRI variation FRI functionality calls, latitude and longitude for each line were input into DIVA-GIS (www.diva-gis.org). The 19 BIOCLIM (www.bioclim.org) climatic variables were extracted for each point. FRI allelic association with these variables was made via t-tests with significance corrected for multiple comparisons by Bonferroni adjustments. The distribution of the climate under each allele was compared by ranking the climate variables and plotting the relative position of each allele relative to its rank.

Results and discussion
(a) Mapping the water use efficiency -flowering time correlation We measured FT and WUE of 195 A. thaliana accessions in a common garden. The genetic correlation between WUE and FT is positive and significant: WUE explains nearly 40 per cent of FT variation (n ¼ 195, r 2 ¼ 0.395, p , 0.0001; electronic supplementary material, figure S1). If this correlation results from many loci independently affecting each phenotype, then recombination between differently adapted lines will break down this favourable correlation. To test the cause of the WUE-FT correlation, we used TK RILs from two phenotypically divergent accessions, TSU-1 (low WUE, short FT) and KAS-1 (high WUE, long FT) [46] (see the electronic supplementary material, figure S1). Experimental crosses induce recombination and break up linkage disequilibrium across these genomes. Despite a large reduction in linkage disequilibrium, FT and WUE remained significantly correlated (n ¼ 304, r 2 ¼ 0.138, p , 0.0001; figure 1a)  , p , 0.0001) than TSU-1 alleles (figure 1c). The genetic correlation between WUE and FT is well documented in agricultural breeding populations and studies of local adaptation in nature [38,[44][45]. High WUE decreases photosynthetic assimilation rates and the amount of fixed carbon available for flowering. As the initiation of flowering is affected, in part, by resource availability [60,61], a physiological connection between WUE and FT is plausible. FRI is a particularly good candidate gene underlying the FT QTL. Derived mutations that reduce expression have been involved in the evolution of spring annual types from the ancestral state of a fully functional FRI and a winter annual life history [47,62]; allelic variation at FRI contributes to variation in FT across diverse accessions [49,[63][64][65]. FRI is also a candidate for WUE [38,66]. Biogeographic analyses have associated lines with functional FRI alleles, such as KAS-1, to regions with lower precipitation; these environments would favour drought adaptation via dehydration avoidance [6,67,68].
To further assess the pleiotropic effects of the FRI locus, we also genotyped 195 A. thaliana accessions at FRI to determine functionality (see the electronic supplementary material, table S1). Consistent with pleiotropy and our observation in the TK RILs  allele (FRI-NIL) to the wild-type progenitor Col-0 (WT) which contains a null fri allele. We phenotyped the three major physiological determinants of WUE: photosynthetic rate (A), leaf internal CO 2 concentration (c i ) and stomatal conductance (g s ). Stomatal conductance, which directly alters leaf water-loss dynamics, also affects A by regulating the supply of CO 2 and thus, c i . For example, low g s reduces A by limiting c i , resulting in increased WUE, decreased growth rate and delayed FT [43] (figure 3a  3c). To determine the physiological mechanism for increased WUE, we modulated c i and repeatedly measured A. Supporting a decrease in g s as the basis for increased WUE, no significant difference in photosynthetic capacity was found between the WT and FRI-NIL while controlling for c i (d.f. ¼ 288, F 1.701 , p ¼ 0.1932; electronic supplementary material, figure S2).
Many functional analyses have found that FRI produces a transcription factor that induces expression of FLC, inhibiting floral development [50,62,63,65,69,70]. To place our analyses in the context of these results, we analysed WUE for WT and FRI-NIL lines which have knocked-out FLC alleles.
Consistent with the epistasis observed to affect FT, FRI confers increased WUE only in the presence of a functional FLC (contrast d.f. ¼ 1, F 44.77 , p , 0.0001), but not the when associated with a null flc allele (contrast d.f. ¼ 1, F 0.79 , p ¼ 0.38; electronic supplementary material, figure S3).
The FRI-NIL (also referred to as 'Sf-2 FRI in Col' or 'Col-FRI') is one of the most used genetic resources in the FT literature [38,49,71]. These studies assume that the FRI-NIL carries a single, narrow, introgression of the Sf-2 genome which contains a functional FRI allele; however, this assumption has never been tested. To assess the size of the Sf-2 introgression, we re-sequenced the FRI-NIL, aligned the reads to the TAIR 9 Columbia genome, called SNPs and mapped SNP density to the reference genome. Many SNPs exist between Sf-2 and Col-0 (data are publically available at http://mus.well.ox.ac.uk/ 19genomes [72]). High SNP density between the FRI-NIL and Col-0 exists solely on proximate Chr. 4 (see the electronic supplementary material, figure S4a,b). The region of elevated SNP density represents a single 1.070 Mb (+10 kb) Sf-2 introgression that contains FRI as well as the other 325 gene models between AT4G00005 and AT4G02710. Although most studies that use the FRI-NIL assume the only genotypic divergence exists at FRI, this is obviously not the case.
To unambiguously determine whether the effects observed in the FRI-NIL were due to FRI, we compared WUE, FT and GR la between WT Col-0 and transgenic lines (Col-0 overexpressing FRI: tr-FRI). Under well-watered conditions, tr-FRI had greater WUE (d.f. ¼ 1, F 57.25 , p , 0.0001), decreased GR la (d.f. ¼ 1, F 22.32 , p , 0.0001) and later FT (d.f. ¼ 1, F 179.1 , p , 0.0001) than WT (figure 3d). As FRI functionality is the only DNA sequence difference between these lines, FRI is pleiotropic and controls covariation of three traits along a vector shown to be adaptive. Our conclusion is supported by QTL, natural accession, NIL and transgenic comparisons.

(d) The population genetics of adaptive pleiotropy
Population genetic models are at odds about the role of pleiotropy in maintaining variation within and among populations. Pleiotropic gene action may cause non-adaptive and adaptive phenotypes to covary, thus reducing the efficacy of correlational selection and permitting the persistence of multiple allelic states within populations [73]. However, where the effects of pleiotropy are more aligned with the direction of selection, within-population variation can be purged by strong directional selection [74]. Therefore, we predicted low levels of within-population variation at FRI, because multivariate selection would favour either a functional (drier habitats) or non-functional allele (wetter habitats). In addition, if variation at FRI can lead to local adaptation, then we predicted increased population structure (across the entire genome) between functional and non-functional FRI classes.
A population genetic test for adaptive pleiotropy is complicated in our study as FRI may cause population structure through both adaptive pleiotropy and allochrony: FRI-NILs and tr-FRI lines flowered at least 28 and 32 days later than Col-0, respectively. All main-raceme Col-0 flowers had been pollinated and produced fruits before any FRI-NIL or tr-FRI lines produced open flowers. In the greenhouse environment, single mutations at FRI can produce a reproductive isolation index near 1.0. However, assortative mating owing to variation at FRI may be tempered in nature as the environment has a profound effect on phenology [71]. To test for evidence of reproductive isolation between accessions and populations that differ at FRI, we first imputed FRI functionality of 1188 accessions [57], then compared the group of individuals with derived, weak alleles (i.e. null Col-0 missense and Ler deletion alleles) to the group of individuals with functional, ancestral-type FRI. We then calculated F ST between FRI allele functional classes in PLINK [59]. F ST values averaged across 216 130 SNPs are significantly greater between the FRI functionality classes than is expected from genomewide subsampling ( p , 0.0001; electronic supplementary material, figure S5). To control for geographical population structure, we divided the global sample into 11 geographical regions according to Horton et al. [57]. Ten of 11 geographical regions showed elevated F ST at FRI compared with a genome-wide sample of sites with the same allele frequencies as the FRI functional variants (see the electronic supplementary material, table S5). These results show that elevated global F ST when sorting by FRI is due to a lack of within-population variation in FRI. Less than 2 per cent of 574 local populations harboured functional variants at FRI.
While extremely low within-population variation is present at FRI, functionally divergent alleles have gone to fixation in geographically proximate populations. Several authors have shown that an abundance of derived null FRI alleles are present in nature, far more than would be expected by chance [75,76]. Here, we demonstrate that these mutations cause a phenotypic leap between drought adaptation strategies which may promote adaptation to novel ecological conditions. Combined, the strong signature of selection, high levels of population structure and lack of within-population variation observed at FRI suggests an adaptive role of this pleiotropy.

(e) FRI and drought adaptation
Previous studies have found that the early flowering, low WUE phenotypes associated with drought escape are adaptive in sites without consistent low soil moisture [36]. Although we did not directly measure selection in this study, we used the large body of work on drought adaptation to infer the adaptive value of specific trait combinations. We predicted that due to the drought escape strategy conferred by derived loss of function mutations at FRI, accessions with these alleles would inhabit environments with consistently wetter growing seasons, relative to accessions with functional FRI alleles. To confirm the allelic association with drought, we generated a climate envelope for both FRI allele classes (see the electronic supplementary material, table S5). Functional alleles tend to be present in areas with lower growing season precipitation than non-functional alleles (t ¼ 23.68, p ¼ 0.0003; electronic supplementary material, figure S6).
We have demonstrated that lines that diverged only at FRI exhibit altered positions along an adaptive phenotypic correlation. Scarcelli et al. [25] found antagonism between the floral morphology traits affected by FRI, and we cannot rule out that a portion of FRI's pleiotropic gene action is maladaptive. However, analyses presented here demonstrate a strong adaptive role of the physiological and phenological phenotypic correlations conferred by FRI. Given our results, it is not surprising that FRI is associated with strong population genetic signatures of diversifying selection [65,75,76]. Studies demonstrating historical selection on FRI invoke the timing of flowering as the phenotype under selection [65].
Our results indicate that the observed signature of selection is not only an effect of FT variation, but may also be due to upstream physiological effects.

Conclusions
We have presented a mechanistic understanding of how FRI alters physiology, phenology and confers local adaptation. Phenology, growth rate and water-use physiology have been mapped to similar genomic loci or correlated in natural or experimental populations [31,38,46,52,66]. Here, we have demonstrated that FRI causes these adaptive correlations to be heritable. Although we present a situation where pleiotropy controls phenotypic variation along a vector known to be adaptive, we have not measured the efficacy of or response to selection in the field. Fitness measures in diverse common gardens with watering treatments would allow for direct inference of the adaptive value of FRI.
To date, most gene annotation and characterization is conducted by forward or reverse genetics whereby a single gene or trait is under consideration. Our results indicate that a more holistic approach to phenotyping and whole plant, integrative approaches for annotating gene function rspb.royalsocietypublishing.org Proc R Soc B 280: 20131043 may reveal complex patterns of pleiotropy among ecologically correlated phenotypes. It is possible that many trait associations are not purely a product of correlational selection, but also affected by adaptive pleiotropy.