Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
This article is a Plant Cell Advance Online Publication. The date of its first appearance online is the official date of publication. The article has been edited and the authors have corrected proofs, but minor changes could be made before the final version is published. Posting this version online reduces the time to publication by several weeks. Hierarchical Nuclear and Cytoplasmic Genetic Architectures for Plant Growth and Defense within Arabidopsis CW Bindu Joseph,a Jason A. Corwin,a Tobias Züst,b,1 Baohua Li,a Majid Iravani,c Gabriela Schaepman-Strub,b Lindsay A. Turnbull,b,2 and Daniel J. Kliebensteina,3,4 a Department of Plant Sciences, University of California at Davis, Davis, California 95616 of Evolutionary Biology and Environmental Studies, University of Zürich, Zurich CH-8057, Switzerland c Department of Natural Resources, Isfahan University of Technology, 83111-84156 Isfahan, Iran b Institute To understand how genetic architecture translates between phenotypic levels, we mapped the genetic architecture of growth and defense within the Arabidopsis thaliana Kas 3 Tsu recombinant inbred line population. We measured plant growth using traditional size measurements and size-corrected growth rates. This population contains genetic variation in both the nuclear and cytoplasmic genomes, allowing us to separate their contributions. The cytoplasmic genome regulated a significant variance in growth but not defense, which was due to cytonuclear epistasis. Furthermore, growth adhered to an infinitesimal model of genetic architecture, while defense metabolism was more of a moderate-effect model. We found a lack of concordance between quantitative trait loci (QTL) regulating defense and those regulating growth. Given the published evidence proving the link between glucosinolates and growth, this is likely a false negative result caused by the limited population size. This size limitation creates an inability to test the entire potential genetic landscape possible between these two parents. We uncovered a significant effect of glucosinolates on growth once we accounted for allelic differences in growth QTLs. Therefore, other growth QTLs can mask the effects of defense upon growth. Investigating direct links across phenotypic hierarchies is fraught with difficulty; we identify issues complicating this analysis. INTRODUCTION A key interest of quantitative genetics and systems biology is to causally link genetic polymorphisms to systems of interacting phenotypes at different levels of biological organization, from individual molecules to the whole organism (Mackay, 2001). One such hierarchical system of particular interest to ecologists and biochemists is the network of links between genetic polymorphisms and changes in chemical defense compounds, a low-order phenotype, with that of growth and fitness, higher order phenotypes (Karban and Baldwin, 1997; Strauss and Agrawal, 1999; Agrawal, 2007; Lankau, 2007). Similar hierarchical systems exist in agronomics and crop breeding, where there is an interest in linking enzyme polymorphisms to nitrogen use efficiency and plant yield (Loudet et al., 2003, 2005). However, the interpretation of such hierarchical systems is usually confounded by their inherent complexity, as the number of 1 Current address: Department of Ecology and Evolutionary Biology, Cornell University, Ithaca, NY 14853. 2 Current address: Department of Plant Sciences, University of Oxford, South Parks Road, Oxford OX1 3RB, United Kingdom. 3 Current address: DynaMo Center of Excellence, University of Copenhagen, Thorvaldsensvej 40, DK-1871 Frederiksberg C, Denmark. 4 Address correspondence to [email protected]. The author responsible for distribution of materials integral to the findings presented in this article in accordance with the policy described in the Instructions for Authors (www.plantcell.org) is: Daniel J. Kliebenstein ([email protected]). C Some figures in this article are displayed in color online but in black and white in the print edition. W Online version contains Web-only data. www.plantcell.org/cgi/doi/10.1105/tpc.113.112615 components that causally affect a phenotype increases strongly with hierarchical level (Fu et al., 2009). For example, linking a promoter polymorphism to variation in the resulting transcript is relatively straightforward because there are a limited number of molecular factors that will influence the abundance of that transcript. By contrast, linking the same promoter polymorphism to a change in yield is dramatically more difficult since it is possible that nearly every aspect of a genome and environment may affect the yield of a plant. This potential for increasing complexity accompanying a shift from molecular to wholeorganism phenotypes can lead to a change in genetic architecture, from the molecular phenotypes having a small number of large-effect loci with moderate epistasis, to the wholeorganism phenotypes having a large number of loci with moderate to small effects and apparent additivity (West et al., 2007; Rowe et al., 2008; Buckler et al., 2009). The idea of different genetic architectures across the phenotypic hierarchy is long established but also linked to significant disagreement that we propose can be resolved by partitioning causality across the phenotypic hierarchy (Fisher, 1930; Wright, 1931). We demonstrate this using the two causally linked phenotypes of growth and chemical defense (Paul-Victor et al., 2010; Kerwin et al., 2011; Züst et al., 2011). Since these phenotypes are at different levels in the same hierarchical genetic system with chemical defense subordinate to growth, they are ideal candidates to empirically test the relation between hierarchical level and genetic architecture. We then explore if these differences in genetic architecture could actually mask the causal link between the two phenotypes. To accomplish this, we chose to compare the architecture of growth and chemical defense in an Arabidopsis thaliana recombinant inbred line (RIL) The Plant Cell Preview, www.aspb.org ã 2013 American Society of Plant Biologists. All rights reserved. 1 of 17 2 of 17 The Plant Cell population by measuring the same plants for both phenotypes, thereby minimizing any confounding effects from environmental variation. Key components of the defense chemistry of Arabidopsis are the glucosinolates, which provide both effective antiherbivore and antimicrobial properties (Lambrix et al., 2001; Kliebenstein et al., 2002; Zhang et al., 2006; Hansen et al., 2008; Clay et al., 2009). The indole glucosinolates play a predominant role in providing defense against aphids, nonhost pathogens, and oviposition by specialized lepidopteran moths (Cui et al., 2002; de Vos et al., 2008; Bednarek et al., 2009; Pfalz et al., 2009; Müller et al., 2010). The aliphatic glucosinolates are the dominant mechanism providing resistance to chewing insects and some adapted pathogens (Lambrix et al., 2001; Kliebenstein et al., 2002; Beekwilder et al., 2008; Hansen et al., 2008; Fan et al., 2011). Importantly for our study, there is significant genetic variation in glucosinolates regulated by natural knockout polymorphisms in enzymatic or regulatory loci (Kliebenstein et al., 2001a; Hansen et al., 2007; Wentzell et al., 2007; Hansen et al., 2008; Chan et al., 2011). This variation in chemical defense has been linked to fitness in the presence of insect herbivores at both the local and continental scale (Mauricio, 1998; BidartBouzat and Kliebenstein, 2008; Züst et al., 2012). In addition, the recreation of some of the natural knockout polymorphisms responsible for this variation and directly testing their influence on higher whole organism phenotypes has shown that natural glucosinolate variation is responsible for at least part of the changes in growth and yield (Mauricio, 1998; Lankau and Strauss, 2008; Paul-Victor et al., 2010; Kerwin et al., 2011; Züst et al., 2011), which is partly due to changes in the circadian clock (Kerwin et al., 2011). Thus, there is ample evidence for the existence of molecular links between chemical defense and growth phenotypes. For this study, we treated growth as whole-organism phenotype as it is considered to be a highly integrative phenotype with minimal epistasis (Meyer et al., 2007; Lisec et al., 2008; Paul-Victor et al., 2010; Züst et al., 2011; Zhang et al., 2012). There are multiple approaches to estimating growth rates, such as size at a given time, the relative growth rate over a given period, or size-corrected growth rates (Paul-Victor et al., 2010; Züst et al., 2011; Paine et al., 2012). In Arabidopsis, rosette size can be readily measured using real-time digital image analysis, and repeated measures of size at different time intervals can be used to calculate growth rates. By measuring all growth descriptors and chemical defense phenotypes in the same experiment on the same individuals, we are able to reduce the amount of environmental variation that may differentially influence both phenotypes and thus be able to directly compare the variation in genetic architecture of growth and defense. When considering the comparative genetic architecture of growth and defense, nuclear genomic variation is only one part of the genetic variation present in any possible cross between two genotypes. Recently, it has been shown that there is considerable genetic diversity in the cytosolic genomes of Arabidopsis, both plastidic and mitochondrial, that could influence growth or glucosinolate production (Moison et al., 2010; Davila et al., 2011). This is especially true since a fraction of glucosinolate synthesis occurs in the cytosolic organelles and these are also the main powerhouses to provide energy for growth (Sønderby et al., 2010a). Studies in other species have shown that genetic variation in the cytosolic genome can influence the structure of genetic diversity in the nuclear genome and may alter the genetic architecture underpinning fitness-related phenotypes (Wade and Goodnight, 2006; Wolf, 2009). In addition to this potential influence on genetic architecture, cytosolic genetic variation could also affect our ability to identify genetic loci in mapping populations (Moison et al., 2010; Tang et al., 2013). To allow us to measure both the nuclear and cytoplasmic genetic architecture of growth and defense in Arabidopsis, we used the Kas 3 Tsu RIL population, which is one of the largest available Arabidopsis RIL populations to date (McKay et al., 2008; Juenger et al., 2010). The Kas 3 Tsu RIL population was generated from a reciprocal cross with approximately half of the resulting lines carrying Kas organelles, while the other half carry Tsu organelles, which allows for the explicit analysis of the influence of cytosolic genetic variation on the genetic architecture of growth and defense (McKay et al., 2008; Juenger et al., 2010). This simple reciprocal RIL structure provides an increase in statistical power for genetic architecture studies over unstructured or more complex structured populations (Mackay, 2001). In this study, we used high-throughput HPLC analysis to measure glucosinolates (Kliebenstein et al., 2001b; Wentzell et al., 2007) and high-throughput digital image analysis to measure the developmental progression in rosette size of 341 lines in the Kas 3 Tsu RIL population of Arabidopsis. Using multiple measures of growth, biomass, and different defense compounds, we were able to demonstrate that growth and defense do have different quantitative genetic architectures with growth displaying a structure more related to an infinitesimal model, which has a large number of genes of small to moderate effect underlying the phenotype, and the glucosinolate defenses containing more large-effect loci (Paul-Victor et al., 2010; Kerwin et al., 2011; Züst et al., 2011). Both the growth and chemical defense phenotypes showed epistatic networks with similar levels of variance partitioned to the epistatic terms, but this variation was partitioned across the loci such that the chemical defenses had more highly connected networks of genetic loci, while growth had fewer detected connections between genetic loci. Significant cytoplasmic effects were present for all traits, but this was almost entirely due to epistatic interactions with nuclear genetic loci. Interestingly, traditional quantitative trait locus (QTL) overlap or genetic correlation approaches could not identify our mechanistically proven link between glucosinolates and growth in this population. We were able to recover this link using analysis of covariance (ANCOVA) where we accounted for the identified growth-associated loci and showed that, as previously identified, this link was strongest when using the sizecorrected growth rate estimate. Thus, there appear to be genetic linkages present within standard RIL populations that are uncoverable using standard statistical methodologies. While the two phenotypic classes display different genetic architectures, it is more appropriate to consider the chemical defense architecture as a subset of a broader genetic architecture governing growth. Future work will be required to understand how this integration from one phenotype to another occurs at a molecular level. Hierarchical Genetic Architecture RESULTS Population and Phenotype Choice and Phenotypic Heritability To investigate the genetic architecture of plant growth and defense, we measured three distinct phenotypic classes on the same individual plants to minimize any influence of environment. We used high-throughput HPLC to measure two related but biochemically distinct classes of metabolic defense in the same plants, the indolic and aliphatic glucosinolates. In this population, we were able to measure 15 aliphatic glucosinolaterelated phenotypes: 10 individual aliphatic glucosinolates and five composite phenotypes (see Supplemental Data Set 1 and Supplemental Figures 1 and 2 online) (Wentzell et al., 2007; Hageman et al., 2012). The Kas-1 parent for this population has a different glucosinolate profile than the Kas that was previously analyzed for glucosinolate content and is most likely a different genotype, which illustrates the importance of measuring the exact parents of any population (Kliebenstein et al., 2001c; Lambrix et al., 2001). We could measure all four known indole glucosinolates and two composite phenotypes known to specifically identify previously cloned indolic loci, providing six indolic glucosinolate phenotypes (see Supplemental Data Set 1 and Supplemental Figures 1 and 2 online) (Reichelt et al., 2002; Pfalz et al., 2009). Additionally, there were three previously unidentified compounds that copurified with the glucosinolates that were measured but were at levels too low to chemically identify (see Supplemental Data Set 1 online). We measured rosette size throughout plant development six times at regular intervals between days 7 and 23 postgermination and used these data to estimate size-standardized growth rate (SGR), giving us seven growth-related traits (see Supplemental Data Set 1 and Supplemental Figures 1 and 2 online) (Paul-Victor et al., 2010; Züst et al., 2011; Paine et al., 2012; Zhang et al., 2012). For all phenotype classes, we estimated both the nuclear and cytoplasmic heritability using all of the per-line measurements (see Supplemental Data Set 1 online). Cytoplasmic heritability is the amount of variation regulated by the cytoplasmic difference between the two reciprocal crosses where the cytoplasmic genome of specific lines is derived from the Kas-1 or Tsu-1 mother. Nuclear heritability is the phenotypic variation ascribable to differences in nuclear genotypes as defined by each inbred line nested within the cytoplasmic background. Using analysis of variance (ANOVA), nuclear genotypic variation significantly altered all phenotypes, except one indolic phenotype (4OH-I3M, which was removed from further analysis). Nuclear genetic variation also had the strongest influence upon phenotypic variation (Figure 1; see Supplemental Data Set 1 online). Interestingly, cytoplasmic genotypic variation significantly regulated phenotypic variation for SGR, two indolic glucosinolates, and two ratios derived from these indolic glucosinolates using the same ANOVA (see Supplemental Data Set 1 online). However, even for these phenotypes, cytoplasmic heritability was at least two log10 orders less than nuclear heritability (see Supplemental Data Set 1 online). At a preliminary level, this indicates that the nuclear genotypic variation is the primary 3 of 17 genetic player in regulating phenotypic variance within this population for defense metabolism and growth, as has been previously noted when investigating populations containing reciprocal crosses (Lisec et al., 2008). The average nuclear heritability for the phenotypic classes was statistically indistinguishable: aliphatic glucosinolates (38.5% 6 3.8%), indolic glucosinolate (40.2% 6 4.1%), and growth phenotypes (40.0% 6 2.1%) (average heritability 6 SE). However, both glucosinolate classes showed a multimodal distribution of nuclear heritabilities, while growth phenotypes displayed a more monomodal distribution (Figure 1; see Supplemental Data Set 2 online). The multimodal distribution of glucosinolates was largely caused by a presence/absence polymorphism at the alkenyl hydroxyl–producing (AOP) structural locus. In the same model, we also tested the effect of block and experiment that, while significant in all cases, were typically one-sixth the variance of the genotype term (0.39 median heritability versus 0.06 median variance explained by the combination of experiment and block). More critically, the genotype 3 experiment term was rarely significant and never explained more than the direct genotype term; thus, we further combined data across experiments. Restriction Enzyme Sequence Comparative Analysis Mapping of the Kas 3 Tsu RIL Population The available map for the Kas 3 Tsu RILs had a number of map gaps larger than 10 centimorgans (cM) that could diminish our ability to detect QTLs in these regions (Mackay, 2001; McKay et al., 2008). To close these gaps and provide more markers throughout the genome, we used restriction enzyme sequence comparative analysis (RESCAN) to generate a high-density map for the 341 Kas 3 Tsu RILs (Monson-Miller et al., 2012; Seymour et al., 2012). From this analysis, we identified a total of 908 good Figure 1. Partitioning of Heritability. Shown are frequency plots of different aspects of heritability ascribable to nuclear genomic variation for the different phenotypic classes: aliphatic glucosinolates (solid; red online), indole glucosinolates (dotted; blue online), and growth (dashed; green online). For each phenotypic class, the bin size is 5% for the frequency plots. [See online article for color version of this figure.] 4 of 17 The Plant Cell quality single nucleotide polymorphisms that were called in at least 59% of the lines. These markers were combined with the 161 markers from the previously published map to create a total set of 1069 markers (see Supplemental Data Set 3 online). Any missing genotypes were imputed using the fill.geno function in the R/qtl package using the single simulation replicate method (Broman et al., 2003). This generated a total map of 519 cM with an average gap size of 0.5 cM and a maximal gap size of 4.5 cM. Several of the gaps are in apparent recombination hotspots on chromosomes two and four (see Supplemental Figures 3 and 4 online). Thus, this map provides a high genetic resolution to accurately map QTLs. QTL Mapping To query the underlying genetic architecture of plant growth and chemical defense, we next mapped the QTLs regulating phenotypic variation in aliphatic glucosinolates, indolic glucosinolates, and growth-related traits within the Kas 3 Tsu population (Figure 2; see Supplemental Data Set 4 online). Aliphatic glucosinolates had the fewest detectable loci (4.1 6 0.5, range 6 to 1), indolic glucosinolate phenotypes had the highest number of QTLs per phenotype (6.5 6 1.5, range 2 to 12), and growth-related traits were intermediate (5.6 6 0.4, range 7 to 4). The number of QTLs per aliphatic glucosinolate phenotype was significantly lower than that observed for growth-related traits or indolic glucosinolates (t test, P < 0.05), while growth-related traits and indolic glucosinolates were statistically similar. The distribution of allelic effects across the QTLs for each phenotypic class also showed that there were differences in the genetic architecture underpinning growth-related traits and defense. Both defense phenotypic classes identified QTLs of moderate to large effect (here defined as >25% additive effect) with aliphatic QTLs being over one-third of the moderate to large effect class and indolic QTLs being approximately one-fifth moderate to large effect QTLs (Figure 3; see Supplemental Data Set 4 online). In distinct contrast, growth-related traits had no loci with effects over 16%, and as such, all growth loci are small-effect loci (Figure 3). Furthermore, there was a difference in the distribution of allelic effects with growth and indolic glucosinolates showing an equal fraction of positive and negative effects from the Kas-1 alleles, whereas aliphatic glucosinolates showed a bias toward negative Kas-1 alleles. The bias toward negative Kas-1 QTL alleles for aliphatic glucosinolates was statistically significant in comparison to both growth traits and indolic glucosinolates (x2, P < 0.05). Using QTL Localization to Partition Phenotypic Classes The distribution of QTLs across the genome was distinct for all three phenotypic classes with surprisingly little overlap (Figure 2; see Supplemental Data Set 4 online). This lack of overlap was readily apparent when clustering the phenotypes using the position and effect of the detected QTLs as input for the algorithm (Figure 2D). The clustering algorithm showed that aliphatic glucosinolate-related phenotypes formed two distinct phenotypic clusters. These two clusters were largely separated by the presence or absence of a QTL at the position of the GSL-Elong (glucosinolate elongation) locus (Figures 2A and 2D). The short-chain glucosinolates (i.e., allyl and butenyl) partitioned into a single cluster with no significant QTLs at GSL-Elong. Both the Kas and Tsu parents produce predominantly 3C glucosinolates, thus limiting variation in the 3C versus 4C glucosinolate allelic state that is the hallmark of the major effect alleles at GSL-Elong (Magrath et al., 1994; Kliebenstein et al., 2001b, 2001c; Kroymann et al., 2003). By contrast, variations in a cluster of long-chain aliphatic glucosinolate phenotypes (i.e., MSOO and MSOH) were associated with a QTL at the position of GSL-Elong. This suggests that despite both parents having a 3C allele at GSL-Elong, there is likely subtle variation between these two alleles and their effect on the production of long-chain glucosinolates. Therefore, the major-effect locus GSL-Elong appears to encode for additional minor-effect allelic variation, as has previously been suggested (Kroymann et al., 2003). It remains to be seen how often major-effect QTLs, such as those for other defense compounds or flowering time, have minor effect allelic classes in what were thought to be functionally equivalent alleles. The individual indolic glucosinolate phenotypes clustered together, while the summation of total indolic glucosinolate accumulation acted as an outlier for this class (Figure 2D). Interestingly, two of the unknown sulfated metabolites that were detected in the HPLC clustered with the major indole glucosinolate cluster, suggesting that they may actually be previously unidentified indolic glucosinolates. Agreeing with this hypothesis, both Unk2 and Unk3 had retention times that were similar to I3M and NMOI3M, and their absorption spectra were highly similar to I3M and NMOI3M, which indicates that they may be modified forms of these compounds. Additionally, the unknown metabolites shared a QTL on chromosome V with opposing allelic effects, suggesting that they may be precursor and product. However, their abundance was much lower than either I3M or NMOI3M, preventing us from conducting any analysis of their chemical structure. However, this does suggest that coassociating phenotypes based on QTL identification might be a useful method to link unknown metabolites to known pathways. This clustering of all the phenotypes showed that growthrelated loci formed a distinct cluster that shared almost no QTLs with the defense-related phenotypes, suggesting that there is no strong link between growth and metabolic defense within this population. Within the growth-related traits, there was a clustering by day such that rosette size at days 15, 19, and 23 shared almost all of the same QTLs within the cluster but had fewer shared QTLs with the early rosette size measures at days 9 and 11. This suggests that growth-related traits QTLs may have some modularity where there is a transition between groups of loci (modules) regulating size at early and late stages. Interestingly, both SGR and rosette size at day 7 had the fewest shared QTLs with the early and late growth modules. While SGR did share several loci with rosette size, there were unique growth QTLs detected by this more direct measure of growth. Hierarchical Genetic Architecture 5 of 17 Figure 2. Clustering of Glucosinolate and Growth QTLs across the Arabidopsis Genome. The distribution of QTLs across the genome for each phenotypic class is presented using a 5-cM sliding window. The genetic distances in graphs from (A) to (C) are scaled to match the genetic plot in part (D). The chromosomes are labeled with Roman numerals at the top of the figure. (A) The distribution of QTLs for 15 aliphatic glucosinolate phenotypes. (B) The distribution of QTLs for six indolic glucosinolate phenotypes. (C) The distribution of QTLs for seven growth phenotypes. (D) Heat map showing the location and effect of loci with LOD scores above the permuted LOD threshold 2, across the five chromosomes. Red indicates a positive effect of the Kas allele, while green (light hue) indicates a positive effect of the Tsu allele. Vertical white lines separate the chromosomes (1 to 5 from left to right). Clustering on the left is based on the absolute Pearson correlation of QTL effects across all significant loci for each phenotype. Phenotypes are colored as to their phenotypic class, with blue for indolic glucosinolate, red (dark) for aliphatic glucosinolate, and green (light) for growth. Three unknown metabolites detected in the glucosinolate HPLC are included. 6 of 17 The Plant Cell Figure 3. Differential QTL Effects per Phenotypic Class. Frequency plots of allelic effect sizes of aliphatic glucosinolate QTLs (solid; red online), indole glucosinolate QTLs (dotted; blue online), and growth QTLs Tsu 2 X Kas (dashed; green online). Allelic effects for each significant QTL are presented as percent effect, by estimating ½X for each significant main = XRIL effect marker. For each phenotypic class, the bin size for the frequency plots is 10% effect size. [See online article for color version of this figure.] Epistatic Architecture of Growth and Defense To compare the epistatic architecture regulating variation in growth and chemical defense within the Kas 3 Tsu population, we first developed statistical models for the QTL regulating each phenotypic class. To do this, we took all of the QTL positions linked to variation in a particular phenotypic class (Figures 2A to 2C; see Supplemental Data Set 4 online), identified the genetic marker closest to the peak, and used them within an ANOVA to test the effects of all main effect loci upon the phenotypes within the phenotypic class. Any markers that showed no significance with any phenotype within a phenotypic class was dropped from the model in a stepwise fashion until the final model contained only statistically significant markers (ANOVA). Individual nonsignificant loci were added back to test for significance, but none were able to improve the model. This generated a main effect model for each phenotypic class that had 11, 12, and 15 QTLs for aliphatic glucosinolate, growth-related traits, and indolic glucosinolates, respectively (see Supplemental Data Set 5 online). To test whether there were detectable effects of the cytoplasmic genomic variation upon the phenotypes after controlling for the main-effect QTLs, we also included cytoplasmic genotype as a factor in the model. Interestingly, none of the phenotypes in any phenotypic class had a detectable main effect of cytoplasmic genetic variation (see Supplemental Data Set 5 online). These phenotypic class-specific models were then used to test all possible pairwise epistatic interactions between significant main effect loci by ANOVA with a false discovery rate (FDR) of 0.2 per model (see Supplemental Data Set 6 online). The largest of our models for the indolic glucosinolates only used 136 of the 314 available degrees of freedom (43%) leaving the majority of the degrees of freedom for the residual variance, which suggests that we are not over fitting the model. Similarly for the aliphatic model, we only used 78 of the 314 available degrees of freedom (25%). All significant interactions were confirmed by rerunning the model containing only the significant terms from the full interaction model. The phenotypic class ANOVA models showed that all three phenotypic classes had significant pairwise epistasis in which all loci could be connected to all other loci through different epistatic networks within the phenotypic class (ANOVA; Figure 4; see Supplemental Data Set 6 online). As expected, the AOP (located at 17 cM on chromosome IV) and methylthioalkylmalate (located at 27 cM on chromosome V) loci that have been previously shown to epistatically regulate aliphatic glucosinolates with numerous other glucosinolates loci were highly interconnected within the aliphatic glucosinolate network (Kliebenstein et al., 2001a; Hansen et al., 2007, 2008; Wentzell et al., 2007; Chan et al., 2011). This supports the validity of these epistatic networks. For all three phenotypic classes, the frequency with which a main-effect QTL altered the phenotypes did not inherently predict the frequency with which it participated in epistatic interactions. In all instances, there were minor loci (i.e., GR.III.56, AL.II.4, and IN.III.63) that participated in epistatic interactions altering the majority of phenotypes in a phenotypic class; conversely, there were examples of dominant main-effect loci (i.e., GR.IV.2, AL.I.80, and IN.III.10) that rarely participated in a pairwise epistatic interaction (Figure 4). While all of the growth-related QTLs could be connected in an epistatic network, there was a distinct dichotomy where the epistatic links identified using SGR were completely distinct from those found for rosette diameter (Figure 4C). This is despite the fact that there is overlap in the main effect loci for both phenotypes (Figure 2D; see Supplemental Data Sets 4 and 5 online). Given that SGR and rosette size were measured on the same plants and all phenotypes showed a similar normal distribution, this is likely not a statistical artifact. Plotting the line means for epistatic pairs that center on the GR.V.82 Hierarchical Genetic Architecture 7 of 17 locus showed that rosette size and SGR show differential responses across the RILs (Figure 5). This supports the concept that SGR, while being estimated directly from the rosette size measurements, is a separate measure of growth in comparison to plant size at any single time. The epistatic networks for the three different phenotypic classes also had different connectivities (Figure 4). Aliphatic glucosinolate phenotypes had 3.5 (60.5) interactions, indolic glucosinolate phenotypes had the most interconnected epistatic network with each main effect QTL participating in 4.3 (60.4) interactions, and growth-related phenotypes had the fewest pairwise interactions, with 2.5 (60.4) interactions when SGR is included as part of the network. Both glucosinolate networks were statistically more connected than the growth-related trait network (t test, P = 0.003 indolic versus growth-related traits and P = 0.045 aliphatic versus growth-related traits) but similarly connected to each other. The mechanistic or statistical basis for this difference in connectivity for growth networks versus metabolic networks remains to be elucidated. Cytonuclear Epistasis Our previous analysis of heritability and main effects suggested that the genetic variation present within the cytosolic (plastidic and mitochondrial) genomes had little to no effect upon phenotypic variation for any of the phenotypes (Figure 1B; see Supplemental Data Sets 1 and 5 online). However, previous research had shown that the variation in cytosolic genome can have an influence on seed germination (Moison et al., 2010) and that it is possible to map epistatic interactions between the cytosolic and nuclear genomes (Tang et al., 2007). Thus, we included the cytosolic genomic variation present in this population as a main-effect term and tested for pairwise epistatic interactions between the cytosolic genome and the nuclear QTLs for each of the phenotypic classes (see Supplemental Data Set 6 online). As previously stated, all P values were adjusted to FDR of 0.2, and all significant interactions were confirmed within a simplified model containing these terms for each phenotype (ANOVA; see Supplemental Data Set 6 online). This showed that the cytosolic genome variation has significant epistasis with nuclear QTLs, especially for the indolic glucosinolates and growth-related traits. In fact, nearly every Figure 4. Epistatic Networks Regulating Glucosinolates and Growth. Epistatic networks were plotted with the nodes representing the loci and the edges showing statistically significant epistatic interactions for the three different phenotypic classes. Loci are named by the phenotypic class and the respective chromosome location in cM, for example, AL. V.8 indicates the aliphatic glucosinolate QTL on chromosome V at 8 cM. The previously cloned AOP and Elong loci are named as such for ease of reference. Cytonuclear interactions are represented by lines connecting nuclear loci to the cytoplasmic node. The sizes of the nodes are proportional to the fraction of phenotypes showing significant main effects at each locus. The color intensity of the edges indicate the percentage of phenotypes per phenotypic class that statistically identify each interaction with solid black being 50 to 75%, dark gray (dark blue online) being 25 to 50% of phenotypes, and light gray (light blue online) being <25% of the phenotypes within that phenotypic class. (A) Aliphatic glucosinolate epistatic network. (B) Indolic glucosinolate epistatic network. (C) Growth epistatic network. SGR epistatic interactions are indicated by dotted lines, while solid lines show epistatic interactions for rosette size. [See online article for color version of this figure.] 8 of 17 The Plant Cell growth-related trait and indolic glucosinolate phenotype identified at least two interactions between cytosolic and nuclear variation. The most common cytonuclear interaction for growth was with the GR.I.96 QTL, while the indolic glucosinolate interactions were spread over more markers but did not overlap with the growth interaction (Figures 4A and 4B). By contrast, there were only two aliphatic glucosinolate phenotypes that identified a single cytonuclear interaction, specifically with AL.III.74 (Figure 4A). Thus, there is a significant fraction of phenotypic variance that depends upon the cytosolic genome variation, but this variance in growth and defense is only detectable after identifying specific nuclear QTLs with which the cytosolic genome epistatically interacts. Plotting the line means for cytonuclear epistatic interactions showed that there were a variety of directionalities for this epistasis and the direction of allelic effects were specific for discrete phenotypes (Figure 6). Differential Genetic Architectures between Phenotypic Classes The three phenotypic classes differed in their number of QTLs and frequency of epistatic and cytonuclear interactions. To further examine these differences, we estimated the total variance explained by the significant terms within each class of terms (ANOVA, main effect, cytoplasmic effect, cytonuclear interaction, and epistatic interaction) across all the phenotypes for each phenotypic class using the simplest model for each phenotype as previously determined. Among the phenotypic classes, the fraction of genetic variance partitioned into the main effects and cytoplasmic effects were similar with main effects being about two-thirds of total variance and cytoplasmic being negligible. Interestingly, there were significant differences between the three phenotypic classes in the fraction of variance that was attributed to two epistatic classes. Aliphatic and indolic glucosinolate phenotypes had about one-third of their variance in the nuclear epistatic component, while growth was only 20% (Figure 7; t test, P < 0.05). Conversely, growth had the highest level of cytonuclear interactions, while aliphatic had negligible cytonuclear interactions, which agrees well with the epistatic network results (Figure 7). This agrees with the different number of identified cytonuclear interactions with aliphatic glucosinolate phenotypes having only one interaction to the five found for indolic glucosinolate phenotypes and three for growth phenotypes (Figure 4). The lower level of epistasis within the growth phenotypes also agrees with the lower connectivity of this network (Figures 4 and 7). This suggests that the genetic architecture for is fundamentally different, and the molecular and evolutionary basis for this remains to be empirically tested. Growth and Defense Relationship In several previous reports, we have mechanistically proven that natural polymorphisms in many glucosinolates genes, including the AOP and Elong loci polymorphic in this population, have causal effects on growth (Paul-Victor et al., 2010; Kerwin et al., Figure 5. Average Phenotypic Values for Loci Showing Epistatic Interactions Affecting Growth. Average phenotypic values for the four allelic classes of genotypes at pairwise QTL found to significantly interact for growth variation. Error bars indicate SE, and each genotype is represented by minimally 60 individuals. Asterisks in each individual graph identify phenotypes showing significant epistasis with P value < 0.05, while a cross sign indicates the phenotypes showing epistasis with P value < 0.1. The bottom letter on the x axis indicates the allelic state at the first QTL in the interaction, while the top letter indicates the allelic state of the second QTL listed for the interaction. The phenotype being plotted is listed in each graph including SGR. All daily growth phenotypes follow a similar pattern; therefore, only representative evenly spaced days are shown. (A) GR.III.56 3 GR.V.82 interaction specific to rosette daily growth but not SGR. (B) GR.IV.39 3 GR.V.82 specific for SGR but not rosette daily growth. [See online article for color version of this figure.] Hierarchical Genetic Architecture 9 of 17 Figure 6. Average Phenotypic Values for Loci Showing Cytonuclear Epistatic Interactions Affecting Growth. Example plots showing the average phenotypic values for the four allelic classes of genotypes at cytonuclear epistatic interactions are shown. Error bars indicate SE, and each genotype is represented by minimally 60 individuals. Asterisks in each individual graph identify phenotypes showing significant epistasis with P value < 0.05, while a cross sign indicates the phenotypes showing epistasis with P value < 0.1. The bottom letter on the x axis indicates the allelic state at the nuclear QTL in the interaction, while the top letter indicates the allelic state of the cytosolic genome for the interaction. The phenotype being plotted is listed in each graph. GLS stands for glucosinolate. (A) Effect of GR.I.96 3 cytoplasmic interaction on growth phenotypes. (B) Effect of GR.V.82 3 cytoplasmic interaction on growth phenotypes. (C) Effect of IN.I. 53 3 cytoplasmic interaction on I3M glucosinolate. (D) Effect of IN.II.59 3 cytoplasmic interaction on I3M glucosinolate. (E) Effect of IN.III.63 3 cytoplasmic interaction on I3M glucosinolate. (F) Effect of In.V.75 3 cytoplasmic interaction on I3M glucosinolate. [See online article for color version of this figure.] 2011; Züst et al., 2011). The AOP locus encodes two genes, AOP2 and AOP3, which contain natural knockout polymorphisms between the Kas and Tsu genotypes due to a local inversion that swaps the functional promoter. This leads to Kas expressing only the AOP2 gene in the rosette tissue and Tsu only expressing the AOP3 gene (Chan et al., 2011). Similarly, the methylthioalkylmalate locus contains functional knockout level polymorphisms within Arabidopsis (Magrath et al., 1994; Kliebenstein et al., 2001a, 2001b; Kroymann et al., 2003). Given that we have a validated mechanistic relationship between glucosinolates variation and growth variation that entails largeeffect polymorphisms in both the natural variation and induced variation, we tested if we could identify this relationship using this standard RIL data set. Using both QTL overlap and genetic correlation tests showed that we could not identify this mechanistic linkage in this standard sized RIL population (Figure 2). Combining this with our previous observation that even large RIL populations have a significant level of false negative error in detecting QTLs led us to hypothesize that the complexity of the growth architecture may hide the effect of glucosinolate variation at the whole organism level (Chan et al., 2011). To directly test this possibility, we conducted an ANCOVA to compare growth and glucosinolate variation where we included genotypic variation at all significant growth QTLs and their epistatic interactions directly in the ANCOVA model. This ANCOVA allows us to emphasize any potential links that may exist between growthrelated phenotypes and glucosinolate phenotypes but that may be hidden by the detected growth QTLs. This ANCOVA showed that we could identify a significant negative relationship between growth and long-chain aliphatic glucosinolates (see Supplemental Data Set 7 online). There was no identifiable relationship between indolic glucosinolates and any growthrelated trait. This was exactly as we had previously found in our mechanistic tests of glucosinolates genes where long-chain glucosinolates had the strongest effect on SGR and there was no significant link to daily growth estimates (Paul-Victor et al., 10 of 17 The Plant Cell infinitesimal model of genetic architecture than the chemical defenses (Fisher, 1918, 1930; Wright, 1931, 1968). These data suggest that chemical defenses will have fewer causative loci than growth, yet it does not prescribe a baseline for the number of causative loci regulating variation in chemical defenses. Recent work has suggested that there are 100s to 1000s of causal polymorphisms regulating variation in these chemical defenses within Arabidopsis (Chan et al., 2010, 2011). Since growth is not a product of glucosinolate metabolism alone, it is not surprising that there are likely many more causal polymorphisms that regulate natural variation in growth within Arabidopsis given that it is a higher order phenotype. Figure 7. Differential Partitioning of QTL Variances across Phenotypic Classes. The proportion of genetic variance explainable by various genetic factors in this population, as measured using type III sums of squares, for aliphatic glucosinolates (red online, left), indole glucosinolates (blue online, center), and growth (green online, right). Error bars show SE, and there are 15 aliphatic glucosinolate traits, six indolic glucosinolate traits, and seven growth traits. Main is the total variance linked to all of the nuclear QTL main effects, cytoplasmic is the proportion variance affiliated with variation in the cytoplasmic genome, cytonuclear epistasis is the proportion variance partitioned to epistatic interactions between nuclear loci and the cytoplasmic genomes, while nuclear epistasis is the fraction of variance explained by interactions between nuclear QTLs. Error variance was removed from this calculation to allow a comparison among the genetic effects. Letters show those phenotypic classes with a statistical difference in the variance fractioned into a given component. [See online article for color version of this figure.] 2010; Kerwin et al., 2011; Züst et al., 2011). Thus, by incorporating allelic variation at detected growth QTLs, we could uncover a previously hidden link between SGR and glucosinolate variation within this population. This indicates that there are likely other mechanistic links between phenotypes that cannot be identified using standard quantitative genetics statistics and population sizes. To identify these links requires more detailed investigations or much larger populations to better estimate the true genetic architecture of growth and defense. DISCUSSION Quantitative Genetics of Growth and Defense Within this population, the genetic architecture of quantitative variation in growth and chemical defense phenotypes showed significant differences. Even though all three phenotypic classes had similar average heritabilities, growth-related phenotypes displayed an approximately normal distribution of heritabilities, while the defense phenotypes displayed a bimodal distribution (Figure 1). The three phenotypic classes detected similar numbers of main effect QTLs, but the estimated allelic effects for each growth locus were significantly smaller than those for the chemical defenses (see Supplemental Data Set 4 online) (t test, P < 0.05). This suggests that growth-related traits follow a more Epistatic Networks and Causality versus Detection Power A key use of modern quantitative genetics is to begin testing and identifying mechanistic links between phenotypes, such as growth and chemical defense, that have had a long theoretical connection and a more recent causal link (Karban and Baldwin, 1997; Mauricio, 1998; Strauss and Agrawal, 1999; Thaler et al., 1999; Paul-Victor et al., 2010; Agrawal, 2011; Kerwin et al., 2011; Züst et al., 2011). We had previously proven that there was a link between glucosinolate genes and growth, but at a first approximation, our standard quantitative genetics data did not identify any common main effect QTLs between these phenotypic classes (Figures 2 to 4) (Kerwin et al., 2011; Züst et al., 2011). A key difference between this work and our previous studies is statistical power: Our studies identifying a mechanistic link focused on single gene manipulations mimicking natural glucosinolate polymorphisms, whereas this experiment used an RIL population with an unknown number of segregating loci. By accounting for the detected growth QTL directly in the ANCOVA statistical test, we could identify the previously known link between growth and glucosinolate variation (see Supplemental Data Set 7 online) (Mauricio, 1998; Lankau and Strauss, 2007, 2008; Lankau and Kliebenstein, 2009; Züst et al., 2011). This link was as predicted to be strongest between long-chain aliphatic glucosinolates and SGR. However, the fact that we were unable to identify a common main effect QTL despite this link between growth and glucosinolates suggests that even this relatively large RIL population is susceptible to an unrecognized rate of false negative QTL detection. Thus, we suggest that the glucosinolate epistatic network is causal with regards to growth but the link is simply hidden by the vast number of other detected and undetected growth QTLs. It is important to point out this observation, since the failure to detect causal links even in large RIL population would suggest that we currently have no way of accurately estimating how many loci are regulating growth and defense metabolism within any given Arabidopsis population. Cytonuclear Interactions and Natural Variation Detection Epistatic interactions between genetic variation in the nuclear and cytoplasmic genomes, cytonuclear interactions, have been found in most species where they have been empirically tested, including plants, insects, fungi, and animals (Burke et al., 1998; Roubertoux et al., 2003; Tao et al., 2004; Zeyl et al., 2005; Dowling et al., 2007, 2010; Etterson et al., 2007; Hierarchical Genetic Architecture Dowling Tang et al., 2013). In this report, we show that there are significant cytonuclear interactions for three phenotypic classes (Figures 4, 6, and 7). The fraction of phenotypic variance regulated by cytonuclear interactions differed among the three phenotypic classes, with growth showing the most variance regulated by cytonuclear interactions and aliphatic glucosinolates the least (Figure 7). Thus, natural variation in cytonuclear interactions can influence a significant fraction of phenotypic variation in traits that determine field fitness within Arabidopsis (Mauricio, 1998; Bidart-Bouzat and Kliebenstein, 2008). Interestingly, the presence of cytonuclear interactions did not depend on the identification of a main effect of the cytoplasmic genome, suggesting that the effects are not due to general changes in the cytosolic genome but instead are due to how the cytosolic and nuclear genomes interact. This could occur because a large fraction of the plants’ nuclear genome encodes proteins that function within either the mitochondria or plastids (Ajjawi et al., 2010). Thus, the fact that the effect of cytosolic genetic variation is largely detectable by interactions with the nuclear genetic variation suggests that variation in cytosolic genomes alters only a subset of the plastid’s or mitochondria’s function and not the operation of the entire organelle. In agreement with this model, the cytonuclear interactions are spread throughout the genome and do not act as master regulatory factors influencing all phenotypes, but instead are specific to distinct phenotypic classes (Figure 4). Thus, trying to identify cytoplasmic effects by solely measuring reciprocal F1 progeny will only identify a small subset of the real phenotypic effects of the cytosolic genomic variation (Moison et al., 2010). Instead, cytoplasmic effects need to be investigated and studied in a more structured population where cytonuclear interactions are directly measured. A major complication to the need to conduct a broader survey of cytonuclear interactions is that the vast majority of mapping populations are not designed to directly test cytosolic genetic variation because they are made using only one cytoplasmic genome. For instance, only two of the myriad of Arabidopsis RIL populations are made with reciprocal crosses necessary to allow variation in the cytosolic genome to be testable (Alonso-Blanco et al., 1998; Loudet et al., 2002; Meyer et al., 2007; McKay et al., 2008; Zhang et al., 2012). In other populations that are currently favored, such as mixed-parent populations (MAGIC, AMPRIL, etc.) (Kover et al., 2009; Huang et al., 2011) or association mapping populations (Atwell et al., 2010; Platt et al., 2010), there is variation in the cytosolic genomes present, but this variation has not been structured and would thus be confounded with the nuclear genomic variation even if it was measured. This confounding structure greatly hinders our ability to tease apart the cytosolic effects from the nuclear genetic background. Based on our observations, it could also greatly inhibit the ability to see either cytosolic or nuclear effects on a phenotype especially in situations of strong cytonuclear epistasis. Thus, there is a need to incorporate cytosolic genetic variation directly into the development of future populations. 11 of 17 Quantitative Genetics of SGR versus Growth A major interest in plant quantitative genetics for ecology, evolution, and agriculture is the study of higher order phenotypes, such as fitness, growth, and yield, which are likely highly integrative measures of numerous underlying networks. An interesting debate that occurs in each field for each integrative phenotype is how to measure fitness, growth, or yield when there are multiple descriptions and mathematical layers to these integrative phenotypes. Our results on the quantitative genetics of rosette size versus the SGR measure suggest that the argument on how to measure a phenotype is in fact flawed and that the true answer is that each measure of a highly integrative phenotype will provide a different view. In our analysis, rosette size and SGR identified only a subset of QTLs and epistatic interactions in common (Figures 2 and 4). If we had analyzed only one or the other description of growth, we would have obtained a picture of the quantitative genetics of growth that would have been severely lacking in its breadth. For instance, the costs of glucosinolate production on growth is most visible on the SGR measure with only hints of an effect on rosette size at later time points (see Supplemental Data Set 7 online) as has been previously found (Paul-Victor et al., 2010; Züst et al., 2011). Thus, measuring multiple aspects of integrative phenotypes, like growth, fitness, and yield, is more desirable than any individual number that can be placed on these phenotypes. In this analysis, we show that growth and chemical defenses in Arabidopsis have different quantitative genetic architecture with growth displaying a more infinitesimal model of variation. We were able to link chemical variation with imposing a variation in growth, suggesting that the identified chemical defense loci are potentially also undetected growth loci. We also revealed the significant effect of genetic variation in the cytosolic genomes upon the different phenotypic classes that was identified as predominantly interacting with nuclear genomic variation. Taken in its entirety, these data show that growth and chemical defenses are genetically linked but that future analysis will be required to develop a more focused image of how the quantitative genetics of these phenotypes are related. Finally, any future experiment attempting to measure the genetic architecture of defense, much less growth quantitative variation in Arabidopsis, requires significantly larger genetic populations of a different structure than are currently available or even being considered. METHODS Growth of the Kas 3 Tsu RIL Population Seeds of the 341 lines of the Kas 3 Tsu recombinant inbred population of Arabidopsis thaliana were obtained from the ABRC (Columbus, OH) (Juenger et al., 2006; McKay et al., 2008; Juenger et al., 2010). We grew a total of four per line, split into two randomized complete blocks per experiment with two independent experiments separated by approximately 3 months. We grew plants in large planting trays with 156 individual wells (bxwxh: 30 3 25 3 100 mm), filled with standard potting soil (Sunshine Mix #1; Sun Gro Horticulture). Prior to sowing, we imbibed seeds in distilled water and cold stratified them at 4°C for 4 d. We placed approximately three to five seeds of a single genotype in the center of a well and covered the trays with a transparent plastic hood to retain 12 of 17 The Plant Cell humidity during germination. Plants from both reciprocal subpopulations where intermixed in the block design to allow direct statistical testing of cytoplasmic effects. We recorded the germination day for all plants. After 1 week, we removed the transparent hoods and surplus plants to leave one seedling per well. We watered plants twice a week with nutrientenriched water (0.5% N-P-K fertilizer in a 2-1-2 ratio; Grow More 4-18-38) and kept them in a climate-controlled chamber at 22°C and a day/night cycle of 10 h/14 h. Genotyping the Kas 3 Tsu RIL Population DNA was extracted from each of the 341 Kas 3 Tsu RILs and rapidly genotyped with next-generation sequencing using the RESCAN method for Arabidopsis as previously described (Monson-Miller et al., 2012). Due to failed samples at the extraction or inappropriate genotypes from the sequencing, 25 of the lines had to be dropped, leaving 316 lines for QTL mapping with 136 RILs with Kas cytoplasm and 180 with Tsu cytoplasm. Libraries for each line were constructed with a unique 5-bp barcode at the 59-end of the adapters, and libraries were pooled in groups of 96 lines. These pools were sequenced on a single lane of an Illumina GAII or HiSeq high-throughput sequencer as well as a single lane of pooled Kas and Tsu genomic libraries. Single nucleotide polymorphisms were called using an in-laboratory pipeline using the BWA and SAM tools (Li and Durbin, 2009; Li et al., 2009). The final set of markers used for mapping at minimum 59% observed genotypic data across the individual RILs. The missing genotypes were imputed using the fill.geno function in the R/qtl package using the single simulation replicate method. Using the data, we tested for segregation distortion at individual loci and pairs of loci but did not detect any significant segregation distortion as previously reported for this population (McKay et al., 2008). Glucosinolate Analysis Thirty-one days after sowing, we harvested plants for glucosinolate analysis. One leaf from the first fully mature leaf pair of each plant was removed and stored in 90% methanol to inhibit enzymatic breakdown of chemical compounds and begin the extraction. We extracted glucosinolates and analyzed them by HPLC according to previously described methods (Kliebenstein et al., 2001b, 2001c). A number of lines failed to grow and were discarded from the analysis. Ten lines with aberrant or impossible glucosinolate profiles based on the known parental genotypes at the glucosinolates loci were removed from the analysis, leaving a total of 316 lines. The removal of these 10 lines was supported by the RESCAN analysis showing that these lines were genetically contaminant lines that had genotypes that were nearly identical to the homozygous Columbia0 accession and are impossible to obtain from the Kas or Tsu germplasm. Growth Recording To record plant growth, we photographed each individual planting tray every second day, starting from day 7 after sowing. We photographed the first block of the experiment in TIFF image format using a Nikon Coolpix E995 digital camera (3.1 megapixel) and the second block in JPEG format using a Fujifilm FinePix S1500 digital camera (10.0 megapixel). Plant size was measured relative to a 1-cm marker present in each tray. Flash was disabled, and we took the pictures under ambient (artificial) light with automatic camera settings for shutter speed, aperture, and focus. Image Processing and Extraction of Individual Leaf Areas We extracted the rosette area of plants from photographs using the image analysis software ENVI (version 4.6; ITT Vis) and supervised maximum likelihood classification, which uses digital number values of the images to separate plant area from the soil background. Pictures were taken so that plants only occupied the central quarter of each picture to minimize any consequences of lens distortion, while the random distribution of plants ensured that any remaining distortion was partitioned randomly. Because the distance of the camera to the trays differed slightly between photographs, the pixel/mm relation differed among images, making direct comparisons of leaf area impossible. In a first step, we thus registered all images to the scale of one reference image, using the image-to-image registration method as implemented in ENVI. This method requires the user to mark a minimum of four points or nodes that are present in both the reference and the sample image. The sample image was then transformed to match the reference based on those nodes. We registered all images to one reference image, placing up to eight nodes at the corners of corresponding wells within the planting trays. In the next step, we classified the images into three categories, namely, leaf, soil, and planting tray, using a maximum likelihood classifier in ENVI. We then selected training areas for the three classes in each image to calculate training class statistics based on the digital number values of the red, green, and blue layers of the image. Based on this training sample, every pixel within an image was assigned to one of the three classes by the maximum likelihood algorithm. As the final step, we marked the area of each well of a planting tray assigning leaf area within a well to a unique plant or rosette identity. The individual plant areas were then defined as the number of leaf pixels within each of the 104 masked wells. Pixel numbers were transformed into rosette areas (mm2) using the 1-cm reference marker present in each image. Growth and size measurements were based on the first 23 d of vegetative growth since plant area measurements were impossible at late stages due to leaf overlap. Size-Standardized Growth Rates To calculate SGR, we first need to fit an appropriate growth function. In this case, we found that a power-law function (Enquist et al., 1998, 1999) provided a good fit to the vegetative growth phase. Power-law growth assumes that the absolute growth rate is given by: dM ¼ aMb dt where M is a measure of size, a is the growth coefficient, and b is the scaling exponent. We modeled untransformed rosette area as a function of days after germination using the closed form solution of the equation above: Areat ¼ 8 > < > : ðArea0 1 2 b þ að1 2 bÞtÞ Area0 expat 1= 12b if b1 9 > = > if b ¼ 1 ; The rosette area of a plant at time t is thus a function of the three parameters: Area0 (estimated rosette area at time = 0), the growth coefficient a, and the scaling exponent b. This equation for area at time t can be fitted to data in the nlme package for the statistical software R (R Development Core Team, 2012), and all three parameters can be estimated (Hautier et al., 2010; Paine et al., 2012). We started with a model that only contained random effects of plant identity on all three parameters, essentially fitting unique growth curves to each individual plant. We then modified this model by adding fixed effects of RIL and experimental block to the model parameters and compared alternative models with different fixed effects using likelihood ratio tests. We extracted RIL-specific parameters from the best model and calculated SGR for each RIL. SGR is given by: SGR ¼ aArearef b 2 1 , where Arearef is a common size at which growth rate is calculated (Paul-Victor et al., 2010; Züst et al., 2011; Paine et al., 2012). The final best model supported a RIL effect on alpha but not on beta, therefore all RILs shared a common beta and choice of reference size did not affect relative rankings of SGR. Hierarchical Genetic Architecture Phenotype Abbreviations For phenotypes within the aliphatic glucosinolates, the abbreviations are as follows: 3-hydroxypropyl glucosinolate (OH3), 3-methylsulfinylpropyl glucosinolate (MSO3), 4-hydroxybutyl glucosinolate (OH4), allyl glucosinolate (Allyl), But-3-enyl glucosinolate (butenyl), 3-methylthiopropyl glucosinolate (MT3), 7-methylsulfinylheptyl glucosinolate (MSOH), 4methylthiobutyl glucosinolate (MT4), 8-methylsulfinyloctyl glucosinolate (MSOO), 8-methylthiooctyl glucosinolate (MT8), sum of three and four carbon glucosinolates (SC), sum of seven and eight carbon glucosinolates (LC), total aliphatic glucosinolate (TotAli), ratio of SC to TotAli (SCpTAli), and ratio of MTO to total 8C glucosinolate (MTOpMTMSOO). For indolic glucosinolates, the abbreviations are as follows: 4-hydroxyindol-3ylmethyl glucosinolate (OHI3M4), indol-3ylmethyl glucosinolate (I3M), 4-methoxy-indol-3ylmethyl glucosinolate (MOI3M4), N-methoxyindol-3ylmethyl glucosinolate (NMOI3M), sum of all indolic glucosinolates (indole), ratio of MOI3M4 to indole (MO4pInd), ratio of TotAli to indole (AlipInd), ratio of SC to indole (SCpInd), and Ratio of LC to indole (LCpInd). There were three putatively sulfated unknown compounds that were detected, but we did not determine their structure and these are abbreviated Unk1, Unk2, and Unk3. The growth phenotypes are abbreviated as follows: rosette size at 7 d after germination (Day7), rosette size at 9 d after germination (Day9), rosette size at 11 d after germination (Day11), rosette size at 15 d after germination (Day15), rosette size at 19 d after germination (Day19), and rosette size at 23 d after germination (Day23). Further information about these abbreviations is shown in Supplemental Data Set 1 online. Estimation of Heritability All 316 lines were represented in every experiment and every block in both experiments creating a perfectly balanced randomized complete block design. All phenotypic data were used to calculate estimates of broad-sense heritability (H) for each glucosinolate and daily growth estimate as H = s2g/s2p, where s2g was estimated for both the RIL nuclear and cytoplasmic genotypes and s2p was the total phenotypic variance for a trait (Liu, 1998). The ANOVA model for each metabolite or growth phenotype in each line (ygmeb) was: ygceb ¼mþCc þGg ðCc Þ þ Ee þBb ðEe Þ þ Cc 3Ee þGg ðCc Þ3Ee þ«gceb , where c = the Kas or Tsu cytoplasm; g = the 1.316 for the 316 RILs, e = experiment 1 or experiment 2, and b = block 1..4 nested within experiment. This allowed cytoplasmic effects to be directly tested in the C term and each RIL genotype (G) nested within the appropriate cytoplasmic class, either Kas or Tsu. Experiment and block nested within experiment were treated as random terms within the model, while genotype and cytoplasm are considered fixed (see Supplemental Data Set 1 online). s2g for the nuclear variation was pulled from the Gg(Cc) term, while s2g for cytoplasmic variation was pulled directly from the Cc term. We were not able to estimate heritability for SGR because of the nonlinear nature of the function. Thus heritability for this term would not be comparable to the heritabilities of other traits. We were able to directly use mean values for the RILs for further analysis as we had a perfect randomized complete block design with no missing values which means that LSmeans are identical to raw means in this instance. All statistical analyses in this article were run within R (R Development Core Team, 2012). 13 of 17 mapping was implemented using the cim function in the R/qtl package with a 10-cM window. Forward regression was used to identify three cofactors per trait. We estimated permutation-derived empirical thresholds using 1000 permutations for each mapped trait. This provided a range of 1.92 to 1.98 as being statistically significant so we set a uniform threshold for all traits and QTLs where a LOD score above 2 was considered significant for further analysis to begin the analysis somewhat conservatively (Churchill and Doerge, 1994; Doerge and Churchill, 1996). Composite interval mapping to assign significance based on the underlying trait distribution is robust at handling normal or near normal trait distributions (Rebai, 1997), as found for most of our phenotypes. The define.peak function implemented in the R/eqtl package was used to identify the location and one-LOD interval of each significant QTL for each trait (Wang et al., 2006). The effectscan function in the R/qtl package was used to estimate the QTL additive effect (R Development Core Team, 2012). QTL clusters were identified using a QTL summation approach where the position of each QTL for each trait was plotted on the chromosome by placing a 1 at the peak of the QTL. This was then used to sum the number of traits that had a detected QTL at a given position using a 5-cM sliding window across the genome (Kliebenstein et al., 2006). The traits were divided into three phenotypic classes: growth, aliphatic glucosinolates, and indolic glucosinolates. Glucosinolates were split into two classes based on previous data showing that aliphatic and indolic glucosinolates are synthesized by different enzymes and regulated by different transcription factors and can therefore be considered distinct pathways (Burow et al., 2010; Sønderby et al., 2010a, 2010b). The summation was conducted within a 5-cM sliding window for the three trait classes. The QTL clusters identified defined genetic positions that were named respective to their phenotypic class and genetic positions with a prefix indicating the phenotypic class followed by the chromosome number and the cM position. For example, AL. I. 80 indicates an aliphatic glucosinolate QTL on chromosome I at 80 cM. The QTLs detected at the previously characterized and cloned GSL-AOP and GSL-Elong loci were named as such (Magrath et al., 1994; Kliebenstein et al., 2001a, 2001b; Kroymann et al., 2003). Phenotype Marker Model ANOVA To directly test the effect of each identified QTL cluster for each phenotypic class, we used an ANOVA model containing the markers most closely associated with each of the significant QTL clusters as individual main effect terms. For each phenotypic class (growth, aliphatic glucosinolate, and indolic glucosinolate), the average phenotype in lines of genotype g at marker m was shown as ygm. The metabolite or growth phenotype for each metabolite in each line (ygm) was as follows: ygm ¼mþ∑2g¼1 ∑m m¼1 Mmg þ«gm , where g = Kas(1) or Tsu(2); m = 1, .,m. The main effect of the markers was denoted as M with growth having a model involving 12 markers (m), aliphatic glucosinolates having 11 markers, and indolic glucosinolate having 15 markers. For each phenotypic class, the cytoplasmic genome was included as an additional marker to test for cytosolic genome effects. We tested all phenotypes per phenotypic class with the appropriate model implemented in the R/car package, which returned all P values, Type III sums of squares for the complete model and each main effect, and QTL main-effect estimates (in terms of allelic substitution values) (Fox and Weisberg, 2011; R Development Core Team, 2012). QTL Analysis We used the above genotyping analysis on the Kas 3 Tsu RIL population to generate a genetic map for the population. To detect glucosinolate and growth QTLs, we used the average phenotype per RIL across all experiments (see Supplemental Data Set 2 online) (Basten et al., 1999; Zeng et al., 1999; Wang et al., 2006). Composite interval QTL Epistasis Analysis To test directly for epistatic interactions between the detected QTLs, we conducted an ANOVA using the marker model generated for each of the three phenotypic classes whereby each marker in the model had 14 of 17 The Plant Cell a significant main effect within the phenotype class. We used this full model per phenotypic class because we had previous evidence that RIL populations have a significant false negative QTL detection issue and wanted to be inclusive of all possible significant loci (Chan et al., 2011). Within the model, we tested all possible pairwise interactions between the markers for each of the three phenotypic classes. For each phenotype, the average value in the RILs of genotype g at marker m was shown as ygm. The model for each metabolite or growth phenotype (ygm) was 2 m m ygm ¼mþ∑2g¼1 ∑m m¼1 Mgm þ∑g¼1 ∑m¼1 ∑n¼mþ1 Mgm Mgn þ«gmn , where g = Kas or Tsu; m = 1, .,m; and n was the identity of the second marker for an interaction. The main effect of the markers was denoted as M with growth having a model involving 12 markers, aliphatic glucosinolates having 11 markers, and indolic glucosinolates having 15 markers, as above. For each phenotypic class, the cytoplasmic genome was included as an additional single-locus marker to test for interactions between the cytosolic and nuclear genomes. Q values, Type III sums of squares for the complete model and each individual term, and QTL pairwise-effect estimates in terms of allelic substitution values were obtained as described for marker model ANOVA (Fox and Weisberg, 2011; R Development Core Team, 2012). All presented significance values were corrected for multiple testing within a model using FDR (<0.20) in the automated script (Storey, 2002; Storey and Tibshirani, 2003). We chose this FDR because we have previously shown that we can readily validate loci even significant at even a nominal P value (Chan et al., 2011). Thus, we feel that this FDR is a good compromise between false positives and false negatives that can both dramatically influence the results. Going to a more stringent FDR would increase our false negatives and remove significant insight (Chan et al., 2011). The main effect and epistatic interactions of the loci in each phenotypic class were visualized using cytoscape.v2.8.3. with interactions significant for <10% of the phenotypes in each phenotypic class excluded from the network analysis (Smoot et al., 2011). For the estimation of epistasis, each phenotype was reanalyzed only with the significant model components to ensure that they remained significant in the simplified model. From this simplified model, variance attributable to each term was obtained. Growth and Glucosinolates ANCOVA To test for a link between growth and glucosinolates that may be conditioned upon the existing genetic architecture regulating growth, we conducted an ANCOVA. The ANCOVA specifically tested for a link between growth and glucosinolates after controlling for the above genetic model found to regulate growth. The model was specifically as follows: 2 m m ygm ¼mþ∑2g¼1 ∑m m¼1 Mgm þ∑g¼1 ∑m¼1 ∑n¼mþ1 Mgm Mgn þxgm þ«gmn , where ygm is the growth of all lines and xgm is the specific glucosinolate content of the same lines. In this model, all markers and interactions from the growth QTL model are included as above where g = Kas or Tsu; m = 1, .,m; and n was the identity of the second marker for an interaction. This was conducted independently for the different growth phenotypes to test if there was a difference among the different estimates of growth. Additionally, we focused on just three glucosinolate phenotypes, total indolic, total longchain aliphatic glucosinolates, and total short-chain glucosinolates, as the QTL mapping data suggested that these could be considered key descriptors of the different glucosinolate phenotypes within this population. The ANCOVA was generated using the R package (R Development Core Team, 2012). Accession Numbers Sequence data from this article can be found in the GenBank/EMBL data libraries under the following accession numbers: AOP2, At4g03060; AOP3, At4g03050; MAM1, At5g20310; and MAM3, At5g203020. Supplemental Data The following materials are available in the online version of this article. Supplemental Figure 1. Distributions of Phenotypes Showing Statistically Normal Distributions. Supplemental Figure 2. Distributions of Non-Normal Phenotypes. Supplemental Figure 3. High-Density Genetic Map. Supplemental Figure 4. Physical to Genetic Distance in Kas 3 Tsu. Supplemental Data Set 1. Phenotypic Heritability. Supplemental Data Set 2. Phenotypic Means for Glucosinolate and Growth Traits. Supplemental Data Set 3. Genotypic Data for 316 RILs Used for Genetic Mapping after Imputation. Supplemental Data Set 4. QTL Information. Supplemental Data Set 5. Single Marker QTL Analysis by ANOVA. Supplemental Data Set 6. Pairwise Epistatic ANOVA q Values and Type III Sums of Squares. Supplemental Data Set 7. Analysis of Covariance of Glucosinolates and Growth. ACKNOWLEDGMENTS This work was supported by the Forschungskredit of the University of Zürich (T.Z.), by Swiss National Science Foundation Grant 31-107531 (L.A.T.) and by U.S. National Science Foundation Grant DBI0820580 (D.J.K.). AUTHOR CONTRIBUTIONS J.A.C., B.L., T.Z., L.A.T., and D.J.K. contributed to the design of the experiment. B.J., J.A.C., T.Z., B.L., G.S.S, and M.I. contributed to the execution of the experiments and collection of the data. All authors contributed to interpretation of the results. B.J., J.A.C., T.Z., L.A.T., and D.J.K. contributed to the writing of the article. Received April 11, 2013; revised May 9, 2013; accepted May 16, 2016; published June 7, 2013. REFERENCES Agrawal, A.A. (2007). Macroevolution of plant defense strategies. Trends Ecol. Evol. (Amst.) 22: 103–109. Agrawal, A.A. (2011). New synthesis—Trade-offs in chemical ecology. J. Chem. Ecol. 37: 230–231. Ajjawi, I., Lu, Y., Savage, L.J., Bell, S.M., and Last, R.L. (2010). Large-scale reverse genetics in Arabidopsis: Case studies from the Chloroplast 2010 Project. Plant Physiol. 152: 529–540. Alonso-Blanco, C., Peeters, A.J.M., Koornneef, M., Lister, C., Dean, C., van den Bosch, N., Pot, J., and Kuiper, M.T.R. (1998). Development of an AFLP based linkage map of Ler, Col and Cvi Arabidopsis thaliana ecotypes and construction of a Ler/Cvi recombinant inbred line population. Plant J. 14: 259–271. Atwell, S., et al. (2010). Genome-wide association study of 107 phenotypes in Arabidopsis thaliana inbred lines. Nature 465: 627–631. Hierarchical Genetic Architecture Basten, C.J., Weir, B.S., and Zeng, Z.-B. (1999). QTL Cartographer, Version 1.13. (Raleigh, NC: North Carolina State University). Bednarek, P., Pislewska-Bednarek, M., Svatos, A., Schneider, B., Doubsky, J., Mansurova, M., Humphry, M., Consonni, C., Panstruga, R., Sanchez-Vallet, A., Molina, A., and SchulzeLefert, P. (2009). A glucosinolate metabolism pathway in living plant cells mediates broad-spectrum antifungal defense. Science 323: 101–106. Beekwilder, J., et al. (2008). The impact of the absence of aliphatic glucosinolates on insect herbivory in Arabidopsis. PLoS ONE 3: e2068. Bidart-Bouzat, M.G., and Kliebenstein, D.J. (2008). Differential levels of insect herbivory in the field associated with genotypic variation in glucosinolates in Arabidopsis thaliana. J. Chem. Ecol. 34: 1026–1037. Broman, K.W., Wu, H., Sen, Ś., and Churchill, G.A. (2003). R/qtl: QTL mapping in experimental crosses. Bioinformatics 19: 889–890. Buckler, E.S., et al. (2009). The genetic architecture of maize flowering time. Science 325: 714–718. Burke, J.M., Voss, T.J., and Arnold, M.L. (1998). Genetic interactions and natural selection in Louisiana iris hybrids. Evolution 52: 1304–1310. Burow, M., Halkier, B.A., and Kliebenstein, D.J. (2010). Regulatory networks of glucosinolates shape Arabidopsis thaliana fitness. Curr. Opin. Plant Biol. 13: 348–353. Chan, E.K., Rowe, H.C., Corwin, J.A., Joseph, B., and Kliebenstein, D.J. (2011). Combining genome-wide association mapping and transcriptional networks to identify novel genes controlling glucosinolates in Arabidopsis thaliana. PLoS Biol. 9: e1001125. Chan, E.K.F., Rowe, H.C., and Kliebenstein, D.J. (2010). Understanding the evolution of defense metabolites in Arabidopsis thaliana using genome-wide association mapping. Genetics 185: 991–1007. Churchill, G.A., and Doerge, R.W. (1994). Empirical threshold values for quantitative trait mapping. Genetics 138: 963–971. Clay, N.K., Adio, A.M., Denoux, C., Jander, G., and Ausubel, F.M. (2009). Glucosinolate metabolites required for an Arabidopsis innate immune response. Science 323: 95–101. Cui, J.P., Jander, G., Racki, L.R., Kim, P.D., Pierce, N.E., and Ausubel, F.M. (2002). Signals involved in Arabidopsis resistance to Trichoplusia ni caterpillars induced by virulent and avirulent strains of the phytopathogen Pseudomonas syringae. Plant Physiol. 129: 551–564. Davila, J.I., Arrieta-Montiel, M.P., Wamboldt, Y., Cao, J., Hagmann, J., Shedge, V., Xu, Y.-Z., Weigel, D., and Mackenzie, S.A. (2011). Double-strand break repair processes drive evolution of the mitochondrial genome in Arabidopsis. BMC Biol. 9: 64. de Vos, M., Kriksunov, K.L., and Jander, G. (2008). Indole-3-acetonitrile production from indole glucosinolates deters oviposition by Pieris rapae. Plant Physiol. 146: 916–926. Doerge, R.W., and Churchill, G.A. (1996). Permutation tests for multiple loci affecting a quantitative character. Genetics 142: 285–294. Dowling, D.K., Friberg, U., Hailer, F., and Arnqvist, G. (2007). Intergenomic epistasis for fitness: Within-population interactions between cytoplasmic and nuclear genes in Drosophila melanogaster. Genetics 175: 235–244. Dowling, D.K., Meerupati, T., and Arnqvist, G. (2010). Cytonuclear interactions and the economics of mating in seed beetles. Am. Nat. 176: 131–140. Enquist, B.J., Brown, J.H., and West, G.B. (1998). Allometric scaling of plant energetics and population density. Nature 395: 163–165. Enquist, B.J., West, G.B., Charnov, E.L., and Brown, J.H. (1999). Allometric scaling of production and life-history variation in vascular plants. Nature 401: 907–911. 15 of 17 Etterson, J.R., Keller, S.R., and Galloway, L.F. (2007). Epistatic and cytonuclear interactions govern outbreeding depression in the autotetraploid Campanulastrum americanum. Evolution 61: 2671– 2683. Fan, J., Crooks, C., Creissen, G., Hill, L., Fairhurst, S., Doerner, P., and Lamb, C. (2011). Pseudomonas sax genes overcome aliphatic isothiocyanate-mediated non-host resistance in Arabidopsis. Science 331: 1185–1188. Fisher, R.A. (1918). The correlation between relatives on the supposition of Mendelian inheritance. Philos. J. 52: 399–433. Fisher, R.A. (1930). The Genetical Theory of Natural Selection. (Oxford, UK: Clarendon). Fox, J., and Weisberg, S. (2011). An R Companion to Applied Regression. (Thousand Oaks, CA: SAGE). Fu, J., et al. (2009). System-wide molecular evidence for phenotypic buffering in Arabidopsis. Nat. Genet. 41: 166–167. Hageman, R.B., Kliebenstein, D.J., and Churchill, G.A. (2012). What can causal networks tell us about metabolic pathways? PLoS Comput. Biol. 8(4): e1002458. Hansen, B.G., Kerwin, R.E., Ober, J.A., Lambrix, V.M., MitchellOlds, T., Gershenzon, J., Halkier, B.A., and Kliebenstein, D.J. (2008). A novel 2-oxoacid-dependent dioxygenase involved in the formation of the goiterogenic 2-hydroxybut-3-enyl glucosinolate and generalist insect resistance in Arabidopsis. Plant Physiol. 148: 2096–2108. Hansen, B.G., Kliebenstein, D.J., and Halkier, B.A. (2007). Identification of a flavin-monooxygenase as the S-oxygenating enzyme in aliphatic glucosinolate biosynthesis in Arabidopsis. Plant J. 50: 902–910. Hautier, Y., Hector, A., Vojtech, E., Purves, D., and Turnbull, L.A. (2010). Modelling the growth of parasitic plants. J. Ecol. 98: 857–866. Huang, X., Paulo, M.-J., Boer, M., Effgen, S., Keizer, P., Koornneef, M., and van Eeuwijk, F.A. (2011). Analysis of natural allelic variation in Arabidopsis using a multiparent recombinant inbred line population. Proc. Natl. Acad. Sci. USA 108: 4488–4493. Juenger, T.E., Sen, S., Bray, E., Stahl, E., Wayne, T., McKay, J., and Richards, J.H. (2010). Exploring genetic and expression differences between physiologically extreme ecotypes: Comparative genomic hybridization and gene expression studies of Kas-1 and Tsu-1 accessions of Arabidopsis thaliana. Plant Cell Environ. 33: 1268– 1284. Juenger, T.E., Wayne, T., Boles, S., Symonds, V.V., McKay, J., and Coughlan, S.J. (2006). Natural genetic variation in whole-genome expression in Arabidopsis thaliana: The impact of physiological QTL introgression. Mol. Ecol. 15: 1351–1365. Karban, R., and Baldwin, I.T. (1997). Induced Responses to Herbivory. (Chicago: University of Chicago Press). Kerwin, R.E., Jiménez-Gomez, J.M., Fulop, D., Harmer, S.L., Maloof, J.N., and Kliebenstein, D.J. (2011). Network quantitative trait loci mapping of circadian clock outputs identifies metabolic pathway-to-clock linkages in Arabidopsis. Plant Cell 23: 471–485. Kliebenstein, D., Lambrix, V., Reichelt, M., Gershenzon, J., and Mitchell-Olds, T. (2001a). Gene duplication in the diversification of secondary metabolism: Tandem 2-oxoglutarate-dependent dioxygenases control glucosinolate biosynthesis in Arabidopsis. Plant Cell 13: 681–693. Kliebenstein, D., Pedersen, D., Barker, B., and Mitchell-Olds, T. (2002). Comparative analysis of quantitative trait loci controlling glucosinolates, myrosinase and insect resistance in Arabidopsis thaliana. Genetics 161: 325–332. Kliebenstein, D.J., Gershenzon, J., and Mitchell-Olds, T. (2001b). Comparative quantitative trait loci mapping of aliphatic, indolic and 16 of 17 The Plant Cell benzylic glucosinolate production in Arabidopsis thaliana leaves and seeds. Genetics 159: 359–370. Kliebenstein, D.J., Kroymann, J., Brown, P., Figuth, A., Pedersen, D., Gershenzon, J., and Mitchell-Olds, T. (2001c). Genetic control of natural variation in Arabidopsis glucosinolate accumulation. Plant Physiol. 126: 811–825. Kliebenstein, D.J., West, M.A., van Leeuwen, H., Loudet, O., Doerge, R.W., and St Clair, D.A. (2006). Identification of QTLs controlling gene expression networks defined a priori. BMC Bioinformatics 7: 308. Kover, P.X., Valdar, W., Trakalo, J., Scarcelli, N., Ehrenreich, I.M., Purugganan, M.D., Durrant, C., and Mott, R. (2009). A multiparent advanced generation inter-cross to fine-map quantitative traits in Arabidopsis thaliana. PLoS Genet. 5: e1000551. Kroymann, J., Donnerhacke, S., Schnabelrauch, D., and MitchellOlds, T. (2003). Evolutionary dynamics of an Arabidopsis insect resistance quantitative trait locus. Proc. Natl. Acad. Sci. USA 100 (suppl. 2): 14587–14592. Lambrix, V., Reichelt, M., Mitchell-Olds, T., Kliebenstein, D.J., and Gershenzon, J. (2001). The Arabidopsis epithiospecifier protein promotes the hydrolysis of glucosinolates to nitriles and influences Trichoplusia ni herbivory. Plant Cell 13: 2793–2807. Lankau, R.A. (2007). Specialist and generalist herbivores exert opposing selection on a chemical defense. New Phytol. 175: 176–184. Lankau, R.A., and Kliebenstein, D.J. (2009). Competition, herbivory and genetics interact to determine the accumulation and fitness consequences of a defence metabolite. J. Ecol. 97: 78–88. Lankau, R.A., and Strauss, S.Y. (2007). Mutual feedbacks maintain both genetic and species diversity in a plant community. Science 317: 1561–1563. Lankau, R.A., and Strauss, S.Y. (2008). Community complexity drives patterns of natural selection on a chemical defense of Brassica nigra. Am. Nat. 171: 150–161. Li, H., and Durbin, R. (2009). Fast and accurate short read alignment with Burrows-Wheeler transform. Bioinformatics 25: 1754–1760. Li, H., Handsaker, B., Wysoker, A., Fennell, T., Ruan, J., Homer, N., Marth, G., Abecasis, G., and Durbin, R.1000 Genome Project Data Processing Subgroup (2009). The Sequence alignment/map format and SAMtools. Bioinformatics 25: 2078–2079. Lisec, J., Meyer, R.C., Steinfath, M., Redestig, H., Becher, M., Witucka-Wall, H., Fiehn, O., Törjék, O., Selbig, J., Altmann, T., and Willmitzer, L. (2008). Identification of metabolic and biomass QTL in Arabidopsis thaliana in a parallel analysis of RIL and IL populations. Plant J. 53: 960–972. Liu, B.-H. (1998). Statistical Genomics: Linkage, Mapping and QTL Analysis. (Boca Raton, FL: CRC Press). Loudet, O., Chaillou, S., Camilleri, C., Bouchez, D., and DanielVedele, F. (2002). Bay-0 x Shahdara recombinant inbred line population: a powerful tool for the genetic dissection of complex traits in Arabidopsis. Theor. Appl. Genet. 104: 1173–1184. Loudet, O., Chaillou, S., Merigout, P., Talbotec, J., and DanielVedele, F. (2003). Quantitative trait loci analysis of nitrogen use efficiency in Arabidopsis. Plant Physiol. 131: 345–358. Loudet, O., Gaudon, V., Trubuil, A., and Daniel-Vedele, F. (2005). Quantitative trait loci controlling root growth and architecture in Arabidopsis thaliana confirmed by heterogeneous inbred family. Theor. Appl. Genet. 110: 742–753. Mackay, T.F.C. (2001). The genetic architecture of quantitative traits. Annu. Rev. Genet. 35: 303–339. Magrath, R., Bano, F., Morgner, M., Parkin, I., Sharpe, A., Lister, C., Dean, C., Turner, J., Ludiate, D., and Mithen, R. (1994). Genetics of aliphatic glucosinolates. I. Side chain elongation in Brassica napus and Arabidopsis thaliana. Heredity 72: 290–299. Mauricio, R. (1998). Costs of resistance to natural enemies in field populations of the annual plant Arabidopsis thaliana. Am. Nat. 151: 20–28. McKay, J.K., Richards, J.H., Nemali, K.S., Sen, S., Mitchell-Olds, T., Boles, S., Stahl, E.A., Wayne, T., and Juenger, T.E. (2008). Genetics of drought adaptation in Arabidopsis thaliana II. QTL analysis of a new mapping population, KAS-1 x TSU-1. Evolution 62: 3014–3026. Meyer, R.C., Steinfath, M., Lisec, J., Becher, M., Witucka-Wall, H., Törjék, O., Fiehn, O., Eckardt, A., Willmitzer, L., Selbig, J., and Altmann, T. (2007). The metabolic signature related to high plant growth rate in Arabidopsis thaliana. Proc. Natl. Acad. Sci. USA 104: 4759–4764. Moison, M., Roux, F., Quadrado, M., Duval, R., Ekovich, M., Lê, D.-H., Verzaux, M., and Budar, F. (2010). Cytoplasmic phylogeny and evidence of cyto-nuclear co-adaptation in Arabidopsis thaliana. Plant J. 63: 728–738. Monson-Miller, J., Sanchez-Mendez, D.C., Fass, J., Henry, I.M., Tai, T.H., and Comai, L. (2012). Reference genome-independent assessment of mutation density using restriction enzyme-phased sequencing. BMC Genomics 13: 72. Müller, R., de Vos, M., Sun, J.Y., Sønderby, I.E., Halkier, B.A., Wittstock, U., and Jander, G. (2010). Differential effects of indole and aliphatic glucosinolates on lepidopteran herbivores. J. Chem. Ecol. 36: 905–913. Paine, C.E.T., Marthews, T.R., Vogt, D.R., Purves, D., Rees, M., Hector, A., and Turnbull, L.A. (2012). How to fit nonlinear plant growth models and calculate growth rates: An update for ecologists. Methods Ecol. Evol. 3: 245–256. Paul-Victor, C., Züst, T., Rees, M., Kliebenstein, D.J., and Turnbull, L.A. (2010). A new method for measuring relative growth rate can uncover the costs of defensive compounds in Arabidopsis thaliana. New Phytol. 187: 1102–1111. Pfalz, M., Vogel, H., and Kroymann, J. (2009). The gene controlling the indole glucosinolate modifier1 quantitative trait locus alters indole glucosinolate structures and aphid resistance in Arabidopsis. Plant Cell 21: 985–999. Platt, A., et al. (2010). The scale of population structure in Arabidopsis thaliana. PLoS Genet. 6: e1000843. R Development Core Team (2012). R: A Language and Environment for Statistical Computing. (Vienna, Austria: R Foundation for Statistical Computing). Rebai, A. (1997). Comparison of methods for regression interval mapping in QTL analysis with non-nomral traits. Genet. Res., Camb. 69: 69–74. Reichelt, M., Brown, P.D., Schneider, B., Oldham, N.J., Stauber, E., Tokuhisa, J., Kliebenstein, D.J., Mitchell-Olds, T., and Gershenzon, J. (2002). Benzoic acid glucosinolate esters and other glucosinolates from Arabidopsis thaliana. Phytochemistry 59: 663–671. Roubertoux, P.L., Sluyter, F., Carlier, M., Marcet, B., MaaroufVeray, F., Chérif, C., Marican, C., Arrechi, P., Godin, F., Jamon, M., Verrier, B., and Cohen-Salmon, C. (2003). Mitochondrial DNA modifies cognition in interaction with the nuclear genome and age in mice. Nat. Genet. 35: 65–69. Rowe, H.C., Hansen, B.G., Halkier, B.A., and Kliebenstein, D.J. (2008). Biochemical networks and epistasis shape the Arabidopsis thaliana metabolome. Plant Cell 20: 1199–1216. Seymour, D.K., Filiault, D.L., Henry, I.M., Monson-Miller, J., Ravi, M., Pang, A., Comai, L., Chan, S.W.L., and Maloof, J.N. (2012). Rapid creation of Arabidopsis doubled haploid lines for quantitative trait locus mapping. Proc. Natl. Acad. Sci. USA 109: 4227–4232. Hierarchical Genetic Architecture Smoot, M.E., Ono, K., Ruscheinski, J., Wang, P.-L., and Ideker, T. (2011). Cytoscape 2.8: New features for data integration and network visualization. Bioinformatics 27: 431–432. Sønderby, I.E., Burow, M., Rowe, H.C., Kliebenstein, D.J., and Halkier, B.A. (2010b). A complex interplay of three R2R3 MYB transcription factors determines the profile of aliphatic glucosinolates in Arabidopsis. Plant Physiol. 153: 348–363. Sønderby, I.E., Geu-Flores, F., and Halkier, B.A. (2010a). Biosynthesis of glucosinolates—Gene discovery and beyond. Trends Plant Sci. 15: 283–290. Storey, J.D. (2002). A direct approach to false discovery rates. J. R. Stat. Soc. Series B Stat. Methodol. 64: 479–498. Storey, J.D., and Tibshirani, R. (2003). Statistical significance for genomewide studies. Proc. Natl. Acad. Sci. USA 100: 9440–9445. Strauss, S.Y., and Agrawal, A.A. (1999). The ecology and evolution of plant tolerance to herbivory. Trends Ecol. Evol. (Amst.) 14: 179–185. Tang, Z., Wang, X., Hu, Z., Yang, Z., and Xu, C. (2007). Genetic dissection of cytonuclear epistasis in line crosses. Genetics 177: 669–672. Tang, Z., Yang, Z., Hu, Z., Zhang, D., Lu, X., Deng, D., and Xu, C. (2013). Cytonuclear epistatic quantitative trait locus mapping for plant height and ear height in maize. Mol. Breed. 31: 1–14. Tao, D.Y., Hu, F.Y., Yang, J.Y., Yang, G.F., Yang, Y.Q., Xu, P., Li, J., Ye, C.R., and Dai, L.Y. (2004). Cytoplasm and cytoplasm-nucleus interactions affect agronomic traits in japonica rice. Euphytica 135: 129–134. Thaler, J.S., Fidantsef, A.L., Duffey, S.S., and Bostock, R.M. (1999). Trade-offs in plant defense against pathogens and herbivores: A field demonstration of chemical elicitors of induced resistance. J. Chem. Ecol. 25: 1597–1609. Wade, M.J., and Goodnight, C.J. (2006). Cyto-nuclear epistasis: Two-locus random genetic drift in hermaphroditic and dioecious species. Evolution 60: 643–659. 17 of 17 Wang, S., Basten, C.J., and Zeng, Z.-B. (2006). Windows QTL Cartographer 2.5. (Raleigh, NC: North Carolina State University). Wentzell, A.M., Rowe, H.C., Hansen, B.G., Ticconi, C., Halkier, B.A., and Kliebenstein, D.J. (2007). Linking metabolic QTLs with network and cis-eQTLs controlling biosynthetic pathways. PLoS Genet. 3: 1687–1701. West, M.A.L., Kim, K., Kliebenstein, D.J., van Leeuwen, H., Michelmore, R.W., Doerge, R.W., and St Clair, D.A. (2007). Global eQTL mapping reveals the complex genetic architecture of transcript-level variation in Arabidopsis. Genetics 175: 1441–1450. Wolf, J.B. (2009). Cytonuclear interactions can favor the evolution of genomic imprinting. Evolution 63: 1364–1371. Wright, S. (1931). Evolution in Mendelian populations. Genetics 16: 97–159. Wright, S. (1968). Evolution and the Genetics of Populations. (Chicago: University of Chicago Press). Zeng, Z.-B., Kao, C.-H., and Basten, C.J. (1999). Estimating the genetic architecture of quantitative traits. Genet. Res. 74: 279–289. Zeyl, C., Andreson, B., and Weninck, E. (2005). Nuclear-mitochondrial epistasis for fitness in Saccharomyces cerevisiae. Evolution 59: 910–914. Zhang, X., Hause, R.J.J., and Borevitz, J.O. (2012). Natural genetic variation for growth and development revealed by high-throughput phenotyping in Arabidopsis thaliana. G3 (Bethesda) 2: 29–34. Zhang, Z.-Y., Ober, J.A., and Kliebenstein, D.J. (2006). The gene controlling the quantitative trait locus EPITHIOSPECIFIER MODIFIER1 alters glucosinolate hydrolysis and insect resistance in Arabidopsis. Plant Cell 18: 1524–1536. Züst, T., Heichinger, C., Grossniklaus, U., Harrington, R., Kliebenstein, D.J., and Turnbull, L.A. (2012). Natural enemies drive geographic variation in plant defenses. Science 338: 116–119. Züst, T., Joseph, B., Shimizu, K.K., Kliebenstein, D.J., and Turnbull, L.A. (2011). Using knockout mutants to reveal the growth costs of defensive traits. Proc. Biol. Sci. 278: 2598–2603. Hierarchical Nuclear and Cytoplasmic Genetic Architectures for Plant Growth and Defense within Arabidopsis Bindu Joseph, Jason A. Corwin, Tobias Züst, Baohua Li, Majid Iravani, Gabriela Schaepman-Strub, Lindsay A. Turnbull and Daniel J. Kliebenstein Plant Cell; originally published online June 7, 2013; DOI 10.1105/tpc.113.112615 This information is current as of June 17, 2017 Supplemental Data /content/suppl/2013/05/21/tpc.113.112615.DC1.html Permissions https://www.copyright.com/ccc/openurl.do?sid=pd_hw1532298X&issn=1532298X&WT.mc_id=pd_hw1532298X eTOCs Sign up for eTOCs at: http://www.plantcell.org/cgi/alerts/ctmain CiteTrack Alerts Sign up for CiteTrack Alerts at: http://www.plantcell.org/cgi/alerts/ctmain Subscription Information Subscription Information for The Plant Cell and Plant Physiology is available at: http://www.aspb.org/publications/subscriptions.cfm © American Society of Plant Biologists ADVANCING THE SCIENCE OF PLANT BIOLOGY