Genetic architecture controlling variation in grain carotenoid composition and concentrations in two maize populations

Genetic control of maize grain carotenoid profiles is coordinated through several loci distributed throughout three secondary metabolic pathways, most of which exhibit additive, and more importantly, pleiotropic effects. The genetic basis for the variation in maize grain carotenoid concentrations was investigated in two F2:3 populations, DEexp × CI7 and A619 × SC55, derived from high total carotenoid and high β-carotene inbred lines. A comparison of grain carotenoid concentrations from population DEexp × CI7 grown in different environments revealed significantly higher concentrations and greater trait variation in samples harvested from a subtropical environment relative to those from a temperate environment. Genotype by environment interactions was significant for most carotenoid traits. Using phenotypic data in additive, environment-specific genetic models, quantitative trait loci (QTL) were identified for absolute and derived carotenoid traits in each population, including those specific to the isomerization of β-carotene. A multivariate approach for these correlated traits was taken, using carotenoid trait principal components (PCs) that jointly accounted for 97 % or more of trait variation. Component loadings for carotenoid PCs were interpreted in the context of known substrate-product relationships within the carotenoid pathway. Importantly, QTL for univariate and multivariate traits were found to cluster in close proximity to map locations of loci involved in methyl-erythritol, isoprenoid and carotenoid metabolism. Several of these genes, including lycopene epsilon cyclase, carotenoid cleavage dioxygenase1 and beta-carotene hydroxylase, were mapped in the segregating populations. These loci exhibited pleiotropic effects on α-branch carotenoids, total carotenoid profile and β-branch carotenoids, respectively. Our results confirm that several QTL are involved in the modification of carotenoid profiles, and suggest genetic targets that could be used for the improvement of total carotenoid and β-carotene in future breeding populations.


Introduction
Selection of yellow grain varieties is noted as a hallmark of maize domestication (Doebley et al. 2006). Yellow pigmentation is attributed to an accumulation of carotenoids in the endosperm resulting from a gain of function mutation in the first, rate-limiting enzyme in the carotenoid pathway, phytoene synthase (y1, psy1) (Palaisa et al. 2004) (Fig. 1). In the 1930s, the discovery of increased nutrition in yellow maize grain (Mangelsdorf and Fraps 1931) led to selection of pigmented grain as a desirable quality trait for both human food and animal feed (Bauernfeind et al. 1981;Weber 1987). It is now widely known that plant-based carotenoids, the source of this yellow pigment in maize grain, can provide dietary vitamin precursors and antioxidants that are essential to human health.
Provitamin A (proVA) carotenoids form the retinyl structural base for vitamin A molecules through cleavage reactions specific to animal metabolism (Fierce et al. 2008;Leuenberger et al. 2001). ProVA carotenoids can be found in many plant-based foods and include β-carotene (2 retinyl groups), β-cryptoxanthin (1 retinyl group) and α-carotene (1 retinyl group). Lutein and zeaxanthin, dihydroxy-xanthophyll conversion products of these proVA carotenoids in plants, are also prevalent in vegetative and seed tissues and are an abundant source of antioxidants. ProVA carotenoids provide vitamin A activity which is involved in immune function, protection of vision systems, and cellular differentiation (Stephensen 2001;Whitcher et al. 2001), and dihydroxy-xanthophylls are believed to aid in the prevention of macular degeneration and visual acuity (Chucair et al. 2007;Seddon et al. 1994), prompting the implementation of various nutritional interventions involving increased consumption of plant-based carotenoids. Consumption of several of these plant metabolites is associated with research outcomes indicative of health improvement (Bouis et al. 2011;Li et al. 2010;Tang et al. 2009). Therefore, from a health and nutritional standpoint both the quantity and quality of carotenoid profiles are important when plantbased foods are being used to satisfy dietary requirements.
Maize grain carotenoid concentrations are among the highest produced in cereals (Howitt and Pogson 2006), and exhibit considerable diversity in the composition of grain carotenoid profiles with respect to the predominant carotenoids (lutein and zeaxanthin), proVA carotenoids (α-carotene, β-carotene and β-cryptoxanthin) and other non-proVA carotenoids (zeinoxanthin) (Harjes et al. 2008).
Genetic analyses conducted with populations segregating in seed color (Chandler et al. 2013;Islam 2004) and carotenoid profiles (Chander et al. 2008;Stevens 2007;Wong et al. 2004) have demonstrated that compositional and concentration differences in seed carotenoids are quantitatively inherited. Biochemical characterization of various maize endosperm color mutants originally classified by Robertson (1975) has helped to identify various maize-specific homologs of carotenoid pathway genes reported in bacteria or model crop species (Fig. 1). Interestingly, concomitant decreases in lutein and zeaxanthin are observed when selection for increased ProVA is imposed with crtRB1 (Yan et al. 2010; data not shown), indicating that the accumulation of carotenoid metabolites can be competitive due to a related biochemical origin or to an altered flux through the pathway.
Genetic mapping studies have also indicated that a significant number of the QTL involved with carotenoid accumulation in grain exhibit pleiotropic effects on multiple carotenoid traits (Harjes et al. 2008;Yan et al. 2010). This suggests that these QTL could control the substrateproduct conversion of multiple metabolites in the pathway, therein revealing potential pathway control points. Reactions involved in glycolysis and isoprenoid biosynthesis appear to have a large impact on the magnitude of carotenoid accumulation in a manner similar to PSY1 (Estevez et al. 2001;Lois et al. 2000;Sandmann et al. 2006). Carotenoid production occurs within plastid organelles, in close proximity to the site of glycolysis. As depicted in Fig. 1, three-carbon molecules from glycolysis are committed to the methyl-erythritol phosphate (MEP) pathway through deoxyxylulose synthase (DXS) and deoxyxylulose reductase (DXR) (Lichtenthaler et al. 1997). The MEP pathway synthesizes five-carbon isoprenoids and subsequently produces isoprene isomers to form a common substrate pool shared by secondary metabolite pathways synthesizing plant sterols, terpenes, gibberellins, and carotenoids (Estevez et al. 2001).
Naturally occurring allelic variation for some of these pathway steps has been shown to affect variation in carotenoids in a wide variety of maize germplasm. Considering recent advances in our understanding of metabolic pathway reactions potentially involved with variation in carotenoid concentration and composition, it is likely that genes involved upstream of carotenoid synthesis will be prime candidates for the regulation of multiple carotenoid traits, while genes involved within carotenoid synthesis will be candidates for compositional differences. In this study, we investigate the genetic architecture underlying carotenoid concentration and compositional differences in F 2:3 maize populations A619 × SC55 and DEexp × CI7, two genetic populations derived from high total carotenoid and high β-carotene parent inbreds. We evaluate carotenoid trait profiles and identify QTL across temperate and subtropical environments, map genes of the carotenoid biosynthesis pathway relative to these QTL, and employ a multivariate approach to QTL mapping of carotenoid traits as a possible method to identify major regulatory control points for carotenoid accumulation.

Genetic materials
Two mapping populations were developed from maize inbred lines selected for superior and complementary carotenoid composition and total carotenoid profiles. This information is based on the multi-year surveys of the Goodman-Buckler Diversity Panel (Harjes et al. 2008).
The first population consists of 103 F 2:3 progeny derived from the cross of DEexp × CI7. DEexp was developed by Dr. Jim Hawk at the University of Delaware by selfing from a southeastern U.S. region Pioneer Hi-bred F 1 hybrid. CI7, developed by the USDA-ARS, is derived from the backcross L317/33-16//L317. Both lines yield high seed β-carotene concentrations, and also have high total carotenoid concentrations relative to lines screened in the Goodman-Buckler Diversity Panel (Supplemental Figure 1). Specific combining effects for β-carotene have been observed in the DEexp × CI7 hybrid, including highparent heterosis for β-carotene in some, but not all, years (T. Rocheford, data not shown).
The second population consists of 227 F 2:3 progeny derived from the cross of A619 × SC55. A619 is an early maturing line developed in Minnesota, derived from the backcross A171/OH43//OH43, and exhibits a high total carotenoid profile relative to most lines in the Goodman Diversity Panels in 2003 and2005 (top 25 and 20 % respectively, Supplemental Figure 1). SC55 was developed in South Carolina and is characterized by proportionally high β-carotene in grain. Favorable specific combining ability for high β-carotene concentration was observed in hybrid seed of the A619 × SC55 cross (Stevens 2007) and other materials (Egesel et al. 2003).
The above F 1 crosses were self-pollinated to create the F 2 generation. The F 2 seed from a single ear for each cross was planted, and plants were hand self-pollinated to create ears of F 2:3 derived seed.

Field evaluation
In 2005, the 103 F 2:3 lines of DEexp × CI7 and 227 F 2:3 lines of A619 × SC55 were grown in two replicates in each of two environments, the University of Illinois Urbana-Champaign and the International Maize and Wheat Improvement Center (CIMMYT), El Batan, Mexico (hereafter noted as Illinois environment and Mexico environment, respectively), using alpha (0, 1) incomplete block designs. The families were planted in single row plots of 5 m length, with 76 cm spacing between rows. Each plot was thinned to a density of approximately 15 plants per 5 m, or 43,000 plants ha −1 . Seven to nine plants were sib-pollinated within each row. Growing seasons varied by environment as follows: Urbana, IL, May-October; El Batan, Mexico, May-November. Fluctuation in average monthly temperature and precipitation recorded at both locations in 2005 is shown in Supplemental Figure 2.
After storage at room temperature for approximately 4 months, seed was bulked for each plot at shelling and an aliquot of approximately 10 g of seed from each plot was stored at −80 °C until carotenoid extraction could be performed on a plot mean basis. Due to field technical complications, the Illinois replicates of population A619 × SC55 were not included this study.
DNA extraction and genotyping DNA extraction was performed as described by Mikkilineni and Rocheford (2003). DNA samples prepared from F 2:3 family seed bulks were used for genotyping. A parental survey using 748 publicly available microsatellite (SSR) markers from MaizeGDB was performed. Pioneer Hi-Bred performed in-kind genotyping on both DEexp × CI7 and A619 × SC55 populations using polymorphic markers identified in the parental survey in addition to several proprietary markers, designated by the prefix "pio". Additional SSR markers genotyped at the University of Illinois were added to the genetic map. Allele-specific markers lcy -MZA, crtRB1-InDel4, and ccd1-5p were used to map three genes implicated in carotenoid synthesis or degradation: lycopene epsilon cyclase (lcy; NM_001153368.1), beta-carotene hydroxylase (crtRB1; NM_001112437), and carotenoid cleavage dioxygenase 1 (ccd1; DQ100346), respectively. Information for functional marker assays designed for this study (lcy and ccd1) and those derived from Yan et al. (2010) (crtRB1) is listed in Supplemental Table 1.

Phenotypic data collection
Carotenoid extraction from finely ground maize kernels was performed at Iowa State University as previously described (Li et al. 2007b). Briefly, the method involves mild heat treatment (50 °C × 15 min) in methanol followed by extraction with tetrahydrofuran and a 5-min saponification with methanolic potassium hydroxide at room temperature. After washing with water, the carotenoids are partitioned into hexane/methylene chloride (5:1 v/v), which is then evaporated to dryness under vacuum. The dried extract is reconstituted in methyl-tert-butyl ether followed by methanol and an aliquot is injected into the HPLC system. An internal standard, β-apo-8′-carotenal, was added during the extraction. Carotenoid separation was performed on a C-30 YMC Carotenoid Column, with gradient separation (methanol/methyl-tert-butyl ether) and photodiode array detection. Carotenoids were quantified using internal standard calibration curves. Measured carotenoid traits included lutein, zeaxanthin, zeinoxanthin, β-cryptoxanthin, α-carotene, trans and cis isomers of β-carotene, phytoene and trans-phytofluene. The summation of all carotenoids with absorption spectra in the visible range (lutein, zeaxanthin, zeinoxanthin, β-cryptoxanthin, α-carotene, total β-carotene) is termed total colored carotenoid in subsequent analyses. Isomers of β-carotene measured in this study include all-trans, 9-cis, 13-cis, and 15-cis β-carotene, and are represented by concentration (μg β-carotene g −1 DW) and proportion (ratio of isomer to total β-carotene). The ratio of colorless to colored carotenoids is defined by the proportion of phytoene plus phytofluene to total colored carotenoid. Proportion of alpha branch carotenoids (lutein, zeinoxanthin, α-carotene) to beta branch carotenoids (zeaxanthin, β-cryptoxanthin, β-carotene) is defined by α:β branch. Provitamin A was calculated according to the number of retinyl groups associated with all measured carotenoids, resulting in the expression ProVA = β-carotene + 0.5*β-cryptoxanthin + 0.5*α-carotene. All other ratios are comprised of absolute concentrations.

Genetic map construction
Genetic maps for both populations were constructed using JoinMap ® Version 3 (Ooijen and Voorrips 2001). Genotype data were screened for segregation distortion; markers with genotype segregation significantly different from the predicted 1:2:1 Mendelian ratio were removed. Maps were created from individual linkage groups using a LOD threshold of P < 0.001 and recombination frequency of 0.49 using Haldane's map function. The final linkage map for DEexp × CI7 consisted of 109 markers, spaced with an average interval length of 15 cM over a total of 1,486 cM. The final linkage map for A619 × SC55 consisted of 117 markers, spaced with an average interval length of 16.2 cM over a total of 1,728.6 cM. Genetic maps for both populations are shown in Supplemental Figure 3.

Phenotypic data analyses
All analyses were carried out using statistical procedures in SAS version 9.2 (SAS Institute 2008). Trait means and ranges by environment and for combined environments were obtained through Proc MEANS. Product-moment correlations of raw trait data were performed using Proc CORR. Analysis of variance for trait data was modeled according to y ijkl = µ + α i + β j(i) + δ k(ij) + γ l + (αγ ) il + e ijkl where y is the carotenoid phenotype (μg g −1 ) of an individual in the ith environment, jth replicate, kth block and lth genotype, μ is the population mean, α is the effect of the ith environment, β is the effect of the jth replication in the ith environment, δ is the effect of the kth block in jth replication of the ith environment, γ is the effect of the lth genotype (or family), (αγ) is the interaction effect of the ith environment and the lth genotype, and ε is the experimental error. The analyses were performed with Proc MIXED and all model effects were considered random. Variance component estimates for genetic variance (σ 2 g ), genotype by environment interaction variance (σ 2 ge ), and error variance (σ 2 ) were performed using Proc VARCOMP. Broad-sense heritability estimates were calculated on an entry mean basis as described by Hallauer and Miranda (1988). Best linear unbiased predictors (BLUPs) of F 2:3 family means for all traits were calculated in each population by environment for further use in principal component and linkage analyses. Residuals were checked for normality and independence prior to the calculation of trait BLUPs. Trait data were modeled according to: y ijkl = µ + β j + γ l + e jl , where β is the effect of the jth replication, γ is the effect of lth genotype (or family), and ε is the experimental error. Block effects were not included as they were not found to significantly account for phenotypic variation when included in models.

Linkage analysis by composite interval mapping
Quantitative trait locus analysis was performed using composite interval mapping (CIM) with PLABQTL software (Utz and Melchinger 1993). This mapping methodology is based on the Haley-Knott regression method for marker intervals (Haley and Knott 1992), and is supplemented by cofactor selection (Jansen and Stam 1994;Zeng 1994). Univariate (single trait BLUP values) and multivariate (PC score) traits were mapped according to the following model: y j = a + B i x ij + k� =i g k z kj where y is the trait value of family j (μg g −1 ), a is the intercept, B i is the genetic effect of the QTL located within the interval between marker i and marker i + 1, x i is a coded variable indicating the genotype of the marker interval (where maximally the value of parent 1 is −1, and parent 2 is 1), k� =i g k z kj is the summation of selected cofactor effects where g k is the partial regression coefficient of the trait value on marker cofactor k and z kj is a coded variable for the genotype at cofactor k.
A LOD threshold equivalent to an experiment-wise Type I error rate of α = 0.25 was used as the initial model selection criterion (Lander and Botstein 1989) for both populations; to obtain the threshold, 1,000 permutations of the trait data were run using a model of specified gene action, cofactor selection (Doerge and Churchill 1996), and a 2 cM scan. This corresponded to a threshold of LOD 3.5 for DEexp × CI7 analyses and LOD 2.9 for A619 × SC55 analyses. Regression models accounting for digenic additive by additive interactions or dominance effects were tested using the marker set selected by stepwise regression. Model fit criteria including Akaike's Information Criterion (AIC) and adjusted R 2 were used to select models accounting for greatest phenotypic variation with fewest estimated parameters. For most traits in both populations, an additive model produced an adequate fit to the data.
Adjusted R 2 values from the final multiple regression model express the proportion of phenotypic variation adjusted by the number of estimated model parameters. Additive effects are noted for each QTL, and reflect the partial regression coefficient of the marker genotype on the phenotype. These effects are expressed as a deviation from the second inbred parents, which in this study are CI7 and SC55. Effects associated specifically with the crtRB1 locus have been previously noted in population DEexp × CI7 by Yan et al. (2010).
The environment effect was significant (P < 0.05) for all traits in the DEexp × CI7 F 2:3 population except α-carotene. Carotenoid concentrations were larger in the Mexico than in the Illinois environment, with the exception of colorless carotenoids (phytoene and phytofluene) which were smaller in Mexico (  with trait distributions from the A619 × SC55 F 2:3 population in the Mexico environment, in which distributions for nearly all carotenoids were skewed toward the direction of the SC55 parent profile, and transgressive segregation was observed for all carotenoids at the higher end of trait distributions. Across both populations, cis isomers of β-carotene constituted a much smaller proportion of total β-carotene relative to the all-trans isomer (50-80 % of total), with relative proportions of: 9-cis, 10-25 %; 13-cis, 12-20 %; 15-cis, 2-6 % (Supplemental Fig 5).

Correlation and principal component analyses for carotenoid concentrations
Pairwise correlations among carotenoid traits were positive and significant (α = 0.05) among all colored carotenoids for the DEexp × CI7 F 2:3 families (Table 1), except for β-carotene and lutein (r = −0.32). Small but significant negative correlations were observed between colorless carotenoids (phytoene and phytofluene) and several colored carotenoids. Correlations among carotenoid concentrations were positive among nearly all colored and colorless carotenoids, and no traits were negatively correlated in the A619 × SC55 population (Table 1).
Four PCs explained the variation in the carotenoid trait matrix for A619 × SC55, with PC1 and PC2 accounting for 49 % (λ = 5.92) and 39 % (λ = 4.73) of the trait variance in the Mexico environment, respectively, through substantial loadings in lutein, zeaxanthin and phytoene. PC3 and PC4 accounted for 7 and 3 % of the variation, respectively, and had substantial loadings for lutein, zeaxanthin and β-carotene. Eigenvectors for the identified A619 × SC55 PCs demonstrated both synergistic and antagonistic effects through positive and negative loadings.
Loadings for PC traits were interpreted in the context of the biochemical pathway. In DEexp × CI7, most of the pathway variations were attributed to lutein, whereas smaller trait correlations appeared to be involved in apportioning carotenoid branch precursors to β-carotene (PC2) or could be involved in competition between phytoene precursors and downstream carotenoids (PC3). A619 × SC55 PCs were interpreted as a description of synergistic fluxes (PC1) or competitive allocation (PC2) of phytoene substrate to the largest carotenoid pools (lutein and zeaxanthin). PC3 described a competition between lutein and β-carotene versus zeaxanthin, whereas PC4 described a competition between α and β branch carotenoids.
Although DEexp and CI7 carotenoid profiles derived from the Mexico environment were very similar (Fig. 2), significant QTL were detected for several traits in the corresponding F 2:3 progeny (Table 3, Supplemental Table 2). Variation in concentrations and trait ratios for colorless carotenoids was attributed to 4-8 main effect QTL which predominantly were located on chromosomes 3 and 4. A larger proportion of the favorable additive effects at these QTL were contributed by DEexp relative to CI7. Eight QTL were detected for β-carotene, several of which individually accounted for more than 10 % of the variation for this trait; most of the positive additive effects for these QTL were also derived from DEexp. Variation in lutein, the carotenoid differing most between inbreds DEexp and CI7, was explained by only two large QTL detected at chromosome 3 (pos. 70, positive CI7 effect of 1.19 μg g −1 ) and chromosome 8 (pos. 66, positive DEexp effect of 2.81 μg g −1 ). Epistatic effects were significant in only two derived traits, α:β branch ratio and colorless:colored carotenoid ratio. While it would be expected that metabolic QTL originating from the same pathway would yield more epistatic interactions, it is likely that the sample size of this population prevented detection of such effects.
Fewer carotenoid QTL were detected for the DEexp × CI7 in Illinois than described above for the Mexico environment. This was attributed to larger trait averages and variances in Mexico than Illinois (Fig. 2). A single, large effect QTL on chromosome 8 (pos. 52-54, R 2 part = 40.3) explained variation in lutein, and although this QTL did not map to the exact pos. as those accounting for variation in lutein in the Mexico environment (Mexico, pos. 66; Illinois pos. 52-54), the magnitude of additive effect and source of parental contribution of the lutein QTL were nearly identical across the two environments, suggesting that the same locus could be operating in both environments. Similarly, QTL for zeaxanthin, β-cryptoxanthin, total colored, colorless:colored ratio, βcar:βcry ratio, α:β branch ratio accounting for trait variation in the DEexp × CI7 population grown in Illinois were also detected in the Mexico environment, and had consistent parent of origin and genetic effects (Supplemental Table 3, Table 3). No loci were found to account for significant variation for α-carotene, β-carotene, phytoene or phytofluene (Supplemental Table 3) for DEexp × CI7 in the Illinois environment.
QTL mapping for PC1, which accounted for 90 % of the variance among the carotenoid trait matrix, revealed two major QTL at chromosomes 3 (pos. 70) and 8 (pos. 66) to account for 37.8 % of the total phenotypic variation at the Mexico environment (Table 5). These effects were similar to those found for lutein in the same environment, and  Table 3). Variation in β-carotene isomer concentrations and proportions mapped to 3-8 loci for all-trans, 9-cis and 13-cis β-carotene, accounting for 25-60.5 % of the phenotypic variation (Supplemental Table 4). No QTL were significantly associated with 15-cis β-carotene in this population.
Genetic effects accounting for variation in four multivariate traits (PCs) were mapped in A619 × SC55. QTL associated with variation in all PC traits consisted of QTL are shown for univariate (single carotenoid concentration) and multivariate (PC) traits mapping to locations of genes known to be in the carotenoid or precursor pathways (overlap delimited by black squares). Correlated trait effects are shown for all PC traits, where colored squares indicate the carotenoids which are best described by a given PC; the presence of only blue squares shows that these carotenoid variation described by a single PC covaries proportionally, whereas presence of blue and red squares indicates an inverse covariation between the carotenoids described by that PC (color figure online) models with 4-5 main effect QTL and few epistatic interactions (Table 5). Models for multivariate traits were found to account for more phenotypic variation (range 31.1-60.4 %) than the univariate carotenoid traits (21.6-56.2 %). Variation in PC1, consisting of significant and positive loadings for phytoene, lutein and zeaxanthin, was explained by five QTL with positive effects predominantly originating from the A619 genome. The QTL at chromosome 9 (pos. 166) with large, pleiotropic effects were observed to affect this PC trait. The PC2 trait, describing antagonism between colorless and colored carotenoids was explained by five QTL, including one observed to have pleiotropic effects on β-branch carotenoids in univariate analysis at chromosome 5 (pos. 76). The PC3 trait, which describes antagonistic variation between zeaxanthin with lutein and β-carotene, was affected by four QTL, one which mapped in close proximity to the crtRB1 locus on chromosome 10, and another which affected β-branch traits on chromosome 2 (pos. 128). Competition between α-and β-branches, described by PC4, was affected by five main effect QTL and two digenic interactions. Positive effects at all of these QTL were small, and were contributed by SC55 (Table 5). Univariate and multivariate traits overlapped extensively, as shown in Supplemental Figures 6 and 7. A range of 3-7 loci accounted for 8.2-36.5 % of the variation in β-carotene isomer concentration and proportion traits in population A619 × SC55 (Supplemental Table 6). As observed in population DEexp × CI7, several loci cosegregated with variation in more than one isomer concentration including: chromosome 5 (pos 72, 13-cis and all-trans), 10 (pos 92, CrtRB1 marker; 13-cis and all-trans). Loci with pleiotropic effects on β-carotene isomer proportion were identified on chromosomes 1 (pos 12; 9-cis, alltrans), 4 (pos 186; 9-cis, all-trans), 5 (pos 72; 9-cis, 15-cis, all-trans) and 6 (pos 34-38; 9-cis, 15-cis, all-trans).
Comparison of detected QTL to known biosynthesis genes A total of 63 main effect QTL for carotenoid univariate traits were found to span 29 marker intervals for the DEexp × CI7 F 2:3 population in the Mexico environment, in contrast to 22 unique main effects for the multivariate principal component traits. For the A619 × SC55 population, a total of 50 main effect QTL across 29 marker intervals were found to explain phenotypic variation in univariate traits, as compared to 19 main effect QTL for multivariate traits across 18 marker intervals. The map locations for univariate and PC traits often overlapped, but QTL that were unique to PC traits were detected in both populations (Supplemental Figure 6 and 7). Many of the detected QTL clusters were found within proximity of mapped carotenoid biosynthesis genes (Andorf et al. 2010; Supplemental Table 6). Although genetic maps for both populations did not provide sufficient resolution to definitively conclude that the known genes and detected QTL mapped within proximity to the same environment (as gauged by chromosome bin assignment), inferences for these gene-phenotype relationships were strengthened by comparing the affected traits with the known metabolic functions of previously identified genes, overlapping detection between univariate and multivariate traits, and detection in previous mapping studies (Fig. 3).
Overall substrate increases to the carotenoid pathway in population DEexp × CI7 were described by PC2, which mapped to chromosome bins also containing genes involved in carbon substrate allocation to the MEP pathway (DXR, bin 3.04) and isoprenoid substrate allocation to the carotenoid pathway (PSY1, bin 6.01). PC3 described the balance between colorless carotenoids and colored pathway intermediates, and mapped near genes involved with carotenoid conversion within the main branch (ZDS, bin 7.02) in addition to conversion of main branch carotenoids to α and β branch carotenoids (LCYβ, bin 5.04). PSY2, an isoform of phytoene synthase in bin 8.07, was also found to map near PC3 QTL. The largest trait variation was observed for lutein, and was represented by the PC1 trait. This PC trait mapped to chromosome bins containing DXR (bin 3.04) and LCYε (bin 8.05), both of which are involved in carotenoid synthesis by committing carbon substrate to the carotenoid pathway, and to the α-branch, respectively.
Multivariate traits for A619 × SC55 also described overall substrate increases to the pathway through PC1 and allocation between colorless substrate and colored intermediates through PC2. These traits mapped close to genes that are involved in the allocation of substrate precursors to the carotenoid pathway (HDS, bin 5.03; GGPPS1, bin 7.04) as well as one that removes carotenoids from the pathway through degradation (CCD1/wc1, bin 9.07). Balance between the accumulation of β-carotene and zeaxanthin, explained by PC3, had associated QTL mapping proximal to genes directly involved in zeaxanthin metabolism including crtRB1(bin 10.06) and ZEP1(bin 2.04) as well as carotenoid conversion by CrtISO1 (bin 4.08). PC4 described competition between the α-and β-pathway branches; although QTL for this trait were found to map within the same chromosome bins containing genes allocating substrate to the entire pathway (GGPPS1, bin 2.08; DXS1, bin 6.05).

Inferences to metabolic functions for detected QTL
The genetic architecture for high β-carotene and total carotenoids as well as various individual and compositional traits was investigated in two F 2:3 populations, DEexp × CI7 and A619 × SC55. Through map location comparisons of the detected QTL with those reported for genes of the MEP, isoprenoid, and carotenoid pathways, we concluded that a large proportion of the effects detected in this study are likely derived from metabolic QTL. Confidence in the association of known biosynthesis genes with detected QTL was highest when the trait affected by the QTL was known to be a product or substrate of the associated biosynthesis gene (Fig. 3). Major effect QTL previously identified in other genetic mapping studies also support the association between variation in specific carotenoid traits with locations of putative metabolic QTL (Fig. 3;Islam 2004;Wong et al. 2004;Chander et al. 2008;Harjes et al. 2008). Metabolic function has been confirmed by association analyses for several of these loci including lcy (Harjes et al. 2008), crtRB1 (Yan et al. 2010) and CCD1 (Kandianis and Rocheford, data not shown). We emphasize, however, that inferences should be confirmed through additional genetic analyses utilizing functional markers that are in complete and near complete linkage disequilibrium with the alleles associated with trait variation (Andersen and Lubberstedt 2003), and/or through the use of genetic mapping with high density, high throughput markers.
Use of trait correlations in QTL analyses has been reported to result in increased power of QTL resolution, increased precision of parameter estimates, and the simplification of simultaneous hypothesis testing for multiple traits (Gilbert and Le Roy 2004). Multitrait mapping approaches using multivariate techniques (Korol et al. 1995) or transformation of correlated traits to canonical variables which are then used univariately (Mangin et al. 1998) have also been successfully employed. As information describing the size and system of effects from pleiotropic QTL could be lost through the estimation of single trait effects (Jiang and Zeng 1995), there was reason to attempt QTL discovery in this system of correlated traits by pairing univariate and multivariate analyses. Principal component (PC) traits identified using the variance-covariance matrix for eight carotenoids were mapped by composite interval analysis, leading to the detection of 2-8 QTL for each principal component. Using this strategy, we were able to identify loci simultaneously affecting multiple univariate carotenoid traits, which was suggestive of a pleiotropic QTL. As such, functional associations for the detected suite of QTL suggest that carotenoid concentrations are heavily influenced by (1) the allocation of carbon substrate to the carotenoid pathway, and (2) removal of carotenoids from the pathway through CCD1-facilitated degradation or ZEP1-mediated conversion. In contrast, β-carotene concentrations are largely attributed to (1) the allocation of carotenoid substrate from the main branch supplied by phytoene through the bifurcation point at lycopene, (2) competition between α-and β-branches and (3) conversion within the β-branch.
Cross population comparisons highlight differences in some of the genetic variation controlling total colored carotenoid concentrations. Population DEexp × CI7 was influenced by the effects of genetic variation in metabolic steps providing substrate precursors to the carotenoid pathway, including those within the MEP pathway associated with DXS and DXR (Estevez et al. 2001;Rodriguez-Concepcion 2006), as well as reactions that alter allocation within the carotenoid pathway including LCYε. PSY1, encoding phytoene synthase which represents the first committed step of the carotenoid pathway (Buckner et al. 1996), mapped near a QTL accounting for variation in phytoene, which provides the precursor substrate of the major colored carotenoids in grain. A large interaction between chromosomes 2 and 3 affecting the total carotenoid trait may have originated from precursor synthesis at DXR and removal of zeaxanthin through ZEP ). These single trait QTL were confirmed in multitrait analyses, in which PC traits detected genetic variation describing control of overall flux to the pathway (as with DXS and DXR), or allocation of substrate to competing pathway branches (as with LCYε and ZEP). Phenotypic variation for total colored carotenoids within the A619 × SC55 population, on the other hand, was dominated by a large effect proximal to the CCD1 locus. Considering the origin of the deleterious effect came from the SC55 allele, this may explain why skewed trait distributions were observed for nearly all carotenoids in the F 2:3 progeny of this cross. Interestingly, multitrait analyses for this population highlighted the potential importance of variation in precursor pathways as compared to minimal detection of QTL associated with biosynthesis pathways upstream of the carotenoid pathway in single trait analyses.
Future testing of biosynthesis genes hypothesized to underlie the QTL detected in this (and other) study will likely conclude that few non-metabolic QTL also affect carotenoid trait variation. Some examples of non-metabolic effects modifying carotenoid concentration and composition have been noted in other plant species; for example, reports of mutations in DNA binding domain proteins in cauliflower (Lopez et al. 2008) and plastid development in tomato (Galpaz et al. 2008) demonstrate that regulatory and developmental factors can also control carotenoid biochemistry in developing curd and fruit, respectively. With available maize genomic sequence, it should be possible to search for maize homologs of these genes for further hypothesis testing.
Modifying carotenoid profiles to address ProVA bioavailability Much of the allure in modifying metabolic flux through the carotenoid biosynthesis pathway in various crops has been driven by the potential to form provitamin A, and particularly β-carotene (Harjes et al. 2008;Welsch et al. 2010;Yan et al. 2010;Ye and Beyer 2000;Babu et al. 2013). As not all proVA carotenoids are equally absorbed and converted to vitamin A (Tanumihardjo et al. 2010;Davis et al. 2008), proVA bioavailability has emerged as a nutritionally valued trait of similar importance to total proVA concentration.
This study describes natural variation in the β-carotene isomers of maize grain, proVA carotenoids which have been shown to differ in bioavailability. β-carotene in foods is often reported as the compilation of one or more β-carotene isomers, including the most common all-trans form, and the less prevalent but more oil-soluble 9-, 13and 15-cis isomers (Schieber and Carle 2005). Cis-isomerization of β-carotene is noted to be an artifact of food processing (Achir et al. 2011;Mordi et al. 1993), but also naturally occurs to varying degrees in foods (Khoo et al. 2011). The isomeric state of dietary β-carotene is nutritionally important as its conformation considerably influences its bioavailability. Trans isomers of β-carotene are preferentially absorbed (Deming et al. 2002;Stahl et al. 1995). Once absorbed, the 9-cis isomer of β-carotene is primarily converted to all-trans forms of vitamin A (Maeda et al. 2011). Regarding provitamin A activities, however, 13-cis and 9-cis β-carotene only maintain 53 and 38 % of the activity conferred by trans β-carotene, respectively (Schieber and Carle 2005).
Results from this study reveal a genetic basis for β-carotene isomerization in maize grain, specifically for traits of concentration and proportion of 9-cis, 13-cis, 15-cis and all-trans β-carotene. As of yet, there is no evidence for enzymatic conversion of trans β-carotene or other carotenoid precursors to cis β-carotene. Considering that carotenoids are antioxidants, it is possible that isomerization of trans β-carotene under light, heat or oxidative stress could give rise to sizable amounts of cis β-carotene. It is also possible that β-carotene isomerization could indeed be facilitated by an isomerase, similar to those which facilitate isomerization of pro-lycopene carotenoids (Chen et al. 2010;Isaacson et al. 2004). This implies that genetic factors influencing cis β-carotene accumulation could include: (1) metabolic QTL affecting synthesis or degradation of trans β-carotene; (2) as of yet unknown metabolic QTL specific to the formation of cis β-carotene and; (3) factors conferring molecular stability of trans β-carotene which could influence the degree to which non-enzymatic trans-cis isomerization occurs, perhaps with changing cellular conditions. Accordingly, we found several QTL with pleiotropic effects for more than one isomer trait to colocalize with known metabolic candidates including ZEP1 (Chr. 2), LCYβ (Chr. 5), LCYε (Chr. 8), and CrtRB1 (Chr. 10).

Recommendations for future carotenoid breeding studies
The QTL results presented here are consistent with the concept that modification of grain carotenoid profiles primarily originates from variation in the carotenoid biosynthesis pathway, or metabolism that directly creates and allocates substrate to the carotenoid pathway. Association mapping of sequence variation in candidate genes from the carotenoid biosynthesis pathway has revealed that carotenoid composition is significantly altered by genes regulating α-and β-branch allocation through lcyε and β-branch conversion through crtRB1. From the associations of mapped QTL effects with map locations of known biosynthesis genes, there is reason to expect that increased total carotenoids can be achieved through altered metabolism, and that biochemical reactions upstream of the carotenoid pathway must be targeted to attain this increase. Characterizing allelic variation representative of MEP and isoprenoid pathways through loci such as DXR should be highly informative and useful in association analysis tests and markerassisted selection studies for total carotenoid grain profiles. Some of the genes in these pathways were detected by PCs representing significant and positive loadings of the major colored carotenoids and phytoene; however, in the case where phytoene and colored carotenoids negatively covary, it may be necessary to include measurement of colorless carotenoids (phytoene and phytofluene) for further genetic analyses.
Factors contributing to efflux from the carotenoid pathway also appear to affect total carotenoid accumulation, through select carotenoids on a specific pathway branch (ZEP) or through non-specific reduction of nearly all carotenoids (CCD1). Compositional differences have been achieved through modification of intra-pathway steps including cyclization reactions (LCYε, LCYβ) and hydroxylation reactions (CRTRB1). Given that most of the phenotypic variation for grain carotenoids in maize exists in lutein, zeaxanthin, zeinoxanthin, β-cryptoxanthin and β-carotene, the steps identified here primarily affect these intermediates and will likely account for most of the phenotypic variation upon selection of the desired alleles. Marker-assisted selection strategies for β-carotene and provitamin A concentrations are already employing variation at lcyε and crtRB1 to attain desired carotenoid profiles. Given the results of this study, there is merit to further investigate the ZEP locus and lcyβ for compositional modification, and to evaluate the influence of environment on the metabolic activity of these pathway enzymes.