Natural Variation of Plant Metabolism: Genetic Mechanisms, Interpretive Caveats, and Evolutionary and Mechanistic Insights

Ecological, evolutionary, and mechanistic insights arise from the study of natural variation in plant metabolism.

Abstract

Combining quantitative genetics studies with metabolomics/metabolic profiling platforms, genomics, and transcriptomics is creating significant progress in identifying the causal genes controlling natural variation in metabolite accumulations and profiles. In this review, we discuss key mechanistic and evolutionary insights that are arising from these studies. This includes the potential role of transport and other processes in leading to a separation of the site of mechanistic causation and metabolic consequence. A reilluminated observation is the potential for genomic variation in the organelle to alter phenotypic variation alone and in epistatic interaction with the nuclear genetic variation. These studies are also highlighting new aspects of metabolic pleiotropy both in terms of the breadth of loci altering metabolic variation as well as the potential for broader effects on plant defense regulation of the metabolic variation than has previously been predicted. We also illustrate caveats that can be overlooked when translating quantitative genetics descriptors such as heritability and per-locus r² to mechanistic or evolutionary interpretations.

The study of quantitative genetics and ecology and evolution in plants has a long history of reliance on the natural variation of metabolic traits. One of the first identified quantitative trait loci (QTLs) in any organism was for the metabolic control of seed color in Phaseolus vulgaris (Sax, 1923). This analysis helped to develop and empirically test some of the foundations of quantitative genetics. Similarly, a significant fraction of ecology and evolutionary theory has focused on the pressures leading to the diversification of plant metabolism and the contravening costs on these defenses (Ehrlich and Raven, 1964; Karban and Baldwin, 1997). These studies have used measurements of metabolite variation to make significant conceptual progress in understanding the underlying pressures without access to the underlying causal genes (Strauss and Agrawal, 1999; Agrawal, 2011; Cook-Patton et al., 2011).

Recent advances in genomics and metabolic profiling have opened new opportunities to study the natural variation of metabolic traits. These include the advent of rapid metabolomic platforms allowing the quantification of hundreds to thousands of metabolites in as many different genotypes (Fiehn, 2001; Meyer et al., 2007; Fiehn et al., 2008). In combination with the ability to sequence and measure the transcriptome of all of these same lines, there is a massive influx of studies reporting on the identification of causal genes controlling the variation in metabolites in numerous species, from crop plants like maize (Zea mays), rice (Oryza sativa), and tomato (Solanum lycopersicum) to models like Arabidopsis (Arabidopsis thaliana) and ecological models like Boechera stricta and Nicotiana attenuata (Fu and Xue, 2010; Hartings et al., 2011; Li et al., 2011, 2015; Kausch et al., 2012; Prasad et al., 2012; Matsuba et al., 2013; Chang et al., 2015; Yan et al., 2015). These studies provide new insights into the mechanistic and evolutionary structures that influence how plant metabolism functions within a broader context. Other reviews have focused on the specific genes being cloned that control metabolite variation and the approaches utilized (Saito et al., 2008; Kliebenstein, 2009, 2012; Kusano et al., 2015; Luo, 2015; Omranian et al., 2015). As such, this review will focus on the broader insights being provided by these new studies into the genetic, mechanistic, and evolutionary processes shaping plant metabolism.

GENETIC ARCHITECTURE OF METABOLOMIC VARIATION

Quantitative genetic studies typically report several descriptors of the measured phenotypes and the candidate loci. These include the heritability of the phenotype, often as broad-sense heritability, a measure of genotypic reproducibility, and the r² of the individual locus linked to a given phenotype, often called the effect size (Lynch and Walsh, 1998). These values are often linked to general conclusions, such as metabolites have high heritability or secondary metabolite loci have higher r² than do primary metabolite loci. However, these values have significant caveats that need to be considered when interpreting the results, which may confuse any ability to make conclusions.

Heritability

Most plant metabolomics studies that focus on natural variation report broad-sense heritability, which is a measure of the phenotype reproducibility within a set of genotypes (Lynch and Walsh, 1998). These studies have shown that the heritability of metabolomic traits displays a wide range in any given population (Fig. 1). In general, however, the heritability of maize metabolic traits appears to be higher than that for rice, with apparently lower Arabidopsis heritabilities (Keurentjes et al., 2006; Rowe et al., 2008; Chan et al., 2010a, 2010b; Yang et al., 2010; Riedelsheimer et al., 2012; Gong et al., 2013; Joseph et al., 2013a, 2013b, 2015; Li et al., 2013; Lipka et al., 2013; Chen et al., 2014; Alseekh et al., 2015; Zhang et al., 2015). These results are similar when using either structured populations like nested association mapping, recombinant inbred line (RIL), or introgression line populations or unstructured genome-wide association (GWA) populations within the same species. While it is tempting to argue that different domestication and selection processes may be influencing the difference in heritability across species, the estimation of heritability is not an absolute value and is influenced by numerous experimental, technical, and quantitative factors (Lynch and Walsh, 1998). For example, the rice analyses exclude residual error variance in the calculation of heritability, while two of the Arabidopsis studies include residual error variance. Recalculating the variance in these two Arabidopsis studies shows that they actually have a highly similar distribution of metabolite heritabilities (Fig. 1). In addition to the calculation choices, there are also biological and experimental differences among the experiments that complicate the comparison. Among three studies in rice, the growth conditions, age at harvest, and metabolite quantification methods all differed (Gong et al., 2013; Chen et al., 2014; Matsuda et al., 2015). Experimental designs range from randomized complete block design (Keurentjes et al., 2008; Chan et al., 2011; Chen et al., 2014; Alseekh et al., 2015) to α-lattice incomplete block design (Riedelsheimer et al., 2012). Thus, it is currently unclear if comparisons of heritability among these studies provide biological insight or simply reflect the technical and experimental differences. Future experiments wherein all technical and experimental differences are controlled would be required to assess if there is any biological difference in heritability among the species, potentially driven by domestication.

Figure 1.

Distribution of broad-sense heritability estimates of metabolic traits across species and methods. Red shows studies involving S. lycopersicum × S. pennellii, green shows studies using Arabidopsis, blue is for maize studies, and peach is for studies involving rice. All maize and Arabidopsis comparisons are intraspecific. Rice studies are intraspecific (Gong et al., 2013) and across subspecies (Chen et al., 2014; Matsuda et al., 2015). Heritability in the Rowe et al. (2008) and Joseph et al. (2014) data sets is shown in their original form, and heritability has been recalculated using solely environmental and genetic variance as in the other studies; the new results are indicated by asterisks.

A series of experiments did estimate the heritability of metabolic, transcriptomic, and physiological traits using the same experimental design, genotypes, and calculations to allow for direct comparison across mechanistic levels. This showed that the heritability of metabolic phenotypes is intermediate between the higher heritability of transcripts and the lower heritability of integrative traits like growth (Keurentjes et al., 2006, 2007, 2008; West et al., 2007; Rowe et al., 2008; Fu et al., 2009; Chan et al., 2010a, 2010b; Joseph et al., 2013a, 2013b, 2015). This could suggest that metabolic genetic variance is, in fact, intermediate between transcripts and integrative traits or that the integrative traits are more responsive to environmental variation leading to lower heritability. A related explanation that combines these options is that vastly more quantitative loci control integrative traits like growth of which the loci for specific metabolic traits are a subset, which would lead to diminishing heritability estimates. This model agrees with work on growth and metabolite traits that found that focusing solely on the growth QTL underestimated the link between metabolic and growth variation (Joseph et al., 2013a, 2013b). Additional studies are required to understand why mechanistically linked traits have varying heritability in the same population.

Effect Size

While heritability is a trait-level descriptor, most studies also provide the estimated effect size of individual loci. This is usually provided as r² or the fraction of total phenotypic variance in a metabolite that is linked to a specific locus. This shows a wide spectrum of effect sizes, where metabolites can be under the control of few loci of large effect or numerous loci of small effect. Across Arabidopsis, rice, and maize, each locus explained on average 15% to 25% of metabolite variation (Fig. 2; Rowe et al., 2008; Ying et al., 2012; Gong et al., 2013; Chen et al., 2014; Dong et al., 2015; Matsuda et al., 2015). These averages, however, hide a wide range of individual locus variation. For example, individual rice loci have been found to explain 35% or more of an individual metabolite’s variation up to nearly 90% (Ying et al., 2012; Chen et al., 2014; Dong et al., 2015; Matsuda et al., 2015). In intraspecific studies, a single locus explained at most over 90% of metabolite variation in Arabidopsis (Rowe et al., 2008) and in rice (Gong et al., 2013) but at most only 62% in maize (Riedelsheimer et al., 2012; Fig. 2). In contrast, most loci found to control carbon and nitrogen metabolism were of small effect size in maize (Zhang et al., 2015). These studies show a wide range of effect sizes for loci linked to variation in metabolic traits.

Figure 2.

Estimates of r² of loci controlling metabolic variation compared across species and methods. Green shows studies using Arabidopsis, blue is for maize studies, and peach is for studies involving rice. All maize and Arabidopsis comparisons are intraspecific. The rice comparisons are intraspecific (Ying et al., 2012; Gong et al., 2013; Dong et al., 2015) and across subspecies (Chen et al., 2014).

The equivalence of r² effects per locus-to-locus effect, as commonly interpreted by mechanistic or molecular biology studies, is not straightforward. The r² of a locus is the variance attributed to that locus divided by the total variance. As such, the calculation of per-locus r² depends on factors that can affect the numerator (number of loci across which the variance is divided, missing loci, overestimates, etc.) as well the denominator (total variance, errors in total variance estimation, etc.). Thus, it is possible to have large-effect loci as defined by r² that have additive effects of 10% or less in metabolite accumulation when comparing the two alleles (Fig. 3). Similarly, if the metabolite shows a large range of variation, it is possible to have small-effect loci per r² that have additive effects of 100% difference in metabolite accumulation between the two alleles (Fig. 3). This leads to a value that is, at best, relative and of use only in that population and that can vary from differences in the number of loci identified (Beavis, 1994). The number of loci found per metabolite shows a wide range of variation across experiments and populations due to replication and statistical methods. In rice, studies have found a range of three to nine loci per metabolite, while in maize, this has ranged from five to 18 (Gong et al., 2013; Li et al., 2013; Chen et al., 2014; Dong et al., 2015; Matsuda et al., 2015; Zhang et al., 2015). While this difference will mathematically lead to the maize loci having smaller r², because there are more loci per metabolite to share the variance, it is not clear if this is a reflection of the biological reality of the genetics controlling metabolite variation in the two species or if there are significant differences in the false-negative error rates leading to fewer detected loci in rice (Joseph et al., 2014). Thus, while r² is a useful quantitative descriptor, it should be handled with care when making mechanistic arguments, as it does not directly scale to additive effects.

Figure 3.

Low correlation between r² and additive effect size estimates for metabolic loci. Intraspecific analysis of nontargeted metabolites in Arabidopsis is indicated by black open circles (r² = 0.006, P = 0.011), genetics of flavonoid abundance in a rice intraspecific comparison is indicated by gray circles (r² = 0.38, P < 2.2e⁻¹⁶), and lipid genetics studied across rice subspecies (japonica and indica) is indicated by black open triangles (r² = 0.43, P = 7.19e⁻⁰⁶). The inset shows a closer look at loci with low effect size (<0.5) and low r² (<0.25).

Clustering of Metabolite Loci

To summarize the identified loci, these studies frequently search for genomic hotspots that alter variation in more metabolites than expected by random chance. Every study identifies hotspots no matter what the species or population utilized, but the number and position of the hotspots can vary across tissues within a specific population, as found in rice (Gong et al., 2013). The position of hotspots can also differ depending upon the metabolite class being measured (Schauer et al., 2006, 2008; Riedelsheimer et al., 2013; Alseekh et al., 2015). The position and frequency of hotspots vary when using different populations from specific rice subspecies; in indica, hotspots were detected on chromosomes 2, 6, and 12, while japonica had hotspots on chromosomes 4 and 12 (Chen et al., 2014). Efforts to use these results to make mechanistic or evolutionary conclusions about the differences, such as arguing that the genetic architecture of metabolic variation is unique between the two rice subspecies, should proceed carefully for a couple of reasons. Different populations within Arabidopsis give different hotspots, which is likely solely associated with the different genes varying in that population regardless of whether the variation alters adaptation (Keurentjes et al., 2006; Rowe et al., 2008; Joseph et al., 2013a, 2013b). Using near-isogenic lines to assess the impact of a nonhotspot region in Arabidopsis showed that variation in this region affected more metabolites than any hotspot found using the associated RILs (Rowe et al., 2008). Thus, the difference between hotspot and nonhotspot regions could be driven by statistical rather than biological issues. As such, while hotspots may generate intriguing lines of research, any broad conclusions should wait for the cloning and characterization of the underlying causal gene(s).

Network Structure

Measuring metabolites across a set of natural genotypes allows investigations into metabolic network properties prior to conducting any locus-specific analysis. This can be done by correlating the variation in average metabolite abundance across the genotypes to look for genetic correlations. A key use of this correlation approach is to help address the significant difficulty presented by the fact that most metabolites measured have no known structure. Significant progress in identifying the structure of unknown metabolites is being made by querying for groups of metabolites that show a high genetic correlation under the assumption that they are chemically related. This approach has been used to identify sets of metabolites produced by the same biosynthetic pathway as found in rice, where using metabolite coaccumulation and structural similarity generated putative biosynthetic networks for amino acids and flavone-O-hexosides (Matsuda et al., 2012, 2015).

These approaches have been extended to other metabolites in both rice and maize to link previously unknown metabolites to each other and to new biosynthetic genes (Gong et al., 2013; Chen et al., 2014; Dong et al., 2014, 2015; Hu et al., 2014; Wen et al., 2014; Hashemi et al., 2015; Kusano et al., 2015; Luo, 2015). Similar approaches have been used to identify and expand pathways for unknown metabolite classes within nonmodel systems like N. attenuata, Barbarea vulgaris, and Capsicum spp. (Dalby-Brown et al., 2011; Wahyuni et al., 2014; Agerbirk et al., 2015; Li et al., 2015). While this approach is very powerful at identifying both new compound structures and new enzymes, it is solely reliant on the presence of genetic variation in the genes controlling the accumulation of these unknown compounds. As such, compounds with low or no genetic variation in their underlying genes will not be amenable to this approach. Fortunately, metabolites that play key roles in adaptation to biotic and abiotic stress are, by default, highly likely to be naturally variable due to both fluctuating and local variation in these selective pressures (Hancock et al., 2011; Züst et al., 2012; Brachi et al., 2015; Kerwin et al., 2015). As such, this approach may be uniquely powerful to identify and classify new metabolic pathways that play key roles in fitness in the wild.

In addition to classifying chemical pathways, it is also possible to use this approach to link metabolic traits with other yield or physiological traits (Lisec et al., 2008; Sulpice et al., 2009; Carreno-Quintero et al., 2012; Shen et al., 2013; Hill et al., 2015). Correlational analyses clustered metabolic traits into large modules, with a strong tendency toward positive correlations among metabolic and yield-associated traits in Solanum pennellii × S. lycopersicum introgression lines (Schauer et al., 2006). This grouping subdivided into three major modules: one containing whole-plant traits and metabolic intermediates, one containing all amino acids, and one including sugars and organic acids that was consistent across multiple years (Schauer et al., 2006, 2008). Similar networks were found using other tomato species (Do et al., 2010; Sauvage et al., 2014). This correlational structure was suggested to indicate network-level competition for photoassimilates that is consistent across tomato species (Schauer et al., 2006). The correlational approach has also been used to rapidly link metabolite variation to insect resistance in nonmodel systems (Kuzina et al., 2011). Thus, the use of genetic correlations has the ability to convey phenotypic insight about metabolite roles in planta.

It is also possible to use this approach to query for factors that alter metabolic network structure. For example, the correlational structure of metabolites in rice was significantly different when comparing the indica versus japonica subspecies, suggesting that the metabolic networks in these two subspecies are dissimilar, potentially in response to unique selective pressures during their separate domestication histories (Hu et al., 2014). In addition to network structure changing in response to long-term selective processes, the structure can change in response to short-term environmental perturbations. In Arabidopsis, the correlational structure shifted when the same population was harvested at different times of day (Chan et al., 2010a). In that case, there were more connections between metabolites at the dawn sampling than at the dusk sampling (Chan et al., 2010a). Thus, it is possible to learn general concepts about the structure of the metabolome using genetic correlations between metabolites even prior to mapping specific loci.

CONDITIONALITY IN METABOLIC VARIATION

The ability to make broad conclusions or identify causal genes using quantitative studies of metabolic variation is greatly influenced by the fact that metabolic abundances measured in these studies are highly conditioned on the environmental, developmental, and genetic variations present within the experiment. While most studies are conducted in a single tissue, a single environment, and without an assessment of dependency on the genetic background, insights are emerging into how these factors effect metabolic variation.

Tissue Specificity

Many metabolites display tissue or ontogenic specificity in their accumulation or synthesis (Moco et al., 2007; Kliebenstein, 2013; Moussaieff et al., 2013; Dong et al., 2015). This developmental specificity limits the feasibility of developing a simple complete picture of the genetic variation in a plant’s metabolome that can be extrapolated across tissues and developmental stages, as each tissue or stage may have completely different genetics. Within rice, only 31% of the metabolites detected in seeds and 15% of the metabolites detected in leaves were shared across both tissues (100 shared metabolites total; Gong et al., 2013). Similarly, there was a large effect of tissue specificity in tomato metabolite QTLs when comparing leaf and fruit loci (Schauer et al., 2006, 2008; Fernie and Schauer, 2009). This complicates the ability to infer when or where a gene may be working to influence a plant’s metabolism.

The tissue specificity of metabolism is partially attributable to differential transcriptional control for the underlying enzyme genes. Transcriptome analysis of tomato fruit tissues found strong spatial variation for the expression of central metabolism and secondary metabolism genes (Matas et al., 2011). Similar results were found in Arabidopsis, with tissue-specific expression of metabolic enzyme genes (Brady et al., 2007; Dinneny et al., 2008). Work with Arabidopsis glucosinolates has shown that ontogenic variation in glucosinolate hydrolysis products maps to the enzymatic loci that produce these products, suggesting that transcriptional variation in these genes controls the ontogenic variation (Wentzell et al., 2008; Wentzell and Kliebenstein, 2008). This has also been shown for the accumulation of tissue-specific metabolites in tomato and rice (Tsai et al., 2012; Gong et al., 2013). This suggests that cis-variation in the promoters of specific enzymatic loci that alters their expression may be one source for tissue specificity in metabolomic variation.

In addition to tissue-specific expression of the enzymes, transporters also play a key role in controlling when or where metabolites may accumulate, allowing the sites of synthesis and accumulation to be separated (Chen et al., 2012; Nour-Eldin et al., 2012; Andersen et al., 2013). A QTL modulating primary metabolism within Arabidopsis was cloned that displayed the potential for metabolite transport, confounding QTL interpretation with regard to tissue specificity (Li and Kliebenstein, 2014). In that study, an AT-HOOK was shown to be the basis of quantitative variation in the tricarboxylic acid cycle. Intriguingly, however, the causal gene was expressed in tissues that were distinct from where the metabolites were measured (Li and Kliebenstein, 2014). Because plant metabolism is highly interconnected across tissues by metabolite transport, it is possible for a gene with a highly localized effect to alter metabolite accumulation in a wider range of tissues. It remains to be seen how frequently cause and effect may be separated across tissues in plant metabolic variation.

Genotype × Environment Interaction

Environmental effects strongly influence metabolism, often interacting with genetic variation. This interaction limits the apparent repeatability of metabolomic variation but is a fundamental property of their biological function. For example, the induction of maize volatiles is highly specific to the attacking herbivore (McCormick et al., 2012). As such, the interaction of genotypic variation with environmental variation should instead be considered fundamental to understanding metabolic variation, and loci that vary across environments are likely important for specific environments and should be studied intensively.

Environmental effects have been studied intensively in tomato. In an interspecific cross between S. pennellii and S. lycopersicum (Eshed and Zamir, 1995), 889 metabolite QTLs were identified through gas chromatography-mass spectrometry of the fruit pericarp using three different years of harvest at a single farm site (Schauer et al., 2006). However, only 5% of the 889 metabolite QTLs were consistently identified across all three tomato harvests (Schauer et al., 2008). In volatile profiling of these same lines, significant season × line interactions were present for nine of the 23 metabolites studied in comparisons of spring and autumn harvests (Tieman et al., 2006). Furthermore, in a GWA analysis of diverse S. lycopersicum and Solanum pimpinellifolium accessions, only 47% of primary metabolic traits studied were stable across 2 years of field cultivation (Sauvage et al., 2014).

Similar environment × genotype interactions have been observed in other systems for metabolite accumulation. In rice, analysis of an indica × japonica cross indicated that only eight out of 29 QTLs associated with fatty acid abundance were consistently identified across two generations in two field trials (Ying et al., 2012). In maize GWA studies, only 17% of the metabolite-locus associations were consistently detectable across at least two of three concurrent field trials (Wen et al., 2014). A study of a high-oil maize RIL population over 2 years of field experiments showed that year and the year × genotype interactions contributed significantly to phenotypic variance (Yang et al., 2010). Identifying environment × genotype interactions does not require conducting field trials, as Arabidopsis GWA studies showed that most identified metabolite-locus associations varied across the time of day of the sampling within a single environment (Chan et al., 2010). Similarly, QTL mapping in the presence or absence of jasmonic acid and GWA studies in the presence or absence of different abiotic stresses identified different metabolic QTLs in Arabidopsis (Kliebenstein et al., 2002; Chan et al., 2011).

Thus, there is a strong interaction of environment with genetics in defining variation within the plant metabolome. One alternative is to analyze QTLs associated with relatively stable ratios between metabolites rather than single compounds (Morreel et al., 2006). Most often, the QTLs or associations that are consistent across environments are the loci followed up for study. However, given that plants evolved to adapt to constantly changing environments, it is likely that the loci that are only found in specific environments may be playing a key role in adaptation to those environments (Kerwin et al., 2015). As such, it is important to move beyond the stable loci and begin to clone and understand the mechanistic basis of environmentally conditional loci.

Epistasis: Genotype × Genotype and Genome × Genome Interactions

Another conditionality that influences the study of metabolite variation is the epistatic interaction of specific loci with the genetic background. In this article, we define epistasis as the interaction of genetic variation at two or more loci that creates a nonadditive or unpredictable change in the trait being studied. Molecular studies often consider epistasis to be evidence for a mechanistic interaction between two genes, but this is not an absolute requirement, as epistasis can also occur between genes in different pathways or even within duplicated genes in a gene family (Segrè et al., 2005; Roguev et al., 2008). In a study of the rice metabolome, 241 metabolites (53% of the total examined) exhibited 3,351 significant pairwise interactions between loci (Chen et al., 2014), and in a study of 16 phenolamides, eight significant pairwise interactions were detected between loci (Dong et al., 2015). In multiple single-nucleotide polymorphism (SNP) models of the loblolly pine (Pinus taeda) metabolome, more SNP effects were identified as dominance effects than as additive effects (Eckert et al., 2012). Twenty-four pairs of epistatic QTLs were detected in the study of high-oil maize RILs, but they accounted for 16% or less of the variance in individual oil phenotypes (Yang et al., 2010). Similarly, there are extensive epistatic interactions found for metabolomic variation in Arabidopsis in both primary and secondary metabolism (Wentzell et al., 2007; Rowe et al., 2008; Joseph et al., 2013a, 2013b; Kerwin et al., 2015). In Arabidopsis, this epistasis is largely higher order, involving the interaction of three or more loci rather than simple interactions of two loci (Wentzell et al., 2007; Rowe et al., 2008; Joseph et al., 2013a, 2013b; Kerwin et al., 2015). Cloning of the underlying loci shows that the epistatic interactions include the interaction of transcription factors with each other and with variation in enzymatic genes (Wentzell et al., 2007; Sønderby et al., 2010). It remains to be seen if higher order epistasis can be detected in other species.

Most studies on natural variation in plants limit their analysis to genetic variation within the nuclear genome. However, there is extensive evidence for genetic variation in the organelles altering adaptive phenotypes and interacting with nuclear loci (Greiner and Bock, 2013). Recent populations have been developed that are allowing quantitative assessment of the role of genetic variation in the organelle in quantitative phenotypes (McKay et al., 2008; Lovell et al., 2015). Analyzing metabolomic variation in the Arabidopsis Kas × Tsu population showed that genomic variation within the organelle altered the accumulation of nearly all metabolites (Joseph et al., 2013a, 2013b, 2015). This analysis also showed that there are extensive epistatic interactions between genetic variation in the organellar and nuclear genomes (Joseph et al., 2013a, 2013b, 2015). It remains to be tested how extensive this may be across different plant species.

Not all studies identify evidence of epistasis. A study of 342 rice metabolites found little evidence for epistasis in a comparison of metabolite abundance between parent accessions and offspring (Matsuda et al., 2015). Similarly, no significant epistatic interactions were detected in a pairwise analysis of genes associated with oil accumulation traits in maize (Li et al., 2013). It remains to be determined what differs between studies that identify epistasis in metabolic variation and those that do not. The differences could be the germplasm being used, the statistical methodologies, the environment, or some blend of these factors. Further studies will be required to quantify the extent of epistasis in plant metabolomic variation.

CAUSAL LOCI CONTROLLING METABOLIC VARIATION

A major goal of all quantitative metabolomic studies is to clone the underlying genes to understand the mechanistic basis of this variation. Recent reports have described the protocols to clone these loci, such as comparing transcriptomic and metabolic variation (Saito et al., 2008; Chan et al., 2011; Kliebenstein, 2012, 2014; Atwell and Kliebenstein, 2013). These correlational/colocalization approaches are greatly speeding up progress in cloning the causal genes, including a broad array of new enzymes, transcription factors, and other genes. This ever-increasing list of cloned loci is vastly beyond any single article’s ability to summarize. As such, we will focus on instances in which the cloned genes are illuminating new and unexpected mechanistic aspects of plant biology.

Interplay of Physiological and Metabolic Variation

A complexity in understanding plant metabolism is defining when one trait ends and another begins. Recent studies are beginning to show that metabolic variation is highly responsive to physiological variation, such as in resource availability and partitioning. A set of 126 maize SNPs associated with major carbon and nitrogen metabolic traits was overrepresented with genes linked to C4 photosynthesis and the regulation of carbon sink-source relationships (Zhang et al., 2015). Whole-plant morphology and growth conditions are also large contributors to tomato metabolic variation relative to the genetics studied (Do et al., 2010). Harvest index, a measure of fruit yield relative to biomass, was identified as the pleiotropic hub of the S. pennellii × S. lycopersicum metabolomic network linked to 46% of the metabolites (Schauer et al., 2006). Similar results were found in other tomato crosses (Prudent et al., 2009; Do et al., 2010). Overall, these results indicate that the interaction between genetics and source-sink dynamics plays a major role in defining central metabolism. Thus, metabolomic variation may often identify key physiological regulators that are important control nodes for the metabolome with pleiotropic effects.

In support of this are recent studies that cloned major metabolic variation QTLs within Arabidopsis. In one of these studies, most metabolic variation was linked to variation in the circadian clock (Kerwin et al., 2011). A key locus controlling this variation is natural variation in the EARLY FLOWERING3 gene that has been linked to altering flowering time, shade avoidance, hypocotyl elongation, circadian clock oscillation, and metabolic variation (Coluccio et al., 2011; Jimenez-Gomez et al., 2011; Kerwin et al., 2011; Undurraga et al., 2012; Anwer et al., 2014; Nieto et al., 2015). In tomato, the transcription factor APETALA2a that controls physiological pathways by regulating hormone synthesis to control fruit ripening also has impacts on phenylpropanoid and carotenoid metabolite accumulation (Karlova et al., 2011). Also in tomato, the ethylene receptor Never-Ripe controls variation in ascorbate and carotenoids but also strongly influences fruit ripening and seed production (Alba et al., 2005). In rice, the transcription factor Rc regulates flavonoid biosynthesis through abscisic acid signaling, with pleiotropic effects on seed dormancy and seed weight (Gu et al., 2011). This raises the question of whether these loci are specific metabolic regulators or are better interpreted as general regulators of physiology whose effect can be measured using metabolic variation. However, it will require analyzing the metabolic consequences of variation in other key physiological genes to tweeze apart the direct and indirect consequences on metabolite accumulation (Fukushima et al., 2009; Li and Kliebenstein, 2014).

Interplay of Metabolic and Defense Variation

The previous results suggest a pleiotropic link wherein variation in physiological or growth regulators influences metabolic traits. The cloning of genes underlying metabolic QTLs is beginning to highlight instances where variation in metabolic genes leads to unexpected effects, particularly on plant defense regulation. A major-effect QTL for aphid susceptibility in maize due to decreased levels of a benzoxazinoid, 2-hydroxy-4,7-dimethoxy-1,4-benzoxazin-3-one glucoside (HDMBOA-Glc), accumulation was caused by natural variation in transposon inactivation of a methyltransferase enzyme (Meihls et al., 2013). Inactivation of this enzyme also altered callose deposition in response to aphid infestation, suggesting a regulatory link between benzoxazinoid accumulation and callose induction (Meihls et al., 2013). A link between defense metabolite accumulation and callose regulation has been found for other methoxylated indolic metabolites with natural variation in their enzyme-encoding genes (Clay et al., 2009; Pfalz et al., 2009).

Supporting the potential for metabolic variation to influence defense regulation was a reanalysis of natural variation in a 2-oxoacid-dependent dioxygenase (AOP2) that controls the production of alkenyl glucosinolates in Arabidopsis and Brassica spp. (Kliebenstein et al., 2001; Li and Quiros, 2003). The presence or absence of this gene was found to alter the periodicity of the circadian clock as well as flowering time in both the laboratory and the field (Kerwin et al., 2011, 2015). A transcriptomic survey of genotypes with or without this enzyme showed that lines with a functional AOP2 enzyme had altered expression of both the biosynthetic and regulatory genes in the jasmonate signaling cascade, leading to decreased jasmonate sensitivity (Burow et al., 2015). As such, detailed studies of the benzoxazinoid and glucosinolate causal genes is beginning to show that metabolic loci, even enzyme-encoding ones, can have unexpected effects on other pathways, indicating potential regulatory influences for these metabolites in planta.

Diversifying Selection and Major-Effect Polymorphisms at Large-Effect Loci

A common conclusion of cloning studies is that secondary metabolite loci are controlled by large-effect presence/absence polymorphisms at the causal loci. This observation has been supported by a number of cloning studies in Arabidopsis, Brassica spp., rice, and maize (de Quiros et al., 2000; Kliebenstein et al., 2001; Li and Quiros, 2003; Hansen et al., 2008; Pfalz et al., 2009; Leckie et al., 2012; Meihls et al., 2013). However, there are also primary metabolite loci with presence/absence polymorphisms. In maize, seed starch and carotenoid contents are both controlled by large-effect polymorphisms in the causal genes (Thévenot et al., 2005; Vignesh et al., 2012; Lipka et al., 2013; Owens et al., 2014). As such, it is not accurate to make a primary versus secondary metabolism split when discussing large-effect or presence/absence polymorphisms in causal loci. Instead, it is more likely the shape of selection on the trait that is critical. In this case, all of these examples are in traits that are under either diversifying or fluctuating selection. In the case of the glucosinolates for Arabidopsis and Brassica spp., the defense traits are responding to fluctuating herbivore populations that create diversifying or balancing selective pressures that likely maintain and possibly even drive the variation (Prasad et al., 2012; Züst et al., 2012; Brachi et al., 2015; Kerwin et al., 2015). In the case of maize, carotenoid color is driven by diversified cultural preferences for white or yellow corn, and starch content is driven by diversifying selection on field corn versus popcorn. Similar to the herbivore pressure in wild Arabidopsis, the diversifying selection applied by human breeders for extremely divergent morphs of maize likely lead to the selection of large-effect presence/absence causal polymorphisms (Springer et al., 2009; Hufford et al., 2012). Thus, before classifying the type of causal polymorphisms expected when working with a metabolic trait, it is probably more important to understand the selective pressure on the metabolite rather than the type of metabolite.

What Genes Alter Variation in Metabolism?

A common question in all natural variation studies, including those on metabolism, is what are the genes that typically cause the phenotypic variation? A reading of the current literature suggests that we have an answer to this question in that we have to assume that if a gene has genetic variation that causes phenotypic variation in a trait, there will be a measurable shift in metabolism. The current set of naturally variable genes validated to impact natural variation include representatives from nearly all types of genes, from enzymes to transcription factors (Thévenot et al., 2005; Vignesh et al., 2012; Angelovici et al., 2013; Lipka et al., 2013; Meihls et al., 2013; Owens et al., 2014). As shown above, these genes can directly influence the metabolite or have potential indirect influences on the metabolite accumulation. As such, it is probably more appropriate to move beyond the question of what type of genes cause metabolic variation in natural populations and on to the more critical question of how these genes influence metabolite variation. Are they direct regulators that have immediate molecular impacts on the pathway regulation or biosynthetic potential? Or, alternatively, do these genes have more distant (sometimes thought of as indirect pleiotropic effects) links to the metabolic pathway whose output is being measured, and if so, how distant are these effects? Association mapping and other systems biology studies of glucosinolates in Arabidopsis have shown that there are dozens to hundreds of genes that can alter accumulation to a level that likely alters fitness (Chan et al., 2011; Li et al., 2014; Brachi et al., 2015; Kerwin et al., 2015). Yet, it is highly unlikely that all of these genes directly interact with the pathway in a molecular context. This generates a system whereby a metabolite’s natural variation can be altered by potentially hundreds of candidate genes, and each gene can conversely alter an array of metabolites (Angelovici et al., 2013). Within this system, the question arises of how selection can identify the proper combination of alleles to optimize fitness at the metabolic level. This mechanistic question is a key topic for the field to begin querying to understand how the metabolic system functions within an individual or species.

STATISTICAL INFLUENCES AND CHOICES

There is a large body of literature and reviews about the technical, statistical, experimental design, and population choices that are required to accomplish a quantitative analysis of natural variation in metabolism (Lynch and Walsh, 1998; Fiehn et al., 2008; Myles et al., 2009; Atwell and Kliebenstein, 2013; Luo, 2015). Thus, we will not provide a detailed step-by-step approach, as this has been done better. Instead, we will work to convey that every step has different options and there is no best choice. Rather, each option/choice introduces a different bias, and it is critical to understand what that bias is to properly interpret the results. This may be best conveyed in the choices of mapping populations available to conduct metabolomics analysis.

A highly popular approach that is gaining momentum is the use of GWA populations that are collections of random wild genotypes (Atwell et al., 2010). This population has the benefit of containing a sampling of the allelic diversity in the species. However, these populations contain significant population structures and, with even hundreds of accessions, have limited capacity to find epistasis or the effect of rare alleles, thus generating an unrecognized false-negative error rate (Chan et al., 2010a, 2010b; Long et al., 2013; Brachi et al., 2015). In contrast, the classical RIL population derived from two parents has the flaw of only having two alleles per gene, but this limitation also provides this population the greatest power to identify epistasis and small-effect loci (Falconer and Mackay, 1996; Mackay, 2014). The nested association mapping population was devised to alleviate the issue of allelic diversity in the RIL design and the rare allele issue in the GWA design. To accomplish this, it combines a set of recursive RIL populations involving multiple parents into a single combined population (Buckler et al., 2009). The true strength of this population lies in its ability to identify moderate-effect additive loci. A similar approach is the multiparent advanced generation intercross (MAGIC) population design, wherein multiple parents are admixed to create a single population (Kover et al., 2009). This population design also works well for moderate-effect, moderate-frequency additive loci, but like a GWA design, it struggles with complex epistasis or small-effect loci (Falconer and Mackay, 1996; Mackay, 2014). Thus, there is not a single population design that is suitable for all studies; instead, the population must be carefully chosen to match the strengths and weaknesses to the goal of the study.

After choosing the population, the next choice is to determine the number of lines and associated bioreplicates. This is then followed by choosing the statistical approach to link genotype with phenotype. However, these choices are highly linked. The rapid explosion of new statistical approaches for quantitative genetics is not a reflection of the flaws in the original approaches but instead a reflection of the mere fact that most experiments do not have sufficient numbers of genotypes to even fractionally sample the complete genotype-to-phenotype matrix. For example, it was recently estimated that a single RIL population would need at least 1,000 to 1,200 independent lines before it could even be determined how many more lines were needed to identify all the possible QTLs (Joseph et al., 2014). This is in a situation when most available RIL populations are maximally 200 or so lines. As such, the newer approaches, like Bayesian, multitrait, or other approaches, are simply trying to maximize the information obtained from significantly underpowered populations. Thus, rather than focusing on the optimal statistical approach, it is better to focus on maximizing the number of lines and replicates used to generate the data input into the statistical algorithms. Maximizing the power of the data will go a longer way toward optimizing the output than any particular statistical algorithm.

CONCLUSION

Plant biology over the past decades has focused largely on the qualitative assessment of small collections of genotypes within limited environments to assess the mechanistic function of one or a few genes. The future, however, will require the quantitative analysis of systematic genotype collections within a range of environments and tissues to understand the functions of entire systems. Metabolomic analysis of natural variation is well positioned to enable these very types of experiments and begin to assess questions that, until now, have been largely overlooked or inaccessible. What is the level of cell or tissue autonomy in metabolism? How many different mechanisms coordinate genetic variation in the organellar and nuclear genomes? What types of selection alter natural genetic variation in the field? All of these questions require the rapid and cheap quantitative ability that metabolomics provides in addition to the large body of enzymatic knowledge about the system. In combination, metabolomics analysis of natural variation should be a key component of future plant studies working to understand how a plant species functions in the wild or the field.

Glossary

QTL

quantitative trait locus

RIL

recombinant inbred line

GWA

genome-wide association

SNP

single-nucleotide polymorphism