- © 2014 American Society of Plant Biologists. All rights reserved.
Abstract
Glycerolipid metabolism of plants responds dynamically to changes in light intensity and temperature, leading to the modification of membrane lipid composition to ensure optimal biochemical and physical properties in the new environment. Although multiple posttranscriptional regulatory mechanisms have been reported to be involved in the process, the contribution of transcriptional regulation remains largely unknown. Here, we present an integrative analysis of transcriptomic and lipidomic data, revealing large-scale coordination between gene expression and changes in glycerolipid levels during the Arabidopsis thaliana response to light and temperature stimuli. Using a multivariate regression technique called O2PLS, we show that the gene expression response is strictly coordinated at the biochemical pathway level and occurs in parallel with changes of specific glycerolipid pools. Five interesting candidate genes were chosen for further analysis from a larger set of candidates identified based on their close association with various groups of glycerolipids. Lipidomic analysis of knockout mutant lines of these five genes showed a significant relationship between the coordination of transcripts and glycerolipid levels in a changing environment and the effects of single gene perturbations.
INTRODUCTION
Glycerolipid metabolism of plants is strongly involved in the response and adaptation to changes in environmental conditions (Moellering and Benning, 2011). Two common environmental parameters affecting glycerolipid metabolism are light intensity and temperature, and the effects of both have been intensively studied in plants over the last four decades (Yoshida and Sakai, 1974; Browse et al., 1981).
One of the most well-known effects of prolonged chilling stress is increased desaturation of glycerolipids, serving as a compensation mechanism for decreased membrane fluidity (Williams et al., 1988; Welti et al., 2002; Tasseva et al., 2004). In agreement with this, cold-tolerant species (Sakamoto et al., 2004; De Palma et al., 2008) and varieties (Horvath et al., 1983) and cold-acclimated plants (Degenkolbe et al., 2012) have been shown to have relatively high levels of desaturated glycerolipids. Changes in glycerolipid saturation levels are also accompanied by other more specific effects, such as a decrease of monogalactosyldiacylglycerols (MGDGs) (Li et al., 2008) and an increase of triacylglycerides (TAGs) connected with the outer chloroplast envelope remodeling mediated by galactolipid:galactolipid galactosyltransferase (Moellering and Benning, 2011). Besides galactolipid:galactolipid galactosyltransferase, a number of proteins have been shown to connect lipid metabolism with cold stress. Common examples include the phospholipases PLDα1 and PLDδ (Welti et al., 2002; Rajashekar et al., 2006; Li et al., 2008) and acyl-CoA binding proteins (Chen et al., 2006; Du et al., 2010a, 2010b). In contrast with the effect of cold stress/treatment, an increased saturation of membrane lipids has been observed as a heat stress–induced adaptational mechanism (Larkindale and Huang, 2004). On the other hand, light intensity affects membrane lipid composition in a different way. Desaturation of acyl chains is not light regulated (Browse et al., 1981), but darkness strongly represses the de novo synthesis of fatty acids (FAs) (Ohlrogge and Jaworski, 1997) and thus also the influx of saturated FAs into glycerolipid biosynthetic pathways. High light has an opposite effect, leading to a surplus of NADPH (Stitt, 1986) and increased FA synthesis. Therefore, as we discussed by Burgos et al. (2011), light affects lipid metabolism mainly by defining energy and carbon availability. The involvement of specific glycerolipids, such as phosphatidylglycerol (PG), in the oligomerization and stabilization of photosynthetic complexes (Frentzen, 2004) further suggests a more complex relationship between light and membrane lipid composition.
Whereas most of the described effects might be attributed to specific posttranscriptional regulatory processes, there are hints that regulation at the transcriptional level also contributes to remodeling of the plant membrane composition. The most important evidence is the tight coexpression of genes involved in lipid metabolism and in its particular biochemical pathways in response to stress (Obayashi et al., 2007; Loraine, 2009; Avin-Wittenberg et al., 2011). Several transcriptional regulators of specific lipid metabolic pathways have also been reported (Baud and Lepiniec, 2010; Bernard and Joubès, 2013), but the details of transcriptional regulation in membrane glycerolipid remodeling are not well understood. Therefore, in this work, we ask three basic questions. (1) To what extent is transcriptional regulation involved in the remodeling of plant membrane lipid composition in response to changes in light and temperature? (2) Are there specific pathways and metabolic processes exhibiting coordination between the expression of the enzymatic genes and the accumulation of certain lipid species? And (3) does this coordination remain in agreement with the effects of single gene perturbations and thus could be useful for gene function prediction?
To answer these questions, we revise the transcriptomic and lipidomic data for Arabidopsis thaliana, described separately in our previous studies (Burgos et al., 2011; Caldana et al., 2011), in a new integrative analysis. Integrative transcriptomic–metabolomic studies, aiming at the identification of significant and biologically relevant coordination between gene expression and metabolite levels (Redestig et al., 2011), proved to be useful in uncovering large-scale organization of metabolic regulation and in highlighting candidates for new enzymes and regulators of plant metabolism (Urbanczyk-Wochniak et al., 2003; Hirai and Saito, 2004; Hirai et al., 2007; Jozefczuk et al., 2010). Here, we use O2PLS, a multivariate regression analysis method developed by Trygg (2002) and Trygg and Wold (2003) and applied in multiple systems-scale studies (Bylesjö et al., 2007, 2009; Consonni et al., 2010; Zamboni et al., 2010; Kusano et al., 2011). O2PLS is an extension of orthogonal projections to latent structures (OPLS) (Trygg and Wold, 2003); whereas OPLS was designed for analysis of a single data set, O2PLS assesses systematic trends across multiple data sets. In principle, O2PLS is a multivariate analysis method like, for example, principal component analysis (PCA), but it parses out the variation in large data sets differently. O2PLS focuses on predictive information, separating the variance common for two data sets (correlated between the data sets) from the variance unique for only one data set (correlated within one data set only) and from analytical noise (residual uncorrelated variance). This is especially important in the case of time-series experiments, where differences between the accumulation of lipids and gene expression changes arise from different dynamics of lipid metabolism and gene regulation and where a large quantity of platform-specific analytical noise is expected (Burgos et al., 2011).
Our study is divided into three major steps. (1) By using O2PLS analysis, we assess the fraction of the predictive variance in both lipid and transcript data and, thus, the extent to which changes in lipid composition are reflected by changes in gene expression and vice versa. (2) We interpret the identified coordination in the context of known regulatory mechanisms and pathway-level regulation of gene expression. (3) Finally, we compare the identified interactions with the effect of individual gene knockouts and evaluate to what degree the environmental transcript–lipid coordination reflects causal transcript–lipid relationships.
The results indicate a significant coordination between gene expression and glycerolipid accumulation in response to changing environmental conditions related to a concerted regulation of major lipid biosynthetic pathways. Furthermore, a correspondence between transcript–lipid abundance coordination and the targeted gene perturbations is shown.
RESULTS
Experimental Setup
Applied treatments include four light regimes: high light (400 μE), control light (150 μE), low light (75 μE), and darkness; and three temperature regimes: cold (4°C), control temperature (21°C), and heat (32°C). These regimes were combined such that all light intensity treatments were performed in control temperature and the cold and heat stress experiments were performed under control light and additionally in darkness, resulting in eight different combinations (Figure 1). In the following sections, the conditions are indicated by the following letter codes: 4-L (cold/normal light), 4-D (cold/darkness), 21-HL (control temperature/high light), 21-L (control temperature/normal light; control conditions), 21-LL (control temperature/low light), 21-D (control temperature/darkness), 32-L (heat/normal light), and 32-D (heat/darkness). For each combination, we obtained time-course data for the first 6 h of the plant response with sampling every 20 min. From strictly quality-filtered probe sets, a subset representing the expression of 480 acyl-lipid metabolism genes was selected based on the ARALIP database classification (http://aralip.plantbiology.msu.edu; version from January 2013) (Li-Beisson et al., 2013). All details on the experimental and analytical procedures are described in Methods as well as by Burgos et al. (2011) and Caldana et al. (2011).
Array of Applied Treatments.
Statistical Model
The basic concepts of O2PLS are key to understanding the results of this study; therefore, first we will introduce the basic terminology and idea of the method. O2PLS is a statistical method allowing the convenient integration of two data sets collected from the same samples and expected to be causally connected, such as transcripts and metabolites (Trygg, 2002; Trygg and Wold, 2003). In an O2PLS model, each data set is decomposed into three variance structures describing predictive, unique, and residual variation of its variables (Figure 2). The predictive structures describe a multivariate relationship between two data sets, allowing reciprocal prediction between them. In our experiment, it allows the estimation of membrane lipid levels from the gene expression and vice versa and might be related to a direct metabolic and regulatory link between gene expression and levels of membrane lipids. The unique structures represent patterns that are not useful to predict the other data set. These might be linked to platform-specific effects, such as baseline bias, or variation that is not reflected in the other data set, such as gene expression regulation having no effect on lipid levels or metabolic changes originating from other than transcriptional regulatory mechanisms. Finally, the residual variance structures represent noise and stochastic effects.
Overview on the O2PLS Model Structures Obtained for Integration of the Transcript and Lipid Data.
Each model structure shows the percentage of the total variance explained.
In O2PLS, each variance structure is composed of a certain number of latent variables and their weights (called loadings), describing the contribution of the latent variables in the variance of each observable variable. Thus, changes of all observable variables (e.g., all measured lipids) in the frame of a certain variance structure (e.g., lipid predictive) are a linear combination of just a couple of latent variables mixed with different proportions.
Here, the O2PLS analysis was performed using a modified version of the algorithm described by Trygg (2002) (see Methods for details). The number of latent variables for each variance structure was identified by choosing the configuration that minimized the generalization error, which was estimated by using a group-balanced Monte Carlo cross-validation (MCCV) (Bylesjö et al., 2007). The optimal setup consisted of eight joint variance components (Figure 3A shows two of them), five transcript-unique components, and six lipid-unique components (Supplemental Figure 1 and Supplemental Table 1). In the following sections, latent variables of the respective model structures will be described by letter codes: J-LV1 to J-LV8 for both transcript and lipid latent variables of the predictive structure, JT-LV1 to JT-LV8 for the transcript latent variables of the predictive structure, JL-LV1 to JL-LV8 for the lipid latent variables of the predictive structure, UT-LV1 to UT-LV5 for the latent variables of the transcript-unique structure, and UL-LV1 to UL-LV6 for the latent variables of the lipid-unique structure. It is important to note that, unlike in PCA, the numbering of O2PLS latent variables is arbitrary and does not relate to the amount of variance explained.
Characteristics of O2PLS Latent Variables.
(A) Exemplary plot of two out of eight identified O2PLS latent variables (LV). Samples are denoted by numbers representing the time points of each experiment (expressed as minutes upon treatment). The plot should be interpreted analogously to PCA. The variances captured by the LV1 (r2X = 0.21 and r2Y = 0.6, where X and Y refer to the transcript and lipid data, respectively) and LV2 (r2X = 0.27 and r2Y = 0.12) separate samples are mostly according to temperature conditions and the time progression of the experiment.
(B) Light-specific latent variables obtained by partial least squares regression of the O2PLS-DA–predicted values (r2X = 0.28 and r2Y = 0.20) on the light gradient and temperature-specific latent variables obtained by partial least squares regression of the O2PLS-DA–predicted values (r2X = 0.32 and r2Y = 0.28) on the temperature gradient. As shown, the light- and temperature-specific latent variables capture the variance of the predictive variance structure, being directly related to the applied temperature treatment and light intensity (note the clear temperature-related separation of samples along the x axis and the light-related separation along the y axis).
(C) Correlations between O2PLS-DA latent variables and the original eight latent variables of the transcript O2PLS predictive model structure. Each green arrow represents one latent variable (numbered next to the arrowhead). The thickness of the arrow represents the proportion of the variance described by each particular latent variable. Additionally, correlations of individual transcripts and lipids are plotted, indicating that variance in both data sets covers a broad spectrum of combinations of light- and temperature-related effects.
Identification of Predictive Transcriptome–Lipidome Variance
The total joint variance between the data sets reaches 61.6% for the transcript-predictive structure and 27.6% for the lipid-predictive structure. Unique variance reaches 20.1% for the transcript data and only 4.8% for the lipid data. The model shows that the environmental stimuli result mostly in systematic and coordinated changes of lipids and lipid metabolism gene expression. A considerable amount of transcript-unique variance also indicates that a significant proportion of the gene expression changes are accompanied by changes in lipid levels in the experimental time scale. On the other hand, the relatively low contribution of lipid-unique variance suggested that almost all systematic effects observed in lipid data could be connected with some changes in transcripts. To verify that the residual structures do not contain systematic variance, these were analyzed by PCA, which indicated no significant treatment- or time-specific effects.
To determine if the lipid metabolic genes have higher predictive power toward lipid profiles than any other set of genes, we performed a permutation test, where an O2PLS model of the same complexity (the same number of variance components for each respective variance structure) was fitted to a randomly chosen set of nonlipid genes of the same size as the lipid genes group (480 genes). After 1000 iterations, none of the random gene sets exceeded the ratio of the joint variance obtained for the lipid genes (P < 0.001). Accordingly, the size of the unique variance structure was higher for the nonlipid genes, with P < 0.05 (Supplemental Figure 2). This result indicated that the changes in expression of lipid-related enzymatic genes were indeed related to the accumulation of glycerolipids and were more significant compared with changes in any other random set of genes.
In the next step, we performed an O2PLS discriminatory analysis (O2PLS-DA) (Bylesjö et al., 2006). O2PLS-DA identifies latent variables discriminating preselected sets of observations within the O2PLS model. Here, O2PLS-DA was performed on predictive variance structures (Figure 3A; see Supplemental Figure 3 for plots of all latent variables of predictive variance structures), and its result was regressed on the gradient on light and temperature conditions (abbreviated as TEMP-LV and LIGHT-LV, respectively; Figure 3B). The new TEMP-LV and LIGHT-LV obtained from the O2PLS-DA model extract a portion of the joint variance, which is related to either the light or temperature gradient. This operation allows representing the transcripts and lipids on a common 2D plane (Figure 3C), where the dimensions describe the light and temperature specificity of the response and the proximity between transcripts and lipids is related to the coordination of their changes.
Large-Scale Coordination of Lipid Biochemical Processes
Applied treatments led to moderate changes in glycerolipid class levels. Darkness resulted in the accumulation of phosphatidylinositol (PI), PG, and phosphatidylethanolamine (PE), reflected by negative correlation of the sum of all molecular species of the respective class with the light-specific latent variable (Figure 4; the generalized changes of whole lipid classes are represented by arrows). In the case of PI, the effect is slightly stimulated by elevated temperature, whereas PG and PE accumulate more pronouncedly in dark/cold conditions. Sulfoquinovosyldiacylglycerol (SQDG), MGDG, and phosphatidylserine, on the other hand, exhibited temperature-specific effects, accumulating under 32°C independently of the light conditions. Fluctuations at the level of whole glycerolipid pools are rather minor; however, major differences occurred between molecular species belonging to the same glycerolipid class. This indicated that most of the significant effects occurred due to a shift in saturation level or acyl chain length (or both) and not in the level of head group chemistry. Indeed, the light effect (represented by LIGHT-LV) correlates significantly with the saturation level of multiple glycerolipids (Supplemental Figure 4B). This concerns in particular MGDGs, phosphatidylcholines (PCs), PEs, and SQDGs, exhibiting gradients of saturation from the desaturated lipid species accumulating in darkness to saturated species increasing in light and highlight conditions. This effect was enhanced by heat and concerns multiple lipid classes. The most extreme effect is exhibited by a group of normally low-abundance species of PC, PE, PI, and SQDG containing fully desaturated 16-C acyl chains. The most representative of them are 34:6 SQDG, PC, and PE, 34:5 PI, and 32:3 PI; the probable source of these species and their connection with temperature are debated by Burgos et al. (2011). The summarized changes of the molecular lipid species containing the same number of carbons in their acyl chains indicate that shorter 16-C chains accumulate in darkness and heat (Supplemental Figure 4C). Heat treatment was positively correlated with molecular species having two 16-C acyl chains, whereas for the control temperature in darkness the combination of 18-C and 16-C was dominant. On the other hand, light was not significantly correlated with any certain number of acyl chain carbons, and cold treatment promoted the accumulation of lipid species with double 18-C acyl chains.
Correlation Loading Plots of the Lipids and Transcripts in the O2PLS-DA Model.
Lipids are denoted as triangles and transcripts as squares. The location of each variable on the plot is described by the correlation of its joint variance projection with light-specific (x axis) and temperature-specific (y axis) latent variables of the O2PLS-DA model. The gradients are from darkness- to high light–specific responses (left to right) on the x axis and from cold- to heat-specific responses (bottom to top) on the y axis. Each plot shows the distribution of all analyzed variables (small gray symbols), but on each of them a different group is highlighted by magnified and color-coded symbols.
(A) Glycerolipids. The total number of double bonds in glycerolipid acyl chains is represented by the color scale from red (six double bonds) to blue (no double bonds). Sum changes of the lipid classes are represented by green arrows. From all plotted lipids, only those with significant correlation (P < 0.01) with one of the latent variables have been plotted.
(B) Transcripts of the FA biosynthesis pathway. Genes belonging to the FA biosynthesis pathway are highlighted in green. Sum changes of all lipid biochemical pathways are plotted as gray arrows. The sum changes of the FA biosynthesis pathway are highlighted by a green arrow. The abbreviations for the pathways are as follows: Cutin, cutin biosynthesis; EuP, eukaryotic phospholipid biosynthesis; FA, FA biosynthesis; FA&TAG deg, FA and TAG degradation; FAE&Wax, FA elongation and wax biosynthesis; LT, lipid trafficking; Oxylipin, oxylipin metabolism; P sig, phospholipid signaling; Pro GSP, prokaryotic galactolipid, sulfolipid, and phospholipid metabolism; Sph, sphingolipid biosynthesis; Suberin, suberin biosynthesis; TAG synth, TAG biosynthesis.
(C) Genes of prokaryotic and eukaryotic galactolipid, sulfolipid, and phospholipid metabolism pathways.
(D) Genes involved in storage oil metabolism.
Parallel changes in gene expression indicate that remodeling of the membrane lipid composition coincides with a coordinated regulation of multiple lipid metabolism pathways (Figures 4B to 4D). The general trends of the pathway gene expression, obtained in the same way as sum changes of the lipid classes, match the latent variables describing major joint variance components (Figure 3C). There are three major transcriptional programs, light-specific/temperature-independent, temperature-specific/light-independent, and intermediate light/temperature-modulated, each specific to a certain subset of lipid biochemical pathways. The major one (intermediate light/temperature-modulated), including FA synthesis and trafficking, prokaryotic glycerolipid synthesis, FA elongation and wax biosynthesis, oxylipin metabolism together with cutin and suberin biosynthetic pathways, follows the J-LV4 pattern, exhibiting a strong positive correlation with light regime and dominated by temperature-modulated downregulation of the gene expression in darkness. FA degradation and both TAG degradation and biosynthesis follow closely J-LV1 and J-LV2, being positively correlated with the temperature gradient. Finally, the last group of pathways, including sphingolipid metabolism, phospholipid signaling, and eukaryotic phospholipid synthesis, follows J-LV3, which covers largely opposite responses to J-LV4 but without a strong modulating effect of temperature.
Relationship between Environmental Coordination and Genetic Control
To estimate the degree to which the observed coordination between transcript and lipid changes is related to the actual metabolic functions of particular genes, we selected five genes, exhibiting various degrees and specificity of response to light and temperature stimuli, and performed lipidomic analysis of their knockout mutant lines.
Mutant Phenotypes
The five genes selected for knockout mutant analysis were those encoding ketoacyl-ACP synthase II (FAB1/KASII); two enzymes of the prokaryotic glycerolipid biosynthesis pathway: digalactosyldiacylglycerol synthase 2 (DGD2) and UDP-sulfoquinovose:DAG sulfoquinovosyltransferase (SQD2); and two enzymes of the eukaryotic glycerolipid biosynthesis pathway: long-chain acyl-CoA synthetase 4 (LACS4) and putative CDP-DAG synthase (CDS2). LACS4 and KASII represent dark-induced genes, SQD2 is light-specific, CDS2 exhibits a heat-specific response, and DGD2 is cold-induced (Figure 5).
Locations of the Selected Gene Candidates on the O2PLS-DA Correlation Loading Space.
The correlation loadings of lipids are denoted as triangles, and those of transcripts are denoted as squares.
T-DNA insertion lines (Alonso and Stepanova, 2003) for these genes were obtained from the Nottingham Arabidopsis Stock Centre collection (http://arabidopsis.info) and are listed in Supplemental Table 2. All selected mutant lines were different from those published previously for these genes, with the exception of LACS4 lines, being identical to those described by Jessen et al. (2011). In wild-type plants, transcripts of all five genes were found to accumulate well within the range of the quantitative RT-PCR detection limit. In the mutant plants, no transcripts were detected in five biological replicates, with the exception of sqd2, where a 116.6-fold decrease in the SQD2 transcript level has been observed (Supplemental Table 3). Samples of the mutant plants, together with the control wild type, were collected in standard nonstress conditions and were subjected to a lipidomic analysis using ultra-performance liquid chromatography coupled with Fourier transform mass spectrometry (Hummel et al., 2011). Obtained features (m/z at a certain retention time) were queried against an in-house lipid database, providing 119 annotated glycerolipid species (Supplemental Data Set 1). Because the mass spectrometry was performed only in positive ionization mode, the obtained data set lacks phosphatidylserine and low-abundance molecular species of PI and PG. In total, 61 of the detected lipids overlapped with those reported by Burgos et al. (2011). Samples from six biological replicates were measured for each mutant and control wild-type line. The significance of the mutant lipidomic phenotype was estimated by two-way ANOVA followed by the Tukey test. The match between mutant lipidomic phenotype and stress-related changes in gene expression was estimated by Spearman correlation analysis (described in detail in Methods).
All of the mutants analyzed exhibited multiple significant changes in their glycerolipid profiles, with independent lines of LACS4 having very similar phenotypes (Figure 6; source data are given in Supplemental Data Set 2). The strongest lipidomic phenotype was obtained in the cds2 mutant. Knockout of CDS2 resulted in decreased general phospholipid:galactolipid ratio, drastic decreases in PI and PG, and depletion in certain PCs and PEs, including highly abundant 34:2 and 34:3 PC species and all molecular species of PE except the low-abundance 38:2 and 38:3 PE. A general decrease in most molecular species of SQDG was also observed, including a 2-fold change in the most abundant 34:3 SQDG. At the same time, a general increase in galactolipid level occurred, including the most abundant 34:6 and 36:6 MGDG and 34:3 and 36:6 digalactosyldiacylglycerol. Taking into account that CDS2 is correlated with many genes involved in storage oil biosynthesis and degradation (Figure 4D), the cds2 mutant exhibits no effect on TAGs. To some extent, the kasII mutant exhibited an opposite effect: a severe depletion of almost all TAGs except the most abundant 54-C molecular species and a general decrease in galactolipids and sulfolipids, balanced by a slight accumulation of several phospholipids. Additionally, a general shift toward shorter acyl chains was seen for galactolipids, marked by a significant decrease in 36-C molecular species of digalactosyldiacylglycerol and MGDG. Considering the relative abundance of measured glycerolipids, the most important effect of LACS4 knockout was the depletion of 36:6 DAG. This change reached a 2-fold decrease and has been observed in both analyzed mutant lines; however, due to the high deviation between replicates of the lacs4.2 line, this result is statistically significant only in lacs4.1. This change was accompanied by a drastic decrease in the content of low-abundance phospholipids and triacylglycerols: 32:3, 34:5, and 34:6 PC, 32:3 PE, and 52:5, 52:6, 52:7, and 52:9 TAG, for which a common structural element is the desaturated 16-C acyl chain. The effect on TAG is very similar to the one in kasII. The dgd2 mutant exhibited an interesting phenotype, showing an accumulation of several saturated PCs, including 32:1, 32:2, 34:1, and 36:1 molecular species, besides a significant accumulation of PE when taking all PE species into account. sqd2 exhibited an expected decrease in SQDG: the most abundant SQDG species decreased up to 2-fold; however, due to the high analytical noise, none of these changes crossed the 0.01 P value threshold. A parallel increase in PG was observed as well as a mixed effect on short-chain TAGs.
Heat Map Representing Results of the Lipidomic Analysis of Selected Knockout Mutant Lines.
The color represents the log2 value of the median mutant/wild-type fold change. Asterisks in the heat map mark the significance of the change (***P < 0.001, 0.001 < **P < 0.01, 0.01 < *P < 0.1). The bar plot parallel to the heat map represents the relative abundance of each lipid species in the frame of the lipid class in wild-type plants (e.g., a value of 60% for MGDG 34:6 means that this molecular species of MGDG represents 60% of the measured MGDG pool).
Match between Environmental Coordination and Genetic Control
The relationship between transcript–lipid correlation in changing environmental conditions and gene function was estimated by matching the results of O2PLS analysis with lipidomic profiles of the selected gene knockout lines. For convenience, we call the transcript–lipid coordination in the O2PLS model “environmental coordination” and the effect of the gene knockout on lipid levels “genetic control.”
The environmental coordination of a gene was calculated as the correlation between its transcript changes and the levels of all the lipids in the frame of the predictive variance structure (and thus the correlation between the joint variance projection of the transcript and the joint variance projections of all lipids). In this way, we estimated the relationship between changes of individual transcripts and lipids irrespective of the unique variation and residual noise. The genetic control was estimated as the effect of the gene knockout on the lipid levels, represented as the mutant to wild-type log2 fold change value. Because the knockout of a gene represents its radical downregulation, lipids depleted in the mutant are treated as positively dependent and those accumulated are treated as negatively dependent. Finally, the relationship between environmental coordination and genetic control of the selected genes was scored by the Spearman correlation coefficient. A significant positive correlation will result if the lipids that positively correlate with the gene of interest are also depleted in the respective mutant line or, analogously, the negatively correlated lipids are accumulated in the mutant.
Among selected genes, KASII, DGD2, SQD2, and CDS2 exhibited significant positive correlation between their environmental coordination and genetic control, whereas LACS4 showed significant negative correlation (Figure 7). Although significant, the correlation coefficients ranged from 0.24 to 0.4, indicating that the major portion of the variance is not explained by the single gene knockout (the significance of the Spearman correlation coefficient has been supported by nonparametric bootstrap analysis; Supplemental Table 4). For KASII, exhibiting the highest correlation, the lipids mostly responsible for the significant match are also those most significantly affected in the kasII mutant (i.e., KASII showed close coordination with the accumulation of highly desaturated phospholipids in darkness), which were conversely deficient in the knockout mutant. The heat-related effect of CDS2 was positively related to the effect of its knockout via the glycerolipid:phospholipid ratio. Inspection of the sum change of the lipid classes (Figure 4A) shows that the accumulation of MGDG and SQDG is associated with heat, whereas PG, PE, and PC accumulation is associated with cold. The same effect observed upon the perturbation of CDS2 expression points to the gene as a strong candidate for acting as a regulator of temperature-related glycerolipid:phospholipid ratio changes. Additionally, molecular species of PC and PE, identified as the most significantly affected lipids by the CDS2 knockout, exhibit the closest correlation with the environmental coordination. A glycerolipid:phospholipid ratio is also affected in DGD2 environmental and genetic phonotypes. In this case, however, the effect of DGD2 knockout coincides with the cold-related accumulation of phospholipids. Yet another scenario was seen for SQD2, which exhibited a significant match between environmental coordination and genetic control for the bulk of lipids, with the exception of two specific lipids, 32:0 and 34:3 PG. Interestingly, LACS4 exhibited a significant but negative match between environmental coordination and genetic control. This was mainly related to the most affected phospholipids in lacs4, 34:6 and 38:6 PC and 36:2 PE, but also occurs for the bulk of nonsignificantly affected lipids, indicating that even slight mutant-induced changes may represent biologically relevant information.
Scatterplots Showing the Match between Environmental Coordination and Genetic Control of the Analyzed Gene Candidates.
Environmental coordination is expressed as Spearman correlation values between joint variance projections of the gene expression and glycerolipids in all applied conditions. Genetic control is described by the log2 value of the median mutant/wild-type fold change. All lipid species are represented by their names plotted in gray; the significant ones, exceeding P > 0.1 in the mutant study, are plotted in black. Above each plot, the Spearman correlation coefficient is provided (bootstrap statistics are given in Supplemental Table 4).
DISCUSSION
We previously described the effects of the various temperature and light treatments on lipid levels independently of the transcript data (Burgos et al., 2011). It is striking, therefore, that all of the described effects are encompassed in the lipid predictive structure of the O2PLS model, which includes only 27% of the original variance. The reason for this is that the lipid predictive variance lacks the high platform-specific technical noise observed in the lipid data set only, leaving only smooth changes clearly related to treatments and response development in time. Investigation of the lipid-unique and residual variance (Supplemental Figures 5 and 6, respectively) indicates that, indeed, all of the nonpredictive variance accounts for changes related neither to the treatment nor to the time dimension of the response. A different situation is observed in the case of transcripts. Here, two major effects discussed by Caldana et al. (2011), a dominance of gene expression changes induced by heat in dark conditions and a halt of circadian changes in cold (described also in Espinoza et al., 2008), are largely encompassed in the unique variance structure. Similarly, the contrasting transcriptional changes in response to 32-L and 4-L conditions were identified as transcript-unique. This demonstrates the advantage of the integrative approach over analysis of the data blocks separately, as it (1) cleaned the data from platform-specific analytical noise, allowing previously hindered biologically interpretable patterns to emerge, and (2) focused the analysis on patterns shared between transcripts and lipids. Here, we discuss the coordination of lipid and gene expression changes, with an emphasis on lipid metabolism gene expression, which was not thoroughly covered by Caldana et al. (2011).
Parallel Changes of Lipids and Transcripts Reflect the Reprogramming of Lipid Metabolism
The applied treatments mostly affected the saturation of acyl chains and to a lesser extent the length of the acyl chains and the pools of the whole lipid classes. This is understandable, considering the short time span of the experiment, which was probably insufficient to cover changes involving the displacement of large carbon pools. As described by Burgos et al. (2011), changes in light intensity, with the accumulation of desaturated lipid species and the decrease of their less desaturated precursors in darkness (in particular MGDG, PC, PE, and SQDG), might be directly related to the inhibition of the de novo synthesis of FAs (Ohlrogge and Jaworski, 1997) accompanied by the continuous activity of desaturases (Browse et al., 1981). Temperature changes, on the other hand, lead to effects that might be related to substrate specificity and shifts in substrate availability of the affected enzymes. This concerns most importantly the desaturases FAD2, FAD3, and SSI2 in heat and the phosphatidic acid binding protein TGD2 in cold treatment. Finally, it is important to note that heat and cold do not trigger the membrane fluidity compensation mechanism during the first 6 h of the stress response, which has been observed 3 d after cold exposure of Arabidopsis rosettes (Welti et al., 2002) and Brassica napus cotyledons (Tasseva et al., 2004). This is an interesting observation considering that the dark-induced change is rapid, showing that plants potentially can modify the saturation level of their membrane lipids within a few hours.
These lipid changes are accompanied by a reorganization of lipid metabolism gene expression along three transcriptional programs, light-specific/temperature-independent, temperature-specific/light-independent, and intermediate light/temperature-modulated, each specific to a certain subset of lipid biochemical pathways (Figures 4B to 4D). To highlight the implications of this observation, we discuss the major pathways in detail.
FA Synthesis
Whereas FA synthesis is known to be light regulated at the level of ACCase by the redox mechanism (Kozaki and Sasaki, 1999), here we observe that genes involved in FA synthesis are also coordinately regulated at the transcriptional level. This involves the dark-induced downregulation of almost all key pathway enzymes, including elements of the FA synthase complex: KASI, KASIII, hydroxyl-ACP dehydrase, and enoyl-ACP reductase ENR1; the biotin carboxylase subunit of ACCase, malonyl-CoA:ACT malonyltransferase, the long-chain acyl-CoA synthetase LACS9, the acyl-ACP synthetase AAE15; and components of the pyruvate dehydrogenase complex: one of the genes for β-PDH (At1g30120) and LPD2 (Figure 4B). Interestingly, two acyl-ACP thioesterases, FatA and FatB, are antagonistically regulated by light. Whereas FatA, exhibiting higher specificity to 18:0-ACP and 18:1-ACP substrates (Salas and Ohlrogge, 2002), is strongly downregulated by dark treatment, FatB, exhibiting higher specificity toward shorter 16:0-ACP substrate, is upregulated in dark conditions. This coincides with the downregulation of steaoryl-ACP desaturase FAB2, potentially further decreasing the flux of ACP-FAs toward FatA. These concerted changes might constitute a mechanism shifting glycerolipid biosynthesis toward shorter acyl chains in response to carbon deficiency in dark conditions, as observed, for example, in seed oil in unfavorable nutrient and light conditions (Li et al., 2006; Ekman et al., 2007). The significant downregulation of the long-chain acyl-CoA synthetase LACS9 suggests that, in addition to the shift toward shorter acyl chains, the FA efflux from the chloroplasts is inhibited, and thus prokaryotic glycerolipid synthesis is promoted. Due to the lack of positional data for the measured lipid molecular species, it is difficult to estimate the change of ratio between lipids originating from eukaryotic and prokaryotic pathways. On the other hand, dark treatment is positively correlated with the occurrence of 34-C lipid species (Supplemental Figure 4C), indicating a larger contribution of shorter 16-C acyl chains.
Prokaryotic and Eukaryotic Glycerolipid Synthesis
Whereas the changes in FA elongation and desaturation suggest a shift toward prokaryotic glycerolipid synthesis, comparison of the expression patterns between genes involved in prokaryotic and eukaryotic glycerolipid synthesis indicates the opposite (Figure 4C). Almost all genes of the prokaryotic pathway are strongly downregulated in response to darkness and upregulated in light conditions. Such coordination of the plastidial glycerolipid synthesis pathway has been observed previously in plants harboring a mutated version of Escherichia coli GDPH gpsAFR insensitive to feedback inhibition (Shen et al., 2010). In these plants, exhibiting 3- to 4-fold higher levels of G3P than the wild type, a coordinated upregulation of the plastidial glycerolipid biosynthetic genes was observed. In darkness, however, Gly-3-P has been shown to significantly decrease both in the normal diurnal cycle (Gibon et al., 2006) and in prolonged night conditions (Usadel et al., 2008). Whereas the correlation between Gly-3-P accumulation and the expression of glycerolipid synthesis genes in darkness gives a first indication of a possible causal relationship, the specific effect of feedback-insensitive GDPH mutants supports the argument that the regulatory mechanisms might be related to Gly-3-P.
On the other hand, the eukaryotic pathway is in large part dark-induced. This concerns, for example, genes involved in eukaryotic diacylglycerol synthesis (putative glycerol-3-phosphate acyltransferase, 1-acylglycerol-3-phosphate acyltransferase 4, and phosphatidate phosphatase 2) and four genes involved in acyl editing: phospholipase A2 (LCAT-PLA) and three 1-acylglycerol-3-phosphocholine acyltransferases (LPLAT, LPLAT1, and LPEAT2). Acyl editing contributes largely to the incorporation of newly synthesized FAs to PC in the sn-2 position and is a key step in the synthesis of polyunsaturated TAG (Bates et al., 2007, 2009; Lu et al., 2009; Bates and Browse, 2012). Here, in conditions of inhibited FA synthesis and in parallel with an increasing pool of desaturated glycerolipids, the activation of the acyl-editing mechanism might play a role in the distribution of the increasing pool of polyunsaturated FAs, which, upon desaturation by FAD2 and FAD3 in PC, return to the pool of acyl-CoA and enter the phospholipid synthesis pathway via 1,2-diacylglycerol-3-phosphate. This observation is supported by the dark-specific upregulation of LACS4 (the only LACS gene up-regulated in response to darkness).
Storage Lipid Metabolism
Although TAGs were not measured, it is possible that the observed coupling of light- and temperature-specific effects in the lipid data set originates from the parallel action of temperature-regulated storage lipid metabolism and mostly light-regulated pathways of FA biosynthesis and glycerolipid metabolism. Genes of all the core enzymes of β-oxidation are specifically upregulated in 21-D and 32-D conditions, including the TAGL lipase SDP1, the acyl-CoA oxidases ACX1, ACX2, and ACX4, and the main isoform of 3-ketoacyl-CoA thiolase (KAT2/PED1) (Germain et al., 2001; Carrie et al., 2007), with the exception of the multifunctional protein MFP2 gene, exhibiting an opposite behavior. This is a surprising observation, taking into account that ACX1,2,3,4 and MFP2 have been shown to be coordinately upregulated during rapid oil breakdown in postgerminative growth (Eastmond and Graham, 2000; Rylott et al., 2001). The dark-related expression of peroxisomal enoyl-CoA hydratase (ECH2) indicates an increased degradation of unsaturated FAs (Goepfert et al., 2006). Similar to other pathways, the expression of genes involved in FA degradation is boosted in heat stress and compromised by cold. Whereas the regulation of β-oxidation genes, including ACX, MFP2, and KAT2, has been described at the transcriptional and posttranscriptional levels during postgerminitive seedling growth (Hayashi et al., 1998; Richmond and Bleecker, 1999; Eastmond and Graham, 2000; Eastmond et al., 2000; Rylott et al., 2003; Goepfert et al., 2005), our observations indicate that the regulation of the β-oxidation pathway genes is also coupled with light signaling. This is likely related to the role of KAT2 in abscisic acid signaling (Jiang et al., 2011), since β-oxidation in darkness is a source of reactive oxygen species (Mittler, 2002). Coordinated changes in the expression of other genes involved in β-oxidation suggest that the transcriptional signal might not be limited to KAT2. Conversely, the key genes involved in TAG biosynthesis exhibit minor light-related effects.
Identified Transcript–Lipid Relationships Are Significantly Linked to the Genetic Control of Membrane Lipid Composition
Lipidomic analysis of the knockout mutant lines indicates a causal relationship between observed transcript changes and changes in glycerolipid levels. Whereas in the case of LACS4 this relationship is negative, indicating a more complex, context-dependent link, KASII, DGD2, SQD2, and CDS2 exhibit direct correlation between their environmental coordination and genetic control.
KASII, responsible for a condensation reaction from palmitoyl-ACP to stearoyl-ACP (Huang et al., 1998), exhibits the most significant match between its environmental coordination and genetic control, coming mostly from the affected phospholipid:glycerolipid ratio. Two previously studied KASII knockout alleles, fab2-1 and fab2-2, exhibited a significant effect on FA abundance, decreasing the level of 18-C FAs and increasing the 16-C (Carlsson et al., 2002). This result was characterized in storage oil composition; however, no glycerolipid analysis has been shown so far. The transcript data show that the upregulation of KASII expression in darkness coincides with the promotion of the incorporation of 16-C acyl chains into glycerolipids, which is an opposite effect to the one expected from the data of Carlsson et al. (2002). This is an interesting observation, suggesting a new putative function for the transcriptional regulation of KASII in changing light conditions. The match between the environmental and genetic phenotypes of DGD2 is mainly due to the accumulation of specific saturated phospholipids, although, in agreement with Kelly et al. (2003), the phospholipid:glycerolipid ratio is not affected in the dgd2 mutant. The accumulation of saturated PE and PC might indicate that temperature-dependent regulation of DGD2 expression is involved in the partial redirection of the de novo–synthesized FAs to phospholipids. In the case of SQD2, the knockout mutant phenotype (across all the SQDG species, although not highly significant) matches the decrease in SQDG content in dark conditions, where SQD2 gene expression is downregulated. A previous study reported the complete absence of sulfolipids in the sqd2 mutant, indicating that our line contains a functional protein, although probably with decreased activity (Yu et al., 2002). In that case, SQD2 provides a perfect example of a clear lipidomic phenotype occurring both upon environmentally induced SQD2 gene expression changes and upon reduction of SQD2 expression/activity upon genetic perturbation. Finally, the phonotype of the cds2 mutant and its heat-specific response indicate that CDS2 might be a key enzyme involved in the temperature-related change in phospholipid:glycerolipid ratio. CDS2 is one of five cytidine diphosphate diacylglycerol synthases (Beisson et al., 2003) and is responsible for transfer of the cytidyl group from cytidine triphosphate (CTP) to phosphatidic acid and the formation of CTP-DAG, a substrate for PI and PG biosynthesis.
Although not perfect, the described match between environmental coordination and genetic control is noteworthy, especially taking into account that (1) our study included only two environmental parameters; (2) gene knockouts are harsh genetic perturbations and often lead to strong pleiotropic effects, making it difficult to associate an observed phenotype with the gene function; and (3) metabolic pathways exhibit a remarkable ability to compensate the effects of single gene perturbations (Papp et al., 2004). It is important to note, however, that the selected genes represent only a limited data set. Thus, the knockout mutant analysis should be regarded as a case study and an attempt to evaluate the usefulness of O2PLS analysis for predicting single gene function and proposing biologically sound hypotheses.
Future Challenges
In contrast to other integrative studies focused on the identification of pairwise metabolite–gene relationships (Nikiforova et al., 2005; Rischer et al., 2006; Hirai et al., 2007), here we described lipid–transcript associations at the systems level and supported our findings by testing a set of gene knockout lines. Such system-scale integration previously has been shown to be a successful strategy for investigating housekeeping metabolic processes and for considering various types of data and different computational methods (Wentzell et al., 2007; Kerwin et al., 2011). We believe that our analysis is an important step in the emerging field of systems biology of plant lipid metabolism, and we expect that this work, complemented by other system-scale studies, such as high-resolution genome-wide association mapping and large-scale functional genomics, will give an even deeper insight into the principles of plant lipid metabolism regulation.
METHODS
Environmental Treatment Experiment
For the details on plant growth conditions, applied treatments, and sampling procedures, see Caldana et al. (2011). The details on microarray and lipidomic analyses are given in Caldana et al. (2011) and Burgos et al. (2011), respectively. Genes involved in lipid metabolism were selected according to the ARALIP database (http://aralip.plantbiology.msu.edu; version from January 2013) (Li-Beisson et al., 2013).
Knockout Mutant Lines: Selection, Genotyping, and Growth Conditions
Arabidopsis thaliana Columbia-0 ecotype (wild-type) plants were used as controls throughout this work. Arabidopsis T-DNA insertion lines, SALK and GABI-KAT (Alonso and Stepanova, 2003; Kleinboelting et al., 2012), were obtained from the Nottingham Arabidopsis Stock Centre collection (http://arabidopsis.info). Plants were selected on plates supplemented with appropriate antibiotics in order to obtain nonsegregating homozygous lines. See Supplemental Table 2 for the list of mutant lines and genotyping and gene expression primers. Quantitative PCR analysis of the mutants lines was performed as described by Brotman et al. (2013) with gene-specific primers after the insertion. The mutant lines (six independent biological replicates for each line) were grown in trays under a short-day regime (22°C, 8 h of light and 16 h of dark). Each tray contained in addition six wild-type plants, and all the plants were randomly placed. Four-week-old plants (rosette leaves) were collected and shock frozen in liquid nitrogen for subsequent lipidomic analysis. All batches (mutants and wild-type plants) were subsequently extracted and processed together.
Lipid Profiling of the Mutant Lines
Samples were processed using ultra-performance liquid chromatography with a C8 reverse-phase column coupled with an Exactive mass spectrometer (Thermo-Fisher; http://www.thermofisher.com) in positive ionization mode. Processing of chromatograms, peak detection, and integration were performed using REFINER MS 7.5 (GeneData; http://www.genedata.com). Processing of mass spectrometry data included the removal of the fragmentation information, isotopic peaks, as well as chemical noise. The obtained features (m/z at a certain retention time) were queried against an in-house lipid database for further annotation (details on compound annotation are given in Supplemental Data Set 3).
Data Integration
O2PLS analysis (Trygg, 2002; Trygg and Wold, 2003) of the transcript and lipid data was performed using the algorithm provided by Bylesjö et al. (2007). All the calculations were performed using R (http://www.R-project.org) (R Core Team, 2013) and the pcaMethods (Stacklies et al., 2007) and pls (Mevik and Wehrens, 2007) packages. In all steps of the O2PLS algorithm that required the extraction of principal components, the singular value decomposition method was used. Below we describe briefly the method principle.
O2PLS is a bidirectional multivariate regression method that aims to separate the covariance between two data sets (it was recently extended to multiple data sets) (Löfstedt and Trygg, 2011; Löfstedt et al., 2012) from the systematic sources of variance being specific for each data set separately. In principle, it identifies a set of components P being common for X and Y and, as such, it represents the joint variance between both data sets. By removing this joint variance, it is possible to identify orthogonal components PX and PY specific for each respective data set. The O2PLS model for X and Y data sets can be written as:
where P and U are score matrices for X and Y, respectively, W and C are the joint variance component matrices, and E and F are the residual matrices. 
and 
are the Y-orthogonal loadings and score matrices, respectively. Analogously, 
and 
are the X-orthogonal loadings and score matrices, respectively. Equation 3 describes the inner relation between U and T, where H is a residual matrix. Predictive equations for 
and 
are then:
In our analysis, the transcript data set (480 genes × 152 samples) and the lipid data set (92 lipids × 152 average values of three biological replicates) were referred to as the X and Y matrices, respectively. Prior to the O2PLS analysis, both lipid and transcript data sets were column-wise mean-centered and scaled to unit variance. Subsequently, in order to ensure that transcripts and lipids have equal weight, both data sets were scaled to give a total sum of squares of 1. In order to avoid overfitting of the model, the optimal number of latent variables for each model structure was estimated using group-balanced MCCV. For each combination of the latent variable’s number (between 4 and 15 for the predictive structures and 1 to 10 for the orthogonal structures), 10-fold MCCV was performed. From these 10 permutations, the average Q2 values were calculated as:
where
For the top 10 models, 10-fold MCCV was performed 50 times and new average Q2 values were calculated. From three best almost equally scored models with the lowest generalization error, the one with the lowest number of latent variables was chosen (Supplemental Table 1). The residual variance structures were analyzed by PCA, showing no treatment-related effects, although several samples in the lipid data set could be described as technical outliers (Supplemental Figure 6).
The O2PLS-DA analysis was performed as described by Bylesjö et al. (2007); briefly, the O2PLS predictive variation [TWT, UCT] was used for a subsequent O2PLS-DA analysis (additional O2PLS-DA analysis of transcript- and lipid-unique variations is shown in Supplemental Figure 7) . O2PLS-DA models with one to eight latent variables were tested, but no significant improvement, in terms of multiple independent light- or temperature-specific effects, was observed for n > 1 (Supplemental Figure 8). In order to improve the visualization of the light- and temperature-specific effects, the O2PLS-DA results (projection of the predictive variation [TWT, UCT] on the first latent variable of the respective O2PLS-DA model) were regressed onto the light and temperature variables using partial least squares regression. The weights for the light and temperature variables were chosen equidistant, meaning that the distance between darkness and low light was set equal to the distance between low light and control light and the distance between normal light and high light.
Generalized Lipid and Transcript Trends
The analytical method for lipid quantification allows extracting only relative information about the lipid changes, and the comparison of abundances of particular lipid species is additionally possible only in the frame of a single glycerolipid class. Therefore, when summing up the changes in the frame of a single lipid class (as in Figure 4A), the same for a particular number of double bonds or carbons in the acyl chains is an approximation (hence, the intensities of the lipid species of different classes are summed). This approximation is performed by within-class normalization of the species average abundance. The averages of the species abundances across all the conditions have been divided by their sum, such that they sum up to 1. Newly obtained values were used to normalize the intensity values across the treatments. As a result, the weight of each class in the computation of general saturation and carbon number trends becomes equal.
Supplemental Data
The following materials are available in the online version of this article.
Supplemental Figure 1. Cross-Validation of the O2PLS Model.
Supplemental Figure 2. Significance of the Transcript and Lipid Changes in the Frame of Lipid Metabolism Pathways.
Supplemental Figure 3. Top Five Latent Variables of the O2PLS Predictive Variance Structure.
Supplemental Figure 4. O2PLS-DA Correlation Loading Plots.
Supplemental Figure 5. Top Four Latent Variables of the Lipid- and Transcript-Unique Variance Structures.
Supplemental Figure 6. PCA of the Lipid and Transcript Residual Variance.
Supplemental Figure 7. O2PLS-DA of the Transcript- and Lipid-Unique Variance Structures.
Supplemental Figure 8. First Five Latent Variables of O2PLS-DA Models Calculated for Light and Temperature Groups.
Supplemental Table 1. Top Scoring O2PLS Models and Their Generalization Error for the Transcript and Lipid Data.
Supplemental Table 2. List of Primer Sequences for Mutant Genotyping and Quantitative PCR Analyses.
Supplemental Table 3. Verification of Gene Knockout by Expression Analysis Using Quantitative RT-PCR.
Supplemental Table 4. Bootstrap Statistics for Figure 7.
Supplemental Data Set 1. Lipidomic Analysis of Selected Knockout Mutant Lines.
Supplemental Data Set 2. Median Fold Change with Respect to the Wild-Type Control (Source Data for Figure 6).
Supplemental Data Set 3. Metabolite Reporting Checklist and Compound Annotation.
Acknowledgments
We thank Ke Xu, Izabela Sierzputowska, and Aenne Eckardt for technical assistance, Asdrubal Burgos for discussions, and Elmien Heyneke for help with writing.
AUTHOR CONTRIBUTIONS
J.S., A.C.-I., and Y.B. performed the analysis. J.S. wrote the article. L.W. and A.C.-I. supervised the work.
Footnotes
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: Jedrzej Szymanski (szymanski{at}mpimp-golm.mpg.de).
↵[OPEN] Articles can be viewed online without a subscription.
↵[W] Online version contains Web-only data.
Glossary
- MGDG
- monogalactosyldiacylglycerol
- TAG
- triacylglyceride
- FA
- fatty acid
- PG
- phosphatidylglycerol
- PCA
- principal component analysis
- MCCV
- Monte Carlo cross-validation
- PI
- phosphatidylinositol
- PE
- phosphatidylethanolamine
- SQDG
- sulfoquinovosyldiacylglycerol
- PC
- phosphatidylcholine
- Received September 22, 2013.
- Revised January 27, 2014.
- Accepted February 17, 2014.
- Published March 18, 2014.
References
