Metabolomic characterization of sunflower leaf allows discriminating genotype groups or stress levels with a minimal set of metabolic markers

Metabolomic characterization of sunflower leaf allows discriminating genotype groups or stress... Introduction Plant and crop metabolomic analyses may be used to study metabolism across genetic and environmental diversity. Complementary analytical strategies are useful for investigating metabolic changes and searching for biomarkers of response or performance. Methods and objectives The experimental material consisted in eight sunflower lines with two line status, four restorers (R, used as males) and four maintainers (B, corresponding to females) routinely used for sunflower hybrid varietal production, respectively to complement or maintain the cytoplasmic male sterility PET1. These lines were either irrigated at full soil capacity (WW) or submitted to drought stress (DS). Our aim was to combine targeted and non-targeted metabolomics to characterize sunflower leaf composition in order to investigate the effect of line status genotypes and environmental condi- tions and to find the best and smallest set of biomarkers for line status and stress response using a custom-made process of variables selection. Results Five hundred and eighty-eight metabolic variables were measured by using complementary analytical methods such as H-NMR, MS-based profiles and targeted analyses of major metabolites. Based on statistical analyses, a limited number of markers were able to separate WW and DS samples in a more discriminant manner than previously published physiological data. Another metabolic marker set was able to discriminate line status. Conclusion This study underlines the potential of metabolic markers for discriminating genotype groups and environmental conditions. Their potential use for prediction is discussed. Keywords Metabolic markers · Metabolomics · Sunflower · Water stress · Maintainer–restorer lines AbbreviationsDS Drought stressed AQC 6-Aminoquinolyl-N-succinimidyl DW Dry weight carbamate FAMES Fatty acid methyl esters AUC Area under curve LASSO Least absolute shrinkage and selec- B Maintainer line tion operator CID Carbon isotope discriminationLC–ESI–QTOF–MS Liquid chromatography–electro- CV Cross-validation spray-ionization–time-of-flight– DAG Days after germination mass spectrometry NMR Nuclear magnetic resonance OSM_POT Osmotic potential Electronic supplementary material The online version of this PCA Principal component analysis article (https ://doi.org/10.1007/s1130 6-019-1515-4) contains PET Petiolaris supplementary material, which is available to authorized users. PLS Partial least squares PLS-DA Partial least squares discriminant * Olivier Fernandez olivier.fernandez@univ-reims.fr analysis R Restorer line Extended author information available on the last page of the article Vol.:(0123456789) 1 3 56 Page 2 of 14 O. Fernandez et al. SLA Specific leaf area Therefore, the B line is widely used for phenotypic and agro- VIP Variable importance in the nomic description of the line. projection Since the introduction of hybrid varieties, sunflower has WW Well-watered undergone an active breeding process (Vear 2016), mainly thanks to molecular marker-assisted selection. Hybrids have been selected with increased resistance to downy mildew (Qi et al. 2016), sclerotinia (Talukder et al. 2014) and water 1 Introduction stress (Marchand et al. 2013; Owart et al. 2014), although sunflower is often cited as moderately drought-tolerant (Hus- Sunflower ( Helianthus annuus L.) is the fourth major crop sain et al. 2018). This selection process will benefit from providing seed for oil production worldwide. In 2016, the recent sequencing of the maintainer inbred line XRQ world production reached 45 MT from 26 Mha, principally (Badouin et al. 2017). As part of these selection efforts, our in Europe (around 70%), Ukraine being the world leader group is currently involved in searching for metabolic mark- (Oilworld 2016; Hussain et al. 2018). Worldwide produc- ers of sunflower performance. A definition of biomarkers tion has increased constantly ever since (Oilworld 2016). (and their sub-category metabolic markers) emerged from Sunflower accounts for more than 50% of total world table- the field of medicine as a characteristic objectively measured oil consumption. Additionally, its high biodegradability to indicate a given biologic, pathologic or pharmacologic makes it suitable for non-alimentary uses such as in paints response (Fernandez et al. 2016). In plant science, meta- and bioplastics. bolic markers have been defined as metabolites or groups Native to North America and introduced into Europe of metabolites that are measured to predict or discriminate in the sixteenth century, sunflower became a major crop plant responses or performance (Fernandez et al. 2016). in this area in the early 1960s. Further development was The use of metabolic markers to predict criteria of plant achieved after the introduction of hybrid varieties in the performance is recent, with pioneering papers dating from early 1980s. Hybrid varieties are based on the use of cyto- the early 2010s (Meyer et al. 2007; Riedelsheimer et al. plasmic male-sterile (CMS) lines (Vear 2016), like many 2012). The possibilities offered by these markers in plant other crops (Chen and Liu 2014). The male sterility used for selection processes were reviewed recently and a pipeline sunflower hybrid production, called PET1-CMS, was first to search and use them has been proposed (Fernandez et al. identified from an interspecific cross between Helianthus 2016). The authors emphasized that the search for meta- petiolaris and H. annuus. It results from the reorganization bolic markers requires a first step of analysis on a small of mitochondrial DNA that generated a new open reading core set of genotypes. The present article investigates this frame ORFH522 co-transcribed with apt1 gene and coding first step, which includes (1) testing the analytic pipeline to a 16 kDa protein. This leads to modified mitochondrial func- establish the dynamic range of targeted metabolites, (2) con- tions and affects pollen development (Balk and Leaver 2001) firming the presence of several secondary or “specialized” through a decline in the mitochondrial membrane integrity metabolites (as defined by Hartmann 2007; Pichersky and and the respiratory control ratio. The mitochondrial pro- Lewinsohn 2011) and (3) investigating which metabolites tein ORFH522 appears to be expressed in all tissues, but are essential for differentiating groups of samples such as, in the deleterious phenotype associated with PET1-CMS has our case, water treatment (well-watered, WW, vs. drought- been thought to be limited to the anthers, and no apparent stressed, DS) and line status (maintainer, B, vs. restorer, R). extra phenotypes have been found in other organs (Horn and These metabolites could later serve as metabolic markers. Friedt 1999; Balk and Leaver 2001). Furthermore, we tested different statistical methods for vari- To complement the mutational effect, a nuclear restora- able selection in order to find the best and smallest sets of tion gene (noted Rf1) is used in sunflower hybrid production. metabolite markers. Indeed, for a given agronomical trait, Restoration genes are nuclear and generally encode tetratri- the deployment of metabolic markers among breeders will copeptides that are thought to transcriptionally control the depend on their cost (Fernandez et al. 2016). CMS mitochondrial gene (Chen and Liu 2014; Igarashi et al. For this purpose, we used a combination of targeted and 2016; Yu et al. 2016). Finally, sunflower hybrid production untargeted metabolomic analyses on sunflower leaf samples is based on crossing a restorer line called R bearing a func- obtained from B or R lines and in WW and DS conditions. tional restoration allele Rf1 (that recovers the PET1-CMS Our results show that a limited number of markers can male-sterility phenotype) to a male-sterile PET1-CMS line clearly differentiate WW from DS samples and in a more called A (carrying a recessive rf1 allele). To maintain this discriminant manner than the physiological data presented in male-sterile line, a maintainer line called B, isogenic to the Blanchet et al. (2018), which are classically used to discrimi- A line, is also used. Each B line carries the rf1 allele but is nate individuals subjected to DS. To our surprise, another male-fertile, as it does not carry the CMS-PET1 cytoplasm. leaf metabolic marker set was able to discriminate B lines 1 3 Metabolomic characterization of sunflower leaf allows discriminating genotype groups or… Page 3 of 14 56 from R ones. Our data underline the potential of metabolic Starch (Hendriks et al. 2003) and protein (Bradford 1976) markers for discriminating genotypes and environmental contents were determined on the pellet. Assays were carried conditions. Their potential use in sunflower breeding for out in 96-well microplates. performance prediction is discussed. Individual free amino acid analysis was carried out using an UPLC separation with fluorescent detection after deri- vatization using 6-aminoquinolyl-N-succinimidyl carbamate 2 Materials and methods (AQC)-tag (a method hereafter referred to as UPLC-Fluo). For lipid analysis, fatty acid methyl esters (FAMES) The protocols used are detailed in Online Resource 1 and were measured after hydrolysis of 20 mg dry weight (DW) summarized here. with 2.5% H SO (v/v) in methanol. GC-FID was per- 2 4 formed using an Agilent 7890 gas chromatograph (Agilent, 2.1 Plant material and growth conditions Santa Clara, California) equipped with a Carbowax column (15 m × 0.53 mm, 1.2 µm; Alltech Associates, Deerfield, IL, The experiment was performed in 2013 in the phenotyping USA) and flame ionization detection. FAMES were identi- platform “Heliaphen” (Gosseau et al. 2018). Eight sunflower fied by comparing their retention times with commercial lines, four B and four R lines, were grown in two conditions fatty acid standards (Sigma, Saint-Quentin Fallavier, France) (WW and DS) with three replicates, leading to a total of 48 and quantified using ChemStation (Agilent). samples. Irrigation was stopped at 38 days after germination (DAG; Schneiter and Miller 1981) for DS plants. Soil evapo- 2.4 H‑NMR analysis of major polar compounds ration was estimated according to Marchand et al. (2013). Both WW and DS plants were weighed four times per day by Polar metabolites were extracted from lyophilized powder the Heliaphen robot to estimate plant transpiration (Gosseau (40 mg DW per biological replicate) with an ethanol–water et al. 2018). At 47 DAG, leaves for metabolomic analyses series (80/20, 50/50, 0/100 v/v) at 80 °C as described in were harvested without their petiole and frozen in liquid Deborde et al. (2009) with modifications. This three-step nitrogen. Two other leaves (mature and young leaves) were extraction process (ethanol–water series) was chosen to take harvested for physiological trait measurements. During the into account the diverse affinities and solubilities of leaf experiment, two samples were excluded before leaf sam- major polar compounds (i.e. sugars, organic acids, amino- pling (excessive irrigation was detected when analysing n fi al acids) for ethanol or water, in order to obtain an accurate Heliaphen readings) and four samples could not be analysed view of these compounds in leaf extracts. The 1D (cpmg because of insufficient powder quantity. This resulted in a and single-pulse) spectra were processed using the NMR- total of 42 samples submitted to metabolic analyses. ProcFlow application v1.1 (Jacob et al. 2017; http://nmrpr ocflo w.org/). For the cpmg dataset, this resulted in 479 nor- 2.2 Physiological trait measurements for plant malized variables corresponding to spectral regions (named phenotyping Unk_ppm:number in Online Resource 2) which included compounds that were annotated later on. The assignments Plant and leaf physiological data are part of a larger dataset of metabolites in the H-nuclear magnetic resonance (NMR) presented in Blanchet et al. (2018). Specific leaf area (SLA) spectra were made by comparing the proton chemical shifts was determined according to Allinne et al. (2009). Both leaf with public or local spectral databases and by spiking the osmotic potential (OSM_POT) and leaf osmotic potential at samples with the corresponding commercial compounds. 2D full turgor (OSM_POT_100) were measured as described experiments were performed on a representative selected in Poormohammad Kiani et al. (2007). To assess carbon extract taken from the WW condition. Quantification of 11 isotope discrimination (CID), samples were oven-dried, identified compounds was performed by using quantified ground, weighed and analysed using a continuous low iso- single-pulse spectra dataset and calibration curves. tope ratio mass spectrometry at the Stable Isotope Platform SHIVA (University of Toulouse, France). 2.5 LC–ESI–QTOF–MS untargeted analysis of semi‑polar metabolites 2.3 T argeted compound measurements Liquid chromatography–electrospray-ionization–time-of- For each sample, about 20 mg fresh weight were extracted flight–mass spectrometry (LC–ESI–QTOF–MS) profiling as in Hendriks et al. (2003). Sucrose, glucose, and fructose of aqueous methanol extracts containing 0.1% formic acid (Jelitto et al. 1992), malate (Nunes‐Nesi et al. 2007), citrate was performed with extracts obtained from 20 mg DW lyo- (Tompkins and Toffaletti 1982) and glucose-6-P (Gibon philized powder. An Ultimate 3000 HPLC (Dionex, Sun- et al. 2002) were determined in the ethanolic supernatant. nyvale, CA, USA) was used to separate metabolites on a 1 3 56 Page 4 of 14 O. Fernandez et al. reversed-phase C18 column using an acetonitrile gradient in Resource 3—Fig. S1. We targeted these compounds because acidified water. Metabolites were detected by using a hybrid they are (1) often considered as putative metabolic mark- quadrupole/time-of-flight mass spectrometer (micrOTOF-Q , ers (Fernandez et al. 2016) and (2) valuable candidates for Bruker Daltonics, Bremen, Germany). Electrospray ioniza- a high-throughput metabolic marker approach, as they are tion in positive mode was used to ionize the compounds. easy and cheap to measure. A quality control sample (QC) was injected after each set The concentrations of these 29 compounds were summed of ten samples. The MS data were processed using XCMS to estimate their contribution to leaf biomass. This yielded (Smith et al. 2006) and R scripts for filtering. A total of about 45% of leaf dry mass. Glucose was found to be the −1 1519 features were detected and reduced to 540 metabolic major soluble sugar. Its concentration (32–45 mg g DW) variables after filtering. The corresponding MS-based vari- was in the same range as that of sucrose, but 8–10 times ables were named using their nominal masses in dalton and higher than fructose depending on the chosen conditions. retention time in seconds in Online Resource 2 (MxxxTyyy). Glutamate, alanine and serine were found to be the most Metabolite identification was performed using the accurate- abundant amino acids. In leaves, linolenic acid (C18:3) was −1 mass data and Orbitrap (Thermo Fisher, Villebon-sur-Yvette, the most abundant fatty acid (7.5–18.6 mg g DW), fol- France) MS and MS/MS data of a representative sample lowed by linoleic acid (Fig. 1). extract. H-NMR profiling was performed on polar extracts to further analyse metabolites from primary metabolism in the 2.6 Statistical analyses millimolar range. Four hundred and seventy-nine regions were observed in the H-NMR cpmg dataset, of which 20 All statistical analyses were performed using the R Soft- compounds were annotated (Online Resource 4). Eleven ware (http://www.r-proje ct.org/), the R package mixOmics identified compounds were measured and quantified with the (Rohart et al. 2017) and the BioStatFlow online tool (bio- H-NMR quantitative single-pulse dataset, but only nine of statflow.org) which is based on R scripts. Two-way ANOVA them were kept in the final dataset to avoid redundancy with with FDR correction was performed to highlight line status targeted spectrophotometric measurements. When summed, or water-treatment effects and interaction. The parameters these compounds represented an additional 5% of the leaf used for partial least squares-discriminant analysis (PLS- dry mass (Online Resource 4). DA) in BioStatFlow were adjusted to a tenfold cross-valida- LC–ESI–QTOF–MS analysis of semi-polar extracts was tion (CV) to generate the model (and calculate the Q ) and performed to analyse specialized metabolites. The most 200-randomized permutations to estimate the robustness of intense peaks that were detected in the sample extracts, the generated model. Some graphical outputs for PLS-DA based on their intensity in the XCMS table generated by a were produced by mixOmics, using the same parameters relative area under curve (AUC) approach, were tentatively than with BioStatFlow. An additional R script from Fu et al. annotated. Orbitrap-MS data were used in order to gain pre- (2017) was used to perform least absolute shrinkage and cision on mass measurement and to perform MS/MS. Online selection operator (LASSO) and sparse partial least square Resource 5 shows the annotation table generated using a (sPLS) selection. Principal component analysis (PCA) and representative spectrum of a leaf extract with annotation of partial least square (PLS) were performed on data mean- the most intense peaks. The two most intense peaks were centred and scaled to unit variance. All statistical analyses annotated as mono and di-caffeoyl quinic acid. With a reten- were performed on the data set in Online Resource 2 or tion time around 17–20 min, several methylated flavonoids subsets of this file. were also detected. Finally, three smaller peaks ranging in retention time from 15 to 17 min were found to putatively represent sunflower sesquiterpenoids. Several peaks after 3 Results 25 min remained elusive. Several metabolite concentrations differed between the 3.1 Sunflower leaf metabolic contents measured conditions, as highlighted by a two-way ANOVA (p < 0.05 by targeted and untargeted approaches with FDR correction, Online Resource 6—Table S1a). In total, 27 metabolites plus starch and protein content were 3.1.1 Difference between DS and WW samples targeted and quantified in sunflower leaf. Major soluble sug- ars (i.e. the ones with the highest content), organic acids The most striking difference was the large increase in each and chlorophylls were quantified with spectrophotometric individual amino acid concentration found for DS sam- analyses. FAMES and free amino acids were measured by ples, with an average increase of 15-fold, (Fig. 1a, Online using GC-FID and UPLC-fluo, respectively. These data are Resource 6—Table S1a). On the other hand, starch, protein presented for the different conditions in Fig.  1 and Online content, linolenic and palmitoleic acids were slightly but 1 3 Metabolomic characterization of sunflower leaf allows discriminating genotype groups or… Page 5 of 14 56 Fig. 1 Concentrations of 27 metabolites measured by targeted meth- Maintainer B lines (white bars) or restorer R lines (black bars). Ver- ods (UPLC-Fluo for amino acids, GC-FID for FAMES, spectropho- tical bars represent standard deviations. Asterisk indicates variables tometry for others) in leaf of B or R sunflower lines cultivated in two that were found significantly different between groups after two-way −1 conditions (WW and DS). Results are expressed in mg  g DW in ANOVA test (p value < 0.05) the four types of samples. a WW (white bars) or DS (black bars). b significantly lower in DS (Fig.  1a; Online Resources 3—Fig (two-way ANOVA test; Online Resources 4 and 6—Tables S1a and 6—Table S1a). Minor differences in starch, pro- S1a). For the other nine compounds identified in H-NMR tein, soluble sugars and GABA were observed between B spectra (amino acids and sugars), excellent correlations and R lines (Fig. 1b and Online Resource 3—Fig S1b) but were found with spectrophotometric and chromatographic none of them was statistically significant (Online Resource targeted methods (data not shown). 6—Table S1a). Finally, only a small group of m/z were significantly Among the variables that were highly significant under different under DS (Online Resource 6—Table  S1a). DS (two-way ANOVA test), most of them were uniden- Four of them were putatively annotated as heliannuol, tified H-NMR spectra regions (Online Resource 6— 3-O-caffeoylquinic acid, tryptophan and phenylalanine. Table S1a). Among them, myo-inositol, glycine betaine The last two were also detected by the UPLC-fluo tar - and trigonelline were significantly higher under DS, geted method. whereas chlorogenate and formate were significantly lower 1 3 56 Page 6 of 14 O. Fernandez et al. 3.1.2 Difference between B and R samples Full data set 1048 variables For the B and R lines, no targeted metabolites were sig- nificantly different. Two unidentified H-NMR variables had a p value < 0.05 (Unk_6.8936 and Unk_3.8733, Online Elimination of redundantvariables (« curation ») Resource 6—Table S1a,). However, except for chlorogenic (RMN annotation, clustering and correlation >,85) acid, most organic acids measured displayed a lower con- centration in R lines leaf samples (Fig. 1b, Online Resource 4). Finally, the rest of the variables that were found signifi- Curateddata cantly different for line status were unidentified MS-based set 588 variables variables (Online Resource 6—Table S1a), except for two putatively annotated flavonoids (Online Resources 5 and 6—Table S1a). Data filtering– variable selection ANOVA –sPLS-LASSO 3.2 Workflow for identifying metabolic markers of water treatment and line status «Top » «Top » Water treatment Line status variables variables The analytical methods allowed the generation of a matrix of 1048 metabolic variables (Online Resource 2). This matrix included 27 targeted metabolites, starch, total protein content Final predictive models and 9 annotated H-NMR variables. The remaining variables PLS-DA were composed of H-NMR unidentified spectral regions and 540 MS-based signatures. The matrix was processed Fig. 2 Description of the statistical analysis pipeline used in this arti- through a three-step biostatistical pipeline to select the more cle relevant variables to discriminate samples according to water treatment and line status: (1) elimination of redundant vari- 3—Fig. S2b), in the 2D space based on the first two latent ables, (2) variable selection for each sample cluster and (3) variables. Predictive ability (Q ) and proportion of vari- final PLS-DA model calculation (Fig.  2).ance (R ) explained by the model were higher than 0.9 and 0.8 in both cases (Table 1), respectively. Each model was 2 2 3.2.1 Elimination of redundant metabolic variables considered as valid as it bore Q and R values above 0.4 and 0.5, respectively (Patil et al. 2016). However, in a high- Since a single metabolite can be encompassed within sev- throughput approach, it is impractical to measure more than eral H-NMR buckets or MS-based ions, we first reduced 500 variables to discriminate or predict cluster differentia- this full data set by hierarchical clustering (BioStatFlow, tion. Therefore, our next step was to test a variable selection Pearson correlation, average linkage as aggregation method). process and to assess the validity of group discrimination Clusters were generated with a correlation threshold of 0.85. with PLS-DA after this selection. PLS-DA was chosen to Within each cluster, MS-based metabolic variables corre- easily compare model performance using Q values. sponding to adducts or isotopes were eliminated while the one with the highest AUC was kept. For H-NMR buckets, 3.2.2 Metabolic variable selection process we used a similar process in order to keep buckets bearing the highest AUC. After this curation process (Fig. 2), the To select variables, we compared three different methods new dataset comprised 588 variables (Online Resource 7). for each condition (DS or line status), a generalised univari- We then tested the discrimination potential of this curated ate method (one-way ANOVA) and two multivariate ones data set on our sample groups using an unsupervised sta- (sPLS and LASSO penalty; Fu et al. (2017); Fig. 2). The tistical approach. PCA was first carried out (Fig.  3). The 588-variable data matrix (Online Resource 7) was submit- first two components displayed in Fig.  3a (water treatment) ted to these methods and subsequent PLS-DAs were per- 2 2 and Fig.  3b (line status) explained 25% of the total vari- formed. We compared the Q and R to assess the quality ability. The separation of our sample groups was incom- of the variable selection process for each resulting PLS-DA plete, although slightly better for DS. We then performed model (Table 1). Since our objective was to find the small- a supervised method (PLS-DA) on this 588-variable data- est possible variable set, we analysed datasets of different set for each type of sample group. Each PLS-DA analysis sizes (90, 50 and 20 variables for water treatment; 35 and 20 was able to discriminate WW from DS samples (Online variables for line status). We dimensioned the first selected Resource 3—Fig. S2a), and B from R lines (Online Resource data set size according to the numbers of variables with a p 1 3 Metabolomic characterization of sunflower leaf allows discriminating genotype groups or… Page 7 of 14 56 Fig. 3 PCA scores plot (PC1 x PC2 plan) generated with the full set DS, orange dots. b Highlighting line types. B, red dots and R, blue of 588 metabolic variables (Online Resource 7) measured in sun- dots. Coloured ellipses represent 95% confidence level. The connect- flower leaf cultivated in a Heliaphen phenotyping platform. a High- ing lines attach each individual point to the centre of the confidence lighting samples with different water treatment. WW, green dots and ellipse Table 1 Comparison of 2 2 Variable selection Condition Data set size Q R Expl var t1/ CV p-value predictive ability (Q ) and year (%) explained variance explained (R ) of the different PLS-DA None Water treatment 588 Variables 0.936 80.2 1.1E−04 models calculated with different Line status 588 Variables 0.916 89 3E−04 selected data sets ANOVA Water treatment 90 Variables 0.964 83.70 3.04E−03 50 Variables 0.96 88.6 9.00E−05 20 Variables 0.974 83.7 2.71E−03 Line status 35 Variables 0.911 75.60 1.12E−03 20 Variables 0.9 76.10 9.00E−05 LASSO Water treatment 90 Variables 0.982 88.90 1.47E−03 50 Variables 0.982 93.1 2.60E−04 20 Variables 0.985 88.90 1.47E−03 Line status 35 Variables 0.973 92 3.29E−03 20 Variables 0.978 94.30 6.00E−05 sPLS Water treatment 90 Variables 0.985 92.90 8.90E−03 50 Variables 0.992 96.40 6.00E−04 20 Variables 0.988 92.90 4.90E−03 Line status 35 Variables 0.97 82.30 1.36E−03 20 Variables 0.934 79.60 5.00E−04 8 Variables 0.96 85.9 6.00E−05 Custom Water treatment Metabolites 6 Variables 0.686 53.9 3.00E−05 Physiological Variable selection conditions, cluster and the number of variables used are indicated. Permutation robust- ness was assessed with 200 CV cycles. The data set providing highest Q was highlighted in bold font value < 0.05 following one-way ANOVA (90 for DS and 35 when using metabolic markers in a high-throughput manner for line status). We then reduced the data set size down to (see discussions on practicality of metabolic markers in Fer- 20, a reasonable number of metabolic variables to measure nandez et al. 2016). For DS, we chose to add an intermediate 1 3 56 Page 8 of 14 O. Fernandez et al. 2 2 data set of 50 variables. Q , R and CV-P-values of indi-3.2.3 Metabolic VIP analyses vidual models are summarized in Table 1. The randomized permutations for validation200 cycles) In PLS-DA, an important feature is the variable importance of each bore a significant p value, thus demonstrating their in the projection (VIP) scores. High VIP-score variables robustness (Table 1). As expected, the resulting models com- strongly contribute to the PLS-DA model. Variables with puted after the selection process displayed a higher Q when VIP scores higher than 1 are listed in Online Resource 6. compared to the previous PLS-DA performed with 588 vari- No matter which variable selection process was applied, ables (Table 1; Online Resource 3—Fig. S2). The ANOVA amino acids were overrepresented in the high VIP-score selection process produced efficient models but with the shortlist, underlying their importance in discriminating lowest Q in all situations (Table 1). sPLS and LASSO selec- DS and WW samples in our experiment (Online Resource tion resulted in more discriminant models, the latter for line 6—Table S1b). Two other variables measured by H-NMR status and the former for water treatment. The most efficient were listed in the VIPs shortlist in nearly all conditions of PLS-DA models are illustrated in Fig. 4: 50 variables for variable selection: inositol and glycine-betaine (Online water treatment (sPLS selection) and 20 variables for line Resource 6—Table S1b). On the other hand, a small num- status (LASSO selection) as well as PCA computed with the ber of LC–MS-based variables had VIP scores higher than same data sets (Online Resource 3—Fig. S3). 1 (Online Resource 6—Table  S1b). For line status dis- crimination, all variables with VIP scores higher than 1 Unk_8_8059 Unk_8_8305 Unk_2_3638 Unk_8_5032 Unk_8_4474 Unk_ Un 2_7992 k_1_4802 Unk_8_4902 UnM2 k_8 98 _2582 T483 UnUn k_2 k_ _3996 4_0776 Unk_2_3755 M275T1188 Unk_3_2019 Proteine Unk_8_5261 M806T1659 UnUn k_4 k_ _3505 4_2778 THR M238T644 GLY M406T600 Unk_3_3693 Un trigk_ onne 4_1049 lline_ SER Un Un k_ k_ 44 _4642 _4986 GL AL U A Tryptophan M263T1611 ch M2 Un lo 33 rk_ ogena T7 452 _4071 te GABA Pa M5 lmit 21T1 olei 31 c_ 7acid PHE Unk_4_2238 Asp M5 Un 20 k_ T1 430 _5076 1 VAL Un M3 k_ UDP 93 4_5754 T1 _l64 ike3 0 LEU ILE Qui_chloro1 Unk_3_6504 Chloro_hydro PRO Unk_6_4072 Qui_chloro2 LY in Sositol glycinebet Unk_4_2413 M707T617 M668T783 Caffeoylquinic_acid M356T617 M723T996 Un M4 k_ 49 3_7209 T779 M465T749 M163T653 Unk_4_6616 M612T7 M3 31 57T550 Unk_4_8664 M713T740 Unk_3_5363 M401T982 Unk_3_7617 Glucose M326T571 M627T664 M779T982 M343T934 M276T563 M518T710 M2M5 69T1 0931 T13 907 M5 M7 M4 17 M5 75 81 T7 T9 17 T1 10 30 T6 804 80 M341T930 M163T550 M359 M2 T9 59 30 T9 M3 30 55T550 - WW M439T1680 M277T930 M3M5 61T9 03 31 T1804 - DS -6 -4 -2 02 46 -1.0 -0.5 0.00.5 1.0 X-variate 1: 9% expl. var Component 1 M515T8 Un08 k_8_2582 UnM4 k_4_1893 26T1013 Unk_4_0245 M429T658 M377T643 M6 M6 51 59 T1 T1 56 01 31 M181T750 M219T624 M248T994 M345T571 M4 Un 89k_ T96_8936 59 M589T839 M233T13 M289T723 M247T873 M303T1508 M711T M247T994 M551T687 M283T1116 M273T M3 1750 46T M2 M5 M2 03T 93 99 1362 T9 T180 458 M567T1141 M521T637 M205T1425 M459T866 M815T1135 M233T1232 M319T1240 M361T1468 M479T765 M579T884 M565T842 -1.0 -0.5 0.00.5 1.0 -4 -2 024 6 Component 1 X-variate 1: 6% expl. var Fig. 4 PLS-DA of metabolic data sets of sunflower leaf on variables and DS (orange dots). b PLS model scores (left) and loadings plot selected from the set of 588 metabolic variables (Online Resource (right) of the 20 best LASSO selected variables discriminating the 7) after a selection process based on sPLS or LASSO. a PLS model two-line types, B maintainer lines (red dots) and R restorer lines (blue scores (left) and loadings plot (right) of the 50 best sPLS selected dots). Coloured ellipses represent 95% confidence level variables discriminating the two water treatments WW (green dots) 1 3 X-variate 2: 7% expl. var X-variate 2: 7% expl. var -6 -4 -2 02 4 -5 05 Component 2 Component 2 -1.0 -0.5 0.0 0.5 1.0 -1.0 -0.5 0.00.5 1.0 Metabolomic characterization of sunflower leaf allows discriminating genotype groups or… Page 9 of 14 56 were unidentified ions or H-NMR spectral regions (Online 4 Discussion Resource 6—Table S1c). 4.1 Sunflower leaf metabolite composition 3.3 Cost‑efficient metabolic markers Sunflower is an important crop that provides most of the Simplicity of measurement and cost-efficiency of metabolic table oil used worldwide. However, few metabolomic data markers are arguably as important as their prediction capac- are available to date concerning both its primary and special- ity (Fernandez et al. 2016). In other words, measuring a set ized metabolism. We now present one of the largest sets of of markers with a (slightly) lower predictive capacity might primary metabolites in adult sunflower leaf, with absolute be relevant if the marker set is easier or cheaper to measure. quantification of 38 metabolites and with several compounds A simple solution is often to replace untargeted methods not quantified by Moschen et al. (2017) using GC–MS. with targeted ones. We estimated the cost-reduction potential Several points can be made about sunflower leaf compo- by a factor of 3–20 (Fernandez et al. 2016). Another pos- sition. Malate, citrate and chlorogenic acid were the major sibility is to measure globally a family of compounds when organic acids (Fig. 1, Online Resource 4) and linolenic acid, they are affected in the same way by a given treatment or linoleic acid and palmitic acid were the major fatty acids condition, like in our case for amino acids in DS samples detected. This is in contrast with the fatty acids in sunflower (Fig. 1a). seed where linoleic acid is the most abundant. Serine, ala- To illustrate this point, we selected metabolic variables nine and glutamate were the major free amino acids (Fig. 1). (from Online Resource 7) known to be simple or cheap to Glucose and sucrose were the major soluble sugars in leaf measure and relevant for water treatment discrimination. but their concentrations were at least eight times higher than Since all free amino acids measured were increased in DS that of fructose. This might be due to some specificity of the samples, we replaced them by a single variable representing fructose metabolism in the Asteraceae family. In sunflower, their sum (hereafter called total free amino acids). Finally, fructose is not metabolized into inulin (a fructose-derived we chose total free amino acids, citrate, glycine-betaine, ino- polymer) but is transported and then accumulated in the stem. sitol, sucrose, glucose, protein and starch. This set of eight For example, Martínez-Noël et al. (2015) found that fructose variables was offered a clear determination of DS and WW was three times more concentrated than any other soluble samples in an unsupervised analysis (PCA, Fig. 5a). Addi- sugar in this organ. This might explain the difference between tionally, the generated PLS-DA model was efficient with glucose and fructose concentrations in our leaf samples. 2 2 Q = 0.96, and R = 0.55 (Table 1, Online Resource 3—Fig Considering the specialized metabolites detected via S4a). We could not perform this approach for line status LC–ESI–QTOF–MS, the peaks presenting the highest inten- since most of their high VIP-score variables were unidenti- sities were putatively annotated (Online Resource 5). They fied metabolic signatures. include compounds from three families: caffeoylquinates, methyl-flavonoids and sesquiterpenoids. These compounds 3.4 Comparison with physiological variables for DS had all been previously detected in sunflower biochemical markers analyses. Caffeoylquinic acid is a compound commonly found in sunflower. It plays a role in lignification and cor - Physiological markers are used to assess the impact of DS relates with leaf age in sunflower (Koeppe et al. 1970). It on plant. In our experiment, SLA, OSM_POT and CID were is the dominant phenolic acid in sunflower florets (Liang measured in young and mature leaves at the end of DS. To et  al. 2013) and is also present in seeds (Karamać et  al. test the quality of our PLS-DA model built with selected 2012; Pedrosa et al. 2000). When present in sunflower oil, metabolic variables, we compared its discriminative capacity caffeoylquinates including oxidized chlorogenic acid can with a PLS-DA model built with this physiological data- generate green-coloured oxidized complexes by reacting set comprising six variables extracted from a larger dataset with sunflower proteins (Wildermuth et al. 2016). This oxi- published in Blanchet et al. (2018). Unsupervised PCA com- dative reaction between chlorogenic acid and proteins partly puted with this dataset showed poor separation of DS and explains why sunflower proteins are still underused in the WW samples (Fig. 5b). Furthermore, the PLS-DA model food industry, despite their qualities such as their cheapness built with these physiological data displayed a Q = 0.68 and and absence of allergens (Wildermuth et al. 2016). Several an R = 0.54 (Table 1, Fig. 4b), but was less efficient than putative methylated flavonoids were also detected (Online those built with the minimal set of eight metabolic variables Resource 5). These compounds have been used as chemot- 2 2 (Q = 0.96, R = 0.55; Table 1, Online Resource 3—Fig S4a). axonomic markers for the Astereaceae family (Emerenciano et  al. 2001). Finally, specific sunflower sesquiterpenoids were also detected, one of which was putatively identified 1 3 56 Page 10 of 14 O. Fernandez et al. Fig. 5 PCA scores plot generated with a an “easy-to-measure” data OSM_POT and CID) measured the day before final sampling. Left, set (total free amino acids, citrate, glycine-betaine, inositol, glucose, scores plot. Right, loadings plot total proteins and starch) and b six physiological variables (SLA, as niveusin. In sunflower, this compound and its derivatives 4.3 Biomarkers of line status are thought to offer potential as insecticides (Prasifka et al. 2015). Leaf samples of R and B lines were discriminated with the metabolic data set mostly through unidentified markers 4.2 Variable selection process measured by LC–ESI–QTOF–MS (Online Resource 6— Table S1c). R lines, which in sunflower breeding are used to Variable selection is necessary in metabolomics, especially restore the CMS phenotype, have a nuclear-encoded Rfl gene when looking for metabolic markers (Fernandez et al. 2016). that might act as a transcriptional activator (Balk and Leaver However, numerous methods can be used for the variable 2001; Chen and Liu 2014). The only known function of the selection process and have already been the subject of dis- Rfl gene is to restore male fertility in CMS plants (Chen and cussion (for review, Grissa et al. 2016). We submitted our Liu 2014) as well as the associated changes restricted to the initial dataset to three variable selection processes: ANOVA, mitochondria of floral tissues linked with this loss of fertil- sPLS and LASSO penalty. ity (i.e. mitochondrial membrane integrity and respiration 1 3 Metabolomic characterization of sunflower leaf allows discriminating genotype groups or… Page 11 of 14 56 ratio). Phenotypes associated with the presence of CMS amounts of glycine-betaine that are too low to significantly or R genes are thought to be limited to floral tissues. The impact the sap osmotic potential. Rather, it might serve as fact that we were able to discriminate R and B lines using a ROS detoxication agent (Giri 2011). In the case of myo- analyses of leaf metabolites suggests that the phenotype is inositol, Taji et al. (2006) suggested it might be involved in not restricted to flowers and that it might affect other plant osmotolerance, or alternatively serve as a secondary messen- tissues and organs. Interestingly, several organic acids were ger involved in phospholipid signalling pathways. Finally, less concentrated in R line samples, although not individ- caffeoylquinates and sesquiterpenoids (a terpene class with ually significantly. This might be due to an effect on the three isoprene units) were also detected as putative mark- mitochondrial metabolism in all organs, but this hypothesis ers of DS versus WW samples (Online Resources 5 and 6). needs to be confirmed. Further annotations of the associated Caffeoylquinates have been associated with DS responses markers would contribute to propose hypotheses about direct in grapevine (Hochberg et al. 2013). Terpenes have been or indirect R gene effects in leaf. Additionally, metabolomic shown to be involved in thermotolerance and antioxidant markers denote intermediate information between genes and effects (Sharkey et al. 2008). Furthermore, terpenes seem to final phenotypes and might capture multilocus-controlled have radical scavenging activity contributing to the mitiga- traits and associate alleles producing the same final phe- tion of oxidative damage during stresses. In sunflower leaf, notype. The latter property would be interesting in breed- genes involved in terpene metabolism have been shown to be ing programs to predict the restoration phenotype of novel upregulated under drought conditions (Moschen et al. 2017). alleles in pre-breeding programs and therefore to identify novel sources of restoration for the PET1. However, further 4.5 Towards a small efficient biomarker dataset biochemical and statistical analyses with more R and B lines are required since PLS-DA may be prone to overfitting. Fernandez et al. (2016) argued that ideal metabolic markers should be easy and cheap to analyse. For this purpose, we tested the discriminant capacity of a small metabolic marker 4.4 Biomarkers of water treatment set composed of eight biochemical variables: total free amino acids, citrate, glycine-betaine, myo-inositol, sucrose, glucose, The discrimination of WW and DS samples using meta- total proteins and starch. An unsupervised PCA clearly sepa- bolic variables was more efficient than the discrimination rated WW and DS samples when these eight biochemical of line status. Amino acids were clearly the best DS mark- variables were used (Fig. 4a), but not with the physiological ers in our dataset, displaying a 5- to 10-fold increase in DS dataset consisting in six common indicators of DS measured sunflower leaves (Fig.  1a). Increases in amino acids under at plant level. Indeed, SLA, OSM_POT and CID (measured DS in sunflower have already been documented, although to in both young and mature leaves) are often used to character- a lesser extent and in a cultivar-dependent manner (Mani- ise the water–stress status of a given crop (Fig. 4b). This was vannan et al. 2007). This feature has also been detected in confirmed when comparing Q values for PLS-DA models other crops such as barley (Lanzinger et al. 2015) and wheat computed with each of these data sets (0.91 and 0.68 respec- (Bowne et  al. 2012). Conversely, Moschen et  al. (2017) tively). However, since amino acids were overrepresented in found that the concentrations of several leaf amino acids our PLS-DA model VIPs, our approach might not be general- were decreased under DS in sunflower (Correia et al. 2005). izable to any given criterion. Indeed, reducing the number of These contradictory results regarding amino acid responses variables was much less efficient in discriminating line status. might be due to water–stress intensity, sampling stage or Furthermore, given the fact that amino acid accumulation differences in nitrogen nutrition. In the present study, the use is not always reported for sunflower experiencing drought, of Heliaphen high-throughput phenotyping platform allowed more studies with various drought scenarios and more lines the application of a precise and reproducible drought sce- will be required to confirm our conclusions. Finding the right nario that is available for more thorough understanding of balance between cost reduction and prediction efficiency of the impact of DS on leaf metabolism. Nevertheless, higher each metabolic marker set is likely an achievable goal in concentrations of individual amino acids such as proline many situations but will certainly require optimisation for and glycine have been detected in DS leaves (Moschen et al. each performance criterion studied. 2017). Amino acids, and especially proline, might partici- pate in osmotolerance under DS, although the case is highly debated for the latter (Szabados and Savouré 2010).5 Conclusions In our dataset, other metabolites appeared as good mark- ers of DS samples, i.e. glycine-betaine and myo-inositol. Metabolic markers are a recent development in science. Glycine-betaine is accumulated in various plants under Applications such as personalized medicine have recently abiotic stress (Giri 2011). Generally, plants accumulate attracted keen interest (Lindon and Nicholson 2014). Their 1 3 56 Page 12 of 14 O. Fernandez et al. use in agronomy as a potential tool for crop breeding is even Compliance with ethical standards more recent (Fernandez et al. 2016). In the present work, we Conflict of interest The authors declare that they have no conflict of show that a limited number of metabolic markers can dis- interest. criminate plant sample groups with different characteristics or treatment applications, especially in the case of DS. This Research involving human and/or animal participants This study did not involve the use of animal or human samples. feature was already noted at early stages of plant develop- ment in maize (Riedelsheimer et al. 2012). The fact that leaves of sunflower lines carrying different alleles of the Open Access This article is distributed under the terms of the Crea- tive Commons Attribution 4.0 International License (http://creat iveco CMS restoration gene were separated by this approach shows mmons.or g/licenses/b y/4.0/), which permits unrestricted use, distribu- that metabolomics can reveal an unsuspected metabolic phe- tion, and reproduction in any medium, provided you give appropriate notype in a given organ. The present work also emphasizes credit to the original author(s) and the source, provide a link to the the importance of variable selection. The pipeline we pro- Creative Commons license, and indicate if changes were made. pose (Fig. 2) may not be optimised for all situations (sam- ple numbers, organ types, analytical approaches…), but will provide a preliminary guideline for future users. Another important point is the specificity of the list of selected mark - References ers towards the selected stress. Indeed, several metabolites Allinne, C., Maury, P., Sarrafi, A., & Grieu, P. (2009). Genetic control could be considered as valid metabolic markers of different of physiological traits associated to low temperature growth in stresses, simply because their concentrations may be signifi- sunflower under early sowing conditions. Plant Science, 177(4), cantly altered under various stress situations. To alleviate 349–359. https ://doi.org/10.1016/j.plant sci.2009.07.002. this bias, these markers should be tested under various stress Badouin, H., Gouzy, J., Grassa, C. J., Murat, F., Staton, S. E., Cottret, L., et al. (2017). The sunflower genome provides insights into oil scenarios (Fernandez et al. 2016). It will indeed be crucial metabolism, flowering and Asterid evolution. Nature, 546(7656), to verify whether such a modelling approach remains valid 148–152. https ://doi.org/10.1038/natur e2238 0. when the predictive metabolomic data and the predicted Balk, J., & Leaver, C. J. (2001). The PET1-CMS mitochondrial muta- phenotypic data are obtained in different experiments. In tion in sunflower is associated with premature programmed cell death and cytochrome c release. The Plant Cell, 13(8), 1803– particular, the possibility to use young plantlets, grown in 1818. https ://doi.org/10.1105/TPC.01011 6. controlled conditions and “metabotyped”, to predict a field Blanchet, N., Casadebaig, P., Debaeke, P., Duruflé, H., Gody, L., Gos- phenotype of interest could be extremely useful especially seau, F., et  al. (2018). Data describing the eco-physiological regarding cost reduction, but will require extreme caution. responses of twenty-four sunflower genotypes to water deficit. Submitted: Data in Brief. Thus, a careful methodology with a clear choice of perfor- Bowne, J. B., Erwin, T. A., Juttner, J., Schnurbusch, T., Langridge, P., mance criteria (see Fernandez et al. 2016), stress scenario, Bacic, A., et al. (2012). Drought responses of leaf tissues from developmental stages and analytical methods will have to be wheat cultivars of die ff ring drought tolerance at the metabolite level. developed to test this hypothesis. Molecular Plant, 5(2), 418–429. https://doi.or g/10.1093/mp/ssr114 . Bradford, M. M. (1976). A rapid and sensitive method for the quantita- tion of microgram quantities of protein utilizing the principle of Acknowledgements We thank Laetitia Fouillen for her help with the protein-dye binding. Analytical Biochemistry, 72, 248–254. lipid analyses and the Heliaphen team (especially Nicolas Blanchet) Chen, L., & Liu, Y.-G. (2014). Male sterility and fertility restoration for plant culture. We also thank Dr. Ray Cooke for language proofread- in crops. Annual Review of Plant Biology, 65(1), 579–606. https ing and editing. Metabolite and lipid analyses were performed at the ://doi.org/10.1146/annur ev-arpla nt-05021 3-04011 9. Bordeaux Metabolome Facility, MetaboHUB. Correia, M. J., Fonseca, F., Azedo-Silva, J., Dias, C., David, M. M., Barrote, I., et al. (2005). Effects of water deficit on the activ - Author contributions OF, NL, YG and AM wrote and corrected the final ity of nitrate reductase and content of sugars, nitrate and free manuscript and designed the experimental procedure, with input from all amino acids in the leaves and roots of sunflower and white other authors. OF performed spectrophotometric and UPLC analysis. TB lupin plants growing under two nutrient supply regimes. performed the LC–MS–MS acquisitions, TB and SB performed the LC– Physiologia Plantarum, 124(1), 61–70. https ://doi.org/10.111 ESI–QTOF–MS annotations, and OF analysed the data. MM and CD ran 1/j.1399-3054.2005.00486 .x. the H-NMR acquisitions and identifications. CD and DJ produced the Deborde, C., Maucourt, M., Baldet, P., Bernillon, S., Biais, B., Talon, NMR absolute quantitative data. OF analysed the H-NMR data. HD pro- G., et al. (2009). Proton NMR quantitative profiling for quality vided support for statistical analysis and g fi ure design. PM provided insight assessment of greenhouse-grown tomato fruit. Metabolomics, on the physiological data. MU provided support for amino acid analysis. 5(2), 183–198. https ://doi.org/10.1007/s1130 6-008-0134-2. Emerenciano, V. P., Militão, J. S. L. T., Campos, C. C., Romoff, P., Funding Olivier Fernandez and Maria Urrutia were funded by ‘Agence Kaplan, M. A. C., Zambon, M., et  al. (2001). Flavonoids as Nationale de la Recherche’ (ANR) through the SUNRISE (ANR-11- chemotaxonomic markers for Asteraceae. Biochemical Systemat- BTBR-0005) and AMAIZING (ANR-10-BTBR-0001) projects respec- ics and Ecology, 29(9), 947–957. https ://doi.org/10.1016/S0305 tively. We acknowledge the MetaboHUB (ANR-11-INBS-0010), PHE- -1978(01)00033 -3. NOME (ANR-11-INBS-0012) and SUNRISE (ANR-11-BTBR-0005) Fernandez, O., Urrutia, M., Bernillon, S., Giauffret, C., Tardieu, F., Le projects for further funding. Gouis, J., et al. (2016). Fortune telling: Metabolic markers of plant 1 3 Metabolomic characterization of sunflower leaf allows discriminating genotype groups or… Page 13 of 14 56 performance. Metabolomics, 12(10), 158. https://doi.or g/10.1007/ Lanzinger, A., Frank, T., Reichenberger, G., Herz, M., & Engel, K.-H. s1130 6-016-1099-1. (2015). Metabolite profiling of barley grain subjected to induced Fu, G.-H., Zhang, B.-Y., Kou, H.-D., & Yi, L.-Z. (2017). Stable bio- drought stress: Responses of free amino acids in differently marker screening and classification by subsampling-based sparse adapted cultivars. Journal of Agricultural and Food Chemistry, regularization coupled with support vector machines in metabo- 63(16), 4252–4261. https ://doi.org/10.1021/acs.jafc.5b011 14. lomics. Chemometrics and Intelligent Laboratory Systems, 160, Liang, Q., Cui, J., Li, H., Liu, J., & Zhao, G. (2013). Florets of sun- 22–31. https ://doi.org/10.1016/j.chemo lab.2016.11.006. flower (Helianthus annuus L.): Potential new sources of dietary Gibon, Y., Vigeolas, H., Tiessen, A., Geigenberger, P., & Stitt, M. b fi er and phenolic acids. Journal of Agricultural and Food Chem- (2002). Sensitive and high throughput metabolite assays for inor- istry, 61(14), 3435–3442. https ://doi.org/10.1021/jf400 569a. ganic pyrophosphate, ADPGlc, nucleotide phosphates, and glyco- Lindon, J. C., & Nicholson, J. K. (2014). The emergent role of meta- lytic intermediates based on a novel enzymic cycling system. The bolic phenotyping in dynamic patient stratification. Expert Opin- Plant Journal, 30(2), 221–235. https ://doi.org/10.1046/j.1365- ion on Drug Metabolism & Toxicology, 10(7), 915–919. https :// 313X.2001.01278 .x.doi.org/10.1517/17425 255.2014.92295 4. Giri, J. (2011). Glycinebetaine and abiotic stress tolerance in plants. Manivannan, P., Jaleel, C. A., Sankar, B., Kishorekumar, A., Plant Signaling & Behavior, 6(11), 1746–1751. https ://doi. Somasundaram, R., Lakshmanan, G. M. A., et al. (2007). Growth, org/10.4161/psb.6.11.17801. biochemical modifications and proline metabolism in Helianthus Gosseau, F., Blanchet, N., Varès, D., Burger, P., Campergue, D., annuus L. as induced by drought stress. Colloids and Surfaces Colombet, C., et al. (2018). Heliaphen, an outdoor high-through- B: Biointerfaces, 59(2), 141–149. https ://doi.org/10.1016/j.colsu put phenotyping platform designed to integrate genetics and crop rfb.2007.05.002. modeling. bioRxiv, 362715. https ://doi.org/10.1101/36271 5 Marchand, G., Mayjonade, B., Varès, D., Blanchet, N., Boniface, Grissa, D., Pétéra, M., Brandolini, M., Napoli, A., Comte, B., & Pujos- M.-C., Maury, P., et al. (2013). A biomarker based on gene expres- Guillot, E. (2016). Feature selection methods for early predictive sion indicates plant water status in controlled and natural environ- biomarker discovery using untargeted metabolomic data. Fron- ments. Plant, Cell and Environment, 36(12), 2175–2189. https :// tiers in Molecular Biosciences, 3, 10. htt ps :/ /doi. org/ 10.33 89/ doi.org/10.1111/pce.12127 . fmolb .2016.00030 . Martínez-Noël, G. M. A., Dosio, G. A. A., Puebla, A. F., Insani, E. M., Hartmann, T. (2007). From waste products to ecochemicals: Fifty years & Tognetti, J. A. (2015). Sunflower: A potential fructan-bearing research of plant secondary metabolism. Phytochemistry, 68(22), crop? Frontiers in Plant Science, 6, 798. https ://doi.org/10.3389/ 2831–2846. https ://doi.org/10.1016/j.phyto chem.2007.09.017. fpls.2015.00798 . Hendriks, J. H. M., Kolbe, A., Gibon, Y., Stitt, M., & Geigenberger, Meyer, R. C., Steinfath, M., Lisec, J., Becher, M., Witucka-Wall, H., P. (2003). ADP-glucose pyrophosphorylase is activated by post- Törjék, O., et al. (2007). The metabolic signature related to high translational redox-modification in response to light and to sugars plant growth rate in Arabidopsis thaliana. Proceedings of the in leaves of Arabidopsis and other plant species. Plant Physiology, National academy of Sciences of the United States of America, 133(2), 838–849. https ://doi.org/10.1104/pp.103.02451 3. 104(11), 4759–4764. https ://doi.org/10.1073/pnas.06097 09104 . Hochberg, U., Degu, A., Toubiana, D., Gendler, T., Nikoloski, Z., Moschen, S., Rienzo, J. A. D., Higgins, J., Tohge, T., Watanabe, M., Rachmilevitch, S., et al. (2013). Metabolite profiling and net- González, S., et al. (2017). Integration of transcriptomic and meta- work analysis reveal coordinated changes in grapevine water bolic data reveals transcription factors involved in drought stress stress response. BMC Plant Biology, 13(1), 184. https ://doi. response in sunflower Helianthus annuus L. Plant Molecular Biol- org/10.1186/1471-2229-13-184. ogy, 94(4–5), 549–564. https: //doi.org/10.1007/s11103 -017-0625-5. Horn, R., & Friedt, W. (1999). CMS sources in sunflower: Different Nunes-Nesi, A., Carrari, F., Gibon, Y., Sulpice, R., Lytovchenko, A., origin but same mechanism? Theoretical and Applied Genetics, Fisahn, J., et al. (2007). Deficiency of mitochondrial fumarase 98(2), 195–201. https ://doi.org/10.1007/s0012 20051 058. activity in tomato plants impairs photosynthesis via an effect on Hussain, M., Farooq, S., Hasan, W., Ul-Allah, S., Tanveer, M., Farooq, stomatal function. The Plant Journal, 50(6), 1093–1106. https :// M., et  al. (2018). Drought stress in sunflower: Physiological doi.org/10.1111/j.1365-313X.2007.03115 .x. effects and its management through breeding and agronomic alter - Oilworld. (2016). https ://www.oilwo rld.biz/. natives. Agricultural Water Management, 201, 152–166. https :// Owart, B. R., Corbi, J., Burke, J. M., & Dechaine, J. M. (2014). Selec- doi.org/10.1016/j.agwat .2018.01.028. tion on crop-derived traits and QTL in sunflower (Helianthus Igarashi, K., Kazama, T., & Toriyama, K. (2016). A gene encoding annuus) crop-wild hybrids under water stress. PLoS ONE, 9(7), pentatricopeptide repeat protein partially restores fertility in e102717. https ://doi.org/10.1371/journ al.pone.01027 17. RT98-type cytoplasmic male-sterile rice. Plant and Cell Physi- Patil, C., Calvayrac, C., Zhou, Y., Romdhane, S., Salvia, M.-V., ology, 57(10), 2187–2193. https ://doi.org/10.1093/pcp/pcw13 5. Cooper, J.-F., et al. (2016). Environmental metabolic footprint- Jacob, D., Deborde, C., Lefebvre, M., Maucourt, M., & Moing, A. ing: A novel application to study the impact of a natural and a (2017). NMRProcFlow: A graphical and interactive tool dedicated synthetic β-triketone herbicide in soil. The Science of the Total to 1D spectra processing for NMR-based metabolomics. Metabo- Environment, 566–567, 552–558. https ://doi.org/10.1016/j.scito lomics, 13(4), 36. https ://doi.org/10.1007/s1130 6-017-1178-y. tenv.2016.05.071. Jelitto, T., Sonnewald, U., Willmitzer, L., Hajirezeai, M., & Stitt, M. Pedrosa, M. M., Muzquiz, M., García-Vallejo, C., Burbano, C., Cuad- (1992). Inorganic pyrophosphate content and metabolites in potato rado, C., Ayet, G., et al. (2000). Determination of caffeic and chlo- and tobacco plants expressing E. coli pyrophosphatase in their cyto- rogenic acids and their derivatives in different sunflower seeds. sol. Planta, 188(2), 238–244. https://doi.or g/10.1007/bf00216819 . Journal of the Science of Food and Agriculture, 80(4), 459–464. Karamać, M., Kosińska, A., Estrella, I., Hernández, T., & Dueñas, Pichersky, E., & Lewinsohn, E. (2011). Convergent evolution in plant M. (2012). Antioxidant activity of phenolic compounds identified specialized metabolism. Annual Review of Plant Biology, 62, 549– in sunflower seeds. European Food Research and Technology, 566. https ://doi.org/10.1146/annur ev-arpla nt-04211 0-10381 4. 235(2), 221–230. https ://doi.org/10.1007/s0021 7-012-1751-6. Poormohammad Kiani, S., Grieu, P., Maury, P., Hewezi, T., Gentzbit- Koeppe, D. E., Rohrbaugh, L. M., Rice, E. L., & Wender, S. H. (1970). tel, L., & Sarrafi, A. (2007). Genetic variability for physiological Tissue age and caffeoylquinic acid concentration in sunflower. traits under drought conditions and differential expression of water stress-associated genes in sunflower (Helianthus annuus L). TAG. Phytochemistry, 9(2), 297–301. https ://doi.org/10.1016/S0031 Theoretical and Applied Genetics, 114(2), 193–207. -9422(00)85138 -9. 1 3 56 Page 14 of 14 O. Fernandez et al. Prasifka, J. R., Spring, O., Conrad, J., Cook, L. W., Palmquist, D. E., Taji, T., Takahashi, S., & Shinozaki, K. (2006). Inositols and their & Foley, M. E. (2015). Sesquiterpene lactone composition of metabolites in abiotic and biotic stress responses. In A. L. Majum- wild and cultivated sunflowers and biological activity against an der & B. B. Biswas (Eds.), Biology of inositols and phospho- insect pest. Journal of Agricultural and Food Chemistry, 63(16), inositides: Subcellular biochemistry (pp. 239–264). Boston, MA: 4042–4049. https ://doi.org/10.1021/acs.jafc.5b003 62. Springer. https ://doi.org/10.1007/0-387-27600 -9_10. Qi, L. L., Foley, M. E., Cai, X. W., & Gulya, T. J. (2016). Genetics Talukder, Z. I., Hulke, B. S., Qi, L., Scheffler, B. E., Pegadaraju, V., and mapping of a novel downy mildew resistance gene, Pl(18), McPhee, K., et al. (2014). Candidate gene association mapping of introgressed from wild Helianthus argophyllus into cultivated Sclerotinia stalk rot resistance in sunflower (Helianthus annuus sunflower (Helianthus annuus L.). TAG. Theoretical and Applied L.) uncovers the importance of COI1 homologs. TAG. Theoretical Genetics, 129(4), 741–752. and applied genetics, 127(1), 193–209. https ://doi.org/10.1007/ Riedelsheimer, C., Czedik-Eysenberg, A., Grieder, C., Lisec, J., Tech-s0012 2-013-2210-x. now, F., Sulpice, R., et al. (2012). Genomic and metabolic predic- Tompkins, D., & Toffaletti, J. (1982). Enzymic determination of cit- tion of complex heterotic traits in hybrid maize. Nature Genetics, rate in serum and urine, with use of the Worthington “ultrafree” 44(2), 217–220. https ://doi.org/10.1038/ng.1033. device. Clinical Chemistry, 28(1), 192–195. Rohart, F., Gautier, B., Singh, A., & Cao, K.-A. L. (2017). mixOmics: Vear, F. (2016). Changes in sunflower breeding over the last fifty years. An R package for ‘omics feature selection and multiple data inte- OCL, 23(2), D202. https ://doi.org/10.1051/ocl/20160 06. gration. PLoS Computational Biology, 13(11), e1005752. https:// Wildermuth, S. R., Young, E. E., & Were, L. M. (2016). Chloro- doi.org/10.1371/journ al.pcbi.10057 52. genic acid oxidation and its reaction with sunflower proteins Schneiter, A. A., & Miller, J. F. (1981). Description of sunflower to form green-colored complexes. Comprehensive Reviews in growth stages 1. Crop Science, 21(6), 901–903. https ://doi. Food Science and Food Safety, 15(5), 829–843. https ://doi. org/10.2135/crops ci198 1.00111 83X00 21000 60024 x. org/10.1111/1541-4337.12213. Sharkey, T. D., Wiberley, A. E., & Donohue, A. R. (2008). Isoprene Yu, Y., Zhao, Z., Shi, Y., Tian, H., Liu, L., & Bian, X. (2016). Hybrid emission from plants: Why and how. Annals of Botany, 101(1), sterility in rice (Oryza sativa L.) involves the tetratricopeptide 5–18. https ://doi.org/10.1093/aob/mcm24 0. repeat domain containing protein. Genetics, 203(3), 1439–1451. Smith, C. A., Want, E. J., O’Maille, G., Abagyan, R., & Siuzdak, G. https ://doi.org/10.1534/genet ics.115.18384 8. (2006). XCMS: Processing mass spectrometry data for metabo- lite profiling using nonlinear peak alignment, matching, and Publisher’s Note Springer Nature remains neutral with regard to identification. Analytical Chemistry, 78(3), 779–787. https ://doi. jurisdictional claims in published maps and institutional affiliations. org/10.1021/ac051 437y. Szabados, L., & Savouré, A. (2010). Proline: A multifunctional amino acid. Trends in Plant Science, 15(2), 89–97. https ://doi. org/10.1016/j.tplan ts.2009.11.009. Affiliations 1,5 1,2,6 1,7 1,3 1,3 Olivier Fernandez  · Maria Urrutia  · Thierry Berton  · Stéphane Bernillon  · Catherine Deborde  · 1,3 1,3,6 4 4 1,3 4 Daniel Jacob  · Mickaël Maucourt  · Pierre Maury  · Harold Duruflé  · Yves Gibon  · Nicolas B. Langlade  · 1,3 Annick Moing Maria Urrutia Annick Moing m.urrutia@enzazaden.es annick.moing@inra.fr Thierry Berton UMR1332 Biologie du Fruit et Pathologie, INRA, thierry.berton@laposte.net Centre INRA de Bordeaux, 71 av Edouard Bourlaux, Stéphane Bernillon 33140 Villenave d’Ornon, France stephane.bernillon@inra.fr UMR AgroImpact, INRA, Estrées-Mons, 80203 Péronne, Catherine Deborde France catherine.deborde@inra.fr Plateforme Métabolome Bordeaux, CGFB, Daniel Jacob MetaboHUB-PHENOME, 33140 Villenave d’Ornon, France daniel.jacob@inra.fr UMR LIPM, INRA, CNRS, Université de Toulouse, Mickaël Maucourt 31326 Castanet-Tolosan, France mickael.maucourt@inra.fr Present Address: Laboratoire RIBP, Université de Reims Pierre Maury Champagne Ardenne, Moulin de la Housse Chemin des pierre.maury@ensat.fr Rouliers, 51100 Reims, France Harold Duruflé Present Address: Enza Zaden Centro de Investigacion S.L., harold.durufle@inra.fr Santa Maria del Aguila, 04710 Almeria, Spain Yves Gibon Present Address: Centre for CardioVascular and Nutrition, yves.gibon@inra.fr UMR INRA-INSERM, Aix-Marseille Univ, INSERM, 13005 Marseilles, France Nicolas B. Langlade nicolas.langlade@inra.fr 1 3 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Metabolomics Springer Journals

Metabolomic characterization of sunflower leaf allows discriminating genotype groups or stress levels with a minimal set of metabolic markers

Loading next page...
 
/lp/springer-journals/metabolomic-characterization-of-sunflower-leaf-allows-discriminating-qLJDLdUpmJ
Publisher
Springer Journals
Copyright
Copyright © 2019 by The Author(s)
Subject
Life Sciences; Biochemistry, general; Molecular Medicine; Cell Biology; Developmental Biology; Biomedicine, general
ISSN
1573-3882
eISSN
1573-3890
D.O.I.
10.1007/s11306-019-1515-4
Publisher site
See Article on Publisher Site

Abstract

Introduction Plant and crop metabolomic analyses may be used to study metabolism across genetic and environmental diversity. Complementary analytical strategies are useful for investigating metabolic changes and searching for biomarkers of response or performance. Methods and objectives The experimental material consisted in eight sunflower lines with two line status, four restorers (R, used as males) and four maintainers (B, corresponding to females) routinely used for sunflower hybrid varietal production, respectively to complement or maintain the cytoplasmic male sterility PET1. These lines were either irrigated at full soil capacity (WW) or submitted to drought stress (DS). Our aim was to combine targeted and non-targeted metabolomics to characterize sunflower leaf composition in order to investigate the effect of line status genotypes and environmental condi- tions and to find the best and smallest set of biomarkers for line status and stress response using a custom-made process of variables selection. Results Five hundred and eighty-eight metabolic variables were measured by using complementary analytical methods such as H-NMR, MS-based profiles and targeted analyses of major metabolites. Based on statistical analyses, a limited number of markers were able to separate WW and DS samples in a more discriminant manner than previously published physiological data. Another metabolic marker set was able to discriminate line status. Conclusion This study underlines the potential of metabolic markers for discriminating genotype groups and environmental conditions. Their potential use for prediction is discussed. Keywords Metabolic markers · Metabolomics · Sunflower · Water stress · Maintainer–restorer lines AbbreviationsDS Drought stressed AQC 6-Aminoquinolyl-N-succinimidyl DW Dry weight carbamate FAMES Fatty acid methyl esters AUC Area under curve LASSO Least absolute shrinkage and selec- B Maintainer line tion operator CID Carbon isotope discriminationLC–ESI–QTOF–MS Liquid chromatography–electro- CV Cross-validation spray-ionization–time-of-flight– DAG Days after germination mass spectrometry NMR Nuclear magnetic resonance OSM_POT Osmotic potential Electronic supplementary material The online version of this PCA Principal component analysis article (https ://doi.org/10.1007/s1130 6-019-1515-4) contains PET Petiolaris supplementary material, which is available to authorized users. PLS Partial least squares PLS-DA Partial least squares discriminant * Olivier Fernandez olivier.fernandez@univ-reims.fr analysis R Restorer line Extended author information available on the last page of the article Vol.:(0123456789) 1 3 56 Page 2 of 14 O. Fernandez et al. SLA Specific leaf area Therefore, the B line is widely used for phenotypic and agro- VIP Variable importance in the nomic description of the line. projection Since the introduction of hybrid varieties, sunflower has WW Well-watered undergone an active breeding process (Vear 2016), mainly thanks to molecular marker-assisted selection. Hybrids have been selected with increased resistance to downy mildew (Qi et al. 2016), sclerotinia (Talukder et al. 2014) and water 1 Introduction stress (Marchand et al. 2013; Owart et al. 2014), although sunflower is often cited as moderately drought-tolerant (Hus- Sunflower ( Helianthus annuus L.) is the fourth major crop sain et al. 2018). This selection process will benefit from providing seed for oil production worldwide. In 2016, the recent sequencing of the maintainer inbred line XRQ world production reached 45 MT from 26 Mha, principally (Badouin et al. 2017). As part of these selection efforts, our in Europe (around 70%), Ukraine being the world leader group is currently involved in searching for metabolic mark- (Oilworld 2016; Hussain et al. 2018). Worldwide produc- ers of sunflower performance. A definition of biomarkers tion has increased constantly ever since (Oilworld 2016). (and their sub-category metabolic markers) emerged from Sunflower accounts for more than 50% of total world table- the field of medicine as a characteristic objectively measured oil consumption. Additionally, its high biodegradability to indicate a given biologic, pathologic or pharmacologic makes it suitable for non-alimentary uses such as in paints response (Fernandez et al. 2016). In plant science, meta- and bioplastics. bolic markers have been defined as metabolites or groups Native to North America and introduced into Europe of metabolites that are measured to predict or discriminate in the sixteenth century, sunflower became a major crop plant responses or performance (Fernandez et al. 2016). in this area in the early 1960s. Further development was The use of metabolic markers to predict criteria of plant achieved after the introduction of hybrid varieties in the performance is recent, with pioneering papers dating from early 1980s. Hybrid varieties are based on the use of cyto- the early 2010s (Meyer et al. 2007; Riedelsheimer et al. plasmic male-sterile (CMS) lines (Vear 2016), like many 2012). The possibilities offered by these markers in plant other crops (Chen and Liu 2014). The male sterility used for selection processes were reviewed recently and a pipeline sunflower hybrid production, called PET1-CMS, was first to search and use them has been proposed (Fernandez et al. identified from an interspecific cross between Helianthus 2016). The authors emphasized that the search for meta- petiolaris and H. annuus. It results from the reorganization bolic markers requires a first step of analysis on a small of mitochondrial DNA that generated a new open reading core set of genotypes. The present article investigates this frame ORFH522 co-transcribed with apt1 gene and coding first step, which includes (1) testing the analytic pipeline to a 16 kDa protein. This leads to modified mitochondrial func- establish the dynamic range of targeted metabolites, (2) con- tions and affects pollen development (Balk and Leaver 2001) firming the presence of several secondary or “specialized” through a decline in the mitochondrial membrane integrity metabolites (as defined by Hartmann 2007; Pichersky and and the respiratory control ratio. The mitochondrial pro- Lewinsohn 2011) and (3) investigating which metabolites tein ORFH522 appears to be expressed in all tissues, but are essential for differentiating groups of samples such as, in the deleterious phenotype associated with PET1-CMS has our case, water treatment (well-watered, WW, vs. drought- been thought to be limited to the anthers, and no apparent stressed, DS) and line status (maintainer, B, vs. restorer, R). extra phenotypes have been found in other organs (Horn and These metabolites could later serve as metabolic markers. Friedt 1999; Balk and Leaver 2001). Furthermore, we tested different statistical methods for vari- To complement the mutational effect, a nuclear restora- able selection in order to find the best and smallest sets of tion gene (noted Rf1) is used in sunflower hybrid production. metabolite markers. Indeed, for a given agronomical trait, Restoration genes are nuclear and generally encode tetratri- the deployment of metabolic markers among breeders will copeptides that are thought to transcriptionally control the depend on their cost (Fernandez et al. 2016). CMS mitochondrial gene (Chen and Liu 2014; Igarashi et al. For this purpose, we used a combination of targeted and 2016; Yu et al. 2016). Finally, sunflower hybrid production untargeted metabolomic analyses on sunflower leaf samples is based on crossing a restorer line called R bearing a func- obtained from B or R lines and in WW and DS conditions. tional restoration allele Rf1 (that recovers the PET1-CMS Our results show that a limited number of markers can male-sterility phenotype) to a male-sterile PET1-CMS line clearly differentiate WW from DS samples and in a more called A (carrying a recessive rf1 allele). To maintain this discriminant manner than the physiological data presented in male-sterile line, a maintainer line called B, isogenic to the Blanchet et al. (2018), which are classically used to discrimi- A line, is also used. Each B line carries the rf1 allele but is nate individuals subjected to DS. To our surprise, another male-fertile, as it does not carry the CMS-PET1 cytoplasm. leaf metabolic marker set was able to discriminate B lines 1 3 Metabolomic characterization of sunflower leaf allows discriminating genotype groups or… Page 3 of 14 56 from R ones. Our data underline the potential of metabolic Starch (Hendriks et al. 2003) and protein (Bradford 1976) markers for discriminating genotypes and environmental contents were determined on the pellet. Assays were carried conditions. Their potential use in sunflower breeding for out in 96-well microplates. performance prediction is discussed. Individual free amino acid analysis was carried out using an UPLC separation with fluorescent detection after deri- vatization using 6-aminoquinolyl-N-succinimidyl carbamate 2 Materials and methods (AQC)-tag (a method hereafter referred to as UPLC-Fluo). For lipid analysis, fatty acid methyl esters (FAMES) The protocols used are detailed in Online Resource 1 and were measured after hydrolysis of 20 mg dry weight (DW) summarized here. with 2.5% H SO (v/v) in methanol. GC-FID was per- 2 4 formed using an Agilent 7890 gas chromatograph (Agilent, 2.1 Plant material and growth conditions Santa Clara, California) equipped with a Carbowax column (15 m × 0.53 mm, 1.2 µm; Alltech Associates, Deerfield, IL, The experiment was performed in 2013 in the phenotyping USA) and flame ionization detection. FAMES were identi- platform “Heliaphen” (Gosseau et al. 2018). Eight sunflower fied by comparing their retention times with commercial lines, four B and four R lines, were grown in two conditions fatty acid standards (Sigma, Saint-Quentin Fallavier, France) (WW and DS) with three replicates, leading to a total of 48 and quantified using ChemStation (Agilent). samples. Irrigation was stopped at 38 days after germination (DAG; Schneiter and Miller 1981) for DS plants. Soil evapo- 2.4 H‑NMR analysis of major polar compounds ration was estimated according to Marchand et al. (2013). Both WW and DS plants were weighed four times per day by Polar metabolites were extracted from lyophilized powder the Heliaphen robot to estimate plant transpiration (Gosseau (40 mg DW per biological replicate) with an ethanol–water et al. 2018). At 47 DAG, leaves for metabolomic analyses series (80/20, 50/50, 0/100 v/v) at 80 °C as described in were harvested without their petiole and frozen in liquid Deborde et al. (2009) with modifications. This three-step nitrogen. Two other leaves (mature and young leaves) were extraction process (ethanol–water series) was chosen to take harvested for physiological trait measurements. During the into account the diverse affinities and solubilities of leaf experiment, two samples were excluded before leaf sam- major polar compounds (i.e. sugars, organic acids, amino- pling (excessive irrigation was detected when analysing n fi al acids) for ethanol or water, in order to obtain an accurate Heliaphen readings) and four samples could not be analysed view of these compounds in leaf extracts. The 1D (cpmg because of insufficient powder quantity. This resulted in a and single-pulse) spectra were processed using the NMR- total of 42 samples submitted to metabolic analyses. ProcFlow application v1.1 (Jacob et al. 2017; http://nmrpr ocflo w.org/). For the cpmg dataset, this resulted in 479 nor- 2.2 Physiological trait measurements for plant malized variables corresponding to spectral regions (named phenotyping Unk_ppm:number in Online Resource 2) which included compounds that were annotated later on. The assignments Plant and leaf physiological data are part of a larger dataset of metabolites in the H-nuclear magnetic resonance (NMR) presented in Blanchet et al. (2018). Specific leaf area (SLA) spectra were made by comparing the proton chemical shifts was determined according to Allinne et al. (2009). Both leaf with public or local spectral databases and by spiking the osmotic potential (OSM_POT) and leaf osmotic potential at samples with the corresponding commercial compounds. 2D full turgor (OSM_POT_100) were measured as described experiments were performed on a representative selected in Poormohammad Kiani et al. (2007). To assess carbon extract taken from the WW condition. Quantification of 11 isotope discrimination (CID), samples were oven-dried, identified compounds was performed by using quantified ground, weighed and analysed using a continuous low iso- single-pulse spectra dataset and calibration curves. tope ratio mass spectrometry at the Stable Isotope Platform SHIVA (University of Toulouse, France). 2.5 LC–ESI–QTOF–MS untargeted analysis of semi‑polar metabolites 2.3 T argeted compound measurements Liquid chromatography–electrospray-ionization–time-of- For each sample, about 20 mg fresh weight were extracted flight–mass spectrometry (LC–ESI–QTOF–MS) profiling as in Hendriks et al. (2003). Sucrose, glucose, and fructose of aqueous methanol extracts containing 0.1% formic acid (Jelitto et al. 1992), malate (Nunes‐Nesi et al. 2007), citrate was performed with extracts obtained from 20 mg DW lyo- (Tompkins and Toffaletti 1982) and glucose-6-P (Gibon philized powder. An Ultimate 3000 HPLC (Dionex, Sun- et al. 2002) were determined in the ethanolic supernatant. nyvale, CA, USA) was used to separate metabolites on a 1 3 56 Page 4 of 14 O. Fernandez et al. reversed-phase C18 column using an acetonitrile gradient in Resource 3—Fig. S1. We targeted these compounds because acidified water. Metabolites were detected by using a hybrid they are (1) often considered as putative metabolic mark- quadrupole/time-of-flight mass spectrometer (micrOTOF-Q , ers (Fernandez et al. 2016) and (2) valuable candidates for Bruker Daltonics, Bremen, Germany). Electrospray ioniza- a high-throughput metabolic marker approach, as they are tion in positive mode was used to ionize the compounds. easy and cheap to measure. A quality control sample (QC) was injected after each set The concentrations of these 29 compounds were summed of ten samples. The MS data were processed using XCMS to estimate their contribution to leaf biomass. This yielded (Smith et al. 2006) and R scripts for filtering. A total of about 45% of leaf dry mass. Glucose was found to be the −1 1519 features were detected and reduced to 540 metabolic major soluble sugar. Its concentration (32–45 mg g DW) variables after filtering. The corresponding MS-based vari- was in the same range as that of sucrose, but 8–10 times ables were named using their nominal masses in dalton and higher than fructose depending on the chosen conditions. retention time in seconds in Online Resource 2 (MxxxTyyy). Glutamate, alanine and serine were found to be the most Metabolite identification was performed using the accurate- abundant amino acids. In leaves, linolenic acid (C18:3) was −1 mass data and Orbitrap (Thermo Fisher, Villebon-sur-Yvette, the most abundant fatty acid (7.5–18.6 mg g DW), fol- France) MS and MS/MS data of a representative sample lowed by linoleic acid (Fig. 1). extract. H-NMR profiling was performed on polar extracts to further analyse metabolites from primary metabolism in the 2.6 Statistical analyses millimolar range. Four hundred and seventy-nine regions were observed in the H-NMR cpmg dataset, of which 20 All statistical analyses were performed using the R Soft- compounds were annotated (Online Resource 4). Eleven ware (http://www.r-proje ct.org/), the R package mixOmics identified compounds were measured and quantified with the (Rohart et al. 2017) and the BioStatFlow online tool (bio- H-NMR quantitative single-pulse dataset, but only nine of statflow.org) which is based on R scripts. Two-way ANOVA them were kept in the final dataset to avoid redundancy with with FDR correction was performed to highlight line status targeted spectrophotometric measurements. When summed, or water-treatment effects and interaction. The parameters these compounds represented an additional 5% of the leaf used for partial least squares-discriminant analysis (PLS- dry mass (Online Resource 4). DA) in BioStatFlow were adjusted to a tenfold cross-valida- LC–ESI–QTOF–MS analysis of semi-polar extracts was tion (CV) to generate the model (and calculate the Q ) and performed to analyse specialized metabolites. The most 200-randomized permutations to estimate the robustness of intense peaks that were detected in the sample extracts, the generated model. Some graphical outputs for PLS-DA based on their intensity in the XCMS table generated by a were produced by mixOmics, using the same parameters relative area under curve (AUC) approach, were tentatively than with BioStatFlow. An additional R script from Fu et al. annotated. Orbitrap-MS data were used in order to gain pre- (2017) was used to perform least absolute shrinkage and cision on mass measurement and to perform MS/MS. Online selection operator (LASSO) and sparse partial least square Resource 5 shows the annotation table generated using a (sPLS) selection. Principal component analysis (PCA) and representative spectrum of a leaf extract with annotation of partial least square (PLS) were performed on data mean- the most intense peaks. The two most intense peaks were centred and scaled to unit variance. All statistical analyses annotated as mono and di-caffeoyl quinic acid. With a reten- were performed on the data set in Online Resource 2 or tion time around 17–20 min, several methylated flavonoids subsets of this file. were also detected. Finally, three smaller peaks ranging in retention time from 15 to 17 min were found to putatively represent sunflower sesquiterpenoids. Several peaks after 3 Results 25 min remained elusive. Several metabolite concentrations differed between the 3.1 Sunflower leaf metabolic contents measured conditions, as highlighted by a two-way ANOVA (p < 0.05 by targeted and untargeted approaches with FDR correction, Online Resource 6—Table S1a). In total, 27 metabolites plus starch and protein content were 3.1.1 Difference between DS and WW samples targeted and quantified in sunflower leaf. Major soluble sug- ars (i.e. the ones with the highest content), organic acids The most striking difference was the large increase in each and chlorophylls were quantified with spectrophotometric individual amino acid concentration found for DS sam- analyses. FAMES and free amino acids were measured by ples, with an average increase of 15-fold, (Fig. 1a, Online using GC-FID and UPLC-fluo, respectively. These data are Resource 6—Table S1a). On the other hand, starch, protein presented for the different conditions in Fig.  1 and Online content, linolenic and palmitoleic acids were slightly but 1 3 Metabolomic characterization of sunflower leaf allows discriminating genotype groups or… Page 5 of 14 56 Fig. 1 Concentrations of 27 metabolites measured by targeted meth- Maintainer B lines (white bars) or restorer R lines (black bars). Ver- ods (UPLC-Fluo for amino acids, GC-FID for FAMES, spectropho- tical bars represent standard deviations. Asterisk indicates variables tometry for others) in leaf of B or R sunflower lines cultivated in two that were found significantly different between groups after two-way −1 conditions (WW and DS). Results are expressed in mg  g DW in ANOVA test (p value < 0.05) the four types of samples. a WW (white bars) or DS (black bars). b significantly lower in DS (Fig.  1a; Online Resources 3—Fig (two-way ANOVA test; Online Resources 4 and 6—Tables S1a and 6—Table S1a). Minor differences in starch, pro- S1a). For the other nine compounds identified in H-NMR tein, soluble sugars and GABA were observed between B spectra (amino acids and sugars), excellent correlations and R lines (Fig. 1b and Online Resource 3—Fig S1b) but were found with spectrophotometric and chromatographic none of them was statistically significant (Online Resource targeted methods (data not shown). 6—Table S1a). Finally, only a small group of m/z were significantly Among the variables that were highly significant under different under DS (Online Resource 6—Table  S1a). DS (two-way ANOVA test), most of them were uniden- Four of them were putatively annotated as heliannuol, tified H-NMR spectra regions (Online Resource 6— 3-O-caffeoylquinic acid, tryptophan and phenylalanine. Table S1a). Among them, myo-inositol, glycine betaine The last two were also detected by the UPLC-fluo tar - and trigonelline were significantly higher under DS, geted method. whereas chlorogenate and formate were significantly lower 1 3 56 Page 6 of 14 O. Fernandez et al. 3.1.2 Difference between B and R samples Full data set 1048 variables For the B and R lines, no targeted metabolites were sig- nificantly different. Two unidentified H-NMR variables had a p value < 0.05 (Unk_6.8936 and Unk_3.8733, Online Elimination of redundantvariables (« curation ») Resource 6—Table S1a,). However, except for chlorogenic (RMN annotation, clustering and correlation >,85) acid, most organic acids measured displayed a lower con- centration in R lines leaf samples (Fig. 1b, Online Resource 4). Finally, the rest of the variables that were found signifi- Curateddata cantly different for line status were unidentified MS-based set 588 variables variables (Online Resource 6—Table S1a), except for two putatively annotated flavonoids (Online Resources 5 and 6—Table S1a). Data filtering– variable selection ANOVA –sPLS-LASSO 3.2 Workflow for identifying metabolic markers of water treatment and line status «Top » «Top » Water treatment Line status variables variables The analytical methods allowed the generation of a matrix of 1048 metabolic variables (Online Resource 2). This matrix included 27 targeted metabolites, starch, total protein content Final predictive models and 9 annotated H-NMR variables. The remaining variables PLS-DA were composed of H-NMR unidentified spectral regions and 540 MS-based signatures. The matrix was processed Fig. 2 Description of the statistical analysis pipeline used in this arti- through a three-step biostatistical pipeline to select the more cle relevant variables to discriminate samples according to water treatment and line status: (1) elimination of redundant vari- 3—Fig. S2b), in the 2D space based on the first two latent ables, (2) variable selection for each sample cluster and (3) variables. Predictive ability (Q ) and proportion of vari- final PLS-DA model calculation (Fig.  2).ance (R ) explained by the model were higher than 0.9 and 0.8 in both cases (Table 1), respectively. Each model was 2 2 3.2.1 Elimination of redundant metabolic variables considered as valid as it bore Q and R values above 0.4 and 0.5, respectively (Patil et al. 2016). However, in a high- Since a single metabolite can be encompassed within sev- throughput approach, it is impractical to measure more than eral H-NMR buckets or MS-based ions, we first reduced 500 variables to discriminate or predict cluster differentia- this full data set by hierarchical clustering (BioStatFlow, tion. Therefore, our next step was to test a variable selection Pearson correlation, average linkage as aggregation method). process and to assess the validity of group discrimination Clusters were generated with a correlation threshold of 0.85. with PLS-DA after this selection. PLS-DA was chosen to Within each cluster, MS-based metabolic variables corre- easily compare model performance using Q values. sponding to adducts or isotopes were eliminated while the one with the highest AUC was kept. For H-NMR buckets, 3.2.2 Metabolic variable selection process we used a similar process in order to keep buckets bearing the highest AUC. After this curation process (Fig. 2), the To select variables, we compared three different methods new dataset comprised 588 variables (Online Resource 7). for each condition (DS or line status), a generalised univari- We then tested the discrimination potential of this curated ate method (one-way ANOVA) and two multivariate ones data set on our sample groups using an unsupervised sta- (sPLS and LASSO penalty; Fu et al. (2017); Fig. 2). The tistical approach. PCA was first carried out (Fig.  3). The 588-variable data matrix (Online Resource 7) was submit- first two components displayed in Fig.  3a (water treatment) ted to these methods and subsequent PLS-DAs were per- 2 2 and Fig.  3b (line status) explained 25% of the total vari- formed. We compared the Q and R to assess the quality ability. The separation of our sample groups was incom- of the variable selection process for each resulting PLS-DA plete, although slightly better for DS. We then performed model (Table 1). Since our objective was to find the small- a supervised method (PLS-DA) on this 588-variable data- est possible variable set, we analysed datasets of different set for each type of sample group. Each PLS-DA analysis sizes (90, 50 and 20 variables for water treatment; 35 and 20 was able to discriminate WW from DS samples (Online variables for line status). We dimensioned the first selected Resource 3—Fig. S2a), and B from R lines (Online Resource data set size according to the numbers of variables with a p 1 3 Metabolomic characterization of sunflower leaf allows discriminating genotype groups or… Page 7 of 14 56 Fig. 3 PCA scores plot (PC1 x PC2 plan) generated with the full set DS, orange dots. b Highlighting line types. B, red dots and R, blue of 588 metabolic variables (Online Resource 7) measured in sun- dots. Coloured ellipses represent 95% confidence level. The connect- flower leaf cultivated in a Heliaphen phenotyping platform. a High- ing lines attach each individual point to the centre of the confidence lighting samples with different water treatment. WW, green dots and ellipse Table 1 Comparison of 2 2 Variable selection Condition Data set size Q R Expl var t1/ CV p-value predictive ability (Q ) and year (%) explained variance explained (R ) of the different PLS-DA None Water treatment 588 Variables 0.936 80.2 1.1E−04 models calculated with different Line status 588 Variables 0.916 89 3E−04 selected data sets ANOVA Water treatment 90 Variables 0.964 83.70 3.04E−03 50 Variables 0.96 88.6 9.00E−05 20 Variables 0.974 83.7 2.71E−03 Line status 35 Variables 0.911 75.60 1.12E−03 20 Variables 0.9 76.10 9.00E−05 LASSO Water treatment 90 Variables 0.982 88.90 1.47E−03 50 Variables 0.982 93.1 2.60E−04 20 Variables 0.985 88.90 1.47E−03 Line status 35 Variables 0.973 92 3.29E−03 20 Variables 0.978 94.30 6.00E−05 sPLS Water treatment 90 Variables 0.985 92.90 8.90E−03 50 Variables 0.992 96.40 6.00E−04 20 Variables 0.988 92.90 4.90E−03 Line status 35 Variables 0.97 82.30 1.36E−03 20 Variables 0.934 79.60 5.00E−04 8 Variables 0.96 85.9 6.00E−05 Custom Water treatment Metabolites 6 Variables 0.686 53.9 3.00E−05 Physiological Variable selection conditions, cluster and the number of variables used are indicated. Permutation robust- ness was assessed with 200 CV cycles. The data set providing highest Q was highlighted in bold font value < 0.05 following one-way ANOVA (90 for DS and 35 when using metabolic markers in a high-throughput manner for line status). We then reduced the data set size down to (see discussions on practicality of metabolic markers in Fer- 20, a reasonable number of metabolic variables to measure nandez et al. 2016). For DS, we chose to add an intermediate 1 3 56 Page 8 of 14 O. Fernandez et al. 2 2 data set of 50 variables. Q , R and CV-P-values of indi-3.2.3 Metabolic VIP analyses vidual models are summarized in Table 1. The randomized permutations for validation200 cycles) In PLS-DA, an important feature is the variable importance of each bore a significant p value, thus demonstrating their in the projection (VIP) scores. High VIP-score variables robustness (Table 1). As expected, the resulting models com- strongly contribute to the PLS-DA model. Variables with puted after the selection process displayed a higher Q when VIP scores higher than 1 are listed in Online Resource 6. compared to the previous PLS-DA performed with 588 vari- No matter which variable selection process was applied, ables (Table 1; Online Resource 3—Fig. S2). The ANOVA amino acids were overrepresented in the high VIP-score selection process produced efficient models but with the shortlist, underlying their importance in discriminating lowest Q in all situations (Table 1). sPLS and LASSO selec- DS and WW samples in our experiment (Online Resource tion resulted in more discriminant models, the latter for line 6—Table S1b). Two other variables measured by H-NMR status and the former for water treatment. The most efficient were listed in the VIPs shortlist in nearly all conditions of PLS-DA models are illustrated in Fig. 4: 50 variables for variable selection: inositol and glycine-betaine (Online water treatment (sPLS selection) and 20 variables for line Resource 6—Table S1b). On the other hand, a small num- status (LASSO selection) as well as PCA computed with the ber of LC–MS-based variables had VIP scores higher than same data sets (Online Resource 3—Fig. S3). 1 (Online Resource 6—Table  S1b). For line status dis- crimination, all variables with VIP scores higher than 1 Unk_8_8059 Unk_8_8305 Unk_2_3638 Unk_8_5032 Unk_8_4474 Unk_ Un 2_7992 k_1_4802 Unk_8_4902 UnM2 k_8 98 _2582 T483 UnUn k_2 k_ _3996 4_0776 Unk_2_3755 M275T1188 Unk_3_2019 Proteine Unk_8_5261 M806T1659 UnUn k_4 k_ _3505 4_2778 THR M238T644 GLY M406T600 Unk_3_3693 Un trigk_ onne 4_1049 lline_ SER Un Un k_ k_ 44 _4642 _4986 GL AL U A Tryptophan M263T1611 ch M2 Un lo 33 rk_ ogena T7 452 _4071 te GABA Pa M5 lmit 21T1 olei 31 c_ 7acid PHE Unk_4_2238 Asp M5 Un 20 k_ T1 430 _5076 1 VAL Un M3 k_ UDP 93 4_5754 T1 _l64 ike3 0 LEU ILE Qui_chloro1 Unk_3_6504 Chloro_hydro PRO Unk_6_4072 Qui_chloro2 LY in Sositol glycinebet Unk_4_2413 M707T617 M668T783 Caffeoylquinic_acid M356T617 M723T996 Un M4 k_ 49 3_7209 T779 M465T749 M163T653 Unk_4_6616 M612T7 M3 31 57T550 Unk_4_8664 M713T740 Unk_3_5363 M401T982 Unk_3_7617 Glucose M326T571 M627T664 M779T982 M343T934 M276T563 M518T710 M2M5 69T1 0931 T13 907 M5 M7 M4 17 M5 75 81 T7 T9 17 T1 10 30 T6 804 80 M341T930 M163T550 M359 M2 T9 59 30 T9 M3 30 55T550 - WW M439T1680 M277T930 M3M5 61T9 03 31 T1804 - DS -6 -4 -2 02 46 -1.0 -0.5 0.00.5 1.0 X-variate 1: 9% expl. var Component 1 M515T8 Un08 k_8_2582 UnM4 k_4_1893 26T1013 Unk_4_0245 M429T658 M377T643 M6 M6 51 59 T1 T1 56 01 31 M181T750 M219T624 M248T994 M345T571 M4 Un 89k_ T96_8936 59 M589T839 M233T13 M289T723 M247T873 M303T1508 M711T M247T994 M551T687 M283T1116 M273T M3 1750 46T M2 M5 M2 03T 93 99 1362 T9 T180 458 M567T1141 M521T637 M205T1425 M459T866 M815T1135 M233T1232 M319T1240 M361T1468 M479T765 M579T884 M565T842 -1.0 -0.5 0.00.5 1.0 -4 -2 024 6 Component 1 X-variate 1: 6% expl. var Fig. 4 PLS-DA of metabolic data sets of sunflower leaf on variables and DS (orange dots). b PLS model scores (left) and loadings plot selected from the set of 588 metabolic variables (Online Resource (right) of the 20 best LASSO selected variables discriminating the 7) after a selection process based on sPLS or LASSO. a PLS model two-line types, B maintainer lines (red dots) and R restorer lines (blue scores (left) and loadings plot (right) of the 50 best sPLS selected dots). Coloured ellipses represent 95% confidence level variables discriminating the two water treatments WW (green dots) 1 3 X-variate 2: 7% expl. var X-variate 2: 7% expl. var -6 -4 -2 02 4 -5 05 Component 2 Component 2 -1.0 -0.5 0.0 0.5 1.0 -1.0 -0.5 0.00.5 1.0 Metabolomic characterization of sunflower leaf allows discriminating genotype groups or… Page 9 of 14 56 were unidentified ions or H-NMR spectral regions (Online 4 Discussion Resource 6—Table S1c). 4.1 Sunflower leaf metabolite composition 3.3 Cost‑efficient metabolic markers Sunflower is an important crop that provides most of the Simplicity of measurement and cost-efficiency of metabolic table oil used worldwide. However, few metabolomic data markers are arguably as important as their prediction capac- are available to date concerning both its primary and special- ity (Fernandez et al. 2016). In other words, measuring a set ized metabolism. We now present one of the largest sets of of markers with a (slightly) lower predictive capacity might primary metabolites in adult sunflower leaf, with absolute be relevant if the marker set is easier or cheaper to measure. quantification of 38 metabolites and with several compounds A simple solution is often to replace untargeted methods not quantified by Moschen et al. (2017) using GC–MS. with targeted ones. We estimated the cost-reduction potential Several points can be made about sunflower leaf compo- by a factor of 3–20 (Fernandez et al. 2016). Another pos- sition. Malate, citrate and chlorogenic acid were the major sibility is to measure globally a family of compounds when organic acids (Fig. 1, Online Resource 4) and linolenic acid, they are affected in the same way by a given treatment or linoleic acid and palmitic acid were the major fatty acids condition, like in our case for amino acids in DS samples detected. This is in contrast with the fatty acids in sunflower (Fig. 1a). seed where linoleic acid is the most abundant. Serine, ala- To illustrate this point, we selected metabolic variables nine and glutamate were the major free amino acids (Fig. 1). (from Online Resource 7) known to be simple or cheap to Glucose and sucrose were the major soluble sugars in leaf measure and relevant for water treatment discrimination. but their concentrations were at least eight times higher than Since all free amino acids measured were increased in DS that of fructose. This might be due to some specificity of the samples, we replaced them by a single variable representing fructose metabolism in the Asteraceae family. In sunflower, their sum (hereafter called total free amino acids). Finally, fructose is not metabolized into inulin (a fructose-derived we chose total free amino acids, citrate, glycine-betaine, ino- polymer) but is transported and then accumulated in the stem. sitol, sucrose, glucose, protein and starch. This set of eight For example, Martínez-Noël et al. (2015) found that fructose variables was offered a clear determination of DS and WW was three times more concentrated than any other soluble samples in an unsupervised analysis (PCA, Fig. 5a). Addi- sugar in this organ. This might explain the difference between tionally, the generated PLS-DA model was efficient with glucose and fructose concentrations in our leaf samples. 2 2 Q = 0.96, and R = 0.55 (Table 1, Online Resource 3—Fig Considering the specialized metabolites detected via S4a). We could not perform this approach for line status LC–ESI–QTOF–MS, the peaks presenting the highest inten- since most of their high VIP-score variables were unidenti- sities were putatively annotated (Online Resource 5). They fied metabolic signatures. include compounds from three families: caffeoylquinates, methyl-flavonoids and sesquiterpenoids. These compounds 3.4 Comparison with physiological variables for DS had all been previously detected in sunflower biochemical markers analyses. Caffeoylquinic acid is a compound commonly found in sunflower. It plays a role in lignification and cor - Physiological markers are used to assess the impact of DS relates with leaf age in sunflower (Koeppe et al. 1970). It on plant. In our experiment, SLA, OSM_POT and CID were is the dominant phenolic acid in sunflower florets (Liang measured in young and mature leaves at the end of DS. To et  al. 2013) and is also present in seeds (Karamać et  al. test the quality of our PLS-DA model built with selected 2012; Pedrosa et al. 2000). When present in sunflower oil, metabolic variables, we compared its discriminative capacity caffeoylquinates including oxidized chlorogenic acid can with a PLS-DA model built with this physiological data- generate green-coloured oxidized complexes by reacting set comprising six variables extracted from a larger dataset with sunflower proteins (Wildermuth et al. 2016). This oxi- published in Blanchet et al. (2018). Unsupervised PCA com- dative reaction between chlorogenic acid and proteins partly puted with this dataset showed poor separation of DS and explains why sunflower proteins are still underused in the WW samples (Fig. 5b). Furthermore, the PLS-DA model food industry, despite their qualities such as their cheapness built with these physiological data displayed a Q = 0.68 and and absence of allergens (Wildermuth et al. 2016). Several an R = 0.54 (Table 1, Fig. 4b), but was less efficient than putative methylated flavonoids were also detected (Online those built with the minimal set of eight metabolic variables Resource 5). These compounds have been used as chemot- 2 2 (Q = 0.96, R = 0.55; Table 1, Online Resource 3—Fig S4a). axonomic markers for the Astereaceae family (Emerenciano et  al. 2001). Finally, specific sunflower sesquiterpenoids were also detected, one of which was putatively identified 1 3 56 Page 10 of 14 O. Fernandez et al. Fig. 5 PCA scores plot generated with a an “easy-to-measure” data OSM_POT and CID) measured the day before final sampling. Left, set (total free amino acids, citrate, glycine-betaine, inositol, glucose, scores plot. Right, loadings plot total proteins and starch) and b six physiological variables (SLA, as niveusin. In sunflower, this compound and its derivatives 4.3 Biomarkers of line status are thought to offer potential as insecticides (Prasifka et al. 2015). Leaf samples of R and B lines were discriminated with the metabolic data set mostly through unidentified markers 4.2 Variable selection process measured by LC–ESI–QTOF–MS (Online Resource 6— Table S1c). R lines, which in sunflower breeding are used to Variable selection is necessary in metabolomics, especially restore the CMS phenotype, have a nuclear-encoded Rfl gene when looking for metabolic markers (Fernandez et al. 2016). that might act as a transcriptional activator (Balk and Leaver However, numerous methods can be used for the variable 2001; Chen and Liu 2014). The only known function of the selection process and have already been the subject of dis- Rfl gene is to restore male fertility in CMS plants (Chen and cussion (for review, Grissa et al. 2016). We submitted our Liu 2014) as well as the associated changes restricted to the initial dataset to three variable selection processes: ANOVA, mitochondria of floral tissues linked with this loss of fertil- sPLS and LASSO penalty. ity (i.e. mitochondrial membrane integrity and respiration 1 3 Metabolomic characterization of sunflower leaf allows discriminating genotype groups or… Page 11 of 14 56 ratio). Phenotypes associated with the presence of CMS amounts of glycine-betaine that are too low to significantly or R genes are thought to be limited to floral tissues. The impact the sap osmotic potential. Rather, it might serve as fact that we were able to discriminate R and B lines using a ROS detoxication agent (Giri 2011). In the case of myo- analyses of leaf metabolites suggests that the phenotype is inositol, Taji et al. (2006) suggested it might be involved in not restricted to flowers and that it might affect other plant osmotolerance, or alternatively serve as a secondary messen- tissues and organs. Interestingly, several organic acids were ger involved in phospholipid signalling pathways. Finally, less concentrated in R line samples, although not individ- caffeoylquinates and sesquiterpenoids (a terpene class with ually significantly. This might be due to an effect on the three isoprene units) were also detected as putative mark- mitochondrial metabolism in all organs, but this hypothesis ers of DS versus WW samples (Online Resources 5 and 6). needs to be confirmed. Further annotations of the associated Caffeoylquinates have been associated with DS responses markers would contribute to propose hypotheses about direct in grapevine (Hochberg et al. 2013). Terpenes have been or indirect R gene effects in leaf. Additionally, metabolomic shown to be involved in thermotolerance and antioxidant markers denote intermediate information between genes and effects (Sharkey et al. 2008). Furthermore, terpenes seem to final phenotypes and might capture multilocus-controlled have radical scavenging activity contributing to the mitiga- traits and associate alleles producing the same final phe- tion of oxidative damage during stresses. In sunflower leaf, notype. The latter property would be interesting in breed- genes involved in terpene metabolism have been shown to be ing programs to predict the restoration phenotype of novel upregulated under drought conditions (Moschen et al. 2017). alleles in pre-breeding programs and therefore to identify novel sources of restoration for the PET1. However, further 4.5 Towards a small efficient biomarker dataset biochemical and statistical analyses with more R and B lines are required since PLS-DA may be prone to overfitting. Fernandez et al. (2016) argued that ideal metabolic markers should be easy and cheap to analyse. For this purpose, we tested the discriminant capacity of a small metabolic marker 4.4 Biomarkers of water treatment set composed of eight biochemical variables: total free amino acids, citrate, glycine-betaine, myo-inositol, sucrose, glucose, The discrimination of WW and DS samples using meta- total proteins and starch. An unsupervised PCA clearly sepa- bolic variables was more efficient than the discrimination rated WW and DS samples when these eight biochemical of line status. Amino acids were clearly the best DS mark- variables were used (Fig. 4a), but not with the physiological ers in our dataset, displaying a 5- to 10-fold increase in DS dataset consisting in six common indicators of DS measured sunflower leaves (Fig.  1a). Increases in amino acids under at plant level. Indeed, SLA, OSM_POT and CID (measured DS in sunflower have already been documented, although to in both young and mature leaves) are often used to character- a lesser extent and in a cultivar-dependent manner (Mani- ise the water–stress status of a given crop (Fig. 4b). This was vannan et al. 2007). This feature has also been detected in confirmed when comparing Q values for PLS-DA models other crops such as barley (Lanzinger et al. 2015) and wheat computed with each of these data sets (0.91 and 0.68 respec- (Bowne et  al. 2012). Conversely, Moschen et  al. (2017) tively). However, since amino acids were overrepresented in found that the concentrations of several leaf amino acids our PLS-DA model VIPs, our approach might not be general- were decreased under DS in sunflower (Correia et al. 2005). izable to any given criterion. Indeed, reducing the number of These contradictory results regarding amino acid responses variables was much less efficient in discriminating line status. might be due to water–stress intensity, sampling stage or Furthermore, given the fact that amino acid accumulation differences in nitrogen nutrition. In the present study, the use is not always reported for sunflower experiencing drought, of Heliaphen high-throughput phenotyping platform allowed more studies with various drought scenarios and more lines the application of a precise and reproducible drought sce- will be required to confirm our conclusions. Finding the right nario that is available for more thorough understanding of balance between cost reduction and prediction efficiency of the impact of DS on leaf metabolism. Nevertheless, higher each metabolic marker set is likely an achievable goal in concentrations of individual amino acids such as proline many situations but will certainly require optimisation for and glycine have been detected in DS leaves (Moschen et al. each performance criterion studied. 2017). Amino acids, and especially proline, might partici- pate in osmotolerance under DS, although the case is highly debated for the latter (Szabados and Savouré 2010).5 Conclusions In our dataset, other metabolites appeared as good mark- ers of DS samples, i.e. glycine-betaine and myo-inositol. Metabolic markers are a recent development in science. Glycine-betaine is accumulated in various plants under Applications such as personalized medicine have recently abiotic stress (Giri 2011). Generally, plants accumulate attracted keen interest (Lindon and Nicholson 2014). Their 1 3 56 Page 12 of 14 O. Fernandez et al. use in agronomy as a potential tool for crop breeding is even Compliance with ethical standards more recent (Fernandez et al. 2016). In the present work, we Conflict of interest The authors declare that they have no conflict of show that a limited number of metabolic markers can dis- interest. criminate plant sample groups with different characteristics or treatment applications, especially in the case of DS. This Research involving human and/or animal participants This study did not involve the use of animal or human samples. feature was already noted at early stages of plant develop- ment in maize (Riedelsheimer et al. 2012). The fact that leaves of sunflower lines carrying different alleles of the Open Access This article is distributed under the terms of the Crea- tive Commons Attribution 4.0 International License (http://creat iveco CMS restoration gene were separated by this approach shows mmons.or g/licenses/b y/4.0/), which permits unrestricted use, distribu- that metabolomics can reveal an unsuspected metabolic phe- tion, and reproduction in any medium, provided you give appropriate notype in a given organ. The present work also emphasizes credit to the original author(s) and the source, provide a link to the the importance of variable selection. The pipeline we pro- Creative Commons license, and indicate if changes were made. pose (Fig. 2) may not be optimised for all situations (sam- ple numbers, organ types, analytical approaches…), but will provide a preliminary guideline for future users. Another important point is the specificity of the list of selected mark - References ers towards the selected stress. Indeed, several metabolites Allinne, C., Maury, P., Sarrafi, A., & Grieu, P. (2009). Genetic control could be considered as valid metabolic markers of different of physiological traits associated to low temperature growth in stresses, simply because their concentrations may be signifi- sunflower under early sowing conditions. Plant Science, 177(4), cantly altered under various stress situations. To alleviate 349–359. https ://doi.org/10.1016/j.plant sci.2009.07.002. this bias, these markers should be tested under various stress Badouin, H., Gouzy, J., Grassa, C. J., Murat, F., Staton, S. E., Cottret, L., et al. (2017). The sunflower genome provides insights into oil scenarios (Fernandez et al. 2016). It will indeed be crucial metabolism, flowering and Asterid evolution. Nature, 546(7656), to verify whether such a modelling approach remains valid 148–152. https ://doi.org/10.1038/natur e2238 0. when the predictive metabolomic data and the predicted Balk, J., & Leaver, C. J. (2001). The PET1-CMS mitochondrial muta- phenotypic data are obtained in different experiments. In tion in sunflower is associated with premature programmed cell death and cytochrome c release. The Plant Cell, 13(8), 1803– particular, the possibility to use young plantlets, grown in 1818. https ://doi.org/10.1105/TPC.01011 6. controlled conditions and “metabotyped”, to predict a field Blanchet, N., Casadebaig, P., Debaeke, P., Duruflé, H., Gody, L., Gos- phenotype of interest could be extremely useful especially seau, F., et  al. (2018). Data describing the eco-physiological regarding cost reduction, but will require extreme caution. responses of twenty-four sunflower genotypes to water deficit. Submitted: Data in Brief. Thus, a careful methodology with a clear choice of perfor- Bowne, J. B., Erwin, T. A., Juttner, J., Schnurbusch, T., Langridge, P., mance criteria (see Fernandez et al. 2016), stress scenario, Bacic, A., et al. (2012). Drought responses of leaf tissues from developmental stages and analytical methods will have to be wheat cultivars of die ff ring drought tolerance at the metabolite level. developed to test this hypothesis. Molecular Plant, 5(2), 418–429. https://doi.or g/10.1093/mp/ssr114 . Bradford, M. M. (1976). A rapid and sensitive method for the quantita- tion of microgram quantities of protein utilizing the principle of Acknowledgements We thank Laetitia Fouillen for her help with the protein-dye binding. Analytical Biochemistry, 72, 248–254. lipid analyses and the Heliaphen team (especially Nicolas Blanchet) Chen, L., & Liu, Y.-G. (2014). Male sterility and fertility restoration for plant culture. We also thank Dr. Ray Cooke for language proofread- in crops. Annual Review of Plant Biology, 65(1), 579–606. https ing and editing. Metabolite and lipid analyses were performed at the ://doi.org/10.1146/annur ev-arpla nt-05021 3-04011 9. Bordeaux Metabolome Facility, MetaboHUB. Correia, M. J., Fonseca, F., Azedo-Silva, J., Dias, C., David, M. M., Barrote, I., et al. (2005). Effects of water deficit on the activ - Author contributions OF, NL, YG and AM wrote and corrected the final ity of nitrate reductase and content of sugars, nitrate and free manuscript and designed the experimental procedure, with input from all amino acids in the leaves and roots of sunflower and white other authors. OF performed spectrophotometric and UPLC analysis. TB lupin plants growing under two nutrient supply regimes. performed the LC–MS–MS acquisitions, TB and SB performed the LC– Physiologia Plantarum, 124(1), 61–70. https ://doi.org/10.111 ESI–QTOF–MS annotations, and OF analysed the data. MM and CD ran 1/j.1399-3054.2005.00486 .x. the H-NMR acquisitions and identifications. CD and DJ produced the Deborde, C., Maucourt, M., Baldet, P., Bernillon, S., Biais, B., Talon, NMR absolute quantitative data. OF analysed the H-NMR data. HD pro- G., et al. (2009). Proton NMR quantitative profiling for quality vided support for statistical analysis and g fi ure design. PM provided insight assessment of greenhouse-grown tomato fruit. Metabolomics, on the physiological data. MU provided support for amino acid analysis. 5(2), 183–198. https ://doi.org/10.1007/s1130 6-008-0134-2. Emerenciano, V. P., Militão, J. S. L. T., Campos, C. C., Romoff, P., Funding Olivier Fernandez and Maria Urrutia were funded by ‘Agence Kaplan, M. A. C., Zambon, M., et  al. (2001). Flavonoids as Nationale de la Recherche’ (ANR) through the SUNRISE (ANR-11- chemotaxonomic markers for Asteraceae. Biochemical Systemat- BTBR-0005) and AMAIZING (ANR-10-BTBR-0001) projects respec- ics and Ecology, 29(9), 947–957. https ://doi.org/10.1016/S0305 tively. We acknowledge the MetaboHUB (ANR-11-INBS-0010), PHE- -1978(01)00033 -3. NOME (ANR-11-INBS-0012) and SUNRISE (ANR-11-BTBR-0005) Fernandez, O., Urrutia, M., Bernillon, S., Giauffret, C., Tardieu, F., Le projects for further funding. Gouis, J., et al. (2016). Fortune telling: Metabolic markers of plant 1 3 Metabolomic characterization of sunflower leaf allows discriminating genotype groups or… Page 13 of 14 56 performance. Metabolomics, 12(10), 158. https://doi.or g/10.1007/ Lanzinger, A., Frank, T., Reichenberger, G., Herz, M., & Engel, K.-H. s1130 6-016-1099-1. (2015). Metabolite profiling of barley grain subjected to induced Fu, G.-H., Zhang, B.-Y., Kou, H.-D., & Yi, L.-Z. (2017). Stable bio- drought stress: Responses of free amino acids in differently marker screening and classification by subsampling-based sparse adapted cultivars. Journal of Agricultural and Food Chemistry, regularization coupled with support vector machines in metabo- 63(16), 4252–4261. https ://doi.org/10.1021/acs.jafc.5b011 14. lomics. Chemometrics and Intelligent Laboratory Systems, 160, Liang, Q., Cui, J., Li, H., Liu, J., & Zhao, G. (2013). Florets of sun- 22–31. https ://doi.org/10.1016/j.chemo lab.2016.11.006. flower (Helianthus annuus L.): Potential new sources of dietary Gibon, Y., Vigeolas, H., Tiessen, A., Geigenberger, P., & Stitt, M. b fi er and phenolic acids. Journal of Agricultural and Food Chem- (2002). Sensitive and high throughput metabolite assays for inor- istry, 61(14), 3435–3442. https ://doi.org/10.1021/jf400 569a. ganic pyrophosphate, ADPGlc, nucleotide phosphates, and glyco- Lindon, J. C., & Nicholson, J. K. (2014). The emergent role of meta- lytic intermediates based on a novel enzymic cycling system. The bolic phenotyping in dynamic patient stratification. Expert Opin- Plant Journal, 30(2), 221–235. https ://doi.org/10.1046/j.1365- ion on Drug Metabolism & Toxicology, 10(7), 915–919. https :// 313X.2001.01278 .x.doi.org/10.1517/17425 255.2014.92295 4. Giri, J. (2011). Glycinebetaine and abiotic stress tolerance in plants. Manivannan, P., Jaleel, C. A., Sankar, B., Kishorekumar, A., Plant Signaling & Behavior, 6(11), 1746–1751. https ://doi. Somasundaram, R., Lakshmanan, G. M. A., et al. (2007). Growth, org/10.4161/psb.6.11.17801. biochemical modifications and proline metabolism in Helianthus Gosseau, F., Blanchet, N., Varès, D., Burger, P., Campergue, D., annuus L. as induced by drought stress. Colloids and Surfaces Colombet, C., et al. (2018). Heliaphen, an outdoor high-through- B: Biointerfaces, 59(2), 141–149. https ://doi.org/10.1016/j.colsu put phenotyping platform designed to integrate genetics and crop rfb.2007.05.002. modeling. bioRxiv, 362715. https ://doi.org/10.1101/36271 5 Marchand, G., Mayjonade, B., Varès, D., Blanchet, N., Boniface, Grissa, D., Pétéra, M., Brandolini, M., Napoli, A., Comte, B., & Pujos- M.-C., Maury, P., et al. (2013). A biomarker based on gene expres- Guillot, E. (2016). Feature selection methods for early predictive sion indicates plant water status in controlled and natural environ- biomarker discovery using untargeted metabolomic data. Fron- ments. Plant, Cell and Environment, 36(12), 2175–2189. https :// tiers in Molecular Biosciences, 3, 10. htt ps :/ /doi. org/ 10.33 89/ doi.org/10.1111/pce.12127 . fmolb .2016.00030 . Martínez-Noël, G. M. A., Dosio, G. A. A., Puebla, A. F., Insani, E. M., Hartmann, T. (2007). From waste products to ecochemicals: Fifty years & Tognetti, J. A. (2015). Sunflower: A potential fructan-bearing research of plant secondary metabolism. Phytochemistry, 68(22), crop? Frontiers in Plant Science, 6, 798. https ://doi.org/10.3389/ 2831–2846. https ://doi.org/10.1016/j.phyto chem.2007.09.017. fpls.2015.00798 . Hendriks, J. H. M., Kolbe, A., Gibon, Y., Stitt, M., & Geigenberger, Meyer, R. C., Steinfath, M., Lisec, J., Becher, M., Witucka-Wall, H., P. (2003). ADP-glucose pyrophosphorylase is activated by post- Törjék, O., et al. (2007). The metabolic signature related to high translational redox-modification in response to light and to sugars plant growth rate in Arabidopsis thaliana. Proceedings of the in leaves of Arabidopsis and other plant species. Plant Physiology, National academy of Sciences of the United States of America, 133(2), 838–849. https ://doi.org/10.1104/pp.103.02451 3. 104(11), 4759–4764. https ://doi.org/10.1073/pnas.06097 09104 . Hochberg, U., Degu, A., Toubiana, D., Gendler, T., Nikoloski, Z., Moschen, S., Rienzo, J. A. D., Higgins, J., Tohge, T., Watanabe, M., Rachmilevitch, S., et al. (2013). Metabolite profiling and net- González, S., et al. (2017). Integration of transcriptomic and meta- work analysis reveal coordinated changes in grapevine water bolic data reveals transcription factors involved in drought stress stress response. BMC Plant Biology, 13(1), 184. https ://doi. response in sunflower Helianthus annuus L. Plant Molecular Biol- org/10.1186/1471-2229-13-184. ogy, 94(4–5), 549–564. https: //doi.org/10.1007/s11103 -017-0625-5. Horn, R., & Friedt, W. (1999). CMS sources in sunflower: Different Nunes-Nesi, A., Carrari, F., Gibon, Y., Sulpice, R., Lytovchenko, A., origin but same mechanism? Theoretical and Applied Genetics, Fisahn, J., et al. (2007). Deficiency of mitochondrial fumarase 98(2), 195–201. https ://doi.org/10.1007/s0012 20051 058. activity in tomato plants impairs photosynthesis via an effect on Hussain, M., Farooq, S., Hasan, W., Ul-Allah, S., Tanveer, M., Farooq, stomatal function. The Plant Journal, 50(6), 1093–1106. https :// M., et  al. (2018). Drought stress in sunflower: Physiological doi.org/10.1111/j.1365-313X.2007.03115 .x. effects and its management through breeding and agronomic alter - Oilworld. (2016). https ://www.oilwo rld.biz/. natives. Agricultural Water Management, 201, 152–166. https :// Owart, B. R., Corbi, J., Burke, J. M., & Dechaine, J. M. (2014). Selec- doi.org/10.1016/j.agwat .2018.01.028. tion on crop-derived traits and QTL in sunflower (Helianthus Igarashi, K., Kazama, T., & Toriyama, K. (2016). A gene encoding annuus) crop-wild hybrids under water stress. PLoS ONE, 9(7), pentatricopeptide repeat protein partially restores fertility in e102717. https ://doi.org/10.1371/journ al.pone.01027 17. RT98-type cytoplasmic male-sterile rice. Plant and Cell Physi- Patil, C., Calvayrac, C., Zhou, Y., Romdhane, S., Salvia, M.-V., ology, 57(10), 2187–2193. https ://doi.org/10.1093/pcp/pcw13 5. Cooper, J.-F., et al. (2016). Environmental metabolic footprint- Jacob, D., Deborde, C., Lefebvre, M., Maucourt, M., & Moing, A. ing: A novel application to study the impact of a natural and a (2017). NMRProcFlow: A graphical and interactive tool dedicated synthetic β-triketone herbicide in soil. The Science of the Total to 1D spectra processing for NMR-based metabolomics. Metabo- Environment, 566–567, 552–558. https ://doi.org/10.1016/j.scito lomics, 13(4), 36. https ://doi.org/10.1007/s1130 6-017-1178-y. tenv.2016.05.071. Jelitto, T., Sonnewald, U., Willmitzer, L., Hajirezeai, M., & Stitt, M. Pedrosa, M. M., Muzquiz, M., García-Vallejo, C., Burbano, C., Cuad- (1992). Inorganic pyrophosphate content and metabolites in potato rado, C., Ayet, G., et al. (2000). Determination of caffeic and chlo- and tobacco plants expressing E. coli pyrophosphatase in their cyto- rogenic acids and their derivatives in different sunflower seeds. sol. Planta, 188(2), 238–244. https://doi.or g/10.1007/bf00216819 . Journal of the Science of Food and Agriculture, 80(4), 459–464. Karamać, M., Kosińska, A., Estrella, I., Hernández, T., & Dueñas, Pichersky, E., & Lewinsohn, E. (2011). Convergent evolution in plant M. (2012). Antioxidant activity of phenolic compounds identified specialized metabolism. Annual Review of Plant Biology, 62, 549– in sunflower seeds. European Food Research and Technology, 566. https ://doi.org/10.1146/annur ev-arpla nt-04211 0-10381 4. 235(2), 221–230. https ://doi.org/10.1007/s0021 7-012-1751-6. Poormohammad Kiani, S., Grieu, P., Maury, P., Hewezi, T., Gentzbit- Koeppe, D. E., Rohrbaugh, L. M., Rice, E. L., & Wender, S. H. (1970). tel, L., & Sarrafi, A. (2007). Genetic variability for physiological Tissue age and caffeoylquinic acid concentration in sunflower. traits under drought conditions and differential expression of water stress-associated genes in sunflower (Helianthus annuus L). TAG. Phytochemistry, 9(2), 297–301. https ://doi.org/10.1016/S0031 Theoretical and Applied Genetics, 114(2), 193–207. -9422(00)85138 -9. 1 3 56 Page 14 of 14 O. Fernandez et al. Prasifka, J. R., Spring, O., Conrad, J., Cook, L. W., Palmquist, D. E., Taji, T., Takahashi, S., & Shinozaki, K. (2006). Inositols and their & Foley, M. E. (2015). Sesquiterpene lactone composition of metabolites in abiotic and biotic stress responses. In A. L. Majum- wild and cultivated sunflowers and biological activity against an der & B. B. Biswas (Eds.), Biology of inositols and phospho- insect pest. Journal of Agricultural and Food Chemistry, 63(16), inositides: Subcellular biochemistry (pp. 239–264). Boston, MA: 4042–4049. https ://doi.org/10.1021/acs.jafc.5b003 62. Springer. https ://doi.org/10.1007/0-387-27600 -9_10. Qi, L. L., Foley, M. E., Cai, X. W., & Gulya, T. J. (2016). Genetics Talukder, Z. I., Hulke, B. S., Qi, L., Scheffler, B. E., Pegadaraju, V., and mapping of a novel downy mildew resistance gene, Pl(18), McPhee, K., et al. (2014). Candidate gene association mapping of introgressed from wild Helianthus argophyllus into cultivated Sclerotinia stalk rot resistance in sunflower (Helianthus annuus sunflower (Helianthus annuus L.). TAG. Theoretical and Applied L.) uncovers the importance of COI1 homologs. TAG. Theoretical Genetics, 129(4), 741–752. and applied genetics, 127(1), 193–209. https ://doi.org/10.1007/ Riedelsheimer, C., Czedik-Eysenberg, A., Grieder, C., Lisec, J., Tech-s0012 2-013-2210-x. now, F., Sulpice, R., et al. (2012). Genomic and metabolic predic- Tompkins, D., & Toffaletti, J. (1982). Enzymic determination of cit- tion of complex heterotic traits in hybrid maize. Nature Genetics, rate in serum and urine, with use of the Worthington “ultrafree” 44(2), 217–220. https ://doi.org/10.1038/ng.1033. device. Clinical Chemistry, 28(1), 192–195. Rohart, F., Gautier, B., Singh, A., & Cao, K.-A. L. (2017). mixOmics: Vear, F. (2016). Changes in sunflower breeding over the last fifty years. An R package for ‘omics feature selection and multiple data inte- OCL, 23(2), D202. https ://doi.org/10.1051/ocl/20160 06. gration. PLoS Computational Biology, 13(11), e1005752. https:// Wildermuth, S. R., Young, E. E., & Were, L. M. (2016). Chloro- doi.org/10.1371/journ al.pcbi.10057 52. genic acid oxidation and its reaction with sunflower proteins Schneiter, A. A., & Miller, J. F. (1981). Description of sunflower to form green-colored complexes. Comprehensive Reviews in growth stages 1. Crop Science, 21(6), 901–903. https ://doi. Food Science and Food Safety, 15(5), 829–843. https ://doi. org/10.2135/crops ci198 1.00111 83X00 21000 60024 x. org/10.1111/1541-4337.12213. Sharkey, T. D., Wiberley, A. E., & Donohue, A. R. (2008). Isoprene Yu, Y., Zhao, Z., Shi, Y., Tian, H., Liu, L., & Bian, X. (2016). Hybrid emission from plants: Why and how. Annals of Botany, 101(1), sterility in rice (Oryza sativa L.) involves the tetratricopeptide 5–18. https ://doi.org/10.1093/aob/mcm24 0. repeat domain containing protein. Genetics, 203(3), 1439–1451. Smith, C. A., Want, E. J., O’Maille, G., Abagyan, R., & Siuzdak, G. https ://doi.org/10.1534/genet ics.115.18384 8. (2006). XCMS: Processing mass spectrometry data for metabo- lite profiling using nonlinear peak alignment, matching, and Publisher’s Note Springer Nature remains neutral with regard to identification. Analytical Chemistry, 78(3), 779–787. https ://doi. jurisdictional claims in published maps and institutional affiliations. org/10.1021/ac051 437y. Szabados, L., & Savouré, A. (2010). Proline: A multifunctional amino acid. Trends in Plant Science, 15(2), 89–97. https ://doi. org/10.1016/j.tplan ts.2009.11.009. Affiliations 1,5 1,2,6 1,7 1,3 1,3 Olivier Fernandez  · Maria Urrutia  · Thierry Berton  · Stéphane Bernillon  · Catherine Deborde  · 1,3 1,3,6 4 4 1,3 4 Daniel Jacob  · Mickaël Maucourt  · Pierre Maury  · Harold Duruflé  · Yves Gibon  · Nicolas B. Langlade  · 1,3 Annick Moing Maria Urrutia Annick Moing m.urrutia@enzazaden.es annick.moing@inra.fr Thierry Berton UMR1332 Biologie du Fruit et Pathologie, INRA, thierry.berton@laposte.net Centre INRA de Bordeaux, 71 av Edouard Bourlaux, Stéphane Bernillon 33140 Villenave d’Ornon, France stephane.bernillon@inra.fr UMR AgroImpact, INRA, Estrées-Mons, 80203 Péronne, Catherine Deborde France catherine.deborde@inra.fr Plateforme Métabolome Bordeaux, CGFB, Daniel Jacob MetaboHUB-PHENOME, 33140 Villenave d’Ornon, France daniel.jacob@inra.fr UMR LIPM, INRA, CNRS, Université de Toulouse, Mickaël Maucourt 31326 Castanet-Tolosan, France mickael.maucourt@inra.fr Present Address: Laboratoire RIBP, Université de Reims Pierre Maury Champagne Ardenne, Moulin de la Housse Chemin des pierre.maury@ensat.fr Rouliers, 51100 Reims, France Harold Duruflé Present Address: Enza Zaden Centro de Investigacion S.L., harold.durufle@inra.fr Santa Maria del Aguila, 04710 Almeria, Spain Yves Gibon Present Address: Centre for CardioVascular and Nutrition, yves.gibon@inra.fr UMR INRA-INSERM, Aix-Marseille Univ, INSERM, 13005 Marseilles, France Nicolas B. Langlade nicolas.langlade@inra.fr 1 3

Journal

MetabolomicsSpringer Journals

Published: Mar 30, 2019

References

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create folders to
organize your research

Export folders, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off