ORIGINAL PAPER

Mass spectrometry-based metabolomic fingerprintingfor screening cold tolerance in Arabidopsis thaliana accessions

Lukas Vaclavik & Anamika Mishra & Kumud B. Mishra &

Jana Hajslova

Received: 7 November 2012 /Revised: 13 December 2012 /Accepted: 21 December 2012 /Published online: 17 January 2013# Springer-Verlag Berlin Heidelberg 2013

Abstract The availability of rapid and reliable tools for moni-toring of plants’ cold tolerance is a prerequisite for researchaimed at breeding of cold-tolerant crop plants. Therefore, wehave tested the capacity of metabolomics-based methodsemploying ultra-high-performance liquid chromatography(UHPLC)–mass spectrometry and direct analysis in realtime–mass spectrometry for high-throughput screening ofcold tolerance in eight differentially cold-tolerant accessionsof Arabidopsis thaliana. Metabolomic fingerprinting of leaftissues was performed in methanolic extracts for (1) 6-week-old non-acclimated (NAC) plants grown at room temperature,(2) NAC plants cold-acclimated (ACC) at 4 °C for 2 weeks,and (3) cold-acclimated plants given sub-zero-temperaturetreatments by slow cooling at −4 °C for 8 h. The generatedchromatograms and mass spectra were processed with the useof multivariate statistical analysis employing principal com-ponent analysis (PCA) and linear discriminant analysis. ThePCA of metabolomic fingerprints classified the investigatedA. thaliana accessions into three categories with low, interme-diate, and high cold tolerance for both the cold-acclimated andthe sub-zero-temperature-treated plants. This indicates thepotential application of metabolomics-based fingerprintingfor measuring cold tolerance in the cold-acclimated state,i.e., without treating plants at freezing temperatures that is

required by currently available methods. Furthermore, weemployed UHPLC coupled to the quadrupole-time-of-flightmass spectrometry to identify characteristic metabolites inACC state and found the abundance of gluconapin andflavon-3-ol glycosides, respectively, in the cold-sensitive andthe cold-tolerant accessions.

Keywords Cold tolerance . Arabidopsis thaliana .

Metabolomic fingerprinting . LC–MS . DART–MS .

Chemometric analysis

Introduction

Tolerance to low temperatures is an important factor that limitsgeographical distribution of wild plant species and adverselyinfluences growth, development, and yield of the importantagricultural crops. In nature, many plant species gain theircold tolerance by acclimation to low but non-freezing temper-atures through highly complex cold acclimation mechanismsand can survive lethal conditions of the harsh winter [1]. Thecold acclimation induces extensive changes in plant transcrip-tome, proteome, and metabolome, as well as compositionaland structural modifications of cytoplasm, membranes, andcell walls [1, 2]. As a result of exposure to sub-zero-temper-atures, multiple mechanisms are initiated to protect leaf tissuesagainst freezing-induced injuries, including enhancement inintracellular solutes and the accumulation of cryoprotectivemetabolites and proteins [3]. The consequent gain in coldtolerance to survive extent of freezing temperatures is largelydifferent for the plants originating from warm and cool cli-mates [4, 5]. Since Arabidopsis (Arabidopsis thaliana) growsover a broad geographic range and is exposed to diversetemperatures, natural accessions of this model plant offer an

L. Vaclavik : J. Hajslova (*)Institute of Chemical Technology, Prague, Department of FoodAnalysis and Nutrition, Technická 3,166 28 Prague, Czech Republice-mail: jana.hajslova@vscht.cz

A. Mishra :K. B. MishraCzechGlobe–Global Change Research Centre AS CR, v.v.i.,Division of Impact Studies and Physiological Analyses,Laboratory of Plant Ecological Physiology, Zámek 136,373 33 Nové Hrady, Czech Republic

Anal Bioanal Chem (2013) 405:2671–2683DOI 10.1007/s00216-012-6692-1

excellent opportunity to study natural variations of adaptivetraits including cold tolerance [5–8]. In fact, it is believed thatthe discoveries made in Arabidopsis are, in general, valuableand applicable to other plant species as well.

A prerequisite condition for the research aiming towardimprovement of breeding strategy of cold-tolerant crops isthe availability of rapid, reliable, and cost-effective toolsthat can measure plants cold tolerance. The classical meth-ods based on the measurements of the plants survival ratesare statistically simple, but the accuracy of these proceduresis always questionable because uncontrolled temperatures inthe growth chamber can strongly influence the results [9].The most accepted and popular method for quantitativemeasurement of plant cold tolerance is the determinationof the lethal temperature (LT50) at which 50 % of electro-lytes are released from the plasma membranes by usingconductivity measurements in freeze–thawed leaves [5,10–12]. In this method, detached leaves of the cold-acclimated or non-acclimated plants are placed in the reac-tion tubes that are cooled inside a programmable coolingbath with controlled temperature and cooling rate, and theleaves are drawn at various sub-zero-temperatures for thaw-ing followed by its conductivity measurements [13, 14]. Analternative non-invasive approach is the analysis offreezing-induced changes in chlorophyll fluorescence(ChlF) emission that is not directed to the thylakoid or cellintegrity but its intrinsic functional and biophysical proper-ties [15]. The LT50 of ChlF parameter maximum yield ofPSII photochemistry (Fv/Fm) has successfully been appliedto quantify cold tolerance in Arabidopsis as well as in otherplant species [16, 17]. Mishra et al. have recently measuredthe chlorophyll fluorescence transients of differentially coldtolerance A. thaliana accessions during progressive coolingstarting from room temperature to −15 °C and found highcorrelation between LT50 measured by electrolyte leakagewith the other chlorophyll fluorescence parameters such asFo( −15 °C )/Fo (4 °C) and [Fs/Fo]−15°C [4]. They alsoemployed advanced statistics-based combinatorial imagingmethods on the measured time-resolved chlorophyll fluores-cence images and found the applicability of methods todetect the cold tolerance at mild sub-zero-temperatures ofaround −4 °C [4]. Cold acclimation, in general, improvesthe plants’ capacity to tolerate freezing temperature byaround −2 to −6 °C in the investigated natural accessionsof A. thaliana and provides better discrimination capacity ascompared with non-acclimated plants by both the electrolyteleakage methods [5] and chlorophyll fluorescence emission[4]. However, none of the above-mentioned methods pro-vide information about the plants cold tolerance capacitywithout cooling them to sub-zero-temperatures.

In the past decade, the area of metabolomics, aiming atcomprehensive analysis of numerous metabolites (low mo-lecular compounds) in a biological sample, have found its

application in diverse research fields [18, 19], includingthose focused on plant response to cold stress [20]. Thisboost of metabolomics-based studies has been enabled bysubstantial advances in analytical platforms facilitating anal-ysis of a wide range of metabolites differing both byphysico-chemical properties and abundance. The techniquesmost frequently employed for metabolomic fingerprintingand profiling of biological samples are nuclear magneticresonance and mass spectrometry (MS) coupled either toseparation techniques employing liquid chromatography(LC), gas chromatography, and capillary electrophoresis[21] or direct (ambient) ionization methods that includedirect infusion electrospray ionization mass spectrometry(DI-ESI–MS) [22], desorption electrospray ionization [23],or direct analysis in real time (DART–MS) [24]. Advanceddata processing software algorithms and multivariate che-mometric tools are typically needed to process complex andmultidimensional metabolomics data for the data mining,pre-selection of the signals related to marker metabolites,and for the overall data interpretation [25].

Both the targeted and the non-targeted metabolomicsworkflows can be used to study metabolic response andtolerance to low temperature stress in Arabidopsis [2].Kaplan et al. employed gas chromatography–mass spec-trometry (GC–MS) and chemometrics to assess the metab-olome of temperature-stressed Arabidopsis plants andidentified numerous metabolites responding to either high-or low-temperature treatment [26]. In similar studies, metab-olome exploration of contrast ecotypes of Arabidopsis withthe use of GC–MS showed significant natural variation forlow-temperature tolerances and preceding acclimatory pro-cesses [27, 28]. The intraspecific variation in cold-temperature metabolic phenotypes of Arabidopsis lyratawas also assessed using targeted metabolomic profiles ofcarbohydrates and free amino acids and non-targeted finger-prints obtained by DI-ESI–MS utilizing time-of-flight(TOF) mass analyzer [29].

In the current study, we have tested the feasibility ofUHPLC–MS technique employing a hybrid triple quadru-pole/linear ion trap (QqQ/LIT) mass analyzer and DART–TOFMS instrumentation for the discrimination of cold tol-erance in A. thaliana accessions originating from differentparts of northern hemisphere. The UHPLC–MS and DART–MS data were processed with multivariate chemometricmethods of principal component analysis (PCA) and lineardiscriminant analysis (LDA) to assess their potential appli-cation for differentiation of cold tolerance of the investigat-ed accessions in: (1) non-acclimated, (2) cold-acclimated for2 weeks at 4 °C, and (3) cold-acclimated plants treated withsub-zero-temperature (−4 °C) for 8 h followed by thawing.Furthermore, we have acquired accurate tandem mass spec-tra (MS/MS) using quadrupole-time-of-flight mass spec-trometry (QTOFMS) to estimate probable elemental

2672 L. Vaclavik et al.

formulae for tentative identification of characteristic markermetabolites.

Experimental

Chemicals

Deionized water was obtained from Milli-Q purificationsystem (Millipore, Bedford, MA, USA). High-performanceliquid chromatography-grade methanol was supplied byMerck (Darmstadt, Germany). Formic acid (≥95 %) andpolyethylene glycol (PEG, average relative molecularweight 600 Da) were provided by Sigma-Aldrich(Steinheim, Germany).

Plant material and growth conditions

Eight A. thaliana accessions: the Tenela (Te, Finland), theRschew (Rsch, Russia), the Columbia-0 (Col-0, geneticallyrelated to Gückingen, Germany), the Niederzenz (Nd,Germany), the Landsberg erecta (Ler, Poland), the Cvi(Cape Verde Islands), and the Coimbra (Co) and the C24from Portugal (for more detail about these accessions, seeHannah et al [5] and A Mishra et al. [4]) were grown in agrowth chamber under controlled temperature (20 °C/18 °C,day/night), light [~90 μmol (photons)m−2s−1], photoperiods(L/D=12:12 h) and humidity (~70 %). The investigatedplants (12 plants of each accession with exception of Ler,for which only six plants were grown) were initially germi-nated in individual pots, and after 2 weeks of their growth,seedlings were re-transplanted in a 6-cm pot filled withequal amount of natural soil. Plants were watered everyalternate day and fertilized with standard fertilizers every2 weeks. Plants were evenly divided into three subsets. Six-week-old plants (first subset) grown at above-mentionedcondition were treated as a non-acclimated (NAC) set.Remaining plants were cold-acclimated (ACC) and keptfor 2 weeks at 4 °C. One subset of the ACC plants wasslowly cooled down in a freezer at a rate of approximately2 °Ch−1 and kept at −4 °C for about 8 h followed by slowheating at 4 °C [sub-zero-temperature treatment (ST-trea-ted)]. The leaf tissues collected from independent plantswere weighted and immediately extracted as describedbelow.

Extraction procedure

Two to three leaf tissues of each accession for NAC, ACC,and ST-treated samples with weight around 150 mg weretransferred into a grinding mortar, frozen with liquid nitro-gen, and homogenized with a pestle. Obtained powder wasextracted with 2.5 mL of methanol for 1 min. An aliquot of

the mixture was placed into a 1 mL Eppendorf tube andcentrifuged for 5 min (4,000 rpm, 20 °C). The supernatantwas passed through a 0.22 μm PTFE filter and transferredinto a 2 mL amber autosampler vial. The concentration ofthe sample matrix in the methanolic extract was adjusted to50 mgmL−1 by the addition of exact volume of the extrac-tion solvent. Prior to instrumental analysis, samples werestored in dark and dry environment at −70 °C. For thepurpose of UHPLC–QTOFMS analysis, a tenfold sampledilution with methanol was performed to avoid the satura-tion of the MS detector.

Ultra-high-performance liquid chromatography–massspectrometry (UHPLC–MS)

Two MS instruments using identical Acquity UltraPerformance LC system (Waters, Milford, MA, USA) wereused in this study. While an AB SCIEX 5500 QTRAPhybrid triple quadrupole/linear ion trap mass spectrometer(AB SCIEX, Concord, Ontario, Canada) was used for themeasurement of UHPLC–MS metabolomic fingerprints ofthe samples, experiments aimed at marker compounds iden-tification were performed with an Synapt G2 HDquadrupole-time-of-flight mass spectrometer (Waters,Milford, MA, USA).

Chromatographic separation was carried out using anAcquity UPLC HSS T3 reversed-phase analytical column(50 mm×2.1 mm i.d., 1.7 μm particle size, Waters, Milford,MA, USA) maintained at 40 °C. Sample volume of 3 μLwith the partial loop injection mode was used in all experi-ments. Autosampler temperature was kept at 10 °C. Themobile phase consisted of 0.1 % aqueous formic acid (A)and 0.1 % formic acid in methanol (B). The followinggradient elution was carried out at 0–6 min eluent B 10–100 % (flow 0.35–0.45 mLmin−1); 6–10 min eluent B100 % (flow 0.45–0.60 mLmin−1); and 10–12 min columnequilibration–eluent B 10 % (flow 0.60 mLmin−1).

The 5500 QTRAP was equipped with an electrospray(ESI) ion source. In the positive ESI, mode parameters ofthe source were as follows: capillary voltage (+5,500 V),nebulizing gas pressure (50 psi), drying gas pressure(60 psi), temperature (500 °C), and declustering potential(80 V). The capillary voltage in negative ESI was −4,500 V;other source settings were the same as for positive ESI.Mass spectrometer was operated using an information de-pendent acquisition method to collect full-scan MS and MS/MS information simultaneously. The two MS experiments(positive and negative polarity) consisted of a survey scanusing enhanced MS scanning (EMS) from m/z 100 to m/z1,000 at 4,000 amus−1 and enhanced product ion (EPI)scanning from m/z 50 to m/z 1,000 at 10,000 Das−1. EPIspectra were collected automatically for the five most in-tense ions in the EMS scan, above a threshold of 3,000

Metabolomics for screening cold tolerance in Arabidopsis thaliana 2673

counts per second (cps), throughout the chromatographicrun. Dynamic background subtraction was activated to au-tomatically acquire EPI spectra of co-eluting compounds.EPI spectra were collected using a collision energy of 25 Vwith a collision energy spread of +/−10 V.

The Synapt G2 HD instrument was operated in negativeESI mode; parameter settings used for the measurementwere as follows: capillary voltage (−1,500 V), cone voltage(−30 V), source temperature (120 °C), desolvation temper-ature (350 °C). Nitrogen was used as desolvation and conegas at flow rate 800 and 10 Lh−1, respectively. To obtainaccurate mass MS/MS spectra of the pre-selected markercompounds, in-cell collision-induced fragmentation wasemployed with the use of linear collision energy gradientfrom 10 to 40 V. Precursor ions and their isotope signalswere filtered with a quadrupole isolation window of 4 Da.Both full MS and the fragmentation mass spectra wereacquired at a rate 5 spectras−1 in the range m/z 50−1,000.The mass resolving power of the instrument was approxi-mately 20,000 full width at half maximum, at m/z 200.

The in-batch sequence of the samples was randomized toavoid any possible time-dependent changes in UHPLC–MSchromatographic fingerprints. Two repeated analyses ofeach sample were performed. Within each sequence, blanksamples (pure methanol) were analyzed.

Direct analysis in real time–mass spectrometry (DART–MS)

The DART–TOFMS system used for measurements con-sisted of a DART ion source (IonSense, Danvers, MA,USA), an AccuTOF LP TOF mass spectrometer (JEOLSAS, Croissy sur Seine, France), and an AutoDART-96autosampler (Leap Technologies, Carrboro, NC, USA).The parameter settings employed in positive and negativeionization modes were as follows: ionization gas (helium)flow (3 Lmin−1), gas temperature (220 °C), discharge needlevoltage (−3,000 V), perforated electrode (±150 V), gridelectrode (±250 V), ion guide voltage (±800 V). Massspectra were recorded in the m/z range 50–1000 at anacquisition rate of 2 spectra s−1. Samples were automaticallytransferred in the front of the DART gun exit on a glass rodof Dip–it sampler (IonSense, Danvers, MA, USA) and des-orbed from its surface in a hot helium gas for 10 s. Tworepeated analyses were performed for each sample. At theend of analysis, mass spectra of PEG (dissolved in metha-nol, 200 μgmL−1) were acquired to enable mass driftcompensation.

Data processing and statistical analysis

For processing of UHPLC–QqQ/LIT MS data (i.e. datamining, alignment, normalization, and PCA), MarkerViewsoftware (version 1.2.1, AB SCIEX, Concord, Ontario,

Canada), was used. Data mining was performed with theuse of an automated algorithm in the retention time (RT)range 0.3–11.0 min and m/z range 50–1000. Extractionintensity threshold was set to 10,000 cps, thus only peakswith signal intensity equal or higher than this value wereextracted. The RT and m/z alignment of the respective peakswas carried out using RT and m/z tolerances of 0.2 min and0.5 Da, respectively. Two separate positive and negativeionization data matrices, comprising lists of peaks for eachsample characterized by RT, m/z, intensity and charge state,were obtained by this way. Prior to PCA, the data were pre-processed using the Pareto scaling that uses square root ofthe standard deviation as the scaling factor. In addition toPCA, t-test was used to study the changes in peak intensitiesof metabolites between sample groups of interest.

Mass spectral data, recorded during the DART ionizationof examined samples, were processed with the Mass Centersoftware (version 1.3, JEOL, Tokyo, Japan). The mass spec-tra were background subtracted and mass drift compensa-tion was carried out. Positive and negative ionization datamatrices were obtained, containing the intensities of manu-ally pre-selected ions across the whole sample set. Thecriterion for selection of the ions was their intensity. Onlyions that provided satisfactory response in extracted ionchronograms (i.e. signal-to-noise of 9 or higher and at least3 data points per desorption peak) were selected. The con-stant row sum transformation of the data (intensity of eachion was divided by the sum of intensities of all selected ionsof each sample) was followed by PCA employingSTATISTICA software (version 8.0, StatSoft, Tulsa, OK,USA).

The stepwise LDA of both UHPLC–QqQ/LITMS andDART–TOFMS data was performed with statistiXL soft-ware package (version 1.8, Broadway–Nedlands, Australia).With regard to high dimensionality of the measured data(especially for those obtained by UHPLC–MS technique),the principal components of the data were used as the inputvariables for LDA in order to avoid model overfitting. Thedata obtained within repeated measurements were averagedprior to LDA. The classification results of LDA modelswere presented in terms of recognition and predictionabilities. Recognition ability represents a percentage ofsuccessfully classified samples in the training set, whileprediction ability is a percentage of correctly classifiedsamples in the test set by using the model developedduring the training step. Validation of the LDA modelswas carried out following the leave-one-out cross-validation (LOOCV).

To evaluate the high resolution (HR) MS/MS mass spec-tra obtained within UHPLC–QTOFMS analyses and to es-timate elemental formulae of pre-selected markercompounds, MassLynx software (version 4.1, Waters,Milford, MA, USA) was used.

2674 L. Vaclavik et al.

Results and discussion

Metabolomic fingerprinting

In order to facilitate the acquisition of information-richmetabolomic fingerprints and to detect as many metab-olites as possible, minimal number of sample prepara-tion steps and generic instrumentation parameter settingswere employed. Considering the diversity in metabo-lites’ physico-chemical properties, we have employedboth positive and negative ionization modes by boththe UHPLC–QqQ/LITMS and the DART–TOFMS mea-surement to enlarge the coverage of the Arabidopsismetabolome.

For UHPLC–MS chromatographic separation, an analyt-ical column with reversed stationary phase containing em-bedded polar groups was used. Compared to conventionalreversed stationary phases (such as C18 or C8), this columnprovides improved retention for highly polar components ofthe plant metabolome. Additionally, usage of stationaryphase with sub-2 μm particles significantly increased thechromatographic resolution and the separation efficiency,thus enabling relatively rapid fingerprinting (injection-to-injection run time 13 min) on a short analytical column withlength of 50 mm. Under optimized conditions, the typicalbase widths of chromatographic peaks were around 10 s.Good reproducibility of both m/z and retention times (RT)represent crucial factors affecting the outcome of the datamining process and overall quality of the LC–MS metabo-lomic data. While the fluctuation of ion m/z values was

minimized by regular instrument calibration, the RT stabilitywas examined in more detail. For this purpose, the RT ofthree peaks (RT 0.6 min, RT 5.2 min, and 9.5 min) observedin positive ionization mode were monitored while measur-ing a sequence of 34 sample injections. Based on these data,typical peak RT variability was found to be below 4 %(relative standard deviation, RSD) that can be consideredas an acceptable value.

The complex UHPLC–QqQ/LITMS total ion current (TIC)chromatograms of extract from NAC Te plants recorded inpositive and negative ionization modes are shown in Fig. 1aand b, respectively. The linear ion trap (LIT) mass analyzer,scanning at a rate of 10,000 Das−1 enabled rapid spectraacquisition with good metabolites peak shapes (> 12 pointsper peak) in full scan MS mode. The utilization of the ion trapdynamic fill time function diminished the risk of the LITdetector saturation by the ions of abundant compounds, whileimproving the sensitivity for the metabolites that occurred atrelatively lower concentrations. Besides recording of full scanmass spectra, simultaneous automatic acquisition of MS/MSfragmentation mass spectra was performed for the ions withintensities higher than the threshold value to provide addition-al spectral information on eluting compounds.

Temperature of ionization represents the key factor affectingthe outcome of the DART–MS-based measurements [30].Therefore, the effect of gas temperature on the mass spectrawas studied in the range of 100–450 °C to optimize theDART–TOFMS system settings. We found that the numberand intensities of detected ions were increased with the rise ofthe temperature up to 220 °C, probably due to the acceleration

Fig. 1 Metabolomic fingerprints of NAC Tenela (Te) extract analyzed by UHPLC–QqQ/LIT MS recorded in positive (A) and in negative (B) ESImodes and by DART–TOFMS mass spectrum recorded in positive (C) and in negative (D) modes

Metabolomics for screening cold tolerance in Arabidopsis thaliana 2675

of liquid sample drying and enhancement of the thermo-desorption process. Similar to the results reported by Harriset al. [24], a significant drop in the signal intensity was ob-served at the temperature settings higher than 250 °C, as aconsequence of thermal degradation of sample componentsand too quick desorption. Therefore, the temperature of ioni-zation was kept at 220 °C while measuring metabolomicfingerprints of the investigated plant samples. Under this set-ting, well-defined desorption peaks could be obtained duringthe exposure of sample to the ionization gas for 10 s. TheDART–TOFMSmass spectra obtained by the analysis of NACTe extract are shown in Fig. 1c, d. No ions withm/z higher than620 (positive mode) and 820 (negative mode) were detected.We further evaluated the repeatability of absolute and relativeion intensities by repeated measurement of fingerprints of thissample (n=12).While relatively high fluctuation was observedfor absolute ion intensities (RSDs up to 55 %), the relativesignal intensities showed acceptable repeatability (RSDsranged from 5.3 to 23.2 %). It is worth to notice that therepeatability depended on the abundance of the respectiveion (the higher the abundance, the better the repeatability).

Data processing and statistical analysis

The data mining and aligning procedures of UHPLC–QqQ/LITMS records yielded the data that were organized into amatrix (m×n) form containing the responses for the n peaks(variables) in each of the m records. For positive and nega-tive ESI mode, data matrices of (180×2,416) and (180×797), respectively, were obtained. Because of large numberof peaks, there was a need to reduce the dimensionality ofthe data to avoid signal redundancy. Dimensionality

reduction was achieved by the exclusion of peaks represent-ing isotopic ions, which were automatically assigned by thesoftware during the data mining based on isotope spacing.The reduced dimensions of the positive and the negativemode matrices were (180×659) and (180×268), respective-ly, and contained only the peaks of monoisotopic ions, i.e.,signals with the lowest m/z value within an isotope pattern.Large differences in the abundance of input variables(peaks) can significantly affect the data variance determinedwithin the treatment by the multivariate analysis. To modifythe weights of respective variables, the Pareto scaling ofthe data was performed [31]. This type of scaling enablesreduction but does not completely eliminate the input varia-bles abundance differences. It provides good results whenapplied to LC–MS data, because it takes into account thatlarger peaks can be more reliable, but all variables are equiv-alent. The dimensions of DART–TOFMS data matrices were(204×35) and (204×27) respectively for positive and negativemodes. The variables collection was performed through in-spection of the mass spectral records and manual selection ofthe ions. Because of poor repeatability of the absolute ionintensities as discussed above, transformation and normaliza-tion via constant row sum was performed prior to multivariateanalysis.

Fig. 2 PCA score plots of positive ionization data measured by UHPLC–QqQ/LIT MS (A) and DART–TOFMS (B) methods ( non-acclimated,acclimated, and sub-zero-temperature-treated sample sets; data were obtained by two repeated analyses of each sample)

Fig. 3 PCA score plots of negative ionization data measured byUHPLC–QqQ/LIT MS for NAC (A), ACC (B), and ST-treated (C)sample set, and by DART–TOFMS for NAC (D), ACC (E), and ST-treated (F) sample sets ( accessions with high tolerance, accessionswith intermediate tolerance, accessions with low tolerance; data wereobtained by two repeated analyses of each sample)

�

2676 L. Vaclavik et al.

Metabolomics for screening cold tolerance in Arabidopsis thaliana 2677

Principal component analysis (PCA)

The PCA represents a highly useful and widely employed toolfor interpretation of complex data sets generated by severalmodern instruments including mass spectrometry. It allowsdimensionality reduction and visualization of the informationpresent in the original data in the form of a few principalcomponents (PCs) while retaining the maximum possible var-iability [32]. We have employed PCA to explore differ-ences in metabolome of A. thaliana accessions withvarying cold tolerance in order to investigate its capac-ity for discriminating cold tolerance in NAC, ACC, andST-treated plant samples.

The initial examination of PCA scores plots obtainedfor the metabolomic data generated by both instrumentaltechniques has explicitly shown that the positive ioniza-tion records can be used to monitor the impact of coldacclimation and sub-zero-temperature-induced changesin the A. thaliana accessions (Fig. 2). As can be seenin Fig. 2a, three well-separated clusters representingNAC, ACC, and ST-treated sample sets were presentin the PCA scores plot of UHPLC–QqQ/LITMS data.However, PCA scores plot of DART–TOFMS datashown in Fig. 2b do not show such an apparent sepa-ration among the samples. Moreover, PCA scores ofcold-acclimated and ST-treated samples from DART–TOFMS data are almost overlapped. To further investi-gate the possible correlation between the cold toleranceof the accessions and metabolome changes measured byPCA, we have employed t test on the data obtained fortwo highly contrasting accessions such as Te (cold-tol-erant) and Co (cold-sensitive). The t test method enabledselection of variables which exhibited statistically sig-nificant change in their abundance. In the ACC and theST-treated samples of the highly cold-tolerant accessionTe, 24 % and 21 % of 659 detected peaks, respectively,showed elevated signal levels as compared with NACplants. While in the highly cold-sensitive accession Co,the abundance of peaks in ACC and ST-treated sampleswas around 20 % and 19 % as compared with its NACcounterparts. There was a high overlap of about 95 %between peaks of accumulated metabolites selected inTe and Co. Only small number of peaks (approximately2 % of detected peaks) showed decrease after acclima-tion and ST treatment in both accessions discussedabove. Therefore, although cold acclimation and ST treat-ments have significant impact on abundance of the metabo-lites, the t test shows that these changes observed underpositive ESI cannot be linked to the cold tolerance of studiedA. thaliana accessions. Due to the lack of chromatographicseparation prior to ionization and detection of sample compo-nents and due to lower sensitivity, the positive-mode DART–TOFMS metabolomic fingerprints provided significantly

lower discrimination potential that resulted in the overlap ofacclimated and sub-zero-temperature-treated sample sets inPCA scores plot (Fig. 2b).

The PCA score plots of metabolomic data recorded innegative ionization mode (Fig. 3) provide much better coldtolerance-dependent separation of samples as compared withthose obtained in the opposite polarity (Fig. 2). Therefore,separate PCAs were performed on each of the three (NAC,ACC, and ST-treated) sample sets to provide better represen-tation of the data and to simplify the interpretation. The scoresplot obtained for UHPLC–QqQ/LITMS data of the respectivesample sets are presented in Fig. 3a–c. It is obvious from thePCA scores of metabolites that NAC accessions are almostseparated into two clusters across the PC1 axis (Fig. 3a). Thefirst cluster is being formed by the accessions with very highcold tolerance and intermediate cold tolerance, while thesecond cluster contained cold-sensitive accessions. Thesedifferences in the PCA scores might be linked to thegenetic variations between the cold-sensitive and thecold-tolerant accessions. Numerous metabolites are in-duced during the cold acclimation that protects the leaftissues during freezing stress of harsh winter. Consequently,the differences in PCA scores of differentially cold-tolerantaccessions in ACC (Fig. 3b) and in ST-treated plants (Fig. 3c)are more pronounced and can separate the accessions intocold-tolerant, intermediate cold-tolerant, and cold-sensitivegroups. The resolution of separation among the three groupswas better in ST-treated accessions because freezing treat-ments probably re-programmed the cold-acclimated metabo-lites to protect the tissues during freezing stress [1, 2]. Inconsistence with the positive mode score plot (Fig. 2b), neg-ative DART–TOFMS data (Fig. 3 d–f) further provided lessapparent separation of sample groups as compared with that ofUHPLC–QqQ/LITMS data.

In order to further investigate the abundant markermetabolites in the cold-tolerant, intermediate cold-tolerant,and cold-sensitive accessions, PCA loadings plot based onUHPLC–QqQ/LITMS fingerprints of acclimated sample setwas reviewed. We have selected four signals with the high-est PC1 and/or PC2 absolute loadings and created the profileplots displaying marker’s abundance across the sample set.As documented in Fig. 4, four markers characteristic for therespective sample groups differing by their cold tolerancewere obtained based on ACC samples set. An additionalconfidence on unique occurrence of the markers in thetested accessions was gained by an automatic extractionof corresponding peaks from the raw data and inspec-tion of blank samples records. A tentative identificationof these unknown metabolites based on accurate MS/MSmeasurements is discussed below. It should be notedthat, while in ST-treated sample set all of the abovemarkers remained characteristic for the respective acces-sion types, in NAC samples, only metabolites with m/z

2678 L. Vaclavik et al.

372.1 (RT 0.5 min) and m/z 376.1 (RT 0.5 min) weredetected. No statistically significant differences in markers’

abundance were observed between accessions with similarcold tolerance.

}

Fig. 4 Profile plots of characteristic markers in ACC sample set ob-served in UHPLC–QqQ/LIT MS records (the y-axis displays abundanceof markers in respective experimental samples). A m/z 372.1 (RT0.5 min), B m/z 376.1 (RT 0.5 min), C m/z 577.2 (RT 2.4 min), and

D m/z 739.2 (RT 1.8 min) ( accessions with high tolerance, acces-sions with intermediate tolerance, accessions with low tolerance; datawere obtained by two repeated analyses of each sample, randomsequence)

Metabolomics for screening cold tolerance in Arabidopsis thaliana 2679

Linear discriminant analysis (LDA)

Finally, we have developed the stepwise forward LDA pre-dictive models using the negative ionization mode data ofthe NAC, ACC, and ST samples to classify cold tolerance ofdifferentially cold-tolerant A. thaliana accessions. To avoidthe LDA model overfitting due to high number of variablespresent in the raw data, we have used the PCs obtainedwithin the PCA as the inputs for LDA. The optimal numberof component scores was obtained using LOOCVobservingthe rate of misclassification (from prediction ability). Theminimum (optimal) value for LDA of UHPLC–QqQ/LITMS and DART–TOFMS data was obtained when tenand seven PCs, respectively, were used. Additionally, thepossibility of LDA model overfitting was also checked byusing the criterion (n−g)/3 (n is the number of objects and grepresents the number of classes), which should not belower than the number of used variables. The classificationrates of LDA models are presented in Table 1. Correct classi-fication (recognition and prediction abilities of 100 %) of allexamined accessions into highly cold-tolerant, intermediatecold-tolerant, and cold-sensitive groups in both training andvalidation step was achieved for the ACC and ST sample setsbased on the UHPLC–QqQ/LITMS data. In line with the

results of PCA, significantly lower classification rates wereobtained for the rest of LDA models.

Considering these results, metabolomics-based finger-print measured by UHPLC–QqQ/LITMS integrated withchemometric tools, such as PCA and LDA, can be poten-tially applicable to predict the cold tolerance of cold-acclimated plants, i.e., without the need to expose them tosub-zero-temperatures prior to measurement.

Tentative identification of marker metabolites

The identification of characteristic biomarkers representsthe most complicated and laborious step of the metab-olomic workflow. The use of a high-resolution instru-ment is typically needed, enabling to acquire accuratemass MS/MS spectra, on the basis of which a correction elemental formula estimation can be performed. TheUHPLC–QqQ/LITMS data, collected at unit resolution,do not provide enough information to accomplish theidentification; therefore, further analyses of selectedplant extracts were performed using an UHPLC–QTOFMS mass spectrometer.

The UHPLC–QTOFMS records were searched for themarker compounds peaks using the information on RT and

Table 1 An overviewof the LDA modelsobtained from the PCsof mass spectral data

Sample set UHPLC–QqQ/LITMS data DART–TOFMS data

No. of inputvariables

Recognitionability (%)

Predictionability (%)

No. of inputvariables

Recognitionability (%)

Predictionability (%)

Non-acclimated(NAC) plants

10 76.6 63.3 7 66.6 60.0

Cold-acclimated(ACC) plants

10 100.0 100.0 7 76.6 66.6

Sub-zero-temperature(ST)-treated plants

10 100.0 100.0 7 80.0 70.0

Table 2 An overview of the marker metabolites detected in A. thaliana accessions

m/z(HRMS)

RT(min)

Ion m/z (HRMS) fragment ions (rel. ab. %) Ion elementalformula

Masserror(ppm)

Tentativeidentification

Marker

372.0419 0.5 [M–H]– 274.9891 (5); 259.0122 (7); 96.9582 (27);95.9512 (12)

C11H18NO9S2 +0.41 Gluconapin S

376.0364 0.5 [M–H]– 341.1089 (25); 274.9888 (7); 259.0105 (5);96.9585 (19); 95.9510 (9)

C15H14N5OS3 +2.39 Unknown I

577.1564 2.4 [M–H]– 431.0970 (81); 430.0893 (29); 285.0399 (58);283.0239 (12)

C27H29O14 +2.11 Kaempferol-3,7-O-dirhamnoside

T

739.2094 2.5 [M–H]– 593.1509 (100); 430.0896 (10); 284.0314 (3) C33H39O19 +1.89 Kaempferol-3-neohesperidoside-7-rhamnoside

T

S characteristic for accessions with low tolerance, I characteristic for accessions with intermediate tolerance, T characteristic for accessions withhigh tolerance

2680 L. Vaclavik et al.

m/z values obtained from the primary UHPLC–QqQ/LITMSdata. All the peaks of interest were successfully found inHRMS chromatograms at almost identical RTs. Also, theMS/MS spectra acquired both under low and high mass re-solving power settings were in good agreement, but differencesin fragmentation patterns were observed only in terms of ionrelative intensities. The overview of pre-selected marker peaksis provided in Table 2. The elemental formula estimation wascarried out with the use of: (1) the accurate full MS mass ofparent ion, (2) the accurate MS/MS masses of fragment ions,and (3) the isotopic profiles of both parent fragment ions. For

calculations, C (n≤50), H (n≤100), N (n≤10), O (n≤20), P(n≤15), and S (n≤5) atoms were considered. The candidateelemental formulae corresponding to the electroneutral mole-cules were subsequently searched in several online databases[MassBank (http://www.massbank.jp/?lang=en), Metlin(http://metlin.scripps.edu/metabo_search_alt2.php), andPubChem (http://pubchem.ncbi.nlm.nih.gov/)]. Of the numer-ous compound hits obtained, only the candidates presence ofwhich was probable in A. thaliana plants were further exam-ined by comparing the experimental fragmentation mass spec-tra (see Fig. 5) with those provided either in the databases

Fig. 5 Positive ESI mode HR MS/MS spectra of characteristic markers (filled upright triangle parent ion). A m/z 372.1 (RT 0.5 min), B m/z 376.1(RT 0.5 min), C m/z 577.2 (RT 2.4 min), D m/z 739.2 (RT 1.8 min)

Metabolomics for screening cold tolerance in Arabidopsis thaliana 2681

(MassBank and Metlin) and/or scientific literature. Using thisprocedure, three of four marker compounds were tentativelyidentified (see Table 2). It should be noted that, for the un-equivocal confirmation of the predicted compound’s identity,the examination of its standard is required. As this was notperformed, the reported identification within this study can beconsidered as tentative only. Two of the marker compoundscharacteristic for highly tolerant Arabidopsis accessions wereidentified as flavon-3-ol glycosides (kaempferol-3,7-O-dir-hamnoside and kaempferol-3-neohesperidoside-7-rhamno-side). In the case of these metabolites, a good match wasobtained during the comparison of experimental MS/MS spec-tra with those provided in MassBank. Additionally, an identi-cal fragmentation pattern was published in the literatures [33,34]. Several studies reported on the relation between this typeof compounds and low temperature tolerance in Arabidopsis[35]. It might be linked to osmotic control through the accu-mulation in vacuoles [36]. The marker with an accurate m/zvalue 372.0419 was tentatively identified as gluconapin(characteristic for sensitive accessions). The experimen-tal fragmentation pattern obtained for the above parention was in good agreement with information availablein the literature [37, 38].

Conclusions

Comprehensive, non-targeted UHPLC–MS metabolomicfingerprinting coupled to chemometric methods was dem-onstrated to be an effective tool in screening of plants coldtolerance. With the use of low-resolution MS instrumentutilizing a QqQ/LIT mass analyzer, information-rich, dis-criminating records were collected for differentially cold-tolerant natural accessions of A. thaliana. We showed thatLDA classification model constructed with the use of neg-ative ESI ionization UHPLC–MS metabolomic data of thecold-acclimated (ACC) and sub-zero-temperature (ST)-trea-ted plants constitutes a reliable discrimination of the acces-sions exhibiting high cold-tolerant (Te, Rsch, and Col),intermediate cold tolerant (Nd and Ler), and cold-sensitive(Co, C24, and Cvi). Ambient mass spectrometry-basedapproach employing DART ionization technique didnot provide sufficient information for sample discrimi-nation. This reveals that UHPLC–MS-based methodscan be potentially used for the discrimination of coldtolerance for the cold-acclimated plants without treatingthem at sub-zero-temperatures (ST) that is required bycurrently available methods (i.e., electrolyte leakage andchlorophyll fluorescence). Finally, we attempted to ten-tatively identify the marker compounds pre-selected dur-ing PCA based on the single MS and MS/MS massspectra obtained within the UHPLC–QqTOFMS analy-sis. Using a combination of experimental data and

information available in chemical, metabolomic, andmass spectral databases, the marker metabolites charac-teristic for contrast accession were tentatively identified.Gluconapin was characteristic for cold sensitive acces-sions, while kaempferol-3,7-O-dirhamnoside and kaemp-ferol-3,7-O-dirhamnoside were characteristic for highlycold tolerant accessions.

Acknowledgments We thank Prof. Dr. Arnd G. Heyer of Universityof Stuttgart (Germany), for providing seeds of the Arabidopsis thalianaaccessions and for critical reading of the manuscript. We thank AnitaZupcanova for assistance with sample preparation. The financial sup-port from the Ministry of Education, Youth and Sports (projectsOC10060 and MSM6046137305) to LV and JH is gratefully acknowl-edged. KBM and AM acknowledges Ministry of Education, Youth andSports project OC08055 and CzechGlobe–Centre project CZ.1.05/1.1.00/02.0073.

References

1. Guy C, Kaplan F, Kopka J, Selbig J, Hincha DK (2008) PhysiolPlant 132:220–235

2. Shulaev V, Cortes D, Miller G, Mittler R (2008) Physiol Plant132:199–208

3. Moellering ER, Bagyalakshmi M, Benning C (2010) Science330:226–228

4. Mishra A, Mishra KB, Höermiller II, Heyer AG, Nedbal L (2011)Plant Signal Behav 6:301–310

5. Hannah MA, Weise D, Freund S, Fiehn O, Heyer AG, Hincha DK(2006) Plant Physiol 142:98–112

6. Zhen Y, Ungerer MC (2008) New Phytol 177:419–4277. Koorneef M, Alonso-Blanco C, Vreugdenhil D (2004) Annu Ev

Plant Biol 55:141–1728. Mitchell-Olds T, Schmitt J (2006) Nature 441:947–9529. Novillo F, Alonso JM, Salinas JR (2004) Proc Natl Acad Sci USA

101:3985–399010. Rohde P, Hincha DK, Heyer AG (2004) Plant J 38:790–79911. Murray MB, Cape JN, Fowler D (1989) New Phytol 113:307–31112. Steponkus PL (1990) Environmental injury to plants. Academic

Press, San Diego13. Azzarello E, Mugnai S, Pandolfi C, Masi E, Marone E, Manusco S

(2009) Trees 23:159–16714. Yamori W, Noguchi K, Terashima I (2005) Plant Cell Environ

28:536–54715. Rumich-Bayer S, Krause GH (1986) Photosynth Res 8:161–17416. Baker NR, Rosenqvist E (2004) J Exp Bot 55:1607–162117. Ehlert B, Hincha DK (2008) Plant Methods. doi:10.1186/1746-

4811-4-1218. Kaddurah-Daouk R, Krishnan KR (2008) Neuropsychopharmacol

34:173–18619. Wishart DS (2008) Drugs R D 9:307–32220. Bundy JG, Davey MP, Viant MR (2009) Metabolomics 5:3–2121. Dunn WB, Ellis DL (2005) Trends Anal Chem 24:285–29422. Castrillo JI, Hayes A, Mohammed S, Gaskell SJ, Oliver SG (2003)

Phytochem 62:929–93723. Chen H, Pan Z, Talaty N, Raftery D, Cooks RG (2006) Rapid

Commun Mass Spectrom 20:1577–158424. Zhou M, McDonald JF, Fernández FM (2011) J Am Soc Mass

Spectrom 21:68–7525. Katajamaa M, Oresic M (2007) J Chromatogr B 1158:318–328

2682 L. Vaclavik et al.

26. Kaplan F, Kopka J, Haskell DW, Zhao W, Schiller KC, Gatzke N,Sung DY, Guy CL (2004) Plant Physiol 136:4159–4168

27. Cook D, Fowler S, Fiehn O, Thomashow MF (2004) Proc NatAcad Sci USA 101:15243–15248

28. Davey MP, Woodward FI, Quick WP (2009) Metabolomics 5:138–149

29. Harris GA, Fernandez FM (2009) Anal Chem 81:322–32930. van der Berg RA, Hoefsloot HCJ, Westerhuis JA, Smilde AK, van

der Werf MJ (2006) BMC Genomics 7:142–15631. Ivosev G, Burton L, Bonner R (2008) Anal Chem 80:4933–494432. Beruetta LA, Alonso-Salces RM, Hérberger K (2007) J

Chromatogr A 1158:196–214

33. Falé PL, Amaral F, Amorim Madeira PJ, Sousa Silva M,Florencio MH, Frazao FN, Serralheiro MLM (2008) FoodChem Toxicol 50:2656–2662

34. Ferreres F, Perereira DM, Valentao P, Andrade PB (2010) RapidCommun Mass Spectrom 24:801–806

35. Korn M, Peterek S, Mock HP, Heyer AG, Hincha DK (2008) PlantCell Environ 31:813–827

36. Chalker-Scott L (1999) Photochem Photobiol 70:1–937. Song L, Morrison JJ, Botting NP, Thornalley PJ (2005) Anal

Biochem 347:234–24338. Rochfort SJ, Trenerry VC, Imsic M, Panozzo J, Jones R (2008)

Phytochem 69:1671–1679

Metabolomics for screening cold tolerance in Arabidopsis thaliana 2683