Introduction ============ As a result of global warming, extreme weather conditions, such as severe drought, are becoming more frequent in many regions of the world ([@B37]). Drought is among the top three stresses limiting plant growth and crop productivity ([@B21]). In most cases, a drought-tolerant genotype is considered one that can maintain relatively high biomass accumulation under drought conditions, and in turn, increase grain yields ([@B15]). Thus, the ability to detect drought tolerance early on during development can assist in the selection of high-yielding individuals for the production of more productive plants. In this regard, plants respond to water stress with changes in their morphological, physiological, and metabolic activities in order to adapt to water stress ([@B16]). Many of the responses of plants to environmental stresses are reflected by changes in the transcriptomes or metabolomes ([@B32]; [@B8]). The metabolome can be defined as the ensemble of metabolites that make up a biological sample, which includes the secondary metabolites in plants ([@B41]). Metabolites perform a range of important functions, such as energy storage, signal transduction, and the production of macromolecules ([@B7]). As metabolites are usually the first reactionary components in the metabolomics process, they are good indicators of biological activity. Currently, there are about two dozen techniques for metabolic profiling, including nuclear magnetic resonance spectroscopy (NMR) ([@B43]), gas chromatography mass spectrometry (GC-MS) ([@B29]), capillary electrophoresis-MS (CE-MS) ([@B18]), high-performance liquid chromatography (HPLC)-MS ([@B14]), liquid chromatography mass spectrometry (LC-MS) ([@B33]), and Fourier transform-MS (FT-MS) ([@B30]). The high sensitivity, reproducibility, and dynamic range make FT-MS a powerful tool for metabolite analysis. There are two principal steps in the interpretation of metabolomics data ([@B12]): one that includes the collection of a large amount of data, and another that involves the statistical analysis of the data. The first step requires that the samples to be compared should be of a similar genetic background, and that the experimental conditions should be homogeneous. This is to ensure that a clear signal or pattern can be obtained. For this reason, we usually use accessions of the same species. Since *Arabidopsis* is one of the model species with the largest number of genetic resources, the accessions used in our study were from three species of the genus: *Arabidopsis thaliana, Arabidopsis halleri*, and *Arabidopsis lyrata*. *A. thaliana* is a typical annual plant which is mainly distributed in southwestern Europe ([@B38]), whereas *A. halleri* and *A. lyrata* are perennial plant species that prefer nutrient-poor sites and grow at lower elevations in comparison with *A. thaliana*. The two latter species are widely distributed in northern, central and eastern Europe, with *A. lyrata* in particular being a common inhabitant of North America ([@B29]). These three species, however, differ in their photosynthetic characteristics ([@B29]). There are approximately 3000 *A. thaliana* accessions that have been fully sequenced, including approximately 300 that are known as natural *A. thaliana* derivatives and are distributed throughout the world ([@B6]). In the present study, a metabolite fingerprinting method based on FT-MS was used to detect differences in the metabolomes of *A. thaliana* accessions grown under well-watered (control) and water-stressed conditions. The aim of this study was to identify an integrated picture of water stress response by using metabolite profiling of *A. thaliana* plants, and investigate the potential of FT-MS to be applied in metabolic fingerprinting studies. Materials and Methods {#s1} ===================== Plant Material and Stress Treatment ----------------------------------- In total, 60 *A. thaliana* accessions were used in this study. The seeds of the accessions were obtained from the Nordborg Crop Centre at Ås, Norway. The seeds were stratified in silica-gel for 4 days to break dormancy at 4°C prior to sowing, which was performed on an automatic greenhouse cabinet (Gavita P-250, Sistec, Czech Republic) under a regime of 14 h light/10 h dark, with a temperature cycle of 25/15°C during the light and dark periods. Seeds were watered three times a week during seedling establishment with 1.3 mL sterile water. One day after sowing, the seedlings were exposed to stress conditions by withholding water until the end of the growing period (7--10 days). The water-stress treatment was performed by stopping watering and then allowing the remaining water in the trays to evaporate ([@B44]). The experimental design included control (well-watered) and water-stressed conditions in this study. All plants were cultivated for 7 days, following which they were harvested in the morning (between 9 and 11 a.m.) when the stomata were completely open ([@B23]). Six individual plants were pooled as one biological replicate to create three biological replicates. Each biological replicate was then divided into two sub-replicates, which were used for metabolomics and transcriptome analyses. The experiment was conducted from November to December in 2013 in the laboratory of Prof. Taina Laajala (University of Helsinki) under a 14 h/10 h light/dark regime at 22°C, and plants were grown using a phytotron greenhouse. The experiments were performed in soil in a hydroponic system. The hydroponic system consisted of glass beakers (300 mm × 220 mm) filled with 1.4 L of aerated standard nutrient solution (Hakaphos-T-plus, The Netherlands) at pH 5.5. The plants were grown in a nutrient solution without any mineral or vitamin addition and no addition of organic matter. The mineral and vitamin contents of the nutrient solution were adapted from the protocol of [@B4]. One day after sowing, the nutrient solution was modified according to [@B4] and was replaced on a weekly basis to avoid nutrient deficiency. Metabolite Extraction --------------------- Plants were rapidly frozen in liquid nitrogen after harvesting and placed in liquid nitrogen immediately after harvest. All tissue samples were freeze-dried for 24 h using a Scanvac CoolSafe TU-135 freeze dryer (Labogene, Denmark) and then lyophilized for a further 24 h in a freeze-drying apparatus (Martin Christ, Göttingen, Germany). Metabolite extraction was performed on the lyophilized samples as follows. Fully ground tissues were collected and immediately placed into 3.8 mL tubes that had been filled with 500 μL of cold chloroform/methanol/water (1:3:1, v/v/v) and agitated for 10 min in a bead mill (Precellys 24, Bertin Technologies, Montigny-le-Bretonneux, France). After centrifugation (13,000 × *g*, 10 min, 4°C), a 400-μL aliquot of each supernatant was transferred into a new tube, and 1.6 mL of water was added to the remaining liquid fraction in order to de-salt the extracts. Subsequently, a 50-μL aliquot of each supernatant was transferred into a new tube, and 1 mL of chloroform was added. Following centrifugation (13,000 × *g*, 5 min), an aliquot (400 μL) of the upper layer was transferred into a new tube for dryness and stored at -80°C until the time of NMR analysis. ^1^H Nuclear Magnetic Resonance Analysis ---------------------------------------- Prior to the NMR analysis, the extracts were allowed to thaw at room temperature and were vortexed vigorously to ensure thorough mixing and a homogeneous solution. Spectra were acquired using a Bruker AVANCE 600 MHz NMR spectrometer (Bruker BioSpin GmbH, Germany) operating at a proton frequency of 600.13 MHz. All one-dimensional (1D) ^1^H NMR spectra were recorded at 300 K using a double-tuned probe (Bruker BioSpin GmbH). For each spectrum, 32 k data points were collected into 64 transients using a spectral width of 20 ppm with 2 s of acquisition time per scan. For each spectrum, an automated phase correction was performed using an adaptive sine-bell weighting function ([@B5]). All spectra were referenced to the internal standard, and the baseline was corrected manually ([@B31]). Using TopSpin 3.0 software (Bruker Biospin), the free induction decays were zero-filled to 64 k data points and multiplied by an exponential line broadening factor of 1 Hz. Prior to multivariate data analysis, all data sets were checked for peak overlap by correlation analysis. Multivariate Data Analysis -------------------------- Multivariate analysis was carried out using SIMCA-P 13 (Umetrics, Sweden). Prior to principal component analysis (PCA) and partial least squares discriminant analysis (PLS-DA), the NMR data were centered and scaled to the Pareto variance ([@B27]). The Pareto scaling method is a procedure used to normalize NMR data, which preserves the important information in the NMR data matrix ([@B27]). The scaled data matrix is subjected to PCA to generate orthogonal components, and the projections of the samples on these components are presented as score plots ([@B26]). In contrast, the supervised PLS-DA is used to find the components that can be used to separate samples from different groups. The PLS-DA model was generated using the scaled data matrix, and its performance is represented by