Anal Bioanal Chem DOI 10.1007/s00216-015-8783-2

RESEARCH PAPER

Metabolic fingerprinting of Lactobacillus paracasei: a multi-criteria evaluation of methods for extraction of intracellular metabolites Kristina B. Jäpelt 1,2 & Nikoline J. Nielsen 1 & Stefanie Wiese 2 & Jan H. Christensen 1

Received: 30 January 2015 / Revised: 6 May 2015 / Accepted: 13 May 2015 # Springer-Verlag Berlin Heidelberg 2015

Abstract An untargeted multi-criteria approach was used to select the best extraction method among freeze-thawing in methanol (FTM), boiling ethanol (BE) and chloroformmethanol (CM) for gas chromatography mass spectrometry (GC-MS) metabolic fingerprinting of Lactobacillus paracasei subsp. paracasei (CRL-431®). The following results were obtained: (i) coverage and efficiency, measured by the number of features extracted and the sum of feature intensities, showed that FTM extraction resulted in the largest compound coverage with a total number of features 8.9×103 ±0.5×103, while merely 6.6×103 ±0.9×103 and 7.9×103 ±0.8×103 were detected in BE or CM, respectively; (ii) the similarity of extraction methods, measured by common features, demonstrated that FTM yielded the most complementary information to BE and CM; i.e. 17 and 33 % of the features of FTM extracted were unique compared to CM and BE, respectively; and (iii) a clear-cut separation according to extraction method was demonstrated by assessment of the metabolic fingerprints by pixel-based data analysis. Indications of metabolite degradation were observed under the elevated temperature for BE extraction. A superior coverage of FTM together with a high repeatability over nearly the whole range of GC-amenable compounds makes this the extraction method of choice for metabolic fingerprinting of L. paracasei.

* Kristina B. Jäpelt [email protected] 1

Analytical Chemistry Group, Department of Plant and Environmental Sciences, Faculty of Science, University of Copenhagen, Thorvaldsensvej 40, 1871 Frederiksberg, Denmark

2

Chr. Hansen A/S, Innovation & Assays, 2970 Hoersholm, Denmark

Keywords Microbial metabolomics . Pixel-based data analysis . Feature detection and extraction . Gas chromatography . Mass spectrometry

Introduction The beneficial health effects of lactobacilli are known to be strain specific, and it has been suggested that the efficacy of the strains is directly related to the ability to survive in the gastrointestinal tract. Thus, the discovery of strains with increased survival following exposure to gastric acid and bile is of great interest. However, the precise molecular mechanism and the strain-dependent factors involved in acid and bile resistance remain poorly understood [1]. Metabolic fingerprinting, defined as the semi-quantitative analysis of the endo-metabolome, that has emerged with the introduction of the ‘omics’ approach can be seen as powerful tools to discriminate strains according to their phenotype and expand the understanding of bacterial acid and bile resistance. Within metabolic fingerprinting of microorganisms, an unsuitable protocol for metabolite extraction can significantly bias the analysis outcome [2] and the optimal extraction method should cover metabolites with a wide range of properties, ensure that no enzymatic and stress-induced degradation of the cell metabolites is occurring and generate reproducible metabolic fingerprints [3, 4]. Several research groups have studied the effects imposed on the metabolic fingerprints of given microorganisms using an extraction method [2, 3, 5–17]. The selection of the optimal extraction method has usually been based on one or a few assessment criteria with the most frequently used being recovery of a sub-set of target metabolites [3, 5, 8, 15–17], coverage as measured by the number of features or metabolites extracted [6, 9, 10, 17], and efficiency and accuracy as measured by semi-quantitative or quantitative

K.B. Jäpelt et al.

comparison of either target metabolites or the full metabolome covered by the analytical platform [2, 6, 9, 10, 13]. For a detailed walk-through of assessment criteria, we refer to Canelas et al. [15]. The decision on method performance will be determined by the assessment criteria selected. Hence, by focussing on a sub-set of metabolites in a targeted manner, the results may be biased towards compounds with specific physicochemical properties and a part of the endo-metabolome will not be covered. A less targeted chemical analysis and untargeted data processing with multiple assessment criteria may therefore be the solution to reduce biases in metabolic fingerprinting studies. In this study, we will therefore apply a multi-criteria approach for selecting the optimal extraction strategy for metabolic fingerprinting of Lactobacillus paracasei and discuss the strengths and weaknesses of each assessment criteria. We will compare boiling ethanol (BE), chloroform-methanol (CM) and cycles of freeze-thawing in methanol (FTM) as these methods have been frequently applied by others [8, 15]. Oximation silylation gas chromatography mass spectrometry (OS-GC-MS) is used frequently for metabolic fingerprinting as it encompass a wide range of metabolites, such as sugars, sugar monophosphates, organic acids, amino acids and alcohols [18], and is therefore the selected method in this study. However, for the analysis of high molecular weight compounds, such as nucleotides, that are not volatilized even after derivatization, it is necessary to complement the analysis with another analytical technique such as liquid chromatography mass spectrometry (LC-MS) [18–20] or capillary electrophoresis (CE) [21]. Unfortunately, the interferences of the detergent Tween-80 in our nutrient medium, and thus cell extracts, challenge LC-MS analysis [20]. Thus, the analysis is limited to OS-GC-MS for analysis of the semi-volatile fraction of the metabolome. With respect to signal processing and data analysis, methods that can entail the entire sample complexity are required, e.g. feature detection and extraction, curve resolution [22] or pixel-based methods [23]. Feature detection, including peak deconvolution, is by far the most frequently employed for metabolic fingerprinting using software like XCMS [24], MZmine [25], MetAlign [26] and Metab [27]. The studies that have used a more untargeted approach to assess the effects of extraction methods on the metabolic fingerprint have all been based on automated peak detection and mass spectral deconvolution software for the data processing [2, 6, 9, 10, 17]. The pixel-based method is an alternative data processing method that treats the chromatograms as images by comparing sample chromatograms pixel by pixel [28]. However, this method has to the authors’ knowledge rarely been used for untargeted analysis of metabolic fingerprints [23]. In this study, the objective was to use a multi-criteria approach for determining the best extraction method for GC-MS metabolic fingerprinting of L. paracasei subsp. paracasei.

The performance of the three extraction methods, FTM, BE and CM, was assessed using (i) coverage and efficiency measured as the number of detected features and the sum of feature intensities, (ii) the similarity of extraction methods as measured by the common features, i.e. the conjoined parts of the features, and (iii) repeatability and differences of the metabolite fingerprints originating from each extraction method as assessed using the pixel-based data analysis combined with principal component analysis (PCA). The two first criteria are evaluated using an in-house feature detection method based on the combined matched Gaussian filter and twofold differentiation [29] and median measure-based thresholds [30]. The last criterion is evaluated using in-house pixel-based workflow previously used in other contexts [23, 28, 31].

Material and methods Organism and inoculum L. paracasei subsp. paracasei (CRL-431®; Chr. Hansen A/S, Denmark) was used throughout this study. Pre-cultures were carried out by inoculating 200 mL MRS broth (Fluka Analytical, Switzerland) with 2 mL stain stocks at 37 °C. After 16 h, the inoculation step was repeated. The final pre-culture was inoculated at 1 %. Cultivation conditions Cultivations were performed anaerobic in a batch fermenter with a working volume of 2 L (BIOSTAT® Bplus; Sartorius Stedim Biotech S.A., France) equipped with continuous data acquisition via BioPAT® MFCS/Da. The cultivations were carried out at 37 °C, with initial pH adjusted to 6.5, nitrogen flow of 0.5 L/min and a stirring speed of 300 rpm. Optical density at 600 nm (OD600) and pH were monitored online. Sampling took place in the exponential phase at an OD600 at ∼4.5. Sampling of pellets A 10-mL aliquot of the cell culture broth was rapidly sampled and centrifuged at 2000g for 5 min at 4 °C. The supernatant was discarded, and the cells were washed with 0.9 % (w/v) saline solution, centrifuged and stored at −80 °C until extraction. Extraction of metabolites All extraction procedures were adapted from literature sources with minor modifications [7, 15]. Centrifugation was carried out at 5000g for 5 min at −10 °C and mixing of samples was performed using a vortex. Extracts were stored at −80 °C until OS-GC-MS analysis.

Metabolic fingerprinting of Lactobacillus paracasei: a multi-criteria evaluation of methods for extraction...

Chloroform-methanol 0.75 mL of 50 % (v/v) pre-cooled aqueous methanol (−40 °C) and 0.75 mL of pre-cooled chloroform (−40 °C) were added to each pellet. Each sample was extracted 40 min at −40 °C (FP50-HL Refrigerated Circulator; Julabo, Germany). During extraction, the samples were frequently mixed. Each sample was centrifuged; the upper water/ methanol phase was collected, and the lower layer was reextracted for 10 min with 0.75 mL of 50 % (v/v) (−40 °C) aqueous methanol. Freeze-thawing in methanol Each pellet was suspended in 0.75 mL of 50 % (v/v) pre-cooled aqueous methanol (−40 °C). The suspension was frozen on dry ice for 5 min, thawed on ice for 5 min and mixed for 30 s. After three freeze-thaw cycles, the sample was centrifuged, the supernatant was collected, and the pellet was re-extracted using one freeze-thaw cycle. Boiling ethanol Ethanol (96 %; VWR chemicals) was preheated in a thermostat block (mini oven, Supertherm™; Denmark) at 90 °C. 0.75 mL of the ethanol was added to the pellet, and the sample was mixed immediately for 30 s and placed in the thermostat. After 5 min, the tube was cooled on ice followed by centrifugation. The supernatant was collected. Generation of metabolic fingerprints Derivatization prior to chromatographic analysis One hundred fifty microlitres of the extract was transferred to a GC vial with insert, and the solvent evaporated under a constant flow of nitrogen at 40 °C. Twenty microlitres of the oximation reagent was added [20 mg/mL methoxyamine hydrochloride (Supelco Analytical, USA) in 10 mL pyridine (Merck, Germany) containing 0.05 mM myristic-d27-acid (98 at.% D; Isotec, USA)] as the surrogate standard. Each vial was mixed and kept at 40 °C for 90 min followed by a reduction in tray temperature to 8 °C. Twenty microlitres of Nmethyl-N-(trimethylsilyl)-trifluoroacetamide (MSTFA) with 1 % chlorotrimethylsilane (TMCS) (Thermo Scientific, USA) was added, and the sample was incubated for 12 min at 46 °C. All sample preparation was performed using an Agilent 7693A Series Automatic Liquid Sampler. Gas chromatography quadrupole mass spectrometry analysis (data set 1) The GC (Agilent 7890A GC Agilent Technologies, USA) was equipped with a DB-5MS column (30 m×0.25 mm, 0.25 μm film thickness; Agilent Technologies) and a 1-m retention gap (fused silica, deactivated; 1 m× 0.25 mm; Agilent Technologies, USA). One-microlitre aliquots were injected in split mode (split ratio of 5:1). The temperature was programmed as 60 °C for 1 min and 10 °C/min to 325 °C. The quadrupole mass spectrometry (Q-MS) (5975C inert XL MSD system; Agilent Technologies, USA) parameters were set to the following condition: scan

range of m/z 40–400, scan rate of 7.31 scans per s, ion source temperature of 230 °C, quadrupole temperature of 150 °C and electron ionization (EI) voltage of −70 eV. The system was controlled by ChemStation E.02.02 (Agilent technologies, USA), and data were exported to the netCDF file formation. Gas chromatography-time-of-flight mass spectrometry analysis (data set 2) GC parameters were identical to those for the gas chromatography quadrupole mass spectrometry (GC-Q-MS) analysis. The BenchTOF-dx TOF MS detector (ALMSCO International, Germany) was controlled by ProtoTOF software (2010 ALMSCO International). Mass spectra were monitored between m/z 40 and 600 at a scan rate of 4 Hz. The ion source and transfer line temperatures were set to 230 and 290 °C, respectively, and EI was carried out at −70 eV. Data were exported to the netCDF file format. Data processing and multivariate data analysis Computing environment NetCDF was used to retrieve relevant data (signal intensities, m/z values, sample names, sample descriptions) in MatLab (R2013a; The MathWorks, MA, USA). In-house programmed functions were used for feature detection based on Danielsson et al. [29] and Ullsten et al. [30]. The pixel-based processing was performed using inhouse programmed functions along with selected published algorithms, all referenced when mentioned. PCA was performed using a PLS toolbox vs. 7.3.1 (Eigenvector Research Inc., USA). Graphics are illustrated using SigmaPlot 13 (USA). Data The effect of consecutive extractions was initially evaluated by extracting biological replicates (n=2) four times followed by OS-GC-MS metabolic fingerprinting of each extract separately (data set 1), and then, the effect of extraction method was evaluated by extracting biological replicates (n= 5) twice after which the extracts were pooled for each biological replicate (data set 2). Data set 1 is comprised of GC-Q-MS scan data of 24 sample extracts (viz., four consecutive extractions of two biological replicates for the three methods), 3 replicate analyses of the quality control (all extracts were pooled, i.e. ‘the reference sample’), 3 derivatization blanks and 2 extraction blanks for each of the three extraction methods. The derivatization blanks were GC vials containing derivatization agent, thus mapping possible derivatization and instrumental background, while the extraction blanks were empty tubes extracted mapping extraction, derivatization and instrumental background. The analytical sequence was randomized with the reference sample analyzed in-between every 15th sample (‘the reference set’). A time window from 6.7 to 27.4 min (scan no. 60 to 9120) was selected for further data analysis

K.B. Jäpelt et al.

leading to a data set of dimensions: 24 samples×m/z 360× 9060 retention time (the ‘training set’) and 3 samples×m/z 360×9060 retention time (the ‘reference set’). Data set 2 comprised of GC-TOF-MS scan data of 21 sample extracts (viz., five biological replicates for each of the three methods plus two additional injections for each extraction method), 7 replicate analyses of the pooled reference sample, 2 derivatization blanks and 2 extraction blanks. The analytical sequence was randomized. The reference sample was analyzed in-between every sixth sample. A time window from 6.4 to 26.7 min (scan no. 1540 to 6410) was selected for data analysis leading to the data set of dimensions: 21 samples×m/ z 560×4870 retention time (the training set) and 3 samples× m/z 560×4870 retention time (the reference set). Feature detection, similarity and repeatability For data set 2, the m/z axis was binned to the nominal mass with bin borders positioned at m/z 0.5 units. For both data set 1 and data set 2, each extracted ion chromatograms (EICs) were filtered using a MatLab implementation of the combined matched Gaussian filter and twofold differentiation [29] and features detected using median measure-based threshold [30]. Filtering was performed prior to removing initial and tailing sections. Thus, artefacts caused by the filtering functions as well as chemically meaningless areas were excluded. The selection of the width-at-base parameter of the filter function was based on visual inspection of raw data peak widths, because the procedure is robust when the filter width is similar to the raw data peak width. The width-at-base parameter was set to 30 and 20 scans for data set 1 and data set 2, respectively. The peak identification threshold was set as the median intensity of the filtered EIC signal (baseline estimate) plus a conservative 15 times the average median absolute deviation (MAD) of the modulus of the filtered EIC signal (noise estimate). The parameters, filter-width-at-base and number of MAD to define the detection threshold, were selected to balance a high number of detected features with an acceptable similarity of replicate detection. For special instrumental settings, i.e. zero insert below a specific ion count, which results in high numbers of spikes in data, it is recommended to use an average threshold across samples within each EIC. Otherwise, the blank samples will generate relatively low thresholds and relatively high numbers of spikes. Because spikes are transformed into peak-like shape by the filter function, the feature detection may be seriously biased. We used the average threshold to mitigate the deleterious effect of zero insert in raw data. To identify left and right zero crossing as a measure of peak width, steps of one scan point and at a maximum of 30 scan points were used. The window search space used to determine the common features was 30 scans, corresponding to a maximum allowed retention time shift of 4.0 and 7.5 s for data set 1 (acquired with GC-Q-MS) and data set 2 (acquired by GC-TOF-MS), respectively. Finally, the repeatability in the

number of extracted features of each extraction method was calculated by subtracting the accumulated standard deviation of the reference set (reflecting the analytical variation and the feature detection variation), from the overall standard deviation. Pixel-based approach The pixel-based approach described by Christensen et al. [31] and reviewed later by Christensen and Tomasi [28] was used in this study with minor modifications. The success of the pixel-based approach requires that the data analysis relies primarily on relevant chemical and biological information. Thus, the largest part of the variation unrelated to intrinsic chemical difference between samples should be removed, e.g. retention time shifts, altered baseline, noise [28]. All data processing was performed on the base peak intensity (BPI) chromatogram, including only the greatest intensity at each scan, to transfer the data matrix from cube format (sample×m/z×retention time) to matrix format (sample×retention time). The chromatographic section from 8.9 to 9.5 min was distorted by an overloaded peak and eliminated from further analysis. Identification of metabolites in our samples was achieved based on the retention time and mass spectrum of that of own standards. The constant baseline offset of the BPI chromatograms was removed by calculating the first derivative of the chromatogram using the SavitzkyGolay algorithm [32] with width of seven data points and a first-order polynomial fit. Subsequently, retention time shifts were corrected by means of the correlation optimized warping (COW) algorithm [33]. The optimal warping parameters, segment length and slack, were obtained using the procedure suggested by Skovet al. [34]. The search space was [30, 100] for segment length, and [1, 4] for slack. The target chromatogram was set to the chromatogram with the maximum cumulative products of correlation coefficient with other chromatograms. The alignment was performed on the full chromatograms to guarantee the flexibility of the endpoints, while the initial and tailing sections were cut off after alignment to avoid effects on further pre-processing [28]. Each BPI chromatogram was normalized to a unit sum of the 1000 variables of the lowest relative standard deviation in the reference set, thereby increasing the importance of peak regions and eliminating contributions from noisy baseline regions. A multipurpose scaling procedure was used to diminish the relative importance of large peaks and of peaks that vary between analytical and biological replicates, i.e. within one extraction method, and to increase the relative importance of smaller peaks and of peaks that do not vary between biological replicates but might vary between extraction methods. This can be done by weighting each pixel by its inverse standard deviation, and in the case where we want to remove biological variation, weighting can be done by the inverse cumulative standard deviation (SD). The effect of scaling each pixel to the analytical SD was demonstrated by Christensen et al. [35] and,

Metabolic fingerprinting of Lactobacillus paracasei: a multi-criteria evaluation of methods for extraction...

furthermore, by Soleimani et al. [36] and Gallotta et al. [37]. Subsequently, pair-wise PCA was conducted on the processed data matrix, e.g. a pair-wise PCA analysis conducted on fingerprinting for all CM and FTM extracts provides model (BE:CM). Thus, a total of three PCA models is prepared.

Results and discussion Coverage and extraction efficiency assessed by repeated extractions (data set 1) Completeness of extraction is often assumed to occur after one or two extraction cycles [6, 7, 15]. However, the assumption is rarely tested. The coverage and extraction efficiency, measured by the number of features and the sum of feature intensities, respectively, is given in Fig. 1. It should be kept in mind that the total number of features does not correspond to the number of metabolites identified as EI will cause fragmentation of the analytes. However, the number of features is a proportional and, thus, indirect measure of the number of metabolites extracted and indicative of the metabolite coverage. The sum of feature intensities is an indirect and proportional measure of extraction efficiency. Extraction efficiency was highest after the 1st cycle followed by a reduction in features after the 2nd extraction cycle and so forth (Fig. 1). The reductions in the sum of feature intensities from the 1st cycle to the 2nd cycle followed a similar trend. A relatively constant level of features and of the sum of feature intensities has been reached after the 2nd extraction cycle for all three extraction methods. The common features, the similarity of GC-MS metabolic fingerprints, demonstrated that the majority of the features from the 2nd cycle were already detected in the Fig. 1 Number of features and sum of feature intensities in consecutive extractions of the three extraction methods: chloroform-methanol (CM), boiling ethanol (BE) and freezethawing in methanol (FTM). Metabolic fingerprints were acquired by GC-Q-MS and evaluated using an in-house feature detection procedure. The extraction blanks were included to show background features and the summed intensity. The boarders of the boxes indicate the range detected for the biological replicates (n=2)

1st cycle (86.2±1.3 % CM, 94.4±0.8 % BE and 94.1± 0.5 % FTM; average±standard deviation from any combination of 1st cycle and 2nd cycle samples (data not shown)). Hence, few new metabolites were extracted during the 2nd extraction cycle. We conclude that two extraction cycles is sufficient in this case; the majority of the metabolites are extracted (i.e. 83.3–87.0 % of the total number of features with the level in extraction blank subtracted) and the extraction efficiencies obtained are reasonably high (i.e. 86.2–92.6 % of the intensity with the level in extraction blank subtracted). Furthermore, limited information is obtained on new metabolites after the consecutive extraction. Complete extraction could be approached by performing a higher number of repeated extractions; however, we would like to keep the sample handling time at a minimum to reduce the risk of metabolic changes.

Effect of extraction method (data set 2) Number of features, similarity and repeatability Most features (8.9×103 ±0.5×103, n=5) and the highest sum of feature intensities were detected (4.6×107 ±4.1×106, n=7) for the FTM extraction (Fig. 2). The numbers of features were 26± 10 % lower for BE and 12±10 % lower for CM extraction. The total intensity of features followed a similar behaviour. The repeatability evaluated proved FTM to be highly repeatable compared to both BE and CM with deviations of 9, 2 and 7 % for the features (relative standard deviations on the number of features, n=5) and 11, 4 and 8 % for the feature intensities (relative standard deviation sum of feature intensities, n=5) for respectively BE, FTM and CM.

K.B. Jäpelt et al. Fig. 2 Number of features and sum of feature intensities in chloroform-methanol (CM), boiling ethanol (BE) and freezethawing in methanol (FTM) extracts (a total of extraction cycles were performed) illustrated as a box plot with the error bars representing 1 standard deviation of the biological replicates (n=5). The metabolic fingerprints were acquired by GC-TOF-MS and evaluated using an in-house feature detection procedure

The common features extracted using the three methods are illustrated in Table 1, i.e. the features extracted with all the extraction methods or two of the extraction methods combined. The metabolite fingerprints of BE and CM shared most of their features with FTM; i.e. 91 and 94 % of features detected in BE and CM, respectively, were also detected in FTM extracts. Extraction using FTM yielded the most complementary information to CM and BE; i.e. 17 % of the features were unique compared to CM features and 33 % were unique compared to BE. However, overall, the majority of the features detected were detected in all the three extraction methods (5608±429 common features out of all features detected with the extraction method). Repeatability and differences in metabolic fingerprints The data pre-processing of the BPI chromatograms in the pixel-based data analysis approach consisted of baseline removal, retention time alignment, normalization to unit sum and scaling of each variable. The effects of preprocessing are shown in Fig. 3. Compared to the original data, the baseline has been removed by calculating the first derivative as the data is centred around zero (Fig. 3b). The COW Table 1 The common features detected for chloroform-methanol (CM), boiling ethanol (BE) and freeze-thawing in methanol (FTM) extraction. The common features, i.e. the features detected in extracts for all three extractions methods, are 5608±429 (c=294). Error indications, representing 1 standard deviation, are calculated based on the number of combinations (c); e.g. for shared features between CM (six extracts) and FTM (seven extracts), a total of 42 combinations exist

CM BE FTM

CM

BE

FTM

7872±839 (c=6) 5657±460 (c=42) 7406±625 (c=42)

6591±852 (c=7) 5999±560 (c=42)

8949±455 (c=7)

algorithm aligned individual pixels in each chromatographic peak (Fig. 3c). Without baseline removal and chromatographic alignment, the PC model would describe these nonchemical variations in the metabolic fingerprints, thereby hampering any relevant interpretation of the model. The normalization removed variation not related to the chemical differences between classes, e.g. variations in injection volume and detector sensitivity, thereby enhancing separation between classes. Finally, scaling ensured that smaller peaks would have a greater chance to influence the model than without scaling to the SD (Fig. 3d). The trade-off of the scaling is a noise enhancement; however, this is not a problem for this application due to the high signal-to-noise ratio in the original raw data. Subsequently, pair-wise PCA was conducted on the processed data matrix, allowing chemical interpretation of differences in the metabolic fingerprint pixel by pixel through inspection of the cumulative sum of the PC1 loadings [28]. As data were scaled to 1/SD, the calculation of the cumulative sum was done after back-scaling by multiplying with SD for each pixel to return to a chromatographic peak shape. The back-scaled cumulative sum of PC1 loadings for each of the PC models is presented in Fig. 4a–c along with the PC1 scores of the corresponding model represented by a bar plot. For interpretation, metabolites with positive loadings are preferentially extracted with the extraction method having positive scores, and vice versa for negative values. No relevant systematic variation was observed on the PC2 loading [28], i.e. no information on differences in the metabolic fingerprints between extraction methods. The standard deviations on the scores indicate the repeatability for each of the extraction methods. The PC1 model scores and loadings for model (BE:CM) and model (BE:FTM) show that metabolites with large positive loadings, such as Ala, 3Glu, Fru, Glu and Un, are

Metabolic fingerprinting of Lactobacillus paracasei: a multi-criteria evaluation of methods for extraction...

Fig. 3 Effects of data processing on the retention time region from 20.7 to 20.9 min of 20 GC-TOF-MS BPI chromatograms (data set 2). The data set includes five biological replicates of each extraction, and for each of the five biological replicates, two were analyzed in duplicate. a BPI chromatograms, b after calculation of the first derivatives, c after retention time alignment using the optimal COW parameters and d after normalization to unit sum and scaling to the cumulative standard deviation

expected as both CM and FTM extraction was accomplished using a 50 % (v/v) aqueous methanol solution. While the FTM extraction is a one-phase system, the CM extraction is a twophase system consisting of the aqueous methanol phase and the relatively nonpolar chloroform phase. Hence, the nonpolar chloroform phase can retain the more nonpolar metabolites; thus, reduced levels of the nonpolar metabolites in the CM extracts would be expected [38]. From evaluation of the PC1 loading in model (CM:FTM) (Fig. 4c), the main differences in the metabolic fingerprints of CM and FTM extracts are high relative concentrations of the majority of the amino acids Ala, Val, Ill, Pro, Asp and Gln and of G6P and low relative concentration of Asn, Cit, FBP and some sugars in FTM extracts compared to CM extracts. No general trend of reduced levels of nonpolar metabolites in the CM extracts was observed. Lower relative concentrations of most of the metabolites in the CM extracts have been observed in previous studies [7, 16]. The poor performance has, to some extent, been suggested to be related to insufficient mixing of the two-phase system [15]. Frequent phase mixing by vortexing was performed during extraction; thus, we do not recon that the transfer to the aqueous methanol phase would be further increased by mixing. The metabolic fingerprint is affected by the extraction procedure

preferentially extracted with BE compared to CM (Fig. 4a) and FTM (Fig. 4b). Conversely, metabolites with negative PC1 loading in both models, such as Gly, Asp, Glut, Gln, Cit, G6P and FBP, are extracted with lower efficiency by BE compared to CM (Fig. 4a) and FTM (Fig. 4b). The lower relative extraction efficiencies of the sugar phosphates G6P and FBP with BE compared to FTM and CM may be due to the use of extensive physical forces during BE extraction. Previous studies demonstrated that the thermo-labile sugar phosphates have low recoveries with BE for Saccharomyces cerevisiae [17] and for Klebsiella oxytoca [14]. In case of degradation, the concentration of thermo-labile metabolites would be reduced, while the degradation products of larger thermo-labile molecule would increase, e.g. monosaccharide and disaccharide originating from polysaccharide degradation or amino acids originating from protein degradation, as suggested by Duportet et al. [2]. No general trend of BE showing higher relative concentrations of amino acids or monosaccharides was observed; however, large positive loadings for the two models are observed for Ala and Glu; thus, the production of these metabolites during the extraction process could be a possibility. PC1 for model (CM:FTM) accounts for only 16 % of the total variance in the data compared to 45 and 37 % for the model (BE:CM) and model (BE:FTM), respectively, demonstrating a larger similarity in the metabolic fingerprint of the CM and FTM extracts (Fig. 4c). This similarity is to be

FTM has been proven superior to both BE and CM for extracting intracellular metabolites of L. paracasei; i.e. a higher number of features were detected in FTM in addition to the largest percentage of unique features (Figs. 2 and 3). The pixel-based analysis demonstrated that the highest number of metabolites was correlated to FTM extraction and a higher repeatability of the method as illustrated by the lower standard deviation of the scores was observed (Fig. 4). Considering the complexity of the metabolome, it is not surprising that FTM did not ensure an optimal extraction of all metabolites. There are limitations to this investigation as the FTM method has been previously questioned due to indications of remaining enzymatic activity [8, 15]. In 2007, Faijes et al. [8] suggested FTM as the preferred method for Lactobacillus based on extraction efficiency of ATP, ADP, AMP, NAD+ and G6P. However, the intracellular pools of ATP was highly dependent on the layout, i.e. temperature and handling time of the FTM method, indicating that the enzymatic activity is not completely eliminated in the first couple of cycles. This was confirmed by Canelas et al. [15]. The methanol and the low temperatures (0 °C) in the thaw cycle will not inactivate the enzymatic activity but will simply slow it down [39]. Thus, factors like temperatures and handling time should be carefully controlled as changes will result in low reproducibility as well as a risk of misleading the metabolic profile. Enzymatic activity was not monitored in this study, but we suggest

K.B. Jäpelt et al. Fig. 4 Bar plot representing the scores for each of the three pairwise PC models, i.e. model (BE:CM), model (BE:FTM) and model (CM:FTM) on the processed data matrix containing, respectively, 13, 14 and 13 training set samples. The following is the abbreviations used in the plot: alanine (Ala), valine (Val), glycine (Gly), aspartic acid (Asp), 5-oxoproline (3Glu), glutamic acid (glut), glutamine (Gln), citric acid (Cit), fructose (Fru), glucose (Glu), unknown (Un), glucose-6phosphate (G6P) and fructose1,6-biphosphate (FBP). a PC1 loadings from model BE vs. CM (model (BE:CM), 45.39 % of the variance in the data explained), b PC1 loadings from model BE vs. FTM (model (BE:FTM), 36.91 % of the variance in the data explained), c PC1 loadings from model CM vs. FTM (model (CM:FTM), 15.98 % of the variance in the data explained)

recording the enzyme activity, i.e. changes in EC, during FTM extraction to keep track of potential metabolic changes in future studies. BE has been widely used for the extraction of metabolites from microbial cells. Studies by Canelas et al. [15] and Li et al. [12] proved BE to be among the preferred methods for extraction of microbial cells due to high efficiency, excellent recoveries of the range of metabolites analyzed and the relatively high number of extracted metabolites. The evaluation of the BE extraction of L. paracasei does not favour this method; it showed the lowest number of extracted features and lowest feature intensities compared to the FTM extracts (Fig. 2), lowest percentage of unique features (Table 1), and as shown using the pixel-based method, there are indications of metabolite degradation in the BE extracts for thermo-labile metabolites such as the phosphorylated sugars (Fig. 4a, b). CM, on the other hand, performed well in respect to the number of extracted features and percentage of unique features compared to BE (see Fig. 2 and Table 1), and the chromatographic profile of the CM extracts has been proven similar to the FTM extracts as demonstrated by the pixel-based analysis (Fig. 4c).

The CM method is technically more demanding, time consuming and possibly prone to experimental error and is not easily automated due to the need of a separation of the aqueous methanol phase from the chloroform phase, making it less desirable to use for high-throughput metabolic fingerprinting. This could explain the low repeatability of the method. However, the method can be the preferred method for other purposes such as for separations of lipids from other compounds, i.e. as seen in van det Werf et al. [18], as it provides increased metabolic coverage if analysis of the chloroform fraction is performed. Strengths of the multi-criteria approach An extraction method that provides the highest recovery for few selected metabolites is not necessarily the same as the method providing the highest metabolite coverage and accuracy of metabolic fingerprints. Thus, the use of small sub-sets of metabolites with similar physicochemical properties may not suggest the optimal method for metabolic fingerprinting [3, 8]. To obtain accurate and precise metabolic fingerprints, it

Metabolic fingerprinting of Lactobacillus paracasei: a multi-criteria evaluation of methods for extraction...

can therefore be argued that the optimization of each step in the analytical protocol should be based on multiple criteria that account both relevant factors and compound groups. However, there can still be a dimension of subjectivity in a multi-criteria approach, as the criteria selected as representative of the extraction method under investigation will affect the conclusion on the optimal method as well on the judgement of which criteria that should have the biggest influence on the final decision. In this case, all the criteria favourized FTM extraction which ease the determination of the optimal method; however, more complicated decision-making can arise from multiple criteria approaches if conflicting criteria arise. Efforts have been made to improve the quality of decision-making when a number of alternatives exist, and multi-criteria decision-making (MCDM) strategies are used to range alternatives using a number of ranking methods in order to find the optimal choice [40]. This could be the way forward in developing the multi-criteria approach. The data processing approach outlined here, of applying feature detection alongside with pixel-based data processing, enabled an untargeted multi-criteria approach for choosing the best extraction method for metabolic fingerprinting of L. paracasei. The feature detection allowed information on coverage and efficiency along with the similarity of extraction methods as measured by the common features, while the pixel-based data analysis revealed which metabolites caused the primary variation between the extraction methods. Thus, the pixel-based data analysis was demonstrated as a viable alternative to feature detection and extraction as there is no need for subsequent feature/peak matching and peak filling between samples [24, 25]. The chromatographic structure of the loadings furthermore eases understanding and interpretation. A particular advantage of pixel-based data processing is the ability to highlight what is big in data —variation wise— and, at the same time, (cor)related to the extraction methods. Furthermore, an extension of the pixel-based analysis to cover the entire unfolded data set (samples×retention time×m/z) instead of the reduced BPI data may lead to retrieval of additional chemical information, the trade-off being the complexity of such an assessment. In BPI, small peaks hidden in the vicinity of large peaks, overlapping peaks and peaks hidden in the baseline will be disregarded. Furbo et al. [41] dealt with this challenge in a study of pixel-based analysis of GC×GCFID data by unfolding the Rt1×Rt2 and then back-folding the loadings after PCA [41].

metabolomics, seeking to cover the full metabolome in an unbiased holistic manner, the analytical method should be expanded to other platforms, such as LC-MS. Acknowledgments We thank Chr. Hansen A/S and University of Copenhagen for financial support.

References 1. 2.

3.

4.

5.

6.

7.

8.

9.

10.

11.

12.

13.

14.

15.

Conclusion For GC-MS metabolic fingerprinting of L. paracasei, cells should be extracted using FTM as the method provided superior coverage and a high repeatability over nearly the whole range of metabolites analyzed. To approach true microbial

16. 17.

Ruiz L, Margolles A, Sanchez B (2013) Bile resistance mechanisms in Lactobacillus and Bifidobacterium. Front Microbiol 4:396 Duportet X, Aggio RBM, Carneiro S, Villas-Boas SG (2012) The biological interpretation of metabolomic data can be misled by the extraction method used. Metabolomics 8:410 De Koning W, Van Dam K (1992) A method for the determination of changes of glycolytic metabolites in yeast on a subsecond time scale using extraction at neutral pH. Anal Biochem 204:118 Gonzalez B, Francois J, Renaud M (1997) A rapid and reliable method for metabolite extraction in yeast using boiling buffered ethanol. Yeast 13:1347 Hiller J, Franco-Lara E, Weuster-Botz D (2007) Metabolic profiling of Escherichia coli cultivations: evaluation of extraction and metabolite analysis procedures. Biotechnol Lett 29:1169 Winder CL, Dunn WB, Schuler S, Broadhurst D, Jarvis R, Stephens GM, Goodacre R (2008) Global metabolic profiling of Escherichia coli cultures: an evaluation of methods for quenching and extraction of intracellular metabolites. Anal Chem 80:2939 Maharjan R (2003) Global metabolite analysis: the influence of extraction methodology on metabolome profiles of Escherichia coli. Anal Biochem 313:145 Faijes M, Mars AE, Smid EJ (2007) Comparison of quenching and extraction methodologies for metabolome analysis of Lactobacillus plantarum. Microb Cell Factories 6:27 Shin MH, Lee DY, Liu K, Fiehn O, Kim KH (2010) Evaluation of sampling and extraction methodologies for the global metabolic profiling of Saccharophagus degradans. Anal Chem 82:6660 Kim S, Lee DY, Wohlgemuth G, Park HS, Fiehn O, Kim KH (2013) Evaluation and optimization of metabolome sample preparation methods for Saccharomyces cerevisiae. Anal Chem 85:2169 Jernejc K (2004) Comparison of different methods for metabolite extraction from Aspergillus niger mycelium. Acta Chim Slov 51:567 Li X, Long D, Ji J, Yang W, Zeng Z, Guo S, Ji Z, Qi G, Chen S (2013) Sample preparation for the metabolomics investigation of poly-gamma-glutamate-producing Bacillus licheniformis by GCMS. J Microbiol Methods 94:61 Meyer H, Liebeke M, Lalk M (2010) A protocol for the investigation of the intracellular Staphylococcus aureus metabolome. Anal Biochem 401:250 Park C, Yun S, Lee SY, Park K, Lee J (2012) Metabolic profiling of klebsiella oxytoca: evaluation of methods for extraction of intracellular metabolites using UPLC/Q-TOF-MS. Appl Biochem Biotechnol 167:425 Canelas AB, ten Pierick A, Ras C, Seifar RM, van Dam JC, van Gulik WM, Heijnen JJ (2009) Quantitative evaluation of intracellular metabolite extraction techniques for yeast metabolomics. Anal Chem 81:7379 Rabinowitz JD, Kimball E (2007) Acidic acetonitrile for cellular metabolome extraction from Escherichia coli. Anal Chem 79:6167 Villas-Boas S, Hojer-Pedersen J, Akesson M, Smedsgaard J, Nielsen J (2005) Global metabolite analysis of yeast: evaluation of sample preparation methods. Yeast 22:1155

K.B. Jäpelt et al. 18.

19. 20.

21.

22.

23.

24.

25. 26.

27.

28. 29.

van der Werf MJ, Overkamp KM, Muilwijk B, Coulier L, Hankemeier T (2007) Microbial metabolomics: toward a platform with full metabolome coverage. Anal Biochem 370:17 Putri SP, Yamamoto S, Tsugawa H, Fukusaki E (2013) Current metabolomics: technological advances. J Biosci Bioeng 116:9 Honore AH, Thorsen M, Skov T (2013) Liquid chromatographymass spectrometry for metabolic footprinting of co-cultures of lactic and propionic acid bacteria. Anal Bioanal Chem 405:8151 Mashego MR, Rumbold K, De Mey M, Vandamme E, Soetaert W, Heijnen JJ (2007) Microbial metabolomics: past, present and future methodologies. Biotechnol Lett 29:1 Amigo JM, Popielarz MJ, Callejon RM, Morales ML, Troncoso AM, Petersen MA, Toldam-Andersen TB (2010) Comprehensive analysis of chromatographic data by using PARAFAC2 and principal components analysis. J Chromatogr 1217:4422 Petersen IL, Tomasi G, Sørensen H, Boll ES, Hansen HCB, Christensen JH (2011) The use of environmental metabolomics to determine glyphosate level of exposure in rapeseed (Brassica napus L.) seedlings. Environ Pollut 159:3071 Smith C, Want E, O’Maille G, Abagyan R, Siuzdak G (2006) XCMS: processing mass spectrometry data for metabolite profiling using nonlinear peak alignment, matching, and identification. Anal Chem 78:779 Katajamaa M, Orešič M (2007) Data processing for mass spectrometry-based metabolomics. J Chromatogr A 1158:318 Lommen A (2009) MetAlign: interface-driven, versatile metabolomics tool for hyphenated full-scan mass spectrometry data preprocessing. Anal Chem 81:3079 Aggio R, Villas-Boas SG, Ruggiero K (2011) Metab: an R package for high-throughput analysis of metabolomics data generated by GC-MS. Bioinformatics 27:2316 Christensen JH, Tomasi G (2007) Practical aspects of chemometrics for oil spill fingerprinting. J Chromatogr A 1169:1 Danielsson R, Bylund D, Markides K (2002) Matched filtering with background suppression for improved quality of base peak chromatograms and mass spectra in liquid chromatography-mass spectrometry. Anal Chim Acta 454:167

30.

Ullsten S, Danielsson R, Backstrom D, Sjoberg P, Bergquist J (2006) Urine profiling using capillary electrophoresis-mass spectrometry and multivariate data analysis. J Chromatogr A 1117:87 31. Christensen J, Mortensen J, Hansen A, Andersen O (2005) Chromatographic preprocessing of GC-MS data for analysis of complex chemical mixtures. J Chromatogr A 1062:113 32. Savitzky A, Golay M (1964) Smoothing + differentiation of data by simplified least squares procedures. Anal Chem 36:1627 33. Tomasi G, van den Berg F, Andersson C (2004) Correlation optimized warping and dynamic time warping as preprocessing methods for chromatographic data. J Chemom 18:231 34. Skov T, van den Berg F, Tomasi G, Bro R (2006) Automated alignment of chromatographic data. J Chemom 20:484 35. Christensen J, Tomasi G, Hansen A (2005) Chemical fingerprinting of petroleum biomarkers using time warping and PCA. Environ Sci Technol 39:255 36. Soleimani M, Farhoudi M, Christensen JH (2013) Chemometric assessment of enhanced bioremediation of oil contaminated soils. J Hazard Mater 254:372 37. Gallotta FDC, Christensen JH (2012) Source identification of petroleum hydrocarbons in soil and sediments from Iguaçu River Watershed, Paraná, Brazil using the CHEMSIC method (CHEMometric analysis of Selected Ion Chromatograms). J Chromatogr A 1235:149 38. Folch J, Lees M, Stanley G (1957) A simple method for the isolation and purification of total lipides from animal tissues. J Biol Chem 226:497 39. More N, Daniel R, Petach H (1995) The effect of low-temperatures on enzyme-activity. Biochem J 305:17 40. Pohekar SD, Ramachandran M (2004) Application of multi-criteria decision making to sustainable energy planning—a review. Renew Sust Energ Rev 8:365 41. Furbo S, Hansen AB, Skov T, Christensen JH (2014) Pixel-based analysis of comprehensive two-dimensional gas chromatograms (color plots) of petroleum: a tutorial. Anal Chem 86:7160

Metabolic fingerprinting of Lactobacillus paracasei: a multi-criteria evaluation of methods for extraction of intracellular metabolites.

An untargeted multi-criteria approach was used to select the best extraction method among freeze-thawing in methanol (FTM), boiling ethanol (BE) and c...
594KB Sizes 0 Downloads 6 Views