Cs

ZD

a

ARRAA

KMTHY

1

tttbptfca(ptd

tiptw

h1

Journal of Chromatography B, 966 (2014) 83–92

Contents lists available at ScienceDirect

Journal of Chromatography B

jou rn al hom ep age: www.elsev ier .com/ locate /chromb

onstruction of a metabolome library for transcription factor-relatedingle gene mutants of Saccharomyces cerevisiae�

anariah Hashim, Shao Thing Teoh, Takeshi Bamba, Eiichiro Fukusaki ∗

epartment of Biotechnology, Graduate School of Engineering, Osaka University, 2-1 Yamadaoka, Suita, Osaka 565-0871, Japan

r t i c l e i n f o

rticle history:eceived 17 September 2013eceived in revised form 17 May 2014ccepted 19 May 2014vailable online 27 May 2014

eywords:

a b s t r a c t

Transcription factors (TFs) play an important role in gene regulation, providing control for cells to adaptto ever changing environments and different physiological states. Although great effort has been taken tostudy TFs through DNA–protein binding and microarray gene expression experiments, the understand-ing of transcriptional regulation is still lacking, due to lack of information that links TF regulatory eventsand final phenotypic change. Here, we focused on metabolites as the final readouts of gene transcriptionprocess. We performed metabolite profiling of 154 Saccharomyces cerevisiae’s single gene knockouts each

etabolomicsranscription factorsierarchical clusteringeast Saccharomyces cerevisiae

defective in a gene encoding transcription factor and built a metabolome library consists of 84 metabo-lites with good reproducibility. Using the metabolome dataset, we obtained significant correlations andidentified differential strains that exhibit altered metabolism compared to control. This work presentsa novel metabolome dataset library which will be invaluable for researchers working on transcriptionalregulation and yeast biology in general.

. Introduction

Cells employ an elaborate and complex gene expression systemhat allows them to reprogram their genetic make-up in respondso different environments and growth demands. At the forefront ofhis control system lies transcription factors (TFs), proteins thatind to specific DNA sequences to further recruit or block RNAolymerase II, resulting in the initiation or suppression of generanscription. For decades researchers have been interested in theunction and regulation of TFs; biotechnologically this knowledgean help in improvement of industrially important microbial strainss demonstrated by global transcriptional machinery engineeringgTME) technique [1,2], and clinically TFs themselves can serve asotential drug targets such as estrogen receptors and c-Myc pro-ein for cancer therapies [3,4], and generally proposed for new drugiscovery [5,6].

Thus, various studies have been conducted to decipherhe intricate regulatory network of TFs, which among othersnvolve determination of DNA-binding domain (DBD) motifs and

rotein–protein interaction, identification of downstream effec-or genes, quantification of transcript and protein abundances, asell as network construction from genome-wide expression data

� This paper is part of the special issue “Metabolomics II” by G. Theodoridis.∗ Corresponding author. Tel.: +81 06 6879 7424; fax: +81 06 6879 7424.

E-mail address: fukusaki@bio.eng.osaka-u.ac.jp (E. Fukusaki).

ttp://dx.doi.org/10.1016/j.jchromb.2014.05.041570-0232/© 2014 Elsevier B.V. All rights reserved.

© 2014 Elsevier B.V. All rights reserved.

using computational methods [7–11]. Despite the huge amountof research, the understanding of global gene regulation by tran-scription factors is not yet complete. While TFs essentially bindto DNA promoter regions to initiate their action, transcriptionalregulatory is not a simple binary on/off control. The number of TFmolecules also plays a part in determining the level of transcrip-tion [12]. In addition, one TF may be involved in various genesand requires a precise set of protein complex and co-activatorsbefore transcription can be initiated. Furthermore, many regulatoryevents that link triggering cues to final phenotypic reprogrammingremain poorly characterized, making it difficult to predict cellu-lar behavior even when the transcriptional machinery is known.The missing link between the change in transcript or protein levelsand phenotype (e.g. growth rate, chemical resistance, productionof secondary metabolites) has to be investigated. Clearly, an alter-native approach to studying TF regulation besides gene expressionprofiling or DNA and protein–protein binding is necessary in orderto gain overall picture of gene regulatory mechanism. In particu-lar, we are interested to know how perturbation in transcriptionfactors affects metabolite levels and ultimately cellular function.

In order to investigate the effects of transcription factor per-turbation toward metabolic profile, we selected metabolomics.Metabolomics, the comprehensive profiling of metabolites con-

tained in a cell or an organism, has been regarded as a highresolution approach, due to its ability to capture subtle changein metabolite levels which often does not manifest in a changein phenotype until at much later stage. For example, Yoshida

8 matog

arovappamtbfpdmngm

mdyrafstnHtt(it

fsgssmdTmaglrg

2

2

fW(u[rgwse

4 Z. Hashim et al. / J. Chro

nd colleagues [13] demonstrated that the difference in yeasteplicative lifespan can be detected from metabolic fingerprintsf exponentially growing yeast cells (∼4 doublings) whereas con-entional method by counting the number of daughter cells from

single mother cell requires at least 20 generations before com-arison can be made. This finding shows that metabolomics is aowerful tool to uncover a complicated phenotype (i.e. lifespan)t earlier stage compared to conventional method. Furthermore,etabolomics has also been used to reveal silent genes, i.e. genes

hat produce no overt phenotype when deleted from the genome,ut have distinct metabolic concentrations [14]. Therefore, weound metabolomics as the best approach to studying metabolichenotype of non-essential gene knockouts of transcription factorsue to: (1) unlike enzymes, TFs have no one-to-one relation withetabolic pathways, and thus hold much more complex regulatory

etwork, and (2) most of these genes are purportedly silent, i.e. theene deletion does not affect growth rate of the organism, leavingetabolic change as the sole indicator of strain condition.So far, studies regarding global transcriptional regulation using

etabolomics approach have been limited. Analysis of condition-ependent TF network using metabolic flux distribution of 119east TF deletion strains was reported previously [15], whicheveals that metabolic flux alteration caused by TF deletion occurslmost exclusively at TCA cycle, and only 23 strains exhibited dif-erential flux ratio change. Amino acid profiling for ∼5000 yeastingle gene deletion strains was also conducted [16], which showshat clustering of functionally related genes can be found for argi-ine biosynthesis and urea cycle pathways but not other pathways.owever, there are no comprehensive reports on metabolites other

han amino acids nor a dedicated metabolite profiling for transcrip-ion factor deletion strains. Although Yeast Metabolome DatabaseYMDB) [17] serves as a database that lists all metabolites containedn yeast, there are no reports on metabolite levels of specific strainypes.

The objective of this research is to build a metabolome datasetor TF-related gene mutants of Saccharomyces cerevisiae. In thistudy, we took advantage of the commercially available yeast sin-le gene deletion library and measured metabolite levels of 154trains each lacking a non-essential gene putatively encoding tran-cription factor and characterize the strains according to theiretabolic profiles. We identified core metabolites and co-factors

eriving from central metabolic pathways such as glycolysis andCA cycle, as well as amino acids and organic acids that are com-only conserved in most organisms. Yeast S. cerevisiae serves as

practical model organism due to the wealth of database forenes/transcripts/proteins and the availability of commercial strainibrary. Metabolome dataset can serve as invaluable inputs to assistesearchers working on transcription factors and yeast biology ineneral.

. Materials and method

.1. Chemicals and reagents

Standard metabolites of the highest purity were purchasedrom NacalaiTesque (Kyoto, Japan), Sigma–Aldrich (MO, USA) and

ako (Osaka, Japan). 1,4-Piperazinediethanesulfonic acid, PIPESDojindo, Kumamoto, Japan) and ribitol (Wako, Osaka, Japan) weresed as internal standards for LC/MS and GC/MS respectively13]. Tributyl amine (TBA) of >98.5% purity used as ion pairingeagent was obtained from Sigma–Aldrich. All other analytical

rade reagents were obtained from Wako (Osaka, Japan). Ultrapureater and/or LC/MS-grade methanol were used for stock solutions,

olvent mixtures and mobile phases. Individual stock solutions ofach standard compound at a concentration of 10–100 mM were

r. B 966 (2014) 83–92

prepared. The working mixtures of standard compounds were pre-pared by diluting these stock solutions.

2.2. Yeast cultivation and sample preparation

154 transcription factor-related (or putatively related) singlegene knockouts from EUROSCARF collection were studied, withBY4742 (MAT ̨ leu2�0 lys2�0ura3�0 his3�1) as the parental back-ground. The mutant strains were purchased from Open Biosystems(Huntsville, AL, USA) which were constructed by replacing the tar-get genes with kanMX cassette that confers resistance to geneticin[18]. A list of strains used in this study is available from Supple-mentary Table S1. These strains (n = 3) were randomly divided intofour batches, each batch included an independent parental strainas control.

Yeast cells from frozen glycerol stock were plated onto YPDagar plates 10 g/L yeast extract, 20 g/L peptone, 20 g/L glucose,20 g/L agar (all from BD, MD, USA except glucose and agar fromNacalai Tesque), with added geneticin G418 (Wako, Osaka, Japan)200 �g/mL for knockout strains and grown at 30 ◦C for 2 days.After two days, a single colony was obtained and re-streaked ona new YPD plate to increase cell number. This plate was used asa master plate for all cultivations. Cultivation was performed inthree steps to reduce variation in cell growth; pre–pre culture,pre-culture and main culture. For pre–pre culture, cells were inoc-ulated from the master plate into 3 mL of culture medium, followedby incubation at 30 ◦C in a rotary shaker (200 rpm) for 18 h. Next,for pre-culture, a portion of pre-pre-culture broth was diluted into15 mL fresh culture medium so that the starting optical density at600 nm (OD600) = 0.01, and incubation with shaking was contin-ued for 18 h. For main culture, the pre-culture broth was dilutedin 15 mL culture medium so that the starting OD600 = 0.1, and incu-bation with shaking was continued until OD600 = 1. SC (syntheticcomplete medium) [19] was prepared as follows: 20 g/L glucose,6.7 g/L DifcoTM yeast nitrogen base without amino acids (BD, MD,USA), 1.92 g/L yeast synthetic drop-out media supplement withouturacil (Sigma–Aldrich), 76 mg/L uracil (Sigma–Aldrich) and used forall liquid cultures, with geneticin G418 added to a final concentra-tion of 200 �g/mL for knockout strains. iMark microplate reader(Bio-Rad, CA, USA) was used to monitor optical density.

For metabolome sampling, a fast filtration method basedon [20] was applied with slight modifications. 5 mL of cul-ture broth at OD600 = 1 (approximately 5 × 107 cells) wererapidly filtered using a0.45 �m-pore size, 25 mm-diameter nylonmembrane (Millipore, MA, USA) under a vacuum filtration.Filter membrane was folded and inserted into a 2-mL samp-ling tube filled with 1 mL single-phase extraction solvent(methanol/chloroform/water = 5/2/2 v/v/v%, with 1.2 �g/mL eachof PIPES and ribitolas internal standards [13]) pre-cooled at −30 ◦C,after which the tube was flash frozen in liquid nitrogen and storedat −80 ◦C until extraction.

For extraction, tubes filled with membrane-bound cells andextraction solvent were placed in a thermomixer (Eppendorf, Ham-burg, Germany) at 4 ◦C, 1200 rpm for 30 min. After that, all liquidextract (900 �L) was transferred to a new tube filled with 400 �Lwater, vortexed and centrifuged at 4 ◦C, 16,100 rcffor 3 min to sep-arate polar and non-polar phases. Next, the upper polar phasewas transferred to a new tube via syringe filtration (0.2 nm PTFEhydrophilic membrane, Millipore, MA, USA) and divided into two,300 �L for LC–MS and 600 �L for GC–MS. The extracts were con-centrated five times from the initial volume under a vacuumcentrifugation system (VC-96R, Taitec, Japan), transferred to glass

vials (Chromacol, Hertfordshire, UK) and ready for UHPLC–MS/MSanalysis. Samples were analyzed within 24 h.

For GC/MS samples, concentrated extracts were lyophilizedovernight, followed by derivatization by oximation and

matog

sr(aA(J3fw

2

ssgtcmasoc

tHqut(0t

2

mStSpahimmeswcTOIma–cpa4i(lta

Z. Hashim et al. / J. Chro

ilylation [21]. The oximation reagent, methoxyamine hydrochlo-ide (Sigma–Aldrich, MO, USA) was first dissolved in pyridineWako, Osaka, Japan) to a concentration of 20 mg/mL and 75 �Ldded to each sample tube containing the lyophilized extracts.fter reaction at 30 ◦C, 1200 rpm for 90 min, 50 �L of N-methyl-N-

trimethylsilyl) trifluoroacetamide (MSTFA) (GL Sciences, Tokyo,apan) was added and the silylation reaction was performed at7 ◦C, 1200 rpm for 30 min. The derivatized samples were trans-erred to glass vials (Chromacol, Hertfordshire, UK) and analyzedithin 24 h.

.3. Growth of yeast strains and sampling procedure

We chose SC as the growth medium to enable all deletiontrains to grow comparably without severe growth defect, whiletill allowing controlled and known nutrient composition [16]. Therowth of yeast strains in SC medium was observed by measuringhe optical density at 600 nm for several time points during mainulture (Supplementary Table S1). Growth rates serve as a generaleasure of the effects of gene deletion. Control strain BY4742 grew

t a maximum specific growth rate of 0.45–0.48 h−1. All deletiontrains grew well in SC medium except for six strains (ino2�, ino4�,pi1�, gcr2�, aft1�, ada2�) that exhibited growth defect of >20%ompared to control.

In microbial metabolome experiments, it is important to ensurehat metabolites are rapidly quenched at the time of sampling.ere, we applied a fast filtration method followed by subse-uently dipping the cells into cold extraction solvent [20], whichsually takes ∼30 s from taking out samples from liquid cul-ure to quenching. Adenylate energy charge, EC, calculated as[ATP] + 0.5[ADP])/([AMP] +[ADP] + [ATP]), was in the range of.72–0.88 (Supplementary Table S1), which is typical of exponen-ially growing cells [22], suggesting that quenching was sufficient.

.4. Ion pairing UHPLC–MS/MS analysis

The analysis platform consists of Shimadzu ultrahigh perfor-ance liquid chromatography (UHPLC) system (Nexera series,

himadzu, Kyoto, Japan) coupled to a triple quadrupole mass spec-rometer (LCMS-8030 with modification to improve sensitivity,himadzu, Kyoto, Japan). We selected ion pairing chromatogra-hy for its ability to retain highly polar metabolites that otherwisere eluted near the void volume in C18 reversed phase LC, and itas better separation and higher signal compared to hydrophilic

nteraction chromatography (HILIC) in the analysis of centraletabolites [23]. The method was developed based on [24,25],odified to match our LC and MS system and tested with sev-

ral types of columns and analytical parameters to optimize peakhape and separation profile. The MS/MS fragment for each analyteas determined using authentic standards. In addition, we suc-

essfully accelerated the analysis time from 36 min [25] to 15 min.he final analytical conditions were as follows, column: l-Column2DS (150 mm × 2.1 mm, 3 �m, Chemicals Evaluation and Research

nstitute Japan); flow rate: 0.3 mL/min; column temperature: 35 ◦C;obile phase A: water containing 10 mM TBA and 15 mM acetic

cid; mobile phase B: methanol; gradient program: 0%B(0–0.5 min) 25%B(7.5 min) – 90%B(11–11.5 min) – 0%B(11.6–15 min); sampleycle time: 15 min; injection volume: 3 �L. The mass spectrometricarameters were: ESI negative mode; desolvation line (DL) temper-ture: 250 ◦C; nebulizer gas flow: 2 L/min; heat block temperature:00 ◦C; other parameters were optimized automatically by flow

njection analysis and auto-tuning. Multiple reaction monitoring

MRM) transition parameters and retention time for each metabo-ite are listed in Supplementary Table S2. One analytical batchypically consisted of around 160 injections (40 h), after which thenalysis was stopped for routine cleaning and preparation of fresh

r. B 966 (2014) 83–92 85

mobile phases, before resumed for the next analytical batch. Allsamples were maintained at 4 ◦C in the autosampler during analy-sis.

2.5. GC/MS analysis

GC/MS was performed on a GCMS-QP2010 Ultra (Shimadzu,Kyoto, Japan) gas chromatograph coupled with quadrupole massspectrometer equipped with an AOC-20i/s autoinjector (Shimadzu,Kyoto, Japan). A CP-SIL 8 CB Low Bleed/MS column (Varian, CA,USA) 30 m × 0.25 mm (0.25 �m) was used for the GC separation. Themass spectrometer was auto-tuned and calibrated prior to analy-sis. 1 �L of sample was injected in split mode with a split ratio of1:25. The inlet temperature was set at 230 ◦C and column flow ratewas 1.12 mL/min (linear velocity 39 cm/s). The column tempera-ture was held at 80 ◦C for 2 min, raised by 15 ◦C/min to 330 ◦C, andheld at 330 ◦C for 6 min. The transfer line and ion source tempera-tures were 250 ◦C and 200 ◦C respectively. Electron ionization (EI)was performed at 70 eV. The mass range of the detector was setto m/z 85–500 and the detector voltage (set by auto-tuning) was0.93 kV.

An alkane standard mix was prepared from 25 �L each ofC8–C20 and C21–C40 alkane standard solutions, diluted with anaddition of 25 �L pyridine and injected at the start of each analyticalrun for calculating retention indices. In addition, a blank pyridinesample was injected every 8 samples for diagnostic purposes (tocheck for column bleed and carryover).

2.6. Metabolite identification and validation procedure

We identified metabolites contained in yeast extracts by match-ing the extracted ion chromatograms with our in-house metabolitelibrary (MS/MS fragment and retention time). The identity waschecked by spiking authentic standards to yeast extract and con-firming that the particular metabolite peak intensity increases withadded concentration. Additional information regarding metaboliteconfirmation is available in Supplementary data.

The stability and reproducibility of the method were evaluatedusing pooled quality control (QC) samples [26–28]. QC sampleswere prepared by pooling unconcentrated yeast extracts from eachsample within the same batch, followed by vacuum concentrationwith the same concentration rate as test samples. The same QCaliquot from one injection vial was used for one analytical batch,injected at least three times at the start of the analytical batch afterthe system has stabilized, then at every sixth injection throughoutthe entire analytical workflow. Additionally a standard mixture of1 �M was routinely injected in the beginning, middle and end ofeach analysis run.

2.7. Dataset construction and data analysis

For LC/MS, peak picking was conducted by LabSolutions(Shimadzu, Kyoto, Japan) followed by manual inspection. Theparameters were set as follows: integration: auto, max peak: 3,width: 5 s; smoothing: standard, counts: 5, width: 1 s; identifi-cation: absolute RT & closest peak, target window: 5%, referencewindow: 5%, process time: ±1 min. To correct for matrix effectcommonly observed in ESI-based LC/MS, the raw peaks were cali-brated using external calibration method. Initially, primary stocksolutions from authentic standards were prepared in water ata concentration of 10–100 mM for each metabolite, from which

standard mixtures of various concentrations were made. Standardmixtures were spiked into yeast extracts and used for making cal-ibration curves. The calibrated peaks (converted to concentration)were then multiplied with a correction factor (peak intensity of

8 matog

sm

ft(rdafq

(lma

MSCf

2

HAdaaawftwwlwtowsm

3

3

aspofo(sfidifAicl

6 Z. Hashim et al. / J. Chro

tandard mixture during calibration/peak intensity of standardixture during actual run).GC/MS raw data files were converted into netCDF (*.cdf)

ormat according to the ANDI (Analytical Data Interchange Pro-ocol) specification using the proprietary software GCMS solutionShimadzu, Kyoto, Japan) before peak detection, baseline cor-ection and retention time alignment using the freely availableata processing tool MetAlign [29]. Data matrices from thelignment were then imported into AIoutput2 ver.1.29 [30]or automated RI-based target compound identification anduantification.

For final metabolome dataset, after peak values correctionLC/MS) and normalization to internal standard ribitol (GC/MS),og2-transformed fold change relative to control strain for each

etabolite were calculated and data was integrated from LC/MSnd GC/MS.

Principal component analysis (PCA) was performed usingetaboAnalyst 2.0 [31] and SIMCA-P+ ver13 (Umetrics, Umeå,

weden). Hierarchical clustering analysis (HCA) was performed byluster 3.0 [32] and viewed on Java Treeview [33]. T-test was per-ormed using MS Excel.

.8. Identification of differential strains

Two parameters were used as a measure for differential strains:otelling’s T2 range and average deviation from mutant median.

strain is defined as differential if it is located outside the confi-ence border of 95% based on T2 statistics on PCA plot or havingn average deviation from mutant median of ≥1.3. In this study,verage deviation from mutant median specifically denotes theverage deviation from the median of mutant measurement, andas employed as an alternative measure of difference, independent

rom the wild-type strain BY4742, since there is a possibility thathe wild-type profile is distorted in some metabolites. Hotelling’s T2

as performed using SIMCA-P+ ver13 (Umetrics, Umeå, Sweden),hile average deviation from mutant median was calculated as fol-

ows: the concentration of each metabolite in individual sampleas divided with the median value of the metabolite concentra-

ion across all mutants, summed over all metabolites, and averagedver the number of metabolites. Division, instead of subtractionas used considering the different magnitudes of metabolite inten-

ity values, to avoid overrepresentation of metabolites with largeagnitude/highly abundant metabolites.

. Results and discussion

.1. Validation of analytical performance and data filtering

The use of QC samples from pooled test extracts to monitornalytical performance has been demonstrated in metabolometudies of urine [26,34] and plasma or serum [35]. QC samples pre-ared from aliquots of test samples provide ‘mean’ representativef all the metabolites contained and thus considered appropriateor the evaluation of reproducibility and sample stability. It wasbserved that QC samples were clustered together on PCA plotFig. 1A), demonstrating that the analysis platform is sufficientlytable throughout the run. Furthermore, we used QC samples tolter out metabolites with poor reproducibility (relative standardeviation, RSD of peak area with a cut-off value 30% [28]), yielding

n a total of 67 metabolites (median RSD of 6.6%, 8.1%, 10.9%, 9.5%or Batch 1–4, respectively) obtained from UHPLC–MS/MS analysis.

dditionally, we obtained 17 more metabolites undetected or hav-

ng poor peak shape in LC–MS by GC–MS (RSD ≤ 30% based on theontrol sample), making a total of 84 metabolites in the final peakist.

r. B 966 (2014) 83–92

3.2. Normalization of batch-to-batch variation

Ion pairing LC–MS has the advantage of wide coverageof metabolites of various species, including polar metabolitesfrom central metabolism, with relatively stable retention time[24,25,36]. However, ion pairing LC–MS lacks reproducibility of dif-ferent batch analysis. Peak intensities tend to deteriorate over time,while there is a need for regular cleaning due to the accumulationof residual ion pairing reagent in the analysis line. As reported pre-viously, day-to-day analytical variation was inevitable, accountingfor the major portion of data variability [28]. Therefore, integra-tion of data from different analytical runs and different batchesneeds thorough consideration and a proper normalization method.It is important to minimize these differences so that true inter-pretations of biological phenomena can be derived, which reflectsstrain differences instead of batch differences. In this study, due toa large number of strains, we performed cultivation and analysisin four batches and included parental strain BY4742 as control ineach batch. To examine the degree of batch-to-batch variation, wefirst plotted only control strains in PCA (Supplementary Fig. 1). Wefound that normalization to an internal standard (calculated as theratio of peak intensity of each metabolite to the peak intensity ofthe internal standard) was not sufficient for eliminating batch-to-batch variation. As shown in Fig. 1A and B, while having excellentintra-batch reproducibility, inter-batch reproducibility was poor.Similarly, a distinct separation among batches can be observedwhen samples from all four batches were combined (Fig. 2A). Thisresult is consistent with a previous study of long-term humanserum metabolomics which reported that longitudinal variationscannot be easily compensated for with internal standards [37].However, the variation was remarkably reduced when normaliza-tion to control strain (calculated as log2-transformed fold-change,i.e. ratio of metabolite in the mutant strain vs. control) was per-formed (Fig. 2B). One control strain from Batch 1 was found to bean outlier and thus omitted from the dataset (Supplementary Fig.1). For LC/MS, because of narrow linear range possibly caused by ionpairing reagent, peaks were first converted to concentrations usingcalibration curves and corrected with a correction factor (see Sec-tion 2.7) before calculation of fold-change, while for GC/MS, peakareas normalized to ribitol were used directly. The final dataset for84 metabolites used in data analysis, can be found from Supple-mentary Table S3.

3.3. Hierarchical clustering analysis

To classify strains according to their metabolic profiles, hierar-chical clustering analysis (HCA) was performed. HCA is a methodwhich finds hierarchy of clusters that share similar characteristicsbased on distance or similarity measure. Compared to other meth-ods such as K-means clustering and self-organizing map (SOM),HCA does not require a priori information of the data structure suchas cluster number, thus it serves as a convenient unsupervised toolfor interpreting complex experimental data [38].

Several algorithms are available for HCA. However, selectionof a proper algorithm is subjective and depends on the purposeof the classification and whether or not the classification resultsatisfies a pre-determined criterion. In this study, we chose theclustering algorithm based on two criteria: (1) ability to keepthe original structure of the data matrix as close as possible, inwhich differential strains with strong characteristics (large foldchange values) and outliers can be distinguished from the rest,and (2) having a reasonably distributed clusters when the hierar-

chical tree is cut at a certain cut-point. We found that Euclideandistance is appropriate for this purpose, as it gives a directmeasure of magnitude and thus was able to separate differen-tial strains. In contrast, Pearson’s correlation provides a relative

Z. Hashim et al. / J. Chromatogr. B 966 (2014) 83–92 87

Fig. 1. Stability and reproducibility of the analysis method (A) within the same analytical run, and (B) between two separate analytical runs as demonstrated by principalcomponent analysis, PCA. The peaks were normalized to internal standard (ratio of peak intensity of each metabolite to the peak intensity of the internal standard) andPareto-scaled (mean-centered and divided with the square root of standard deviation). QC samples injected periodically were clustered together, showing that the methodi ate rua and

s

dtmcm

F(NE

s stable and reproducible within the same analysis run. However, when two separlternative normalization procedure is necessary. Data was taken from (A) Batch 1tatistics.

istance measure independent with magnitude, which made it fail

o isolate differential strains. When comparing different linkage

ethods, single and centroid could not generate an appropriateut interpretation, with many small distorted clusters having fewembers. Average and complete linkages performed comparably,

ig. 2. Comparison of two normalization methods for combining data from all four batchlog2-transformed fold-change of metabolites relative to control). (A) Was Pareto-scaled

umbers 1–4 indicate batch number. Batch-to-batch variation was greatly reduced and dllipse indicates 95% confidence border based on Hotelling’s T2 statistics.

ns were combined, batch separation can be observed along PC1 (B), indicating that(B) Batch 1 and 2. Ellipse indicates 95% confidence border based on Hotelling’s T2

but the former had an overall structure closer to the original data

matrix. Ward’s minimum linkage was efficient in finding compact,homogenized clusters, but incapable of filtering outliers. There-fore, we chose Euclidean distance with average linkage for ourstudy.

es, (A) normalization to internal standard, and (B) normalization to control strainto improve normal distribution while no additional scaling was performed for (B).ata was evenly distributed when normalization to control strain was employed (B).

88 Z. Hashim et al. / J. Chromatogr. B 966 (2014) 83–92

Fig. 3. Heat map showing metabolite changes in 154 transcription factor-related mutant strains analyzed in this study. Clustering was based on hierarchical clusteringanalysis, HCA using log2normalized to parental strain (fold-change) dataset. The clustering parameters were as follows: Euclidean distance and average linkage for strainclustering; Pearson’s r and average linkage for metabolite clustering. The more differential a strain is, the farther it is located from the center and closer to the outer hierarchy.

Z. Hashim et al. / J. Chromatogr. B 966 (2014) 83–92 89

Table 1List of differential strain clusters obtained from hierarchical clustering. Strains were listed in hierarchical importance (i.e. from outer to inner hierarchy).

Group Cluster Members Correlation

1: Differential, no clusters – arg82�, ada2�, sef1�, swi6�, aft1�, cbf1�, gcr2�, ada3�, sin3�, gln3�, stb5�, rpn4� –

2: Differential andformed clusters

1 ino2�, ino4�, opi1�, ric1� 0.852 pho23�, stp2� 0.863 mks1�, rrn10�, rtg3�, rtg1� 0.874 mac1�, sum1� 0.895 cst6�, sds3� 0.926 sfl1�, usv1�, azf1�, rlm1� 0.927 sas2�, ccr4� 0.918 gcn4�, hap3�, met31� 0.899 yap1�, hir1�, tec1� 0.91

10 ixr1�, hal9�, rim101�, tuf1� 0.9111 ace2�, dal81�, tea1�, ash1� 0.9212 ppr1�, skn7� 0.95

3: Not differential andformed clusters

13 ecm22�, hap4�, aro80�, fzf1�, arg80�, uga3� 0.9514 met28�, oaf1� 0.9515 bas2�, bas1�, ime1�, lys14�, leu3�, cad1�/yap2�, dal80�, yap6�, mot3�, pho4�, stp1� 0.9616 hap2�, hap5�, thi2� 0.9517 msn1�, aft2�, arr1�, yap5�, cin5�, crz1�, pdr8�, cha4�, msn4�, rgt1�, pdr3�, msn2�,

cat8�, sko1�, yap7�, xbp1�, wtm2�, mig2�, pdr1�, sut1�, ume6�, tye7�, yrr1�0.96

18 nrg1�, yap3� 0.9519 adr1�, hac1�, smp1�, yrm1� 0.9520 gat3�, gat2�, put3�, gat1�, gis1�, gat4�, spt23�, mal33�, dal82�, rsf2�, zap1�, mig1�,

mig3�, mal13�, rgm1�0.94

21 cup9�, argr2�, gal3�, gal4�ace1�/cup2�, mga1�, gal80�, ssn6�, sut2�

0.96

22 sir3�, sir4�, sir1�, zds2� 0.9623 hms1�, fkh2�, yhp1� 0.9524 rox1�, hir2�, hir3� 0.9625 ezl1�, ndt80�, fkh1�, rfx1�, phd1�, flo8�, wtm1�, zds1�, yox1�, kar4�, rph1�, swi4�,

swi5�0.94

26 hcm1�, mbp1� 0.9527 sir2�, tsp1� 0.94

–

m(ttifdjTcpstncmop(lbwl(

tetmi

Table 2List of differential strains based onaverage deviation from mutantvalues andHotelling’s T2 statistics.

No. Strain Average deviationfrom mutanta

Hotelling’s T2

rangebGroupc

1 arg82� 3.16 96.77 12 ada2� 1.85 69.95 13 aft1� 1.50 40.62 14 gcn4� 1.44 11.31 25 azf1� 1.43 22.83 26 swi6� 1.37 54.94 17 usv1� 1.37 19.70 28 ino2� 1.37 35.02 29 sum1� 1.37 25.15 2

10 mac1� 1.32 13.86 211 ada3� 1.32 50.20 112 gln3� 1.31 15.62 113 rlm1� 1.30 15.04 214 tea1� 1.30 5.19 215 cbf1� 1.30 12.82 116 sin3� 1.30 13.68 117 sas2� 1.30 6.96 218 ace2� 1.29 13.76 219 cst6� 1.28 30.21 220 dal81� 1.26 7.75 2

a

∑n

i=1Xi/Mi

n , X: metabolite concentration in individual mutant, M: median ofmetabolite concentration across all mutants, i, . . ., n: metabolite ID. This value was

4: Not differential, no clusters – gzf3�, sip4�, not3�, uaf30�

Using metabolome dataset of log2 fold-change values (Supple-entary Table S3), hierarchical clustering analysis was performed

Fig. 3). Here, the farther a strain is located on the outer hierarchy,he more differential it is relative to control. As illustrated in Table 1,he transcription factor deletion strains can mainly be categorizednto four groups: (1) differential, no clusters, (2) differential andormed clusters, (3) not differential and formed clusters, and (4) notifferential, no clusters. The cut-point for “differential” on HCA is

ustified by average deviation from mutant median and Hotelling’s2 values (Table 2). Moreover, clustering of metabolites based onorrelation on Fig. 3 revealed that metabolites that share similarathway (e.g. amino acids biosynthesis) or having similar chemicaltructure (e.g. nucleotides, sugar phosphates) tended to be groupedogether or closely positioned with each other. Five metabolites,amely arginine, histidine, guanine, inosine and guanosine indi-ated significant changes (p < 0.05) in more than 70% of theutant strains, which could possibly be a unique characteristic

f BY4742 derivatives, whereas pyroglutamate, fructose 6-hosphate and melibiose were altered in only 2 out of 154 mutantsSupplementary Table S5). In terms of between-metabolite corre-ations, the highest correlation (Pearson’s r 0.922) was observedetween 1,3-bisphosphoglycerate and phosphoenolpyruvatehich are two intermediates of glycolysis. Other strong corre-

ations were mainly exhibited by amino acids and nucleotidesTable 3).

As summarized in Table 1, a total of 27 strain clusters andwo sets of no-cluster were obtained. Particularly we are inter-

sted in independently differential strains (Group 1) and clustershat revealed significantly altered metabolic profiles (Group 2). A

ore detailed discussion on these clusters and differential strainss described in the next section.

averaged over replicate number (3) for each strain, and ≥1.3 is defined as differential.b Based on PCA with eleven significant principal components (SIMCA-P+ ver13,

Umetrics, Umeå, Sweden). Italicized values indicate less than 95% confidence range(<21.85).

c Refers to group designation on Table 1.

90 Z. Hashim et al. / J. Chromatog

Table 3Between-metabolite correlations (≥0.6).

Positive correlation Negative correlation

Metabolites Pearson’s r Metabolites Pearson’s r

1,3-BPG-PEP 0.922 Phenylalanine-NADP −0.653Phosphate–glucose 0.902 Methionine-NADP −0.643F1,6-F2,6P 0.858 Phenylalanine-F2,6P −0.629Methionine–tyrosine 0.838 Methionine-F2,6P −0.608Tyrosine–phenylalanine 0.837 Citrate-PEP −0.601Alanine–glycine 0.813 Phenylalanine-GDP −0.601Glutamate-2-oxoglutarate 0.802Isoleucine–phenylalanine 0.794Methionine–isoleucine 0.776NADP-F2,6P 0.776Isoleucine–tyrosine 0.775Phenylalanine–tryptophan 0.775CTP–GTP 0.770CTP–UTP 0.768Glycine–leucine 0.761Uracil–inositol 0.753UMP-2-isopropylmalate 0.752Methionine–phenylalanine 0.751AMP–ADP 0.7453-Aminoethanol-glucose 0.744Inosine–guanosine 0.731GTP–ATP 0.717R5P–Ru5P 0.713Uracil–melibiose 0.701Asparagine–methionine 0.698Proline–methionine 0.695Alanine–leucine 0.6862-Aminoethanol-phosphate 0.679F6P-F1P 0.678CDP–ADP 0.674Asparagine-tyrosine 0.673Tyrosine-asparagine 0.673Serine–aspartate 0.648Serine–asparagine 0.636Pyruvate–UMP 0.634NAD–ATP 0.633GMP–ADP 0.615Asparagine–aspartate 0.614Tyrosine–tryptophan 0.611Arginine–histidine 0.610GAP–DHAP 0.604UMP–PEP 0.602GMP–AMP 0.602Guanine–guanosine 0.601Methionine–aspartate 0.601

Abbreviations: 1,3BPG: 1,3 bisphosphoglycerate; PEP: phosphoenolpyruvate; F1,6P:fructose 1,6-bisphosphate; F2,6P: fructose 2,6-bisphosphate; R5P: ribose 5-phosphate; Ru5P: ribulose 5-phosphate; F6P: fructose 6-phosphate; F1P: fructose1-phosphate; GAP: glyceraldehyde 3-phosphate; DHAP: dihydroxyacetone phos-phate; NAD: nicotinamide adenine dinucleotide; NADP: nicotinamide adeninedinucleotide phosphate; ATP: adenosine 5′-triphosphate; ADP: adenosine 5′-d ′ ′

G5

3g

ddwwwGasliT

iphosphate; AMP: adenosine 5 -monophosphate; GTP: guanosine 5 -triphosphate;DP: guanosine 5′-diphosphate; GMP: guanosine 5′-monophosphate; CTP: cytidine′-triphosphate; CDP: cytidine 5′-diphosphate; UMP: uridine 5′-monophosphate.

.4. Identification of differential strains and functionally relatedenes

From HCA (Fig. 3 and Table 1), twelve strains were found to haveifferential and unique metabolic profiles. 36 more strains showedifferential profile and formed clusters. HCA result is in accordanceith differential analysis as depicted in Table 2, in which strainsith high average deviation from mutant median and T2 valuesere located at the outer hierarchy of HCA, and thus belong toroup 1 or 2. Within these two groups, six strains (gcr2�, aft1�,da2�, ino2�, ino4�, opi1�) showed growth defect of> 20% and

ix (rpn4�, arg82�, sin3�, swi6�, cbf1�, ada3�) had somewhatower maximum specific growth rate compared to control, suggest-ng that altered metabolism is likely to be related with poor growth.o test whether the metabolic phenotype is a function of growth

r. B 966 (2014) 83–92

rate, we calculated Pearson’s correlation coefficient between max-imum specific growth rate (�) and average deviation from mutantmedian and Hotelling’s T2 range, and between � and each metabo-lite in these twelve strains. As a result, average deviation frommutant median and Hotelling’s T2 were not correlated with growthrate (correlation coefficients were 0.17 and −0.08 respectively),indicating that metabolic profile alteration was directly due tothe gene deletion and not indirectly influenced by poor growth.Almost all metabolites also showed no correlation with maximumspecific growth rate, with the exception of four metabolites i.e.trehalose, glyceraldehyde 3-phosphate (GAP) and its isomer dihy-droxyacetone phosphate (DHAP), and 2-oxoglutarate, in which thecorrelation coefficients were −0.72, 0.72, 0.64, and 0.68 respec-tively.

In this study, we identified arg82� as the most differential strainwith the highest values of average deviation from mutant medianand Hotelling’s T2 (Table 2). Arg82p is an inositol polyphosphatemultikinase (IPMK), a global regulator involved in the regulation ofarginine-, phosphate-, and nitrogen-responsive gene [39]. Argininelevel was increased (FC = 1.5) in arg82�, in agreement with a pre-vious study that reported mutation in ARG82 leads to constitutiveproduction of the arginine biosynthetic enzymes encoded by theARG1, ARG3, ARG5,6 and ARG8 genes [40]. Deletion of ARG82 alsocaused altered levels of various metabolites, mainly amino acidsbut includes TCA cycle intermediates and co-factors such as UTPand ATP. Most apparent was 370-fold change in citrate level. HowArg82p regulates citrate metabolism requires further investiga-tions. More importantly, arg82� showed a great level of metabolicalteration, with an average deviation from mutant median of 3.16vs. 1.85 in the second most differential strain ada2�. Arg82p hasbeen described as a global regulator [40], thus our finding sug-gests that ARG82 may take part in more metabolic regulations thanpreviously reported.

We also observed differential profile in ada2� and ada3�.Ada2p and Ada3p are dual function regulators involved in regu-lation of many other transcription factors, and are a componentof SAGA, SLIK and ADA complexes [39]. We found that ada2�exhibited larger variation in metabolic profile compared to ada3�(average deviation from mutant median of 1.85 and 1.32 respec-tively). While both shared similar metabolic pattern in somemetabolites such as amino adipic acid, inosine, orotate, succinate,malate (increased) and alpha-glycerophosphate, glutathione, CMP,pyruvate, nicotinate (decreased), only ada2� showed a markedchange in citrate, 1,3-bisphosphoglycerate and phosphoenolpyru-vate levels which were not observed in ada3�. Moreover, we didnot find a strong correlation in metabolic profile of ada2� andada3� (correlation coefficient 0.59), suggesting that there are addi-tional pathways exclusive to only one of the genes causing differentdegree and configuration of metabolic regulation.

In addition, hierarchical clustering analysis based onmetabolome dataset allows for the screening of a specific group ofstrains with related functions. On top of the list is ino2�, ino4�,opi1�, ric1� cluster (correlation 0.85). This cluster is characterizedby marked increase in trehalose, succinate and citrate, as wellas decrease in proline, guanosine and 1,3-bisphosphoglycerate(Fig. 4A). Ino2p, Ino4p and Opi1p engage in inositol metabolismand glycerolipids regulation [41–43]. Interestingly, ino2�, ino4�and opi1� all showed growth defect of >20% compared to control,with remarkably high correlation in metabolic profile, whilethe association of Ric1p with these three regulators has not beendescribed yet. Similarly, mks1�, rtg3�, rtg1� and rrn10� formed acluster with high correlation (0.87). This cluster share a similar fea-

ture of decreased amino adipic acid, glutamate and 2-oxoglutarate(Fig. 4B). WhileMks1p, Rtg1p and Rtg3phave been previouslydemonstrated to be involved in mitochondrial retrograde responsein yeast [44–46], there was no report about the involvement of

Z. Hashim et al. / J. Chromatogr. B 966 (2014) 83–92 91

F ters (Ar assocp ofiles

Rfottiowuf

btlisgtptptruobrsmarthcml(ot

4

org

ig. 4. Heat map of metabolite fold-change in two representative differential clusrn10�). For example, in Cluster 1, although Ino2p, Ino4p and Opi1p genes have beenathways has not been reported. Clustering of strains that share similar metabolic pr

rn10p. Our findings demonstrate the possibility of novel geneunction exploration based on metabolic phenotype that cannly be unraveled by metabolomics. Further studies regardinghese uncharacterized roles of transcription factors would be aopic for future studies. Comprehensive discussion on biologicalnterpretations of all differential genes and clusters is not the aimf our current study and thus omitted in this report. Our goalas to construct a yeast metabolome dataset and demonstrate itstility in providing significant correlations which can be practicalor future studies.

Overall, we found more metabolic change in terms of the num-er of correlations and significantly altered metabolites comparedo previous reports [15,16], and thus demonstrated the high reso-ution approach of metabolomics employed in this study. However,t is important to note that transcriptional regulation is condition-pecific and tightly controlled. In this study, only a ‘standard’rowth condition is examined. Therefore, investigation of regula-ions that are activated only during a specific condition should beerformed under a defined experimental set-up and might involveime-series profiling. Moreover, some transcription factors can takeart in multiple pathways, resulting in compounding effects inhe final metabolic profile, thus un-assembling of these pathwaysequires thorough consideration and further experiments. Here, wesed BY4742 derivatives which carry multiple gene deletions. Usef a prototrophic strain collection [47] may yield less biascausedy auxotrophic markers and feasible with minimal media, therebyeducing the confounding factors posed by additional nutrientupplementation. Nonetheless, we demonstrated that useful infor-ation regarding transcriptional regulation based on metabolite

lteration can be obtained from our dataset. For future studies,esearchers can select only cluster of interest and conduct fur-her experiments under a more defined condition based on theints provided by metabolite profile similarity. In this study, onlylusters with high correlation (≥0.85) and significantly differentialetabolic profiles were discussed mainly. Further examination of

ess differential clusters but share highly similar metabolic patterni.e. clusters in Group 3 in Table 1) might also reveal other previ-usly unknown correlations and lead to a better understanding ofranscriptional regulations.

. Conclusions

Metabolome dataset provides useful insights into the effectsf transcription factor deletion toward metabolic pathwayearrangement. We focused on metabolites as the final readouts ofene transcription process to help delineate the complex network

[

[

) Cluster 1 (ino2�, ino4�, opi1�, ric1�) and (B) Cluster 3 (mks1�, rtg1�, rtg3�,iated with inositol and phospholipid regulation, the role of Ric1pin these regulatoryproves to be useful in finding both known and new gene/TF-metabolite correlations.

of transcriptional regulation by first constructing a metabolomelibrary which can be used to infer the regulatory effects of TF.We next showed that significant correlations among TFs and dif-ferential strains can be revealed by hierarchical clustering anddifferential analysis. We obtained both previously reported andpossibly new correlations. We demonstrated how the informa-tion can be used to open the doors to deeper investigation. Ourmetabolome dataset does not only provide information about keymetabolites but also represents a useful resource for future trans-criptional regulation studies.

Acknowledgements

The authors thank Dr. Yukio Mukai from Nagahama Institute ofTechnology, Japan for help with the strains and discussions. Z.H. isfinancially supported by The Hitachi Scholarship Foundation Japan.This study represents a portion of the dissertation submitted by Z.H.to Osaka University in partial fulfillment of the requirements for herPh.D.

Appendix A. Supplementary data

Supplementary data associated with this article can befound, in the online version, at http://dx.doi.org/10.1016/j.jchromb.2014.05.041.

References

[1] H. Alper, J. Moxley, E. Nevoigt, G.R. Fink, G. Stephanopoulos, Science 314 (2006)1565.

[2] H. Alper, G. Stephanopoulos, Metab. Eng. 9 (2007) 258.[3] S. Sommer, S.A. Fuqua, Semin. Cancer Biol. 11 (2001) 339.[4] J.E. Delmore, G.C. Issa, M.E. Lemieux, P.B. Rahl, J. Shi, H.M. Jacobs, E. Kastritis,

T. Gilpatrick, R.M. Paranal, J. Qi, M. Chesi, A.C. Schinzel, M.R. McKeown, T.P.Heffernan, C.R. Vakoc, P.L. Bergsagel, I.M. Ghobrial, P.G. Richardson, R.A. Young,W.C. Hahn, K.C. Anderson, A.L. Kung, J.E. Bradner, C.S. Mitsiades, Cell 146 (2011)904.

[5] P. Blancafort, D.J. Segal, C.F. Barbas, Mol. Pharmacol. 66 (2004) 1361.[6] Y. Shi, Drug Discov. Today 12 (2007) 440.[7] T.I. Lee, N.J. Rinaldi, F. Robert, D.T. Odom, Z. Bar-Joseph, G.K. Gerber, N.M. Han-

nett, C.T. Harbison, C.M. Thompson, I. Simon, J. Zeitlinger, E.G. Jennings, H.L.Murray, D.B. Gordon, B. Ren, J.J. Wyrick, J.-B. Tagne, T.L. Volkert, E. Fraenkel,D.K. Gifford, R.A. Young, Science 298 (2002) 799.

[8] Z. Hu, P.J. Killion, V.R. Iyer, Nat. Genet. 39 (2007) 683.[9] J. Reimand, J.M. Vaquerizas, A.E. Todd, J. Vilo, N.M. Luscombe, Nucl. Acids Res.

38 (2010) 4768.10] W. Zheng, H. Zhao, E. Mancera, L.M. Steinmetz, M. Snyder, Nature 464 (2010)

1187.11] D. Dikicioglu, E. Karabekmez, B. Rash, P. Pir, B. Kirdar, S.G. Oliver, BMC Syst. Biol.

5 (2011) 148.

9 matog

[

[

[

[

[

[

[

[

[

[

[

[[

[[

[

[[[

[

[[[

[

[

[

[

[

[[[[

2 Z. Hashim et al. / J. Chro

12] S. MacArthur, X.-Y. Li, J. Li, J.B. Brown, H.C. Chu, L. Zeng, B.P. Grondona, A. Hech-mer, L. Simirenko, S.V.E. Keränen, D.W. Knowles, M. Stapleton, P. Bickel, M.D.Biggin, M.B. Eisen, Genome Biol. 10 (2009) R80.

13] R. Yoshida, T. Tamura, C. Takaoka, K. Harada, A. Kobayashi, Y. Mukai, E. Fukusaki,Aging Cell 9 (2010) 616.

14] L.M. Raamsdonk, B. Teusink, D. Broadhurst, N. Zhang, A. Hayes, M.C. Walsh, J.a.Berden, K.M. Brindle, D.B. Kell, J.J. Rowland, H.V. Westerhoff, K. van Dam, S.G.Oliver, Nat. Biotechnol. 19 (2001) 45.

15] S.-M. Fendt, A.P. Oliveira, S. Christen, P. Picotti, R.C. Dechant, U. Sauer, Mol. Syst.Biol. 6 (2010).

16] S.J. Cooper, G.L. Finney, S.L. Brown, S.K. Nelson, J. Hesselberth, M.J. MacCoss, S.Fields, Genome Res. 20 (2010) 1288.

17] T. Jewison, C. Knox, V. Neveu, Y. Djoumbou, A.C. Guo, J. Lee, P. Liu, R. Mandal,R. Krishnamurthy, I. Sinelnikov, M. Wilson, D.S. Wishart, Nucl. Acids Res. 40(2012) D815.

18] G. Giaever, A.M. Chu, L. Ni, C. Connelly, L. Riles, S. Véronneau, S. Dow, A.Lucau-Danila, K. Anderson, B. André, A.P. Arkin, A. Astromoff, M. El-Bakkoury,R. Bangham, R. Benito, S. Brachat, S. Campanaro, M. Curtiss, K. Davis, A.Deutschbauer, K.-D. Entian, P. Flaherty, F. Foury, D.J. Garfinkel, M. Gerstein,D. Gotte, U. Güldener, J.H. Hegemann, S. Hempel, Z. Herman, D.F. Jaramillo, D.E.Kelly, S.L. Kelly, P. Kötter, D. LaBonte, D.C. Lamb, N. Lan, H. Liang, H. Liao, L. Liu,C. Luo, M. Lussier, R. Mao, P. Menard, S.L. Ooi, J.L. Revuelta, C.J. Roberts, M. Rose,P. Ross-Macdonald, B. Scherens, G. Schimmack, B. Shafer, D.D. Shoemaker, S.Sookhai-Mahadeo, R.K. Storms, J.N. Strathern, G. Valle, M. Voet, G. Volckaert,C. Wang, T.R. Ward, J. Wilhelmy, E.A. Winzeler, Y. Yang, G. Yen, E. Youngman,K. Yu, H. Bussey, J.D. Boeke, M. Snyder, P. Philippsen, R.W. Davis, M. Johnston,Nature 418 (2002) 387.

19] C. Kaiser, S. Michaelis, A. Mitchell, Methods in Yeast Genetics: A Cold SpringHarbor Laboratory Course Manual, 1994 ed., Cold Spring Harbor, NY, USA,1994.

20] C.A. Crutchfield, W. Lu, E. Melamud, J.D. Rabinowitz, Methods Enzymol. 470(2010) 393.

21] U. Roessner, C. Wagner, J. Kopka, R.N. Trethewey, L. Willmitzer, Plant J. 23 (2000)131.

22] L. Coulier, R. Bas, S. Jespersen, E. Verheij, M.J. van der Werf, T. Hankemeier, Anal.Chem. 78 (2006) 6573.

23] W. Lu, B.D. Bennett, J.D. Rabinowitz, J. Chromatogr. B 871 (2008) 236.24] B. Luo, K. Groenke, R. Takors, C. Wandrey, M. Oldiges, J. Chromatogr. A 1147

(2007) 153.

[[[[

r. B 966 (2014) 83–92

25] J.M. Buescher, S. Moco, U. Sauer, N. Zamboni, Anal. Chem. 82 (2010) 4403.26] H.G. Gika, G.A. Theodoridis, J.E. Wingate, I.D. Wilson, J. Proteome Res. 6 (2007)

3291.27] H.G. Gika, E. Macpherson, G.A. Theodoridis, I.D. Wilson, J. Chromatogr. B 871

(2008) 299.28] H.G. Gika, G.A. Theodoridis, M. Earll, I.D. Wilson, Bioanalysis 4 (2012) 2239.29] A. Lommen, Anal. Chem. 81 (2009) 3079.30] H. Tsugawa, Y. Tsujimoto, M. Arita, T. Bamba, E. Fukusaki, BMC Bioinform. 12

(2011) 131.31] J. Xia, R. Mandal, I.V. Sinelnikov, D. Broadhurst, D.S. Wishart, Nucl. Acids Res.

40 (2012) W127.32] M.J.L. de Hoon, S. Imoto, J. Nolan, S. Miyano, Bioinformatics 20 (2004) 1453.33] A.J. Saldanha, Bioinformatics 20 (2004) 3246.34] M.A. Kamleh, T.M.D. Ebbels, K. Spagou, P. Masson, E.J. Want, Anal. Chem. 84

(2012) 2670.35] W.B. Dunn, D. Broadhurst, P. Begley, E. Zelena, S. Francis-McIntyre, N. Anderson,

M. Brown, J.D. Knowles, A. Halsall, J.N. Haselden, A.W. Nicholls, I.D. Wilson, D.B.Kell, R. Goodacre, Nat. Protoc. 6 (2011) 1060.

36] W. Lu, M.F. Clasquin, E. Melamud, D. Amador-Noguez, A.a. Caudy, J.D. Rabi-nowitz, Anal. Chem. 82 (2010) 3212.

37] E. Zelena, W.B. Dunn, D. Broadhurst, S. Francis-McIntyre, K.M. Carroll, P. Begley,S. O’Hagan, J.D. Knowles, A. Halsall, I.D. Wilson, D.B. Kell, Anal. Chem. 81 (2009)1357.

38] S. Sharma, Applied Multivariate Techniques, John Wiley & Sons, Inc., New York,NY, USA, 1996.

39] J.M. Cherry, E.L. Hong, C. Amundsen, R. Balakrishnan, G. Binkley, E.T. Chan, K.R.Christie, M.C. Costanzo, S.S. Dwight, S.R. Engel, D.G. Fisk, J.E. Hirschman, B.C.Hitz, K. Karra, C.J. Krieger, S.R. Miyasato, R.S. Nash, J. Park, M.S. Skrzypek, M.Simison, S. Weng, E.D. Wong, Nucl. Acids Res. 40 (2012) D700.

40] F. Messenguy, E. Dubois, Mol. Cell. Biol. 13 (1993) 2586.41] J.A. Graves, S.A. Henry, Genetics 154 (2000) 1485.42] S.A. Henry, S.D. Kohlwein, G.M. Carman, Genetics 190 (2012) 317.43] J. Hoppen, A. Repenning, A. Albrecht, S. Geburtig, H.-J. Schüller, Yeast 22 (2005)

601.

44] I. Dilova, S. Aronova, J.C.-Y. Chen, T. Powers, J. Biol. Chem. 279 (2004) 46527.45] R.A. Butow, N.G. Avadhani, Mol. Cell 14 (2004) 1.46] J.J. Tate, K.H. Cox, R. Rai, T.G. Cooper, J. Biol. Chem. 277 (2002) 20477.47] M. Mülleder, F. Capuano, P. Pir, S. Christen, U. Sauer, S.G. Oliver, M. Ralser, Nat.

Biotechnol. 30 (2012) 1176.