Development of a gas chromatography/mass spectrometry based metabolomics protocol by means of statistical experimental design

ORIGINAL ARTICLE

Development of a gas chromatography/mass spectrometry basedmetabolomics protocol by means of statistical experimental design

Anders P. H. Danielsson • Thomas Moritz •

Hindrik Mulder • Peter Spegel

Received: 11 August 2010 / Accepted: 28 January 2011 / Published online: 11 February 2011

� Springer Science+Business Media, LLC 2011

Abstract Metabolomics is a growing research field

where new protocols are rapidly developed and new

applications discovered. Common applications include

biomarker discovery and elucidation of drug metabolism.

However, the development of such protocols rarely

includes a systematic optimization followed by validation

with real samples. Here a GC/MS-based protocol using

methoximation followed by silylation with N-tert-butyldi-

methylsilyl-N-methyltrifluoroacetamide (MTBSTFA) for

analysis of blood plasma metabolites is thoroughly devel-

oped and optimized from derivatization to detection with

statistical design of experiments (DOE). Validation was

performed with blood plasma samples and proved the

methodology to be efficient, rapid and reliable with a total

of 51 analyses performed in 24 h, with linear responses,

low detection limits and good precision. The obtained

chromatograms were much cleaner, due to the absence of

glucose overloading, and the data was found to drift less

with MTBSTFA derivatisation than with MTBSTFA

derivatisation.

Keywords Metabolomics � Gas chromatography �Mass spectrometry � MTBSTFA � Blood plasma � DOE

1 Introduction

Metabolomics aims for an unbiased quantification of all

metabolites in a biological sample. As the metabolite levels

reflects both genetic and environmental effects, metabolo-

mics is an exceptionally useful technique for discovery of

potential biomarkers for environmental factors, medical

treatments, and diseases (Chorell et al. 2009; Fiehn 2002).

However, the metabolome comprises a very large and

heterogeneous group of metabolites with concentrations

differing several orders of magnitude. Thus, great care

must be taken in the development of metabolomics meth-

ods to attain a wide selectivity, high efficiency and high

sample throughput while at the same time avoid introduc-

ing excessive biases.

To date, no single analytical method exists that is

capable of simultaneously measuring all members of the

metabolome. Several techniques, including nuclear mag-

netic resonance spectroscopy (NMR) (Zhang et al. 2008),

near-infrared spectroscopy (NIR) (Cozzolino et al. 2006),

gas chromatography (GC) (Fiehn 2008; Jiye et al. 2005),

liquid chromatography (LC) (Zelena et al. 2009), and

capillary electrophoresis (CE) (Lapainis et al. 2009), with

the latter three being coupled to mass spectrometry (MS),

have been applied in metabolomics. Out of these tech-

niques, the combination of a chromatographic separation

with mass spectrometric detection offers a somewhat

higher sensitivity than the pure spectroscopic techniques,

although sample preparation generally becomes more

complicated. Among these techniques, GC/MS, applied in

the present investigation, offers the highest separation

Electronic supplementary material The online version of thisarticle (doi:10.1007/s11306-011-0283-6) contains supplementarymaterial, which is available to authorized users.

A. P. H. Danielsson � H. Mulder � P. Spegel (&)

Unit of Molecular Metabolism, CRC 91:12, Entrance 72,

UMAS, Lund University Diabetes Centre, Clinical Research

Centre, SE-205 02 Malmo, Sweden

e-mail: [email protected]

A. P. H. Danielsson

Analytical Chemistry, Faculty of Engineering LTH,

Lund University, Lund, Sweden

T. Moritz

Umea Plant Science Center, Swedish University of Agricultural

Sciences, Umea, Sweden

123

Metabolomics (2012) 8:50–63

DOI 10.1007/s11306-011-0283-6

efficiency and also yields the cleanest chromatograms due

to the absence of interfering peaks from peptides, proteins

and other large non-volatile substances.

Metabolite derivatisation is an important element of GC/

MS-based metabolomics. Great care must be taken in the

development of derivatisation conditions, as it will influ-

ence both the limit of detection (LOD), the sensitivity, and

the selectivity of the method. The most common derivati-

zation procedure applied to GC/MS-based metabolomics

involves a two-step reaction, comprising a methoximation

step and a silylation step (Gullberg et al. 2004). Methoxi-

mation reduces the number of sugar tautomers, thus

reducing the number of peaks generated from a single

sugar metabolite and thereby enhances both separation and

quantification (Asres and Perreault 1997; Fiehn et al. 2000;

Schweer 1982). Additionally, methoximation protects

a-ketoacids from decarboxylation (Fiehn et al. 2000;

Tam and Normanly 1998). The silylation reagent most

commonly applied in metabolomics, N-methyl-N-tri-

methylsilyltrifluoroacetamide (MSTFA), yields volatile

trimethylsilyl (TMS) derivatives of a vast number of

metabolites including large multifunctional metabolites

such as sugars and their derivatives (Begley et al. 2009;

Danielsson et al. 2010; Gullberg et al. 2004). Besides

MSTFA, N-tert-butyldimethylsilyl-N-methyltrifluoroaceta-

mide (MTBSTFA), has also been applied in metabolomics.

The large size of this silylation reagent reduces both the

yield and the volatility of glucose resulting in a complete

elimination of glucose and other large carbohydrates from

the analysis. In blood plasma analysis, glucose is found at

high concentrations which generally severely overloads the

column and complicating detection and quantification of

other metabolites eluted in the same retention interval.

Furthermore, the tert-butyldimethylsilyl (TBDMS) deriv-

atives have been shown to offer both improved hydrolytic

stability and improved thermal stability over the corre-

sponding TMS derivatives (Rodrıguez et al. 2003; Yu et al.

2007) and generate an intense [M-57]? mass fragment

(Fiehn et al. 2000) providing additional information for

identification.

Although MTBSTFA derivatization has several unique

features in comparison to MSTFA derivatization in meta-

bolomics analysis, the derivatization and chromatographic

parameters for this derivatization reagent have not yet been

thoroughly optimized. In contrast, there are several devel-

oped metabolomics protocols based on MSTFA (Begley

et al. 2009; Danielsson et al. 2010; Gullberg et al. 2004;

Jiye et al. 2005). These protocols are aimed at plant cells

(Gullberg et al. 2004), adherent cell cultures (Danielsson

et al., 2010) and blood plasma (Begley et al. 2009; Jiye

et al. 2005).

In the present investigation, a metabolomics protocol for

blood plasma analysis based on methoximation and

MTBSTFA derivatization followed by GC/MS is devel-

oped and optimized using statistical design of experiments

(DOE) (Araujo and Brereton 1996a, b, c). DOE is a multi-

variate statistical optimization tool that enables efficient

experimental planning and identification of both linear

effects, interaction effects and higher order non-linear

effects with a minimum number of experiments. The two-

step derivatization method, injection onto the GC, the

chromatographic settings and the mass spectrometer set-

tings were optimized aiming at enhancing the limit of

detection (LOD), the sample throughput and the peak

capacity. Finally, the developed method was applied to a

set of blood plasma samples, and the performance and

validity of the developed protocol was assessed.

2 Experimental

2.1 Chemicals

Methoxyamine hydrochloride, N-tert-butyldimethylsilyl-

N-methyltrifluoroacetamide and N-methyl-N-trimethylsi-

lyltrifluoroacetamide were from Aldrich (Steinhein,

Germany). Pyridine, heptane, methanol, methyl stearate and

the alkane standard mixtures (C8–C20 and C21–C40) were

from Fluka (Buchs, Switzerland). The stable isotope-labeled

internal standards; 13C3–15N alanine, 13C4-succinate, 13C6–

phenylalanine, 13C3-serine, 2H7-cholesterol, 13C16-palmitate

and 13C5-a-ketoisovalerate were from Cambridge Isotope

Laboratories, Inc. (Andover, MA). The stable isotope-

labeled standards; 13C9–15N-tyrosine and 13C18-oleate were

from Isotec (Sigma-Aldrich, St. Louis, MA). Water was

purified using a Purelab Ultra water purification system

(Elga, Gothenburg, Sweden). The metabolites used in the

study were all purchased from Sigma-Aldrich.

2.1.1 Blood samples

The subjects, four males and two females, were all Cauca-

sian and non-obese (body mass index 22.4 ± 2.4 ranging

from 18 to 24). They were 33.8 ± 6.9 (ranging from 28 to

46) years of age, non-smokers and all healthy. None were

taking any prescribed medication. Blood samples were

drawn after an overnight fast and the subjects were all sitting

down during the whole procedure. Blood samples were

immediately centrifuged at ?4�C and plasma was separated.

The blood plasma was subsequently pooled and frozen on

dry ice. The samples were kept at -80�C until analysis.

2.2 Instrumentation

GC/MS was performed on an Agilent 6890N gas chro-

matograph (Agilent, Atlanta, GA) equipped with an

Development of a Metabolomic Protocol 51

123

Agilent 7683B autosampler (Agilent) and coupled to a

Leco Pegasus III TOFMS electron impact, time-of-flight

mass spectrometer (Leco Corp.). Two columns were used

in the study, a 10 m (ID 180 lm, phase thickness 0.18 lm)

and a 30 m (ID 250 lm, phase thickness 0.25 lm), both

DB5-MS (J&W Scientific, Folsom, CA). The initial

method employed a splitless injection with the injector

temperature set to 270�C with a purge vent time of 60 s.

The flow rate through the column was 1 ml/min, with a

three-step temperature program starting with an initial

isocratic temperature of 70�C kept for 2 min, followed by a

gradient rate of 30�C/min reaching a final temperature of

320�C kept for a duration of 5 min. The ionization voltage

was set at 70 eV and the mass spectra were recorded

between 50 and 800 m/z, throughout the study. Data were

acquired using the Leco ChromaTof software v. 3.31 (Leco

Corp.). Retention indexes (RIs) were calculated based

on the elution times of a homologous series of n-alkanes

(C8-C40).

2.3 Metabolite cocktail

36 metabolite standards were derivatized and analyzed

individually to construct a database (NIST MS Search 2.0)

containing retention indexes and mass spectra. The

metabolites were selected to cover a broad range of func-

tionalities, metabolic pathways, and retention indexes.

Next, a polar and a non-polar model mixture were created

from the 36 metabolite standards, isotope labeled internal

standards and methyl stearate. The combination of both

these standard mixtures is herein referred to as the

metabolite cocktail (Table 1) and is used throughout the

model optimizations conducted in this paper.

2.4 Raw data pre-treatment and experimental designs

Peak integration was performed either directly in Leco

ChromaTof 3.31 or after export as NetCDF files to

MATLAB 7.0 (Mathworks, Natich, MA) using a hierar-

chical multivariate curve resolution (H-MCR) script

(Jonsson et al. 2006). Peak identification was performed

with NIST MS Search 2.0 using mass spectra and retention

index from several libraries. For the TBDMS derivatives

the database constructed here was used and all metabolites

were quantified using only their [M-57]? mass fragment

(Table 1). For the TMS derivatives, the NIST mass spectra

library, a library developed at the Max Planck Institute in

Golm (http://csbdb.mpimp-golm.mpg.de/), and libraries

developed in house, both at Umea Plant Science Centre

(UPSC) and Lund University Diabetes Centre (LUDC)

were used.

Experimental designs were created in ModdeTM 8.01

(Umetrics, Umea, Sweden). All responses were centered

and scaled to unit variance (UV) and projections to latent

structures (PLS) were used to calculate the models. Prior to

evaluation, the models were optimized in a five-step

procedure aimed at reaching the highest possible cross-

validated predictive power (Q2Y), which compares the

cross-validated residuals to the total residual for the model

(Wold 1978). In the first step, normalization, when appli-

cable, of the responses was performed. In the second step,

responses deviating from the normal distribution were

transformed, to avoid erroneous influence on the model. In

the third step, responses with a replicate variance larger

than half of the total variance of the response were con-

sidered irreproducible and were therefore deleted. In the

fourth step responses exhibiting considerable lack-of-fit,

i.e. responses where the model error variance was equal or

larger than the replicate variance, were deleted. In the fifth

and final step, linear factors, factor interactions and qua-

dratic factors which reduced Q2Y were identified. These

factors and interactions that mainly modeled noise were

considered insignificant for the model and consequently

deleted. The models were evaluated from coefficient and

contour plots.

Orthogonal projections to latent structures (OPLS) was

performed on centered and UV-scaled data in Simca P?

12.0 (Umetrics, Umea, Sweden).

2.5 Derivatization design

A D-optimal design with a quadratic model consisting of 8

factors was constructed to explore the MTBSTFA deriva-

tization conditions using the metabolite cocktail as sample

(Table 2). Specifically, the temperature, duration and sol-

vent composition used for both methoximation and silyla-

tion were varied. The aim was to obtain the maximum yield

of metabolite derivatives, and minimum number of arti-

facts in the shortest possible derivatization time.

In total 45 runs were performed including three center

point runs. The responses were the peak areas for the

metabolites in the metabolite cocktail, normalized to the

peak area of the underivatizable methyl stearate, and

the number of artifact peaks detected by Leco ChromaTof

in the chromatograms. The artifact peaks were quantified as

the number of non-metabolite peaks between the first and

last eluted metabolite.

2.6 Injection design

A three-level full factorial design with a quadratic model

was generated to find the optimum settings for the injector

using the metabolite cocktail as sample (Table 2). The

design included 2 factors and the responses consisted of the

peak areas of the metabolites in the cocktail. The injection

volume was left unchanged at 1 ll throughout the study, to

52 A. P. H. Danielsson et al.

123

Table 1 Metabolite and isotope

labeled standard derivatives

investigated in the present study

with their [M-57]? fragments

and retention indexes (RIs)

a Underivatizable standardb Quantification mass

Analyte Abbreviation RI [M-57]?

Pyruvate MEOX TBDMS PYR 1254.7 174

2-ketobutyrate MEOX TBDMS 2KBA 1305.5 18813C5-a-ketoisovalerate MEOX TBDMS peak1 AKIX1 1321.6 20713C5-a-ketoisovalerate MEOX TBDMS peak2 AKIX2 1336 207

a-ketoisovalerate MEOX TBDMS peak1 AKI1 1321.6 202

a-ketoisovalerate MEOX TBDMS peak 2 AKI2 1336 202

Lactate 2TBDMS LAC 1493.1 261

Alanine 2TBDMS ALA 1541.4 26013C3-15N alanine 2TBDMS ALAX 1541.5 264

Glycine 2TBDMS GLY 1562.1 246

Valine 2TBDMS VAL 1662.4 288

Leucine 2TBDMS LEU 1702.1 302

Isoleucine 2TBDMS ILE 1735.7 302

Threonine 2TBDMS THR2 1747.5 29013C4-succinate 2TBDMS SUCX 1761.1 293

Succinate 2TBDMS SUC 1761.1 289

Proline 2TBDMS PRO 1775.6 286

Undecanoate TBDMS FA11 1783.8 243

Dodecanoate TBDMS (Laurate) FA12 1884.3 257

Trans-4-hydroxyproline 2TBDMS T4HP2 1970.1 302

Methionine 2TBDMS MET 1976.2 320

Serine 2TBDMS SER 1998.5 390

Alpha ketoglutarate MEOX 2TBDMS AKG 2011.7 346

Threonine 3TBDMS THR3 2032.4 404

Phenylalanine 2TBDMS PHE 2103.5 336

Malate 3TBDMS MAL 2117.8 419

Methyl Stearatea MEST 2134.5 298b

Trans-4-hydroxyproline 3TBDMS T4HP3 2196 416

Cysteine 3TBDMS CYS 2217.1 406

Phosphoenolpyruvate 3TBDMS PEP 2239.6 453

Hexadecanoate TBDMS (Palmitate) FA16 2288.3 313

Heptadecanoate TBDMS (Margarate) FA17 2389.5 325

Glyceraldehyde 3-phosphate MEOX 3TBDMS GA3P 2346.5 484

Dihydroxyacetone phosphate MEOX 3TBDMS peak1 DHAP1 2370.2 484

Dihydroxyacetone phosphate MEOX 3TBDMS peak2 DHAP2 2390.2 484

Lysine 3TBDMS LYS 2390.4 431

(9Z)-Octadec-9-enoate TBDMS (Oleate) FA18U 2468.6 339

Histidine 3TBDMS HIS 2609.6 440

Citrate 4TBDMS CIT 2632.2 591

Isocitrate 4TBDMS ISOC 2647.2 591

3-Phosphoglycerate 4TBDMS 3PGA 2647.6 585

Tryptophan 2TBDMS TRP 2708.9 375

Serotonin 2TBDMS SERO 2720.5 347

N-Acetyl 5-hydroxytryptamine TBDMS NAHT 2775.7 275

Tricosanoate TBDMS FA23 2999.2 411

L-5 Hydroxytryptophan 3TBDMS L5HT 3187.6 505

Cholesterol TBDMS CHO 3493.8 443

Development of a Metabolomic Protocol 53

123

minimize liner contamination that is expected from

excessive injection of non-metabolite matrix from complex

biological samples.

2.7 Gas chromatography design

A three-level central composite design (CCF) with a qua-

dratic model was calculated for each column. The two

models were then merged to facilitate the interpretation.

For each column 5 factors and 8 responses were investi-

gated using the metabolite cocktail as sample (Table 2).

The aim was to optimize the separation efficiency, peak

symmetry, detection limits and analysis time. To achieve

this, all parameters of the temperature program were

optimized, including the initial temperature and its duration

and the temperature gradient rate. The final temperature

and its isocratic duration were maintained at 320�C and

5 min, respectively, due to their insignificant influence on

the selected responses. In addition, also the volumetric flow

rate was varied to minimize band broadening and hence

optimize the separation efficiency.

The injector purge vent time was included also in this

design to allow for the investigation of sample loading

effects on the chromatographic performance.

To obtain a representative response for the separation

efficiency, the peak capacity was calculated from the peak

width at 10% of the peak height (w0.1) for three analytes

spanning a broad retention window, valine 2TBDMS,

Table 2 Summary of the models for derivatization and GC/MS optimization

Study Design and model Factors Low High Responses

Methoximation D-optimal design, quadratic model Amount Acetonitrile

Methoximation (MAC)a0% 25% Peak areas

Artifact peaks

Methoximation Temperature

(MTE)

20�C 80�C

Methoximation Duration

(MDU)

1 h 17 h

Silylation Amount Heptane in Silylation

(SHP)b0% 75%

Amount Acetonitrile in

Silylation (SAC)b0% 75%

Amount MTBSTFA in

Silylation (STB)b25% 100%

Silylation Duration (SDU) 0.5 h 4 h

Silylation Temperature (STE) 20�C 100�C

Injection Three-level full factorial design,

quadratic model

Injector Temperature (IJT) 200�C 320�C Peak areas

Injector Purge Vent Time

(PVT)

5 s 115 s

Chromatography Central composite face (CCF) design,

quadratic model

Injector Purge vent time (PVT) 45 s 115 s Peak capacity for valine (PxVc)

Initial Gradient Temperature

Duration (ITD)

2 min 6 min Peak capacity for cholesterol

(PxCc)

Initial Gradient Temperature

(ITE)

60�C 90�C Peak capacity for (9Z)-Octadec-

9-enoate (PxOc)

Temperature Gradient Rate

(TGR)

10�C/

min

40�C/

min

Average peak capacity (PxMc)

Volumetric Flow Rate (VFR) 1 ml/

min

3 ml/

min

Total analysis time (Txc)

Sample throughput per 24 h

(STxc)

Asymmetry factor for valine

(AxVc)

Peak height for cholesterol

(HxC)c

Mass

spectrometry

Three-level full factorial design,

quadratic model

Data Acquisition rate (ACQ) 10 Hz 50 Hz [M-57] ? Peak area

Ion source temperature (IST) 130�C/

min

250�C/

min

Ratio of low and high m/z-

fragment

a Pyridine is used as the additional solvent in the methoximation stepb Formulation factorc x refers to the column length and has a value of either 10 or 30 m

54 A. P. H. Danielsson et al.

123

oleate TBDMS and cholesterol TBDMS, representing the

early, intermediate and late portion of the chromatogram,

respectively, according to Eq. 1. (Pous-Torres et al. 2008)

Pc ¼ 1þtrðchoÞ � trðvalÞ

w0:1ð1Þ

The elution times for cholesterol TBDMS (tr(chol)) and

valine 2TBDMS (tr(val)) defined the retention window.

Additionally an average peak capacity was calculated.

The asymmetry factor (Kirkland, 1977) (As) for valine

2TBDMS was included as a response as this peak was

observed to suffer from asymmetric band broadening,

whereas this type of peak distortion was weaker or absent

for later eluting metabolites. As was determined according

to Eq. 2,

As ¼B10%

A10%ð2Þ

where A10% and B10% are the distances at 10% of the peak

height, measured from a line perpendicular to the baseline

from the peak apex, to the peak front and back respectively.

The noise level was not found to be significantly

affected by the parameter settings and therefore the peak

height of cholesterol TBDMS was included as a response

reflecting the method detection limit. Cholesterol TBDMS

was chosen as it is the last eluted metabolite derivative in

our cocktail and should therefore be most affected by the

longitudinal band broadening which is of primary concern

in GC separations.

2.8 Mass spectrometer design

A three-level full factorial design with a quadratic model

was generated to find the optimal settings of the mass

spectrometer using the metabolite cocktail as sample

(Table 2). The responses studied were the area of the

[M-57]? fragment peak and the ratio between the [M-57]?

fragment peak area and a low weight high intensity frag-

ment peak area. Peak areas of the [M-57]? response were

normalized using the [M-57]? peak area showing a close to

zero reproducibility as it was assumed that for such a peak

the instrumental variability considerably overpowered the

systematic variation caused by the varied factors.

2.9 Method validation

Pooled blood plasma from 6 healthy individuals was

diluted to relative concentrations of 0.1, 0.2, 0.4, 0.6, 0.8

and 1.0 (v/v, plasma/plasma ? water). A total of 12 mea-

surements from four preparations measured in triplicates

were performed at each concentration to assess both

instrumental and preparation variability. The total volume

of diluted plasma in each extraction tube was 100 ll. Prior

to extraction 3.75 lg of each stable isotope labeled stan-

dard were added to each of the 24 tubes. Extraction of the

metabolites was performed according to a previously

developed protocol (Jiye et al. 2005). To each tube 900 ll

methanol/water mixture (8:1 v/v) was added and the sam-

ples were rapidly mixed and kept on ice for 10 min. All

tubes were then vigorously extracted using a multitube

vortexer (VX-2500 Multi Tube Vortexer, VWR, West

Chester, PA) operating at 30 Hz for 3 min. Subsequent the

extraction tubes were centrifuged for 10 min at 175309g at

4�C. 200 ll of the supernatant was then transferred to a GC

vial and evaporated to dryness (miVac Duo concentrator;

Genevac, Ipswich, UK). Samples were derivatized and

analyzed with the optimized protocol developed here.

The results were evaluated with regards to linear range,

intra-day precision, limit of detection (LOD) and limit of

quantification (LOQ). The peak areas were normalized to

the peak areas of the corresponding stable isotope labeled

standard, if available. If no corresponding stable isotope

labeled standard was available, a stable isotope-labeled

standard with similar properties was selected (only per-

formed for determination of linear range and precision).

From the normalized peak areas, a linear model was cre-

ated and analyzed in Modde 8.0.1. The models were tested

for lack-of-fit and also cross-validated to thoroughly

determine the linear range. The precision was calculated as

the average relative standard deviation (RSD) for 10

sample runs at 2 different concentrations; 0.1 and 0.4 (v/v,

plasma/plasma ? water).

The LOD and LOQ were estimated as 3 and 10 times,

respectively, the standard deviation for the signal-to-noise

ratio (S/N) for a 10 times diluted sample of pooled blood

plasma.

2.10 Sample drift comparison

Unnormalized data from the method validation runs were

further analyzed using OPLS to relate the peak area to the

sample run order. An OPLS model was accordingly created

for blood plasma samples derivatized with MSTFA

according to a protocol developed elsewhere (Jiye et al.

2005) aiming at comparing the drift pattern using these two

derivatization protocols.

3 Results and discussion

3.1 Derivatization

Initially, a PLS model was calculated from the normalized

peak areas of all metabolites in the cocktail. In the loading

plot the metabolites were found to cluster roughly

Development of a Metabolomic Protocol 55

123

according to their chemical functionality. However, the

model predictive power, Q2Y, was poor which probably

was due to the heterogeneity of the responses. Guided by

the chemical functionality of the metabolites and their

positions in the loading plot, groups of metabolites show-

ing similar behavior were constructed and modeled. All

models are summarized in Supplementary material S1.

From the linear effects of the three studied methoxi-

mation step factors, it was obvious that the settings of the

methoximation duration and temperature had the largest

influence on the derivatization yield (Supplementary

material S1). For most metabolites, the methoximation

duration and temperature were negatively correlated to the

peak areas. Furthermore, the methoximation temperature

was positively correlated to the number of artifact peaks.

Generally, addition of acetonitrile to the methoximation

solution did not improve the yield.

Next, the linear effects in the silylation step were

investigated. It was found that the most influential factors

were the silylation duration and temperature. The silylation

temperature was positively correlated with 3-phospho-

glycerate 4TBDMS and several amine containing analytes,

whereas it was negatively correlated with succinate

2TBDMS, containing two carboxylic acid functionalities.

The silylation duration was positively correlated with

cysteine 4TBDMS, cholesterol TBDMS and serotonin

2TBDMS but was negatively correlated with isocitrate

4TBDMS, lactate 2TBDMS, succinate 2TBDMS and fatty

acid TBDMS esters. Thus, a low silylation time and tem-

perature was beneficial for carboxylic acids, whereas the

opposite improves the yield of amines. The composition of

the silylation solvent was generally insignificant.

Next, the contour plots were evaluated to characterize the

effects of significant factor interactions and quadratic

effects. For the modifications of the methoximation solvent,

the a-keto acids pyruvate MEOX TBDMS, 2-ketobutyrate

MEOX TBDMS and a-ketoglutarate MEOX 2TBDMS

showed a negative correlation to the interaction between the

added amount of acetonitrile and the methoximation tem-

perature. A moderate temperature of 40–55�C and no ace-

tonitrile was optimal. Interestingly, the amount of

acetonitrile in the methoximation solvent also showed a

dependence on the silylation temperature for both leucine

2TBDMS and the number of artifact peaks. The yield for

leucine 2TBDMS and the number of artifacts increased with

a low fraction of acetonitrile and a high silylation temper-

ature. Only trans-4-hydroxyproline 3TBDMS benefitted

from a high fraction acetonitrile. Although there are a few

benefits of using more complex reaction mixtures, the

absence of strong effects motivates their removal as a much

simplified method will be the result. Therefore, the amount

of pyridine in the methoximation step and the amount of

MTBSTFA in the silylation step was set to 100%.

The interaction between the methoximation time and

temperature was significant for numerous metabolites and

the contour plots generally showed that the yield was

Fig. 1 Contour plots describing

the effect of methoximation

temperature and duration on the

normalized peak areas (as

numerical values in the plots)

for (a) 2-ketobutyrate MEOX

TBDMS, (b) a-keto isovalerate

MEOX TBDMS,

(c) phosphoenolpyruvate

3TBDMS, and (d) histidine

3TBDMS. These contour plots

clearly illustrate the complexity

involved in optimizing a

derivatization method for

metabolome analyses

56 A. P. H. Danielsson et al.

123

increased with settings in the lower range for both factors.

However, there were some exceptions. As noted above the

a-keto acids are sensitive to the settings of methoximation

temperature and only a-ketoisovalerate MEOX TBDMS

benefited from a lower setting while the rest in this group

had local maxima for the methoximation temperature in

the range of 35–70�C (Fig. 1a, b). Phosphoenolpyruvate

3TBDMS followed the general trend with a substantial

decrease in peak area with higher settings of the methox-

imation time and temperature (Fig. 1c). Even so, the

methoximation temperature was set to room temperature,

approximately 20�C, which was optimal for most metab-

olites and resulted in only moderate deviations from the

optimum yield for some of the a-keto acids.

With respect to the methoximation settings, the majority

of metabolites benefited from a short duration, whereas two

amino acids with amine-containing side chains, histidine

3TBDMS (Fig 1d) and tryptophan 2TBDMS, were found

to have maxima between 5 and 10 h. The settings of the

methoximation step has previously been shown to affect

also metabolites lacking aldehyde or ketone functionality.

(Gullberg et al. 2004) Speculatively, this effect may be

related to the solubility of the amino acids in the deriva-

tization solvents and reagents. The methoximation duration

was set to 4 h, which allowed for a sufficient yield also for

histidine 3TBDMS and tryptophan 2TBDMS.

Evaluation of the interactions of the silylation temper-

ature with other factors supports the employment of a high

silylation temperature. This factor was therefore set to

100�C, which is within the range of previous studies that

have reported 60–120�C (Birkemeyer et al. 2003; Buscher

et al. 2009; Ewald et al. 2009; Fiehn et al. 2000). The

silylation duration was set to 170 min, which allowed for

sufficient yield of amines, carboxylic acids and cholesterol

TBDMS.

Quantification of metabolites yielding several different

derivatives is difficult as the derivatives may have very

different properties and the ratio between derivatives may

vary with time (Kanani and Klapa 2007). However, of all

investigated metabolites, only two generates multiple

derivatives in the form of unsilylated and monosilylated

amines. The levels of the unsilylated amines were very low

and likely not affecting the quantification of these amino

acids to any greater extent. For the same reason, these

derivatives could not be accurately modeled due to their

low signal-to-noise ratios. With MSTFA, monosilylated

and disilylated amines are commonly observed. The partial

derivatization of threonine and trans-4-hydroxyproline

observed in the present investigation is probably caused by

steric hindrance from hydroxyl bound TBDMS which

decreases the yield of derivatization of the amine group.

The present models support this order of derivatization,

with carboxylic acids generally reacting fastest, followed

by hydroxyls and amines. Thus, in comparison with TMS

derivatization, TBDMS derivatization may reduce both

the formation of multiple derivatives and reduce the

-0,8

-0,6

-0,4

-0,2

-0,0

0,2

0,4

0,6

0,8

-0,2 0,0 0,2 0,4 0,6 0,8 1,0

wc[

2]

wc[1]

IJT

PVT

PVT*PVT

IJT*PVT

Fig. 2 PLS loading plot from

the injection study, showing the

influence of the factors (filledtriangle) and factor interactions

(filled diamond) on the

metabolite derivatives peak

areas (filled circle). The left

contour plot shows the behavior

of the thermally unstable

phosphoenolpyruvate 3TBDMS

and the right contour plot shows

the behavior of the late eluting

cholesterol TBDMS.

PVT = purge vent time and

IJT = injector temperature

Development of a Metabolomic Protocol 57

123

desilylation rate, due to its higher resistance to hydrolysis

(Rodrıguez et al. 2003; Yu et al. 2007).

3.2 Injection

A PLS model was calculated using all metabolite peak

areas as responses. The model yielded a total explained

variation of 95% (R2Y = 0.954) and a cross-validated

predictive power of 93% (Q2Y = 0.934) with 0.846 B

R2Y B 0.982 and 0.933 B Q2Y B 0.967 for all individual

responses (Supplementary material S2).

The loading scatter plot for the first two PLS compo-

nents (Fig. 2), the contour plots and the coefficient plot for

each of the analytes showed that, for all analytes the peak

area was positively correlated with the purge vent time. At

this point, an injector purge vent time of 115 s was chosen

as optimum and possible column overloading effects will

be assessed in the optimization of the GC-settings.

The effect of increased injector temperature was for

most analytes positively correlated to the yield and there

was also a clear trend towards an increasingly positive

effect of injector temperature with increasing retention

time. For the three latest eluting analytes, having

low volatility, the injector temperature was in fact the

dominating effect. However, for phosphoenolpyruvate

3TBDMS, threonine 2TBDMS and trans-4-hydroxyproline

2TBDMS, an increase in injector temperature decreased

the peak areas, suggesting a degradation of these com-

pounds at higher temperatures. This is illustrated in Fig. 2

for phosphoenolpyruvate 3TBDMS in comparison to cho-

lesterol TBDMS. For these amino acids only the partially

silylated 2TBDMS derivatives demonstrated this behavior

in contrast to their more abundant 3TBDMS derivatives,

clearly showing the improved thermal stability of silylated

amines over unsilylated amine.

The difference in yield in the investigated temperature

interval was for some analytes very large. Increasing the

injector temperature from 200 to 320�C increased the

peak area of cholesterol TBDMS and decreased the peak

area for phosphoenolpyruvate 3TBDMS with 90 and

80%, respectively. An injector temperature of 270�C was

found being a good compromise, resulting in an

approximately 45% reduction in peak area of these two

analytes with the remaining metabolites close to their

maximum values.

3.3 Chromatography

A PLS model was calculated, displaying a total explained

variation of 93% (R2Y = 0.932) and a cross-validated

predictive power of 88% (Q2Y = 0.884) with 0.719 B

R2Y B 0.989 and 0.564 B Q2Y B 0.979 for all studied

responses (Supplementary material S3).

It was found that an increase in injector purge vent time

greatly reduced the peak capacity for the shorter column,

whereas the effect on the longer column was much smaller

(Fig. 3). The main factors affecting the peak asymmetry

factor on the 10 m column were purge vent time and flow

rate, whereas these factors were insignificant on the 30 m

column. It was also observed that the range for the peak

asymmetry factor (Kirkland 1977) calculated for valine

2TBDMS was much greater on the 10 m column (0.4 B

As B 1.4) than on the 30 m column (0.7 B As B 1.2). The

peak height of cholesterol TBDMS was increased for both

columns primarily by an increase in temperature gradient

rate and/or purge vent time. However, improving this

response by increasing these parameters would severely

reduce the peak capacity and yield asymmetric peaks,

especially on the shorter column. Although the 10 m col-

umn may generate faster analyses, only a 17% increase in

the number of analyses in 24 h was obtained compared to

the 30 m column.

Fig. 3 Contour plots describing the influence of the injector purge

vent time and temperature gradient rate on the average peak capacity

for the 30 m column (a) and 10 m column (b). All factors not

presented in the plots are assigned center values

58 A. P. H. Danielsson et al.

123

Due to a greatly higher average peak capacity, superior

peak symmetry and decreased sensitivity to the factor

settings only the 30 m column was further optimized. The

purge vent time was set to 115 s, to optimize the chro-

matography for a high sample load.

The flow rate and the gradient rate affected both the

analysis time and the peak capacity. The effect of the flow

rate on the analysis time was moderate whereas its effect

on peak capacity was very pronounced. The flow rate was

consequently set to the lower setting of 1 ml/min yielding

high peak capacity and only slightly slower analysis.

The interaction between the temperature gradient rate

and the injector purge vent time was significant for the

peak capacity. At higher gradient rates, an increased purge

vent time decreases the peak capacity, whereas at a gra-

dient rate lower than 25�C/min, this effect is much less

pronounced. The gradient rate was consequently set to

25�C/min.

Increasing the initial temperature duration decreased the

peak capacity for the early portion of the chromatogram

while it was increased for the late portion of the chromato-

gram, suggesting an influence on the sample zone focusing.

However, as the average peak capacity was unaffected, the

initial isothermal duration was set to the lowest investigated

value, 2 min to increase the sample throughput.

The initial temperature was negatively correlated with

analysis time and the peak capacity for the early portion of

the chromatogram, but was positively correlated with the

peak capacity for the late portion of the chromatogram. At

the highest investigated temperature, 2 min could be cut off

the analysis time due to decreased retention during the initial

isothermal period. The interaction between the initial tem-

perature and the temperature gradient rate was significant

for the analysis time, but the contour plot showed that at the

chosen gradient rate of 25�C/min, the setting of the initial

temperature was unimportant. The interaction between the

initial temperature and the purge vent time was significant

for peak asymmetry for valine 2TBDMS, but as the peak

asymmetry factor for valine 2TBDMS was acceptable

regardless of the settings, this effect was not further con-

sidered. The same interaction also reduced the peak capacity

for the intermediate and late portion of the chromatogram,

whereas it was insignificant for the early part. A low setting

of the initial temperature could therefore be motivated.

In conclusion, the 30 m capillary column was optimal

together with a 115 s purge vent time, a flow rate of 1 ml/min,

an initial temperature of 60�C, an initial isothermal time of

2 min, and a gradient rate of 25�C/min. With these settings,

51 analyses, including an oven equilibration time of 10 min,

can be performed in 24 h. In comparison, the maximum

number of analyses achievable with the 30 m column is 68.

However, with the fastest method the average peak capacity is

severely decreased to 119 compared to the proposed method

with which a peak capacity of 220 is achievable.

-0,8

-0,6

-0,4

-0,2

-0,0

0,2

0,4

0,6

0,8

-0,2 -0,1 -0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0

wc[

2]

wc[1]

IST

Acq

Acq*Acq

IST*Acq

Late eluting

Early eluting

Fig. 4 PLS loading plot of the

model for the normalized peak

area of the [M-57]? fragments.

The plot shows that all

responses are positively

correlated with the ion source

temperature (IST), whereas the

three latest eluting metabolites

are strongly and negatively

correlated with the acquisition

rate (ACQ). The contour plots

of cholesterol TBDMS and

dihydroxyacetone phosphate

MEOX 3TBDMS illustrates the

different behavior of these two

clusters

Development of a Metabolomic Protocol 59

123

3.4 Mass spectrometry

The [M-57]? fragment of a-ketoisovalerate MEOX

TBDMS peak 1 was used for normalization as it was found

to have a very low reproducibility of 0.005, indicating that

the random error greatly overpowered the effects of the

factors. The model obtained after normalization yielded

R2Y = 0.686 and Q2Y = 0.582 with 0.507 B R2Y B

0.930 and 0.400 B Q2Y B 0.904 for all studied responses

(Supplementary material S4). It was found that the ion

source temperature was positively correlated with the

[M-57]? fragment peak areas of all metabolite derivatives

(Fig. 4). The acquisition rate had a low influence on the

majority of responses. However, for the late eluting

metabolites, this factor became increasingly important with

a negative contribution to the responses. This effect is

probably related to the increased peak width with

increasing elution times.

Although the ion source temperature should be kept high

according to the above results, it is likely to affect the

fragmentation pattern. To study whether this occurs in the

present investigation, a ratio of the unnormalized [M-57]?

fragment, and a low m/z-fragment of the metabolite

derivatives was modeled. A lower value of this response

indicates a more excessive fragmentation. The model

yielded R2Y = 0.887 and Q2Y = 0.822 with 0.598 B

R2Y B 0.999 and 0.401 B Q2Y B 0.999 for all studied

responses (Supplementary material S5). Interestingly, the

PLS loading plot shows distinct clustering of early, inter-

mediate and late eluting metabolites (Fig. 5).

The acquisition rate affected the ratio for the early and

late eluting metabolites, whereas this effect was insignifi-

cant for the metabolites with intermediate elution times.

The ratio was furthermore, for all metabolites, declining

with increasing ion source temperature. In general, proto-

cols for GC/MS-based metabolomics apply an ion source

temperature ranging between 200 and 250�C and increas-

ing the temperature further might hamper the construction

of common mass spectra databases. For these reasons, the

ion source temperature was set to 250�C. The acquisition

rate was set to 20 Hz.

3.5 Method validation

To properly assess the linearity criteria a lack-of-fit test,

which compares the pure error (random error) to the model

error, i.e. the error in the linear fit (Araujo, 2009), was

included in the linear range determination. For a truly

linear response, these errors are of the same magnitude.

However, if the model error is significantly larger than the

pure error, the model has lack-of-fit and the values do not

strictly follow a linear function. For some of the chosen

metabolites, lack-of-fit due to curvature was detected

(Table 3). The curvature was detected at the highest

Fig. 5 PLS loading plot of the

peak area ratio between low and

high m/z fragments. Contour

plots of alpha ketoisovalerate

MEOX TBDMS (early eluting),

dihydroxyacetone phosphate

MEOX 3TBDMS (intermediate

eluting) and cholesterol

TBDMS (late eluting) illustrate

the different behavior of the

three clusters

60 A. P. H. Danielsson et al.

123

concentration and was therefore concluded to be due to

sample overloading. For these analytes the linear range was

accordingly decreased. All metabolites investigated were

found to be linear at a relative plasma concentration B0.8

(v/v, plasma/plasma ? water). Each linear function was

also cross-validated to further ensure that the response is

truly linear in the determined range.

The LOD and LOQ were estimated for the analytes for

which a stable isotope-labeled internal standard was available

(Table 3). The LOD and LOQ were found to be in the range

of 0.05 and 2.8 lM and 0.17 and 9.3 lM, respectively, for all

metabolites except for the latest eluting metabolite, choles-

terol, having a LOD and LOQ of 14.1 and 47 lM, respec-

tively. The RSD of the peak areas, reflecting the precision of

the method, was below 5% for most analytes, although some

metabolites exhibited comparatively high standard deviations

especially at the lowest concentration (Table 3). This is most

likely due to that the LOD for these metabolites are approa-

ched at high dilutions of the plasma. Cholesterol, on the other

hand, being highly abundant in the plasma has a higher pre-

cision at the lowest plasma concentration. Speculatively, this

distinct behavior of cholesterol results from a slight over-

loading at higher plasma concentrations.

A chromatogram, illustrated both by the TIC and over-

laid SIMs of the investigated metabolites, illustrates the

high intensity of the [M-57]? fragments that greatly

facilitates the identification and quantification of the

metabolites (Fig. 6).

3.6 Sample drift comparison

Finally, an orthogonal projections to latent structures

(OPLS) model was calculated to relate the metabolite peak

area to the run order. The model had one predictive and 5

orthogonal components and yielded R2Y = 0.95 and

Q2Y = 0.81. The model indicated that mainly metabolites

containing carboxylic acid and hydroxyl functionalities

were drifting, with a slight decrease with the run order. The

drift was overall very small, as indicated by cholesterol

TBDMS possessing the strongest decay with run order

(Area = -0.002(run order) ? 1.5, R2 = 0.02). Likewise,

a model was calculated on replicates of the same set of

plasma samples derivatized with MSTFA (Gullberg et al.

2004; Jiye et al. 2005). This model required one predictive

and two orthogonal components, yielding R2Y = 0.90 and

Q2Y = 0.76. In that model, several metabolite peak areas

were found to increase with run order. Among these

metabolites, amino acids with disilylated amines were

over-represented indicating a progressing silylation.

Glycine 3TMS, having the least sterically hindered amine,

was found having the strongest positive loading, and a plot

of the raw data indicated a pronounced alteration in peak

area over time (Area = 0.11(run order) ? 0.1, R2 = 0.57).

Thus, it is evident that the application of a bulkier silyla-

tion reagent reduces the number of disilylated amines

and that this in turn reduces the drift in the metabolomics

data.

Table 3 Linear range, LOD and precision calculated from a set of blood plasma samples

Metabolite R2 Q2 Linear rangea Internal standard LOD (lM) LOQ (lM) %RSD (0.1)c %RSD (0.4)c

Alanine 0.990 0.989 0.1–0.8 13C3-15N alanine 0.11 0.37 0.8 0.4

Cholesterol 0.973 0.971 0.1–1.0 2H7-cholesterol 14.1 47 1.0 5.6

Citric acid 0.987 0.986 0.1–0.8 13C4-succinic acid NAb NAb 6.6 4.6

Cysteine 0.949 0.943 0.1–0.8 13C3-serine NAb NAb 27.8 6.5

Glycine 0.958 0.953 0.1–0.8 13C5-a-ketoisovaleric acid NAb NAb 12.2 4.4

Isoleucine 0.976 0.973 0.1–0.8 13C3-15N alanine NAb NAb 10.6 2.3

Isocitric acid 0.950 0.946 0.1–1.0 13C4-succinic acid NAb NAb 16.1 6.4

Lactic acid 0.964 0.961 0.1–1.0 13C4-succinic acid NAb NAb 5.2 3.3

Leucine 0.980 0.978 0.1–0.8 13C3-15N alanine NAb NAb 9.6 2.3

Lysine 0.958 0.953 0.1–0.8 13C9-15N-tyrosine NAb NAb 7.4 3.1

Methionine 0.951 0.946 0.1–0.8 13C3-serine NAb NAb 9.2 2.2

Oleic acid 0.961 0.958 0.1–0.8 13C18-oleic acid 2.80 9.3 6.1 2.7

Phenylalanine 0.995 0.995 0.1–1.0 13C6-phenylalanine 0.05 0.17 1.4 0.5

Serine 0.995 0.995 0.1–1.0 13C3-serine 0.05 0.17 1.2 0.5

Threonine 0.957 0.953 0.1–0.8 13C3-serine NAb NAb 9.5 2.7

Tyrosine 0.989 0.988 0.1–1.0 13C9-15N-tyrosine 0.07 0.23 2.9 1.4

Valine 0.962 0.958 0.1–0.8 13C3-15N alanine NAb NAb 6.3 1.2

a The linear range is expressed as relative concentrations 0.1, 0.4 and 1.0 (v/v, plasma/plasma ? water)b No stable isotope labeled internal standard availablec Precision expressed as relative standard deviation calculated at relative concentrations 0.1 and 0.4 (v/v, plasma/plasma ? water)

Development of a Metabolomic Protocol 61

123

4 Conclusion

Optimizing a metabolomics protocol is complicated by the

heterogeneity and vastly differing concentration ranges of

the metabolites. In an effort to cope with this obstacle when

optimizing a metabolomics protocol, sets of metabolites

were selected that spanned a large portion of the metabo-

lome. Thus, the individual metabolites investigated here

should not be considered being single metabolites but

rather representatives of whole groups of metabolites

having similar properties. This means that the metabolite

cocktail, although containing 36 unique metabolites, is a

less complex model of a real plasma sample, where several

hundred metabolites normally are detected. Moreover,

generating the whole set of responses from all of the

metabolites would result in very large data sets, which

would complicate both data generation and interpretation.

For this reason, parts of the optimization were only per-

formed on a selected set of representative metabolites. The

optimization therefore does not deal with specific co-elu-

tion problems that commonly are encountered in meta-

bolomics analyses. Nevertheless, responses related to the

overall method performance, such as peak capacity, anal-

ysis time, efficiency, and detection limits, still are valuable

measures, which represent the ability of the method to

handle real metabolomics samples.

With DOE, the influence of all major parameters in the

metabolomics protocol, spanning derivatization and GC/

MS analysis, could be investigated. The results clearly

reflected the heterogeneity of the metabolome, with several

metabolites showing inversely correlating responses. The

response surfaces calculated with DOE were extremely

beneficial in finding the optimal parameters resulting in

sufficient recovery of all metabolites. Although a general

metabolomics protocol is presented, the models calculated

here may also serve as guidelines to aid the selection of

parameters for targeted metabolomics, also known as

metabolite profiling. The proposed protocol for blood

plasma analysis, based on MTBSTFA derivatization, also

offers a different selectivity to that of the more frequently

used protocols based on MSTFA. Thus, it may serve both

as a standard operating procedure for some applications

and as a complementary approach for other metabolomics

protocols based on MSTFA. The optimal protocol found in

the present investigation suggested that methoximation

should be performed at 20�C for 4 h in 100% pyridine. The

silylation should be performed at 100�C for 170 min with

pure MTBSTFA added the vial. Injection of the samples

should be performed at 270�C, employing an injector purge

vent time of 115 s. The chromatography should be per-

formed on a 30 m capillary column with a flow rate of

1 ml/min, an initial temperature of 60�C, an initial iso-

thermal time of 2 min, and a gradient rate of 25�C/min.

The ion source temperature should be 250�C, with an

acquisition rate of 20 Hz for the mass spectrometer.

Acknowledgments This work was supported by grants from Swedish

Research Council (14196-06-3), the Crafoord Foundation, Lars Hierta,

Fredrik and Ingrid Thuring, Ake Wiberg, Albert Pahlsson, O.E. and

Edla Johansson Foundations, Knut and Alice Wallenberg Foundation,

and the Royal Physiographic Society. Support from Inga and John Hain

Foundation to PS is acknowledged.

References

Araujo, P. (2009). Key aspects of analytical method validation and

linearity evaluation. Journal of Chromatography B, 877,

2224–2234.

Araujo, P. W., & Brereton, R. G. (1996a). Experimental design I.

Screening. Trends in Analytical Chemistry, 15, 26–31.

Fig. 6 Chromatogram from pooled blood plasma illustrating the high

intensity of the [M-57]? fragments from the metabolites. a TIC

(b) SIMs from the [M-57]? fragments of the investigated metabolites

62 A. P. H. Danielsson et al.

123

Araujo, P. W., & Brereton, R. G. (1996b). Experimental design II.

Optimization. Trends in Analytical Chemistry, 15, 63–70.

Araujo, P. W., & Brereton, R. G. (1996c). Experimental design III.

Quantification. Trends in Analytical Chemistry, 15, 156–163.

Asres, D. D., & Perreault, H. (1997). Monosaccharide permethylation

products for gas chromatography - mass spectrometry: How

reaction conditions can influence isomeric ratios. CanadianJournal of Chemistry, 75, 1385–1392.

Begley, P., Francis-McIntyre, S., Dunn, W. B., et al. (2009).

Development and performance of a gas chromatography-time-

of-flight mass spectrometry analysis for large-scale nontargeted

metabolomic studies of human serum. Analytical Chemistry, 81,

7038–7046.

Birkemeyer, C., Kolasa, A., & Kopka, J. (2003). Comprehensive

chemical derivatization for gas chromatography-mass spectrom-

etry-based multi-targeted profiling of the major phytohormones.

Journal of Chromatography A, 993, 89–102.

Buscher, J. M., Czernik, D., Ewald, J. C., Sauer, U., & Zamboni, N.

(2009). Cross-platform comparison of methods for quantitative

metabolomics of primary metabolism. Analytical Chemistry, 81,

2135–2143.

Chorell, E., Moritz, T., Branth, S., Antti, H., & Svensson, M. B.

(2009). Predictive metabolomics evaluation of nutrition-modu-

lated metabolic stress responses in human blood serum during

the early recovery phase of strenuous physical exercise. Journalof Proteome Research, 8, 2966–2977.

Cozzolino, D., Flood, L., Bellon, J., Gishen, M., & Lopes, M. D. B.

(2006). Combining near infrared spectroscopy and multivariate

analysis as a tool to differentiate different strains of Saccharo-

myces cerevisiae: A metabolomic study. Yeast, 23, 1089–1096.

Danielsson, A. P. H., Moritz, T., Mulder, H., & Spegel, P. (2010).

Development and optimization of a metabolomic method for

analysis of adherent cell cultures. Analytical Biochemistry, 404,

30–39.

Ewald, J. C., Heux, S. p., & Zamboni, N. (2009). High-throughput

quantitative metabolomics: Workflow for cultivation, quenching,

and analysis of yeast in a multiwell format. Analytical Chem-istry, 81, 3623–3629.

Fiehn, O. (2002). Metabolomics—the link between genotypes and

phenotypes. Plant Molecular Biology, 48, 155–171.

Fiehn, O. (2008). Extending the breadth of metabolite profiling by gas

chromatography coupled to mass spectrometry. Trends inAnalytical Chemistry, 27, 261–269.

Fiehn, O., Kopka, J., Trethewey, R. N., & Willmitzer, L. (2000).

Identification of uncommon plant metabolites based on calcu-

lation of elemental compositions using gas chromatography and

quadrupole mass spectrometry. Analytical Chemistry, 72,

3573–3580.

Gullberg, J., Jonsson, P., Nordstrom, A., Sjostrom, M., & Moritz, T.

(2004). Design of experiments: An efficient strategy to identify

factors influencing extraction and derivatization of Arabidopsisthaliana samples in metabolomic studies with gas chromatogra-

phy/mass spectrometry. Analytical Biochemistry, 331, 283–295.

Jiye, A., Trygg, J., Gullberg, J., et al. (2005). Extraction and GC/MS

analysis of the human blood plasma metabolome. AnalyticalChemistry, 77, 8086–8094.

Jonsson, P., Johansson, E. S., Wuolikainen, A., et al. (2006). Predictive

metabolite profiling applying hierarchical multivariate curve

resolution to GC-MS Data-A potential tool for multi-parametric

diagnosis. Journal of Proteome Research, 5, 1407–1414.

Kanani, H. H., & Klapa, M. I. (2007). Data correction strategy for

metabolomics analysis using gas chromatography-mass spec-

trometry. Metabolic Engineering, 9, 39–51.

Kirkland, J. J. (1977). Sampling and extra-column effects in high-

performance liquid chromatography; influence of peak skew on

plate count calculations. Journal of Chromatographic Science,15, 303–316.

Lapainis, T., Rubakhin, S. S., & Sweedler, J. V. (2009). Capillary

electrophoresis with electrospray ionization mass spectrometric

detection for single-cell metabolomics. Analytical Chemistry, 81,

5858–5864.

Pous-Torres, S., Baeza-Baeza, J. J., Torres-Lapasio, J. R., & Garcıa-

Alvarez-Coque, M. C. (2008). Peak capacity estimation in

isocratic elution. Journal of Chromatography A, 1205, 78–89.

Rodrıguez, I., Quintana, J. B., Carpinteiro, J., Carro, A. M., Lorenzo,

R. A., & Cela, R. (2003). Determination of acidic drugs in

sewage water by gas chromatography-mass spectrometry as tert.-

butyldimethylsilyl derivatives. Journal of Chromatography A,985, 265–274.

Schweer, H. (1982). Gas chromatography–mass spectrometry of

aldoses as O-methoxime, O-2-methyl-2-propoxime and O-n-

butoxime pertrifluoroacetyl derivatives on OV-225 with meth-

ylpropane as ionization agent: I. Pentoses. Journal of Chroma-tography A, 236, 355–360.

Tam, Y. Y., & Normanly, J. (1998). Determination of indole-3-

pyruvic acid levels in Arabidopsis thaliana by gas chromatog-

raphy-selected ion monitoring-mass spectrometry. Journal ofChromatography A, 800, 101–108.

Wold, S. (1978). Cross-validatory estimation of the number of

components in factor and principal components models. Tech-nometrics, 20, 397–405.

Yu, Z., Peldszus, S., & Huck, P. M. (2007). Optimizing gas

chromatographic-mass spectrometric analysis of selected phar-

maceuticals and endocrine-disrupting substances in water using

factorial experimental design. Journal of Chromatography A,1148, 65–77.

Zelena, E., Dunn, W. B., Broadhurst, D., et al. (2009). Development

of a robust and repeatable UPLC-MS method for the long-term

metabolomic study of human serum. Analytical Chemistry, 81,

1357–1364.

Zhang, S., Nagana Gowda, G. A., Asiago, V., Shanaiah, N., Barbas,

C., & Raftery, D. (2008). Correlative and quantitative 1H NMR-

based metabolomics reveals specific metabolic pathway distur-

bances in diabetic rats. Analytical Biochemistry, 383, 76–84.

Development of a Metabolomic Protocol 63

123

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.