Quality control for plant metabolomics: reporting MSI-compliant studies.

Abstract

The Metabolomics Standards Initiative (MSI) has recently released documents describing minimum parameters for reporting metabolomics experiments, in order to validate metabolomic studies and to facilitate data exchange. The reporting parameters encompassed by MSI include the biological study design, sample preparation, data acquisition, data processing, data analysis and interpretation relative to the biological hypotheses being evaluated. Herein we exemplify how such metadata can be reported by using a small case study - the metabolite profiling by GC-TOF mass spectrometry of Arabidopsis thaliana leaves from a knockout allele of the gene At1g08510 in the Wassilewskija ecotype. Pitfalls in quality control are highlighted that can invalidate results even if MSI reporting standards are fulfilled, including reliable compound identification and integration of unknown metabolites. Standardized data processing methods are proposed for consistent data storage and dissemination via databases.

Full text

Page 1

TECHNIQUES FOR MOLECULAR ANALYSIS
Quality control for plant metabolomics:
reporting MSI-compliant studies
Oliver Fiehn1,*, Gert Wohlgemuth1, Martin Scholz1, Tobias Kind1, Do Yup Lee1, Yun Lu1, Stephanie Moon2and Basil Nikolau2
1Davis Genome Center, University of California, Davis, CA 95616, USA, and
2Department of Biochemistry, Biophysics and Molecular Biology, Iowa State University, Ames, IA 50011, USA
Received 22 October 2007; revised 26 November 2007; accepted 27 November 2007.
*For correspondence (fax þ1 530 754 9658; e-mail ofiehn@ucdavis.edu).
Summary
The Metabolomics Standards Initiative (MSI) has recently released documents describing minimum para-
metersforreportingmetabolomicsexperiments,inordertovalidatemetabolomicstudiesandtofacilitatedata
exchange. The reporting parameters encompassed by MSI include the biological study design, sample
preparation, data acquisition, data processing, data analysis and interpretation relative to the biological
hypotheses being evaluated. Herein we exemplify how such metadata can be reported by using a small case
study – the metabolite profiling by GC-TOF mass spectrometry of Arabidopsis thaliana leaves from a knockout
allele of the gene At1g08510 in the Wassilewskija ecotype. Pitfalls in quality control are highlighted that can
invalidate results even if MSI reporting standards are fulfilled, including reliable compound identification and
integration of unknown metabolites. Standardized data processing methods are proposed for consistent data
storage and dissemination via databases.
Keywords: GC/MS, standard operating procedure, abiotic stress, wounding, BinBase, SetupX.
Introduction
Plant metabolism is known to be highly flexible in terms of
both metabolite abundances and identities. By differentially
quantifying metabolite data among biological samples that
are differentiated by genetic, environmental, spatial or
developmental parameters, it is envisioned that metabolo-
mics data will yield information concerning the genetic,
environmental, spatial or developmental control of meta-
bolism (Fiehn, 2002). However, there are a number of tech-
nical and infrastructure issues that currently limit this vision.
There are many ways to conduct plant biochemical or
physiological studies, and there are also a variety of valid
methods and techniques to study metabolic alterations on a
comprehensive scale. No single analytical technology or
protocol facilitates the vision of comprehensive metabolite
analysis (van der Greef et al., 2004). Inherent in this techno-
logical barrier is thefact that metabolite concentrations span
up to sevenordersof magnitude (e.g.comparingabundance
of plant hormones to transport compounds such as
sucrose). Moreover, while core biochemical pathways are
certainly conserved across species, this is not true for the
tremendous variety of secondary metabolites in plants, their
pathway structures, regulatory mechanisms and spatial
specializations. Due to this metabolic flexibility, metabolo-
mes cannot easily be computed from genomes, and there-
fore the basic question of the extent of an organism’s
metabolome cannot be addressed as such. Consequently,
the ‘plant metabolome’ is a poorly defined, variable
entity that must be determined from empirical data.
Repositories need to be developed that host the comple-
ment of plant metabolite data integrated with accu-
rately curated metabolic information (Rhee et al., 2003;
Zhang et al., 2005). Initial databases have been released
(http://lab.bcb.iastate.edu/projects/plantmetabolomics/, http://
csbdb.mpimp-golm.mpg.de/csbdb/gmd/gmd.html,
fiehnlab.ucdavis.edu:8080/m1/main_public.jsp) to accom-
plish this goal.
While there is no single best way to conduct metabolomic
studies, there are a number of pitfalls and known problems
http://
ª 2008 The Authors
Journal compilation ª 2008 Blackwell Publishing Ltd
691
The Plant Journal (2008) 53, 691–704 doi: 10.1111/j.1365-313X.2007.03387.x

Page 2

that need to be carefully avoided, and detailed guidelines
and practice protocols have been published previously
(Fiehn, 2006; Lisec et al., 2006). We present here some
additional pitfalls and solutions for a number of problems
encountered when using the select technique of GC/MS-
based metabolite profiling, for example sample carry-over
effects or matrix problems. Similar quality issues are
relevant for other analytical methods such as LC/MS, but
to maintain clarity, we focus here on the comparatively
mature method of GC/MS. Other problems are independent
of the analytical chemistry method, for example compre-
hensive and reproducible extraction of metabolites despite
diverse structure and abundances, or difficulties in correctly
annotating analytical signals as genuine plant metabolites
and in storing and quantifying peaks that lack structural
elucidation. Ultimately, a combination of metabolite-profil-
ing techniques will need to be combined to yield a detailed
overview of plant metabolism.
In order to enable researchers to re-use metabolomic
data, repeat experiments or evaluate the validity of claims, it
is mandatory that any metabolomic study provides clear
descriptions of the biological design of the experiment and
that all technical parameters are presented in great detail.
Information about data origins is also called ‘metadata’.
Such metadata are usually reported in method sections, but
may also be found as experimental design details in the
bodyofplantresearchpapers.Itisoftendifficulttoextractall
the necessary information from a journal-published report
regarding how a plant metabolomic study was actually
performed. For instance, details on plant growth conditions
that some authors may find unimportant may have been left
out, such as details on the watering regimes or pathogen
protection used in a greenhouse. Metabolism readily
responds to environmental conditions. In order to enable
comparisons across studies, sufficient details on environ-
mental parameters need to be reported even for studies
focusing on metabolic effects of genotype variations. Sim-
ilarly, authors often only refer to technical methods by
literature citations. This may be a reasonable option once
procedures are standardized and validated to a great extent.
However, currently metabolomics studies often involve a
variety of methods, and therefore readers have to accumu-
late technical details from several sources of information. In
addition, the actual use of technical methods sometimes
deviates from previously published method descriptions,
which may not be clear in the final reports. In summary,
therefore, a lack of metadata on plant biology as well as
missing information on technical methods make the final
results and conclusions less useful and harder to reproduce.
The ever growing number of technical methods that are
employed in a single plant metabolomic study requires that
reporting structures be formalized. After a series of initial
drafts,a consultationperiod andconferenceworkshops(e.g.
the 4th International Conference on Plant Metabolomics,
2006), the Metabolomics Society released a number of
documents on proposed minimum standards for use in
metabolomic reports (Fiehn et al., 2007a), among them
standards on plant biology context (Fiehn et al., 2007b)
and on reporting of methodological details of chemical
analysis (Sumner et al., 2007). These documents are
open for criticism and continuing improvement using
the open-accessMSI website
sourceforge.net/). The number of parameters and the level
of detail required to be given may change over time, but the
principal idea is that all metadata are compiled and reported
inasingledocumentthatwouldalsotypicallyberequiredby
leading plant science journals, called ‘minimal reporting
standards’. These standards are called minimal because, for
certain biological studies, it is important to gather further
information,forexampleontheoriginandstoragehistoryof
plant seeds or on air circulation specifications in climate
chambers that ultimately might have a confounding influ-
ence on metabolite levels. Even such compilations of
‘minimal’ metadata usually extend far beyond typical
method sections (also called experimental procedures).
As yet, these ‘minimal standards’ lack defined and pub-
lished instances of actual studies. We here outline how
metadataforplantmetabolomicsshouldbereportedusinga
small case study profiling a well-characterized Arabidopsis
genotype. In addition, we present quality controls and
methods on plant metabolite profiling by GC/MS that have
been developed over the past 7 years, highlighting practical
problems that need to be addressed when applying estab-
lished methods to plant tissues for which these methods
were not developed and validated. In addition, we exemplify
how novel injection techniques can improve standard GC/
MS-based metabolite profiling. We further outline how
automatic data processing and the use of standardized
methods support a GC/MS-based plant metabolomic data-
base. While this database can so far only handle a specific
GC/MS instrument, the algorithms and concepts for the data
processing can and should be applied more widely to
facilitate data exchange and comparison of data across
studies.
(http://msi-workgroups.
Reporting experimental methods
The Metabolomics Standards Initiative has proposed a
variety of parameters along the general metabolomic
workflow. It is a common misconception that metabolomics
approaches focus on a specific machine or utilize a specific
best type of extraction procedure. This is generally not the
case, as indicated by the vast set of methods and combina-
tion of these methods that are published in plant meta-
bolomics. Instead, metabolomics workflows need to be
described in detail, covering the plant biology context,
chemical analysis methods, data processing methods, and
the statistics and data interpretation methods.
692 Oliver Fiehn et al.
ª 2008 The Authors
Journal compilation ª 2008 Blackwell Publishing Ltd, The Plant Journal, (2008), 53, 691–704

Page 3

An example of how such extended experimental proce-
dures can be reported is given below, following point by
point the recommendations outlined by the Metabolomics
Standards Initiative. It is important to note that, by following
these minimal reporting requirements, studies become
‘MSI-compliant’, but such compliance does not imply that
the studies are validated studies or potentially more valu-
able with respect to scientific findings. MSI compliance only
means that a sufficient amount of metadata is given to
enable readers or users of the data to reproduce how the
study was performed. However, the full complement of
method details extends the space requirements allowed by
many peer-reviewed journals. It is generally in accordance
with MSI compliance to submit metadata as supplementary
information in file formats, and to give only a brief overview
over the most important parameters in the printed text.
To illustrate this reporting structure and provide an
example of a good practice document, we present herein a
small metabolomics study from an NSF2010 project that is
exploring the use of metabolomics data to decipher
functions of multiple genes whose functions are currently
unknown. This study was designed as a means of validat-
ing the experimental approach, and hence used a T-DNA
mutant stock that has undergone extensive characteriza-
tion in previous studies (Bonaventure et al., 2003). This
Arabidopsis mutant stock is in the Wassilewskija (Ws)
ecotype background. It carries a T-DNA disruption in the
gene At1g08510, which codes for a fatty acyl-ACP thioest-
erase (FATB). The metabolomics experiment sought to
determine the effect of the mutation on the metabolome in
conjunction with an environmentally induced perturbation
(plant wounding). All data generated by the NSF2010
consortium are publicly available and can be downloaded
from a dedicated server (http://www.plantmetabolomics.
org). GC–TOF data can also be downloaded directly from
the participating laboratory (http://fiehnlab.ucdavis.edu:
8080/m1/main_public.jsp). A number of technical details
involving the data acquisition instrument and the data-
base annotations are cutting-edge technology, and there-
fore may require additional investment. In such cases,
we have added comments on the basic reasons why
we chose these technologies and what substitutes could
be used.
Plant context metadata
Plant biology parameters are separated into the physical
object of the study (BioSource) and general growth condi-
tions, followed by treatment and harvest parameters (Fiehn
et al., 2007b). Species names follow the National Center for
Biotechnology Information (NCBI) taxonomy, and nomen-
clature for plant features must conform to the classifications
and relationships given by the Plantontology consortium
(http://www.plantontology.org/). Unitsmustbe specifiedbut
are not yet standardized. For example, the amount of plant
material used for the study is often given in terms of milli-
grams of fresh weight, but dry weight information might be
preferable. Both alternatives are MSI-compliant, but ‘fresh
weight’ comparisons would probably not be reasonable for
drought-stress experiments. While in principle, MSI-com-
pliant data can be submitted as flow text, it is more advan-
tageous to present data in categories or tables as given
below to facilitate comparisons between studies and
to provide data inputs and outputs that are more easily
machine-readable.
BioSource Species: Arabidopsis thaliana.
Genotype: Wassilewskija (Ws) and Wassilewskija (Ws)
fatB knockout (At1g08510).
Organ: Leaf.
Organ specification: Rosette leaves, aerial portion.
Amount: 50 mg fresh weight per sample, of which 20 mg
were used for extraction.
Growth Support: Fourteen to sixteen seeds were sown on
20–25 ml of sterile Murashige and Skoog basal salt mixture
(MS medium) containing 0.1% w/v sucrose and 1· liquid
vitamin solution (Sigma, http://www.sigmaaldrich.com/)
containing 15 g l)1bacto agar (BD) in 100 · 100 · 15 mm
square Falcon Petri dishes (Thermo Fisher Scientific; http://
www.thermofisher.com). Seeds were arranged on the
plates in a single horizontal line 1 cm from the top of the
plate. Prior to sowing, seeds were sterilized by treating for
1 min at room temperature with 300 ll of 50% v/v ethanol;
this solution was then removed and replaced by 300 ll of a
solution consisting of 1% v/v Tween-20 (Thermo Fisher
Scientific) and 50% v/v bleach (Clorox; http://www.clorox.
com), and incubated at room temperature for 10 min. The
seeds were then washed with three changes of 0.3 ml of
sterile water. After sowing the seeds, the plates were
wrapped using micropore tape (3 M Health Care; http://
www.3m.com), and then stored horizontally for 4 days at
4?C in the dark. On the 5th day, plates were moved to the
growth room, and held in a vertical position in Plexiglass
holders for 12 days.
Location: Controlled-environment facility at Iowa State
University, Nikolau laboratory.
Plot design: Each genotype and replicate were grown on
individual plates and placed randomly in the Plexiglass
holders.
Light period: 24 h day at 82 lmol m)2s)1(light source
Sylvania; http://www.sylvania.com), F34CW/SS/ECO/RP).
Humidity: Day 100%, night 100%.
Temperature: Day 24?C, night 24?C.
Watering regime: No further watering, plates remained
closed.
Nutritional regime: MS medium without further fertilizers.
Date of plant establishment: 25 September 2006.
Quality control in metabolomics 693
ª 2008 The Authors
Journal compilation ª 2008 Blackwell Publishing Ltd, The Plant Journal, (2008), 53, 691–704

Page 4

Treatment Abiotic treatment: On 11 October 2006 at 11:30
a.m. (13th day after plant establishment), plates were
uncovered in the growth room and leaves were wounded
by piercing using an 18-gauge sterile needle. Plates were
then re-covered with their lids and wrapped with micro-
pore tape (3M Health Care). All plants, including the
non-wounded plants, were exposed to the air during
the wounding period. Plates were then placed vertically
in Plexiglass holders under the conditions described
above.
Dose: Ten punches.
Duration: 3 min wounding period; 2 h response period
before harvest.
Harvest Date: 11 October 2006.
Time: 1:30 p.m.
Growth stage: Boyes 1.1–1.4.
Metabolism quenching method: Immersion in liquid
nitrogen within 1 min after harvest.
Harvest method: Petri plates were opened and the aerial
portions of the plants were cut.
Storage: )70?C for 1 day, then shipping on dry ice and
storage at )80?C for 2 weeks.
Chemical analysis metadata
Following the recommendations of the Metabolomics Stan-
dards Initiative (Sumner et al., 2007), chemical analysis
metadata are separated into information on ‘sample pro-
cessing and extraction’, ‘chromatography’, ‘mass spectro-
metry’, ‘metabolite identification’ and ‘quality control’.
DetailsgivenhererefertothefatBwoundingstudybyGC/MS.
Sample processing and extraction Tissue processing: Fro-
zen tissues were kept in 2 ml round-bottomed Eppendorf
tubes equipped with one 3 mm diameter steel ball, and
homogenized using a Retsch (http://www.retch-us.com) ball
mill for 30 sec at 25 sec)1.
Extraction: Ground tissue powder was kept in liquid
nitrogen between homogenization and extraction. The
extraction solvent was prepared by mixing isopropanol/
acetonitrile/water at the volume ratio 3:3:2 and degassing
this mixture by directinga gentle streamof nitrogen through
the solvent for 5 min. The solvent was cooled to )20?C prior
to extraction. Randomly processing all samples of the study,
1 ml of cold solvent per 20 mg of ground tissue was added,
vortexed for 10 sec, and shaken at 4?C for 5 min to extract
metabolites and simultaneously precipitate proteins. After
centrifugation at 12 800 g for 2 min, 90% of the supernatant
was removed, taking care not to remove any residues from
the pellet.
Extract concentration: The supernatant was separated
into two equal aliquots and concentrated to dryness in a
Centrivap cold trap vacuum concentrator (http://www.
labconco.com) at room temperature for 4 h.
Extract clean-up: In order to fractionate complex lipids
and waxes, the residue was re-suspended in 500 ll 50%
aqueous acetonitrile and centrifuged at 12 800 g for 2 min.
The supernatant was transferred to a 1.5 ml Eppendorf
tube and concentrated to dryness in a vacuum concen-
trator.
Extract storage: Dried extracts can be kept under nitrogen
at )80?C for up to 4 weeks. In the study presented here,
extracts were immediately derivatized for GC–TOF mass
spectrometry.
Chromatography Sample preparation: A mixture of internal
retention index (RI) markers was prepared using fatty acid
methyl esters of C8, C9, C10, C12, C14, C16, C18, C20, C22,
C24, C26, C28 and C30 linear chain length, dissolved in
chloroform at concentrations of 0.8 mg ml)1(C8–C16) or
0.4 mg ml)1(C18–C30). Aliquots (2 ll) of this RI mixture
were added to the dried extracts, then 10 ll of a solution of
20 mg ml)1of 98% pure methoxyamine hydrochloride (CAS
number 593–56-6, Sigma) in pyridine (silylation grade;
Pierce; http://www.pierce.net.com) was added to protect
aldehyde and ketone groups, and the mixture was shaken at
30?C for 90 min. Then 90 ll of N-methyl-N-trimethylsilyl-
trifluoroacetamide with 1% trimethylchlorosilane (MSTFA/
1% TMCS) (1 ml bottles; Pierce) was added for trimethyl-
silylation of acidic protons and shaken at 37?C for 30 min.
The reaction mixture was transferred to a 2 ml clear glass
auto-sampler vial with micro-insert (Agilent; http://www.
agilent.com) and closed using a 11 mm T/S/T crimp cap
(MicroLiter; http://www.microliter.com).
Auto-injector: A Gerstel automatic liner exchange system
with a MPS2 dual rail multi-purpose sampler and two
derivatization stations was used in conjunction with a
Gerstel CIS cold injection system (Gerstel; http://www.
gerstelus.com). For every 10 samples, a fresh multi-baffled
liner was inserted (Gerstel catalog number 011711-010-00)
using the Gerstel Maestro 1 software version 1.1.4.18.
Before and after each injection, the 10 ll injection syringe
was washed three times with 10 ll ethyl acetate. Each 1 ll
sample was filled using 39 mm vial penetration at 1 ll sec)1
filling speed, injecting 0.5 ll at a 10 ll sec)1injection speed
at an initial temperature of 50?C which was ramped by
12?C sec)1toafinaltemperatureof250?Candheldfor3 min.
The injector was operated in split-less mode, opening the
split vent after 25 sec. Samples were injected between
2–24 h after derivatization using randomized sequences
controlled by the laboratory information management and
database system, SetupX (Scholz and Fiehn, 2007).
Chromatography instrument: An Agilent 6890 gas chro-
matograph controlled using Leco ChromaTOF software
verson 2.32; http://www.leco.com.
694 Oliver Fiehn et al.
ª 2008 The Authors
Journal compilation ª 2008 Blackwell Publishing Ltd, The Plant Journal, (2008), 53, 691–704

Page 5

Separation column: A 30 m long, 0.25 mm internal dia-
meter Rtx-5Sil MS column with 0.25 lm 95% dimethyl/5%
diphenylpolysiloxanefilm
integrated guard column was used (Restek; http://www.
restek.com).
Separation parameters: 99.9999% pure helium with
built-in purifier (Airgas; http://www.airgas.com) was used
at a constant flow of 1 ml min)1. The oven temperature was
held constant at 50?C for 1 min, and then ramped at
20?C min)1to 330?C, and held constant for 5 min.
andan additional10 m
Mass spectrometry Instrument: A Leco Pegasus IV time-of-
flight mass spectrometer controlled using Leco ChromaTOF
software version 2.32, and operated by Yun Lu on 5 April
2007.
Sample introduction: The transfer line temperature
between gas chromatograph and mass spectrometer was
set to 280?C.
Ionization: Electron impact ionization at 70 V was
employed, with an ion source temperature of 250?C.
Data acquisition: After 290 sec solvent delay, filament 1
was turned on and mass spectra were acquired at mass
resolving power R = 600 from m/z 85-500 at 20 spectra per
second and 1550 V detector voltage without turning on the
mass defect option. Recording ended after 1200 sec. The
instrument performed auto-tuning for mass calibration
using FC43 (perfluorotributylamine) before starting analysis
sequences.
Metabolite
index window ? 2000 U (around ? 2 sec retention time
deviation).
Second parameter: Mass
additional confidence criteria as detailed below (Data
processing).
Metabolite library: Importantly, GC/MS peaks can only
be annotated as identified according to MSI standards if at
least two independent parameters are recorded and
matched: mass spectra and retention index, in this
instance. Hence, spectra that are just matched to library
entries without retention time (or retention index) infor-
mation cannot be considered identified. All signals that are
exported by the BinBase database (Fiehn et al., 2005) are
reported by the quantification ion, a unique database
identifier, retention index and the complete mass spectrum
encoded as string. Database entries are named using the
Fiehn library, which currently includes 713 unique metab-
olites and 1197 unique spectra, for which names, structure
graphs and codes, retention indices and database referen-
ces are available at http://fiehnlab.ucdavis.edu/Metabolite-
Library-2007/. Mass spectra for the library itself are
commercially available from the instrument manufactur-
ers. Alternative libraries of GC/MS metabolite spectra are
freely available at http://csbdb.mpimp-golm.mpg.de/csbdb/
identifications Firstparameter:Retention
spectralsimilarity plus
gmd/msri/gmd_msri.html, and apply a different set of GC/
MS conditions.
Quality controls: Daily quality controls were used. These
comprised two method blanks (involving all the reagents
and equipment used to control for laboratory contamina-
tion) and four calibration curve samples spanning one
order of dynamic range and consisting of 31 pure refer-
ence compounds. These quality control calibration sam-
ples were injected at 0.5 ll injection volumes and a split
ratio of 1/5 because potential problems with the stability of
trimethylsilylated amino acids are first apparent in split
mode. Intervention limits were established to ensure basic
validation of the instrument for metabolite profiling.
Data processing
According to the recommendations of the Metabolomics
Standards Initiative (Sumner et al., 2007), metadata are
required to detail how resulting GC–TOF mass spectrome-
try profiles were investigated. The details of this procedure
are critical for achieving highly reliable and consistent
results. Unbiased analysis seeks to (i) find all signals that
can be distinguished from background noise or signals with
very similar retention times, and then (ii) report signal
intensities for all detected signals in all chromatograms. In
subsequent steps, signals can be identified using mass
spectral libraries and then investigated for differential
regulation using statistics software. Each of these steps
requires great diligence if data are to be stored and dis-
seminated permanently. Details presented here involve
elaborate algorithms in order to reduce the number of false-
positive and false-negative peak detections and to avoid
misidentification of metabolites. Freely available software
programs such as AMDIS (Stein, 1999), mzmine (Katajamaa
et al., 2006) or SpectConnect (Styczynski et al., 2007)
accomplish parts of these processes but are not tailored
towards a database approach.
File formats: Files were pre-processed directly after data
acquisition and stored as ChromaTOF-specific *.peg files, as
generic *.txt result files, and additionally as generic ANDI
MS *.cdf files. It is recommended that such generic *.cdf
files are made available for published data sets to facilitate
data sharing.
Pre-processing details: ChromaTOF version 2.32 was
used for data pre-processing without smoothing, with
3 sec peak width, baseline subtraction just above the noise
level, and automatic mass spectral deconvolution and
peak detection at signal/noise levels of 10:1 throughout
the chromatogram. Apex masses were used for quantifi-
cation. Resulting *.txt files were exported to a data server
with absolute spectra intensities, and further processed
using the BinBase algorithm (Fiehn et al., 2005; see
Figure S1). This algorithm used the settings: validity of
Quality control in metabolomics 695
ª 2008 The Authors
Journal compilation ª 2008 Blackwell Publishing Ltd, The Plant Journal, (2008), 53, 691–704

Page 6

chromatogram (<10 peaks with intensity >10^7 counts
sec)1), unbiased retention index marker detection (MS
similarity >800 and exceeding thresholds for ion ratio
abundances for high m/z marker ions), retention index
calculation by 5th order polynomial regression. Spectra
were cut to 5% base peak abundance, and matched to
database entries from most- to least-abundant spectra
using the following matching filters: retention index
window ?2000 U (equivalent to about ?2 sec retention
time), validation of unique ions and apex masses (unique
ion must be included in apex masses and present at >3%
of base peak abundance), mass spectrum similarity that
must fit criteria dependent on peak purity and signal/noise
ratios (Table S1), optional ion ratio settings to distinguish
peaks with high similarity, and a final isomer filter
(annotating the isomer spectrum with the closest RI fit).
In any metabolomic report, is important to state the
confidence or quality thresholds that were used to identify
signals as genuine (known) metabolites. Failed spectra
were automatically entered as new database entries if
signal-to-noise ratio >25, purity <1.0 and presence in the
biological study design class was >80%. This filter ensured
that (i) signals were reported that had never been detected
previously in any other sample, but (ii) only signals are
reported that can be assumed to be biologically relevant
using relatively abundant and pure signals and ensuring
that these are positively detected in most of the biological
replicates.
All thresholds reflect settings for ChromaTOF version
2.32. Signal intensities were reported as peak heights using
the unique ion as default, unless an alternative quantifica-
tion ion was manually set in the BinBase administration
software Bellerophon. A quantification report table was
produced for all database entries that were positively
detected in more than 80% of the samples of a study design
class (as defined in the SetupX database). This procedure
results in 10–30% missing values, which could be caused by
true negatives (compounds that were below the detection
limit in a specific sample) or false negatives (compounds
that were present in a specific sample but that did not match
quality criteria in the BinBase algorithm). A subsequent
post-processing module was employed to automatically
replace missing values from the *.cdf files using the open-
access mzmine software (Katajamaa et al., 2006) with the
following parameters. For each positively detected metab-
olite, the average retention time was calculated. For each
chromatogram and each missing value, the intensity of the
quantification ion at this retention time was extracted by
seeking its maximum value in a retention time region of
?1 sec and subtracting the minimum (local background)
intensity in a retention time region of ?5 sec around the
peak maximum. The resulting report table therefore did not
have any missing values. Replaced values were labeled as
‘low confidence’ by color coding.
Statistics
The Metabolomics Standards Initiative has released mini-
mum requirements for reporting data transformations and
statistics (Goodacre et al., 2007). Here we describe common
data evaluation methods by separation into data transfor-
mation and statistics.
Data transformation: Result files were transformed by
calculating the sum intensities of all structurally identified
compounds for each sample (i.e. those signals that had
been positively identified in the data pre-processing
schema outlined above), and subsequently dividing all
data associated with a sample by the corresponding
metabolite sum. The resulting data were multiplied by a
constant factor in order to obtain values without decimal
places. Intensities of identified metabolites with more than
one peak (e.g. for the syn- and anti-forms of methoxi-
mated reducing sugars) were summed to only one value
in the transformed data set. The original non-transformed
data set was retained. The general concept of this data
transformation is to normalize data to the ‘total metabolite
content’, but disregarding unknowns that might poten-
tially comprise artifact peaks or chemical contaminants.
Caution must be exercised when comparing classes of
samplesthatmight have
metabolite concentrations (e.g. cold-acclimatized plants
versus plants grown at room temperature). Alternative
means of normalization may utilize the amount of mate-
rial used for the analysis (e.g. the mass of the extracted
tissue) and additional internal standards that would
accountfor losses during
injections.
Statistics: Statistical analyses were performed on all
continuous variables using Statistica software version 7.1
(StatSoft; http://www.statsoft.com). Univariate statistical
analysis for multiple study design classes was performed
by breakdown and one-way ANOVA. F statistics and P-values
were generated for all metabolites. Data distributions were
displayed by box–whisker plots, giving the arithmetic mean
value for each category and the standard error as box and
whiskers for 1.96 times the category standard deviation to
indicate the 95% confidence intervals, assuming normal
distributions. Multivariate statistical analysis was per-
formed by unsupervised principal component analysis
(PCA) to obtain a general overview of the variance of
metabolic phenotypes in the study, by entering metabolite
values without study class assignments. In addition, super-
vised partial least-square (PLS) statistical analysis was
performed, which requires information about the assigned
study classes. Three plots were obtained for each PCA and
PLS model. The first is a scree plot for the Eigen values of
the correlation or covariance matrix. This is considered as a
simple quality check and should have a steep descent with
an increasing number of Eigen values. Second, 2D score
biologicallyvery different
sample preparationand
696 Oliver Fiehn et al.
ª 2008 The Authors
Journal compilation ª 2008 Blackwell Publishing Ltd, The Plant Journal, (2008), 53, 691–704

Page 7

scatter plots were generated for at least the first three
dimensionless principal components, and 3D plots are
generated to better distinguish metabolic phenotypes.
Third, loading plots were generated for each vector in
PCA or PLS, showing the impact of variables on the
formation of vectors. Metabolites near the coordinate
center had no separation power; conversely, variables far
away from the coordinate center were important for
building PCA and PLS models. Variables that are located
close to each other are strongly correlated.
Quality control measures in GC/MS-based metabolite
profiling
Extraction parameters have a direct impact
on metabolic data
The extraction conditions proposed above involved inacti-
vation of enzymes by a )20?C cold protein precipitationstep,
concomitant with dissolving small molecule metabolites
into an excess of a ternary solvent mixture of lipophilic and
hydrophilic organic solvents with water. It is important to
note that other extraction conditions can be applied, for
example using stronger lipophilic solvents such as chloro-
form (Gullberg et al., 2004; Weckwerth et al., 2004) instead
of isopropanol. Older protocols involved sequential use of
hydrophilic and lipophilic solvents, e.g., hot aqueous
methanol followed by a second chloroform extraction and a
solvent-based fractionation schema (Fiehn et al., 2000a;
Roessner et al., 2001). Quantitative results from metabolite
profiles will differ according to the extraction protocol used.
For example, if fractionation schemes are employed, semi-
polar metabolitesmay be preferentiallypresent in one or the
other fraction, depending on the composition of the total
matrix. However, metabolites may be found in different
fractions depending on the plant species: Arabidopsis leaf
extracts yield chlorophyll in the lipophilic phase, whereas a
considerable fraction of chlorophyll is found in the polar
phasein leaf extractsof Cucurbita maxima when performing
fractionated extractions using methanol/water and metha-
nol/chloroform protocols (Fiehn et al., 2000a; Roessner
et al., 2001). It is a general problem for comparisons
between extraction methods that no accepted benchmark
profiles are available. While metabolomics means making
compromises, each protocol has to be validated by the
extent that known metabolites can be reproducibly and
quantitatively extracted from the plant matrix. While some
metabolites (such as glucose-6-phosphate) must be present
in Arabidopsis extracts to validate a certain protocol, the
presence of signals for unknown metabolites may always be
explainable by presumed artifact formation or degradation.
One way to overcome this problem is to assess recovery
rates obtained by comparing external calibration curves to
spiked standard additions of a select number of pure refer-
ence compounds. However, such recovery rates cannot be
used as true surrogates for the accuracy of assessing in vivo
concentrations, because external and internal calibration
curves may vary due to extraction and matrix effects. Use of
internal isotope-labeledcompounds or plant material grown
under isotope-labeled nutrient conditions may partly rem-
edy this situation. A simpler way is to compare frequency
distributions of the technical reproducibility of metabolite
profiles as a discriminator between alternative extraction
protocols. A non-normal distribution of precision values
may be expected, with a high number of metabolites giving
low error rates for repetitive analysis of technical replicates
(<15% relative standard deviation) and a low number of
compounds with high or very high errors (>40% technical
errors). In general, the median technical reproducibility
determined over all detectable metabolites should be
smaller than20%relativestandarddeviation.Despitesimilar
values for overall precision, the frequency distribution of
quantitative reproducibility may differ for different solvent
mixtures (Figure 1). For technical replicate extractions of
young soybean seeds, we found that a ternary extraction
mixture of water–chloroform–methanol (Weckwerth et al.,
2004) yielded good quantitative precision for most analytes,
whereas a water–isopropanol–acetonitrile mixture resulted
in a bimodal distribution for soybean, with a range of com-
pounds around 40–55% relative standard deviation. Such a
bimodal frequency distribution was not found for extraction
of Arabidopsis rosette leaves. Therefore, method validation
studies must be carried out for every new plant organ that
is subjected to metabolite-profiling studies. As a rule of
thumbforsuchmethoddevelopmentandvalidationstudies,
a list of expected metabolites should be generated for the
target plant organ, including preliminary assessments of
abundance and potential detectability by a specific meta-
bolite-profiling technique. For example, ADP/ATP cannot be
0
10
20
30
40
50
60
0–5
11–15 21–2531–35 41–45 51–55
>60
% Relative standard deviation
Count of metabolic peaks
Figure 1. Precision of quantitative analysis for GC–TOF peaks reported by
BinBase for young soybean seeds. Frequency distributions are given in
increments of 5% relative standard deviations, comparing cold methanol/
chloroform/water extractions (blue line) or the cold acetonitrile/isopropanol/
water protocol (red line), using nine technical replicates each.
Quality control in metabolomics 697
ª 2008 The Authors
Journal compilation ª 2008 Blackwell Publishing Ltd, The Plant Journal, (2008), 53, 691–704

Page 8

analyzed by GC–TOF mass spectrometry, whereas AMP
levels are routinely determined. Special attention should be
given to compounds that are prone to be degraded by
residual enzymatic activity, heat or oxidation, for example
sugar phosphates, cysteine and ascorbate.
Additionally, the choice of extraction solvent mixtures
may be made in a rational design by calculating relative
solubility. Here we employed predictions using COSMOfrag
and the COSMO-RS software (Klamt et al., 2002). Around
1000 compounds from various chemical classes were
selected from the metabolic pathways database KEGG
(http://www.genome.jp/ligand) database, and relative solu-
bilities were calculated for 18 single solvents and ternary
mixtures of methanol, chloroform and water. The worst
solvent was dimethyl sulfoxide, which was incapable of
dissolving around 80% of the selected substances. Calcula-
tions for pyridine showed excellent solubility for a number
of carbohydrates, as also shown experimentally (Modi et al.,
2000), justifying its frequent use as a solvent for derivatiza-
tions in GC–TOF metabolite profiling. As expected, it was
found that solvents such as diethyl ether, benzene or
dichloroethane are strong but very selective solvents, and
that solvent mixtures such as methanol:chloroform:water
(3:1:1) can dissolve a very broad range of biologically active
compounds and are superior to any solvent by itself.
The sample preparation method involves all steps includ-
ing harvest, quenching of metabolism, storing tissues,
homogenizing representative samples and extraction con-
ditions. Each step should be carefully considered and not all
potential pitfalls can be described here. Once the sample
preparation method has been validated, details of the
procedures must be documented in written form, resulting
in a ‘standard operating procedure’. However, an important
part of the quality control process is to ensure that use of the
protocol yields reproducible data sets, independently of the
individualwhocarriedoutthesamplepreparations.Training
of new laboratory members (or efforts involving partnering
laboratories) must be documented by technical replicate
analyses to be confident that the standard operating proce-
dure is sufficiently clear and concise.
Quality of GC/MS-based metabolite profiles depends
on rigorous control of injector conditions
Metabolomics poses a challenge to quality control: in GC/
MS or LC/MS-based metabolomics, the instrument is
physically exposed to the sample, unlike, for example, in
NMR-based studies. Hence, sample constituents may
remain inside the chromatograph or mass spectrometer,
leading to cross-contamination or sensitivity losses, or
otherwise adversely affecting data quality (Hajslova et al.,
1998). Using replicate washing steps eliminates contami-
nation by the auto-sampler itself, but other parts of the
instrument may be heavily exposed to the sample. For
example, it is common to use reverse-phase columns in
liquid chromatography (HPLC): plant waxes will show
severe and sometimes irreversible adsorption under stan-
dard HPLC conditions, such that obtaining high-quality
quantitative metabolite profiles is very difficult. Similarly,
waxes or membrane lipids are not volatile enough for
GC/MS analysis, and would therefore be retained within the
injector, either in the liner, the injector body or even within
the first centimeters of the column. Over time, these con-
taminations result in severe problems that have an impact
on data quality, from formation of catalytically active sites to
absorption effects by pyrolytic particles in the liner.
One way to remove such involatile compounds is by lipid/
polar fractionation schemes as described previously (We-
ckwerth et al., 2004). However, apart from the problem of
splitting semi-polar compounds into two fractions, sterols,
plant hormones and free fatty acids are all found in the
lipophilic fractions, which (due to time constraints) are often
notanalyzedinseparateanalyticalruns.Inaddition,merging
results from two separate data files is a challenge in itself.
These problems can be avoided if liners are continuously
exchanged using additional instrument hardware. Two
automatic liner-exchange devices are commercially avail-
able, a direct thermo-desorption device (DTD) with crimp-
capped liners that contain micro-inserts (ATAS; http://
www.atasgl.com) (Denkert et al., 2006; Meyer et al., 2007;
Sanz et al., 2004), or a liner exchange/cold-injection system
(ALEX-CIS, Gerstel) (David et al., 2006). Both systems use
cold injection followed by heat ramping of the injector to
ensurethatonlyvolatilecomponentsreachtheinjectorbody
and column. Cold-injection techniques prevent the violent
explosions of injection drops found for classic hot injectors
(Grob and Biedermann,2000), and thushelp to improvedata
quality by reduced degradation of thermolabile compounds.
Involatilematrixcomponentsremaininsidethemicro-inserts
or exchangeable liners, which are disposed of after use.
Automatic liner-exchange injectors facilitate the injection
ofplantextracts,includinglipidssuchassterolsandfreefatty
acids. Without liner exchange, full metabolome injections in
a hot split/split-less injector result in increasing signal
intensities for low-abundance unsaturated free fatty acids,
as observed in an analysis of blank controls (Figure 2) run
subsequent to GC/MS analysis of derivatized plant extracts.
Despite liner exchange, however, every injection into a GC/
MSinletsysteminevitablyintroducesdirt,whichisdeposited
in the liner, septum, injector body, injector bottom plate and
injection syringe, and ultimately the first few centimeters of
the guard column. Even without automatic liner exchange, it
has been recommended that liners are manually exchanged
every 20 samples (Koek et al., 2006). Matrix deposits in the
linerandtheinjectorsystemhavedetrimentalconsequences
for the accuracy of quantitative analysis of amino acids.
Usingthestandardprotocol,acidicprotonsofallcompounds
are trimethylsilylated (TMS) to render the otherwise invola-
698 Oliver Fiehn et al.
ª 2008 The Authors
Journal compilation ª 2008 Blackwell Publishing Ltd, The Plant Journal, (2008), 53, 691–704

Page 9

tile metabolites amenable to gas chromatography (Halket
and Zaikin, 2003). It has been argued that quantitative amino
acid analysis is less accurate by TMS derivatization and GC/
MS analysis than by classic amino acid analysis, e.g. by
phtaldialdehydederivatizationandHPLCfluorescencedetec-
tion (Noctor et al., 2007). In other reports, formation of
N-TMSgroupswassaidtobedependentonthereactiontime
and temperature used during the derivatization (Gullberg
et al., 2004; Kanani and Klapa, 2007). While we cannot
confirm the latter observation, all these reports clearly
suggest that the signal intensities of the various TMS
derivatives are not robust for several amino acids.
In our experience, this lack of robustness and the change
in N-TMS derivatization status are directly related to the
status of the injector and total matrix deposition. If the
injector body, column, injector plate, liner and syringe are
kept scrupulously clean, the intensity ratio of the various
trimethylsilylated forms of amino acids can be kept con-
stant. Only if dirt accumulates or if the sample itself contains
too much matrix, will the relative peak ratios and overall
abundance of TMS-derivatized amino acids be negatively
affected. Figure 3 shows that this effect is not observed for
compounds that only bear hydroxyl- or carboxyl moieties,
such as sterols or sugars, and that some amino compounds
are more affected than others. When comparing clean
injector conditions to injector conditions for which dirt
problems were recognized (Figure 3), amino acids such as
asparagine, glutamate and serine are severely decreased,
especially at lower concentrations, whereas putrescine or
glycine are not affected. Other amino acids, such as alanine,
urea, valine and isoleucine, are also negatively affected to a
lesser extent. It is important to note that these effects are
most prominent under split-injection conditions (here at
split ratio 1/5), but are much less severe under split-less
conditions. While the differences in adverse responses
between structurally related amino acids cannot be readily
explained, all observations suggest that there is a delicate
balance between the formation and degradation of N-TMS
bonds during the injection, in which the N-TMS formation is
favored by clean and inert surfaces and by longer periods in
the reactor (i.e. split-less conditions). In order to limit matrix
effects caused by the sample matrix itself (e.g. complex
lipids), clean-up before derivatization is recommended.
Soluble metabolites should be re-dissolved in 50% aqueous
acetonitrile to remove them from insoluble complex lipids
and plant waxes that otherwise may hamper the N-TMS
stability for some amino compounds. We have found that
large amounts of sterols and free fatty acids could be
completely recovered using this clean-up step, and that
N-silylation was greatly improved for samples with high
amounts of complex lipids.
Alternatively, amino acids could be separately analyzed
using the far more stable tertiary butyldimethylsilyl deriva-
tization (Fiehnet al.,2000b;Glassopet al.,2007).However,if
adequate quality control measures are taken as discussed
above, trimethylsilylation can still be used as a universal,
mild and quick process for metabolite profiling, avoiding the
need to double or triple the number of analytical runs per
plant sample.
Data processing requires filtering of noisy and
inconsistent signals
The main task of quality control in unbiased (non-target)
metabolomics is to detect all metabolic signals and subse-
0.00
0.25
0.50
0.75
1.00
0 255075 100
% QC dilution series
Ratio poor /good injector conditions
asn
glu
ser
gly
putr.
urea
ala
sorbitol
cholesterol
Figure 3. Normalized peak intensities for nine of the 31 quality control
standards thatareinjected daily atfour concentrations.Theratioof theresults
obtained under poor and good injector conditions is given.
Linoleic acid
770rt (s) 745 755
1250
Counts sec–1
Counts sec–1
Linolenic acid
770
Linoleic acid
rt (s) 745
650
Linolenic acid
2nd
blank
5th
blank
755
Figure 2. Rice leaf extracts including lipids injected into an Agilent hot split/
splitless injector (left panel) and a Gerstel automatic liner exchange/cold
injection system (right panel). Ion traces m/z 335 + 337 are given. The
accumulation of matrix deposits without liner exchange causes increasing
levels of background signals in intermittently injected blank controls. This
cross-contamination reaches half the intensity of endogenous low-abun-
dance free fatty acid levels after as few as five samples.
Quality control in metabolomics 699
ª 2008 The Authors
Journal compilation ª 2008 Blackwell Publishing Ltd, The Plant Journal, (2008), 53, 691–704

Page 10

quently remove or combine redundant signals, filter out
noisy peaks, select consistent peaks and annotate these in
an unambiguous way as known and unknown metabolites
as recommended by the MSI document on chemical anal-
ysis (Sumner et al., 2007). A variety of commercial and free
software tools are available that perform one or several of
these tasks, but none of the tools are sufficient to annotate
known metabolites in large and diverse datasets. Mostly,
software tools assist in analyzing three-dimensional data
files such as ‘retention time * m/z value * intensity’ by
aligningchromatogramsusing
(Smith et al., 2006). However, while such tools are helpful in
detecting ‘biomarkers’ of metabolic differences among
related experimental classes in small-scale studies, align-
ments fail when comparisons between very different
samples are attempted (such as different plant organs), or
when studies with hundreds of samples are carried out that
may comprise larger shifts in retention times. The single
best way to robustly refer to locked retention times in
GC/MS is by using a grid of marker compounds, the classic
retention index. Alternatively, some GC/MS instrument
software offers instrument control of carrier gas flow to
adjust absolute retention times to a single marker com-
pound. Hence, while peak picking and alignment software
tools can result in MSI-compliant reports, comparisons of
data sets across studies and data exchange between labo-
ratories is much enhanced if internal retention markers are
used. We suggest using C8–C30 fatty acid methyl esters
because these compounds are absent in plants and gen-
erate abundant ions at high m/z values under electron
impact mass spectrometry, which facilitates automatic
recognition. Resulting retention indices can be converted to
the community standard of Kovacs indices if required.
Quality control in data processing means that the
incoming data are deconvoluted and filtered according to
consistency criteria. Depending on the complexity of the
samples, as well the peak detection and mass spectral
deconvolution software that is used, between 300 and 800
spectra are reported per chromatogram. How many of
these spectra represent authentic metabolic signals? While
important advances have been made in the quality of this
mass spectral deconvolution (Kind et al., 2007; Stein, 1999),
there will inevitably be false-positive and false-negative
spectral reports: very-low-abundance compounds adjacent
to very abundant peaks may be compromised in spectral
quality, or the presence of two very closely co-eluting
compounds may not be recognized, leading to export of
only one spectrum that comprises ions of both peaks. It is
not straightforward to filter these spectra into a unified and
consistent table of quantified metabolites. No public
repositories exist to date that would allow researchers to
carry out the task of unambiguous peak detection and
verification. A combination of mass spectral deconvolution
and alignment-type data filtering meets some of the
similarity recognition
necessary conditions to obtain unified and high-quality
data-processing results from instrument-independent input
files (Styczynski et al., 2007). For *.peg file formats (as
produced by Leco GC–TOF mass spectrometers), an algo-
rithm has been published that removes noisy and incon-
sistent spectra using a multiple filter system and that
annotates correct peaks even if mass spectral similarity is
compromised (Fiehn et al., 2005). In addition to the MSI
criteria of retention index and mass spectral similarity
matches, further peak metadata are exploited, such as peak
abundance, mass spectral purity, unique ions and apex
masses (Figure S1). Quality thresholds for peak annota-
tions are given in Table S1.
Documentation and designs of the resulting database
BinBase are available for free download at http://fiehnlab.
ucdavis.edu/staff/wohlgemuth/binbase/. Open-source code
will be available later this year. Importantly, the algo-
rithm also recognizes novel signals that had never been
detected in other samples, and enters these spectra into
the database if they are detected in at least six samples and
at least 80% of the samples of a class of the biological study
design.
Naming signals as authentic metabolites must not be
done on the basis of mass spectral matches alone, as
required by the MSI document on chemical analysis.
Instead, reference compounds or reference libraries must
confirm matching for at least two independent parameters;
for GC/MS, these are retention time and mass spectrum.
Known artifact spectra such as column bleed, phtalates or
polysiloxanes should be recognized and discarded from
use in statistics. Compounds are identified using the Fiehn
library of 713 unique metabolites as outlined above in
order to annotate spectra with correct metabolite names
and associated identifiers such as PubChem and KEGG
numbers. In addition to mass spectra as encoded in the
result files, the complete library is available as quadrupole
and time-of-flight mass spectra from the corresponding
instrument vendors. Result data sets are available from the
SetupX and NSF2010 metabolomics databases (http://
www.plantmetabolomics.org and http://fiehnlab.ucdavis.
edu:8080/m1/main_public.jsp), including replacement of
missing values using unprocessed *.cdf (ANDI MS) files
as detailed above. For any reporting of metabolomic data,
it is important to state the way in which missing values
have been replaced so that users may be aware how
confident the eventual statistical outputs are. While guess
estimates for missing values can be computed, the
approach proposed here is more favorable as it investi-
gates the authentic data acquisition files instead of assign-
ing putative values. It is important to note that the number
of missing values not only depends on the samples and
peak abundances, but first and foremost on the stringency
of quality criteria: if lax criteria for retention index windows
and mass spectral matching are used, the number of false-
700 Oliver Fiehn et al.
ª 2008 The Authors
Journal compilation ª 2008 Blackwell Publishing Ltd, The Plant Journal, (2008), 53, 691–704

Page 11

positive matches will increase and falsely mask actual
misannotations. Conversely, if the criteria are too strict, too
many false-negative peak annotations will result. We
suggest resolving this dilemma by adjusting peak criteria
based on peak metadata (e.g. peak purity and peak
abundance), by strictly annotating only one result entry
per detected peak (avoiding double annotations), and by
replacing the resulting missing data using the raw data
instead of assigning values. Lower-confidence results are
then labeled.
Presenting results of metabolite profiling by statistical
comparisons
The metabolomic workflow is completed by statistical
analysis of the quantitative data and presentation of the
results in form of graphs, which support the subsequent
biochemical or physiological interpretations. Pre-existing
knowledge shouldbe integrated
hypotheses to facilitate the interpretation of differential
regulation of metabolite levels. Discovery of unexpected
metabolic changes may then advance our knowledge of
regulation of metabolic pathways. We exemplify such
hypothesis building by discussing background information
on the current knowledge about the gene At1g08510, which
encodes FATB, an acyl-ACP (acyl carrier protein) thioester-
ase. We wish to emphasize that the data concerning the fatB
knockout allele presented herein are not sufficient to prove
the lipid-signaling hypothesis presented below; this would
require additional supporting evidence. Instead, this case is
given here as an example indiacting that the physiological
relevance of gene disruptions may often only be revealed
under testing under various environmental stress situations,
such as physical wounding.
Acyl-ACP thioesterases catalyze hydrolysis of the thio-
ester bond that links the acyl chain to the ACP carrier used in
de novo fatty acid biosynthesis. Prior characterizations of
the Arabidopsis fatB knockout mutant indicated that a
change in growth rate was associated with a 40–50%
decrease in the total amount of saturated fatty acids in
various tissues compared to wild-type, specifically affecting
the fatty acid composition of extra-plastidial phospholipids,
sphingolipids, waxes in leaves and stems, and triacylgly-
cerolsinseeds (Bonaventureet al., 2003).Incuticular waxes,
the fatB mutation leads to an 80% reduction in hexadecane
derivatives, concomitant with a reciprocal increase in unsat-
urated C18 fatty acids (Bonaventure et al., 2004b). In further
investigations, it was found that total fatty acid turnover
(both synthesis and breakdown of lipids) was increased in
order to maintain a steady export of palmitate from the
plastids(Bonaventure et al., 2004a).It was hypothesized that
cellular sensors might exist that regulate lipid synthesis
rates and plastidial palmitate export. It should be noted that,
in these earlier studies, lipid analyses were performed after
as initialworking
transmethylations, hence obstructing information about the
contents of free fatty acids. Moreover, these earlier studies
targeted analysis specifically towards lipid constituents but
disregarded potential impact on other metabolic pathways.
For the fatB knockout mutant case study, a working hypoth-
esis can thus be formulated that known changes in metab-
olite levels, such as the sphingolipid pools, might affect
signaling pathways, specifically under stress situations. This
impact on lipid-signaling pathways might subsequently
influence unrelated metabolic pathways and lead to meta-
bolic phenotypes that are more distinct under stress than
under normal environmental conditions. Consequently, the
study design included a simple physical stress imposed on
both the fatB knockout mutant and its Ws ecotype parental
genotype, using unstressed plants as negative controls for
this difference in wounding response.
Because different statistical tests use specific data treat-
ments and have a set of unique theoretical assumptions, we
recommend that a variety of statistical tests be performed in
order to verify metabolic differences. The methods and
minimal requirements of statistical tests have been com-
pared previously (Goodacre et al., 2007). It is important to
note that there is no single best statistical method or
combination of methods, but the methods used should
reflect the hypotheses and class design of the study.
Nevertheless, both univariate and multivariate tests, as well
as supervised and unsupervised tools, should be used.
Supervised methods are multivariate tests that develop
classification models based on a priori knowledge of the
study design, and hence tend to emphasize differences
between classes. Supervised methods comprise decision
trees, linear discriminant analysis or regression methods
such as partial least-square analysis (PLS). Unsupervised
tools (such as the classic principal components analysis,
PCA) are mainly used to compress data by linear combina-
tions to derive principal vectors based on the variance
inherentin thedata set.Therefore,PCAgraphswill resemble
PLS plots if the biological variance in a data set is far greater
than unrelated variance (e.g. noise), whereas PLS will focus
on biologically significant variance that is useful for building
model vectors that discriminate between classes. Super-
vised model building is at risk of data over-fitting, especially
when a low number of samples and a high number of
variables (here metabolites) are used. Univariate data anal-
ysis may be presented after analysis of variance (ANOVA) by
significance thresholds (e.g. P < 0.05); however, data are
often further validated by reducing the number of potential
falsepositivesusingBonferroniadjustments(i.e. bydividing
the P value by the overall number of tested variables to
obtain stricter P values).
The finding of general differences in metabolic responses
should be exemplified by multivariate and univariate data
analysis (Figure 4). The overall extent of differences in
metabolic phenotypes can be assessed by multivariate
Quality control in metabolomics 701
ª 2008 The Authors
Journal compilation ª 2008 Blackwell Publishing Ltd, The Plant Journal, (2008), 53, 691–704

Page 12

statistics such as PLS. The PLS vector with the highest score
distinguished the fatB mutant under wound response from
itswild-typecounterpart(Figure 4,leftpanel),indicatingthat
the physiological roles of the fatB gene involve control of
wound responses. Vector 2 emphasized generic differences
between control and wound responses for both genotypes,
whereas vector 3 best separated the fatB knockout mutant
from the wild-type phenotypes under both treatments.
Therefore, each parameter of the study design could be
dissected and led to specific metabolic differences that
discriminated the metabolic phenotypes.
Univariate statistical analysis is used to investigate
changes for individual compounds. A total of 369 metabolic
signals were reported by BinBase after final data processing,
of which 64% were found to be significantly altered in the
fatB knockout allele compared to the wild-type after wound-
ing (P < 0.05, one-way ANOVA). In comparison, only 11% of
the metabolites were found to be significantly altered under
control conditions. In accordance with the unaltered export
rates of palmitate in fatB mutants (Bonaventure et al.,
2004a), we found that concentrations of endogenous free
fatty acids (both unsaturated and saturated) remained
unchanged under control conditions, including those of
the potential precursors for membrane lipids such as
monopalmitin and monostearin. However, upon wounding
stress, the metabolic response in the fatB mutant greatly
differed from that of the wild-type control (Figure 4, right
panel). Free palmitate levels were only slightly decreased
upon wounding in the fatB knockout mutant, while short-
chain free fatty acids such as lauric and capric acids, as well
asmonoacylglycerols, weresignificantlyreducedby30–66%
2 h after wounding in the mutant, but not in the Ws control
genotype. Further metabolic responses in the fatB knockout
mutant included up-regulation of the hydroxylated short-
chain fatty acid 2-hydroxyvaleric acid. Low-abundance odd-
chain free fatty acids such as pelargonate, but also free
stearate and the C16/C18 ratio, remained unchanged under
all conditions. In addition to these specific differences, a
number of wound responses were found to be similar
between both genotypes, for example up-regulation of
amino acids and increased levels of ascorbate. Therefore,
the fatB gene may be assumed not to be the only regulatory
gene in wound responses.
It should be noted that use of so-called box–whisker plots
(Figure 4, right panel) gives more information than classic
bar diagrams because mean values, error bars, quartiles and
outliers are compared. Therefore, the use of box–whisker
plots rather than bar diagrams is encouraged. Due to the
number of detected peaks, not all metabolites can be
displayed on graphs. Some may be mentioned in the text
whileothermetabolicchangesmightbeleftunspecified.The
graph presentations, tables or text discussions of metabolic
differences are usually targeted to support a certain inter-
pretation that authors wish to convey. However, it is impor-
tant that data are not presented in a way that distorts the
wealth of detected differences. We therefore present here a
plot of tryptophan levels (Figure 4, right panel) in addition to
the lipid focus of the study to support the notion that some
wounding responses were generic and unaltered in the fatB
knockout genotype compared to its Ws genetic background,
and we mention a variety of other metabolite classes for
whichsignificantdifferenceswereobserved.Additionally,all
significant differences should be given at least in the form of
processed data, for example in supplementary tables. For
ourtestcase,wegoonestepfurtherbyallowingscientiststo
download all processed and original data from http://
www.plantmetabolomics.org
edu:8080/m1/main_public.jsp to verify statistical differences
that are claimed here or to perform further tests.
As stated above, this example data set can be interpreted
as indicating that metabolic remodeling in the wound
response is severely affected in the fatB mutant, and this
largely outweighs the metabolic differences between the
genotypes under control conditions. This observation sup-
ports the hypothesis that control of physiological stress
conditions (such as the wounding response) may be a
functional role for this gene; however, the lack of supple-
mentary lines of evidence (e.g. time-course studies, gene
expression or studying putative signal receptors and signal
molecules) does not allow us to draw further conclusions on
the validity of the lipid-signaling hypothesis (Bonaventure
et al., 2004a,b).
orhttp://fiehnlab.ucdavis.
P L S v e c t o r 1
P L S ve c t or 2
Ws_wounded
Ws_control
FatB_control
FatB_wounded
P L S v e c t o r 3
FatB
ctrl.
FatB
wd.
Ws
ctrl.
Ws
wd
2000
3000
4000
5000
Tryptophan
FatB
ctrl.
FatB
wd.
Ws
ctrl.
Ws
wd
500
800
1100
1400
1-Monopalmitin
Figure 4. Example of statistical representationof
genotype and treatment effects on Arabidopsis
metabolism. Left panel: partial least-square mul-
tivariate statistics on overall metabolic pheno-
types of fatB wounded plants and Wassilewskija
controls. Right panels: examples of metabolites
responding to wounding in a generic and a
FATB-specific way, using univariate analysis of
variance.
702 Oliver Fiehn et al.
ª 2008 The Authors
Journal compilation ª 2008 Blackwell Publishing Ltd, The Plant Journal, (2008), 53, 691–704

Page 13

Conclusions
The value of metabolomic data is directly proportional to the
quality of the associated metadata. Thus, metadata that
define parameters associated with the source of biological
materials, the protocols for metabolite extraction, identifi-
cation and quantification, and the procedures for data
analysis and interpretation, are essential for integrating
metabolomics datasets. Uniformity in such metadata will
provide a means of querying complex datasets using a
specific hypothesis that can be assessed from the collected
datasets (e.g. stress associated or organ-specific metabo-
lites). Queries could then compare studies across species
to distinguish species-specific from generic responses. In
essence, the usefulness of metabolomics depends on the
quality of the acquired data as well as the level of detail used
in describing the studies.
The examples given here should show that quality control
extends far beyond automatic routines, such as the internal
calibrationofamassspectrometer.Qualitycontrolisinstead
an attitude of diligence that seeks to optimize procedures.
Here we have focused on several aspects of the metabolo-
mic workflow, specifically on GC–TOF metabolite profiling.
We have outlined a good-practice document, including
consistent data-processing algorithms and database que-
ries. However, we have not covered all pitfalls in a compre-
hensive manner, and more observations and improvements
will alter our standard operating procedures and algorithms.
Theusefulnessof plantmetaboliteprofilingwillbenefitfrom
establishing open access MSI-compliant repositories.
In a similar manner as presented here for GC–TOF
metabolite profiling, quality controls and procedures should
be developed for LC/MS-based approaches, direct-infusion
MS lipid fingerprints and NMR. The Metabolomics Society
has established an initiative to foster work in this direction,
and MSI documents are publicly available at the initiative’s
public web page (http://msi-workgroups.sourceforge.net/).
Comments on any of the suggested minimal reporting
standards are welcome, and may be sent to Msi-work-
groups-feedback@lists.sourceforge.net.
Acknowledgments
This work presented here has been mainly funded by research grant
5R01ES13932 from the US National Institute of Environmental
Health Sciences and grant MCB-0520140 from the US National
Science Foundation. Additional contributions and projects by cor-
porate sponsors Agilent (Santa Clara, CA), Leco (St Joseph, MI),
Monsanto (St Louis, MO) and DuPont-Pioneer (Wilmington, DE)
helped in improving the protocols, technologies and databases. We
are grateful for the solubility calculations performed using COS-
MOfrag by Andreas Klamt (COSMOlogic GmbH and Co. KG,
Leverkusen, Germany). We further appreciate the diligence of all the
contributors with regard to the minimum reporting standards
within the Metabolomics Standards Initiative. We acknowledge the
contribution by Katayoon Dehesh (University of California at Davis),
who provided seeds of the fatB knockout mutant first described by
Bonaventure et al. (2003).
Supplementary Material
The following supplementary material is available for this article
online:
Figure S1. The BinBase algorithm for annotation of GC–TOF mass
spectra.
Table S1. Spectrum similarity criteria in BinBase.
This material is available as part of the online article from http://
www.blackwell-synergy.com.
Pleasenote:Blackwellpublishing arenotresponsible forthecontent
or functionality of any supplementary materials supplied by the
authors. Any queries (other than missing material) should be
directed to the corresponding author for the article.
References
Bonaventure, G., Salas, J.J., Pollard, M.R. and Ohlrogge, J.B. (2003)
Disruption of the FATB gene in Arabidopsis demonstrates an
essential role of saturated fatty acids in plant growth. Plant Cell,
15, 1020–1033.
Bonaventure, G., Ba, X.M., Ohlrogge, J. and Pollard, M. (2004a)
Metabolic responses to the reduction in palmitate caused by
disruption of the FATB gene in Arabidopsis. Plant Physiol. 135,
1269–1279.
Bonaventure, G., Beisson, F., Ohlrogge, J. and Pollard, M. (2004b)
Analysis of the aliphatic monomer composition of polyesters
associated with Arabidopsis epidermis: occurrence of octadeca-
cis-6,cis-9-diene-1,18-dioate as the major component. Plant J. 40,
920–930.
David,F.,Devos, C.,Joulain,D., Chaintreau, A.and Sandra,P. (2006)
Determination of suspected allergens in non-volatile matrices
using PTV injection with automated liner exchange and GC-MS.
J. Sep. Sci. 29, 1587–1594.
Denkert, C., Budczies, J., Kind, T., Weichert, W., Tablack, P.,
Sehouli, J., Niesporek, S., Konsgen, D., Dietel, M. and Fiehn, O.
(2006) Mass spectrometry-based metabolic profiling reveals
different metabolite patterns in invasive ovarian carcinomas and
ovarian borderline tumors. Cancer Res. 66, 10795–10804.
Fiehn, O. (2002) Metabolomics – the link between genotypes and
phenotypes. Plant Mol. Biol. 48, 155–171.
Fiehn, O. (2006) Metabolite profiling in Arabidopsis. In Methods in
Molecular Biology (Salinas, J. and Sanchez-Serrano, J.J., eds).
Totowa, NJ: Humana Press, pp. 439–447.
Fiehn, O., Kopka, J., Dormann, P., Altmann, T., Trethewey, R.N. and
Willmitzer, L. (2000a) Metabolite profiling for plant functional
genomics. Nat. Biotechnol. 18, 1157–1161.
Fiehn, O., Kopka, J., Trethewey, R.N. and Willmitzer, L. (2000b)
Identification of uncommon plant metabolites based on calcula-
tion of elemental compositions using gas chromatography and
quadrupole mass spectrometry. Anal. Chem. 72, 3573–3580.
Fiehn, O., Wohlgemuth, G. and Scholz, M. (2005) Setup and anno-
tation of metabolomic experiments by integrating biological and
mass spectrometric metadata. LNCS, 3615, 224–239.
Fiehn, O., Robertson, D., Griffin, J. et al. (2007a) The Metabolomics
Standards Initiative (MSI). Metabolomics, 3, 175–178.
Fiehn, O., Sumner, L.W., Rhee, S.Y., Ward, J., Dickerson, J., Lange,
B.M., Lane, G., Roessner, U., Last, R. and Nikolau, B. (2007b)
Minimum reporting standards for plant biology context infor-
mation in metabolomic studies. Metabolomics, 3, 195–201.
Quality control in metabolomics 703
ª 2008 The Authors
Journal compilation ª 2008 Blackwell Publishing Ltd, The Plant Journal, (2008), 53, 691–704

Page 14

Glassop, D., Roessner, U., Bacic, A. and Bonnett, G.D. (2007)
Changes in the sugarcane metabolome with stem development.
Are they related to sucrose accumulation? Plant Cell Physiol. 48,
573–584.
Goodacre, R., Broadhurst, D., Smilde, A.K. et al. (2007) Proposed
minimum reporting standards for data analysis in metabolomics.
Metabolomics, 3, 231–241.
van der Greef, J., Stroobant, P. and van der Heijden, R. (2004) The
role of analytical sciences in medical systems biology. Curr. Opin.
Chem. Biol. 8, 559–565.
Grob, K. and Biedermann, M. (2000) Video-taped sample evapora-
tion in hot chambers simulating gas chromatography split and
splitless injectors – II. Injection with band formation. J. Chroma-
togr. A, 897, 247–258.
Gullberg, J., Jonsson, P., Nordstrom, A., Sjostrom, M. and Moritz,
T. (2004) Design of experiments: an efficient strategy to identify
factors influencing extraction and derivatization of Arabidopsis
thaliana samples in metabolomic studies with gas chromatogra-
phy/mass spectrometry. Anal. Biochem., 331, 283–295.
Hajslova, J., Holadova, K., Kocourek, V., Poustka, J., Godula, M.,
Cuhra, P. and Kempny, M. (1998) Matrix-induced effects: a critical
point in the gas chromatographic analysis of pesticide residues.
J. Chromatogr. A, 800, 283–295.
Halket, J.M. and Zaikin, V.G. (2003) Derivatization in mass spec-
trometry – 1. Silylation. Eur. J. Mass Spectrom., 9, 1–21.
Kanani, H.H. and Klapa, M.I. (2007) Data correction strategy for
metabolomics analysis using gas chromatography–mass spec-
trometry. Metab. Eng. 9, 39–51.
Katajamaa, M., Miettinen, J. and Oresic, M. (2006) MZmine: toolbox
for processing and visualization of mass spectrometry based
molecular profile data. Bioinformatics, 22, 634–636.
Kind, T., Tolstikov, V., Fiehn, O. and Weiss, R.H. (2007) A compre-
hensive urinary metabolomic approach for identifying kidney
cancer. Anal. Biochem. 363, 185–195.
Klamt, A., Eckert, F., Hornig, M., Beck, M.E. and Burger, T. (2002)
Prediction of aqueous solubility of drugs and pesticides with
COSMO-RS. J. Comput. Chem. 23, 275–281.
Koek, M.M., Muilwijk, B., van der Werf, M.J. and Hankemeier, T.
(2006) Microbial metabolomics with gas chromatography/mass
spectrometry. Anal. Chem. 78, 1272–1281.
Lisec, J., Schauer, N., Kopka, J., Willmitzer, L. and Fernie, A.R.
(2006) Gas chromatography mass spectrometry-based metabo-
lite profiling in plants. Nat. Protocols, 1, 387–396.
Meyer, R.C., Steinfath, M., Lisec, J. et al. (2007) The metabolic sig-
nature related to high plant growth rate in Arabidopsis thaliana.
Proc. Natl Acad. Sci. USA, 104, 4759–4764.
Modi, A.T., Streeter, J.G. and McDonald, M.B. (2000) Relative effi-
ciency of ethanol and pyridine as extractants of low molecular
weight carbohydrates from soybean seed axes. Seed Sci. Tech-
nol. 28, 193–200.
Noctor, G., Bergot, G.L., Mauve, C., Thominet, D., Lelarge-Trouv-
erie, C. and Prioul, J.L. (2007) A comparative study of amino
acid measurement in leaf extracts by gas chromatography–time
of flight–mass spectrometry and high performance liquid
chromatography with fluorescence detection. Metabolomics, 3,
161–174.
Rhee, S.Y., Beavis, W., Berardini, T.Z. et al. (2003) The Arabidopsis
Information Resource (TAIR): a model organism database pro-
viding a centralized, curated gateway to Arabidopsis biology, re-
search materials and community. Nucleic Acids Res. 31, 224–228.
Roessner, U., Luedemann, A., Brust, D., Fiehn, O., Linke, T.,
Willmitzer, L. and Fernie, A.R. (2001) Metabolic profiling allows
comprehensive phenotyping of genetically or environmentally
modified plant systems. Plant Cell, 13, 11–29.
Sanz, J., Soria, A.C. and Garcia-Vallejo, M.C. (2004) Analysis of
volatile components of Lavandula luisieri L. by direct thermal
desorption–gas chromatography–mass spectrometry. J. Chro-
matogr. A, 1024, 139–146.
Scholz, M. and Fiehn, O. (2007) SetupX – a public study design
database for metabolomic projects. Pacific Symposium on Bio-
computing, 12, 169–180.
Smith, C.A., Want, E.J., O’Maille, G., Abagyan, R. and Siuzdak, G.
(2006) XCMS: processing mass spectrometry data for metabolite
profiling using nonlinear peak alignment, matching, and identi-
fication. Anal. Chem. 78, 779–787.
Stein, S.E. (1999) An integrated method for spectrum extraction and
compound identification from gas chromatography/mass spec-
trometry data. J. Am. Soc. Mass Spectrom. 10, 770–781.
Styczynski,M.P.,Moxley, J.F.,Tong,L.V.,Walther,J.L.,Jensen,K.L.
and Stephanopoulos, G.N. (2007) Systematic identification of
conserved metabolites in GC/MS data for metabolomics and
biomarker discovery. Anal. Chem. 79, 966–973.
Sumner, L.W., Amberg, A., Barrett, D. et al. (2007) Proposed mini-
mum reporting standards forchemical analysis.Metabolomics,3,
211–221.
Weckwerth, W., Wenzel, K. and Fiehn, O. (2004) Process for the
integrated extraction identification, and quantification of metab-
olites, proteins and RNA to reveal their co-regulation in bio-
chemical networks. Proteomics, 4, 78–83.
Zhang, P., Foerster, H., Tissier, C.P., Mueller, L., Paley, S., Karp, P.D.
and Rhee, S.Y. (2005) MetaCyc and AraCyc. Metabolic pathway
databases for plant research. Plant Physiol. 138, 27–37.
704 Oliver Fiehn et al.
ª 2008 The Authors
Journal compilation ª 2008 Blackwell Publishing Ltd, The Plant Journal, (2008), 53, 691–704

Page 15

import
validation
marker det. RI calculation
postmatchingexport
pass
purity
+ s/n
pass
pass
pass
pass
purity
+ s/n
pass
s/n
seq.
BIN
ratio
sim
uni
RI
s/n
MS
class
newBIN
discard
iso