Doping Control Using High and Ultra-High Resolution

Mass Spectrometry Based Non-Targeted Metabolomics-A

Case Study of Salbutamol and Budesonide Abuse
Agneta Kiss1☯, Marianna Lucio2☯, Aurélie Fildier1, Corinne Buisson3, Philippe Schmitt-Kopplin2,4, Cécile
Cren-Olivé1*
1 EquipeTRACES, Institut des Sciences Analytiques-UMR, Villeurbanne, France, 2 Helmholtz Zentrum München, German Research Center for Environmental
Health, Neuherberg, Germany, 3 Département des analyses, Agence Française de Lutte contre le Dopage (AFLD), Châtenay-Malabry, France, 4 Chair of
Analytical Food Chemistry, Technische Universität München, Freising-Weihenstephan, Germany

Abstract

We have detected differences in metabolite levels between doped athletes, clean athletes, and volunteers (non
athletes). This outcome is obtained by comparing results of measurements from two analytical platforms: UHPLC-
QTOF/MS and FT-ICR/MS. Twenty-seven urine samples tested positive for glucocorticoids or beta-2-agonists and
twenty samples coming from volunteers and clean athletes were analyzed with the two different mass spectrometry
approaches using both positive and negative electrospray ionization modes. Urine is a highly complex matrix
containing thousands of metabolites having different chemical properties and a high dynamic range. We used
multivariate analysis techniques to unravel this huge data set. Thus, the several groups we created were studied by
Principal Components Analysis (PCA) and Partial Least Square regression (PLS-DA and OPLS) in the search of
discriminating m/z values. The selected variables were annotated and placed on pathway by using MassTRIX.

Citation: Kiss A, Lucio M, Fildier A, Buisson C, Schmitt-Kopplin P, et al. (2013) Doping Control Using High and Ultra-High Resolution Mass Spectrometry
Based Non-Targeted Metabolomics-A Case Study of Salbutamol and Budesonide Abuse. PLoS ONE 8(9): e74584. doi:10.1371/journal.pone.0074584
Editor: Francisco J. Esteban, University of Jaén, Spain
Received February 12, 2013; Accepted August 1, 2013; Published September 18, 2013
Copyright: © 2013 Kiss et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits
unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: The authors would like to thank the Rhone-Alpes Regional Council for the PhD funding, the Procope Program for the Franco-German
collaboration, and the French Anti-Doping Agency for the urine samples. The funders had no role in study design, data collection and analysis, decision to
publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
* E-mail: [email protected]
☯ These authors contributed equally to this work.

Introduction our understanding of the biochemical transformations. Among

these, two are believed to have played a key role: the
The molecular diversity of the human urinary metabolome is introduction of high (Time of Flight mass spectrometry, TOF)
very well reflected by the existing databases [1,2] displaying and ultra-high resolution techniques (Fourier Transform Ion
thousands of metabolites classified in as much as 70 different Cyclotron Resonance mass spectrometry, FT-ICR/MS) and the
structural classes [3-5]. The majority of these metabolites are development of algorithms capable of handling the thousands
small molecular weight compounds having molecular weights of signals generated by such analytical platforms. Indeed,
between 100 and 800 Dalton (Da) and small peptide fragments techniques such as FT-ICR/MS are becoming more and more
[6] with very different physico-chemical properties (solubility, available and their advantages can be now fully exploited.
polarity, proton affinity, etc). Despite this astonishing chemo- Thus, molecular formulae generation based on exact masses
diversity, for quite some time, mainly targeted studies, often and relevant database search are now possible due to the high
restricted to a particular chemical family or to compounds resolution and accuracy of this type of technique [7,8]. Of equal
having similar properties have been applied into doping control importance, the recent advances in the pre-processing [9],
studies. Moreover, the metabolites and the changes were often mathematical modeling [10,11] and the statistical analysis [10]
regarded in an univariate manner and the correlations between lead to more comprehensive biological interpretation of the
them were often disregarded. metabolomics data [12]. Up to present, this strategy has been
Nowadays, the progresses made in analytical fields like applied in a variety of fields, including drug discovery [13,14],
sample preparation, chemical analysis and data processing nutrition [15,16], toxicology [17], clinical trials [18] and more
offer a wider view of the metabolome and greatly contribute to recently chemical submission [19]. It is of particular interest in

PLOS ONE | www.plosone.org 1 September 2013 | Volume 8 | Issue 9 | e74584

 Non-Targeted Metabolomics for Doping Control

Figure 1. Study design and general strategy.

doi: 10.1371/journal.pone.0074584.g001

areas like doping control as new approaches are needed to and differences. Finally, we compare the data obtained by
fight the ongoing development of performance-enhancing direct injection into the FT-ICR mass spectrometer with the one
methods [20-23]. obtained by including a separation step in the analysis
In this study, this strategy was applied to real urine samples framework. Thus, the LC-QTOF data set was compared with
collected from volunteers (V), high-level clean athletes (CA) the FT-ICR/MS data set by means of PCA-CCA (Principal
and athletes declared positive for the use of beta-2-agonists Components Analysis-Canonical Correlation Analysis).
(salbutamol, S) or glucocorticoids (budesonide, B). The choice
of the doping agents was based on three criteria: the high
Sample Collection
frequency detection in sport [24], the common medical use in
The protocol was approved by the Institutional Review Board
asthma treatment and respiratory conditions [25] and the
excretion through urine [26,27]. The main purposes of this of the University of Lyon, F-69622, forty-three bd du 11
paper are to explore (i) the metabolic differences between novembre, Villeurbanne, and the Institutional Review Board of
volunteers and sportsmen and (ii) the metabolic changes the French Anti-Doping Agency, 143, avenue Roger Salengro,
induced by the intake of budesonide and salbutamol. Châtenay-Malabry (according to the Declaration of Helsinki).
All subjects gave written informed consent before the study
Materials and Methods commenced. The investigation was conducted in accordance
with the ethical principles of Good Clinical Practice. The
Study design and general strategy experiment includes 11 urine samples coming from clean
Firstly, the spectra acquired with FT-ICR/MS were athletes, 9 urine samples collected from volunteers and 27
preprocessed and prepared for analysis. Several groups urine samples provided by the French Anti-Doping Agency
(Figure 1) were selected and treated independently via (AFLD, Chatenay-Malabry) and coming from athletes declared
mathematical and statistical methods such as Principal positive for illicit substance consumption. Among the 27
Component Analysis (PCA, unsupervised) and Orthogonal positive samples, 17 samples belong to salbutamol doped
Partial Least Square (OPLS, supervised). The two models athletes and 10 samples belong to budesonide doped athletes.
involving budesonide and salbutamol treated athletes were The urine samples coming from volunteers were collected in
then compared in order to emphasize the metabolic similarities comparable conditions and by following the AFLD guidelines.

PLOS ONE | www.plosone.org 2 September 2013 | Volume 8 | Issue 9 | e74584

 Non-Targeted Metabolomics for Doping Control

All urine samples were aliquoted and stored at -24°C until following conditions: nebulizer pressure 1 bar, dry gas flow 8 L/
analysis, as recommended for stability reasons [27]. min, source temperature 120 °C, source temperature 200 °C,
capillary voltage 3500 V. Mass spectrometric data were
Sample Preparation collected over the range 80-1000 m/z in profile mode at 1
A “quality control” (QC) sample was prepared as spectra/second. A sodium formate solution was injected for 0.1
recommended in different publications [28-30] by mixing equal min at the beginning of each run for internal calibration.
volumes (50 µL) of each of the samples as they were being
aliquoted. This “pooled” urine was used to provide a Data Analysis
representative mean sample containing all the analytes that will The data analysis procedure followed in this paper consists
be encountered during the analysis and it was mainly used to of: internal calibration, peak identification, peak list alignment,
assess the quality of the data. At the beginning of the batch, 3 mathematical and statistical analysis and database
methanol samples were run to ensure that the analytical annotations.
system had come to equilibrium. The QC samples were The raw data (FT-ICR/MS and UHPLC-QTOF) were
injected once at the beginning, in the middle and at the end of independently pre-processed with the proprietary software
the batch. For the FT-ICR analysis, all the samples were DataAnalysis (Bruker Daltonik, Bremen). The FT-ICR/MS
diluted 50 times in methanol and vortexed for 2 minutes at acquired spectra were internally calibrated using a reference
10000 rpm. The supernatant was then taken and injected. For list: solvent impurities for the positive mode and fatty acids for
the UHPLC-QTOF analysis the urine samples were injected as negative mode in order to account for deviations occurring
neat. during the batch. Calibrations errors were found to be inferior to
0.2 ppm regardless of the ionization mode. Next, the peaks
Experimental Conditions having a signal-to-noise ratio superior to four (4) were exported
Fourier Transform Ion Cyclotron (FT-ICR MS). High- to individual peak lists. The peak lists were then aligned over
resolution mass spectra were acquired on a Bruker (Bremen, the entire mass range through an in-house software with a
Germany) SolariX Fourier Transform Ion Cyclotron Resonance tolerance of 1 ppm [31]. Profile Analysis (Bruker Daltonics) was
MS equipped with a 12 Tesla superconducting magnet and an used to internally calibrate and align the UHPLC-QTOF data in
Apollo II electrospray source. The high magnetic field FT-ICR both m/z and retention time directions Tolerances of ±0.05 Da
instrument combines the highest resolution (400.000 at m/z and ±0.2 min were used, respectively.
400) and the best mass accuracy (<200 ppb) and thus enables Prior to data analysis the FT-ICR/MS and UHPLC-QTOF
the direct conversion of the experimental mass into elementary data sets were Pareto-scaled. Modeling and statistical analysis
composition. The urine samples were infused with the micro- were done with SIMCA-P+12 (Umetrics, Umeå, Sweden) and
electrospray source at a flow rate of 120 mL/h The The in MATLAB (v.10, Mathworks, USA). Several techniques were
nebulizer gas pressure and the drying gas pressure were fixed used in order to extract the information contained in the
at 20 psi and 15 psi, respectively. The source temperature was multivariate data sets. Firstly, Principal Component Analysis
250°C. Data was acquired in both positive and negative (PCA) was used in order to have a first overview of the FT-
electrospray modes. The mass spectrometer was externally ICR/MS data and to detect general trends and relations
calibrated at the beginning of the batch. The calibration solution between the spectra and the variables (m/z). Orthogonal Partial
consisted of arginine clusters (m/z of 173.10440, 347.21607, Least Squares regression (OPLS), was used in order to select
521.32775 and 695, 43943) dissolved in methanol at a those signals related to the nature of samples e.g. treated,
concentration of 10 mg/L. Calibration errors in the relevant clean, volunteer and athlete (supervised technique). As
mass range were always below 0.1 ppm (part per million). showed in Figure 1, the models were built for each comparison.
Liquid chromatography coupled to quadrupole time of These models were firstly validated by the 7 fold cross
flight mass spectrometry (UHPLC-QTOF-MS). The UHPLC- validation procedure implemented in the software and then with
MS platform consisted of a chromatographic system suitable CV-ANOVA (Cross Validation ANOVA), in order to exclude the
for high resolution separations (Dionex Ultimate 3000) coupled possible presence of over-fitting. Pertinence indicators such as:
with a quadrupole time-of-flight (QTOF) mass detector p-value, R2(Y) and Q2(cum), were subsequently reported for
equipped with electrospray ionization (microTOF QII, Bruker each model indicating the significance, the goodness of the fit
Daltonics). The chromatographic separation was conducted on and the goodness in the model prediction. Moreover potential
an Acquity C18 BEH column (100mm x 2.1 mm, 1.7 µm). The discriminating variables were chosen by examining the S-plot
column oven temperature was set to 50°C, injection volume at (which combines the modeled covariance and correlation,
5 µL and flow-rate at 0.5 ml/min without a split. The eluents variables with high value in magnitude and reliability can be
used were: (A) water acidified with 0.1% formic acid and (B) candidate for the class separation) and VIP values (variable
acetonitrile with 0.1% formic acid. The gradient was set as influence in the projection). Additionally graphical methods
follows: 100% (A) over 2 min followed by a linear increase from such as SUS-plot (Shared and Unique Structures) were used
0 to 100% (B) over 30 min. An isocratic cleaning step at 100% for model comparison. Canonical Correlation Analysis
(B) over 5 min and column equilibration step at 100% (A) over preceded by PCA was done in MATLAB in order to compare
15 min were used after each sample. Each urine sample was the output of the two platforms (UHPLC-QTOF and FT-ICR).
injected three times in order to assure reproducibility. Detection The masses characterizing the metabolic distinction were
was performed in negative electrospray mode under the submitted to MassTRIX using Homo sapiens as reference

PLOS ONE | www.plosone.org 3 September 2013 | Volume 8 | Issue 9 | e74584

 Non-Targeted Metabolomics for Doping Control

Figure 2. Distribution of the signals as a function of the mass range.

doi: 10.1371/journal.pone.0074584.g002

species and a maximum error of 1 ppm. This web server

assigns a potential annotation from KEGG (http://
www.genome.jp/kegg/), LipidMaps (http://www.lipidmaps.org/)
and HMDB [4,5] to each m/z value. As second step, MassTRIX
calls the KEGG/API (http://www.genome.jp/kegg/soap/) to
generate the pathway maps of the annotated masses [6].

Results and Discussion

FT-ICR/MS
The alignment of the FT-ICR/MS peak lists lead two matrices
corresponding to the two ionization modes. Initially the positive
and the negative matrices counted 23.246 signals and 20.830
signals, respectively. The signals appearing in less than 5% of
the samples were associated to noise and were excluded from
further analysis. Thus, the positive and negative matrices Figure 3. UHPLC-QTOF/MS chromatogram.
underwent a serious downsizing to 63% and 89% of the initial doi: 10.1371/journal.pone.0074584.g003
number of signals, respectively. The majority of the remaining
signals are situated between 150 and 500 Da and its
unreliable and were not included in the statistical analysis. 104
distribution is shown in Figure 2.
variables were detected in all the samples. The majority of the
detected compounds have molecular weights between 100 and
UHPLC-QTOF/MS
700 Da. Since the distribution of the detected signals over the
The positive mode (Figure 3) yielded 47133 variables. Some
mass range is sensitive to the acquisition parameters of the
of these are metabolites while others are fragments, adducts,
isotopes and noise. 40457 variables are present in less than mass spectrometer, a direct comparison of two distributions
five samples. As before, these variables were considered (UHPLC-QTOF and FT-ICR) would not be appropriate.

PLOS ONE | www.plosone.org 4 September 2013 | Volume 8 | Issue 9 | e74584

 Non-Targeted Metabolomics for Doping Control

Figure 4. Principal Component Analysis showing clusters corresponding to the 4 groups (B, S, V, CA).
doi: 10.1371/journal.pone.0074584.g004

A detailed analysis of the data emphasises a wide range of applied to the totality of the samples, a clustering of some
polarity. Approximately 16% of the compounds are eluted samples becomes apparent by eye (see Figure 4). The 3D
before 20% acetonitrile and the vast majority of compounds are score plot shows that the (clean) volunteers (V) are more
eluted before reaching 75% acetonitrile. dispersed along the second principal component (Figure 4)
suggesting a larger variability in the metabolism or the
Unsupervised Statistical Analysis interference of different factors (diet, sport, medication, etc.).
The PCA is used in order to get a first glance of the The loadings plots are extremely complex and difficult to
properties of the FT-ICR data. As a result of this analysis interpret due to the presence of many different groups. In order

PLOS ONE | www.plosone.org 5 September 2013 | Volume 8 | Issue 9 | e74584

 Non-Targeted Metabolomics for Doping Control

Table 1. Significance values for the OPLS models (CA vs annotated compounds are shown in Figure 6a (white columns).
B), (CA vs S) and (CA vs V). A large part of the uptrend metabolites were identified as
belonging to steroid hormone biosynthesis, pentose and
glucuronate interconversions, starch and sucrose metabolism,
ESI+ ESI- ABC transporters, galactose metabolism and some amino
p R2 Q2 p R2 Q2 acids metabolism (arginine, purine, hystidine and proline). The
FT-ICR/MS models metabolites having downtrend relative intensities are situated
CA vs V 0.003 0.99 0.711 0.007 0.959 0.563 on the starch and sucrose and pentose and glucuronate
CA vs S 0.02 0.988 0.467 0.03 0.979 0.464 interconversions pathways.
CA vs B 0.07 0.954 0.389 0.06 0.697 0.29 In the negative mode, among the 11% of the discriminating
LC-QqToF models signals, almost 4% have an uptrend behavior with respect to
CA vs V 0.003 0.99 0.69 0.007 0.93 0.55 the group of CA. As in the positive mode the steroid hormone
CA vs S 0.03 0.97 0.47 0.03 0.98 0.5 biosynthesis and pentose and glucuronate inetconversion
CA vs B 0.08 0.90 0.28 0.06 0.7 0.14 pathways appear to have different trends in the two groups
doi: 10.1371/journal.pone.0074584.t001 (Figure 6b, white columns). Phenylalanine metabolism seems
to be another important pathway since as much as 8
to enhance the differences between the groups and isolate the compounds have a downtrend profile with respect to CA.
masses that discriminate them, supervised statistical methods Unlike the positive mode which counts 18% discriminant
were used. Models opposing the clean athletes and the signals annotated by MassTRIX, the negative mode, which is
volunteers (CA vs V), the clean athletes and the budesonide less universal, counts around 9% annotated discriminant
treated athletes (CA vs B) and the clean athletes and the signals.
salbutamol treated athletes (CA vs S) were created by using
the OPLS statistical method. The values of significance for the (S vs CA) and (B vs CA): Salbutamol/Budesonide
FT-ICR models and UHPLC-QTOF models are reported in treated athletes vs clean athletes
Table 1. As previously, the OPLS models were built for each of the
two comparisons and for each of the two ionization modes. The
(CA vs V): clean athletes vs volunteers comparison between the clean athletes and the salbutamol
The aim of this study is to explore all the possible differences treated athletes (S vs CA) shows good performance indicators,
between the two groups from a global point of view and to irrespective of whether we look to the positive or to the
emphasize the extent to which these differences interfere with negative mode. On the contrary, the model opposing the clean
metabolomics studies [32]. The differences between the athletes and the budesonide treated athletes (B vs CA) is
metabolism of the non-athletes and that of high level athletes weaker and the p-value for the cross-validation is above the
have already been the subject of many research papers. Thus, considered threshold (0.05) (see Table 1). This observation
studies comparing the effects of endurance [33], sports type could be explained either by a weaker effect of medication or
[34], effort degree and duration [35,36], can be found in
by an insufficient number of samples, as seen elsewhere.
literature. As the present study involves real urine samples, the
Among the 1602 discriminant masses (~7%) selected in
emphasized contrast will originate not only in the physical
positive electrospray, salbutamol was found as a highly
effort, but also in any dietary, lifestyle or medical differences.
discriminant signal (Figure 7b). Its presence was confirmed by
The OPLS model comparing the 10 volunteers to the 11
exact mass measurement (error 0.1ppm) and the assessment
athletes controlled negatively by the AFLD, resulted in
of the isotopic profile and testifies to the feasibility of this
excellent R2(Y) (0.99), Q2(cum) (0.711) and p-values. (0.003).
approach.
These values prove a strong relationship between the matrix of
Figure 8a shows a higher number of downtrend then uptrend
the measured variables (m/z signals) and the class
signals with respect to salbutamol treated athletes (black
membership in both positive and negative modes (Table 1).
Several discriminant signals were emphasized and selected columns). Also, the metabolic pathways that seem to be
based on the S-plot and the VIP values. An explanatory modified are those involving protein digestion, amino acids
example of discriminant profiles having downtrend or uptrend (arginine, proline, tryptophan, and lysine) [37-41] and pentose.
behaviors with respect to CA are shown in Figure 5a (creatine) According to the same figure, budesonide intake plays an
and Figure 5b (xanthosine), respectively. These signals were important role in the biological reactions belonging to pathways
identified in the UHPLC-QTOF data under the form [M+H]+ such as steroid hormone biosynthesis. This could be explained
Thus, for the positive mode of the FT-ICR data, about 14,3% by the fact that budesonide is a glucocorticoid itself.
of the total number of variables were selected as discriminant In negative mode (Figure 8b) the trend is reversed and only
signals based on their contributions in the covariance and in a few uptrend variables are highlighted. The number of
the correlation. Among these, close to 50% appear as uptrend annotated compounds remains nevertheless lower than in
with respect to CA. positive mode and the majority of the compounds are issued
The discriminant signals are loaded into MassTRIX and the from the model concerning the salbutamol treated athletes
metabolic pathways presenting the highest number of (~0.54% from the totality of examined variables).

PLOS ONE | www.plosone.org 6 September 2013 | Volume 8 | Issue 9 | e74584

 Non-Targeted Metabolomics for Doping Control

Figure 5. CA vs. V discriminant profiles: (a) creatine and (b) xanthosine.

doi: 10.1371/journal.pone.0074584.g005

Similarities and differences between salbutamol and Legitimacy of including volunteers in the experimental
budesonide treated athletes setup
SUS-plot is a graphical tool used to compare two statistical As seen previously, a clear difference between the
models (Figure 9). It indirectly finds the shared and the unique volunteers and the clean athletes can be made at metabolic
features of the groups that constitute the models. In brief, the level by using a global approach. In order to study the
SUS-plot is obtained by plotting the loading vector of a first legitimacy to associate the volunteers to clean athletes, two
model against the loading vector of the second model.
models were compared in positive electrospray: (BS vs CA)
Pathways like ABC transporters appear as modified by both
and (BS vs CA+V).
treatments to a similar degree. Steroid hormone biosynthesis
pathway is modified mainly by budesonide and only in a small As it can be seen in Figure 6a, the number of downtrend
degree by salbutamol. In turn, pentose and glucose pathway is compounds annotated in MassTRIX is basically not influenced
manly modified by salbutamol and to a smaller extent by by the introduction of volunteers in the group of clean athletes
budesonide. Surprisingly, fructose and mannose metabolisms (grey and black segments). However, for the uptrend
do not appear to be modified when the two treatments are metabolites, the metabolic differences between the athletes
considered separately. In contrast, only when the two types of and the volunteers seem to have a greater impact on the
samples are considered as a single group (BS vs CA, Figure 1) comparison between treated (BS, budesonide and salbutamol
they appear to be modified. Arachidonic acid and vitamin B6 together) and clean individuals (V+CA, athletes and volunteers
metabolisms are modified by budesonide only, while linoleic
together). This is, for example, the case of steroid hormone
acid metabolism or biosynthesis of unsaturated fatty acids are
biosynthesis pathway. In figure 6a, we can see that when
specific to salbutamol treatment (Figure 8a and 8b).
considering (BS vs CA) there are no uptrend annotations (black
segment). However, a small segment of uptrend metabolites
appears when associating the volunteers to the group of clean
athletes (grey segment).

PLOS ONE | www.plosone.org 7 September 2013 | Volume 8 | Issue 9 | e74584

 Non-Targeted Metabolomics for Doping Control

Figure 6. Metabolic pathways corresponding to discriminating signals (ESI+) annotated in MassTRIX (a). Metabolic
pathways corresponding to discriminating signals (ESI-) annotated in MassTRIX (b).
doi: 10.1371/journal.pone.0074584.g006

Canonical Correlation Analysis for the comparison of ionization suppression phenomena. By considering a tolerance
two analytical techniques of 5 mDa, we found that about 21% of the signals detected in
As stated previously, the two analytical platforms were used FT-ICR are common to both data sets Thus, approximately
independently to obtain the molecular fingerprints of each urine 15% of the discriminating signals whose profiles were visually
sample: UHPLC-QTOF/MS and FT-ICR/MS and the models checked were detected in both data sets and almost 57% came
resulted in similar performance indicators (Table 1). The latter from the FT-ICR data (see Table S1 for discriminating
has the advantage of having a very high resolution and 200 compounds as annotated by MassTRIX).
ppb precision, while the former one includes a possible Canonical Correlation Analysis preceded by Principal
separation step before the mass spectrometer which limits Components Analysis (PCA-CCA) was used in order to

PLOS ONE | www.plosone.org 8 September 2013 | Volume 8 | Issue 9 | e74584

 Non-Targeted Metabolomics for Doping Control

Figure 7. S vs. CA discriminant profiles: (a) pantothenic acid and (b) salbutamol.
doi: 10.1371/journal.pone.0074584.g007

compare the multivariate information contained in the two data principal components. Negatively or positively correlated m/z
sets. PCA was applied prior to the CCA in order to deal with variables can be emphasized by considering these principle
the multicolinearity present in this type of data, to reduce the components (loadings). One such example is presented in
dimensionality, and thus, ease the interpretation of the results. Figure 10. The two correlated signals (m/z=217.97 and m/
As an example, this method was used for the two data sets z=541.32) have down and up trend profiles, respectively. They
corresponding to the model opposing the volunteers and the prove to contribute greatly to the first principal components
clean athletes (V vs CA) in negative mode. For each of the two showing a clear discrimination between the two groups.
data sets the 5 first principal components (loadings) were Thus, the CCA shows that, globally, the information
selected for further analysis. In each case, the principal contained in the LC-TOF data set is the same with the one
components considered for CCA analysis accounted for about contained in the FT-ICR data set.
50% of the total variance.
The analysis yielded five canonical functions explaining Conclusions
96.6%, 71.5% 57.1%, 28.9% and 14.7% of the total variance
shared by the two data sets. The full model, as well as the The non-targeted analysis of forty-six human urine samples
association of functions (f2,f3,f4,f5) and (f3,f4,f5) are was done on two analytical platforms: UHPLC-QTOF/MS and
considered statistically significant since the associated p- FT-ICR/MS. Given the different origins of the samples
values are less than 0.05 (Table 2). On the contrary, the (V=Volunteers, CA=Clean Athletes, S=Salbutamol and
associations of function 4 and 5 and function 5 alone are not B=Budesonide) several groups were created. These groups
statistically significant and therefore not interpreted. were firstly studied separately and then compared in order to
As for the OPLS models, the inspection of the canonical emphasize possible similarities or differences between the
coefficients and structure coefficients reveal highly correlated samples. The statistical models created by using the OPLS

PLOS ONE | www.plosone.org 9 September 2013 | Volume 8 | Issue 9 | e74584

 Non-Targeted Metabolomics for Doping Control

Figure 8. Metabolic pathways corresponding to discriminating signals (ESI+) annotated in MassTRIX (a). Metabolic
pathways corresponding to discriminating signals (ESI-) annotated in MassTRIX (b).
doi: 10.1371/journal.pone.0074584.g008

method show good predictability for two models: CA vs V and variables. Significant advances towards annotation
S vs CA and lead to the selection of several discriminating identification and interpretation have been made thanks to the

PLOS ONE | www.plosone.org 10 September 2013 | Volume 8 | Issue 9 | e74584

 Non-Targeted Metabolomics for Doping Control

Figure 9. Up and downtrend features shared by the two models in the positive and negative mode.
doi: 10.1371/journal.pone.0074584.g009

Figure 10. Negatively correlated variables emphasized by Canonical Correlation Analysis.

doi: 10.1371/journal.pone.0074584.g010

high-resolution and high-precision measurements with the FT- Table 2. Significance parameters for the Canonical
ICR/MS and due to MassTRIX. More precisely, a significant Correlation Analysis.
number of the discriminated signals were annotated and placed
on metabolic pathways.
The Canonical Correlation Analysis showed a strong f1 f2 f3 f4 f5

correlation between the UHPLC-QTOF/MS data and the FT- squared canonical correlations
0.96 0.715 0.571 0.289 0.147
ICR data. coefficients
associated p-values 6.84e-07 0.008 0.043 0.144 0.109
The non-targeted metabolomics approach proved to be a
doi: 10.1371/journal.pone.0074584.t002
helpful strategy even when real urine samples coming from a
heterogeneous population were considered. In order to confirm

PLOS ONE | www.plosone.org 11 September 2013 | Volume 8 | Issue 9 | e74584

 Non-Targeted Metabolomics for Doping Control

the results presented in this paper and take into account the (DOC)
large inter-human variability, the same approach should be
repeated on a larger cohort of individuals. Further studies are Author Contributions
also needed in order to investigate the effect of other
stimulants and doping substances. Conceived and designed the experiments: AK CC PS.
Performed the experiments: AK AF. Analyzed the data: AK ML.
Supporting Information Contributed reagents/materials/analysis tools: CB. Wrote the
manuscript: AK.
Table S1. List of compounds as annotated by MassTrix.

References
1. Kanehisa M, Goto S (2000) KEGG: Kyoto Encyclopedia of Genes and 19. Zhang J, Yan L, Tian M, Huang Q, Peng S et al. (2012) The
Genomes. Nucleic Acids Res 28: 27-30. doi:10.1093/nar/28.7.e27. metabonomics of combined dietary exposure to phthalates and
PubMed: 10592173. polychlorinated biphenyls in mice. J Pharm Biomed Anal 66: 287-297.
2. Kanehisa M, Goto S, Sato Y, Furumichi M, Tanabe M (2012) KEGG for doi:10.1016/j.jpba.2012.03.045. PubMed: 22502909.
integration and interpretation of large-scale molecular data sets. 20. Kiss A, Jacquet AL, Paisse O, Flament-Waton MM, de Ceaurriz J et al.
Nucleic Acids Res 40: D109-D114. doi:10.1093/nar/gkr988. PubMed: (2011) Urinary signature of anabolic steroids and glucocorticoids in
22080510. humans by LC-MS. Talanta 83: 1769-1773. doi:10.1016/j.talanta.
3. Wishart DS, Knox C, Guo AC, Eisner R, Young N et al. (2009) HMDB: 2010.10.030. PubMed: 21238782.
a knowledgebase for the human metabolome. Nucleic Acids Res 21. Boyard-Kieken F, Dervilly-Pinel G, Garcia P, Paris AC, Popot MA et al.
37(Database issue): D603–10. PubMed: 18953024 (2011) Comparison of different liquid chromatography stationary
4. Wishart DS, Tzur D, Knox C, Eisner R, Guo AC et al. (2007) HMDB: phases in LC-HRMS metabolomics for the detection of recombinant
the Human Metabolome Database. Nucleic Acids Res 35 (Database growth hormone doping control. J Sep Sci 34: 3493-3501. doi:10.1002/
issue):D521-6. PubMed: 17202168 jssc.201100223. PubMed: 21710695.
5. Wishart DS, Tzur D, Knox C, Eisner R, Guo AC et al. (2007) HMDB: 22. Badoud F, Guillarme D, Boccard J, Grata E, Saugy M et al. (2011)
the Human Metabolome Database. Nucleic Acids Res 35: D521–D526. Analytical aspects in doping control: Challenges and perspectives.
doi:10.1093/nar/gkl923. PubMed: 17202168. Forensic Sci Int 213: 49-61. doi:10.1016/j.forsciint.2011.07.024.
6. Good DM, Zürbig P, Argilés A, Bauer HW, Behrens G, et al. (2010) PubMed: 21824736.
Naturally occurring human urinary peptides for use in diagnosis of 23. Anizan S, Bichon E, Duval T, Monteau F, Cesbron N, et al. (2012) Gas
chronic kidney disease. Mol Cell Proteomics 9: 2424-2437. doi: chromatography coupled to mass spectrometry-based metabolomic to
10.1074/mcp.M110.001917. PubMed: 20616184. screen for anabolic practices in cattle: identification of 5α-androst-2-
7. Suhre K, Schmitt-Kopplin P (2008) MassTRIX: mass translator into en-17-one as new biomarker of 4-androstenedione misuse J Mass
pathways. Nucleic Acids Res 36: W481-W484. doi:10.1093/nar/gkn194. Spectrom 47: 131–40.
PubMed: 18442993. 24. World Anti-Doping Agency website. Available: http://www.wada-
8. Wägele B, Witting M, Schmitt-Kopplin P, Suhre K (2012) MassTRIX ama.org. Accessed 2013 September 6.
Reloaded: Combined Analysis and Visualization of Transcriptome and 25. Bryan J (2007) Ventolin remains a breath of fresh air for asthma
Metabolome Data. PLOS ONE 7: e39860. doi:10.1371/journal.pone. sufferers, after 40 years. Pharm J 279: 404-405.
26. Ventura R, Segura J, Bergés R, Fitch KD, Morton AR et al. (2000)
0039860. PubMed: 22815716.
Distinction of inhaled and oral salbutamol by urine analysis using
9. Fonville JM, Richards SE, Barton RH, Boulange CL, Ebbels TMD
conventional screening procedures for doping control. Ther Drug Monit
(2010) The evolution of partial least squares models and related
22: 277-282. doi:10.1097/00007691-200006000-00008. PubMed:
chemometric approaches in metabonomics and metabolic phenotyping.
10850394.
J Chemom, 24: 636-649. doi:10.1002/cem.1359.
27. Deventer K, Mikulcíková P, Van Hoecke H, Van Eenoo P, Delbeke FT
10. Trygg J, Holmes E, Lundstedt T (2007) Chemometrics in
et al. (2006) Detection of budesonide in human urine after inhalation by
metabonomics. J Proteome Res 6: 469-479. doi:10.1021/pr060594q.
liquid chromatography-mass spectrometry. J Pharm Biomed Anal 42:
PubMed: 17269704.
474-479. doi:10.1016/j.jpba.2006.05.016. PubMed: 16842962.
11. Waterman DS, Bonner FW, Lindon JC (2009) Spectroscopic and
28. Gika HG, Macpherson E, Theodoridis GA, Wilson ID (2008) Evaluation
statistical methods in metabonomics. Bioanalysis 1: 1559-1578. doi: of the repeatability of ultra-performance liquid chromatography-TOF-
10.4155/bio.09.143. PubMed: 21083104. MS for global metabolic profiling of human urine samples. J
12. Jansson J, Willing B, Lucio M, Fekete A, Dicksved J et al. (2009) Chromatogr B Analyt Technol Biomed Life Sci 871: 299-305. doi:
Metabolomics reveals metabolic biomarkers of Crohn’s disease. PLOS 10.1016/j.jchromb.2008.05.048. PubMed: 18579458.
ONE 4(7): e6386. doi:10.1371/journal.pone.0006386. PubMed: 29. Gika HG, Theodoridis GA, Wilson ID (2008) Liquid chromatography
19636438. and ultraperformance liquid chromatography-mass spectrometry
13. Xu EY, Schaefer WH, Xu Q (2009) Metabolomics in pharmaceutical fingerprinting of human urine - Sample stability under different handling
research and development: Metabolites, mechanisms and pathways. and storage conditions for metabonomics studies. J Chromatogr A
Curr Opin Drug Discov Devel 12: 40-52. PubMed: 19152212. 1182: 314-322.
14. Wishart DS (2008) Applications of metabolomics in drug discovery and 30. Sangster T, Major H, Plumb R, Wilson AJ, Wilson ID et al. (2006) A
development. Drugs R D 9: 307-322. doi: pragmatic and readily implemented quality control strategy for HPLC-
10.2165/00126839-200809050-00002. PubMed: 18721000. MS and GC-MS-based metabonomic analysis. Analyst 131: 1075-1078.
15. McNiven EM, German JB, Slupsky CM (2011) Analytical metabolomics: doi:10.1039/b604498k. PubMed: 17003852.
nutritional opportunities for personalized health. J Nutr Biochem 22: 31. Lucio M, Fekete A, Schmitt-Kopplin P (2010) High resolution tools offer
995-1002. doi:10.1016/j.jnutbio.2011.05.016. PubMed: 21999844. to follow bacterial growth on a molecular level In: Handbook of
16. Manach C, Hubert J, Llorach R, Scalbert A (2009) The complex links Molecular Microbial Ecology I Metagenomics and Complementary
between dietary phytochemicals and human health deciphered by Approaches. Hoboken: Wiley-Blackwell.
Metabolomics. Mol Nutr Food Res 53: 1303–1315. doi:10.1002/mnfr. 32. Lucio M, Fekete A, Weigert C, Wägele B, Zhao X et al. (2010) Insulin
200800516. PubMed: 19764066. sensitivity is reflected by characteristic metabolic fingerprints--a Fourier
17. Roux A, Lison D, Junot C, Heilier JF (2011) Applications of liquid transform mass spectrometric non-targeted metabolomics approach.
chromatography coupled to mass spectrometry-based metabolomics in PLOS ONE 5(10): e13317. doi:10.1371/journal.pone.0013317.
clinical chemistry and toxicology: A review. Clin Biochem 44: 119-135. PubMed: 20976215.
doi:10.1016/j.clinbiochem.2011.08.1091. PubMed: 20800591. 33. Kohler M, Walpurgis K, Thomas A, de Maree M, Mester J et al. (2010)
18. Nordström A, Lewensohn R (2010) Metabolomics: Moving to the Clinic. Effects of endurance exercise on the urinary proteome analyzed by 2-D
J Neuroimmune Pharmacol 5: 4-17. doi:10.1007/s11481-009-9156-4. PAGE and Orbitrap MS. Proteomics Clin Appl 4: 568-576. PubMed:
PubMed: 19399626. 21137074.

PLOS ONE | www.plosone.org 12 September 2013 | Volume 8 | Issue 9 | e74584

 Non-Targeted Metabolomics for Doping Control

34. Yan B, J A, Wang GJ, Lu HL, Huang XP, et al. (2009) Metabolomic 38. del Barrio AS, García-Calonge MA, Fernández-Quintela A, Simón E,
investigation into variation of endogenous metabolites in professional Portillo MP et al. (1995) Effects of the Beta-Adrenergic Agonist
athletes subject to strength-endurance training. J Appl Physiol 106: Salbutamol and Its Withdrawal on Protein-Metabolism Of Lambs Ann
531-538. PubMed: 19036890. Nutr Metab 39: 317-324. doi:10.1159/000177879. PubMed: 8585702.
35. Burns SF, Oo HH, Tran AT (2012) Effect of Sprint Interval Exercise on 39. Bastos ML, Carvalho F, Remião F, Mendes ME, Ferreira MA et al.
Postexercise Metabolism and Blood Pressure in Adolescents. Int J (1997) Changes in taurine levels in response to repeated administration
Sport Nutr Exerc Metab 22: 47-54. PubMed: 22248500. of the beta 2-agonist salbutamol in lambs. J Vet Pharmacol Ther 20:
36. Broad EM, Maughan RJ, Galloway S DR (2011) Effects of Exercise 33-37. doi:10.1046/j.1365-2885.1997.00046.x. PubMed: 9049947.
Intensity and Altered Substrate Availability on Cardiovascular and 40. Martineau L, Horan MA, Rothwell NJ, Little RA (1992) Salbutamol, a
Metabolic Responses to Exercise After Oral Carnitine Supplementation beta 2-adrenoceptor agonist, increases skeletal muscle strength in
in Athletes. Int J Sport Nutr Exerc Metab 21: 385-397. PubMed: young men. Clin Sci 83: 615-621. PubMed: 1335400.
21813919. 41. van Baak MA, Mayer LH, Kempinski RE, Hartgens F (2000) Effect of
37. Paetsch PR, Greenshaw AJ (1991) Beta-Adrenergic Effects On Plasma salbutamol on muscle strength and endurance performance in
And Brain Large Neutral Amino-Acids Are Unaltered by Chronic nonasthmatic men. Med Sci Sports Exerc 32: 1300-1306. doi:
Administration of Antidepressants. J Neurochem 56: 2027-2032. 10.1097/00005768-200007000-00018. PubMed: 10912897.

PLOS ONE | www.plosone.org 13 September 2013 | Volume 8 | Issue 9 | e74584