Serum metabolomics reveals many novel metabolic markers of heart failure, including pseudouridine and 2-oxoglutarate

Abstract

There is intense interest in the identification of novel biomarkers which improve the diagnosis of heart failure. Serum samples
from 52 patients with systolic heart failure (EF<40% plus signs and symptoms of failure) and 57 controls were analyzed by
gas chromatography – time of flight – mass spectrometry and the raw data reduced to 272 statistically robust metabolite peaks.
38 peaks showed a significant difference between case and control (p<5×10−5). Two such metabolites were pseudouridine, a modified nucleotide present in t- and rRNA and a marker of cell turnover, as
well as the tricarboxylic acid cycle intermediate 2-oxoglutarate. Furthermore, 3 further new compounds were also excellent
discriminators between patients and controls: 2-hydroxy, 2-methylpropanoic acid, erythritol and 2,4,6-trihydroxypyrimidine.
Although renal disease may be associated with heart failure, and metabolites associated with renal disease and other markers
were also elevated (e.g. urea, creatinine and uric acid), there was no correlation within the patient group between these
metabolites and our heart failure biomarkers, indicating that these were indeed biomarkers of heart failure and not renal
disease per se. These findings demonstrate the power of data-driven metabolomics approaches to identify such markers of disease.

Full text

Page 1

Serum metabolomics reveals many novel metabolic markers of heart
failure, including pseudouridine and 2-oxoglutarate
Warwick B. Dunn,a,bDavid I. Broadhurst,bSasalu M. Deepak,c,dMamta H. Buch,cGarry McDowell,e
Irena Spasic,a,bDavid I. Ellis,bNicholas Brooks,dDouglas B. K ell,a,b,* and Ludwig Neysesc,**
aThe Manchester Centre for Integrative Systems Biology, Manchester Interdisciplinary Biocentre, The University of Manchester, 131 Princess St.,
Manchester, M1 7DN, UK
bSchool of Chemistry, Manchester Interdisciplinary Biocentre, The University of Manchester, 131 Princess St.,
Manchester, M1 7DN, UK
cDivision of Cardiovascular and Endocrine Sciences, The University of Manchester, Room 1.302, Stopford Building, Oxford Road,
Manchester, M13 9PL, UK
dDepartment of Cardiology, South Manchester University Hospital, Southmoor Road, Wythenshawe, Manchester, M23 9LT, UK
eDepartment of Clinical Biochemistry, Manchester Royal Infirmary, Oxford Road, Manchester, M13 9WL, UK
Received 7 December 2006; Accepted 23 April 2007
Thereis intense interest in the identification of novel biomarkers which improve the diagnosis of heart failure. Serum samples
from 52 patients with systolic heart failure(EF < 40% plus signs and symptoms of failure) and 57 controls wereanalyzed by gas
chromatography – time of flight – mass spectrometry and the raw data reduced to 272 statistically robust metabolite peaks. 38
peaksshowed a significant differencebetween caseand control (p< 5·10)5). Two such metaboliteswerepseudouridine, a modified
nucleotidepresent in t- and rRNA and a marker of cell turnover, aswell asthetricarboxylic acid cycleintermediate2-oxoglutarate.
Furthermore, 3 further new compounds were also excellent discriminators between patients and controls: 2-hydroxy, 2-
methylpropanoic acid, erythritol and 2,4,6-trihydroxypyrimidine. Although renal diseasemay beassociated with heart failure, and
metabolites associated with renal disease and other markers were also elevated (e.g. urea, creatinine and uric acid), there was no
correlation within thepatient group between thesemetabolites and our heart failurebiomarkers, indicating that thesewereindeed
biomarkers of heart failure and not renal disease per se. These findings demonstrate the power of data-driven metabolomics
approaches to identify such markers of disease.
KEY WORDS: heart failure; metabolomics; biomarkers; pseudouridine; 2-oxoglutarate.
1. Introduction
Biomarkers are essential tools in diagnosing disease
and monitoring progression as well as response to
therapy. In heart failure, BNP has proved its usefulness
as a diagnostic marker and there are some limited data
on monitoring response to therapy (Troughton et al.,
2000; Zaphiriou et al., 2005). Other markers such as
TNF alpha, IL-1beta and IL-6 havealso been shown to
be elevated and are likely to reflect peripheral processes
distinct from cardiac overload and/or hypertrophy
which is predicated by BNP secretion (Pan et al., 2004).
Thereisintenseinterest in theidentification of further
biomarkers in the syndrome of heart failure. Distinct
patterns of the ensemble of several such markers may
eventually help in identifying specific classes of the
syndrome with improved predictive power in terms of
diagnosis, prognosis and treatment options.
In a first step towardsthisgoal, weherepresent a new
approach to identifying biomarkers by focusing on
metabolic changes during heart failure. Metabolomics is
the data-driven study of the different patterns of
metabolites within living organisms, in which we seek
measurements that are as comprehensive as possible.
Conceptually, metabolomics lends itself ideally to
studying heart failure and other diseases. In contrast
say to malignancy, where tissue is readily available for
transcriptomic and proteomic analysis, routine labora-
tory tests for heart failure are limited to serum samples.
Furthermore, metabolic alterations have been well doc-
umented in the heart (Neubauer et al., 1997; Finck and
K elly, 2006) and peripheral tissues (Drexler et al., 1992;
DeSousa et al., 2000) in heart failureand other diseases
(Ores ˇ ic ˇet al., 2006; van der Greef et al., 2006). Meta-
bolomics is also expected on theoretical grounds to be
morediscriminatory thanproteomics(Oliveret al., 1998;
Raamsdonk et al., 2001; K ell, 2004; K ell, 2006a, b).
The present metabolomic study of serum from
patients with heart failure and appropriate controls
detected 272 candidate metabolite peaks, of which 38
showed highly significant differences between cases and
controls. At least two of these metabolites, pseudouri-
dine and 2-oxoglutarate, have the potential, alone or
together, to improveon or add to BNP in terms of their
sensitivity and specificity as biomarkers.
To whom correspondence should be addressed.
*E-mail: dbk@manchester.ac.uk
**E-mail: ludwig.neyses@CMMC.NHS.uk
Metabolomics, Vol. 3, No. 4, December 2007 (? 2007)
DOI: 10.1007/s11306-007-0063-5 413
1573-3882/07/1200-0413/0 ? 2007 Springer ScienceþBusiness Media, LLC

Page 2

2. Methods
2.1. Participants
Thestudy wasapproved by thelocal ethicscommittee
for clinical studies. Patients attending the heart failure
clinic at the South Manchester University Hospital
Regional Cardiothoracic Centre and Heart Transplant
Unit were invited to participate in the study by pro-
viding samples of venous blood. 52 patients were
recruited with an established diagnosisof left ventricular
dysfunction (LVD, EF < 40%) and heart failure.
Ejection fraction wasdetermined echocardiographically.
All patients belonged to NY HA (New Y ork Heart
Association) classes II-IV. The age-matched control
group were mainly taken from an ENT clinic, though
some were from an ophthalmology clinic and others
were the healthy partners of the patients. They com-
prised subjects with no history of cardiac disease, and
were examined clinically to this end at recruitment.
Subjects with acuteor chronic inflammatory conditions,
malignancies and significant respiratory pathology were
excluded. See Table 1 for the patient characteristics.
2.2. Sample collection
Venous blood was collected from the antecubital
veins by venepuncture, left to coagulate for 10 min,
followed by centrifugation (2200 g,
supernatant serum was then aliquoted into 2.5 mL
cryotubes and stored at ) 80?C until required.
15 min).The
2.3. Sample preparation and GC-TOF-MS analysis
Serum samples were prepared for Gas Chromatog-
raphy-Time of Flight-Mass Spectrometry (GC-tof-MS)
analysis as described (K enny et al., 2005; O?Hagan
et al., 2005; Underwood et al., 2006). 200 lL serum was
spikedwith100 lL internal
(0.18 mg.mL)1succinic d4 acid; Sigma-Aldrich, Gill-
ingham, UK ) and vortex mixed for 15 s. Protein pre-
cipitation was performed by addition of 600 lL
acetonitrile and vortex mixing for 15 s followed by
centrifugation (13,385 g, 15 min). The supernatant was
transferred to an Eppendorf tube and lyophilised
(HETO VR MAX I vacuum centrifuge attached to a
HETO CT/DW 60E cooling trap; Thermo LifeSciences,
Basingstoke, UK ). To increase metabolite volatility and
thermal stability a two-stage chemical derivatisation
procedure was performed. 50 lL 20 mg.mL)1O-meth-
ylhydroxylaminesolution wasadded and heated at 40?C
for 90 min followed by addition of 50 lL MSTFA (N-
acetyl-N-(trimethylsilyl)-trifluoroacetamide) and heating
at 40?C for 90 min. 20 lL of a retention index solution
(3 mg.mL)1n-decane, n-dodecane, n-pentadecane, n-
nonadecane, n-docosanedissolved in hexane) wasadded
and the samples wereanalysed using an Agilent 6890 N
gas chromatograph and 7683 autosampler (Agilent
Technologies, Stockport, UK ) coupled to a LECO
Pegasus III electron impact time-of-flight mass spec-
trometer (LECOCorporation,
employing previously described optimised instrumental
conditions for serum (O?Hagan et al., 2005). Initial data
processing of raw data was undertaken using LECO
ChromaTof v2.25 software to construct a data matrix
(metabolite peak vs. sample no.) including response
ratios (peak area metabolite/peak area succinic-d4
internal standard) for each metabolite peak in each
sample. The number of peaks in this matrix is 1467,
based on a series of studies of serum samples additional
to but including those in the present sample set (and
note that some metabolites can produce more than one
derivative). This matrix was then further reduced by
removing any peak which had more than 60% missing
values. Inthisway if any classwasnot being consistently
matched with a given peak, that peak was considered
not robust enough for further statistical analysis.
standard solution
StJoseph, USA)
2.4. Pro-BNP
N-terminal pro-BNP was measured on the Roche
E170 immunoassay analyser using the manufacturer?s
chemiluminescence kit according to their instructions.
2.5. Urea and creatinine
Urea and creatinine were measured enzymatically in
serum using a Roche analyser according to the manu-
facturer?s instructions.
Table 1
Summary of patient characteristics. IQR = inter-quartile range
Cases HF
(n = 52)
Controls
(n = 57)
Age,± 1SD, range
Sex M:F
Median NY HA Class (IQR)
Median Ejection Fraction (IQR)
Mean Ejection Fraction (IQR)
Etiology ischemic:non-ischemic
Bean BMI ± 1SD
Hypertensive: normotensive
DM:nonDM
Smoker:non-smoker
Na+ , mean,± 1SD, Range
K + , mean,± 1SD, range
Urea, mean,± 1SD, range
Creatinine, mean,± 1SD, range
Hemoglobin, mean,± 1SD, range 142, 13, 110–182 144, 14, 111–152
Beta-blockers Y :N
ACE inhibitors Y :N
Diuretics Y :N
68, 10, 46–86
43:9
2 (0)
25(5)
27(6)
33:19
27(5)
8:44
9:43
5:47
141, 3, 133–145
4.2, 0.7, 3.3–5.2 4.2, 0.3, 3.7–5.2
12, 8, 3 – 56
139, 57, 70–352
67, 9, 44–87
22:35
0
26(5)
15:42
4:53
8:49
140, 3, 132–146
6, 2, 3–12
81, 18, 51–142
44:4
43:8
44:7
7:52
6:50
7:49
DM = Diabetes mellitus. The p-value for the difference in creatinine
between cases and controls was 2.5·10)12. Spearman?s rank correla-
tion coeffi cient between creatinine and BNP was 0.66 with a signifi-
cance p-value of 8·10)13
W. B. Dunn et al./Heart failure metabolomics
414

Page 3

2.6. Statistical analyses
Univariatestatistical analysis was performed in order
to assess the characteristics of each independent
metabolite peak generated using the aboveprotocol. As
theexperimental design employed in this study was that
of a matched case–control study (Rothman and
Greenland, 1998), single-factor analysis of class distri-
bution was performed for each metabolite peak. Before
choosing which significance test to apply, each metab-
olitepeak was checked for within-class kurtosis, and for
within-class goodness of fit to a normal distribution
using the Lilliefors test. For a given metabolite peak, if
either control or case samples had kurtosis > 3, or
failed the Lilliefors test (where the null hypothesis, that
thesampleset hasa normal distribution, isrejected if the
p-value< 0.05) then the non parametric rank-based
Wilcoxon rank sum (Mann–Whitney) test was used
(Hollander and Wolfe, 1973); otherwise the classical
one-way ANOVA test was used. For each metabolite,
the null hypothesis (that the sample metabolite con-
centrations from both classes came from populations
with thesamemean) wastested. Generally if a p-valueof
less than 0.01 is calculated then the two sample popu-
lationsfor that metabolitearedeemed to besignificantly
different; however, when many parallel tests are per-
formed carehas to betaken regarding TypeI errors (i.e.
falsely rejecting the null hypothesis). To help reduce the
possibility of Type I errors the critical p-value was
modified using Bonferroni correction. Bonferroni cor-
rectionisconsidered to beoverzealousinitsreduction of
TypeI errors (Bland and Altman, 1995; Perneger, 1998;
Broadhurst and K ell, 2006), leading to the conclusion
that any peak found to havea p-valuebelow this level is
clearly significant. Nonetheless, Type I errors are an
ever-present danger in this type of study (Wacholder
et al., 2004). Thus, with 272 statistically usable peaks
detected (note that some metabolites have more than
one peak, though not all metabolites are identified), the
modified critical (p-) value was set conservatively to
0.01/200, or, 5·10)5.
All statistical analyses were carried out using the
Matlab? scripting language (http://www.mathworks.
com/). All algorithms used are implemented such that
any missing values are ignored. Scripts are available
upon request.
3. Results
Of the 272 metabolite peaks produced by the pre-
processing protocol, 38 showed a significant difference
between case and control (p-value< 5.10)5) (figure 1,
Table 2), indicating changes in several metabolic path-
ways. Wenotethat therewas somegender bias between
the cases and controls (Table 1), but that this is not
responsible for the discriminating biomarkers observed
(Table 2) (cf. (K irschenlohr et al., 2006) and see also
(Ransohoff, 2005; Broadhurst and K ell, 2006) for
reviews). There was also inevitable bias in the use of
pharmaceutical drugs between cases and controls
(Table 1) but this did not affect the distribution of the
biomarkers detected (seelater, and all relevant data and
metadata will be made available electronically as sup-
plementary information). Analysis of these peaks indi-
cated that themetabolitesthat they represent camefrom
many different parts of metabolism including energy
metabolismand nitrogen metabolism. Whileit is known
that metabolites such as urea, uric acid and creatinine
are often raised in heart failure, reflecting impaired
renal excretion (Smith et al., 2006), a number of ?novel?
metabolites which have not previously been implicated
in heart failurewerealso observed to beto bepresent at
considerably different levels in cases and controls. Five
such metabolites were identified chemically and consid-
ered to be of particular biological interest: pseudouri-
dine, 2-oxoglutaric acid, 2-hydroxy, 2-methylpropanoic
acid, erythritol, and 2,4,6-trihydroxypyrimidine. Fig-
ure 1a demonstrates the ROC curves for these metabo-
lites, as well as for BNP which is presently considered a
?gold standard? heart failurebiomarker (Zaphiriouet al.,
2005). It isevident that pseudouridineand 2-oxoglutaric
acid are at least as good indicators of heart failure as is
BNP, while 2-hydroxy,2-methylpropanoic acid, erythri-
tol and 2,4,6-trihydroxypyrimidine are almost as good.
The area under the ROC curve and the statistical
p-value are not entirely unrelated, but they do never-
theless measure different properties; ROC curves are
moresensitiveto theactual classdistributions across the
total range, rather than being encoded as an estimateof
this distribution as calculated by variance-based signif-
icance tests. The relationship between these two values
for a series of metabolites is illustrated in figure 2b.
Figure 1 highlights the ability of the technologies
employed to identify metabolites from the mass spectra
produced, in this example, that of pseudouridine. There
are also two ?unknowns? (peaks with mass spectra very
different from any of those in currently available mass
spectral libraries) which have areas under the ROC
curve greater than 0.85 (one being the third highest),
which may betaken to indicatethat metabolomicsisstill
a youthful science (Harrigan and Goodacre, 2003; van
der Greef et al., 2004).
Figure 3a demonstrates the corresponding ?box and
whisker? plots for thetop 5 novel metabolites plus BNP.
A model combining pseudouridine and 2-oxoglutaric
acid discriminates visually between case and control
with only 3 falsenegatives (figure 3b). Furthermore, the
patients misclassified by this model differ from those
misclassified using BNP.
Bias or the influence of uncontrolled variables can
present a major headache in many omics studies,
including those seeking biomarkers (Ransohoff, 2005;
Broadhurst and K ell, 2006; K irschenlohr et al., 2006;).
In a case–control study such asthepresent one, it occurs
W. B. Dunn et al./Heart failure metabolomics
415

Page 4

in particular when variables other than thoseof primary
interest are unevenly distributed between the cases and
thecontrols. It is then possible, and certainly possibleto
argue, that the changed distribution of any proposed
biomarkers is not due to the presence of the disease but
to the differential distribution of these ?confounding?
variables. One immediate potential source of bias arises
from the straightforward fact that ?cases? are (presum-
ably considerably) more likely to be taking pharma-
ceutical medication than are controls. In the present
work, however, it transpired that whilethedistributions
of thosetaking particular classes of medication between
casesand controlsisnot at all even, simpleinspection of
the data shows that quite a considerable fraction of the
Figure 1. GC-tof-MS analysis of pseudouridine. (a) Total Ion Chromatogram as a function of time obtained by GC-tof-MS for a diseased
subject. The inset shows an enlarged single ion chromatogram for pseudouridine (m/z 357) from a subject with heart failure (full line) and a
control subject (dotted line) showing thedifferencein peak area and thereforeconcentration between thetwo classes. (b) Mass spectra of a peak
identified as pseudouridine obtained from (i) clinical serum samples and (ii) an authentic standard. Note that several metabolites appear more
than once because one metabolite can form multiple derivatives during sample preparation. For (b) the ordinate is in count normalised to the
highest peak while the abscissa is the mass-to-charge ratio.
W. B. Dunn et al./Heart failure metabolomics
416

Page 5

Table 2
Discriminatory metabolite peaks
Metabolite
Median value C
(Female, Male)
Median abs. dev. C
(Female, Male)
Median value HF
(Female, Male)
median abs. dev. HF
(Female, Male)
p-value
M–W test
ROC Convex
hull area
Pseudouridine *
(0.048, 0.048)
(0.0077, 0.0047)
(0.094, 0.11)
(0.015, 0.037)
1.78E-15
0.96
N-BNP
(8.5, 7.5)
(2.5,4.5)
(291, 305)
(125,241)
9.99E-16
0.93
2-Oxoglutaric Acid *
(0.028, 0.03)
(0.0034, 0.0044)
(0.045, 0.054)
(0.0065, 0.0091)
1.61E-13
0.93
Unknown
(0.058, 0.059)
(0.02, 0.016)
(0.16, 0.12)
(0.11, 0.08)
4.59E-05
0.89
2-hydroxy,2-methylpropanoic acid *
(0.039, 0.036)
(0.0063,0.0059)
(0.19, 0.11)
(0.13, 0.042)
7.09E-10
0.88
Erythritol *
(0.0099, 0.011)
(0.0017, 0.0027)
(0.015, 0.022)
(0.0018, 0.0071)
1.37E-10
0.88
Uric acid*
(0.015, 0.014)
(0.0039, 0.0039)
(0.031, 0.026)
(0.0079, 0.01)
9.06E-11
0.87
Unknown
(0.016, 0.016)
(0.0047, 0.0034)
(0.036, 0.034)
(0.016, 0.013)
7.75E-10
0.87
Erythritol *
(0.074, 0.076)
(0.014, 0.018)
(0.11, 0.16)
(0.017, 0.048)
4.37E-10
0.87
Unknown
(0.0039, 0.0055)
(0.0014, 0.0017)
(0.0054, 0.0096)
(0.0023, 0.004)
1.16E-08
0.87
2,4,6-Trihydroxypyrimidine
(0.013, 0.015)
(0.0018, 0.0028)
(0.023, 0.026)
(0.0037, 0.0064)
2.13E-10
0.87
Threonic acid
(0.034, 0.033)
(0.0058, 0.0067)
(0.056, 0.092)
(0.0049, 0.031)
5.68E-10
0.85
Uric acid *
(0.41, 0.56)
(0.11, 0.15)
(0.83, 0.87)
(0.089, 0.17)
8.83E-10
0.85
Myo-inositol *
(0.55, 0.57)
(0.062, 0.097)
(0.7, 0.86)
(0.14, 0.25)
1.42E-08
0.84
Para-hydroxyphenylacetic acid
(0.0028, 0.0033)
(0.00079, 0.0011)
(0.0064, 0.0056)
(0.0022, 0.0023)
5.07E-08
0.83
Unknown
(0.0029, 0.0027)
(0.00072, 0.0012)
(0.0047, 0.0055)
(0.0015, 0.0023)
4.92E-07
0.83
X ylitol or ribitol
(0.024, 0.022)
(0.0054, 0.0074)
(0.038, 0.046)
(0.018, 0.016)
1.39E-08
0.83
Unknown
(0.021, 0.015)
(0.0065, 0.004)
(0.012, 0.009)
(0.0067, 0.0042)
2.41E-05
0.82
Uric acid*
(1.5,2)
(0.56, 0.66)
(3.2,3.4)
(0.43, 0.6)
1.82E-07
0.82
X ylose or isomer
(0.024, 0.03)
(0.0054, 0.011)
(0.038, 0.048)
(0.018, 0.017)
6.87E-08
0.81
Disaccharide
(0.0056, 0.0082)
(0.0021, 0.0025)
(0.0086, 0.013)
(0.0022, 0.005)
3.68E-09
0.80
Para-cresol
(0.088, 0.13)
(0.04, 0.041)
(0.15, 0.26)
(0.026, 0.12)
1.09E-06
0.80
Disaccharide
(0.0025, 0.0027)
(0.00041, 0.00029)
(0.0032, 0.0051)
(0.00065, 0.0015)
5.09E-06
0.80
Unknown
(0.0061, 0.0055)
(0.0013, 0.0024)
(0.012, 0.0092)
(0.0034, 0.0028)
6.69E-06
0.79
Unknown
(0.067, 0.072)
(0.017, 0.02)
(0.08, 0.13)
(0.022, 0.044)
1.60E-05
0.79
Indole-3-acetic acid *
(0.043, 0.054)
(0.0098, 0.022)
(0.08, 0.091)
(0.033, 0.026)
5.25E-06
0.79
Glyceric acid
(0.091, 0.094)
(0.012, 0.014)
(0.12, 0.14)
(0.02, 0.043)
1.81E-05
0.78
2,4-Dihydroxybutanoic Acid
(0.014, 0.018)
(0.003, 0.0031)
(0.029, 0.034)
(0.0035, 0.013)
3.00E-09
0.78
Unknown
(0.038, 0.064)
(0.011, 0.03)
(0.099, 0.1)
(0.022, 0.03)
2.37E-05
0.77
Sucrose *
(0.019, 0.019)
(0.011, 0.011)
(0.035, 0.062)
(0.012, 0.043)
1.78E-05
0.77
Unknown
(0.025, 0.029)
(0.004, 0.0037)
(0.041, 0.037)
(0.0037, 0.0047)
2.26E-06
0.77
Tetronic acid
(0.035, 0.027)
(0.0095, 0.0068)
(0.043, 0.048)
(0.011, 0.018)
4.59E-05
0.76
Dehydroascorbic acid
(0.22, 0.21)
(0.08, 0.07)
(0.15, 0.1)
(0.078, 0.061)
3.59E-05
0.76
Peaks arelisted when thereis significant discrimination between case(HF) and control, i.e. thearea under theROC (receiver operator characteristic) curve‡0.75 and theWilcoxon rank sum (Mann–
Whitney = MW) test p-values< 5·10)5. Median absolute deviation (abs.dev.) is used in place of the usual standard deviation as it is less sensitive to outliers and non-normal distributions.
Identification of themetaboliteswasperformed initially by massspectral library searchesemploying NIST/EPA/NIH (02) and an author?s(WBD) prepared massspectral libraries(requiring a similarity
and reverse match score > 700 for a peak to be identified). Further definitive identification (indicated by *) was performed by analysis of authentic standards with identical analytical conditions and
identification was confirmed if theretention index (± 10) and mass spectra (similarity and reversematches> 750) of metabolitepeak in sampleand standard wereequivalent (as illustrated in figure 1
for pseudouridine). C = Control; HF = Heart failure
W. B. Dunn et al./Heart failure metabolomics
417

Page 6

controls were in fact also receiving the medications that
onemight expect in heart failurepatients (Table 1), and
we give the data for beta-blockers, ACE inhibitors,
diuretics and statins in figure 4 (and for these and other
metadata in tabulated form in the Supplementary
information). It is equally clear from the data shown in
figure 4 that the two main markers are still much larger
in the disease than in the control patients, although for
instance some of the patients taking ACE inhibitors do
have values of Psi lower than those who are not (fig-
ure 4c), possibly providing evidence of the effi cacy of
these drugs. By contrast, diuretics have no (or rather
even a negative) effect in terms of lowering Psi (fig-
ure 4e) or 2OG (figure 4f) in patients, either indicating
that excretion has less of a controlling influence on the
steady-state values of these metabolites or (perhaps less
plausibly) that thedrugshappen to havebeen given only
to patients with lower values of these metabolites in the
first place.
Another kind of potential for bias or misinterpre-
tation can come where A causes B and A causes C and
it might then be interpreted that C causes B as they
p-value (Mann-Whitney test)
Area under ROC convex hull
'105'
'2030'
'122'
'130'
'196'
'126'
'145'
'146'
'136'
'Unknown'
'152'
'Unknown'
'Unknown'
'Unknown'
'Unknown'
'Urea'
'178'
'182'
'330'
'190'
'Pyruvic acid *'
'Pyruvic acid *'
'Malonic acid *'
'222'
'224'
'226'
'Para-cresol'
'240'
'249'
'Malonic acid *'
'Glyceric acid'
'Unknown'
'289'
'361'
'Unknown'
'322'
'Erythritol *'
'Erythritol *'
'Unknown'
'351'
'Tetronic acid'
'384'
'388'
'Xylitol or ribitol '
'Xylose or isomer'
'Monosaccharide'
'491'
'2-Oxoglutaric Acid *'
'628'
'Unknown'
'Indole-3-acetic acid *'
'Uric acid*'
'Unknown'
'Unknown'
'Pseudouridine *'
'941'
'Unknown'
'2,4,6-Trihydroxypyrimidine'
'Unknown'
'Uric acid *'
'Disaccharide'
'Unknown'
'2022'
'Unknown'
'Uric acid*'
'Threonic acid'
'2,4-Dihydroxybutanoic Acid'
'3077'
'Unknown'
'-1'
'N-BNP'
1E-15 1E-13 1E-11 1E-9 1E-7 1E-5 0.001 0.1
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
(a)
(b)
Figure 2. Receiver-operator characteristic (ROC) values for various metabolite peaks (a) ROC curves for BNP and five of the metabolites
determined in the present study, (b) relationship between area under ROC curve and p-values for the various metabolites. * indicates that the
metabolite has been chemically identified by analysing authentic chemical standards and determining that retention indices (± 10) and mass
spectra (match> 750) of metabolites and authentic standardsaresimilar. In (b), thecolour of thesymbol indicates whether themetabolitelevels
are greater (blue) or smaller (red) in the disease than the control.
W. B. Dunn et al./Heart failure metabolomics
418

Page 7

Figure 3. Metabolomics biomarkers for heart failure. (a) Box-whisker plots of the distribution of the most discriminating chemically identified
markers in heart failure described in this study. HF: Heart failure; C: Controls. In the ordinates, N-BNP is in units of pmol/L while the other
metabolites are unitless and described as a response ratio (peak area-metabolite/peak area-internal standard). The lower and upper lines of the
‘‘box’’ arethe25th and 75th percentiles of thesample. Thedistancebetween thetop and bottomof thebox is theinterquartilerange. Thelinein
the middle of the box is the sample median. If the median is not centred in the box, that is an indication of skewness. The ‘‘whiskers’’ are lines
extending aboveand below thebox. They show theextent of therest of thesample(unlessthereareoutliers). Assuming no outliers, themaximum
of thesampleisthetop of theupper whisker. Theminimumof thesampleisthebottomof thelower whisker. By default, an outlier isa valuethat
is morethan 1.5 times theinterquartilerangeaway from thetop or bottom of thebox. Thenotches in thebox area graphic confidenceinterval
about the median of a sample. A side-by-side comparison of two notched box plots provides a graphical way to determine which groups have
significantly different medians. (b) Relationship between 2-oxoglutarate and pseudouridine in controls (open circles) and heart failure patients
(closed circles). The BNP concentration from the same individuals is encoded via the size of the symbols (small = lower, large = higher), and
serves to illustrate that the three markers visually misclassify different samples and thus give different information.
W. B. Dunn et al./Heart failure metabolomics
419

Page 8

will tend to be correlated (Pearl, 2000; Mackay, 2003).
As mentioned above, it is, for instance, known that
patients with heart failure may tend to have problems
with renal function (Tsutamoto et al., 2006), and that
the impaired renal function might then be ?responsible?
for an inference that biomarker X is a biomarker for
heart disease when it is ?really? a biomarker for
impaired renal function (and known serum biomarkers
for renal dysfunction include creatinine and urea). The
concentration of any metabolite is clearly a function of
its synthesis, degradation, sequestering and excretion,
such that changes in its excretion might automatically
be expected to raise its concentration. Of course almost
any small molecule might then be raised due to renal
dysfunction, but the specificity that we see for our
biomarkers relative to other molecules makes this kind
of explanation a priori unlikely. There are several other
reasons why we do not consider that our new heart
failure biomarkers such as Psi and 2OG fall into this
category of renal dysfunction co-variates. First, these
Psi_BBLOCKERS
NY
0
0.1
0.2
0.3
0.4
0.5
E
N
I
D
I
R
U
O
D
U
E
S
P
BETA-BLOCKERS
2OG_BBLOCKERS
NY
0
0.05
0.1
0.15
BETA-BLOCKERS
E
T
A
R
A
T
U
L
G
O
X
O
- 2
Psi_ACE_INHIBS
NY
0
0.1
0.2
0.3
0.4
0.5
ACE INHIBITORS
E
N
I
D
I
R
U
O
D
U
E
S
P
2OG_ACE_INHIBS
NY
0
0.05
0.1
0.15
E
T
A
R
A
T
U
L
G
O
X
O
- 2
ACE INHIBITORS
Psi_DIURETICS
NY
0
0.1
0.2
0.3
0.4
0.5
E
N
I
D
I
R
U
O
D
U
E
S
P
DIURETICS
2OG_DIURETICS
NY
0
0.05
0.1
0.15
E
T
A
R
A
T
U
L
G
O
X
O
- 2
DIURETICS
(a)(b)
(c)(d)
(e) (f)
Figure 4. Lack of effect of pharmaceutical medications in increasing the values of pseudouridine (a, c, e, g) or 2-oxoglutarate (b, d, f, h) in
patientsrelativeto controls. Themedicationsinvolved werebeta-blockers(a, b), ACE inhibitors(c, d), Diuretics(e, f) and statins(g, h), although
becauseof thenumbersof samplesinvolved thesewerenot stratified further into individual drugs. Closed circles patients, open squares controls.
Drug status indicated by N/Y . Data points have been ?jittered?.
W. B. Dunn et al./Heart failure metabolomics
420

Page 9

patients do primarily have, and indeed presented with,
heart failure. A combined aetiology that goes renal
failure causes our biomarkers causes heart failure is not
realistic. Secondly, as before (K enny et al., 2005), to
distinguish this type of phenomenon it is appropriate
to look within as well as between classes. Thus
although they are reasonably well correlated (see also
(Niwa et al., 1998)), the relationship between Psi and
urea and between Psi and creatinine (both measured
enzymatically in the same serum samples) is little dif-
ferent between cases and controls (figure 5a, b), and it
is not suggested that the controls are suffering from
renal failure. Similarly, within a class, there is no
relationship at all between 2OG and these same renal
function markers (figure 5c, d). It is then even harder
to argue somehow that both Psi and 2OG should
therefore be renal dysfunction markers. This contrasts
markedly with the close relationship between the two
known renal dysfunction biomarkers urea and creati-
nine both within and between classes (figure 5e).
Finally, we mention again the relative lack of impact of
diuretics (figure 4e, f) on our biomarkers relative to
their apparent effect on creatinine (figure 5f).
Uric acid has been suggested as a marker of car-
diovascular mortality and its level is known to increase
in patients with heart failure (Leyva et al., 1998a, b;
Anker et al., 2003 Sakai et al., 2006), albeit not, it is
claimed, when adjusted for diuretic use (Culleton et al.,
1999b). It is probably a measure of impaired oxidative
capacity (Leyva et al., 1997, b) and it is of interest that
it too was observed as being among the most signifi-
cant biomarkers in our metabolomics data (figure 2,
Table 2). In this case, the GC method showed three
separate derivatives to be discriminatory (figure 2) and
the data for the two more significant ones are given in
figure 5g. Again, there is little relationship between uric
acid and creatinine within a class (i.e. controls or cases,
figure 5h) (and the uric acid levels in this case seem
little influenced by diuretic usage; data not shown).
The same is true for BNP against creatinine and urea
(figure 5i, j).
4. Discussion
Theprincipal findingsof thisstudy arerepresented by
the identification of a fingerprint of serum metabolites
that characterizeheart failure, and stressthepotential of
pseudouridineand2-oxoglutaric acid (a-ketoglutaratein
the traditional nomenclature)
markers of heart failure. Furthermore, this study high-
lights the potential of metabolomic analysis and, more
generally, of the systems biology approach (K ell, 2004,
2006a), in the evaluation of a disease condition.
An advantage of the metabolomic approach in the
identification of novel biomarkersisthat it isnot limited
by our current knowledge, and instead seeks to measure
the maximum number of metabolites in a given sample;
thetotal number in thenativehuman metabolic network
is unknown (although leaving asidecombinatorial lipids
a number around 3000 is a reasonable starting estimate
(K ell, 2006b; Duarte et al., 2007)) and a more restricted
number, a ?metabolic profile?, is usually obtained. The
most common analytical technologies employed in
metabolic profiling applications (Dunn and Ellis, 2005)
involve chromatographic separation (gas chromatogra-
phy, liquid chromatography and capillary electropho-
resis) followed by sensitivemass spectrometric detection
(so-called hyphenated techniques). Gaschromatography
– time of flight-mass spectrometry (GC-tof-MS), which
was used in this study, probably provides the highest
one-dimensionalchromatographic
detection from nM to mM concentrations, and with
metabolite identification being provided via electron-
impact ionisation mass spectrometry. Two-dimensional
GCxGC can reveal yet more metabolites (O?Hagan
et al., 2007).
Pseudouridine is a modified nucleoside that is found
in ribosomal and transfer RNA and is produced post-
transcriptionally (Charette and Gray, 2000; Ofengand,
2002) such that it is then ?fixed? inside these macromol-
ecules. It is not further metabolised other than by RNA
hydrolysis, and its appearance in blood and urine is
therefore considered to be an excellent measure for
as novel diagnostic
resolution,with
Psi_STATINS
NY
0
0.1
0.2
0.3
0.4
0.5
E
N
I
D
I
R
U
O
D
U
E
S
P
STATINS
2OG_STATINS
NY
0
0.05
0.1
0.15
STATINS
E
T
A
R
A
T
U
L
G
O
X
O
- 2
(g)(h)
Figure 4. continued
W. B. Dunn et al./Heart failure metabolomics
421

Page 10

RNA degradation and thus of cell turnover. Tumour
cells exhibit an unusually high turnover, and conse-
quently it has also been proposed as a tumour marker
(e.g. (Higley et al., 1982; Salvatore et al., 1983; Russo
et al., 1984; Tamura et al., 1987; Amuro et al., 1988; X u
et al., 2000; Zheng et al., 2005)) where it can have sig-
nificant prognostic value (Pane et al., 1993). In heart
failureits raised level may in part reflect theremodelling
process in the heart itself (Morgan et al., 1987), the
exact dynamics of which are unknown, but a substan-
tially raised turnover of peripheral cells including
skeletal muscle seems a more likely explanation.
Remodelling processes in peripheral muscle are well
described in heart failure(Drexler et al., 1992; DeSousa
et al., 2000) and arethought to bein part responsiblefor
symptoms such as fatigue and shortness of breath.
Conceptually, a biomarker for peripheral cell turnover
would therefore be a valuable addition to a matrix
of markers classifying subgroups of patients. It is of
interest that the differences in pseudouridine between
Psi_UREA
5 10 15 20 25 30 35 40
UREA
5 10 15 20 25 30 35 40
UREA
5 10 15 20 25 30 35 40
UREA
0
0.1
0.2
0.3
0.4
0.5
E
N
I
D
I
R
U
O
D
U
E
S
P
Psi_creat
50 100 150 200 250 300 350
CREATININE
50 100 150 200 250 300 350
CREATININE
0
0.1
0.2
0.3
0.4
0.5
E
N
I
D
I
R
U
O
D
U
E
S
P
2OG_UREA
0
0.05
0.1
0.15
E
T
A
R
A
T
U
L
G
O
X
O
- 2
2OG_creat
0
0.05
0.1
0.15
E
T
A
R
A
T
U
L
G
O
X
O
- 2
CREAT_UREA
50
100
150
200
250
300
350
E
N
I
N
I T
A
E
R
C
CREAT_DIURETICS
NY
50
100
150
200
250
300
350
E
N
I
N
I T
A
E
R
C
DIURETICS
(a)(b)
(c)(d)
(e)(f)
Figure 5. Imperfect relationships between novel and existing heart failurebiomarkers and known renal dysfunction biomarkers. Pseudouridine
against (a) urea and (b) creatinine. 2-oxoglutarateagainst (c) urea and (d) creatinine. (e) urea vs creatinine. (f) creatinineas a function of diuretic
usage. (g) thetop two peaksfromuric acid areplotted against eachother, also showing that occasionally individual peakscan ?disappear?. (h) uric
acid vs creatinine. BNP against (i) creatinineand (j) urea. Notethat when urea is shown thescalehasbeen increased for visual clarity, leading to
the disappearance of 3 samples (2 case, 1 control) whose urea levels were greater than 60. All data are in the supplementary information. Cases
closed circles, controls open squares, as in Figure 4(f) data are jittered for clarity.
W. B. Dunn et al./Heart failure metabolomics
422

Page 11

cases and controls also increased somewhat with age,
although the levels in the controls were essentially
unchanged with age, aswasalso truefor creatinine(data
not shown, see also (Culleton et al., 1999a)).
We note that 2,4,6-trihydroxypyrimidine is often
displayed and described as its tautomer malonylurea
or barbituric acid. Its origin in mammals has not been
fully elucidated (http://www.genome.ad.jp/dbget-bin/
show_pathway?map00240± C02067), but the fact that
we identify it as a marker in an important syndrome
such as heart failure points to the value of further
research into this issue.
We note, too, that it is also possible that pseudouri-
dine and other metabolites whose concentrations are
URIC_ACID_PEAKS
0
0
1
2
3
4
5
1
D
I
C
A

C
I
R
U
URIC ACID 2
URIC_ACID_CREAT
0
1
2
3
4
5
1
D
I
C
A

C
I
R
U
CREATININE
BNP_CREATININE
0
0
500
1000
1500
2000
2500
3000
3500
4000
P
N
B
CREATININE
BNP_UREA
0
0
500
1000
1500
2000
2500
3000
3500
4000
P
N
B
UREA
1.4
1.21
0.80.6
0.4 0.2
350300 250
200150100
50
350300 250
200
150100
50
100
80 60
4020
(g)(h)
(i) (j)
Figure 5. continued
Table 3
Bootstrap-calculated Pearson?s correlation coeffi cients (r)
2-Oxoglutaric Acid *N-BNPUric acid* UreaCreatinine *
Pseudouridine*
0.005)0.409
0.297)0.926
)0.070–0.505
)0.454–0.399
0.007–0.751
0.219 ± 0.110
2-Oxoglutaric Acid*
0.018)0.444
0.029–0.448
)0.108–0.304
)0.168–0.296
0.561 ± 0.177
0.268 ± 0.109
N-BNP
)0.344–0.125
)0.245–0.674
0.017–0.494
0.299 ± 0.139
0.247 ± 0.108
)0.122 ± 0.119
Uric acid*
)0.186–0.431
0.068–0.497
)0.086 ± 0.218
0.103 ± 0.104
0.164 ± 0.263
0.138 ± 0.159
Urea
)0.430–0.248
0.489 ± 0.188
0.041 ± 0.118
0.207 ± 0.116
0.291 ± 0.112
)0.103 ± 0.173
Creatinine *
2-Oxoglutaric Acid *
N-BNP
Uric acid*
Urea
Creatinine *
If r = 0 this indicates that there is no correlation; r = 1 indicates that the two compared measurements are perfectly positively correlated;
r = )1 indicates perfect negative correlation. The bootstrap procedure (see (Efron and Gong, 1983; Efron and Tibshirani, 1993)) involved
repeatedly (p times) choosing n randomsampleswith replacement fromthediseaseonly data set (wheren = thetotal number of availabledisease
samples; thus a particular data point from this original data set could appear multipletimes in a given bootstrap sampleset) and analyzing each
samplethesameway – in this caseusing Pearson?scorrelation coeffi cient equation. For thistablep wasset to 5,000. Theupper right quadrant of
the table shows the mean of the 5,000 r values calculated for each pair-wise metabolite correlations, together with + /) the bootstrap standard
error. The lower left quadrant describes the same correlation data but in terms the 95% confidence interval for r for each pair-wise metabolite
correlation (calculated using thebiascorrected and accelerated percentilemethod; see(DiCiccio and Efron, 1996)). Thecorrelation data indicates
that all the metabolites are positively correlated with each other (i.e. r> 0 for all comparisons). However, there are no strong correlations
(r> 0.7). However, even the relatively high correlation between Pseudouridine and N-BNP (rest = 0.561 + /) 0.177) is questionable upon
inspection of such a small data set
W. B. Dunn et al./Heart failure metabolomics
423

Page 12

raised in heart failure are not merely innocent marker
metabolites but participate in the pathophysiology of
thesyndrome, e.g. by inducing cellular alterations in the
heart or in peripheral tissues. Such may be the case for
uric acid (Sakai et al., 2006).
Pseudouridine is also related to creatinine in our
overall dataset. This might be expected, given that a
reduction in renal function is part of the pathophysiol-
ogy of heart failure and indeed, reduced renal function
hasbeen considered to contributeto theincreasein BNP
in heart failure (Tsutamoto et al., 2006).
Conversely, reduced renal function contributes to the
progression of heart failure. The mechanisms of the
intimate link between renal and cardiac function,
sometimes referred to as the ?cardiorenal syndrome?
(Bongartz et al., 2005), are unclear, but our data raise
the interesting hypothesis that pseudouridine or the
other metabolites identified by us contribute to the
deterioration in renal function in heart failure. More
research is needed, but our present data clearly provide
novel and testable hypotheses to address these issues.
Overall, however, when we look at the heart failure
class alone, the correlation between BNP and creatinine
was weak (Table 3), as it was between any proposed
heart failure marker and any renal marker, and it
appears unlikely that a decrease in renal function is the
dominant explanation for theincreasein pseudouridine,
2OG and the other heart markers.
Aswith other marker studies, it cannot becompletely
excluded a priori that drug therapy influenced the level
of metabolites. Our analysis makes this highly unlikely,
but the ultimate proof of measuring the markers in
patients left untreated for several weeks is clearly not
ethical and this question may therefore be addressed
further by a combination of animal experiments and
indirect clinical investigations in future, e.g. by mea-
suring these parameters in patients with similar medi-
cation but other diseases, such as hypertension. By
definition, however, it will be impossible to identify a
non-heart failuregroup with identical combinationof all
drugs in a large patient cohort and without the use of
inferential methods some questions may remain unre-
solved in the human situation, in close analogy to BNP
in human heart failure.
2-oxoglutarate is a major intermediate of the tricar-
boxylic acid cycle (also known as citric acid or K rebs
cycle) that occupiesa central placein energy metabolism
and isoneof the12major precursorsfor thesynthesisof
most biochemical substances (Csete and Doyle, 2004)).
It was also substantially raised in the heart failure
patients. In recent years, it hasbecomeincreasingly clear
that heart failure is characterized by alterations in
energy metabolism. In humans, thereduction of theflux
rate from creatine phosphate to ATP is characterized
best as it is amenable to in vivo NMR measurements
(Neubauer et al., 1997). The most straightforward
hypothesis therefore is that the raised levels of 2-oxo-
glutaratereflect a decreased flux through theK rebscycle
and overflow of some metabolites into the circulation.
(In bacteria, 2-oxoglutarate is a classical product of
?overflow metabolism? in microbes (Neijssel and Tem-
pest, 1976) where as a partial oxidation product of
carbohydrate metabolism it signifies an insuffi cient oxi-
dative capacity.) Recent findings using metabolomics
to investigate alterations during cardiac ischemia
showed decreases in several constituents of the K rebs
cycle (though 2-oxoglutarate was not
named) (Sabatine et al., 2005). This suggests that the
metabolic state of the heart and/or peripheral tissues is
at least in part reflected in serum metabolites, which can
beharnessed as markers of disease. Further mechanistic
studies regarding this issue are warranted.
The origin of the sugar alcohol erythritol in serum is
unclear. Although there are a few reports of its mea-
surement in serum and urine (e.g. (Bultitude and
Newham, 1975; Pitka ¨ nen, 1972; Schoots et al., 1979;
Roboz et al.,1984, 1990;Verhoeven et al., 2001; Wame-
link et al., 2005)) it is not entirely certain whether its
origin is exogenous/dietary (as well as being a compo-
nent of various plant and animal tissues it is a permitted
food additive and pharmaceutical congener), by bio-
synthesis within human tissues (it does not yet appear in
the human metabolic network reconstruction (Duarte
et al., 2007)), or from gut microflora (a significant con-
tributor to the metabolome (Nicholson et al., 2005)).
However, as a reducing compound (Jauniaux et al.,
2005) it shares with 2-oxoglutarate the potential impli-
cation of an impaired oxidative capacity. Consistent
withthis, two other sugar alcohols(xylitol, myo-inositol)
also appear in the list of most discriminatory metabo-
lites (Table 2). The biological significance of 2-hydroxy,
2-methyl propanoic acid is unknown, although it too
may be of microbial origin (see (Hierro et al., 2005) for
some closely related metabolites).
In conclusion, an unbiased, hypothesis-generating
(K ell and Oliver, 2004) metabolomics strategy has
uncovered numerous novel candidate markers for heart
failure. Targeted analyses of these(using a method with
greater analytical precision) may be expected to prove
even more discriminatory, and a mechanistic biochemi-
cal and longitudinal analysis of their role as prognostic
indicators and in the aetiology of heart failure is clearly
warranted.
specifically
Acknowledgments
DBK thanks the BBSRC, EPSRC, RSC and BHF and
LN thankstheMRC and theBHF for financial support.
We thank the referees for some very thoughtful and
helpful comments. DBK thanks Prof Phil Baker for a
useful discussion. This is a contribution from the
Manchester Centre for Integrative Systems Biology
(www.mcisb.org).
W. B. Dunn et al./Heart failure metabolomics
424

Page 13

References
Amuro, Y ., Nakaoka, H., Shimomura, S., Fujikura, M., Y amamoto,
T., Tamura, S., Hada, T. and Higashino, K . (1988). Serum
pseudouridine as a biochemical marker in patients with hepa-
tocellular carcinoma. Clin. Chim. Acta 178, 151–158.
Anker, S.D., Doehner, W., Rauchhaus, M., Sharma, R., Francis, D.,
K nosalla, C., Davos, C.H., Cicoira, M., Shamim, W., K emp,
M., Segal, R., Osterziel, K .J., Leyva, F., Hetzer, R., Ponikowski,
P. andCoats, A.J. (2003). Uric acidandsurvival in chronic heart
failure: validation and application in metabolic, functional, and
hemodynamic staging. Circulation 107, 1991–1997.
Bland, J.M. and Altman, D.G. (1995). Multiple significance tests: the
Bonferroni method. BMJ 310, 170.
Bongartz, L.G., Cramer, M.J., Doevendans, P.A., Joles, J.A. and
Braam, B. (2005). The severe cardiorenal syndrome: ?Guyton
revisited?. Eur. Heart. J. 26, 11–7.
Broadhurst, D. and K ell, D.B. (2006). Statistical strategies for avoid-
ing false discoveries in metabolomics and related experiments.
Metabolomics 2, 171–196;DOI: 10.1007/s11306-006-0037-z.
Bultitude, F.W. and Newham, S.J. (1975). Identification of some
abnormal metabolites in plasma from uremic subjects. Clin.
Chem. 21, 1329–1334.
Charette, M. and Gray, M.W. (2000). Pseudouridine in RNA: what,
where, how, and why. IUBMB Life 49, 341–351.
Csete, M. and Doyle, J. (2004). Bow ties, metabolism and disease.
Trends Biotechnol. 22, 446–450.
Culleton, B.F., Larson, M.G., Evans, J.C., Wilson, P.W., Barrett, B.J.,
Parfrey, P.S. and Levy, D. (1999). Prevalence and correlates of
elevated serum creatinine levels: the Framingham Heart Study.
Arch. Intern. Med. 159, 1785–1790.
Culleton, B.F., Larson, M.G., K annel, W.B. and Levy, D. (1999).
Serum uric acid and risk for cardiovascular disease and death:
the Framingham Heart Study. Ann. Intern. Med. 131, 7–13.
DeSousa, E., Veksler, V., Bigard, X ., Mateo, P. and Ventura-Clapier,
R. (2000). Heart failure affects mitochondrial but not myofibr-
illar intrinsic properties of skeletal muscle. Circulation 102,
1847–1853.
DiCiccio, T.J. and Efron, B. (1996). Bootstrap confidence intervals.
Stat. Sci. 11, 189–228.
Drexler, H., Riede, U., Munzel, T., K onig, H., Funke, E. and Just, H.
(1992). Alterations of skeletal muscle in chronic heart failure.
Circulation 85, 1751–1759.
Duarte, N.C., Becker, S.A., Jamshidi, N., Thiele, I., Mo, M.L., Vo,
T.D., Srvivas, R. and Palsson, B.Ø. (2007). Global reconstruc-
tion of the human metabolic network based on genomic and
bibliomic data. Proc. Natl. Acad. Sci. 104, 1777–1782.
Dunn, W.B. and Ellis, D.I. (2005). Metabolomics: current analytical
platforms and methodologies. Trends Anal. Chem. 24, 285–294.
Efron, B. and Gong, G. (1983). A leisurely look at the bootstrap, the
jackknife, and cross-validation. American Statistician 37, 36–48.
Efron, B. and Tibshirani, R.J. (1993). Introduction to the Bootstrap.
Chapman and Hall, London.
Finck, B.N. and K elly, D.P. (2006). PGC-1 coactivators: inducible
regulators of energy metabolism in health and disease. J. Clin.
Invest. 116, 615–622.
Harrigan, G.G. and Goodacre, R. (Eds) (2003). Metabolic Profiling:
its Role in Biomarker Discovery and Gene Function analysis.
K luwer Academic Publishers, Boston.
Hierro, E., Ordonez, J., Bruna, J. M., Pin, C., Fernandez, M. and dela
Hoz, L. (2005). Volatile compound generation in dry fermented
sausages by the surface inoculation of selected mould species.
Eur. Food Res. Technol. 220, 494–501.
Higley, B., DeMello, J. Jr., Oakes, D.J. andGiles, G. R. (1982). Serum
pseudouridine concentration as a marker of gastrointestinal
cancer. Clin. Oncol. 8, 305–312.
Hollander, M. and Wolfe, D.A. (1973). Nonparametric Statistical
Methods. Wiley, New Y ork.
Jauniaux, E., Hempstock, J., Teng, C., Battaglia, F. C. and Burton,
G.J. (2005). Polyol concentrations in the fluid compartments of
the human conceptus during the first trimester of pregnancy:
maintenance of redox potential in a low oxygen environment.
J. Clin. Endocrinol. Metab. 90, 1171–1175.
K ell, D.B. (2004). Metabolomicsand systemsbiology: making senseof
the soup. Curr. Op. Microbiol. 7, 296–307.
K ell, D.B. (2006). Metabolomics, modelling and machine learning in
systems biology: towards an understanding of the languages of
cells. The 2005 Theodor Bu ¨ cher lecture. FEBS J. 273, 873–894.
K ell, D.B. (2006). Systems biology, metabolic modelling and meta-
bolomics in drug discovery and development. Drug Disc. Today
11, 1085–1092.
K ell, D.B. and Oliver, S.G. (2004). Here is the evidence, now what is
the hypothesis? The complementary roles of inductive and
hypothesis-driven sciencein the post-genomic era. Bioessays 26,
99–105.
K enny, L.C., Dunn, W.B., Ellis, D.I., Myers, J., Baker, P. N., The
GOPEC Consortium and K ell, D.B. (2005) Novel biomarkers
for pre-eclampsia detected using metabolomics and machine
learning. Metabolomics. 1, 227–234 - online DOI: 10.1007/
s11306–005-0003-1.
K irschenlohr, H.L., Griffi n, J.L., Clarke, S.C., Rhydwen, R., Grace,
A.A., Schofield, P.M., Brindle, K . M. and Metcalfe, J. C. (2006).
Proton NMR analysisof plasma isa weak predictor of coronary
artery disease. Nat. Med. 12, 705–710.
Leyva, F., Anker, S., Swan, J.W., Godsland, I.F., Wingrove, C.S.,
Chua, T.P., Stevenson, J.C. and Coats, A.J. (1997). Serum uric
acid as an index of impaired oxidative metabolism in chronic
heart failure. Eur. Heart. J. 18, 858–865.
Leyva, F., Anker, S.D., Godsland, I.F., Teixeira, M., Hellewell, P.G.,
K ox, W.J., Poole-Wilson, P.A. and Coats, A.J. (1998). Uric acid
in chronic heart failure: a marker of chronic inflammation. Eur.
Heart. J. 19, 1814–1822.
Leyva, F., Chua, T.P., Anker, S.D. andCoats, A.J. (1998). Uric acid in
chronic heart failure: a measure of the anaerobic threshold.
Metabolism 47, 1156–1159.
Mackay, D.J.C. (2003). Information Theory, Inference and Learning
Algorithms. Cambridge University Press, Cambridge.
Morgan, H.E., Gordon, E.E., K ira, Y ., Chua, H.L., Russo, L.A.,
Peterson, C.J., McDermott, P.J. and Watson, P. A. (1987).
Biochemical mechanisms of cardiac hypertrophy. Annu. Rev.
Physiol. 49, 533–543.
Neijssel, O.M. and Tempest, D.W. (1976). Theroleof energy-splitting
reactions in the growth of Klebsiella aerogenes NCTC 418 in
aerobic chemostat culture. Arch. Microbiol. 110, 305–311.
Neubauer, S., Horn, M., Cramer, M., Harre, K ., Newell, J. B., Peters,
W., Pabst, T., Ertl, G., Hahn, D., Ingwall, J. S. and K ochsiek,
K . (1997). Myocardial phosphocreatine-to-ATP ratio is a pre-
dictor of mortality in patients with dilated cardiomyopathy.
Circulation 96, 2190–2196.
Nicholson, J.K ., Holmes, E. and Wilson, I.D. (2005). Gut microor-
ganisms, mammalian metabolism and personalized health care.
Nat. Rev. Microbiol. 3, 431–438.
Niwa, T., Takeda, N. and Y oshizumi, H. (1998). RNA metabolism in
uremic patients: accumulation of modified ribonucleosides in
uremic serum. Technical note. Kidney Int. 53, 1801–1806.
O?Hagan, S., Dunn, W.B., Broadhurst, D., Williams, R., Ashworth,
J.A., Cameron, M., K nowles, J. and K ell, D. B. (2007). Closed-
loop, multi-objective optimisation of two-dimensional gas
chromatography (GCxGC-tof-MS) for serum metabolomics.
Anal. Chem. 79, 464–476.
O?Hagan, S., Dunn, W.B., Brown, M., K nowles, J.D. and K ell, D.B.
(2005). Closed-loop, multiobjective optimisation of analytical
instrumentation: gas-chromatography-time-of-flight mass spec-
trometry of the metabolomes of human serum and of yeast
fermentations. Anal. Chem. 77, 290–303.
W. B. Dunn et al./Heart failure metabolomics
425

Page 14

Ofengand, J. (2002). Ribosomal RNA pseudouridines and pseudo-
uridine synthases. FEBS Lett. 514, 17–25.
Oliver, S.G., Winson, M.K ., K ell, D.B. and Baganz, F. (1998). Sys-
tematic functional analysis of the yeast genome. Trends Bio-
technol. 16, 373–378.
Ores ˇ ic ˇ , M., Vidal-Puig, A. and Ha ¨ nninen, V. (2006). Metabolomic
approaches to phenotype characterization and applications to
complex diseases. Expert Rev. Mol. Diagn. 6, 575–585.
Pan, J.P., Liu, T.Y ., Chiang, S.C., Lin, Y .K ., Chou, C.Y ., Chan, W.L.
and Lai, S.T. (2004). The value of plasma levels of tumor
necrosis factor-alpha and interleukin-6 in predicting theseverity
and prognosis in patients with congestive heart failure. J. Chin.
Med. Assoc. 67, 222–228.
Pane, F., Savoia, M., Fortunato, G., Camera, A., Rotoli, B.,
Salvatore, F. and Sacchetti, L. (1993). Serum pseudouridine in
the diagnosis of acute leukaemias and as a novel prognostic
indicator in acute lymphoblastic leukaemia. Clin. Biochem. 26,
513–520.
Pearl, J. (2000). Causality: Models, Reasoning and Inference. Cam-
bridge University Press, Cambridge.
Perneger, T.V. (1998). What?s Wrong with Bonferroni Adjustments.
Bmj 316, 1236–1238.
Pitka ¨ nen, E. (1972). Serumpolyol pattern and urinary polyol excretion
in diabetic and in uremic patients. Clin. Chim. Acta 38, 221–230.
Raamsdonk, L.M., Teusink, B., Broadhurst, D., Zhang, N., Hayes, A.,
Walsh, M., Berden, J.A., Brindle, K .M., K ell, D.B., Rowland,
J.J., Westerhoff, H.V., van Dam, K . and Oliver, S.G. (2001). A
functional genomicsstrategy thatusesmetabolomedatatoreveal
thephenotypeof silent mutations. Nature Biotechnol. 19, 45–50.
Ransohoff, D.F. (2005). Bias as a threat to the validity of cancer
molecular-marker research. Nat. Rev. Cancer 5, 142–149.
Roboz, J., K appatos, D. and Holland, J.F. (1990). Polyol concentra-
tions in serum during hemodialysis. Clin. Chem. 36, 2082–2086.
Roboz, J., K appatos, D.C., Greaves, J. and Holland, J. F. (1984).
Determination of polyols in serum by selected ion monitoring.
Clin. Chem. 30, 1611–1615.
Rothman, K .J. and Greenland, S. (1998). Modern Epidemiology.
(second ed.). Lippincott, Williams & Wilkins, Philadelphia.
Russo, T., Colonna, A., Salvatore, F., Cimino, F., Bridges, S. and
Gurgo, C. (1984). Serumpseudouridineasa biochemical marker
in the development of AK R mouse lymphoma. Cancer Res. 44,
2567–2570.
Sabatine, M.S., Liu, E., Morrow, D.A., Heller, E., McCarroll, R.,
Wiegand, R., Berriz, G.F., Roth, F.P. and Gerszten, R.E.
(2005). Metabolomic identification of novel biomarkers of
myocardial ischemia. Circulation 112, 3868–3875.
Sakai, H., Tsutamoto, T., Tsutsui, T., Tanaka, T., Ishikawa, C. and
Horie, M. (2006). Serum level of uric acid, partly secreted from
the failing heart, is a prognostic marker in patients with con-
gestive heart failure. Circ. J. 70, 1006–1011.
Salvatore, F., Russo, T., Colonna, A., Cimino, L., Mazzacca, G. and
Cimino, F. (1983). Pseudouridinedetermination in blood serum
as tumor marker. Cancer Detect. Prev. 6, 531–536.
Schoots, A.C., Mikkers, F.E.P., Cramers, C.A.M.G. and Ringoir, S.
(1979). Profiling of uremic serum by high-resolution gas chro-
matography-electron-impact, chemical ionization mass spec-
trometry. J. Chromatogr. 164, 1–8.
Smith, G.L., Shlipak, M.G., Havranek, E.P., Foody, J.M., Masoudi,
F.A., Rathore, S.S. and K rumholz, H.M. (2006). Serum urea
nitrogen, creatinine, and estimators of renal function: mortality
in older patients with cardiovascular disease. Arch. Intern. Med.
166, 1134–1142.
Tamura, S., Fujioka, H., Nakano, T., Hada, T. and Higashino, K .
(1987). Serum pseudouridine as a biochemical marker in small
cell lung cancer. Cancer. Res. 47, 6138–6141.
Troughton, R.W., Frampton, C.M., Y andle, T.G., Espiner, E.A.,
Nicholls, M.G. and Richards, A.M. (2000). Treatment of heart
failure guided by plasma aminoterminal brain natriuretic pep-
tide (N-BNP) concentrations. Lancet 355, 1126–1130.
Tsutamoto, T., Wada, A., Sakai, H., Ishikawa, C., Tanaka, T.,
Hayashi, M., Fujii, M., Y amamoto, T., Dohke, T., Ohnishi, M.,
Takashima, H., K inoshita, M. and Horie, M. (2006). Relation-
ship between renal function and plasma brain natriuretic
peptide in patients with heart failure. J. Am. Coll. Cardiol. 47,
582–586.
Underwood, B.R., Broadhurst, D., Dunn, W.B., Ellis, D.I., Michell,
A.W., Vacher, C., Mosedale, D.B., K ell, D.B., Barker, R.,
Grainger, D.J. and Rubinsztein, D. C. (2006). Huntington?s
disease patients and transgenic mice have similar pro-catabolic
serum metabolite profiles. Brain 29, 877–886.
van der Greef, J., Hankemeier, T. and McBurney, R.N. (2006). Met-
abolomics-based systems biology and personalized medicine:
moving towards n = 1 clinical trials?. Pharmacogenomics 7,
1087–1094.
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.
Verhoeven, N.M., Huck, J.H., Roos, B., Struys, E. A., Salomons,
G.S., Douwes, A.C., van der K naap, M. S. and Jakobs, C.
(2001). Transaldolasedeficiency: liver cirrhosisassociated with a
new inborn error in the pentose phosphate pathway. Am. J.
Hum. Genet. 68, 1086–1092.
Wacholder, S., Chanock, S., Garcia-Closas, M., El Ghormli, L. and
Rothman, N. (2004). Assessing the probability that a positive
report is false: an approach for molecular epidemiology studies.
J. Natl. Cancer. Inst. 96, 434–442.
Wamelink, M.M., Smith, D.E., Jakobs, C. and Verhoeven, N.M.
(2005). Analysis of polyols in urine by liquid chromatography-
tandem mass spectrometry: A useful tool for recognition of
inbornerrorsaffecting polyol metabolism. J. Inherit. Metab. Dis.
28, 951–963.
X u, G., Schmid, H.R., Lu, X ., Liebich, H.M. and Lu, P. (2000).
Excretion pattern investigation of urinary normal and modified
nucleosides of breast cancer patients by RP-HPLC and factor
analysis method. Biomed Chromatogr 14, 459–63.
Zaphiriou, A., Robb, S., Murray-Thomas, T., Mendez, G., Fox, K .,
McDonagh, T., Hardman, S.M., Dargie, H.J. and Cowie, M.R.
(2005). Thediagnostic accuracy of plasma BNP and NTproBNP
in patients referred from primary care with suspected heart
failure: resultsof theUK natriuretic peptidestudy. Eur. J. Heart
Fail. 7, 537–541.
Zheng, Y .F., K ong, H.W., X iong, J.H., Lv, S. and X u, G.W. (2005).
Clinical significanceand prognostic valueof urinary nucleosides
in breast cancer patients. Clin Biochem 8, 24–30.
W. B. Dunn et al./Heart failure metabolomics
426