University of Calgary PRISM: University of Calgary's Digital Repository Graduate Studies The Vault: Electronic Theses and Dissertations 2013-02-01 Multivariate NMR analysis of human disease models Duggan, Gavin Duggan, G. (2013). Multivariate NMR analysis of human disease models (Unpublished doctoral thesis). University of Calgary, Calgary, AB. doi:10.11575/PRISM/27043 http://hdl.handle.net/11023/540 doctoral thesis University of Calgary graduate students retain copyright ownership and moral rights for their thesis. You may use this material in any way that is permitted by the Copyright Act or through licensing that has been assigned to the document. For uses that are not allowable under copyright legislation or licensing, you are required to seek permission. Downloaded from PRISM: https://prism.ucalgary.ca
Transcript
University of Calgary
PRISM: University of Calgary's Digital Repository
Graduate Studies The Vault: Electronic Theses and Dissertations
2013-02-01
Multivariate NMR analysis of human disease models
Duggan, Gavin
Duggan, G. (2013). Multivariate NMR analysis of human disease models (Unpublished doctoral
thesis). University of Calgary, Calgary, AB. doi:10.11575/PRISM/27043
http://hdl.handle.net/11023/540
doctoral thesis
University of Calgary graduate students retain copyright ownership and moral rights for their
thesis. You may use this material in any way that is permitted by the Copyright Act or through
licensing that has been assigned to the document. For uses that are not allowable under
copyright legislation or licensing, you are required to seek permission.
In the late 1990s, the field of metabolic profiling evolved into metabolomics
following the general move towards systems biology and other omics techniques.
Using sensitive, analytical platforms such as NMR, metabolomics aims to gather an
unbiased, broad perspective of the active biochemistry in biofluids. The result was
an explosive growth in the data available to study short term physiological effects,
followed perforce by the application of multivariate pattern-recognition techniques
to aid in its interpretation.
Given the sensitive and comprehensive nature of the technique, it quickly became
apparent that any number of artifactual or spurious relationships appear in the
results. To alleviate those concerns, a variety of improved experimental designs,
analytical techniques, and validation paradigms can be applied. Starting with a
basic experimental design, the aim of this work is to explore the ability of properly
validated metabolomics to provide useful information about the metabolic shifts
seen in established animal models of insulin resistance, a human disease with
increasing medical significance. Different two-factor experimental designs are used
to refine the results of this early study, validate the resulting hypothesis and
reinforce its interpretation.
Having seen significant differences in ostensibly identical batches of animals in the
first three experiments, further analysis of the differences are performed.
Techniques for comparing batch models, as a form of multivariate hypothesis
validation, are evaluated and the ability of statistical techniques to predict or
ameiliorate these “batch effects” is studied. Finally, a rat model of vitamin C
deficiency, another condition with ongoing pathological implications in the third
world, is studied using the same metabolomic techniques. The identified metabolic
shifts are subjected to a complete pathway analysis, the context of which provides a
potentially interesting insight into the regulation of an important human oxidative
damage control mechanism.
iii
ACKNOWLEDGEMENTS
Research is never a solitary venture, and all of the work presented herein was
accomplished with significant assistance. I would like to thank Dr. Vogel and Dr.
Weljie for the ongoing guidance, patience, and encouragement. Additionally, many
collaborators were instrumental to the success of each project.
At every stage, Dr. Jane Shearer has provided every measure of pivotal assistance.
The experimental design for chapters 2, 3, and 4 were developed in whole or in part
with her assistance. Her expertise in animal management was essential to the
acquisition of high quality data, upon which any analytical techniques are
dependant.
The vitamin-C analysis performed in chapter 6 is based on the Gulo -/- mouse
model developed by Dr Frank Jirik. All of the animal handling and fluid acquisition
was performed by Joan Miller and other members of Dr Jirik’s lab.
Over the years, my teachers, friends, and family have provided years of support and
encouragement without which I could never have started, let alone completed, this
endeavour. I’d like to thank all of them, but most of all Sarah, with whose help a
great many more things are possible, and everything is more enjoyable.
iv
DEDICATION
To the many great teachers I’ve had – most especially
Pierre Poitras
Ruth Benda
Morris Tchir
for their willingness and ability to inspire, motivate, and believe.
v
TABLE OF CONTENTS
Abstract ............................................................................................................................... ii Acknowledgements ............................................................................................................ iii Dedication .......................................................................................................................... iv
Table of Contents .................................................................................................................v
List of Tables .................................................................................................................... vii List of Figures ................................................................................................................... vii LIST of Abbreviations ..................................................................................................... viii Citations for previously published chapters ....................................................................... ix
1.5 Where does metabolomics break down? .................................................................47
1.6 Research Scope ........................................................................................................49
CHAPTER 2 - METABOLOMIC PROFILING OF DIETARY-INDUCED INSULIN RESISTANCE IN THE FAT–FED MOUSE ...........................................................52
CHAPTER 3 - DIFFERENTIATING SHORT- AND LONG-TERM EFFECTS OF DIET IN THE OBESE MOUSE USING 1H-NUCLEAR MAGNETIC RESONANCE METABOLOMICS...................................................................................................70
CHAPTER 4 METABOLOMIC RESPONSE TO EXERCISE TRAINING IN LEAN AND DIET-INDUCED OBESE MICE .............................................................................77
CHAPTER 6 - METABOLIC PROFILING OF VITAMIN C DEFICIENCY IN GULO−/− MICE USING PROTON NMR SPECTROSCOPY ...............................................108
Figure 6.5: First component OPLS loadings of WT vs KO mice ............................................... 118
Figure 6.6 KEGG pathway diagram for serine and glycine metabolism .................................... 123
viii
LIST OF ABBREVIATIONS
1H Single-proton hydrogen 13C Carbon-13 DNA Deoxyribonucleic acid HSQC Heteronuclear Single Quantum Coherence, an NMR Spectroscopy technique KEGG The Kyoto Encyclopedia of Genes and Genomes MS Mass Spectrometry mTOR Mammalian Target of Rapamycin protein NMR Nuclear Magnetic Resonance Spectroscopy OPLS Orthogonal PLS PCA Principle Component Analysis PI3K Phosphatidylinsositol 3-kinase family of enzymes PLS Projection to Latent Structures -DA Discriminant Analysis RNA Ribonucleic Acid TCA Tricarboxylic acid cycle
ix
CITATIONS FOR PREVIOUSLY PUBLISHED CHAPTERS
Chapter 2 – (Shearer et al., 2008)
Shearer, J, Duggan, G., Weljie, A., Hittel, D. S., Wasserman, D. H., & Vogel, H. J. (2008). Metabolomic profiling of dietary-induced insulin resistance in the high fat-fed C57BL/6J mouse. Diabetes, Obesity & Metabolism, 10(10), 950–958. doi:10.1111/j.1463-1326.2007.00837.x
Chapter 3 (Duggan et al., 2011a)
Duggan, G.E., Hittel, D.S., Hughey, C.C., Weljie, A., Vogel, H.J., and Shearer, J. (2011). Differentiating short-and long-term effects of diet in the obese mouse using 1H-nuclear magnetic resonance metabolomics. Diabetes, Obesity and Metabolism 13, 859–862.
Chapter 4 –(Duggan et al., 2011b)
Duggan, G.E., Hittel, D.S., Sensen, C.W., Weljie, A.M., Vogel, H.J., and Shearer, J. (2011b). Metabolomic response to exercise training in lean and diet-induced obese mice. Journal of Applied Physiology 110, 1311–1318.
Chapter 6 – (Duggan et al., 2011c)
Duggan, G.E., Joan Miller, B., Jirik, F.R., and Vogel, H.J. (2011). Metabolic profiling of vitamin C deficiency in Gulo-/- mice using proton NMR spectroscopy. Journal of Biomolecular NMR 49, 165–173.
10
Chapter 1 - Introduction
1.1 Preamble and Context
Systems biology has been an important renaissance in the analysis of living systems,
human or otherwise. Where the study of individual elements yields insight to their
properties, examining the connections between gives them context (Aristotle, 1985).
Connections, and the networks they form, represent the true combinatorial
complexity of life, and only by studying the interactions between elements can the
entirety of a process, ergo its function, be elucidated (Oliver et al., 1998; Kell, 2006;
Trewavas, 2006; van der Greef et al., 2007).
Perhaps the most visible manifestation of this revolution has been the proliferation
of -omics technologies. In the last two decades of the previous century, high-
throughput sequencing techniques gave rise to the study of Genomics, with the
promise of completely characterising our biology by studying the genetic blueprint
from which it is constructed (Collins et al., 1998). The focus of genetic
experiments shifted from individual genetic loci, identified by their relationship to
some phenomenon, to an unbiased characterisation of an entire DNA sequence
which could be mined and analyzed a priori (McElheny, 2012). This shift in
perspective was not the precursor of systems biology but rather the harbinger of its
rise to prominence in the scientific and popular awareness; the study of interactions
in biological systems dates back decades (Woodger, 1929; Von Bertalanffy, 1950;
Pauling et al., 1971), if not longer (Descartes, 1643).
Because of their relatively static nature of genetic material, genomics was a natural
first frontier for this form of automated, unbiased analysis of biological state
(Lander et al., 2001). Unfortunately, pure1 genomics also suffers because the same
1 “Pure” genomics referring to the study of exclusively the DNA sequence, which
we now know (in large part due to genomics’ inability to explain many biological
11
relative intransience of its subject matter: genetic material ostensibly does not
change over the course of an organism’s lifespan. Life is dynamic by definition, not
the least because of its inseparable relationship with its environment, and hence it
was desirable to develop techniques for the study of other, more dynamic biological
entities in the same unbiased fashion.
Proteins were the next step in the analysis of ongoing changes in an organism’s
state; whether receptors or enzymes, they are the primary means by which a living
cell recognizes and reacts to changes in its external or internal state. Moreover,
protein’s relatively large mass, segmental structure, and chemical composition made
them amenable to examination by mass spectrometry: the field of Proteomics was
the result, fast on the heels of genomics (James, 1997). The mechanisms of
production and expression being inexorably linked to the operation of protein
networks within the cell, transcriptomics followed almost immediately by
leveraging microarray technology to analyse RNA molecules in parallel (Clark et
al., 2002; Bertone et al., 2004).
Like genomics though, protein-centric -omics failed to capture many temporal facets
of an organism’s state (Vitzthum et al., 2005; Mai et al., 2011; Zhu et al., 2011).
Protein expression being an energetically expensive, tightly regulated, and
physically constrained process, the time scale on which protein levels change is still
relatively slow compared to many physiological and environmental changes (Bino
et al., 2004; Scalbert et al., 2009). The smaller, chemical-scale molecules with
which proteins interact represent the largest source of energy, entropy, and (as a
combination of the two) dynamic information content in a cell (Bino et al., 2004;
van der Greef et al., 2007).
phenomena) represents only some of the information contained within genetic
material, hence the development of epi-genomics.
12
Small molecule metabolites act as long distance, concentration- based signals and
effectors of change within and outside the cell. They turn over on a much faster
range of time scale within a cell, making them a closer indicator of environmental
response, and changes in small-molecule concentrations can affect the activity level
of enzymes in a cell via mechanisms such as allosteric modification or classical
Michaelis-Menten enzyme kinetics (Raamsdonk et al., 2001; Fiehn, 2002; Goodacre
et al., 2004; Kell, 2006). As a result, metabolites are both products of, and
information conduits within, protein function networks, which makes them
extremely informative targets of investigation.
Metabolites are also relatively easy to collect. Whether they’re exported from the
cell, the result of membrane-local processes, or released from the cellular membrane
by experimental permeation, small molecules provide an accessible window into the
state of, or interactions between, biological processes (Fiehn, 2002). The ability of a
single protein molecule to catalyse changes in numerous metabolites enhances this
accessibility further by decreasing the need for sensitivity; when multiple proteins
form a metabolic cascade as in the case of insulin signalling, the result is an
exponential growth in the metabolic information content (Fell, 1997).
As such, it’s no coincidence studies of metabolism are some of the oldest forms of
biological investigation, a mainstay of medical diagnosis and an important (arguably
indispensable) level at which to study systems biology (Bollard et al., 2005;
Nicholson, 2006). Historically though, the only means available for quantifying the
presence or concentration of molecular entities was with highly specific assays
dependant on their chemical properties (Fell, 1997). Starting in early 1980s, early
systems biology efforts used NMR to fish potential biomarkers from serum samples
using e.g. lipid peak widths (Gadian and Radda, 1981; Bottomley, 1985;
Dawidowicz, 1987; Cerdan et al., 1990). As with genomics and proteomics, the
inflection point in the advancement of metabolic analysis was the advent of
technical platforms which can segregate and quantify multiple metabolites’
concentration simultaneously from a single sample (Gadian and Radda, 1981). With
13
the adoption of sufficiently high-resolution instrumentation, it became possible to
deconstruct a single sample’s composite spectrum into components (Nicholson and
Wilson, 1987; Fiehn, 2002). By analyzing the resulting concentrations in parallel,
and without a priori knowledge of their significance to a particular phenomenon, the
field of metabolomics was conceived.
Within a decade, metabolomics showed great promise in agriculture (Dixon et al.,
2006), disease research (Griffin, 2006), nutrition (Zivkovic and German, 2009),
medical diagnostics (Vitzthum et al., 2005), pharmaceutical development (Frank
and Hargreaves, 2003; Malet-Martino and Holzgrabe, 2010; Robertson and Reily,
2012), among other fields. On its own or as part of a larger systems biology
approach, the ability to capture a comprehensive metabolic snapshot from biofluids
was extremely enticing. Once a standard methodology was established, it was
applied to a vast array of diseases or biological conditions.
Like any nascent field of study (arguably any active field) however, metabolomics is
not without its faults and limitations (Ransohoff and Feinstein, 1978). There have
been notable failures in experiments’ accuracy (Bantscheff et al., 2007; Zhu et al.,
2011), and numerous studies purporting to yield breakthrough insights have failed to
have useful implications. The same automated collection of large data sets which
allowed for the shift in focus from individual metabolites to comprehensive
metabolic state also overwhelmed classical analysis tools. Instrumental procedures
required refinement to ensure the data was properly annotated (Brown et al., 2005),
and novel statistical techniques had to be developed or adapted to the inherent
character of metabolomics data. Biological interpretation of the resulting numerical
inferences has proven difficult and has often been left unfinished (Robertson and
Reily, 2012).
Perhaps most importantly, the exploratory nature of metabolomics methods and the
ability to uncover potential patterns in data presents a strong temptation to label
interesting hypotheses as unequivocal conclusions or game-changing insights. The
14
same trait which makes metabolomics so valuable, its ability to examine the
interconnected gestalt of a biological organism, also makes it difficult to segregate
any one element.
As a result, there is perhaps no such thing as a “simple” metabolomics study and
strongly meaningful conclusions are difficult to come by (Nicholls, 2012). For the
past decade, there has been, and continues to be, a need to properly develop, refine,
and reinforce metabolomics methods in order to make them more robust and
meaningful.
1.2 What is metabolomics?
Varied implementations, definitions, and labels notwithstanding, the term
metabolomics has had a reasonably consistent core meaning in the decade since its
inception (Nicholson and Wilson, 1987; Oliver et al., 1998). Ostensibly unbiased
selection of targets has always been the basis for the -omic nomenclature, and other
facets stem from the interaction of goals and technical limitations. For the purposes
of this thesis, the process of analyzing samples can be broken down into four steps
which, while varied in their implementation, are consistent in all metabolomics
methodologies.
Small molecule focus. While the study of metabolic-scale small molecules is
advantageous for a number of reasons, the technical limitations of both NMR or MS
instruments are an unavoidable motivator. In either case, the maximum structural
complexity which can be uniquely deconvolved from a complex spectrum of
multiple known compounds is far smaller than what can be identified in an
unknown but pure sample. Both platforms center on the detection of molecular
elements, returning spectra whose features correspond to the presence of a particular
molecular substructure. The physical limitations of mass/charge ratios and
fragmentation (in the MS paradigm), or spin relaxation times (in the NMR
paradigm), are dependent on increasing but finite instrumental limits. As a result,
while “metabolites” in the technical sense may comprise a larger range of molecular
15
entities, the focus of “metabolomics” is usually restricted to molecules below a
fairly low given molecular weight/structural complexity, usually only a few hundred
g/mol.
Nevertheless, exceptions to this principle do occur, when investigators trade off
specificity for the ability to handle larger or more complicated metabolic
mixtures/identities. Examples include the field of lipidomics, which analyses lipid
concentrations in a manner analogous to metabolomics’ perspective of smaller
molecules. Notwithstanding the order of magnitude difference in size,
metabolomics and lipidomics are often referred to collectively because of the similar
techniques used.
Unbiased selection. Within the scope of metabolic entities, the essence of systems
biology is the goal of characterising as much as possible of the system before
identifying which elements are relevant to a particular problem. Regardless of
platform or processing technique, the pivotal aspect is that data is not excluded from
acquisition, but instead by topical analysis.
Simultaneous Acquisition. An experiment which measured numerous metabolites
individually and then combined them into a single dataset for analysis could be
considered “metabolomics”, but due to practical constraints this is almost never the
case. The number of molecular species which need to be captured for the
experiment to be considered unbiased precludes anything but parallel, simultaneous
acquisition using one of the instruments outlined below.
Numerical analysis of concentrations. The true value of high-throughput
metabolite capture is the structured, numerical format of the data. The size of the
resulting matrix dictates the application of both statistical analysis and pattern
recognition techniques in order to draw biological conclusions. Biological states
being highly dynamic, it is the consistent differences (labelled “variance”) between
time points or treatments which usually convey the necessary information;
16
biological processes being stochastic in nature and varied in circumstance, context-
free analysis of concentrations is much harder to translate into practical application.
Taken together, these four aspects represent the commonality between almost all
metabolomics experimentation, and the definition given above. Beyond that,
metabolomics has seen a wide variety of implementations. Even the nomenclature
has been fragmented, with groups attempting first to modify the terminology, and
then trying to spin off sub-fields below the useful linguistic resolution. Taken as a
whole, however, metabolomics has produced a remarkably expansive body of
insight into numerous problems.
1.3 What is metabolomics used to study?
Given that changes in small-molecule metabolism are a major response mechanism
of living organisms at every scale, it’s unsurprising that efforts have been made to
apply metabolomics to a wide range of problems and perspectives.
1.3.1 Disease Models
Bacteria are easy to manipulate and devoid of ethical concerns, leading to numerous
studies of bacterial metabolism (Mashego et al., 2007). Graham-negative bacteria,
in particular, have been particularly interesting because of their tendency to express
numerous metabolites into their growth medium, resulting in a metabolic footprint
in contrast to extraction of intracellular state, referred to as a fingerprint. From a
pharmacological perspective, metabolomics has been used to study both established
and potential antibiotic agents for both efficacy and method of action. Identification
of bacterial gene functionality by knockout (Idle and Gonzalez, 2007) has helped
annotation efforts, and analysis of flux facilitated by 13C labelling is an important
bacterial platform (Zamboni and Sauer, 2009). Interactions between gut microflora
17
and human metabolism have been highlighted by numerous studies (Nicholls et al.,
2003; Wikoff et al., 2009) as both common and potentially important for human
health. Where bacterial culture interacts pathologically with mammalian biology,
metabolomics has provided potential for both diagnostic and functional insight
(Stringer et al., 2011).
Moving up the evolutionary ladder, yeast, plant, and animal species have been
studied extensively via metabolomics. Agricultural biotechnologies were among the
first common applications for metabolomics (Hall et al., 2002; Dixon et al., 2006)
and continue to be integrated with other systems biology to improve crop science
(Sakurai et al., 2011). Both plants and microbiotic metabolomics have been
leveraged to study extracted and expressed natural products (Oksman-Caldentey and
Inzé, 2004). Animal health studies have been conducted to improve livestock,
husbandry, and ecological conditions (Moore, Kirwan, Doherty, & Whitfield, n.d.).
1.3.2 Epidemiology, toxicology, and in situ pathology
Of course, the potential for non-invasive metabolic profiling has the highest
potential in areas related to human health. By applying metabolomics methods,
studies of nutrition have supplemented work in agriculture to examine the human
effects of energy intake, vitamin deficiency, nutritional supplements, dietary
composition, etc. In the realm of translational drug research, pharmacodynamics is
a high-value target which is difficult to evaluate. Systematic factors like efficacy,
off-target effects, and dosing calibration can be facilitated by whole-body metabolic
profiling (Kaddurah-Daouk et al., 2008). These studies of drug dynamics are
dependant to a degree on characterisation of the recipient’s disease state, which has
been a primary focus of metabolomics since its inception.
Regardless of the application, characterising disease state can be broken down,
under various nomenclature, into diagnostic and explanatory studies. The
dichotomy is artificial but pragmatic, with the former indicating a desire solely to
differentiate between outcomes, where the latter focusing on the ostensible
18
biochemical mechanism. Modern clinical practice being primarily a mapping of
disease to established treatment options, diagnostic power is held in high regard in
those circles. Developing treatments, understanding of disease, and other
epidemiological advances hinge on explaining the underlying mechanisms of
disease however. Black box diagnostic tools can be developed without assigning
biochemical function to metabolic shifts, and indeed can be useful for tracking
disease progress or physiological response to interventions (pharmalogical (Malet-
Martino and Holzgrabe, 2010), environmental (Cartee et al., 1989; Boulé et al.,
2001; Miccheli et al., 2009), etc) (Madsen et al., 2010), but are still a dead end vis-
a-vis future development.
As such the potential mechanistic insights bear consideration in the design of an
experiment and selection of analytical tools. Whether or not the practitioners
understood the implications at the biochemical level, metabolism has always been
one of the most important levels at which pathology is understood, studied, and
diagnosed. Genetic diseases which disrupt molecular machinery were among the
earliest conditions studied by targeted metabolic profiling (Wilcken et al., 2003)
because of the direct links. The technique has since been applied to many
conditions, some only indirectly related to metabolism, including almost every entry
in the WHO’s top causes of human mortality: cardiovascular disease (Stamler et al.,
1993; Bonora et al., 1996; Roussel et al., 2007), Alzheimer’s (Barba et al., 2008;
Wood, 2012), cancer (Abate-Shen and Shen, 2009; Fan et al., 2009; Bathe et al.,
2011), diabetes (Bain et al., 2009; Lanza et al., 2010), and HIV (Pendyala et al.,
2007; Ghannoum et al., 2011). Even psychiatric and neurological conditions, whose
connections to systematic biochemistry are inherently difficult to quantify, “are
linked to disturbances in metabolic pathways related to neurotransmitter systems
(dopamine, serotonin, GABA and glutamate); fatty acids such as arachidonic acid-
cascade; oxidative stress and mitochondrial function.” (Quinones and Kaddurah-
Daouk, 2009)
19
As a class of diseases, cancer is probably the most metabolically studied condition
in the past three decades, because of its systematic effects, growth-related nature,
pathological complexity, population prevalence, and available funding (Carroll and
Kritchevsky, 1994; Abate-Shen and Shen, 2009; Chan et al., 2009; Bathe et al.,
2011; Mai et al., 2011). Cancerous cells undergo a variety of metabolic shifts,
depending on the origin, location, and genetic causes of the growth. The glycolytic
upregulation known as the Warburg effect is almost universal (Weljie and Jirik;
Vander Heiden et al., 2009), while other changes accompanying angiogenesis (Li,
2000), disregulation of apoptosis, nutrient scavenging, and oxidative stress vary
depending on cancer type. As a result, metabolomics has been applied with varying
degrees of success to the understanding and/or diagnosis of pancreatic (Bathe et al.,
2011), prostate (Abate-Shen and Shen, 2009), breast (Claudino et al., 2007;
Giskeødegård et al., 2010), oral (Tiziani et al., 2009), lung (Hilario et al., 2003),
kidney (Kind et al., 2007; Kim et al., 2009a), and brain cancers (Griffin and
Kauppinen, 2007) among others (Chung and Griffiths, 2008; Kim et al., 2008;
Spratlin et al., 2009).
1.3.3 Insulin resistance related disease
Type II Diabetes Mellitus, and the related phenomena of insulin resistance,
metabolic syndrome, and obesity, are among the most studied health conditions in
the field (Griffin and Nicholls, 2006; Sébédio et al., 2009). Ostensibly caused by a
developed resistance to insulin which causes disregulation of sugar levels in the
body, diabetes exhibits crippling pathology which is eventually fatal (Brownlee and
Cerami, 1981). Worse, occurrences of metabolic syndrome are growing
exponentially in North America. Increases are partially due to an aging population,
but the prevalence in children is increasing along with childhood obesity . After
onset of acute diabetes, treatment requires a life-long regimen of insulin to prevent
nervous, coronary, renal, optical, hematologic, and skeletal (Heath et al., 1980;
Laakso et al., 1992; Bouillon et al., 1995; Kelley et al., 2002; Wong et al., 2007;
Wang et al., 2011) failures, among others.
20
Figure 1.1: mTOR/AKT pathway diagram
The mammalian target of Rapamycin pathway, showing relationships
between PI3K, mTOR and AKT actors believed to be involved in diabetic
pathology. Figure from Wikimedia commons, created by Charles Betz and
released under Creative Commons License v3.0.
Mechanistically, acquired insulin resistance has a complicated etiology which has
impeded discovery of better treatments or cures. Changing insulin sensitivity in
muscle and fat cells causes chronic hyperglycaemia, but sugar levels are
reciprocally a causal factor. Known mechanisms include overloading mTOR or
PI3K effectors, but non-enzymatic chemical activities such as Schiff-base
glycosylation have also been identified as part of downstream diabetic pathology,
and postulated as part of causal mechanism (Brownlee and Cerami, 1981). Such a
process would elude any protein or expression based analysis, but could influence
amino acid and sugar metabolism in a noticeable way. Complicating any analysis is
the influence of genetic and environmental factors on susceptibility, with both
variable and interacting second-order sensitivities (Christian and Stewart, 2010).
21
Insulin related conditions have therefore been a worthwhile target for analysis with
metabolomics. Overall nutrition and dietary health, both inexorably linked to
metabolic syndrome, have since been studied in this manner (Zivkovic and German,
2009), along with the various causes of obesity (Griffin, 2006). Urine from
established diabetic cases has also been the focus of NMR analysis (Jankevics et al.,
2009; Lanza et al., 2010), but the extended and unpredictable development of
insulin resistance makes direct examination of its onset difficult. In place of
prohibitively large human cohorts, effective animal models for insulin resistance
allow for more stringent controls and an induced onset. The C57BL/6 mouse strain
offers genetic homogeneity, human analogue metabolism, and strongly inducible
insulin resistance using a high-fat diet, making it an ideal model for the study of
insulin related disease.
All told, metabolomics can and has been used to study a wide variety of diseases,
human health conditions, and industrial applications. Direct and indirect influences
on metabolism can both be detected, with almost any significant change in biology
being expressed in biofluid metabolome of a living organism. The field, however, is
not without limitations and failures.
1.4 How is metabolomics studied?
Notwithstanding the variety of definitions and implementations, most metabolomics
experiments follow a standard methodology which can be broken down into stages
based (to first approximation) on the physical location in which they’re conducted.
While the nomenclature defined here is not universal, the approach is derived from
the work of Douglas Kell and made popular by a number of academic labs.
Organizations such as the metabolomics society and NIH special interest groups
have established standard operating procedures to facilitate deployment and
consistency (Fiehn et al., 2007; Goodacre et al., 2007), but every stage provides
Outlier detection Multivariate/Univariate model construction
Model Optimizaton Model Validation
Interpretation Variable Selection and model reconstruction Variable importance
Variable identity confirmation Pathway Analysis
Hypothesis generation Figure 1.2: Overview of Metabolomics experiments.
1.4.1 Collection
Collection of samples is, nominally, the same as for any experiment in metabolism.
Samples are manipulated or selected from some larger population in such a way as
to focus on the results of some treatment of interest.
While solid state experiments have been performed, fluid samples make up the vast
majority of studies because of the simplicity in their handling and acquisition with
either NMR or MS instrumentation. Serum or urine are the most common
collection medium for vertebrate biology because of their relative ease in
acquisition, but tissue extracts and more exotic samples such as amniotic or
cerebrospinal fluid have been used (Dunne et al., 2005). Plant and bacterial extracts
have also yielded significant insights, with the latter (Bock, 1982) providing both
fingerprint (cellular lysate) and footprint (media content) analysis.
23
Regardless of fluid extraction method, the selection or manipulation of samples to
optimize the information content is critical (Jonsson et al., 2004, 2005; Broadhurst
and Kell, 2006). The complexity and variance inherent to biological systems
implies that each sample provides limited perspective, a problem which is
exacerbated by the large number of variables (metabolites) examined (Hastie et al.,
2009). In higher order systems, such as mammalian biology, many factors such as
environmental inconsistency, genetic inhomogenetiy, and experimental handling all
introduce confounding, undesireable, unrelated, or unexpected variance (“noise”)
which must be filtered out using replicates and experimental design (Dumas et al.,
2006; Dunn et al., 2011).
Information content is maximized when samples are selected or manipulated such
that the only sources of variation present are those of interest. Doing so is
exceptionally difficult in humans, because of extreme inconsistencies in diet,
exercise, genetics, and disease and health (Bijlsma et al., 2006; Scalbert et al., 2009)
Every factor affects metabolite levels either directly, as a result of uptake in
nutrients, indirectly via biological response, or both. Among the established
epidemiological methods for coping with human inconsistencies, several are directly
applicable to metabolomics:
Experimental control structures
When the selective application of treatment is not ethical, for example, structured
sample cohorts control variance by selecting samples from a larger pool in a
balanced way with respect to both primary attributes (e.g. disease state) and
unrelated variables like age, race, weight, etc. Similar samples are then arranged
into blocks which can be contrasted in order to identify and eliminate the effect of
those nuisance factors (Addelman, 1969). Further, blocking has application to
both hypothesis-generation or biomarker validation studies although the latter brings
with it other issues of sample size and logistics (Pepe et al., 2008).
24
Longitudinal studies use each individual as its own control, sampling biological
state from each both before and after one or more treatments of interest, or along
some time series. The differential of each sample in the time dimension then
becomes the scope of the experiment, with consistent change of a sample relative to
time zero rejecting a null hypothesis. Of note, longitudinal studies can only
compensate for any variance between individuals prior to the beginning of the
experiment. Differences in response between individuals will still appear as
inconsistencies in the final result.
Careful use of animal models of human biology are particularly powerful in
metabolomics. The ability to control or manipulate sample biology using diet,
medication, genetic alteration, etc. means smaller intra-class noise and higher inter-
class signal. Care must be taken when using models though, as the extreme
sensitivity of the method again makes it possible to detect many artifactual changes;
even strain differences have been shown to differ in metabolomic profile (Griffin,
2006).
Crossover studies are a form of longitudinal study, frequently employed in
mammalian contexts, which provide evidence for the causal nature of a change in
metabolism by switching the treatment status of some samples midway through the
course of the experiment. Typically, half of the control samples switch to undergo
treatment or induction of disease, while half of the “affected” samples are switched
to a control protocol. The three-point time-course controls for many factors and
shows more clearly how the manipulated variable is responsible for the change in
response.
Of all the factors which can be controlled in animal models, gut bacteria is an
extremely important one (Nicholls et al., 2003; Wikoff et al., 2009). Mice can be
raised in a sterile environment if necessary, which ostensibly eliminates
gastrointestinal microbial influences, but normal laboratory conditions quickly
induce bacterial effects. Whether such factors are significant to a disease varies, but
25
they have a clear impact on serum or urine metabolome regardless (Nicholls et al.,
2003). Given that, adjustments or controls for gut culture differences, such as
cohabitation acclimatization periods, standardized diets, and/or sterile handling, are
worth considering as both pros and cons of any animal model study (Bell et al.,
1991; Barnard et al., 2009; Wikoff et al., 2009).
1.4.2 Quantification
As with any experiment, translation of real-world biological samples into numerical
data requires appropriate instrumentation. Because of their ability to (A)
differentiate metabolites based single-atom chemical structure differences, and (B)
provide quantitative2 measurements of each metabolite’s presence in solution, two
instrumental paradigms are commonly applied:
Nuclear Magnetic Resonance (NMR) quantifies metabolites on the basis of their
component atoms’ spin state reaction to RF fields in a strong magnetic field (Canet,
1996). After processing a Fourier transform into frequency domain, NMR analysis
of a pure compound yields a quantized spectrum (Figure 2a) with peaks whose
frequency shift relative to a standard reflects the structural identity of an atom
therein.
2 The term quantitation or quantification is used in some metabolomics literature to refer to a combination of metabolite identification and concentration measurement with linear, homoscedastic accuracy. From an information theory perspective however, any measurement of physical properties, and their capture in a numerical matrix (as opposed to qualitative descriptors) is a quantification process. It is this numerical capture, regardless of any metabolite identification which precedes or follows it, which is referred to here as quantification: the second fundamental and necessary step in all metabolomics analysis.
26
Figure 1.3: Examples of NMR and MS spectra
The popularity of NMR stems from several factors. The fact that NMR probe
hardware never comes into contact with, or interacts chemically, samples improves
the reproducibility of the technique. Moreover, because the RF and magnetic fields
in an ideal NMR probe are homogeneous throughout the sample, every molecule in
the sample undergoes approximately the same process, and exhibits the same
response. The result is highly coherent signal, absent artifacts resulting from
(a)
(b)
27
interaction with the matrix or other molecules as a result of the instrument: the
response (in each peak) is linear with atomic concentration (e.g. proton
concentration in 1H NMR) and strictly additive in peak height, even across multiple
molecular species which overlap in the spectrum (Nicholson and Wilson, 1989;
Nicholson et al., 1995; Canet, 1996). In addition to yielding extremely consistent
data, these properties allow the use of known patterns to deconvolve a spectrum of
multiple compounds resulting from complex mixtures such as biofluids.
The quantum-mechanical details of NMR which yield these properties are usually
abstracted for the purposes of metabolomics, but remain important insofar as they
dictate instrumental allocation and properties of the resulting data: central to both is
the inverse-squared relationship between NMR sensitivity (signal-to-noise ratio) and
time spent scanning each sample (Canet, 1996). Because only atoms with non-zero
quantum spin are detectable by NMR, proton (1H) resonance spectra comprise the
vast majority of metabolomics experiments -- 1H’s high abundance in natural (non-
labelled) metabolites leads to high sensitivity and hence low required acquisition
times. Of the other elements available for NMR detection, 13C NMR has been
performed (Fan et al., 2009; Shaykhutdinov et al., 2009; Zamboni and Sauer, 2009),
but its low natural abundance leads to low sensitivity, so its utility is mostly
restricted to 2D 1H-13C experiments whose purpose is to identify unknown spectral
elements (Duggan et al., 2011c).
Specificity in NMR, in the form of accurate spectral-peak-to-metabolite assignment,
stems from the strength of the magnet used (Nicholson and Wilson, 1989). A well
maintained and shimmed 400 MHz magnet (that is, a magnet with a sufficiently
powerful field to cause 1H nuclear spins to precess at 400 MHz) can generate
sufficiently narrow spectral peaks to differentiate numerous compounds, but the 600
MHz magnet used for this work was much closer to the norm, and a necessary
baseline against which to demonstrate improvements (Nicholson et al., 1995;
Duggan et al., 2011c). With proper compensation for water signal and shimming to
28
optimize peak shape, concentrations as low as 1-2 uM were clearly discernible for
up to one hundred metabolites in practical circumstances.
NMR-based analysis was the primary analytical platform used in work for this
thesis. For comparison, three other technologies used for the quantification of
metabolomics samples are Mass Spectrometry (MS), and Fourier Transform
Infrared spectroscopy (FT-IR). Of the two, the former is far more common due to
its increased sensitivity, up to one hundred fold above that of NMR, and is usually
coupled to some form of chromatographic separation in the form of Gas
Chromatography (GC), Liquid Chromatography (LC), etc (Weckwerth and
Morgenthal, 2005). The GC- or LC- component of the hybridized instrument serves
to increase the accuracy and resolution of the ensuing MS instrumental readings
(Dunn et al., 2011). By segregating the contents of a sample on the basis of some
physical property, the chromatographic step allows MS to be applied in serial to
some smaller subset of the metabolites therein, while still ensuring coverage of the
entire molecular range.
Mass spectrometry itself operates by modifying the charge state of molecules in the
analyte, then separating and identifying the resulting ions on the basis of their mass
(a magnetic field exerting a constant force, and hence variable acceleration on them
within the detector). Like NMR, the result is a spectrum (Figure 2b) of peaks
separated by chemical identity and proportional in height to concentration. With
MS however, the relationship is not linear or additive, owing to frequent chemical
interactions between ions, the sample matrix, and/or the instrument (Pigini et al.,
2006; Bantscheff et al., 2007). Separation of metabolites on some orthogonal basis
prior to MS analysis notwithstanding, the result is data which is more sensitive but
less quantitatively reliable. Internal standards and better separation using more
precise detectors have been employed to eliminate such errors, which are not
necessarily as obvious as mis-assignment or changes in mean accuracy but can
include more nefarious artefacts such as heteroskedastic or inconsistent sensitivity;
the bane of statistical analysis.
29
Optical/vibrational spectroscopies such as FT-IR, Near Infrared (NIR) or Ramen
Spectroscopy have also been employed, with the significant advantage that IR-based
instrumentation is often orders of magnitude less expensive than either MS or NMR
based platforms. While this renders them attractive as an end-point application
(Dunn and Ellis, 2005) for some sample types, the absorbance of IR by water makes
the normal aqueous biofluid samples less amenable to full range (FT) IR analysis.
Limiting the spectrum to NIR analysis has yielded some improvement in the
analysis of aqueous samples when the experiment is focused on single compounds
with dominating concentrations (Lafrance et al., 2004). In general however, IR
based techniques indicate only the presence of gross structural features somewhere
in molecular structure, hence cannot differentiate between e.g. structural isomers or
identify low-concentration metabolites, which limits their use as a wide-perspective
platform (McGovern et al., 2002; Kaderbhai et al., 2003; Lafrance et al., 2004).
Regardless of the instrument used, sample preparation specific to the instrument is
performed prior to spectral analysis (Lu et al., 2008; Dunn et al., 2011). Steps
include removal of unrelated macromolecules such as proteins and lipids, chemical
derivitization to render metabolites volatile, and/or protection from bacterial decay.
The variations, based on experimental goals, are an important example of the many
sources of error and inconsistency inherent to the complex sample processing
pipeline required. Inconsistent sample preparation and handling can result in “batch
effects”, which differentiate samples based on acquisition date, instrument, analyst
or other irrelevant factor (Villas-Bôas et al., 2007; Alam et al., 2009; Zelena et al.,
2009; Draisma et al., 2010). Because of the wide scope and high specificity of the
methods used, subtle variations in batch processing are sometimes captured.
Metabolomics has shown to be at least as reproducible as other –omics methods
(Dumas et al., 2006), but the presence of batch effects is something which must be
considered, notwithstanding standard operating procedures (SOPs) and other
attempts to achieve consistent acquisition (Villas-Bôas et al., 2007; Alam et al.,
2009). Given the complexity of the biological systems in question, it would seem
30
unavoidable refining instrument sensitivity and sample processing invariably results
in other inconsistencies being more prominent.
As such, the character of measurements taken and the nature of errors are influential
on the resulting analysis. Neither NMR nor MS can be treated as a black box which
outputs absolute concentrations without regard to instrumental artefacts.
Spectral processing
Two important steps for dealing with procedural artefacts are variable normalization
and assignment (Broadhurst and Kell, 2006; Goodacre et al., 2007). Either could be
considered part of the modeling phase, but the focus on compensating for
differences between reality and quantified data suggests they are better considered
pre-modeling steps.
Normalization, collectively, refers here to an array of pre-processing techniques
designed to justify assumptions used in the statistical modeling phase, in particular
regarding the distribution of variance in metabolite concentrations (Craig et al.,
2006; Sysi-Aho et al., 2007; Torgrip et al., 2008; Warrack et al., 2009). Separate
from the scaling and mean centering performed as part of the statistical analysis,
whose goal is to make each variable numerically comparable; normalization
addresses the unavoidable differences in quantification of samples.
Differences in sample dilution are the most common quantification artefact, which
can result in scale differences across all measured metabolites (Craig et al., 2006;
Torgrip et al., 2008). To compensate, the measured values for each sample are
usually divided by some value based on an estimation of that sample’s “dilution
factor” (DF) (Craig et al., 2006; Dieterle et al., 2006). An obvious but relatively
unavoidable caveat of this approach is that differences in biofluid concentration can
occur in vivo, especially in fluids such as urine.
31
Nevertheless, various DF estimation algorithms may result in better final results.
Historically, the total concentration of all metabolites in a sample was used as a
simple estimation of its DF, with some manual curation to mitigate the effects of
dynamic range (such as removing glucose, lactate, or other high-concentration
metabolites) (Weljie et al., 2006). More recently, other algorithms have been
proposed that obviate the need for manual interference or awkward threshold
effects. Normalization relative to an average spectrum, rank transforms, internal
standards, or scaling the effect of each variable on DF estimation by its variance
across spectra have all been applied in the past (Dieterle et al., 2006; Sysi-Aho et al.,
2007); other techniques drawn from e.g. microarray analysis have been adapted and
may be prove valuable (Li and Wong, 2001; Bolstad et al., 2003; Shurubor et al.,
2005).
Other sources of error which NMR quantification can introduce include baseline
distortions, peak shifts, and metastable peaks. Spectrum-wide phenomena are
handled by subtracting a spline fit to the baseline (Weljie et al., 2006). Metastable
peaks, like urea, and peak shifts caused by salts/pH inconsistencies in the sample,
are part of the larger problem of assigning chemical identities to measured variables:
As a consequence of analog-to-digital processing, after Fourier transformation NMR
spectral peaks are each captured in an unknown number of variables (Nicholson and
Wilson, 1989). The same phenomenon occurs in MS data, regardless of
chromatographic separation, but for different reasons. As a result, the exact number
of data points per peak varies across the spectrum based on peak width, spectral
resolution, and alignment between the two. Peak overlaps, peak shifts, and
occasionally variable peak widths exacerbate the problem of aggregating and
attributing data points to specific resonance peaks, which is necessary before
deconvolution of peaks (and multiplets) into chemical identities is possible (Åberg
et al., 2009). Unlike normalization, some assignment tasks can be performed after
statistical modeling (i.e. only for variables identified as significant), but others (such
as dealing with peak shifts) cannot so it remains primarily an issue of accurate
32
spectrum-wide processing. Proposed approaches to simplify assignment include
binning (Davis et al., 2007; Anderson et al., 2008, 2010), targeted profiling (Weljie
et al., 2006), and correlation based analysis (Cloarec et al., 2005; Cho et al., 2008).
By summing all of the points within fixed regions of the chemical shift dimension
(either evenly spaced bins, or buckets believed to segregate metabolites based on a
priori knowledge) binning reduces the number of variables which facilitates both
modeling and assignment (Beckwith-Hall et al., 1998; Davis et al., 2007; Anderson
et al., 2010). Drawbacks include boundary effects, which introduce errors when one
or more windows incompletely cover a spectral peak. The unpredictable shift in
peaks necessitates human curation, or some kind of automated peak detection
algorithm. While peak detection is relatively simple at high signal-to-noise ratios, it
becomes increasingly difficult as concentrations (and peaks) get smaller; peak
boundaries become intractable and assignment quality suffers.
More sophisticated approaches to assigning spectral elements are based on the
correlation between peaks stemming from the same molecule, an important
implication of the aforementioned linear relationship between peak height and
molecular concentration. Targeted Profiling is a manual assignment strategy which
will be discussed in depth as it pertains to individual applications (Weljie et al.,
2006). Developed by Chenomx Inc, it depends heavily on software libraries to
inform the user of known correspondences between spectral peaks and particular
metabolites. The primary advantage of such an approach is that the additive
property of NMR spectra is leveraged to deconvolve overlapping peaks, recovering
more metabolite-specific concentration information. The result is a much simplified
dataset with theoretically minimal loss of information.
Other correlation-based assignment strategies include statistical total correlation
spectroscopy (STOCSY) (Cloarec et al., 2005), which simplifies spectra by filtering
the spectral autocorrelation matrix before applying automated peak detection or
presenting the same data for manual assignment.
33
1.4.3 Modeling
Having collected samples and quantified the contents of each into a vector (usually
proportional to concentration values, with some error characteristic of the
instrument used for quantitation), numerical analysis techniques are required to
draw meaningful inferences about samples, and from there the population.
Regardless which approach is used to analyse sample variations, certain terminology
and steps are consistent. A training set of samples is selected from among the
samples collected, and used to train a model of the variation seen, with the hope that
the distribution of each variable (metabolite concentrations) in the training set will
reflect its overall distribution in the greater populations (Trygg et al., 2007; Hastie et
al., 2009). The structure of the model is optimized using a testing dataset of some
kind, and then its accuracy must be confirmed on validation data before its utility
can be ascertained.
Training a metabolomics models occurs in a number of ways, but the most general
schema differentiates between supervised and unsupervised techniques (Goodacre et
al., 2004; Hastie et al., 2009) – semi-supervised learning techniques have seen little
application, if any, in metabolomics. Unsupervised analysis looks to model the
most prevalent sources of variation within the population, while supervised analysis
looks for specific patterns. Patterns of interest may be those related to a response
measure such as change in bodyweight, or a manipulated property of each sample
such as dosage. Other experiments look for the difference between discrete sample
classes, in which case the experiment is referred to as a discriminant or
classification problem (Hastie et al., 2009). In all cases, a supervisory variable of
some kind is used to inform the model training process what patterns to prioritize.
Because successful quantification often results in dozens, if not hundreds or
thousands of variables, multivariate considerations are significant to any
metabolomics study (Broadhurst and Kell, 2006). Application of univariate
techniques with adjustment for multiple hypothesis testing can avoid some
34
problems, but quantitation of NMR spectra provides no guarantee of linear
independence between metabolite concentrations (in fact, metabolites are often co-
linear for both biological and instrumental reasons). Thus, consideration of
metabolites on an individual basis discards important context for each, and falls prey
to Bellman’s curse of dimensionality (Broadhurst and Kell, 2006).
A number of analytical approaches have been applied to metabolomics data in the
past, drawn from various machine learning, signal processing, and chemometric
disciplines (Holmes and Antti, 2002; Goodacre et al., 2004). The related techniques
of probabilistic neural networks and Support Vector Machines have been applied
separately to the problems of assignment and biological modeling (Truong et al.,
2004; Xu et al., 2006; Mahadevan et al., 2008; Giskeødegård et al., 2010), however
they suffer from the problem that the resulting model structure is difficult to
interpret by mapping back to the original biological context. As stated by Madsen
et al.:
For non-linear methods it is often impossible to interpret the model
biologically, since statistical significance of model coefficient size
and its direction may not be directly related to the importance of the
corresponding variables for the model building. In our experience,
we have yet to discover an example where non-linear methods
indeed provide superior classification results (e.g. in an independent
follow-up study and not by cross-validation testing) compared to
linear methods to justify this significant disadvantage in biological
interpretation (Madsen et al., 2010)
Tree based hierarchical clustering, rule based inference, soft independent modelling
of class analogies, ANOVA-simultaneous component analysis (Smilde et al., 2005),
and genetic algorithms (Gilbert et al., 1997; Hageman et al., 2008; Koza et al.,
2008) have also been applied to both problems, but suffer from overfitting when the
35
number of samples is small – as previously mentioned, metabolomics
instrumentation can usually quantify several hundred metabolites or more, so the
number of samples present is usually much smaller than the number of variables.
Exception is that of agricultural, bacterial, or plant metabolomics, where replicates
are easier to come by and population variance smaller (Catchpole et al., 2005;
Fukusaki and Kobayashi, 2005; Tikunov et al., 2005). Markov-chain based
modeling, especially as a fitting tool for Bayesian estimators, have been applied to
the spectral processing correspondence problem (Kim et al., 2006; Rubtsov and
Griffin, 2007), but very little work has been done on its application to metabolic
interpretation; the best examples build on top of other techniques and as yet still
suffer from interpretation issues (Gavai, 2009; Mcateer et al., 2009; Franken et al.,
2012; Krumsiek et al., 2012).
Projection-based methods are a family of analytical techniques which handle both
smaller sample sizes and numerous co-linear variables well (Hastie et al., 2009). As
a result, they been the most popular and successful tool employed to analyse
metabolomics (Trygg et al., 2007). Projection techniques identify weighted linear
combinations of variables (hyperplanes) of maximal covariation within the Hilbert
space defined by the measured concentrations. Doing so intrinsically identifies so-
called latent variables: underlying combinations of variables which drive the
information content of the experiment. In projection based methods, each latent
variable is attributed to a perpendicular eigenvector of minimal variation, and hence
maximal information content. Among other things, this helps to solve the problem
of highly co-linear variables. The result is the identification of patterns of consistent
co-variation between metabolites which simplifies the difference between samples,
and aids both identification and application of the model.
Two projection-based techniques, principle component analysis (PCA) and
projection to latent structures (PLS), constitute the vast majority of the analysis
done in metabolomics, especially as it applies to human health science (Trygg et al.,
2007; Madsen et al., 2010). PCA and PLS correspond approximately to the
36
unsupervised and supervised versions of the same algorithm, and share a number of
facets in both data preparation and results interpretation:
Analysis with projection-based models begins by concatenation of the samples'
concentration vectors into a matrix, at which point the importance of consistent
treatment of samples in earlier steps becomes obvious. The coherence implied by
samples’ row-wise unification is only true to the extent that quantified
measurements are either accurate to a biological comparison or can be rendered
comparable using some mathematical equilibration. In either case, exploratory data
analysis or strong assumptions must be used to justify the use of projection
techniques because of their dependence on (at least) a center-weighted distribution
for in each variable (Hastie et al., 2009).
Before the resulting data matrix can be used to train a model of the sample variance,
a number of “column-wise” adjustments are required. Mean centering and scaling
to comparable measured second moment (ie: variance, either unit variance or, in
some cases, unit root variance “Pareto” scaling) are almost universal because they
render variables comparable within the framework of a normal probability
distribution. Where necessary, additional transforms such as logarithm or cube root
transforms can further coerce the data into an approximately normal distribution
(Trygg et al., 2007).
One final point of note is an inversion of the nomenclature with regard to dependant
and independent variables in a PCA or PLS model. The nature of metabolomics
experiments is that sample class or treatment is manipulated and the effects on
metabolite levels measured, making metabolites dependant variables. In both PLS
and PCA analysis however, the metabolite concentrations are usually labelled X
variables, and the sample treatment is the response (Y) variable when used
(Umetrics, 2006).
Principle Component Analysis (PCA)
37
PCA is a relatively simple form of factor analysis which uses no supervisory
variables: it identifies patterns in the measured metabolite data without regard to
descriptive sample characteristics (Wold et al., 1987). Each component is an
orthogonal hyperplane, also referred to as a dimension, which represents a dominant
or recurring pattern of correlated variables in the X-matrix. The structured variation
in the samples is then estimated using the normal vectors as an orthogonal basis of
“loadings”; a set of correlated changes orthogonal from any other component.
Figure 1.4: Example loading plots
1D and 2D plots showing the loadings, for one and two components,
representing patterns of correlated change seen in the samples.
Using the orthogonal loadings as basis vectors, the difference of each spectrum from
the mean can be reconstructed uniquely using a weighted combination.
Reconstruction weights comprise scalar “scores” in each dimension, with separation
in each score dimension representing modeled differences between samples.
-0.45
-0.40
-0.35
-0.30
-0.25
-0.20
-0.15
-0.10
-0.05
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
Allantoin
Trimethyla
Creatine
Citrate
3-M
ethylxa
Mannose
Hippurate
Fumarate
Alanine
Phenylacet
Tryptophan
Glutamine
Threonine
N,N-Dim
eth
Lysine
Asparagine
Glycolate
Uracil
2-Hydroxyi
Carnitine
Guanidoace
Isocitrate
Caprylate
N-Carbamoy
Ethanolami
N-Isovaler
Xylose
4-Hydroxyp
Urea
Ascorbate
Taurine
w*c[1]P
Var ID (Primary)SIMCA-P+ 12.0.1 - 2012-09-29 21:32:35 (UTC-7)
These metabolites increase with high-fat feeding compared with controls
(chow fed). Values are stated as mean (uM) ± s.e., n = 10/dietary treatment,
p values for each individual metabolite are listed.
61
Metabolite Control High fat P Value
Glycine 97 ± 16 42 ± 16 0.002
Lysine 88 ± 12 60 ± 5 0.005
Suberate 8 ± 2 4 ± 1 0.005
Acetate 148 ± 29 88 ± 14 0.014
Leucine 58 ± 8 41 ± 5 0.019
Valine 73 ± 11 51 ± 6 0.023
TMAO 13 ± 3 7 ± 3 0.038
Hippurate 79 ± 21 48 ± 11 0.042
Arginine 44 ± 15 22 ± 3 0.049
Phenylalanine 30 ± 4 23 ± 2 0.064
Lactate 3390 ± 483 2546 ± 532 0.103
Isobutyrate 9 ± 1 7 ± 1 0.114
Succinate 33 ± 22 8 ± 3 0.117
Tyrosine 29 ± 4 24 ± 3 0.149
Isoleucine 38 ± 10 26 ± 7 0.166
Uridine 6 ± 2 5 ± 1 0.202
Allantoin 39 ± 7 31 ± 7 0.250
Tryptophan 8 ± 2 6 ± 1 0.290
Glutamate 43 ± 24 29 ± 5 0.334
Fumarate 4 ± 1 3 ± 1 0.373
Formate 19 ± 6 15 ± 3 0.374
Pyruvate 72 ± 21 60 ± 12 0.499
Benzoate 139 ± 49 115 ± 34 0.570
Choline 13 ± 3 11 ± 3 0.625
Proline 65 ± 7 64 ± 5 0.906
Table 2.2 Negative contributors to multivariate analysis.
These metabolites decrease with high-fat feeding compared with controls
(chow fed). Values are stated as mean (uM) ± s.e., n = 10/dietary treatment,
p values for each individual metabolite are listed.
62
Model generation
The importance of the multivariate model is the ability to identify the set of
metabolites most directly responsive to the high-fat diet. In OPLS models, variation
in the x component represents changes in metabolite concentration, while y is
correlated to variation in either diet (model 1) or body mass (model 2). Other
sources of variation can be isolated in subsequent orthogonal components so that
their impact is not interpreted as part of the diet response but rather alternate sources
of variation (batch, exercise, instrument instability, etc.).
Model 1 – diet
Chow and high-fat diets resulted in clearly distinguishable spectra. Acetate, formate
and pyruvate were elevated, while uridine, creatine, tyrosine and the ketone body 3-
hydroxybutyrate were simultaneously depressed. The variable importance of each
metabolite in model 1 (Figure 2.2) is shown in Figure 2.3. The most significant
metabolites included amino acids (Lys, Gly, Leu, Phe and Val) and energy
metabolites (citrate, acetate, glycerol, suberate and lactate). Model 1 (Figure 2.2)
had one sample (of 20) outside the 95% confidence interval, which is within
acceptable tolerances.
An orthogonal partial least squares discriminant analysis scores plot shows the
separation between samples along each of the two model components in model 1.
The first (horizontal) component is the variation associated by the model with
metabolite-based interclass differences. The second, vertical component represents
subject variation unrelated to diet. n = 10 mice/dietary treatment
63
Figure 2.2: OPLS-DA Scores Plot
An orthogonal partial least squares discriminant analysis scores plot shows the separation between samples along each of the two model components in model 1. The first (horizontal) component is the variation associated by the model with metabolite-based interclass differences. The second, vertical component represents subject variation unrelated to diet. n = 10 mice/dietary treatment
Figure 2.3: (VIP) from model 1
Variable Importance Plot shows the relative contribution that each metabolite makes to the difference between classes in model 1. Metabolites with VIP scores above 1.0 are considered to be strong contributors; anything below 0.5 is considered non-significant. Only compounds with an error smaller than their magnitude (95% confidence in their significance) are shown.
64
Model 2 – mass
A second OPLS model, created from the same metabolite data, used the mass (y
variable) for regression (figure 4, r2 = 0.89, q2 = 0.52). The purpose of this second
model was to compare the metabolic changes associated with the cause (diet) with
those associated with the effect (change in mass). Validation was performed on this
model as well as scores, loadings and cross-validation statistics. The mass
measurements of the chow-fed animals were clustered in a small range
(35.4 ± 4.3 g). However, the high fat–fed animals were more disperse (42.8 ± 9.1 g),
showing some overlap with the chow-fed mass range.
Figure 2.4 SUS plot comparing loadings for model 1 vs 2
Co-efficients on the diagonal react similarly in both models, whereas off-
axis metabolites show a differential change. Those in the top-right corner are
elevated in high-fat animals and high-mass animals and vice versa. Citrate,
which falls to the right of the x axis, shows elevated concentrations in fat-fed
animals, but does not change in concert with body mass. Conversely,
Alanine and benzoate concentrations are higher in animals with elevated
body mass, but do not respond directly to diet.
65
Upon inspection, the high-fat animals showed a bimodal distribution and as a result,
the body mass model (model 2) could not separate animals into diet classes. All
animals with higher body mass were among those fed a high-fat diet (80%);
however, not all animals on a high-fat diet exhibited a significant increase in body
mass (20%). It was, however, a good cross-validation predictor of body mass based
on metabolic profile. Model 2 showed substantially different compounds of
importance compared with model 1. Of possible interest was the substantial drop in
the significance of citrate and suberate. In contrast, the significance of alanine, 2-
oxoglutarate, benzoate and formic acid all increased in this model.
Internal validation
Model 1 cross-validation accuracy (YPredCV) was 18/20 or 90%. If Hotelling’s
95% confidence interval was used to exclude outliers from this prediction, the
accuracy improved but the sample size shrinks. Overall, the sample size is too small
for the model to serve as a classifier (predictor for third-party samples), but is
sufficiently robust for biological interpretation.
External validation
Because cross validation cannot detect sampling bias, particularly in small sample
sets, additional EV on samples was performed. Seven of 10 EV samples fell within
the model’s expected parameter range. The body mass of the EV samples predicted
using model 2 correlated reasonably well with the experimental mass of the animals,
with a correlation co-efficient of 0.28. It is important to restate that these predictions
are not intended to describe the power of the model as a clinical or externally
predictive tool. Rather, they are intended to verify that the model is not overfit to the
small number of samples used. Given the limited sample size to build the models
(n = 10/dietary treatment), a 70% membership is reasonable.
66
2.5 Discussion
The C57BL/6J mouse is a commonly employed experimental model of metabolic
disease as it readily develops obesity and insulin resistance when fed a high-fat diet
(Leiter, 1993; Raab et al., 2005; Alevizos et al., 2007; Clee and Attie, 2007). The
model is advantageous over genetically induced mouse models (e.g. db/db and
ob/ob) as obesity occurs gradually and is environmentally induced. Specifically,
high-fat feeding results in a decline in whole body glucose disposal, increased fatty
acid utilization, impaired glucose tolerance and cardiac dysfunction (Fueger et al.,
2004a; Park et al., 2005; Shearer et al., 2005; Wu et al., 2006; Ayala et al., 2007;
Rao et al., 2007). Here, we present a first step in validating 1H-NMR-based
metabolic profiling techniques to detect dietary-induced insulin resistance in this
model.
Results showed NMR-based spectroscopy to be a predictor of dietary-induced
insulin resistance in an experimental model of limited sample size. EV showed that
both diet and body mass could be predicted with reasonable success. The variable
importance plot, a measure of which metabolites significantly discriminate chow
and high fat–fed animals, shows lysine, glycine, citrate, leucine, suberate and
acetate to be altered. These metabolites are all involved in energy metabolism and
may be representative of the perturbations taking place with insulin resistance.
Of interest, results show the amino acids leucine, glycine and lysine to be depressed
in high fat–fed animals compared with chow-fed animals. There has been renewed
interest in the role of leucine in dietary-induced insulin resistance (Layman and
Walker, 2006; Bain et al., 2009; Newgard et al., 2009). This amino acid interacts
with insulin signalling through the rapamycin (mTOR) pathway and is important in
protein synthesis and substrate selection. Analogous to the present findings, plasma
leucine has been shown to be depressed with high-fat feeding in the Sprague–
Dawley rat with a 22% reduction in levels following 4 weeks of dietary
manipulation(Calles-Escandon et al., 1984). In humans, plasma leucine levels are
67
depressed with type 2 diabetes compared with healthy volunteers but restored with
6 weeks of rosiglitazone treatment (Van Doorn et al., 2007). Of note, these results
are for plasma and not for urine that typically shows elevated branched-chain amino
acid levels with the disease (Van Doorn et al., 2007).
Other altered metabolites include glycine and citrate that were elevated with high-fat
feeding compared with controls. Glycine is a key metabolite in nucleic acid
synthesis and is known to fluctuate with changes in energy status. Starvation results
in large increases in glycine, signalling energy transition and energy conservation
(Adibi, 1968). Likewise, citrate is produced from the Trichloroacetic Acid (TCA)
cycle and has been previously shown to be increased in alloxan diabetic rats as well
as ketoacidotic humans with type 1 diabetes (DeVilliers et al., 1966). Endogenously
produced from fatty acid and glucose metabolism, concentrations are known to be
sensitive to insulin and glucose levels (Piloquet et al., 2003). Citrate becomes
elevated with high-fat feeding as a consequence of hyperglycaemia, insulin
resistance, a heavy reliance on fatty acid utilization and a decreased liver clearance
(Sjostrom, 1937). Previous reports show this metabolite to be increased with
hyperglycaemia and lowered upon insulin administration in both experimental
animal models and humans (Penttila and Pollanen; Natelson et al., 1948; Pincus et
al., 1948; Todorow and Dikow, 1960). Despite these findings, the use and
applicability of citrate as a disease marker in metabolomic studies are controversial
and should be interpreted with caution (Robertson, 2005).
Correlations between the degree of insulin resistance induced by high-fat feeding
and metabolomic profile could not be performed in the present study as insulin
resistance and metabolite profiling were performed on separate subsets of animals
because of limited serum volume. However, results show metabolite profiling to be
a predictor of body mass. This correlation may be weaker than expected as only
80% of the animals on high-fat diet became obese, the other 20% maintained body
masses in the range of chow-fed animals (>38 g). Such variation in dietary-induced
obesity in inbred rodent strains has been previously reported and is thought to arise
68
from subtle polygenic differences between animals (Schemmel et al., 1970; Levin et
al., 1987, 1997; West et al., 1992). Additionally, the present study examined
metabolites from a single serum sample. Given this, we could not determine which
tissues were primarily responsible for the observed metabolomic alterations with
high-fat feeding.
Examination of the literature shows the use of 1H-NMR spectroscopy to detect
complex disease states in humans to be highly speculative at best (Kirschenlohr et
al., 2006; Roussel et al., 2007). Mixed results are, in part, because of the complex
aetiology of these conditions and the heterogeneous nature of the populations
sampled. To date, the majority of studies have examined freezer samples of
sera/plasma collected on a large number of patients and show poor assessment of
disease in metabolomic profiles. A more appropriate strategy is to examine samples
collected under more defined conditions. An excellent example of this can be found
in the work of Salek et al. (Salek et al., 2007) who compared spectral profiles in
urine from humans, rats and mice with type 2 diabetes. In this study, metabolomic
profiles of unmedicated human subjects, the db/db mouse and fa/fa rat (genetically
induced models of diabetes caused by leptin receptor mutation) were analysed.
Changes in metabolites involved in nucleotide, methylamine metabolism and TCA
cycle intermediates were common between the animal and the human models. There
were a greater number of metabolites altered in the animal vs. human models, an
expected finding considering that these models exhibit extreme type 2 diabetic
phenotypes. Likewise, these models had greater changes in the number, severity and
type of metabolites with diabetes compared with the present study that employs a
milder, dietary-induced model of the disease. Overall, the study showed that control
and disease samples from each model could be clearly distinguished, data that may
eventually provide novel biomarkers for tracking type 2 diabetes in urine.
Metabolomic technologies encompassing 1H-NMR allow the examination of a large
number of metabolites from small volumes of serum. To the authors’ knowledge,
this is the first examination of mouse serum from chow and high fat–fed C57BL/6J
69
mice. Of note, this predictive ability was generated on a small subset of
experimental animals as they most closely reflect numbers used in common
laboratory experiments. Animals in the present study were fasted, negating any
direct effect of the diets on metabolomic profiles. Limitations of this study include
the measurement of whole serum and the presence of anaesthesia. The specific
effects of anaesthesia on metabolite profiles are not known. However, given the
volume of blood collected from each animal (1 ml), this was unavoidable.
In conclusion, results show NMR-based spectroscopy to be a predictor of dietary-
induced insulin resistance and body mass in mice. Such results are a first step in
validating the technique for further experimental use. It is important to note that we
are not suggesting that metabolomic profiling be used as a surrogate for the
detection of insulin resistance or insulin clamp studies. Indeed, future work will
need to establish the relationships between the severity of insulin resistance and
metabolomic profile.
70
Chapter 3 - Differentiating short- and long-term effects of diet in the obese
mouse using 1H-nuclear magnetic resonance metabolomics
G. E. Duggan, D. S. Hittel2, C. C. Hughey3, A. Weljie1, H. J. Vogel1, J. Shearer
Published: 2011-07-26 in Diabetes, Obestity and Metabolism
3.1 Abstract
This study determined whether targeted metabolomic profiling of serum, using 1H
nuclear magnetic resonance, could be employed to distinguish the effects of obesity
from those of diet in mice. Following weaning, littermates were randomly divided
into two diet groups: chow and high fat. After 12 weeks of dietary manipulation, fat-
fed animals were obese and hyperglycaemic. Mice from each treatment either
maintained their current diet or switched to the opposite diet for a final week.
Differences in metabolite levels were determined using orthogonal projection to
latent structures and cross-validated discriminant analysis. The short- and long-term
effects of each diet could be clearly distinguished. Short-term diet effects are the
major contributor to the metabolic profile, underscoring the need for controls
beyond the standard fast before serum collection. This work shows the importance
of dietary controls when attempting to isolate obesity-related changes and highlights
the ability of metabolomics to identify subtle changes when experiments are
properly structured.
3.2 Introduction
Obesity results from excess caloric consumption and a chronic mismatch between
energy consumption and expenditure. These conditions induce systemic changes in
metabolism, altering a number of biochemical pathways. While the study of
individual metabolites has been a staple of nutrition research, technological
advancements now allow for the simultaneous measurement of numerous
metabolites in small sample volumes. These ‘metabolomic’ studies are seeing an
71
expanded use as a tool for the study of obesity (Kim et al., 2009b). The technique is
relatively inexpensive, rapid, highly sensitive and provides an abundance of data. To
date, the technology has been proposed to monitor disease progression, identify
biomarkers and generate new therapeutic targets for obesity. Despite these perceived
opportunities, significant challenges remain.
One such caveat when applying metabolomics to the study of obesity is
distinguishing metabolites that are altered as a consequence of diet versus those
changing as a result of obesity (Zivkovic and German, 2009). Excess caloric intake,
independent of obesity, results in a number of metabolic changes. Overfeeding
(+50%) in healthy males for 5 days increases hepatic glucose production, insulin
secretion and alters incretin response (MORGAN and LAZAROW, 1962). All these
physiological changes would be expected to alter an individual's metabolic profile.
Given this, we sought to employ a factorial study design to distinguish the specific
effects of diet from those of obesity. We test both the sensitivity and specificity of
the technique by employing a ‘diet switch’ in the C57BL/6J mouse.
3.3 Methods
Procedures were approved by the University of Calgary Animal Care and Use
Committee and abide by the Canadian Association for Laboratory Animal Science
guidelines for animal experimentation. Animals were maintained in a humidity-
controlled room with a 12-h light : dark cycle. Following weaning (3 weeks), male
C57BL/6J littermates were randomly segregated into two groups and maintained in
microisolator cages for 1 week. Following this acclimation period, animals received
either chow (C) or high-fat diets (H; 60%) for 12 weeks (58M1, 58R3—TestDiet;
Purina, Richmond, IN, USA) (n = 20–24). At the end of 12 weeks, half of the chow-
fed animals (n = 10–12 per class) were switched to a diet of high fat (C-H) while the
remainder continued to receive chow feed for a final week (C-C). At the same time,
one half of the obese animals previously fed a high-fat diet were switched to the
chow diet (H-C). Animals remaining on a high-fat diet for the final week were
72
labelled H-H. Animal weight was measured every week during initial and final
dietary treatments (figure S1, Table S1).
On the day of the experiment, animals were fasted for 6 h prior to being weighed
and euthanized (pentobarbital). Whole blood (~1 ml) was obtained by immediate
cardiac puncture to minimize anaesthetic effect on metabolism. Blood was placed
on ice and allowed to clot for 30 min. Samples were then centrifuged for 10 min at
3000 rpm and sera collected prior to storage at −80 °C. Blood glucose was assessed
using One Touch (Lifescan, Bunaby, BC, USA). Non-esterified fatty acids were
measured spectrophotometrically (Wako NEFA C kit; Wako Chemicals, Richmond,
VA, USA). Immunoreactive insulin was assayed with a double-antibody method
(MORGAN and LAZAROW, 1962). Following blood collection, a small section of
the left liver lobe was isolated, cleaned and rapidly frozen in liquid nitrogen prior to
triglyceride analysis. Liver triglycerides were determined by a standard kit (Point
Scientific, Canton, MI, USA). Abundant data on the chow- and high–fat-fed
C57BL/6J mouse have been previously published (Fueger et al., 2007; Shearer et
al., 2008). Serum metabolites were prepared as previously described (Appendix S1)
(Shearer et al., 2008). Briefly, one-dimensional nuclear overhauser effect
spectroscopy spectra were acquired using an automated Case sample changer on a
600-MHz Bruker Ultrashield Nuclear Magnetic Resonance (NMR) spectrometer
(Bruker Biospin, Karlsruhe, Germany). Processed spectra were imported into
Chenomx 4.6 for quantification (Chenomx Inc, Edmonton, Alberta, Canada). Urea,
which is often excluded from spectral profiling, was included notwithstanding its
semiquantitative nature but caution is advised in its interpretation. For measures of
body mass, blood glucose and insulin, a two-way analysis of variance (anova) was
performed to detect statistical differences (p < 0.05). Differences within the anova
were determined using Tukey's post hoc test. All data are reported as means ± s.e.
For metabolomics data, all concentration measurements were exported into simca-p
(UMetrics, Umeå, Sweden) for multivariate analysis. The simca-p software suite
was used to construct an orthogonal projection to latent structures discriminant
73
analysis (OPLS-DA) model of both traditional and diet-switched samples. Here, the
relationships between dietary treatment and NMR spectra were assessed.
3.4 Results
Metabolomics is highly sensitive and broad in scope, but must be coupled with
proper experimental design to ensure specificity. This is especially true when
quantifying complex and subtle biochemical effects such as those occurring in
obesity. In this study, we set out to identify whether metabolomics could
differentiate the short- and long-term effects of chow and high-fat diets. Employing
OPLS-DA, we show clear separation between classes (Figure 3.1A). The major
finding of this work shows that the final week's diet was the predominant
contributor to the metabolic profile (Figure 3.1B). When variation in weight gains
was considered by comparing outliers to distribution of body weights within the
class, the association became even stronger.
The study design allowed us to distinguish the specific effects of diet from those of
obesity in the mouse. The fat-fed C57BL/6J mouse is an established method for
studying obesity (Park et al., 2005). Data on body mass, plasma values and liver
triglycerides are shown in Table 3.1. Exposure to a fat-enriched diet (H-H) for an
extended period results in long-term obesity and hyperglycaemia. Switching from a
high-fat diet to a chow diet (H-C) had no impact on obesity, but normalized plasma
glucose levels. Unexpectedly, the diet switch protocol (H-C), although brief in
duration, normalized liver triglyceride levels. Previous studies from our laboratory
have shown a 50% reduction in whole-body glucose disposal in the high-fat-fed
C57BL/6J mouse using a hyperinsulinaemic–euglycaemic clamp as well as
dysregulation of a number of energy- and signalling-related pathways (Shearer et
al., 2008).
74
Figure 3.1 OPLS
Results
(A) Orthogonal projection to latent structures
discriminant analysis (OPLS-DA) scores
showing separation
between C-C and H-H samples, the two groups employed in the majority of obesity studies. (B) Predicted scores for diet-switched samples. Comparison with (A) shows that C-H samples fall into the same range as H-H samples while H-C samples were more similar to C-C samples. This shows how the
final week diet masks the initial obesogenic diet of interest. Upon reinspection, the C-H outliers were found to have gained a significantly higher weight during the initial dietary treatment which may explain their differentiated metabolism. (C) Shared and unique structure plot comparing individual metabolite changes in the simple and diet-switched studies. OPLS coefficients are shown on the x-axis for animals remaining on the same diet throughout (H-H vs. C-C, no diet switch). The y-axis shows a second model comparing (diet switch controlled) H-C animals to the same C-C baseline. Significant metabolites (as determined by a t-test on loadings in each model) are highlighted in black. Metabolites found on the diagonal are responsive to obesity (valine, leucine and glutamine), whereas off-axis metabolites (pyruvate and alanine) change as a result of diet. Urea (marked with an asterisk) should be regarded as a semiquantitative measurement because of peak suppression in the nuclear magnetic resonance spectrum; nevertheless its biochemical significance merits some consideration, albeit with caution.
Body mass, plasma glucose, insulin, non-esterified fatty acids (NEFAs) and liver triglyceride levels are shown for C57BL/J mice. Different superscripts indicate significantly different classes within each row (p < 0.05). Some indicators vary with initial 12-week diet, whereas others align to the final week's diet. Data represent means ± s.e., n = 9–12 animals per treatment.
Comparison of all four treatments in this study allowed us to unmask the short-term
changes of diet from those specifically related to obesity. Individual metabolites
changing in response to the diet switch are shown in Figure 3.1C. Metabolites found
near the diagonal (e.g. valine, leucine, isoleucine, glutamine and glutamate)
responded similarly regardless of the treatment in the final week. These compounds
show bona fide changes associated with obesity. Numerous laboratories, including
our own, have previously identified these metabolites to be altered with
obesity(Shearer et al., 2008; Newgard et al., 2009; Connor et al., 2010), but the
results are often confounded with changes in other metabolites. Although the
metabolomics platforms utilized in these studies are sensitive, they often lack
dietary controls resulting in low specificity.
Examination of off-axis metabolites in Figure 3.1C showed that several compounds
responded differently when the final week diet returned to a normal fat content (H-
76
C). These transient changes result from the diet and not obesity per se. For example,
glucose levels were lowered but succinate was elevated with H-C. Metabolites
involved in short-term energy regulation (e.g. creatine, ornithine and taurine) were
also clustered. All these metabolites are influenced by the animals' energy intake
and general health, but revert to normal levels in the H-C animals furthering
suspicion in their biological relevance to obesity. Pyruvate, alanine and lactate
(lower right quadrant in Figure 3.1C) were elevated in H-H animals (and C-H
animals), which suggests a suppression of carbohydrate metabolism on a high-fat
diet. As this ‘metabotype’ does not persist in H-C animals, it may result from
hyperglycaemia that was resolved with the diet switch in this group.
Limitations of this study include no information on energy consumption and energy
balance in individual animals. Analysis of body mass as a function of metabolomic
profile also reveals the presence of outliers (n = 2), animals that did not gain the
expected weight when placed on a high-fat diet (figure S2). The presence of these
animals did not alter the model outcomes. Finally, the diet switch results in slight
weight loss in both C-H and H-C. The effects of this loss on the terminal
metabolomic profile cannot be discounted.
3.5 Conclusions
In this study, NMR was employed as a highly sensitive tool to discriminate the
effects of diet from those of obesity in the mouse. Novel findings show
metabolomic profile to be predominantly influenced by recent diet, despite fasting
prior to serum collection. Short-term diet effects were mainly centred around
changes in energy metabolism and glucose utilization, whereas obesity-related
changes were focused on amino acids and large non-polar molecules. This
underscores the importance of dietary controls when profiling obese states,
especially when highly sensitive techniques are employed.
77
Chapter 4 Metabolomic response to exercise training in lean and diet-induced
obese mice
Gavin E. Duggan, Dustin S. Hittel, Christoph W. Sensen, Aalim M. Weljie, Hans J. Vogel, and Jane Shearer
Published: 2011-01-23 in the Journal of Applied Physiology
4.1 Abstract
Exercise training is a common therapeutic approach known to antagonize the
metabolic consequences of obesity. The aims of the present study were to examine
1) whether short-term, moderate-intensity exercise training alters the basal
metabolite profile and 2) if 10 days of mild exercise training can correct obesity-
induced shifts in metabolic spectra. After being weaned, male C57BL/6J littermates
were randomly divided into two diet groups: low fat (LF) or high fat (HF). After 12
wk of dietary manipulation, HF animals were obese and hyperglycemic compared
with LF animals. Mice from each group were further divided into sedentary or
exercise treatments. Exercise training consisted of wheel running exercise (2 h/day,
10 days, 5.64 m/min). After exercise training, animals were rested (36 h) and fasted
(6 h) before serum collection. Samples were analyzed by high-resolution one-
dimensional proton NMR. Fifty high- and medium-concentration metabolites were
identified. Pattern recognition algorithms and multivariate modeling were used to
identify and isolate significant metabolites changing in response to HF and exercise
training. The results showed that while exercise can mitigate some of the abnormal
patterns in metabolic spectra induced by HF diet feeding, they cannot negate it. In
fact, when the effects of diet and exercise were compared, diet was a stronger
predictor and had the larger influence on the metabolic profile. External validation
of models showed that diet could be correctly classified with an accuracy of 89%,
whereas exercise training could be classified 73% of the time. The results
demonstrate metabolomics to effectively characterize obesity-induced perturbations
in metabolism and support the concept that exercise is beneficial for this condition.
78
4.2 Introduction
Obesity is reaching epidemic proportions in the population. The health implications
of obesity are well known and include an increased mortality from cardiovascular
disease, type 2 diabetes, and cancer (Stamler et al., 1993; Bonora et al., 1996;
Giovannucci and Michaud, 2007). A common therapeutic approach to treating
obesity is exercise. Exercise increases caloric expenditure, alters patterns of
substrate utilization, and enhances whole body insulin sensitivity (Yamanouchi et
al., 1995). However, the volume, intensity, and specific mechanisms underlying the
beneficial health effects of exercise are debatable and often difficult to assess.
A largely unexplored yet highly sensitive tool to examine the effects of exercise
training is metabolomics. Metabolomics is the quantitative assessment of small low-
molecular-weight metabolites in response to physiological or pathophysiological
stimuli. While the study of metabolites has been a staple of exercise physiology
(e.g., blood lactate), advances in technology now allow for the simultaneous
quantification of numerous metabolites in a small sample volume. This unbiased,
systematic approach, or “metabolomic profile,” is powerful in that it permits the
investigator to evaluate the fluctuations of a biochemical pathway and its individual
components. Unlike transcripts or proteins, which often take hours, days, or even
weeks to change, metabolites are the end products of a reaction and closely reflect
underlying physiological processes (Weljie et al., 2006). For this reason, the outputs
of metabolomics are highly sensitive indicators of an organism's physiological or
disease status (Glassbrook et al., 2000; Glassbrook and Ryals, 2001).
Using the common laboratory mouse, the present study aimed to determine 1)
whether short-term, moderate-intensity exercise training alters basal, whole body
metabolite profiles and 2) if exercise training can correct obesity-induced shifts in
metabolic spectra. Our laboratory has previously characterized the metabolomic
profile of diet-induced obesity in this model (Shearer et al., 2008). The results of the
present study showed that 10 days of low-intensity exercise training was sufficient
79
to normalize blood glucose in obese animals and induce global metabolomic shifts
in both lean and obese animals. When diet and exercise treatments of rested animals
were compared, diet rather than exercise training had the most pronounced effects
on the metabolomic profile, with obese animals displaying perturbations in
branched-chain amino acids (BCAAs), large noncharged amino acids, and numerous
mediators of insulin signaling.
4.3 Methods
Mouse maintenance
Procedures were approved by the Animal Care and Use Committee of the
University of Calgary and abided by the Canadian Association for Laboratory
Animal Science guidelines for experimentation. Animals were maintained in a
humidity-controlled room with a 12:12-h light-dark cycle. After being weaned (3
wk of age), male C57BL/6J littermates were randomly segregated into two groups
and maintained in microisolator cages for 1 wk. After this acclimation period,
animals received either a low-fat (LF) diet or a high-fat (HF) diet for 12 wk (58R3,
TestDiet, Purina, Richmond, IN). The energy density of the LF diet (5001
Laboratory Rodent Diet, Purina) was 234.0 g/kg of energy as protein, 45.0 g/kg of
energy as fat, and 499.0 g/kg of energy as carbohydrate. The HF diet contained
197.4 g/kg of energy as protein, 358.0 g/kg of energy as fat, and 358.2 g/kg of
energy as carbohydrate. The primary source of fat in this diet was lard. Both diets
met all nutritional requirements of adult mice. To encourage the development of
diet-induced obesity, food and water were provided ad libitum throughout the
experiment. Food consumption was not monitored or controlled throughout the
study.
Exercise training
After 12 wk of dietary manipulation, mice were randomly assigned to either
sedentary (SED) or exercise treatment (EX) groups (n = 9–12 mice/treatment).
80
Exercise was performed on a rotating treadmill wheel system (model 80800A-10,
Lafayette Instruments, Lafayette, IN) during daylight hours (9–11 AM).
Acclimation to exercise consisted of 15 min of treadmill exercise at 5.64 m/min in
EX animals followed by a 3-day rest period. Exercise was then performed at a
constant intensity for a defined duration (5.64 m/min, 2 h/day) for 10 consecutive
days. Exercise intensity was chosen so that obese mice could complete the entire
protocol. To account for stress induced by animal handling, SED animals were also
placed in a stationary treadmill for both acclimation and exercise treatments.
Animal experimentation.
At the end of the exercise period, animals were rested for 36 h. This rest period was
chosen to negate any direct effects of exercise. Previous work (Cartee et al.,
1989)has shown that 18 h of rest is sufficient for the acute effects of exercise on
insulin responsiveness to subside. This rest included a 6-h fasting period to negate
the effects of postprandial food absorption. Animals were then weighed and
anesthetized (pentobarbital). Whole blood (~1 ml) was obtained by a cardiac
puncture, placed on ice, and allowed to clot for 30 min. Samples were then
centrifuged for 10 min (3,000 rpm), and sera were collected before storage at
−80°C. Blood glucose was assessed in anesthetized mice (One Touch, Lifescan,
Burnaby, BC, Canada). Nonesterified fatty acids (NEFAs) were measured
spectrophotometrically (Wako NEFA C kit, Wako Chemicals, Richmond, VA).
Immunoreactive insulin was assayed with a double-antibody method (MORGAN
and LAZAROW, 1962). Abundant data on LF and HF diet-fed C57BL/6J mice have
been previously published (Fueger et al., 2005; Ayala et al., 2007).
Metabolite sample preparation.
Serum samples of ~0.3 ml were stored at −80°C before data acquisition. Samples
were thawed and filtered twice using 3-kDa NanoSep microcentrifuge filters, which
had been prewashed to reduce preservative contamination. The filtrate was
81
transferred to clean microfuge tubes; the final sample volume ranged from 100 to
300 ul. Samples were brought to 450 ul by the addition of 140 ul phosphate buffer
containing dimethyl silapentane sulfonate (DSS; final concentration: 0.5 mM), 40 ul
sodium azide, and distilled H2O. The final sample pH was adjusted to 7.0 ± 0.01.
Spectrum acquisition.
One-dimensional nuclear overhauser effect spectroscopy (NOESY) spectra were
acquired, in two batches, using an automated NMR case sample changer on a 600-
MHz Bruker Ultrashield spectrometer. The NOESY pulse sequence had a mixing
time of 100 ms and a water presaturation pulse. Samples were individually shimmed
to ensure a half-height line width of ~0.8 Hz for the major DSS peak, which was
calibrated to 0.0 ppm. Spectra were acquired with 768 scans, zero padded, and
Fourier transformed to 64,000 points. Standard postprocessing of the spectra
included deletion of the water region, B-spline baseline correction, reference
deconvolution, and calibration of the DSS peak. A 1H13C heteronuclear single
quantum coherence (HSQC) spectrum of one sample, chosen at random, was
acquired for peak assignment and verification.
Metabolite concentration profiling.
Processed spectra were imported into Chenomx software (version 4.6) for
quantification (Edmonton, AB, Canada). Fifty compounds were profiled based on
chemical shift assignments verified by 1H13C two-dimensional HSQC. Spectra were
randomly ordered for fitting in the Chenomx Profiler to avoid progressive bias.
Compounds were fit from the highest initial concentration to the lowest, with an
iterative reinspection. After all samples had been quantified, Perl scripts were used
to detect inconsistencies in peak assignment for further reinspection. Each
measurement was normalized to the mean sample concentration by dividing each
profiled spectral concentration by the total concentration of all profiled metabolites
in that sample (Weljie et al., 2006). For univariate reporting and analysis, profiled
concentrations were scaled back to the mean sample dilution.
82
Multivariate analysis.
Projection to Latent Structures (PLS) is a family of pattern recognition tools used to
model differences between treatment groups. Normalized concentration measures
were imported into SIMCA-P software (Umetrics, Sweden) for multivariate pattern
analysis. To evaluate the interaction and separation of diet- and exercise-related
effects, PLS was used to characterize the metabolomic changes seen within the four
treatment groups. Centered and scaled PLS coefficients were used to evaluate the
relative importance of metabolites in each model. Because of its ability to further
isolate changes in metabolism specific to each treatment, Orthogonal PLS (OPLS)
was subsequently used to isolate diet-related changes in each exercise group. An
OPLS model comparing LF and HF diet-fed animals in the SED treatment was
created (Y = 0 and Y = 1, respectively). An identical model was constructed for the
EX animals, and the resulting model structures were compared.
Model validation
The strength and reliability (“goodness of fit”) of models derived from
metabolomics data were assessed by comparing the percentage of variation captured
in known (modeling) and unknown (testing) samples. Here, two samples from each
treatment group or 26% of the samples collected in the study were randomly
selected and blinded to the investigator. Samples were then examined for fit into
previously described OPLS models. Cross validation was then performed to test
each model's ability to predict the class of samples not used in its creation.
Jackknifing was used to calculate SEs and t-value 95% confidence intervals for PLS
coefficients.
Network analysis
To establish the relationships between metabolites and insulin signaling, metabolite
loadings were exported into the Ingenuity Pathway Analysis (IPA) tool (Ingenuity
Systems, Redwood City, CA; http://www.ingenuity.com). A separate IPA was
83
performed for SED and EX animals; in each case, only loadings for metabolites
changing significantly (P < 0.05) in response to the HF diet were included in the
IPA.
Protein determination
Total protein and phosphorylation of Akt and mammalian target of rapamycin
(mTOR) were determined in skeletal muscle (gastrocnemius) and liver lystates.
Lysates were prepared in Laemmli buffer and separated by SDS-PAGE. Samples
were resolved on 4–12% bis-Tris SDS-PAGE gels (Invitrogen, Carlsbad, CA)
followed by electrophoretic transfer to polyvinylidene difluoride membranes
(Millipore, Billerica, MA). Membranes were blocked in 2% nonfat milk diluted in
Tris-buffered saline (TBS) containing 0.05% Tween 20. Membranes were probed
with primary antibodies overnight at 4°C and then incubated with secondary
antibodies for 1 h at room temperature. Phospho-Akt (Ser473) and phospho-mTOR
(Ser2448) were normalized to total Akt and mTOR as well as GAPDH, which was
run as a secondary loading control (data not shown). Primary antibodies were as
follows: total Akt, phospho-Akt (Ser473), total mTOR, and phospho-mTOR (Ser2448)
(all from Cell Signaling Technology, Boston, MA) and GAPDH (Abcam,
Cambridge, MA). All antibodies were diluted in 2% nonfat milk diluted in TBS
containing 0.05% Tween 20. Membranes were washed in TBS containing 0.05%
Tween 20. Densitometry was performed using GeneTools (Syngene, Fredrick, MD).
Statistical analysis
Differences between body weight, blood glucose, insulin, and NEFAs were
determined using two-way ANOVA. A significance level of P < 0.05 was used, and
differences with ANOVA were determined using a Tukey's post hoc test. Two-way
ANOVA was used to calculate individual F-test significance for each metabolite.
Metabolites were considered to be significant if they had a P value of <0.05 in both
OPLS models or a univariate significance of P < 0.05 for either factor. All data are
reported as means ± SE.
84
4.4 Results
Animal characteristics
After dietary manipulation, HF diet-fed animals became obese, with average
weights of 30.8 ± 0.9 and 46.6 ± 1.5 g for LF and HF diet-fed animals, respectively
(P < 0.05). These data confirm those of other studies (Ayala et al., 2007)
demonstrating that HF diet feeding in the C57BL/6J mouse causes significant
obesity and insulin resistance, disposing of ~50% less glucose compared with chow-
fed animals. Animal characteristics are shown in Table 4.1: Change in body mass,
glucose, NEFA and Insulin levels. Exercise training resulted in a slight weight loss
in both LF and HF diet-fed animals (P < 0.05). This finding was unexpected given
the short duration and mild nature of the EX regime. Fasting blood glucose levels
were elevated in HF diet-fed SED animals but not HF diet-fed EX animals,
indicating the EX protocol was effective in abolishing diet-induced hyperglycemia
(P < 0.05). Fasting NEFA levels were not different between groups. Plasma insulin
levels were greater in HF diet-fed animals compared with LF diet-fed animals
regardless of the exercise intervention (P < 0.05). However, no differences were
noted between HF diet-fed SED animals and HF diet-fed EX animals.
85
LF-SED LF-EX HF-SED HF-EX
Animals/Group 10 10 10 9
Initial Body Mass (g) 32.1±1.3 29.4±1.2 47.8±1.8 * 45.4±2.7 *
Final Body Mass (g) 32.5±1.3 28.6±0.8 47.7±1.5 * 39.9±1.8 *†
Percent Change Mass 1.3±1.0 -2.6±1.6 † 0.0±1.1 -11.4±2.5 *†
Table 4.1: Change in body mass, glucose, NEFA and Insulin levels
Values are means ± SE. NEFA, Non-esterified fatty acids ; LF, low fat; HF, high fat; SED, sedentary. Body weight was assessed before and after the exercise intervention. * P < 0.05, significance between LF and HF diets within a treatment (SED or EX); † P < 0.05 between SED and EX groups within a diet.
Model validation
Test samples used for validation were blinded to the investigator and not used in
model creation. While n = 9–12 samples were obtained, only n = 7–8 samples are
reported in the models. The remaining samples were used for validation purposes.
The results showed that there was a greater variation of metabolites due to diet than
exercise treatment. This can be seen in the greater spread between samples in the
score plot (Figure 4.1A). Cross-validation tests of model quality demonstrated diet
class prediction with 89% accuracy in EX animals and 73% in SED animals. Given
the limited sample size used in model creation (n = 7–8 samples/treatment) and the
complex physiology of the diet and exercise regimes, these validation scores
indicate that the generated PLS models were veridical.
Model interpretation.
Although diet and exercise manipulations resulted in distinct separation of
treatments, significant interactions between groups were visible (Figure 4.1A). To
86
separate these interactions, the data was modeled with OPLS (Figure 4.1B).
Analysis of the resulting loadings (Figure 4.1B) showed the response of both SED
and EX animals to HF diet feeding. Metabolites on the dashed line of identity in
Figure 4.1B responded similarly to HF diet feeding, regardless of exercise.
Conversely, metabolites closer to the top left or bottom right corner of Figure 4.1B
responded oppositely to the HF diet depending on the animal's exercise regime.
Individual metabolites.
To determine the individual metabolites changing in response to exercise, diet, and
the diet-exercise interaction, two-way ANOVA was performed. Individual
concentrations of significant metabolites in each treatment group are shown in Table
4.2. A metabolite was considered significant if it had a univariate ANOVA factor P
< 0.05 or a multivariate P value of <0.05 for both EX and SED animals. Statistical
differences and interactions for each metabolite in Table 4.2 are shown in Table 4.3.
A column plot showing the magnitude and direction of change for individual
metabolites for SED and EX animals is shown in Figure 4.2A.
When exercise and diet treatments were directly compared, diet had the most
profound effects on the metabolite profile. Given the animals were rested for 36 h
before serum collection, this finding was not unexpected. Of note, the HF diet
caused a lower concentration in BCAAs Figure 4.2B) as well as many large polar
amino acids, including phenylalanine and tyrosine. In addition, taurine and
methionine, key metabolites involved in lipid homeostasis and insulin sensitivity,
were lower in HF diet-fed animals (Table 4.2).
87
Figure 4.1: Class separation and comparison of SED vs EX response
A: two-dimensional score plot showing the Projection to Latent Structures (PLS) model's ability to separate samples based on diet (horizontal dimension) and exercise (vertical dimension). Multivariate scores represent each sample's composition; the corresponding axes of the score plot represent how much of each pattern is seen in a sample. Validation scores for the PLS model for both diet and exercise were 44%, 72%, and 27% for R2
X, R2Y, and Q2
Y. The model was generated with n = 7–8 samples/treatment. LF, low-fat; HF, high fat diet; SED, sedentary
88
treatement; EX, exercise treatment. B: comparison of metabolic responses in EX and SED animals subjected to a HF diet. Metabolites on the dashed line of identity responded similarly regardless of exercise, e.g., metabolites further to the right rose more in SED animals and did not respond in EX animals (glucose and ornithine). Multivariate significance of each response is indicated by the open circles (significant in one population) or solid circles (both populations) [by Orthogonal PLS (OPLS)]. OPLS validation scores of diet for SED animals were 79%, 95%, and 73%, whereas EX animals had scores of 73%, 98%, and 89% for R2
X, R2Y, and Q2
Y, respectively. The plot represents 7–8 animals/treatment. TMAO, trimethylamine oxide.
Table 4.2 Individual metabolite changes in response to exercise and diet
Values are means ± SE (in uM); n = 7–8 samples/treatment. Data were categorized by treatment groups for all significant metabolites. Metabolites were considered significant if any factor ANOVA P values or both Orthogonal Projection to Latent Structures (OPLS) coefficient P values were <0.05. Statistical data for all metabolites are shown in Table 3.
Table 4.3 Univariate and multivariate significance scores
n=7-8 samples/treatment. * P < 0.05; † < 0.05
90
Figure 4.2 Significance of changes in metabolite levels
(A) column plot representing the magnitude and directionality of change for significant metabolites (OPLS diet coefficients) in SED animals (solid bars) and EX animals (shaded bars). Each bar represents the difference between HF and LF (CH) diet-fed animals within a diet. For example, acetate was lower in both SED and EX animals; however, the magnitude of the change was greater in the EX group. Nonsignificant coefficients are indicated by the open bars for comparison. The plot represents 7–8 animals/treatment.
(B) metabolic responses of branched-chain amino acids in LF or HF diet-fed animals subjected to either SED or EX conditions. *Significant difference between dietary treatments (P < 0.05). The plot represents 7–8 animals/treatment. Data are means ± SE.
91
Network analysis
To leverage the system-wide perspective of metabolomics and to further investigate
the relationships between individual metabolites, IPA (Ingenuity Systems;
http://www.ingenuity.com) was used. While significantly changed metabolites were
relevant to a number of biological networks and functions (Supplemental Material,
Supplemental Fig. S1), pathways related insulin signaling were of particular interest
and were highlighted.1 In both SED and EX samples, numerous metabolites
clustered with the functional groupings associated with insulin, proinsulin, and p38
MAPK. This metabolic “neighborhood,” and the changes seen in each sample
group, are shown in Supplemental Figs. S2 and S3 of the Supplemental Material.
Protein content and phosphorylation.
Total and phosphorylated levels of Akt and mTOR were evaluated in skeletal
muscle and the liver. Levels of total Akt, mTOR, and GAPDH did not change with
genotype or treatment and were used as loading controls. All graphical data are
shown as values normalized for the respective total protein. Protein levels of mTOR
(Ser2448) were elevated in the liver with HF diet feeding in SED animals but not in
EX animals (Figure 4.3A). No differences in the phosphorylation of this protein
were observed in skeletal muscle (Figure 4.3B). Examination of phospho-Akt
(Ser473) revealed that HF diet feeding increased the phosphorylation of this protein
in the liver of SED animals (Figure 4.3C). EX lowered the levels of HF diet-induced
phospho-Akt (Ser473) with no differences between LF and HF diet-fed animals. In
skeletal muscle, no differences in phospho-Akt (Ser473) were observed in SED
animals. However, EX resulted in an increase in phospho-Akt (Ser473) in HF diet-
fed animals but not in LF diet-fed animals, implying a higher relative exercise
training intensity in these animals (P < 0.05; Figure 4.3D).
92
Figure 4.3 mTOR and Akt western blot responses in liver and muscle
Total and site-specific phosphorylation of mammalian target of rapamycin (mTOR) and Akt in skeletal muscle and liver lysates. A and B: phosphorylation of mTOR (Ser2448) in the liver (A) and skeletal muscle (B). As mTOR (Ser2448) is phosphorylated by Akt, phosphorylation of this protein was also assessed. C and D: phosphorylation of Akt (Ser473) in the liver (C) and skeletal muscle (D). Values were normalized to their respective total protein contents and are expressed in arbitrary units. Data are means ± SE; n = 6–8 animals/treatment. *P < 0.05, LF vs. HF diet-fed animals within SED or EX conditions; **P < 0.05 vs. all other treatments.
93
4.5 Discussion
Exercise training effectively enhances the rates of energy expenditure and substrate
flux, creating an ideal situation for large-scale metabolomic profiling. While single-
metabolite measures have been a staple of physiological analysis for over a century,
using magnetic resonance-based strategies in combination with targeted metabolite
profiling represents a relatively novel tool. The goals of the present study were to
determine whether short-term, moderate-intensity exercise training could alter basal
metabolite profiles. A secondary aim was to examine whether exercise training
could correct obesity-induced shifts in metabolic spectra.
The results demonstrate that 1H NMR profiling can clearly distinguish between SED
and EX in both LF and HF diet-fed animals. Additional findings demonstrate that
while exercise can mitigate some of the abnormal patterns in metabolic spectra
induced by HF diet feeding, they cannot negate it. In fact, when the effects of diet
and exercise training were compared in the basal state, diet was a stronger predictor
and had the larger influence on the metabolic profile. To test the reliability of the
models generated in this study, external validation of the models was performed.
Here, two samples from each treatment group or 26% of the samples collected in the
study were randomly selected and blinded to the investigator. Samples were then
examined for fit into PLS models. External validation samples were not used in
model creation. The results demonstrated the diet could be correctly classified with
an accuracy of 89%, whereas exercise training could be classified 73% of the time.
These predictive percentages are high considering the limited sample sizes and the
complex physiology of HF diet feeding and exercise training. As such, we can
conclude that metabolomic profiling can create biologically relevant models to
predict responses to both diet and exercise treatments.
In the present study, the common laboratory mouse was chosen to examine the
effects of diet and exercise on the basal metabolomic profile. In addition to offering
strict control over genetic background, diet, exercise, and age, the HF diet-fed
94
C57BL/6J mouse also recapitulates human obesity. Much like the human condition,
obesity occurs over a prolonged time period and is multifactoral (as opposed to a
single genetic mutation) and results in impaired insulin sensitivity (Fueger et al.,
2004a). When fed a HF diet for 12 wk, C57BL/6J mice experience considerable
insulin resistance, as measured by a hyperinsulinemic-euglycemic clamp (Fueger et
al., 2004b; Shearer et al., 2008). Network analysis of significant metabolites in both
diet and exercise treatments showed them to converge on insulin signaling
(Supplemental Material, Supplemental Figs. S1 and S2). Amino acids are of
particular interest as they are intimately involved in cell signaling, gene expression,
and protein phosphorylation (Wu, 2009). Despite near-identical concentrations in
the diets, BCAAs were lower with HF diet feeding but largely unaffected by
exercise. These findings are in agreement with studies in Sprague-Dawley rats
demonstrating a 22% reduction in BCAA levels after 4 wk of HF diet feeding
(Calles-Escandon et al., 1984) as well as in humans, where plasma leucine levels
were depressed in type 2 diabetes and restored with 6 wk of rosiglitazone treatment
(Van Doorn et al., 2007). It remains unknown whether differences in BCAAs with
obesity stem from differential dietary consumption or metabolism. Recent work by
Newgard et al. (Newgard et al., 2009) has shown a strong relationship between a
BCAA metabolite model (principal component analysis score) and insulin resistance
(homeostatic model assessment). As BCAAs have been implicated in insulin
signaling, in part through activation of the mTOR signaling cascade, phospho-
mTOR (Ser2448) and phospho-Akt (Ser473) were examined in the liver and skeletal
muscle. The results showed increased phosphorylation of both proteins in the liver
with HF diet feeding in SED animals. Exercise training eliminated these differences.
No differences in skeletal muscle phospho-mTOR (Ser2448) were noted. However,
there was an increase in skeletal muscle phospho-Akt (Ser473) with EX in HF diet-
fed animals. This differential response is likely due to a greater relative exercise
training intensity in this group compared with LF diet-fed animals. No direct
relationships between BCAAs and either mTOR or Akt phosphorylation could be
established from the present data.
95
Despite animals resting for 36 h before sample collection, serum lactate levels were
also lower in HF diet-fed animals in both SED and EX conditions. The lower lactate
levels in HF diet-fed animals may reflect a diminished turnover of carbohydrate
stores. Additional metabolites declining with HF diet feeding in both SED and EX
were taurine and its precursor, methionine. This semiessential amino acid has been
implicated in the regulation of lipid homeostasis, insulin secretion, glucose uptake,
and antioxidant defence (Franconi et al., 2006). The fact that this pathway was
downregulated in HF diet-fed animals highlights that their obese, insulin-resistant
phenotype was largely unaltered by the exercise regime. Supplementation of taurine
to animal and human diabetes has shown it to reduce disease severity (Franconi et
al., 2006; Wu, 2009). Besides individual metabolites, clusters of metabolites that are
closely related and react similarly can be useful in indentifying points of
disregulation. Taken individually, the significance of some of these metabolites
might be overlooked, but their concerted pattern may be of interest. In the present
study, arginine, citrulline, and proline are all closely related via nitric oxide
synthesis to vasculature regulation and have been implicated in insulin delivery,
sensitivity, and glucose uptake (Wu and Meininger, 2009). These metabolites
clustered in the OPLS model (Figure 4.1B). The results showed all of these
metabolites were elevated in HF diet-fed animals, regardless of exercise; coupled
with the one-directional nature of arginine synthesis from ornithine, its depletion in
EX animals invites further study.
Exercise also had profound effects on a number of individual metabolites. In HF
diet-fed animals undergoing EX, there was a normalization of circulating glucose,
indicating a therapeutic benefit. Exercise training has been shown to increase
glucose transporter (GLUT4) expression and glucose tissue utilization in this model
(Fueger et al., 2004b, 2004c). Additional metabolites changing in response to
exercise included glutamine and acetate. Primarily produced by skeletal muscle,
resting levels of glutamine declined after EX. The ratio of glutamine to glutamate,
which is often used as an indicator of training, was also altered, with a lower value
96
of 4.6 in the LF diet-fed EX group and a high value of 9.9 in the HF diet-fed SED
group. This fact is reinforced by the network analysis, where glutamate was
centrally featured in both SED and EX response to a HF diet (Supplemental Figs. S2
and S3). Finally, acetate is a precursor to acetyl-CoA, which is metabolized by the
tricarboxylic acid cycle. Lower concentrations in EX animals are indicative of a
greater reliance on glucose and enhanced accumulation of glycogen in tissues
(Imoto and Namioka, 1983a, 1983b).
As metabolomic profiling can quantify entire spectral profiles, metabolite clusters,
and individual metabolites, this technology has typically been used to examine
disease states including cardiovascular disease (Kirschenlohr et al., 2006), arthritis
(Weljie et al., 2007), and cancer (Chan et al., 2009). It is only recently that
metabolomics has been used to examine health, obesity, insulin resistance, and
exercise training. Of note, a study by Yan et al. (Yan et al., 2009) specifically
examined the effects of exercise training. Analysis of sera from professional rowers
undergoing rigorous training showed the technology to successfully detect
differences between control subjects and athletes as well as differences in training
duration and years of athlete experience. Metabolomic profiling was also sensitive
enough to discriminate changes in metabolites occurring with 1 versus 2 wk of
exercise training. Indeed, metabolomics technology may eventually be used to
determine individual responses to exercise training, overtraining, and nutritional
supplementation.
Limitations of the present study include the measurement of metabolites by NMR.
This technology is not as sensitive as mass spectroscopy and detects only
metabolites at higher concentrations within the metabolome. Ideally, both
methodologies would be included. Additional limitations include the presence of
anesthesia, a factor that could undoubtedly alter metabolomic profiling in these
animals. However, given the large volume of blood collected from animals, this
cofounder was unavoidable. Other limitations include the measurement of serum
and not tissue metabolites. Although we can confirm LF and HF diet-fed animals
97
had differential responses to the exercise training regime, we cannot pinpoint the
source or tissue responsible for these differences. It is likely that differences in
relative exercise intensity and skeletal muscle and hepatic metabolism were all
contributors to the observed metabolomic profiles. In summary, we show that
metabolomics is capable of discriminating prior exercise training in a basal state in
lean and obese mice. Our results also show that, when compared, diet rather than
exercise training is the predominant determinant of the resting metabolic profile.
98
Chapter 5 Comparison of multiple high-fat mouse metabolomics experiments
5.1 Introduction
Chapters 3 and 4 are variations on the theme presented in chapter 2, analysing the
effects metabolic of high fat diet in a susceptible mouse model of insulin resistance.
By introducing additional factors, such as a short term diet change or exercise
intervention, we can establish connections between our metabolomic findings and
the known etiology of the disease. They formed a natural progression in
experimental design, driven by the exposure of potential confounders or
opportunities in earlier studies and allowing an exploration of the potential of
metabolomics.
Moreover, since each experiments contains, at its core, the same experiment
comparing chow-fed animals to those on a high fat diet, we may be able to draw
some additional information from comparisons of the results. To the extent they are
the same, we can take the similarities as evidence that the models are, at least,
reproducible and more likely biologically drive. To the extent the results differ they
may also provide insight to the limitations of the process applied.
It bears noting that the studies published as chapters 2-4 were conducted over a span
of 5 years, during which the rapidly growing field of metabolomics has progressed.
The author’s knowledge of metabolomics, statistics, and biochemistry has also
advanced during the same window, and some of the results published in those
chapters are incomplete by modern standards. The fact that those chapters were
written with an eye to translational acceptance by a non-technical audience gave rise
to some incompleteness, but most of the limitations were not yet well characterised
when published. Thus, while a complete reanalysis of the resulting data is not
within the scope of this work, a comparison of the results should take into account
their limitations.
99
5.2 Comparison of resulting models
In terms of specific metabolites, a number of compounds were consistently
identified as existing at different concentrations in high-fat and chow-fed animals.
Both individual compounds and combinations thereof, especially as identified by
pathway analysis (either curated or automated) are of interest; all have some history
in the insulin resistance literature, and many have been investigated further in the
time since publishing.
5.2.1 Branched Chain Amino Acids
The most conserved response to the introduction of a high fat diet, and presumably
the induction of insulin resistance, was a drop in the level of branched chain amino
acids. Across all three studies, three large non-polar amino acids (leucine,
isoleucine, and valine) were consistently depleted in animals on a high fat diet,
regardless of any diet-switch or exercise considerations.
The disruption of BCAA levels in insulin resistant individuals (Luetscher Jr, 1942;
Harper et al., 1984; Garlick and Grant, 1988; Marchesini et al., 1991), and the
interplay between their level has long been established (Forlani et al., 1984; Brooks
et al., 1986), so this was both reassuring and interesting as we hypothesized that
correlated shifts in other metabolites could suggest potential mechanisms of action.
As previously discussed, leucine in particular has been the focus of much research
in insulin resistance as it shows significant changes in numerous models and human
studies. In light of this, and its strong recurrence as a signal in these metabolomic
studies, it may be worth examining models with leucine as a supervisory variable to
draw out strongly correlated mechanistic shifts in metabolism.
100
Figure 5.1 BCAA mechanism of IR induction proposed by Newgard et al
When introduced as part of, or in supplement to, a high fat diet, branched
chain amino acids overwhelm amino acid catabolism and result in an
increase of ketone bodies. The outcome is modified availability of
acylcarnitine transport and single carbon flux to gluconeogenesis.
Reproduced from (Newgard et al., 2009)
One of the most prominent metabolomic analyses since the publishing of our initial
findings was that of Newgard et al in 2009, which reinforced and clarified the amino
acid signature associated with the onset of insulin resistance. Newgard’s analysis
was particularly elegant because it combined human studies with an animal model
while studying not just animo acids but fatty acid, derivatives (triglycerides and
acylcarnitines, in particular), hormone levels and phosphorylation assays similar to
those performed in chapter 4. Backed by a strong correlation with the Homeostasis
101
Modeling Assessment score, an established metric, and the interaction with
rapamycin (an mTOR activation inhibitor) they used supplementation of BCAA and
general-composition amino acids diet to tease out a specific interaction between the
branched chain acids and the high fat diet. In particular, the use of measured
respiratory energy requirement would constitute another potentially strong
supervisory variable for future metabolomic studies of insulin resistance.
5.2.2 Aromatic amino acids
Phenylalanine and tyrosine are not traditionally grouped in with branched chain
amino acids, but both have some history of with insulin resistance and were
included in Newgard’s amino acid signature (Newgard et al., 2009; Friedrich, 2012).
In our datasets tyrosine did not show significant changes in response to the high fat
diet, but phenylalanine showed a significant drop in response, correlated with
increase in bodymass.
Therefore, it is particularly interesting that switching animals to a lower fat chow
diet for the final week of their treatment appears to abolish this effect suggesting it
may be a symptom of insulin resistant animals’ inability to process dietary sources.
Newgard hypothesizes that the change in phenylalanine level may be because of
oversaturation of the large neutral amino acid transporter, while in 2011 Wang et al
noted in a large human cohort that phenylalanine increases in serum up to a decade
before the onset of diabetes (Wang et al., 2011). One of the more difficult
etymological and etiological challenges poised by diabetes is its existence as a
relatively late endpoint, before which many biological systems have been altered
and/or disrupted (Li et al., 2010). As such, the diet-switch protocol should likely be
expanded to investigate the response of specific amino acids and isolate diverse
involvements. The possibility of clamping phenylalanine or other potential
signalling amino acids (Castellino et al., 1987; Chevalier et al., 2004), rather than
glucose in a metabolomics context would likely prove informative.
102
5.2.3 Single-carbon amino acids
Alanine was another amino acid whose response different in the presence of a final-
week chow diet, but was otherwise elevated in multiple high-fat fed populations.
Glycine, in contrast was depressed in both HC and HH diets, but did not change
significantly in either exercised or sedentary high-fat animals in chapter 4. Since
glycine is known to provide a single carbon source for several nucleic acid synthetic
pathways, these results may provide a small insight to the hypothetical
overburdening of amino acid catabolism. Under such a disruption, glycine levels
would be elevated not because of lowered uptake but a surplus of free amino acids,
resulting in decreased flux through nucleosynthetic pathways.
5.2.4 Carnitine and Ketone bodies
While a number of energy-related metabolites were consistently lowered in high fat
animals (acetate, TMAO), the consistent increase in ketone bodies such as
hydroxybutyrate may shed more light on the biochemical disruptions associated
with insulin resistance. Energy metabolites represent a difficult area in which to
draw strong conclusions given the differences in diet and energy source, even under
the diet switch. Ketone bodies, on the other hand, factor heavily into both amino
acid synthesis and fatty-acid metabolism (Salek et al., 2007), and occupy a central
position in a hypothetic link between branched chain amino acid metabolism, fatty
acid metabolism, and a shift in energy usage towards carbohydrates (Honors et al.;
Newgard et al., 2009; Friedrich, 2012). The fact that 3-hydroxybutyrate levels
remained elevated notwithstanding the diet switch, and responsed identically in
exercised and sedentary animals but did not change in correlation with bodymass
highlights its value as a potential central indicator for this pathway, and reinforces
the hypotheses put forward about the BCAA/fatty acid metabolism link.
On a similar note, the level of several carnitine metabolites also showed significant
differentiation in high-fat fed animals. While their large size prevents the
measurement of acylcarnitines using NMR instrumentation, it’s of note that the
103
level of the smaller acetylcarnitine molecule was higher in two of our four HF
sample sets, but lower in the diet-switched HF samples, lower in the sedentary HF
animals and showed a muted decrease in exercised animals. According to
Friedman, serum acylcarnitine levels were elevated in several HF diet studies, which
suggests that the pool of carnitine used for mitochondrial transport may be depleted
or destroyed by oxidative radicals in the insulin resistant state, but the inconsistent
responses seen here are perplexing.
5.2.5 Choline
One final metabolite of interest which showed a consistent change in HF fed
animals was Choline, which was significantly depressed in all HF datasets
notwithstanding a final week diet shift. In contrast, the change in phosphocholine
levels was not consistent but went up in some models and down in others. Choline
plays a smaller role in the literature of insulin resistance (Zhang et al., 2008) and
does not factor into the BCAA-mediated shift in fatty acid metabolism proposed
earlier. Indeed, the paper published as chapter 2 represents the primary publication
directly linking choline to insulin resistance with metabolomics, but other
publications (Mooradian, 1987; Brenner et al., 2000) have linked this salt with
insulin resistance and the CDP-choline pathway in non-metabolomics contexts so its
role as symptom, upstream effector or side effect bear further investigation.
5.3 Limitations and Omissions
Several caveats bear mentioning as they apply to all of the datasets analysed. While
the small number of samples was potentially informative vis-à-vis the predictive
power of metabolomics in a typical experimental cohort, it would have been more
valuable to ascertain the true metabolic effect of insulin resistance and the ability of
a metabolomics model to identify insulin resistance in each singular, new sample
outside of the training data. As such, a larger dataset would have been useful for
both increasing power and better characterising the magnitude of the effect based on
diet feeding, etc.
104
While some tissue-specific assays were performed to characterise the level of
phosphorylation in mTOR, only system-wide serum levels were analysed in a
metabolomics context. Since many of the proposed causes of insulin resistance
occur specifically in the liver as a result of hepatic steatosis (Friedrich, 2012),
especially when it comes as a result of dietary manipulation, analysis of tissue
extracts would have been worthwhile, even given the difficulties of sample
extraction.
Finally, while steps were taken to eliminate confounders and bias, such as
transferring sedentary animals to an immobile treadmill and controlling for (or at
least identifying correlations with) age and susceptibility (bodymass change), by no
means was every potential confounder eliminated. Changes in bodymass and
animal health (in response to diet shock) are known to cause substantial metabolic
changes, so experiments should be designed to differentiate between these potential
confounders and a causal relationship with insulin resistance. Moreover, the
direction of the causality between that resistance and several of the changes must be
established using the aforementioned amino acid clamps, supplementation, or some
similar technique.
5.4 Estimtaing batch effects and bias
In addition to the similarities between batches, there are notable differences in what
should hypothetically be identical models. This similarity in experimental structure
(insofar as the CC-HH comparison was performed for at least some samples)
suggests that, in addition to systematic biases which may be present, there is a
substantial “batch effect” present which differentiates animals on the basis of
process defects rather than experimental treatment.
Notwithstanding our extensive precautions to standardize the acquisition and
processing of samples, including precise pH adjustment, randomization of sample
processing, subsequent randomization of sample quantification, controlled baseline
adjustment and dilution normalization, PCA analysis of all CC or HH samples from
105
the three experiments showed substantial differences, even when they were re-
normalized to the same mean concentration.
Within batches, cross validation of the PLS-DA models was repeated using 5-fold
rather than 7-fold batches to avoid problems with small sample sizes, and CV-
ANOVA was used to minimize over-fitting. Nevertheless PLS-DA models of each
batch yielded similar differential metabolites, with models trained on one batch
having much lower prediction accuracy on other batches than expected based on
cross-validation metrics. This held for models quantified with either targeted
profiling or binning.
5.4.2 Variable importance estimators
Because the number of samples used for modeling each experiment is relatively
small (the total number of samples used in all three experiments is fairly small,
compared to the complexity of the effect in question), it is possible that differences
between experimental datasets (replicates or batches, when used for comparative
analysis) are caused by insufficient coverage of population variance.
In addition to cross-validation as a means to optimize model structure, a commonly
applied technique for estimating the coverage of a dataset is resampling based
estimates of the variance in each variable. By retraining a model with a variety of
subsets from the training data, techniques such as jackknifing or bootstrapping
provide an estimation of the consistency in each model parameter: assuming the
distribution of each parameter is normal (or at least, known) the mean and variance
of each loading value can be used to calculate confidence intervals or other
significance tests for the probability that a particular loading value is non-spurious
and will be conserved in future experimental replicates.
106
Serum Samples
Spectra
Metabolite concentrations
Random
Assortment
Profiling
NMR
Batch 1 Batch 2
OPLS OPLS
Model 1 Model 2
Compare
Loadings
Reliability
estimation
Jackknifing
Model 1 with
Confidence
Figure 1
Figure 5.2 OPLS loadings comparison schema
As a baseline, each experimental dataset was divided into two subgroups
randomly (balanced for treatment), which were then modelled separately and
compared in light of the confidence estimates based on batch #1. The same
procedure was then repeated, but with each batch representing a separate
experimental dataset. The efficacy of confidence estimates in the former,
artificially segregated batches, was compared to its efficacy in true
experimental replicates.
By implementing OPLS in GNU-R with a nested resampling framework (Figure
Figure 5.2 OPLS loadings comparison schema), we calculated 95% confidence
intervals for each loading in the models constructed for each CC/HH experimental
107
dataset above. We compared the loadings from experimental batches to each other,
and also subdivided each batch into two halves and modelled each separately with
the same technique.
Figure 5.3 Loadings replication vs significance estimation for mouse replicates
Font size being proportional to the significance estimate based on batch one,
it is visible that the “volcano plot” effect visible on the right dominates over
the consistency between replicates.
While the confidence intervals were successful in predicting the value of loadings
for models build on the other “half” of the same batch, they were much less
effective at predicting the outcome of other, independent, batches. The magnitude
of the loadings had a dominating impact on their significance (Figure 5.3). As such,
we conclude that the distribution of loadings in models resampled from a particular
batch does not match the distribution of loadings in other batches, reinforcing the
existence of bias which can’t be captured by resampling for exactly this reason.
Even when extensive experimental controls are used, it is necessary to acquire
samples in multiple experimental replicates and then combine them together before
modeling in order to account for experimental drift/bias.
-0.3
-0.2
-0.1
0.0
0.1
0.2
X2.Hydroxybutyrate
X2.Oxoglutarate
X3.Hydroxybutyrate
X3.Methyl.2.oxovalerate
Acetate
Alanine
AsparagineAspartate
Benzoate
Carnitine
Choline
Citrate
Creatine
Formate
Fumarate
Glucose
Glutamate
Glutamine
Glycerol
Glycine
Hippurate
Isobutyrate
Isoleucine
Lactate
Leucine
Lysine
Methionine
O.Acetylcarnitine
Ornithine
Phenylalanine
Proline
Propylene.glycol
Pyruvate
Serine
Succinate
Taurine
Threonine
Trimethy lamine.N.oxide
Tryptophan
Tyrosine
Urea
Valine
-0.2 -0.1 0.0 0.1 0.2 0.3
Mouse B1 vs. Mouse B2 (opls:P 1/3 comps)
Batch 1 loading
Batch 2 loading
2
4
6
-0.2 -0.1 0.0 0.1 0.2
pvalue vs Loading
Rat C1 [ 3 ]
Batch 1 loading
sigval
distance
0.05
0.10
0.15
0.20
0.25
108
Chapter 6 - Metabolic profiling of vitamin C deficiency in Gulo−/− mice using
proton NMR spectroscopy
Gavin E. Duggan1, B. Joan Miller2, Frank R. Jirik2 and Hans J. Vogel1
Published: 2011-03-1 in the Journal of Biomolecular NMR, Vol 49, pp165-173
6.1 Abstract
Nutrient deficiencies are an ongoing problem in many populations and ascorbic acid
is a key vitamin whose mild or acute absence leads to a number of conditions
including the famously debilitating scurvy. As such, the biochemical effects of
ascorbate deficiency merit ongoing scrutiny, and the Gulo knockout mouse provides
a useful model for the metabolomic examination of vitamin C deficiency. Like
humans, these animals are incapable of synthesizing ascorbic acid but with dietary
supplements are otherwise healthy and grow normally. In this study, all vitamin C
sources were removed after weaning from the diet of Gulo−/− mice (n = 7) and wild
type controls (n = 7) for 12 weeks before collection of serum. A replicate study was
performed with similar parameters but animals were harvested pre-symptomatically
after 2–3 weeks. The serum concentration of 50 metabolites was determined by
quantitative profiling of 1D proton NMR spectra. Multivariate statistical models
were used to describe metabolic changes as compared to control animals; replicate
study animals were used for external validation of the resulting models. The results
of the study highlight the metabolites and pathways known to require ascorbate for
proper flux.
6.2 Introduction
Vitamin C deficiency in humans results in scurvy, a debilitating condition that is
well known historically as it affected sailors that were circumnavigating the globe.
In 1747, in what was probably the first clinical trial, the Scottish naval surgeon
James Lind discovered that lemon and orange juice could be used to prevent or
109
reverse the effects of scurvy. It took until the 1800s however for this cure to become
widely accepted (Harvie, 2005). The chemical structure of the active ingredient
vitamin C was conclusively established in 1932 (Svirbely and Szent-Györgyi,
1932). Most animals, including mice, are capable of synthesizing ascorbic acid to
satisfy their endogenous needs; it is produced through a well-defined metabolic
pathway from D-glucose and D-galactose. Guinea pigs, apes and humans however
have lost the ability to synthesize ascorbic acid, because they do not have a
functional L-gulonolactone oxidase gene (GULO), which codes for the enzyme that
catalyses the final step in the biosynthesis of vitamin C (CHATTERJEE et al.,
1961). Fortunately many fresh fruits and vegetables contain high concentrations of
ascorbic acid and hence humans and simians can acquire a sufficient supply of this
vitamin through regular consumption of appropriate foods (Nishikimi and Yagi,
1991; Naidu, 2003). It is noteworthy that humans still possess a nonfunctional
GULO pseudogene on chromosome 8p21.1, with several introns missing and
numerous insertions, and it has been estimated that most primates lost the ability to
synthesize their own vitamin C over 40 million years ago (Nishikimi et al., 1994).
Similarly, guinea pigs still contain the highly mutated remnants of an ancestral
GULO gene (Nishikimi et al., 1992).
Without access to environmental sources of vitamin C, humans suffer system-wide
failures in health; long term, severe shortages lead to scurvy, which is invariably
fatal if untreated. Some researchers have referred to this ‘hypoascorbemia’ as a
‘public inborn error of metabolism’ (Stone, 1979). Interestingly several studies have
shown that gene therapy can restore the ability of human cells to produce ascorbic
acid endogenously (Ha et al., 2004; Li et al., 2008).
High levels of vitamin C are thought to have beneficial effects on various diseases
ranging from cancer, all the way to the common cold (for a review see (Naidu,
2003)). In order to facilitate studies of the effects of vitamin C on the progression of
various diseases in mouse models, a Gulo−/− knockout strain has been developed
(Maeda et al., 2000). Like humans, this mouse strain needs to obtain vitamin C from
110
its food or drinking water. As a result, they suffer several deficiencies, including
partial loss of neutrophil function and increased oxidative stress during development
(Vissers and Wilkie, 2007; Harrison et al., 2010). By crossing Gulo knockout mice
with other disease model strains, the protective effects of ascorbate can be
investigated in a controlled environment. In this work we have used the Gulo−/−
strain to determine the effects of vitamin C deficiency on the metabolic profile of
these mice.
Proton Nuclear Magnetic Resonance (1H NMR) analysis of serum has the advantage
of quantitative, simultaneous measurements for many metabolites (Schicho et al., 0;
Weljie et al., 2007; Shearer et al., 2008; Dunn et al., 2011). This approach facilitates
interpretation and visualization of the resulting metabolic snapshot and is widely
used for mouse model studies (Griffin, 2006). The spectral area derived from each
metabolite can be segregated and attributed to individual metabolites. Targeted
profiling provides a significant advantage in metabolite identification by virtue of
correlated and corroborated grouping in peaks based on latent biochemical
knowledge. We have used the Chenomx Profiler approach (Weljie et al., 2006) to
identify specific metabolites and translate peak intensities into metabolite
concentrations using a library of spectral profiles. The resulting metabolite
concentrations provide a simplified dataset on which to perform multivariate
statistical analysis, which is easier to interpret and assess for significance.
These data were further analyzed using Orthogonal Projection to Latent Structures
(OPLS), a multivariate analysis technique that can separate the metabolic variation
in samples into noise and two classes of structured variation: patterns of change
which are correlated to the onset of vitamin deficiency, and coherent patterns of
change which are however not correlated to a particular phenomenon (Trygg and
Wold, 2002). As such, OPLS can be used to identify and remove variation between
sample batches and highlight metabolic changes of interest.
111
6.3 Experimental procedures
Mouse maintenance
Procedures were approved by the University of Calgary Animal Care and Use
Committee and abided by the Canadian Association for Laboratory Animal Science
guidelines for animal experimentation. Animals were maintained in a humidity-
controlled room with a 12-h light:dark cycle. Following weaning (3 week of age),
male Gulo−/− (KO) and Gulo +/+ (WT) mice (on a C57BL/6 background) were
maintained in microisolator cages for one week. Following this acclimation period,
all animals were switched to a vitamin C deficient diet (58R3—TestDiet, Purina,
Richmond, IN). The diet met all other nutritional requirements of adult mice. Food
and water were provided ad libitum throughout the experiment.
Animal experimentation
Animal mass was recorded on the date of vitamin C restriction and weekly
thereafter. Following 12 weeks of experimental observation, WT animals (n = 7) in
study 1 (model training) were healthy and KO animals (n = 7) showed signs of
vitamin C deficiency. WT animals in study group 2 (model testing) were harvested
after 11 (n = 4) or 19 days (n = 4). KO animals in group 2 were harvested after
19 days (n = 4). Final mass was recorded for all animals before euthanasia by
carbon dioxide inhalation (Weljie et al., 2007). Whole blood (~1 mL) was obtained
by a cardiac puncture and placed on ice and allowed to clot for 30 min. Samples
were then centrifuged for 10 min (3,000 rpm) and sera collected and divided into
two aliquots. Serum was stored at −80 °C until ready for data acquisition.
Metabolite sample preparation
Serum samples (0.2–0.3 mL) were thawed and filtered twice using 3 kDa NanoSep
microcentrifuge filters, pre-washed to reduce preservative contamination. Filtrates
were then transferred to clean microfuge tubes; final sample volume ranged from
112
150 to 250 uL. Samples were brought to 650 uL by addition of 140 uL of phosphate
buffer containing dimethyl-silapentane-sulfonate (DSS, final concentration
0.5 mM), 40 uL of sodium azide to limit bacterial growth, and D20. Sample pH was
then adjusted to a range of 7.0 ± 0.01.
Spectrum acquisition
One-dimensional 1H NMR spectra with optimal water suppression were acquired
using a standard pulse sequence (Bruker pr1d_noesy pulse definition) (Nicholson et
al., 1995). Spectra were acquired in two batches, using an automated NMR Case
sample changer on a 600 MHz Bruker Ultrashield Plus spectrometer. The pulse
sequence had a mixing time of 100 ms, a Z-gradient water presaturation pulse, and a
3 s acquisition time. Samples were individually shimmed to ensure half-height line
width of approximately 0.8 Hz for the major DSS peak calibrated to 0.0 ppm.
Spectra were acquired with 512 scans, zero-filled and Fourier transformed to 64 k
points. Standard post-processing of the spectra included deletion of the water region,
B-spline baseline correction, reference deconvolution and calibration of the DSS
peak. 1H,13C heteronuclear single quantum coherence (HSQC) spectra of two
samples, chosen at random, were acquired for peak assignment and verification.
Metabolite concentration profiling
Processed spectra were imported into Chenomx Profiler version 5 (Edmonton,
Alberta, Canada) for quantification. A standard set of fifty (50) compounds were
profiled based on chemical shift assignments verified by the 1H,13C HSQC results.
These metabolites represent an average of 84% of the total spectral area, excluding
the water, lactate and glucose areas. Spectra were randomly ordered for fitting in
Chenomx Profiler to avoid progressive bias. Compounds were fit from highest
initial concentration to lowest, with an iterative reinspection. Each measurement
was normalized to the mean sample concentration, dividing profiled spectral
concentrations by the total concentration of all profiled metabolites in that sample
(Weljie et al., 2006).
113
Multivariate statistical analysis
Normalized concentration values were imported into SIMCA-P software (Umetrics,
Sweden) for multivariate pattern analysis. Orthogonal Projection to Latent
Structures (OPLS) was used to capture artifactual differences between Batch 1 and 2
(irrespective of strain) and to compensate for batch variation. Hierarchical OPLS
was then applied to the residuals of Batch 1 samples to isolate changes between KO
and WT animals. 50 metabolite concentrations were used as X variables; animal
body mass at time of harvesting was used as the supervisory (Y) variable (Schicho
et al., 0).
Cross validation
A seven-fold rotating cross-validation (CV) was used to estimate the robustness of
the OPLS model. In each iteration, a model based on ~85% of Batch 1 samples was
used to predict the response (final body mass) for the other 15%. As a measure of
model strength (Eastment and Krzanowski, 1982), the average ratio of total sum of
squared errors (Q2Y) was compared to the percentage of Y variance captured in the
total model (R2Y). The distribution of predicted body-mass was compared to that of
actual training sample body masses.
External validation
The model constructed from Batch 1 samples were used to predict body-mass for
batch two samples. Predicted body-mass is compared to the actual strain and actual
mass of animals to determine the predictive ability of the model on external samples
not used in the model’s construction.
Model interpretaton
Primary component OPLS coefficients for the Batch 1 model, multiplied by the
weighting, were used to interpret the response pattern in affected animals (Trygg et
114
al., 2007). 95% confidence intervals were determined from CV jack-knifing and t-
value distributions (Efron et al., 1993).
6.4 Results
Animal characteristics
Wild-type animals in Batch 1 (training samples) showed no visible signs of vitamin
C deficiency after 12 wks. In contrast, Gulo−/− animals exhibited significant signs
of scurvy, including joint bruising and osteodegeneration (Hirschmann and Raugi,
1999). Batch 2 (testing) animals were harvested prior to the onset of symptoms, for
external validation purposes.
Animal body mass was reasonably consistent within strains, with a significant
difference (P < 0.05) induced between them when vitamin C was removed from the
diet (Figure 6.1). The body mass of wild-type animals in Batch 1 increased steadily
after the weaning period finished. The mass of KO mice lagged in the first month,
tailed off in the second and fell steeply in the last 4 weeks. Batch 2 KO mice already
had significantly lower body mass than WT mice when harvested after 3 weeks.
Figure 6.1: Effect of Gulo -/- knockout on mass gain
115
Statistical validation
The multivariate OPLS model of Batch 1 animals was supervised using the animals’
body mass after 12 weeks as the Y variable. Seven-fold cross validation was used,
with CV predictions of each training sample’s mass when left out of the model
construction. Overall 96.8% of the variance in body mass (Table 6.1 Cross
validation metrics for model training) could be accounted for, with 71% accuracy
(mean normalized sum of squares across cross validation iterations). Given the
relatively small number of samples in the training data set, good agreement between
predicted and actual body mass (Figure 6.2) indicated a strong model of the
metabolic changes between WT and KO mice when vitamin C was withheld.
Model Samples R2Y (Coverage) Q2Y (Quality)
KO vs WT (Y=mass) Batch 1 96.8% 71% (very good) Batch differences All 96.5% 96% (excellent)
Table 6.1 Cross validation metrics for model training
Figure 6.2: Cross-validation predicted vs actual body mass for batch 1 mice.
Average predicted (white) vs actual (dark) bodymass. Predictions were made for samples left out of each CV model. Intraclass overlap and interclass separation are a measure of model quality. One sample, believed to be a knockout, fell between the expected range for both classes. Re-typing showed it to be an unexpected heterozygote
116
External validation
A principle components analysis (PCA) model of all samples showed that the
differences between Batches 1 and 2 were more significant than differences in KO
and WT mice (Supplementary Figure S1). This is quite common in metabolomics
studies, where sensitive techniques such as serum NMR-profiling can detect
differences in animal environment (Bollard et al., 2005) but in this case is caused by
differences in animal age. An OPLS model of inter-batch differences showed
excellent quality measures (Table 6.1), indicating the model’s ability to put
Figure 6.3: WT vs KO hierarchical OPLS scores
WT (black) vs KO (white) OPLS model scores of modeling samples after removal of intrabatch variation using hierarchical OPLS. First component (horizontal) score indicates the degree of vitamin C deficiency perceived by the model. The one modeling sample (HET/Triangle) misclassified by the model was the heterozygous mutant (See Fig. 2)
117
samples on a common basis for comparison, independent of age (Figure 6.3). Based
on these analyses, the Batch 1 model was able to significantly (P < 0.05, Figure 6.4)
differentiate WT and KO mice in Batch 2—pre-symptomatic early harvesting of
Batch 2 samples notwithstanding.
Figure 6.4: External validation results
Predicted (light) and actual (dark) masses for batch 2 samples not used in
any model construction. Standard error bars show good separation between
predictions for samples, notwithstanding the fact that they were harvested
pre-symptomatically
Model interpretation
OPLS coefficients were multiplied by the scaling weights (Trygg et al., 2007) for
interpretation of the metabolite changes between strains. Jackknifing was used to
calculate 95% confidence intervals, and only metabolites with significantly
increased or decreased concentrations are reported (Figure 6.5). Several groups of
metabolites with similar biochemical functions (Kanehisa et al., 2010) were seen to
respond in concert.
118
Figure 6.5: First component OPLS loadings of WT vs KO mice
Values indicate relative change in compounds in knockout mice after
12 weeks of vitamin C deficiency
119
6.5 Discussion
An estimated 15–30% of low-income North Americans suffer from varying degrees
of vitamin C deficiency (Frikke-Schmidt and Lykkesfeldt, 2009), of which even
mild shortfall undermines the body’s ability to defend against damaging oxidizing
byproducts. Vitamin C deficiency has been implicated in joint trauma (Padh, 1990),
heart disease (Heinig and Johnson, 2006; Cangemi et al., 2007), cancer (Carroll and
Kritchevsky, 1994), as well as motor control and degenerative conditions
(Sauberlich, 1994). Linus Pauling proposed that high doses of vitamin C would
prevent the common cold, a claim that has been hotly debated (Harri and others,
2009), and some investigators have even suggested that infusion of extremely high
doses of vitamin C may have beneficial effects as a prodrug in the treatment of some
cancers (Chen et al., 2005, 2008; Cabanillas, 2010). Its popular appeal as a panacea
may be grounded in some degree of biochemical fact since ascorbic acid is a known
cofactor for over a dozen biological pathways (Padh, 1990), and a prominent
antioxidant in many others. Vitamin C is one of the nutrients being studied in the
European Prospective Investigation into Cancer and Nutrition (EPIC) population
study (e.g. (Gonzalez and Riboli, 2010)).
Its significance notwithstanding, the mechanisms by which vitamin C affects
various conditions have not been fully characterised (Naidu, 2003). Broad-spectrum
metabolic state profiling, in conjunction with known networks of metabolic
functionality, gives a contextualized view of changes and the Gulo knockout mouse
provides an ideal platform for the study of vitamin C deficiency (Y et al., 2003;
Harrison et al., 2010). The ability of this mutant to be crossed with other disease
models also provides abundant avenues for further investigation. Herein, we have
used 1D proton NMR of serum from Gulo−/− mice to study the differences in
metabolism induced when vitamin C was withheld after weaning.
Because direct analysis of NMR spectra can be difficult, 64,000 spectral intensity
points for each serum sample were translated to corresponding concentrations for 50
120
metabolites. This translation to named compounds prior to analysis was
instrumental in model interpretation, refinement, and establishment of significance
in differences between knockout and wild-type animals. In this study, the animals’
body mass was used as a supervisory measure of signs of vitamin C deficiency
(Figure 6.1). While the training cohort was small, testing of the models supported
their validity as the predicted values were in good agreement with actual values for
Batch 2 samples (Figure 6.2). Moreover, despite the early harvesting of the external
validation samples, prior to the onset of symptoms, the model showed a significant
(P < 0.05) ability to separate KO and WT samples from the second batch of unseen
animals (Figure 6.4). Taken together with the confidence intervals for the metabolite
changes, we have evidence that the model accurately describes the changes induced
by vitamin C shortfalls and not other phenomena or random variation.
The model loadings (Figure 6.5) reflect the significant changes between WT and
KO mice. The difficulty of interpretation stems from the complexity of the
interactions within the animal systems; no single metabolite can both capture and
segregate the multiple responses to various oxidative stresses. Some are ubiquitous
indicators of non-specific pathology, while others are involved in numerous
pathways. By viewing changes in the context of the total response the loadings can
highlight specific pathway segments of interest. In addition to literature sources, the
Kyoto Encylopedia of Genes and Genomes (Kanehisa et al., 2010) and the Human
Metabolome Database (Wishart et al., 2007) were used to identify functional groups
of metabolites with significant changes. Several functional groups were associated
with changes in vitamin C availability, which together highlight major shifts in
metabolism: an increase in glutathione production to compensate for ascorbic acid’s
loss, and/or an increase in glycerophospholipid metabolism for either energetic or
precursor-provision reasons.
Energy regulation (Pyruvate, Acetate, Oxaloacetate, Succinate, Fumarate)
121
The TCA cycle’s central role in energy metabolism features prominently in many
metabolomic profiles, and central TCA players can be affected by so many
responses (e.g. injury, diet, uptake, repair) that direct attribution is often difficult.
The known role of ascorbate in activating prolyl 4-hydroxylase to increase succinate
creation from 2-oxoglutarate raises some questions about succinate’s perceived
increase despite the decreasing levels of fumarate, pyruvate, and oxaloacetate. This
may imply an ascorbate-shortage induced shunt from either pyruvate or succinate
(or both) via alanine (which demonstrated a significant increase) and aspartate.
While aspartic acid exhibited no consistent increase in concentration there was a
large increase in variance indicating possible increases in flux; the changes in both
lysine and phenylalanine levels could be attributed to disruption of aspartic acid
production from alanine and oxaloacetate.
Carnitine biosynthesis
Two of the most significant and coherent changes seen are the decrease in carnitine
levels and an increase in lysine. Carnitine can act as a transporter of fatty acids into
the mitochondria, and the twofold role of ascorbate in its synthesis is well
established (Padh, 1990). Ascorbic acid acts as a cofactor for hydroxylation of
trimethyl-lysine and butyrobetaine, both precursors for the production of carnitine.
The direct implications of this can be seen in the significant drop in carnitine and O-
acetylcarnitine in KO mice, concurrent with a significant increase in lysine and
betaine. Whether it is a cause or effect requires further investigation: the resulting
shift in aspartate-to-lysine conversion would result in increased flux towards
cysteine and glutathione production, or an increase in lipid metabolism could result
from other oxidative stress responses (Hopps et al., 2010) (see below).
Glutathione synthesis
As part of the GSH/ascorbate redox chain the significant change in glutathione
levels is of immediate interest. In the absence of dehydroascorbic acid, GSSG
conversion to GSH is impeded and the normal cycle of reductive protection is cut
122
off from the terminal outlet of vitamin E quinones (Winkler et al., 1994). While
increased production in GSH would eventually be overwhelmed it could serve as a
short-term compensatory measure, and a number of metabolic shifts support the
possibility of increased glutathione production in these knockout mice.
Cystathionine-β-synthase (CBS) is the committed enzyme step in production of
cysteine from methionine, and is regulated by the oxidative state of its heme group.
The decrease in methionine levels supports the notion that an absence of ascorbate
results in oxidation of CBS and upregulation of cysteine production, one of two
required precursors for the increased production of glutathione. The other precursor
for glutathione production, glycine, is also seen to increase while serine (its
precursor) decreased significantly (Figure 6.6). Significant depletion of serine is
also supported by a potential increase in glycerophospholipid metabolism (via
conversion to O-phosphatidylserine, see below).
123
Figure 6.6 KEGG pathway diagram for serine and glycine metabolism
Glycerophospholipid metabolism
The additional role of ascorbate as a primary regulator of glycerophosphocholine-
cholinephosphodiesterase enzymes (Sok, 1998) may explain the differential levels
of choline, phosphocholine, and glycerol in our results. If the loss of ascorbate’s
stabilizing action on GPC-cholinediesterase gave rise to an increase in free choline,
the corresponding increase in betaine would feed the glutathione production above.
It seems reasonable that the concomitant release of diacylglycerols into the cell may
be an evolutionary response. DAG is a known secondary messenger, activating
124
protein kinase C. Moreover, release of those free lipids may be related to the reports
of insulin resistance in some cases of long term, mild, ascorbate deficiency. Further
investigation of intermediate energy metabolism using other techniques (e.g.:
lipidomics, hyperinsulinaemic clamp), might be used to distinguish between these
possibilities. Regardless, it suffices to say that the implications of low ascorbate
levels on lipid metabolism seem both profound and far-reaching.
The one metabolite of high significance that cannot be directly ascribed to any
particular ascorbate-related reaction within the scope of this study is 1,3-
dimethyluric acid. The importance of uric acid as a secondary antioxidant (Ames et
al., 1981), particularly in primates that lack the GULO gene (Sevanian et al., 1991),
suggests a number of possible functions however currently no literature linking the
two phenomena is available. A number of dimethylxanthine-related interactions
(Lee, 2000) have been described in the literature as having ascorbic acid
associations but in light of the numerous possibilities there is insufficient evidence
in these results to extrapolate. The lack of significant change in some related
metabolites, such as allantoin, further complicates the issue.
6.6 Conclusions
Our results highlight the ability of the Gulo knockout mouse to present a clear
picture of the metabolic changes induced by vitamin C deficiency, and of 1D proton
NMR to capture those changes. Quantitative profiling simplified the NMR results to
a manageable set of identified metabolites, and PLS was used to identify coherent
patterns (statistical models) of change between wild-type and KO mice.
Orthogonalization of those PLS loadings clarified the results and allowed
interpretation of the models with respect to vitamin C deficiency induced changes.
They also allowed comparison of model loadings between two sample batches, and
validation of the resulting models by virtue of accurately predicting the condition of
unknown animals. Finally, our results are in good agreement with several
125
metabolites known to require ascorbate for enzymatic activation and proper flux.
Almost all of the metabolic changes seen in the animal can be attributed to a few
pathways of known importance, which provides a coherent picture of how the
absence of ascorbate’s reducing action upregulates production of other antioxidants,
primarily glutathione, and possibly uric acid. The links between those pathways and
the perceived changes underscores the ability of NMR-based metabolomics to
capture the impact of complex diseases in a quantitative way.
126
Chapter 7 Conclusions and Future Work
Metabolomics as a field continues to evolve to meet the expectations of clinical,
pharmaceutical, biochemical and agricultural interests. This rapid growth can be
seen in the development of novel applications for metabolic profiling in the last five
years, but also in the frequent questions raised by end users about the interpretation,
implications and limitations of past results.
In this early, developmental phase questions answered with metabolomics
techniques often present opportunities for further study, new biochemical
hypotheses or variations on technical themes which might improve the information
available. These novel avenues for exploration have comprised the crux of the work
performed herein, both in terms of specific innovation but more importantly in
addressing the larger concept that an “engineered” approach to experimental design
may be both possible and appropriate to a given biological problem.
There is a saying about those who fail to study history, and thus a proper
examination of metabolomic techniques began with a baseline experiment on a well-
established system of interest. Dr Shearer’s work on the medically pressing issue of
insulin resistance, using mouse models to control as many factors as possible,
generated high quality serum samples. Analysis via 1H NMR (with select HSQC
spectra to calibrate our chemical assignment) provided successful insight into the
power of metabolomics. Unseen in the published results were the various
incremental steps required to optimize acquisition procedures, including pulse
sequence, relaxation times, sample concentration, and proper methods for targeted
profiling. The resulting data was an early example of physiological experimentation
which built on early work by Dr Weljie and has since been expanded on by us and
others (Zhao et al., 2010; Kim et al., 2011). The significance of branched-chain
amino acids, a previously known class of signalling metabolites (Newgard et al.,
2009), was the key biochemical assertion made as a result, which was later held up
in our other studies.
127
Incidental to the biochemical implications, several other phenomena became
apparent during the unfolding investigative process. To calibrate the profiling
process better, the same set of samples were profiled multiple times and the results
compared. The variance revealed the existence of a significant acclimatization
and/or training period in each individual’s use of the Chenomx software, and a
heteroskedastic nature of human error in the profiling. The latter is likely an
unavoidable consequence of the univariate scaling technique used: a consistent error
variance in the profiling is exacerbated when the variances are multiplied by the
inverse of the metabolite’s average concentration. This became relevant in later
chapters looking at the use of variance-based estimation techniques.
Even with this early study design though, limitations were readily apparent.
Reviewers noted the potentially confounded short and long term effects of diet, and
the presence of energy metabolites’ responses in any kind of health-affecting
intervention is established in the literature. The diet switch study described in
chapter three showed that the aforementioned amino acids, with established
physiological signalling implications, responded similarly regardless of the final
week’s diet and the normalization of glucose levels, liver triglycerides. This was in
contrast, as suspected, to the changes in immediate energy metabolites, which fell
off-axis in the SUS plot comparing regular- and highfat-chow fed metabolic
changes.
Notwithstanding its well established use in other fields, there have been few
metabolomics studies using similar controls (Miccheli et al., 2009; Scalbert et al.,
2009; Zivkovic and German, 2009; Westerhuis et al., 2010). The complexity of
normal multivariate results is compounded, exponentially, by the introduction of
four different sample treatments. Because no linear order exists between treatments,
a simple OPLS model can’t capture both relationships, but the SUS-based approach
presents a simple and intuitive way to present the results. By using two models and
then juxtaposing the two related phenomena in a graphical form, the relationships
between samples are easy to grasp.
128
The concept of data “purity”, protected by proper controls and unadulterated by
other effects, can be augmented and complemented by improving an experiment’s
perspective and depth. Once the data is focused to the condition under study, the
quality can still be improved by taking metabolomics beyond the realm of simple
case-controls.
In the case of insulin resistance, a phenomenon with astronomical cost implications
and diabolical complexity (Bain et al., 2009), diagnostic potential is undeniably
important but arguably secondary to insight about the physiological mechanisms
and interactions. The effects of exercise are known to interact with both causes and
effects of insulin resistance in complex ways (Atalay and Laaksonen, 2002), making
it an interesting subject for two-factor analysis. If metabolomics could identify and
isolate those effects, and interactions, of exercise and high-fat feeding then it could
provide a perspective unavailable to one-factor experiments.
When analysing the resulting metabolite concentration matrix, a PLS model with
both feeding and exercise as Y-variables gave a serendipitous model but the contrast
of two one-factor OPLS models was clearer (Figure 4.1a). By separating the factors
as seen in chapter 4, this technique leveraged the experimental cohorts better in the
same way that analysis of age- or gender-paired controls provides contrast analysis.
Doing so also establishes an important continuity and context, allowing for easier
comparison with other experiments.
Biochemically, the initial experiments with HF-diet induced changes highlighted
energy metabolism and signalling as major shifts associated with the diet, which the
non-metabolomic glycemic clamp technique suggests also caused insulin resistance
in the susceptible mice. The two-factor analysis shows that, while the changes some
energy metabolites persist when exercised, the hyperglycaemia, the changes in
signalling amino acids, and shifts in other energy metabolite were ablated. It’s
worth noting that the significance estimates for the loadings in this study were
generated with a jackknife-based technique similar to that studied in chapter five.
129
They should therefore be viewed with a degree of skepticism. Nevertheless, they do
show a degree of biochemical logic which is encouraging.
Given the assumption that exercise mitigates the onset of insulin resistance, the
results suggest a potential method of activity at the regulatory level, and highlight a
potential difficulty in solely glucose-based non-metabolomic diagnostic measures.
Considering the relative ease of manipulating exercise and dietary fat content in
humans, this suggests that a potential next step would be human studies using
calibrated diet or health measures. While withholding treatment in humans is often
an ethical non-starter, the impact of controlling caloric intake and fat/protein
balance is more acceptable. The obvious, and unavoidable, limitation is the long
onset time of any insulin-resistance related condition, but it would appear that
metabolomics investigation could extract meaningful data from the non-pathological
initial changes seen in a months-scale crossover timeline. Unfortunately, the strong
glucose-screening effects of the kidneys make the easier urine-based studies
unlikely to bear useful information in this case.
Interesting alternate applications of the same methodology could include the two
factor analysis of antibiotic treatments in sepsis models(Izquierdo-García et al.,
2011; Lacy, 2011; Stringer et al., 2011) or plaque formation in Alzheimer’s
(Irizarry, 2004; Barba et al., 2008). These types of pathological conditions with
unknown etiology and muddled biochemistry benefit from the contrast provided by
a second factor. Moreover, both conditions are amenable to either simple case-
control structure or a more sophisticated dose/response model using treatment
magnitude as the supervisory variable.
In general, it seems the suitability of a particular condition to metabolomic analysis
will likely depend on the breadth of biochemical disruption caused by the
experimental manipulation. Treatments or conditions which induce changes in
immune response, primary energy, or other generic systems will require adaptation
to clarify the resulting interpretation. Two factor analysis is one potential for doing
130
so, however the use of PLS is not without its drawbacks. One difficulty which arose
in comparing HF-induced changes in sedate and exercised animals was the scaled
nature of the loadings resulting from the OPLS (or any PLS-derived) model. The
use of weighted coefficients aids interpretation by controlling the spread of the
loadings over a fixed range, but renders them harder to compare because absolute
values are unknown. Instead, the comparison has to be done in a geometric context,
with the relative slope of point clusters providing context for other changes.
The same drawback became more significant when the comparison of multiple
mouse/diet models was performed in chapter 5. While the orthogonalization of the
model into single predictive components makes it easier to compare studies, the
different numerical ranges spanned by loadings still led to the possibility of non-unit
“slope” appearing between models. Such a trend would render either angle- or
distance-from-identity metrics of conservation invalid, and hence may be a reason
why variance-based estimates of significance were less accurate than expected.
Unfortunately, the scattering of a small number of metabolites at the extremes of the
loading ranges is usually insufficient to estimate the slope effect, if it exists.
Possible solutions would be further application of median-based slope estimators, or
forcing the loadings into the same range, but doing so would require significant
validation as a method. It may also be possible that the use of weights to scale
loadings may render the different loadings ranges meaningful and/or consistent.
Investigating the latter would still require investigation, but could be an important
step towards the accurate interpretation of OPLS models. Doing so would likely
require a properly designed control experiment with known sample variances, but
could be done on inexpensive bacterial studies.
Other possible reasons abound for the unexpected failure of jackknifing as a
significance predictor. Since it appears unlikely that practitioners will (or even
should) cease the direct interpretation of PLS loadings, their relative interpretability
being the primary reason for their use, further investigation of both jackknifing and
131
other significance estimators remains an important area of methodological research.
Extending the comparison of batches to projection coefficients from PLS, as they
pertain to each Y variable, should be a relatively simple and immediate next step; it
may prove that variance eliminated from the predictive components by
orthogonalization is damaging the ability of jackknifing, and that simple PLS in fact
yields better results. In general, the interpretation of PLS-based terms is
insufficiently understood by metabolomics practitioners.
In addition to being our first use of multi-factor metabolomics, the study of exercise
and high fat diet marked our first effective application of network analysis tools
such as the IPA™ suite, from Ingenuity Systems software(Ingenuity, 2011), and
KEGG pathway diagrams (Kanehisa et al., 2010). Because energy metabolites are
directly implicated in insulin-related pathology, and tightly coupled with many
processes at the cellular and systematic level, the pathway analysis didn’t provide
significant benefit in this case. Its subsequent application to the vitamin C study in
chapter 6 however, yielded much greater insight. While the postulated method of
interaction of vitamin C deficiency with the regulation of glutathione synthesis is
only a hypothesis, there was significant literature support for such a relationship.
Such an interpretation would not have been possible without a graphical network-
based perspective, which is increasingly becoming a minimum, and mandatory,
element of any serious metabolomics investigation of biochemistry.
The interaction between network analysis and significance prediction occurs at
multiple levels. While good numerical estimates of significance can aid
interpretation of the results in a network context by paring down the relevant
entities, the biochemical implications and connections contribute a large margin to
the final confidence in the results. Metabolite concentration shifts which were not
flagged as numerically significant can be interpreted with increased confidence
when they’re bracketed by complementary changes in network neighbours. In the
same way that Chenomx’s targeted profiled uses a priori molecular structural
knowledge to enhance the interpretation of the binned spectra, biochemical structure
132
data can enhance interpretation of multivariate results. The IPA™ suite represents a
first generation attempt to automate the application of this information, but as yet is
still very raw.
One area of research which represents a major frontier for metabolomics in the
future, and on which such next-generation biochemical interpretation tools will
likely be based, is the integration of data from multiple sources. A manual
incarnation of this integration can be seen in several of the studies presented here:
the antibody work, glycaemic clamps, and additional measures such as non-
esterified fatty acids (NEFA) gathered by Dr. Shearer gave important supporting
evidence to the multivariate metabolomics results. Automating this process is an
oft-discussed goal for metabolomics (Westerhuis et al., 2010), whether it’s single
additional variables such as this, or entire independent multivariate platforms such
as mass spectrometry (Dettmer et al., 2007; Amtmann and Armengaud, 2009;
Connor et al., 2010).
The naïve approach of concatenating input matrices discards far too much
information about the source properties of the data, but hierarchical multi-block PLS
presents an established tool which is more likely to accommodate the differences.
Hierarchical modeling can present a danger of over-fitting, further reinforcing the
need for effective confidence estimation at the base level, but also presents several
interesting possibilities. The use of residuals from the base models can be used to
factor out known issues (Schicho et al., 2010), while the use of base-scores as
variables for a top level model could theoretically integrate data from diverse
sources, effectively normalising and balancing each source. Unfortunately, at this
time there doesn’t appear to be any research into “carrying” variable confidence
from base models to higher orders.
Several other options exist for addressing the problem of integrating multisource
data, and developing techniques for application in metabolomics specifically will be
much more an exercise in statistical literature mining and testing than novel
133
methodological research. Examples include block-scaling, an intermediate option
which is offered by the commonly used SIMCA software package, but which is as
yet unexplored in a metabolomics context. As always, a properly structured
estimation for concentration distributions, trained with probabilistic integration
methods, represents the most flexible and powerful option but would require
significant development work and potentially larger sample sizes. One significant
advantage to such an approach would be the potential for relatively direct
integration of both molecular structure and biochemical network information,
making numerical and biological modeling an interactive two-way street.
Time Series: The Future of Metabolomics
Having established methods for improving sample manipulation, acquisition,
comparison, and interpretation, there remains potential for improvement at every
stage of metabolomics pipeline. A major frontier and overarching theme to many
potential improvements is the concept of time series analysis. Samples which can
be compared longitudinally to others taken from the same biological context
(individual, animal, etc), prior to manipulation, offer the strongest form of controls
in a multivariate setting. Time series have been an increasingly important part of
the metabolomics literature and will likely continue to grow in significance for
several reasons.
In humans, the potential for variation in environmental exposure, dietary variation,
and genetic background is well established. Not all of this variation can be
eliminated or calibrated by comparison to a baseline, but the ability of self-controls
to focus on the changes relative to treatment in a particular window increases both
the purity and focus of the data. Moreover, the animal samples examined herein
have shown that even genetically controlled model organisms such as mice are
subject to variations in infection, dietary consumption, and other random factors
which differentiate individuals. Severity of onset in manipulated pathology, and the
benefits of pharmalogical interventions, are likewise unpredictable and among the
134
most significant potential areas of systems biology application, so time series
analysis would likely benefit all such studies.
Acquiring and processing time series data will require, or at least benefit from, a
number of innovations or engineering improvements. At the sample collection and
instrumental acquisition level, the ability to analyse smaller volumes will be an
important development. Mass spectrometry presents an appealing sensitivity on this
front, but smaller bore NMR probes or capillary-NMR may eventually provide
complementary means to analyse small volume samples. The ability to quantify
samples in the 5-10uL range would facilitate both human and animal experiments.
In murine studies, it could allow temporal perspectives of the same animal by
collecting a non-fatal serum volume at each time point. In humans, the
physiological effects of repeated blood draw could be minimized, as well as the
application of metabolomics to cerebrospinal fluid from lumbar punctures, and
potentially the development of a portable, continuous-sampling apparatus. Such a
device could be derived from the insulin pump technology currently in public use,
inverted to provide an intimate perspective into human activity. Beyond
establishing a baseline, continuous sampling would allow analysis at a much higher
temporal resolution. This could eliminate, or at least capture and calibrate, many of
the problematic sources of variation such as food and drug metabolism, diurnal
biological cycles, and shifts caused by changes in energy expenditure.
From a numerical perspective, tools exist for the analysis of time-series data,
however additional assumptions are often required about the nature (transient,
periodic, monotonic) of the trends involved. Moreover, training and testing will be
required to validate their application in metabolomics data, and the results will once
again need to be translated for consumption by end users. Visualization, whether
customised or standardized, will likely remain the most powerful tool available for
doing so.
135
One final area of investigation which would benefit from time series analysis, but
remains a distant goal for metabolomics is that of measuring biochemical flux
within both cells and individuals. Very little quality work (Lee et al., 2006; Zamboni
and Sauer, 2009) has been done on the measurement of flux by metabolomic
methods, but it remains an important stepping stone for both biochemical
investigation and pharmacodynamics efforts. It may be that additional work is
required to clarify the relationships between cellular mechanics and biofluid-level
effects, before assertions can be made about flux based on metabolomics
measurements, but a higher temporal resolution between sampling would greatly
increase their utility.
136
Closing Thoughts
As an insider, it often seems that metabolomics has had a difficult adolescence,
failing to gain traction with external audiences. While there have been numerous
applications, the integration with other types of analysis has been slow coming.
Moreover, there is no clear roadmap for the integration of metabolomics with
clinical or pharmaceutical efforts. Achieving that acceptance will be as much
political as scientific, but a foundation of clearer, more reliable data will be
necessary.
With some more reflection however, it’s clear that in chasing its potential as a
method the field’s growth has been even more accelerated than other areas of
systems biology; the growing pains of the past decade are to be expected. With this
work, we’ve showed that metabolomics is comprised of a pipeline with the potential
for improvements at each step. By adapting established experimental structures to
metabolomics, it’s clear that proper controls and well thought out comparisons yield
robust results and clearer biological insight. Nevertheless, the need for continued
research into data processing, confidence estimation, and visualization techniques
remains clear. Combining those innovations will invariably lead to more efficient
diagnostic techniques, more insight into disease causes, greater acceptance of the
results, and connections between the metabolite pipeline and those of other systems
biology fields. Only then will metabolomics have fully matured as an investigative
field.
137
Chapter 8 Bibliography
Abate-Shen, C., and Shen, M.M. (2009). Diagnostics: The prostate-cancer metabolome. Nature 457, 799–800.
Åberg, K.M., Alm, E., and Torgrip, R.J.O. (2009). The correspondence problem for metabonomics datasets. Anal Bioanal Chem 394, 151–162.
Addelman, S. (1969). The Generalized Randomized Block Design. The American Statistician 23, 35–36.
Adibi, S.A. (1968). Influence of dietary deprivations on plasma concentration of free amino acids of man. J Appl Physiol 25, 52–57.
Alam, T.M., Alam, M.K., McIntyre, S.K., Volk, D.E., Neerathilingam, M., and Luxon, B.A. (2009). Investigation of Chemometric Instrumental Transfer Methods for High-Resolution NMR. Anal. Chem. 81, 4433–4443.
Alevizos, I., Misra, J., Bullen, J., Basso, G., Kelleher, J., Mantzoros, C., Stephanopoulos, G., and others (2007). Linking hepatic transcriptional changes to high-fat diet induced physiology for diabetes-prone and obese-resistant mice. CELL CYCLE-LANDES BIOSCIENCE- 6, 1631.
Alsberg, B.K., Kell, D.B., and Goodacre, R. (1998). Variable selection in discriminant partial least-squares analysis. Analytical Chemistry 70, 4126–4133.
Amantonico, A., Urban, P.L., and Zenobi, R. (2010). Analytical techniques for single-cell metabolomics: state of the art and trends. Anal Bioanal Chem 398, 2493–2504.
Ames, B.N., Cathcart, R., Schwiers, E., and Hochstein, P. (1981). Uric acid provides an antioxidant defense in humans against oxidant- and radical-caused aging and cancer: a hypothesis. Proc. Natl. Acad. Sci. U.S.A 78, 6858–6862.
Amtmann, A., and Armengaud, P. (2009). Effects of N, P, K and S on metabolism: new knowledge gained from multi-level analysis. Current Opinion in Plant Biology 12, 275–283.
Anderson, P.E., Mahle, D.A., Doom, T., Reo, N.V., DelRaso, N.J., and Raymer, M.L. (2010). Dynamic adaptive binning: an improved quantification technique for NMR spectroscopic data. Metabolomics.
Anderson, P.E., Reo, N.V., DelRaso, N.J., Doom, T.E., and Raymer, M.L. (2008). Gaussian binning: a new kernel-based method for processing NMR spectroscopic data for metabolomics. Metabolomics 4, 261–272.
Bantscheff, M., Schirle, M., Sweetman, G., Rick, J., and Kuster, B. (2007). Quantitative mass spectrometry in proteomics: a critical review. Analytical and Bioanalytical Chemistry 389, 1017–1031.
Barba, I., Fernandez‐Montesinos, R., Garcia‐Dorado, D., and Pozo, D. (2008). Alzheimer’s disease beyond the genomic era: nuclear magnetic resonance (NMR) spectroscopy‐based metabolomics. Journal of Cellular and Molecular Medicine 12, 1477–1485.
Barnard, D.E., Lewis, S.M., Teter, B.B., and Thigpen, J.E. (2009). Open- and Closed-Formula Laboratory Animal Diets and Their Importance to Research. Journal of the American Association for Laboratory Animal Science : JAALAS 48, 709.
Bathe, O.F., Shaykhutdinov, R., Kopciuk, K., Weljie, A.M., McKay, A., Sutherland, F.R., Dixon, E., Dunse, N., Sotiropoulos, D., and Vogel, H.J. (2011). Feasibility of Identifying Pancreatic Cancer Based on Serum Metabolomics. Cancer Epidemiol Biomarkers Prev 20, 140–147.
Beckwith-Hall, B.M., Brindle, J.T., Barton, R.H., Coen, M., Holmes, E., Nicholson, J.K., and Antti, H. (2002). Application of orthogonal signal correction to minimise the effects of physical and biological variation in high resolution 1H NMR spectra of biofluids. Analyst 127, 1283–1288.
Beckwith-Hall, B.M., Nicholson, J.K., Nicholls, A.W., Foxall, P.J.D., Lindon, J.C., Connor, S.C., Abdi, M., Connelly, J., and Holmes, E. (1998). Nuclear Magnetic Resonance Spectroscopic and Principal Components Analysis Investigations into Biochemical Effects of Three Model Hepatotoxins. Chem. Res. Toxicol. 11, 260–272.
Bell, J.D., Sadler, P.J., Morris, V.C., and Levander, O.A. (1991). Effect of aging and diet on proton NMR spectra of rat urine. Magn. Reson. Med. 17, 414–422.
139
Von Bertalanffy, L. (1950). An outline of general system theory. British Journal for the Philosophy of Science 1, 134–165.
Bertone, P., Stolc, V., Royce, T.E., Rozowsky, J.S., Urban, A.E., Zhu, X., Rinn, J.L., Tongprasit, W., Samanta, M., Weissman, S., et al. (2004). Global Identification of Human Transcribed Sequences with Genome Tiling Arrays. Science 306, 2242–2246.
Bijlsma, S., Bobeldijk, I., Verheij, E.R., Ramaker, R., Kochhar, S., Macdonald, I.A., Van Ommen, B., and Smilde, A.K. (2006). Large-Scale Human Metabolomics Studies: A Strategy for Data (Pre-) Processing and Validation. Anal. Chem. 78, 567–574.
Bino, R.J., Hall, R.D., Fiehn, O., Kopka, J., Saito, K., Draper, J., Nikolau, B.J., Mendes, P., Roessner-Tunali, U., Beale, M.H., et al. (2004). Potential of metabolomics as a functional genomics tool. Trends in Plant Science 9, 418–425.
Bock, J.L. (1982). Analysis of serum by high-field proton nuclear magnetic resonance. Clinical Chemistry 28, 1873–1877.
Bollard, M.E., Stanley, E.G., Lindon, J.C., Nicholson, J.K., and Holmes, E. (2005). NMR-based metabonomic approaches for evaluating physiological influences on biofluid composition. NMR Biomed. 18, 143–162.
Bolstad, B.M., Irizarry, R.A., Åstrand, M., and Speed, T.P. (2003). A comparison of normalization methods for high density oligonucleotide array data based on variance and bias. Bioinformatics 19, 185–193.
Bonora, E., Targher, G., Branzi, P., Zenere, M., Saggiani, F., Zenti, M.G., Travia, D., Tonoli, M., Muggeo, M., and Cigolini, M. (1996). Cardiovascular risk profile in 38-year and 18-year-old men. Contribution of body fat content and regional fat distribution. Int. J. Obes. Relat. Metab. Disord 20, 28–36.
Bottomley, P. (1985). Noninvasive study of high-energy phosphate metabolism in human heart by depth-resolved 31P NMR spectroscopy. Science 229, 769–772.
Bouillon, R., Bex, M., Herck, E.V., Laureys, J., Dooms, L., Lesaffre, E., and Ravussin, E. (1995). Influence of age, sex, and insulin on osteoblast function: osteoblast dysfunction in diabetes mellitus. JCEM 80, 1194–1202.
Boulé, N.G., Haddad, E., Kenny, G.P., Wells, G.A., and Sigal, R.J. (2001). Effects of exercise on glycemic control and body mass in type 2 diabetes mellitus. JAMA: The Journal of the American Medical Association 286, 1218–1227.
Brenner, R.R., Bernasconi, A.M., and Garda, H.A. (2000). Effect of experimental diabetes on the fatty acid composition, molecular species of phosphatidyl-choline
140
and physical properties of hepatic microsomal membranes. Prostaglandins, Leukotrienes and Essential Fatty Acids 63, 167–176.
Brindle, J.T., Antti, H., Holmes, E., Tranter, G., Nicholson, J.K., Bethell, H.W.L., Clarke, S., Schofield, P.M., McKilligin, E., Mosedale, D.E., et al. (2002). Rapid and noninvasive diagnosis of the presence and severity of coronary heart disease using 1H-NMR-based metabonomics. Nature Medicine 8, 1439–1445.
Broadhurst, D., Goodacre, R., Jones, A., Rowland, J.J., and Kell, D.B. (1997). Genetic algorithms as a method for variable selection in multiple linear regression and partial least squares regression, with applications to pyrolysis mass spectrometry. Analytica Chimica Acta 348, 71–86.
Broadhurst, D.I., and Kell, D.B. (2006). Statistical strategies for avoiding false discoveries in metabolomics and related experiments. Metabolomics 2, 171–196.
Brooks, D.C., Bessey, P.Q., Black, P.R., Aoki, T.T., and Wilmore, D.W. (1986). Insulin stimulates branched chain amino acid uptake and diminishes nitrogen flux from skeletal muscle of injured patients. Journal of Surgical Research 40, 395–405.
Brown, M., Dunn, W.B., Ellis, D.I., Goodacre, R., Handl, J., Knowles, J.D., O’Hagan, S., Spasić, I., and Kell, D.B. (2005). A metabolome pipeline: from concept to data to knowledge. Metabolomics 1, 39–51.
Brownlee, M., and Cerami, A. (1981). The Biochemistry of the Complications of Diabetes Mellitus. Annual Review of Biochemistry 50, 385–432.
Cabanillas, F. (2010). vitamin c and cancer: what can we conclude-1,609 Patients and 33 Years later? PR Health Sciences Journal 29,.
Calles-Escandon, J., Cunningham, J., and Felig, P. (1984). The plasma amino acid response to cafeteria feeding in the rat: influence of hyperphagia, sucrose intake, and exercise. Metabolism 33, 364–368.
Canet, D. (1996). Nuclear magnetic resonance: concepts and methods (Chichester ; New York: Wiley).
Cangemi, R., Angelico, F., Loffredo, L., Del Ben, M., Pignatelli, P., Martini, A., and Violi, F. (2007). Oxidative stress-mediated arterial dysfunction in patients with metabolic syndrome: Effect of ascorbic acid. Free Radic. Biol. Med 43, 853–859.
Carroll, K.K., and Kritchevsky, D. (1994). Nutrition and disease update: Cancer (The American Oil Chemists Society).
141
Cartee, G.D., Young, D.A., Sleeper, M.D., Zierath, J., Wallberg-Henriksson, H., and Holloszy, J.O. (1989). Prolonged increase in insulin-stimulated glucose transport in muscle after exercise. Am J Physiol Endocrinol Metab 256, E494–E499.
Castellino, P., Luzi, L., Simonson, D.C., Haymond, M., and DeFronzo, R.A. (1987). Effect of insulin and plasma amino acid concentrations on leucine metabolism in man. Role of substrate availability on estimates of whole body protein synthesis. Journal of Clinical Investigation 80, 1784.
Catchpole, G.S., Beckmann, M., Enot, D.P., Mondhe, M., Zywicki, B., Taylor, J., Hardy, N., Smith, A., King, R.D., and Kell, D.B. (2005). Hierarchical metabolomics demonstrates substantial compositional similarity between genetically modified and conventional potato crops. Proceedings of the National Academy of Sciences of the United States of America 102, 14458–14462.
Cerdan, S., Künnecke, B., and Seelig, J. (1990). Cerebral metabolism of [1,2-13C2]acetate as detected by in vivo and in vitro 13C NMR. J. Biol. Chem. 265, 12916–12926.
Chan, E.C.Y., Koh, P.K., Mal, M., Cheah, P.Y., Eu, K.W., Backshall, A., Cavill, R., Nicholson, J.K., and Keun, H.C. (2009). Metabolic profiling of human colorectal cancer using high-resolution magic angle spinning nuclear magnetic resonance (HR-MAS NMR) spectroscopy and gas chromatography mass spectrometry (GC/MS). J. Proteome Res 8, 352–361.
CHATTERJEE, I.B., KAR, N.C., GHOSH, N.C., and GUHA, B.C. (1961). Biosynthesis of L-ascorbic acid: missing steps in animals incapable of synthesizing the vitamin. Nature 192, 163–164.
Chen, Q., Espey, M.G., Krishna, M.C., Mitchell, J.B., Corpe, C.P., Buettner, G.R., Shacter, E., and Levine, M. (2005). Pharmacologic ascorbic acid concentrations selectively kill cancer cells: Action as a pro-drug to deliver hydrogen peroxide to tissues. PNAS 102, 13604–13609.
Chen, Q., Espey, M.G., Sun, A.Y., Pooput, C., Kirk, K.L., Krishna, M.C., Khosh, D.B., Drisko, J., and Levine, M. (2008). Pharmacologic doses of ascorbate act as a prooxidant and decrease growth of aggressive tumor xenografts in mice. PNAS 105, 11105–11109.
Chevalier, S., Gougeon, R., Kreisman, S.H., Cassis, C., and Morais, J.A. (2004). The hyperinsulinemic amino acid clamp increases whole-body protein synthesis in young subjects. Metabolism 53, 388–396.
Cho, H.-W., Kim, S.B., Jeong, M.K., Park, Y., Ziegler, T.R., and Jones, D.P. (2008). Genetic algorithm-based feature selection in high-resolution NMR spectra. Expert Systems with Applications 35, 967–975.
142
Christian, P., and Stewart, C.P. (2010). Maternal Micronutrient Deficiency, Fetal Development, and the Risk of Chronic Disease. The Journal of Nutrition 140, 437 –445.
Chung, Y.-L., and Griffiths, J.R. (2008). Using Metabolomics to Monitor Anticancer Drugs. In Oncogenes Meet Metabolism, G. Kroemer, D. Mumberg, H. Keun, B. Riefke, T. Steger-Hartmann, and K. Petersen, eds. (Springer Berlin Heidelberg), pp. 55–78.
Clark, T.A., Sugnet, C.W., and Ares, M. (2002). Genomewide Analysis of mRNA Processing in Yeast Using Splicing-Specific Microarrays. Science 296, 907–910.
Claudino, W.M., Quattrone, A., Biganzoli, L., Pestrin, M., Bertini, I., and Di Leo, A. (2007). Metabolomics: available results, current research projects in breast cancer, and future applications. Journal of Clinical Oncology 25, 2840.
Clee, S.M., and Attie, A.D. (2007). The genetic landscape of type 2 diabetes in mice. Endocrine Reviews 28, 48–83.
Cloarec, O., Dumas, M.-E., Craig, A., Barton, R.H., Trygg, J., Hudson, J., Blancher, C., Gauguier, D., Lindon, J.C., Holmes, E., et al. (2005). Statistical total correlation spectroscopy: an exploratory approach for latent biomarker identification from metabolic 1H NMR data sets. Anal. Chem 77, 1282–1289.
Collins, F.S., Patrinos, A., Jordan, E., Chakravarti, A., Gesteland, R., and Walters, L. (1998). New Goals for the U.S. Human Genome Project: 1998-2003. Science 282, 682–689.
Connor, S.C., Hansen, M.K., Corner, A., Smith, R.F., and Ryan, T.E. (2010). Integration of metabolomics and transcriptomics data to aid biomarker discovery in type 2 diabetes. Mol. BioSyst. 6, 909.
Craig, A., Cloarec, O., Holmes, E., Nicholson, J.K., and Lindon, J.C. (2006). Scaling and normalization effects in NMR spectroscopic metabonomic data sets. Anal. Chem. 78, 2262–2267.
Davis, R.A., Charlton, A.J., Godward, J., Jones, S.A., Harrison, M., and Wilson, J.C. (2007). Adaptive binning: An improved binning method for metabolomics data using the undecimated wavelet transform. Chemometrics and Intelligent Laboratory Systems 85, 144–154.
Dawidowicz, E.A. (1987). Dynamics of Membrane Lipid Metabolism and Turnover. Annual Review of Biochemistry 56, 43–57.
Defernez, M., and Kemsley, E.K. (1997). The use and misuse of chemometrics for treating classification problems. TrAC Trends in Analytical Chemistry 16, 216–221.
143
Descartes, R. (1643). Discours de la méthode; Méditations y Description du corps humain. Oeuvres Philosophiques, Editions Alquie, París 1967,.
Dettmer, K., Aronov, P.A., and Hammock, B.D. (2007). MASS SPECTROMETRY-BASED METABOLOMICS. Mass Spectrom Rev 26, 51–78.
DeVilliers, D.C., Dixit, P.K., and Lazarow, A. (1966). Citrate metabolism in diabetes: I. Plasma citrate in alloxan-diabetic rats and in clinical diabetes. Metabolism 15, 458–465.
Dieterle, F., Ross, A., Schlotterbeck, G., and Senn, H. (2006). Probabilistic Quotient Normalization as Robust Method to Account for Dilution of Complex Biological Mixtures. Application in 1H NMR Metabonomics. Anal. Chem. 78, 4281–4290.
Dixon, R.A., Gang, D.R., Charlton, A.J., Fiehn, O., Kuiper, H.A., Reynolds, T.L., Tjeerdema, R.S., Jeffery, E.H., German, J.B., Ridley, W.P., et al. (2006). Applications of Metabolomics in Agriculture. J. Agric. Food Chem. 54, 8984–8994.
Van Doorn, M., Vogels, J., Tas, A., Van Hoogdalem, E.J., Burggraaf, J., Cohen, A., and Van Der Greef, J. (2007). Evaluation of metabolite profiles as biomarkers for the pharmacological effects of thiazolidinediones in Type 2 diabetes mellitus patients and healthy volunteers. British Journal of Clinical Pharmacology 63, 562–574.
Draisma, H.H.M., Reijmers, T.H., Van der Kloet, F., Bobeldijk-Pastorova, I., Spies-Faber, E., Vogels, J.T.W.E., Meulman, J.J., Boomsma, D.I., Van der Greef, J., and Hankemeier, T. (2010). Equating, or Correction for Between-Block Effects with Application to Body Fluid LC−MS and NMR Metabolomics Data Sets. Anal. Chem. 82, 1039–1046.
Duarte, N.C., Becker, S.A., Jamshidi, N., Thiele, I., Mo, M.L., Vo, T.D., Srivas, R., and Palsson, B.Ø. (2007). Global reconstruction of the human metabolic network based on genomic and bibliomic data. PNAS 104, 1777–1782.
Duggan, G.E., Hittel, D.S., Hughey, C.C., Weljie, A., Vogel, H.J., and Shearer, J. (2011a). Differentiating short-and long-term effects of diet in the obese mouse using 1H-nuclear magnetic resonance metabolomics. Diabetes, Obesity and Metabolism 13, 859–862.
Duggan, G.E., Hittel, D.S., Sensen, C.W., Weljie, A.M., Vogel, H.J., and Shearer, J. (2011b). Metabolomic response to exercise training in lean and diet-induced obese mice. Journal of Applied Physiology 110, 1311–1318.
Duggan, G.E., Joan Miller, B., Jirik, F.R., and Vogel, H.J. (2011c). Metabolic profiling of vitamin C deficiency in Gulo-/- mice using proton NMR spectroscopy. Journal of Biomolecular NMR 49, 165–173.
144
Dumas, M.-E., Maibaum, E.C., Teague, C., Ueshima, H., Zhou, B., Lindon, J.C., Nicholson, J.K., Stamler, J., Elliott, P., Chan, Q., et al. (2006). Assessment of analytical reproducibility of 1H NMR spectroscopy based metabonomics for large-scale epidemiological research: the INTERMAP Study. Anal. Chem. 78, 2199–2208.
Dunn, W.B., Broadhurst, D.I., Atherton, H.J., Goodacre, R., and Griffin, J.L. (2011). Systems level studies of mammalian metabolomes: the roles of mass spectrometry and nuclear magnetic resonance spectroscopy. Chem Soc Rev 40, 387–426.
Dunn, W.B., and Ellis, D.I. (2005). Metabolomics: Current analytical platforms and methodologies. TrAC Trends in Analytical Chemistry 24, 285–294.
Dunne, V.G., Bhattachayya, S., Besser, M., Rae, C., and Griffin, J.L. (2005). Metabolites from cerebrospinal fluid in aneurysmal subarachnoid haemorrhage correlate with vasospasm and clinical outcome: a pattern-recognition 1H NMR study. NMR in Biomedicine 18, 24–33.
Eastment, H.T., and Krzanowski, W.J. (1982). Cross-Validatory Choice of the Number of Components from a Principal Component Analysis. Technometrics 24, 73–77.
Efron, B., and Gong, G. (1983). A Leisurely Look at the Bootstrap, the Jackknife, and Cross-Validation. The American Statistician 37, 36–48.
Efron, B., Tibshirani, R., and Tibshirani, R.J. (1993). An introduction to the bootstrap (CRC Press).
Eriksson, L., Trygg, J., and Wold, S. (2008). CV‐ANOVA for significance testing of PLS and OPLS® models. Journal of Chemometrics 22, 594–600.
Fan, T.W., Lane, A.N., Higashi, R.M., Farag, M.A., Gao, H., Bousamra, M., and Miller, D.M. (2009). Altered regulation of metabolic pathways in human lung cancer discerned by 13C stable isotope-resolved metabolomics (SIRM). Mol Cancer 8, 41.
Fell, D. (1997). Understanding the control of metabolism. (Portland Press Ltd.).
Fiehn, O. (2002). Metabolomics – the link between genotypes and phenotypes. Plant Molecular Biology 48, 155–171.
Fiehn, O., Robertson, D., Griffin, J., Van der Werf, M., Nikolau, B., Morrison, N., Sumner, L.W., Goodacre, R., Hardy, N.W., Taylor, C., et al. (2007). The metabolomics standards initiative (MSI). Metabolomics 3, 175–178.
145
Forbes, V.E., Palmqvist, A., and Bach, L. (2006). The use and misuse of biomarkers in ecotoxicology. Environmental Toxicology and Chemistry 25, 272–280.
Forlani, G., Vannini, P., Marchesini, G., Zoli, M., Ciavarella, A., and Pisi, E. (1984). Insulin-dependent metabolism of branched-chain amino acids in obesity. Metabolism 33, 147–150.
Franconi, F., Loizzo, A., Ghirlanda, G., and Seghieri, G. (2006). Taurine supplementation and diabetes mellitus. Current Opinion in Clinical Nutrition & Metabolic Care 9, 32.
Frank, R., and Hargreaves, R. (2003). Clinical biomarkers in drug discovery and development. Nature Reviews Drug Discovery 2, 566–580.
Franken, H., Seitz, A., Lehmann, R., Häring, H.U., Stefan, N., and Zell, A. (2012). Inferring disease-related metabolite dependencies with a bayesian optimization algorithm. Evolutionary Computation, Machine Learning and Data Mining in Bioinformatics 62–73.
Friedrich, N. (2012). Metabolomics in diabetes research. J Endocrinol 215, 29–42.
Frikke-Schmidt, H., and Lykkesfeldt, J. (2009). Role of marginal vitamin C deficiency in atherogenesis: in vivo models and clinical studies. Basic Clin. Pharmacol. Toxicol 104, 419–433.
Fueger, P.T., Bracy, D.P., Malabanan, C.M., Pencek, R.R., Granner, D.K., and Wasserman, D.H. (2004a). Hexokinase II overexpression improves exercise-stimulated but not insulin-stimulated muscle glucose uptake in high-fat-fed C57BL/6J mice. Diabetes 53, 306–314.
Fueger, P.T., Bracy, D.P., Malabanan, C.M., Pencek, R.R., and Wasserman, D.H. (2004b). Distributed control of glucose uptake by working muscles of conscious mice: roles of transport and phosphorylation. Am J Physiol Endocrinol Metab 286, E77–E84.
Fueger, P.T., Hess, H.S., Bracy, D.P., Pencek, R.R., Posey, K.A., Charron, M.J., and Wasserman, D.H. (2004c). Regulation of Insulin-Stimulated Muscle Glucose Uptake in the Conscious Mouse: Role of Glucose Transport Is Dependent on Glucose Phosphorylation Capacity. Endocrinology 145, 4912–4916.
Fueger, P.T., Lee-Young, R.S., Shearer, J., Bracy, D.P., Heikkinen, S., Laakso, M., Rottman, J.N., and Wasserman, D.H. (2007). Phosphorylation Barriers to Skeletal and Cardiac Muscle Glucose Uptakes in High-Fat–Fed Mice Studies in Mice With a 50% Reduction of Hexokinase II. Diabetes 56, 2476–2484.
146
Fueger, P.T., Shearer, J., Bracy, D.P., Posey, K.A., Pencek, R.R., McGuinness, O.P., and Wasserman, D.H. (2005). Control of muscle glucose uptake: test of the rate-limiting step paradigm in conscious, unrestrained mice. J Physiol 562, 925–935.
Fukusaki, E., and Kobayashi, A. (2005). Plant metabolomics: potential for practical operation. Journal of Bioscience and Bioengineering 100, 347–354.
Gadian, D.G., and Radda, G.K. (1981). NMR Studies of Tissue Metabolism. Annual Review of Biochemistry 50, 69–83.
Garlick, P.J., and Grant, I. (1988). Amino acid infusion increases the sensitivity of muscle protein synthesis in vivo to insulin. Effect of branched-chain amino acids. Biochemical Journal 254, 579.
Gavai, A.K. (2009). Bayesian networks for omics data analysis (Wageningen Universiteit (Wageningen University)).
Ghannoum, M.A., Mukherjee, P.K., Jurevic, R.J., Retuerto, M., Brown, R.E., Sikaroodi, M., Webster-Cyriaque, J., and Gillevet, P.M. (2011). Metabolomics Reveals Differential Levels of Oral Metabolites in HIV-Infected Patients: Toward Novel Diagnostic Targets. OMICS: A Journal of Integrative Biology.
Gilbert, R.J., Goodacre, R., Woodward, A.M., and Kell, D.B. (1997). Genetic Programming: A Novel Method for the Quantitative Analysis of Pyrolysis Mass Spectral Data. Analytical Chemistry 69, 4381–4389.
Giovannucci, E., and Michaud, D. (2007). The role of obesity and related metabolic disturbances in cancers of the colon, prostate, and pancreas. Gastroenterology 132, 2208–2225.
Giskeødegård, G.F., Grinde, M.T., Sitter, B., Axelson, D.E., Lundgren, S., Fjøsne, H.E., Dahl, S., Gribbestad, I.S., and Bathen, T.F. (2010). Multivariate Modeling and Prediction of Breast Cancer Prognostic Factors Using MR Metabolomics. J. Proteome Res. 9, 972–979.
Glassbrook, N., Beecher, C., and Ryals, J. (2000). Metabolic profiling on the right path. Nat. Biotechnol 18, 1142–1143.
Glassbrook, N., and Ryals, J. (2001). A systematic approach to biochemical profiling. Current Opinion in Plant Biology 4, 186–190.
Gonzalez, C.A., and Riboli, E. (2010). Diet and cancer prevention: Contributions from the European Prospective Investigation into Cancer and Nutrition (EPIC) study. European Journal of Cancer 46, 2555–2562.
147
Goodacre, R., Broadhurst, D., Smilde, A.K., Kristal, B.S., Baker, J.D., Beger, R., Bessant, C., Connor, S., Capuani, G., Craig, A., et al. (2007). Proposed minimum reporting standards for data analysis in metabolomics. Metabolomics 3, 231–241.
Goodacre, R., Vaidyanathan, S., Dunn, W.B., Harrigan, G.G., and Kell, D.B. (2004). Metabolomics by numbers: acquiring and understanding global metabolite data. TRENDS in Biotechnology 22, 245–252.
Van der Greef, J., Martin, S., Juhasz, P., Adourian, A., Plasterer, T., Verheij, E.R., and McBurney, R.N. (2007). The Art and Practice of Systems Biology in Medicine: Mapping Patterns of Relationships. J. Proteome Res. 6, 1540–1559.
Griffin, J.L. (2006). Understanding mouse models of disease through metabolomics. Current Opinion in Chemical Biology 10, 309–315.
Griffin, J.L., and Kauppinen, R.A. (2007). A metabolomics perspective of human brain tumours. FEBS Journal 274, 1132–1139.
Griffin, J.L., and Nicholls, A.W. (2006). Metabolomics as a functional genomic tool for understanding lipid dysfunction in diabetes, obesity and related disorders. Pharmacogenomics 7, 1095+.
Ha, M.N., Graham, F.L., D’Souza, C.K., Muller, W.J., Igdoura, S.A., and Schellhorn, H.E. (2004). Functional rescue of vitamin C synthesis deficiency in human cells using adenoviral-based expression of murine l-gulono-γ-lactone oxidase. Genomics 83, 482–492.
Hageman, J.A., Van den Berg, R.A., Westerhuis, J.A., Van der Werf, M.J., and Smilde, A.K. (2008). Genetic algorithm based two-mode clustering of metabolomics data. Metabolomics 4, 141–149.
Hall, R., Beale, M., Fiehn, O., Hardy, N., Sumner, L., and Bino, R. (2002). Plant metabolomics: the missing link in functional genomics strategies. The Plant Cell Online 14, 1437–1440.
Harper, A.E., Miller, R.H., and Block, K.P. (1984). Branched-chain amino acid metabolism. Annual Review of Nutrition 4, 409–454.
Harri, H., and others (2009). Vitamin C supplementation and the common cold-was Linus Pauling right or wrong?
Harrison, F.E., Meredith, M.E., Dawes, S.M., Saskowski, J.L., and May, J.M. (2010). Low ascorbic acid and increased oxidative stress in gulo(-/-) mice during development. Brain Res 1349, 143–152.
Harvie, D.I. (2005). Limeys: The Conquest of Scurvy (The History Press Ltd).
148
Hastie, T., Tibshirani, R., and Friedman, J.H. (2009). The elements of statistical learning: data mining, inference, and prediction (Springer).
Heath, H., Melton, L.J., and Chu, C.-P. (1980). Diabetes Mellitus and Risk of Skeletal Fracture. New England Journal of Medicine 303, 567–570.
Heinig, M., and Johnson, R.J. (2006). Role of uric acid in hypertension, renal disease, and metabolic syndrome. Cleve Clin J Med 73, 1059–1064.
Hilario, M., Kalousis, A., M\üller, M., and Pellegrini, C. (2003). Machine learning approaches to lung cancer prediction from mass spectra. Proteomics 3, 1716–1719.
Hirschmann, J.V., and Raugi, G.J. (1999). Adult scurvy. J. Am. Acad. Dermatol 41, 895–906; quiz 907–910.
Holmes, E., and Antti, H. (2002). Chemometric contributions to the evolution of metabonomics: mathematical solutions to characterising and interpreting complex biological NMR spectra. Analyst 127, 1549–1557.
Honors, M.A., Davenport, B.M., and Kinzig, K.P. Effects of consuming a high carbohydrate diet after eight weeks of exposure to a ketogenic diet. Nutr Metab (Lond) 6, 46–46.
Hopps, E., Noto, D., Caimi, G., and Averna, M.R. (2010). A novel component of the metabolic syndrome: the oxidative stress. Nutr Metab Cardiovasc Dis 20, 72–77.
Imoto, S., and Namioka, S. (1983a). Acetate-Glucose Relationship in Growing Pigs. J ANIM SCI 56, 867–875.
Imoto, S., and Namioka, S. (1983b). Nutritive Value of Acetate in Growing Pigs. J ANIM SCI 56, 858–866.
Ingenuity, I. (2011). IPA (Ingenuity Pathway Analysis).
Irizarry, M.C. (2004). Biomarkers of Alzheimer Disease in Plasma. NeuroRX 1, 226–234.
Izquierdo-García, J.L., Nin, N., Ruíz-Cabello, J., Rojas, Y., De Paula, M., López-Cuenca, S., Morales, L., Martínez-Caro, L., Fernández-Segoviano, P., Esteban, A., et al. (2011). A metabolomic approach for diagnosis of experimental sepsis. Intensive Care Medicine 1–10.
James, P. (1997). Protein identification in the post-genome era: the rapid rise of proteomics. Quarterly Reviews of Biophysics 30, 279–331.
149
Jankevics, A., Liepinsh, E., Liepinsh, E., Vilskersts, R., Grinberga, S., Pugovics, O., and Dambrova, M. (2009). Metabolomic studies of experimental diabetic urine samples by 1H NMR spectroscopy and LC/MS method. Chemometrics and Intelligent Laboratory Systems 97, 11–17.
Jentzmik, F., Stephan, C., and Jung, K. (2010a). Reply to Arun Sreekumar, Laila M. Poisson, Thekkelnaycke M. Rajendiran, et al.’s Letter to the Editor re: Florian Jentzmik, Carsten Stephan, Kurt Miller, et al. Sarcosine in Urine after Digital Rectal Examination Fails as a Marker in Prostate Cancer Detection and Identification of Aggressive Tumours. Eur Urol 2010;58:12-8. European Urology 58, e31–e32.
Jentzmik, F., Stephan, C., Lein, M., Miller, K., Kamlage, B., Bethan, B., Kristiansen, G., and Jung, K. (2011). Sarcosine in Prostate Cancer Tissue is Not a Differential Metabolite for Prostate Cancer Aggressiveness and Biochemical Progression. The Journal of Urology 185, 706–711.
Jentzmik, F., Stephan, C., Miller, K., Schrader, M., Erbersdobler, A., Kristiansen, G., Lein, M., and Jung, K. (2010b). Sarcosine in Urine after Digital Rectal Examination Fails as a Marker in Prostate Cancer Detection and Identification of Aggressive Tumours. European Urology 58, 12–18.
Jonsson, P., Bruce, S.J., Moritz, T., Trygg, J., Sjöström, M., Plumb, R., Granger, J., Maibaum, E., Nicholson, J.K., Holmes, E., et al. (2005). Extraction, interpretation and validation of information for comparing samples in metabolic LC/MS data sets. Analyst 130, 701–707.
Jonsson, P., Gullberg, J., Nordstr\öm, A., Kusano, M., Kowalczyk, M., Sj\östr\öm, M., and Moritz, T. (2004). A strategy for identifying differences in large series of metabolomic samples analyzed by GC/MS. Anal. Chem 76, 1738–1745.
Kaddurah-Daouk, R., Kristal, B.S., and Weinshilboum, R.M. (2008). Metabolomics: a global biochemical approach to drug response and disease. Annu. Rev. Pharmacol. Toxicol. 48, 653–683.
Kaderbhai, N.N., Broadhurst, D.I., Ellis, D.I., Goodacre, R., and Kell, D.B. (2003). Functional genomics via metabolic footprinting: monitoring metabolite secretion byEscherichia coli tryptophan metabolism mutants using FT-IR and direct injection electrospray mass spectrometry. Comp. Funct. Genom. 4, 376–391.
Kadoglou, N.P.E., Iliadis, F., Angelopoulou, N., Perrea, D., Ampatzidis, G., Liapis, C.D., and Alevizos, M. (2007). The anti-inflammatory effects of exercise training in patients with type 2 diabetes mellitus. European Journal of Cardiovascular Prevention & Rehabilitation 14, 837–843.
150
Kanehisa, M., Goto, S., Furumichi, M., Tanabe, M., and Hirakawa, M. (2010). KEGG for representation and analysis of molecular networks involving diseases and drugs. Nucleic Acids Res 38, D355–360.
Kanehisa, M., Goto, S., Sato, Y., Furumichi, M., and Tanabe, M. (2011). KEGG for integration and interpretation of large-scale molecular data sets. Nucleic Acids Research 40, D109–D114.
Kardia, S.R., Chu, J., and Sowers, M.F.R. (2006). Characterizing variation in sex steroid hormone pathway genes in women of 4 races/ethnicities: The study of women’s health across the nation (SWAN). The American Journal of Medicine 119, S3–S15.
Kell, D.B. (2006). Systems biology, metabolic modelling and metabolomics in drug discovery and development. Drug Discovery Today 11, 1085–1092.
Kelley, D.E., He, J., Menshikova, E.V., and Ritov, V.B. (2002). Dysfunction of Mitochondria in Human Skeletal Muscle in Type 2 Diabetes. Diabetes 51, 2944–2950.
Kemsley, E.K., and Tapp, H.S. (2009). OPLS filtered data can be obtained directly from non-orthogonalized PLS1. Journal of Chemometrics 23, 263–264.
Kim, H.-J., Kim, J.H., Noh, S., Hur, H.J., Sung, M.J., Hwang, J.-T., Park, J.H., Yang, H.J., Kim, M.-S., Kwon, D.Y., et al. (2011). Metabolomic Analysis of Livers and Serum from High-Fat Diet Induced Obese Mice. J. Proteome Res. 10, 722–731.
Kim, K., Aronov, P., Zakharkin, S.O., Anderson, D., Perroud, B., Thompson, I.M., and Weiss, R.H. (2009a). Urine metabolomics analysis for kidney cancer detection and biomarker discovery. Molecular & Cellular Proteomics 8, 558–570.
Kim, S.B., Wang, Z., and Duran, C.M. (2006). A bayesian approach for the alignment of high-resolution nmr spectra. In Proceedings of the INFORMS Artificial Intelligence and Data Mining Workshop, Pittsburgh, PA, USA, pp. 1–6.
Kim, S.-H., Yang, S.-O., Kim, H.-S., Kim, Y., Park, T., and Choi, H.-K. (2009b). 1H-nuclear magnetic resonance spectroscopy-based metabolic assessment in a rat model of obesity induced by a high-fat diet. Anal Bioanal Chem 395, 1117–1124.
Kim, Y.S., Maruvada, P., and Milner, J.A. (2008). Metabolomics in biomarker discovery: future uses for cancer prevention. Future Oncology 4, 93–102.
Kind, T., Tolstikov, V., Fiehn, O., and Weiss, R.H. (2007). A comprehensive urinary metabolomic approach for identifying kidney cancer. Analytical Biochemistry 363, 185–195.
151
Kirschenlohr, H.L., Griffin, J.L., Clarke, S.C., Rhydwen, R., Grace, A.A., Schofield, P.M., Brindle, K.M., and Metcalfe, J.C. (2006). Proton NMR analysis of plasma is a weak predictor of coronary artery disease. Nat Med 12, 705–710.
Kohavi, R. (1995). A study of cross-validation and bootstrap for accuracy estimation and model selection. In International Joint Conference on Artificial Intelligence, pp. 1137–1145.
Koza, J.R., Streeter, M.J., and Keane, M.A. (2008). Routine high-return human-competitive automated problem-solving by means of genetic programming. Information Sciences 178, 4434–4452.
Kramer, M.A. (1991). Nonlinear principal component analysis using autoassociative neural networks. AIChE Journal 37, 233–243.
Krumsiek, J., Suhre, K., Illig, T., Adamski, J., and Theis, F.J. (2012). Bayesian Independent Component Analysis Recovers Pathway Signatures from Blood Metabolomics Data. Journal of Proteome Research 11, 4120–4131.
Laakso, M., Edelman, S.V., Brechtel, G., and Baron, A.D. (1992). Impaired Insulin-Mediated Skeletal Muscle Blood Flow in Patients With NIDDM. Diabetes 41, 1076–1083.
Lacy, P. (2011). Metabolomics of sepsis-induced acute lung injury: a new approach for biomarkers. American Journal of Physiology-Lung Cellular and Molecular Physiology 300, L1–L3.
Lafrance, D., Lands, L.C., and Burns, D.H. (2004). In vivo lactate measurement in human tissue by near-infrared diffuse reflectance spectroscopy. Vibrational Spectroscopy 36, 195–202.
Lander, E.S., Linton, L.M., Birren, B., Nusbaum, C., Zody, M.C., Baldwin, J., Devon, K., Dewar, K., Doyle, M., FitzHugh, W., et al. (2001). Initial sequencing and analysis of the human genome. Nature 409, 860–921.
Lanza, I.R., Zhang, S., Ward, L.E., Karakelides, H., Raftery, D., and Nair, K.S. (2010). Quantitative metabolomics by 1H-NMR and LC-MS/MS confirms altered metabolic pathways in diabetes. PloS One 5, e10538.
Lawton, K.A., Berger, A., Mitchell, M., Milgram, K.E., Evans, A.M., Guo, L., Hanson, R.W., Kalhan, S.C., Ryals, J.A., and Milburn, M.V. (2008). Analysis of the adult human plasma metabolome. Pharmacogenomics 9, 383–397.
Layman, D.K., and Walker, D.A. (2006). Potential importance of leucine in treatment of obesity and the metabolic syndrome. The Journal of Nutrition 136, 319S–323S.
152
Lee, C. (2000). Antioxidant ability of caffeine and its metabolites based on the study of oxygen radical absorbing capacity and inhibition of LDL peroxidation. Clin. Chim. Acta 295, 141–154.
Lee, J.M., Gianchandani, E.P., and Papin, J.A. (2006). Flux balance analysis in the era of metabolomics. Briefings in Bioinformatics 7, 140–150.
Leiter, E.H. (1993). Obesity genes and diabetes induction in the mouse. Critical Reviews in Food Science & Nutrition 33, 333–338.
Levin, B.E., Dunn-Meynell, A.A., Balkan, B., and Keesey, R.E. (1997). Selective breeding for diet-induced obesity and resistance in Sprague-Dawley rats. Am J Physiol Regul Integr Comp Physiol 273, R725–R730.
Levin, B.E., Triscari, J., Hogan, S., and Sullivan, A.C. (1987). Resistance to diet-induced obesity: food intake, pancreatic sympathetic tone, and insulin. Am J Physiol Regul Integr Comp Physiol 252, R471–R478.
Li, C., and Wong, W.H. (2001). Model-based analysis of oligonucleotide arrays: model validation, design issues and standard error application. Genome Biology 2, research0032.1.
Li, L.O., Hu, Y.-F., Wang, L., Mitchell, M., Berger, A., and Coleman, R.A. (2010). Early Hepatic Insulin Resistance in Mice: A Metabolomics Analysis. Molecular Endocrinology 24, 657–666.
Li, Y., Shi, C.-X., Mossman, K.L., Rosenfeld, J., Boo, Y.C., and Schellhorn, H.E. (2008). Restoration of Vitamin C Synthesis in Transgenic Gulo −/− Mice by Helper-Dependent Adenovirus-Based Expression of Gulonolactone Oxidase. Human Gene Therapy 19, 1349–1358.
Lu, W., Bennett, B.D., and Rabinowitz, J.D. (2008). Analytical strategies for LC–MS-based targeted metabolomics. Journal of Chromatography B 871, 236–242.
Luetscher Jr, J.A. (1942). The metabolism of amino acids in diabetes mellitus. Journal of Clinical Investigation 21, 275.
Lutz, U., Bittner, N., Lutz, R.W., and Lutz, W.K. (2008). Metabolite profiling in human urine by LC-MS/MS: method optimization and application for glucuronides from dextromethorphan metabolism. Journal of Chromatography B 871, 349–356.
Madsen, R., Lundstedt, T., and Trygg, J. (2010). Chemometrics in metabolomics--A review in human disease diagnosis. Analytica Chimica Acta 659, 23–33.
153
Maeda, N., Hagihara, H., Nakata, Y., Hiller, S., Wilder, J., and Reddick, R. (2000). Aortic wall damage in mice unable to synthesize ascorbic acid. Proc. Natl. Acad. Sci. U.S.A 97, 841–846.
Mahadevan, S., Shah, S.L., Marrie, T.J., and Slupsky, C.M. (2008). Analysis of metabolomic data using support vector machines. Analytical Chemistry 80, 7562–7570.
Mai, P.L., Wentzensen, N., and Greene, M.H. (2011). Challenges Related to Developing Serum-Based Biomarkers for Early Ovarian Cancer Detection. Cancer Prev Res 4, 303–306.
Malet-Martino, M., and Holzgrabe, U. (2010). NMR techniques in biomedical and pharmaceutical analysis. J Pharm Biomed Anal.
Mamas, M., Dunn, W.B., Neyses, L., and Goodacre, R. (2010). The role of metabolites and metabolomics in clinically applicable biomarkers of disease. Archives of Toxicology 85, 5–17.
Marchesini, G., Bianchi, G.P., Vilstrup, H., Capelli, M., Zoli, M., and Pisi, E. (1991). Elimination of infused branched-chain amino-acids from plasma of patients with non-obese type 2 diabetes mellitus. Clinical Nutrition 10, 105–113.
Mashego, M.R., Rumbold, K., De Mey, M., Vandamme, E., Soetaert, W., and Heijnen, J.J. (2007). Microbial metabolomics: past, present and future methodologies. Biotechnology Letters 29, 1–16.
Mcateer, J.G.P., Skerrett, S.J., Liggitt, D., and Frevert, C.W. (2009). A Bayesian integration model of high-throughput proteomics and metabolomics data for improved early detection of microbial infections. In Pacific Symposium on Biocomputing 2009: Kohala Coast, Hawaii, USA, 5-9 January 2009, p. 451.
McElheny, V. (2012). Drawing the Map of Life: Inside the Human Genome Project (Basic Books).
McGovern, A.C., Broadhurst, D., Taylor, J., Kaderbhai, N., Winson, M.K., Small, D.A., Rowland, J.J., Kell, D.B., and Goodacre, R. (2002). Monitoring of complex industrial bioprocesses for metabolite concentrations using modern spectroscopies and machine learning: Application to gibberellic acid production. Biotechnology and Bioengineering 78, 527–538.
Miccheli, A., Marini, F., Capuani, G., Miccheli, A.T., Delfini, M., Cocco, M.E.D., Puccetti, C., Paci, M., Rizzo, M., and Spataro, A. (2009). The Influence of a Sports Drink on the Postexercise Metabolism of Elite Athletes as Investigated by NMR-Based Metabolomics. J Am Coll Nutr 28, 553–564.
154
Mooradian, A.D. (1987). Blood-brain barrier choline transport is reduced in diabetic rats. Diabetes 36, 1094–1097.
MORGAN, C.R., and LAZAROW, A. (1962). Immunoassay of insulin using a two-antibody system. Proc. Soc. Exp. Biol. Med 110, 29–32.
Næs, T., Isaksson, T., Fearn, T., and Davies, T. (2004). Interpreting PCR and PLS solutions. A User-Friendly Guide to Multivariate Calibration and Classification 1, 39–54.
Naidu, K.A. (2003). Vitamin C in human health and disease is still a mystery ? An overview. Nutrition Journal 2, 7.
Natelson, S., Pincus, J.B., and Lugovoy, J.K. (1948). RESPONSE OF CITRIC ACID LEVELS TO ORAL ADMINISTRATION OF GLUCOSE. I. NORMAL ADULTS AND CHILDREN 1. Journal of Clinical Investigation 27, 446–449.
Newgard, C.B., An, J., Bain, J.R., Muehlbauer, M.J., Stevens, R.D., Lien, L.F., Haqq, A.M., Shah, S.H., Arlotto, M., and Slentz, C.A. (2009). A Branched-Chain Amino Acid-Related Metabolic Signature that Differentiates Obese and Lean Humans and Contributes to Insulin Resistance. Cell Metabolism 9, 311–326.
Nicholls, A.W. (2012). Realising the potential of metabolomics. Bioanalysis 4, 2195–2197.
Nicholls, A.W., Mortishire-Smith, R.J., and Nicholson, J.K. (2003). NMR Spectroscopic-Based Metabonomic Studies of Urinary Metabolite Variation in Acclimatizing Germ-Free Rats. Chem. Res. Toxicol. 16, 1395–1404.
Nicholson, J.K. (2006). Global systems biology, personalized medicine and molecular epidemiology. Molecular Systems Biology 2,.
Nicholson, J.K., Foxall, P.J., Spraul, M., Farrant, R.D., and Lindon, J.C. (1995). 750 MHz 1H and 1H-13C NMR spectroscopy of human blood plasma. Anal. Chem 67, 793–811.
Nicholson, J.K., and Wilson, I.D. (1987). High resolution nuclear magnetic resonance spectroscopy of biological samples as an aid to drug development. Prog Drug Res 31, 427–479.
Nicholson, J.K., and Wilson, I.D. (1989). High resolution proton magnetic resonance spectroscopy of biological fluids. Progress in Nuclear Magnetic Resonance Spectroscopy 21, 449–501.
Nishikimi, M., Fukuyama, R., Minoshima, S., Shimizu, N., and Yagi, K. (1994). Cloning and chromosomal mapping of the human nonfunctional gene for L-gulono-
155
gamma-lactone oxidase, the enzyme for L-ascorbic acid biosynthesis missing in man. Journal of Biological Chemistry 269, 13685–13688.
Nishikimi, M., Kawai, T., and Yagi, K. (1992). Guinea pigs possess a highly mutated gene for L-gulono-gamma-lactone oxidase, the key enzyme for L-ascorbic acid biosynthesis missing in this species. Journal of Biological Chemistry 267, 21967–21972.
Nishikimi, M., and Yagi, K. (1991). Molecular basis for the deficiency in humans of gulonolactone oxidase, a key enzyme for ascorbic acid biosynthesis. Am. J. Clin. Nutr 54, 1203S–1208S.
Oksman-Caldentey, K.M., and Inzé, D. (2004). Plant cell factories in the post-genomic era: new ways to produce designer secondary metabolites. Trends in Plant Science 9, 433–440.
Oliver, S.G., Winson, M.K., Kell, D.B., and Baganz, F. (1998). Systematic functional analysis of the yeast genome. In European Symposium of Life Sciences Research in Space (Oser), p. 583.
Osten, D.W. (1988). Selection of optimal regression models via cross-validation. Journal of Chemometrics 2, 39–48.
Padh, H. (1990). Cellular functions of ascorbic acid. Biochem. Cell Biol 68, 1166–1173.
Park, S.-Y., Cho, Y.-R., Kim, H.-J., Higashimori, T., Danton, C., Lee, M.-K., Dey, A., Rothermel, B., Kim, Y.-B., Kalinowski, A., et al. (2005). Unraveling the temporal pattern of diet-induced insulin resistance in individual organs and cardiac dysfunction in C57BL/6 mice. Diabetes 54, 3530–3540.
Pauling, L., Robinson, A.B., Teranishi, R., and Cary, P. (1971). Quantitative Analysis of Urine Vapor and Breath by Gas-Liquid Partition Chromatography. PNAS 68, 2374–2376.
Pendyala, G., Want, E.J., Webb, W., Siuzdak, G., and Fox, H.S. (2007). Biomarkers for neuroAIDS: the widening scope of metabolomics. Journal of Neuroimmune Pharmacology 2, 72–80.
Penttila, I.., and Pollanen, O. Effect of Insulin and Tolbutamide on Blood Citric Acid in Rabbits, Scandinavian Journal of Clinical & Laboratory Investigation, Informa Healthcare.
Pepe, M.S., Feng, Z., Janes, H., Bossuyt, P.M., and Potter, J.D. (2008). Pivotal Evaluation of the Accuracy of a Biomarker Used for Classification or Prediction: Standards for Study Design. JNCI J Natl Cancer Inst 100, 1432–1438.
156
Pigini, D., Cialdella, A.M., Faranda, P., and Tranfo, G. (2006). Comparison between external and internal standard calibration in the validation of an analytical method for 1-hydroxypyrene in human urine by high-performance liquid chromatography/tandem mass spectrometry. Rapid Communications in Mass Spectrometry 20, 1013–1018.
Piloquet, H., Ferchaud-Roucher, V., Duengler, F., Zair, Y., Maugere, P., and Krempf, M. (2003). Insulin effects on acetate metabolism. Am J Physiol Endocrinol Metab 285, E561–E565.
Pincus, J.B., Natelson, S., and Lugovoy, J.K. (1948). RESPONSE OF CITRIC ACID LEVELS TO ORAL ADMINISTRATION OF GLUCOSE. II. ABNORMALITIES OBSERVED IN THE DIABETIC AND CONVULSIVE STATE 1. Journal of Clinical Investigation 27, 450–453.
Pohjanen, E., Thysell, E., Lindberg, J., Schuppe-Koistinen, I., Moritz, T., Jonsson, P., and Antti, H. (2006). Statistical multivariate metabolite profiling for aiding biomarker pattern detection and mechanistic interpretations in GC/MS based metabolomics. Metabolomics 2, 257–268.
Quinones, M.P., and Kaddurah-Daouk, R. (2009). Metabolomics tools for identifying biomarkers for neuropsychiatric diseases. Neurobiology of Disease 35, 165–176.
Raab, R.M., Bullen, J., Kelleher, J., Mantzoros, C., and Stephanopoulos, G. (2005). Regulation of mouse hepatic genes in response to diet induced obesity, insulin resistance and fasting induced weight reduction. Nutr Metab (Lond) 2, 15.
Raamsdonk, L.M., Teusink, B., Broadhurst, D., Zhang, N., Hayes, A., Walsh, M.C., Berden, J.A., Brindle, K.M., Kell, D.B., Rowland, J.J., et al. (2001). A functional genomics strategy that uses metabolome data to reveal the phenotype of silent mutations. Nature Biotechnology 19, 45–50.
Ransohoff, D.F. (2005). Bias as a threat to the validity of cancer molecular-marker research. Nature Reviews Cancer 5, 142–149.
Ransohoff, D.F., and Feinstein, A.R. (1978). Problems of spectrum and bias in evaluating the efficacy of diagnostic tests. New England Journal of Medicine 299, 926–930.
Rao, A.R., Motiwala, H.G., and Karim, O.M.A. (2008). The discovery of prostate-specific antigen. BJU Int 101, 5–10.
Rao, R., Hao, C.-M., Redha, R., Wasserman, D., McGuinness, O., and Breyer, M. (2007). Glycogen synthase kinase 3 inhibition improves insulin-stimulated glucose
157
metabolism but not hypertension in high-fat-fed C57BL/6J mice. Diabetologia 50, 452–460.
Robertson, D.G. (2005). Metabonomics in Toxicology: A Review. Toxicol. Sci. 85, 809–822.
Robertson, D.G., and Reily, M.D. (2012). The Current Status of Metabolomics in Drug Discovery and Development. Drug Development Research 73, 535–546.
Robertson, D.G., Watkins, P.B., and Reily, M.D. (2010). Metabolomics in Toxicology: Preclinical and Clinical Applications. Toxicological Sciences 120, S146–S170.
Rocconi, R.P., Matthews, K.S., Kemper, M.K., Hoskins, K.E., Huh, W.K., and Straughn, J.M., Jr (2009). The timing of normalization of CA-125 levels during primary chemotherapy is predictive of survival in patients with epithelial ovarian cancer. Gynecol. Oncol. 114, 242–245.
Roussel, R., Mentre, F., Bouchemal, N., Hadjadj, S., Lievre, M., Chatellier, G., Menard, J., Panhard, X., Le Henanff, A., Marre, M., et al. (2007). NMR-based prediction of cardiovascular risk in diabetes. Nat Med 13, 399–400.
Rubtsov, D.V., and Griffin, J.L. (2007). Time-domain Bayesian detection and estimation of noisy damped sinusoidal signals applied to NMR spectroscopy. Journal of Magnetic Resonance 188, 367–379.
Sakurai, T., Akiyama, K., and Saito, K. (2011). Analytical Platforms and Databases Ranging from Plant Transcriptomics to Metabolomics. Medicinal Plant Biotechnology 222.
Salek, R.M., Maguire, M.L., Bentley, E., Rubtsov, D.V., Hough, T., Cheeseman, M., Nunez, D., Sweatman, B.C., Haselden, J.N., Cox, R.D., et al. (2007). A metabolomic comparison of urinary changes in type 2 diabetes in mouse, rat, and human. Physiol. Genomics 29, 99–108.
Sauberlich, H.E. (1994). Pharmacology of vitamin C. Annu. Rev. Nutr 14, 371–391.
Scalbert, A., Brennan, L., Fiehn, O., Hankemeier, T., Kristal, B.S., Ommen, B., Pujos-Guillot, E., Verheij, E., Wishart, D., and Wopereis, S. (2009). Mass-spectrometry-based metabolomics: limitations and recommendations for future progress with particular focus on nutrition research. Metabolomics 5, 435–458.
Schemmel, R., Mickelsen, O., and Gill, J.L. (1970). Dietary Obesity in Rats: Body Weight and Body Fat Accretion in Seven Strains of Rats. J. Nutr. 100, 1041–1048.
158
Schicho, R., Nazyrova, A., Shaykhutdinov, R., Duggan, G., Vogel, H.J., and Storr, M. (0). Quantitative metabolomic profiling of serum and urine in DSS-induced ulcerative colitis of mice by 1H NMR spectroscopy. Journal of Proteome Research 0,.
Sébédio, J.-L., Pujos-Guillot, E., and Ferrara, M. (2009). Metabolomics in evaluation of glucose disorders. Current Opinion in Clinical Nutrition and Metabolic Care 12, 412–418.
Sellick, C.A., Knight, D., Croxford, A.S., Maqsood, A.R., Stephens, G.M., Goodacre, R., and Dickson, A.J. (2010). Evaluation of extraction processes for intracellular metabolite profiling of mammalian cells: matching extraction approaches to cell type and metabolite targets. Metabolomics 6, 427–438.
Sevanian, A., Davies, K.J., and Hochstein, P. (1991). Serum urate as an antioxidant for ascorbic acid. Am. J. Clin. Nutr 54, 1129S–1134S.
Shaykhutdinov, R., MacInnis, G., Dowlatabadi, R., Weljie, A., and Vogel, H. (2009). Quantitative analysis of metabolite concentrations in human urine samples using <sup>13</sup>C<sup>1</sup>H NMR spectroscopy. Metabolomics 5, 307–317.
Shearer, J., Duggan, G., Weljie, A., Hittel, D.S., Wasserman, D.H., and Vogel, H.J. (2008). Metabolomic profiling of dietary-induced insulin resistance in the high fat-fed C57BL/6J mouse. Diabetes Obes Metab 10, 950–958.
Shearer, J., Fueger, P.T., Bracy, D.P., Wasserman, D.H., and Rottman, J.N. (2005). Partial gene deletion of heart-type fatty acid-binding protein limits the severity of dietary-induced insulin resistance. Diabetes 54, 3133–3139.
Shurubor, Y.I., Paolucci, U., Krasnikov, B.F., Matson, W.R., and Kristal, B.S. (2005). Analytical precision, biological variation, and mathematical normalization in high data density metabolomics. Metabolomics 1, 75–85.
Smilde, A.K., Jansen, J.J., Hoefsloot, H.C.J., Lamers, R.-J.A.N., Greef, J. van der, and Timmerman, M.E. (2005). ANOVA-simultaneous component analysis (ASCA): a new tool for analyzing designed metabolomics data. Bioinformatics 21, 3043–3048.
Sok, D.-E. (1998). Ascorbate-Induced Oxidative Inactivation of Zn2+-Glycerophosphocholine Cholinephosphodiesterase. Journal of Neurochemistry 70, 1167–1174.
Spratlin, J.L., Serkova, N.J., and Eckhardt, S.G. (2009). Clinical applications of metabolomics in oncology: a review. Clinical Cancer Research 15, 431–440.
159
Sreekumar, A., Poisson, L.M., Rajendiran, T.M., Khan, A.P., Cao, Q., Yu, J., Laxman, B., Mehra, R., Lonigro, R.J., Li, Y., et al. (2009). Metabolomic profiles delineate potential role for sarcosine in prostate cancer progression. Nature 457, 910–914.
Sreekumar, A., Poisson, L.M., Rajendiran, T.M., Khan, A.P., Cao, Q., Yu, J., Laxman, B., Mehra, R., Lonigro, R.J., Li, Y., et al. (2010). Re: Florian Jentzmik, Carsten Stephan, Kurt Miller, et al. Sarcosine in Urine after Digital Rectal Examination Fails as a Marker in Prostate Cancer Detection and Identification of Aggressive Tumours. Eur Urol 2010;58:12-8. European Urology 58, e29–e30.
Stamler, J., Vaccaro, O., Neaton, J.D., and Wentworth, D. (1993). Diabetes, other risk factors, and 12-yr cardiovascular mortality for men screened in the Multiple Risk Factor Intervention Trial. Diabetes Care 16, 434–444.
Stone, I. (1979). Homo sapiens ascorbicus, a biochemically corrected robust human mutant. Medical Hypotheses 5, 711–721.
Stringer, K.A., Serkova, N.J., Karnovsky, A., Guire, K., Paine, R., and Standiford, T.J. (2011). Metabolic consequences of sepsis-induced acute lung injury revealed by plasma 1H-nuclear magnetic resonance quantitative metabolomics and computational analysis. Am J Physiol Lung Cell Mol Physiol 300, L4–L11.
Svirbely, J.L., and Szent-Györgyi, A. (1932). The chemical nature of vitamin C. Biochemical Journal 26, 865.
Sysi-Aho, M., Katajamaa, M., Yetukuri, L., and Orešič, M. (2007). Normalization method for metabolomics data using optimal selection of multiple internal standards. BMC Bioinformatics 8, 93.
Tikunov, Y., Lommen, A., De Vos, C.H.R., Verhoeven, H.A., Bino, R.J., Hall, R.D., and Bovy, A.G. (2005). A novel approach for nontargeted data analysis for metabolomics. Large-scale profiling of tomato fruit volatiles. Plant Physiology 139, 1125–1137.
Tiziani, S., Lopes, V., and Günther, U.L. (2009). Early stage diagnosis of oral cancer using 1H NMR-based metabolomics. Neoplasia (New York, NY) 11, 269.
Todorow, J.T., and Dikow, A.L. (1960). Changes of blood citric acid level during glucose tolerance test in normals and diabetic and hyperthyreotic patients. Clinica Chimica Acta 5, 762–765.
Torgrip, R.J.O., Åberg, K.M., Alm, E., Schuppe-Koistinen, I., and Lindberg, J. (2008). A note on normalization of biofluid 1D 1H-NMR data. Metabolomics 4, 114–121.
160
Trewavas, A. (2006). A Brief History of Systems Biology “Every object that biology studies is a system of systems.” Francois Jacob (1974). Plant Cell 18, 2420–2430.
Truong, Y., Lin, X., and Beecher, C. (2004). Learning a complex metabolomic dataset using random forests and support vector machines. In Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 835–840.
Trygg, J., Holmes, E., and Lundstedt, T. (2007). Chemometrics in metabonomics. J. Proteome Res 6, 469–479.
Trygg, J., and Wold, S. (2002). Orthogonal projections to latent structures (O-PLS). Journal of Chemometrics 16, 119–128.
Umetrics (2006). Multi- and Megavariate Data Analysis, Part 1, Basic Principles and Applications (MKS Umetrics AB).
Vander Heiden, M.G., Cantley, L.C., and Thompson, C.B. (2009). Understanding the Warburg effect: the metabolic requirements of cell proliferation. Science Signalling 324, 1029.
Vial, R., Balart, L., Arroyave, G., and others (1974). Effect of exercise and physical fitness on serum lipids and lipoproteins. Atherosclerosis 20, 1–9.
Villas-Bôas, S.G., Koulman, A., and Lane, G.A. (2007). Analytical methods from the perspective of method standardization. In Molecular Genetics of Recombination, A. Aguilera, and R. Rothstein, eds. (Springer Berlin Heidelberg), pp. 487–513.
Vinzi, V.E. (2010). Handbook of Partial Least Squares: Concepts, Methods and Applications (Springer).
Vissers, M.C.M., and Wilkie, R.P. (2007). Ascorbate deficiency results in impaired neutrophil apoptosis and clearance and is associated with up-regulation of hypoxia-inducible factor 1alpha. J. Leukoc. Biol 81, 1236–1244.
Vitzthum, F., Behrens, F., Anderson, N.L., and Shaw, J.H. (2005). Proteomics: From Basic Research to Diagnostic Application. A Review of Requirements & Needs†. J. Proteome Res. 4, 1086–1097.
Wang, T.J., Larson, M.G., Vasan, R.S., Cheng, S., Rhee, E.P., McCabe, E., Lewis, G.D., Fox, C.S., Jacques, P.F., Fernandez, C., et al. (2011). Metabolite profiles and the risk of developing diabetes. Nature Medicine 17, 448–453.
161
Warrack, B.M., Hnatyshyn, S., Ott, K.H., Reily, M.D., Sanders, M., Zhang, H., and Drexler, D.M. (2009). Normalization strategies for metabonomic analysis of urine samples. Journal of Chromatography B 877, 547–552.
Weckwerth, W., and Morgenthal, K. (2005). Metabolomics: from pattern recognition to biological interpretation. Drug Discovery Today 10, 1551–1558.
Weljie, A.M., Dowlatabadi, R., Miller, B.J., Vogel, H.J., and Jirik, F.R. (2007). An inflammatory arthritis-associated metabolite biomarker pattern revealed by 1H NMR spectroscopy. J. Proteome Res 6, 3456–3464.
Weljie, A.M., and Jirik, F.R. Hypoxia-induced metabolic shifts in cancer cells: Moving beyond the Warburg effect. The International Journal of Biochemistry & Cell Biology In Press, Corrected Proof,.
Weljie, A.M., Newton, J., Mercier, P., Carlson, E., and Slupsky, C.M. (2006). Targeted profiling: quantitative analysis of 1H NMR metabolomics data. Anal. Chem 78, 4430–4442.
West, D.B., Boozer, C.N., Moody, D.L., and Atkinson, R.L. (1992). Dietary obesity in nine inbred mouse strains. Am J Physiol Regul Integr Comp Physiol 262, R1025–R1032.
Westerhuis, J., Van Velzen, E., Hoefsloot, H., and Smilde, A. (2010). Multivariate paired data analysis: multilevel PLSDA versus OPLSDA. Metabolomics 6, 119–128.
Wiklund, S., Johansson, E., Sjostrom, L., Mellerowicz, E.J., Edlund, U., Shockcor, J.P., Gottfries, J., Moritz, T., and Trygg, J. (2007). Visualization of GC/TOF-MS-Based Metabolomics Data for Identification of Biochemically Interesting Compounds Using OPLS Class Models. Anal. Chem. 80, 115–122.
Wikoff, W.R., Anfora, A.T., Liu, J., Schultz, P.G., Lesley, S.A., Peters, E.C., and Siuzdak, G. (2009). Metabolomics analysis reveals large effects of gut microflora on mammalian blood metabolites. Proceedings of the National Academy of Sciences 106, 3698.
Wilcken, B., Wiley, V., Hammond, J., and Carpenter, K. (2003). Screening newborns for inborn errors of metabolism by tandem mass spectrometry. New England Journal of Medicine 348, 2304–2312.
Winkler, B.S., Orselli, S.M., and Rex, T.S. (1994). The redox couple between glutathione and ascorbic acid: a chemical and physiological perspective. Free Radic. Biol. Med 17, 333–349.
162
Wishart, D.S., Jewison, T., Guo, A.C., Wilson, M., Knox, C., Liu, Y., Djoumbou, Y., Mandal, R., Aziat, F., Dong, E., et al. (2012). HMDB 3.0--The Human Metabolome Database in 2013. Nucleic Acids Research 41, D801–D807.
Wishart, D.S., Tzur, D., Knox, C., Eisner, R., Guo, A.C., Young, N., Cheng, D., Jewell, K., Arndt, D., Sawhney, S., et al. (2007). HMDB: the Human Metabolome Database. Nucleic Acids Res 35, D521–526.
Wold, S., Esbensen, K., and Geladi, P. (1987). Principal component analysis. Chemometrics and Intelligent Laboratory Systems 2, 37–52.
Wold, S., Trygg, J., Berglund, A., and Antti, H. (2001). Some recent developments in PLS modeling. Chemometrics and Intelligent Laboratory Systems 58, 131–150.
Wong, M. -c., Chung, J.W.Y., and Wong, T.K.S. (2007). Effects of treatments for symptoms of painful diabetic neuropathy: systematic review. BMJ 335, 87–87.
Wood, P. (2012). Lipidomics of Alzheimer’s disease: current status. Alzheimer’s Research & Therapy 4, 1–10.
Woodger, J.H. (1929). Biological principles: A critical study (Taylor & Francis).
Wu, G. (2009). Amino acids: metabolism, functions, and nutrition. Amino Acids 37, 1–17.
Wu, G., and Meininger, C.J. (2009). Nitric oxide and vascular insulin resistance. BioFactors 35, 21–27.
Wu, J.J., Roth, R.J., Anderson, E.J., Hong, E.-G., Lee, M.-K., Choi, C.S., Neufer, P.D., Shulman, G.I., Kim, J.K., and Bennett, A.M. (2006). Mice lacking MAP kinase phosphatase-1 have enhanced MAP kinase activity and resistance to diet-induced obesity. Cell Metabolism 4, 61–73.
Xu, Y., Zomer, S., and Brereton, R.G. (2006). Support vector machines: a recent method for classification in chemometrics. Critical Reviews in Analytical Chemistry 36, 177–188.
Xuan, J., Pan, G., Qiu, Y., Yang, L., Su, M., Liu, Y., Chen, J., Feng, G., Fang, Y., and Jia, W. (2011). Metabolomic profiling to identify potential serum biomarkers for schizophrenia and risperidone action. Journal of Proteome Research 10, 5433–5443.
Y, I., Y, O., and M, N. (2003). The Whole Structure of the Human Nonfunctional L-Gulono-.GAMMA.-Lactone Oxidase Gene-the Gene Responsible for Scurvy-and the Evolution of Repetitive Sequences Thereon. J Nutr Sci Vitaminol 49, 315–319.
163
Yamanouchi, K., Shinozaki, T., Chikada, K., Nishikawa, T., Ito, K., Shimizu, S., Ozawa, N., Suzuki, Y., Maeno, H., Kato, K., et al. (1995). Daily Walking Combined With Diet Therapy Is a Useful Means for Obese NIDDM Patients Not Only to Reduce Body Weight But Also to Improve Insulin Sensitivity. Dia Care 18, 775–778.
Yan, B., A, J., Wang, G., Lu, H., Huang, X., Liu, Y., Zha, W., Hao, H., Zhang, Y., Liu, L., et al. (2009). Metabolomic investigation into variation of endogenous metabolites in professional athletes subject to strength-endurance training. J Appl Physiol 106, 531–538.
Zamboni, N., and Sauer, U. (2009). Novel biological insights through metabolomics and 13C-flux analysis. Current Opinion in Microbiology 12, 553–558.
Zelena, E., Dunn, W.B., Broadhurst, D., Francis-McIntyre, S., Carroll, K.M., Begley, P., O’Hagan, S., Knowles, J.D., Halsall, A., Wilson, I.D., et al. (2009). Development of a Robust and Repeatable UPLC−MS Method for the Long-Term Metabolomic Study of Human Serum. Anal. Chem. 81, 1357–1364.
Zhang, S., Nagana Gowda, G.A., Asiago, V., Shanaiah, N., Barbas, C., and Raftery, D. (2008). Correlative and quantitative 1H NMR-based metabolomics reveals specific metabolic pathway disturbances in diabetic rats. Analytical Biochemistry 383, 76–84.
Zhao, X., Fritsche, J., Wang, J., Chen, J., Rittig, K., Schmitt-Kopplin, P., Fritsche, A., Häring, H.-U., Schleicher, E., Xu, G., et al. (2010). Metabonomic fingerprints of fasting plasma and spot urine reveal human pre-diabetic metabolic traits. Metabolomics 6, 362–374.
Zhu, C.S., Pinsky, P.F., Cramer, D.W., Ransohoff, D.F., Hartge, P., Pfeiffer, R.M., Urban, N., Mor, G., Bast, R.C., Moore, L.E., et al. (2011). A Framework for Evaluating Biomarkers for Early Detection: Validation of Biomarker Panels for Ovarian Cancer. Cancer Prev Res 4, 375–383.
Zivkovic, A.M., and German, J.B. (2009). Metabolomics for assessment of nutritional status. Curr Opin Clin Nutr Metab Care 12, 501–507.
