Metabolic phenotyping in health and disease.

Abstract

Analyzing metabolites (small molecules <1 kDa) in body fluids such as urine and plasma using various spectroscopic methods provides information on the metabotype (metabolic phenotype) of individuals or populations, information that can be applied to personalized medicine or public healthcare.

Full text

Page 1

Leading Edge
Forum
714 Cell 134, September 5, 2008 ©2008 Elsevier Inc.
A major challenge for 21st century medi-
cine is to understand the relationships
and interactions between genetic varia-
tions and environmental triggers of dis-
ease (Nicholson, 2006). In response to
this challenge, systems biology powered
by genomics, proteomics, bioinformat-
ics, and now metabolomics and metabo-
nomics (Box 1), is providing a new logical
framework to elucidate disease etiol-
ogy and to uncover latent connections
between seemingly disparate disease
states (Loscalzo et al., 2007). In particu-
lar, genome-wide association studies
(GWAS) enable genetic features to be
linked to many pre-disease conditions
(Kruglyak, 2008). But GWAS can explain
only a small proportion of incidence
variation and may not translate between
different populations—for example, hav-
ing two copies of an FTO gene variant
is associated with significantly higher
levels of obesity and type 2 diabetes in
Western populations but not in the Han
Chinese (Li et al., 2008a). Furthermore,
the relationships between individual
genomic and phenotypic variations in
response to drug treatment is still poorly
understood, and it is unlikely that genetic
information alone will be able to direct
personalized drug therapy.
The metabotype or metabolic pheno-
type (Box 1) provides a readout of the
metabolic state of an individual and is
the product of genetic and environmen-
tal (diet, lifestyle, gut microbial activity)
contributions under a particular set of
conditions (Nicholson, 2006). Analyzing
metabolites (small molecules <1 kDa) in
body fluids such as urine and plasma
using various spectroscopic methods
provides knowledge of the metabotype.
Such metabolic profiles provide infor-
mation that cannot be obtained directly
from the genotype, gene expression pro-
files, or even the proteome of an individ-
ual. These metabolic profiles provide a
top-down “systems level” readout of the
biochemistry, physiological status, and
environmental exposure of individuals
and populations that can be exploited
in personalized medicine and public
healthcare (Nicholson, 2006). Metabo-
lism-driven approaches should prove
highly tractable as they combine the
gathering of systemic information based
on minimally invasive analysis with high-
throughput capabilities.
Metabolome Variation and Metabolic
Profiling
Factors such as gender, age, diet, gut
microbiota, physical activity, latent dis-
ease, medication, hormone, and stress
levels modify the metabotypes of both
individuals and populations (Figure 1),
and hence the prevalence and risk of
disease (Nicholson, 2006; Daviglus et
al., 2004; Li et al., 2008b). Metabolome-
wide association studies (MWAS) have
the capacity to link human metabotype
variations to disease risk factors in the
general population (Holmes et al., 2008).
Xenobiotics such as plasticizers, food
preservatives, pesticides, and plant
secondary metabolites (caffeine, fla-
vanoids, phytoestrogens) also contrib-
ute to metabotype variations as does gut
Metabolic Phenotyping in Health
and Disease
Elaine Holmes,1 Ian D. Wilson,2 and Jeremy K. Nicholson1,*
1Department of Biomolecular Medicine, Division of Surgery, Oncology, Reproductive Biology and Anaesthetics, Faculty of Medicine, Impe-
rial College London, South Kensington, London SW7 2AZ, UK
2Department of Safety of Medicines, AstraZeneca Pharmaceuticals, Mereside, Macclesfield, Cheshire SK10 4TG, UK
*Correspondence: j.nicholson@imperial.ac.uk
DOI 10.1016/j.cell.2008.08.026
Analyzing metabolites (small molecules <1 kDa) in body fluids such as urine and plasma using
various spectroscopic methods provides information on the metabotype (metabolic phenotype)
of individuals or populations, information that can be applied to personalized medicine or public
healthcare.
Box 1. Glossary
Conditional metabolic phenotype: A characteristic metabolite profile reflecting the host
genome and its interaction with environmental factors, diet, and the gut microbiome.
Co-metabolite and co-metabolome: Compound or set of compounds derived from interac-
tions of more than one genome in symbiotic systems.
Metabolome: A quantitative description of all endogenous low-molecular-weight compo-
nents (<1 kDa) in a biological sample such as urine or plasma. Each cell type and biological
fluid has a characteristic set of metabolites that reflects the organism under a particular set
of environmental conditions and that fluctuates according to physiological demands. The
metabolome can be divided into the primary metabolome (controlled by the host genome) and
the co-metabolome (dependent on the microbiome).
Metabonome: Theoretical combinations, sums, and products of the interactions of multiple
metabolomes (genome, symbiotic, parasitic, environmental, and co-metabolic) in a complex
system.
Metabonomics: The quantitative measurement over time of the metabolic responses of an
individual or population to drug treatment or other intervention.
Microbiome: The consortium of microorganisms, bacteria, protozoa, and fungi that live com-
mensally or symbiotically with a host.
Pharmacometabonomics: Prediction of the quantitative outcome of a drug intervention in an
individual based on a pre-dose mathematical model of their metabolic state.
Xenometabolome: Characteristic profile of nonendogenous compounds (drugs and their
metabolites, pollutants, dietary components, herbal medicines) in a biofluid.

Page 2

Cell 134, September 5, 2008 ©2008 Elsevier Inc. 715
microbiome activity through the produc-
tion of co-metabolites (Li et al., 2008b;
see Essay by Turnbaugh and Gordon,
page 708 of this issue). The composition
of the microbiome (Box 1) has an impact
on a variety of disorders from intestinal
disease to obesity and cancer. Under-
standing variation of the co-metabolome
(Box 1) will shed light on how an imbal-
ance in our microbiota may contribute to
disease.
Most analytical platforms for meta-
bolic profiling are based on spectro-
scopic techniques, e.g., nuclear mag-
netic resonance (NMR) spectroscopy
or mass spectrometry (often combined
with chromatographic separation), to
permit extensive and rapid analysis of
small molecule metabolites, generally
with minimal sample preparation (Nich-
olson et al., 2002). These approaches
provide relatively high-throughput analy-
ses at low cost per sample. Multiparam-
eter datasets containing quantitative
information on a comprehensive range of
classes of small molecules can be pro-
duced, often in a nontargeted manner.
Given the complexity and high-density
nature of the spectral data, many com-
puter-based preprocessing and multi-
variate modeling techniques have been
developed to facilitate the analysis and
interpretation of these data, in terms of
the response of an organism to physio-
logical or pathological events; such bio-
informatics techniques can also be used
to probe gene-environment interactions.
Mathematical modeling of metabolic
data enables diagnostic characteriza-
tion allowing identification of potential
biomarkers, which may contribute to
a better understanding of pathological
processes and potentially to the elucida-
tion of new drug targets (Nicholson et al.,
2002).
Discovery of new metabolic biomark-
ers using untargeted metabolic profiling
could be translated into clinical tools for
application to personalized medicine. For
example, we have demonstrated proof-
of-principle of pharmacometabonomics
(Box 1), in which a mathematical model of
a group of pre-treatment metabolite pro-
files in rats was used to predict post-drug
interventional outcomes such as toxicity
and metabolism of the xenobiotic itself
(Clayton et al. 2006). Human variability
is greater than that of experimental ani-
mals and this poses problems as well as
advantages for modeling, but in principle
the greater the variation in the metabolic
starting point the greater the variation in
resulting observed responses. Clearly, in
a clinical setting, the complex analytical
instrumentation used in the discovery
phase of this type of work would not be
appropriate or affordable for the mea-
surement of these markers in patients.
Implementation of this type of monitor-
ing will depend on evolving technolo-
gies such as “lab-on-a-chip” methods,
sensors and electronic “noses,” which
can detect complex molecular finger-
prints (Westhof et al. 2007) that can be
tuned to detect specific biomarkers (or
pathways) rather than providing compre-
hensive metabolic profiles. Ultimately,
the goal is to have handheld “intelligent”
devices in the clinic, possibly incorpo-
rating microfluidics, that would detect
selected proteins, mRNAs and metab-
olites simultaneously in a pin-prick of
blood. An emerging area that illustrates
how this might be achieved is in breath
analysis using ion mobility spectrometry
(IMS), which has shown promise for the
diagnosis of infections, lung cancer and
sarcoidosis based on detecting a range
of low molecular weight volatile metabo-
lites in exhaled breath (Westhof et al.,
2007).
Another potential strategy is to mea-
sure the chiroptical properties of metab-
olites, an interesting approach given that
different organisms use different forms
of optically active isomers. Indeed, sym-
biotic organisms including man contain
both L- and D-forms of amino acids
reflecting their human and gut microbial
sources. Thus “chiral metabonomics”
using chiroptical spectroscopy or chiral
chromatographic derivitization meth-
ods can be used to differentiate small
molecules originating from mammalian
and microbial metabolism, potentially
helping to elucidate how such symbi-
otic interactions may go awry in intes-
tinal diseases. For example, L-lactate
is a mammalian product of anaerobic
metabolism, but gut bacteria produce
excessive D-lactate in many gastroin-
testinal disorders causing systemic aci-
dosis that cannot be detected with the
standard lactate dehydrogenase assay,
which only detects the L-lactate stereoi-
somer. Methods designed to quantitate
chiral metabolic components will help to
deconvolute human and symbiont con-
tributions to the metabotype and their
relative importance in the etiology of the
condition or response under study.
Applying a Systems Biology
Approach to Medicine
Systems biology is providing many new
tools to describe, model, and visualize
the integrated action of regulatory net-
works at all levels of biological organiza-
Figure 1. Interactions between Genes and the Environment and Their Effects on Health
Metabolic phenotypes (metabotypes) can be measured by profiling small molecules (<1 kDa) in biofluids
(usually urine or blood) by spectroscopic techniques such as NMR. Metabolic phenotypes are influenced
by intrinsic and environmental factors that determine health status and disease risk of an individual or
group. Measuring and modeling the profile of all metabolites (metabolome) in an individual may provide
insights into disease risk factors and etiology, information that could be used for personalized medicine.

Page 3

716 Cell 134, September 5, 2008 ©2008 Elsevier Inc.
tion (Loscalzo et al., 2007). In “top-down”
systems biology we study how biocom-
partments (subcellular, cellular, tissue,
organ) interact in space and time using
minimally invasive methods that capture
the properties of systemic homeostasis
and its dysregulation (Nicholson, 2006).
The efficient integration of “omics” data
and platform outputs via bioinformatics
should provide a deeper understanding
of multilevel connections including gene-
gene and gene-environment interactions
involved in disease development. The
next decade will see the translation of
these tools and approaches into predic-
tive and preventive “systems medicine.”
By extracting lists from gene expres-
sion profiles and metabolic profiles
associated with disease or therapeutic
intervention, disparate biological events
can be connected. Several approaches
have been applied to link genotypes and
phenotypes in animal models. Exam-
ples include reverse genetics, in which
a single gene is ablated and the pheno-
typic consequences ascertained, and
forward genetics, which exploits subtle
allelic variation relying on identification
of mutant strains (and underlying genes)
from the phenotype. Of particular inter-
est are methods that investigate gene-
gene interactions directly by analyzing
and interpreting microarray expression
data against a background of traditional
genetic mapping techniques to estab-
lish associations between quantitative
trait loci and expression data (eQTLs).
Three susceptibility genes for obesity
were identified by an eQTL approach that
predicted transcriptional responses to a
single gene perturbation in mice (Schadt
et al., 2005). This principle was subse-
quently extended to the use of metabolic
profiles (mQTL) as the basis for mapping
loci associated with quantitative changes
in the plasma of a rat strain susceptible to
type 2 diabetes (Dumas et al., 2007). An
alternative approach to forward genetics
has recently been proposed that involves
identification of gene networks that are
perturbed by susceptibility loci that may
lead to the development of disease. This
approach allows the emergent properties
of gene relationships to be discovered
(at least in experimental animals) and the
validation of new genes associated with
disease risk such as those associated
with obesity (Chen et al., 2008).
Tools for modeling metabolic networks
include those in which nodal architec-
tures (corresponding to genes, gene
products, proteins, or phenotypes) are
connected to demonstrate interactions
between those nodes. More sophisti-
cated approaches using hierarchical
modularity network strategies reflect the
clustering of nodes and interconnection
probabilities (Loscalzo et al., 2007). An
alternative approach to integrating mul-
tiple omics datasets is to coanalyze the
disparate matrices to identify those data
that co-vary and those that are unique.
It is important to elucidate how different
levels of cellular nanomachinery interact
in space and time to control metabolism
and how they are disrupted in disease. A
variety of advanced mathematical tools
including linear projection or Bayes-
ian (conditional probabilistic) methods
with bidirectional orthogonal filters (to
remove unwanted noise components
in the data) can be used to identify co-
variation between metabolite profiles
and transcriptomic, proteomic, or meta-
genomic data matrices and so give new
insight into multilevel interactions in
whole organisms and how they relate to
disease (Nicholson, 2006).
Predicting the Outcome of
Therapeutic Interventions
A key goal of modern medicine is to
develop personalized therapies tailored
to an individual’s biology. Patient strati-
fication based on biology (genetic and/
or phenotypic) is usually viewed in terms
of maximizing drug safety and effi-
cacy. Individuals respond differently to
therapeutic interventions, and adverse
drug reactions are a cause for concern,
especially idiosyncratic toxicity that is
revealed when large populations are
exposed to new therapeutics. Variations
among individuals in response to thera-
pies are influenced by differences in the
conditional metabolic phenotype (Box 1).
In the mathematical modeling of these
interindividual variations, the conditional
metabolic phenotype can be thought of
as a starting point (pretreatment) for an
individual in a conceptual multivariate
metabolic space that reflects the com-
bination of many physical, chemical,
genetic, and environmental influences.
It is this starting position (irrespective
of relative contributions of the individual
“vector” components) that determines
the outcome of an intervention. If the
factors that predispose individuals to
such idiosyncratic events could be deter-
mined, then at-risk individuals could be
screened out of a patient population
helping to retain pharmaceuticals that
are of value to the majority of patients.
Most personalized approaches to
drug treatment so far have been based
on measuring genotype variation and
polymorphisms in drug metabolizing
enzymes, such as cytochrome P450
and N-acetyl transferases (Nebert et
al., 2003). However, pharmacogenomic
predictions of drug metabolism and
toxicity have proved disappointing. We
have demonstrated a pharmacometa-
bonomic approach to predicting the
interventional outcome of drugs based
on mathematical models of pre-dose
metabolic profiles (Clayton et al., 2006).
We used a projection-to-latent structure
(PLS) modeling method to predict liver
toxicity in rats given a threshold toxic
dose of the analgesic acetaminophen.
There was a significant association
between the pre-dose metabolite profile
in rat urine and the post-dose outcome
regarding urinary excretion of acetamin-
ophen metabolites and the severity of
liver damage. Although preliminary, this
work shows that there is potential for
applying pharmacometabonomics non-
invasively for screening human popula-
tions. However, a better understanding is
needed about the relationships between
endogenous metabolic status and drug
metabolism outcomes for a wider range
of drugs. Ultimately, a judicious mixture
of pharmacogenomics and pharmaco-
metabonomics will be required to meet
the personalized healthcare challenge of
individualized drug therapy.
Molecular Epidemiology and
Metabolome-Wide Association
Studies
There is considerable geographical
metabotype variation between popula-
tion groups linked to ethnicity and life-
style, and this diversity may be closely
related to disease risk and incidence
(Holmes et al., 2008). There are well-
documented disparities in genetic back-
ground and diet between geographically
dispersed populations, and there are
also regional variations in the incidences

Page 4

Cell 134, September 5, 2008 ©2008 Elsevier Inc. 717
and concentrations of foreign (xenobi-
otic) metabolites. These xenobiotics also
contribute to the human metabotype
and its regional variation in the form of
the foreign compound profile or xeno-
metabolome (Box 1) of an individual or
population.
The goal of MWAS is to perform large-
scale metabotyping of populations using
spectroscopic methods and to relate
these metabotypes to disease risk fac-
tors. The primary advantage of this
approach is that the resulting biomark-
ers are genuine metabolic end-points
and investigations into these pathway
perturbations may yield new therapeutic
targets or insights into dietary interven-
tions. This has been demonstrated in
human populations: a biomarker (for-
mate) identified in the human urinary
metabolome inversely correlates with
blood pressure (Holmes et al., 2008).
Dietary salt intake is strongly linked to
blood pressure; formate is involved in
chloride-ion exchange in the kidney via
the chloride-formate exchanger and the
SLC26 transporters (a series of intercon-
nected renal ion exchangers involved
in Na+ and Cl− transport) (Sindic et al.,
2007). In this case, there is a clear con-
nection between the MWAS-derived
metabolic biomarker and physiological
evidence pointing to its role in blood
pressure regulation. Thus, MWAS stud-
ies have the potential to provide new
insights into disease mechanisms and
pathophysiology that may ultimately lead
to new drug targets. As there are many
hundreds of major epidemiological data
sets and sample repositories worldwide
with known long-term outcomes, there is
a great opportunity to use retrospective
MWAS to identify new prognostic bio-
markers. Given the complementarity of
MWAS and GWAS, future molecular epi-
demiological surveys should be able to
capture both genetic and environmental
information providing fresh insights into
changing disease patterns around the
world and informing personalized and
public healthcare strategies.
Despite the potential for metabolic phe-
notyping, there remain many challenges
both analytical and biological. The sheer
scale of screening required to obtain use-
ful patient stratification information or
epidemiological insights is still formida-
ble, and there is still no established ana-
lytical exploration strategy. However, we
do know that a range of techniques both
“shotgun” and targeted will be required to
give the metabolome coverage needed
to discover new biomarker combinations
for disease risk factors in the general
population. But these are mainly techni-
cal challenges and will inevitably fall with
the march of analytical discovery. Per-
haps the greatest challenge of all will be
data visualization, that is, capturing the
richness of mathematical models of the
metabolic phenotypes of vast numbers of
people so that their biological messages
can be understood by the medical practi-
tioners of the future.
REFEREnCEs
Clayton, T.A., Lindon, J.C., Antti, H., Charuel, C.,
Hanton, G., Provost, J.P., Le Net, J.L., Baker, D.,
Walley, R.J., and Everett, J.E. (2006). Nature 440,
1073–1077.
Chen, Y., Zhu, J., Lum, P.Y., Yang, X., Pinto, S.,
MacNeil, D.J., Zhang, C., Lamb, J., Edward,
S., Sieberts, S.K., et al. (2008). Nature 452,
429–435.
Daviglus, M.L., Stamler, J., Pirzada, A., Yan, L.L.,
Garside, D.B., Liu, K., Wang, R., Dyer, A.R., Lloyd-
Jones, D.M., and Greenland, P. (2004). JAMA 292,
1588–1592.
Dumas, M.E., Wilder, S.P., Bihoreau, M.T., Barton,
R.H., Fearnside, J.F., Argoud, K., D’Amata, L.,
Wallis, R.H., Blancher, C., Keun, H.C., et al. (2007).
Nat. Genet. 39, 666–672.
Holmes, E., Loo, R.L., Stamler, J., Bictash, M., Yap,
I.K., Chan, Q., Ebbels, T., De Iorio, M., Brown, I.J.,
Veselkov, K.A., et al. (2008). Nature 493, 396–401.
Kruglyak, L. (2008). Nat. Rev. Genet. 9,
3114–3118.
Li, H., Wu, Y., Loos, R.J., Hu, F.B., Liu, Y., Wang, J.,
Yu, Z., and Lin, X. (2008a). Diabetes 57, 264–268.
Li, M., Wang, B., Zhang, M., Rantalainen, M.,
Wang, S., Zhou, H., Zhang, Y., Shen, J., Pan, X.,
Zhang, M., et al. (2008b). Proc. Natl. Acad. Sci.
USA 105, 2117–2122.
Loscalzo, J., Kohane, I., and Barabasi, A.L.
(2007). Mol. Syst. Biol. 3, 124.
Nebert, D.W., Nebert, J., and Vessell, E.S. (2003).
Am. J. Pharmacogenomics 3, 361–370.
Nicholson, J.K. (2006). Mol. Syst. Biol 2, 52.
Nicholson, J.K., Connelly, J., Lindon, J.C.,
and Holmes, E. (2002). Nat. Rev. Drug Disc. 1,
153–161.
Schadt, E.E., Lamb, J., Yang, X., Zhu, J., Ed-
wards, S., Guhathakurta, D., Sieberts, S.K.,
Monks, S., Reitman, M., Zhang, C., et al. (2005).
Nat. Genet. 37, 710–717.
Sindic, A., Chang, M.H., Mount, D.B., and Rome-
ro, M.F. (2007). Curr. Opin. Nephrol. Hypertens.
16, 484–490.
Westhof, M., Litterst, P., Freitag, L., and Baum-
bach, J.I. (2007). J. Physiol. Pharmacol. 58 (Suppl
5), 739–751.