Metabolic phenotyping, nowadays most often termed ‘metabolomics’, is becoming increasingly
applied in molecular epidemiology, particularly in concert with genomics. Since Jeremy
Nicholson and colleagues coined the term ‘metabonomics’ in 1999,
1
over 15 000 publications have appeared under this conceptual and technological umbrella
(a Pubmed search on 15 August 2016 at [http://www.ncbi.nlm.nih.gov/pubmed/] with metabonomics
or metabolomics or lipidomics). Most of the published works have been, and still continue
to be, methodologically oriented
2
and thereby bear little direct relevance to applied epidemiology. Particularly the
spectroscopy-based chemometric approaches–typically aiming at classification of individuals
with or without a particular disease–have for a long time (mis)guided metabolomics
research.
3–9
Many of the limitations of these types of multivariate metabolomics applications are
currently well understood: overtraining of classification models with high numbers
of variables (typically spectral data points), cross-sectional study settings with
very small numbers of individuals and no independent replication.
5
,
7
,
8
However, some lack of clarity still remains, partly related to some misplaced conceptions
as to the scope of truly personalized medicine.
10–13
Individual diagnostics of polygenic diseases, when both the disease liability
14
and the metabolic phenotypes
15–17
are continuous, fundamentally preclude diagnostic models that would provide both high
sensitivity and high specificity.
4
,
5
,
7
,
18–20
For example, conditions like autism, long considered rigid disease classifications,
clearly involve a somewhat arbitrary division of a continuously distributed underlying
liability,
21
limiting attempts at improved binary classification. In addition, many metabolomics
applications have ignored confounding in data analyses and interpretations, though
it is well established in observational epidemiology that confounding–by lifestyle
and socioeconomic factors, or by baseline health status, treatment and medication
effects–is prone to affect many associations.
22
,
23
Getting quantitative and molecular
Recent technological developments resulting in increased numbers of quantitative molecular
applications of metabolomics triggered the idea for this themed issue in metabolic
phenotyping in epidemiology. The pivotal role of absolute quantification of identified
molecular entities in epidemiology and genetics is evident from a multitude of recent
applications.
6
,
24–28
The data analysis protocols in mass spectrometry (MS) often build on a quantitative
logic, i.e. identification, assignment and evaluation of specific molecular signals.
Though most of the MS-based studies have been small or moderate in numbers from the
epidemiological perspective, multiple interesting studies have recently been published,
24
,
26
,
29–31
including comparisons between two common commercial MS platforms.
32
On the other hand, a large number of the applications of nuclear magnetic resonance
(NMR) spectroscopy in metabolomics have been spectroscopy-based chemometric approaches.
5
,
33
Multiple highly implausible conclusions, for example in relation to coronary heart
disease and cancer diagnostics, have been published.
3
,
7
However, during the past 20 years NMR-based lipoprotein subclass profiling has been
commercialized and become visible in epidemiology and clinical applications.
34
,
35
NMR spectroscopy can also be used as a general method to quantify multiple molecular
constituents in serum and in other biofluids.
6
,
28
,
36–38
However, few applications apart from focused lipoprotein profiling,
35
,
39
have been published in epidemiological contexts: only one quantitative serum metabolomics
platform being systematically applied in metabolic profiling studies of more than
a thousand people.
6
,
28
,
40
The pros and cons of NMR and MS for metabolic profiling have been extensively covered
in multiple reviews.
2
,
6
,
41–44
MS-based lipidomics in epidemiology is reviewed in this issue by Meikle and co-workers,
45
with the anticipation that further progress is in sight from large, well-characterized
cohorts. In addition, a compelling reminder of the current possibilities of in vivo
metabolic phenotyping, and a vision of how to reach large-scale neurochemical profiling
in epidemiological research, is provided in this issue by McKay and Tkáč.
46
Combining MS and NMR and notes on replication and causality
So far only a few epidemiological studies have combined MS and NMR methodologies.
These are important for increasing the number of metabolic measures studied and also
to validate biomarker findings by different technologies. Recently Wahl and colleagues
47
combined over 400 quantitative measures from serum MS and NMR platforms for over 1600
participants, when studying a multi-omic signature of body weight change in a population-based
cohort. Würtz and colleagues
15
corroborated NMR-based cardiovascular biomarker associations with MS in two population
cohorts with 671 and 2289 individuals. In this issue, Vogt and coauthors
48
characterize associations between serum 25-hydroxyvitamin D concentrations and those
of 415 metabolite and lipid measures, quantified by both NMR- and MS-based platforms
in 1726 people from a population-based study, KORA F4. Importantly, they also replicated
the majority of their findings in an independent population-based study with 6759
individuals for the NMR-based measures and 609 for the MS-based measures. In another
study in this issue, Nelson and co-workers
49
also characterize associations between serum 25-hydroxyvitamin D and serum metabolites,
using primarily a non-targeted MS-based approach with eventually 940 compounds identified
in eight mutually exclusive chemical classes; lipids, amino acids, xenobiotics, peptides,
co-factors and vitamins, carbohydrates, energy metabolites and nucleotides. This is
a remarkable metabolic coverage from a technical point of view. Nevertheless, the
results are weakened by the lack of replication. Comparison of the results from Nelson
and co-workers
49
with those of Vogt and coauthors
48
illustrates the importance of biologically independent data; for example, the association
of 3-carboxy-4-methyl-5-propyl-2-furanpropanoic acid (CMPF) with 25-hydroxyvitamin
D is similarly positive and strong in all three cohorts; however, eicosapentaenoate
(EPA) and docosahexaenoate (DHA), which show strong positive associations in the ATBC
study,
49
do not associate with 25-hydroxyvitamin D in KORA F4.
48
Possible explanation for this discrepancy is the finding in the ATBC study that retinol
influenced the associations between 25-hydroxyvitamin D and EPA as well as DHA.
49
There are of course many possible confounding factors that complicate interpretations
of cross-sectional epidemiological studies, particularly when aiming to infer causality,
22
which Mendelian randomization analyses can attempt to circumvent. A recently published
study, applying the Mendelian randomization framework
50–53
to investigate potential causality in the association between 25-hydroxyvitamin D
and schizophrenia, gives an exemplar of how, in certain cases, potential confounding
by lifestyle factors can be minimized.
54
The study findings suggested that associations between schizophrenia and 25-hydroxyvitamin
D may not be causal, and, therefore the evidential basis for vitamin D supplementation
as a candidate approach for preventing schizophrenia is weakened. Similarly, a recent
extensive Mendelian randomization study provides no support for a causal role for
25-hydroxyvitamin D in the risk of coronary artery disease.
55
On the other hand, in this issue Ong and co-workers
56
found that genetically lowered 25-hydroxyvitamin D levels were associated with higher
susceptibility to ovarian cancer. Along the same lines, genetically lowered 25-hydroxyvitamin
D levels have been found to be associated with increased susceptibility to multiple
sclerosis.
57
Thus, in relation to these outcomes, vitamin D sufficiency may be important in delaying
onset of or preventing the disease and might thus merit further investigations in
long-term randomized controlled trials. These recent applications of the Mendelian
randomization framework are interesting examples of causal epidemiology and a welcome
reminder that observational association studies, such as those regarding 25-hydroxyvitamin
D and circulating lipids and metabolites in this issue,
48
,
49
are often prone to confounding and reverse causation.
22
Furthermore, an important note here is that replication of observational associations
does not mean causation, as previously illustrated in the case of sex hormone-binding
globulin and systemic metabolism.
58
In relation to these key challenges, also in this issue, Swerdlow and co-workers
59
present important guidance on how to plan and interpret Mendelian randomization studies
on disease biomarkers: very relevant in the current era of a wealth of genetic data
from genome-wide studies and the increasing number of quantitative metabolomics studies
in epidemiology.
The use of Mendelian randomization analysis or randomized controlled trials to assess
causality calls for certain prerequisites to be feasible.
53
,
60
,
61
One interesting example in this issue is the study by Wang and co-workers
62
on the effects of hormonal contraception on systemic metabolism. The difficulties
in randomizing women to use hormonal contraception or placebo, and randomizing to
hormonal or non-hormonal contraception, are obvious. Use of Mendelian randomization
analysis to assess the causal effects of oestrogen and progestin is currently limited
by the lack of genetic variants that are reliable and consistent proxies for oestrogen
and progesterone exposure. Furthermore, even were such variants to be identified,
they might not mimic the effects of taking exogenous hormones. To address these difficulties
and strengthen causal inference Wang et al.
62
integrated findings from cross-sectional and longitudinal study settings. They assessed
a comprehensive molecular profile of 75 metabolic measures and 37 cytokines in up
to 5841 women within three population-based cohorts. Women using combined oral contraceptive
pills (COCPs) were compared with those who did not use hormonal contraception. Metabolomics
profiles were also reassessed for 869 women after 6 years to uncover the metabolic
effects of starting, stopping and persistently using hormonal contraception. The extensive
metabolic measurements allowed multiple novel findings on the systemic effects of
COCPs. Perhaps of greater importance to public health, persistent use of COCPs did
not appear to produce cumulative effects over time and the metabolic perturbations
were reversed upon discontinuation.
62
The study by Wang and co-workers
62
applied a quantitative serum NMR metabolomics platform; one that has commonly been
used in epidemiology and genetics.
6
,
40
The platform provides some 150 primary concentration measures. These include a fine-grained
lipoprotein subclass profiling, and quantification of circulating fatty acids, amino
acids, gluconeogenesis-related metabolites and many other molecules from multiple
metabolic pathways, in addition to multiple biomarkers already routinely used in epidemiology.
15–17
,
27
,
28
,
63–67
The primary measures can also be used to calculate many derived measures and metabolic
ratios with potential biological importance. In addition to the study by Wang and
co-workers,
62
three other papers in this issue also apply this methodology, namely the above-mentioned
study on serum 25-hydroxyvitamin D by Vogt and co-authors
48
and two by Würtz and co-workers which look at metabolic profiling of alcohol consumption
68
and metabolic signatures of birthweight in adulthood.
69
Alcohol consumption in nearly 10 000 young adults (in a cross-sectional setting in
three population-based cohorts) was associated with a complex metabolic signature,
comprising both favourable and adverse effects in relation to the risk of cardiovascular
disease and type 2 diabetes.
68
As in the study by Wang and co-workers,
62
Würtz et al.
68
also complemented the cross-sectional study setting with a longitudinal set-up to
be able to better assess the potential causality of the associations. In fact, the
metabolic changes associated with the changes in alcohol intake (during 6-year follow-up
in 1466 individuals) matched well with the cross-sectional results, increasing evidence
that these changes were, at least partly, due to alcohol consumption. The results
of Würtz et al.
68
can be interpreted along the lines of recent extensive Mendelian randomization analyses,
70
suggesting that reduction of alcohol consumption, even for light to moderate drinkers,
is beneficial for cardiovascular health, with the obvious implication that Mendelian
randomization studies of the influence of alcohol on the metabolome would be a natural
extension of this work. The study looking at the metabolic signatures of birthweight
in adulthood in 18 288 people is one the largest epidemiological metabolomics studies
published to date.
69
These data show that lower birthweight is adversely associated with a wide range of
established and emerging circulating cardiometabolic biomarkers in adulthood. However,
the magnitudes of metabolic aberrations were weak (although statistically significant)
and the authors questioned their public health relevance.
69
The pattern of metabolic deviations associated with lower birthweight resembled the
metabolic signature of higher adult body mass index (R2 0.77) with 1 kg lower birthweight
being associated with similar metabolic aberrations as caused by 0.92 units higher
body mass index in adulthood. The authors suggested that shared underlying metabolic
pathways may be involved and concluded that birthweight is only a weak indicator of
metabolic risk in adulthood.
69
Quantitative molecular data–the base for a multitude of statistical and clinical appoaches
When metabolomics gets quantitative,
5
,
6
it no longer matters if the technology is based on MS or NMR (or whatever methodology),
the output just becomes a list of molecular concentrations, the length and details
of which depend on the method in question. This marks a fundamental distinction from
diagnostics-oriented spectroscopy-based chemometric approaches; moving from (potentially
thousands of) spectral data points to (typically up to a few hundred) identified molecular
measures dramatically simplifies the statistical analyses and interpretations of epidemiological
data. Of course, biological and clinical appreciation would need to be integrated
into the metabolomics study rationale to change the search for non-existent binary
disease states in the case of polygenic outcomes.
5
,
6
,
9
Although the transformation from spectra to quantitative molecular measures is a challenge
in itself, it is not discussed here since the topic falls into the domain of analytical
chemistry and a plentiful literature is available.
6
,
34
,
36
,
42
,
71–75
Nevertheless from the epidemiology perspective, with quantitative metabolomics data,
i.e. with that list of molecular concentrations, all is back to basics and business
as usual, apart from the fact that the list of molecular variables is longer than
it has usually been in epidemiological studies. All standard statistical approaches
can be applied in a straightforward manner, including adjustments for potential confounding
factors.
6
,
22
,
24
,
25
,
28
Of course, as with any quantitative molecular data, the analyses are by no means limited
to standard approaches but, for example, multivariate non-linear approaches,
76
,
77
network analyses,
77
,
78
pathway approaches
79
,
80
and integration of multi-omic data
26
,
47
,
75
,
81–86
are all feasible. In fact, this is in contrast to spectral-based approaches in which
these types of analyses, aiming for detailed biological understanding, would mostly
be impossible to perform and interpret at the molecular level. Fearnley and Inouye,
87
in their review in this issue, survey epidemiological studies that leverage metabolomics
and multi-omics to gain insight into disease mechanisms. Whereas they emphasize the
role of quantitative metabolic data in biomarker identification and in understanding
the metabolic underpinnings of diseases, they also underline limitations and discuss
potential solutions in relation to statistical power issues with respect to sample
sizes and limited coverage of relevant metabolites. They advocate a conceptual shift
from metabolite concentrations towards experiments and graph-theoretical analyses
based on the reactions themselves, and envision the identification of subgroups of
individuals enriched for variation in relevant subregions of a reaction network in
population-level epidemiological studies.
87
In another review in this issue, Sattar and colleagues discuss the applications and
use of metabolomics in cardiometabolic intervention studies and trials.
88
They express their concerns regarding the small scale and focus on surrogate outcomes
in most metabolomics studies in the area. Advancing a list of recommendations for
future biomarker studies, they call for multi-expertise research coalitions to work
together for rigorous experimental study designs, with an early focus on truly relevant
clinical questions:
88
the latter an issue recently elaborated in relation to clinical research in general
by Ioannidis.
89
They also present an interesting and critical discussion on the potential role of
metabolomics in predicting drug responses, and elaborate this in the case of statins.
Their points are well made from the clinical point of view and very valuable to consider.
Generally, the idea of individual metabolic phenotypes is alluring
90–92
and might in some cases provide additional and predictive value, for not only drugs
but dietary substances as well.
41
,
93
However this does not necessarily translate into clinical relevance or applicability
and, as the review by Sattar and colleagues
88
indicates, applications of metabolomics in clinical trials are scarce. Recently Würtz
and co-workers published a proof-of-concept study17–a, “natural”, clinical trial of
statin effects
94
–in which they overcame the lack of metabolomics data in randomized controlled trials
by using serially collected blood samples in population-based cohorts, in which a
subset of individuals had started to use statins during follow-up. To verify that
the observed metabolic changes were actually due to the effects of statins, the analyses
were corroborated via Mendelian randomization analyses using a genetic variant in
the HMGCR gene as an unconfounded proxy for the pharmacological action of statins.
In fact, this type of combination of metabolomics data with genetic data in a large
number of individuals readily extends to studies of all drugs with established genetic
proxies mimicking their pharmacological action. With increasing numbers of extensive
metabolomics and genetic data becoming available, we anticipate that comprehensive
metabolic profiles of drug targets are likely to augment drug development in preclinical
stages. Applications of two-sample Mendelian randomization would allow the gene-risk
factor and gene-outcome associations to be taken from different data sources.
95
An early example in drug research is the Consortium for Metabonomic Toxicology, a
collaboration that involved several pharmaceutical companies in applications of metabolomics
to preclinical drug safety studies; this consortium, via the measurement of a dataset
of NMR spectra of rodent urine and serum samples, built a predictive system for liver
and kidney toxicity.
96
However, these approaches have not been widely adopted by the pharmaceutical industry.
This reluctance was likely due to the fact that the methodologies originally used
were spectroscopy-based and chemometrics-driven and thereby not sufficiently sensitive,
quantitative, molecular-specific or platform-independent to permit routine or widespread
implementation.
97
With the recent shift and developments in metabolomics towards quantitative molecular
methodologies,
6
,
32
it may well be worthwhile to revisit this approach. In biomedical omics applications
in general, it is also worth noting that there have been intensive developments regarding
proteomics methods during the past few years, and their quantitative applications
are likely soon to become feasible in large-scale studies to complement the information
from other omics domains.
98–100
Commercial products and peer-reviewed publications
Intellectual property rights for most of the quantitative metabolomics platforms,
particularly those applied in most epidemiological and clinical applications to date,
are owned by companies.
6
,
32
This is not surprising as such, since innovations are, manifestly, often developed
by such enterprises. Ioannidis has recently reflected upon biomedical innovations
and scientific peer review.
89
,
101
He makes the point that ’stealth research’ is prone to create ambiguity about what
evidence can be trusted in a mix of (possibly) ground-breaking ideas, aggressive corporate
announcements and mass media hype.
89
Even more recently, he has called for scientific peer-reviewed articles as a requirement
for technologies aiming to affect health care at large.
101
In this sense–of publishing peer-reviewed scientific articles–the current key metabolomics
methodologies in epidemiology are reasonably represented.
6
,
32
In fact, it appears that many companies operating in the metabolomics arena have made
scientific peer-reviewed articles part of their business strategy. This is logical
in light of the great potential for commercial benefit provided by independent technological
validation by the scientific community. This is something that cannot be achieved
by patents or intellectual property rights alone; high-quality peer-reviewed publications
are hard to copy or buy. However what still partly remains, even in the peer-reviewed
scientific literature, is ungrounded hype and expectations of omics sciences transforming
not only biological knowledge, but also medicine and public health.
102
In relation to clinical applications, we should also keep in mind the problematic
pathway from the announcement of new biomarkers to their integration into predictive
and diagnostic models or validated target identification.
18
,
20
,
103–107
Is the future for metabolic phenotyping in epidemiology precarious?
In the history of science and medicine, remarkable leaps in progress have often been
made due to novel physical or chemical technologies. The new omics methodologies have
already taken a big step, particularly in combination with genomic data, but also
generally in biomedical sciences. Quantitative molecular technologies have been developed,
many of which are ready for integration into large-scale epidemiological studies.
There is much interest in ‘multiomic’ science, in addition to epidemiological and
potential clinical applications. However, considerable hype has also been generated,
often in connection with the concept of personalized or precision medicine and clinical
diagnostic applications. We suggest that for the future meaningful development of
metabolic phenotyping in epidemiological and (potentially) clinical settings, much
of the early metabolomics literature should be put behind us, just as we needed to
discard a large (and largely meaningless) ‘candidate gene’ literature with the arrival
of genome-wide association studies.
108
We should embrace some of the basic requirements of good molecular epidemiology: molecular
identification and quantification, large-scale studies, independent biological replication
of results, and an appreciation of confounding with the aim of understanding causation.
We should combine different metabolomics methodologies and multiomics combinations
to triangulate findings, explore multiple angles and put findings into a biological
perspective. Furthermore we should apply appropriate statistical tools (noting caveats
related to multiple testing) and, maybe most importantly, apply self-critical interpretation
of statistical results, biological implications and potential clinical applicability.
Last, we should carefully reflect–against all the hopes and hypes of personalized
medicine–on the fundamentals of biological processes in polygenic conditions, and
the epidemiological connotations of this with respect to the articulation of population
and individual perspectives.
10
,
11
As demonstrated by many excellent contributions in this themed issue, large-scale
metabolic phenotyping, together with many other omics technologies, are already here
to enrich epidemiology and eventually to make irreversible headway in the era of systems
epidemiology.
Funding
MAK has received funding from the Sigrid Juselius Foundation and strategic research
funding from the University of Oulu, Finland. MAK and GDS work in a Unit that receives
funds from the University of Bristol and UK Medical Research Council (MC_UU_12013/1).
The funders had no role in study design, decision to publish, or preparation of the
manuscript.