Metabolomics enables precision medicine: “A White Paper, Community Perspective”

Metabolomics and the Central Dogma

Recognition of the Central Dogma of molecular biology had a major impact in the life sciences. Indeed, an appreciation for the fact that life arises from chromosomal DNA being transcribed into RNA, which is in turn translated into functional proteins, affords a molecular explanation for the basis of life. It also provided molecular insights into the inherent variation in human susceptibility to diseases and for the differential efficacy of therapeutic drugs in patients. However, the Central Dogma fails to consider that the actual function of genes, transcripts, and proteins is specifically to control the small molecule composition of cells. It is these small molecules that carry out the main work of functioning cells, including regulating the activity of the macromolecules in a complex feedback circuit (Fig. 1). Indeed, it is the interactions of the small molecules with the macromolecular components of the cell that comprise the main determinants of cell function and dysfunction.

Importantly, the metabolome is dynamic and rich, arising only partially as the product of our own gene-encoded proteins. It also arises from the metabolic products of the microbes within us, the air we breathe, the food we eat, and the water we drink (Nicholson et al. 2004; Goodacre 2007; Scalbert et al. 2009; Lindon and Nicholson 2014). The metabolome is thus, incompletely defined and cannot be predicted from knowledge of the human genome, transcriptome, or proteome alone. Yet, despite this complexity and its associated technical challenges to comprehensive quantification, we posit that the metabolome offers the most revealing real time insights toward understanding human disease heterogeneity and variation in response to treatment, and does so at a systems level. It captures important influences on human health that go beyond the genome.

Recent technological breakthroughs have now enabled broad and confident quantification of hundreds of metabolites or relative quantification of thousands of distinct metabolites in complex biological mixtures (e.g., blood and urine). This capability advances a powerful new analytical reality of untargeted metabolite profiling. With the continuing scale-up of current technologies, bolstered by further technological advances on the near horizon, we posit that comprehensive metabolite profiling, applied to samples from large initiatives like the NIH precision medicine initiative cohort (PMI), will provide perhaps one of the most valuable tools for patient stratification, further enabling the promise of precision medicine. Accordingly, to obtain optimal return on investment in large precision medicine initiatives, it is critical that metabolomic analyses be included. While the inclusion of metabolomics could be staged over time, we urge that samples for current studies be collected in a manner that meets metabolomic profiling needs (sample collection, preparation, and storage).

Metabotyping is currently used to screen for inborn errors of metabolism

Metabolism data and inclusion of a targeted metabolic profile in blood has led to the identification of many inborn errors of metabolism (IEMs), a group of monogenetic metabolic disorders that can be lethal in newborns, or result in irreversible organ damage, if not diagnosed and treated swiftly. However, if IEMs are recognized by early screening, many can be controlled with life-saving nutritional supplements and dietary interventions. Using current metabolite profiling platforms, which can now survey thousands of metabolites in microliter quantities of neonatal blood, we anticipate an enormously expanded scope of IEM diagnosis and the discovery of previously unrecognized genetic diseases in the near-term (Schulze et al. 2003; Vernon 2015; Yoon 2015).

Metabotypes for genotypes—the metabolome provides a readout for functions of genetic variants

Over 100 years ago, Archibald Garrod conjectured that “inborn errors of metabolism” are “merely extreme examples of variations of chemical behavior which are probably everywhere present in minor degrees” and that this “chemical individuality predisposition to and immunities from the various mishaps which are spoken of as diseases”. Population based studies collected demographic, health and life‐style related information from thousands of individuals from the general population, and bio‐banked samples of blood, urine, and other body fluids are now analyzed using genomics, transcriptomics, proteomics, metabolomics and other large scale omics technologies (Kastenmüller et al. 2015; Sanseau et al. 2012; Shin et al. 2014; Suhre et al. 2011a, b, 2016; Suhre and Gieger 2012).

These studies have proven Garrod’s conjecture in many instances, showing how genetic predisposition interacts through intermediate metabolic phenotypes with environmental factors and lifestyle choices in the pathogenesis of complex disorders. Figure 2 gives an example of a genetically influenced metabotype (GIM). To date, over 150 such GIMs have been discovered (Kastenmüller et al. 2015) and are now being used to dissect the genetic and environmental factors that trigger complex disorders (Fig. 3). We anticipate that deep metabolic phenotyping of the precision medicine initiative cohort program (and other precision medicine initiatives) will confidently identify fundamental factors in the development of major diseases, lead to the discovery of new biomarkers, and reveal novel targets for clinical intervention. A comprehensive blueprint of human metabolic pathways and linkage to genes and their expression, such as the global human metabolic network (Thiele et al. 2013) that had been created by the systems biology community, would inform strategies for modifying dysregulated metabolites in a rational and targeted manner, potentially using drugs that are pre-existing and safe—as suggested by recent genome-wide association study (GWAS) findings (Sanseau et al. 2012).

Fig. 2

Example of a genetically influenced metabotype (GIM). Fatty acid desaturase 1 (FADS1) catalyzes the desaturation of dihomolinolenoyl-CoA to arachidonoyl-CoA (C20:3 to C20:4). Minor allele homozygotes (7.6 % of the population) of the rs174548 variant convert C20:3 to C20:4 poly-unsaturated fatty acids (PUFAs) about 50 % slower than homozygous carriers of the major allele (52.6 % of the population). The FADS locus has been implicated in multiple GWAS with different cancers, Crohn’s disease and cardiovascular disease traits. This example shows how genetic variance in metabolic traits can be linked to complex disorders to provide a functional understanding of the underlying disease mechanism.

Figure from Suhre et al. 2016

Fig. 3

The metabolic trait is an intermediate phenotype that links the genome, lifestyle and environmental factors to the clinical endpoint. The general concept (a) and an example using information from actual genome-wide association studies with metabolic traits (b). The association of a genetic variant is strongest with its closest intermediate phenotype [IP; for example, the association of fatty acid desaturase 1 (FADS1) with its product–substrate pair], while the association with the clinical end point may be hard to detect at a level of genome-wide significance in a GWAS (P = 0.021 for FADS1 with coronary heart disease). The ensemble of all genetic associations with metabolic traits defines our metabolic individuality and thereby our predisposition to disease. Deep metabolic phenotyping of large precision medicine initiatives allows to identify key factors for the development of complex disorders, which can then serve as biomarkers and targets for clinical intervention.

Figure reproduced from Suhre and Gieger 2012

In summary, the metabolome provides a read out for functions of genetic variants associated with human disease and great mechanistic insights about pathways implicated in disease, disease heterogeneity, disease progression and variation in response to treatment.

Pharmacometabolomics: a detailed biochemical roadmap for defining disease heterogeneity and drug response variation

With significant funding from the National Institute of General Medical Sciences (NIGMS), the Pharmacometabolomics and Pharmacogenomics Research Networks (PGRN and PMRN) have interacted closely over the past 8 years and pioneered how genetic and metabolic data alone, or in combination can inform about treatment outcomes and mechanisms that underlie variation in response to treatments. Over ten classes of therapies were studied in patients to illustrate the concept and its generalizability in human studies for precision medicine approach. This along with few other studies established key concepts and foundations for this new field Pharmacometabolomics (also known as Pharmacometabonomics) Table 1. A historic study in animals by Imperial researchers and their pharmaceutical consortium revealed how metabolomics data at baseline can inform about drug metabolism and toxicity (Clayton et al. 2006).

Pharmacometabolomics evolved as a field that determines the so-called “metabotype” or metabolic state of an individual as affected by environmental, genetic, and enteric microbiome influences (Fig. 1) to study drug responses and to understand treatment outcomes. Metabolic profiles at baseline prior to treatment were shown to inform about disease heterogeneity and treatment outcomes. Also metabolic profiles provide tools for mapping global effects of drugs on metabolism, and for identifying pathways and networks implicated in the mechanisms of a drug’s action and the basis for variation in drug responses.

Examples of totally novel insights about mechanisms of variation of response to drugs used for treatment of neuropsychiatric and cardiovascular diseases were highlighted in recent reviews (Kaddurah-Daouk and Weinshilboum 2014, 2015). Metabolic profiles were shown to provide insights about variation of response to antipsychotics, statins, antidepressants, antihypertensives, antiplatelet therapies, and development of side effects to treatment. A first validation study in humans mapped effects of three antipsychotics in patients with schizophrenia, compared their effects on metabolism and defined signature at baseline implicated in treatment outcomes (Kaddurah-Daouk et al. 2007). The mapping of depressed patient’s metabotypes and the trajectory of biochemical changes induced by SSRI antidepressants (serotonin reuptake inhibitors) have since begun to explain the biochemical basis for delayed response, responses to placebo, and resistance to treatment in major depression. Ethnic basis for variation in response to antihypertensives (beta blockers and thiazide) and gender differences in response to antiplatelet therapies (aspirin) provided first examples of how metabolomics data combined with genetic data could start to define factors that contribute to variation in response to treatment. An important role for the gut microbiome was highlighted by the variation of response to statins (Kaddurah-Daouk et al. 2011b) and for acetaminophen (Clayton et al. 2006, 2009; Winnike et al. 2010). These studies as well as several others applied to the study of cancer chemotherapies (Backshall et al. 2011; Miolo et al. 2016; Stebbing et al. 2012) and others under pharmacometabonomics (for review see Everett et al. 2013; Everett 2015; Huang et al. 2015) all provide support for importance of inclusion of metabolomics data in precision medicine initiatives (Table 1). Metabolomics data captures influences on human health that go beyond genetic makeup.

Metabolomics also has a great potential to be introduced directly into the clinic, as a key profiling and phenotyping platform used to predict patients’ responses to different treatments. This can lead to development of decision support tools for patients and clinicians for the purpose of selecting or recommending optimal treatment regimens (to be used in broad sense of the word, thus including also lifestyle changes). The expectation is that, by using the ‘personalized profiles,’ we will be able to circumvent the commonly applied treatment-failure approach and, thus, ultimately contribute to better patient outcomes (Fig. 4).

Fig. 4

Precision medicine approach using metabolomics as compared to treatment-failure evidence-based medicine approach in clinical practice. ‘Personalized profile’ based on metabolomics as well as other clinical and lifestyle data will be used to predict the patients’ responses to specific treatments and thus help select the best treatment regimens

The gut microflora influence human metabolism and a metabolic profile informs about gut microbiome activity

The human gut microbiome is known to possess metabolic activity considered comparable to that of the liver. The composition of the gut microbiome has wide ranging local and systemic effects. As such, it has been implicated in local disorders, such as inflammatory bowel disorders and an array of systemic conditions, including, but not limited to, disorders of the nervous, cardiovascular, and immune systems (Table 2). Microbiome-associated metabolomics has been employed to investigate and characterize a range of clinically relevant features. In this regard, it is generally considered useful to characterize both the (1) stool/gut metabolome and microbiome, in combination with (2) a readout of the host’s systemic metabolome (e.g., biofluids such as blood, urine and saliva). Clinically-relevant and notable applications of metabolomics to gut microbiome activity are noted below in Table 2.

Thus, the metabolomics task group recommends at least one urine and one fecal sample to be collected from a subset of individuals enrolled in the PMI Cohort and other similar initiatives to evaluate gut microbiome metabolic activity. These samples would ideally be acquired following an overnight fast. Further samples, collected from a subset of the population at selected serial time points would be extremely useful for monitoring responses to environmental perturbations, e.g. how the gut microbiome is affected by cognate drug treatments.

Nutrition-associated metabolic phenotyping

Metabolomics in a nutritional context has already been employed to investigate and characterize relevant features of the metabolic phenotype. Furthermore, diet remains a key environmental factor that needs to be characterized in order to understand its relationship to disease, and to develop a clear and confident public health message in relation to disease prevention. Disease nutrition relationships through metabolomics is not only useful for increasing our understanding of pathology, but also for understanding how the knowledge of metabolism can be employed as a key factor in disease prevention, as well as to optimize human function and performance.

Assessing dietary input accurately is difficult-most studies rely on surrogate methods based on questionnaires, which are notoriously unreliable (Dhurandhar et al. 2015). A metabolic profile of blood and urine contains biomarker information that can, in fact, provide a far more accurate estimate of dietary inputs, which can also correct misreporting on questionnaires. There is ample evidence for the importance of diet in health and disease. Obesity and adiposity have numerous effects on health including CVD and some cancers. For example, adiposity is a significant risk factor for colorectal cancer, which is higher in males than females, and is also linked to diet. Diets rich in red or processed meats increase the risk of colorectal cancer in particular (http://www.cancer.gov/about-cancer/causes-prevention/risk/obesity/obesity-fact-sheet#q6).

Therefore, nutritional metabolomics investigations fall into three major experimental categories: (A) applications to identify dietary biomarkers (B) applications to study metabolic responses to dietary interventions and (C) applications to study diet-related diseases. All of these applications can deliver further information on the impact of one of the key modifiable environmental factors (diet). Some notable applications of this approach are enumerated in Table 3.

Stable isotope tracing in mechanism-based human health

Although tracer methodologies provide more interpretable information and power than profiling (Fan et al. 2012), this is impractical in a large study of this kind. However, such studies for a limited subset (for example, cancer) using tissue biopsies including liquid biopsies (leukocytes) can be envisaged, in which the sample is incubated with an appropriate tracer, and the metabolism of the tissue or cells is analyzed. Small amounts of tissue can now be evaluated in this manner, owing to the technological advances in mass spectrometry in particular. These approaches become especially valuable in longitudinal studies, for example pre- and post-therapy.

In summary, we have taken large steps over the last 15 years to allow the holistic study of metabolites in humans to be performed. We are now at a stage, where robust, reproducible large-scale studies of the role of metabolites as a determinant of human subject variation are becoming practical as goals of the PMI Cohort and other large initiatives. Through the combination of multiple methodologies and platforms, we can currently provide broad coverage of the metabolome and anticipate the future establishment of standardized approaches to be applied at to-be-established high-throughput metabolite profiling facilities worldwide. New developments in technology are expected to allow for an ever-broadening scope of metabolite coverage, dramatic enhancements in sample throughput, establishment of normal human ranges for diverse blood-borne metabolites and more facile ways to move samples from the home to the clinic.

Computational medicine

Metabolic phenotyping has the potential to generate high-volumes of complex spectral data. The process of translating this data into actionable medical information, be it diagnostic, mechanistic, or patient stratification, requires significant computational power and expertise. There are currently several computational frameworks in general use in the metabolomics community; however, none of these methods are particularly focussed on the needs of precision medicine (Xia et al. 2013). As a first step, large-scale metabolic modeling permits to integrate and simulate multiple ‘omics’ data types with metabolic networks (Aurich and Thiele 2016), and it also enables to integrate dietary and genetic information in addition to metabolomic data (Heinken and Thiele 2015). The underlying metabolic models are based on human biochemistry and as such provide mechanistic links between genes, proteins, and metabolites (Aurich and Thiele 2016). There are now studies becoming available that demonstrate the potential of metabolic modeling for precision medicine (Yizhak et al. 2014), particularly when combined with personalized metabolomic data (Aurich et al. 2016). We anticipate that in the immediate future computational medicine will make a quantum leap in practical functionality, combining the robust requirements of classical epidemiology with the flexibility of modern “big data” machine learning algorithms. It will be necessary to map non-linear relationships between metabolomic, genomic, and other omic data, and combine the resulting profiles with bedside clinical metadata, producing the next generation of clinical expert systems.

Compressing such large amounts of high-throughput data into clinically coherent tools will be a major bottleneck in the development of omic precision medicine. Moreover, an omic based prognostic/diagnostic signature cannot be considered a viable “test” without an associated robust computational model. Indeed, the Institute of Medicine (IOM) report Evolution of Translational Omics (Institute of Medicine 2012) defines an omic’ test as “an assay composed of or derived from multiple molecular measurements and interpreted by a fully specified computational model to produce a clinically actionable result”. Thus, the development of rigorous statistical protocols together with task-specific computational models is as important to this field of research as managing the rapid advances in technology.

Recommendations for precision medicine initiatives

• Collect biofluids by applying metabolomics community agreed-upon standard operating procedures. Include members of the metabolomics community in working groups for establishment of optimal sample collection, preservation, sample processing, data acquisition, and omics’ analysis protocols.
• Preserve samples by establishing a reliable and robust biobanking system with strict inventory control. Samples should include plasma (EDTA, or heparin) and preferably also serum or flash frozen samples.
• Consider the additional collection of stool samples and urine samples for sub studies related to nutrition, gut microflora, and other scientifically relevant applications.
• Establish resources and protocols for the distribution of biobanked samples to NIH-supported research programs or similarly funded programs.
• Establish funding opportunities to support personalized medicine challenges using biobanked samples and both untargeted and targeted strategies for large-scale metabolic phenotyping studies.
• Employ NIST Standard Reference material (SRM-1950) plasma samples for standardization and quality control of plasma metabolomics studies over time and across analytical sites.
• Develop NIST Standard Reference material for urine and serum for standardization and quality control of urine and serum metabolomics studies. Develop other standard reference materials to provide additional options.
• Establish rigorous statistical/epidemiological protocols and enable the development of the next generation of computational medicine tools.