Jeremy R. Everett

Medway Metabonomics Research Group, University of Greenwich, Kent, UK

Variable patient responses to drugs are a key issue for medicine and for drug discovery and development. Personalized medicine, that is the selection of medicines for subgroups of patients so as to maximize drug efficacy and minimize toxicity, is a key goal of twenty-first century healthcare. Currently, most personalized medicine paradigms rely on clinical judgment based on the patient's history, and on the analysis of the patients' genome to predict drug effects i.e., pharmacogenomics. However, variability in patient responses to drugs is dependent upon many environmental factors to which human genomics is essentially blind. A new paradigm for predicting drug responses based on individual pre-dose metabolite profiles has emerged in the past decade: pharmacometabonomics, which is defined as “the prediction of the outcome (for example, efficacy or toxicity) of a drug or xenobiotic intervention in an individual based on a mathematical model of pre-intervention metabolite signatures.” The new pharmacometabonomics paradigm is complementary to pharmacogenomics but has the advantage of being sensitive to environmental as well as genomic factors. This review will chart the discovery and development of pharmacometabonomics, and provide examples of its current utility and possible future developments.


Many patients experience little or no efficacy, or even suffer toxicity, when prescribed drugs today. A 1998 study by Pomeranz et al. estimated that in US hospitals in 1994, over 2 million patients had serious adverse drug reactions (ADRs) requiring hospitalization, producing permanent disability, or in an estimated 106,000 cases, that led to death (Lazarou et al., 1998). This is a shocking state of affairs given the advances in twenty-first century medicine. It is estimated that the cost to the US economy of ADRs is between $30 and $100 billion per year (Lee et al., 2014).

There is thus a clear need to be able to personalize medicine to ensure that patients are prescribed medications that will be both efficacious and free of noxious side-effects. Personalized medicine has many definitions including “Application of genomic and molecular data to better target the delivery of healthcare, facilitate the discovery and clinical testing of new products, and help determine a person's predisposition to a particular disease or condition” (Abrahams et al., 2005). It is also known as precision medicine, stratified medicine, or individualized medicine. A simpler definition would be “the use of genomic, molecular, and clinical information to select medicines that are more likely to be both effective and safe for that patient” (Everett et al., 2016). Personalized medicine has a long history, as all good physicians and clinicians will tailor their treatments and medication prescription to the needs of the individual patient. However, since the sequencing of the human genome completed, there has been a growing interest in the analysis of human genetic variations, particularly single nucleotide polymorphisms (SNPs), and the correlation of those variations with drug efficacy and safety. There have also been significant developments in the association of genetic variation with differing metabolite profiles or metabotypes (Holmes et al., 2008) between individuals, in genome-wide association studies (GWAS).

Pharmacogenomics is the study of how genetic variation modulates drug responses between individuals and evidence has accumulated of the involvement of over 2000 genes in drug responses (Salari et al., 2012). However, the successful use of pharmacogenomics in the clinic has been limited (Urban and Goldstein, 2014) and recent reviews of the use of pharmacogenomics in randomized clinical trials in cardiovascular disease (Joseph et al., 2014), type-2 diabetes (Maruthur et al., 2014), and depression (Perlis, 2014) have failed to show clear value.

In principle there are several reasons why pharmacogenomics studies on their own may struggle to predict drug responses in humans: (1) drug absorption, metabolism, and excretion will be subject to environmental factors such as diet, the use of alcohol, the taking of other medications, and the status of the patient's microbiome (Holmes et al., 2012); (2) the detection of upstream genetic differences in a patient indicates that there may be alterations in the patient's downstream metabolic phenotype, not that there necessarily will be: there is not always a fixed relationship between altered genotype and expression of phenotype, and (3) the issue of phenoconversion, induced by drug co-administration (Shah and Smith, 2015), where a genetic extensive metabolizer can be converted into a phenotypic poor metabolizer and thus confound a pharmacogenomics analysis.

In this situation, the use of metabolic profiling to predict drug efficacy and safety has a number of notable advantages. Firstly, the metabolic phenotype reflects the actual physiological status of the patient in real time, not some future possible state. Secondly, metabolic profiling has the ability to be sensitive to both genetic and environmental factors, including the status of the gut microbiome, that are critical for phenotype expression.

Metabolic profiling of biological fluids, tissues and other samples using various technologies has a history that goes back at least several decades (Lindon and Wilson, 2016). The use of these approaches increased significantly in the 1980s with the advent of advanced pulsed Fourier transform nuclear magnetic resonance (NMR) spectroscopy (Lindon et al., 1999) and hyphenated mass spectrometry (MS) (Theodoridis et al., 2011) analytical technologies capable of profiling dozens to hundreds of metabolites in biological fluids such as urine or blood plasma. Early applications were established in the investigation of drug metabolism (Everett et al., 1984), toxicology (Holmes et al., 1992), inborn errors of metabolism (Iles et al., 1985) and the understanding of disease states (Bales et al., 1984). Metabolic profiling is now termed metabonomics or metabolomics (Lindon et al., 2007).

Metabonomics1 has the following interventional definition: “the quantitative measurement of the multiparametric metabolic response of living systems to pathophysiological stimuli or genetic modification” (Lindon et al., 2000). The alternative term metabolomics was coined by Fiehn et al. (Fiehn, 2002) and given the following observational definition: “a comprehensive analysis in which all the metabolites of a biological system are identified and quantified.” The latter definition is potentially less useful due to both its observational nature and the near impossibility of identifying, let alone quantifying, all the metabolites in a complex biological system. Originally the terms were distinct with metabonomics being used for studies of biofluids and tissues, particularly using NMR detection methodologies, and metabolomics being used for studies of plant and cellular metabolites, particularly by MS. The two terms are nowadays used inter-changeably: we will use the original term metabonomics throughout.

The two main technologies used for metabolic profiling studies are NMR spectroscopy and MS, the latter usually in a hyphenated mode with a separation technology such as gas chromatography (GC), high performance liquid chromatography (HPLC), or ultra performance liquid chromatography (UPLC). The key features of these technologies are briefly summarized in Boxes 1, 2 and the interested reader is referred to consult further literature (Lindon et al., 1999, 2007; Theodoridis et al., 2011; Dona et al., 2016).

Box 1. NMR Spectroscopy

Nuclear magnetic resonance (NMR) spectroscopy is the most powerful method for the elucidation of the structure of small molecules in solution and it has an important role in the detection, identification, and quantification of metabolites in biological samples, especially in biological fluids. In a typical one-dimensional NMR experiment a sample of a biofluid in a glass tube would be inserted into a probe placed in a strong [usually 14.1 Tesla (T) or above] and highly homogeneous magnetic field that would induce an alignment of NMR-active nuclei with the magnetic field direction. A short (typically microseconds duration) radiofrequency pulse of the correct power and frequency for a given NMR-active nucleus is then applied to the sample which causes the NMR-active nuclei to move out of alignment with the magnetic field. The relaxation of these excited nuclei back to their ground state induces an oscillating electric current in the receiver coils of the probe of the spectrometer that decays, typically over a few seconds. Fourier transformation of this time-domain free induction decay signal gives rise to the familiar frequency-domain NMR spectrum in which signal intensity is plotted as a function of nuclear resonance frequency. The most commonly studied NMR-active nucleus is the proton, 1H, which is the most sensitive non-radioactive nucleus and is the workhorse of metabonomics studies. Through the use of a reference standard, typically 3-(trimethylsilyl)-2,2′,3,3′-tetradeuteropropionic acid (TSP) or deuterated forms of 4,4-dimethyl-4-silapentane-1-sulfonic acid (DSS) or its sodium salt, the NMR frequencies of the protons / hydrogen atoms are converted into a dimensionless chemical shift measured in parts per million relative to the reference. This chemical shift is constant, no matter what the magnetic field strength.

The 1H NMR spectrum of a biofluid has some remarkable properties:

1. Each chemically non-equivalent hydrogen atom in a metabolite will resonate at its own individual chemical shift which is determined by the chemical nature of that hydrogen and its neighboring atoms in the metabolite and by the solution environment of the sample.

2. Each individual hydrogen atom in a metabolite gives rise to a signal that has an area proportional to the relative concentration of that metabolite in the sample i.e., given certain provisos, 1H NMR spectroscopy is a fully quantitative technology and the signal area for a CH2 group in a metabolite will be exactly double that of the signal area for a CH moiety in that same metabolite. NMR spectroscopy is thus an excellent, quantitative, non-selective detector, and the 1H NMR spectrum of a biofluid will generally equate to the sum of the NMR spectra of all the component metabolites, in proportion to their concentrations in the sample.

3. The nuclei of neighboring, non-equivalent hydrogen atoms in a metabolite will spin-couple to one another, giving rise to 1H NMR signal splittings (also known as J-couplings) that are critical for metabolite identification and structure elucidation. These signal splittings or multiplicities obey an n + 1 rule in first order spectra, where n = the number of equivalent neighboring protons. For instance in the 1H NMR spectrum of lactic acid, CH3-CH(OH)-CO2H, the CH3- methyl group protons will resonate as a 2-line, 1:1 intensity doublet (1 + 1 = 2) at 1.33 ppm due to it having one neighboring hydrogen on the adjacent CH group. By contrast the CH proton will resonate as a 4-line, 1:3:3:1 intensity quartet (3 + 1 = 4) as it has three equivalent hydrogen neighbors. The intensity patterns obey Pascal's triangle. The size of the splittings is also informative and depends on the distance in bonds between the coupling hydrogens and their chemical, stereochemical, and conformational properties. Typically, 2-bond couplings, 2JH, H between non-equivalent hydrogens on the same carbon atom are larger in size than 3-bond couplings, 3JH, H, between hydrogens on adjacent carbons, and coupling over 4-bonds or more are generally much smaller.

Many nuclei including 12C and 16O are NMR-inactive and this may appear to be a disadvantage but in fact, this helps simplify the NMR spectra of metabolites, which would otherwise be hugely complex and difficult to interpret. In metabolite identification, use is made however of the fact that the 1.1% natural abundance 13C isotope is NMR-active, and although insensitive to direct detection, can be readily observed via indirect detection through the 1H nucleus. This is important as the chemical shift range in 13C NMR is 20 times that of 1H NMR and thus much more sensitive to subtle changes in metabolite structure. Two-dimensional NMR experiments (see below) that correlate proton chemical shifts with those of directly bonded or more remote carbons are particularly important in metabolite identification.

Huge advances have been made in NMR spectroscopy in the past several decades due to the development of:

1. Higher magnetic fields enabling greater sensitivity and higher signal dispersion.

2. 2-dimensional NMR spectroscopy enabling the spreading out of the NMR spectra of complex samples such as biofluids into a second frequency dimension and the automated correlation of through-bond or even through-space connectivities between NMR-active nuclei, critical to metabolite identification.

3. Cryo-cooled probes giving much higher spectral sensitivity due to the reduction in thermal noise.

4. Automated and highly stable, digital NMR spectrometers giving unmatched reproducibility, quality, and throughput.

The Achilles heel of NMR spectroscopy has been and still is low sensitivity, due to the fact that the signal in NMR comes only from the tiny fraction of nuclei that are in excess in the nuclear magnetic ground state at spin equilibrium. In general the detection limits are in the range mM to uM whereas MS-based techniques like LC-MS can detect metabolites in the range mM down to pM.

Box 2. Mass Spectrometry (MS)


MS is used as the other main detection technology for metabolic profiling experiments. Mass spectrometers measure the mass-to-charge ratios of charged molecular ions or molecular ion fragments following their ionization in an ion source. In general, MS detection follows a hyphenated, on-line separation step such as GC, HPLC, or UPLC, so that fewer components are introduced into the mass spectrometer at the same time, reducing ionization suppression effects.

After separation by HPLC or UPLC, metabolites are ionized, typically using an ionization technology such as electrospray and often utilizing both positive and negative ion generation and detection. Electrospray ionization (ESI) is a soft ionization method that results in few fragment ions. In the positive ionization mode, ESI+, it will usually give protonated molecular ions [M + H]+, or salt or solvent adducts thereof as the most intense ions in the spectrum. In cases where metabolites are unstable, for instance exhibiting a tendency to dehydrate, it is common to observe protonated, dehydrated molecular ion fragments, [M + H − H2O]+. Correspondingly in negative ion mode the ES- spectrum will consist largely of deprotonated molecular ions or ion complexes.

Separation of metabolites by GC will generally require derivatisation of the metabolites in order to make them volatile enough to travel through the GC column. This can introduce issues of variable metabolite derivatisation both in terms of the efficiency of derivatisation between different metabolites and the possibility that some individual metabolites may exhibit mono-, di-, or poly-derivatisation, resulting in multiple molecular species and spectral interpretation complexity. In GC-MS mode, electron impact ionization is often used, which is a hard ionization technique generating a much greater degree of fragmentation in the mass spectrum and enabling the use of MS libraries for metabolite identification.

The hyphenated mass spectra of a biofluid sample will have the following characteristics:

1. Elution of metabolites from the separation technology at characteristic retention times, allowing confirmation of metabolite identity based on chromatographic retention time, assuming that authentic reference standards of the metabolite are available.

2. Presence of molecular ion, molecular ion adduct or fragment ions that can be characteristic for the metabolite giving rise to them, in terms of mass-to-charge ratios.

3. If experiments are conducted under conditions of high resolution, molecular formula information can be derived on the molecular ions and fragments.

4. Quantitation of a metabolite can be challenging unless a reference standard of the metabolite is available.

5. Sensitivity of metabolite detection is very high: mM down to pM and much superior to NMR spectroscopy.

6. Run-to-run reproducibility is lower than for NMR spectroscopy, due to direct introduction of samples into the spectrometer and the variability that that generates in the spectrometer, particularly in ionization efficiency.

MS is thus highly complementary to NMR spectroscopy and the two technologies are often most powerful when used in concert. Although MS is not a universal detector like NMR spectroscopy, its high sensitivity is a key factor that dictates it choice in many metabolic profiling experiments.

Irrespective of the detection technology used, multivariate statistical analysis (MVA) methods will probably be needed in order to analyse the complex spectra from a metabolic profiling experiment and to determine statistically significant differences between the spectra of, for instance, different groups of patients. These MVA methods are in two main classes: (i) unsupervised methods such as principal components analysis (PCA), where the class of the samples e.g., patients dosed with drug X or patients dosed with placebo, is unknown to the MVA algorithm and (ii) supervised methods, such as projection to latent structures (PLS), where the class of the samples is used as an input to the algorithm. In the case of supervised MVA methods, care must be exercised to avoid over-fitting of the data, and external validation of results with an independent cohort of samples is best practice. Analysis of metabolic profiling data will often start by using an unsupervised method and especially PCA to provide: (i) an overview of the spread of the samples in metabolic space, (ii) a visualization of any separation in metabolic space between subgroups of samples, and (iii) the identification of any significant outliers in the data. In PCA, the algorithm will take linear combinations of the input variables, such as NMR signal intensities, to form a series of principal components (PCs). Each successive PC will explain a decreasing amount of the variance in the data set and will be orthogonal to the preceding PCs. The PCA scores plot shows the relationships between the samples in the study across typically two or three PCs. The loadings plot shows which variables in the input data are contributing the variance observed between the samples. See Figure 1 below for an example of the use of PCA. The interested reader can find further information in the recent literature (Lindon and Nicholson, 2008; Robinette et al., 2013).