Metabolomics: The Stethoscope for the Twenty-First Century Open Access

• Metabolomics encompasses the systematic identification and quantification of all metabolic products from the human body.

• This field could provide clinicians with novel sets of diagnostic biomarkers for disease states, in addition to quantifying treatment response to medications at an individualized level.

• Outstanding issues preventing widespread clinical use are scalability of data interpretation, standardization of sample handling practice, and e-infrastructure.

In recent years, the novel field of metabolomics (synonymous with “metabonomics” and “metabolic profiling”), which included in the “omics” subclass of biological studies, has gained increasing attention from both clinical and academic health circles. Metabolomics refers to the “systematic identification and quantification of the small molecule metabolic products (the metabolome, which consists of 40,000 metabolites in humans [1]) of a biological system at a specific point in time” [2]. It has emerged as a fertile area for research and development given its inherently vast translational promise. The fundamental paradigm underpinning clinical metabolic phenotyping is that any localized metabolic, physical, or histological perturbation in the human body will result in global changes characterized in biological samples. These changes are statistically connected to both the disease process and complex gene-environment interactions. Through this, predictions regarding disease risk and treatment responses may be estimated at an individual level [2]. These benefits have tangible health and socioeconomic benefits as effective provision of care improves patient safety and increases the cost-effectiveness of delivered therapies [3].

Further to the fundamental tenants of the field, this technology embraces several concepts popular within contemporary healthcare. Metabolomics feeds into the relatively novel concept of systems medicine [4-6], which supports the importance of viewing each patient as a distinct combination of biochemical, physiological, and environmental interactions. In turn, this holistic appreciation facilitates the paradigm shift in care provision towards a “P4” approach, consisting of prediction, prevention, personalization, and participation [7]. As such, metabolomics is a foundation exemplar of precision medicine [8] and fulfils the needs of offering the powerful strengths of mathematics and analytics to healthcare (iatromathematics) [9]. However, despite the vast translational potential of this field, there is relatively scarce knowledge of this field outside research circles, particularly in comparison to other “omic” based ventures as such the Human Genome Project [10]. As a consequence, the uptake of this potentially disruptive innovation may be hamstrung by a lack of clinician awareness of its pending utility within clinical practice. This gap in collective knowledge may subsequently result in substandard care with respect to both individual and population health as a whole. This review aims to highlight the technology underpinning metabolic profiling, identify potential applications of metabolomics in clinical practice, and discuss the translational challenges that metabolomics faces.

The process of attaining relevant metabolomic data requires 3 steps [11]:

1.Sample collection

2.Sample preparation

3.Sample analysis

4.Sample collection

Metabolic profiling can be performed on both in vivo and in vitro samples [12]. These include a range of samples consisting of cells, fluids, or tissues. In practice, biofluids are amongst the easiest to acquire and work with. Such samples include serum, plasma, urine, saliva, or faecal content. All samples require precise handling as metabolic pathways are highly susceptible to exogenous environmental factors, which can lead to inaccurate results upon analysis. Therefore, for example, maintaining a low temperature as well as ensuring consistent sample extraction is vital for both modalities.

The compounds within a given sample are usually of a complex nature. These sample constituents may be separated to allow for siphoning of pertinent analytes from others that cannot be resolved by the detector [13]. The preparation of the sample is highly dependent on the detection modality that is subsequently used. Of the 2 principles modalities, nuclear magnetic resonance (NMR) spectroscopy and mass spectroscopy (MS) [14], NMR techniques often omit a separation phase.

Sample pretreatment depends on whether a targeted or non-targeted study is being performed [15]. Pertinently for non-targeted studies, minimal pretreatment is desired so that metabolites are not lost. There are a number of pretreatment separation modalities used with MS that hold distinct advantages in characterizing particular aspects of a metabolome.

The principle separation modalities used with MS include the following: gas chromatography allows for gas phase separation of molecules [16]. It is most useful for analysis of trace amounts of volatile compounds. High-performance liquid chromatography [17] uses chromatographic columns, which are filled with microparticles, to allow for high pressure elution of the sample, allowing for increased chromatographic separation. It is the most commonly used analytical technique given its versatility and ability to retain a number of compounds.

Capillary electrophoresis [18] allows for the electrokinetic separation of the sample. It is another versatile technique which enables separation of a wide range of analytes ranging from small inorganic ions to larger proteins.

It should be noted that there are many further pre-analytical considerations of vital importance, which include the collection of samples from either fasting or fed states, the importance of circadian rhythms, the consideration of which additives are added to collection tubes as well as the potential for sample haemolysis. These are outside the scope of this review but are covered in detail in other reports [19, 20].

As mentioned earlier, metabolic profiling is based upon 2 principle analytical modalities: NMR spectroscopy and MS. Both are able to simultaneously identify and quantify information on a wide range of molecules. Moreover, both require only a small amount of sample.

NMR spectroscopy relies upon the ability of spin active nuclei to absorb and re-emit pulsed electromagnetic radiation when placed within a magnetic field [21, 22]. The frequency pattern, a signature which is a consequence of the interaction of the nuclei with the electromagnetic field, provides information regarding the molecular structure, motion, and chemical environment. In biological samples, hydrogen is the most commonly targeted nucleus, due to its abundance in biological samples. Other atoms such as carbon and phosphorus may also be targeted. NMR is perceived to be a highly reproducible and rapid platform which crucially offers a non-destructive method of sample analysis, which may be either fluid or solid. It offers exact quantification of a wide range of chemical structures and remains the only tool that can provide atom-centred information. Practically, it is relatively low cost per sample, requires minimal sample preparation, and offers prompt throughput (15 min per sample). Moreover, samples subjected to NMR analysis may be subsequently further analysed by MS. Deficiencies of this technology are that it is insensitive, requires high user skills, and that there is an initial high start-up cost in order to acquire the instruments.

In comparison, mass spectrometry is a destructive analytical process which relies upon the formation of gas phase ions, which are subsequently separated by their mass/charge ratio. The ions then hit a detector, which accounts for the number of ions for each mass/charge ratio [23]. This is subsequently analysed and compared against available mass spectral databases in order to predict the molecular identity of the constituents. MS is a highly sensitive method of sample analysis, which can be used for targeted and non-targeted analyses. However, the sensitivity and accuracy of the detection are highly dependent upon the experimental conditions and the instrument settings.

There is a need for rapid and accurate statistical tools that can process the complexity and volume of the vast amount of data that is generated. Different metabolomic features may be used as the input for data analysis. These include spectral peak areas, metabolite concentrations, and spectral bin areas [24]. A number of univariate and multivariate statistical approaches can be performed, which also focus upon data pre- and post-processing tasks such as signal extraction/peak detection, noise reduction, correction of run order drifts or batches, and peak fitting. As a class, they are known as chemometric methods [25].

Univariate methods analyse the metabolomic features independently [26]. They are often easier to interpret as they employ more commonly used and understood statistical approaches. However, this analysis does not consider the presence of interactions between different metabolic features. Confounding variables such as gender, diet, or BMI are not accounted for. This leads to an increased probability of incorrect results. The choice of statistical analysis is particularly important given the number of features that are simultaneously analysed, thus increasing the risk of a false-positive statistical result, known as the multiple testing problem. However, it can also be argued that situations which harbour significant confounders are best solved by careful cohort stratification from the outset.

In contrast to univariate analysis, multivariate analysis considers all the imputed metabolomic features and attempts to identify relationships between them [27]. These methods can be classified into 2 groups, supervised and unsupervised methods. Unsupervised methods are able to effectively detect data patterns with biological variables, with the most common unsupervised method being principal component analysis. Supervised methods identify patterns within variables of interest while down-weighting other sources of variance. The most common supervised statistical method is partial least squares regression analysis.

Although there are admittedly limited clinical applications for metabolomics currently, there are a burgeoning number of potential applications that could disrupt clinical practice in the near future. As noted, the conceptual appeal of the field is to provide biomarkers, which may be either predictive, diagnostic, or prognostic.

Metabolomics offers a particularly broad set of oncological applications, particularly with respect to providing serum [28] or imaging-based biomarkers of cancer. Metabolic profiling will assist in the diagnosis of several tumour types. This has been prominently noted with breast cancer in particular. There have been over 30 endogenous metabolites which have been noted in breast cancer specimens, including tCHo levels (resulting from increased phosphocholine), low glycerophosphocholine, and low glucose [29]. Moreover, Bathen et al. [30] stated that a malignant phenotype can be reliably detected against normal tissue with sensitivity and sensibility between 83 and 100% for tumour size, lymph node status, hormone status, and histology. Metabolic signatures have also been mapped for ovarian cancer [31], lung cancer [32], endometrial cancer [33], and colorectal cancer.

There has also been particular focus upon harnessing metabolomics to guide oncological surgery. Inglese et al. [34] highlighted the potential of combining 3D mass spectrometry imaging with unsupervised neural network-based techniques to precisely determine the extent of malignant tissue clusters pre-operatively. Intra-operatively, rapid ionization mass spectrometry has been successfully coupled to electrosurgical tools to allow for near real-time characterization of margins during cautery-led tumour dissection [35, 36]. This device, known as the “intelligent knife” (iKnife) has been tested in vivo and has garnered promising post-operative histopathological support [37]. However, current limitations with respect to the iKnife include the costly start-up costs as well as the need for robust, histologically specific mass spectral libraries.

In addition to diagnosis, metabolic profiling can be used in prognosticating clinical outcomes. NMR-based techniques have predicted which samples of human glioma cell cultures would be drug sensitive and drug resistant prior to treatment with either chemotherapy or hormonal therapy. Glunde et al. [38] demonstrated that a decrease in tCHO signal equates to a promising response to chemotherapy and may be an early marker of therapeutic effect, which is detectable prior to changes in conventional imaging in breast, brain, or prostate cancer. Moreover, micrometastases were predicted in a study of patients with breast cancer [39], who were noted to have higher levels of plasma glucose proline, lysine, phenylalanine, N-acetylcysteine, and lower lipid levels. Furthermore, Tenori et al. [40] demonstrated that pretreatment serum samples for patients with metastatic breast cancer are predictive for overall survival, time to progression and treatment toxicity, according to serum phenylalanine, glutamate, and glucose levels respectively.

Quantifying the individual risk of type 2 diabetes has also been a key target for the field of metabolomics [41]. Wang et al. [42] suggest that the metabolite 2-aminoadipic acid (2-AAA) is a marker of diabetes risk and a potential modulator of glucose homeostasis in humans. They performed a nested case-control study over 12 years on 188 individuals who had developed diabetes and 188 propensity-matched normoglycemic controls identified from the Framingham Heart Study. It was demonstrated that those with 2-AAA concentrations in the top quartile had a >4-fold risk of developing diabetes. It was observed to be elevated up to 12 years before the onset of a clinically appreciable disease state. Moreover, it was noted that 2-AAA concentrated were not well correlated with other metabolite biomarkers of diabetes, which suggests that 2-AAA is from a distinct pathophysiological pathway. Furthermore, 2-AAA treatment enhanced insulin secretion from a pancreatic β cell line as well as from murine and human islets. However, no human trials with 2-AAA analogues have been conducted.

There appears to be a use for metabolomics in the production of diagnostic biomarkers of both inflammatory and non-inflammatory rheumatological conditions [43]. Ouyang et al. [44] demonstrated that serum from patients with systemic lupus erythematosus had significant reductions in valine, tyrosine, phenylalanine, lysine, isoleucine, histidine, glutamine, and alanine amongst others in comparison to patients with rheumatoid arthritis and healthy controls. Kim et al. [45] have shown that over 20 metabolites are potential biomarkers to discriminate rheumatoid arthritis from other conditions such as ankylosing spondylitis, Behçet’s disease, and gout. In addition to these inflammatory conditions, Adams et al. [46] demonstrated that collagen degradation products contribute to an osteoarthritis (OA) signature in blood with identification possible between healthy controls, early OA, and end-stage OA. However, the authors did note that these differences may be secondary to age-related chondrocyte changes.

In neurological disorders, the use of metabolomics is rapidly increasing [47]. In the field of Alzheimer’s disease, identified metabolites are consistently associated with cognition, dementia, and particular lifestyle factors, which suggests that there may be novel targets for the prevention of cognitive decline and dementia [48]. In Parkinson’s disease, metabolomics has deepened the knowledge about alterations in biochemical pathways involved in Parkinson’s disease pathogenesis [49]. Metabolomics enables neurodegenerative disease stratification [50] as well as a diagnosis of disease severity in multiple sclerosis [51]. Another promising area is the use of metabolomics in the early diagnosis of traumatic brain injury (TBI) as well as the patient specific prognosis. Recently, a metabolic signature was identified in the serum of patients after TBI that may be indicative of a disrupted blood-brain barrier, offering a new avenue towards more precise TBI patient stratification in clinical practice [52].

NMR-based metabolomic assays of urine and plasma have demonstrated 3 particular urinary metabolites which are correlated with lung function, in cohorts of patients with and without chronic obstructive pulmonary disease [53]. Of the 3 metabolites, trigonelline was noted to have the strongest correlation with baseline pulmonary function. Increased hippurate and formate were associated with better lung function. It has also been determined that children with asthma and allergic rhinitis may be differentiated from age-matched controls based upon the alkane and aldehyde content of their breath condensate [54].

Dawiskiba et al. [55] undertook a profiling study with inflammatory bowel disease (IBD) and concluded that there is no clear separation between ulcerative colitis and Crohn’s disease; however, they did note evidence of characteristic signatures between active IBD and IBD in remission. These results were contrasted by Williams et al. [56] who stated that it is possible to distinguish Crohn’s disease from ulcerative colitis in a similarly sized age- and gender-matched cohort, based upon choline, lipoprotein, and N-acetylated glycoprotein levels. Between IBD and healthy individuals, there is evidence of increased phenylalanine in the serum samples of patients with IBD, whereas higher glycine and lower acetoacetate levels were observed in urinary samples [57]. With respect to viral hepatitis, urinary metabolomics have shown 11 discriminant metabolites distinguishing patients with hepatitis B from those with cirrhosis or liver cancer [58].

Metabolic profiling of atherosclerosis, the predominant underlying process of cardiovascular disease, has been a longstanding population health priority [59]. Wurtz et al. [60] demonstrated that serum docosahexaenoic acid, glutamine, and tyrosine may be potential predictors for atherosclerosis development. Brindle et al. [61] published a seminal NMR-based study which noted promising separation between blood samples between patients who have coronary heart disease against healthy participants. Sabatine et al. [62] demonstrated that there is an increase in lactic acid and metabolites involved in AMP-mediated skeletal muscle catabolism in both healthy patients as well as those with inducible ischaemia. However, the rise in the citric acid cycle products was seen only in patients with inducible ischaemia. Further metabolic separation was noted by Vallejo et al. [63] in patients who were diagnosed with acute coronary syndrome against healthy controls both acutely and after 6 months. With respect to heart failure, Dunn et al. [64] demonstrated that pseudouridine, 2-oxoglutarate, 2-hydroxy 2-methylpropanoic acid, erythritol, and 2,4,6-trihydroxypyrimidine were noted as potential biomarkers when discriminating between patients with heart failure and healthy individuals.

Given the relative simplicity, safety, and non-invasiveness of the approach, the application of metabolomic analyses is particularly suitable to the field of paediatrics. There is particular use in the diagnosis inborn errors of metabolism [65]. Metabolic profiling has shown promise in detecting putative gain-of-function of prevalent mutations in genes encoding metabolic enzymes such as isocitrate dehydrogenase (IDH). Metabolomics approaches have shown that mutated IDH1 and IDH2 proteins catalyse an additional reaction resulting in the formation of 2-hydroxyglutarate (2-HG), which is a metabolite absent when IDH is not mutated. This “new” metabolite inhibits α-ketoglutarate-dependent dioxygenases that play a key role in regulating the epigenetic state of cells. On the basis of these findings, phase I trials with selective inhibitors of mutated IDHs have been performed in patients with advanced haematologic malignancies, which have demonstrated an objective response rate ranging from 31 to 40% with durable responses (>1 year). Furthermore, IDH inhibitors have demonstrated early signs of activity in solid tumours with IDH mutations, including cholangiocarcinomas and low-grade gliomas [66].

Successful translation requires widespread understanding of the underpinning science, acceptance of standard operating procedures, training in the interpretation of results, and presence of reliable metabolomic data libraries. As such, the transition from a research tool to a viable clinical tool requires further co-ordination across a variety of disciplines on a local level. Currently, issues are related to equipment design, experimental validation, standardization of methods, and interpretability in a reliable and reproducible fashion.

Translation hinges upon several hurdles. The first is the standardization of experimental procedures and equipment, particularly through the adoption of standard operating procedures [22]. Data quality and reproducibility requires standardization. In keeping with this, the NIHR-MRC National Phenome Centre has prioritized the standardization of sample collection and analysis [67]. Targeted metabolomic profiling relies on commercial kits which offer highly standardized measurements on a restricted set of known metabolites. While coverage of the metabolic space is limited, the reduced cost means large cohorts can be characterized. This can lead to valuable insights into disease aetiology. Moreover, owing to their affordability and their highly standardized protocol, these kit-based approaches have gained popularity.

Two such suppliers, Biocrates and Metabolon, offer quantification of a few hundred metabolites in a fast and semi-quantitative manner. In so-called “ring trials,” Siskos et al. [68] have demonstrated a median inter-laboratory coefficient of variation of 7.6% when using different instrumentation but a common protocol. Inter-laboratory reproducibility of targeted platforms [68], which underlines the potential of the Biocrates platform to monitor human serum in plasma-targeted metabolite profiling. The study also provided a better understanding of the platform error model. These are particularly useful for data scientists carrying out meta-analysis and working towards integrating datasets. Several groups have elected to apply this molecular phenotyping technique to large cohorts. A pertinent example is the KORA cohort, which has been extensively used by Jourdan et al. [69] where 1,614 subjects from KORA S4 (aged 54–75 years at the time of examination) and 3,061 subjects in KORA F4 (aged 31–82 years) have had their serum metabolite profile evaluated and correlated to the body fat-free mass index (FFMI), with a particular focus on branched-chain amino acids. Lending weight to previous findings, recently, nearly 10,000 patients have been profiled to test for a link between diabetes type-2 onset and high levels of circulating branched-chain amino acids [70]. Dutch researchers identified genetic contributors to serum metabolite variation based on Biocrates profiling of nearly 7,500 patients [71]. These studies are testament of how metabolite profiling can be used efficiently to support the understanding of disease and how the technique can inform clinical decision in routine practice.

The second major hurdle is in data analysis and interpretation. As noted, the 2 major analytical approaches, NMR and MS, have significantly different requirements. Firstly, for the purpose of data collection, there are sample size and time issues. Data processing is also troublesome as 1 must account for correction of experimental artefacts, equipment differences, batch correction, and normalization. This leads to major differences in the standardization of data for analysis. NMR is intrinsically more reproducible given standard conditions, partly because the sample does not come into physical contact with the detector. MS, despite higher sensitivity, suffers from poorer reproducibility due to changes in instrument response over time and across operators or laboratories [72, 73]. As such, significant efforts are required in mass spectrometry to obtain reproducible data necessary for translation to the clinic.

The third step is the interpretation of data, which requires advanced statistical approaches, such as machine learning. A key stumbling block is the expertise required to fit and interpret these models robustly on a local level if this technology should permeate into routine clinical work. There have been major strides in recent years in both understanding the requirements for data interpretation in this area (eliminating noise, normalization, and non-linear approaches to data analysis) and in the development of new algorithms for metabolite identification [74] and prediction, predominantly through machine learning techniques [75, 76]. In some cases, such as the aforementioned iKnife, this process has been reduced to a binary decision: tumour or not tumour. However, other applications (e.g., analysis of biofluids, such as urine) will probably require standardized application of computational pipelines that result in reproducible and interpretable outputs. This is one of the major aims of the EU-2020 program, PhenoMeNal [77], which is a European e-infrastructure to provide interoperable tools and workflows for modern and large-scale clinical metabolomics. Aside to the incorporation of machine learning, achieving biologically consistent annotation in untargeted metabolomics remains a major challenge within the field metabolomics. The current gold standard for metabolite identification is centred around matching detected features with an authenticated standard, which has been attained using the same methodology and the same equipment. This rigidity understandably leads to several practical challenges when attempting to apply this approach to large datasets. Widely used approaches to overcome this is to use spectral libraries, which are often incomplete, or alternative computational methods, which can match multiple identities with a single feature [78]. This is perceived to be another major hurdle in the widespread dissemination of the technology currently.

The fourth hurdle is centred around the informatic and e-infrastructural support that is required. Many of the human metabolomics data are, as of yet, incomplete and are housed among fractured sets of databases. Initiatives such as PhenoMeNal aim to provide clinical researchers with well-tested and reproducible workflows to perform standard analysis of clinical samples. This includes the computational clustering of metabolomics data for patient stratification as well as workflows for biochemical pathway enrichments. In addition, PhenoMeNal also provides easily accessible computational infrastructures that can sustain large-scale analysis with sufficient compute and storage resources.

The translation of new metabolomics knowledge into health-related practices will require the development of an ethical framework that ensures compliance and enforcement of patient rights. Initiatives such as Tryggve (https://neic.no/tryggve), PhenoMeNal (http://phenomenal-h2020.eu/home/), and Elixir (https://www.elixir-europe.org/) are paving the way towards state-of-the art-scalable cross-border e-infrastructures for efficient, safe, and e-compliant storage of sensitive personal data metabolic phenotyping data. These initiatives are also strongly contributing to establish educational competences to ensure Findable, Accessible, Interoperable, Reusable (FAIR) Data Principles accomplishment in the analysis, sharing, and reuse of sensitive personal data. If governed appropriately, the ultimate direction of these frameworks will allow next-generation big data analysis and artificial intelligence capabilities to combine and coordinate the improved analysis of these metabolomics datasets.

Metabolomic profiling offers a deeper understanding of metabolic and physiological function. Although it is argued by some that metabolomics has been a vital part of medicine for some time now as well-established laboratory biochemical techniques are essentially targeted quantitative metabolomics. However, it is the novel approach to untargeted analysis using emerging quantitative methods that is unique. Moreover, metabolomics offers a paradigm shift in our perception of disease states as we move away from looking for single molecule disease biomarkers and replace this with a search for more complex and dynamic patterns of metabolite concentrations. There is great potential for metabolomics platforms and technology to be harnessed into routine clinical practice to support clinical decisions and to empower patients through a clearer understanding of underlying disease dynamics. While the routine utilization of metabolic phenotyping at a patient and population level requires formalized evidence of effectiveness and safety through large-scale randomized or pragmatic clinical trials, the current wave of evidence is advocating the massive benefits of this approach. Ultimately, metabolomics can offer a tangible route through which to translate the depth of innovation in addressing human health and disease through the aid of big data analytics and the enhanced translation of precision mathematics to precision medicine. By doing so, such a multifaceted and wide-ranging technology assessing metabolism can fulfil the 200-year-old legacy of René Laennec’s original diagnostic device to render metabolomics the stethoscope of the 21st century.

This work was supported by the European Commission-funded H2020 project PhenoMeNal (Grant No. 654241). M.C. acknowledges support of CIBERHD (CIBER de enfermedades hepáticas y respiratorias, Madrid) and of the Icrea Academia and 2017-SGR-1033 (AGAUR, Generalitat de Catalunya). K.K. acknowledges support from the Åke Wiberg Foundation.

The authors declare no conflict of interest.