Applying metabolomics to cardiometabolic intervention studies and trials: past experiences and a roadmap for the future

Metabolomics and lipidomics are emerging methods for detailed phenotyping of small molecules in samples. It is hoped that such data will: (i) enhance baseline prediction of patient response to pharmacotherapies (beneficial or adverse); (ii) reveal changes in metabolites shortly after initiation of therapy that may predict patient response, including adverse effects, before routine biomarkers are altered; and( iii) give new insights into mechanisms of drug action, particularly where the results of a trial of a new agent were unexpected, and thus help future drug development. In these ways, metabolomics could enhance research findings from intervention studies. This narrative review provides an overview of metabolomics and lipidomics in early clinical intervention studies for investigation of mechanisms of drug action and prediction of drug response (both desired and undesired). We highlight early examples from drug intervention studies associated with cardiometabolic disease. Despite the strengths of such studies, particularly the use of state-of-the-art technologies and advanced statistical methods, currently published studies in the metabolomics arena are largely underpowered and should be considered as hypothesis-generating. In order for metabolomics to meaningfully improve stratified medicine approaches to patient treatment, there is a need for higher quality studies, with better exploitation of biobanks from randomized clinical trials i.e. with large sample size, adjudicated outcomes, standardized procedures, validation cohorts, comparison witth routine biochemistry and both active and control/placebo arms. On the basis of this review, and based on our research experience using clinically established biomarkers, we propose steps to more speedily advance this area of research towards potential clinical impact.


Metabolomics, lipidomics and pharmacometabolomics
Metabolomics is defined as the study of the metabolome: 1,2 the small molecule complement of a biological system (including drug-or microbiome-related metabolites). 3,4 Mass spectrometry (MS) and proton nuclear magnetic resonance ( 1 H-NMR) spectroscopy, utilizing targeted or untargeted methods, are the most commonly used techniques. 2,[5][6][7] Lipidomics, a subset of metabolomics, is the study of triglycerides, sphingomyelins, phosphatidylcholines and others, using MS methods optimized for lipids. 8,9 Epidemiologically, metabolomic methods of phenotyping serum and plasma samples in patients and populations are attractive because they provide a large number of quantitative/semi-quantitative measures relevant to current health status and future health outcomes. The metabolite profile may be influenced by genotype, individual phenotypes, different environmental factors (e.g. diet, activity, smoking, medical treatments) and differences in the microbiome. 10 Therefore, metabolomics may provide greater insight into mechanisms of drug action, particularly in different subgroups of patients. 11 Pharmacometabolomics, the application of metabolomics (and lipidomics) to the study and prediction of variation in drug response, is a relatively new direction. 10,12 One potential application is that the baseline metabolite profile (prior to pharmacotherapy) can predict drug response, in terms of both efficacy and safety, helping to stratify patients most likely to benefit from therapy or helping to select dose or type of therapy. 12,13 Alternatively, it may be possible to use on-treatment changes in metabolites, shortly after therapy has been initiated, to identify good versus poor responders or those susceptible to adverse drug reactions. These patients can then, potentially, be offered alternative therapy or offered therapy at a more appropriate dose, so-called early intervention pharmacometabolomics. 14 For example, prediction of drug-induced liver injury (at baseline or early post-dose) is particularly relevant to clinical trials since it is estimated that 40% of drug candidates that are discontinued in the clinical trial phase are as a result of hepatotoxicity. 15,16 The ability to use pharmacometabolomics to predict drug-induced liver injury due to variation in metabolism of paracetamol was demonstrated by Clayton et al., 2009. 17 This was the first example of pharmacometabolomics being used in humans. 12,18 The value of randomized trial and intervention study biobanks The advantage of randomized controlled trials (RCT) over observational studies is that they provide unconfounded estimates of the effect of an intervention. RCTs with appropriate clinical outcomes, for example cardiovascular disease (CVD) events linked to low-density lipoprotein-cholesterol (LDL-c) lowering, are described as having the highest level of evidence for evaluating causal pathways which integrate biomarkers. 19 Even so, many trials are susceptible to weaknesses such as small sample size (with consequent low statistical power to detect effects of the intervention) and possible bias with, for example, openlabel design. Clinical trials can be designed to investigate the effect of an intervention on surrogate outcomes and metabolic pathways, or on clinically relevant endpoints. Endpoint-driven trials are typically very large and can allow the examination of treatment effects according to various baseline characteristics and, where there are unexpected adverse or beneficial outcomes, novel predictors can also be investigated. The clinical community would • Clinical studies can also be used to identify novel predictors of important outcomes, in the same way as high-quality observational studies.
• Metabolomics in pre-treatment samples can be tested to predict response to therapy (both desired and undesired), but current data in cardiometabolic trials are limited.
• Early intervention metabolomics may identify changes in metabolites soon after initiation of therapy, which may predict later response to therapy or outcomes (both desired and undesired) but, again, current data from cardiometabolic trials are limited.
• Future studies of cardiometabolic trials would benefit from better design (with greater exploitation of placebo-controlled study samples and adjudicated outcomes), larger sample sizes, standardization of procedures, validation of biomarkers in new cohorts and better testing of predictive abilities against current benchmarks to demonstrate real clinical utility. welcome biomarkers that can predict variation in response to therapy. The hope is that more people who would benefit could be appropriately targeted whereas those who would potentially suffer net harm would be spared such therapy: so-called stratified (or personalized) medicine. 11,18,20 The use of a single (or a combination of) technique(s) that provides good coverage of clinically relevant measures of small molecules, such as metabolomics, is thus highly attractive. Even more so if those methods are quality controlled, validated and provide robust identifications and absolute quantitation. 21 However, applications to patient care must be pragmatic and they must meaningfully and cost-effectively guide care to be truly impactful.
Another value of major trials is that they typically offer rich phenotyping of all participants including conventionally measured clinical biomarkers of interest. Therefore, the added information obtained through more costly or novel biomarker approaches, such as metabolomics, can be readily measured against conventional biomarkers. This provides a realistic benchmark to ensure that the novel biomarker(s) add value over and above routine measures and are cost effective. 22,23 Aside from randomized comparisons, the availability of adjudicated outcomes, as well as prospective follow-up in trials, also provides the platform for post hoc observational analyses. It is with this background that researchers value stored biobanks from such large-scale clinical intervention studies to test novel biomarkers for clinical value.

Examples of new knowledge from use of routine biomarkers in clinical intervention studies
To anticipate where metabolomics might be impactful, and to make a useful comparison with existing metabolomics data, evidence for conventional biomarker use in clinical trials (including uncontrolled intervention studies of single arms of these trials) ought to be considered.
So far, the clinical impact of routine biomarkers to predict the degree of treatment response in many fields has been relatively modest. 24 For example, some investigators have argued that following initiation of statin therapy it might be advantageous to track what benefit the therapy is having on an individual patient's risk profile. Although initial trial data putatively supported an early reduction in high-sensitivity C-reactive protein (hsCRP) change as a predictor of subsequent cardiovascular benefit in addition to LDL-lowering, 23,25 further studies from our group and others refuted this notion. [26][27][28] Additionally, preliminary evidence has suggested that liver function tests may predict the degree of CVD benefit by statins, but this observation remains unconfirmed. 29 There are also suggestive trial findings of differential effects based on baseline phenotypes with other lipid-lowering agents, for example that the fibrate class of drugs lowers cardiovascular risk more in those with high triglyceride and low HDL-c, and that ezetimibe prevents cardiovascular disease more in those with diabetes than those without. 30,31 Whether more detailed lipidomic/ metabolomic phenotyping at baseline, or on-treatment, can better inform on benefits/risks of such therapies are questions of interest.
Regarding type 2 diabetes mellitus (T2DM), patients are heterogeneous in terms of weight, insulin sensitivity, beta cell function and renal and liver function. We have limited information on which baseline phenotypes predict treatment responses to a wide range of glucose-lowering drugs with very different mechanisms. Ongoing work, such as in the MASTERMIND study (interventional cross-over study), 32 will attempt to relate baseline phenotypes (including metabolomics) to differential treatment responses to three different classes of oral hypoglycaemic agents. The potential to enrich existing diabetes trial biobanks with metabolomics also exists and is likely to be pursued, with multiple trials now reaching completion.
This article concentrates on metabolomic gains from trials (including uncontrolled intervention studies), but the utility of trial datasets or biobanks with the availability of serially recorded routine biochemistry measures from individuals with adjudicated events should not be underestimated. These have been used to inform on predictors of events or to identify patterns of disease. For example, by exploiting the availability of serial 6-monthly liver function tests from the West of Scotland Coronary Prevention Study (WOSCOPS) (observational nested case control study, n ¼ 946, 4.9 years follow-up), we noted that plasma levels of the liver enzyme alanine aminotransferase (ALT) and fasting triglyceride were the only significant independent predictors of the development of T2DM. As previous data had linked ALT levels to liver fat, these serial data were amongst the first to suggest that hepatic fat accumulation may increase before diabetes development. 33 Such repeated measure datasets and biobanks provide excellent resources to better understand the evolution of novel metabolomics pathways, with potential for linking such changes to events of interest. Similarly, trials have been used to provide more support for cardiac biomarkers, such as N-terminal-pro-B-type natriuretic peptide (NT-proBNP), as independent predictors of CVD events. [34][35][36] Potential of metabolomics in drug trials In terms of the investigation of mechanisms of drug action or pharmacometabolomics, it is thought that changes in metabolite concentrations after drug administration may        be more pronounced (amplified) than changes in the transcriptome or proteome. [37][38][39] The fact that drugs are: (i) transported using metabolite transporters (particularly membrane transporters); 37 (ii) commonly metabolized using (phase I and II) enzymes used in the metabolism of endogenous metabolites; 40 and that they (iii) target enzymes, receptors and transporters that evolved for endogenous metabolites 11 suggests that metabolite concentrations are important predictors of local drug concentration and therefore drug response, and understanding of the pathways involved in drug metabolism and the effect of drug metabolites on endogenous metabolism is important in understanding drug mode of action as well as factors effecting drug efficacy (ultimately concentration of active drug at the site of action) and side effects. 40 Measurement of only a few metabolites, for example routine or emerging biochemical biomarkers, is often inadequate to interrogate drug effects, since most drugs affect multiple interconnected metabolic pathways resulting in multiple metabolic changes. 11 The greater coverage offered by metabolomics offers research potential. For this reason, metabolomics has the potential for clinical applications, but better quality studies are required to test this notion. 41,42 Genomics and pharmacogenomics studies have clearly demonstrated that genetics alone cannot explain all the variation in drug response. 10 Pharmacoproteomics also has potential, but so far has been described as lagging behind genomics. 20 Metabolomics has been described as 'potentially superior' to genomics since it takes into account exogenous information [environmental influences (e.g. diet/ other pharmacotherapies) or interaction with the microbiome] in addition to genetic influences. 10,15,18 Existing examples of use of metabolomics in trials and intervention studies We now describe early examples of the use of metabolomics and/or lipidomics to investigate the effect of cardiometabolic disease-related pharmacotherapies on metabolite profiles in trials (including uncontrolled single-arm interventional studies), or to predict variation in drug response, with some specific examples from the cardiometabolic arena (Table 1). We excluded studies investigating only drug effects on lipoprotein profiles (as previously reviewed [43][44][45].

Strengths of metabolomic studies in trials to date
All of the studies used state-of-the art analytical methods ( Table 1). Many of the studies investigating pharmacometabolomics or effects of pharmacotherapies on the metabolome were well designed and included randomization and control arms (Table 1). Many included important endpoints (such as LDL-c lowering after statin therapy); however, only one had a hard endpoint (CVD outcome in glipizide-treated patients). 46 A number of studies included routinely measured biomarkers for comparison. [46][47][48][49][50][51][52][53][54][55] The majority of studies also adjusted for multiple statistical comparisons (Table 1). Some performed internal model validation 55,56 and two studies performed external model validation. 57,58 Some studies also combined metabolomics data with genomic and proteomic data. 55,57,59 Hypothesis-generating/post hoc studies of the effect of cardiometabolic drugs on metabolomics biomarkers The majority of studies that have measured metabolomics in samples from clinical intervention studies are the result of post hoc or hypothesis-generating studies (Table 1), as often recognized by the authors themselves. In most cases, published data to date are from studies that are small in size (the largest being a study of 272 adults 51 ), consequently they may be limited in power. Despite this some have shown some promise for the future of metabolomics, and the clear need for hypothesis-generating research has been highlighted. 60,61 However, we also need to acknowledge that a higher standard of research is now required to advance the field to a degree where real-world clinical applications of metabolomics can be considered. We therefore highlight examples of weaknesses in existing research that might be improved in the future.
Most studies have investigated how a specific drug may alter metabolic profile (mode of actions studies), but little else (Table 1). Only a few have linked any specific metabolite to clinical outcomes, or to prediction of benefit or harm, primarily due to lack of power. Often only surrogate markers for clinical outcomes were used: LDL-c lowering as a marker for statin-related CVD reduction, 62-65 ALT as a marker for liver injury, 59 platelet aggregation as a marker for anti-platelet effects 57,58 or low cardiac output as a marker for myocardial protection, 52 for example. Whereas these studies are of interest, they are currently limited in terms of clinical translation. 66 Few studies compared metabolite-based biomarkers with routine measures (Table 1). [47][48][49][50][51][52]55 For example, in a lipidomic study of rosuvastatin's effect in healthy volunteers, a number of lipids were found to be decreased. 8 Some decreases will be entirely dependent on increased clearance of LDL particles as a result of rosuvastatin therapy and therefore, in isolation, such data provide limited additional insight into mechanisms of rousuvastatin or clinical implications of treatment. In a previous lipidomics study of the effect of simvastatin in men with mixed dyslipidaemia, reductions in fatty acids (FAs) were associated with increased clearance of LDL particles; no reductions in FAs were observed in non-responders who did not have a decrease in LDL-c. 67 Again, such findings are hypothesisgenerating and require further detailed and robust study.
Approximately half of included studies were post hoc analysis of interventional studies used for observational investigations (no longer placebo-controlled) or are based on analysis of single-arm interventional studies (Table 1). 8,[49][50][51][57][58][59][62][63][64][65] Although they are included in this narrative review, it should be noted that evidence from such studies is potentially biased or confounded. For example, in an investigation of the effect of rosuvastatin on lipidomic profile, changes were compared with baseline concentration only, 8 and in a study of metformin's effect on metabolite profile, metformin-treated individuals were compared with individuals with no therapy. 68 Atenolol is a commonly used selective adrenergic B1 receptor antagonist that is known to increase the risk of hyperglycaemia, impaired fasting glycaemia (IFG) and diabetes mellitus 50 (Table 1). In a nested casecontrol study of the Framingham Offspring Study, a combination of five amino acids was found to modestly predict the risk of diabetes or insulin resistance. 69 Using a hypothesis-driven approach to determine if these same amino acids predicted risk of IFG after atenolol therapy, other researchers performed amino acid analysis (described as targeted metabolomics) in 122 adults treated with atenolol for a mean of 9 weeks as part of a nested case-control study of an interventional study. 50 They demonstrated that the same amino acids could predict individuals most likely to develop IFG after atenolol treatment. The authors suggested that treatment with drugs such as atenolol may be an environmental trigger in these at-risk individuals. However, as the study only included participants from the atenolol arm (with no control arm), it is not possible to speculate on how many individuals would have developed IFG anyway, nor whether usual predictors of diabetes would have given the same or better information on diabetes prediction.
Some studies included only healthy individuals (for example studies of fenofibrate 49 and aspirin 57,58 ). Changes in metabolite concentrations in healthy individuals may differ from changes in metabolite concentration in intended recipients of pharmacotherapy. 51 Other studies were limited by the availability of time-matched samples. In a study of ximelagatran toxicity, samples from cases of drug-induced liver injury were not available from controls at the same time points, so causal inferences cannot be made with any confidence. 59 In some studies, patients in different treatment arms were not well matched at baseline, for example gender imbalance in a study of atenolol in African Americans versus Caucasians, 51 and age imbalance between healthy volunteers and patients with T2DM in a study of rosiglitazone. 70 There is also the strong likelihood of publication bias, with negative studies being very slow to publish or left unpublished. 71 Hence, most of the studies published to date should, at best, be described as hypothesisgenerating, and few as yet have added clinically meaningful insights.
Finally, a recent observational study took advantage of a parallel genetic analysis to show almost identical changes in lipoprotein and lipidomic profiles with statin use (versus non-use) and corresponding genetic instruments (HMGCR rs12916 variant known to mimic HMGCR inhibition). 72 No robust changes were observed for the metabolites measured (amino acids, ketones, glycolysis-and gluconeogenesisrelated metabolites) in the statin-treated or the rs12916 variant groups, suggesting minimal pleiotropic effects. This study is therefore an exemplar for future combined genetic/ phenotyping studies, so called 'natural' or in silico clinical trials, to determine metabolomics effects of drugs. 73

Examples of metabolomics studies limited by lack of clinical utility
As is the case for many drugs, there is clinical variation in the effects of statins on LDL-cholesterol. In one study, decreases in LDL-c ranged from less than 5% to more than 60% on simvastatin 40 mg daily (however, LDL-c lowering was as expected in the majority, i.e. > 30% in > 75% of participants, and > 20% in > 90% of participants) even when concordance with therapy was apparently accounted for. 65 Investigators have therefore sought to determine whether metabolomics can help predict degree of statin efficacy in terms of LDL-cholesterol lowering (Table 1). [62][63][64][65]74,75 Here, preliminary evidence indicates secondary bile acids and amino acids produced by the gut microbiome may contribute to the inter-individual variation in LDL-c lowering response. [62][63][64] Baseline concentrations of secondary bile acids and other metabolites have been found to predict the circulating statin concentration and the degree of LDL-c lowering by statins, 62,63,65 but more robust studies are needed to confirm these findings.
Whereas these studies are of academic interest, our view is that the use of pharmacometabolomics in prediction of statin LDL-c-lowering response in clinical practice is likely to be impractical, given the huge numbers of patients eligible for statin therapy, the low cost of treatment, the safety of statins and the LDL-c lowering that is likely to occur in the vast majority. Patients routinely start on a standard statin dose (e.g. atorvastatin 20 mg/day for primary prevention and atorvastatin 80 mg/day for secondary prevention) unless there is potential for drug interaction or there appears a high risk of adverse events. 76 If a patient is found to have responded poorly to statin therapy, i.e. a lack of LDL-c reduction (or better non-HDL-c change) using routine biochemistry analysis, 77 adherence and diet/lifestyle changes are discussed and the dose can be increased. It is unlikely that the current cost of predicting patient response using metabolomics, even if this were possible with a high degree of accuracy, whether at baseline or early post intervention, would be clinically effective or cost effective in this regard. Similar limitations have been outlined for the use of pharmacogenomics on the prediction of optimal statin therapy. 78 Interestingly, the Clinical Pharmacogenetics Implementation Consortium has issued guidelines for using SLCO1B1 to inform simvastatin dose, but only when genetic data are already available. 79 In our view, metabolomics measurements to predict response to therapy may well have a place with medicines where risk/benefit is more closely balanced, and where response to therapy is more variable. We believe that there is an important distinction to be made between cheap, safe and clinically effective pharmacotherapies such as statins and more expensive, risky pharmacotherapies that would more likely benefit from tailoring, such as oncotherapy. In such situations a strong case can be made for detailed phenotyping to allow selection of individuals most likely to benefit.
Certainly there is a variation in response to statins and this does correlate with efficacy in terms of CVD risk reduction. 80 However, in many cases where expected response to statin therapy is not observed, poor adherence to therapy or discontinuation of therapy is the usual cause. 81,82 Such cases will not be helped by pharmacometabolomics. It is also important to note that although on-treatment reduction in LDL-cholesterol is only one measure of statin effectiveness, it is a simple, cheap and effective measure of response (and adherence) to statin therapy. 83 Most importantly, we note that there are currently no lipid-modifying therapies that can compete with statins in terms of both clinical and cost effectiveness. There is no evidence that a patient who does not respond well to a statin would respond better to ezetimibe or bile acid sequestrants. Withholding statin therapy based on pharmacometabolomics and treating with such alternatives would likely be unethical. New LDL-cholesterol lowering therapies such as PCSK9 inhibitors may be more widely prescribed in future (currently limited to patients with familial hypercholesterolaemia and those receiving statin therapy for secondary prevention who do not achieve LDL-c targets). 80 These agents have to be injected, are hugely expensive and have not yet been proven to reduce CVD risk. Giving patients such therapies based on pharmacometabolomics results is again unlikely, in our view. It remains to be shown (in clinical trials) that pharmacometabolomics is superior to monitoring of LDL-cholesterol response, and we would argue that given the low costs and effectiveness of statin therapy, and the low cost of monitoring, this is unlikely to be the case.
Despite our reservations concerning pharmacometabolomics for the purpose of predicting statin efficacy, we agree that pharmacometabolomics of statin therapy may have a role in identifying potential biomarkers of myopathy (Table 1). 9,75,84 UPLC/MS lipidomics identified changes in phosphatidylethanolamine, long-chain fatty acids, phosphatidylcholine and cholesterol esters, which differed after simvastatin (80 mg/day) compared with atorvastatin (40 mg/day) for 8 weeks in a placebo-controlled study. These drug-specific changes correlated with muscle expression of arachadonate 5-lipoxygenase-activating protein, involved in pro-inflammatory pathways. The authors postulated that the metabolomic approach may provide a sensitive marker of statin-induced changes in muscle and could be used to identify patients at risk of myopathy. The availability of a placebo control group is certainly a strength of the study. However, the study included a small number of participants (37 in total) and results were not corrected for false discovery rate (FDR). Moreover, the study did not include any participants with previous myopathy on a statin and a study in this population would be valuable. Even if some markers are eventually validated as useful predictors of myopathy, the health economic benefits of metabolite measurement would need to be established. Nevertheless, this is an interesting area of future research and may be applicable, in future, to high-risk individuals such as those with impaired hepatic or renal function, those on concomitant drug therapy known to increase the risk of myopathy or those with a family history of statin intolerance or muscle disease. 83 Clearly, for other drugs/interventions where immediate optimal therapy is more critical, or where risk/benefit considerations are more complicated, pharmacometabolomic-guided therapy may be more valuable.

Analytical challenges
There are many challenges in metabolomics and lipidomics (Table 1). 10,85 Most result from the complex mixture of metabolites found in biological samples such as serum and urine. Metabolites have a diverse range of physiochemical properties; therefore no single method can detect all of them. 21 There is a limited ability to detect low concentration metabolites due to limitations in sensitivity, particularly with 1 H-NMR metabolomics, narrowing the detectable dynamic range (ratio of highest versus lowest concentration: e.g. pM to mM). Problems with metabolite identification are common, partly due to limited availability of standards needed to confirm metabolite identity. Absolute quantitation, particularly in untargeted MS methods, is also problematic, again partially due to limited availability of stable isotope-labelled standards needed for quantification by MS due to variability in ion suppression and instrument stability. Inter-batch variation is an issue in MS metabolomics of studies with a large number of samples. 86 Artefactual results due to variation in sample collection, storage and preparation are also known to be a problem for both metabolomics and routine clinical chemistry methods. 5 High cost can be an issue, in particular for MS-based studies, and this often means smaller sample sizes with reduced power are often used, limiting the information gained. 85 The benefits and limitations of MS and 1 H-NMR metabolomics have been highlighted in a number of studies, and are generally described as complementary. 21,87,88 Some of these problems can be overcome, and considerable effort has gone into optimizing available methods. Primarily, the analytical platform that naturally best meets the requirements to answer a specific research question should be chosen. Decisions on which analytical technique to use will often be pragmatic, based on instrument availability, costs, sample volumes or the need for absolute quantitation, high sensitivity, high throughput or improved coverage. 21,89,90 Problems resulting from lack of coverage of the metabolome can be minimized by employing multiple analytical techniques in tandem as demonstrated in several of the studies described. 55,59,[62][63][64][65]68 Alternatively, multiple extraction techniques or columns and mobile phases can be employed in LC-MS. However, this can become costly and time consuming. 89 Additionally, the drug of interest may direct the analytical technique chosen, for example the mechanism of action of lipid-lowering therapies may be more usefully probed using lipidomic methods rather than methods optimized for polar metabolites. However, the bile acid and simvastatin example demonstrates that this is not always the case. 62,64 Problems with identification will be improved as spectral databases such as MassBank and the Human Metabolome Database (HMDB) continue to grow. 91,92 Metabolites can also be identified by following up studies with more detailed methods such as two-dimensional NMR or high mass accuracy MS with fragmentation (MS n ). 40,89 Methods are emerging which allow absolute quantitation of a large number of metabolites by MS by including large numbers of labelled internal standards [Biocrates (current maxmum 180 metabolites and lipids) and Mass Isotopomer Ratio Analysis of U-13 C-Labeled Extracts (MIRACLE)] 93,94 Inter-batch variation in MS can be corrected for statistically using pooled quality control samples and statistical normalization methods, 9,86 although this approach is only satisfactory for research (not clinical) purposes. Artefacts resulting from variation in sample handling can be minimized by the implementation of strict standard operating procedures. 5 Advances in automation (robotic sample preparation, handling and delivery as well as automated data processing) are allowing metabolomics to be used in large-scale studies. 91 This has contributed to the initially high costs associated with metabolomics being reduced. 91 Costs can be further reduced by choosing high-throughput methods, for example, the 1 H-NMR method described by Soininen et al., 2015, which has been used in a large number of observational studies. 42 However, it should be noted that this method quantifies only 18 metabolites due to limited sensitivity of 1 H-NMR, the other metabolic measures being lipoprotein and lipid measures. 21,42 Other methods that have been used in large-scale observational studies include the Metabolon method. 95 This method is comparatively expensive and provides only relative (not absolute) quantification. 95 Accepting that every technique, including those in routine biochemistry, has its limitations, we expect that many of these issues will not necessarily be overcome but minimized. There will always be trade-off between sensitivity, metabolite coverage and precision. 89

Interpretative challenges
There are multiple challenges in the interpretation of metabolomics studies (Table 1). There is the clear potential risk of false-positives and over-fitting models, particularly if there are a large number of 'metabolite measures' in relation to sample size. 96 Another statistical issue relates to the analysis of multiple biomarkers in data that are highly correlated, and avoiding co-linearity in prediction models.
False-positive results can be minimized by appropriate statistical correction for multiple comparisons and by validation of models. 91,96,97 Over-fitting of models can be reduced by cross-validation of the model (using an excluded test set) or, preferably, by validation in an independent external dataset (a separate follow-up study analysed separately). 91,98 Co-linearity issues can been addressed using multivariate methods, such as principal component analysis (PCA) and orthogonal partial least squares discriminant analysis (OPLS-DA). 99 However, such methods are not always appropriate, depending on the research aims. 97,98 Highly correlated metabolites can also be removed from iterative models. 9 More complicated statistical strategies for dealing with 'omic data (such as elastic net) have been reviewed. 100 As clinical trial samples are very valuable and limited, careful use is essential. Most importantly, the studies need to be backed up by relevant clinical or mechanistic questions to avoid generating a series of results that, although of academic interest, do not advance clinical utility or provide mechanistic insights. This can be achieved by interdisciplinary work. 22 Many of these obstacles (both analytical and interpretative) are not unique to metabolomics, and also apply to emerging chemistry-based markers, proteomics, transcriptomics and genomics. 19,79,101 Therefore, both analytical technologies and computational methods are constantly advancing and evolving. It should be stressed that despite these limitations, studies have shown the power of metabolomics in mechanism of action studies to generate novel hypothesis. 10 How can real gains be made?
The issues described above suggest that considerable work is required to better test the clinical utility of metabolomics in clinical trial datasets. We now provide a list of recommendations, in part mirroring a roadmap for implementation of proteomic biomarkers which is equally valid for metabolic biomarkers. 22 i. Analysis of existing larger trials as well as trials including a wide range of age groups, ethnicities and other characteristics, or at least representing the population the drugs are expected to be prescribed for, rather than only healthy volunteers. ii. Use of data from existing trials with hard endpoints and not just surrogate endpoints. This allows testing whether pharmacometabolomics can predict the outcomes that really matter in clinical practice. iii. Use of data from trials with a control arm: either placebo-controlled or a new therapy compared with the current gold-standard pharmacotherapy. iv. Comparison of predictive abilities of potential new biomarkers versus gold-standard conventional biomarkers. Such an approach enables proper assessment of incremental utility gained by novel biomarkers over and above established tests. v. Standardization of methods, including collection and storage of samples, to lessen the effect of measurement artefacts on data. vi. Implementation of fully validated methodology with absolute quantification, batch correction and confident identification where possible, to yield data that can be trusted, compared between studies and interpreted (biologically) more easily. vii. Whenever possible, validation of any novel findings in an independent cohort to give added confidence that results are robust and generalizable.
viii. Economic analysis to estimate whether any incremental benefits justify the increase in costs or to determine realistic prices to enable clinical consideration. ix. Dissemination, communication and working with a range of stakeholders: scientists, patients, clinicians, clinical scientists, regulatory bodies, expert committees, statisticians, health economists, pharmaceutical/biotechnology companies and biobanks. This ensures studies are relevant, well designed and conducted, and useful.

Conclusion
Metabolomics has been applied to biobanked samples from clinical trials in recent years, but selected studies have typically been of a small scale and focused on surrogate outcomes, with the consequence that published evidence has indicated at best only modest potential. If such findings can be extended and properly validated, there remains a potential for metabolomics to aid better tailoring of drugs to patients (beyond those enrolled in clinical trials). Accordingly, pharmacometabolomics has been described as a 'potential gateway' to stratified medicine. 12,13 However, the usefulness of pharmacometabolomics in clinical trials will likely vary from drug to drug, depending on how well safety or efficacy can be predicted from simpler tests, what other options are currently available and the cost of sample analysis. 12,102 Current evidence in this area remains exciting but is largely at the hypothesis-generating or proofof-concept stage. As such, further assessment of its use is required in a larger number of robust studies or trials. For real advances to be made, investigators with metabolomics expertise need to work with clinicians and statisticians to: (i) develop rigorous experimental study designs; (ii) to identify the best biobanks to exploit; and (iii) to carefully outline the key clinical questions that could be usefully addressed by metabolomics at the very beginning of the project. Funding N.R. is supported by the UPBEAT RCT mother-child study; stratifying and treating obese pregnant women to prevent adverse pregnancy, perinatal and longer term outcomes