Metabolomics in cardiovascular medicine: Not personalised, not diagnostic

Dona and colleagues1 recently published a review on metabolomics in cardiovascular research in this journal. They discussed the current status of metabolomics applications in cardiovascular medicine. I agree with many of the authors’ points. However, they gave a rather optimistic view of personalised medicine and presented a multivariate statistical example of metabolomics-based diagnostics. These are concepts that are often unrealistically interpreted. I would like to reflect on related challenges and caveats. Polygenic traits, such as common cardiovascular outcomes, are quantitative traits and, as a result, are continuous by nature.2 The distributions of metabolic phenotypes are also continuous. This is a background that poses a fundamental limitation for both risk assessment and diagnostics.3,4 At the population level, we may identify several biomarkers that relate to an outcome.5 However, the substantial phenotypic overlap between those with the disease (cases) and those without the disease (controls) precludes such cut-off values for diagnostic models that would provide both high sensitivity and high specificity.4,6 While these fundamentals are well recognised in the cardiovascular research arena,5,6 they are continually ignored in many metabolomics applications.3,4,7 At the core of the problem is the lack of quantitative molecular data and the generally accepted convention of expressing the biological motivation as the black-and-white classification of individuals into those with and without a disease. The common solution of applying ‘multivariate chemometrics’ (e.g. orthogonal partial least squares discriminant analysis, OPLS-DA, as illustrated by Dona and colleagues1) typically misguides the analyses and interpretations, particularly when the often (almost) perfect (i.e. highly implausible) classification results are not critically evaluated in light of the analysis method used and with respect to the underlying biology of the application.3,4,7 The key hurdles of this type of multivariate metabolomics applications are currently well identified: overtraining of the classification models with a high number of variables (typically spectral data points); very small numbers of individuals; and the lack of independent biological replications.4,7 It would be beneficial for the increasing numbers of medical scientists who are using metabolomics to have the caveats of multivariate analyses critically communicated to them. Yet, quantitative metabolomics data, as are also called for by Dona and colleagues,1 are directly amenable to standard statistical methods (e.g. linear regression models) with appropriate variable adjustments and replications.5 In order to appraise the true value of metabolomics in cardiometabolic research, we should first tenaciously reject implausible diagnostic applications of metabolomics, and second abandon the current logic of ‘one common risk model for all individuals’. I would anticipate that moving towards the specific molecular characterisation of people as well as disease subgroups would give us the means to improve cardiovascular risk assessments via metabolomics. Nevertheless, the abovementioned fundamental limitations would restrict the applicable results to population subgroups, thereby meeting the challenges of personalised medicine only halfway. The author(s) declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: MAK is a shareholder of Brainshake Ltd (www.brainshake.fi), a company offering nuclear magnetic resonance-based metabolic profiling. The author(s) disclosed receipt of the following financial support for the research, authorship and/or publication of this article: Sigrid Juselius Foundation and Strategic Research Funding from the University of Oulu, Finland, and UK Medical Research Council (grant number MC_UU_12013/1).