Combining clinical chemistry with metabolomics for metabolic phenotyping at population levels.

Introduction: Untargeted metabolic phenotyping (metabolomics/metabonomics), also known as metabotyping, has been shown to be able to discriminate reliably between different physiological or clinical conditions. However, we believe that standard panels of routinely collected clinical and clinical chemistry data also have the potential to provide assay panels that complement metabotyping.

Objectives: To test the above hypothesis and evaluate the use of multivariate statistical analyses to provided panels of clinical/clinical chemistry data measurements that predict the age, sex and body mass index (BMI) of 977 normal subjects and compare these predictions with results acquired by metabotyping on the same healthy individuals.

Methods: Metabotyping involved serum metabolomics using gas chromatography-mass spectrometry (GC-MS) and liquid chromatography-mass spectrometry (LC-MS) previously reported in our HUSERMET study (Dunn et al., 2015), while clinical chemistry data were obtained in clinic for 19 measurements assessing liver and kidney function, blood pressure, serum glucose, cations, as well as lipids. Multivariate analyses involved using support vector machines, random forest and partial least squares, to predict sex, age and BMI. These models used as inputs: (i) the clinical chemistry data alone; (ii) three metabolomics datasets; (iii) combinations of clinical chemistry with the metabolomics data. Model predictions were rigorously validated using 1,000 bootstrapping re-sampling coupled with permutation tests.

Results: Multivariate statistical analyses on the clinical chemistry data obtained for these healthy participants could be used to predict: their sex, based on creatinine; their age, based on systolic blood pressure, total serum protein and serum glucose; as well as BMI using alanine transaminase, total cholesterol (Total-c) to high-density lipoprotein cholesterol (HDL-c) ratio and diastolic blood pressure. Combining clinical chemistry and metabolomics data sets enhanced the predictions of these characteristics. Moreover, this powerful combination allowed for quantitative predictions of age and BMI.

Conclusion: Multivariate statistical analysis on clinical chemistry data from the HUSERMET study obtained similar predictions of age, sex or BMI, compared to metabotyping using GC-MS and LC-MS. These predictions from clinical chemistry data were between 71 and 85% accurate (depending on the MVA used) and compared favourably with metabolomics (71-91 depending on analytical method). Combining clinical chemistry and metabolomics data sets enhanced the predictions of these characteristics to 77-93% accuracy, suggesting that this augmentation of methods may be a useful approach in the search for clinical biomarkers.