EMBRACING THE UNCORRELATED: DYNAMIC TIME WARPING FOR THE EXPLORATION OF HIGH DIMENSIONAL TIME-SERIES METABOLOMICS DATA FROM THE HUMET STUDY

Abstract

The data analyzed has been collected in the framework of the HuMet study via different metabolomics approaches to assess normal metabolic variation within 15 healthy young men selected to be as homogenous as possible. After comprehensive phenotyping (clinical chemistry, DEXA-scanning, indirect calorimetry) all volunteers underwent four distinct nutritional challenges (with intermittent fasting) over a time period of 4 days in a controlled environment. For the present analysis, metabolites within blood samples collected at 0, 15, 30, 45, 60, 90, and 120 minutes following a 75 g oral glucose tolerance test, were identified and quantified using liquid chromatography tandem-mass spectrometry (LC-MS/MS) as performed by Metabolon, Inc. Time-series metabolomics data, though rare, is often plagued by the ‘curse of dimensionality`. Low observation occurrence coupled with high variable number has restricted the applicability of many temporal analytic methods and increased reliance on feature selection to cope. Non-correlated metabolites naturally are the first to fall victim to aggressive dimension-reducing techniques despite an unexplored translational potential. Interestingly, when utilizing temporal alignment algorithms to correct for lag asynchrony between uncorrelated metabolite series, significant latent trend-shape similarities are uncovered. Within this present project, dynamic time warping (DTW), a temporal alignment measure popular within automated speech recognition, was used to establish a metabolite alignment dissimilarity index with glucose monotonicity. The index was used to evaluate the shape-dependent temporal relationships of 189 metabolites as possible glucose-response variables within a 2 lag parameter. Metabolites with high initial correlation coefficients to glucose (rs > 0.50) were removed, and the remaining indexed metabolites scoring above the median were classified into various combinations to assess their capability to produce subject clusters with high correlation to subject clusters predicated by glucose alone (cophenetic coefficient: c >0.60). The results identified promising targets, as 3-metabolite and 4-metabolite combinations, all with an average rs ≈ -0.10, displayed pronounced efficacy at replicating subject clusters of the glucose-response curves (c ≈ 0.70). This new approach thus defines novel metabolites and clusters in the multi-metabolite dynamics of the glucose response. Acknowledgements : Metabolon, Inc.