Longitudinal Metabolomics Data Analysis Informed by Mechanistic Models

<b>Background</b>: Metabolomics measurements are noisy, often characterized by a small sample size and missing entries. While data-driven methods have shown promise in terms of analyzing metabolomics data, e.g., revealing biomarkers of various phenotypes, metabolomics data analysis can significantly benefit from incorporating prior information about metabolic mechanisms. This paper introduces a novel data analysis approach to incorporate mechanistic models in metabolomics data analysis. <b>Methods</b>: We arranged time-resolved metabolomics measurements of plasma samples collected during a meal challenge test from the COPSAC₂₀₀₀ cohort as a third-order tensor: <i>subjects</i> by <i>metabolites</i> by <i>time samples</i>. Simulated challenge test data generated using a human whole-body metabolic model were also arranged as a third-order tensor: <i>virtual subjects</i> by <i>metabolites</i> by <i>time samples</i>. Real and simulated data sets were coupled in the <i>metabolites</i> mode and jointly analyzed using coupled tensor factorizations to reveal the underlying patterns. <b>Results</b>: Our experiments demonstrated that the joint analysis of simulated and real data had better performance in terms of pattern discovery, achieving higher correlations with a BMI (body mass index)-related phenotype compared to the analysis of only real data in males, while in females, the performance was comparable. We also demonstrated the advantages of such a joint analysis approach in the presence of incomplete measurements and its limitations in the presence of wrong prior information. <b>Conclusions</b>: The joint analysis of real measurements and simulated data (generated using a mechanistic model) through coupled tensor factorizations guides real data analysis with prior information encapsulated in mechanistic models and reveals interpretable patterns.