Characterizing human postprandial metabolic response using multiway data analysis

Abstract

Analysis of time-resolved postprandial metabolomics data can enhance our knowledge about human metabolism by providing a better understanding of similarities and differences in postprandial responses of individuals, with the potential to advance precision nutrition and medicine. Traditional data analysis methods focus on clustering methods relying on summaries of data across individuals or use univariate methods analyzing one metabolite at a time. However, they fail to provide a compact summary revealing the underlying patterns, i.e., groups of subjects, clusters of metabolites, and their temporal profiles. In this study, we analyze NMR (Nuclear Magnetic Resonance) spectroscopy measurements of plasma samples collected at multiple time points during a meal challenge test from 299 individuals from the COPSAC2000cohort. We arrange the data as a three-way array:subjectsbymetabolitesbytime, and use the CAN-DECOMP/PARAFAC (CP) tensor factorization model to capture the underlying patterns. We analyze thefasting statedata to reveal static patterns of subject group differences, and thefasting state-corrected postprandial data to reveal dynamic markers of group differences. Our analysis demonstrates that the CP model reveals replicable and biologically meaningful patterns capturing certain metabolite groups and their temporal profiles, and showing differences among males according to their body mass index (BMI). Furthermore, we observe that certain lipoproteins relate to the group difference differently in the fasting vs. dynamic state in males. While similar dynamic patterns are observed in response to the challenge test in males and females, the BMI-related group difference is only observed in males in the dynamic state.

Revealing static and dynamic biomarkers from postprandial metabolomics data through coupled matrix and tensor factorizations

“…No statistically significant BMI-related group difference is observed using the patterns extracted by the ACMTF model. Previously, analysis of fasting data in females using PCA showed a weak BMI group difference while the analysis of T 0 -corrected data using a CP model did not reveal any BMI-related component (Yan et al, 2023). Hence, we observe that when jointly analyzing these data sets, the dynamic part blurs the weak signal in the fasting data.…”

“…In our analysis, we include 161 features 2 and 292 subjects (140 males, 152 females) as in (Yan et al, 2023). Yan et al have shown that BMI groups behave differently in terms of metabolic responses among males compared to females (Yan et al, 2023).…”

“…Body composition contains body muscle mass and fat mass which are estimated by bioelectrical impedance using a Tanita scale (TANITA MC-780MA), and consequently computed variables such as body fat percentage, muscle to fat ratio, fat mass index and fat free mass index. For more details about the challenge test, data acquisition and meta data, see (Yan et al, 2023).…”

“…To preserve the multiway structure of time-resolved data, recent work arranges postprandial data as a third-order tensor with modes: subjects, metabolites, and time, and use tensor factorizations, i.e., extensions of matrix factorizations to multiway arrays (Smilde et al, 2004;Acar and Yener, 2009;Kolda and Bader, 2009), to reveal the underlying patterns in subjects, metabolites and time modes (Li et al, 2023;Yan et al, 2023;Fujita et al, 2023). For instance, Li et al use the CANDECOMP/PARAFAC (CP) (Harshman, 1970;Carroll and Chang, 1970) tensor model to analyze simulated challenge test data generated by a human whole-body metabolic model, and demonstrate that the analysis of fasting signals (i.e., T 0 data) using PCA and the analysis of dynamic data (i.e., T 0 -corrected data obtained by subtracting fasting signals from postprandial data) using CP reveal metabolic differences at fasting vs. dynamic states (Li et al, 2023).…”

“…For instance, Li et al use the CANDECOMP/PARAFAC (CP) (Harshman, 1970;Carroll and Chang, 1970) tensor model to analyze simulated challenge test data generated by a human whole-body metabolic model, and demonstrate that the analysis of fasting signals (i.e., T 0 data) using PCA and the analysis of dynamic data (i.e., T 0 -corrected data obtained by subtracting fasting signals from postprandial data) using CP reveal metabolic differences at fasting vs. dynamic states (Li et al, 2023). Such an analysis of time-resolved metabolomics data collected during a meal challenge test has revealed static and dynamic biomarkers of a BMI (body mass index)-related group difference in males (Yan et al, 2023). While these studies show promise in static and dynamic biomarker discovery, the separate analyses using PCA and CP models generate two different segmentations of the subject group.…”