Temporal response characterization across individual multiomics profiles of prediabetic and diabetic subjects

Abstract

Longitudinal deep multiomics profiling, which combines biomolecular, physiological, environmental and clinical measures data, shows great promise for precision health. However, integrating and understanding the complexity of such data remains a big challenge. Here we utilize an individual-focused bottom-up approach aimed at first assessing single individuals' multiomics time series, and using the individual-level responses to assess multi-individual grouping based directly on similarity of their longitudinal deep multiomics profiles. We used this individual-focused approach to analyze profiles from a study profiling longitudinal responses in type 2 diabetes mellitus. After generating periodograms for individual subject omics signals, we constructed within-person omics networks and analyzed personal-level immune changes. The results identified both individual-level responses to immune perturbation, and the clusters of individuals that have similar behaviors in immune response and which were associated to measures of their diabetic status.

Figure 1

Cohort description. Summary distributions across sexes for (a) Age, (b) observation window, and (c) visits for different conditions. (d) Proportion of time series from different data modalities.

Figure 2

Workflow. Following the initial parsing of multiple omics datasets (i), our workflow has two main branches: (ii) single subject analysis and (iii) multi-subject similarity analysis, with examples of the output shown and relevant figures and tables.

Figure 3

Single individuals’ multiomics clusters. Two examples of Lag 1 classification outcomes are shown for (a) Subject ZKFV71L and (b) Subject ZTMFN3O. In these examples the information is summarized as follows: Left panel: the cluster of groups/subgroups for Lag 1 class are shown in the visits time frame. The visit time points have been labeled by healthy status, where H: Healthy, W: Weight gain/loss, Im: Immunization, In: Infection. Middle panel: the community structure of visits within each subgroup, where the community structure is based on our visibility-graph-based community detection algorithm. Right panel: corresponding autocorrelations for the time series shown.

Figure 4

Similarity analysis across individuals. (a) The k-means based community structure of the subjects’ similarity network (nodes represent subjects and weighted edges omics showing similar temporal behavior across individuals). (b)–(f) Distributions of five types of measures in the 4 network communities by gender: (b) BMI; (c) DI, disposition index; (d) SSPG, steady-state plasma glucose; (e) Matsuda index and (f) isrMax, maximum insulin secretion rate.