Multi-omics microsampling for the profiling of lifestyle-associated changes in health

Abstract

Current healthcare practices are reactive and use limited physiological and clinical information, often collected months or years apart. Moreover, the discovery and profiling of blood biomarkers in clinical and research settings are constrained by geographical barriers, the cost and inconvenience of in-clinic venepuncture, low sampling frequency and the low depth of molecular measurements. Here we describe a strategy for the frequent capture and analysis of thousands of metabolites, lipids, cytokines and proteins in 10 μl of blood alongside physiological information from wearable sensors. We show the advantages of such frequent and dense multi-omics microsampling in two applications: the assessment of the reactions to a complex mixture of dietary interventions, to discover individualized inflammatory and metabolic responses; and deep individualized profiling, to reveal large-scale molecular fluctuations as well as thousands of molecular relationships associated with intra-day physiological variations (in heart rate, for example) and with the levels of clinical biomarkers (specifically, glucose and cortisol) and of physical activity. Combining wearables and multi-omics microsampling for frequent and scalable omics may facilitate dynamic health profiling and biomarker discovery.

Fig. 1. Overview of the microsampling multi-omics workflow and stability analysis.

a, The samples were collected using microsampling devices, and then multi-omics data (proteomics, metabolomics, lipidomics, cytokine and so on) were acquired. b, Outline of the primary microsampling analyses. c, The coefficient of variation (CV) distribution for proteins, metabolites and lipids across all the samples in the stability analysis. d, The percentage of analytes is significantly affected by storage duration, temperature and interactions (linear regression). The red line shows the expected proportion of nominally significant results at the alpha level of 5% (P = 0.05). e, The Spearman correlations between microsamples and intravenous blood samples (n = 34) for metabolites and lipids, respectively. Source data

Fig. 2. The overview of Ensure shake study and molecular response to Ensure shake.

a, The study design and overview of the Ensure shake study. b, The summary of multi-omics data from the microsamples. c, Responses of metabolites, lipids and cytokines/hormones after Ensure shake consumption (two-sided Wilcoxon rank test). d, The clustering of dysregulated molecules following Ensure shake consumption. e, Amino acid response to Ensure shake consumption. f, Response of three dysregulated carbohydrates to Ensure shake consumption. g, Acylcarnitine response to Ensure shake consumption. h, Cytokine/hormone response to Ensure shake consumption. The points are represented by mean ± s.d. Source data

Fig. 3. Metabolic phenotyping based on the multi-omics response to the Ensure shake.

a, The visualization of the AUC metric for each analyte used in a metabolic score. b, The analytes used in calculating each of the six metabolic scores are shown. c, Participants were grouped into five groups on the basis of six metabolic scores. d, Five participant examples for each group. Source data

Fig. 4. Overview of the study design, sample collection and data acquisition for the 24/7 study.

a, One participant was closely monitored using wearable devices and high-frequency microsampling (approximately hourly) across 7 days. Microsamples were then analysed for internal multi-omics data measurements. b, Molecular information was detected from the high-frequency microsamples. c, Wearable data from the smartwatch (sleep and step count) and Dexcom (CGM glucose). Legend defines the status of sleep (note that REM = rapid eye movement), the category of consumed foods, and the day/night period at every record. The yellow background represents the daytime (6:00 to 18:00). d, The internal molecules were grouped into 11 clusters using fuzzy c-means clustering.

Fig. 5. Wearable and internal multi-omics data reflect the individual physiological status and circadian rhythm analysis of multi-omics data.

a, Four molecules reflect the participant’s lifestyle. b, Heatmap to show the rhythmic molecules. c, Three clusters that have strong rhythmic patterns. d, Lipid class distributions of lipids in three clusters. Cer: ceramide; SM: sphingomyelin; DAG: diacylglycerol; LPE: lysophosphatidylethanolamine. e, Examples for each cluster. The yellow background represents the daytime (6:00 to 18:00). Source data

Fig. 6. CGM and internal molecule causal association network.

a, The CGM glucose subnetwork from the whole network. b, Three examples are shown to represent the causal relationships between CGM glucose and internal molecules. NS, not significant. Source data

Extended Data Fig. 1. Stability analysis of the microsampling approach.

a, The study design of the protein, metabolite, and lipid stability analyses in microsamples. b, The partial R² distribution for proteins, metabolites, and lipids. The most affected protein, metabolite, and lipid by storage duration (c), temperature (d) and interaction effect (e), respectively. The icons used in this figure are from iconfont.cn.

Extended Data Fig. 2. Metabolic phenotyping separates samples and subjects.

a, tSNA plot using all samples from all the participants. Colors represent the participants. b, tNSA plots for 6 participants. Colors represent the timepoints. The timepoints are also labeled on the plot. c, Silhouette plots for consensus clustering with group numbers 2, 3, 4, and 5. When the group number is 2, the Silhouette width achieves the highest value, so the group number is set as 2 for subsequent analysis. d, Heatmap plot showing differential clustering of molecular features in various samples compared to baseline (0 min) for each participant. Green represents low distance, and red represents high distance. e, The SSPG values for participants (only 13 participants) in group 1 and group 2.

Extended Data Fig. 3. Wearable and multi-omics data can reflect the individual’s health status.

a, The Spearman correlations between all the clusters from all molecules in 24/7 study using the Fuzzy c-means clustering. b, The mosaic plot shows the molecules’ classes for 11 clusters. c, The maximum modularity observed in our correlation network community analysis for cluster 1 was 0.689 at iteration 72 of community pruning. d, The molecule detection from the correlation network. The molecules have more connections inside than outside and are grouped as a module. e, Cluster 1 and Module 1_4 from it. f, Peak detection from the module using the peak detection algorithm. g, The heatmap to show the association between modules and nutrition. h, MS² spectra matching plots for 1,2,3-benzenetrlol sulfate, Hydroxyphenyllactic acid, and Salicylic acid, respectively.

Extended Data Fig. 4. Wearable and multi-omics data can reflect the individual’s health status.

a, The correlation plot between Caffeine intensity and sleep score. b, The MS² spectra matching plot for Caffeine. c, Molecules that were upregulated from Wednesday to Friday. d, Molecules that were downregulated from Wednesday to Friday. Source data

Extended Data Fig. 5. Circadian rhythm analysis for internal molecules.

a, Consistence scores versus circadian rhythm p-values (-log10). b, Heatmap to show all the circadian molecules. c, Spearman correlation plot to show the correlations between 5 clusters from circadian molecules. d, The components of all 5 clusters. e, Cluster 4 contains 1 cytokine, 3 lipids, 2 metabolites, and 22 proteins. f, Cluster 3 contains 76 lipids, 1 metabolic panel, and 7 metabolites. Source data

Extended Data Fig. 6. Lipid enrichment results for lipids in clusters 1–3.

Red represents cluster 1, dark green represents cluster 2 and purple represents cluster 3.

Extended Data Fig. 7. Wearable data and internal molecule causal association network.

a, Example association network between wearable data and internal molecules. b, Node and edge distribution of association network.