Phenotyping obesity through a two-dimensional tree structure reveals cardiometabolic heterogeneity

Abstract

Obesity, a major public health challenge, is characterized by substantial phenotypic heterogeneity. Here, we employ the discriminative dimensionality reduction tree (DDRTree) method to routine clinical data from 18,733 Chinese individuals with obesity enrolled in the nationwide China Cardiometabolic Disease and Cancer Cohort (4C) study. We identify five distinct metabolic phenotypes, among which the phenotype characterized by hyperglycemia and insulin resistance exhibits a higher risk of glycemic deterioration, while the phenotype characterized by hypertension and dyslipidemia demonstrates an elevated risk of microvascular and macrovascular diseases. These findings are validated in an independent prospective cohort. Additionally, we reveal distinctive metabolomic features that contribute to the heterogeneity of obesity in the 4C study. To translate our findings into practice, we develop a user-friendly online tool to assess event risks in the obese population. Overall, our analysis illustrates the underlying phenotypic variations influencing subsequent obesity-related outcomes, emphasizing the importance of precision medicine in obesity management.

Keywords: DDRTree; cardiovascular disease; diabetes; heterogeneity; obesity.

Graphical abstract

Figure 1

A visual representation of the phenotypic characteristics of 18,733 participants with obesity in the 4C Study (A) DDRTree was used to reduce the ten clinical variables (HbA1c, HOMA-IR, WHR, total cholesterol, HDL cholesterol, triglycerides, ALT, creatinine, SBP, and DBP, residualized for age and gender into a non-linear tree structure [N = 18,733]). The clinical variable values are overlaid on the tree structure to visualize the distribution of ten clinical variables over the reduced tree structure. Each point in the figure represents one individual. (B) Attribution of numbers to five phenotypes based on five branches (from 1 to 5). (C) Linear regression (N = 18,733) estimates (with 95% CI) between the DDRTree dimensions and the ten clinical variables, showing the association between the clinical variables and the dimensions. (D) Spatial autocorrelation (N = 18,733 for each spatial correlation analysis) of the ten clinical variables. The Moran’s I-statistic is shown on the x axis, with higher values representing clinical variables that are more strongly autocorrelated; all values were at p < 0.0001. CREAT, creatinine; TC, total cholesterol; TG, triglycerides. See also Tables S3 and S4 and Figure S4.

Figure 2

Visualizing the heterogeneity in disease progression in the 4C Study (A) Probability of incident T2DM (N = 1,022) over a 5-year period (N = 9,828). (B) Predicted probability of incident CVD (N = 475) over a 5-year period (N = 14,412). (C) Probability of CKD (N = 779) over a 5-year period (N = 11,926). For all outcomes (A–C), probabilities were generated from Cox proportional hazard models constructed with DDRTree dimensions. (D) HRs (95% CI) of DDRTree dimensions for each outcome from Cox proportional hazard models. (E) Spatial autocorrelation of three outcomes. Moran’s I-statistic is shown on the x axis, with higher values representing variables that are more strongly autocorrelated; all values were at p < 0.0001. All predictions are from the models with DDRTree dimensions. See also Figures S3 and S4.

Figure 3

Visualizing the heterogeneity in disease progression in the XJ_2011 Study A mapping function was used to position individuals with obesity in the XJ_2011 study (N = 1,653) onto the 4C (reference) tree. Each dot represents the position of one individual from the XJ_2011 study. (A) Predicted probability of incident T2DM (cases/N = 55/912). (B) Predicted probability of incident CVD (cases/N = 33/1,653). (C) Predicted probability of CKD (cases/N = 60/423). The probability of CVD outcome (B) was generated from Cox proportional hazard model constructed with DDRTree dimensions. For the T2DM and CKD outcomes (A and C), probabilities were generated from logistic regression models constructed with DDRTree dimensions. (D) HR (95% CI) and odds ratios (ORs) (95% CI) of DDRTree dimensions for each outcome form Cox proportional hazard and logistic regression models. (E) Spatial autocorrelation of three outcomes. Moran’s I-statistic is shown on the x axis, with higher values representing variables that are more strongly autocorrelated; all values were at p < 0.0001. All predictions are from the models with DDRTree dimensions. See also Figure S5.

Figure 4

Distribution of metabolites across the phenotypic tree (A) Association of 140 metabolites with dimension 1 and dimension 2 in the 4C study (N = 453). The x axis represents the linear regression estimates between dimension 1 and each metabolite, while the y axis represents the linear regression estimates between dimension 2 and the metabolites. Metabolites with significant associations were labeled with their names. (B) An enrichment overview of diabetes-related metabolites (decreases on both x axis and y axis or increases on the both axes). (C) An enrichment overview of cardiovascular-related metabolites (decreases on the x axis and increases on the y axis or increases on the x axis and decreases on the y axis). (D) An enrichment overview of chronic kidney disease-related metabolites (significantly changed on the x axis but remained constant in y axis). BCAA, branched-chain amino acid; LCFA, long-chain fatty acid; MCFA, medium-chain fatty acid; MUFA, monounsaturated fatty acid; PUFA, polyunsaturated fatty acid; SFA, saturated fatty acid; VLCFA, very long-chain fatty acid; PBAs, primary bile acids.