Integrating untargeted metabolomics and machine learning to reveal an aberration of sphingolipid metabolism in cardiometabolic HFpEF

Figures

• 

  Fig. 1. Flowchart of the overall study design. The analytical pipeline consisted of untargeted metabolomic profiling of serum samples from the discovery cohort, screening and functional enrichment of differential metabolites, WGCNA, SHAP, candidate biomarker selection via multiple machine learning algorithms, diagnostic performance evaluation using ROC curves, and external validation by ELISA in an independent validation cohort

• 

  Fig. 2. Metabolomic profiling of the discovery cohort. (A) Donut chart illustrating the class distribution of all detected metabolites in positive ion mode. (B) Donut chart illustrating the class distribution of all detected metabolites in negative ion mode. (C) PCA score plot showing global separation of metabolic profiles between the MetS and HFpEF groups. (D) Volcano plot of differential metabolites: each dot represents a single metabolite. Red dots denote significantly upregulated metabolites, green dots denote significantly downregulated metabolites, and grey dots indicate metabolites without significant intergroup differences. The x-axis corresponds to log₂FC, where a larger absolute value reflects greater intergroup abundance disparity; the y-axis represents -log₁₀(P-value) for statistical significance, and dot size is proportional to the VIP value. (E) Hierarchical clustering heatmap of differential metabolites: rows correspond to metabolites and columns correspond to samples. Red indicates relatively high abundance, while blue indicates relatively low abundance. The left dendrogram shows metabolite clustering, and the top dendrogram shows sample clustering. Abbreviations: GP,…

• 

  Fig. 3. Screening and functional enrichment analysis of differential metabolites. (A)Top 20 differential metabolites ranked by FC. (B)Top 20 differential metabolites ranked by VIP. (C) Comparison of the proportional abundance of SM species relative to total detected metabolites between the MetS and HFpEF groups. (D) Bubble diagram for KEGG pathway enrichment of differential metabolites. The x-axis shows enrichment factor, and the y-axis lists annotated KEGG pathways. Bubble size corresponds to the number of differential metabolites enriched in each pathway, and bubble color represents the significance of enrichment (P-value)

• 

  Fig. 4. Results of machine learning–based screening of candidate biomarkers. (A) Cross-validation curve of the Elastic Net model. The x-axis represents log(λ), and the y-axis denotes model accuracy. Red dots indicate the average accuracy at each λ value, and the red line illustrates the trend of accuracy across varying λ. (B) Feature importance ranking plot of the Random Forest model: the x-axis shows feature importance scores, and the y-axis lists metabolite nomenclature. (C) Independent ROC curves of each candidate metabolite for discriminating MetS and HFpEF. (D) ROC curve of C24:1 Sphingomyelin for binary classification, with an AUC of 0.999 (95% CI: 0.996–1.000). (E) Correlation heatmap between WGCNA modules and clinical indicators: clinical markers are displayed on the x-axis, and co-expression modules on the y-axis. Red cells indicate positive correlations and blue cells indicate negative correlations; values within each cell represent P-values, with P < 0.05 defined as statistically significant correlation. (F) SHAP summary plot illustrating the magnitude and direction of predictive contribution of each metabolite feature in the Random Forest model. The x-axis corresponds to SHAP…

• 

  Fig. 5. External validation results based on clinical applicability. (A)Bar chart presenting the mean serum sphingomyelin concentration in the MetS and HFpEF validation subgroups (n = 30 per group). Error bars denote standard deviation (SD). *P < 0.05 by independent Student’s t-test, indicating statistically significant intergroup differences in SM concentration. (B) ROC curve of total sphingomyelin for differentiating MetS and HFpEF in the validation cohort; the corresponding AUC value and 95% confidence interval (CI) are annotated in the plot