Metabolomic Atlas of Cardiovascular Diseases

Figures

• 

  Figure 1. Metabolomic Profiling and Machine Learning Classification of Cardiovascular Diseases in UK Biobank (A) Distributional differences of 9 major metabolite groups between non-cardiovascular disease (blue) and cardiovascular disease (yellow) populations, with red indicating up-regulation and blue indicating downregulation. (B and C) Performance comparison of 3 machine learning algorithms (Logistic Regression, Random Forest, and XGBoost) in classifying cardiovascular disease vs non-cardiovascular disease using metabolomics data alone and in combination with clinical parameters. The study included 27,950 cardiovascular disease patients and 216,617 non-cardiovascular disease controls. (D and E) SHapley Additive exPlanations analysis revealing key metabolites and clinical factors driving model performance. AUC = area under the curve; CVD = cardiovascular disease; SHAP = SHapley Additive exPlanations.

• 

  Figure 2. Global Metabolic Landscape and Interdisease Similarity Heatmaps (A) Classes. Heatmap of disease-wise mean z-scores (cardiovascular disease vs non-cardiovascular disease) for 325 nuclear magnetic resonance features across cardiovascular disease classes. Columns (diseases) are ordered by hierarchical agglomerative clustering (complete linkage) of a composite interdisease similarity S=ωJ+(1−ω)R, where J is the Jaccard overlap of significantly altered metabolites (|meanZ| >0.5) and R is the Spearman rank correlation of disease-specific mean z-score profiles rescaled to [0,1]; ω=0.5 in the primary analysis. The top dendrogram summarizes interdisease similarity. Rows are individual features, annotated by biomarker group (right bar). Colors encode mean z-scores (red = higher, blue = lower vs non-cardiovascular disease). (B) Subclasses. Same display as panel A but for cardiovascular disease subclasses. Columns (subclasses) are ordered by the same composite similarity S. Compared with classes, the subclass map reveals finer-grained heterogeneity, with tight ischemic and hypertensive-renal microclusters and more distinct rheumatic/valvular patterns. For visualization purposes only,…

• 

  Figure 3. Metabolomics-Only Discrimination Across Cardiovascular Disease Classes and Subclasses With Internal Geographic Hold-Out Validation (A) Classes—ranking. Area under the curve values for cross-sectional discrimination of individual cardiovascular disease classes vs non-cardiovascular disease using metabolomics data alone, ranked in descending order (XGBoost; England/discovery). (B) Classes—distribution. Proportion of cardiovascular disease classes across area under the curve ranges for metabolomics-only models (<0.6, 0.6-0.7, 0.7-0.8, 0.8-0.9, >0.9). (C) Classes—internal geographic hold-out validation. Area under the curve values for the same metabolomics-only models evaluated in Scotland + Wales (independent validation cohort, n = 21,961). (D) Subclasses—ranking. Area under the curve values for cross-sectional discrimination of individual CVD subclasses vs non-cardiovascular disease using metabolomics only, ranked in descending order (XGBoost; England/discovery). (E) Subclasses—distribution. Proportion of cardiovascular disease subclasses across area under the curve ranges for metabolomics-only models. (F) Subclasses—internal geographic hold-out validation. Area under the curve…

• 

  Figure 4. Sankey Diagram Visualization of Relationships Between Key Metabolites, Clinical Parameters, and Cardiovascular Diseases (A) Sankey diagram depicting the associations between top-ranked metabolites and clinical parameters with major cardiovascular disease classes. The width of the connecting flows represents the magnitude of SHapley Additive exPlanations values, indicating the strength of the relationship between each feature and disease category. (B) Detailed Sankey diagram showing the connections between key metabolites and clinical parameters with specific cardiovascular disease subclasses. In both diagrams, metabolites are shown in blue, clinical parameters in red, and disease categories in green.

• 

  Figure 5. Evaluation of Cross-Sectional Discriminative Patterns Using Selected Metabolites Across Cardiovascular Disease Classes and Subclasses (A) Area under the curve values for the top 15 cardiovascular disease classes in the England cohort (n = 222,606), ranked from highest to lowest performance. (B) Area under the curve values for the top 15 cardiovascular disease subclasses in the England cohort. (C) Comprehensive area under the curve values for all cardiovascular disease classes in the internal geographic hold-out validation cohort (Scotland and Wales), demonstrating robust generalizability of metabolite-based cross-sectional discrimination. Abbreviation as in Figure 1.

• 

  Figure 6. Identification and Validation of Consistent Metabolic and Clinical Predictors Across Cardiovascular Disease Spectrum (A) Frequency distribution of top-ranked features identified by SHAP analysis in at least 10 different cardiovascular conditions, stratified by feature type (metabolites or clinical parameters) and disease cross-sectional discrimination level (cardiovascular disease classes or subclasses). Color coding represents different cardiovascular disease categories: chronic rheumatic heart diseases (dark red), hypertensive diseases (yellow), ischemic heart diseases (green), other forms of heart disease (blue), cerebrovascular diseases (purple), diseases of arteries, arterioles and capillaries (magenta), and pulmonary heart disease and diseases of pulmonary circulation (orange). (B) Box plots showing the distribution of the 5 most frequently recurrent metabolites (IDL_CE, IDL_C, LA_pct, LDL_C_pct, and LA) across different cardiovascular disease classes compared to non-cardiovascular disease controls. (B) Statistical comparisons between each cardiovascular disease subclass and the non-cardiovascular disease control group were performed using the 2-sided Mann-Whitney U test.…

• 

  Central Illustration. A Metabolomic Atlas of Cardiovascular Disease Heterogeneity This systematic exploration leverages the UK Biobank (N = 244,567) and machine learning to map the metabolic architecture across 87 cardiovascular disease phenotypes. The resulting “metabolomic cartography” identifies 21 core “cross-cardiovascular disease” features—primarily convergent lipid-related disturbances—that remain robust across independent geographic cohorts and major medication classes. By uncovering diagnosis-independent structures like the “Ischemic Core,” this foundational scaffold provides a reproducible landscape to disentangle intrinsic metabolic signals from treatment effects in medically managed populations. ICD-10 = International Classification of Diseases-10th Revision; IDL_CE = cholesteryl esters in intermediate-density lipoprotein; LA = linoleic acid; SHAP = SHapley Additive exPlanations; other abbreviation as in Figure 1.