Harnessing Large‐Scale Multi‐Omics Data for Risk Prediction and Deep Phenotyping of Valvular Heart Diseases in the General Population

Figures

• 

  FIGURE 1. Overview of study design: Diverse data from the UK Biobank are utilized, including six categories: sociodemographic data, lifestyle data, physical measurement data, genetic data, metabolic data, and protein data. Predictor variables are generated after completing field selection, data cleaning, and preprocessing of missing data. Response variables are derived from hospitalization, death, self‐reported, and primary care data, which are standardized to ICD‐10 codes. After temporal alignment of the independent and dependent variables, a 10‐year survival time observation period is constructed, and multiple models are fitted to carry out disease prediction, risk assessment, and bioinformatics analyses. This approach provides ideas for predicting valvular disease risk at baseline, promoting precise risk stratification, and identifying mechanistic pathways.

• 

  FIGURE 2. Clinical prediction model of valvular disease risk in the primary cohort (Clin model). (A) ROC curves and precision‐recall curves of valvular disease prediction models. (B) Performance metrics of the Clin model (using the top 25 important features) in the independent test dataset. (C) SHAP visualization plot of the top 25 important features selected. The width of the horizontal bars represents the degree of contribution to disease prediction—the wider the range, the greater the contribution. The color of the horizontal bars indicates the magnitude of the feature values, represented by a gradient from blue (low) to red (high), with the color bar on the right showing the specific color coding. The direction of the x‐axis indicates the likelihood of developing the disease (right side) or remaining healthy (left side). DBP Diastolic Blood Pressure, SBP Systolic Blood Pressure, BMI Body Mass Index, HbA1c Glycated Hemoglobin A1c.

• 

  FIGURE 3. Multi‐omics valvular disease risk biomarkers. (A) Top ranked features from omics models. The C‐index values reported for the Pro model correspond to the full protein‐only LASSO model including all selected protein predictors. (B) SHAP visualization plot of the top 25 important features selected based on the Pro model, covering VHD, AVS, and MVR. (C) The ROC curves and precision‐recall curves of the valvular disease prediction models constructed using only the four most important proteins identified from the Pro model. TG/TL in IDL% Triglycerides to Total Lipids in IDL percentage. CE/TL in IDL% Cholesteryl Esters to Total Lipids in IDL percentage. C/TL in Medium LDL% Cholesterol to Total Lipids in Medium LDL percentage. TG/TL in Medium LDL% Triglycerides to Total Lipids in Medium LDL percentage. PL/TL in Medium LDL% Phospholipids to Total Lipids in Medium LDL percentage. Avg Diam for VLDL Particles Average Diameter for VLDL Particles. TL in CML & XL‐VLDL Total Lipids in Chylomicrons and Extremely Large VLDL. FC in Very Small VLDL Free Cholesterol in Very Small VLDL. PL in Very Small VLDL Phospholipids in Very Small VLDL. TG in CML & XL‐VLDL Triglycerides in Chylomicrons and…

• 

  FIGURE 4. Clustering risk overview in the primary cohort. (A) The feature contribution plots corresponding to the three clustering subtypes, covering VHD, AVS, and MVR. (B) Comparison of the event‐free probability over time across the three clusters generated for each condition: VHD, AVS, and MVR.

• 

  FIGURE 5. GO Enrichment analysis and protein—protein interaction network of valvular diseases after LASSO‐based screening. (A–C) The GO enrichment analysis for VHD, AVS and MVR. (D–E) The protein—protein interaction network for VHD, AVS and MVR. For the proteins screened in the valve disease model, a protein‐protein interaction (PPI) network was constructed using the STRING database or the average absolute SHAP value. Based on the existing biological information, the score‐threshold was set to 400 when exploring protein‐protein interactions using the STRING database. Additionally, in the PPI network constructed based on the average absolute SHAP values, the size and color of the proteins are determined by their number of connections. Proteins with more connections to other proteins are displayed larger and in a more intense color.

• 

  FIGURE 6. Mendelian randomization analyses between proteomics and valvular disease after LASSO screening. After identifying suitable protein GWAS data in the GWAS catalog (https://www.ebi.ac.uk/gwas/home) and IEU databases (https://gwas.mrcieu.ac.uk), all proteins for AVS and MVR were selected based on their covariate‐adjusted importance rank for the subsequent Mendelian randomization analysis. In cases where evidence of horizontal pleiotropy was present, the MR‐Egger results were used. The instrumental variable threshold for all proteins was set to p < 5e‐08. (A) Forward Mendelian randomization results with protein as the exposure and AVS as the outcome. (B) Backward Mendelian randomization results with AVS as the exposure and proteins as the outcome. (C) Forward Mendelian randomization results with proteins as the exposure and MVR as the outcome. (D) Backward Mendelian randomization results with MVR as the exposure and proteins as the outcome. Proteins with (*) are potentially druggable targets (Therapeutic Target Database, at http://db.idrblab.net/).