CardiOmicScore
Summary¶
CardiOmicScore is a multitask deep learning framework built on UK Biobank data that generates disease-specific proteomic (ProScore, from 2,920 proteins via the MetNet-paired network ProNet) and metabolomic (MetScore, from 168 metabolites via MetNet) risk scores for six cardiovascular diseases — coronary artery disease (CAD), stroke, heart failure (HF), atrial fibrillation (AF), peripheral artery disease (PAD), and venous thromboembolism (VTE). Combined with polygenic risk scores (PRS) and clinical predictors in Cox proportional hazards models, ProScore and MetScore materially improve discrimination, calibration, and net benefit up to 15 years before disease onset [1].
Study Design¶
UK Biobank participants were split by omics availability into a Metabolomics-only cohort (N=220,859; used to train/validate MetNet), a Proteomics-only cohort (N=19,086; used to train/validate ProNet), and an untouched Multiomics cohort (N=24,287, with genomics, metabolomics, and proteomics) reserved purely for evaluating the added value of combining omics layers. Each development cohort was further split into training/validation sets (England and Wales) and a held-out geographic testing set (Scotland) to assess regional generalizability. MetNet and ProNet are multitask neural networks that jointly predict risk for all six CVDs from their respective omics panel, rather than fitting six separate models — this lets the models learn shared risk representations across cardiovascular outcomes. Final risk models combine PRS, MetScore, ProScore, and one of three predefined clinical predictor sets (AgeSex; Clin, 25 clinical/questionnaire variables; PANEL, Clin plus blood count) via Cox proportional hazards (CPH) regression.
Performance¶
- Sole predictors: ProScore alone achieved C-index 0.69 (VTE) to 0.82 (PAD); MetScore alone achieved 0.64 (VTE) to 0.74 (PAD); PRS alone was substantially weaker (0.52–0.60). Both ProScore and MetScore stratified Kaplan-Meier survival for all six CVDs (log-rank P < 0.001), while PRS-based stratification was significant only for CAD, stroke, AF, and VTE. Top-tertile ProScore carried 1.68–2.18-fold higher CVD risk than the bottom tertile; top-tertile MetScore carried up to 1.70-fold higher risk (CAD).
- Added value over clinical models: adding omics scores to clinical-only Cox models significantly increased the C-index for every CVD, from +0.005 (PRS added to the AgeSex model for stroke) up to +0.102 (PRS+ProScore added to AgeSex for PAD). ProScore alone contributed the largest single-score gain (+0.040 to +0.101 across CVDs), ahead of MetScore (+0.009 to +0.051). Gains persisted, though attenuated, even when added to the richer PANEL clinical set (+0.003 to +0.049).
- Clinical utility: PANEL-based Cox models were well calibrated across all six CVDs; adding omics scores (especially ProScore) further improved standardized net benefit across decision thresholds, with the full PRS+MetScore+ProScore model improving area-under-net-benefit-curve (A-NBC) by 35.4 (stroke) to 212.4 (AF) relative to the PANEL-only model.
Availability¶
Code: github.com/YanLuoCityU/cardiomicscore. An interactive demo accepting clinical and omics inputs to predict individual CVD risk is deployed at yanluocityu.github.io/cardiomicscore-website.
See Also¶
- Multimodal Cardiovascular Risk Prediction — the broader paradigm of combining clinical, omics, and polygenic predictors for CVD risk, of which CardiOmicScore is a large-scale, six-disease implementation.
- PanMETAI — a similarly-motivated tabular-omics diagnostic model (NMR metabolomics plus clinical/protein features) applied to pancreatic cancer rather than cardiovascular disease.