AI-based multiomics profiling for personalized prediction of cardiovascular disease: A prospective UK Biobank study

Genomics, metabolomics and proteomics offer complementary insights into the risk of cardiovascular diseases (CVDs), yet current risk prediction models lack the capability to comprehensively integrate such multiomics data and clinical information. Leveraging the in-depth data from 24,308 individuals in the UK Biobank, we developed a novel multitask deep learning model to simultaneously learn disease-specific, personalized proteomic (ProScore) and metabolomic (MetScore) risk scores for the nine most common CVD events by profiling 168 metabolites and 2,920 proteins of participants. Experiments demonstrated that, ProScore and MetScore could be used not only as the sole predictor of CVDs within 16.6 years (mean C-index range: 0.67–0.83 for ProScore and 0.62–0.73 for MetScore), but also showed complementary predictive power to clinical predictors. When combined with clinical data, these omics signatures significantly enhanced cardiovascular risk prediction, improving the mean delta C-index by 0.016–0.094 across various CVDs. Important CVD-related proteins, such as NT-proBNP, NPPB, and HAVCR1, and metabolites, including creatinine, albumin, and glycoprotein acetyls, were also identified by our model. Our findings suggest that incorporating multiomics profiling into clinical practice has the potential to enhance personalized risk assessments and enable earlier, more targeted interventions for various CVDs. The mechanistic insights provided by our models could guide the development of novel biomarkers and therapeutic targets and inform the repurposing of existing drugs, paving the way for precision medicine for primary prevention of CVDs.