Multimodal Deep Learning Reveals the Modular Genetic Architecture of Cardiovascular Aging.

Age is the dominant risk factor for cardiovascular disease, yet individuals of the same chronological age can differ markedly in the organs and biological pathways through which cardiovascular vulnerability emerges. We used deep learning to estimate biological age from four cardiovascular data streams in more than 100,000 UK Biobank participants - 12-lead electrocardiograms, cardiac magnetic resonance imaging, carotid ultrasound, and retinal fundus photographs - representing electrical, structural, macrovascular, and microvascular domains. The resulting biological age gaps showed limited overlap across modalities in participants with complete phenotyping, and phenome-wide analyses linked each axis to distinct disease profiles. Genome-wide association analyses identified 38 independent loci, with largely modality-specific signals involving electrophysiologic, myocardial structural, vascular regulatory, and ocular/microvascular pathways. Cross-trait LD score regression and polygenic risk scores further supported partial genetic separation of the four axes, with modality-specific associations for atrial fibrillation, heart failure, peripheral arterial disease, hypertension, diabetes, and diabetic retinopathy in UK Biobank and broadly concordant patterns in All of Us. Integration with myocardial single-cell transcriptomic data nominated distinct cellular contexts for these genetic signals. These findings suggest that AI-derived cardiovascular age is not a single biomarker of systemic senescence, but a family of related phenotypes that decompose cardiovascular aging into measurable electrical, myocardial, macrovascular, and microvascular modules.