Imaging genetics of the beating heart: disentangling static and dynamic cardiac phenotypes

Abstract Genome-wide association studies (GWAS) have identified loci linked to traits such as ejection fraction and stroke volume, but these conventional metrics provide limited resolution into the spatial and temporal complexity of cardiac motion. Here, we introduce a spatiotemporal graph autoencoder applied to dynamic cardiac meshes, which disentangles anatomical and motion-related factors while explicitly modeling temporal correlations across the cardiac cycle. Trained separately for each cardiac chamber, the model captures subject-specific contraction and relaxation patterns in a compact latent space. Applying GWAS to latent representations from over 46,000 UK Biobank participants, we uncovered associations beyond those detectable with standard phenotypes. Dynamic features highlighted conduction and contractility genes (HCN4, MYH11), static traits emphasized developmental regulators (PITX2, HAND2), and shared associations involved canonical sarcomeric genes (TTN, BAG3). Atrial-specific signals, including HMGA2 and BCAR1, were linked to left atrial dynamics. Compared with ejection fraction and electrocardiogram-derived traits, our phenotypes captured complementary and sometimes distinct genetic signals. These findings demonstrate that spatiotemporal modeling of dynamic cardiac meshes offers a powerful framework for genetic discovery, enabling higher-fidelity characterization of cardiac motion and revealing novel genetic axes of atrial and ventricular function.