Applying unsupervised deep learning to four-dimensional cardiac MRI from 78,113 UK Biobank adults, researchers extracted latent motion representations of left-ventricular movement and mapped them to 39 genome-wide significant loci — roughly two-thirds (approximately 26) invisible to conventional imaging indices like ejection fraction or wall thickness. Critically, the motion-specific variants didn't point to cardiomyocytes but instead converged on fibroblasts, endocardial cells, endothelial cells, and autonomic neurons, validated against single-cell transcriptional atlases. Environmental exposures including smoking and air pollution also linked distinctly to motion profiles. Findings replicated across All of Us, FinnGen, and GPMap cohorts.
This preprint, not yet peer-reviewed, represents a genuinely paradigm-shifting reframing of cardiac genomics. The field has long been constrained by reductive summary metrics that collapse complex myocardial dynamics into single numbers, systematically blinding us to non-cardiomyocyte biology. The finding that stromal and vascular cell-type regulatory programmes — not muscle itself — dominate motion genetics forces a reconsideration of how cardiac disease risk is encoded. Fibroblast and endothelial contributions to mechanical behavior have been underappreciated therapeutically. The cohort size is exceptional, and multi-biobank replication strengthens confidence. Limitations include the observational design, predominantly European ancestry in discovery cohorts, and the interpretive challenge of latent deep-learning features. If peer review confirms these findings, this methodology could fundamentally reshape how cardiac genomic risk scores are constructed and which cell types are prioritized as drug targets.