Deep learning representations and proteome-wide Mendelian randomization identify causal mediators of myocardial fibrosis
Authors
Affiliations (1)
Affiliations (1)
- Departments of Medicine, Genetics, Computer Science, and Biomedical Data Science Stanford University
Abstract
Cardiac fibrosis is a central pathological process in heart failure, yet the molecular mechanisms governing its spatial organization remain poorly defined. We developed an artificial intelligence (AI)-based phenotyping approach to decode the spatial organization of cardiac fibrosis from routine cardiac magnetic resonance imaging (MRI). We applied convolutional variational autoencoders (VAEs) and distributional metrics to native T1 maps from 50,239 UK Biobank participants. VAE-derived features predicted mortality with greater precision than standard T1 measures (C-index 0.614 vs. 0.547; likelihood ratio p = 2.9x10-3), identifying spatial fibrosis patterns as independent prognostic indicators. Through genome-wide association studies, we identified genetic loci underlying T1 distribution metrics, implicating oxidative stress pathways (SOD2, GSS) and calcium signaling (CAMK2D, CALU). Pathway enrichment revealed distinct biological processes: T1 distributions reflected metabolic and coagulation activity, while spatial VAE dimensions reflected extracellular matrix organization and complement regulation. Mendelian randomization identified cathepsin S (CTSS) and extracellular matrix protein 1 (ECM1) as causal mediators with near-certain colocalization evidence (PP.H4>0.88), validated in an independent Icelandic cohort (n=35,559). FKBPL demonstrated causal effects on both T1 distributional and spatial features. Published preclinical studies show CTSS inhibition reduces collagen deposition and ventricular stiffening, ECM1 stabilizes extracellular matrix and prevents fibrosis, and FKBPL peptides attenuate fibroblast activation. These findings highlight tractable pathways for therapeutic modulation of myocardial fibrosis.