Deep learning imaging-informed genomics and phenome-wide analyses reveal a mitral–ventricular–aortic axis in sporadic mitral valve prolapse and mitral annular disjunction
H Nicholls, T S Tugba Sanver, H X Hao Xu, S F Stefano Figliozzi, S D M Silvana di Maio, F D Francois Deharo, K M Kedar Madan, V S Vaselios Stylianidis, C P Carlotta Poggi, A D Alexandra Drompalou, S E P Steffen E Petersen, A Y Alistair Young, P G M Pier-Giorgio Masci, N A Nay AungAbstract
Background
Non-syndromic mitral valve prolapse (MVP), affecting ~3% of adults, is the most prevalent valvular abnormality worldwide. Although often benign, a subset develops progressive mitral regurgitation, heart failure (HF), and malignant ventricular arrhythmias (MVA). MVP frequently coexists with mitral annular disjunction (MAD) and disproportionate left ventricular (LV) remodelling and dysfunction, suggesting a broader cardiovascular phenotype with unexplored genetic drivers.
Purpose
To evaluate MVP as a cardiovascular disorder, we aimed to: 1) build and validate a convolutional neural network (CNN) for automated detection of MVP/MAD from cardiovascular magnetic resonance (CMR) in the UK Biobank; 2) perform genome-wide association study (GWAS) of CMR-defined MVP/MAD to establish their genetic architecture; 3) perform phenome-wide-association study (PheWAS) and longitudinal analysis.
Methods
A quality-controlled 2D nnU-Net CNN was developed to automatically detect MVP/MAD from cine 3-chamber CMR in a clinical dataset. The network was tested in 55,353 UK Biobank participants with CMR data. GWAS was performed in European-ancestry participants (51,384 controls); loci underwent functional annotation, colocalisation, Mendelian randomisation (MR), and polygenic risk scores (PRS) for PheWAS. Cox proportional hazards models tested associations of CMR-defined MVP with incident atrial fibrillation (AF) and HF adjusting for traditional risk factors, left atrial and ventricular volumes and function.
Results
The CNN achieved sensitivity/specificity of 0.99/0.57, reflecting a sensitivity-optimised design to maximise case detection, identifying 2,191 MVP(+) and 2,475 MAD(+) cases for GWAS. Seven loci reached genome-wide significance (P<5x10-8) across both MVP and MAD (10.2% and 9.8% SNP-heritability respectively). MVP had three novel loci (SMAD3, MIR548F3, HDGFL1). CDC25A, SGCD, and LOC100507065 were specific to MAD. Only SMAD3 associated with both traits. Bioinformatic analysis prioritised familial aortopathy genes (SMAD3 and TGFB2) and HF genes (RBM20 and HDGFL1). SMAD3 colocalised with HF and genetically predicted protein levels of SMAD3 were causally linked to MVP in MR. In PheWAS, PRS-MVP associated with hypertension and aortic aneurysm/dissection (P=6.78x10-30). Cox models showed that CMR-identified MVP independently predicted incident AF (HR 1.58, 95%CI 1.28–1.95; P=2.36x10-5) and HF (HR 1.43, 95%CI 1.00–2.00; P=0.048).
Conclusion
Imaging-informed genomics delineates the genetic architecture of MVP and MAD and implicates pathways linking mitral degeneration with ventricular-aortic remodelling. PheWAS identifies strong associations with aortic aneurysm and dissection, while longitudinal analyses demonstrate that MVP independently predicts incident AF and HF. These findings support MVP as a mitral-ventricular-aortic condition with important implications for risk stratification and clinical decision-making.Fig1_MVP_MAD_Genomic_Phenotypic_Assoc.png