Skip to content

Pirruccello JP, Chaffin MD, Chou EL, et al. Deep learning enables genetic analysis of the human thoracic aorta. Nat Genet. 2022;54:40-51. PMID 34837083

One-paragraph summary

A deep learning model was trained to measure ascending and descending thoracic aortic diameter from 4.6 million cardiac magnetic resonance images in the UK Biobank, converting an expensive manual measurement into a scalable quantitative trait. Genome-wide association studies were then run in 39,688 individuals, identifying 82 loci associated with ascending and 47 associated with descending thoracic aortic diameter, of which only 14 overlapped between the two segments. Transcriptome-wide association analyses, rare-variant burden testing, and human aortic single-nucleus RNA sequencing were used together to prioritise causal genes, including SVIL, which was strongly associated with descending aortic diameter. A polygenic score for ascending aortic diameter was associated with incident thoracic aortic aneurysm in 385,621 UK Biobank participants at a hazard ratio of 1.43 per standard deviation (95% CI 1.32–1.54, P = 3.3 × 10⁻¹²).

Key findings

  • 82 ascending and 47 descending diameter loci from a single biobank, versus a prior literature dominated by rare Mendelian genes — thoracic aortic size is a common polygenic trait.
  • Only 14 of the loci were shared between ascending and descending aorta, establishing that aortic genetics is segment-specific and tracks the embryological boundary between second-heart-field and neural-crest-derived aorta.
  • Ascending-diameter polygenic score predicted incident TAA at HR 1.43 per SD in 385,621 participants — a genotype-only signal of clinically relevant magnitude in an unselected population.
  • Multi-modal gene prioritisation (TWAS + rare-variant burden + single-nucleus RNA-seq) nominated SVIL among others, demonstrating that single-cell atlases and GWAS are complementary rather than parallel efforts.
  • Methodologically: deep learning removed the phenotyping bottleneck that had capped aortic genetics at case-control designs on small ascertained cohorts.

Limitations

  • Ancestry. UK Biobank is predominantly European-ancestry; the resulting loci and polygenic score are not calibrated for other populations, a limitation the same group later stated explicitly for the derived AORTA Gene score (PMID 37662232).
  • Healthy-volunteer bias. UK Biobank imaging participants are healthier than the general population, and the aortic diameter distribution is correspondingly narrow — the GWAS is powered on normal variation, not on disease.
  • Diameter is a proxy. The trait measured is aortic size, not wall integrity or dissection risk. Whether loci for diameter are loci for rupture remains open, and the parallel case-control TAAD GWAS (PMID 37308786) has not been formally reconciled with these diameter loci.
  • Cross-sectional imaging. A single time point per participant means growth rate — the clinically actionable phenotype — was not a trait in this analysis.
  • Prediction is not yet conditional on diameter. The polygenic score predicts aneurysm, but no analysis demonstrates incremental value in patients whose aortic diameter is already known, which is the only scenario in which it would be used.

Why it matters

This paper is the hinge between the Mendelian era and the polygenic era of thoracic aortic disease. It made three things possible at once: a genome-wide view of aortic size, a population-scale polygenic score, and a template — deep learning as a phenotype factory — that the same group immediately reapplied to spatially resolved LVOT/root/ascending diameters (PMID 35902171), to a clinical prediction model (PMID 36378208), to a 1.1-million-variant polygenic score improving diameter prediction over clinical factors (PMID 37662232), and to velocity-encoded flow phenotypes in 47,223 participants (PMID 37205587). Its most consequential single finding is the least headline-friendly one: the near-independence of ascending and descending aortic genetics, which argues that "thoracic aortic aneurysm" names several distinct diseases and that segment-specific risk models — and possibly segment-specific thresholds — are the correct unit of analysis.

Cited by wiki pages

  • omics and emerging science
  • biomarkers