Deep-learning-derived glaucoma-related endophenotypes enable novel genome-wide genetic and functional discovery

Area of Science:

• Ophthalmology
• Genetics
• Artificial Intelligence

Background:

• Primary open-angle glaucoma (POAG) is a major cause of irreversible blindness with an incompletely understood genetic basis.
• Previous genome-wide association studies (GWAS) were limited by imprecise phenotypes derived from electronic health records.

Purpose of the Study:

• To identify novel genetic loci associated with POAG using a machine learning (ML) framework and precise endophenotypes.
• To elucidate the genetic architecture and pathogenic mechanisms underlying POAG.

Main Methods:

• Developed a disease-trained, task-transfer ML framework to derive glaucoma-related damage patterns from optical coherence tomography (OCT) scans.
• Applied ML-derived OCT endophenotypes to GWAS in UK Biobank participants across European, African, and Asian ancestries, followed by cross-ancestry meta-analyses.
• Conducted extensive functional analyses, including Bayesian colocalization and Mendelian randomization, to identify high-confidence effector genes.

Main Results:

• Identified 36 (EUR) and 43 (cross-ancestry) novel lead genome-wide association study (GWAS) loci for POAG.
• Over two-thirds of identified loci overlapped with known POAG associations, validating the ML approach.
• Discovered 21 novel loci and 11 high-confidence effector genes, five novel to glaucoma, implicating Wnt-mediated outflow and retinal ganglion cell vulnerability.

Conclusions:

• The ML-driven strategy effectively identified novel POAG genetic associations and potential therapeutic targets.
• Findings provide mechanistic insights into POAG pathogenesis at a cell-type resolution.
• This generalizable ML approach can accelerate discovery for complex diseases.