Large-scale machine-learning-based phenotyping significantly improves genomic discovery for optic nerve head...

Area of Science:

• Ophthalmology
• Genetics
• Artificial Intelligence

Background:

• Accurate phenotyping is crucial for genome-wide association studies (GWASs).
• Expert labeling of ocular features is time-consuming and can be variable.
• Machine learning (ML) offers a potential solution for automated phenotyping.

Purpose of the Study:

• To develop and validate an ML model for predicting glaucomatous optic nerve head features from fundus photographs.
• To utilize the ML model for a large-scale GWAS of vertical cup-to-disc ratio (VCDR).
• To identify novel genetic associations with VCDR and glaucoma.

Main Methods:

• Developed an ML model to predict VCDR from color fundus photographs.
• Applied the ML model to 65,680 individuals in the UK Biobank (UKB).
• Conducted a GWAS using ML-derived VCDR and compared results with a manually labeled GWAS.

Main Results:

• The ML-based GWAS identified 299 independent genome-wide significant (GWS) hits in 156 loci.
• Replicated 62 of 65 previously identified GWS loci and discovered 93 novel loci.
• Pathway analyses suggested biological relevance for novel loci, including genes in neuronal biology and Mendelian ophthalmic disease genes.
• ML-based GWAS significantly improved polygenic prediction of VCDR and primary open-angle glaucoma in an independent cohort.

Conclusions:

• ML-based phenotyping is a viable and efficient alternative to manual labeling for large genetic studies.
• The study significantly expanded the understanding of genetic factors influencing VCDR and glaucoma.
• The findings enhance the potential for polygenic risk prediction in glaucoma.