Genome-wide association study for refractive error imputed from fundus images via deep learning: a proof-of-concept...

To evaluate whether deep learning-predicted phenotypes from fundus photographs can serve as surrogate phenotypes for genome-wide association studies (GWAS), using spherical equivalent as a model trait. We assessed concordance between genetic associations derived from measured and AI-imputed values.

Cross-sectional analysis integrating deep learning-based prediction with genetic association in a population-based cohort.

This study included 9850 participants from the Nagahama Study. A Swin Transformer was fine-tuned to predict spherical equivalent from fundus photographs. Model training was performed on participants without genotype data who had not undergone cataract surgery. The trained model was then applied to the remaining participants to generate AI-imputed spherical equivalent values. GWAS were conducted separately for measured and AI-imputed spherical equivalent. Concordance between GWAS results was assessed by overlap of significant Single-nucleotide polymorphisms (SNPs) and correlation of effect size estimates.

The model achieved an MAE of 0.78 diopters (D) and a Pearson's correlation coefficient (r) of 0.92 on an independent test set. The GWAS using AI-imputed spherical equivalent captured all genome-wide significant SNPs identified in the measured spherical equivalent GWAS. Effect-size estimates derived from AI-imputed GWAS were highly concordant with those from measured GWAS across the genome (r = 0.96).

Deep learning-imputed spherical equivalent from fundus images can serve as a reliable surrogate phenotype for genetic discovery. This approach enables large-scale and cost-efficient GWAS for phenotypes that are difficult or expensive to ascertain in cohorts.