Abstract
Purpose :
Despite extensive genome-wide association studies (GWAS) on refractive error, these efforts capture only a partial fraction of heritability. Many genomic cohorts lack refractive data but possess fundus images, currently underutilized. Recent deep learning advances facilitate precise prediction of refractive errors from fundus images. A broader GWAS on refractive errors using imputed values from fundus images could deepen our understanding of myopia's pathophysiology. Our aim is to evaluate the feasibility by comparing the effect sizes of single nucleotide polymorphisms (SNPs) on directly measured spherical equivalent (SE) and imputed SE from fundus images using deep learning.
Methods :
We utilized second visit data of the Nagahama Study, comprising 9,850 individuals. We excluded subjects with high blood pressure and abnormal cardio ankle vascular index, resulting in the inclusion of 7,301 participants. Among them, 3,796 individuals had genome-wide SNP data determined by DNA microarrays, and the other 3,505 individuals did not. Firstly, using the latter dataset, we developed a deep learning model based on a swin transformer model pre-trained with ImageNet, to predict SE from color fundus images. In this process, the data was divided into training and validation dataset at a ratio of 4:1. Secondly, this model was applied to the 3,505 images with SNP data. Finally, GWAS was conducted for both measured SE and imputed SE, adjusting for age and sex. We looked up 438 refractive error susceptibility SNPs (Hysi, P. G. et al., Nature Genetics, 2020) from both results, and the effect sizes were compared.
Results :
Included 7,301 subjects exhibited a mean age of 56.3±12.4 years, an average SE of -1.57±2.89 diopters (D), with a gender distribution comprising 5,178 females and 2,123 males. The developed model showed a mean absolute error of 0.72D. Among 438 SNPs, 326 SNPs were present in our dataset. The correlation coefficient for 326 SNPs between βmeasuredSE and βimputedSE was calculated to be 0.95 [95% confidence interval:0.94-0.96].
Conclusions :
Myopia susceptibility SNPs exhibited comparable correlations with both directly measured SE and imputed SE. This suggests the viability of conducting GWAS for refractive errors on datasets that includes genomic information and fundus images but lack refractive data. The application of this approach holds promise for enhancing the genomic exploration of myopia.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.