Abstract

Pigmentation varies widely across humans and is shaped by melanin quantity, type, and spatial distribution. Retinal pigmentation protects against light-induced damage, yet its genetic and evolutionary bases remain unclear. We developed a deep learning framework (DeepGRP) to quantify retinal pigmentation from high-resolution fundus images and conducted a genome-wide association study (GWAS), identifying 42 signals, including 26 previously unidentified loci, with single-nucleotide polymorphism–based heritability of 21.4%. Single-nucleus assay for transposase-accessible chromatin by sequencing and RNA sequencing of human fetal retinal tissues revealed key cellular contributors, including retinal pigment epithelium and photoreceptor cells. Among candidate genes, ARHGAP18 emerged as a previously unrecognized regulator of melanogenesis. Evidence of polygenic adaptation in Europeans suggests selection driven by snow-reflected light at high latitudes. A polygenic risk score for retinal pigmentation correlated with a 4.8-fold higher risk of myopia and a 1.5-fold lower risk of skin cancer. These findings demonstrate the power of deep learning for large-scale ocular phenotyping and reveal insights into the genetic and evolutionary architecture of retinal pigmentation.