Structure-aware retinal disentanglement reveals the genetic architecture of ocular and systemic diseases

Deep learning effectively extracts retinal phenotypes but often functions as an entangled black box, obscuring specific genetic mechanisms and hindering clinical interpretability. To resolve this, we present the Unsupervised Ophthalmic Feature Extraction (UOFE) framework. Our approach feeds full fundus images, optic disc masks, and vessel masks into three isolated structure-aware autoencoders. Crucially, these streams share no parameters and are optimized separately to explicitly encode the macular background, optic disc, and retinal vasculature. Models were trained on 53,600 EyePACS images, utilizing a median smoothing kernel to disentangle the background, optic disc/cup, and retinal vasculature, with generalizability rigorously confirmed across 15 public datasets. Evaluated against a dimension-matched baseline trained to reconstruct the entire fundus image, UOFE demonstrated superior biological disentanglement by achieving high structural consistency between left and right eyes. In a GWAS of 75,010 UK Biobank participants, applying strict linkage disequilibrium clumping, UOFE identified 255 independent genomic loci. This represents an 8-fold increase over the monolithic baseline (31 loci) and uncovers highly novel signals compared to established global methods. Biologically, a double dissociation emerged: glaucoma mapped primarily driven by optic disc features ([Formula: see text]), while AMD mapped to background features. Clinically, vascular embeddings provided powerful incremental prognostic value for diabetic retinal abnormalities (Hazard Ratio 2.38, 95% CI: 2.17-2.61, [Formula: see text]). By successfully isolating specific anatomical architectures, UOFE transforms retinal imaging into an interpretable precision window, uncovering localized genetic signals and prognostic biomarkers lost in traditional entangled representations.