Abstract / Summary
Glaucoma endophenotyping is limited by disease heterogeneity and single-modality structural imaging. We introduce FORGE (Functionally guided OCT Representation for Glaucoma Endophenotyping), a cross-modal contrastive framework that uses visual field (VF) signals as privileged supervision during training, enabling OCT-only inference through a learned null embedding. In 5,372 paired examinations at Mass Eye and Ear, FORGE resolved 15 macular RNFL endophenotypes that fuzzy c-means (FCM) clustering partitioned into 9 clinically distinct phenotypic clusters, improving clustering quality over the OCT-only baseline by 22%. Transferred without refitting to 74,077 UK Biobank (UKB) and 3,119 Mass General Brigham Biobank images, they reproduced their defining RNFL-thickness signatures (Lin's concordance 0.74). Treating each of the 15 continuous endophenotypes as a quantitative trait, genome-wide association analysis in 39,381 UKB European participants identified 53 LD-independent loci, 23 unreported for glaucoma or intraocular pressure, replicating in 711 external participants (sign concordance 75%). Endophenotypes learned with VF-guidance had increased genetic discovery power as compared with an identically trained ablation (median heritability Z 6.15 versus 4.76) and resolved retinal ganglion-cell enrichment.