Abstract
Purpose :
Glaucoma is a complex and heterogenous disease with structural and functional variation that might reflect underlying pathogensis. We aim to discover hidden structural patterns of optical coherence tomography (OCT) derived macular ganglion cell complex (GCC) thickness using unsupervised deep learning (DL) among subjects with high polygenic risk (PRS) for primary open angle glaucoma (POAG). Subsequently we used these endophenotypes to discover novel genomic loci.
Methods :
POAG PRS was constructed for UK Biobank subjects with LDpred2 using genome-wide association study (GWAS) summary statistics from largest cross-ancestry meta-analysis. Macular OCTs from participants with highest PRS were segmented using Topcon Advanced Boundary Segmentation algorithm to extract GCC layer thickness. A hybrid unsupervised artificial intelligence method that combines deep convolutional autoencoders, manifold learning, and Gaussian Mixture Model (GMM) clustering grouped semantically similar GCC thickness patterns. Optimal number of clusters was determined using probabilistic model selection method (AIC/BIC). We performed GWAS for each GMM-identified feature dimension using linear mixed models adjusted for age, sex, and principal components of ancestry.
Results :
We identified 11 optimal macular GCC patterns using 26806 OCT scans from 13403 UKBB participants in the top half of the POAG PRS (Fig1). Among all patterns, P1, 7 and 8 had the highest proportion of subjects with self-report or ICD-code for glaucoma (6.25%, 6.37%, 6.53%) and IOPcc > 24 mmHg (5.83%, 5.70%, 5.84%) while pattern 9 had most myopic refractive error (SE -0.3 ±2.4 D). GWAS was performed with hard-call genotypes for 424155 variants after quality control filters (Fig2). There was no evidence of unaccounted stratification (lambdaGC 1.000 to 1.011). We identified numerous genomic regions associated (P < 1e-05) with at least one feature dimension related to neurologic disease and development (P1, gene: KIRREL3; P3, genes: DENND4B, ARHGEF28; C7 ZDHHC7), immune system and autoimmune disease (P7 HIVEP3, P8 MARK1, P10 MUSK), metabolomics/obesity (P11 CFAP77) and ocular development (P10 MAF).
Conclusions :
Here we demonstrate that DL can be used to identify hidden OCT patterns which in turn can be used to identify unexpected genetic associations. Future work is needed to ensure these patterns are disease specific and further explore the impact of identified genetic variants.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.