Abstract
Primary open-angle glaucoma (POAG) is a leading cause of irreversible blindness, yet its genetic architecture remains largely unexplained. Traditional genome-wide association studies (GWAS) often rely on imprecise phenotypes from electronic health records or simplified global averages of ocular measurements. In this study, we present a disease-trained, task-transfer machine learning (ML) framework designed to learn complex glaucoma-related damage patterns from a clinical repository of 18,985 optical coherence tomography (OCT) scans. By applying these ML-derived endophenotypes to 47,908 UK Biobank participants, we performed a cross-ancestry GWAS that identified 43 independent genetic loci, 21 of which are novel to glaucoma. Functional analyses prioritized 11 high-confidence gene effectors, providing mechanistic insights into Wnt-mediated outflow dysfunction and retinal ganglion cell vulnerability. This study demonstrates a powerful, generalizable strategy for accelerating the discovery of disease mechanisms and therapeutic targets for complex diseases.
Introduction
Primary open-angle glaucoma (POAG) is characterized by the progressive degeneration of retinal ganglion cells (RGCs) and the optic nerve. While heritability is high, current genetic variants explain only about 9% of POAG's variance. A significant bottleneck in genetic discovery is the reliance on binary "case vs. control" labels or global endophenotypes, such as mean retinal nerve fiber layer (RNFL) thickness. These metrics often fail to capture the spatially heterogeneous nature of glaucomatous damage, which typically manifests in specific quadrants or patterns rather than uniform thinning. To overcome this, we developed a deep learning approach to extract high-dimensional, disease-relevant features from retinal imaging to power more sensitive genetic association studies.
Methodology: Task-Transfer Machine Learning
The core of our approach is a disease-trained, task-transfer ML framework. Unlike previous models trained on healthy populations to recognize standard anatomy, our models were trained specifically to recognize patterns of glaucomatous decay.
Feature Extraction
We utilized two distinct deep learning architectures to learn latent representations of RNFL and ganglion cell complex (GCC) thickness maps:
- Unsupervised Autoencoder (AUTO): Learns to compress and reconstruct thickness maps, capturing the most significant variations in retinal structure.
- Self-supervised Momentum Contrastive Learning (MOCO): Learns by comparing different views of the same data, identifying robust features that distinguish individual scans.
Task Transfer and GWAS
These models were first trained on a clinical dataset of 8,323 glaucoma patients from Massachusetts Eye and Ear (MEE). Once the models learned to identify "glaucomatous signatures," they were transferred to the UK Biobank (UKB) population. We generated scores for 47,908 UKB participants, effectively using the ML models as "digital experts" to quantify glaucoma-like structural patterns in a large, genotyped cohort. These scores served as the quantitative traits for a cross-ancestry GWAS spanning European, Asian, and African ancestries.
Results and Genetic Discovery
The ML-derived endophenotypes proved significantly more powerful for genetic discovery than traditional global averages. Our cross-ancestry meta-analysis identified 43 LD-independent GWAS loci that reached genome-wide significance.
- Novelty: Of the 43 identified loci, 21 were novel and had not been previously associated with glaucoma. This suggests that the ML models are capturing unique biological signals related to the spatial distribution of retinal thinning that simple averages overlook.
- Ancestral Consistency: Despite the smaller sample sizes of African and Asian cohorts in the UK Biobank, the effect sizes were remarkably consistent across ancestries (Pearson r = 0.998 between European and cross-ancestry results), highlighting the fundamental biological relevance of these traits.
Functional Insights and Causal Genes
To move from genetic associations to biological mechanisms, we integrated the GWAS results with functional genomic data, including tissue expression (GTEx), retinal eQTLs (EyeGEx), and single-cell RNA-seq datasets. This pipeline prioritized 11 high-confidence effector genes, five of which are novel to glaucoma research.
The findings converged on two primary pathogenic axes:
- Outflow Dysfunction: We identified genes like RSPO2, a Wnt signaling potentiator. RSPO2 is active in the eye's drainage tissues (the trabecular meshwork), suggesting a role in regulating intraocular pressure.
- Neuronal Vulnerability: Genes such as EIF3E and DYNC1I2 were linked to the resilience of Retinal Ganglion Cells. These genes represent potential targets for neuroprotective therapies aimed at making RGCs more resistant to pressure-induced damage.
Discussion and Future Directions
This study demonstrates that disease-specific machine learning can refine phenotypes for genetic discovery in ways that traditional clinical metrics cannot. By training on diseased eyes, the model learned to ignore "normal" anatomical variation and focus on the structural signatures of pathology.
While the results are promising, limitations remain. The UK Biobank is predominantly European, and while the cross-ancestry results were consistent, larger non-European datasets are required for comprehensive validation. Additionally, future work will focus on integrating longitudinal data to determine if these ML scores can predict disease progression before it becomes clinically apparent.
In conclusion, our ML-driven strategy provides a scalable framework for identifying the genetic underpinnings of complex diseases, offering a roadmap for discovering novel therapeutic targets in glaucoma and beyond.
Summary updated