Novel Genetic Factors Associated with Primary Open-Angle Glaucoma Identified Using Artificial Intelligence

Abstract

Purpose : To evaluate the predictive power of an artificial intelligence (AI) construct to predict primary open-angle glaucoma (POAG) from single nucleotide polymorphisms (SNPs) and to identify potentially novel genetic factors associated with glaucoma.

Methods : We included 1012 participants of the Ocular Hypertensive Treatment Study (OHTS) from multiple ancestries. A total of 817 participants remained without evidence of POAG (controls) and 195 participants eventually developed POAG (cases) over ~15 years. We developed a case-control genome-wide association study (GWAS) to identify SNPs that were independently associated with glaucoma and selected top scored SNPs (Fig. 1). We then developed a machine learning model composed of feature subset selection and learning to identify the most promising subset of SNPs that were highly predictive of POAG development. We finally developed an independent machine learning classifier to predict glaucoma from the discovered subset of SNPs (Fig. 1). We evaluated the accuracy of the model using cross-validation of the receiver operating characteristics curves.

Results : The top 1000 highly scored SNPs, out of ~1 million, were selected for the downstream analysis. Machine learning discovered 63 SNPs, out of 1000 (Table 1), that were collectively predictive of glaucoma with an AUC of 0.88 (95% CI: 0.86 – 0.90) without using any additional glaucoma endophenotype features. The AUC of the model using 750 subjects from only (self-reported) European race was 0.93 (95% CI: 0.90 – 0.95). The discovered SNPs were mapped to 45 independent genes, of which eight were previously known to be associated with glaucoma traits and 37 genes remain as potential candidates for development of the POAG.

Conclusions : The combined statistical modelling and machine learning frameworks achieved a high accuracy in predicting glaucoma. Successful development of this learning model may assist clinicians in using DNA testing to identify individuals with ocular hypertension who are at-risk of glaucoma development and future vision loss. Independent datasets are desirable to further validate the findings in this study.

This is a 2021 ARVO Annual Meeting abstract.

 

Figure 1. The pipeline of the proposed AI construct for predicting POAG from genetic data and discovering novel genetic factors associated with POAG.

 

Table 1. List of SNPs and mapped genes that machine learning discovered as highly predictive of POAG.