Comparison of multiple polygenic risk score derivation approaches for population-level prediction of primary open-angle glaucoma

Abstract

Purpose : Recent genome-wide association studies (GWAS) have identified hundreds of risk loci for primary open-angle glaucoma (POAG). Using these GWAS, polygenic risk scores (PRS) can be generated to indicate an individual’s relative risk of POAG in the population. In the future, these PRS may guide targeted population screening for POAG, helping to reduce late presentation and blindness. There are multiple methods for deriving PRS from GWAS and it remains uncertain which method is best for POAG prediction. In this study, we compared the performance of multiple PRS methods in an independent population cohort.

Methods : We used a multi-trait analysis (MTAG) to estimate PRS weights based on GWAS data for POAG (N=216,257), intraocular pressure (IOP; N=138,863), and vertical cup-disc ratio (VCDR; N=65,680). Multiple POAG PRS were constructed using the following methods: LDpred-2, PRS-CS, PRSice-2 and Lassosum. LDpred-2 uses adjusted SNP effect size estimates from GWAS summary statistics to calculate the PRS. PRS-CS estimates posterior SNP effect sizes from summary statistics using a continuous shrinkage prior. PRSice-2 uses LD-informed pruning and P-value thresholding to select significant SNPs for inclusion in the PRS calculation, while Lassosum uses penalized regression (LASSO) in its PRS calculation approach. The predictive power of each PRS for POAG prevalence was evaluated in an independent cohort, the EPIC-Norfolk study (N_cases = 227, N_controls = 6,171), using the area under the receiver operating characteristic curve (AUC) metrics and comparing the POAG prevalence in the top 10% of each PRS.

Results : While all PRS methods have reasonably strong predictive performance, the LDpred-2 PRS had the highest AUC in the EPIC-Norfolk validation cohort. The AUCs for each PRS together with age and sex were: LDpred-2 78%, PRS-CS 75%, PRSice-2 74% and Lassosum 76% in EPIC-Norfolk. The odds ratios comparing the prevalence of POAG in the top 10% of PRS with the remaining 90% of the cohort were: LDpred-2 3.75 (95% CI 2.72- 5.11), PRS-CS 2.33 (95% CI 1.63-3.27), PRSice-2 2.62 (95% CI 1.85-3.63), Lassosum 2.42 (95% CI 1.69-3.38).

Conclusions : Our independent, population-level validation results support the utility of a PRS approach for identifying subsets of the general population at high risk of POAG. We identified the LDpred-2 method as potentially superior.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.