Polygenic Risk Prediction for Normal-Tension Glaucoma

Abstract

Purpose

Normal-tension glaucoma (NTG) is a subtype of glaucoma characterized by optic nerve damage in the setting of normal intraocular pressure. Polygenic risk scores (PRSs) have shown potential to assist with risk prediction in glaucoma, but to date no comprehensive studies have evaluated the predictive ability of PRSs for NTG.

Methods

We utilized genome-wide association study (GWAS) summary data for NTG from a European cohort to estimate the variant weights and construct PRSs. The PRSs were computed using both the SBayesRC and clumping and thresholding (C+T) methods in 317 European ancestry NTG cases and 634 controls from the National Institutes of Health All of Us dataset. To validate our findings, we used the Genetics of Glaucoma (GOG) dataset for NTG cases (n = 89) and the QSkin Sun and Health Study (QSkin) dataset for controls (n = 267).

Results

We applied the SBayesRC method, which incorporates genome functional annotation, to compare results across both studies. Logistic regression was performed to assess the association between PRSs and NTG. SBayesRC analysis demonstrated that the NTG PRS was significantly associated with NTG, yielding an odds ratio per standard deviation of 1.53 (95% confidence interval [CI], 1.32–1.77; P = 6.86 × 10⁻⁹) in the All of Us dataset and 1.83 (95% CI, 1.42–2.38; P = 4.01 × 10⁻⁶) in the combined GOG and QSkin dataset. The C+T method produced results similar to those for SBayesRC.

Conclusions

Despite the limited sample size of current NTG GWASs, our findings suggest that NTG-specific PRSs hold promise for risk prediction. Future large-scale GWASs for NTG may enable the development of clinically relevant PRSs, improving early detection and personalized risk assessment for this challenging phenotype.

Primary open-angle glaucoma (POAG) consists of two main subtypes: high-tension glaucoma (HTG) and normal-tension glaucoma (NTG). NTG is a progressive optic neuropathy characterized by normal intraocular pressure (IOP), typically below 21 mmHg,¹ and exhibits similar structural and functional changes as observed in HTG, including changes in the anterior chamber angle, thinning of the peripapillary retinal nerve fiber layer, glaucomatous optic nerve damage, and visual field defects.² However, different mechanisms may underlie the pathogenesis of NTG compared to that of high-pressure POAG. NTG may be underdiagnosed, because the IOP measurement does not raise clinical suspicion.³

Current treatments for NTG primarily focus on lowering IOP to slow disease progression, suggesting that IOP still plays a significant role in the pathogenesis of NTG. These treatments are most effective when the disease is diagnosed early, yet many cases of NTG are not detected until the disease is at an advanced stage.⁴ Thus, early detection of NTG is essential for slowing disease progression and preventing blindness.⁵

Various genes have been found to be involved in NTG, suggesting a genetic predisposition for NTG development.^6–9 Polygenic risk scores (PRSs) aggregate the effects of genetic variants genome, including those with modest effect size, to quantify the overall genetic risk. Previous studies have shown that POAG PRS can effectively identify individuals at high risk for POAG.^6,10–13 To our knowledge, no comprehensive studies have evaluated the effectiveness of NTG-specific PRSs in predicting NTG risk. This study aimed to assess whether an NTG-specific PRS can predict diseases risk.

Materials and Methods

Figure 1 presents a schematic overview of the study design, highlighting both the study baseline and the target population. In this study, we generated PRSs for NTG based on data from European cohorts. We first constructed PRSs using clumping and thresholding.¹⁴ We then employed SBayesRC v0.2.3,¹⁵ a method that integrates genome-wide association study (GWAS) summary statistics with functional genomic annotations, to optimize variant association weight estimates and improve polygenic predictions across both datasets. We initially applied the NTG PRSs to the NTG datasets and controls, using data from the National Institutes of Health (NIH) All of Us project.¹⁶ Results were validated using the Genetics of Glaucoma (GOG) study data as cases and QSkin Sun and Health Study (QSkin) as controls.

Figure 1.

Schematic overview of the study process. We used the GWAS of NTG to estimate the effect sizes of genome-wide significant variants. These estimates were then used to calculate PRSs in the target populations from the All of Us, GOG, and QSkin datasets. PRS models were subsequently developed and evaluated within these cohorts.

Baseline Data

Discovery NTG GWAS

We utilized previously published GWAS data for the NTG cases and controls from European ancestry data from several sources: International Genetics of Glaucoma Consortium (IGGC; 3247 cases and 47,997 controls), UK Biobank (UKBB) (2184 cases and 7000 controls), Canadian Longitudinal Study on Aging (CLSA; 755 cases and 3000 controls), FinnGen (81,756 cases and 326,434 controls), and a structural measurement of optic nerve integrity through vertical cup-to-disc ratio (n = 97,939). The results from this multi-trait analysis of GWAS served as baseline data for generating PRSs in the target population. More detailed information is available in Torres et al.¹⁷

Target Data

All of Us

To obtain the NTG cases and controls, we used the All of Us dataset. This program began enrolling participants in May 2018, recruiting individuals ages 18 years or older from over 340 recruitment sites. We used the version 7 data release, made available in the Researcher Workbench on April 20, 2023. This release includes genotyping array samples, with 245,394 short-read whole genome sequencing (srWGS) samples. For WGS, DNA samples were first sheared using an ultrasonicator (Covaris, Woburn, MA, USA) and then size selected using AMPure XP beads (Beckman Coulter, Brea, CA, USA). Pooled libraries were loaded onto a NovaSeq 6000 system (Illumina, San Diego, CA, USA). Data from the initial sequencing run were used to perform quality control (QC) on individual libraries and to remove non-conforming samples from the pipeline. The details of All of Us sequencing and stringent QC metrics implemented at the NIH All of Us Genome Centers and Data and Research Center are described in a prior publication and in Reference 18. We utilized the ACAF_threshold_v7.1 call set, which includes variants with a population-specific allele frequency greater than 1% or a population-specific allele count greater than 100 in at least one ancestry subgroup. Genotype QC was performed by the All of Us research group prior to data acquisition. For genetic ancestry inference, we used a tsv file from the All of Us data that provides categorical genetic ancestry for all participants. The ancestry categories correspond directly to those used in the Human Genome Diversity Project and the 1000 Genomes Project. In this study, we focused solely on the European population, due to the low number of NTG cases for other ancestries in the All of Us and GOG datasets. Additional information on QC and ancestry identification is available in related publications.¹⁶ In the All of Us dataset, we identified 744 participants with NTG. Among them, 380 were defined as female assigned at birth. These participants were identified using the Systematized Nomenclature of Medicine (SNOMED) Clinical Terms and International Classification of Diseases (ICD) codes, including ICD10CM-H40.12, ICD 10 CD-H40.123, ICD9CM-365.12, ICD10CM-H40.1232, ICD10CM-H40.1231, ICD10CM-H40.122, ICD10CM-H40.121, ICD10CM-H40.129, and ICD10CM-H40.1290. We then selected participants with European ancestry. The final selection included 317 NTG who that had genetic data with European ancestry and 634 randomly selected controls with European ancestry who had not been diagnosed with NTG or POAG.

GOG and QSkin Dataset

We used existing data from the GOG dataset, a large-scale research project involving 6068 participants, with data collection beginning in 2019. The study focuses on Australian participants recruited through a media campaign targeting individuals with glaucoma or a family history of glaucoma and through direct mail-outs to those prescribed glaucoma medications.¹⁹ We identified 89 participants with European ancestry who had both genetic data and self-reported NTG.

To select controls, we used data from the QSkin database. The QSkin study consists of a cohort of 43,794 individuals ages 40 to 69 years, sampled from the population of Queensland, Australia, in 2011. For this study, we included participants from both QSkin I and QSkin II. The objective of the QSkin cohort was to investigate skin cancer and melanoma in a population with the highest reported incidence of these diseases worldwide.²⁰ We excluded participants with a history of glaucoma or a family history of glaucoma. From the remaining participants, we randomly selected 267 controls of European ancestry.

Genotyping for both the GOG and QSkin cohorts was performed using the Illumina Infinium Global Screening Array. The GOG and QSkin datasets were merged, cleaned, and imputed together. Extensive QC measures were applied to the genetic data, excluding individuals with >3% genotype missingness and variants with missing call rates > 3%, as well as removing variants that violated Hardy–Weinberg equilibrium P < 1 × 10⁻⁶ or had a low minor allele frequency of <0.01. Only individuals of European ancestry were included based on principal components (PCs) analysis, and imputation was performed using Trans-Omics for Precision Medicine (TOPMed).²¹

Methods for PRS Construction

We applied the following two PRS methods and used meta-analysis summary statistics as discovery data to derive variant association weights.

P-Value–Based Clumping and Thresholding

We first identified overlapping single-nucleotide polymorphism (SNPs) between the discovery and target datasets and then the C+T approach was applied. We used a range of P-value thresholds: 5e-8, 5e-7, 5e-6, 5e-5, 5e-4, 0.001, 0.01, 0.05, 0.1, 0.5, and 1, referred to as study level 1 through study level 11. This range of thresholds was selected based on established practices in PRS development and reflects varying levels of association strength, from genome-wide significance to more inclusive cutoffs that may capture additional polygenic signals. The use of multiple thresholds enables a systematic evaluation of PRS performance across different levels of stringency, allowing us to identify the optimal threshold that best balances signal inclusion and noise. This strategy is widely used in PRS studies to assess how polygenic architecture influences prediction across diverse datasets.^22,23

For clumping, we utilized a randomly sampled set of 4990 unrelated individuals with White British ancestry from the UKBB as the linkage disequilibrium (LD) reference. We extracted variant weights and corresponding allele information from the GWAS summary statistics for NTG for PRS calculation at multiple P-value thresholds. For this analysis, we used PLINK 2.0 for the GOG and QSkin datasets and biocontainer/plink2.3_jan2020 for the All of Us database in the Jupyter notebook with dsub job.

SBayesRC

To investigate improvement in the polygenic prediction of NTG, we employed SBayesRC, a Bayesian method that combines GWAS summary statistics with functional genomic annotations. This approach is designed for scalable genome-wide variant analysis and enhances the detection of likely causal variants by integrating functional annotations, which impact both the likelihood of identifying causal variants and better estimating the distribution of their effect sizes.¹⁵ We used healthy European individuals from the UKBB as the testing dataset. The European UKBB served as an appropriate reference panel for our analysis as it aligns with the ancestry of our study population. This ensured that the LD structure was accurately represented, which is crucial for the robustness of our association studies and for improving the accuracy of PRS calculation. The multiple testing correction was applied using the Bonferroni method with the formula P < 0.05/N tests. In this case, the N tests include 11 C + T (CT) tests and one SBayesRC test, resulting in a significance threshold of 0.004. Analyses were performed using R 4.3.1 (R Foundation for Statistical Computing, Vienna, Austria).

Assessment Criteria

After calculating NTG PRS scores for the two target datasets, all scores were standardized based on the mean and standard deviation of the PRS, and logistic regression with 95% confidence intervals (CIs) was performed for each. In the C+T method, for each P-value threshold, the odds ratios (ORs) and 95% CIs were presented. Similarly, for the SBayesRC method, ORs with 95% CIs and the P-values were reported. Additionally, we conducted a sensitivity analysis of the SBayesRC results using four models: The first model adjusted the PRS for age, the second adjusted the PRS for sex, the third adjusted the PRS for both age and sex, and the final model adjusted the PRS for age, sex, and 10 PCs. Using the R package pROC, the area under the receiver operating characteristic (AUC) curve was calculated to assess how well the model stratified individuals into different groups, typically those with or without the disease. A high AUC for a PRS indicates effective discrimination between individuals at high and low risk of developing the disease. We calculated the AUC and its 95% CI for each threshold to compare predictive performance across datasets, aiming to determine which could better predict the model.

Results

Among the All of Us and GOG participants, 142 (44%) and 31 (34%), respectively, were male. The mean ages of participants were 76.61 ± 9.42 years for All of Us cases and 65.87 ± 8.88 years for the GOG dataset. The body mass index (BMI) in cases was 27.9 compared to 29.19 in controls. In contrast, the BMI was 25.37 in the GOG cases and 26.83 in the QSkin controls. Detailed information is presented in Table 1.

Table 1.

Characteristics of NTG and Selected Controls From the All of Us, GOG, and QSkin Databases

	Case		Control
Variables	All of Us	GOG	All of Us	QSkin
Country	United States	Australia	United States	Australia
Ancestry	European	European	European	European
Genotyping platform	Illumina GDA	Illumina GSA	Illumina GDA	Illumina GSA
Samples, n	317	89	634	267
Sex, n (%)
Males	142 (0.44)	31 (0.34)	248 (0.39)	120 (0.44)
Females	153 (0.48)	58 (0.65)	368 (0.58)	147 (0.55)
Unknown	22 (0.06)	0 (0)	18 (0.02)	0
Age, mean ± SD	76.61 ± 9.42	65.87 ± 8.88	60.63 ± 17.21	57.57 ± 7.85
BMI, mean ± SD	27.96 ± 6.14	25.37 ± 4.45	29.19 ± 7.64	26.83 ± 4.91

GDA, Global Diversity Array; GSA, Global Screening Array. More information, especially regarding genetic data, can be found in the following: All of Us (All of Us Research Program Genomics Investigators¹⁶), GOG (Gharahkhani et al.¹⁹), and QSkin (Pandeya et al.³¹).

The prevalence of NTG among individuals in the top 10% of PRSs was 5.89% in the GOG and QSkin cohorts, and it was 5.36% in the All of Us cohort. In contrast, the prevalence among individuals in the bottom 10% of PRSs was 2.41% in the All of Us cohort and was 0.84% in the GOG and QSkin cohorts. The C+T results in Figure 2A show that, at the first threshold level (P < 5e-8), the All of Us data yielded an OR of 1.52 per standard deviation change in NTG PRS (95% CI, 1.32–1.76; P = 8.35e-09). In the GOG and QSkin study, at the same threshold, the NTG PRS resulted in an OR of 2.34 (95% CI, 1.79–3.14; P = 1.95e-09). As seen, the 95% CIs for the GOG and QSkin dataset are wider and do not overlap with those of All of Us, likely due to the smaller sample size. More detailed information is available in Figure 2 and Supplementary Tables S1 and S2. Although the prediction levels across all thresholds were very close, the highest threshold corresponds to the first level (S1, genome-wide significant variants) in both the All of Us and the GOG and QSkin results. Figure 2 and Supplementary Tables S1 and S2 provide additional details on NTG PRSs, including their P-value ranges, ORs, and 95% CIs. The SBayesRC predictions for NTG are illustrated in Figure 3 for both studies. For the All of Us data, the OR was 1.53 (95% CI, 1.32–1.77; P = 6.86e-09). In the GOG and QSkin study, the OR was 1.83 (95% CI, 1.42 to 2.38; P = 4.01e-06). Table 2 shows the ORs for the SBayesRC results after adjusting for age, sex, and 10 PCs. We discuss possible reasons for the different ORs for All of Us and for GOG and QSkin in the Discussion section. Table 2 shows the ORs for the SBayesRC results after adjusting for age, sex, and 10 PCs. The ORs and their CIs slightly increased to OR = 1.59 (95% CI, 1.35–1.90; P = 6.25e-08) in All of Us study and OR = 2.07 (95% CI, 1.42–3.11; P = 2.1e-04) in the GOG and QSkin study.

Figure 2.

Association of PRS with NTG across multiple P-value thresholds in the All of Us, GOG, and QSkin datasets. The figure presents ORs and 95% CIs for NTG associated with PRSs calculated at a range of P-value thresholds in each dataset. The thresholds used include 5e-8, 5e-7, 5e-6, 5e-5, 5e-4, 0.001, 0.01, 0.05, 0.1, 0.5, and 1, corresponding to study levels 1 through 11 and labeled as S1 to S11 in the figure.

Figure 3.

SBayesRC results for NTG across the All of Us, GOG, and QSkin datasets. This figure displays the polygenic risk prediction results generated using SBayesRC, a summary-based Bayesian regression method. It compares model performance and variant contributions across the three datasets, highlighting the consistency and generalizability of SBayesRC-derived PRSs across diverse populations.

Table 2.

ORs and Adjusted CIs for SBayesRC in the All of Us, GOG, and QSkin Datasets

	All of Us		GOG and QSkin
Variable	OR (Adjusted CI)	P	OR (Adjusted CI)	P
Crude	1.53 (1.32–1.77)	7.08e-09	1.83 (1.42–2.38)	4.01e-06
PRS adjusted for age	1.54 (1.31–1.82)	2.25e-07	2.03 (1.48–2.84)	1.76e-05
PRS adjusted for sex	1.57 (1.35–1.82)	1.76e-09	1.85 (1.43–2.41)	3.50e-06
PRS adjusted for age and sex	1.55 (1.32–1.84)	1.71e-07	2.15 (1.54–3.06)	1.04e-05
PRS adjusted for age, sex, and 10 PCs	1.59 (1.35–1.90)	6.25e-08	2.07 (1.42–3.11)	2.1e-04

Figure 4 displays AUC values and 95% CIs for all thresholds for both studies. The AUC for SBayesRC in the All of Us dataset was 0.61 (95% CI, 0.58–0.66, P = 1.39e-09), and it was 0.66 (95% CI, 0.60–0.73; P = 1.09e-06) for the GOG and QSkin study. More detailed information on the AUC values and their CIs can be found in Figure 4 and Supplementary Tables S1 and S2.

Figure 4.

AUC with 95% CIs for NTG polygenic risk prediction across various P-value thresholds in the All of Us, GOG, and QSkin datasets. The figure presents AUC values with corresponding 95% CIs for each dataset, demonstrating the ability of PRSs to distinguish NTG cases from controls. PRSs were calculated using a range of P-value thresholds: 5e-8, 5e-7, 5e-6, 5e-5, 5e-4, 0.001, 0.01, 0.05, 0.1, 0.5, and 1, which correspond to study levels 1 through 11 and are labeled as S1 to S11 in the figure.

Discussion

In this study, we found that PRS is a significant predictor of NTG. However, increasing sample sizes may enhance the predictive accuracy of NTG PRS, improving its utility in identifying individuals at risk. Although NTG is influenced by multiple factors, the precise cause remains a topic of ongoing debate. Various theories have been proposed to explain its development, but no single risk factor has been definitively identified as the primary cause of the condition.²⁴ Previous studies have suggested various contributors to the pathogenesis of NTG-related optic neuropathy, including vascular dysregulation, impaired ocular blood flow, higher pressure gradient across the lamina cribrosa, impaired cerebrospinal fluid circulation, and immune and genetic influences.¹

Previous studies have demonstrated that glaucoma PRS can be valuable for predicting the progression of POAG in both at-risk individuals and existing POAG patients.^25,26 PRS is also an effective tool for risk stratification, enhancing screening and guiding treatment decisions.²⁷ In this study, we found that the current NTG PRSs show potential in predicting NTG. However, given the limited NTG GWASs compared to those available for POAG, we believe that increasing the NTG sample sizes could lead to more accurate NTG predictions in the future. To our knowledge, this is the first study to evaluate the PRS of NTG in relation to NTG. Given the more complex pathophysiology of NTG compared to POAG, we believe that PRS could play a significant role in addressing the challenges of early detection, prevention, and disease progression in NTG. Previous reports suggest that high-risk individuals with POAG may experience delays in receiving necessary treatments.¹⁶ This issue is further complicated by the characteristics of NTG. Therefore, enhanced risk-stratification tools for identifying high-risk NTG patients, such as family members of those affected by glaucoma, are urgently needed.

The prevalence of NTG varies across different geographic regions. NTG has a notably higher prevalence in East Asian populations compared to European ancestry³; however, most genetic studies and GWAS summary results have come from European populations. Given the higher prevalence of NTG in Asian populations, predicting NTG risk in countries with higher prevalence could be crucial. This study represents an initial effort to predict NTG patients. By incorporating data from various regions, we could optimize NTG prediction. Although PRS holds promise as a valuable tool for the future, especially in diverse populations and for developing better methods of predicting across various ethnic groups and separating gene–environment effects, it is important to recognize that allele frequencies (AFs) and LD patterns vary across populations. These differences can affect both GWAS effect sizes and identifying causal variants. Additionally, the transferability of PRS becomes more complex in populations with admixture ancestry.²⁸ To improve the generalizability of PRS, especially in light of epidemiological reports showing significant differences in NTG prevalence across populations, we need more diverse genomic data for high-prevalence populations to obtain more conclusive and population-specific results.

A recent study on POAG suggests that the accuracy of PRS is expected to improve significantly in the near future.²⁵ In the long term, incorporating a broader range of ancestries in future GWAS, expanding diverse reference genomes, and improving prediction methods tailored to various ethnic groups will further refine its effectiveness.²⁶

We selected the C+T method and the SBayesRC method because they represent two widely used and complementary approaches for PRS construction. The C+T method is a simple, computationally efficient, and well-established baseline approach, making it a suitable choice for initial PRS analysis.²² However, a common criticism of clumping is that it involves selecting an arbitrarily defined correlation threshold, which may affect the robustness of the results.²⁹ In contrast, SBayesRC is a more advanced Bayesian method that leverages functional annotations and the LD structure of the model to improve prediction accuracy. Its main advantage is that it allows the inclusion of a larger number of SNPs (∼7 million) in the predictive model and provides more accurate effect size estimates by incorporating LD reference panels and functional annotations.¹⁵ By including both methods, we aimed to evaluate PRS performance from both traditional and annotation-informed perspectives and to assess their utility in predicting NTG risk within our dataset.

In our study, we observed consistent results across two different datasets, suggesting the robustness of our findings. The concordance between C+T and SBayesRC results further reinforces this conclusion. Differences in the effect size estimates between the two studies may be attributed to chance (due to the extremely small number of NTG cases in GOG; n = 89) and to differences in the definition of NTG (self-reports in GOG vs. ICD10 diagnosis in All of Us) across the datasets. The smaller dataset exhibits higher variability, which could contribute to the inflated ORs seen in the GOG study, whereas the wider CIs indicate increased uncertainty in the estimates from the GOG and QSkin study. Additionally, the GOG cohort may represent more severe cases, as it is not a population-based study but specifically targeted individuals with glaucoma, with a high proportion (>50%) of the cases having a family history of glaucoma.¹⁹ Furthermore, although the All of Us study used whole-genome sequencing, GOG relied on imputed genotype data, which could, to some extent, contribute to the differences between the two datasets.

We also compared the AUC as a quantitative measure to evaluate the discriminatory performance of PRSs.³⁰ The AUC with CIs between the two studies seemed to be quite different. These discrepancies are likely due to the reasons mentioned above—especially the smaller sample size in GOG, which may inflate AUC estimates and widen confidence intervals. Moreover, differences in recruitment strategy and population characteristics (e.g., case enrichment in GOG vs. population-based sampling in All of Us) may lead to higher apparent predictive performance in GOG. Collectively, these factors likely explain the higher AUC and wider CIs observed in the GOG and QSkin dataset. Further research with larger, more diverse populations and harmonized diagnostic criteria is needed to refine PRS accuracy and improve generalizability.

This study evaluated NTG PRSs for predicting disease risk and validates the results across two different studies. We believe that the primary results of this study are promising; nonetheless, we face limitations due to the smaller effective sample size of NTG. As the NTG GWAS sample size improves, the predictive ability of the NTG PRS may also improve and could be potentially useful for NTG prediction. We believe that increasing the sample size, particularly across diverse ancestries, will lead to more robust and accurate findings.

In the GOG cohort, most participants are European, with a limited number with other ancestry. In the All of Us dataset, other ancestry groups have a limited sample size, and we need a larger population to perform PRS analysis on NTG in these groups. Consequently, these analyses were excluded. However, previous GWAS studies on African populations have also had limited sample sizes due to the low prevalence of NTG, which is a relatively rare disease among African populations. Larger, multi-ancestries cohorts are needed for better risk variant identification and polygenic score prediction of NTG.

Conclusions

Our findings demonstrate that NTG-specific PRSs can effectively predict NTG risk. These results highlight the potential of PRS as a valuable tool for identifying individuals at higher genetic risk. However, to enhance predictive accuracy and clinical utility, larger GWASs are necessary. Future studies should prioritize increasing cohort diversity and refining PRS methodologies to improve NTG risk stratification and early detection.