Decoding Primary Open-Angle Glaucoma: A Multi-Omics Approach to Identify Druggable Effector Genes

Abstract

Purpose

Genomewide association studies (GWAS) have identified numerous primary open angle glaucoma (POAG) risk loci, yet most reside in non-coding regions with unclear function. Mapping these loci to effector genes can elucidate disease mechanisms, identify functionally conserved variants, improve cross-ancestry risk prediction by reducing population-specific noise, and uncover shared therapeutic targets.

Methods

Here, we integrate European POAG GWAS with six types of multi-omics molecular Quantitative Trait Locis (xQTLs) using multi-trait colocalization to identify candidate effector variants and evaluate their cross-population relevance using genetic risk score (GRS) analysis, and their therapeutic potential through drug target prioritization.

Results

We identified 25 POAG effector variants colocalized with at least one xQTLs. In non-European populations, effector variants showed stronger effect size correlations with Europeans than non-colocalized variants (Pearson r² = African 0.85 vs. 0.71; East Asian 0.81 vs. 0.69; and Latin American 0.91 vs. 0.75). Effector variants also had smaller allele frequency variations across populations (average interquartile range [IQR] = 0.15 vs. 0.20). The genetic risk score based on effector variants performed comparably to the genome-wide significant single-nucleotide polymorphism (SNP)-based GRS in non-European populations. Drug prioritization identified zinc, copper, sunitinib, probucol, and astemizole as potential common therapeutic agents for POAG and its subtypes.

Conclusions

Our findings offer deeper insight into the molecular mechanisms underlying glaucoma and effector variants for developing more robust GRS models and broadly effective therapeutic strategies for POAG.

Glaucoma is a leading cause of irreversible vision loss worldwide, primarily driven by progressive retinal ganglion cell (RGC) degeneration and optic nerve damage, often in association with elevated intraocular pressure (IOP).¹ Early detection of glaucoma remains a major clinical challenge, as many patients are not diagnosed until the disease has progressed to an advanced stage, contributing substantially to the global burden of vision impairment.² Primary open-angle glaucoma (POAG) is one of the most prevalent forms of glaucoma, affecting approximately 65 million individuals worldwide.¹ Recent genomewide association studies (GWASs) have elucidated the genetic architecture of POAG, identifying over 100 susceptibility loci linked to disease risk.^3–7 However, many identified variants lie in non-coding regions, making their functional effects difficult to interpret. Moreover, most GWAS analyses have been conducted in European populations, which limits the translatability of these findings to other ethnic groups.^8–11 In addition, POAG encompasses clinically distinct subtypes, including high-tension and normal-tension glaucoma, which may involve divergent pathogenic mechanisms.^3–7 Therefore, identifying reliable and druggable targets that are effective across diverse ancestral backgrounds and POAG subtypes is essential for the development of equitable and precision-based therapeutic strategies.

Effector gene—defined as those mediating the biological consequences of trait-associated genetic variants—are central to elucidating the functional relevance of GWAS findings.¹² Given that the majority of POAG-associated variants reside in non-coding regions, linking these variants to their downstream effector genes facilitates biological interpretation and uncovers the molecular pathways underlying disease susceptibility. Furthermore, effector gene-based prioritization enables the identification of functionally conserved variants, which may enhance the accuracy and transferability of GRS across diverse ancestral populations by minimizing the influence of population-specific, non-causal variants.¹³ Critically, as genes constitute the primary targets of pharmacological intervention, systematic identification of effector genes offers a valuable framework for therapeutic discovery and drug repurposing in POAG.

Identifying functional genes through the integration of omic dataset—such as expression, splicing, or protein quantitative trait loci (eQTLs, sQTLs, and pQTLs)—with GWAS summary statistics has become an increasingly established strategy in elucidating the biological mechanisms underlying complex traits. Colocalization and Mendelian randomization are commonly used analytical approaches in this context.^14,15 Although these studies have shed light on the functionality of certain POAG-associated loci, they often focus on a single or limited number of data types, missing the broader biological context that emerges from integrating multiple omics layers. A multi-omics strategy can offer a more comprehensive view of variant functionality, improve cross-population relevance, and deepen our understanding of POAG pathogenesis.^12,16 Previous colocalization methods, such as “COLOC,”¹⁷ “HyPrColoc,”¹⁸ or “MOLOC,”¹⁹ are limited by their ability to incorporate only a small number of molecular phenotypes and often assume a single causal variant per locus, reducing power to detect complex regulatory mechanisms. A recently developed Omics PlEiotRopic Association (OPERA)²⁰ method addresses these limitations by jointly integrating multi-omics xQTL data with GWAS summary statistics using reference linkage disequilibrium (LD), thereby improving the resolution and power to identify effector genes.

In this study, we build on these insights by combining POAG GWAS results with six types of omics QTLs using the OPERA method, followed by functional annotation of the resulting effector variants. We then conducted comprehensive comparison between effector variants and non-colocalized variants and assess the broader applicability of these findings through genetic risk score (GRS) analysis in multiple ancestral groups. Finally, we leverage the NeDRex platform²¹ to evaluate potential drug targets of the identified effector genes (Fig. 1). By using an integrated multi-omics framework and emphasizing cross-population relevance, this work seeks to advance our understanding of POAG's molecular underpinnings, guide more effective precision medicine strategies, and ultimately improve intervention outcomes for those at risk of this vision-threatening disease.

Figure 1.

Analysis work flow. (A) POAG GWAS summary data was colocalized with six QTLs utilizing OPERA method. Significant independent loci at the P < 5e–08 threshold were used as the primary analysis, whereas independent loci at the 1e-05 and 1e-04 thresholds were used for sensitivity analysis. Colocalized loci from IOP-dependent and IOP-independent GWAS summary statistics were utilized for drug target analysis. (B) The allele frequencies of effector and non-colocalized SNPs was queried across 28 populations in the Ensembl database. The annotations and associated traits of effector and non-colocalized SNPs was queried from FUMA. We compared the function and distribution between the effector and non-colocalized SNPs. (C) Effector SNPs at 5e–08, 1e–05, 1e–04 thresholds, all independent SNPs and non_colocalized SNPs were utilized to calculate GRS in the UK Biobank validation cohort. The GRS model was regulated by age, age², sex, and principal components. (D) The PAOG effector genes, IOP_dependent genes, IOP_independent genes were used to perform Nedrex drug target analysis.

Methods

Study Design

An overview of the analytical steps and approaches taken are described in Figure 1.

In the first step, POAG GWAS summary statistics were colocalized with six QTL datasets using the OPERA method. Genomewide significant, independent loci (P < 5e–08) were used for the main analysis, whereas loci meeting more relaxed thresholds (P < 1e–05 and P < 1e–04) were included in sensitivity analysis. Additionally, in secondary analysis, separate colocalization analysis were performed using IOP-dependent and IOP-independent GWAS summary statistics to identify effector genes relevant to different POAG subtypes.

In the second step, we conducted a comprehensive comparison between effector variants (those with evidence of colocalization) and non-colocalized variants. We compared their allele frequencies, effect sizes on POAG, and functional annotations to assess heterogeneity across populations.

In the third set of analysis, we constructed GRS using effector variants and non-colocalized variants separately in the UK Biobank cohort. We then evaluated the predictive performance of these GRS models for POAG across different ethnic groups.

In the final step, we performed drug target prioritization using the NeDRex platform based on the identified effector genes for POAG, including both IOP-dependent and IOP-independent genes. This aimed to identify shared therapeutic targets that may be effective across POAG subtypes.

GWAS Summary Statistics of POAG and Its Subtypes

We used stage 1 GWAS summary statistics from Gharahkhani et al.,³ derived from a European population comprising 16,677 POAG cases and 199,580 controls. In this stage, all POAG cases were defined based on International Classification of Diseases, Ninth Revision (ICD-9) and ICD, Tenth Revision (ICD-10) criteria. To investigate genetic signals of IOP-dependent and IOP-independent, we incorporated GWAS summary data from Huang et al.,²² who used genomic structural equation modeling to dissect the distinct components of POAG. This study used the IOP GWAS from the European-ancestry participants of the UK Biobank (n = 97,644) and the POAG GWAS from Gharahkhani et al.³ with FinnGen participants excluded as inputs for genomic structural equation modeling, from which IOP-dependent and IOP-independent POAG GWAS were derived. We did not use GWAS datasets related to POAG severity because large-scale GWAS studies of POAG severity are currently lacking. Variant coordinates were based on the GRCh37 reference genome, and variants on sex chromosomes or missing rsID, base pair, or allele frequency data were excluded. Details of all GWAS datasets are provided in Supplementary Table S1.

xQTLs Summary Statistics

Six types of omics QTL data were integrated with POAG GWAS summary statistics for colocalization analysis. The methylation QTLs (mQTLs) came from GWAS on whole blood methylation data (Brisbane Systems Genetics Study, n = 614 and Lothian Birth Cohorts, n = 1366), 52,916 methylation sites were involved.²³ In total, 118,152 QTLs related to chromatin accessibility (caQTLs) were obtained from adult prefrontal cortex tissue of 135 individuals with schizophrenia and 137 controls.²⁴ A total of 12,794 sQTLs were identified from RNA splicing analyses of RNA-sequencing (RNA-seq) data from 2,865 brain cortex samples of unrelated individuals of European ancestry.²⁵ For gene expression, eQTLs included both whole blood data from GTEx (V8, n = 670, and 19,270 genes) and retinal data from the Eye Genotype Expression (EyeGEx; n = 453 and 17,374 genes) database.^26,27 The 2 datasets share 13,930 genes in common; GTEx eQTLs have 5340 genes that are unique, and EyeGEx eQTLs have 3444 unique genes. Protein quantitative trait loci including 4907 proteins were based on plasma protein levels measured in 35,559 Icelanders with genotype and phenotype data.²⁸ Variant coordinates were mapped to the GRCh37 genome build. Variants on sex chromosomes or lacking rsID, base pair position, or allele frequency were excluded. For pQTLs, only proteins with at least one genomewide significant association (P < 5e–08) were included, resulting in 1198 proteins retained for analysis. Full details of the xQTL datasets are provided in Supplementary Table S2.

OPERA Colocalization

We used OPERA, a method that combines multiple types of xQTL summary data with GWAS data to detect pleiotropic associations of molecular phenotypes with a complex trait at a GWAS locus, with the primary aim of providing mechanistic interpretation of the GWAS signal.²⁰ OPERA is a Bayesian generalization of the SMR and HEIDI approach²⁹ to a multi-omics model, in which the molecular phenotypes are considered as multiple exposures and the trait is considered as the outcome. Because OPERA requires predefined genomic regions, at first, we used the 66 significant independent loci (P < 5e–08) from Gharahkhani et al.³ as the primary input for the main analysis. Recognizing that loci below this threshold may still be functional—particularly in non-European populations—we also conducted sensitivity analysis by applying less stringent significance thresholds (P < 1e–05 and P < 1e–04) and LD clumping (r² < 0.01, window size = 1 MB, reference panel = 1000 Genomes Phase 3 EUR) via PLINK2. For the secondary analysis, colocalization of IOP-dependent and IOP-independent loci was similarly evaluated at P < 5e–08 threshold.

OPERA first computes marginal posterior probabilities of association (PPA) for each exposure and derives joint PPAs for combinations of exposures. The HEIDI test is then used to distinguish pleiotropic signals from those driven by LD. Colocalization was performed within ±1 MB of each index variant. As suggested by the original study, which also represents the default setting of this method, loci with PPA >0.9 and HEIDI P value > 0.01 were designated as effector variants, indicating multi-omics support and minimal LD confounding.

Variants Effect Size, Allele Frequency, and Functional Annotation

GWAS summary statistics for POAG in African, East Asian, European, and Latin American populations were obtained (see Supplementary Table S1). The effect sizes (β) of effector variants and non-colocalized variants were then compared across populations. We calculated the Pearson correlation coefficient (r²) between the effect sizes (β) of variants in non-European ancestries and those in the European population, separately for colocalized and non-colocalized variants. Fisher's z-test was then used to compare the r² values between the two groups.

We analyzed allele frequencies of both colocalized and non-colocalized loci across 26 populations using data derived from the Ensembl database (http://www.ensembl.org/Homo_sapiens/). For each locus, allele frequency variation across populations was quantified using interquartile range (IQR). Functional annotation for both sets of loci was conducted using FUMA³⁰ in both main and sensitivity analysis; ANNOVAR³¹ were used for region-based annotation; the CADD score, derived from 63 annotations, estimates variant deleteriousness, with scores above 12.37 suggesting potentially harmful variants.³² Additionally, the pleiotropic effects of each variants were queried from the GWAS Catalog.³³

POAG Definition and Genetic Risk Score Construction

The UK Biobank cohort was used to assess the discriminatory power of effector variants in distinguishing POAG cases across diverse ancestral populations. In the UK Biobank, glaucoma cases were identified through multiple sources, including hospitalization records (ICD-10 = H40.X and H42.X and ICD-9 = 365.X), surgical procedures (OPCS-4 = C60–C62 and C65 and OPCS-3 = 1574 and 1512), self-reported non-cancer illness (field 20002, code 1277), eye disorder history (field 6148), and prescriptions for glaucoma-related eye drops (field 20003). The earliest recorded date from these sources was used as the date of diagnosis. POAG cases were defined based on specific ICD codes (ICD-10 = H40.1, H40.8, and H40.9 and ICD-9 = 365.1). Controls were individuals who underwent eye screening and confirmed no history of glaucoma. The UK Biobank POAG cohort included 27,955 individuals of European ancestry (prevalence = 1.8%), 624 of African ancestry (prevalence = 3.0%), 757 of Asian ancestry (prevalence = 2.2%), and 459 of other ancestries (prevalence = 1.9%).

The GRS for POAG were then generated and their discriminative abilities were assessed. Scores were computed using the “score” function in PLINK version 2.0.³⁴ Variant effect sizes were used as weights. The GRS_effector was constructed using independent, significant effector variants that colocalized with POAG, selected at thresholds of P < 5e–08, P < 1e–05, and P < 1e–04. As a comparison, we also computed GRS for non-colocalized loci at P < 5e–08 (GRS_noncolocalized) and for all significant loci (GRS_all). Logistic regression was performed with POAG as the outcome and GRS, age, age², sex, and ancestry principal components as covariates. All GRSs were standardized (mean = 0 and variance = 1) to minimize bias from outliers. Model performance was evaluated using Nagelkerke R², area under the curve (AUC) of the GRS. The DeLong test was used to compare the AUCs between different models.

Drug Target Analysis

We performed drug target analysis using the NeDRex²¹ platform, which constructs heterogeneous biological networks, identifies disease modules, and prioritizes drugs based on their relevance to disease mechanisms. To explore potential therapies for different glaucoma pathways, we analyzed three gene sets: genes corresponded with POAG GWAS effector genes (POAG gene set), IOP-dependent effector genes (IOP-dependent gene set), and IOP-independent effector genes (IOP-independent gene set). Genes included in the effector gene sets were defined as the nearest genes, according to FUMA single-nucleotide polymorphism (SNP) annotation, to variants that showed colocalization with at least one type of xQTL. Disease models were constructed for each gene set using the multi-Steiner trees (MuST) algorithm, which identified connected subnetworks from the gene-gene layer of the integrated network, potentially reflecting underlying disease mechanisms.

Drug Repurposing

After obtaining the POAG gene, IOP-dependent gene, and IOP-independent gene models, we applied the TrustRank algorithm to prioritize candidate drugs for repurposing based on these disease gene models. TrustRank ranks nodes in a network based on how well they are connected to a trusted set of seed nodes. During the drug repurposing prioritization process, only approved drugs were considered. The damping factor of TrustRank was set to 0.85, which is the default value recommended for this algorithm. We selected the top 50 drugs ranked by the TrustRank score for inclusion in the final results.

Results

Multi-Omics Colocalization Identifies 25 Effector Variants With POAG GWAS

Details of the data resources are listed in the Methods section. In the main analysis (colocalization at P < 5e–08 threshold), OPERA identified 25 loci associated with at least one type of omics data using independent loci at the P < 5e–08 threshold (Table 1; Supplementary Tables S3, S4), accounting for 38% of all significant loci. Among these, 5 loci were linked to 2 omics types, whereas rs41543317 and rs6687545 were associated with 3. The mQTLs and sQTLs explained the largest number of the colocalized loci (17 and 8, respectively), followed by GTEx whole blood eQTLs (4 loci) and EyeGEx retinal eQTLs (2 loci). No loci were explained by pQTLs or caQTLs (Fig. 2A).

Table 1.

Effector Variants at P < 5e–08 Threshold

RS ID	CHR	Position	Effect Allele	Other Allele	Beta	Nearest Gene	xQTLs
rs41543317	17	44087500	A	G	−0.0823	MAPT	mQTL, sQTL, eQTL
rs6687545	1	38071195	T	C	0.0742	GNL2/RSPO1	mQTL, sQTL, eQTL
rs1139795	22	19867771	T	C	−0.1435	TXNRD2	mQTL, sQTL
rs2790049	1	165743523	A	G	0.3049	TMCO1	eQTL, sQTL
rs12208086	6	36586070	A	G	0.1083	SRSF3/MIR3925	mQTL, eQTL
rs111439095	13	76254433	A	G	−0.0859	LMO7	mQTL
rs1336980	9	129377855	C	G	−0.1232	LMX1B	mQTL
rs1555214	6	45928400	T	C	0.0775	CLIC5	eQTL
rs1577488	1	101078400	A	G	−0.1276	GPR88/LINC01349	mQTL
rs1649068	10	60304864	A	C	0.1056	BICC1	mQTL
rs17135357	6	602983	A	C	−0.1228	EXOC2	mQTL
rs2027312	9	129920463	A	G	−0.0781	RALGPS1	mQTL
rs2472494	9	107695539	T	C	0.1894	ABCA1	eQTL, mQTL
rs2579989	6	51460154	T	G	−0.2067	TFAP2B/PKHD1	mQTL
rs2745572	6	1548369	A	G	0.1258	FOXF2/FOXCUT	mQTL
rs4653159	1	36579215	A	G	0.1031	COL8A2	eQTL
rs58073046	11	120248493	A	G	−0.1727	ARHGEF12	sQTL
rs581796	11	86355565	T	C	−0.1576	ME3	mQTL
rs6475604	9	22052734	T	C	−0.2686	CDKN2B−AS1	sQTL
rs6495085	15	74221313	C	G	−0.1156	LOXL1	mQTL
rs6602453	10	10840849	A	C	0.0816	CELF2	sQTL
rs935328	15	57538801	A	T	0.0953	TCF12	mQTL, sQTL
rs9494457	6	136474794	A	T	−0.1164	PDE7B	eQTL
rs9649390	7	115801846	T	C	0.0671	TFEC/TES	mQTL
rs9913911	17	10031183	A	G	0.1607	GAS7	mQTL

CHR and Position denote the chromosome number and genomic position, respectively. Effect allele is the allele associated with the reported effect size (Beta), whereas Other allele is the alternative allele. Beta represents the estimated effect size from GWAS. Nearest gene shows the closest gene(s) to the variant. The xQTLs refers to molecular quantitative trait loci associated with the variants.

Figure 2.

POAG GWAS summary statistics colocalized with six xQTLs. (A) Illustrates the intersection of multi-omics QTLs in opera analysis of effector SNPs (P < 5e–8). The top bar chart represents the number of shared SNPs across different QTL combinations, while the bottom matrix indicates the combination of QTLs. The left bar chart shows the total SNP count in each QTL dataset. (B) Shows the proportion of effector SNPs explained by xQTLs in the main analysis (P < 5e–8) and sensitivity analysis (P < 1e–4, P < 1e–5). (C) Shows the number of SNPs associated with different number of QTLs in the main analyses (P < 5e–8) and sensitivity analysis (P < 1e–4, P < 1e–5).

In the sensitivity analysis, OPERA identified 40 and 52 loci colocalized with POAG at P < 1e–05 and P < 1e–04 thresholds, respectively, representing 18% and 6% of all independent loci (Supplementary Tables S5, S6; Supplementary Fig. S1). Although the proportion of explained loci was lower than in the main analysis, the overall distribution of xQTL types was consistent: mQTLs explained most of the colocalized loci, followed by sQTLs, whole blood eQTLs, and retinal eQTLs. At the relaxed thresholds, one locus colocalized with caQTLs at P < 1e–05 (rs10824305, PPA = 0.932), and one locus colocalized with pQTLs at P < 1e–04 (Fig. 2B; rs80312472, PPA = 0.993). Notably, the sensitivity analysis revealed more variants associated with three or more omics types. For example, rs58621819 and rs80312472 were associated with 4 omics types at P < 1e–05, whereas rs58621819 and rs7738466 were associated with 4 omics types at P < 1e–04 (Fig. 2C). The rs58621819 is colocalized with whole blood and retinal eQTLs, caQTLs, and mQTLs (P < 1e–05, PPA = 0.900, P < 1e–04, PPA = 0.915) as well as rs80312472 (PPA = 0.915) and rs7738466 (PPA = 0.915) with whole blood eQTLs, retinal eQTLs, mQTLs, and pQTLs.

In the secondary analysis, OPERA analysis of IOP-dependent and IOP-independent loci (P < 5e–08) identified 31 and 9 colocalized loci, respectively, accounting for 41% and 53% of significant loci. These variants were linked to five xQTL types—mQTLs, sQTLs, whole blood eQTLs, retinal eQTLs, and caQTLs—with distributions similar to those observed in the main and sensitivity analysis (see Supplementary Figs. S1, S2; Supplementary Tables S7, S8). Similar to POAG, the majority of the variants were colocalized with methylation.

Different Profile of Effector Variants and Non-Colocalized Variants

The effect sizes of 25 effector and 41 non-colocalized variants were compared across POAG GWAS summary statistics from African, East Asian, European, and Latin American populations (Supplementary Table S9). For Latin American versus European populations, the Pearson correlation coefficient (r²) of variant β values at effector loci was 0.91 (P < 0.001), higher than that at non-colocalized loci (r² = 0.75, P < 0.001). A Fisher's Z-test confirmed that this difference between the r²s was statistically significant (P = 0.03). Similarly, the correlation between African and European populations was higher at effector loci (r² = 0.85, P < 0.001) than at non-colocalized loci (r² = 0.71, P < 0.001), although the difference was not statistically significant (P = 0.17). For East Asian versus European populations, the r² at effector loci was 0.81 (P < 0.001), compared to 0.69 (P < 0.001) at non-colocalized loci, showing no significant difference (P = 0.14; Supplementary Fig. S3; Supplementary Table S10).

Diversity in allele frequencies was further analyzed. First, for both colocalized and non-colocalized variants, allele frequency variation was assessed using the IQR across 26 regional populations representing African, European, South Asian, East Asian, and American ancestries (Fig. 3; Supplementary Table S11). Comparison of IQRs showed that non-colocalized variants exhibited greater variability (mean IQR = 0.20, median IQR = 0.19) than effector variants (mean IQR = 0.15, median IQR = 0.14), with the difference reaching boundary statistical significance (Wilcoxon rank-sum test, P = 0.056; Supplementary Fig. S4).

Figure 3.

The distribution of effector variants’ and non_colocalized variants’ allele frequency across 26 populations. Heatmap shows the allele frequency distribution of variants across global populations. Each row represents a variant and each column corresponds to a population group. Color intensity reflects normalized allele frequency values (scaled 0–1). The bar plot on the left displays the interquartile range (IQR) of allele frequencies for each variant. Populations are grouped by continental ancestry: African, European, South Asian, East Asian, and American. Variants are grouped by effector variants and non_colocalized variants.

To further characterize the variants, a functional comparison was conducted. Functional annotation revealed that a greater proportion of effector variants were located in intronic regions compared with intergenic regions (65% vs. 20%; Supplementary Table S12), whereas non-colocalized variants were more frequently found in intergenic than intronic regions (49% vs. 44%; Supplementary Table S13). A χ² test indicated a significant difference in the distribution of functional annotations between the two groups (P < 0.001). Sensitivity analysis using a relaxed threshold (P < 1e–05) showed consistent trends: 57% of effector variants were intronic and 24% intergenic, compared to 44% and 42% for non-colocalized variants, respectively (χ² test P < 0.001; Supplementary Fig. S5; Supplementary Tables S14, S15). Although the mean Combined Annotation Dependent Depletion (CADD) scores were slightly higher for effector variants (6.22 ± 6.15) than for non-colocalized variants (5.72 ± 7.51), this difference was not statistically significant (P = 0.78; Supplementary Fig. S6).

To evaluate pleiotropic effects, variants and their previously reported trait associations were further examined using the GWAS Catalog. It was found that 112 traits were associated with effector variants and 441 with non-colocalized variants. Effector variants were predominantly linked to glaucoma-related traits (61.3%), followed by metabolism/endocrinology (7.7%), neuroscience (7.1%), cancer (5.8%), and immunity/blood traits (3.4%; Supplementary Table S16). In contrast, non-colocalized variants were associated with a broader range of traits, including glaucoma (24.1%), metabolism/endocrinology (19.8%), immunity/blood (15.1%), cardiovascular (12.6%), and neuroscience (4.6%; Supplementary Fig. S7; Supplementary Table S17).

GRS of Effector Variants

Five sets of GRSs were generated using data from the UK Biobank. GRS_all was constructed with 66 independent SNPs that were genome-wide significantly in GWAS for POAG. The GRS_effector was constructed using independent, significant effector variants that colocalized with POAG, selected at thresholds of P < 5e–08 (25 SNPs), P < 1e–05 (40 SNPs), and P < 1e–04 (52 SNPs). GRS_noncolocalized were computed for non-colocalized loci at P < 5e–08 (41 SNPs). SNPs used in each GRS are demonstrated in Supplementary Table S18. Overall, the GRSs demonstrated better predictive performance in European and African populations, but lower performance in Asian populations.

In Europeans, the GRS_all model achieved the highest AUC (0.622), followed closely by the GRS_effector model at P < 1e–04 (AUC = 0.619). The AUC of the GRS_all model was significantly higher than that of the GRS_effector model (DeLong test = P < 0.001). Both models showed significantly higher AUCs compared to the GRS_noncolocalized model (AUC = 0.618), with the DeLong test = P < 0.001 versus GRS_all and P = 0.029 versus GRS_effector. However, GRS_effector models constructed using more stringent thresholds (P < 1e–05 and P < 5e–08) performed similarly (AUC = 0.617 for both models) to the GRS_noncolocalized model. In South Asians, the GRS_effector model at P < 5e–08 and GRS_all achieved the highest AUC (0.611). The difference in AUC between the GRS_all and GRS_effector P < 5e–08 models was not significant (DeLong test = P = 0.966). Both the GRS_all and GRS_effector at P < 5e–08 models outperformed the GRS_noncolocalized model, with statistically significant improvement for GRS_all (DeLong test = P = 0.004) and a borderline significance for GRS_effector at P < 5e–08 (DeLong test = P = 0.053). In African ancestry and East Asians, no statistically significant differences were observed among the GRS models. In summary, in the European population, both GRS_all and GRS_effector showed significantly higher AUCs than GRS_non-colocalized. In South Asians, GRS_all and GRS_effector were comparable and marginally exceeded GRS_non-colocalized. No significant AUC differences were observed among models in East Asian or African populations (Fig. 4A; Supplementary Table S19).

Figure 4.

AUC and Nagelkerke R² Comparison of GRS Models. (A, B) Shows the AUC and Nagelkerke R² of GRS comparison across different ethnicities using various SNP selection GRS models.

Nagelkerke R² was calculated to assess the proportion of POAG risk variation explained by each GRS in logistic regression models, providing a standardized measure of model explanatory power for binary outcomes. Nagelkerke R² ranges from 0 to 1, with values closer to 0 indicating minimal explanatory contribution of the GRS beyond the null model, and values closer to 1 indicating near-complete discrimination between cases and controls.³⁵ The differences in Nagelkerke R² across models within each ancestry group were minimal. The European GRS model showed the highest R² (0.0835–0.0848), followed by models in African (0.0710–0.0723), East Asian (0.0688–0.0699), and South Asian (0.0645–0.0662) populations (Fig. 4B).

Drug Target Analysis Identified 16 Drugs Shared Between POAG Subtypes

We performed drug target analysis using NeDRex for genes associated with POAG effector variants, IOP-dependent effector variants, and IOP-independent effector variants. The top 50 ranked drugs were selected. The resulting drug–target network is shown in Figure 5.

Figure 5.

Disease model and drug rank derived from NeDRex. (A, B, C) Show the POAG, IOP_dependent, and IOP_independent model derived by MUST, combined with its targeting 50 top-ranked drugs from TrustRank. Squares represent genes, hexagons represent drugs, and edges indicate interactions between them. (D) Illustrates the overlap of target drugs among IOP-dependent effector module, IOP-independent effector module, and POAG effector model.

Figure 5.

Continued.

First, disease models were constructed using 19, 9, and 13 effector genes derived from the colocalization analysis for POAG, IOP-dependent, and IOP-independent mechanism, respectively (Supplementary Tables S20, S21, S22). ABCA1 and TMCO1 were shared across all three input sets. Using the MuST algorithm, these gene sets were expanded to three gene-gene interaction networks, which contained 51, 31, and 53 genes, respectively. In addition to the original shared genes, NCOR1 was newly identified as a shared gene across all expanded gene sets.

Then, drug repurposing analysis were conducted according to the gene networks. Based on TrustRank scores, the top three drugs targeting POAG effector genes were caffeine (TrustRank score = 7.55 × 10⁻⁵), dipyridamole (TrustRank score = 7.29 × 10⁻⁵), and flavoxate (TrustRank score = 6.57 × 10⁻⁵; Supplementary Table S23). For IOP-dependent genes, the highest-ranked candidates were glutamic acid (TrustRank score = 1.67 × 10⁻⁴), topiramate (TrustRank score = 1.66 × 10⁻⁴), and butabarbital (TrustRank score = 1.66 × 10⁻⁴; Supplementary Table S24). In contrast, the top drugs associated with IOP-independent genes were fostamatinib (TrustRank score = 6.2 × 10⁻⁵), sunitinib (TrustRank score = 5.6 × 10⁻⁵), and nintedanib (TrustRank score = 4.17 × 10⁻⁵; Supplementary Table S25). Notably, 16 drugs—including sunitinib, zinc, and copper—were shared across all three gene sets (Table 2; see Figs. 5D, 6), whereas an additional 30 drugs, such as caffeine and NADH, were shared between 2 sets. These overlaps suggest the existence of common molecular mechanisms and highlight the potential for shared therapeutic strategies across different POAG subtypes.

Table 2.

Drugs Sharing Across POAG, IOP_dependent, and IOP_independent Gene Sets

	Rank_	Score_	Rank_	Score_	Rank_	Score_
Drug	POAG	POAG	IOP_dependent	IOP_dependent	IOP_independent	IOP_independent
Zinc Chloride	6	5.68E–05	23	1.79E–05	19	1.84E–05
Fostamatinib	8	5.19E–05	9	6.23E–05	1	6.20E–05
Sunitinib	9	4.58E–05	10	5.50E–05	2	5.60E–05
Nintedanib	10	3.73E–05	11	4.64E–05	3	4.17E–05
Midostaurin	12	3.64E–05	15	3.52E–05	4	3.91E–05
Bosutinib	13	3.35E–05	12	4.45E–05	5	3.74E–05
Copper	14	3.00E–05	22	2.43E–05	8	2.79E–05
Crizotinib	15	2.54E–05	14	3.53E–05	7	2.81E–05
Dasatinib	16	2.31E–05	16	3.29E–05	9	2.57E–05
Glyburide	18	2.08E–05	17	3.07E–05	14	2.06E–05
Axitinib	20	1.91E–05	25	1.75E–05	16	1.95E–05
Probucol	21	1.90E–05	18	2.93E–05	18	1.87E–05
Vandetanib	22	1.88E–05	19	2.90E–05	11	2.18E–05
Neratinib	23	1.79E–05	20	2.81E–05	17	1.91E–05
Erlotinib	24	1.68E–05	27	1.70E–05	20	1.82E–05
Astemizole	26	1.16E–05	38	1.08E–05	29	1.08E–05

Score represents the TrustRank algorithm score for the drug.

Rank indicates the drug's rank based on this score.

Figure 6.

The common drugs among three gene sets model. Illustrates the common drugs among IOP-dependent model, IOP-independent model, and POAG module. The picture shows the key structures of drugs, drug type, and the relevant researches on this drugs’ application in glaucoma.

Discussion

Rather than identifying additional genome-wide significant loci, as many previous GWAS studies have aimed to do, our study focused on prioritizing functionally relevant variants among established POAG-associated loci as post-GWAS analysis. By integrating multi-omics data with POAG GWAS results, we identified 25 effector variants among 66 genome-wide significant independent loci, enhanced the biological interpretation of POAG GWAS findings. These prioritized effector variants showed stronger cross-population consistency than POAG-associated variants without colocalization evidence. They demonstrated stronger correlations in POAG effect sizes across populations, smaller allele-frequency differences, and a higher proportion of intronic variants. These findings suggest that the prioritized variants may have greater functional relevance and potential trans-ethnic transferability. Consistently, a genetic risk score constructed from these effector variants showed a degree of generalizability in non-European populations. These effector SNPs also indicate novel therapeutic targets. Our drug target analysis identified 17 candidate drugs with potential relevance to different POAG and its subtypes, including IOP-dependent and IOP-independent POAG. Some candidates, such as lipid-lowering agents, are supported by previous epidemiological and clinical evidence, whereas others, including anticancer drugs and histamine receptor antagonists, are supported by emerging experimental or mechanistic evidence but require further clinical validation in POAG. Together, these findings provide post-GWAS evidence for functionally relevant POAG effector variants and genes, support their potential cross-population transferability, and highlight candidate therapeutic opportunities for future validation.

Our approach integrated POAG GWAS summary statistics with six types of xQTL omics data for joint colocalization analysis. Unlike previous studies that used limited omics types, our multi-omics strategy provided a more comprehensive view of the molecular mechanisms underlying POAG. For example, Hamel et al. applied eCAVIAR and enloc to assess colocalization between POAG/IOP GWAS loci and cis-eQTLs/sQTLs, identifying colocalization in 58% of tested loci.¹⁴ Similarly, Yuan et al. used SMR and Bayesian colocalization to link plasma proteins to POAG, identifying eight potential therapeutic targets.¹⁵ Compared with these studies, our integration of a broader spectrum of omics data revealed loci associated with multiple molecular traits, enabling a more complete understanding of disease biology. Our multi-omics colocalization strategy integrates six molecular phenotypes spanning the hierarchical gene regulatory cascade—from upstream DNA regulation (caQTLs for chromatin accessibility and mQTLs for DNA methylation) through transcriptional (eQTLs) and post-transcriptional (sQTLs) processing to downstream protein abundance (pQTLs). These datasets, each previously validated for causal variant identification in complex traits, provide a robust foundation prioritizing candidate effector genes.^14,15,20 This layered coverage provides distinct advantages over previous single-omics approaches. First, it enables mechanistic triangulation: when a variant colocalizes with multiple molecular traits along the same regulatory axis, it provides orthogonal evidence of coordinated biological regulation rather than spurious association due to LD. Second, as demonstrated in the OPERA methodology paper,³⁵ joint multi-omics analysis increases discovery power by reducing multiple-testing burden. This multi-layer convergence strengthens causal inference through biological replication, yielding a refined set of candidate effector genes with enhanced cross-ancestry conservation and greater potential for therapeutic translation.¹⁹ Moreover, its use of the HEIDI test helped reduce spurious associations by accounting for cross-exposure linkage between omics layers.

Although additional QTL datasets could theoretically expand effector gene discovery, two constraints guided our selection. First, OPERA becomes computationally infeasible with >10 distinct molecular phenotypes.²⁰ Second, tissue relevance outweighs quantity: although GTEx demonstrates high cis-eQTL sharing across bulk tissues,³⁶ ocular tissue QTLs remain unavailable at scale and nonocular tissues likely miss eye-specific regulatory mechanisms relevant to POAG. We therefore prioritized regulatory layer diversity (epigenetic → transcriptional → protein) within OPERA’s optimal range rather than incorporating numerous nonocular tissue eQTLs that would increase computational burden without guaranteed biological relevance.

In the colocalization analysis, we could not find SNPs with P < 5e–08 colocalized with caQTLs and pQTLs (see Fig. 2B) but in SNPs with less stringent threshold (P < 1e–05 or P < 1e–04). Consistent with the OPERA framework’s observation that most colocalized signals arise from upstream regulatory layers (e.g., methylation and expression) rather than downstream protein abundance. Two factors are likely to contribute to this pattern. First, proteins represent a distal molecular phenotype in the central dogma; their abundance is influenced not only by genetic variation but also by post-transcriptional regulation, translation efficiency, and protein degradation—processes that dilute genetic effects and necessitate substantially larger sample sizes for robust detection. Our pQTL dataset, although derived from 35,559 individuals, covered only 1198 proteins in final analysis, far fewer than the >10,000 proteins detectable in human plasma³⁷; whereas caQTLs were severely underpowered (N = 272). Second, nearly all xQTL datasets originated from nonocular tissues (e.g., blood and brain cortex), potentially missing regulatory mechanisms specific to trabecular meshwork or RGCs. Future studies incorporating larger, ocular tissue-derived pQTL/caQTL resources may uncover additional protein-mediated biological pathways.

For mQTL analysis, we used whole blood data rather than ocular tissue-specific data. The mQTLs exhibit a degree of tissue specificity and using whole blood data may limit our ability to identify methylation signals directly related to retinal or ocular tissues.³⁸ However, POAG has a multifactorial etiology, involving IOP, vascular circulation, and other factors. Many blood circulation related traits are involved in the development and progression of POAG. For example, studies have shown that a lower oxygen consumption rate in peripheral blood mononuclear cells is a novel biomarker for POAG.³⁹ In addition, a Mendelian randomization study identified a causal relationship between peripheral blood high-density lipoprotein 3 cholesterol and POAG.⁴⁰ Several studies have also reported alterations in retinal vascular morphology and function in patients with POAG.^41,42 As an important circulation-related indicator, blood DNA methylation has the potential to provide insights into the etiology of POAG. Previous studies have shown partial overlap in DNA methylation patterns between blood and multiple ocular tissues. Specifically, blood methylation profiles show a relatively high correlation with retinal pigment epithelium/choroid tissues, suggesting that peripheral blood may capture some ocular tissue-related epigenetic signals.⁴³ In addition, ocular tissues, particularly the trabecular meshwork, are much more difficult to obtain than whole blood, and consequently, methylation studies in these tissues are relatively limited. Taken together, we therefore chose whole-blood mQTL data because of its larger sample size and greater robustness.

We noted variants rs41543317 and rs6687545 colocalized with eQTL, mQTL, and sQTL data, indicating their involvement in multiple layers of molecular regulation. SNP rs41543317 is linked to the MAPT gene, which encodes microtubule-associated protein tau. Tau promotes microtubule assembly and stability and plays a role in establishing and maintaining neuronal polarity.⁴⁴ Abnormal phosphorylation of tau can lead to neurotoxic aggregates.⁴⁵ MAPT is a well-established risk locus for both Alzheimer's and Parkinson's diseases and is shared across POAG, optic neuropathy traits, and neurodegenerative disorders.⁴⁶ In glaucoma surgical specimens, reduced levels of normal tau and accumulation of abnormal tau have been observed in the optic nerve head and retina.⁴⁷ In addition, rs6687545 is associated with GNL2 and RSPO1. GNL2 encodes G protein nucleolar 2, which is involved in ribosome biogenesis.⁴⁸ Loss of GNL2 disrupts the timely cell cycle exit of retinal progenitor cells due to dysregulated expression of cell cycle regulators.⁴⁹ Whereas RSPO1 encodes a ligand for leucine-rich repeat-containing G-protein-coupled receptors and has been shown to negatively regulate the TGF-β pathway.^50,51 Because TGF-β signaling promotes trabecular meshwork fibrosis—a key pathological feature in glaucoma—RSPO1 may represent as a potential therapeutic target for glaucoma.⁵²

In the European population, the AUC of the GRS_ALL model was significantly higher than that of other models. Because the colocalized variants were derived from European-based datasets, GRS_effector did not outperform GRS_ALL. However, GRS_effector (P < 1e–04) still showed a significantly higher AUC than GRS_noncolocalized, indicating that effector variants possess stronger functional relevance than non-colocalized variants in Europeans. In the South Asian population, GRS_effector (P < 5e–08) also outperformed GRS_noncolocalized, and the difference between GRS_ALL and GRS_effector was not significant. In the African and East Asian populations, the differences among GRS models were not statistically significant. Although the predictive performance of GRS_effector remains modest and insufficient for clinical use, it did not underperform GRS_all in non-European populations, and the performance gap between these models narrowed compared with Europeans. GRS_effector also yielded higher AUCs than GRS_non-colocalized across most ancestries. Although limited by small non-European sample sizes, these patterns tentatively suggest that effector variants—prioritized through multi-omics colocalization—may enrich for functionally relevant signals with greater cross-ancestry conservation. Validation in larger, diverse cohorts should be conducted before applying such a filtering strategy to achieve more equitable polygenic risk prediction.

In the drug target analysis, we first constructed disease modules centered on effector genes and in the following drug repurposing analysis, 16 shared candidate therapeutics were identified across POAG and the two subtypes (see Fig. 5). Probucol, a lipid-lowering drug targeting ABCA1, was found to be a potential therapeutic drug for the POAG and the two subtypes. Previous studies have shown that ABCA1 protects optic nerve cells by inhibiting apoptosis and regulating cholesterol metabolism.^53–55 It was also reported that dysregulation of cholesterol transport may play a role in the pathogenesis of POAG⁴⁰ and a lipid lowering drug may be a potential glaucoma medicine by lowering IOP and promoting RGC survival.^56,57 Furthermore, multiple tyrosine kinase inhibitors (TKIs)—including sunitinib, fostamatinib, and bosutinib—were prioritized, and they have been shown to promote RGC survival following optic nerve injury.^58–60 Interestingly, these drugs target distinct genes (e.g., EGFR, SPRK1, and BTK) that might belong to the same functional pathway—revealing pathway-level convergence rather than identical gene targets. Although these TKIs may cause substantial systemic toxicity when used as anticancer agents, local intraocular delivery via eye drops or intravitreal injection can reduce systemic exposure. Sustained-release formulations may further improve safety while maintaining their therapeutic potential for ocular diseases.^59,60 A sustained intraocular delivery formulation of sunitinib is currently being evaluated in a phase 2 clinical trial (ClinicalTrials.gov Identifier: NCT04085341). Astemizole, as a histamine receptor antagonist targeting both EGFR and MAPT, has been shown to reduce IOP and to exert protective effects on the optic nerve through its anti-inflammatory and antioxidant properties.^61,62 Although no clinically approved drugs currently target TMCO1, it represents a promising candidate for glaucoma and its subtype therapies, TMCO1 is expressed in multiple tissues in the POAG eye, including the trabecular meshwork and retina.⁶³ Genetic variation disrupting TMCO1 function is strongly associated with elevated glaucoma risk, whereas inhibition of TMCO1 expression by si-RNA significantly attenuated drug-induced IOP elevation in a rat model.⁶⁴ For other shared drugs, zinc and copper levels have been associated with glaucoma, with zinc typically decreased and copper elevated in both the blood and aqueous humor of affected individuals.^65,66 NADH was found as an intervention for both the IOP-dependent and IOP-independent sets. NADH may offer neuroprotection in glaucoma by enhancing oxidative phosphorylation, buffering metabolic stress, increasing mitochondrial size and motility, and suppressing neuronal firing rates.^67,68

Therapeutics shared across all three glaucoma subtypes tend to exhibit broader, dual mechanisms that combine neuroprotection with IOP modulation (e.g., probucol and astemizole). In contrast, candidates restricted to individual subtypes demonstrate more targeted actions. For example, caffeine was enriched in the POAG and IOP-dependent modules but was absent from the IOP-independent module—suggesting pathway relevance to IOP-mediated mechanisms rather than IOP-independent pathways. This pattern aligns with observational evidence that caffeine elevates IOP in patients with glaucoma.^69,70 However, recent Mendelian randomization analysis indicated no causal relationship between dietary caffeine intake and baseline IOP or glaucoma.⁷¹ This suggests caffeine-related pathways may modulate IOP-dependent mechanisms without driving IOP-independent neurodegeneration. Conversely, paclitaxel and docetaxel—which target MAPT—were prioritized in both POAG and IOP-independent modules. As mentioned previously, MAPT is predominantly related to neurodegenerative disorders.⁴⁴ Although a normal-tension glaucoma mouse model (OPTN E50K) exhibits an Alzheimer’s-like proteomic profile,⁷² no direct experimental evidence yet demonstrates whether MAPT modulation can reverse glaucomatous neurodegeneration through IOP-independent mechanisms.

Our study has several limitations. First, most xQTL data were derived from nonocular tissues, reducing tissue specificity. The current OPERA method does not support large-scale integration of multiple xQTL datasets. Future studies using larger, tissue-relevant, and more comprehensive multi-omics data, together with advanced colocalization methods, may improve functional locus discovery and biological interpretation. Second, most GWAS to date are of European ancestry, limiting discovery and transferability. We therefore use colocalization to nominate effector genes for GRS derivation and assess cross-ancestry performance. However, the African and Asian subsamples in the UK Biobank remain small, reducing power for validation. As larger, ancestry-diverse GWAS and sequencing resources become available, population-specific signals can be resolved and more universally predictive GRS developed. Third, DNA methylation level changes with aging.⁷³ The mQTLs used in this study were derived from cohorts with mean ages of 70 and 79 years, which may introduce potential bias when extrapolating colocalization results to younger populations. However, mQTL genetic effects are reported to be relatively stable over time,⁷⁴ and the POAG GWAS cohorts in this study are also largely older (mean ages at diagnosis = approximately 55–70 years). Thus, the age distributions are broadly comparable, supporting the reliability of our colocalization results. Future work with age-stratified xQTL data will clarify age-dependent regulatory mechanisms in POAG. Finally, the biological relevance of loci identified through OPERA and the therapeutic potential of candidate drugs prioritized by NeDRex remain to be experimentally validated. Currently, clinically used glaucoma medications do not target genetically validated POAG gene targets. Further experimental studies are required to evaluate the efficacy of gene-targeted therapies.

Conclusions

By integrating multi-omic data, our study identified 25 POAG effector loci linked to gene expression, splicing, and methylation regulation, offering deeper insight into the molecular mechanisms underlying glaucoma. The reduced population diversity observed in both allele frequency and effect size among effector variants suggests their potential utility in developing more robust genetic risk prediction models and broadly effective therapeutic strategies for POAG. As large-scale sequencing efforts continue to expand, the discovery of additional glaucoma-associated variants and tissue-specific QTLs is expected. Incorporating these data into colocalization analysis will facilitate the identification of new effector variants and further elucidate the complex biological processes underlying POAG.