Multiethnic Genome-Wide Association Study of Diabetic Retinopathy Using Liability Threshold Modeling of Duration of Diabetes and Glycemic Control

Diabetic retinopathy (DR) is a leading cause of blindness (1). Established risk factors include longer duration of diabetes (DoD) and poor glycemic control (2). Genetic factors are also implicated, with heritability of 52% for proliferative DR (PDR) (3,4). Several candidate gene and genome-wide association studies (GWAS) have been conducted (5–11). Although several polymorphisms have been suggested to be associated with DR, few have been convincingly replicated (10,12–15).

There are several reasons why studies have not yielded consistent findings. The genetic effects are likely modest, and identification requires large sample sizes. Previous studies have not consistently accounted for the strongest two covariates, DoD and glycemic control. Liability threshold (LT) modeling is one way to incorporate these covariates while also increasing statistical power (16). Finally, previous genetic studies have largely examined individual variants. Techniques that examine top GWAS findings collectively for variants that cluster in biological networks based on known protein–protein interactions have the potential to identify variants where there is insufficient power to detect their individual effects.

The purpose of this study was to identify genetic variants associated with DR by 1) assembling a large sample size through inclusion of multiple ethnicities, 2) incorporating DoD and glycemic control via LT modeling, and 3) collectively examining variants that cluster in biological networks.

All studies conformed to the Declaration of Helsinki tenets and were Health Insurance Portability and Accountability Act compliant. Written informed consent was obtained from all participants. Institutional Review Board/Ethics Committee approval was obtained by each individual study.

The discovery sample, encompassing 7 African American and 8 European cohorts, arose from a consortium of 11 DR studies for a total of 3,246 Europeans and 2,611 African Americans (6–8,12,13,17,18). Inclusion criteria for the discovery stage were 1) type 2 diabetes, and 2) European or African American ethnicity. Type 2 diabetes was defined as a fasting plasma glucose (FPG) ≥126 mg/dL (7.0 mmol/L) or a hemoglobin A_1c (HbA_1c) ≥6.5% (48 mmol/mol) (19) with onset of the diabetes after 30 years of age. Table 1 summarizes the DR phenotyping protocols and covariates by discovery cohort. Phenotyping protocols have been previously described (4,20–29), and additional details are in the Supplementary Data.

The analysis plan prespecified four DR case-control definitions with varying Early Treatment Diabetic Retinopathy Study (ETDRS) score thresholds for case and control subjects (Table 2) (30). The primary case-control definition compared any DR to no DR (ETDRS ≥14 vs. ETDRS <14, henceforth referred to as the any DR analysis). There were three secondary case-control definitions. The first compared patients with PDR to those without PDR (ETDRS ≥60 vs. ETDRS <60, henceforth the PDR analysis). The second compared those with nonproliferative DR (NPDR) or worse to those without DR (ETDRS ≥30 vs. ETDRS <14, henceforth the NPDR analysis). The third compared those with PDR to those without DR (ETDRS ≥60 vs. ETDRS <14, henceforth the extremes of DR analysis). The rationale for the four definitions is in the Supplementary Data. Table 1 shows the available samples by cohort and ETDRS score thresholds. Supplementary Table 1 summarizes the mean values for glycemic control and DoD.

The genotyping platforms and numbers of single nucleotide polymorphisms (SNPs) genotyped are summarized in Supplementary Table 2. Details about quality control, imputation, and data filtering are in the Supplementary Data. Supplementary Fig. 1 provides a flowchart of the discovery and replication analyses. For the four main case-control definition analyses, we performed each of the analyses 1) without incorporating DoD and glycemic control using EIGENSOFT (16,31) and 2) with LT modeling of DoD and glycemic control using LTSCORE (16). LT modeling details are in the Supplementary Data. Both the EIGENSOFT and LTSCORE tests were implemented in LTSOFT version 2.0 (see Web Resources in the Supplementary Data). For the discovery analyses, we ran principal components (PC) analysis with EIGENSTRAT using only typed SNPs and five PCs, separately by ethnicity and case-control definition (32). We computed association analyses for each of the seven African American and eight European cohorts separately and then meta-analyzed by ethnicity. Meta-analysis was performed using inverse-variance weighting, accounting for both effective sample size (defined as 4/[1/N_case + 1/N_control]) and allele frequency (33). We also performed multiethnic (Europeans and African Americans together) meta-analyses for the any DR and PDR analyses using inverse-variance weighting and a sensitivity analysis of the any DR meta-analyses in African Americans and Europeans (see Supplementary Data). Because we included rare variants in this GWAS, we also tested the robustness of the top associations (P < 5 × 10⁻⁸) by performing two additional tests: 1) a Fisher exact test on case or control subjects aggregated across all cohorts tested per variant and on each cohort separately, and 2) an inverse variance-weighted meta-analysis across cohorts using the ln of the odds ratio (OR) as the effect size (34) without adjusting for covariates.

The P value thresholds for genome-wide significance were based on empirically determined thresholds for different ancestral populations that account for the GWAS multiple testing burden, as well as population-specific linkage disequilibrium (LD) patterns (35):

•  1. P < 3.24 × 10⁻⁸ for SNPs ascertained in African ancestry populations

•  2. P < 5.0 × 10⁻⁸ for SNPs ascertained in European ancestry populations

•  3. P < 3.24 × 10⁻⁸ for SNPs ascertained in multiethnic meta-analyses

We further corrected these thresholds for additional multiple testing from examination of four case-control definitions, each with and without covariate incorporation, for eight tests total. This yielded the following P value thresholds for our study:

•  4. P < 3.75 × 10⁻⁹ for SNPs ascertained in African ancestry populations

•  5. P < 6.25 × 10⁻⁹ for SNPs ascertained in European ancestry populations

•  6. P < 3.75 × 10⁻⁹ for SNPs ascertained in multiethnic meta-analyses

We note that correction for eight tests is conservative because the case-control definitions are not completely independent. We did not apply further multiple testing correction for the different ancestries analyzed.

Eight European, eight Asian, and four Hispanic replication cohorts provided summary statistics on SNPs with P < 1 × 10⁻⁵ in the discovery analyses (Table 3). Their phenotyping/genotyping protocols have been previously described, and details are in the Supplementary Data (6–8,12,13,17,18). The rationale for including additional ethnicities in the replication phase is that high transethnic genetic correlations have been documented for type 2 diabetes and other traits/diseases and support the use of multiethnic studies to increase sample size (36). Supplementary Table 3 summarizes the replication cohorts’ mean values for HbA_1c, FPG, and DoD. Replication was in silico with existing genotyping. LT modeling was not applied to the replication cohort analyses. The replication cohorts used standard covariate adjustment in their regression models. Replication meta-analysis was also performed using inverse-variance weighting, first individually by each ethnicity (Europeans, Hispanics, and Asians) followed by all cohorts combined. Replicated genome-wide significance had to meet the aforementioned thresholds after meta-analysis of the discovery and replication results.

To identify significantly enriched protein networks among the loci with the highest statistical evidence for association with DR, we applied the Disease Association Protein-Protein Link Evaluator (DAPPLE) to our discovery GWAS (37). It has been shown that top associated loci, despite not being genome-wide significant, tend to cluster in biological networks (37,38). For this reason, we examined the top 1,000 loci from the discovery GWAS in the two monoethnic analyses (European and African American) and for each of the four case-control definition analyses that incorporated DoD and glycemic control (eight network analyses in total). Our threshold for significance was therefore P < 0.00625 (0.05 corrected for eight tests). We used the publically available version of DAPPLE, and the protocol is outlined in the Supplementary Data. This methodology has been used successfully with previous GWAS to identify protein networks with biological relevance (37–39).

To further support the protein–protein interaction results from the DAPPLE analysis, we applied gene set enrichment analysis (GSEA) using Meta-Analysis Gene-Set Enrichment of variaNT Associations (MAGENTA) (40) to the set of genes significantly enriched for protein–protein interactions in the DAPPLE analysis (details in Supplementary Data).

To understand to what extent genetic determination of DR might reflect enrichment for type 2 diabetes or glycemic control genes, we computed a correlation between case status in the any DR analysis and the sum of the β*risk allele (for quantitative glycemic traits) or logOR*risk allele (for type 2 diabetes) of the trait-associated SNPs for each cohort and each trait (see Supplementary Data for details).

Supplementary Fig. 2 shows the PC analysis. We observed little inflation in the association statistic distribution (Supplementary Fig. 3), indicating no significant population stratification as a confounder. Supplementary Fig. 4 shows the Manhattan plots for the any DR analyses. Supplementary Tables 4–25 show the top 10 SNPs for independent loci with the lowest P values for each discovery analysis, including the sensitivity analyses (full results are available on the Type 2 Diabetes Knowledge Portal [http://www.type2diabetesgenetics.org/], both on the downloads page and fully integrated into the portal modules).

Table 4 shows SNPs that met the traditional nominal threshold for genome-wide significance of P < 5 × 10⁻⁸ from the discovery analyses. All of the SNPs in Table 4 were either from the PDR or extremes of DR analyses; Fig. 1 shows the QQ and Manhattan plots for the PDR and extremes of DR analyses. The results for the associations in Table 4 are shown for each cohort separately in Supplementary Table 26. Results for these SNPs after meta-analysis with replication samples both combined and separated by ethnicity are shown in Table 5 and Supplementary Table 27, respectively.

Using the corrected significance thresholds, only one SNP in the discovery meta-analyses met genome-wide significance: rs142293996 for the extremes of DR analysis incorporating DoD and glycemic control in Europeans (P = 2.1 × 10⁻⁹). The association was not significant without adjusting for covariates based on a Fisher exact test (Supplementary Table 28). This is an intronic variant in the nuclear VCP-like (NVL) gene, which encodes a member of the ATPases associated with diverse cellular activities (AAA) superfamily (41). The NVL gene is widely expressed in vivo with highest expression in retina (https://www.proteinatlas.org/ENSG00000143748-NVL/tissue#top).

We tested whether this association was a significant cis-expression quantitative trait locus (eQTL) in the Genotype-Tissue Expression (GTEx) Project release v7 (see Supplementary Data for eQTL analysis details). This variant, rs142293996, lies in the 22nd intron of NVL and is in LD (r² = 0.62) with variant rs41271487 in the 24th intron of NVL. rs41271487 is a significant eQTL (P = 6.4 × 10⁻⁶; effect size 1.27) in the GTEx spinal cord cervical c-1 tissue, targeting calpain 2 (CAPN2), a calcium-activated neutral protease (Supplementary Fig. 5). Common variants in the intron or regulatory region of CAPN2, 527–576 kb upstream of the DR association, are associated with variation in serum α-carotene levels (42), a vitamin A precursor required for sight, supporting a functional role for this gene. Based on the eQTL analysis, increased expression of CAPN2 is associated with decreased risk of DR (Supplementary Fig. 6). CAPN2 is expressed in the retina (https://www.proteinatlas.org/ENSG00000162909-CAPN2/tissue).

When examined in the replication analyses (which included a more diverse population), the direction of effect in the replication cohorts for rs142293996 was the same, but the meta-analysis P value was not genome-wide significant (P = 4.10 × 10⁻⁶).

In African Americans, the SNP with the lowest P value was rs115523882 from the PDR analysis (P = 5.37 × 10⁻⁹). This was short of the 3.75 × 10⁻⁹ threshold for significance in African Americans. We could not reproduce this finding in the replication cohorts. This variant is located near the GOLIM4 gene, which helps process proteins and mediates protein transport. The SNP rs115523882 specifically changes a motif that is a binding site for Nlx3, a transcription factor in blood, suggesting it plays a regulatory role. This variant is mainly present in people of African ancestry (minor allele frequency [MAF] = 0.0393) and not common in other ethnic groups, suggesting we may have had insufficient power to replicate it.

Of note, there was one SNP, rs184340784, suggestively associated with DR (P = 3.52 × 10⁻⁸) in the extremes of DR analysis without covariates in African Americans that was not present in our replication cohorts (due to low MAF) and thus could not be replicated. Neither rs115523882 nor rs184340784 was analyzed for eQTL activity in GTEx due to their low MAF (MAF < 0.01 in GTEx tissues).

Table 6 and Supplementary Table 29 show the discovery variants with P < 1 × 10⁻⁵ that achieved a nominal P < 0.05 in the complete replication sample or in one of the replication ethnicities, respectively, and had the same direction as the discovery samples. None of these variants achieved genome-wide significance after discovery and replication meta-analysis, as defined above.

One protein network from the African American PDR analysis was significant (P = 0.0009) for average binding degree within the network (Fig. 2). The aforementioned top-ranked SNP (rs115523882) could not be included in the DAPPLE analysis because its nearby gene (GOLIM4) is not in the protein database. The significant protein network includes genes with primary roles in inflammation including IFNG, IL22RA1, CFH, and SELL. IFNG encodes interferon-γ, which is highly expressed in ocular tissues from patients with PDR (43). IL22RA1 encodes the IL-22 receptor, and CFH encodes complement factor H; both proteins are suspected to play a role in PDR (44,45). SELL encodes l-selectin, which is expressed at higher levels in lymphocytes from patients with DR and associated with increased endothelial adhesion (46). We did not identify any statistically significant protein networks for any of the other case-control definitions in African Americans or Europeans.

We examined the 41 genes in the significant network identified by the DAPPLE analysis via GSEA using MAGENTA. The genes showed a significant (16.5-fold) enrichment of low association P values in the African American PDR analysis (P < 1 × 10⁻⁶) (Supplementary Fig. 7 and Supplementary Table 30) and to a lesser extent in African American extremes of DR analysis (P = 2 × 10⁻⁴) (Supplementary Table 30), suggesting new DR associations of modest effects in African Americans (Supplementary Table 31). No significant gene set enrichment was found for the PDR and extremes of DR analyses in Europeans.

The results of the correlation analysis between type 2 diabetes/glycemic trait-associated SNPs and DR case status are shown in Supplementary Table 32. The Z score for type 2 diabetes was +2.256 (P = 0.024). The correlation coefficient R was positive, indicating that a greater burden of SNPs that increase type 2 diabetes risk is correlated with having DR. However, this Z score was not significant after correcting for the six hypotheses (six traits) tested.

We extracted results from our discovery meta-analysis for the variants with the lowest association P values from previously published DR GWAS or large candidate gene studies (Supplementary Table 33). There were three variants that were nominally significant (P < 0.05) in our sample and had the same direction of effect as in the previously published studies. Two of the variants, rs9896052 and rs6128, were from previous studies for which samples overlapped with some samples in our discovery meta-analysis and therefore do not represent independent replication (10,20). Variant rs1399634, originally found in Chinese patients (P = 2 × 10⁻⁶), was nominally significant in our European discovery cohort (P = 0.0124). Meta-analysis of the original study and our cohorts was performed using the same method as our discovery and replication meta-analyses and was short of genome-wide significance (OR 1.47; P = 9.63 × 10⁻⁸).

To our knowledge, this study represents the largest GWAS performed for DR. The discovery analysis included 3,246 Europeans and 2,611 African Americans. The replication analysis included 18,545 Europeans, 16,453 Asians, and 2,710 Hispanics. Despite the relatively large sample size, we did not identify any individual variants that were associated at a genome-wide significant level after meta-analysis with multiethnic replication cohorts. However, among the most significant results in the African American PDR analysis, we did identify a statistically significant enrichment for a network of genes using DAPPLE, which was corroborated by GSEA using MAGENTA.

In the discovery meta-analyses, several variants from the PDR and extremes of DR analyses achieved nominal genome-wide significance of P < 5 × 10⁻⁸, but the only variant to achieve genome-wide significance after conservative multiple testing correction was rs142293996 in the European analysis for extremes of DR (P = 2.1 × 10⁻⁹). It is notable that the variants with the most significant findings came from the two case-control definitions that have PDR as their case definition. This is consistent with the fact that PDR has a higher heritability than overall DR (4). Although the most strongly associated variants in the discovery analyses (rs142293996 in NVL in Europeans and rs115523882 in GOLIM4 in African Americans) did not reach genome-wide significance with replication, it is still possible that they do play a role in DR pathogenesis. NVL is highly expressed in the retina, and the implicated variant is in LD with an eQTL acting on CAPN2 with functional implications in neural tissue. The eQTL variant falls in a binding site of a transcription factor (47). The GOLIM4 variant also has a known regulatory role.

We could not replicate the association with rs142293996 when we used the Fisher exact test, although the Fisher exact test did not allow for covariate incorporation. There is potential for inflated false-positive rate when standard association methods are applied to rare (e.g., MAF <1%) variants in imbalanced (e.g., case fraction <10%) case-control cohorts at modest sample sizes (48). However, most cohorts in this study did not have case fraction <10%. Larger sample sizes will help determine the confidence in these top associations.

There was one variant suggestively associated in the extremes of DR discovery analysis in African Americans, rs184340784, which was not present in any replication data sets. The T allele of this variant has a frequency of 0.0023 in African populations and 0 in European, East Asian, South Asian, and Hispanic populations in the 1000 Genomes phase 3 panel. In the discovery analysis, the P = 3.52 × 10⁻⁸ was shy of the genome-wide significance threshold of 3.75 × 10⁻⁹ for variants discovered from the African ancestry analyses. This variant is within an intronic region upstream of adherens junctions–associated protein 1 (AJAP1), which has its highest expression in brain frontal cortex but is also expressed in the retina (https://www.proteinatlas.org/ENSG00000196581-AJAP1/tissue).

In the DAPPLE analysis, we did find that the top signals for the PDR analyses in African Americans analysis were enriched for a biologic network. The advantage of DAPPLE is that it can identify a protein pathway that may not be evident solely from the primary individual variant GWAS. The presence of an underlying network among the top loci suggests there are likely true associations within top findings that have yet to reach genome-wide significance due to limited power. Multiple pathways including inflammatory pathways are implicated by this network. To confirm biological significance, these results will need to be followed up with functional in vitro studies.

The DAPPLE results were corroborated by the MAGENTA GSEA in the African American PDR and extremes of DR analyses. This network of genes, however, was not enriched for in Europeans. This could either be due to technical differences (e.g., the number of African American cases is approximately threefold larger than the number of European cases) or due to biological reasons. For example, we found that the allele frequencies of the most significant variant per gene for 40% of these protein-interacting genes are rare in Europeans (MAF <0.2%), whereas they are common in African Americans (MAF >1%), according to the Genome Aggregation Database (see Web Resources in the Supplemental Data).

In the analysis between type 2 diabetes/glycemic trait SNPs and DR case status, only type 2 diabetes variants were significantly associated with DR prior to, but not after, multiple testing correction. One previous study examined aggregate effects of 76 type 2 diabetes–associated variants in Asian patients (49). Participants in the top tertile of type 2 diabetes risk score were 2.56-fold more likely to have DR compared with lowest tertile participants. Our study’s result showed the same direction of effect as in the prior study, with type 2 diabetes risk-raising alleles increasing DR risk. The prior study did not examine glycemic traits. Our inability to detect a correlation for glycemic traits may be due to the small amount of glycemic variance captured by these variants. In European patients, HbA_1c SNPs explain ∼5% of HbA_1c variance (50).

We were unable to replicate findings from previous studies (6–8,12,13,17,18). We did have the same direction of effect in our European discovery sample for rs1399634 (LRP2), which was initially reported in an Asian population. However, the meta-analysis was shy of genome-wide significance. The overall lack of replication of previous reports’ findings is not surprising, given the heterogeneity in phenotyping, case-control definitions, ethnicities, and analytic approaches, although we did try to match our case-control definitions to the original studies’ definitions.

There are many potential reasons why we were unable to identify replicable, significant associations from our discovery GWAS. First, the genetic risk in DR development may be quite small in proportion to the nongenetic risk factors. Therefore, even though we assembled the largest sample, it may not be sufficient to detect very modest effects. There was heterogeneity between the discovery and replication cohorts that could contribute to inability to replicate. The discovery cohort included individuals with type 2 diabetes, whereas the replication cohorts included individuals with either type 1 or type 2 diabetes. It is not known definitively whether genetic variants for DR differ between type 1 and type 2 diabetes. Clinically, DR phenotypes are similar in patients with type 1 and type 2 diabetes, so we hypothesize that at least some of the genetic risk is shared. However, we cannot be certain of this, and heterogeneity of diabetes type might have contributed to lack of replication. The discovery cohort included individuals who were of either European or African American descent, whereas the replication cohorts included individuals of European, Hispanic, or Asian descent. This heterogeneity could also have led to lack of replication. Europeans were represented in both the discovery and replication phases, but even our European discovery analysis has limited power. Power calculations show that our discovery GWAS for the any DR analysis in Europeans had 100% power to detect a variant with an MAF of 0.40 with a heterozygous genotypic relative risk of 1.5 with a P value <5 × 10⁻⁸, whereas the power decreases to 5% for the same variant with genotypic relative risk of 1.2.

We attempted to harmonize the phenotypes as much as possible, but there were some limits to complete harmonization, particularly for cohorts with limited-field or no photography. Misclassification of participants because of limited DR ascertainment could have biased the results to the null. Although we did use LTSCORE modeling to account for DoD, we may have had some misclassification bias because we did not have a minimum DoD for control subjects (i.e., some control subjects could have developed DR with longer DoD), which would also bias our result toward the null. We only had one HbA_1c measure. Repeated HbA_1c measures would reflect long-term glycemia more accurately.

In summary, we have executed the largest GWAS of DR to date. There were no genome-wide significant findings, but analysis of protein–protein interaction networks point to possible candidate pathways for PDR in African Americans. Future studies examining DR genetics would benefit from a greater international collaboration encompassing larger samples that would allow strict case-control definitions that define a minimal DoD without sacrificing power. Furthermore, these studies should focus case definitions on the advanced forms of DR—PDR and diabetic macular edema—and incorporate more refined phenotyping, particularly optical coherence tomography for diabetic macular edema. Finally, whole-genome sequencing might reveal a role for very rare variants, particularly for the DR phenotypic extremes.

Acknowledgments. The authors thank the staff at the Icelandic Heart Association and the Age, Gene/Environment Susceptibility - Reykjavik Study (AGES) participants who volunteered time and allowed the authors to contribute data to this international project; the staff and participants of the Atherosclerosis Risk in Communities (ARIC) study for important contributions; and all of the patients recruited in the Genetics of Diabetes and Audit Research Tayside Study (GoDARTS) and other European and African American cohorts. The authors especially thank the Health Informatics Centre in the School of Medicine, University of Dundee, for help with data access; the staff and participants of the Jackson Heart Study (JHS); and the other investigators, staff, and participants of the Multiethnic Study of Atherosclerosis (MESA) study for valuable contributions (a full list of participating MESA investigators and institutions can be found at https://www.mesa-nhlbi.org/). A full list of principal Cardiovascular Health Study (CHS) investigators and institutions can be found at CHS-NHLBI.org.

Funding. This study has received support from the following organizations for this research: Research to Prevent Blindness, Inc., National Eye Institute (EY-16335, EY-22302, EY-11753, and R01-EY-023644 and core grant EY-001792), National Institutes of Health (NIH) (R01-HG-006399), Massachusetts Lions Eye Research Fund, Alcon Research Institute, American Diabetes Association (1-11-CT-51), and Harvard Catalyst. AGES was supported by the NIH through the Intramural Research Program of the National Institute on Aging (ZIAAG007380) and the National Eye Institute (ZIAEY00401), NIH contract number N01-AG-1-2100, Hjartavernd (the Icelandic Heart Association), the Althingi (Icelandic Parliament), and the University of Iceland Research Fund.

The funders had no role in collection, management, analysis, or interpretation of data and were not involved in the preparation, writing, or approval of the article or the decision to submit the article for publication.

ARIC is carried out as a collaborative study supported by National Heart, Lung, and Blood Institute (NHLBI) contracts HHSN268201100005C, HHSN268201100006C, HHSN268201100007C, HHSN268201100008C, HHSN268201100009C, HHSN268201100010C, HHSN268201100011C, and HHSN268201100012C and R01-HL-087641, R01-HL-59367, and R01-HL-086694; National Human Genome Research Institute contract U01-HG-004402; and NIH contract HHSN268200625226C. ARIC has been funded in whole or in part with federal funds from the National Heart, Lung, and Blood Institute, NIH, Department of Health and Human Services (contract numbers HHSN268201700001I, HHSN268201700002I, HHSN268201700003I, HHSN268201700004I, and HHSN268201700005I). Infrastructure was partly supported by the NIH and NIH Roadmap for Medical Research (grant UL1-RR-025005). Funding support for “Building on GWAS for NHLBI-diseases: the U.S. CHARGE consortium” was provided by the NIH through the American Recovery and Reinvestment Act of 2009 (5RC2HL102419). The Australian Genetics of Diabetic Retinopathy Study was supported by the Australian National Health and Medical Research Council (NHMRC) (595918), Canberra, Australia, and the Ophthalmic Research Institute of Australia. J.E.C. is supported by a Practitioner Fellowship from the NHMRC. K.P.B. is supported by a Senior Research Fellowship from the NHMRC. The Blue Mountains Eye Study (BMES) was supported by the Australian NHMRC (project grant identification numbers 974159, 211069, and 302068) and Centre for Clinical Research Excellence in Translational Clinical Research in Eye Diseases (grant 529923). The BMES GWAS and genotyping costs were supported by the Australian NHMRC (project grant identification numbers 512423, 475604, and 529912) and the Wellcome Trust (U.K.) as part of the Wellcome Trust Case Control Consortium 2 (WTCCC2) (085475/B/08/Z and 085475/08/Z). The Cardiovascular Health Study was supported by NHLBI contracts HHSN268201200036C, HHSN268200800007C, HHSN268201800001C, N01-HC-55222, N01-HC-85079, N01-HC-85080, N01-HC-85081, N01-HC-85082, N01-HC-85083, N01-HC-85086, and N01-HC-75150 and grants U01-HL-080295 and U01-HL-130114, with additional contribution from the National Institute of Neurological Disorders and Stroke. Additional support was provided by National Institute on Aging grant R01-AG-023629.

The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH. Subjects included in the present analysis consented to the use of their genetic information.

This work was supported by Tenovus Scotland (2015 T15/40 to W.M.). GoDARTS was supported by Chief Scientist Office Scotland and Diabetes UK. The genotyping costs were granted by the Wellcome Trust for WTCCC2 samples and by the Innovative Medicines Initiative for SUMMIT (SUrrogate markers for Micro- and Macrovascular hard endpoints for Innovative diabetes Tools) samples. The SUMMIT consortium was supported by the European Union’s Seventh Framework Programme (FP7/2007-2013) for the Innovative Medicine Initiative under grant agreement IMI/115006 (to the SUMMIT consortium). FinnDiane was supported by grants from the Folkhälsan Research Foundation, the Wilhelm and Else Stockmann Foundation, the Liv och Hälsa Foundation, Helsinki University Central Hospital Research Funds (EVO), the Novo Nordisk Foundation (NNF14SA0003), and the Academy of Finland (134379, 275614, and 299200). The JHS is supported by the NHLBI and the National Institute on Minority Health and Health Disparities and is conducted in collaboration with Jackson State University (contracts HHSN268201300049C and HHSN268201300050C), Tougaloo College (HHSN268201300048C), and the University of Mississippi Medical Center (HHSN268201300046C and HHSN268201300047C).

The views expressed in this manuscript are those of the authors and do not necessarily represent the views of the NHLBI, the NIH, or the U.S. Department of Health and Human Services.

MESA and the MESA SHARe (SNP Health Association Resource) project are conducted and supported by the NHLBI in collaboration with MESA investigators. Support for MESA is provided by NHLBI contracts HHSN268201500003I, N01-HC-95159, N01-HC-95160, N01-HC-95161, N01-HC-95162, N01-HC-95163, N01-HC-95164, N01-HC-95165, N01-HC-95166, N01-HC-95167, N01-HC-95168, N01-HC-95169, UL1-TR-000040, UL1-TR-001079, UL1-TR-001881, and DK-063491. Additional funding is provided by the Intramural Research Program of the National Eye Institute (ZIAEY000403). Funding for SHARe genotyping was provided by NHLBI contract N02-HL-64278. Genotyping was performed at Affymetrix (Santa Clara, CA) and the Broad Institute of MIT and Harvard (Boston, MA) using the Affymetrix Genome-Wide Human SNP Array 6.0. The Singapore Epidemiology of Eye Diseases study was supported by the National Medical Research Council (NMRC), Singapore (grants 0796/2003, 1176/2008, 1149/2008, STaR/0003/2008, 1249/2010, CG/SERI/2010, CIRG/1371/2013, and CIRG/1417/2015), and Biomedical Research Council, Singapore (08/1/35/19/550 and 09/1/35/19/616). C.-Y.C. is supported by an award from the NMRC (CSA-SI/0012/2017).

NMRC had no role in the design or conduct of this research.

The Starr County Health Studies were supported, in part, by the State of Texas and NIH grants EY-012386, DK-047487, and DK-073541. The Korean Study of Diabetic Retinopathy was supported by National Research Foundation of Korea grants funded by the Korea government (NRF-2017R1A2B2011436 and NRF-2012R1A1A2008943). For the Wake Forest School of Medicine Study (WFU), genotyping services were provided by the Center for Inherited Disease Research. The Center for Inherited Disease Research is fully funded through a federal contract from the NIH to Johns Hopkins University, contract HHSC268200782096C. This work was supported by NIH grants R01-DK-087914, R01-DK-066358, R01-DK-053591, DK-070941, and DK-084149, Wake Forest School of Medicine grant M01-RR-07122, and Venture Fund. Work for this manuscript was supported in part by the Genetics of Latinos Diabetic Retinopathy (GOLDR) Study grant EY14684. This study was supported by the National Eye Institute of the NIH (EY-014684 to Y.-D.I.C. and J.I.R.) and ARRA (American Recovery and Reinvestment Act of 2009) Supplement (EY014684-03S1 and EY014684-04S1), National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) DK-063491 to the Southern California Diabetes Endocrinology Research Center, the Eye Birth Defects Foundation Inc., the National Science Council, Taiwan (NSC 98-2314-B-075A-002-MY3 to W.H.-H.S.), and the Taichung Veterans General Hospital, Taichung, Taiwan (TCVGH-1003001C to W.H.-H.S.). The provision of genotyping data was supported in part by the National Center for Advancing Translational Sciences, CTSI grant UL1-TR-001881 and NIDDK Diabetes Research Center grant DK-063491 to the Southern California Diabetes Research Center. The Family Study of Nephropathy and Diabetes (FIND) study was supported by grants U01-DK-57292, U01-DK-57329, U01-DK-057300, U01-DK-057298, U01-DK-057249, U01-DK-57295, U01-DK-070657, U01-DK-057303, and U01-DK-57304 and, in part, by the Intramural Research Program of the NIDDK. Support was also received from NHLBI grants U01-HL-065520, U01-HL-041654, and U01-HL-041652. This project has been funded in whole or in part with federal funds from the National Cancer Institute, NIH, under contract N01-CO-12400 and the Intramural Research Program of the NIH, National Cancer Institute, Center for Cancer Research. This work was also supported by the National Center for Research Resources for the General Clinical Research Center grants: Case Western Reserve University, M01-RR-000080; Wake Forest University, M01-RR-07122; Los Angeles Medical Center, Harbor-University of California, M01-RR-00425; College of Medicine, University of California, Irvine, M01-RR-00827–29; Health Sciences Center, University of New Mexico, M01-RR-00997; and the Frederic C. Bartter Award, M01-RR-01346. Computing resources were provided, in part, by the Wake Forest School of Medicine Center for Public Health Genomics.

The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

M.I.M. is a Wellcome Senior Investigator supported by Wellcome awards 090532, 106130, 098381, and 203141 and a National Institute for Health Research Senior Investigator.

The views expressed in this article are those of the authors and not necessarily those of the National Health Service, National Institute for Health Research, or the Department of Health.

Duality of Interest. J.Z.K. is employed by Sun Pharmaceutical Industries, Inc. P.-H.G. has received investigator-initiated research grants from Eli Lilly and Company and Roche, is an advisory board member for AbbVie, AstraZeneca, Boehringer Ingelheim, Cebix, Eli Lilly and Company, Janssen, Medscape, Merck Sharp & Dohme, Novartis, Novo Nordisk, and Sanofi, and has received lecture fees from AstraZeneca, Boehringer Ingelheim, Eli Lilly, ELO Water, Genzyme, Merck Sharp & Dohme, Medscape, Novo Nordisk, and Sanofi. B.L.Y. is a full-time employee of Genentech, Inc., and holds stock and stock options in Roche. No other potential conflicts of interest relevant to this article were reported.

Sun Pharmaceutical Industries, Inc., is not in any way involved in this study.

Author Contributions. S.P., E.J.R., A.V.Se., S.D., L.K.S., A.L.P., and L.S. contributed to the writing of the manuscript. R.P.I., R.A.J., M.C., X.L., C.-Y.C., M.C.Y.N., A.V.Sm., G.S.T., Y.-D.I.C., J.Z.K., L.M.D., W.M., S.M.H., M.I., D.N., J.K., Y.H., Y.J., J.A., A.L., K.S., K.H.P., X.G., E.I., K.D.T., S.G.A., J.R.S., B.I.F., I.-T.L., W.H.-H.S., M.K., A.T., S.H., M.M., D.-A.T., R.M.-C., R.V., M.I.M., L.G., E.A., V.L., E.A., A.M., A.S.F.D., H.M.C., I.T., N.S., P.-H.G., S.M., C.L.H., A.P., C.J.C., H.H., P.M., J.E.C., E.Y.C., A.D.P., M.A.G., C.P., D.W.B., B.L.Y., D.S., M.F.C., J.J.W., K.P.B., T.Y.W., B.E.K.K., R.K., J.I.R., and S.K.I. reviewed and edited the manuscript. S.P., R.P.I., R.A.J., C.-Y.C., M.C.Y.N., A.V.Sm., G.S.T., Y.-D.I.C., J.Z.K., W.M., S.M.H., M.I., J.K., J.A., A.L., K.S., K.H.P., X.G., B.I.F., S.H., M.M., D.-A.T., R.M.-C., I.T., N.S., P.-H.G., S.M., C.L.H., A.P., C.J.C., H.H., P.M., J.E.C., E.Y.C., A.D.P., M.A.G., C.P., D.W.B., B.L.Y., D.S., M.F.C., J.J.W., K.P.B., T.Y.W., B.E.K.K., R.K., J.I.R., S.K.I., A.L.P., and L.S. collected and researched data. S.P., R.P.I., R.A.J., M.C., X.L., C.-Y.C., M.C.Y.N., A.V.Sm., E.J.R., A.V.Se., S.D., G.S.T., Y.-D.I.C, J.Z.K., L.M.D., L.K.S., W.M., S.M.H., M.I., D.N., J.K., Y.H., Y.J., J.A., A.L., K.S., S.H., M.M., I.T., N.S., B.L.Y., K.P.B., A.L.P., and L.S. performed the analysis. L.S. is the guarantor of this work and, as such, had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Parts of this study were presented in abstract form at the 42nd Annual Meeting of the Macula Society, Bonita Springs, FL, 13–16 February 2019.