Genome-wide association study of advanced age-related macular degeneration identifies a role of the hepatic lipase gene (LIPC)

Case and Control Sample Development.

Our initial study consisted of 1,057 unrelated cases with geographic atrophy or neovascular AMD and 558 unrelated controls without AMD who were phenotyped based on clinical examination and ocular photography (6, 9, 11). The AMD grade in the worse eye was used in the analyses (17). All individuals were of European ancestry (Methods and Table S1). To enhance the power of our study (18), we included additional unrelated control samples that were genotyped on the same platform in the same laboratory, and we conducted stringent quality control to ensure the technical and population compatibility of these datasets (Methods and Table S2). The final discovery sample consisted of 979 unrelated cases and 1,709 controls. Because the additional controls were unscreened for AMD status, we evaluated their impact on established associations with AMD. We examined 159 SNPs that were in linkage disequilibrium (LD) with the most positively associated variants, and of these SNPs, 137 showed an improvement in the χ² with the addition of these controls. After quality control, the average ratio of the final study χ² to the initial study χ² was 1.82—nearly identical to the expected improvement in χ² based on theoretic power calculations of 1.84.

Genome-Wide Association Discovery Phase.

Association was assessed using logistic regression as implemented in PLINK (19). We plotted our genome-wide association findings from the final dataset in quantile–quantile (QQ) plots. The top end of the QQ plot (Fig. S1A) is dominated by the strong associations of previously reported SNPs, which were excluded to more fully examine the remaining results (Fig. S1B). The regions of previously reported association are near CFH (rs572515 proxy for rs1061170 with r² = 0.654 and P < 10⁻⁵⁵; rs10737680 proxy for CFH rs1410996 with r² = 1 and P < 10⁻⁴⁷), CFI (rs7690921 proxy for rs10033900 with r² = 0.403 and P < 10⁻⁴), CFB/C2 (rs522162 proxy for rs641153 with r² = 1 and P < 10⁻⁶), and the ARMS2/HTRA1 locus (rs10490924 with P < 10⁻⁵⁹) (Table S3). There were no adequate proxies for the C3 locus. Although no other SNPs reached genome-wide significance of P < 5 × 10⁻⁸ (20), we did identify several SNPs of interest in new regions with P values between 10⁻⁴ and 10⁻⁶ in our discovery scan (Table 1), including rs493258 (LIPC) with P = 4.5 × 10⁻⁵ as discussed in more detail below.

Replication Phases.

For all SNPs with P < 10⁻³ in our GWAS, we received results for those SNPs from the Michigan/Penn/Mayo (MPM) scan (21). For all SNPs with P < 10⁻³ in the MPM scan, we sent the results for those SNPs from our GWAS to the MPM group. We excluded overlapping samples and we combined the GWAS results for advanced cases as equally weighted Z scores. We then genotyped SNPs with P < 10⁻⁴ or better in our independent local replication sample of advanced cases and controls from Tufts Medical Center, Tufts University School of Medicine, and Massachusetts General Hospital (Tufts/MGH). These samples were obtained from individuals unrelated to those in our GWAS. We also provided replication data for the MPM study using this cohort, in addition to sharing data from our discovery GWAS sample. For replication in independent samples, subsets of promising SNPs were tested in samples from Johns Hopkins University (JHU), Columbia University (COL), University of Utah (UT), Hopital Intercommunal de Creteil (FR-CRET), and Washington University (WASH-U). In total, there were 5,789 cases and 4,234 controls in the replication cohorts (Table S1). The most significantly associated SNPs from our overall analyses are presented in Tables 1 and 2. We did not find evidence for other reported candidate gene associations with AMD (22), including TLR3, SERPING1, LRP6, PEDF, VEGF, TLR4, CX3CR1, ELOVL4, PON1, and SOD2 (Table S4).

Results of Combined Scan and Replication Analysis.

With additional replication, the SNP that we identified on chromosome 15, rs493258, showed significant association with P = 1.61 × 10⁻⁸ (independent of the MPM GWAS data P = 4.92 × 10⁻⁷). This SNP is located 35 kb upstream of the LIPC gene on chromosome 15q22 and is in LD (r² = 0.33; D′ = 0.847) with the previously described functional variant (rs10468017) reported to influence LIPC expression and serum HDL levels (23, 24).

To evaluate whether or not this association represented an effect of that same functional variant, we genotyped rs10468017 in our replication samples (Table 2). The resultant P value was 1.34 × 10⁻⁸; the average minor-allele frequency (T) among cases was 25.8%, and the average frequency for this allele among controls was 30.0%. Logistic regression analyses conditional on allelic dosage of each SNP indicated that rs493258 is not associated with AMD conditional on rs10468017, but rs10468017 is still associated with AMD (P < 0.001) conditional on association at rs493258. Therefore, we concluded that the observed association with AMD is likely caused by this functional variant that has been identified in genetic studies of HDL levels (25). The estimated odds ratio (OR) for rs10468017 from the combined discovery plus replication data was 0.82 (95% confidence interval = 0.77–0.88), showing that the minor allele T, previously associated with reduced LIPC expression and higher HDL, is also associated with a reduced risk of AMD. This SNP, like those in the complement pathway loci, is observed to have a consistent OR for geographic atrophy (advanced dry OR = 0.73; Tufts/MGH only) and neovascular AMD (wet OR = 0.77; Tufts/MGH only) when each disease subtype is compared with controls. We found no significant gene–gene interactions between this SNP and seven previously established genetic loci when evaluating their effect on risk of advanced AMD. For rs10468017, there was modest evidence for heterogeneity under the Breslow–Day test for our replication samples (P = 0.022); the UT sample (Breslow–Day P = 2.7 × 10⁻³) lacked association, and the FR-CRET sample (Breslow–Day P = 0.040) had higher than expected association (Table S5). If these two samples were excluded, the association was stronger (P = 6.45 × 10⁻⁹). These samples were not discounted based on modest heterogeneity, particularly given the fact that both samples showed association to complement loci; such heterogeneity could be attributed to yet to be discovered genotype or phenotype correlations and ascertainment differences between these samples.

We then evaluated the top SNPs for advanced wet and dry AMD phenotypes separately and found a significant difference between the two advanced types only for the ARMS2/HTRA1 locus (OR 1.38, P = 0.003). This locus on chromosome 10 is also more strongly related to progression to neovascular AMD (26). When comparing these subtypes separately with controls across the genome, we did not observe any associations with P < 10⁻⁶, and we did not observe excess of SNPs with P < 10⁻⁴ beyond the number expected by chance. Our scan did not have sufficient statistical power to conclusively highlight weak or moderate subtype-specific SNP associations between these two advanced forms of AMD.

Given the association of AMD to a SNP known to influence serum HDL levels, we evaluated other HDL loci (25). Among the best proxy SNPs for these other HDL-associated variants, no consistent trend between HDL-lowering alleles and increased AMD risk was observed (Table 3). When our discovery results were combined with the replication samples, the meta P was 1.41 × 10⁻³ for the cholesterylester transfer protein (CETP) SNP, rs3764261. However, the HDL-increasing allele, A, was in the direction of increased risk for AMD, opposite to the protective effect seen with the functional HDL-increasing LIPC variant. The variant, rs12678919, in the lipoprotein lipase precursor gene (LPL) with meta P = 0.07 as well as rs1883025 in the ATP-binding cassette, subfamily A member 1 (ABCA1) gene with meta P = 9.73 × 10⁻⁴, were also not genome-wide significant, and they showed discordant effects between the HDL-increasing allele and protective or risk effects for AMD.

Other Loci.

In our discovery scan, the SNP rs9621532 in the synapsin III gene, a little more than 100 kb upstream of the tissue inhibitor of the metalloproteinase 3 (TIMP3) locus in the MPM scan (21), had P = 0.045. Combined with our replication cohorts, results were in the direction of a protective effect for the minor C allele (OR = 0.62, P = 3.71 × 10⁻⁹). Of note, the SNPs at the LIPC locus were not on the list of top SNPs from the MPM scan, and the SNP rs9621532 was not on our list of top SNPs. TIMP3 has been found to be a matrix-bound angiogenesis inhibitor, and mutations in the gene have been shown to induce abnormal neovascularization (27). Our study of the TIMP3 locus and AMD in 1997, inspired by the association of this gene to Sorsby fundus dystrophy that has similar phenotypic features to advanced AMD, did not find evidence of association or linkage between AMD and TIMP3 among 38 AMD families. However, as we stated, “the possibility that a subset of cases could be caused by the TIMP3 locus could not be ruled out” by this relatively small study (28).

Several regions in our genome-wide scan other than LIPC continued to show association in replication, although not at levels of genome-wide significance (Table 1). SNP rs11755724 on chromosome 6 showed a protective effect of the A allele with P = 9.88 × 10⁻⁷ and OR = 0.87. This SNP is in the intron of RAS responsive element binding protein 1 isoform (RREB1), which encodes a transcription factor that binds to the RAS-responsive elements of gene promoters (29). The SNP rs13095226 on chromosome 3 (P = 2.50 × 10⁻⁶ and OR = 1.24 for the C allele) is an intronic SNP in the COL8A1 gene, which encodes one of the two alpha chains of type VIII collagen, a major component of the multiple basement membranes in the eye, including Bruch's membrane and the choroidal stroma (30). Bruch's membrane is located directly below the retinal pigment epithelium and plays a central role in the pathogenesis of AMD (31).

CNVs.

We used CANARY (32) to evaluate 554 common segregating copy-number polymorphisms (CNPs) in advanced cases and controls from the phenotyped sample, and observed no CNPs with P < 0.001 with the exception of the previously reported deletion variant at CFHR1 that showed strong association (P = 1 × 10⁻¹⁹) (33). This observation is explained by other previously reported CFH variants associated with AMD, particularly the intronic SNP rs1410996 (6), and does not constitute an independent association. We also used Birdseye to examine the potential role of rare CNVs (32). After stringent QC, we observed a set of 150 rare deletions and 278 rare duplications greater than 100 kb. However, these events were not enriched across the genome among cases compared with controls nor were there specific regions showing association of a rare CNV (Fisher's exact test P < 0.01). We also did not find evidence for any rare CNVs disrupting genes at the six confirmed associated loci.

Immunoprecipitation Real-Time PCR and Immunohistochemistry of LIPC.

We examined sequence conservation for LIPC in evolution and determined approximately 75% amino acid identity across mammals (Fig. S2A). To investigate whether the LIPC protein was expressed in the eye and if expression levels change over time, we performed immunoprecipitation using a monoclonal LIPC antibody and quantitative RT-PCR. Despite LIPC being considered a hepatic enzyme, we were able to detect low levels of LIPC protein in bovine retinas (Fig. S2B) and LIPC RNA in the retinas of C57BL/6 mice across a broad range of ages (Fig. S2 C–D). Consistent with our expression data from mouse retina, we were also able to amplify LIPC message from total RNA isolated both from human retinal epithelial transformed cells (RPE19) and human eyes. Next, using a rabbit polyclonal antibody raised against amino acids 91–160 of human LIPC, we were able to show strong labeling of human macula and RPE/choroid and weaker labeling of peripheral retina at the appropriate molecular weight of ∼55 kDa observed in human liver (Fig. 1). Bands of mobility ∼80 kDa likely represented glycosylated hepatic lipase. This same antibody was then used to immunolocalize LIPC in sections of perfusion-fixed monkey retina (Fig. 2). Prominent labeling of all retinal neurons was observed, especially in retinal ganglion cells of the central retina. However, Mueller glial cells were unlabeled, and RPE cells appeared to exhibit significant fluorescence signal beyond intrinsic autofluorescence. Thus, our immunoblot and immunohistochemical data corroborate the findings of Tserentsoodol et al., (34) regarding the existence of intraretinal proteins (e.g., ABCA1 transporter, apoA1, SR-BI, and SR-BII) that mediate lipid transport and processing.

Tufts/MGH Study-Sample Descriptions.

This study conformed to the tenets of the Declaration of Helsinki and received approval from Institutional Review Boards; informed consent was signed by all participants. Tufts participants were unrelated and had geographic atrophy, neovascular disease, or no sign of AMD, based on fundus photography and ocular examination (17). Subjects were derived from ongoing AMD study protocols as described previously (6, 9, 11, 48–51). The Tufts/MGH replication dataset included DNA samples from unrelated Caucasian individuals not included in the GWAS, including 868 advanced AMD cases and 410 examined controls identified from the same Tufts cohorts as well as 379 unexamined controls (52).

Genotyping, Analysis, and Replication.

Genotyping was performed at the Broad Institute using the Affymetrix SNP 6.0 GeneChip and the Sequenom MassARRAY system for iPLEX assays (32). We started with 1,057 cases and 558 examined controls and studied 906,000 genotyped SNPs and 946,000 CNVs using the Affymetrix 6.0 GeneChip, which passed QC filters. Then, 43,562 SNPs were removed for low call rate, 4,708 were removed for failing the Hardy–Weinberg test at 10⁻³, 8,332 SNPs were removed because of failing a differential missing test between cases and controls at 10⁻³, and 126,050 SNPs were removed for having allele frequency less than 1%. We removed 73 individuals for lower than expected call rate, resulting in 1,006 cases and 536 controls. All QC steps were performed using PLINK (19). We conducted a preliminary χ² association analysis to determine the extent to which population stratification and other biases were affecting the samples and observed λ ∼ 1.05, indicating that the samples were generally matched well for population ancestry but with some minor inflation remaining (explanation and visual representation in Fig. S3). We added Myocardial Infarction Genetics Consortium (MIGen) shared controls (53), which were genotyped on the same Affymetrix 6.0 GeneChip product, and conducted population stratification analyses using multidimensional scaling in PLINK (19). These analyses identified 27 cases, 12 AMD controls, and 223 MIGen controls for a total of 262 individuals who were outliers in the principal component analysis. The final genomic control λ for the logistic regression included seven significant (for prediction of phenotype status) principal components as covariates, and was 1.036 for 632,932 SNPs. This dataset was used for our official GWAS analysis.

SNPs with P < 10⁻³ from our GWAS discovery sample (n = 720 SNPs excluding previously associated regions) were exchanged with the MPM study group, which enabled us to use the two scans as primary replication efforts, and enhanced the power of each study. We performed genotyping of all SNPs with combined P < 10⁻⁴ using Sequenom iPLEX with our Tufts/MGH replication sample. Focusing on sites that continued to show association with P < 10⁻⁴ after this local replication, we performed a third stage of replication in which our collaborators at JHU and COL genotyped SNPs with ABI 7900 Taqman genotyping, samples from WASH-U were genotyped using the Sequenom platform, and samples from UT and FR-CRET were sent for genotyping to the Broad Institute (Table S1).

CNVs.

Using the Birdsuite (32) family of algorithms, we called rare CNV and common CNV in 459 controls and 838 cases after removing those individuals failing SNP clustering that had an excessive number of singleton CNVs or an excessive length of total length of singleton CNVs. The study was well-powered to detect CNV events (Fig. S4). For rare variants, we identified all singleton CNVs that were larger than 100 kb, had logarithm of the odds scores greater than 10, and relied on at least 10 probes. We also looked at a smaller set if rare variants (>20 kb) appeared in <1% of individuals, and we used PLINK's permutation function to assess if there were any regions that had significant differences between cases versus controls. For common variants, we examined 554 variants that had at least one allele <99%. We used logistic regression in PLINK to test if copy number significantly predicted case-control status.

See SI Methods for methods related to PCR and imuunohistochemistry.