Genome-wide analysis of disease progression in age-related macular degeneration

Abstract

Family- and population-based genetic studies have successfully identified multiple disease-susceptibility loci for Age-related macular degeneration (AMD), one of the first batch and most successful examples of genome-wide association study. However, most genetic studies to date have focused on case–control studies of late AMD (choroidal neovascularization or geographic atrophy). The genetic influences on disease progression are largely unexplored. We assembled unique resources to perform a genome-wide bivariate time-to-event analysis to test for association of time-to-late-AMD with ∼9 million variants on 2721 Caucasians from a large multi-center randomized clinical trial, the Age-Related Eye Disease Study. To our knowledge, this is the first genome-wide association study of disease progression (bivariate survival outcome) in AMD genetic studies, thus providing novel insights to AMD genetics. We used a robust Cox proportional hazards model to appropriately account for between-eye correlation when analyzing the progression time in the two eyes of each participant. We identified four previously reported susceptibility loci showing genome-wide significant association with AMD progression: ARMS2-HTRA1 (P = 8.1 × 10⁻⁴³), CFH (P = 3.5 × 10⁻³⁷), C2-CFB-SKIV2L (P = 8.1 × 10⁻¹⁰) and C3 (P = 1.2 × 10⁻⁹). Furthermore, we detected association of rs58978565 near TNR (P = 2.3 × 10⁻⁸), rs28368872 near ATF7IP2 (P = 2.9 × 10⁻⁸) and rs142450006 near MMP9 (P = 0.0006) with progression to choroidal neovascularization but not geographic atrophy. Secondary analysis limited to 34 reported risk variants revealed that LIPC and CTRB2-CTRB1 were also associated with AMD progression (P < 0.0015). Our genome-wide analysis thus expands the genetics in both development and progression of AMD and should assist in early identification of high risk individuals.

Introduction

Age-related macular degeneration (AMD) is a heritable neurodegenerative disease and a leading cause of blindness in the elderly population in the United States. The disease is progressive and irreversible in affecting central vision. The progress starts with appearance of drusen and advances to late AMD, which has two major forms: wet AMD [choroidal neovascularization (CNV)] and dry AMD [geographic atrophy (GA)] (1).

Patients have different AMD progression rates. Some with early AMD maintain good vision for a long time without advancing to late AMD, while others quickly develop vision-threatening late AMD. Thus, it is critical to identify AMD progression-associated genetic variants to better understand the pathobiology of the disease. However, the large-scale genome-wide studies conducted so far have mainly focused on dichotomous affected/unaffected phenotypes but ignored progression phenotypes, which can be more crucial for understanding the pathways for AMD and for identifying potential therapies. For example, over the past few years, case–control genome-wide association studies (GWASs) have successfully detected multiple single nucleotide polymorphisms (SNPs) associated with AMD susceptibility (2–4), in addition to the loci in two well-replicated genes, CFH in chromosome 1 (5–7) and ARMS2 in chromosome 10 (8,9). A few of the reported AMD risk variants in CFH and ARMS2 appear to have significant effect on AMD progression (3,10–14). C3, COL8A1, CFB and RAD51B have also been reported to exhibit association with AMD progression in both uni- and multi-variable studies, whereas C2 and C9 are associated with AMD progression based on uni-variable studies (15). However, these studies investigated only a small number of variants. To our knowledge, no GWAS has been performed so far for discovering risk variants for AMD progression.

Using data from a National Eye Institute-funded large multi-center randomized clinical trial of oral supplementation with micronutrients—the Age-related Eye Disease Study (AREDS) (1), which was also designed to evaluate the risk factors for the development and progression of AMD (1), we conducted a genome-wide scan to test for association between time-to-late AMD (either CNV or GA) and genetic variants. Previous studies considered individuals with both eyes free of late AMD at a baseline time and typically defined progression at the level of the subject when assessing genetic effects (10,16). In such studies, the subject was considered as having “progressed” as soon as one eye progressed to late AMD. As a consequence, the progression time of the second eye to progress was ignored. Instead of defining progression time at the subject level, we defined AMD progression time for each eye and therefore made more complete and efficient use of the available data. To account for the association in the progression times in the two eyes within a subject, robust variance estimates were used for the regression coefficients in the Cox proportional hazards model, treating each subject as a cluster.

Results

Study data characteristics

The demographic and clinical characteristics of the 2721 Caucasian participants from AREDS are described in Table 1. About 26% participants were free of AMD at enrollment and the rest entered the study with certain severity level of the disease. The participants were aged between 55- and 81-years old at enrollment (mean ± SD = 68.7 ± 4.9). Women comprised 56% (N = 1527) of the cohort. Most participants received education higher than high school (67%). About 47% of participants were never smokers (N = 1272) and another 47% were former smokers (N = 1288). The mean follow-up time was 10.3 (SD = 1.7) years. The 5017 eyes of the 2721 participants were free of late AMD at baseline and the majority of eyes had low baseline severity score (N = 3125, 62% for baseline severity score between 1 and 3). The rest 425 eyes were already at late AMD at baseline and were excluded from the progression analysis. More participants were randomized to placebo (N = 842, 31%) or antioxidants alone (N = 850, 31%) than zinc alone (N = 507, 19%) or antioxidant plus zinc (N = 522, 19%).

Table 1.

Baseline characteristics of the AREDS cohort

Subject-level variables	N = 2721 subjects
Enrollment age, year
Mean ± SD	68.7 ± 4.9
Range	[55.3, 81.2]
Female (N, %)	1527 (56)
Follow-up time, (mean ± SD)	10.3 ± 1.7
Education (N, %)
≤High school	906 (33)
>High school	1814 (67)
Missing	1 (0)
Smoking (N, %)
Never smoked	1272 (47)
Former smoker	1288 (47)
Current smoker	161 (6)
Treatment (N, %)
Placebo	842 (31)
Antioxidants alone	850 (31)
Zinc	507 (19)
Antioxidants+zinc	522 (19)
AMD categories (N, %)
1	694 (26)
2	631 (23)
3	930 (34)
4	466 (17)
Eye-level variables	N = 5017 eyes
Baseline AMD severity score at eye-level
Mean ± SD	3.0 ± 2.3
1–3 (N, %)	3125 (62)
4–6 (N, %)	1293 (26)
7–8 (N, %)	599 (12)

Baseline AMD severity and progression

It has been shown in Ding et al. (14) that the AMD progression rate increases as the eye’s baseline AMD severity increases. Ding et al. (14) categorized the eyes that were free of late AMD at baseline into three groups based on their baseline severity score: 1–3 (low), 4–6 (middle) and 7–8 (high). In these 5017 eyes from AREDS, only 3% eyes with baseline severity 1–3 progressed to late AMD by the end of their follow-up, while 40% eyes with baseline severity 4–6 progressed and 81% eyes with baseline severity 7–8 progressed. In addition, within each of the three groups (1–3, 4–6 and 7–8), the progression rate increases as their fellow eyes’ baseline AMD severity increases. Moreover, it has been found that the genetic effect on AMD progression can decrease drastically after adjusting for baseline severity (14), which is not surprising given the progression is determined based on this severity score. Thus, we considered two statistical models in our GWAS: (i) with baseline severity score included for controlling its effect on AMD progression, and (ii) without adjusting for baseline severity score to avoid masking certain moderate genetic effects.

GWAS of AMD progression

In the modeling of AMD progression, based on the univariable Cox models, baseline age, smoking status, education level and baseline severity score (treated as a continuous variable) were selected as covariates. In addition, the first two principal components were also included as covariates to account for any effect of population stratification.

Inspection of genomic control values (λ_BL=1.02; λ_noBL=1.01 where BL stands for including baseline severity in the model and noBL is otherwise) suggested that the false positive rate is well-controlled (Supplementary Material, Fig. S1). As shown in Figure 1, rs10922109 in the CFH region (P_BL=1.33 × 10⁻⁸; P_noBL=3.46 × 10⁻³⁷) and rs2284665 in the ARMS2-HTRA1 region (P_BL=6.28 × 10⁻¹³; P_noBL=8.08 × 10⁻⁴³) reach genome-wide significance (P < 5 × 10⁻⁸) in both GWASs with and without controlling for baseline severity (Table 2; Supplementary Material, Figs S2–S5). The Kaplan-Meier (KM) plots depict clear separations between different genotypes for these two SNPs either in the overall or stratified (by baseline severity) plots (Fig. 2A and B). In addition, as shown in Figure 1B, rs116503776 in the C2-CFB-SKIV2L region (P_noBL=8.07 × 10⁻¹⁰) and rs2230199 in the C3 region (P_noBL=1.20 × 10⁻⁹) reach genome-wide significance without controlling for baseline severity (Table 2; Supplementary Material, Figs S6 and S7), and these two SNPs show separable curves in the overall KM plots (Fig. 2C and D). However, in the baseline severity stratified plots, rs116503776 (C2-CFB-SKIV2L) (Fig. 2C) does not show separable KM curves with low baseline severity while rs2230199 (C3) (Fig. 2D) does not show separable KM curves with high baseline severity. Minor alleles at rs2284665 (ARMS2-HTRA1; HR_BL = 1.48; HR_noBL = 2.06) and rs2230199 (C3; HR_BL = 1.16; HR_noBL = 1.45) were associated with faster progression. On the contrary, minor alleles at rs10922109 (CFH; HR_BL = 0.70; HR_noBL = 0.43) and rs116503776 (C2-CFB-SKIV2L; HR_BL = 0.74; HR_noBL = 0.56) were associated with slower progression. These four loci (CFH, ARMS2-HTRA1, C2-CFB-SKIV2L and C3) were reported to be associated with AMD risk in the consortium case–control studies (2,3). The hazard ratios for these four loci from our progression analysis were all consistent with the odds ratios from Fritsch et al. (3) (Table 2). Moreover, in these four loci, the proportion of minor allele copies was significantly different between progressors and non-progressors (Supplementary Material, Table S1).

Figure 1.

Manhattan plots of GWAS results of AMD progression. (A) GWAS results with adjustment for baseline severity score, and (B) GWAS results without adjustment for baseline severity score. Summary of genome-wide association results using Cox model with robust variance for paired eyes. The red horizontal line is the conservative significance level (P = 5 × 10⁻⁸) and the blue horizontal line is the suggestive significance level (P = 1 × 10⁻⁵).

Table 2.

List of loci associated with AMD progression identified in AREDS

SNP	Chr	Position	Major/ minor allele	MAF	Gene	With BL severity		Without BL severity		Fritsche et al. case–control results
						HR	P-value	HR	P-value	P-value
Significant loci reported also in consortium case–control studies
rs2284665	10	124 226 630	G/T	0.30	ARMS2-HTRA1	1.48	6.3 × 10−13	2.06	8.1 × 10−43	4.0 × 10−697
rs10922109	1	196 704 632	C/A	0.33	CFH	0.70	1.3 × 10−8	0.43	3.5 × 10−37	9.6 × 10−618
rs116503776	6	31 930 462	G/A	0.12	C2-CFB-SKIV2L	0.74	0.001	0.56	8.1 × 10−10	1.2 × 10−103
rs2230199	19	6 718 387	C/G	0.24	C3	1.16	0.014	1.45	1.2 × 10−9	3.8 × 10−69
Marginally significant novel loci
rs79069165	5	75 184 914	T/C	0.08	POC5/SV2C	1.52	7.4 × 10−8	1.19	0.062	0.724
rs74320127	15	64 075 102	T/G	0.05	HERC1	1.71	5.3 × 10−8	1.15	0.279	0.232
rs56072732	2	237 519 496	C/T	0.06	ACKR3	1.34	0.002	1.71	6.4 × 10−8	0.497

MAF, minor allele frequency; BL, baseline severity; HR, hazard ratio relative to the minor allele (minor allele/major allele). Bolded SNPs are also the lead variants in Fritsche et al. (3)

Figure 2.

KM curves for SNPs from the four known genes identified by case–control studies. (A)ARMS2-HTRA1 (rs2284665), (B)CFH (rs10922109), (C)C2-CFB-SKIV2L (rs116503776) and (D)C3 (rs2230199). For all the KM plots (Figures 2, 3,5–7), the 1st column shows the KM curves of total samples. The second to fourth columns show the KM curves of samples with baseline severity 1–3, 4–6 and 7–8, respectively. The bottom left legend shows the number of samples for each genotype category and the percentage of how many these samples progressed to late AMD. The genotypes were derived from dosages: common homozygotes when dosage ≤ 0.5, heterozygotes when 0.5 < dosage < 1.5 and rare homozygotes when dosage ≥ 1.5.

In addition, we found rs79069165 in the POC5/SV2C region (P_BL = 7.41 × 10⁻⁸) and rs74320127 in the HERC1 region (P_BL = 5.29 × 10⁻⁸) marginally reach 5 × 10⁻⁸ level in the GWAS with baseline severity adjustment (Table 2; Supplementary Material, Figs S8 and S9), and rs56072732 in the ACKR3 region (P_noBL = 6.40 × 10⁻⁸) marginally reach 5 × 10⁻⁸ level in the GWAS without baseline severity adjustment (Table 2; Supplementary Material, Fig. S10). These three SNPs are relatively rare (i.e. 0.05 < MAF < 0.10) and have not been identified by previous case–control studies. Figure 3 indicates they have more significant effects for subjects with moderate or high baseline severity than low baseline severity, ignoring the rare homozygous group with sample size < 10 (in all three cases). In HERC1 (rs74320127) and ACKR3 (rs56072732), the proportion of minor allele copies was significantly different between progressors and non-progressors (Supplementary Material, Table S2). All other genes reaching the suggestive significance level (1 × 10⁻⁵) are listed in Supplementary Material, Tables S3 and S4.

Figure 3.

KM curves for SNPs from the three novel genes. (A)POC5/SV2C (rs79069165), (B)HERC1 (rs74320127) and (C)ACKR3 (rs56072732).

In addition to the GWAS of progression to any type of late AMD, we conducted two additional GWAS to investigate the effects of genetic variants on progression to different late AMD subtypes: (i) Progression to CNV (including 508 eyes progressed to CNV and 3912 censored eyes) and (ii) Progression to GA (including 615 eyes progressed to GA and 3912 censored eyes). The eyes that progressed to CNV + GA (N = 18) contributed to both subtype analyses. Excluding these 18 eyes led to very similar results (not shown here). We did not adjust for baseline severity in these two subtype GWAS, to avoid missing moderate genetic effects given the sample size is smaller in this subtype analysis. Two SNPs in the TNR and ATF7IP2 regions were found to be associated with progression to CNV (rs58978565 in the TNR region: P = 2.31 × 10⁻⁸, MAF = 0.36 and HR = 1.51; rs28368872 in the ATF7IP2 region: P = 2.95 × 10⁻⁸, MAF = 0.12 and HR = 1.69; Fig. 4A;Table 3; Supplementary Material, Figs S11 and S12) at the level of 5 × 10⁻⁸, but not associated with progression to GA (rs58978565 in the TNR region: P = 0.98, MAF = 0.35 and HR = 1.00; rs28368872 in the ATF7IP2 region: P = 0.03, MAF = 0.12 and HR = 1.26; Fig. 4B;Table 3). Both loci did not reach the genome wide significance level in the previous main GWAS without baseline severity (rs58978565 in the TNR region: P = 0.0016; rs28368872 in the ATF7IP2 region: P = 8.23 × 10⁻⁶). The variant rs58978565 in the TNR region is not associated with progression to GA even at a conservative significance level of 0.05, which indicates that TNR is specifically associated with progression to CNV. We refit the models by adjusting for baseline severity on TNR (rs58978565) and ATF7IP2 (rs28368872) and obtained similar results (Table 3). The KM plot (Fig. 5A) shows a clear separation for rs58978565 in the TNR region for progression to CNV. Such separation for rs58978565 was not observed for progression to GA (Fig. 5B). For rs28368872 in the ATF7IP2 region, the plots (Fig. 6A) shows clear separations for progression to CNV for eyes with baseline severity >4, whereas the separations are moderate for progression to GA (Fig. 6B). Moreover, in TNR (rs58978565), the number of minor allele copies is correlated with the proportion of CNV progression but not with the GA progression (Supplementary Material, Table S5). In ATF7IP2 (rs28368872), the number of minor allele copies is correlated with the proportion of both CNV and GA progression (Supplementary Material, Table S5).

Figure 4.

Manhattan plots of GWAS results of CNV and GA progression. (A) progression to CNV and (B) progression to GA. The baseline severity was not included as a covariate. Summary of genome-wide association results using Cox model with robust variance for paired eyes. The red horizontal line is the conservative significance level (P = 5 × 10⁻⁸) and the blue horizontal line is the suggestive significance level (P = 1 × 10⁻⁵).

Table 3.

Results for rs58978565 in TNR and rs28368872 in ATF7IP2

SNP	Chr	Position	Major/minor allele	Gene	AMD subtypes	MAF	With BL severity		Without BL severity
							HR	P-value	HR	P-value
rs58978565	1	175 345 602	C/CAGAGT	TNR	GA	0.35	0.99	0.83	1.00	0.98
					CNV	0.36	1.34	1.2 × 10−4	1.51	2.3 × 10−8
rs28368872	16	10 585 350	G/A	ATF7IP2	GA	0.12	1.31	0.02	1.26	0.03
					CNV	0.12	1.75	8.0 × 10−9	1.69	2.9 × 10−8

MAF, minor allele frequency; BL, baseline severity; HR, hazard ratio relative to the minor allele (minor allele/major allele).

Figure 5.

KM curves for rs58978565 in TNR. It was associated with progression to CNV at significance level of 5 × 10⁻⁸, but not with progression to GA.

Figure 6.

KM curves for rs28368872 in ATF7IP2. It was associated with progression to CNV at significance level of 5 × 10⁻⁸, but not with progression to GA.

We queried pathway databases (GO and KEGG) using DAVID (17) for 36 genes (Supplementary Material, Table S6) in the nine narrow AMD locus regions (seven loci identified in the main GWAS in Table 2 and two additional loci identified in the CNV GWAS in Table 3). Novel identified loci are POC5/SV2C, HERC1/DAPK2, ACKR3/COPS8, TNR and ATF7IP2/EMP2. We found biological pathways enriched across AMD progression associated loci, such as the complement pathway, the protein regulation pathway and the immune response pathway (Supplementary Material, Table S7), highlighting novel genes (ACKR3, EMP2 and DAPK2) that connect multiple pathways.

Connection with existing GWAS findings for AMD risk

We examined whether the 34 AMD risk variants reported in the Fritsche et al. (3) case–control study were also associated with late AMD progression (Table 4). We found that rs10922109 (CFH), rs116503776 (C2-CFB-SKIV2L), rs3750846 (ARMS2-HTRA1), rs2043085 (LIPC), rs2230199 (C3) and rs72802342 (CTRB2-CTRB1) were also associated with late AMD progression at P < 0.0015 (Bonferroni correction for 34 tests). The variant rs10922109 is the most significant SNP in the CFH region in both our late AMD progression study and the consortium case–control study. Most of the 34 SNPs (24 from with baseline severity analyses and 25 from without baseline severity analyses) have the same odds ratio and hazard ratio directions.

Table 4.

Results for 34 AMD risk variants reported in the latest case–control study conducted by the International AMD Genomics Consortium

SNP	Chr	Position	Major/ minor allele	MAF	Gene	With BL severity		Without BL severity		Fritsche et al. case–control results
						HR	P-value	HR	P-value	OR	P-value
Known SNPs identified in consortium case–control studies
rs10922109	1	196 704 632	C/A	0.33	CFH	0.70	1.3 × 10−8	0.43	3.5 × 10−37	0.38	9.6 × 10−618
rs62247658	3	64 715 155	T/C	0.45	ADAMTS9-AS2	0.93	0.1798	0.97	0.5813	1.14	1.8 × 10−14
rs140647181	3	99 180 668	T/C	0.02	COL8A1	1.62	0.0549	1.31	0.2420	1.59	1.4 × 10−11
rs10033900	4	110 659 067	C/T	0.50	CFI	0.98	0.7502	0.86	0.0045	1.15	5.4 × 10−17
rs62358361	5	39 327 888	G/T	0.01	C9	1.04	0.8757	1.64	0.0170	1.8	1.3 × 10−14
rs116503776	6	31 930 462	G/A	0.12	C2-CFB-SKIV2L	0.74	0.0011	0.56	8.1 × 10−10	0.57	1.2 × 10−103
rs943080	6	43 826 627	T/C	0.48	VEGFA	1.06	0.1927	0.93	0.2250	0.88	1.1 × 10−14
rs79037040	8	23 082 971	T/G	0.46	TNFRSF10A	1.04	0.3570	1.05	0.3666	0.9	4.5 × 10−11
rs1626340	9	101 923 372	G/A	0.21	TGFBR1	0.92	0.2422	0.81	0.0024	0.88	3.8 × 10−10
rs3750846	10	124 215 565	T/C	0.30	ARMS2-HTRA1	1.44	1.8 × 10−11	2.04	5.3 × 10−42	2.81	6.5 × 10−735
rs9564692	13	31 821 240	C/T	0.27	B3GALTL	0.84	0.0023	0.87	0.0237	0.89	3.3 × 10−10
rs61985136	14	68 769 199	T/C	0.37	RAD51B	0.87	0.0139	0.86	0.0079	0.9	1.6 × 10−10
rs2043085	15	58 680 954	T/C	0.37	LIPC	0.88	0.0114	0.83	0.0012	0.87	4.3 × 10−15
rs5817082	16	56 997 349	C/CA	0.24	CETP	0.90	0.0745	0.89	0.0632	0.84	3.6 × 10−19
rs2230199	19	6 718 387	C/G	0.26	C3	1.16	0.0139	1.45	1.2 × 10−9	1.43	3.8 × 10−69
rs429358	19	45 411 941	T/C	0.11	APOE	0.95	0.5298	0.82	0.0261	0.7	2.4 × 10−42
rs5754227	22	33 105 817	T/C	0.12	SYN3-TIMP3	0.91	0.2080	0.88	0.1203	0.77	1.1 × 10−24
rs8135665	22	38 476 276	C/T	0.20	SLC16A8	0.99	0.8491	1.08	0.2515	1.14	5.5 × 10−11
rs11884770	2	228 086 920	C/T	0.27	COL4A3	0.98	0.6881	0.95	0.4445	0.9	2.9 × 10−8
rs114092250	5	35 494 448	G/A	0.02	PRLR-SPEF2	0.60	0.0122	0.59	0.0280	0.7	2.1 × 10−8
rs7803454	7	99 991 548	C/T	0.20	PILRB-PILRA	1.00	0.9929	1.09	0.1677	1.13	4.8 × 10−9
rs1142	7	104 756 326	C/T	0.36	KMT2E-SRPK2	1.01	0.8829	1.04	0.5327	1.11	1.4 × 10−9
rs71507014	9	73 438 605	GC/G	0.43	TRPM3	1.01	0.8196	1.00	0.9360	1.1	3.0 × 10−8
rs10781182	9	76 617 720	G/T	0.32	MIR6130-RORB	1.15	0.0104	1.14	0.0180	1.11	2.6 × 10−9
rs2740488	9	107 661 742	A/C	0.27	ABCA1	0.99	0.8419	0.95	0.4548	0.9	1.2 × 10−8
rs12357257	10	24 999 593	G/A	0.24	ARHGAP21	1.08	0.2005	1.18	0.0057	1.11	4.4 × 10−8
rs3138141	12	56 115 778	C/A	0.21	RDH5-CD63	1.18	0.0416	1.19	0.0338	1.16	4.3 × 10−9
rs61941274	12	112 132 610	G/A	0.02	ACAD10	1.12	0.5593	0.94	0.7565	1.51	1.1 × 10−9
rs72802342	16	75 234 872	C/A	0.07	CTRB2-CTRB1	0.79	0.0570	0.67	0.0008	0.79	5.0 × 10−12
rs11080055	17	26 649 724	C/A	0.48	TMEM97-VTN	0.89	0.0267	0.97	0.5847	0.91	1.0 × 10−8
rs6565597	17	79 526 821	C/T	0.38	NPLOC4-TSPAN10	1.03	0.5769	1.06	0.2877	1.13	1.5 × 10−11
rs67538026	19	1 031 438	C/T	0.45	CNN2	0.90	0.0475	0.90	0.0684	0.9	2.6 × 10−8
rs142450006	20	44 614 991	TTTTC/T	0.13	MMP9	0.77	0.0021	0.77	0.0029	0.85	2.4 × 10−10
rs201459901	20	56 653 724	T/TA	0.06	C20orf85	1.07	0.4688	0.96	0.7440	0.76	3.1 × 10−16

Bolded SNPs are those SNPs reaching the Bonferroni corrected significance level = 0.0015 (either with BL severity adjusted or not) in the GWAS of AMD progression. MAF, minor allele frequency; BL, baseline severity; OR, odds ratio; HR, hazard ratio relative to the minor allele (minor allele/major allele). Model was adjusted for baseline age, education, smoking status and first two principal components in addition to the BL severity.

Of the 34 reported loci (3), rs42450006 upstream of MMP9 was previously identified in the case–control study to be exclusively associated with CNV but not with GA. That was the first variant identified to be specific to one subtype of AMD. We also confirmed the association for this variant with AMD progression: rs42450006 is associated with CNV progression (P = 0.0006) but not significantly with GA progression (P = 0.17; Table 5; Fig. 7). The MMP9 signal for CNV is consistent with previous evidence that upregulation of MMP9 induces neovascularization (18) and interacts with vascular endothelial growth factor (VEGF) signaling in the retinal pigment epithelial (19).

Table 5.

Results for rs142450006 in MMP9

SNP	Chr	Position	Major/minor allele	Gene	AMD subtypes	MAF	With BL severity		Without BL severity
							HR	P-value	HR	P-value
Significant loci reported also in consortium case–control studies
rs142450006	20	44 614 991	TTTTC/T	MMP9	GA	0.14	0.83	0.09	0.86	0.17
					CNV	0.13	0.66	0.0008	0.66	0.0006

MAF, minor allele frequency; BL, baseline severity; HR, hazard ratio relative to the minor allele (minor allele/major allele).

Figure 7.

KM curves for rs142450006 in MMP9. It was identified by the case–control study to be exclusively associated with CNV.

Discussion

We conducted the first GWAS for AMD progression by using a robust Cox proportional hazards regression model. The study of AMD progression complements earlier case–control studies by providing details about the process of AMD development and progression. Instead of using the subject-level analysis used in previous AMD progression papers (10,16), we performed the analysis at the eye-level to make full use of the data. Our results show that the progression of one eye is strongly dependent on its fellow eye’s progression. The subject-level analysis approach only considers the first progressed eye and ignores the correlation between two eyes. In contrast, we used the robust variance estimate in the Cox model to appropriately account for the correlation between eyes within a subject when assessing the genetic effects on AMD progression, which is consistent with the approach in Ding et al. (14).

In this study, we considered two sets of models: with and without adjustment for baseline severity. The baseline non-genetic risk factors are typically included in the GWAS so that the obtained genetic effect has been adjusted for other important risk factors. Thus, we considered the adjustment for baseline severity in addition to baseline age and smoking status. However, in our study the baseline severity is a special risk factor, given the response variable is determined by the longitudinal observations of this severity measurement. Therefore, inclusion of baseline severity can potentially mask the effects of other risk factors, including the genetic effect. Thus, we also considered the GWAS without adjustment for baseline severity.

We identified four progression loci with genome-wide significant association (P < 5 × 10⁻⁸) (Table 2). All of them were reported in the Fritsche et al. (3) case–control study, near ARMS2-HTRA1, CFH, C2-CFB-SKIV2L and C3, and they showed consistent effect directions as in the Fritsche et al. case–control study (Table 4). Of the lead SNPs in the four significant loci, rs10922109 (CFH), rs116503776 (C2-CFB-SKIV2L) and rs2230199 (C3) are also the lead SNPs in their corresponding loci in the Fritsche et al. (3) case–control study. The previous AMD progression studies (20,21) also showed the association of CFH and ARMS2 with AMD progression. In addition, C3 and C2 regions were reported to be associated with AMD progression in Seddon et al. (15), which analyzed 10 SNPs from the AREDS data. Thus, our GWAS results reinforced the understanding of these four gene regions on AMD progression. Although these four susceptibility genes have been consistently identified in case–control studies over years (2–4), their exact functional effects on AMD are still unclear. Our finding of the association between these genes and AMD progression can potentially help researchers better understand their functions. Furthermore, the four loci could be used to generate a genetic risk score (GRS) for the prediction of AMD progression. We also identified three additional loci that were marginally significant (P < 1 × 10⁻⁷) with AMD progression in the POC5/SV2C, HERC1 and ACKR3 regions (Table 2). These three loci were not identified by previous case–control studies or progression studies. They are relatively rare variants (i.e. 0.05 < MAF < 0.1). Thus, larger sample size studies are needed to verify their significance with AMD progression.

We also identified that the locus at TNR was specifically associated with progression to CNV at the P < 5 × 10⁻⁸ level, but not to GA even at a significance level of 0.05 (Table 3). The locus at ATF7IP2 was also found to be associated with progression to CNV at the level of 5 × 10⁻⁸, but not to GA at P = 5 × 10⁻⁸. Both loci did not reach genome wide significance level in the GWAS of any type of AMD progression. This finding indicates that the subtypes of AMD progression may have different genetic determinants.

Next, we examined whether the 34 loci reported in Fritsche et al. (3) were also associated with AMD progression (Table 4). In addition to the four loci (i.e. ARMS2-HTRA1, CFH, C2-CFB-SKIV2L and C3) associated at P < 5 × 10⁻⁸, LIPC and CTRB2-CTRB1 were also found to be associated with AMD progression (P < 0.0015, Bonferroni corrected significance level). Of the 34 loci, rs42450006 upstream of MMP9, which is the first locus identified to specifically associated with one AMD subtype (i.e. CNV) in a case–control study, was also found to be specifically associated with CNV progression in our study.

In addition to the AREDS data, AREDS2, a subsequent randomized clinical trial of additional oral supplements, was conducted for patients who were at high risk for developing late AMD. However, the AREDS2 data have quite different characteristics from the AREDS data (14) (see Table 1 in Ding et al.). The average age of AREDS2 participants is 2.8 years older than in the AREDS participants. The AREDS2 eyes are much more severe than the AREDS eyes at baseline. The majority (53%) of subjects have severity score of 7–8 in AREDS2. In contrast, only 12% subjects have severity score of 7–8 in AREDS. Thus, the AREDS2 subjects could progress to late AMD more easily due to high baseline severity. The genetic effects may not play a substantial role in the progression under this circumstance. We examined whether the four loci identified in the AREDS data are also associated with AMD progression in the AREDS2 data. In Supplementary Material, Table S8, the results show that only the variant at ARMS2-HTRA1 is significant (P < 0.0125, Bonferroni corrected significance level) with and without adjustment for baseline severity.

Our findings successfully demonstrate the benefit of GWAS of AMD progression as a complement of the case–control GWAS. The genetic study of AMD progression provides new insights for AMD risk investigation, which is helpful for researchers to better understand the AMD pathobiology. The confirmation of existing AMD risk related genes and identification of additional novel genes offer a potentially larger range of targets for AMD therapies. In our previous study (14), we derived a GRS) based on 34 risk variants from a case–control study for AMD risk (3) to predict AMD progression. In future, it will be more ideal to use the risk variants identified in this AMD progression GWAS to derive a GRS to establish prediction models for AMD progression.

Materials and Methods

Study population and phenotype definition

The study subjects are from the AREDS study, which was a multicenter, controlled, randomized clinical trial of treatment effects of antioxidant vitamins and/or zinc with 12 years follow up to assess the risk factors and clinical course of the progression of AMD and age-related cataracts. The AREDS data were part of a larger data set used in the previous consortium case–control studies (2,3). In this report, we confined our analyses to the AMD data and used Caucasian participants (validated with principal component analysis) with genotype data and at least one follow-up visit.

In AREDS, the participants were phenotyped based on fundus photographs (graded by a central reading center), best corrected visual acuity and ophthalmologic evaluations. The AMD severity was measured using a validated severity score based on centralized grading of stereoscopic fundus photographs obtained at each semi-annual/annual follow-up visit (22). The severity score ranges from 1 to 12, with 12 being the most severe stage. Late AMD is defined as the severity score is equal to 9 (non-central GA) or higher (10: central GA, 11: CNV, and 12: CNV and central GA). For each eye that is free of late AMD at baseline (i.e. baseline severity score <9), the time-to-progression is defined as the time (in years) from the baseline visit to the first visit when the severity score reaches 9 or higher. If the eye’s severity score is <9 by the end of follow-up, time-to-progression is treated as censored at the last visit. The covariates we have considered in the study are: gender, baseline age, education level (≤high school, >high school), baseline smoking status (never, former, current), and treatment group (placebo, antioxidants alone, zinc, antioxidants plus zinc).

Genotype data

DNA samples from consenting subjects in AREDS were collected and genotyped centrally by the International AMD Genomics Consortium (3). A custom-modified HumanCoreExome array by Illumina was used to obtain the genotypes followed by imputation with the 1000 Genomes Project reference panel (Phase I). Detailed information about genotype calling and imputation was described previously in (3). In this study, we use SNPs with minor allele frequency (MAF) > 0.05. Instead of using genotypes (0: no minor allele; 1: one copy of the minor allele; 2: two copies of the minor allele), we use dosages that can be any number between 0 and 2 in the analysis. The final data set includes 8 974 355 SNPs (265 096 genotyped and 8 709 259 imputed SNPs).

Statistical analysis

We use a Cox proportional hazards regression model to test for association between time-to-late AMD and SNPs. In order to account for the intra-subject correlation, a robust variance is used, which is calculated by grouped jackknife using the R function coxph with the ‘cluster’ option (23), where each subject is treated as a group/cluster (24). This approach has also been used in Sardell et al. (21) and Ding et al. (14) to analyze AMD progression, which we will refer to as the ‘robust Cox model’. The model formulation involves specification of a marginal Cox proportional hazards model for each eye which can be written as:

$${\mathrm{\lambda }}_{ij}\left(t|G_{\mathrm{i}},X_{ij},{PC}_i\right)={\mathrm{\lambda }}_0\left(t\right)\mathrm{e}\mathrm{x}\mathrm{p}\left\{G_i\alpha +X_{ij}\beta +{PC}_i\gamma \right\}$$

where $t$ is the progression time, ${\mathrm{\lambda }}_{ij}\left(t\right)$ is the hazard function of progression time for the jth eye (j = 1: left, 2: right) in the ith subject, ${\mathrm{\lambda }}_0\left(t\right)$ is the baseline hazard function (we assume the baseline hazard function to be the same for left and right eyes), $G_i$ (ranging from 0 to 2) is the SNP dosage for the ith subject, representing the copies of minor alleles, $X_{ij}$ are the covariates for the jth eye of ith subject and ${PC}_i$ is the principal component(s) of ancestry for subject i. Note that the covariates $X_{ij}$ may be defined based on either subject-level or eye-level features. The variables that show significant association with AMD progression in the uni-variable robust Cox model with P < 0.1 are considered as covariates in the GWAS analysis.

In addition, we utilize DAVID (17) to perform gene-annotation enrichment analysis to highlight the most relevant GO terms and KEGG pathways associated with a list of AMD progression associated genes. This is to test whether the number of AMD progression associated genes included in some relevant GO terms or KEGG pathways is more than expected by chance. A modified Fisher’s exact test is used for this enrichment analysis.

Data availability

All the phenotype data of AREDS participants required in this report were obtained from the public available website dbGap (accession: phs000001.v3.p1). The genotype data on AREDS participants has been reported earlier (3) and is available from dbGap (accession phs001039.v1.p1).

Supplementary Material

Supplementary Material is available at HMG online.