A large genome-wide association study of age-related macular degeneration highlights contributions of rare and common variants.

Advanced age-related macular degeneration (AMD) is the leading cause of blindness in the elderly, with limited therapeutic options. Here we report on a study of >12 million variants, including 163,714 directly genotyped, mostly rare, protein-altering variants. Analyzing 16,144 patients and 17,832 controls, we identify 52 independently associated common and rare variants (P < 5 × 10(-8)) distributed across 34 loci. Although wet and dry AMD subtypes exhibit predominantly shared genetics, we identify the first genetic association signal specific to wet AMD, near MMP9 (difference P value = 4.1 × 10(-10)). Very rare coding variants (frequency <0.1%) in CFH, CFI and TIMP3 suggest causal roles for these genes, as does a splice variant in SLC16A8. Our results support the hypothesis that rare coding variants can pinpoint causal genes within known genetic loci and illustrate that applying the approach systematically to detect new loci requires extremely large sample sizes.

Advanced age-related macular degeneration (AMD) is a neurodegenerative disease and the leading cause of vision loss among the elderly affecting 5% of those >75 years of age[1,2]. The disease is characterized by reduced retinal pigment epithelium (RPE) function and photoreceptor loss in the macula. Advanced AMD is classified as wet (choroidal neovascularization, CNV, when accompanied by angiogenesis) or dry AMD (geographic atrophy, GA, when angiogenesis is absent). These advanced stages of disease are typically preceded by clinically asymptomatic earlier stages[3]. Advanced AMD is estimated to affect 10 million patients worldwide, reaching >150 million for earlier stages[4]. At present, understanding of disease biology and therapies remains limited[5].

Genetic variants can help uncover disease mechanisms and provide entry points into therapy. Analyses of common variation have uncovered numerous risk loci for many complex diseases (see Web Resources) including 21 loci for AMD[6-12]. However translation into biological insights remains a challenge, since the functional consequences of disease-associated common variants are typically subtle[13] and hard to decipher.

With advances in sequencing technology, genetic analyses are gradually extending to rare variants, which often have more obvious functional consequences[14,15] and can thus accelerate translation into biological understanding[14,16]. For example, identifying multiple disease-associated coding variants (particularly knock-out alleles) in the same gene provides strong evidence that disrupting gene function leads to disease[17]. So far, studies that implicate specific rare variants in complex diseases either rely on special populations[8,18,19], on targeted examinations of a few genes[7,9-11,20,21], or on genome-wide assessments of relatively modest numbers of individuals[22-25]. In contrast, systematic analyses of common variation are now available in hundreds of thousands of phenotyped individuals[26,27]. Thus, there remains considerable uncertainty about the relative role of rare variants in complex disease and about the sample sizes and study designs that will enable systematic identification of these variants[16].

Here, we set out to systematically examine common and rare variation of AMD in the International AMD Genomics Consortium (IAMDGC). The preceding largest study of AMD examined ∼2.4 million variants including ∼18,000 imputed or genotyped protein-altering variants using meta-analysis[6]. Customizing a chip for de novo centralized genotyping, we analyze >12 million variants including 163,714 directly typed protein-altering variants in 43,566 unrelated subjects of predominantly European ancestry. Our study constitutes a detailed simultaneous assessment of common and rare variation in a complex disease and a large sample, setting expectations for other well-powered studies.

Results

The study data and genomic heritability

We gathered advanced AMD cases with GA and/or CNV, intermediate AMD cases, and control subjects across 26 studies (Supplementary Table 1). While recruitment and ascertainment strategies varied (Supplementary Table 2), DNA samples were collected and genotyped centrally. Making maximal use of genotyping technologies, we utilized a chip with (i) the usual genome-wide variant content, (ii) exome content comparable to the exome chip (adding protein-altering variants from across all exons), and a specific customization to add (iii) protein-altering variants detected by our prior sequencing of known AMD loci (see Methods) and (iv) previously observed and predicted variation in TIMP3 and ABCA4, two genes implicated in monogenic retinal dystrophies. After quality control, we retained 439,350 directly typed variants including a grid of 264,655 primarily non-coding (93%) common variants (frequency among controls >1%) and 163,714 protein-altering variants (including 8,290 from known AMD loci), mostly rare (88% with frequency among controls ≤1%). Imputation to the 1000 Genomes reference panel enabled examining a total of 12,023,830 variants (Supplementary Table 3A). Our final data set included a total of 43,566 subjects consisting of 16,144 advanced AMD patients and 17,832 control subjects of European ancestry for our primary analysis, as well as 6,657 Europeans with intermediate disease and 2,933 subjects with Non-European ancestry (Supplementary Table 3B, Supplementary Figure 1).

Altogether, our genotyped markers accounted for 46.7%[28] of variability in advanced AMD risk in the European ancestry subjects (95% confidence interval [CI] 44.5% to 48.8%). Regarding AMD subtypes, estimates for CNV (h[2] = 44.3%, CI 42.2% to 46.5%) and GA (h[2] = 52.3%, CI 47.2% to 57.4%) were similar; bivariate analyses[29] showed genetic correlation of 0.85 (CI 0.78 to 0.92) between disease subtypes.

Thirty-Four Susceptibility Loci for AMD

We first conducted a genome-wide single variant analysis of the >12 million genotyped or imputed variants (applying genomic control λ=1.13) comparing the 16,144 advanced AMD patients and 17,832 controls of European ancestry (full results online; see Web resources). We obtained >7000 genome-wide significant variants (P ≤ 5×10–8, Supplementary Figure 2). Sequential forward selection (Supplementary Figure 3) identified 52 independently associated variants at P ≤ 5×10–8 (Supplementary Table 4, Supplementary File 1). These are distributed across 34 locus regions (Figure 1A), each extending across the identified and correlated variants, r[2]≥0.5, ±500kb (Supplementary Table 5). The 34 loci include 16 loci that reached genome-wide significance for the first time (novel loci, Table 1) and include genes with compelling biology like extra-cellular matrix genes (COL4A3, MMP19, MMP9), an ABC transporter linked to HDL cholesterol (ABCA1), and a key activator in immune function (PILRB). Also included are 18 of the 21 AMD loci that reached genome-wide significance previously[6,9] (known loci, Table 1), between-study heterogeneity was low, particularly for the new loci (Supplementary Note 1, Supplementary Table 6, 7).

Figure 1

Genome-wide search reveals 34 loci and genes with rare variant burden for AMD

(a) We conducted a genome-wide single variant association analysis for >12 million variants in 16,144 advanced AMD patients versus 17,832 controls. Shown is the Manhattan Plot exhibiting P-values for association highlighting novel (P < 5×10–8 for the first time, green) and known (blue) AMD loci (see Table 1). (b) We computed independent effect size (Odds Ratios) of each of the 52 identified variants (Supplementary Table 4). Shown are these effect sizes versus the frequency of the AMD risk increasing allele and a 80% power curve. (c) We conducted a genome-wide gene-based test for disease burden based on the protein-altering variants testing 17,044 RefSeq genes by the variable threshold test[51]. Shown is the Manhattan Plot with P-values, the red horizontal line indicating genome-wide significance (P ≤ 0.05/17,044 = 2.9×10–6) and the yellow line indicating AMD-locus-wide significance (given 703 genes in the 34 AMD loci, P ≤ 0.05/703 = 7.1×10–5). No gene outside the 34 loci is genome-wide significant; 14 genes are AMD-locus-wide significant (blue), four remain significant after locus-wide conditioning (bold letters, Supplementary Table 11).

Table 1

Thirty-four loci for age-related macular degeneration

Our genome-wide single-variant association analysis identified 34 loci for advanced AMD with genome-wide significance (P < 5 × 10–8) based on logistic regression in 16,144 cases and 17,832 controls of European ancestry. Shown are P-values and effect sizes (Odds Ratios, OR) for the variant with the smallest P-value per locus (lead variant) and the number of independent signals per locus (see Supplementary Table 4)

Lead Variant	Chr	Positiona	Major/minor allele	Locus nameb	# Signalsc	MAF			Association
						Cases	Controls	OR	P
KNOWN (previously reported with genome-wide significance, P < 5 × 10–8)
rs10922109	1	196,704,632	C/A	CFH	8	0.223	0.426	0.38	9.6 × 10–618
rs62247658	3	64,715,155	T/C	ADAMTS9-AS2	1	0.466	0.433	1.14	1.8 × 10–14
rs140647181	3	99,180,668	T/C	COL8A1	2	0.023	0.016	1.59	1.4 × 10–11
rs10033900	4	110,659,067	C/T	CFI	2	0.511	0.477	1.15	5.4 × 10–17
rs62358361	5	39,327,888	G/T	C9	1	0.016	0.009	1.80	1.3 × 10–14
rs116503776	6	31,930,462	G/A	C2/CFB/SKIV2L	4	0.090	0.148	0.57	1.2 × 10–103
rs943080	6	43,826,627	T/C	VEGFA	1	0.465	0.497	0.88	1.1 × 10–14
rs79037040	8	23,082,971	T/G	TNFRSF10A	1	0.451	0.479	0.90	4.5 × 10–11
rs1626340	9	101,923,372	G/A	TGFBR1	1	0.189	0.209	0.88	3.8 × 10–10
rs3750846	10	124,215,565	T/C	ARMS2/HTRA1	1	0.436	0.208	2.81	6.5 × 10–735
rs9564692	13	31,821,240	C/T	B3GALTL	1	0.277	0.299	0.89	3.3 × 10–10
rs61985136	14	68,769,199	T/C	RAD51B	2	0.360	0.384	0.90	1.6 × 10–10
rs2043085	15	58,680,954	T/C	LIPC	2	0.350	0.381	0.87	4.3 × 10–15
rs5817082	16	56,997,349	C/CA	CETP	2	0.232	0.264	0.84	3.6 × 10–19
rs2230199	19	6,718,387	C/G	C3	3	0.266	0.208	1.43	3.8 × 10–69
rs429358	19	45,411,941	T/C	APOE	2	0.099	0.135	0.70	2.4 × 10–42
rs5754227	22	33,105,817	T/C	SYN3/TIMP3	1	0.109	0.137	0.77	1.1 × 10–24
rs8135665	22	38,476,276	C/T	SLC16A8	1	0.217	0.195	1.14	5.5 × 10–11
NOVEL (reported with genome-wide significance, P < 5 × 10–8, for the first time)
rs11884770	2	228,086,920	C/T	COL4A3	1	0.258	0.278	0.90	2.9 × 10–8
rs114092250	5	35,494,448	G/A	PRLR/SPEF2	1	0.016	0.022	0.70	2.1 × 10–8
rs7803454	7	99,991,548	C/T	PILRB/PILRA	1	0.209	0.190	1.13	4.8 × 10–9
rs1142	7	104,756,326	C/T	KMT2E/SRPK2	1	0.370	0.346	1.11	1.4 × 10–9
rs71507014	9	73,438,605	GC/G	TRPM3	1	0.427	0.405	1.10	3.0 × 10–8
rs10781182	9	76,617,720	G/T	MIR6130/RORB	1	0.328	0.306	1.11	2.6 × 10–9
rs2740488	9	107,661,742	A/C	ABCA1	1	0.255	0.275	0.90	1.2 × 10–8
rs12357257	10	24,999,593	G/A	ARHGAP21	1	0.243	0.223	1.11	4.4 × 10–8
rs3138141	12	56,115,778	C/A	RDH5/CD63	1	0.222	0.207	1.16	4.3 × 10–9
rs61941274	12	112,132,610	G/A	ACAD10	1	0.024	0.018	1.51	1.1 × 10–9
rs72802342	16	75,234,872	C/A	CTRB2/CTRB1	1	0.067	0.080	0.79	5.0 × 10–12
rs11080055	17	26,649,724	C/A	TMEM97/VTN	1	0.463	0.486	0.91	1.0 × 10–8
rs6565597	17	79,526,821	C/T	NPLOC4/TSPAN10	1	0.400	0.381	1.13	1.5 × 10–11
rs67538026	19	1,031,438	C/T	CNN2	1	0.460	0.498	0.90	2.6 × 10–8
rs142450006	20	44,614,991	TTTTC/T	MMP9	1	0.124	0.141	0.85	2.4 × 10–10
rs201459901	20	56,653,724	T/TA	C20orf85	1	0.054	0.070	0.76	3.1 × 10–16

Chr = Chromosome; MAF = minor allele frequency; OR = Odds Ratio

Chromosomal position is given based on NCBI RefSeq hg19;

The locus name is a label of the region using the nearest gene(s), but does not necessarily state the responsible gene;

number of independent variants in this locus; hg19 = human genome reference assembly (version 19)

Most associated variants are common (45 out of 52) with fully conditioned odds ratios (OR) from 1.1 to 2.9 (Figure 1B, Supplementary Table 4) with two interacting variants (Supplementary Note 2). We also observed seven rare variants with frequencies between 0.01% and 1% and ORs between 1.5 and 47.6 (Figure 1B, Supplementary Table 4). All of these variants were also rare in Non-European ancestries (Supplementary Table 8, extended association results on Non-European in Supplementary File 2). All seven rare variants are located in/near complement genes: four previously described non-synonymous (CFH:Arg1210Cys, CFI:Gly119Arg, C9:Pro167Ser, C3:Lys155Gln)[7-11]; three others (CFH: rs148553336, rs191281603, rs35292876) described here for the first time including two with the rare allele decreasing the disease risk. . To ensure validity of our results, we verified associations of lead variants in sensitivity analyses that relied on alternate association tests, adjusted for age, gender, or ten ancestry principal components, or were restricted to population-based controls or controls ≥ 50 years of age (data not shown). Altogether, our genome-wide single variant analysis nearly doubles the number of AMD loci and variants.

Prioritizing variants within 52 association signals

It is often challenging to translate common variant association signals into mechanistic understanding of biology; two key challenges are (i) variants with similar signals because of linkage disequilibrium and (ii) subtle functional consequences. Without narrowing lists of candidate variants, follow-up functional experiments are complicated. To prioritize among nearby variants, we computed each variant's ability to explain the observed signal and derived, for each of the 52 signals, the smallest set of variants that included the causal variant with 95% probability [30,31]. The 52 credible sets each included from 1 to >100 variants (total of 1,345 variants, Supplementary File 3). Twenty-seven (of 52) sets were small with ≤10 variants (19 with ≤5 variants, Supplementary Table 9); seven sets included only one variant. Among the 205 variants with >5% probability of being causal, we observe 11 protein-altering (all non-synonymous) variants (versus 2 expected assuming 1% protein-altering variants, P for enrichment = 8.7×10–6, Supplementary Table 10). We recognize that the analysis has limitations [for example, when causal variants when the signal is due to a combination of multiple variants, as in the counter example in Supplementary Figure 4].

Rare Variant Association Signals

Analysis of rare variants that alter peptide sequences (non-synonymous), truncate proteins (premature stop), or affect RNA splicing (splice site) can help to identify causal mechanisms – particularly when multiple associated variants reside in the same gene[16,32]. We examined the cumulative effect of rare protein-altering variants in each ancestry group. Genome-wide, no signal was detected with P ≤ 0.05/17,044 = 2.9×10–6 outside the 34 AMD loci (Figure 1C). Within the 34 loci, we found 14 genes with significant disease burden (P < 0.05/703 genes = 7.1×10–5, Supplementary Table 11). To eliminate settings where a rare variant burden finding is a linkage disequilibrium shadow of a nearby common variant, we re-evaluated each burden signal conditioning on nearby single variants (from Supplementary Table 4). Four of the 14 genes retained P < 0.05/703 = 7.1×10–5 in this analysis (CFH, CFI, TIMP3, SLC16A8; conditioned P = 1.2×10–6, 1.0×10–8, 9.0×10–8, or 3.1×10–6, respectively, Table 2). Sensitivity analyses provide similar (excluding previously sequenced subjects) and extended results (prioritizing variants with high predicted functionality, Supplementary Note 3, Supplementary Table 12).

Table 2

Four genes with a significant rare variant burden within the 34 AMD loci independent from other identified variants

We computed a gene-based burden test of rare protein-altering variants comparing 16,144 advanced AMD cases and 17,832 controls. Shown are P-values from the variable threshold test (up to 100 million permutations) and Odds Ratios from the collapsed burden test, both adjusted for the other identified variants in the respective locus (locus-wide conditioning). Four genes (among the 703 genes in the 34 AMD locus regions) showed a significant (P < 0.05/703 = 7.1 × 10–5) burden. Details about the corresponding rare variants underlying the observed burden can be found in Supplementary File 4. Results for the 14 genes that show significant burden within the 34 AMD loci without locus-wide conditioning are shown in Supplementary Table 11. Rare variants were defined here as variants with minor allele frequency in cases and controls < 1% in each of the ancestries, European, Asian, and African.

Gene	Optimal Threshold for Rare Variants Count (%)	Number of Variants below Optimal RAC	Summed Rare Allele Count (Frequency [%])		Pa	Odds Ratio
		Total (Exome Chip Base + Custom)	Cases N = 16,144	Controls N = 17,832
CFH	10 (0.015%)	37 (9+28)	88 (0.273%)	38 (0.107%)	1.2 × 10–6	2.94
CFI	46 (0.068%)	43 (17+26)	213 (0.660%)	82 (0.230%)	1.0 × 10–8	2.95
TIMP3	14 (0.021%)	9 (1+8)	29 (0.0898%)	1 (0.00280%)	9.0 × 10–8	31.21
SLC16A8	648 (0.954%)	9 (7+2)	487 (1.51%)	392 (1.10%)	3.1 × 10–6	1.40

RAC = rare allele count;

P-values are from the variable threshold test conditioned on other identified variants in the locus (locus-wide conditioned).

Several interesting patterns emerge, many of which we owe to our chip design. First, three of the four rare variant burden signals (CFH, CFI, TIMP3) are due to variants with frequency <0.1%, all genotyped (Supplementary File 4). Many human genetic studies have used frequency thresholds of 1% to 5% as a working definition of “rare”, but our data suggests that trait associated variants with clear function may often be much rarer – necessitating very large sample sizes for analysis. In two genes (CFH, CFI), the rare burden was detected because we enriched arrays with variants from previous sequencing of AMD loci[10] (54 of 80 variants). The burden findings in CFH (new, Supplementary Note 4) and CFI[9] together with variants CFH:Arg1210Cys and CFI:Gly119Arg[7,9], corroborate a causal role for these genes in AMD etiology.

The third signal (TIMP3) was in a gene previously associated with Sorsby's fundus dystrophy, a rare monogenic disease with early onset at <45 years of age but with clinical presentation strikingly similar to AMD[33,34]. Because the majority of Sorsby's alleles disrupt cysteine-cysteine bonds in TIMP3, we arrayed all possible cysteine disrupting sites together with other previously described Sorsby's risk alleles [33,34]. The nine rarest TIMP3 variants were cumulatively associated with >30-fold increased risk of disease. TIMP3 resides in an established AMD locus[5,35] targeted in previous sequencing efforts[32,35] that were too small to evaluate rare variation on this scale (1 variant in 17,832 controls versus 29 variants in 16,144 cases). Interestingly, although Sorsby-associated TIMP3 variants typically occur in exon 5, four of the unpaired cysteine residues we observed map to other exons – perhaps because unpaired cysteines in different locations impair protein folding in different ways. AMD cases with these rare TIMP3 risk alleles still exhibited higher counts of AMD risk alleles across the genome than controls, suggesting that TIMP3 is not a monogenic cause of AMD but contributes to disease together with alleles at the other risk loci. Our finding illustrates a locus where complex and monogenic disorders arise from variation in the same gene, similar to MC4R and POMC in obesity[36] or UMOD in kidney function[37]. In a similar approach, we analyzed 146 rare protein-altering variants in ABCA4, a gene underlying Stargardt disease[38], but found no association (P=0.97).

The rare variant burden signal in SLC16A8 was primarily driven by a putative splice variant (c.214+1G>C, rs77968014, minor allele frequency among controls, CAF = 0.81%, OR = 1.5, imputed with R[2]=0.87, Supplementary File 4). This is not a burden from multiple rare variants, but a single variant emerging as significant due to the reduced multiple testing from gene-wide testing (single variant association P = 9.1×10-6, conditioned on rs8135665 P = 1.3×10-6). This variant is interesting as it is predicted to disrupt processing of the encoded transcript (as +1 G variant, Human Splicing Finder 3.0). SLC16A8 encodes a cell membrane transporter, involved in transport of pyruvate, lactate and related compounds across cell membranes[39]. This class of proteins mediates the acidity level in the outer retinal segments, and SLC16A8 gene knock-out animals have changes in visual function and scotopic electroretinograms, but not overt retinal pathology[40]. Interestingly, a progressive loss of SLC16A8 expression in eyes affected with GA was reported with increasing severity of disease[41]. In summary, our chip design and our large data set enabled us not only to detect interesting features of AMD genetics, but also to provide guidance for future investigations on rare variants.

From Disease Loci to Biological Insights

Many analyses can further narrow the list of candidate genes in our loci. We annotated the 368 genes closest to our 52 association signals (index variant and proxies, r[2] ≥0.5, ±100kb, Supplementary File 5), noting among these the genes those that contained associated credible set variants (Supplementary File 3) or a rare variant burden (Table 2) – these are the highest priority candidates, consistent with previous analysis of putative cis-regulatory variants[42]. We further checked whether genes were expressed in retina (82.6% of genes) or RPE/choroid (86.4%, Supplementary File 6). We sought relevant eye phenotypes in genetically modified mice (observed in 32 of the 368 queried genes, Supplementary File 7). We tagged genes in biological pathways enriched across loci, such as the alternative complement pathway, HDL transport, and extracellular matrix organization and assembly (Supplementary Table 13) – highlighting genes that connect multiple pathways (COL4A3/COL4A4, ABCA1, MMP9, and VTN). We also highlighted genes that were approved or experimental drug targets (31 of the 368 queried, Supplementary File 8). Finally, we prioritized genes where at least one of the credible set variants (Supplementary File 3) was protein-altering or located in a putative functional region (promoter, 3′/5′ UTR).

All this information is summarized in the gene priority score table (Supplementary File 9, Supplementary Note 5, Supplementary Table 14), which uses a simple customizable scoring scheme to assign priority: the scheme using equal weights for each column assigns highest scores (Figure 2A, Supplementary Table 15) to genes such as master regulators of immune function (PILRB), matrix metalloproteinase genes (MMP9, MMP19), genes involved in in lipid metabolism (ABCA1, GPX4), an inhibitor of the complement cascade (VTN), another collagen gene (COL4A3), a gene causing a developmental monogenic disorder (PTPN11), and a retinol dehydrogenase (RDH5). Six of these are current drug targets (ABCA1, MMP19, RDH5, PTPN11, VTN, GPX4). In the known AMD loci, the highest scores per locus included the usual suspects (CFH, CFI, CFB, C3, and APOE) as well as TIMP3 and SLC16A8 (Figure 2B). This summary of evidence is not amenable to formal statistical enrichment testing, but may help prioritize genes for follow-up functional experiments.

Figure 2

Genes with top priority based on biological and statistical evidence combined

We queried 368 genes in the 34 narrow AMD regions (index and proxies, r[2] ≥0.5, ±100kb) for biological (red; expression in retina/RPE/choroid, Supplementary File 6; ocular mouse phenotype, Supplementary File 7), statistical, (blue; ≥1 credible set variant in gene ±50 kb, Supplementary File 3; rare variant burden, Table 2), putative functional (green; ≥ 1 credible set variant in gene ±50 kb being protein-altering, 5′/3′ UTR, other exonic, or putative promoter, Supplementary File 3), and molecular (magenta; enriched molecular pathway, drug target) evidence. We here focus on the gene(s) with the highest gene priority score (GPS) per locus (full list of genes in Supplementary File 9). Shown are (a) the 16 genes with highest GPS in the 15 novel AMD loci (one novel locus without any gene), and (b) the 25 genes with highest GPS in the 18 known AMD loci. Colored fields indicate yes and GPS counts number of colored fields per row.

Commonalities and differences of advanced AMD subtypes

Previously identified risk variants all contribute to the two advanced AMD subtypes, CNV and GA. We compared association signals between our 10,749 cases with CNV and 3,235 cases with GA. Four of the 34 lead variants show significant difference (Pdiff < 0.05/34 = 0.00147) between disease subtypes (in the loci ARMS2/HTRA1, CETP, MMP9, SYN3/TIMP3, Figure 3A, Supplementary Table 16). Variant rs42450006 upstream of MMP9 was the only one that was specific to one subtype, being exclusively associated with CNV (frequency in controls = 14.1%; ORCNV = 0.78 vs. ORGA = 1.04; Pdiff = 4.1×10–10), but not with GA (PGA=0.39, Supplementary Note 6). The MMP9 signal for neovascular disease fits well with prior evidence: upregulation of MMP9 appears to induce neovascularization[43] and interacts with VEGF signaling in the RPE[44]. VEGF currently provides an effective therapy for patients with CNV, but the struggle to keep vision continues. Beyond confirming a shared genetic predisposition of the two subtypes, our data identifies – for the first time – one variant that is specific to one subtype.

Figure 3

Comparison of advanced AMD subtypes and intermediate versus advanced AMD

We compared associations of the 34 lead variants across different AMD phenotypes. Shown are effect sizes (log Odds Ratio) per minor allele in controls as well as 95% confidence intervals (widths and heights of diamonds). (a) Comparison of neovascular disease (10,749 CNV cases vs. 17,832 controls) and GA (3,235 GA cases vs. 17,832 controls) identified four variants (in loci MMP9, ARMS2/HTRA1, CETP, and SYN3/TIMP3) with significantly different association comparing CNV with GA (Pdiff < 0.05/34, marked in red, see also Supplementary Table 16). (b) Comparison of intermediate AMD (6,657 cases vs. 17,832 controls) with advanced AMD (16,144 cases vs. 17,832 controls) identifies 24 variants with nominally significant (P < 0.05, marked in red) association with intermediate AMD (Pbinomial = 4.8 × 10–24), all of which have the same effect direction and less extreme effect sizes compared to advanced AMD (Supplementary Table 17).

Commonalities and differences of advanced and early AMD

We evaluated our association signals in 6,657 individuals with intermediate AMD, defined as having more than five macular drusen greater than 63μm and/or pigmentary changes in the RPE. Examining all genotyped variants[28], we found a correlation of 0.78, indicating substantial overlap between genetic determinants of advanced and intermediate AMD (95% CI 0.69 to 0.87). Among our 34 index variants, 24 showed nominally significant association (Pintermediate ≤ 0.05) with intermediate AMD (2 expected, Pbinomial = 4.8×10–24); all had ORs in the same direction but smaller in magnitude (Figure 3B, Supplementary Table 17). The other 10 variants showed no association with intermediate AMD (Pintermediate > 0.05), despite sufficient power (Supplementary Table 18). Interestingly, these 10 variants point to 7 extra-cellular matrix genes (COL15A1, COL8A1, MMP9, PCOLCE, MMP19, CTRB1/2, ITGA7, Supplementary Table 19), based on which one may hypothesize that the extra-cellular matrix points to a disease subtype without early stage manifestation or with extremely rapid progression. If confirmed, a group of rapidly progressing patients or without early symptoms might eventually derive maximum benefit from genetic diagnosis and future preventive therapies.

An Accounting of AMD Genetics

To account for progress made here in understanding AMD genetics, we estimated the proportion of disease risk explained by our 52 independent variants and compared it to our initial estimates of heritability obtained by examining all genotyped variants. We computed a weighted risk score of the 52 variants[45] and modeled a population risk score distribution (see Materials and Methods). Individuals in the highest decile of genetic risk have a 44-fold increased risk of developing advanced AMD compared to the lowest decile; of these, 22.7% are predicted to have AMD in an elderly general population above 75 years of age with ∼5% disease prevalence (Figure 4A, Supplementary Table 20). Altogether, the 52 variants explain 27.2% of disease variability (Figure 4B, also highlighting results based on other prevalence assumptions), including a 1.4% contribution from rare variants. The 52 identified variants thus explain more than half of the genomic heritability; the balance might be attributed to additional variation not studied here, or to genetic interaction with environmental factors such as smoking, diet or sunlight exposure.

Figure 4

Variance explained and absolute risk of disease based on the 52 identified variants

(a) Absolute disease risk (=proportion of affected) by genetic risk score intervals (deciles and top 10 percentiles in embedded bar plot) based on our cases-control-data weighted to model a general population with 5% disease prevalence (see also Supplementary Table 20). (b) Shown is disease liability explained by the 52 identified variants (bars) compared to the genomic heritability based on all genotyped variants (red lines) assuming disease prevalence of 1%, 5%, or 10%, respectively.

Discussion

We set out to improve our understanding of rare and common genetic variation for macular degeneration biology, to guide the development of therapeutic interventions and facilitate early diagnosis, monitoring and prevention of disease. We systematically examine rare variation (through direct genotyping) and common variation (through genotyping and imputation) for AMD in a study designed to discover >80% of associated protein-altering variants with an allele frequency of >0.1% and >3-fold increased disease risk (or >0.5% frequency and >1.8-fold increased disease risk). Our study provides a simultaneous assessment of common and rare variation enabling us to understand the relative roles of rare and common variants and the scientific insights to be gained from rare variation.

Rare protein-altering variants are an attractive target for genetic studies because most of these variants are expected to damage gene function. Furthermore, observing that many rare variants in a gene are, together, associated with a change in disease risk strongly suggests that the gene is causally implicated in disease biology and – further – suggests the consequences of mimicking or blocking gene action using a drug. Our study demonstrates that when rare variants are systematically assessed genome-wide, significant signals can be assigned to single rare variants as well as to rare variant burden in individual genes.

Our study also demonstrates the challenges of these analyses. For three of the genes where we identified a rare variant burden, the accumulated evidence was spread across very rare variants with frequencies <0.1% in controls. Most of these variants derived from sequencing AMD patients. This emphasizes the value of a hybrid approach with direct targeted sequencing of patient samples for variant discovery, followed by genotyping in larger samples for association analysis. Another conclusion is about required sample sizes: although such rare variants are expected to exist in nearly all genes, no rare variant burden was observed in most of the 34 loci we studied. For these loci, identifying causal mechanisms through the study of rare protein-altering variants will require a combination of more sequencing and even larger sample sizes. While our findings of rare variant burden are predominantly from targeted enrichment, the knowledge about effect sizes and frequencies of contributing variants illustrates that applying the approach genome-wide to detect new loci requires extremely large sample sizes. In our view, a recent estimate that sequencing of 25,000 patients will be needed to identify genes where rare variants have a substantial impact on disease risk is likely to be optimistic, particularly given the fact that effect sizes for AMD risk alleles appear to be larger than for many other complex traits [16].

In addition to corroborating previous reports of rare variants that disrupt genes in the complement pathway and lead to large increases in disease risk, our study also includes two unexpected rare variant findings. First, we show that a putative splice variant in SLC16A8 can greatly increase the risk of age-related macular degeneration – providing strong evidence that the gene is directly involved in disease biology. SLC16A8 is a lactate transporter expressed[39] specifically by the RPE; a deficit of lactate transport results in acidification of the retina and photoreceptor dysfunction in Slc16a8 knock-out mice[40]. Second, we show a >30-fold excess of rare TIMP3 mutations among putative cases of macular degeneration. TIMP3 is an especially attractive candidate that has been the subject of previous, underpowered, genetic association studies.

While it has been hypothesized that studies of rare and low frequency genetic variants will greatly increase the proportion of genetic risk that can be explained, our results don't support this. Our study and others successfully identify many low frequency disease risk alleles, and these provide clues about disease biology, but our results also show that common variants make a much larger contribution to disease risk. Common variants also suggest interesting leads and pathways for future analysis (Supplementary Table 15, Figure 2A), including attractive candidates such as immune regulators (PILRB), genes implicated in mouse ocular phenotypes (MMP9, MMP19, COL4A3, PTPN11, GPX4, and RDH5), and proven drug targets (ABCA1, MMP19, RDH5, PTPN11, VTN, GPX4). In a literature search, we identified no previous candidate gene association studies targeting our novel loci, although several model organism, cellular, and functional studies evaluated potential links between genes in these loci and AMD (highlights of this search in Supplementary Table 15) and a few loci were nominally associated and proposed as candidates in prior genome-wide searches [46,47]. As richer functional annotations of the genome[48] become available in diverse cell types, systematic assessment of overlap between these and our loci should clarify disease biology.

Our study also suggests additional important observations. While our results show that the majority of genetic risk is shared between GA and CNV, we also identify – for the first time – a variant that is specific to one advanced AMD subtype: a genetic variant near MMP9 is specific to CNV, a candidate gene also supported by prior gene expression analyses in the Bruch's membrane of patients with neovascular disease[49]. Future efforts extending to longitudinal data might help improve the dissection of pure CNV and pure GA and their genetic make-up even further. If substantiated, the fact that nearly all disease associated variants modulate risk of both CNV and GA has potentially significant therapeutic consequences. It implies that individuals at high risk of CNV are also at high risk of GA. This suggests that therapeutic strategies which mitigate CNV but not GA will only provide temporary relief to patients – who are likely to remain at high risk of developing GA and may still require future interventions to prevent it.

Therefore, our findings have several important implications for future studies of rare variation in human complex traits. First, they clearly emphasize the need for very large sample sizes in population studies: the functionally most interesting variants we identify have frequencies in the range of 0.01 – 1.0% and, despite their strong impact on disease risk, could only be implicated using 10,000s of individuals. Second, they illustrate the value of hybrid approaches, where sequencing is used to detect interesting variants and custom arrays and imputation are used to examine these variants in very large samples. Since all the large effect rare variants we identify reside in or near GWAS loci, as with most complex trait associated rare variants [7-11,20,21,23,50], focused studies around GWAS loci may continue to be a cost-effective compromise. Third, our analysis of cysteine variants in TIMP3 illustrates not only the potential for targeted variant discovery but the critical need to understand the consequences of rare variants when analyzing them together. While very large samples will be needed, our results also show that the effort to extend genetic studies to rare variants is worthwhile as these variants can pinpoint causal genes and advance our understanding of disease biology.

Online Methods

Study data and phenotype

In the International AMD Genomics Consortium (IAMDGC), we gathered 26 studies with each including (i) advanced AMD cases with GA and/or CNV in at least one eye and age at first diagnosis ≥ 50 years, (ii) intermediate AMD cases with pigmentary changes in the RPE or more than five macular drusen greater than 63μm and age at first diagnosis ≥ 50 years, or (iii) controls without known advanced or intermediate AMD. Recruitment and ascertainment strategies varied by study (Supplementary Tables 1 and 2, Supplementary Note 7). All groups collected data according to the Declaration of Helsinki principles. Study participants provided informed consent and protocols were reviewed and approved by local ethics committees.

DNA and chip design

We gathered DNA samples of more than 50,000 individuals. Groups with very limited amounts of available DNA contributed aliquots after whole-genome amplification (8% of subjects).

We utilized a custom-modified HumanCoreExome array by Illumina, Inc., which includes (i) tagging variants across the genome (genome chip content) and (ii) a catalogue of protein-altering variants (exome chip content). Our customization of the array included three additional tiers to enrich for variants from 22 AMD loci implicated by our previous genome-wide association analysis[6] based on 19 index variants with genome-wide significance, 3 with consistent effect direction in the replication stage and 4×10–7 ≤ P ≤ 2×10–6) by selecting (iii) tagging variants (pair-wise tagging r[2] < 0.8) from Phase I 1000G/HapMap[52,53] common variants (minor allele frequency, MAF, ≥ 1 % in European or East Asian individuals) using Tagger implemented in Haploview[54] within ±100kb of the 22 index variants expanded to cover all correlated variants (r[2] [EUR] > 0.5) and the complete gene (transcript ±1 kb), (iv) protein-altering variants within 500 kb of the 22 index variants as identified from public general population data bases (dbSNP[55], the NHLBI Exome Sequencing Project[56], the Phase I 1000 Genomes Project, see Web Resources), and (v) protein-altering variants within the 500 kb of the 22 index variants identified by re-sequencing AMD case-control study data (targeted re-sequencing of 2,335 AMD cases and 789 controls[10,57] and whole-genome sequencing 60 AMD cases and 60 controls; G. Abecasis and A. Swaroop). The customization further included (vi) the 1,000 top independent (> 2 Mb distant) variants from the previous analysis and additional 100 top variants from each the previous CNV only and the previous GA only analysis, (vii) and 375 variants in ABCA4, including known variants causing Stargardt disease[58], benign variants, and those of unknown significance, as well as 10 known and 44 predicted cysteine mutations in TIMP3, motivated by the known variants causing Sorsby's fundus dystrophy[33,34] (also B. Weber, personal communication).

Annotation

Variant identifiers were based on NCBI dbSNP v137. Chromosomal position and functional annotation of the variant was based on the NCBI Reference Sequence Human Genome Build 19 (RefSeq hg19)[59] and SeattleSeq Annotation 138[60] (see Web Resources). We particularly focus on protein-altering variants including non-synonymous coding variants (missense, stop loss, in-frame insertion/deletion, frameshift, premature stop codon) and splice sites. We converted the description of splice site variants to HGVS nomenclature using Mutalyzer version 2.0.beta-33[61] (see Web Resources).

Genotypes

We genotyped all subjects centrally at the Center for Inherited Diseases Research (CIDR), Johns Hopkins University School of Medicine, Baltimore, MD, USA. From the 569,645 genotyped variants, our quality control excluded poorly genotyped variants as evidenced by genotype call rates < 98.5% (5.8%), deviations from Hardy-Weinberg equilibrium with P < 10–6 (0.34%), variants that mapped at multiple genome locations (0.25%) or variants failing other criteria, resulting in 521,950 (91.6%) variants passing all quality criteria. After excluding monomorphic variants (15.8%), we yielded 264,655 common variants distributed across autosomes, sex chromosomes, and mitochondria, as well as 163,714 directly genotyped protein-altering variants including 8,290 from previously implicated AMD loci (Supplementary Table 3A). For these variants, genotype call rates averaged 99.9% (99.1% for subjects with amplified DNA).

We phased the autosomal and X-chromosomal genotype data using SHAPEIT (200 states, 2.5 Mb windows)[62], then imputed genotypes based on the 1000 Genomes Project[63] reference panel (1000G Phase I, version 3, SHAPEIT2 Reference) using MINIMAC[64] (reference-based 2.5 Mb chunks, 500 kb buffer regions). We then merged study variants that were excluded during imputation (not found in the reference panel) back into the final data set. We excluded common variants (CAF ≥ 1%) with bad imputation quality, R[2] < 0.3, and adopted a more stringent exclusion criterion for rare variants (CAF < 1%), R[2] < 0.8, for the initial identification of lead variants. This yielded a total of 12,023,830 genotyped (439,350) or imputed (11,584,480) quality-controlled variants (Supplementary Table 3A).

Analyzed subjects

Using the genomic information for subject-level quality control, we excluded duplicated and related individuals (kinship coefficient (x003D5) ≥ 0.0884, i.e. 3rd degree relatives or closer)[65], subjects with discrepancies between reported gender and sex chromosomal information or with atypical sex chromosome configurations[66], or subjects with genotyping call rates < 98.5%; we derived ancestry based on the first two principal components using autosomal genotyped variants together with genotype information of the samples from the Human Genome Diversity Project (HGDP)[67]. Our final data set contained 43,566 successfully genotyped unrelated subjects including 16,144 advanced AMD cases and 17,832 controls of European ancestry, 6,657 intermediate AMD cases of European ancestry, and 2,933 subjects (advanced AMD or controls) of Asian or African ancestries (Supplementary Table 3B).

Genomic heritability and genomic correlation

Combined contribution of genotyped variants to disease was evaluated using a variance-component based heritability analysis[68]. This analysis used genotypes to build a similarity matrix, summarizing the overall genetic kinship between each pair of individuals, and then examined the correspondence between genetic and phenotypic similarity. We estimated the explained variance on all genotyped, autosomal variants using restricted maximum likelihood (REML) analysis implemented in GCTA[28] (see Web Resources). We jointly estimated the contributions of rare (MAF in controls < 1 %) and common (MAF in controls ≥ 1%) genotyped variants by first separately calculating their genetic relationship matrices before adding both to the model. Obtained estimates of variance explained were transformed from the observed scale to the liability scale assuming various levels of disease prevalence[68].

We estimated the genomic correlation between disease sub-phenotypes using bivariate REML analyses implemented in GCTA and only included common (MAF in controls ≥ 1%) genotyped variants [29]. We compared 10,749 cases with CNV versus 3,325 cases with GA (excluding the 2,070 cases with mixed CNV and GA) and we compared 6,657 intermediate AMD cases with 16,144 advanced AMD cases. For both analyses, we used the control subjects as reference and avoided shared controls between traits by randomly splitting the 17,832 unrelated European control individuals into two sub-samples of 8,916 individuals.

Genome-wide single variant association analysis

Single-variant association tests analyzing the 16,144 advanced AMD cases and 17,832 controls of European ancestry were based on the Firth bias-corrected likelihood ratio test[69], which is recommended for genetic association studies that include rare variants[70], as implemented in EPACTS (see Web Resources). Analyses were adjusted for two principal components and source of DNA (whole-blood or whole-genome amplified DNA). Allele dosages of the imputed data were utilized, Sensitivity analyses were conducted to evaluate the influence of alternative association tests, alternative covariate adjustment including age or sex, or up to 10 principal components instead of two, as well as the influence of restricting to population-based controls, or to controls aged 50 years or older. Genomic control correction[71] was used to account for potential population stratification using all genotyped variants with minor allele count ≥ 20 outside of 20 previously described AMD loci[6,9]. As usual for genome-wide association studies, we considered P-values ≤ 5 × 10–8 as genome-wide significant.

To identify independently associated variants, we adopted a sequential forward selection approach: We first computed single variant association for each of the > 12 million variants. Then we selected the variant with the smallest P-value and its flanking ±5 Mb region, repeating the process until no genome-wide significant variant (P ≤ 5 × 10–8) was left yielding a number of 10 Mb regions. Within each of these large regions, we re-analyzed each variant conditioning on the top variant, and repeated this process by adding the previously identified genome-wide significant variant(s) within the respective 10 Mb region. This yielded one or more independently associated genome-wide significant variant(s) per 10 Mb region.

A locus region was defined by a genome-wide significant variant and its correlated variants (r[2]≥ 0.5) ± 500kb; overlapping locus regions were merged to one locus, so some loci contained more than one index variant (details in Supplementary Figure 3).

In order to derive independent effect sizes (log odds ratios) for all identified variants, we computed a fully conditioned logistic regression model including all identified variants.

Bayesian approach to prioritize variants

n order to summarize the statistical evidence of a variant for its association strength, we computed the Bayes factor for each variant, which is a measure of the strength of the association that is comparable irrespective of variant frequency or study sample size. It provides the probability of the genotype configuration at a variant (in cases and controls) under the alternative hypothesis (association) divided by the probability of the genotype configuration under the null hypothesis (no association). It is computed using the association results per variant [72]. The posterior probability of each variant is then computed as the Bayes factor relative to the sum of all variants' Bayes factors across one locus region and can be thought of as the relative strength of evidence in favor of each SNP studied in the respective region. This assumes that there is one causal variant per region and that the causal variant is in the analyzed data set.

Expanding to loci with multiple association signals and thus a single alleged causal variant per signal, we used the association results per SNP obtained by conditioning on the other independent variants at that locus for computing the Bayes factor.

We derived 95% credible sets of variants per signal, which is the minimal set of variants, for which the sum of the posterior probabilities accumulates beyond 95%. This approach was recommended for fine-mapping of association signals and for prioritizing variants[73]. Assuming that there is only one causal variant in an association signal and that the causal variant is contained among the analyzed variants, such a credible set of variants contains the causal variant with 95% probability.

We annotated functionality of the variants in each of the 95% credible sets (see above).

Gene-based burden analysis

Single variant analyses have limited power to depict rare variants with association. Gene-based burden tests evaluating accumulated association from multiple rare variants per gene have been shown to complement such analyses and improve power to detect a burden of disease. We computed the burden of disease using the variable threshold test[51] as implemented in EPACTS. These analysis assume that all variants in a gene either increase or decrease disease risk. When variants with opposite directions of effect reside in the same gene, power will be reduced. An analysis with SKAT and SKAT-O, which both allow for variants with opposite directions of effect to reside in the same gene, did not identify additional signals (data not shown).

We focused this analysis on protein-altering variants, since we assumed that the other (not protein-altering) variants would outnumber these predicted deleterious variants by far and would thus dilute a disease burden from the deleterious variants. Assuming a negative selection against such deleterious variants that cause their frequency to be low across ancestries, we restricted our rare variant definition to variants with MAF < 1% (cases and controls combined) in each of our ancestry groups (African, Asian, and European). We utilized the genotypes of these rare protein-altering variants if genotyped directly, or rounded imputed allele dosages to the next best genotype if imputed; imputed variants were restricted to those of highest imputation quality (RSQ >= 0.8).

We assessed statistical significance by adaptive permutation testing with variable thresholds (up to 100 million permutations; minimal P-value = 1 × 10–8)[51]. When rare variants appear on a haplotype associated with disease through a common variant allele already identified for AMD, the rare variant burden would depict a mere shadow of the already identified variant. Therefore, we repeated the variable threshold test conditioned on the variant(s) identified in the respective locus by single variant analysis (locus-wide conditioning), to unravel a gene-based burden of rare variants independent of risk variants identified in single variants tests.

First, we searched for rare variant disease burden genome-wide applying a genome-wide Bonferroni-corrected significance threshold of 0.05 / 17,044 = 2.9 × 10–6 (17,044 genes genome-wide with at least 1 variant included in the analysis, i.e. with ≥ 1 rare protein-altering variant). In a second view on this, we focused on our 34 identified AMD loci and here applied a significance threshold based on the 703 genes overlapping with the locus regions (P < 0.05 / 703 = 7.1 × 10–5). Odds ratio estimates of the burden were derived by logistic regression using the Wald test on the collapsed burden. As there was an overlap of the sequenced subjects with the chip data subjects, we conducted a sensitivity analysis for the burden test excluding overlapping subjects (see Supplementary Note 8).

Follow-up queries for genes underneath the association signals

In order to derive information for all genes underneath our 52 identified association signals (spread across the 34 AMD loci), we built a gene list containing all genes that overlapped with a more narrow definition of locus regions: We have been using a particularly comprehensive definition of the locus region during the signal identification step (index variants and proxies, r[2] ≥ 0.5, ±500kb), to avoid far-reaching linkage disequilibrium that may generate shadow signals (particularly in the light of strong associations in the CFH, C3, C2/CFI, and ARMS2/HTRA1 loci) and to optimally differentiate independent signals within locus. We have also used this wide locus region definition for the rare variant burden test again to fully correct for independent signals in the respective wider locus regions and to be conservative in the multiple testing corrections for the AMD-locus-wide burden test search. However, this wide definition is less adequate when prioritizing genes around the identified signals under the assumption that most protein-altering or regulating variants exert their effects in cis[42]. We thus focused the gene list for further queries to a more narrow locus region definition (index variants and proxies, r[2] ≥ 0.5, ± 100kb) and yield 368 overlapping RefSeq genes (Supplementary File 5).

Gene expression

For the 368 genes in our gene list (see above), we sought to obtain gene expression in relevant tissues, retina, RPE, and choroid, in two independent data sets (see details in Supplementary Note 9). A consensus rating of gene expression observed in the two labs was derived as follows: Expression of a gene in one set of tissues (retina or RPE/choroid) was inferred, if both labs detected expression in the respective set of tissues; if at least one of the labs did not observe expression, the gene was considered as not expressed; gene expression of all other genes (one lab observing expression and the other with missing, or both labs with missing data) was regarded as missing.

Mouse model phenotypes

For the 368 genes in our gene list, we queried the Mouse Genome Informatics (MGI)[74] and the International Mouse Phenotyping Consortium (IPMC)[75] data bases (see Web Resources), and manually curated results by information from published literature. We determined whether a gene exhibited a relevant eye-phenotype (i.e. retina, RPE, or choroid phenotypes) in established genetic mouse models (knock-out, knock-in, or trans-genic mice).

Enrichment for molecular pathways

For the 368 overlapping genes, we performed functional enrichment analysis using INRICH[76] with default settings unless stated otherwise (see Supplementary Note 10). Target intervals of this analysis were the narrow AMD locus regions (index variants and proxies, r[2] ≥ 0.5, ± 100kb, Supplementary Table 5). Since there is no consensus approach to pathway analysis, we queried multiple data bases: (i) Kyoto Encyclopedia of Genes and Genomes (KEGG)[77], (ii) Reactome[78], and (iii) Gene Ontology (GO) Consortium[79] (see Web Resources). For example, while KEGG is a manually curated database on metabolic pathways, GO also includes automatic annotations and more comprehensive set of cellular processes and molecular functions.

Drug pathways and targets

In order to derive information on whether the product of a gene among the 368 genes in our gene list was a direct drug target, we searched the DrugBank database (Version 4.1) which contains 4,207 drug targets (= genes) and 7,740 drugs [80](see Web Resources).

Explained variability in disease liability

Based on the 52 identified AMD variants, we estimated the explained proportion of disease liability explained by these variants (see Web Resources)[81] using the log Odds Ratio estimates from the model including all 52 identified variants (fully conditioned) to derive independent effect sizes. We compared this proportion explained by the 52 variants with the earlier derived genomic heritability based on all genotyped variants (see above).

Genetic risk score and relative and absolute genetic risk of AMD

For each individual, we computed a genetic risk score (GRS) as the effect size weighted sum of the AMD risk increasing alleles for all 52 independent variants divided by the sum of all effect sizes. To derive a a realistic genetic risk score distribution, we modeled a general population based on our case-control data, which requires an assumption on the prevalence of advanced AMD(see Supplementary Note 11). For this modeled general population, we derived the GRS distribution and its deciles. For the weighting, the log Odds Ratios for each of the 52 variants were derived from the fully adjusted model (including all 52 variants) to assure independence of effect sizes.

We derived relative risk estimates (as Odds Ratios) for each GRS decile with the first decile as reference. This relative risk estimate is independent of the prevalence except that the decile to form the genetic risk groups used the GRS distribution as expected in a general population (which requires a prevalence assumption). We also computed absolute risk estimates per GRS decile as the proportion of advanced AMD cases applying the weights and prevalence assumptions as described above.