Age-Related Maculopathy: A Genomewide Scan with Continued Evidence of Susceptibility Loci within the 1q31, 10q26, and 17q25 Regions

Families

The families in the University of Pittsburgh cohort were ascertained from a pool of 35,000 individuals, presumed to be affected by ARM, by a combination of approaches, including recruitment of families from within the retina service of the Department of Ophthalmology at the University of Pittsburgh and mailings sent to patients with ARM (on the basis of billing ICD-9 codes) from multiple retina and ophthalmic practices across the United States (Weeks et al. 2000, 2001). Patients who were interested in participating in the study forwarded a self-addressed postcard to the investigators so that they could be contacted by phone and mail. Through the initial family members, other family members were informed of the study and encouraged to participate. The informed consent process was approved by the institutional review board of the University of Pittsburgh. The families with ARM in the Duke University and Vanderbilt University cohorts were selected for the disease in a similar fashion to that used by Weeks et al. (2000, 2001). Probands were identified from the retina services of each institution, and family members were brought to the examination centers for standardized clinical evaluations and fundus photographs. Families were genotyped in three different sets (see below for details); table 2 shows the number of families included at each of the three genotyping stages.

Documentation of the retinal features and other ocular pathology were obtained from direct clinical examinations (of local subjects when possible), eye care records (from all eye care providers), and fundus photographs. Every effort was made to document the presence or absence of drusen; the type, size, number, and confluence of drusen; the presence of pigment epithelial figures; choroidal neovascular membranes; geographic atrophy; pigment epithelial detachments (both serous and drusenoid); pigment epithelial tears; and laser treatments (for choroidal neovascular membranes). Otherocular pathology and surgery, including potential confounding diagnoses, were also recorded.

Evaluations

Clinical records were evaluated separately from fundus photographs by a single investigator (M.B.G.), and evaluations were done in separate sessions. Prior to review, each fundus photograph was masked with a unique series of letters and numbers by the clinical coordinating staff to conceal the identity of the individual. As original eye records, which often cannot be masked, were obtained, these were graded separately and then were compared with the photographs to reconcile discrepancies and determine a final grade. If the documentation was insufficient to clearly characterize the individual’s condition, additional records and/or photographs were sought. In the cases from Vanderbilt and Duke, the clinical records and fundus photographs were reviewed by a single investigator (M.B.G.) on separate days and then were reconciled. With only a few exceptions, the individual grading classifications (affected, probably affected, and possibly affected or unknown) were the same for clinical investigators from Duke (E.A.P.) and Vanderbilt (A.A.) and for M.B.G. Among the clinical investigators, there were no discrepancies in the diagnoses of ARM. Of the total 1,616 individuals submitted for genotyping in any or all of the three genomewide scans, 1,177 individuals had photographs and 1,560 individuals had records. There were 415 individuals who had only records available and 46 who had only photographs available.

Genotyping Methods

Genomic DNA was extracted from leukocytes obtained from whole blood collected into EDTA vacutainer tubes. The samples from the University of Pittsburgh were extracted with the use of a simple salting-out procedure, as described elsewhere (Miller et al. 1988), and the samples from our collaborators at Duke and Vanderbilt Universities were extracted with the use of the Puregene DNA purification method (Gentra Systems). The samples were diluted in 1 times TE buffer and sent as coded aliquots either to the National Heart, Lung, and Blood Institute (NHLBI) Mammalian Genotyping Service (MGS) or to the Center for Inherited Disease Research (CIDR) (supported by the National Institutes of Health [NIH], including the National Eye Institute). Both genotyping centers use primarily tri- and tetranucleotide repeat polymorphisms, but each institution uses their own genotyping technology, customized primers, and controls (see NHLBI MGS and CIDR Web sites). Genotyping was done without any knowledge of diagnostic status, but family structure information was used to help detect and resolve genotyping errors.

Classification of Subjects for Linkage Analyses

There are two fundamental questions with regard to the status of each participant in a genetic study of ARM: (1) How confident can one be that an individual isaffected with a macular degeneration? (i.e., diseaseseverity status) and (2) Given an affected individual, how confident can one be of the diagnosis? (i.e., diagnosis status). Three classifications were used for the disease-severity status: definitely affected, probably affected, and possibly affected or unaffected. For the diagnosis status, the following four classifications were used: ARM, ARM or another condition, another condition that is not ARM, and no evidence of ARM (from either photographs or records). Each subject was classified for both disease-severity status and diagnosis status. After all members of a given family had been evaluated, the entire family was reviewed for consistency of the diagnostic criteria and was assessed as to final eligibility for the genomewide genotyping efforts.

We established three clinical grades of ARM for genetic analyses. Under the most stringent model (model A), individuals were classified as “affected” only if they were clearly affected with ARM on the basis of extensive/coalescent drusen, pigmentary changes (including pigment epithelial detachments), and/or the presence of endstage disease (geographic atrophy and/or choroidal neovascular membranes). Model B analyses classified as “affected” all those considered affected under model A and those who were considered to be probably affected with ARM, on the basis of moderate to extensive soft drusen, extensive hard drusen with pigmentary changes, and suspected pigment epithelial detachments. Model C classified as “affected” all of those considered affected under models A and B and those whose diagnosis of ARM was less clear. Under model C, individuals were considered affected if they were definitely and probably affected with ARM or with a related maculopathy (having insufficient evidence to rule out another type ofmacular disease). Model C also included individuals with endstage disease (choroidal neovascular membranes, disciform scarring, and/or geographic atrophy), in the absence of any other documentation of macular pathology. These individuals were considered to be definitely affected, but the determination of whether the etiology was ARM or another maculopathy was considered ambiguous. Our diagnostic models were nested; for example, individuals considered affected under modes B and C, but not under model A, were coded as “unknowns” in the linkage analyses performed with model A diagnoses.

Families eligible for genomewide scans were required to have at least two affected individuals. Efforts were also made to enroll an unaffected informative family member for comparison. The average size of the families (including affected, unknown, or unaffected individuals) ascertained at the Pittsburgh, Duke, and Vanderbilt sites was 2.58 people. Of 628 families (totaling 1,616 individuals) submitted for analysis, 1,159 individuals were definitely affected with ARM and 261 were probably affected with ARM. Thus, 75.69% of affected individuals were included within model A, and 92.17% of all individuals were treated as affected in any of the models (A, B, or C).

Statistical Analyses: Error Checking

Our genotyping data were subjected to several quality checks in order to help ensure accurate genotyping. Genotypes at each locus were checked for Mendelian inconsistencies with the use of our program, PedCheck (O’Connell and Weeks 1998), which uses genotype elimination to identify the more subtle inconsistencies. We compared the number of homozygotes observed with the number of homozygotes expected on the basis of the estimated allele frequencies; a disparity in these numbers may indicate the misnaming of heterozygous genotypes as homozygous. We also examined the genotypes of several putative MZ twin pairs to assay the quality of our genotyping data.

We used the relationship-estimation program PREST (McPeek and Sun 2000) to check the accuracy of our assumed relationships between individuals against the realized level of allele sharing across the whole genome. This led to the omission of four small families, in which relationship errors could not be resolved, from our current set of families.

Allele Frequency Estimation

We estimated allele frequencies at each locus by simple allele counting in all individuals (disregarding relationships), within each data set (MGS, CIDR1, and CIDR2) separately. This approach generates unbiased estimates when applied to randomly ascertained families (Broman 2001). However, our allele-frequency estimates can be biased by the fact that the vast majority of our genotypes occur in affected individuals. Thus, if there were a strong association of a particular allele occurring more often in affected individuals than in the general population, the frequency of that allele would be overestimated. This potential effect is conservative, in that it would bias us away from linkage (Ott 1992; Göring and Terwilliger 2000). We found elsewhere (Weeks et al. 2000) that the use of more precise allele-frequency estimates that take relationships into account (obtaining these estimates is computationally time consuming) generated essentially identical results to those obtained by our simpler (and much faster) estimation approach.

We used data set–specific marker-allele frequencies in all our analyses, because the genotyping centers used different technologies that resulted in slightly different allele labels at many loci. Additionally, the technology used by CIDR to genotype the CIDR2 data set had changed since they first performed genotyping for the CIDR1 data set. Rather than laboriously attempting to establish the exact correspondence between every allele in each of the three data sets, we took a simplerapproach, employing data set–specific marker-allelefrequencies in all our analyses, especially because the capabilities of Allegro (Gudbjartsson et al. 2000) had been extended graciously (at our request) to handle correctly the data-set–specific allele frequencies.

Linkage Analysis

We performed the following analyses:

• 1.

  Single-point and multipoint LOD scores under heterogeneity: Since many of our pedigrees show inheritance patterns that seem to be consistent with dominant inheritance, we chose, a priori, to compute LOD scores under a single simple dominant model (disease allele frequency = 0.0001; penetrance vector = [0.01 0.90 0.90]) and to allow for heterogeneity under an admixture model (Hodge et al. 1983). Only two disease phenotypes were used: “affected” and “unknown”; no one was given a “normal” phenotype, owing to the complexities of the ARM phenotype. The single-point and multipoint LOD scores were computed by Allegro(Gudbjartsson et al. 2000) with data set–specific marker allele frequencies (as described above). Note that the majority of our families are so small that the family-specific LOD score maximizes at either θ=0 or θ=.50. With a desire to keep the number of models used in the analyses as low as possible, we elected not to employ age-dependent penetrance models in our study.

• 2.

  Single-marker and multipoint “model-free” methods: We computed single-point and multipoint “model free” LOD scores, using the S_all statistic under the linear model with Allegro (Gudbjartsson et al. 2000). Data set–specific marker-allele frequencies were used.

• 3.

  Power to detect linkage: Although power estimates for complex diseases may not be very helpful, we presented simulations elsewhere (Weeks et al. 2000) that indicated that, if the genetic effect is large enough, we will have high power to detect linkage. However, as might be expected, the power depends strongly on the percentage of families linked to the region of interest.

• 4.

  False-positive rates: We simulated genomewide autosomal marker data, using data set–specific allele frequencies, on all the pedigrees for each of the diagnostic models. To do this, we used Allegro to simulate marker genotypes for exactly the same people genotyped and the same disease phenotypes as in the original real data, under the assumption that the set of markers was segregating independently of ARM disease status. We then analyzed each simulated replicate, using Allegro in exactly the same manner in which we had analyzed the real data, and recorded the results. We used 1,000 replicates to compute the empirical probability of observing a LOD score greater than or equal to a given threshold anywhere in the genome.

• 5.

  Ordered subsets analysis (OSA): To explore the relationship of relevant covariates to genetic heterogeneity, we applied OSAs, which use quantitative covariates to identify potentially more homogeneous subgroups from the overall sample (Hauser and Boehnke 1998; Watanabe et al. 1999; Ghosh et al. 2000; Hauser et al. 2001, 2004). In the OSA approach, families are first ranked by an appropriate family-level covariate. Then, families are added into the calculations one at a time in rank order, from one extreme to the other. After each family is added, a maximum LOD score (MLS) is computed for the current contiguous subset of the families; here we use the nonparametric LOD score based on the S_all sharing measure, with data set–specific allele frequencies. The OSA-LOD statistic is the largest MLS of all ordered subsets of the families. Three different OSA-LOD statistics are obtained—by adding families from lowest to highest rank, by adding families from highest to lowest rank, and by finding the best contiguous subset (“optimal slice”) with the largest MLS.

We tested whether the best covariate-defined subset yields a significant increase in the LOD score, as compared with what one would obtain with randomly ordered families. In other words, the P value measures how often an OSA-LOD greater than or equal to the observed OSA-LOD would occur in a subset of the randomly ranked families, on a chromosomewide basis, given our family-level LOD scores. Other simulation studies have shown that this approach has an appropriate type I error rate and can lead to increased power in the presence of heterogeneity (Hauser et al. 2001).

The family-level covariates we used were (1) mean age at onset, (2) mean pack-years, (3) proportion of affecteds with at least one APOE-2 allele, and (4) proportion of affecteds with at least one APOE-4 allele. Age at onset is the earliest age, as reported by the participant, either at which symptoms were attributed to ARM or at diagnosis. If this is not available, then it is the age at the earliest records/photos. Pack-years are self reported and are defined as the product of years and packs, where “packs” is the number of cigarette packs smoked per day, with 20 cigarettes per pack. Nonsmokers were assigned a pack-years value of zero, and cigar and pipe smokers were assigned missing values. We computed the family-specific values, using only the covariate values of the affected members in each pedigree. Therefore, it is possible for a family to have different covariate values under the different diagnostic models if who is affected differs across diagnostic models. Also, note that the OSA program only includes families with known covariate values, so the number of families used varies across covariates and diagnostic models.

Third Scan as Independent Confirmation

If one treats the third genome scan as an independent replication set, then we can directly address the question of whether a peak implicated in our earlier work is replicated in our third genome scan. However, because some of the families used in our earlier studies were expanded and regenotyped for the third genome scan, this required that we recompute our results on the first two genome scans after removing these families. This did not alter our earlier conclusions, based on the combined analyses of the first two genome scans (Weeks et al. 2001), that four regions (1q31, 9p13, 10q26, and 17q25) all show multipoint HLOD or S_all scores of ⩾2.0 (under at least one model). However, when we examine the results on the third genomewide screen, only two (10q26 and 17q25) of these four regions appear to have been replicated with LOD scores >1.0. In the 10q26 region, we obtain an S_all of 1.50 at D10S212 under model B (and an S_all of 1.20 and 1.21 under models A and C, respectively, at or near D10S212). In the 17q25 region, we obtained an HLOD of 1.20 at D17S784 under model B (whereas the HLOD reaches a maximum of 0.85 at D17S784 under model A). When interpreting these results, it is important to remember that genome scans on small data sets can show remarkable variability in both evidence of linkage and location estimates (Cordell 2001).

Pooled Analyses

We computed both single-point and multipoint linkage statistics (HLOD and S_all) on our combined set of families; summary descriptions of the numbers of families and numbers of affected relative pairs are provided in table 2. Table 3 displays those markers that had elevated single-point or multipoint statistics ⩾1.5, and figure 1 presents the multipoint curves for chromosomes 1, 2, 10, and 17. On chromosome 1, the 1q31 region had a maximum multipoint HLOD of 2.72 at D1S1660 under model C; the corresponding genomewide empirical P=.061. Elevated HLOD peaks were also seen under models A and B (fig. 1). On chromosome 2q14.3, under model A, there is a significant single-point S_allof 3.23 (marginal single-point empirical P=.0) at D2S1328 (table 2) and a multipoint S_all of 2.42 (genomewide empirical P=.12) at the same location (fig. 1). Under model B, the single-point S_all at D2S1328 is 2.18, but the multipoint curve reaches a maximum of only 1.35, whereas there is no linkage signal under model C. On chromosome 10q26, a multipoint S_all peak is seen consistently across all three models, with a maximum S_all of 2.69 (genomewide empirical P=.06) at D10S1230 under model C. The 17q25 region contains our highest LOD score, with elevated HLODs across all three models and a maximum multipoint HLOD of 3.53 between D17S784 and D17S928 under model B (fig. 1). The genomewide empirical P value for an HLOD of 3.53 is .007.

On the basis of our simulation study, the genomewide empirical P value for a multipoint LOD score (either S_all or HLOD) of 3.0 is in the range of .03 to .04. For a multipoint LOD score of 2.5, the genomewide empirical P value is in the range of .09 to .11. Note that HLOD and S_all have approximately the same null distribution, since both involve estimating two free parameters—for HLOD, one maximizes across location and α (where α is the fraction of linked families); for S_all, one maximizes across location and δ (see Kong and Cox [1997] for a definition of δ). The genomewide empirical P value measures the chance of seeing a LOD score greater than or equal to the threshold anywhere in the genome and, thus, is larger than the marginal P value evaluated at a single point in the genome.

Exploration of Covariate Effects

The formation of subsets on the basis of relevant covariates with the OSA approach, although exploratory in nature, may result in more genetically homogenous subsets. Tables 4, 5, 6, and 7 present our results after we formed subsets on the basis of (1) mean age at onset, (2) mean pack-years, (3) proportion of affecteds with at least one APOE-2 allele, and (4) proportion of affecteds with at least one APOE-4 allele. With regard to the regions on chromosomes 1, 2, 10, and 17 implicated by the conventional linkage analyses described above, only the 10q26 region has interesting OSA-LODs (table 5). Under both models A and B, the OSAs reveal a subgroup of ∼160 highest smoking (mean pack-years ⩾26) families, with elevated OSA-LODs of 3.87 and 4.25, respectively.

With the use of mean pack-years, the optimal slice method generates large OSA-LODs between 4.63 and 5.66, under all three models, on chromosome 20 near D20S478 (table 5). Under models B and C, the permutation-based P value is quite small, and the optimal slice includes families with ∼27–43 mean pack-years.

When we stratify on the proportion of affecteds with at least one APOE-2 allele or on the proportion with at least one APOE-4 allele, only modest OSA-LODs are observed (tables 6 and 7). For APOE-4, however, we do observe, across all three diagnostic models, elevated OSA-LODs in the same region on chromosome 7 (table 7).