Evidence for Association between Multiple Complement

Genetic Epidemiology 31: 224–237 (2007)

Evidence for Association between Multiple ComplementPathway Genes and AMD

Valentin Dinu,1� Perry L. Miller,2–4 and Hongyu Zhao5,6

1Program in Computational Biology and Bioinformatics, Yale University, New Haven, Connecticut2Department of Molecular, Cellular, and Developmental Biology, Yale University, New Haven, Connecticut

3Center for Medical Informatics, Yale University, New Haven, Connecticut4Department of Anesthesiology, Yale University, New Haven, Connecticut

5Department of Epidemiology and Public Health, Yale University, New Haven, Connecticut6Department of Genetics, Yale University, New Haven, Connecticut

In this paper we explore the use of biological knowledge to supplement statistical analysis in identifying genes associatedwith disease. It has been previously found that the 402H variant in complement factor H (CFH) is associated with risk fordeveloping age related macular degeneration (AMD). By focusing on the single nucleotide polymorphisms (SNPs) in thecomplement pathway, we were able to use the genotype data from a recently published AMD genome wide associationstudy to identify two additional genes, C7 and MBL2, as potentially associated with subtypes of AMD. Two SNPs situatedin introns of C7 and MBL2 could help differentiate between two forms of AMD: wet (more severe form of AMD) and dry(milder form of AMD). We identified a C7 haplotype associated with protection against developing wet AMD amongindividuals with homozygous CFH risk allele 402H (p-value 0.001 for wet AMD versus dry AMD, odds ratio (OR) 0.16, OR95% CI 0.05–0.49) as well as among individuals with at least one CFH risk allele (p-value 0.007 for wet AMD versus dryAMD, OR 0.35, OR 95% CI 0.16–0.77). The fact that the statistical scores for the C7 and MBL2 SNPs were significant (lowfalse discovery rate) at the pathway level, but not significant at the genome level suggests that focusing at the pathway levelcan be beneficial for identifying SNP signals that would be lost at the genome-wide level. Genet. Epidemiol. 31: 224–237, 2007.r 2007 Wiley-Liss, Inc.

Key words: age related macular degeneration (AMD); complement pathway; false discovery rate (FDR); pathway-baseddisease association; single nucleotide polymorphisms (SNP)

Contract grant sponsor: National Library of Medicine; Contract grant numbers: P20 LM07253 and T15 LM07056; Contract grant sponsor:NIGMS; Contract grant numbers: GM59507; Contract grant sponsor: NIH; Contract grant numbers: U24 NS051869.�Correspondence to: Valentin Dinu, Center for Medical Informatics, Yale University School of Medicine, PO Box 208009, New Haven, CT06520-8009. E-mail: [email protected] 15 September 2006; Accepted 28 November 2006Published online 31 January 2007 in Wiley InterScience (www.interscience.wiley.com).DOI: 10.1002/gepi.20204

INTRODUCTION

Age related macular degeneration (AMD)affects the macula, the region of the retinainvolved in central vision. AMD is the leadingcause of blindness in the developed world. It isbelieved that more than 30% of persons over75 years of age have their vision affected by someform of AMD, and about 6% of the persons in thisage group are affected by the most advancedforms of AMD [Friedman et al., 2004]. In theUnited States, for example, about 10 millionpersons have their vision affected by AMD andabout 1.75 million have the advanced forms ofAMD [Friedman et al., 2004]. The early and

intermediate stages of AMD are characterized bythe deposition of drusen (debris-like material) inthe macula. The most advanced stages of AMDare termed geographic atrophy (dry AMD) andchoroidal neovascularization (wet AMD). DryAMD is caused by the accumulation of largedrusen, hyper- and hypo-pigmentation of theretinal pigment epithelium (RPE), and atrophy ofthe RPE. Dry AMD can also progress to wet AMD.Wet AMD is present in about 10% of AMD casesbut accounts for about 90% of all blindness fromthe disease. It is characterized by choroidalneovascularization, RPE detachments, and/orfluid or blood leakage into the retina. Wet AMDleads to a rapid deterioration of vision.

r 2007 Wiley-Liss, Inc.

The etiology of AMD is not well understood andthere is no cure for AMD. It is believed that AMDis a complex disease, caused by the interactionof multiple genes and environmental factors. Overthe past decade, linkage and association studieshave identified AMD susceptibility loci on almostall chromosomes, most recently reviewed inHaddad et al. [2006]. Recent evidence points tothe strong association of a complement factor H(CFH) variant, Y402H, with AMD [Edwards et al.,2005; Haines et al., 2005; Klein et al., 2005]. Thisassociation has since been confirmed in manyother studies and is believed to have an AMDpopulation-attributable risk of about 60% [Thak-kinstian et al., 2006]. CFH is the major inhibitor ofthe alternative complement pathway. These find-ings have focused the attention on the role playedby inflammation, and especially by the comple-ment pathway, in the etiology of AMD. Followingthis line of thought, Gold et al. [2006] analyzed theassociation between AMD and complement factorB (BF) and C2, two paralogous complementpathway genes located 500 base pairs (bp) aparton chromosome 6. The authors identified one riskand two protective haplotypes in that genomicregion.

In this paper, we describe an analysis focused onthe complement pathway for its association withAMD, using a single nucleotide polymorphism(SNP) data set previously used in a genome wideAMD association study [Klein et al., 2005]. Wefound that, although the signals from individualgenes (other than CFH) in this pathway were notstrong enough to stand out during the initialgenome-wide screening [Klein et al., 2005], afocused pathway analysis led to the identificationof two additional genes, C7 and MBL2, thatpotentially may play a role in affecting risk, andmore importantly, affecting the chance that anindividual is progressing to wet AMD, the mostsevere form of this disease.

MATERIALS AND METHODS

DATA SET

The data set contains 116,204 genome-wideSNPs genotyped with the Affymetrix 100K GeneChip for a matched case-control study of 146Caucasian individuals, comprising 50 controlsand 96 cases with advanced AMD: 50 patientswith neovascular (wet) AMD and 46 with geo-graphic atrophy of the macula (dry) AMD [Kleinet al., 2005]. An initial analysis of the data set,

identifying a mutation in complement factor H(CFH) on chromosome 1 as strongly associatedwith AMD, was described previously [Klein et al.,2005].

SINGLE MARKER DISEASE ASSOCIATION

For each SNP, we performed Armitage’s Test ofTrend [Armitage, 1955] to investigate associationwith disease status. We used an additive model,with the following multipliers: 0 for homozygousnon-risk alleles, 1 for heterozygous alleles, 2 forhomozygous risk alleles. We chose as risk allelethe one with a higher allele count in the casegroup. For each SNP, we computed the statistic forfour comparisons: control versus case (wet1dry),control versus wet, control versus dry and dryversus wet.

SELECTION OF SNPS IN THE COMPLEMENTPATHWAY

We identified a set of 46 genes involved (orannotated as similar to genes involved) in thecomplement pathway using the complementpathway annotation from KEGG and the annota-tion of genes identified in NCBI human genomebuild 35 version 1. More specifically, we selectedall the genes identified in the KEGG diagram forthe complement pathway (Fig. 1) to form a firstsubset of genes. We then formed a second subsetof genes by retrieving all the genes annotated inthe NCBI human genome whose name or descrip-tion contained the word ‘‘complement’’, and wemanually eliminated those genes that were notrelated to the complement pathway. The union ofthe two subsets consisted of 46 genes associatedwith the complement pathway.

We then selected the genotyped SNPs in closevicinity to each of the 46 complement genes. Fromthe initial set of 116,204 SNPs we identified thoseat most 10,000 bp upstream from the gene start to10,000 bp downstream from the gene end position.A total of 94 SNPs met this criterion.

STRATIFICATION BASED ON CFH RISKALLELE

A non-synonymous coding mutation at SNPrs1061170 (Y402H in the amino-acid sequence) incomplement factor H (CFH) had been previouslyidentified as increasing the risk of AMD severalfold [Klein et al., 2005; Zareparsi et al., 2005]. ThisSNP was not available in the set of 116,204 SNPsinitially genotyped using the Affymetrix 100 KGeneChip [Klein et al., 2005], but was later

225Complement Pathway Association with AMD

Genet. Epidemiol. DOI 10.1002/gepi

separately genotyped for all 146 individuals. Weused the genotype information for the rs1061170SNP to divide the overall sample into two subsets:one subset of 48 individuals homozygous for theCFH risk alleles (CC at the rs1061170 locus), andanother subset of 95 individuals with at least oneCFH allele that was non-risk associated (CT orTT at the rs1061170 locus). Three individuals hadno call at the rs1061170 locus. The trend teststatistic and the corresponding p-values werecalculated for the two subsets separately. For the48 individuals with homozygous CFH risk alleles,the SNPs around C7 on chromosome 5 hada strong ability to differentiate between thedry AMD group and wet AMD group of patients(see results). We used a C7 SNP for the nextstratification.

STRATIFICATION BASED ON CFH AND C7RISK ALLELES

We used the genotype information for thers1061170 SNP (from CFH) and rs2876849 (C7)to divide the overall sample of 146 individualsinto two subsets: one subset of 30 individualswith homozygous CFH and C7 risk alleles (CC atrs1061170 and AA at rs2876849), and anothersubset of 113 individuals with at least one CFHor C7 allele that was non-risk associated (CT orTT at rs1061170; AT or TT at rs2876849). The trendtest statistic and the corresponding p-values werecalculated for the two subsets separately.

C7 HAPLOTYPE ANALYSIS

We downloaded the HapMap genotypes forthe CEU population (90 individuals) for the10 C7 SNPs genotyped using the Affymetrix100 k GeneChip. We used Haploview 3.2 toidentify the haplotypes and their frequencies inthe HapMap data as well as in the AMD studypopulation.

PROBABILITY OF DEVELOPING AMD (DRY ORWET) CONDITIONAL ON JOINT GENOTYPESAT THREE SNPS: RS1061170 SNP (FROM CFH),RS2876849 (FROM C7), AND RS930508 (FROMMBL2)

Bayes’ rule was used to compute the conditionalprobability of developing the dry or wet form ofthe disease, conditioned on the joint genotype atthe three SNPs. To account for the fact that thedata set was enriched for AMD (particularly wetAMD), we used the following AMD data for thegeneral population: among persons over 75 yearsof age, about 30% have AMD; and amongall AMD patients, about 10% have the wet form.For example, the calculation of the probabilitythat someone develops the wet form of AMDconditioned on the joint genotype G at the threeSNPs is: P(wet AMD|G) 5 P(wet AMD and G)/P(G) 5 P(G|wet AMD)�P(wet AMD)/[P(G|wetAMD)�P(wet AMD)1P(G|dry AMD)�P(dryAMD)1P(G|no AMD)�P(no AMD)]. For exam-ple, let G be the joint ‘‘homozygous risk’’ genotype

Fig. 1. Complement pathway adapted from the Kyoto Encyclopedia of Genes and Genomes (KEGG) online database. Highlighted with

a double oval is CFH, identified previously [Klein et al., 2005], and with a single oval the newly identified candidates C7 and MBL2(a.k.a. MBL).

226 Dinu et al.


combination at all three SNPs: {G 5 CC atrs1061170 (CFH), AA at rs2876849 (C7) and CCat rs930508 (MBL2)}. Then, using Table I, P(G|wet AMD) 5 20/48, P(G|dry AMD) 5 1/44, P(G|AMD-free) 5 2/46. Also, according to our assum-ptions about AMD occurrence in individualsover 75, P(no AMD) 5 0.7, P(dry AMD) 5 0.27and P(wet AMD) 5 0.03. As a result, P(wetAMD|G) 5 0.25. 95% confidence intervals (CI)were computed by assessing first the confidenceintervals for the penetrance of each genotype:P(G)71.96 SQRT[P(G)� (1–P(G))/n], n being thenumber of individuals successfully genotyped atall three loci. The lower and upper boundariesfor P(G) were then used to calculate the CI forthe Bayesian probabilities.

STATISTICAL SIGNIFICANCE OFINTERACTION AMONG THREE SNPS

We used a permutation test to estimate thesignificance for the interaction of the three SNPs.For each permutation, we randomly permuted thelabels of the 48 individuals with homozygous‘‘risk’’ alleles at the CFH SNP, randomly distribut-ing them into control, dry AMD and wet AMDgroups. We then computed the trend test statisticfor dry AMD versus wet AMD individuals. Weidentified the SNP with the highest test statistic(equivalent to the C7 SNP in the unpermuted dataset), and used it to select a subset of individualswith homozygous ‘‘risk’’ alleles at this top scoringSNP. Using this subset, we again computed thetrend test statistic for dry AMD versus wet AMDpatients and selected the highest score (equivalentto the MBL2 SNP in the unpermuted data set). Werepeated this procedure 100,000 times, and thencalculated the fraction of times when the valuesobtained through random permutation were

larger than those observed for the C7 and MBL2SNPs.

FALSE DISCOVERY RATE

Since a large number of statistical tests wereperformed (there are m 5 116,204 genotyped SNPson the array, or m 5 94 genotyped SNPs in thevicinity of complement pathway genes), weneeded to adjust the individually calculatedp-values in statistical inference. While the firststudy of this very same data set [Klein et al., 2005]used the stringent Bonferroni adjustment, wechose to use instead the false discovery rate(FDR) in this article for statistical inference. Theaim of the FDR procedure [Benjamini and Hoch-berg, 1995] is to control, at a desired level a (suchas 0.05), the proportion of type I errors (falsepositives) among all significant results. Whenmultiple genes may be involved in diseaseetiology, this procedure may yield higher statis-tical power compared to family-wise error rate(FWER) procedures, such as Bonferroni, whichcontrol at a desired level a the probability ofhaving at least one type I error. The FDR algo-rithm that we used was similar to a previouslydescribed algorithm [Storey and Tibshirani, 2003]and is detailed in the appendix.

RESULTS

AMD ASSOCIATION RESULTS FOR 94COMPLEMENT PATHWAY SNPs USINGALL 146 SUBJECTS

We identified 46 genes involved in the comple-ment pathway (Fig. 1), and then identified 94SNPs (from 25 out of 46 genes) genotyped withthe Affymetrix 100 k GeneChip that werewithin 10,000 bp of a gene’s genomic position.

TABLE I. The number of individuals in each disease category differentiated by genotypes at the CFH, C7 andMBL2 SNPs

Individuals with homozygousCFH SNP risk alleles

Individuals without homozygousCFH SNP risk alleles

rs1061170(CFH)

CC CC CC CC CT or TT CT or TT CT or TT CT or TT

rs2876849(C7)

AA AA AT or TT AT or TT AA AA AT or TT AT or TT

rs930508(MBL2)

CC CG or GG CC CG or GG CC CG or GG CC CG or GG

Control 2 1 2 1 13 11 10 6Dry AMD 1 4 6 4 12 5 11 1Wet AMD 20a 1 1 4 12 4 5 1

aThe first column, in bold, contains a disproportionately large number of individuals with wet AMD.



The remaining 21 genes had no genotyped SNPsin their vicinity.

For each SNP, we computed the Armitage TrendTest statistic [Armitage, 1955] for four comparisons:control versus case (wet AMD1dry AMD), controlversus wet AMD, control versus dry AMD anddry AMD versus wet AMD (Fig. 2). (In addition tothe trend test, we also computed the p-values forallele and genotype Pearson’s w2 test and alleleFisher’s exact test. These results, not shown, werevery similar to those from the trend test.) In additionto the expected large association peak around CFHon chromosome 1, we found potentially diseaseassociated peaks around C7 on chromosome 5, C5on chromosome 9, and MBL2 on chromosome 10.The p-values for these genes were less significantthan those for CFH, with values between 0.01and 0.001. The chance of having three SNPs thatall show statistical evidence at this level froma random sample of 94 SNPs is rather small(p-value 5 0.01, calculated as described in theMaterials and Methods section using a permutationtest with 100,000 permutations of the individuals’labels). This is a conservative estimate of the

statistical significance as many of the SNPs withinthe original 94 SNP set are dependent on each other,reducing the effective number of comparisons[Nyholt, 2004]. For example, if two SNPs arecompletely redundant, then these two SNPs shouldbe considered as one instead of two in an overallmultiple comparison adjustment scheme. Therefore,the presence of dependence among the SNPs makesthe findings at three independent significant lociless likely by chance than when the SNPs areindependent of each other.

The region on chromosome 5p where C7 islocated has been associated with AMD in at leasttwo previous linkage studies [Abecasis et al., 2004;Schick et al., 2003]. The region on chromosome 9qwhere C5 is located has also been linked to AMDby at least three other studies [Abecasis et al.,2004; Iyengar et al., 2004; Majewski et al., 2003].

SAMPLE STRATIFICATION BASED ON CFHGENOTYPE

We then explored whether there are any othercomplement pathway genes, apart from CFH, that

Fig. 2. Statistical significance for all 146 individuals at the 94 complement pathway SNPs. A strong association peak is present around

CFH. Smaller peaks around C7, C5 and MBL2 can be observed. Shown in this figure are the �log 10 of the p-values from the trend test

versus the genomic location.

228 Dinu et al.


might help differentiate between the wet (moresevere) and dry (less severe) forms of AMD, orexplain the presence of AMD in individualswithout the CFH risk allele. To this end, westratified the subjects according to the CFH Y402Hgenotype. We identified the strongest associationwhen dividing the sample into the following twosubsets: one subset of 48 individuals homozygousfor the CFH risk alleles (six controls, 15 dry AMD,27 wet AMD having CC at the rs1061170 locus, i.e.,402H in the CFH amino-acid sequence), andanother subset of 95 individuals with at leastone CFH allele that was non-risk associated (42controls, 30 dry AMD, 23 wet AMD having CTor TT at the rs1061170 locus). Three individuals(two controls, one dry AMD) had no call at thers1061170 locus and were excluded from furtheranalysis. The trend test scores and the correspond-ing p-values were calculated for the two subsetsseparately (Fig. 3). For the 48 individuals withhomozygous CFH risk alleles, the SNPs aroundC7 on chromosome 5 had a strong ability todifferentiate between the dry AMD group andwet AMD group of patients (Fig. 3A and B). Forthe 95 individuals that did not have the homo-zygous CFH risk genotype (Fig. 3C), there were2 peaks–one peak around CFH on chromosome 1(possibly due to the presence of one CFH riskallele in some individuals) and one peak aroundC5 on chromosome 9.

C7 HAPLOTYPE ANALYSIS

The top 2 scoring SNPs (rs2329434 andrs2876849) in the dry AMD versus wet AMDanalysis among the 402H homozygous indivi-duals are situated in an intron of the C7 gene,close to exon 16 of C7, near a region of highconservation. They are 376 bp apart and in stronglinkage disequilibrium (LD). (r2 5 0.95, using theCEU HapMap build 21 data.) In the CEU HapMapdata there are two main haplotypes spanning thislocation: H1 (A at both loci) in 78.3% of indivi-duals and H2 (T at both loci) in 20.8% ofindividuals (Table II). The proportions were verysimilar for the AMD study population: 78.4% H1and 21.6% H2. The H1 and H2 haplotypes were,however, unevenly distributed among the CFH402H homozygotes. H2 was encountered only in11.1% individuals with wet AMD and almost fourtimes that proportion, 43.3%, among the dry AMDpatients. The control subjects in this group alsohad a high percentage of H2, at 33.3% (Table II).The H2 haplotype is associated with protection

against developing wet AMD among individualswith homozygous 402H (p-value 0.001 for wetAMD versus dry AMD, odds ratio (OR) 0.16, OR95% CI 0.05–0.49) as well as among individualswith at least one 402H allele (p-value 0.007 for wetAMD versus dry AMD, OR 0.35, OR 95% CI0.16–0.77). The protective role appears strongestin the individuals with the homozygous 402Hrisk allele.

SAMPLE STRATIFICATION BASED ON CFHAND C7 GENOTYPE

We further stratified the set of 146 individualsinto two subsets, based on the presence of riskalleles at both CFH (rs1061170) and C7 (rs2876849)SNPs. The two subsets were one subset of 30individuals with homozygous CFH and C7 riskalleles (three controls, five dry AMD, 22 wet AMDwith CC at rs1061170 and AA at rs2876849), andanother subset of 113 individuals with at least oneCFH or C7 allele that was non-risk associated (CTor TT at rs1061170; AT or TT at rs2876849). For the30 individuals with homozygous risk alleles atboth CFH and C7 SNPs, the SNPs around MBL2on chromosome 10 had a strong ability todifferentiate between the dry AMD group andwet AMD group of patients (Fig. 4A). For the 113individuals that did not have the homozygousCFH and C7 risk genotypes (Fig. 4B), there wasone peak around CFH on chromosome 1 (possiblydue to the presence of one CFH risk allele in someindividuals).

USE OF FALSE DISCOVERY RATE TO ASSESSSTATISTICAL CONFIDENCE IN THE ABILITYOF SNPs RS2876849 (FROM C7) AND RS930508(FROM MBL2) TO DIFFERENTIATE BETWEENWET AND DRY FORMS OF AMD AMONGPATIENTS WITH CFH RISK ALLELES

It had been previously established using thesame data set and a very stringent statisticaladjustment of the p-values (namely, Bonferronicorrection), that only two SNPs in the vicinity ofCFH had enough statistical evidence to beassociated with AMD. The Bonferroni correction,however, is known to be conservative in thissetting due to LD among the SNPs. In addition,the Bonferroni correction at the genome level doesnot incorporate biological knowledge into accountin statistical inference. By focusing on the comple-ment pathway and performing patient stratifica-tion first on CFH risk alleles and then on putativeCFH and C7 risk alleles, we determined that C7



230 Dinu et al.


Fig. 3. (a) Statistical significance at the 94 complement pathway SNPs for the subset of 48 individuals with CFH 402H homozygous risk

alleles. A peak can be observed around the C7 gene on chromosome 5. (b) Enlargement of (3) for the 10 SNPs around C7 on chromosome

5. (c) Statistical significance at the 94 complement pathway SNPs for the subset of 95 individuals with at least one non-risk CFH 402Y

allele.

TABLE II. The C7 H2 haplotype (T at both rs2329434 and rs2876849) confers protection against wet AMD among AMDpatients with the CFH 402H (AMD risk) genotype

Protection by H2 against wetAMD

rs1061170 (CFH Y402H) genotype C7 haplotype H1 (%) H2 (%) p-value OR 95% CI OR

Two SNP genotypes AA TTCEU HapMap 78.3 20.8AMD study sample 78.4 21.6Category (hap count)

CC (CFH 402H recessive risk model) Controls (12) 66.7 33.3 0.05 0.25 0.06–1.09Dry AMD (30) 56.7 43.3 0.001 0.16 0.05–0.49Wet AMD (54) 88.9 11.1

CC or CT (CFH 402H dominant risk model) Controls (62) 80.6 19.4 0.35 0.66 0.27–1.58Dry AMD (74) 68.9 31.1 0.007 0.35 0.16–0.77Wet AMD (88) 86.4 13.6

CT (402H heterozygous) Controls (50) 84 16 0.84 1.12 0.35–3.59Dry AMD (44) 77.3 22.7 0.58 0.73 0.24–2.25Wet AMD (34) 82.4 17.6

TT (CFH 402Y homozygous, reduced AMD risk) Controls (34) 70.6 29.4 0.39 0.48 0.09–2.60Dry AMD (16) 81.2 18.8 0.89 0.87 0.12–6.21Wet AMD (12) 83.3 16.7

The number of inferred haplotypes for each patient category is given in parenthesis. The p-values (w2 statistic) and odds ratios are calculatedagainst the risk of developing wet AMD. Highlighted with bold characters are statistically significant results.



and MBL2 SNPs could also potentially be asso-ciated with the disease. In particular, the twoSNPs we focused on, rs2876849 (from C7) andrs930508 (from MBL2), had relatively high statis-tical scores when comparing between the dry andwet forms of AMD. We used an FDR algorithmsimilar to one previously proposed [Storey andTibshirani, 2003] to assess the statistical signifi-cance of the observed associations. The detailedsteps employed by the FDR algorithm are pre-sented in the appendix. Briefly, 1,000 randompermutations of patient labels were performedto estimate FDR at various cutoff points (-log10

from the p-values derived from the trend testsfor dry AMD versus wet AMD) at the complementpathway level. The trend test statistic of thers2876849 SNP (C7 intron) was second highestamong the 94 SNPs for the 48 individuals withhomozygous CFH risk alleles. At the pathwaylevel, the significance of the SNP score wasestimated to have a .063 FDR. To assess thebenefit of focusing on the complement-pathwaySNPs only versus using the information for theentire genome, we also estimated the FDR at thegenome-wide level, which was 0.945. Therefore,these results would not be considered significantwithout the pathway focus. Similarly, the statisticfor the rs930508 SNP (from MBL2) was highestamong the 94 SNPs for the 30 individualshomozygous for CFH and C7 risk haplotype. Atthe pathway level, the FDR for this score wasestimated to be .008. The genome-level FDR wasestimated to be 0.995.

We also performed FDR to account for thefact that we focused on only one comparison,dry AMD versus wet AMD, out of fourpossible comparisons: control versus case (wetAMD1dry AMD), control versus wet AMD,control versus dry AMD and dry AMD versuswet AMD. We recorded the fraction of randomiterations when the most significant trend testacross all four comparisons (not only dry versuswet) was higher then the desired cutoff.We computed the FDR for the C7 SNP rs2876829cutoff at 0.25 and for the MBL2 SNP rs930508cutoff at 0.018.

We were able to identify the C7 and MBL2 peakswith statistical significance due to our focusedanalysis of the 94 SNPs related to the complementpathway. Had we chosen instead to use all the116,204 genotyped SNPs, the scores for the twoSNPs would have been ‘‘lost in noise.’’ Thissuggests that focusing at the pathway level canbe beneficial.

C7 AND MBL2 SNPS POTENTIALLY DIFFER-ENTIATE BETWEEN DRY AND WET FORMSOF AMD FOR INDIVIDUALS WITH HIGHAMD RISK CFH ALLELES

We focused on the C7 and MBL2 genes to learnif we could see a genetic predisposition leadingto either wet AMD or dry AMD. We selectedhigh scoring SNPs from C7 (rs2876849) and MBL2(rs930508), as well as the CFH SNP (rs1061170),and created a table to differentiate betweencontrols, dry and wet AMD (Table I). The anno-tations for the three SNPs are listed in Table III.As it can be observed from the first column ofTable I, having the risk genotype at all three SNPsgreatly increases the risk for wet AMD (Po0.0001,Fisher’s exact test). Having the homozygous riskalleles at the CFH locus, but not a homozygousrisk allele at either C7 or MBL2 SNP resulted ina relatively higher risk for dry AMD and smallerrisk for wet AMD (columns 2–4 of Table I). Asexpected, not having the homozygous risk alleleat the CFH locus corresponded to a lower AMDrisk (columns 5–8 of Table I).

Using the data from Table I, we used a logisticregression model to study the differentiationbetween the dry and wet forms of AMD basedon all the possible interactions between the threeloci (CFH, C7 and MBL2). All the individual locias well as the two- and three-way interactionterms had p-values below 0.05: 0.003 (CFH), 0.006(C7) and 0.001 (MBL2) for the individual loci;0.004 (CFH�C7), 0.014 (CFH�MBL2) and 0.002(C7�MBL2) for the two-way interactions; and0.049 (CFH�C7�MBL2) for the three-way inter-action of the three loci.

PROBABILITY OF DEVELOPING AMD (DRY ORWET) CONDITIONAL ON JOINT GENOTYPESAT THREE SNPs: RS1061170 SNP (FROM CFH),RS2876849 (FROM C7), AND RS930508 (FROMMBL2)

On the basis of this data set we proceeded toestimate the AMD risk conferred by the jointgenotype defined by the three SNPs. To accountfor the fact that the data set was enriched for AMD(particularly wet AMD), we used the followinginformation on AMD prevalence in the generalpopulation: among persons over 75 years of age,about 30% have AMD; and among all AMDpatients, about 10% have the wet form. Usingthese data, we estimated the risk of developingAMD by Caucasians over the age of 75 conferredby each genotype combination (Table IV). These

232 Dinu et al.


Fig. 4. (a) Statistical significance at the 94 complement pathway SNPs for the subset of 30 individuals with homozygous risk alleles atboth rs1061170 (CFH non-synonymous coding SNP) and rs2876849 (C7 intron). A peak can be observed around the MBL2 gene on

chromosome 10. (b) Statistical significance at the 94 complement pathway SNPs for the subset of 113 individuals with at least one non-

risk CFH and C7 alleles.


estimations suggest that having the homozygousrisk genotypes at all three SNPs–CFH, C7 andMBL2—results in about 25% chance of developingwet AMD (Table IV, column 1), while theother genotype combinations result in a muchlower chance of wet AMD of only 1–6% (Table IV,columns 2–7).

DISCUSSION

In this paper, we have taken a complementpathway-focused approach to identifying genesassociated with AMD. Although the resultsobtained are statistically significant at the path-way level and intriguing, they require further

validation in other studies. As both SNPs identi-fied in C7 and MBL2 are located in introns of theirrespective genes, a logical next step would beto re-sequence the genomic regions around C7and MBL2 in search of non-synonymous codingSNPs found in strong LD with the identifiedSNPs, as was the case with CFH [Klein et al.,2005], or non-coding regulatory SNPs that could,for example, affect intron splicing.

It is interesting to note that at least two previousAMD linkage studies have identified the 5pregion that contains C7 as associated with AMD.Abecasis et al. [2004] identified a peak in the same5p region that had the highest genome-wide score(LOD 5 2.55, P 5 0.0003) in their multipoint ana-lysis. As the authors noted, their finding was in

TABLE III. Annotation for the three AMD-associated SNPs

Chromosome 1 5 10

Gene symbol CFH C7 MBL2Gene starta 193352832 40945356 54195147Gene enda 193448257 41018798 54201466dbSNP Id rs1061170 rs2876849 rs930508SNP positiona 193390894 41015547 154198304Position w.r.t. genea exon intron intronWet AMD risk genotypeb 1 (CC) 1 (AA) 1 (CC)Genotype 1 (AA) CC AA CCGenotype 2 (AB) CT AT CGGenotype 3 (BB) TT TT GGp(A) – allele A frequency 0.4286c 0.7857d 0.7143d

q(B) – allele B frequency 0.5714c 0.2143d 0.2857d

H(2pq) – heterozygosity 0.4898c 0.3367d 0.4082d

aNCBI human genome build 35 version 1 (May 2004).bFrom Affymetrix for Caucasians.cFrom dbSNP.dGenotype 1 confers extra risk for wet AMD.

TABLE IV. AMD risk conferred by joint genotype across the three SNPs in Caucasian individuals over 75 years of age(95% CI given in parentheses)

rs1061170(CFH)

CC CC CC CC CT or TT CT or TT CT or TT CT or TT

rs2876849(C7)

AA AA AT or TT AT or TT AA AA AT or TT AT or TT

rs930508(MBL2)

CC CG or GG CC CG or GG CC CG or GG CC CG or GG

No AMD 0.62 0.377 0.448 0.36 0.709 0.835 0.683 0.931(0.358–1) (0.208–1) (0.277–1) (0.201–1) (0.559–0.969) (0.626–1) (0.521–0.992) (0.618–1)

Dry AMD 0.125 0.608 0.542 0.581 0.264 0.153 0.303 0.063(0.072–0.471) (0.335–1) (0.335–1) (0.324–1) (0.208–0.361) (0.115–0.229) (0.231–0.44) (0.042–0.127)

Wet AMD 0.255a 0.015 0.009 0.059 0.027 0.012 0.014 0.006(0.147–0.959) (0.009–0.083) (0.006–0.024) (0.033–0.288) (0.021–0.037) (0.009–0.019) (0.011–0.02) (0.004–0.013)

aColumn 1 has a particularly high risk for wet AMD.

234 Dinu et al.


proximity to an AMD linkage region that hadbeen identified previously [Schick et al., 2003].The study performed by Abecasis et al. [2004]identified five regions of modest AMD linkage,located at 1q, 2p, 5p, 9q and 22q, covering a totalregion of roughly 3% of the entire genome. Byfocusing on the genes in the complement pathway,we identified four AMD association peaks (ifwe include the C5 peak on chromosome 9), onchromosomes 1q, 5p and 9q and 10q. Threeof the four peaks we identified matched regionsfrom the study by Abecasis et al. [2004]. Theprobability of such an occurrence just by chanceis low (p-value 5 10�4). These previous findingscould provide possible supporting evidence forour results. Both our data set and the one usedby Abecasis et al. [2004] were generated usingCaucasian subjects and both were enriched in theadvanced stages of the disease. The fact thatmany of the other previous AMD linkage studiesidentified different genomic regions could par-tially be explained by genomic heterogeneity ofthe disease, variation in phenotype measurement(some studies used a more inclusive 1–15, 1–16 or1–6 AMD grading scale), or random association.

As mentioned previously, a recent study identi-fied one risk and two protective haplotypes in theBF/C2 region on chromosome 6 [Gold et al., 2006].The authors found the largest protective effectin the CFH 402H homozygotes. In our study weidentified the same trend for the C7 H2 haplo-type: the largest protective effect was in the 402Hhomozygotes. The individuals at highest risk fordeveloping AMD due to their homozygous CFHrisk genotype who had not developed wet AMDhad a high frequency of the protective alleles atthe C7 locus (33.3% in controls and 43.3% in dryAMD versus 11.1% in wet AMD). The protectionagainst developing wet AMD conferred by the C7H2 haplotype among CFH 402H homozygousAMD patients appears slightly more pronounced(OR 0.16, OR 95% CI 0.05–0.49) than the resultfound for the BF protection (OR 5 0.27) againstAMD development in 402H homozygotes [Goldet al., 2006].

C7 is one of the five components (C5b, C6, C7,C8, C9) that assemble in the final membraneattack complex (MAC) of the complement path-way. The terminal MAC is responsible for cell lysisand is also associated with the release of growthfactors such as vascular endothelial growth factor(VEGF), transforming growth factor b2 (TGF-b2),basic fibroblast growth factor (bFGF) and platelet-derived growth factor (PDGF) [Benzaquen et al.,

1994; Bora et al., 2005]. Recently, Bora andcolleagues found that the activation of thecomplement pathway, specifically of MAC, isessential for the development of laser-inducedchoroidal angiogenesis in mice [Bora et al., 2005].Previous studies have identified C7 mutationsthat are associated with reduced complementhemolytic activity [Barroso et al., 2006; Nishizakaet al., 1996; Schirinzi et al., 2006]. At least one ofthese studies identified a complement deficiencyassociated mutation in exon 16 of C7, which is inproximity to the location of the C7 H2 haplotype[Nishizaka et al., 1996]. Conceivably, the protec-tion against developing wet AMD observed forthe C7 H2 haplotype could be due to a reducedability to form the terminal MAC.

Our analysis of the complement pathway couldhave been more complete with a more compre-hensive set of SNPs. Overall, we found only 94SNPs among the 116,204 genotype SNPs that werein close vicinity to the 46 complement pathwayrelated genes. This means that, on average, abouttwo SNPs were identified in the vicinity of eachgene. Only 25 out of 46 genes, however, hadgenotyped SNPs in their vicinity, while theremaining 21 genes had no SNPs nearby. Sinceno SNPs situated in the vicinity of the BF/C2 genecluster were genotyped using the Affymetrix100K GeneChip, we could not verify the findingsof Gold et al. [2006]. Future genotyping using SNPmicroarrays of higher density (such as theAffymetrix 500k GeneChip or the Illumina Hu-manHap550) or custom arrays may lead to ahigher number of complement pathway SNPs forfurther dissection of this pathway’s relationshipwith AMD.

Our analysis was focused on the complementpathway SNPs that could differentiate betweenthe dry and wet forms of AMD. As a result, wefocused on the C7 and MBL2 SNPs and we did notperform a more in-depth analysis of the associa-tion peak surrounding C5. The C5 SNPs did notshow as strong an association as the C7 and MBL2SNPs when differentiating between the dry andwet forms of AMD. The region on chromosome 9qwhere C5 is located, however, has previously beenlinked to AMD by at least three other studies[Abecasis et al., 2004; Iyengar et al., 2004;Majewski et al., 2003]. Further analysis of theassociation of the C5 locus with AMD should beperformed to understand better the nature of itspossible involvement in the etiology of AMD.

The identification of the two SNPs in C7 andMBL2 suggests the potential benefit of statistical



analysis capitalizing on biological knowledge,such as known pathways. None of the SNPsidentified in this pathway-based analysis (apartfrom the CFH SNPs, previously described [Kleinet al., 2005]), were classified in the top 500 out ofthe 116,204 SNPs genotyped (Table V). A genome-wide analysis would have most likely missedthem, but, by focusing on the pathway itself, wewere able to identify two potentially interestingSNPs located in proximity to complement path-way genes. In addition to biological pathways,other types of data, such as protein–proteininteractions and linkage peaks identified pre-viously, may also help researchers to conductmore focused analysis to increase the likelihoodof elucidating how genetic variants interact witheach other to affect individuals’ susceptibility todiseases.

ACKNOWLEDGMENTS

We thank Josephine Hoh for providing the AMDdata set.

ELECTRONIC DATABASEINFORMATION

See also HapMap: http://www.hapmap.org;KEGG Database: http://www.genome.jp/kegg/pathway/hsa/hsa04610.html; National Eye Insti-

tute website. http://www.nei.nih.gov/health/maculardegen/armd_risk.asp

REFERENCESAbecasis GR, Yashar BM, Zhao Y, Ghiasvand NM, Zareparsi S,

Branham KE, Reddick AC, Trager EH, Yoshida S, Bahling J andothers. 2004. Age-related macular degeneration: a high-resolu-tion genome scan for susceptibility loci in a populationenriched for late-stage disease. Am J Hum Genet 74:482–494.

Armitage P. 1955. Tests for linear trends in proportions andfrequencies. Biometrics 11:375–386.

Barroso S, Rieubland C, Jose Alvarez A, Lopez-Trascasa M, BartPA, Nunez-Roldan A, Sanchez B. 2006. Molecular defects of theC7 gene in two patients with complement C7 deficiency.Immunol 118:257–260.

Benjamini Y, Hochberg Y. 1995. Controlling the false discoveryrate: a practical and powerful approach to multiple testing. J RStatisti Soci Ser B (Method) 57:289–300.

Benzaquen LR, Nicholson-Weller A, Halperin JA. 1994. Terminalcomplement proteins C5b-9 release basic fibroblast growthfactor and platelet-derived growth factor from endothelialcells. J Exp Med 179:985–992.

Bora PS, Sohn JH, Cruz JM, Jha P, Nishihori H, Wang Y,Kaliappan S, Kaplan HJ, Bora NS. 2005. Role of com-plement and complement membrane attack complex in laser-induced choroidal neovascularization. J Immunol 174:491–497.

Edwards AO, Ritter R, Abel KJ, Manning A, Panhuysen C,Farrer LA. 2005. Complement factor H polymorphism andage-related macular degeneration. Science 308:421–424.

Friedman DS, O’Colmain BJ, Munoz B, Tomany SC, McCarty C, deJong PT, Nemesure B, Mitchell P, Kempen J. 2004. Prevalenceof age-related macular degeneration in the United States. ArchOphthalmol 122:564–572.

TABLE V. P-values and Ranks for the following SNPs: rs380390 (CFH intron), rs2876849 (C7 intron) and rs930508 (MBL2intron)

Chromosome 1 5 10

Gene symbol CFH C7 MBL2

rs380390a rs2876849b rs930508b

DBSNP Id P-valuec Rankd P-valuec Rankd P-valuec Rankd

Trend test controlversus both

3.1E-07 1 9.0E-01 96645 2.0E-01 21480

Trend test controlversus dry AMD

2.3E-04 14 2.3E-01 24498 6.0E-01 65105

Trend test controlversus wet AMD

4.1E-07 2 1.5E-01 15675 8.7E-02 9040

Trend test dry AMDversus wet AMD

9.0E-02 9464 1.1E-02 977 2.6E-01 28063

aThe SNP situated in the CFH intron (and in high linkage disequilibrium with rs1061170) had Bonferroni-adjusted significant p-values intwo out of four categories.bThe P-values for the C7 and MBL2 SNPs are not ranked high at the whole genome level and would have been missed without focusingfirst on the complement pathway.cFrom trend test.dOut of 116,204 SNPs genotyped on the Affymetrix 100K chip.

236 Dinu et al.


Gold B, Merriam JE, Zernant J, Hancox LS, Taiber AJ, Gehrs K,Cramer K, Neel J, Bergeron J, Barile GR et al. 2006. Variation infactor B (BF) and complement component 2 (C2) genes isassociated with age-related macular degeneration. Nat Genet38:458–462.

Haddad S, Chen CA, Santangelo SL, Seddon JM. 2006. Thegenetics of age-related macular degeneration: a review ofprogress to date. Surv Ophthalmol 51:316–363.

Haines JL, Hauser MA, Schmidt S, Scott WK, Olson LM, Gallins P,Spencer KL, Kwan SY, Noureddine M, Gilbert JR et al. 2005.Complement factor H variant increases the risk of age-relatedmacular degeneration. Science 308:419–421.

Iyengar SK, Song D, Klein BE, Klein R, Schick JH, Humphrey J,Millard C, Liptak R, Russo K, Jun G et al. 2004. Dissection ofgenomewide-scan data in extended families reveals a majorlocus and oligogenic susceptibility for age-related maculardegeneration. Am J Hum Genet 74:20–39.

Klein RJ, Zeiss C, Chew EY, Tsai JY, Sackler RS, Haynes C,Henning AK, Sangiovanni JP, Mane SM, Mayne ST and others.2005. Complement factor H polymorphism in age-relatedmacular degeneration. Science 308:385–389.

Majewski J, Schultz DW, Weleber RG, Schain MB, Edwards AO,Matise TC, Acott TS, Ott J, Klein ML. 2003. Age-relatedmacular degeneration—a genome scan in extended families.Am J Hum Genet 73:540–550.

Nishizaka H, Horiuchi T, Zhu ZB, Fukumori Y, Volanakis JE. 1996.Genetic bases of human complement C7 deficiency. J Immunol157:4239–4243.

Nyholt DR. 2004. A simple correction for multiple testingfor single-nucleotide polymorphisms in linkage disequili-brium with each other. Am J Hum Genet 74:765–769.

Schick JH, Iyengar SK, Klein BE, Klein R, Reading K, Liptak R,Millard C, Lee KE, Tomany SC, Moore EL and others. 2003. Awhole-genome screen of a quantitative trait of age-relatedmaculopathy in sibships from the Beaver Dam Eye Study. Am JHum Genet 72:1412–1424.

Schirinzi R, Lantin JP, Fremeaux-Bacchi V, Schifferli JA, Trende-lenburg M. 2006. [Combined-heterozygous deficiency ofcomplement c7 in a patient with recurrent meningitis.] MedKlin (Munich) 101:655–658.

Storey JD, Tibshirani R. 2003. Statistical significance for genome-wide studies. Proc Natl Acad Sci USA 100:9440–9445.

Taylor J, Tibshirani R, Efron B. 2005. The ’miss rate’ for theanalysis of gene expression data. Biostatistics 6:111–117.

Thakkinstian A, Han P, McEvoy M, Smith W, Hoh J, MagnussonK, Zhang K, Attia J. 2006. Systematic review and meta-analysisof the association between complementary factor H Y402Hpolymorphisms and age-related macular degeneration. HumMol Genet 15:2784–2790.

Zareparsi S, Branham KE, Li M, Shah S, Klein RJ, Ott J, Hoh J,Abecasis GR, Swaroop A. 2005. Strong association of theY402H variant in complement factor H at 1q32 with suscept-ibility to age-related macular degeneration. Am J Hum Genet77:149–153.

APPENDIX

The FDR algorithm that we used was similar to aalgorithm described previously [Taylor et al.,2005]. The main modifications were in step 7,where we adjusted FDRj to make FDR monotonic,as proposed previously [Storey and Tibshirani,2003], and in step 6, where we used differentpercentiles to compute p0 5 m0/m, the true pro-portion of null hypotheses among the m tests. Forvarious values of the cut-point C, we estimatedFDR by randomly permuting the labels of theindividuals:

1. For each genotype SNP, compute the statisticsTi (trend test) for each SNP i 5 1, 2,y,m anddetermine the min and the max values: Tmin

and Tmax.2. Choose k (e.g., k 5 50) cutoff points C1, C2,y,

Ck, equally spaced over the interval [Tmin,Tmax].

3. For each Cj, let Rj 5P

i I(TiZC), the observednumber of rejected null hypotheses; I( ) is theindicator function, I(true) 5 1, I(false) 5 0.

4. For b 5 1, 2, y, B (e.g., B 5 1000), randomlypermute the labels of the patients. For eachpermutation b compute the statistics for all themarkers, T1b, T2b,y,Tmb.

5. For each Cj, let Vj 5 (1/B)i,b I(TibZCj), theestimated number of false positives.

6. Estimate p0 5 m0/m, the true proportion of nullhypotheses among the m tests. Let q0.5 5 (1/B)P

b T½m=2�b, the half-point of the permuted Ti

distributions. Let p0 5 2P

i I(Tioq0.5)/m, andset p0 5 min (p0, 1).

7. For each Cj, estimate FDRj 5p0Vj/Rj (orFDRj 5 0 if Rj 5 0). To make FDR monotonous,adjust FDRj as proposed previously [Storeyand Tibshirani, 2003] by setting FDRj ¼ min iojFDRi.

8. Determine the cutoff level C0.05 that controlsFDR at a 0.05 level: C0.05 5 min (Cj|FDRjr0.05).



Documents

Evidence for Association between Multiple Complement