C2 and CFB Genes in Age-Related Maculopathy and Joint Action with CFH and LOC387715 Genes

Background Age-related maculopathy (ARM) is a common cause of visual impairment in the elderly populations of industrialized countries and significantly affects the quality of life of those suffering from the disease. Variants within two genes, the complement factor H (CFH) and the poorly characterized LOC387715 (ARMS2), are widely recognized as ARM risk factors. CFH is important in regulation of the alternative complement pathway suggesting this pathway is involved in ARM pathogenesis. Two other complement pathway genes, the closely linked complement component receptor (C2) and complement factor B (CFB), were recently shown to harbor variants associated with ARM. Methods/Principal Findings We investigated two SNPs in C2 and two in CFB in independent case-control and family cohorts of white subjects and found rs547154, an intronic SNP in C2, to be significantly associated with ARM in both our case-control (P-value 0.00007) and family data (P-value 0.00001). Logistic regression analysis suggested that accounting for the effect at this locus significantly (P-value 0.002) improves the fit of a genetic risk model of CFH and LOC387715 effects only. Modeling with the generalized multifactor dimensionality reduction method showed that adding C2 to the two-factor model of CFH and LOC387715 increases the sensitivity (from 63% to 73%). However, the balanced accuracy increases only from 71% to 72%, and the specificity decreases from 80% to 72%. Conclusions/Significance C2/CFB significantly influences AMD susceptibility and although accounting for effects at this locus does not dramatically increase the overall accuracy of the genetic risk model, the improvement over the CFH-LOC387715 model is statistically significant.


Introduction
Age-related maculopathy (ARM), also known as age-related macular degeneration (AMD), is a devastating disorder and a major public health issue.ARM poses one of the greatest threats to vision in the elderly of developed countries and an estimated 1.75 million individuals over 40 years old in the United States suffer vision loss from the disease with an estimated increase to 2.95 million individuals by 2020 [1].ARM is a degenerative disorder primarily, but not exclusively, affecting the central macular region of the retina.It is characterized by formation of drusen, pigment epithelial changes, atrophic degenerative changes, and formation of choroidal neovascularization.
In the present study, we do not try to distinguish between the genes and variants in this region but use S69A as a tagging SNP; given the extensive LD in the region, especially between S69A and the HTRA1 promoter variant, S69A can serve as a reasonable proxy for the genetic risk contributed by this region.In fact, a recent finemapping effort in this region does suggest that S69A is more likely, than the HTRA1 promoter variant, to be causally responsible for the impact of this locus on ARM [54].
CFH is now widely accepted as an important ARM susceptibility gene, harboring variants and haplotypes associated with increased and reduced disease risk.Functional studies suggest that CFH inhibits the activation of the alternative complement cascade and complements have been found in the drusen of ARM patients [55][56][57][58][59].It is therefore logical to ask whether other genes involved in the alternative complement pathway may influence the risk.This task was partly tackled by Gold et al. [60] who found ARM-associated variants in the complement component receptor B (CFB, GeneID 629) gene and the adjacent complement component 2 (C2, GeneID 717) gene on chromosome 6p21.Both genes play a role in complement pathways: CFB in the alternative pathway and C2 in the classical pathway.As was the case for CFH and LOC387715, this finding also seems robust and has been replicated in two casecontrol cohorts [27,61] and one family cohort [61].However, because of the strong LD across the C2/CFB region, distinguishing between the genes and identifying true functional variants has proven challenging.Recently two studies [62,63] reported significant associations between ARM and variants in the complement component 3 (C3, GeneID 718) gene on chromosome 19p13.C3 plays an important role in activation of both the classical and the alternative complement pathways and the plasma complement C3a des Arg levels are significantly elevated in ARM cases compared to controls [64].A fourth recent study [65] also found ARM associated variants in the C7 (GeneID 730) and MBL2 (GeneID 4153) complement pathway genes by complement pathway focused analysis of an earlier genome-wide association scan [6].
In the present study, we investigated four SNPs in the C2/CFB region, rs9332739 and rs547154 in C2 and rs4151667 and rs2072633 in CFB, in case-control and family cohorts of white subjects.Only rs547154, an intronic SNP in C2, was significantly associated with ARM in our data.Subsequently, rs547154 was used as a tag for this region in multifactor analyses of the joint effect of the three genomic regions (CFH, LOC387715, and C2/ CFB) on ARM susceptibility.

Phenotyping, study participants and quality control
Because of the complexity and ambiguity in the ARM phenotype, we have previously defined three affection status models (types A, B, and C) [66,67].For clarity we restrict our analyses here to unaffected (or normal) individuals and type A affected individuals.The type A model is our most stringent and conservative diagnostic model and individuals classified as type A ARM affected are clearly affected with ARM based on extensive and/or coalescent drusen, pigmentary changes (including pigment epithelial detachments) and/or the presence of end-stage disease (geographic atrophy [GA] and/or choroidal neovascular [CNV] membranes).Unaffected individuals were those for whom eye-care records and/or fundus photographs indicated either no evidence of any macular changes (including drusen) or a small number (,10) of hard drusen (#50 mm in diameter) without any other retinal pigment epithelial (RPE) changes.Individuals with evidence of large numbers of extramacular drusen were not classified as unaffected and therefore not included in the analyses.
No family member was considered unaffected but was considered of unknown phenotype if not affected with type A ARM.
Using only the subset of white participants, our data include 611 ARM families, 187 unrelated cases and 168 unrelated controls.The ARM families consist of 1,524 genotyped individuals (569 males and 955 females) and, in terms of genotyped affected relative pairs, the families include total of 501 sib pairs, 7 half sib pairs, 60 cousin pairs, 13 parent-child pairs, and 38 avuncular pairs; Pedstats (version 0.6.8)[68] was used to get summary counts of the family data.See Table 1 for other characteristics of the subjects.Before analyzing the family data, PedCheck (version 1.1) [69] was used to check for Mendelian inconsistencies.Since it can be extremely difficult to determine who exactly has the erroneous genotype within small families [70], we set genotypes of problematic markers to missing for every individual within each family containing a Mendelian inconsistency; this needed to be done for only one SNP (rs859705 in part 3) in one family.
Informed consent was obtained from all participants under research protocols that have been reviewed and approved in accordance with the Declaration of Helsinki and the Guidelines for Human Subjects Protection issued by the Office of Human Subjects Research (National Institutes of Health) by the University of Pittsburgh IRB (#9506133) and the University of California-Los Angeles IRB (#10-06-096-01).

Association analyses and LD estimation
Case-Control data.Using the set of unrelated cases and controls, SNP-disease allelic and genotypic associations were tested using the Fisher's exact test as implemented in R (version 2.2.1) [71].For significantly associated SNPs the strength of the association was estimated by crude odds ratios (ORs) and population attributable risks (PARs).To calculate the PARs we used the general formula: PAR = P f (OR21)/(1+P f (OR21)), where P f is the prevalence of the risk or protective factor (genotype) in the general population as estimated from the controls.The ORs were calculated using logistic regression models in R. Confidence intervals (CIs) for the ORs and PARs were derived using the asymptotic normal distribution of ln(OR) and ln(1-PAR), respectively.Haplotypic associations of 2-and 3-SNP moving window haplotypes in the C2/CFB locus were evaluated using the haplo.ccfunction of the haplo.statspackage (version 1.2.2) [72] of R.This function implements a score test for global test of association between binary traits and haplotypes and accounts for ambiguous linkage phase by the EM algorithm; empirical P-values were generated using 10,000 replicates.Allele and genotype frequencies were estimated by direct counting and deviations from Hardy-Weinberg equilibrium (HWE) were tested, in cases and controls separately, using the exact test as implemented in R Genetics package (version 1.2.1) [73].Haploview (version 3.32) [74] was used to estimate the LD across the C2/CFB region, both D' and r 2 were calculated separately in cases and controls.
Family data.When incorporating cases from the families into the analyses, the CCREL method (version 0.3) [75] was used to test SNP-disease allelic, genotypic and 2-and 3-SNP haplotypic associations.The CCREL method permits testing for association with the use of related cases and unrelated controls simultaneously and, briefly, it accounts for biologically related subjects by calculating an effective number of cases such that individuals are assigned weights that are used to construct a composite likelihood, which is then maximized iteratively to form likelihood ratio tests.For the CCREL analyses, type A-affected family members were assigned the phenotype ''affected'', unrelated controls the phenotype ''normal'' and family members not affected with type A ARM the phenotype ''unknown''.

Multifactor and gene-gene interaction analyses
To build predictive models of the genetic risk of ARM contributed by the CFH, LOC387715, and C2/CFB loci, we applied both logistic regression and the new generalized multifactor dimensionality reduction (GMDR) method (version 0.7) [76].The GMDR method, unlike the original MDR method [77], permits adjustment for covariates and better handles data with unequal numbers of cases and controls, and can be used to analyze both qualitative (e.g.binary) and quantitative traits via different link functions.Both methods only handle unrelated individuals.Therefore, to make use of more of our data, we combined one type A affected person picked at random from each of the 611 ARM families with the data of unrelated cases and controls.We consider this to be appropriate to do since the association results suggest the effects of the genes to be similar in both groups.
Logistic regression.For each pair of loci, we first followed the modeling strategy proposed by North et al. [78] for two-factor genetic risk models.A series of logistic regression models were fitted to the data in order to find a parsimonious model for the joint effects of each pair of loci.Models allowing for additive effects (ADD1, ADD2, and ADD-BOTH), models incorporating dominance effects (DOM1, DOM2, and DOM-BOTH), and three interaction models (ADD-INT, ADD-DOM, and DOM-INT) were fitted.We fit three-factor models of the joint effect of all three loci and test, using a likelihood ratio test (LRT), whether accounting for the protective effects at C2/CFB significantly improves the fit of a model with CFH and LOC387715 effects only.Since, for each pair of loci, the two-factor analyses implicated additive models as the most parsimonious and to keep the number of parameters as small as possible we only fit three-factor additive models without interaction (ADD1, ADD2, ADD3, ADD12, ADD13, ADD23, and ADD123).The models are compared by the Akaike information criterion (AIC); the most parsimonious model has the lowest AIC and a model is considered to provide a significantly better fit to the data if it has AIC more than 2 units lower than the comparison model [78].Details regarding coding of genotypes in the models are available in the supporting information (Text S1).
GMDR.Just as in the case of logistic regression, when using the GMDR method, one needs to be aware of the risk of overfitting, especially in the case of small sample sizes.The GMDR method, however, uses cross-validation to guard against overfitting.We applied the method to our data in order to identify three-locus genotypes associated with increased and decreased disease risk.For comparison we also present and discuss the CFH and LOC387715 two-factor model.We performed both crude analysis and analyzed the data while adjusting for age, gender, and cigarette smoking.We used 5-fold leave-one-out cross-validation and exhaustive search of all possible one-to three-locus models in the GMDR analyses.In the adjusted analysis age (in years) was the age at the time blood was drawn (i.e.DNA donated), and cigarette smoking was a binary variable (ever vs. never smoked).The smokers smoked on averaged 40.45 (standard deviation [SD] 32.96; range 0.23-207.00)pack-years (years6packs/day smoked) of cigarettes.The sample in the adjusted analysis includes fewer observations (557 cases and 118 controls fully typed at all three SNPs) than the sample in the unadjusted analysis (640 cases and 142 controls fully typed at all three SNPs) because of missing information.We compared both the sensitivity = TP/(TP+FN) and the specificity = TN/(TN+FP) of the models, where TP = number of true positives, TN = number of true negatives, FP = number of false positives, and FN = number of false negatives.As a single measure of the accuracy of the models we used the balanced accuracy = (sensitivity+specificity)/2 rather than the accuracy = (TP+TN)/(TP+TN+FP+FN) because number of cases and controls is unequal.The average sensitivity, specificity, and balanced accuracy over the testing sets of all five cross-validations are reported.As a measure for the appropriateness of the models, the sensitivity, specificity, balanced accuracy, and Pvalue are reported for all models when applied to the whole dataset.

Interaction with cigarette smoking
In a logistic regression framework we tested, using a LRT and the combined data of unrelateds and one type A affected from each family, whether cigarette smoking interacts with the SNPs at the three genes.The genotypes were coded in additive way, as in the logistic regression analysis above, and cigarette smoking as ever vs. never smoked.

Results of association analyses
The genotype distributions of the 4 SNPs typed in C2 and CFB and the Y402H variant in CFH and the S69A variant in LOC387715 are in HWE in both our cases and controls (Table 2).Of the 4 SNPs typed in the C2/CFB region, only rs547154, an intronic SNP in C2, is significantly associated with ARM (Table 2) in both our case-control (P-value of genotypic test 0.00007) and family data (P-value of genotypic test 0.00001), which is also significant after adjusting for multiple testing of 4 tests (Bonferroni corrected 0.05 significance level is 0.0125).The haplotypic association tests show that haplotypes spanning the entire C2/CFB locus are significantly associated with ARM (Table 2).Although LD between rs547154 and the SNPs in CFB (Figure 1) is not strong, in neither cases nor controls, these results are not sufficient to rule out either C2 or CFB as an ARM candidate gene, because of limited number of SNPs investigated.Individuals carrying the protective allele at C2 are at 0.22 (95% CI 0.10 to 0.48) times less risk of having ARM compared to controls as estimated with a crude OR.The corresponding PAR is -18% (95% CI -28% to -8%).Detailed results of marginal association of Y402H in CFH and S69A in LOC387715 are in Table 2 and the supporting information (Text S2).

Results of multifactor analyses
Logistic regression.First we fitted two-factor genetic risk models for each pair of loci and found that an additive model without interaction was the most parsimonious in all cases (Table 3).Three-factor additive model was then fitted in order to test whether the three-factor model provided better fit to the data than any two-factor models (Table 4).The three-factor model of CFH, LOC387715, and C2 SNPs coded in additive fashion was the most parsimonious and fitted significantly better (P-value of LRT 0.002) than the next-best model (which modeled CFH and LOC387715 additive effects only).
GMDR.The two-factor GMDR unadjusted and adjusted models (Figure 2A and B) classify everyone with a homozygous (TT) LOC387715 risk genotype as cases and everyone with the homozygous (GG) LOC387715 non-risk genotype as controls.On the other hand, individuals heterozygous (GT) at LOC387715 need to have at least one CFH risk allele (C) to be classified as cases.When comparing the unadjusted two-factor model (Figure 2A) to the unadjusted three-factor model (Figure 2C), the most dramatic change is in the upper left most cell (CC-GG CFH-LOC387715 joint genotype): 76 cases in that cell that were wrongly classified as controls by the two-factor model while 71 are correctly (5 wrongly) classified in the three-factor model.This increases the sensitivity from 63% to 73%, but comes at a cost of decreased specificity (80% to 72%).Now looking more closely at the three-factor models (Figures 2C  and D), the results of the GMDR analyses suggest that having at least one copy of the protective allele (T) at C2/CFB may reduce the risk contributed by CFH and LOC387715 risk genotypes.For example in the unadjusted model (Figure 2C), individuals with the CT-GT and CT-GG two-locus genotypes at CFH and LOC387715 and without the C2/CFB protective allele are classified as cases while those with the protective allele are classified as controls.In the adjusted model (Figure 2D), however, individuals with the CT-GT and CC-GG as well as TT-TT and CC-GT two-locus genotypes at CFH and LOC387715, are classified as controls if they carry the C2/CFB protective allele but cases otherwise.Note that the difference between the three-factor unadjusted and adjusted models is not due to the smaller dataset used in the adjusted analysis.To make sure this was not the case, we ran unadjusted analysis on the smaller dataset and arrived at the same model as in the unadjusted analysis.The predictive models presented in Figure 2 seem sensible as the predicted high-risk two-and threelocus genotypes group together.The predictive accuracy of the three-factor model measured by sensitivity, specificity, and balanced accuracy is .70% of both the unadjusted and adjusted models (Table 5).In the unadjusted analysis all five crossvalidations suggest that the classification scheme classifies individuals significantly better than random (P-values ,0.05) and in the adjusted analysis the classification is significantly better in all but one cross-validation experiment (Table 5).Both models provide excellent fit to the whole data (P-values ,0.0001).In Table S2, we present the joint genotype and relative genotype frequencies in cases and controls, which provides a complementary view of the same findings as in Figure 2.
Logistic regression vs. GMDR.In both the logistic regression and GMDR analyses, the best fitting one-factor model is the model with LOC387715 only (Tables 4-5; in the logistic regression, the LOC387715 model has the lowest AIC of all one-factor models and, in the GMDR results the LOC387715 model has the highest balanced accuracy (in both the unadjusted and adjusted analyses).However, the difference between the CFH and LOC387715 one-factor models is, very small, and, as the GMDR analyses show, the difference lies in the sensitivity and specificity rather than the overall balanced accuracy measure (Table 5).The three-factor model of CFH, LOC387715, and C2/ CFB effects is implicated as the best model in both the regression and GMDR analyses (Tables 4-5).The logistic regression analyses suggest that accounting for C2/CFB effects significantly improves the two-factor model of CFH and LOC387715 only (P-value 0.002).The GMDR analyses show that this improvement is due the increases sensitivity but the balanced accuracy increases only from 71% to 72% (Table 5).The GMDR analyses also suggest that adjusting for age, gender, and cigarette smoking does not dramatically improve the fit of the models.In fact, all models (1-, 2-, and 3-factor) have approximately the same balanced accuracy irrespective of whether adjustment is made (Table 5).

Results of gene-cigarette smoking interaction analysis
Cigarette smoking does not significantly interact with any of the three variants investigated in our data.The p-values of LRTs are 0.24, 0.99, and 0.43 for Y402H in CFH, S69A in LOC387715, and IVS10 in C2, respectively.For all three genes the most parsimonious models, according to AIC, are the models with only the additive gene effect and no smoking effect (results not shown).

Discussion
We have replicated the association of one C2 variant (rs547154) with ARM in both our case-control and family datasets and we have shown that accounting for the effects of C2/CFB significantly improves the fit of the logistic regression model in comparison to the two-factor model of joint additive effects of CFH and LOC387715 (Table 2 and 4).Interestingly, both of the nonsynonymous coding changes, E318D (rs9332739) in C2 and L9H (rs415667) in CFB, identified by Gold et al. [60] are insignificant in both of our datasets.However, as these variants (rs9332739 and rs415667) are quite rare, power to detect these variants is low.Even so, our independent confirmation of the statistically significant effect of this locus in ARM in two datasets, including family-based data, further supports the contribution of this locus to the genetic susceptibility of ARM.
As mentioned above, accounting for the effect of the C2/CFB locus significantly improved the fit of a logistic regression model of additive effects of CFH and LOC387715 variants.To further understand this, we built predictive models of these three loci using the new generalized multifactor dimensionality reduction method (GMDR), and found that addition of C2/CFB to the model increased sensitivity (from 63% to 73%).However, the specificity is lowered (from 80% to 72%) and so the balanced accuracy only increases from 71% for the two-factor CFH-LOC387715 model to 72% for the three-factor CFH-LOC387715-C2 model in the unadjusted analysis (Table 5).If it were considered more important to identify cases than controls correctly, while maintaining a reasonable specificity, the three-factor model would be the better choice.
Since our associated variant (rs547154) in C2/CFB is rare, it is expected that accounting for the effect of this locus, using rs547154 as a tag, would not markedly improve the overall prediction accuracy of the genetic risk model with CFH and LOC387715 effects only, even though the effect may be strong.Although, positive associations in the C2/CFB region have been found and replicated primarily for rare variants [27,60,61], we cannot exclude the possibility that the true causal variant(s) in this region may be common, especially since not all known common SNPs have been typed in C2/CFB studies (Figure S1).Obviously, a genetic risk model of CFH, LOC387715, and C2/CFB effects could be quite different from our model presented here if the C2/CFB causal variant(s) were common, as then the rare rs547154 would be a bad proxy.
Another concern regarding correctness of the three-factor model is the small sample size for the 'protective' GT genotype at IVS10 (rs547154) in C2 (Figure 2 and Table S1), although it is important to remember that cross-validation does guard against over-fitting due to small sample sizes or a large number of parameters.The least stable classifications in Figure 2C are those cells in which the height of the bars is similar or number of individuals is low.In such cases, the classification rule can change if only a few individuals were added to that cell.For example, if we had only one additional control with a CC-GG-GG CFH-LOC387715-C2 genotype (upper left most cell, left panel in Figure 2C), then individuals with this genotype combination would have been classified as controls instead of cases.To construct our original unrelated data set, we picked one case at random from each of the families.Figure S2 examines the sensitivity of our three-factor analyses when we randomly re-pick one case from each family.We created 10 other combined data-sets (overlap among cases from the families ranges from 57% to 66%) and ran the GMDR method.The figure clearly shows that only classifications corresponding to the rare GT genotype at IVS10 (rs547154) in C2 are changed across samples, while the classifications corresponding to the common GG genotype are robust.
Accounting for covariates (age, gender, and cigarette smoking) failed to improve the prediction accuracy of the genetic risk models (Table 5).In fact, for the one-and two-factor models, the adjusted analyses arrived at the same high-risk (and low-risk) genotype combinations as the unadjusted analyses.The difference in sensitivity, specificity, and balanced accuracy between the two analyses is solely due to different number of individuals used in each set of analyses (because of incomplete smoking information).
In the three-factor model, genotypes were grouped differently depending on whether unadjusted or adjusted analyses were performed (Figure 2) and, as mentioned in the results section, this difference is not solely due to different number of individuals used in each set of analyses.The one-factor models of CFH and LOC387715 did worse than the higher-factor models (balanced accuracy 64% and 68%, respectively), although, when considering they only model genetic effects at one locus, both models perform amazingly well.The GMDR method selected the LOC387715 model as the best of all the one-factor models.However, depending on what the goals of using a prediction model are, one could easily choose the CFH model as the best one-factor model.For example, the sensitivity of the CFH model is much higher than of the LOC387715 model (85% vs. 71%), but this increased sensitivity comes at a cost of low specificity (44% vs. 65%).
In their original report on the C2/CFB locus in ARM, Gold et al. [60] did not include LOC387715 variants and, using a genetic algorithm search approach, they arrived at a genetic risk model of two CFH variants and three C2/CFB variants.The sensitivity and specificity of their model were 74% and 56%, respectively (which results in balanced accuracy of 65%).Interestingly, our threefactor model, which includes LOC387715 effects, provides a better prediction accuracy (balanced accuracy 72%), similar specificity (73%), and better specificity (72%) than their more complicated five-factor model.Furthermore, even our simpler two-factor model of CFH and LOC387715 effects also provides better prediction accuracy (balanced accuracy 71%).
We believe that a word of caution must be provided with regard to the possible use of these predictive models in clinical situations.It must be understood that the models presented in this paper and by others are based on comparison of extreme phenotypes (those with advanced forms of ARM and age-matched controls with minimal or no clinical findings).This does not address the determination of ARM risk for individuals for whom mild to moderate retinal findings are present.Secondly, odds ratios based on case-control association studies are not comparable to prospective, population-based relative risk assessments that still need to be done for ARM.Finally, one must always consider the composition of the population that may be subjected to molecular genetic screening.If we are considering the general population for whom the risk of ARM-related vision loss is less than 1% over their lifetime, then the current genetic models have inadequate levels of specificity to avoid a high percentage of false positive results.However, for individuals from high-risk cohorts for whom the prevalences of the high-risk variants are known, molecular diagnostic testing may be sufficiently discriminating of relative risk, though it is unclear how such knowledge would affect individual behavior or preventive treatments at this time.In summary, we have confirmed the likely influence of the C2/ CFB locus on ARM and shown that accounting for the effects at this locus can likely further stratify individuals as being at high or low risk of developing ARM.The important role the classical and/ or alternative complement pathways seem to have in the diseasepathology of ARM should now encourage investigators to not only look at more complement pathway genes, but also to establish the biological mechanism behind the influence of LOC387715 (or HTRA1) on the development of the disorder.Then, once either LOC387715 or HTRA1 has been convincingly shown to be the true ARM susceptibility gene on 10q26, it is likely that we will see similar trends in discoveries of genes involved in the same pathway as either of those genes.
allele MAF = minor allele frequency HWE = Hardy-Weinberg equilibrium The haplotypic P-values correspond to the haplotypes of the SNP in the same row as the P-value and the next one or two SNPs for the 'With 2 SNPs' and 'With 3 SNPs' P-values, respectively Genotype counts in unrelated cases and controls are available in Table S1 doi:10.1371/journal.pone.0002199.t002

Figure 1 .
Figure 1.Linkage disequilibrium (LD) across the C2/CFB region in unrelated cases and controls.The darker the boxes the higher the r 2 .The top number in each box is r 2 and the bottom number is D'.Locations of the SNPs within the genes are shown.Red lines/boxes show the locations of exons in C2 and green lines/boxes the locations of exons in CFB.doi:10.1371/journal.pone.0002199.g001

Figure 2 .
Figure 2. Results of GMDR analyses.A: Results of unadjusted GMDR analysis for the best two-factor model.B: Results of adjusted GMDR analysis for the best two-factor model.C: Results of unadjusted GMDR analysis for the three-factor model.D: Results of adjusted GMDR analysis for the threefactor model.Dark grey and light grey boxes correspond to the high-and low-risk genotype combinations, respectively.The black and white bars within each box correspond to cases and controls, respectively.The top number above each bar is number of individuals and the bottom number is the sum of scores for the corresponding group of individuals (cases or controls with particular three-locus genotype).The heights of the bars are proportional to the sum of scores in each group.doi:10.1371/journal.pone.0002199.g002

Table 3 .
Results of fitting two-factor logistic regression models.

Table 4 .
Results of fitting three-factor logistic regression models.Factor 1 is Y402H in CFH, Factor 2 is S69A in LOC387715, and Factor 3 is rs547154 in C2.Detailed model definitions are given in the 'Materials and Methods-Multifactor and interaction analyses' section and Text S1.AIC difference is the difference from the AIC of the best fitting model.Most parsimonious model is in bold.Model with best fit (lowest AIC) has AIC difference = 0. doi:10.1371/journal.pone.0002199.t004