The Framingham Study is a multi-generation community based study that began as a study of cardiovascular disease in 1948 among 5209 individuals comprising the Original cohort [24]. The Framingham Offspring Study includes 5216 spouses and offspring of the Original cohort members who have been followed since 1971 [25]. Surviving members of this cohort were examined and classified for lens opacities during exams conducted between May 1989 and December 1991 as part of the Framingham Offspring Eye Study (FOES) [26]. Starting in 1999, dementia-free surviving members of both Original and Offspring cohorts were invited to have a magnetic resonance imaging (MRI) examination of the brain. Repeat measurement data were obtained from a second MRI scan performed 2.6 years and 6.2 years later on average in the Original and Offspring cohorts, respectively. For the molecular genetic analysis, we focused on 1249 members of the Offspring cohort for whom ocular, MRI and GWAS data were publicly available (http://www.ncbi.nlm.nih.gov/gap. Accessed 2012 July 2). Subsets of Framingham Study subjects included in each component are shown in Fig. S1. This study was approved by the Boston University Institutional Review Board.

Ophthalmic examinations in the FOES were performed by two experienced certified ophthalmologists who evaluated the slit lamp photographs obtained through a dilated pupil using a standardized grading system [27]. Both eyes were examined in all study subjects. Nuclear opacification was classified by comparison to a set of standard photographs (grades 0 to 3) according to a validated ordinal rating system [28]. Cortical opacification was measured using a grid to assess opaque area in one-eighth wedges from the circumference of the lens cortex. An ordinal rating of cortical cataract was assigned based on the total number of one-eighth wedges occupied by cortical cataract. Persons with ≥4 positive wedges were assigned a rating of 4 (Fig. S2). Posterior subcapsular cataract (PSC) was assessed by measuring the vertical width of posterior opacification recorded by slit lamp photography (range: 0 to 7 mm). We selected the worse measurement from the right and left eyes. Thirty-three subjects with missing ophthalmic examination data or with a history of cataract surgery were excluded from the analysis.

Details of the MRI scan protocol and procedures for deriving structural volumes and measures of degeneration have been described previously [29]. In brief, volumetric brain MRI measures were obtained using analysis of digital images from the MRI scans. Participants were imaged on a Siemens 1 or 1.5-T MR machine (Siemens Medical, Erlangen, Germany). Digital images were processed by a central laboratory blinded to demographic and clinical identifiers. Quantified MRI measures were derived for frontal lobar volume (FBV), occipital brain volume (OBV), parietal brain volume (PBV), temporal brain volume (TBV), hippocampal volume (HPV), lateral ventricular volume (LVV), temporal horn volume (THV), and white matter hyperintensity volume (WMHV). HPV information was unavailable for 43 (2%) of the subjects at the time of the study.

Subjects were administered a neuropsychological test battery using standard protocols and trained examiners. Details of the tests administered and normative values for the Framingham Original and Offspring cohorts have been published [30], [31]. Cognitive test data were selected from the exam at the time the baseline MRI scan was performed.

Heritability (h²) in the narrow sense for a quantitative trait is the ratio of variance contributed by the additive genetic effects to the total phenotypic variance and can be estimated from family data by analysis of intraclass correlations (r). Heritability based on analysis of data obtained among siblings is calculated using the formula: h² = 2r. Co-heritability is the ratio of analogous covariance components between two traits [32]. It can be estimated from analysis of cross-trait sibling correlations. A cross-trait sibling correlation is the average of the correlation of the first trait in sibling A with the second trait in sibling B and the correlation of the second trait in sibling A with the first trait in sibling B. By contrast, self correlation is the correlation between two traits within individuals. The design for estimating self and cross-trait sibling correlations is illustrated in Fig. S3.

Self and cross-trait sibling correlations of volumetric brain MRI and cataract phenotypes with AD were estimated using the FCOR program in the Statistical Analysis for Genetic Epidemiology (S.A.G.E) software (version 6.0) [33]. FCOR calculates correlations for dichotomous and quantitative traits in extended families using the Pearson product-moment estimator with generalized weights. In addition, FCOR estimates multivariate familial correlations with their asymptotic standard errors without assuming multivariate normality of the traits by using a second-order Taylor series expansion and replacing all correlation parameters with their respective estimates [33], [34].

Raw measurements showing significant self and cross-trait correlations were further analyzed by adjusting for sex and age at exam. Age at blood draw (for DNA extraction) was substituted for age at eye exam because this information was not available for a large proportion of subjects at the time the data were analyzed. This was deemed reasonable since the age- and sex-adjusted residuals for age at blood draw and age at eye exam were highly correlated among individuals who had information for both variables (r = 0.96, P<10⁻⁴) and the blood draw and eye exam occurred during the same Framingham Study exam cycle. Correlations were also computed for the annual rate of change in the selected MRI traits which was calculated as the difference in the measures between the baseline and second MRI evaluations divided by the number of years between examinations. Because the distributions of many traits were multi-modal, age and sex-adjusted residuals for each trait were derived and normalized, forcing the marginal distribution of the trait to be approximately normally distributed under the null hypothesis, by taking the inverse standard normal transformation of the empirical quantile that was obtained using the formula [r(y)-1/3] divided by (n+1/3), where r(y) is the rank of residuals, y [20]. This approach has been shown to be valid in family-based association tests even with rare variants [20]. Normalized residuals were used for subsequent analyses. The significance threshold accounting for 38 independent tests was set at P = 0.0013.

All available participants were genotyped at Affymetrix (Santa Clara, CA) using the Affymetrix GeneChip® Human Mapping 500K Array Set and 50K Human Gene Focused Panel®. Phenotype and genome-wide association study (GWAS) data including raw genotyping calls and genotypes imputed using the HapMap 2 and 3 reference populations were obtained from the dbGaP website (http://www.ncbi.nlm.nih.gov/gap. Accessed 2012 July 2).

Prior to analysis, SNPs with a call rate less than 98%, SNPs with a minor allele frequency (MAF) less than 2%, and SNPs not in Hardy-Weinberg equilibrium (P<10⁻⁶) were excluded. Individuals with SNP call rates below 98% among the remaining SNPs, or whose gender as determined by analysis of X-chromosome data using PLINK [35] was inconsistent with the self-reported gender were also excluded. The 354,310 SNPs remaining among 7,966 subjects were then pruned to remove one from each pair of SNPs in high linkage disequilibrium (r²>0.8) using PLINK with a window size of 1500 SNPs and a step size of 150. After applying these criteria, 187,657 genotyped autosomal SNPs remained. Familial relationships were verified by examining segregation of marker genotypes within families using the MARKERINFO program in S.A.G.E. After confirming these relationships, marker genotypes showing Mendelian inconsistency were set to missing.

Population substructure was examined among those with self-reported ethnicity of white or Caucasian, first among the complete sample and again within a set of unrelated individuals consisting of one randomly selected individual from each extended pedigree. The complete sample of individuals was used for investigation of population structure using the pruned and cleaned genotypes from the Affymetrix 500K array and the CEU HapMap phase 2 reference data. SNP loadings were prepared using the smartpca script implemented in the EIGENSTRAT package [36]. This initial run was performed to confirm the presence of a single ethnic-specific cluster delineated based on self-declared ethnicity. For evaluation of within-cluster substructure, a set of SNP loadings were prepared using the pruned and cleaned genotypes from the unrelated sample of individuals with the same smartpca script implemented in the EIGENSTRAT package. The SNP weights for each eigenvector from the unrelated sample were then applied to all remaining family members to compute principal components for all individuals in the sample. We did not find any significant association between principal components and any cataract or MRI phenotypes in univariate analyses.

The SNP pruning strategy described above was employed to reduce computational burden associated with bivariate analysis in extended families. SNPs were coded under an additive model as 0, 1, or 2 with respect to the number of the reference alleles. The coded SNP values were included as covariates in regression models along with variables for polygenic effects and random effects to account for the various constellations of relative pairs. Association of each SNP with a bivariate outcome comprising age and sex-adjusted normalized residuals of quantitative measures for one cataract trait (ycc) and one MRI trait (ymri) was evaluated using the bivariate extension of the two-level Haseman-Elston (tHE) regression method developed for general pedigree data implemented in the RELPAL program in S.A.G.E [37], [38]. The regression model is denoted as: ycc_ik,ymri_ik ≃ βx_ik+bz_ik+e_ik for individual i in pedigree k; x, z and e represent the SNP, polygenic and random effects, respectively; and β and b are coefficients. This approach uses an iterative generalized squares algorithm which can accommodate various family structures and incorporate both individual-level and pedigree-level covariates. We demonstrated previously by simulation that this regression model has high power (>99%) for a SNP with a large effect size (>1.3), even with very rare variant (MAF<0.001), and produces few false positives in large extended families [39]. Some of the pedigrees in the Framingham Study have similar pedigree structures to those included in the simulations.

Nominal P values for association were determined using first-level Wald tests computed by RELPAL. The most significant results (i.e., SNPs with P<10⁻⁵) in the genome-wide scan were investigated further using data for all known SNPs in the implicated genes including those that were genotyped but had been dropped due to pruning. Additional SNPs in the HapMap 2 and 3 reference panels were imputed using MaCH [40]. All SNPs within 100 kb of top-ranked SNPs not located within genes were analyzed. A quantitative estimate between 0 and 2 representing the dose of the minor allele was used in the analysis instead of imputed genotypes to incorporate the uncertainty of the imputation estimates. We excluded imputed SNPs with low MAF (<2%), not in Hardy-Weinberg equilibrium (P<10⁻⁶) and low imputation quality (RSQ<0.8). A total of 186,192 genotyped SNPs and 1,465 imputed SNPs for a selected region were analyzed for association. Based on this number of SNPs, the threshold for genome wide significance was determined to be 2.66×10⁻⁷.

The possibility of a functional role of genes identified by bivariate GWAS in lens opacity and neurodegeneration was evaluated statistically by testing models including the genome wide significant SNP, a SNP from APP or PSEN1, and an interaction term. The two most significant APP genotyped and uncorrelated SNPs (pairwise LD<0.5), each with MAF>0.1, from the bivariate GWAS were tested for interaction with the top CTNND2 SNP rs17183619. No PSEN1 SNPs met these criteria and hence were not tested for interaction. The proportion of genetic variance in the bivariate trait explained by rs17183619 was estimated using methods proposed by So and colleagues [41], [42].

Eyes from a 68 year-old male and 71 year-old female with neuropathologically-confirmed AD and from two normal male controls ages 68 and 70 were procured from the Boston University Alzheimer Disease Center and National Disease Research Interchange (Philadelphia, PA). Eyes were fixed in 4% paraformaldehyde, embedded in paraffin, and sectioned at 8 um. Sections underwent antigen retrieval using citrate buffer pH 6.0 diluted to the manufacturer's instructions and heated to 90°C for 20 minutes. Tissue sections were immunostained with δ-catenin antibody (SC-81793, Santa Cruz Biotechnology, Inc., Santa Cruz, CA) at 1∶100 dilution and processed by conventional immunohistochemistry (LabVision Autostainer 360, Freemont, CA). Negative controls were run for each antibody. Tissue sections were visualized by brightfield photomicroscopy.

To identify quantitative measures of brain damage that are the best surrogate markers for presymptomatic AD and associated with lens opacity, we examined correlations among these traits within the same individual and between siblings (Table S1, Fig. S3). Although most of the MRI traits were significantly correlated with AD among all Framingham Study subjects who had at least one MRI exam, only TBV, LVV, and THV were significantly co-heritable with future AD after correction for multiple testing (Table S2). Lens opacity information was available for 1249 individuals who participated in the Framingham Eye Offspring Study (FEOS), had at least one brain MRI exam and were on average 51.3 years old at the time of the ophthalmic examination (Table S1, Fig. S1). At the time of the first MRI exam, none of these subjects were demented and only nine (0.7%) had a previous stroke. The mean interval between the eye exam and the first MRI exam was 9.3 years. Our analysis also revealed a significant sib-sib correlation of AD with cortical cataract (CC,  = 0.12, P<10⁻⁴) and posterior subcapsular cataract (PSC,  = 0.11, P<10⁻⁴), but not with nuclear cataract (NC) (Table S2). Heritability and co-heritability estimates were much greater for CC and several MRI traits, particularly THV and LVV, after adjustment for age and sex, and variable normalization (Table 1, Table S3). Taken together, these results suggest that CC and several AD-related brain changes have shared genetic liability or are associated with another variable under genetic influence, and that development of CC may foreshadow development of AD-related brain changes in later life.

Since CC was much more strongly co-heritable with THV than LVV (Table 1), a genome-wide search was conducted for association of common variants with the bivariate outcome of CC and THV. For this analysis, we used the set of 186,192 genotyped SNPs that remained after applying quality control procedures and a pairwise linkage disequilibrium (LD) cutoff of 0.8 (see Methods). The quantile-quantile (QQ) plot showed that the distribution of P values for the majority of SNPs (i.e., P>10⁻⁵) met expectation under the null hypothesis (genomic inflation factor lambda = 1.036), whereas more SNPs than expected under the null had P values<10⁻⁵ suggesting several true associations within this set of SNPs (Fig. S4). While none of the genotyped SNPs were genome-wide significant (threshold P = 2.7×10⁻⁷), suggestive association (P<5×10⁻⁶) was observed for CC-THV with three SNPs in CTNND2 (Table S4). This region was further examined by testing association of CC-THV with 1,465 accurately imputed SNPs (average correlation between actual and imputed genotypes RSQ = 0.9). Genome-wide significance was attained with CTNND2 SNPs rs17183619, rs13155993, and rs13170756 (Fig. 1A). These SNPs are in high LD (pairwise r²>0.9) and the results are supported by strong associations (P<10⁻⁵) with 19 additional CTNND2 SNPs (Fig. 1B, Table S5). The association of rs17183619 with CC-THV (_CC = −0.41, _THV = −0.18, P = 1.3×10⁻⁷) was stronger than the association with either CC ( = −0.24, P = 1.1×10⁻⁴) or THV ( = −0.14, P = 0.017) considered as separate outcomes. The effect direction (indicated by β value sign) was consistent in univariate and bivariate models indicating that the minor allele (G) of rs17183619 is associated with a decrease in both CC and THV. These results suggest that a variant in CTNND2 influences a process that is more precisely correlated with by degeneration in both the lens and brain than in either tissue considered separately. The proportion of the total co-heritability of CC and THV explained by rs17183619 is 19%.

Analysis of a model including amyloid precursor protein (APP) SNP rs2096488, CTNND2 SNP rs17183619, and a term for their interaction revealed evidence of a synergistic effect on the bivariate outcome of CC-THV (interaction P = 0.038) (Table 2). There was also significant interaction in the univariate model for CC (P = 0.0015), but not for THV.

To further explore the influence of variation in CTNND2 on processes occurring in the eye and brain, the top three SNPs from the GWAS of CC-THV were tested for association with bivariate measures of cortical cataract and performance on the Boston Naming Test (BNT), Immediate Recall (LMI) and Delayed Recall (LMD) portions of Logical Memory subtest of the Wechsler Memory Scale, Hooper Visual Organization Test (HVOT), Wide Range Achievement Tests (WRAT) or the Trail Making Test–B (TMTB) which are among the most heritable neuropsychological tests administered to FHS participants (h²>0.7). SNP associations were genome-wide significant with CC-BNT (P<2.4×10⁻⁷) and CC-WRAT (P<1.0×10⁻⁸), nearly genome wide significant with CC-LMI (P<6.0×10⁻⁷), and highly significant with CC-LMD (P<1.3×10⁻⁶), CC-HVOT (P<1.9×10⁻⁶) and CC-TMTB (P<3.7×10⁻⁶). These SNPs were not associated with any of the cognitive tests considered independently of CC (P>0.02).

The top-ranked CTNND2 SNPs (P<10⁻⁵) are located in the region encompassing exons 14–16, and the three most significant SNPs flank exon 14 (Fig. 1C). Bioinformatic analysis of this region revealed that the top-ranked SNP, rs17183619, is 249 base pairs upstream of a rare non-synonymous coding SNP (rs61754599) that results in a glycine to arginine missense substitution at residue 810. This missense mutation is predicted to have a deleterious effect on the protein structure (Fig. 2A). The sequence surrounding G810R is highly conserved across several species and bears a signature stretch of basic residues which are required for nuclear function of δ-catenin, the protein encoded by CTNND2 (Fig. 2B).

Transfection of the human CTNND2 cDNA with the G810R mutation into HEK293 cells that stably express the human APP Swedish mutation (APP_swe) showed that the distribution of mutant δ-catenin is altered predominantly at or near the cell surface (Fig. 3A). Transfected cells expressing the G810R mutation had a significantly more elaborate network of protrusions (P = 0.006, Fig. 3B), suggesting that the mutated CTNND2 gene may alter interactions of the plasma membrane with the underlying actin cytoskeleton. Moreover, transfection with the mutant δ-catenin resulted in a significant and specific increase in Aβ_1–42 secretion (P = 0.02, Fig. 3C).

We detected δ-catenin immunoreactivity in the central and bow regions of the lens epithelium and underlying anterior, subequatorial, and deep supranuclear cortical regions of the lens (Fig. 4) and retina (Fig. S5). The tissue localization pattern in postmortem lens from subjects with neuropathologically confirmed AD, but not lens from non-AD controls, was notable for abnormally increased δ-catenin immunoreactivity in the lens epithelium and cortex regions of the lens examined. Moreover, we detected pathological accumulation of δ-catenin immunoreactive product in AD lenses that was detectable as compacted cytosolic deposits with a striking basolaminar preponderance in the anterior and equatorial epithelium and as heterogeneously distributed punctate deposits in the superficial, subequatorial, deep cortical and supranuclear lens subregions. The presence of these pathological δ-catenin immunoreactive accumulations in AD lenses comprises a presumptive biological substrate for locally increased light scattering, lenticular opacification, and frank cataract.