Two males, one control and one AD, were heterozygous for a previously unreported C->T (Arg->Stop) missense mutation at the first position of codon 479 (exon 11) of LHCGR (Figure 3).

Studies that aim to identify genetic interactions are best served by the multi-analytic approach to data analysis practiced here and, for family-based data, by 30. In the absence of biochemical evidence, putative genetic interactions are often difficult to accept intellectually. Multiple lines of statistical support, which represent the identification of a significant genetic interaction using different statistical and biological approaches, increase statistical and intellectual confidence in the biological reality of an interaction. In this study, LR and MDR provided distinct statistical/computational methods for the detection of significant genetic interactions, while LD analysis provided a distinct biological approach to the problem of interaction detection. A majority of AD-associated interactions reported here lack intra-study corroboration, as they were only identified by one analytical method (Table 3). As such, these interactions are not well supported. On the other hand, lhcgr2/APOE (the only interaction supported by all 3 analytic methods at the modified FDR level) is a better candidate for biochemical study. In addition, more support for lhcgr2/APOE is derived from the association of lhcgr2 with differences in AoO. The different analyses are not, of course, statistically independent of one another, as the same dataset is being analyzed in each case. Nevertheless, because they are not independent, failure to identify the same interaction using distinct, robust analytical methods seriously impugns the biological reality of a putative interaction.

The most significant interaction associated with AD in this study includes lhcgr2, a polymorphism located in intron 1 of LHCGR. Intronic polymorphisms are frequently implicated in increased disease susceptibility 114041 and intronic mutations have the potential to alter protein sequence dramatically. An intronic mutation can cause the disarrangement of an existing splice site or introduce a cryptic splice site, resulting in the addition/removal of hundreds of amino acids or premature termination of translation. A second mechanism by which intronic mutations might affect disease etiology is through their effect on the miRNAs that may be derived from the mutated intron. ~10–30% of spliced intronic material is exported to the cytoplasm, where it has the potential to act as an miRNA and alter protein expression 42. It follows that mutation of intronic sequence could decrease or increase the complementarity of an intron-derived miRNA to its target mRNA, thereby causing a relative decrease or increase in expression of the encoded protein. Indeed the 15 bases immediately upstream of lhcgr2 are complementary to APOE mRNA (Figure 4A), indicating this polymorphism may cap an intron-derived miRNA.

Examination of the sequence surrounding lhcgr2 and alignment of human and mouse (Mus musculus) LHCGR intron 1 indicate lhcgr2 may be located within a cryptic 3' acceptor splice site (Figure 4B). Acceptor splice sites are characterized by two conserved sequence patterns: a pyrimidine-rich sequence, known as the polypyrimidine tract, and the proximate terminal 'AG' of the intron 43. Although the length of the polypyrimidine tract and its distance from the end of the intron are variable, the terminal 'AG' is invariant 4344. Polymorphism lhcgr2 is located within a pyrimidine-rich region of intron 1 (Figure 4B). The lhcgr2 'C-allele' increases the CT-content of the surrounding 19 nucleotides to 79%, and, more locally, a 'C' at lhcgr2 forms a contiguous sequence of 7 Cs and Ts. Four bases downstream of this pyrimidine-rich tract is a 3' acceptor site consensus sequence, C

AG

LD between lhcgr1 and LHB loci was commonly detected in several AD and C datasets (Table 2), including 6 instances in the 'All male' dataset. The frequent identification of LD between lhcgr1 and LHB loci indicates that specific LHB haplotypes are found most frequently in individuals that are hetero- or homozygous for the LQ-insertion at lhcgr1. We questioned whether our data supported the idea that the LQ-insertion at lhcgr1 is, in general, in LD with LHB loci. To examine this possibility, we compared the frequency of genotypes at all 6 LHB loci between individuals with no LQ-inserts and those with 1 or 2 LQ-inserts (Figure 5). In both the group of all males and the group of all samples, χ²-tests revealed that the genotype counts of a majority of LHB loci were markedly different from one another in the LQ-insert and no-LQ-insert strata (Figure 5). This suggests: (1) a non-random force, such as selection, is driving the non-random association of LQ-insert alleles with specific LHB haplotypes in the North American Caucasian population – i.e., certain variants of LHB may be better suited to the LQ-insert variant of LHCGR; (2) given the extent of LD between lhcgr1 and LHB loci in C and AD samples, instances of significant LD between lhcgr1 and LHB loci in AD samples are not indicative of AD-associated interactions.

The National Cell Repository for Alzheimer's Disease (NCRAD; University of Indiana, Bloomington, IN) provided total DNA samples from 200 control (negative for AD and other neurodegenerative diseases, 50 male and 50 female) and late-onset AD patients (negative for other neurodegenerative diseases, 50 male and 50 female). All samples were obtained from North American Caucasian subjects. All samples were derived from individuals >= 75 years of age, and all AD samples were acquired from individuals whose AoO was = 75. Mean age of the control group was 84.73 w 4.61 years, while mean age of the AD group was 81.95 ± 5.69 years. Among the AD samples, AoO was 79.18 ± 3.47 years for males and 80.26 ± 5.07 years for females. DNA was obtained from 40 additional samples (10 male samples – 5 control, 5 AD and 30 female samples – 6 control, 24 AD) to test for significant interaction between APOE and an LHCGR locus (lhcgr2, see below). Direct sequencing of APOE, LHB (promoter, signal, and coding regions), and LHCGR (exons 1, 10, and 11) was performed using the primer pairs listed in Table 4. Cycle sequencing products were run on an ABI 3730 XL DNA analyzer at the University of Wisconsin Biotechnology Center (Madison, WI) and the resultant chromatograms were analyzed with FinchTV v1.4 (Geospiza, Seattle, WA). This study was carried out with IRB approval from the Health Sciences Institutional Review Board of the University of Wisconsin.

Similar to the analytic paradigm suggested by 30 for family-based data, we chose to analyze our case-control dataset using an array of analytic methods, testing for interactive as well as main effects, and treating the convergence of results from distinct analyses as the best evidence of association. Allele and genotype counts were used in the following analyses: (1) χ² tests of allele and genotype counts to test for main effects of individual polymorphisms; (2) tests of each locus for Hardy-Weinberg Equilibrium (HWE); (3) tests of combinations of two or three loci for linkage disequilibrium (LD); (4) tests for gene-gene interactions using multifactor dimensionality reduction (MDR); (5) tests for interactions using logistic regression (LR), and; (6) tests for association of polymorphisms with age of onset using one-way ANOVA. Additionally, to control for heterogeneity we stratified the dataset according to gender and applied the same 6 analyses. Finally, for each bi-allele locus, four genotype models were analyzed in tests for interactive effects: co-dominant, allele 1 dominant, allele 2 dominant, and over-dominant. This schema enabled us to: (1) address the possibility of heterozygote advantage, and; (2) test both alleles for dominance, as we had no a priori knowledge of which allele might carry risk.

For each sample, genotype and demographic data were entered into a MySQL relational database, enabling the quick identification of samples meeting an array of criteria. APOE genotype and 14 previously reported polymorphisms of LHB and LHCGR were scored. For each polymorphism, allele and genotype frequencies of the AD and control groups were calculated. Additionally, both groups were stratified by gender and gender-specific allele and genotype frequencies were calculated. Four separate genotype models were used in tests for main and interactive effects of bi-allele loci. For example, the following models would be used for a locus that varied between alleles B and b: (1) co-dominant (BB vs. Bb vs. bb); (2) B dominant [(BB + Bb) vs. bb]; (3) b dominant [BB vs. (Bb + bb)], and; (4) over-dominant [(BB + bb) vs. Bb]. For the tri-allele APOE, an 'ε4 dosage' model (genotypes grouped by the number of ε4 alleles) and an 'ε4 positive' model (ε4 allele present or not) were used in analyses.

The program Genetic Data Analysis (GDA) 46 was used to test each polymorphic locus for HWE. Minitab 47 was used to test for the association of individual polymorphisms with AD (χ² tests of allele and genotype counts) and AoO in the AD groups (ANOVA). In all tests for LD, genotypes were preserved in order to prevent significant deviations from HWE at a single locus from contributing to the measure of LD. We considered a number of theoretical issues when designing our analytical approach to detect gene-gene interactions (see Supplementary Information) and ultimately chose the combination of LD, MDR and LR analyses. Pairwise tests for LD were performed using the program PyPop, where the p-value reported here is derived from the difference between the likelihood of the inferred haplotype frequencies and the likelihood of the data if the two loci are assumed to be in linkage equilibrium 48. We reported the D' measure of LD, as this is an intuitive metric that represents the estimated proportion of maximum possible LD exhibited by the sample data. MDR was performed using MDR Software 49, which output the best 1-, 2-, 3-, and 4-factor models for a given dataset. 10-fold cross-validation was used. Given the weight APOE carries as a single factor, MDR was also run using APOE-free datasets in order to detect any interactions that did not include APOE. An interaction model was considered significant if it was selected as the best model in 5 or more of the CV runs and exhibited a testing accuracy of >0.5 in 7 or more CV runs. Pairs of loci exhibiting significant LD (p <= 0.02) and significant multi-locus models discovered using MDR were input as disease models in LR analyses performed in 47. This form of LR model selection was necessary, as a lack of several multi-locus combinations made backward model selection impossible and the lack of significant main effects in most loci studied made forward model selection impractical.

To account for multiple tests, testwise α levels were corrected using modified FDR (see Supplementary Information). Because a multi-locus combination was only tested with LR if LD and/or MDR analyses were suggestive of its association with AD, only a subset of the total array of possible LR tests were actually performed and the total, male, and female datasets were subjected to a different total number of tests: 182, 252, and 190 tests, resulting in modified FDR α levels of 0.0082, 0.0077, and 0.0081, respectively.

We searched for genetic association of single loci with AD using standard χ² and HWE tests and considered that subsequent tests for gene-gene interactions may identify interactions whose loci may or may not produce significant main effects on their own. There are a number of analytical issues to consider when searching for multiple interacting genetic (or environmental) factors associated with a disease of complex etiology. For one, genetic or trait heterogeneity among the samples (e.g., AD samples from males or females, or, with or without hypertension) has the potential to confound data analysis. Stratification of the dataset is the most straightforward method used to control for heterogeneity. For example, one might examine the effects of APOE heterogeneity by splitting control and AD groups into ε4 positive and negative strata and asking: Is the frequency of an SNP of interest significantly higher in ε4-positive AD samples than in ε4-positive controls, but not in the comparable ε4-negative comparison? If so, the data suggest an interactive effect between the SNP and ε4. Significantly, a non-stratified comparison of AD and control groups in such a case might lead a researcher to conclude the SNP has no association with AD, or, conversely, that the SNP is associated with AD in ε4 positive and negative individuals. Though more mathematical methods to control for heterogeneity exist, the majority of them are not applicable to case-control data.

In studying the genetics of a complex disease, it is important to consider the possibility that gene-gene or gene-environment interactions produce interactive effects that provide significant explanatory power, even in the absence of single factor, main effects 53. A number of methods allow researchers to test for gene-gene interactions using case-control data. A traditional method is logistic regression (LR), which, given a dataset, models the probability of a discrete outcome (in our case, AD or not) on n factors and their interactions, each qualified by a coefficient estimated using Maximum Likelihood Estimation. Attractively, LR produces an Odds Ratio (OR), which provides the researcher with an intuitive measure of how a particular array of genetic and/or environmental factors affects the likelihood of developing the disease. The major shortcoming of LR is the so-called 'curse of dimensionality', which refers to poor coefficient estimation resulting from too few or no examples of various multi-factor combinations in the dataset if sample size is too small or number of factors too large 54.

An alternative to LR analysis for the detection of interactions is multifactor dimensionality reduction (MDR) 55, which is advantageous for several reasons. A chief, practical advantage is that the researcher can easily test disease models that include interaction terms whose components lack significant main effects. This is critical, since it is increasingly apparent that genetic interactions, in the absence of main effects, frequently contribute to the susceptibility of an individual to complex diseases like AD 56. Also important, MDR is not "cursed" by dimensionality: the method is robust even when the input dataset lacks examples of various, multi-factor combinations. Finally, unlike LR, MDR analysis automatically measures the predictive accuracy and validity of a selected model through partitioning of the original dataset into training and testing subsets 5758.

Another approach to gene-gene interaction detection is to calculate linkage disequilibrium (LD) amongst combinations of the loci under investigation. Significant LD among 2 or more loci indicates non-random segregation of the loci in question, which implies that at least one multi-allele combination at these loci is overrepresented. Logically, any multi-allele combination enriched in the case but not the control group is considered to contribute increased susceptibility to the disease, and is reflected by a measurement of significant LD in the case group only 59. Williams et al. 60 used this approach in a study of polymorphisms associated with hypertension, discovering 16 combinations of 7 loci in 5 genes that exhibited significant LD in the case group only. Significantly, none of these loci were associated with a main effect on hypertension, indicating LD analysis has the ability to detect potentially significant gene-gene interactions in the absence of main effects.