Genetic association is an increasingly popular method to identify genetic determinants of common diseases. Traditional single-locus association tests evaluate the marginal effects of each marker. It is to be expected, however, that the genetic susceptibility of complex traits would result from the interplay of several factors, including gene × gene interactions. As such, analytical approaches that consider single-nucleotide polymorphism (SNP) interaction effects have the potential to provide more power, especially when susceptibility genes have small or undetectable marginal effects.

The number of possible SNP pairs grows rapidly with respect to the number (n) of SNPs in a study, following n(n - 1)/2, and testing for the entire probability space for SNP × SNP interaction can become computationally unfeasible or cumbersome. Some promising two-stage approaches have been proposed [1, 2] to alleviate the problem, and we were particularly interested in comparing the performance of three of those by using replicate 1 of the simulated data set from Problem 3 of the Genetic Analysis Workshop 15 (GAW15).

The two-stage strategies that we use rely on the selection of SNPs based on their marginal single-locus test results obtained in the first stage. The simultaneous design is an approach that will test for interaction effects only between SNPs with p-values that fall below a pre-determined marginal significance threshold. We here compare the performance of the simultaneous design with thresholds of p = 0.05 and p = 0.1. It is expected that the more permissive threshold would offer greater detection power for an underlying model with weaker marginal signals.

We also evaluated the performance of the conditional two-stage design. The conditional approach tests for interaction effects of SNPs that reach global significance (after multiple comparison adjustment) at the first stage, in pairwise combination with all of the SNPs in the data set. This approach, as compared to the simultaneous design, has the advantage of including interactions where one of the SNPs would not show any evidence for a main effect.

Interaction effects have previously been modeled by logistic regression as described by Cordell [3]. We were interested in comparing a model that includes both dominance and additive interaction terms as compared to modeling only the additive effects. In previous work published by North et al. [4], it was suggested that the additive-only model was sufficient to detect most interaction models, whereas the combined additive and dominance effect model had the disadvantage of increasing the degrees of freedom of the logistic model and increasing the number of interaction terms to test for in each model construction, thereby imposing an unnecessary additional multiple comparison burden.

To obtain a sample of cases and controls, we randomly chose one case from each simulated affected sib pair from Replicate 1 of GAW15 Problem 3. Our sample consisted of 1500 cases with rheumatoid arthritis and 2000 controls. Available covariates for controls included sex, lifetime smoking, DR alleles and age. We used only sex, smoking, and DR alleles as significant covariates in our adjusted model. We ran our interaction models once without covariates and once with the covariates. We used the genome-wide 10 K simulated SNP chip set with 9187 polymorphic SNPs. There are no missing data and no errors in this data set. Analyses were performed without knowledge of the simulated answers. Computations were made with SAS v.9.1.3 (SAS Institute Inc., Cary, NC, USA) on WinXP and Sun (SAS code available upon request, see http://www.statgen.org).

The simulated data set contained five genetic loci associated with the case-control status (Loci A to F). According to the simulation schema of Problem 3, a simulated genetic interaction effect between Locus A on chromosome 16 and Locus C on chromosome 6 was modulated by the DR alleles. In Stage 1, we were able to detect the simulated Loci C on chromosome 6, D on chromosome 6, and E on chromosome 18, which passed Bonferroni correction in both logistic regression models with adjustment for covariates and without adjustment for covariates (Fig. 1). Locus F on chromosome 11 was detected by the additive and dominance model. Locus A on chromosome 16 was only identified by the logistic model with covariates in Stage 1 of the simultaneous design with a p-value < 10^-3. Overall, we detected SNP × SNP interactions involving C*C, F*F and F*chromosome 1 loci.

The simultaneous method optimizes the search for interaction effects between SNPs that provide some indication of marginal associations. This situation is most appropriate for disease models with additive polygenic effects. The advantage of the conditional design, on the other hand, is that it can capture interactions with loci that have no marginal effects, and as such is expected to be valid over a wider range of disease models. With our data, neither approach was able to capture the real underlying interaction between Loci A and C. Posterior testing for interaction between these loci revealed a p-value of 0.0015 with the additive model adjusted for covariates [Eq. (4)] and a p-value of 0.0001 with the additive model without adjusting for covariates, which was insufficient to be detected after multiple testing adjustment. We mention here that we used only the first replicate of GAW15 Problem 3 simulated data due to computational constraints, but the analysis of more replicates may have provided different results.

An important advantage of the conditional design in a real data set, is that Step 1 should provide only a small number of globally significant SNPs (i.e., two or three), which are then tested in pairwise combination with the full complement of SNPs in the study. As such, the conditional design should generate fewer interaction tests than the simultaneous design. In the GAW15 simulated data set, however, over 400 SNPs remained significant after adjustment for multiple testing in Stage 1 and caused a special situation in which the conditional design exploded into many interaction tests and lead to the physical proximity interaction effects that we detected with both approaches. It is not clear why there were so many significant results on chromosome 11 and it would be interesting to validate those results by testing the other replicates. The use of logistic regression for the detection of interactions has the disadvantage of performing poorly with high dimensionality, which can lead to false positives and decreased power. In the present study for instance, the conditional approach with additive and dominance effects required the estimation of 16,279,364 parameters from 32,154,500 observations. Another potential problem with logistic regression is the phenomenon of the separation or quasi-separation of data. This occurs in the fitting process of the logistic model when the likelihood converges while at least one parameter estimate diverges to infinity. Quasi-complete separation of data, however, was only observed in 0.3% of the tests that we performed with the additive model with covariates in this study. High correlations between predictors, referred to as multicollinearity, may also decrease the power of the logistic regression. This may occur with SNPs that are in high LD. Multicollinearity due to linkage disequilibrium could, however, be addressed statistically by combining SNPs that are in high LD into haplotypes.

The significant interactions found by the full logistic regression model with both dominance and additive effects [Eq. (3)] were all captured by the additive effect coefficient (i _aa ). We conclude from our results that using a model with only the additive genetic effects of two SNPs should capture most of the underlying interactions and will reduce the multiple testing otherwise imposed by the four interaction terms. A large portion of the interactions that we detected resulted from SNP pairs that were within 1000 kb of each other. Our results clearly highlight the need to include an additional step or procedure to account for locus proximity in the evaluation of interaction effects. All remaining interactions without exception were false positives that included a SNP with exceptionally high marginal signals from chromosome 11 or chromosome 6 (Fig. 1). Those SNPs with excessively high marginal signals (p-values < 10^-200) are likely to be responsible for the false-positive signals. This finding draws attention to the importance of directing further work toward the evaluation of the impact of strong marginal SNPs in interaction modeling.