A comparison of methods sensitive to interactions with small main effects

Numerous genetic variants have been successfully identified for complex traits, yet these genetic factors only account for a modest portion of the predicted variance due to genetic factors. This has led to increased interest in other approaches to account for the "missing" genetic contributions to phenotype, including joint gene-gene or gene-environment analysis. A variety of methods for such analysis have been advocated. However, they have seldom been compared systematically. To facilitate such comparisons, the developers of the multifactor dimensionality reduction (MDR) simulated 100 data replicates for each of 96 two-locus models displaying negligible marginal effects from either locus (16 variations on each of six basic genetic models). The genetic models, based on a dichotomous phenotype, had varying minor allele frequencies and from two to eight distinct risk levels associated with genotype. The basic models were modified to include "noise" from combinations of missing data, genotyping error, genetic heterogeneity, and phenocopies. This study compares the performance of three methods designed to be sensitive to joint effects (MDR, support vector machines (SVMs), and the restricted partition method (RPM)) on these simulated data. In these tests, the RPM consistently outperformed the other two methods for each of the six classes of genetic models. In contrast, the comparison between other two methods had mixed results. The MDR outperformed the SVM when the true model had only a few, well-separated risk classes; while the SVM outperformed the MDR on more complicated models. Of these methods, only MDR has a well-developed user interface.