Efficient identification of epistatic effects in multifactorial disorders

Complex diseases are typically caused by the combined effects of multiple genetic variations. When epistatic effects are present, these genetic variations show stronger effects when considered together than when considered individually. To identify groups of single nucleotide polymorphisms (SNP) that can be of use in the explanation of multifactorial conditions, we look for pairs of SNPs with a high value of information interaction. We show that standard classification methods or greedy feature selection methods do not perform well on this problem. We propose a computationally efficient method that uses the information interaction value as a figure of merit, and compare it with state of the art methods (BEAM and SNPHarvester) in artificial datasets simulating epistatic interactions. The results show that the method is powerful and efficient, and more effective at detecting pairwise epistatic interactions than existing alternatives. We also present results of the application of the method to the WTCCC breast cancer dataset. We found 89 statistically significant pairwise interactions with a p-value lower than 10−3. Somewhat unexpectedly, almost all the SNPs involved in pairs with high value of information interaction also have moderate or high marginals, a result that may imply that the search for more complex interactions may be more effectively conducted by looking only at SNPs which, by themselves, have correlations with the condition under study.