Support vector machines with L1 penalty for detecting gene-gene interactions.

Interactions among genetic variants are likely to affect risk for human complex diseases, and their identification should increase the power to detect disease-associated variants and elucidate biological pathways underlying diseases. We propose a two-stage approach: model selection with support vector machines identifies the most promising single nucleotide polymorphisms and interactions; logistic regression ensures a valid type I error by excluding non-significant candidates after Bonferroni correction. Simulation studies for case-control data suggest that our method powerfully detects gene-gene interactions. We analyze a published genome-wide case-control dataset, where our method successfully identifies an interaction term, which was missed in previous studies.