COE: A General Approach for Efficient Genome-Wide Two-Locus Epistasis Test in Disease Association Study

The availability of high-density single nucleotide polymorphisms (SNPs) data has made genome-wide association study computationally challenging. Two-locus epistasis (gene-gene interaction) detection has attracted great research interest as a promising method for genetic analysis of complex diseases. In this article, we propose a general approach, COE, for efficient large scale gene-gene interaction analysis, which supports a wide range of tests. In particular, we show that many commonly used statistics are convex functions. From the observed values of the events in two-locus association test, we can develop an upper bound of the test value. Such an upper bound only depends on single-locus test and the genotype of the SNP-pair. We thus group and index SNP-pairs by their genotypes. This indexing structure can benefit the computation of all convex statistics. Utilizing the upper bound and the indexing structure, we can prune most of the SNP-pairs without compromising the optimality of the result. Our approach is especially efficient for large permutation test. Extensive experiments demonstrate that our approach provides orders of magnitude performance improvement over the brute force approach.