Inference of High-Order Epistatic Interactions Using Generalized Relevance Learning Vector Quantization with Parametric Adjustment

Single Nucleotide Polymorphism (SNP) is an important type of variation in the genome which is frequently associated to particular traits or to underlying biological mechanisms. SNPs can operate alone or in groups, through epistatic interactions. In this work, we present a method to identify relevant high-order epistatic SNPs interactions involving 3, 4, and 5 SNPs. We use a pattern classification algorithm based on LVQ, called Generalized Relevance LVQ (GRLVQ). We show that by performing a careful algorithm analysis and a fine tunning of the parameters, the method is able to consistently obtain excellent results with low computational cost, when 3, 4 or 5 SNPs are considered in the epistasis. To the best of our knowledge, these results are far better than any other in the current literature.