Detecting epistatic interactions in genomic data using Random Forests

Epistatic interactions can play an important role in the genetic mechanisms that control phenotypic variation. However, identifying these interactions in high dimensional genomic data can be very challenging due to the large computational burden induced by the high volume of combinatorial tests that have to be performed to explore the entire search space. Random Forests Decision Trees are widely used in a variety of disciplines and are often said to detect interactions. However, Random Forests models do not explicitly detect variable interactions. Most Random Forests based methods that claim to detect interactions rely on different forms of variable importance measures that suffer when the interacting variables have very small or no marginal effects. The proposed Random Forests based method detects interactions using a two-stage approach and is computationally efficient. The approach is demonstrated and validated through its application on several simulated datasets representing different data structures with respect to genomic data and trait heritabilities. The method is also applied to two high dimensional genomics data sets to validate the approach. In both cases, the method results were used to identify several genes closely positioned to the interacting markers that showed strong biological potential for contributing to the genetic control for the respective traits tested. Contacthawlader.almamun@csiro.au