Detection of epistatic interactions with Random Forest

Abstract

In order to elucidate the influence of genetic factors on phenotype variation, non-additive genetic interactions (i.e., epistasis) have to be taken into account. However, there is a lack of methods that can reliably detect such interactions, especially for quantitative traits. Random Forest was previously recognized as a powerful tool to identify the genetic variants that regulate trait variation, mainly due to its ability to take epistasis into account. However, although it can account for interactions, it does not specifically detect them. Therefore, we propose three approaches that extract interactions from a Random Forest by testing for specific signatures that arise from interactions, which we termed 'paired selection frequency', 'split asymmetry', and 'selection asymmetry'. Since they complement each other for different epistasis types, an ensemble method that combines the three approaches was also created. We evaluated our approaches on multiple simulated scenarios and two different real datasets from different \textit{Saccharomyces cerevisiae} crosses. We compared them to the commonly used exhaustive pair-wise linear model approach, as well as several two-stage approaches, where variants are pre-selected prior to interaction testing. The Random Forest-based methods presented here generally outperformed the other methods at identifying meaningful genetic interactions both in simulated and real data. Further examination of the results for the simulated and real datasets established how interactions are extracted from the Random Forest, and explained the performance differences between the methods. Thus, the approaches presented here extend the applicability of Random Forest for the genetic mapping of biological traits.