RDAforest: Identifying Environmental Drivers of Polygenic Adaptation

Identifying environmental gradients driving genetic adaptation is one of the major goals of ecological genomics. We present RDAforest, a methodology that leverages the predominantly polygenic nature of adaptation and harnesses the versatility of random forest regression to solve this problem. Instead of computing individual SNP-environment associations, RDAforest seeks to explain the overall genetic covariance structure based on multiple environmental predictors. By relying on random forest instead of linear regression, this method can detect non-linear and non-monotonous dependencies as well as all possible interactions between predictors. It also incorporates a novel procedure to select the best predictor out of several correlated ones, and uses jackknifing to model uncertainty of genetic structure determination. Lastly, our methodology incorporates delineation and plotting of “adaptive neighbourhoods”—areas on the landscape that are predicted to harbour differentially adapted individuals. Such maps can be used as a guide for planning conservation and ecological restoration efforts. We demonstrate the use of RDAforest in two simulated scenarios and one real dataset (North American grey wolves).