Predictive power of randomly selected features.

We fit 500 models with 500 subsets of 25 randomly selected features from the uncorrelated 4800 features set. Test (top panel) and train (bottom) AUC values for the 500 models are shown. In green the AUC values for test and train obtained with the best performing models of the no population structure control (nPSC) pipeline, and in red the test and train values of the best performing models with population structure control (PSC). Randomly selected features can show high accuracy most likely due to the correlation of AET outcome and the phylogeny. (TIF)