Fig. 3

Performance of feature selection methods on simulated data with low gene expression. Noise was introduced to mimic genes with no or very low expression, representing (a) 10% or (b) 30% of the total features. Performance of a random forest classifier was evaluated using classification accuracy on the test set as well as the percentage of identified gene pairs that contained at least one signal gene. All simulations were performed five times and data are presented as mean ± SEM