Fig. 3

ROC data and accuracies for DREAM6, Model 1. a–d: ROC data for the four penalty functions applied: ℓ₁, Adaptive Lasso, ℓ_0.8 and Elastic net. The diagonal, black dashed line represents the characteristic that can be expected from a random classifier. The upper left corner of each ROC curve represents the optimal classification. The thin dotted lines show the changes with respect to the same run penalized by ℓ₁. Mean curves were omitted to focus on the differences between ℓ₁ and the extended methods. Both, Adaptive Lasso and ℓ_0.8 lead to mostly horizontal changes compared to ℓ₁, which corresponds to an increased specificity and a constant sensitivity. e–g: Absolute run-wise changes for extended methods vs. ℓ₁ and Adaptive Lasso against ℓ_q. Boxplot whiskers extend to twice the interquartile range at most. The boxplots emphasize the gain in specificity and consequently in accuracy