Skip to main content

Table 3 AUC performance in the simulated target-prediction experiments

From: Accurate and efficient target prediction using a potency-sensitive influence-relevance voter

Method Average AUC (\(1\,\upmu \hbox {M}\)) Average AUC (\(5\,\upmu \hbox {M}\)) Average AUC (\(10\,\upmu \hbox {M}\))
Training without random negatives
 PS-IRV 0.88 0.84 0.83
 SVM 0.84 0.85 0.85
 RF 0.84 0.80 0.79
Training with random negatives
 PS-IRV 0.98 0.98 0.97
 SVM 0.98 0.98 0.98
 RF 0.98 0.98 0.98
  1. Models were trained using a tenfold cross-validation protocol and tested on the corresponding test set augmented with 9000 randomly selected ChEMBL molecules
  2. In the top panel, models were trained in the standard way, without random negatives. In the bottom panel, the training set was supplemented with 1000 random negatives
  3. Adding random negatives dramatically improves the performance of all methods
  4. Best results are in italics