Fig. 5

AUC versus balanced accuracy of the models for the differently seeded networks. Each dot presents the external evaluation of one final model. All models were trained with the same conformation ensemble but different seeds for the initialization. The captions of the subplots denote the amount of oversampling done, e.g. 1–16 denotes 1 conformation for the negative class and 16 conformations for the positive class. The grey dots represent the models trained by different oversampling strategies, whereas the blue dots represent the 10 best models as reported in the Tox21 challenge leaderboard. The grey area depicts the area where both, the AUC and balanced accuracy, are higher than 0.5