Fig. 1

Summary of the benchmarking results for the MoleculeNet datasets. Error bars represent the standard error of the mean (N = 50), while the asterisks denote whether the difference is significant (one indicates α < 0.05, two α < 0.01). The statistical tests with Bonferroni correction are carried out with respect to WCE or to the best performing loss function. We define the differences between loss functions within LightGBM as performance comparisons, while classifier comparisons refer to the benchmarking of the best loss function against the classifiers from Jiang et al. a Loss function comparison on the HIV dataset. b Comparison between the best loss function and the best models from Jiang et al. on the HIV dataset c Loss function comparison on the Tox21 dataset. d Comparison between the best loss function and the best models from Jiang et al. on the Tox21 dataset. e Loss function comparison on the MUV dataset. f Comparison between the best loss function and the best models from Jiang et al. on the MUV dataset