Fig. 2: Performance evaluation of ImageMol using the benchmark datasets.

The performance was evaluated in a variety of drug discovery tasks, including molecular properties (that is, drug metabolism, toxicity, brain penetration) and molecular target profiles (that is, HIV and BACE). a–c, FPR, false positive rate; TPR, true positive rate The AUC values are given in each panel. a, ROC curves of ImageMol across eight datasets (BBBP, Tox21, HIV, ClinTox, BACE, Side Effect Resource (SIDER), maximum unbiased validation (MUV) and ToxCast) with scaffold split and random scaffold split. b, ROC curves of Chemception⁴⁶ and ImageMol on HIV and Tox21 datasets with the same experimental set-up as for Chemception, which is a classical CNN for predicting molecular images. c, ROC curves of Chemception, ADMET-CNN¹², QSAR-CNN⁴⁷ and ImageMol on five CYP isoform validation sets (PubChem data set II). ADMET-CNN and QSAR-CNN are the latest molecular image-based drug discovery models. d, The ROC-AUC performance of sequence-based, graph-based and fingerprint-based models and ImageMol across six classification datasets (BBBP, Tox21, BACE, ClinTox, SIDER and ToxCast) with random scaffold split. For each type of method the maximum value is selected for display. e, The ROC-AUC performance across four regression datasets (FreeSolv, ESOL, lipophilicity (Lipo) and QM7) with random scaffold split. For each type of method the maximum value is selected for highlighting. For aesthetic presentation, the results of FreeSolv and QM7 are scaled down by a factor of 10 and 100, respectively. MAE, mean absolute error. f, The ROC-AUC performance of fingerprint-based (MACCS-based and FP4-based) methods and ImageMol across five major CYP isoform validation sets (PubChem data set II). g, The ROC-AUC performance of sequence-based and graph-based models across CYP450 datasets with balanced scaffold split.