Graph convolutional network-based structure-activity relationships of wild type and T315I mutant ABL kinase inhibitors

Mutations in the BCR-ABL kinase are a major obstacle for designing new drugs against Chronic Myeloid Leukemia. Particularly the T315I mutation is responsible for 15−20% of all clinically relevant mutations1. T315I is capable of modifying the ATP-binding pocket geometry, interrupting critical protein−drug interactions required for inhibitory activity1. Classification models using Graph Convolutional-Based Neural Network (kGCN)2 were generated for the activity of known compounds against the wild type and mutant ABL1,3. The models produced during the hyperparameters optimization were validated in a5-fold cross-validation and evaluated using the area under the curve of a ROC curve, Matthews correlation coefficient and Accuracy. The selected model for each endpoint were considered predictive (external validations: AUC = 0.916/1.00; MCC = 0.750/1.00; ACC = 0.875/1.00 respectively for the wild type and mutant). The SAR analysis highlighted the importance of the change from benzimidazole to benzothiazole and the substitutions in the benzene ring.