Fig 1.
Workflow of this research that includes datasets preparation, CYP2C9 structure, dynamics and ligand binding analyses, machine learning to train different predictive models and in vitro identification of new drug inhibitors of CYP2C9.
Fig 2.
Chemical space of the training and test sets as described by the principal component analysis (PCA).
The first two components, and their representation in % of the total variance are shown. (A). PCA of the training set’ inhibitors of PubChem vs. ChEMBL data sets. (B) PCA of the external test set’ inhibitors of PubChem vs. ChEMBL data sets. (C) PCA of the training and external test sets’ inhibitors. (D) PCA of the training and external test sets’ non-inhibitors.
Table 1.
Performances of the optimized RF models with MOE and IE descriptors on the training set (cross-validation CV) and the external validation set.
Table 2.
Performances of the optimized SVM models with MOE and IE descriptors on the training set (cross-validation CV) and the external validation set.
Table 3.
Comparison of the performances of the final RF and SVM models and other recent models on the external validation set.
Table 4.
Statistical significance (p-values) of the accuracy hypothesis between our final RF and SVM models and other recent models on the external validation set.
Fig 3.
Inhibition effect of sertindole, cloperidone, asapiprant, ticagrelor, duvelisib, dasatinib, bifeprofen and piriqualone on HepG2 cells expressing human CYP2C9.
The enzymatic activity with respect to the control is shown as a function of the drug concentration. HepG2 cells were treated with the drugs at the indicated concentrations for 24 h. Similar results were observed in three independent experiments. The bar graphs were obtained with GraphPadPrism v. 5.03 and represent mean ± SD of triplicate determinations.
Table 5.
In vitro inhibition assays of 18 tested compounds.
Fig 4.
Docking conformations of vatalanib and ticagrelor in the binding pocket of CYP2C9.
(A) Two poses of vatalanib (in salmon) docked into the crystal structure of CYP2C9 (PDB ID 5XXI) and the two co-crystallized molecules of losartan (PDB ID 5XXI) (in yellow). (B) The best pose of vatalanib (in salmon) docked into the MD5 structure of CYP2C9 and the superposed co-crystallized structure of losartan of the PDB ID 5XXI (in yellow). (C) The best pose of ticagrelor (in salmon) docked into the crystal structure of CYP2C9 PDB ID 5XXI and one of the two co-crystallized molecules of losartan (PDB ID 5XXI) (in yellow). (D) The best pose of ticagrelor (in salmon) docked into the MD4 structure of CYP2C9 and the superposed co-crystallized structure of the CYP2C9 inhibitor 2QJ (PDB ID 4NZ2) (in gray). Helices F and I of CYP2C9 are noted. The MD4 and MD5 structures correspond to CYP2C9 conformations generated from MD simulations of CYP2C9 bound to losartan and apo CYP2C9, respectively.
Fig 5.
Metabolism assays using recombinant CYP2C9 supersomes for the identification of CYP2C9-produced metabolites.
The suggested metabolite structures for abemaciclib, tarafenacin, cloperidone and vatalanib are shown.