Fig. 6.

A. ROC-AUC curve showing the cross-validation performance of the logistic regression model to predict the state (active/inactive) of cyclin-dependent kinases (CDKs). This model is trained on structural residue features (bond angles, residue depths, fluctuations, and solvent accessibilities), and aligned according to Caretta’s multiple structure alignment of the CDKs. B. Structure of active proline-rich tyrosine kinase 2 (PYK2), co-crystallized with the ATP$\gamma$S ATP-mimetic kinase inhibitor. The catalytically important DFG motif is labelled and the fifteen most predictive residues for active/inactive state determination are colored in blue and represented as sticks.