Fig. 2

Explanation of the GCN model’s prediction of lapatinib and other inhibitors in EGFR, JAK, and PIM models. (A) Grad-CAM preferences of lapatinib from the latest graph convolutional layer for both positive and negative classes. Circles are centered at each atom, with green ones for the positive class and orange for the negative class. The larger the circle, the more the atom contributes to the prediction of the model at a specific class. (B) Preferences for different inhibitors within and across families. Within the same family, conserved attention on similar environments is visualized, and family-sensitive pre-moieties can be seen by comparing cross-family inhibitors. (C) Crystallized complexes of the pan-EGFR inhibitor lapatinib (deep blue, PDB ID: 1XKK), pan-JAK inhibitor tofacitinib (light blue, PDB ID: 3EYG), and pan-PIM inhibitor LGH-447 (purple, PDB ID: 5DWR) demonstrated three different modes of kinase inhibition