Fig. 5

Demonstration of a simple structure–based drug design in REINVENT 4 using a crystal structure for PDK1 (PDB ID 2XCH). The cumulative number of hits identified over 50 epochs are shown a for reinforcement learning starting from the prior (RL, black) or from a transfer learning agent (TL-RL, red). The diversity of the hits generated is compared using principal component analysis (PCA) based on 2D RDKit descriptorsb and by counting the number of distinct hit and not–hit generic scaffolds c. For the PCA plot, we show hits as colored circles and include the convex hulls of all generated compounds as polygons b. d The predicted binding pose in the PDK1 binding site (based on PDB 2XCH) for the best scoring idea from each method are shown with a stick representation, contrasted with the native ligand in cyan. The docking scores for the poses are as follows: \(-10.1\) kcal/mol (RL) and \(-10.1\) kcal/mol (TL-RL). The protein is represented as a cartoon with key binding site residues (ALA 162/green, LYS 111/blue, GLU166/red, GLU209/red, ASN 210/blue) shown in a stick representation, with a transparent binding site surface overlaid. 2D inserts show the structure of the ligands. Hits are defined as molecules with a docking score \(\le -8\) kcal/mol and QED \(\ge 0.7\)