Feature-similarity protein classifier as a ligand engineering tool

Kinases have been often targeted in drug therapy aimed at blocking signaling pathways. However, the conservation of protein structure across homologs often leads to uncontrolled cross-reactivity. On the other hand, sticky packing defects in proteins are typically not conserved across homologs, making them ligand-anchoring sites potentially important to enhance selectivity. Thus, we introduce a hierarchical clustering of PDB-reported kinases according to packing differences. This kinome partitioning is highly correlated with proximity relations arising from the pharmacological profiling of kinases. A variable packing sensitivity is observed for individual drugs, with highly promiscuous ligands being the most insensitive to packing differences. Our classifier enables a strategy to design selective inhibitors.