Chemically similar drugs often bind biologically diverse protein targets, and proteins with similar sequences or structures do not always recognize the same ligands. How can we uncover the pharmacological relationships among proteins, when drugs may bind them in defiance of bioinformatic criteria? Here we consider a technique that quantitatively relates proteins based on the chemical similarity of their ligands. Starting with tens of thousands of ligands organized into sets for hundreds of drug targets, we calculated the similarity among sets using ligand topology. We developed a statistical model to rank the resulting scores, which were then expressed in minimum spanning trees. We have shown that biologically sensible groups of targets emerged from these maps, as well as experimentally validated predictions of drug off-target effects.