Rationalization of the Performance and Target Dependence of Similarity Searching Incorporating Protein−Ligand Interaction Information

Interacting fragments (IFs) derived from protein-ligand complex crystal structures have previously been utilized to complement conventional two-dimensional (2D) similarity searching. In many instances, the (indirect) incorporation of three-dimensional (3D) interaction information through the use of IFs has further improved the search performance of fingerprints of unmodified ligands. However, for a number of targets, changes in the relative performance of conventional fingerprints and IF-based representations have also been observed, and reasons for these effects have thus far remained elusive. Herein, we analyze target-ligand systems that display different similarity search phenotypes. We study protein-ligand interactions and the resulting IF information at the molecular level of detail in order to better understand systematic differences in the search performance of unmodified ligands and IFs. The results show that the degree of conservation of IFs isolated from multiple crystallographic ligands is a major determinant of similarity search performance, regardless of whether IFs consist of coherent substructures or disjoint fragments. Conserved IFs focus similarity search calculations on 2D pharmacophore elements, as revealed by 3D interaction analysis. This leads to increased hit rates compared to unmodified reference ligand representations and to the recognition of smaller and less complex, yet structurally diverse hits. On the basis of these findings, one can predict for which targets the inclusion of IF information will likely result in improved 2D similarity search performance.