Fragment-Based Analysis of Ligand Dockings Improves Classification of Actives

We describe ADChemCast, a method for using results from virtual screening to create a richer representation of a target binding site, which may be used to improve ranking of compounds and characterize the determinants of ligand receptor specificity. ADChemCast clusters docked conformations of ligands based on shared pairwise receptor ligand interactions within chemically similar structural fragments, building a set of attributes characteristic of binders and nonbinders. Machine learning is then used to build rules from the most informational attributes for use in reranking of compounds. In this report, we use ADChemCast to improve the ranking of compounds in 11 diverse proteins from the Database of Useful Decoys-Enhanced (DUD-E) and demonstrate the utility of the method for characterizing relevant binding attributes in HIV reverse transcriptase.