Target-specific native/decoy pose classifier improves the accuracy of ligand ranking in the CSAR 2013 benchmark.

As part of the CSAR 2013 benchmark exercise, we have implemented a hybrid docking and scoring workflow to rank 10 steroid ligands of an engineered digoxigenin-binding protein. Schrödinger's Glide docking software was used to generate poses for each steroid ligand and rank them according to both standard docking precision (SP) and extra docking precision (XP) scoring functions. The unique component of our approach was the use of a target-specific pose classifier trained to discriminate nativelike from decoy poses. To build the classifier, a single cognate ligand with a known native pose (PDB code 4J8T) was docked multiple times into its target protein, and the generated poses were divided into two classes (nativelike and decoy) using a root-mean-square deviation threshold of 2 Å. All of the poses were characterized by the MCT-Tess descriptors of the protein-ligand interface, and random forest (RF) models were trained to discriminate the two classes of poses on the basis of their descriptors. The consensus pose classifier was then applied to the Glide-generated poses of each CSAR ligand in order to filter out those poses predicted as decoys and rerank the remaining ones using both XP and SP scoring functions. The best-scoring pose for each ligand following this filtering step was used for final ligand ranking. Overall, the ranking accuracy for the 10 ligands evaluated by the Spearman correlation coefficient was 0.64 for SP and 0.52 for XP but reached 0.75 for SP/RF consensus scoring (ranked third in the CSAR 2013 benchmark exercise). This study reconfirms that target-specific pose scoring models are capable of enhancing the reliability of structure-based molecular docking by discarding decoy poses.