(A) The best-ranked pose for PPARγ (PDB code: 2PRG) is presented, where the native ligand poses are in gray and the predicted ligand poses are in color. The MedusaScore (B) and ranking energy (C) are plotted against the corresponding RMSD from the native pose. The MedusaScore alone cannot distinguish the near-native poses from the decoys. For example, the pose with the lowest MedusaScore has an RMSD larger than 10 Å. Using the ranking energy (Methods), we are able to separate near-native poses from decoys. For the LXR β ligand binding domain (PDB code: 1PQC), the predicted pose (D), the MedusaScore (E), and the ranking energy (F) are shown. By ranking the poses using the ranking energy, the near-native poses can be unambiguously selected. (G) The best-ranked pose (colored) and the native pose (gray) of Estrogen receptor (PDB code: 3ERT) are partially symmetric. For P38 MAP Kinase (PDB code: 1BMK), the native ligand pose (H) and the predicted best-ranked pose (I) show distinct interactions between ligand and receptor.

The Percentage of recovered known binding ligands is plotted as a function of percentage of library screened using various VS strategies. For reference, we also plot the enrichment curves for ideal VS, where all ligands are ranked ahead of decoys, and random VS, where all molecules are ranked randomly. For VEGFr2, we find all the 15 top-raked ligands are known VEGFr2 binders (ideal VS performance). For the other targets, there are seven, six, and three known binders among the top 10 ligands for CDK2, HIVPR, and HIVRT, respectively.

(A) The ligand binding surface (PDB ID: 1UVS) features several deep sub-pockets. Successful prediction requires simultaneous fitting of the ligand to all sub-pockets. (B) The schematic energy landscape with and without VDW repulsion between the ligand and receptor side chains. By turning off the VDW repulsion between ligand and receptor side chains, the rugged energy landscape is smoothed.

The input ligand conformation has ideal bond lengths and angles and only dihedral angles of rotatable bonds are changed during the procedure. To evaluate whether a ligand conformation has clashes, we only compute the VDW repulsions between atom pairs whose pairwise distances are determined by at least one rotatable bond. Two atoms are denoted as non-local if their pairwise distance is determined by more than one rotatable bond and as local if their distance is governed by only one rotatable bond. For each local atom pair, we determine the minimal VDW repulsion energy by enumerating the corresponding dihedral angle. We consider a non-local atom pair clashing if the VDW repulsion energy is larger than 0.6 kcal/mol, and a local atom pair clashing if the VDW repulsion energy is at least 2.4 kcal/mol more than its minimum VDW repulsion energy. A carbon-carbon pair (both local and non-local) is considered clashing if their distance is bellow 3 Å. A ligand conformation is determined as clashing if one or more atom pairs are clashing. For ligands with small degrees of freedom, we enumerate all possible non-clashing conformations. For ligands with large degrees of freedom, we stop the rotamer generation once either the maximum number of rotamers (set to 1000) is generated or the total number of trials (set to 10⁶) is reached.

Each docking simulation starts with the generation of the STROLL, followed by the clustering of ligand rotamers (Methods) with a kRMSD cutoff of 2 Å. The N_C cluster centroids are used as the representative ligand conformations for the initial coarse-docking (CD), where the ligand is kept rigid and the ligand-binding pocket is rapidly sampled (Methods). Since only a small number of the ligand rotamers is used and the ligand is kept rigid during coarse-docking, the near-native pose is not necessarily the lowest energy pose, but usually among the low-energy poses. We sort the N_C poses according to binding energies. We group similar poses together if the RMSD of two ligand poses is smaller than their kRMSD plus 2 Å. The grouping of similar poses will reduce the necessary number of the more expensive calculations of fine-docking. We select the top N_F (~10% N_C) groups of lowest energy poses for the next round of fine-docking (FD). For each group of coarse-docked poses, we perform fine-docking to enrich the ligand rotamers and to minimize the total energy (Methods). During fine-docking, we add the ligand rotamers within 2.0 Å kRMSD to the initial centroid rotamer in order to enrich the ligand conformation. For each of the members in a group, we perform the enrichment step. We only select the lowest energy poses for further minimization (Method). Therefore, we will have N_F poses for each MedusaDock run.

For 1PMN ligand docked to 1PMV receptor, the backbone has severe clashes with the native pose (A). For 1PQC ligand docked to 1PQ6 receptor (B), the lowest-RMSD pose (colored) has a RMSD of only 2.2 Å. (C) The MedusaScore of the best-ranked pose (BRP) in crossing-docking is compared to that of self-docking. (D) In cross-docking, the MedusaScore of the best-ranked near-native pose (BRNN) is also close to that of the best-ranked poses.

(A) Inclusion of the native rotamer in STROLL increases the number of near-native poses sampled by MeduasDock. (B) The prediction accuracy in terms of ranking near-native poses as the top one, top 10, and sampled does not depends on whether the native ligand rotamers are included. (C) The histogram of the number of near-native poses sampled. (D) The MedusaScore of the best-ranked poses from self-docking simulations is close to that of the minimized poses around the input X-ray crystallographic structure.

During coarse-docking, the ligand is kept fixed, with only rigid-body motion allowed. Iterative receptor side chain repacking and rigid-body docking are performed to identify the lowest binding energy pose.