1000 Model Poses Pdf Download
Download === https://urlin.us/2tqQ4A
1000 Model Poses Pdf Download
Besides the ML-based approaches mentioned above, we also utilized several classical methods for comparison, including the docking scores from Surflex-Dock (or AutoDock Vina or Glide SP), the Vina scores extracted from the NNscore features, empirical SF X-Score  and more robust Prime-MM/GBSA . For X-Score, the FixPDB and FixMol2 utilities were first utilized to prepare the protein and ligand files, respectively, and then the average score of the three individual SFs available in X-Score was employed for rescoring the binding poses. Prime-MM/GBSA was executed with the prime_mmgbsa utility implemented in Schrödinger. The rescoring was conducted with the variable-dielectric generalized Born (VSGB) solvation model and OPLS2005 force field.
The top1 and top3 success rates of the models trained on PDBbind-ReDocked-Refined and tested on PDBbind-CrossDocked-Core-s, based on A all poses, B re-docked poses, and C cross-docked poses. The error bars represent the standard deviation of the random sampling of 1000 redundant copies with replacements, and the dotted line indicates the ceiling of the success rate
The top1 and top3 success rates of the models trained on PDBbind-ReDocked-Refined and tested on the A cross-docked and D re-docked poses in PDBbind-CrossDocked-Core-s, the B cross-docked and E re-docked poses in PDBbind-CrossDocked-Core-g, and the C cross-docked and F re-docked pose in PDBbind-CrossDocked-Core-v. The error bars represent the standard deviation of the random sampling of 1000 redundant copies with replacements, and the dotted line indicates the ceiling of the success rate
Taken together, it seems that those ML-models trained on the re-docked poses can be well generalized to the re-docked or cross-docked poses generated by the same docking program. For the pose space defined by other docking programs, their performance is limited, especially for the predictions of cross-docked poses. Hence, a feasible strategy is to enlarge the training set, either through the augmentation and the diversification of the pose space for a certain complex or through the involvement of more complexes in the training set.
To address the issue left in the previous section, we try to enlarge our training set by introducing the cross-docked poses into the training set, thus creating the PDBbind-CrossDocked-Refined set. At first, we also tried to include the native pose of each cross-docked complex (cross-native pose), which was generated through the alignment of two crystal structures regardless of the possible steric conflicts, in the training set just as we have conducted for the re-docked poses. Although they exert little effects on other test sets (e.g., PDBbind-ReDocked-Core), the performance on CASF-docking is awfully poor, as shown as Fig. 7. Except ECIF_XGB, the other models can gain prominent improvements when removing those native poses from the training set, suggesting that these models have learnt incorrect information from the cross-native poses. We guess that two reasons may majorly account for the higher sensitivity of this dataset to these incorrect cross-native poses. Firstly, the poses in CASF-docking were manually preprocessed and clustered, so that they are uniformly distributed in each RMSD window; secondly, CASF-docking owns more poses for a certain complex than the other test sets and even the training set (at most 100 vs at most 20). As for the minor influence on the ECIF features, we guess that this type of pure atomic pairwise counts-based features may be insusceptible to the possible conflicts between the protein and ligand because it only relies on the counts within the predefined distance, while the NNscore (containing several interaction-pairwise counts) and Vina (some physics-based energy terms) features are obviously not the case. Anyway, we will not include these cross-native poses in the following experiments, and we also do not recommend this type of poses to be invo