Multiple initial conformations re-scored by Convex-PL with a regression model (solvent modeled with points on a grid) and automatically chosen by similarity.
AutoDock Vina with in-house modifications, RDKit 2017, SciPy, PyMOL 1.8.4, unreleased version of Convex-PL scoring function
1000 conformations for each molecule in RDKit, sequence identity for proteins >= 99
BACE 1 - 19 conformations were generated in RDKit and clusterized with hierarchical clustering. Each 100 clusters were selected for docking. Receptors downloaded from RCSB with help of their search by sequence feature.
Vina exhaustiveness 10, number of poses 350-450, Convex-PL rescoring with typization system 14 and solvent model with points on a grid, 0.5A clusterization. Regression part of Convex-PL was trained with 5.2 cutoff of the scores and with regularization 7000. Poses with rmsd > 3.0 between the atoms of maximum common substructures were rejected (if there were poses with rmsd < 3.0).
We ran docking experiments with 100 conformations of each ligand except BACE_20, and did the rescoring with plain version of Convex-PL.
No
Yes