CNN docking with affinity model ranking
docking performed with gnina commit b3fa6ae13fc6b42924f49b2d751d68f1bc14bc08 available from https://github.com/gnina/gnina, conformer generation performed with rdkit via https://github.com/dkoes/rdkit-scripts/rdconf.py.
rdconf.py was used with the defaults except --maxconfs=1 was specified. Water was left in the receptor structure for docking.
for each SMILES input, a single conformer was generated via rdkit using the UFF force field. All compounds were docked by gnina using the default convolutional neural network scoring function.
Reference ligands were used to define a box, and then 8 Angstroms of padding were added to each face of that box to define the search space for docking. The random seed was set to 0 and the exhaustiveness was set to 50. The default number of final conformers was used, which is 9. --gpu and --cnn_scoring were enabled. Compounds were ranked by the affinity predicted by the default convolutional neural network affinity model. To be added to the final prediction set, a pose was required to differ from poses already in the set by more than 0.5 RMSD in order to ensure pose diversity.
Compounds were docked using gnina's default convolutional neural network model. Docked poses were ranked by their predicted affinity. The best-scoring pose was chosen, and subsequent poses were chosen in order of rank if they differed from existing poses in the chosen set by greater than 0.5 RMSD until at most 5 poses were chosen.
No
Yes