1146-1-t2v3c-blind_CNN_PoseProtocol.txt

Name

Blind CNN

Software

rdkit/gnina/smina

System Preparation Parameters

Assumed pH to be 7.4

System Preparation Method

I used RDkit to determine which ligand in Pocketome was closest to a given smile via the rdFMCS function to find the maximum common substructure (MCS).
This was determined by counting the number of bonds in the MCS and storing which ligand had the most.
This was done with the provided smile_comp.py script, and accompanying comp_jobs & bace_lig_smiles.txt files.
The output of the script gives the PDB id, PDB lig name, and SMARTS pattern of the MCS for each BACE ligand.
1 conformer was generated with rdconf.py (https://github.com/dkoes/rdkit-scripts/rdconf.py).
The aligned pose was generated via obfit to align the conformer to the MCS via the SMARTS pattern from above.
A CNN was trained with 1 round of iterative training as described by this poster: http://bits.csb.pitt.edu/files/gnina2018_poster.pdf
The CNNmodelfile and weights are provided with this submission (*.model and *.caffemodel respectively)

Pose Prediction Parameters

cnn_model=default2018_paul.model #caffe model file for our Convolutional Neural Network
cnn_weights=it1_.0_iter_1064000.caffemodel #caffe weights from 1 round of iterative training of cross-docked poses generated from Pocketome
autobox_add=20 $amount of angstroms to increase the search space generated by the autobox
num_modes=10 #maximum number of poses to generate (default 9)

Pose Prediction Method

Docking was performed smina(a fork of Autodock Vina) using the gnina software package with the above parameters (defaults for the rest), and using the CNN to rescore the given poses.
The top 5 poses are submitted with this protocol, as scored by the CNNscore.
The commands to run this protocol are given in the provided blind_cnn_jobs.txt file.
Note that the CNN method predicts an affinity, after using the CNNscore to evalutate the poses sampled by Autodock Vina

Answer 1

No

Answer 2

Yes