hvjjb-LigandScoringProtocol.txt

Name

orig_aff_aff

Software

smina/gnina

Parameters

--score_only --cnn_scoring --cnn_model ~/git/models/affinity/affinity.model --cnn_weights ~/git/models/affinity/affinity_full_iter_100000.caffemodel

Method


Protocols are named: [orig|crystals]_[bind|aff]_[bind|aff]
which corresponds to the pose prediction method (orig or crystal),
pose selection method, and pose scoring method.

orig:
smina was used to dock to a selection of reference
receptors (same poses as stage1).

crystals:
In addition to the original poses, smina was used to re-dock the provided
stage2 complexes. Additionally, smina was used to minimize the crystal pose
to generate a low rmsd pose at a local minimum of the Autodock Vina scoring function.

No cross-docking to these receptors was performed.

For both pose selection and scoring, and convolutional neural network
model was used to score the smina generated poses. This CNN model is
substantially improved from stage1 as it is now trained using the much
larger PDBbind database and includes an affinity prediction output in
addition to the bind/not-bind classification output (which is trained to
distinguish between low and high RMSD poses).

Pose selection
"bind":
For each ligand, across all the poses generated against all considered receptors,
the pose with the best binding classification output from the CNN model is selected.
"aff":
For each ligand, across all the poses generated against all considered receptors,
the pose with the best affinity prediction output from the CNN model is selected.


Pose scoring
"bind":
The selected ligand is scored using the binding classification output, which
ranges from 0 (bad) to 1 (good). When attempting to correlate this score
to binding affinities rather than ranking (i.e., Pearson or RMSD metrics),
a logit transformation should be applied first.

"aff":
The selected ligand is scored using the affinity prediction output.
This is a "pK" value where larger, more positive numbers indicate higher
affinity binders (e.g., 9 == 1nm). Although we label this a score, it is
appropriate to evaluate it using direct comparisons (e.g. RMSD).

hvjjb-PosePredictionProtocol.txt

Name

orig_aff_aff

Software

smina/gnina

System Preparation Parameters

This field intentionally left blank, but not really since the validator won't accept a blank field.

System Preparation Method

Waters were stripped from receptors. Protonation was performed with openbabel.

Pose Prediction Parameters

--exhaustiveness 50 --seed 0

Pose Prediction Method


Protocols are named: [orig|crystals]_[bind|aff]_[bind|aff]
which corresponds to the pose prediction method (orig or crystal),
pose selection method, and pose scoring method.

orig:
smina was used to dock to a selection of reference
receptors (same poses as stage1).

crystals:
In addition to the original poses, smina was used to re-dock the provided
stage2 complexes. Additionally, smina was used to minimize the crystal pose
to generate a low rmsd pose at a local minimum of the Autodock Vina scoring function.

No cross-docking to these receptors was performed.

For both pose selection and scoring, and convolutional neural network
model was used to score the smina generated poses. This CNN model is
substantially improved from stage1 as it is now trained using the much
larger PDBbind database and includes an affinity prediction output in
addition to the bind/not-bind classification output (which is trained to
distinguish between low and high RMSD poses).

Pose selection
"bind":
For each ligand, across all the poses generated against all considered receptors,
the pose with the best binding classification output from the CNN model is selected.
"aff":
For each ligand, across all the poses generated against all considered receptors,
the pose with the best affinity prediction output from the CNN model is selected.


Pose scoring
"bind":
The selected ligand is scored using the binding classification output, which
ranges from 0 (bad) to 1 (good). When attempting to correlate this score
to binding affinities rather than ranking (i.e., Pearson or RMSD metrics),
a logit transformation should be applied first.

"aff":
The selected ligand is scored using the affinity prediction output.
This is a "pK" value where larger, more positive numbers indicate higher
affinity binders (e.g., 9 == 1nm). Although we label this a score, it is
appropriate to evaluate it using direct comparisons (e.g. RMSD).