Home

Awesome

Analysis of Conformer Generation Procedures for Structure Based Drug Discovery

It is assumed RDKit is installed (e.g. pip install rdkit-pypi). We used version 2022.9.

Raw Data

Platinum 2017 dataset, downloaded March 10, 2023

wget https://comp3d.univie.ac.at/fileadmin/user_upload/p_comp3d/datasets/platinum-dataset-2017-01-sdf.zip
unzip platinum-dataset-2017-01-sdf.zip

Separate into indvidual files and generate smiles:

./split_platinum.py --sdf platinum_dataset_2017_01.sdf

PDBbind 2020 You will need to register and download the refined tarball at http://www.pdbbind.org.cn/download.php

tar xvfz PDBbind_v2020_refined.tar.gz 
./make_refined_smiles.py

Note that 77 structures failed to be parsed with RDKit. We ignore these (newer version of RDKit are less problematic).

COD (Crystallography Open Database) Original source is http://www.crystallography.net/cod/ Processed files (from https://chemrxiv.org/engage/chemrxiv/article-details/6356b98e2e0c63bd4d3a35f4) can be downloaded:

wget http://bits.csb.pitt.edu/files/cod.tgz
mkdir cod
cd cod
tar xvfz ../cod.tgz

DUDE Downloaded from https://dude.docking.org/

wget https://dude.docking.org/db/subsets/all/all.tar.gz
tar xvfz all.tar.gz
mv all DUDE

Crossdocking benchmark Downloaded from (original source https://doi.org/10.1002/pro.3784)

wget https://bits.csb.pitt.edu/files/gnina1.0_paper/crossdocked_ds_data.tar.gz
tar xvfz crossdocked_ds_data.tar.gz
./make_wierbowski_smiles.py
curl https://raw.githubusercontent.com/gnina/models/master/data/Gnina1.0/ds_cd_input_pairs.txt | sed 's/carlos_cd/wierbowski_cd/g' | sed 's/PDB_Structures\///g' > ds_cd_input_pairs.txt

RDKit Conformer Generation

First we randomly sample a max of 250 conformers for each smile. The commandlines can be generated with:

for f in platinum2017/*.smi refined-set/*/*.smi  wierbowski_cd/*/*.smi cod/*/*.smi
do 
 echo ./gen_rdkit_conformers.py --smi $f
done

If you are very patient you can remove the echo, or alternatively divide up the commands using your favorite batch processing technique (we use the genrd.slurm script to distribute across our SLURM cluster).

This results in files named 4EB8_rdkit_250_dg_unmin.sdf.gz and 4EB8_rdkit_250_etkdg_unmin.sdf.gz.
This script also uses the obrms utility from OpenBabel to calculate the RMSD of each generated conformer and puts the result in .txt file.

DMCG Conformer Generation

Checkout code and install (we are using commit 23e90f7231eac9a065093ee44b378cbf65181a7c).

git clone https://github.com/DirectMolecularConfGen/DMCG.git
cd DMCG
pip install .

You need to download the Large Drugs model weights. This file is called checkpoint_94.pt.

DMCG has several dependencies (e.g. torch_geometric, torch_sparse, torch_scatter and DGL) that may need to be installed. Make sure you can run the dmcg.py script on a test smile:

dmcg.py --smi test.smi --out test.sdf.gz --maxconfs 25 --eval-from checkpoint_94.pt

Setup generation commands the same as with rdkit:

for f in platinum2017/*.smi refined-set/*/*.smi wierbowski_cd/*/*.smi cod/*/*.smi
do 
 echo ./dmcg.py --smi $f  --out ${f%.smi}_dmcg_250_unmin.sdf.gz --maxconfs 250 --eval-from /tmp/checkpoint_94.pt
done

This script also creates a .txt file with the RMSDs to the reference calculated by obrmsd.

Energy Minimization

We next energy minimize the generated conformers using the UFF forcefield. The resulting _min.sdf.gz file is unordered.

for f in platinum2017/*unmin*sdf.gz refined-set/*/*unmin*.sdf.gz wierbowski_cd/*/*unmin*.sdf.gz cod*/*/*unmin.sdf.gz
do 
 echo ./minimize_sdf.py --sdf $f
done

RMSDs to the reference conformation (if available) and all cross RMSDS (the RMSD between every pose in the file with every other pose - this is used for filtering) are also calculated with obrms and put into .txt and .npy files. Note that a few cod conformers take a huge amount of time to calculate RMSDs for because of their symmetries.

Bioactive Conformation Analysis.

See the notebook bioactive.ipynb. We evaluate the effect the size of the conformational ensemble has on bioactive conformation identification as well as various methods for constructing ensembles (e.g., energy minimization and RMSD filtering).

Pharmacophore Matching

DUDE Conformer Generation

The gen_dude_conformers.py script will generate the following conformers:

#label actives
for f in DUDE/*/actives_final.ism
do 
  sed -i 's/$/_active/' $f
done

for i in DUDE/*/*.ism
do 
  echo ./gen_dude_conformers.py --smi $i
done

Pharmacophore Elucidation

Use Pharmit to identify the interacting features on the crystal ligand. This uses SMARTS expressions to identify features and distance and count based heuristics to enable them if they are interacting appropriately with the protein. Note we set each feature radius to 1.0 and remove any other constraints on the features.

For all identified interacting features, we then enumerate every possible combination with at least three features. Rather than worrying about identifying a single best pharmacophore, we will simply test them all and take the best result.

for i in DUDE/*
do 
  pharmit pharma -receptor $i/receptor.pdb -in $i/crystal_ligand.mol2 -out $i/pharmit.json
  ./genqueries.py $i/pharmit.json
done

Note: Using this procedure only two interaction features were identified for the cah2 target, so the remaining two non-interacting features were added to ensure there are at least 3 to choose from.

Pharmacophore Search

Setup parameters for building databases.

cd DUDE
for d in * ; do for i in 1 5 10 25 50 100 200; do echo "$d $i filtered_2.0"; echo "$d $i filtered_1.5"; echo "$d $i filtered_1.0"; echo "$d $i filtered_0.5"; echo "$d $i unranked"; done; done > params
mv params ../
cd ..

Count the number of conformers (sdsorter is available here: https://sourceforge.net/projects/sdsorter/)

for sdf in DUDE/*/*.sdf.gz
do
echo $sdf
sdsorter -printCnt $sdf > ${sdf%.sdf.gz}.cnt
sdsorter -reduceconfs 1 -printCnt $sdf >> ${sdf%.sdf.gz}.cnt
done

Build and search in parallel across a cluster.

sbatch -a 1-3605 get_pharma_res.slurm

Analysis

See pharmacophore_analysis.ipynb

Docking

First generate input conformations. We will limit ourselves to a maximum ensemble size of 5 due to the cost of docking. We generate these small ensembles:

for i in refined-set/*/unmin.sdf.gz wierbowski_cd//*unmin.sdf.gz do echo ./setup_docking_confs.py $i; done > makedock

Remove water and other hetatoms from receptor structures of PDBbind.

for i in refined-set/*/*_protein.pdb
 do
 grep -v HETATM $i > ${i%_protein.pdb}_rec.pdb
done

Create a file containing all the docking commands:

./setup_docking_cmds.py > alldock

We run this using dodock.slurm.

To perform docking with smina, can modify the command lines as follows:

sed 's/gnina/smina/g' alldock | sed 's/_docked/_sdocked/g' > salldock

Analysis

See docking_analysis.ipynb