Structure-Based_docking
Scripts to handle structure-based molecular docking, single-receptor virtual screening, consensus results of multiple receptor docking, and etc.
There are 2 major folders with primary running scripts (A_docking_scripts and B_collect_scripts), 1 folder with miscellaneous scripts (Z_misc_scripts), and 1 folder with example scripts to run those primary scripts (Examples) :
/A_docking_scripts -- Setting up and running of docking
|------ /1_prep_ligands - scripts to prepare ligand for docking
|------ /2_virtual_screen - scripts to run docking
|------ /3_conformer_gen - generate 3D conformers
|------ /CSD_confgen - CSD's generator
/B_collect_scripts -- Collecting the docking results
/Z_misc_scripts -- collection of less used scripts to manage docking results
/Examples
|------ /1_prep_ligands - prepare ligand library
|------ /2_virtual_screen - running docking
|------ /3_single_VS_rslt - collect docking result for 1 structure
|------ /4_consensus_VS_rslt - consensus of multiple docking results
|------ /5_extract_VS_data - extract a subset of ligands from library
|------ /6_print_select_2d - print out docking result in 2D
|------ /7_search_analogs_substruct - substructure search with SMILES patterns
|------ /8_search_analogs_FP - search SMILES patterns with fingerprints
#######################################################################################
Steps to Prepare Ligand Library for Docking - 19.07.28
Prepare initial ligand library
-
- Download library subsets in SMILES format.
-
- check total number of ligands, collect all SMILES from different subsets into one single file, the shuffle the SMILES to randomize the order
> cat */*/*smi | grep -v 'smiles' > all.smi
> sort -uR all.smi > shuffled.smi
> cat shuffled.smi | wc -l
-
- to the shuffled SMILES file, split it into a number of library subsets with each subset has ~100,000 ligands (a '.' is needed before 'smi' and after 'new_name'). This subset number will need to be precalculated by knowing how many ligands are there (from previous step, 'cat shuffled.smi | wc -l')
> split --numeric-suffixes=1 \
--additional-suffix=".smi" \
--number=30 \
shuffled.smi "new_name."
one issue with these shuffled SMILES files -- the first line is somehow always missing some characters, rendering the first line (molecule) of all smiles files to be bad and any conversion will fail. But that only affect 75 molecules out of ~ 6M, ignore for now.
-
for lead-like (300 < MW <= 400) and fragment-like (MW <= 300) libraries, keep each subset file at 100,000 mol max.
-
for drug-like (MW > 400) library, keep each subset file at 80,000 mol max.
-
The above procedures are now summarized in a shell script:
> lig_library_split.csh
[ input smiles file ]
[ output prefix ]
[ maximum smiles in each piece ]
e.g. x.csh lead.smi enm19.lead 100000
- If the available library is only in SDF format, the following script will clean it up to remove known salts and split into shuffled fragment/lead/drug-like sets of SMILES files based on molecular weight.
- Beware of the size of the SDF library, since reading into RDKit will blow up the size alot.
> sdf_library_2_smi.py
[ sdf filename ]
> generate:
- lead.smi (300 < l <= 400) | drug.smi (> 400)
- frag.smi (150 < f <= 300) | small.smi (<=150)
- .csv (physical properties of molecules)
Warning: SDF file should not exceed 1.3GB; that will use ~ 25GB RAM
#################################
Prepare ligand library for Glide docking
-
from Schoridnger/2016-03+, the '-r 1' flag is decrepated. Before, it controls the ring conformation "add input ring conformation if available". 18.08.22 added property filter to remove useless compds (known reactives and PAINS). while removing known reactives (acyl chloride, aldehyde, etc) is useful for fragment-like library, there are very few reactives in lead-like and drug-like libraries.
-
This setup will use OPLS3e force field to generate the 3D structure for the input ligand. Use the universal SDF.gz format. It generates one(1) most preferred charged state/tautomer for each ligand at pH=7.2 +/- 0.3, becuase most screening are against targets in the cytosol (pH~7.2) or blood/extraceullar domain (pH~7.4). For targets in specialized organelles (peroxisome, mitocondria, lysosome, etc) that have lower pH, a new set of ligands with charged state/tautomer appropiate for that pH should be generated. It also generates the stereoismers, tho it is recommended to use input with pre-defined chiral centers available.
-
Athough removing PAIN compounds is a good idea, the SMARTS patterns used in the substructure recognition are greedily "dirty"(1) (2) (3) and can flag down okay compounds as false positives. This has happened to the lead compound in Dyrk1A type-II inhibitor project, which turns out it is a false positive. Though I now do not recommend doing a PAINS pre-filtering, it is better to understand the key features (1) (2) and pick out potential PAINS by eye in the hit-pick stage, then suggest watching these guys out during testing in orthogonal assays.
time ${SCHRODINGER}/ligprep -WAIT -LOCAL \
-i 2 -epik -We,-ph,7.2,-pht,0.3,-ms,1 \
-s 1 -t 1 \
-bff 16 \
# -f $homedir/all-1.sch.propt_filter.cflt \
-ismi $smi_dir/SDIN.smi \
# -isd /sc/hydra/projects/schlea02a/8_lib/zinc_lead13/glide/SDIN.sch.sdf.gz \
-osd SDIN.sch.sdf.gz \
-HOST localhost:$schcpu \
-NJOBS $schnjob
echo $!
## Use Schrodinger Canvas to filter certain pains ligands
# $SCHRODINGER/utilities/canvasSearch -WAIT -LOCAL \
# -isd SDIN.sdf.gz \
# -osd SDIN.sch.sdf.gz \
# -osd2 SDIN.fail.sdf.gz -filter \
# -file $homedir/all-1.sch.pains_filter.cflt \
# -JOB JBNAME \
# -HOST localhost:$schcpu
- To run this in batch mode on a HPC cluster, use the shell script run_ligprep_lsf.csh to modify the template LSF script and submit the jobs:
> A_docking_scripts
|---- /1_prep_ligands
|---- run_ligprep_lsf.csh
|---- ligprep.tmpl.lsf
#################################
Prepare ligand library for OpenEye Docking
- OpenEye preparation requires multiple steps. First with 'fixpka' to generate one(1) most relevant tautomoer/charged state for the ligand as pH=7.4, then 'flipper' to generate multiple stereoisomers (tho it is recommended to use starting ligands with their chiral center predefined), then use 'omega2' to generate conformer library. For conformer generation, recommend to get -maxconfs 300, since for larger molecules (like sorafenib) the default 200 maximum conformer is not enough to cover the conformational space of ligands with more rotatable bonds.
- Although OMEGA uses the older MMFF94 force field with electrostatics modification by the original developers of force fields (Bayly, Darden are with OE!), the generated small molecule conformations are actually similar to crystallographic poses and appropiate for use in docking. For atoms with unique hybridization state (e.g. "the almost aromatic" ring NC1=NN=CCS1 in lead compound in Dyrk1A type-II inhibitor), conformation will look bad in comparison to those generated by the all-new OPLS3e or QM calculation. But given the extreme efficiency and quality for vast majority of the ligands, results from OMEGA is still very much usable.
fixpka -in input.smi -out input.oe.smi
flipper -in input.oe.smi -out input.oe.ism
omega2 \
-mpi_np $cpu \
-in input.oe.ism \
-out input.oeb.gz \
-prefix prefix \
-maxconfs 300 \
-progress none \
-strict false
- To run this in batch mode on a HPC cluster, use the shell script run_omega_lsf.csh to modify the template LSF script and submit the jobs:
> A_docking_scripts
|---- /1_prep_ligands
|---- run_omega_lsf.csh
|---- omega.tmpl.lsf
#######################################################################################
Docking
> ${OPENEYE}/bin/make_receptor
> ${OPENEYE}/bin/fred -mpi_np $cpu
-
OpenEye docking has 2 steps, (1) generating the rigid receptor grid file, and (2) rigid docking of ligand conformers to receptor grid. For the most parts the default settings by OpenEye will work for 95% of the cases becuase they really really optimized everything. As for the scoring function ChemGauss4, it is a combination of a knowledge-based Gaussian potential for shape complementary plus a modified ChemScore/PFP/GoldScore(?) empirical-based chemical feature scoring.
-
Grid generation - this is done through the command-line GUI make_receptor. Default setting will work for almost all cases. May want to incresae the size of the docking pocket. Can setup directional H-bond (donor or acceptor) or SMARTS-defined spherical constraints if these interactions are known and critical.
-
Docking - this is done through the command-line program fred. The default setup will work for 99% of the case. Write out the docking results to sdf.gz format, not the oeb.gz format. Be sure to change the -hitlist_size to 0 (save all), otherwise only 500 (default) will be saved and the rest are lost.
-
To run this in batch mode on a HPC cluster, use the shell script fred_submit_lsf.csh to modify the template LSF script and submit the jobs:
> A_docking_scripts
|---- /2_virtual_screen
|---- fred_submit_lsf.csh
|---- fred_docking.repeat.lsf
#################################
> ${SCHRODINGER}/glide glide-grid_generation.inp
> ${SCHRODIMGER}/glide glide-dock.inp
-
Schrodinger's Glide docking has 2 steps, (1) generating the rigid receptor grid file, and (2) flexible ligand docking to receptor grid. Most of the inital setup is done through Schrodinger's Maestro GUI, but certain settings can only be done via altering the input files written out by the GUI, and then run the modified input files in command-line mode. As for the scoring function GlideScore, it is an empirical-based scoring function, plus the options to consider H-bond-like Aromatic H and Halogen bonding interactions.
-
Grid generation - the default vdw scaling of 1.0x is alright, but I have found a softened vdw scaling (0.75x - 0.80x) to be optimal for most cases (kinases, transporers, etc.). But really, need to try a couple vdw scalings (0.60x - 1.20x, in 0.05 increment) on knonwn ligands to test what is the best vdw scaling factor for the particular receptor. If there is no knonwn ligand available, then a 0.80x vdw scaling is a safe bet. Cannot edit this in the GUI, have to edit it in the input text file written out by the GUI. Always turn on Aromatic Hand Halogen bonds options. If you know certain interaction must happen, e.g. kinase inhibitors H-bonding the backbone amide of hinge residue, type-II kinase inhibitors occupying the hydrophobic DFG-pocket, then you can setup contraint location (spherical, or directional for H-bond) and preferred ligand types in SMARTS patterns.
-
Docking - the default Standard Precision (SP) setting works for most cases. Do not use High-throughput virtual screen (HTVS) setting to "save time" as the manual recommended. Tried it and this setting failed to capture many known ligands even in the top 50% ranking. Extreme Precision (XP) is useless and slow for most cases as it uses very harsh and rigid vdw scaling, which fails in most cases too. Only ever use it if you are working on a congeneric series of a scaffold that is developed based on a crystal structure. Always turn on Aromatic H and Halogen bonds options. Can use constraints previously defined in Grid generation by requiring X of n constraints must be fulfilled. Also, save the docking results with no receptor pose, and in SDF format, don't use MAE format.
-
To run this in batch mode on a HPC cluster, use the shell script glide_submit_lsf.csh to modify the template LSF script and submit the jobs. Schrödinger does not write out a readable score file (.rept) like it used to after ~ ver.2016-09, and the pose files are in .gz format. The data will need to be cleaned up to get the more compact .bz2 format and to extract the scores from the pose files. Run the glide_clean_folder.csh with a list of the docking folder name (concatinate the *.clean files generated during the initial submission step) to get the /1_data folder with all the cleaned data.
> A_docking_scripts
|---- /2_virtual_screen
|---- glide_submit_lsf.csh # control glide submission
|---- glide-dock.template.lsf # docking LSF template
|---- glide-dock.HTVS_SP.template.in # docking input template
|---- glide-grid_template.in # grid input template
|
|---- glide_clean_folder.csh # summarize batch docking folders
|---- glide_clean_score.py # make readable Glide score file
#################################
> ${SCHRODINGER}/ifd
-
Induced-Fit Docking (IFD) from Schrodinger performs flexible ligand docking to flexible receptor. It combines softened docking with receptor minimization, and optional loop modeling. Most of the setup is done through Maestro IFD GUI.
-
The initial rigid docking uses a much softened vdw radii (recommend 0.5-0.6x vdw scaling), followed by receptor "side chain" minimization, followed by regular rigid Glide docking (recommend 0.8x vdw scaling) and rescoring for the final result. User can define which residue's side chain can be minimized (default: those within 5A distance from initial ligand docking pose), or user-defined residues. User can also designate which residue side chain is "removed" in the initial docking to avoid steric hindrance or overwhelming bias in the intial docking (see the supplementary of article (Lazarus Ung JACS 2019 ULK4), where IFD for prodrug R788 required such maneuver).
-
Also, a hidden option to use "Flexible Loops" is accessible by modifying the input text file written out by IFD. This is described in the manual but is not available in the Meastro IFD GUI. It is useful for cases where binding site involves a flexible loop, e.g. P-loop of protein kinase.
#######################################################################################
Collecting Results of Standard Rigid-Receptor Docking
> B_collect_scripts/5_general_screen_get_top.py
-score [ score file(s) ] # accept multiple files "*.txt,x.txt"
-sdf [ sdf files(s) ] # same prefix as files in -score
-top [ number of top mol to output: int ]
-dock [ docking software: fred | sch | etc ]
-outpref [ output prefix for sdf, png, txt files ]
Optional:
-hmax [ histogram max score: Def: < fred: -14 | sch: -10 > ]
-hmin [ histogram min score: Def: < fred: -2 | sch: -3 > ]
-coll [ collect X times top MOL in memory (Def: 2) ]
-exclude [ <SMARTS pattern> (smt-clean) ] removal filter
-select [ <SMARTS pattern> (smt-selec) ] selection filter
e.g.> *.py -score "*docked.txt.bz2" -sdf "*docked.sdf.bz2"
-top 1000 -dock sch -outpref single_VS.summary
-hmax "-16.0" -hmin "-2.0" # need double quote to worl
-exclude 'C(=O)[O-]|S(=O)(=O)[O-]|P(=O)(O)[O-]' # no acidic moieties
result > single_VS.summary.sch_top1k.sdf, single_VS.summary.sch_top1k.txt,
single_VS.summary.sch_top1k.histo.png
- Collect the top scoring results from one or multiple docking files and generate a histogram to show the distribution of the docking scores (.png). The score files (.txt) and the docking pose files (.sdf) can be in zipped format but must have the same prefix, i.e.:
test.sch_docked.txt.bz2 test.sch_docked.sdf.bz2
#################################
> B_collect_scripts/7_general_docking_cluster.py
[ Docking input(s): sdf ] - accept multiple files "*.sdf,x.sdf"
[ Tanimoto cutoff: float ] - 0.4 works for lead-/drug-like molecules
[ output prefix ]
[ receptor structure: pdb ]
Optional:
[ -top <no. of mol>: int ]
[ -dl DayLight Fingerprint
-ec ECFP_4 Fingerprint (Default)
-ms MACCS 166-bit Key ]
# also output a table of clustered molecules in PDF
e.g.> *.py vina.sdf,fred.sdf,ehits.sdf
0.4 output.clust-04 recpt.pdb.bz2 -dl
result > output.clust-04.sdf, output.clust-04.pse,
output.clust-04.pdf
- Cluster the supplied molecules using a fingerprint-based chemical similarity method with a user-defined Tanimoto cutoff value. A value of 0.40 works for most cases of lead-/drug-like molecules (MW 300-500+), while fragment-like may need a lower value, e.g. 0.37. The clustered results are placed into a PyMOL session file and the docking receptor is used. The clustered molecules are also output into a PDF file.
#################################
> B_collect_scripts/9_consensus_best_pose.py
[ List of Score Files ] ** SDF and Score files must have same prefix
[ Score Column in Score File ] ** Standard numbering
[ Output Prefix ]
[ Threshold Top Ranking to Count: int % ]
[ Threshold Number of model for Consensus Result: int ]
[ Running mode ]
## [ Running mode ]: One of the following options, -c is recommended
# Analyze ALL models
[-a single best rank]
# Minimal number of model to pass the threhold number and % ranking
[-b average of rank | -c single best rank]
e.g.> *.py score_file.list 2 consensus_result.cons_50-2 50 2 -c
result > consensus_result.cons_50-2.sdf, consensus_result.cons_50-2.txt
- Perform a consensus of results from a set of dockings to a single receptor. A molecule has to be in the top % of number of model to be considered in the consensus result. For each molecule, the highest ranking score and pose is saved. The final consensus is the ranking of the single best pose or the average of rank, tho single best rank is better overall. SDF and Score files used must have same prefix.
#################################
Collecting Results of Schrodinger Induced-Fit Docking (IFD)
> B_collect_scripts/3_glide_IFD_mae2pdb.py
[ ligand file: sdf ] # full path to file
[ directory of IFD result ] # full path to directory
[ output filename prefix ]
e.g.> *.py atp.sdf InducedFit_1 test_IFD
result > test_IFD.pse
*** start outside of the parent IFD folder,
*** result files will be in the IFD folder
- Consolidate Schrodinger's IFD top results scattered within the IFD working directory and put them into a PyMOL session file. The IFD scores are within a result CSV file, which lists the IFD docking pose file associating with the ligand used. This script extracts those information to file the corresponding .MAE files, converts them back into .PDB files, and collects them into the PyMOL session file.
#######################################################################################
Required software / packages
OpenEye # 2018+
fred
omega2
flipper
fixpka
make_receptor
Schrödinger # 2019-01+
ligprep
glide
glide_sort
ifd
prepwizard
structconvert
csh/tcsh # shell
python # 3.6.8+
numpy # 1.16.2+
pandas # 0.24.2+
matplotlib # 3.0.3+
seaborn # 0.9.0+
rdkit # 2019.09.02
tabulate # 0.8.3+
pathos # 0.2.3+
tqdm # 4.31.1+
argparse # 1.1
html #
gzip #
bz2 #
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent