Merging, linking and placing compounds by stitching them together like a reanimated corpse
Project description
Fragmenstein
Stitched molecules
Fragmenstein: Merging, linking and placing compounds by stitching bound compounds together like a reanimated corpse.
Fragmenstein can perform two different tasks:
- Combine hits (merging and linking) based on their atomic overlap
- Place a given followup molecule based on one or more parent hits
NB. Whereas docking uses pre-generates comformers and finds the best pose that best matches the parent (if set-up to do so), Fragmenstein creates a monstrous comformer from the parent(s) and then minimises it, optionally in the protein. Hence why we call it a 'placement' not docking tool.
Quick links:
- For manuscript data see manuscript data repository
- For authors see Authors
- For command line interface see Command line interface
Example of multiple applications:
Classes
Like Frankenstein's creation it may violate the laws of chemistry. Trigonal planar topologies may be tetrahedral, bonds unnaturally long etc. This monstrosity is therefore then energy minimised with strong constraints within the protein.
There are four main classes —named after characters from the Fragmenstein book and movies:
Monster
makes the stitched together molecules indepent of the protein — documentationIgor
uses PyRosetta to minimise in the protein the fragmenstein monster followup — documentationVictor
is a pipeline that calls the parts, with several features, such as warhead switching —documentationLaboratory
does all the combinatorial operations with Victor (specific case)
NB. In the absence of pyrosetta
(which requires an academic licence), all bar Igor
work and
alternative Victor classes need to be used, for example
Wictor
(RDkit minimisation only), `OpenVictor (using OpenMM).
Additionally, there are a few minor classes.
One of these is mRMSD
, a multiple RMSD variant which does not superpose/align and bases which atoms
to use on coordinates —documentation
The class Walton
performs geometric manipulations of compounds, to set them up to demonstrate
features of Fragmenstein (like captain Walton, it does not partake in the plot, but is key to the narration)
There are two module hosted elsewhere:
Rectifier
from molecular_rectifier is a class that corrects mistakes in the molecule automatically merged byMonster
.Params
from rdkit to params module parameterises the ligands
Combine
It can also merge and link fragment hits by itself and find the best scoring mergers. For details about linking see linking notes. It uses the same overlapping position clustering, but also has a decent amount of impossible/uncommon chemistry prevention.
Monster:
from fragmenstein import Monster
monster = Monster(hits=[hits_a, hit_b])
monster.combine()
monster.positioned_mol #: RDKit.Chem.Mol
Victor:
from fragmenstein import Victor
import pyrosetta
pyrosetta.init( extra_options='-no_optH false -mute all -ex1 -ex2 -ignore_unrecognized_res false -load_PDB_components false -ignore_waters false')
victor = Victor(hits=[hits_a, hit_b],
pdb_filename='foo.pdb', # or pdb_block='ATOM 1 MET ...'
covalent_resi=1) # if not covalent, just put the first residue or something.
victor.combine()
victor.minimized_mol
The PyRosetta init step can be done with the helper function:
Igor.init_pyrosetta()
The two seem similar, but Victor places with Monster and minimises with Igor. As a result it has energy scores
victor.ddG
Fragmenstein is not really a docking algorithm as it does not find the pose with the lowest energy within a given volume. Consequently, it is a method to find how faithful is a given followup to the hits provided. Hence the minimised pose should be assessed by the RMSD metric or similar and the ∆∆G score used solely as a cutoff —lower than zero.
For a large number of combination:
from fragmenstein import Laboratory
lab = Laboratory(pdbblock=pdbblock, covalent_resi=None)
combinations:pd.DataFrame = lab.combine(hits, n_cores=28)
Place
Here is an interactive example of placed molecules.
It is rather tolerant to erroneous/excessive submissions (by automatically excluding them)
and can energy minimise strained conformations.
Three mapping approaches were tested, but the key is that hits are pairwise mapped to each other by means of one-to-one atom matching based upon position as opposed to similarity which is easily led astray. For example, note here that the benzene and the pyridine rings overlap, not the two pyridine rings:
RDkit only and OpenMM
PyRosetta is needed for the pocket-centric minimisation. Two alternatives are available:
Wictor
(without): stops at the RDKit minimisationOpenVictor
(with OpenMM): uses OpenMM to minimise in the protein
Whereas the PyRosetta steps operate via Igor, OpenVictor uses Fritz. OpenMM is a lot slower than PyRosetta on CPU only, but is free, open source and potentially more accurate.
Igor is a much larger class as it needs to disable rotamer sampling and other things, which is not an issue in OpenMM.
A further detail is that openMM is already parallel,
therefore when using with Laboratory
request only one core.
from fragmenstein import Laboratory, OpenVictor
Laboratory.Victor = OpenVictor
lab = Laboratory(pdbblock=MPro.get_template())
combinations: pd.DataFrame = lab.combine(hits,
n_cores=1, # 1 core unless $OPENMM_CPU_THREADS is set
timeout=600, # 2 minutes
combination_size=2, # pairwise
max_tasks=0) # 0 is no chunking
Examples
Monster:
from fragmenstein import Monster
monster = Monster(hits=[hits_a, hit_b])
monster.place_smiles('CCO')
monster.positioned_mol
Victor:
from fragmenstein import Victor, Igor
Igor.init_pyrosetta()
victor = Victor(hits=[hits_a, hit_b], pdb_filename='foo.pdb')
victor.place('CCO')
victor.minimized_mol
For a lengthier example see example notes or documentation.
Demo data
Some demo data is provided in the demo
submodule.
from fragmenstein.demo import MPro, Mac1
pdbblock: str = Mac1.get_template()
for hitname in Mac1.get_hit_list():
Mac1.get_hit(hitname)
...
To use SAR-COV-2 MPro as a test bed, the following may be helpful:
fragmenstein.MProVictor
, a derived class (ofVictor
), with various presents specific for MPro.fragemenstein.get_mpro_template()
, returns the PDB block (str) of MProfragemenstein.get_mpro_molblock(xnumber)
, returns the mol block (str) of a MPro hit from Fragalysisfragemenstein.get_mpro_mol(xnumber)
, as above but returns aChem.Mol
instance.
For the matched sets of derivative hits to reference hits see the manuscript's data repository.
Other features
Installation
Fragmenstein and dependencies
Python 3.6 or above. Install from pipy
python -m pip install fragmenstein
Requires Pyrosetta
:warning: PyRosetta no longer runs on CentOS 7 due to old kernel headers (cf. blog post).
Pyrosetta requires a password to be downloaded (academic licence) obtained by https://els2.comotion.uw.edu/product/pyrosetta. This is a different licence from the Rosetta one. The username of the Rosetta binaries is formatted variant of "academic user", while the PyRosetta is the name of a researcher whose name bares an important concept in protein folding, like boltzmann + constant (but is not that). Pyrosetta can be downloaded via a browser from http://www.pyrosetta.org/dow. Or in the terminal via:
curl -u 👾👾👾:👾👾👾https://graylab.jhu.edu/download/PyRosetta4/archive/release/PyRosetta4.Release.python38.linux/PyRosetta4.Release.python38.linux.release-NNN.tar.bz2 -o a.tar.bz2
tar -xf a.tar.bz2
cd PyRosetta4.Release.python38.linux
sudo pip3 install .
or using conda
or using install_pyrosetta
from the pyrosetta-help
package.
pip install pyrosetta-help
PYROSETTA_USERNAME=👾👾👾 PYROSETTA_PASSWORD=👾👾👾 install_pyrosetta
The PYROSETTA_USERNAME
and PYROSETTA_PASSWORD
are environment variables,
which should not be shared publicly (i.e. store them as private environmental variables
in your target application).
Origin
Fragmenstein was created to see how reasonable are the molecules of fragment mergers submitted in the COVID moonshot project, because after all the underlying method is fragment based screening. This dataset has some unique peculiarities that potentially are not encountered in other projects.
Command line interface
The strength of Fragmenstein is as a python module, but there is a command line interface.
This allows different levels of usage.
The top level is the fragmestein pipeline
, which does the whole thing,
namely it
- place the reference hits against themselves and gets the PLIP interactions
- combines the hits in given combination size, while skipping blacklisted named compounds.
- searches in SmallWorld the top N mergers
- places them and
- ranks them based on a customisable multiobjective function, which takes into account the PLIP interactions along with number of novel atoms (increase in risk & novelty).
This in effect reflects the pipeline I commonly use.
usage: fragmenstein pipeline [-h] -t TEMPLATE -i INPUT [-o OUTPUT] [-r RANKING] [-c CUTOFF] [-q QUICK] [-d SW_DIST] [-l SW_LENGTH] [-b SW_DATABASES [SW_DATABASES ...]] [-s SUFFIX] [-n N_CORES] [-m COMBINATION_SIZE] [-k TOP_MERGERS] [-e TIMEOUT] [-x MAX_TASKS] [-z BLACKLIST] [-j WEIGHTS] [-v]
# n_cores is optional and set to all cores by default. Here is doing something fancier, for sake of example.
# `--n_cores` sets number of cores you want to use, this simply is to get the number of cores:
export N_CORES=$(cat /proc/cpuinfo | grep processor | wc -l);
# make template
fragmenstein utils minimize --template holo_protein.pdb --constraint_weight 10 --cycles 15 --output minimised.pdb -first
# run
fragmenstein pipeline \
--template minimised.pdb \
--hits filtered.sdf \
--n_cores $(($N_CORES - 1)) \
--suffix _pairs \
--max_tasks 5000 \
--sw_databases REAL-Database-22Q1.smi.anon MculeUltimate-20Q2.smi.anon \
--combination_size 2 \
--timeout 600;
The values for the pipeline command are:
template
: The template, a polished PDB. The template must not contain a ligand in the site of interest, as Fragmenstein accepts other ligands (e.g. metals, cofactors etc.) and it is best to use a PyRosetta minimised template (i.e. one that has been through the ringer already).hits
: The hits in sdf format. These need to have unique names.output
: The output foldersuffix
: The suffix for the output files. Note that due tomax_tasks
there will be multiple sequential files for some steps.quick
: Does not reattempt "reanimation" if it failed as the constraints are relaxed more and more the more deviation happens.blacklist
: A file with a lines for each molecule name to not perform (sayhitA–hitZ
)cutoff
: The joining cutoff in Ångström after which linkages will not be attempted (default is 5Å)sw_databases
: See SmallWold or the SmallWorld API in Python for what datasets are available (e.g. 'Enamine-BB-Stock-Mar2022.smi.anon').sw_length
: How many analogues for each query to keepsw_dist
: The distance cutoff for the SmallWorld searchmax_tasks
: To avoid memory issues, the pipeline performs a number of tasks (controlled viamax_tasks
) before processing them, to disable this use--max_tasks 0
.weights
: This is a JSON file that controls the ranking
This will minimise the first chain only stripping waters and heterogens
Specific cases
fragmenstein monster combine -i hit1.mol hit2.mol >> combo.mol
fragmenstein monster place -i hit1.mol hit2.mol -s 'CCO' >> placed.mol
fragmenstein victor combine -i hit1.mol hit2.mol -t protein.pdb -o output >> combo.mol
fragmenstein victor combine -i hit1.mol hit2.mol -s 'NCO' -n molname -t protein.pdb -o output >> placed.mol
fragmenstein laboratory combine -i hits.sdf -o output -d output.csv -s output.sdf -c 24
Authors
Author | Role | Homepage | Department | Badges |
---|---|---|---|---|
Matteo Ferla | main developer | WCHG | Wellcome Centre for Human Genetics, University of Oxford | |
Rubén Sánchez-Garcia | discussion/code | Stats | Department of Statistics, University of Oxford | |
Rachael Skyner | discussion/editing/code | |||
Stefan Gahbauer | discussion | |||
Jenny Taylor | PI | WCHG | Wellcome Centre for Human Genetics, University of Oxford | |
Charlotte Deane | PI | |||
Frank von Delft | PI | CMD | Diamond Lightsource / CMD, Oxford | |
Brian Marsden | PI | CMD | CMD, Oxford |
See Also
- ChemRXiv preprint — TBA
- Steph Wills's fragment network merges repo contains useful filtering algorithms
- Fragmenstein is used in Schuller et. al. 2021
- Figures for the upcoming manuscript are in a separate repo
- The conversion of a rdkit Chem.Mol that cannot be sanitised to an analogue that can is done by the molecular rectifier package
- The conversion of a rdkit Chem.Mol to a PyRosetta residue type (a "params file") is done via the rdkit-to-params package
- The pipeline demo colab notebook uses Brian Shoichet's SmallWorld webapp, interfaced via its API in Python
- The playground demo colab notebook features a JSME widget — JSME is a popular JS only molecular editor
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.