Molpert is a library for graph-based perturbation of RDKit molecules. It supports atom and bond property modifications, insertions and deletions and combinations thereof. A plethora of settings provide fine grained control over how perturbations are executed. The library can be used stochastically, applying a random eligible perturbation, or deterministically, enumerating all eligible perturbations. Perturbation eligibility is determined by arbitrary user-provided constraints.
Ensure the following dependencies are installed:
- RDKit
- Boost. You already have this if you installed the RDKit. If you'd like to build the Python bindings make sure Boost.Python is installed.
- CMake. Only necessary to build the Python bindings.
The following instructions are for GNU+Linux. For alternative operating systems you'll have to adapt these commands slightly.
git clone https://github.com/AlanKerstjens/Molpert.gitMolpert is header-only. If all you care about is the C++ API you are done. If you want the Python bindings too you must build them.
export MOLPERT="$(pwd)/Molpert"
mkdir ${MOLPERT}/build && cd ${MOLPERT}/build
cmake ..
make installTo be able to import the library from Python add ${MOLPERT}/lib to your ${PYTHONPATH}. Consider doing so in your bash_profile file. Otherwise you'll have to manually extend ${PYTHONPATH} everytime you open a new shell.
export PYTHONPATH="${PYTHONPATH}:${MOLPERT}/lib"CMake will try to find the dependencies for you. To avoid problems ensure you build Molpert with the same Boost and Python versions that you used to build the RDKit. If CMake finds a different Boost or Python installation you'll need to point it to the correct one, as described here and here.
CMake will search for the RDKit in the active Anaconda environment (if you have one) and at ${RDBASE} if set. If neither of these are the case you need to specify the path to the RDKit yourself. Replace the above CMake command with the one below, substituting the <placeholder/path> with your path.
cmake -DRDKit_ROOT=<path/to/rdkit> ..This section provides a high level overview of how to use Molpert and its capabilities. It's not intended to serve as comprehensive documentation, and boilerplate code is omitted. Example code snippets are provided for C++ and Python (in that order).
To start using Molpert include or import the relevant headers or package.
#include "MoleculePerturber.hpp"import molpert as mptPerturbations are specifications of how a molecule ought to be modified. They are represented as classes that inherit from the MolecularPerturbation base class. There are 8 perturbation types of interest to the average user. Each derived class has a corresponding entry in the MolecularPerturbation::Type enum.
| Derived class | MolecularPerturbation::Type |
Classification |
|---|---|---|
AtomicNumberChange |
AtomicNumberChange_t |
Decorative |
FormalChargeChange |
FormalChargeChange_t |
Decorative |
ExplicitHydrogensChange |
ExplicitHydrogensChange_t |
Decorative |
BondTypeChange |
BondTypeChange_t |
Decorative |
AtomInsertion |
AtomInsertion_t |
Topological |
AtomDeletion |
AtomDeletion_t |
Topological |
BondInsertion |
BondInsertion_t |
Topological |
BondDeletion |
BondDeletion_t |
Topological |
After construction a perturbation can be executed by calling its operator(). Said operator has two overloads: one to perturb mutable RDKit::RWMols in place, and one to perturb copies of immutable RDKit::ROMols. In Python the RDKit makes no distinction between RDKit::ROMol and RDKit::RWMol, exposing a single mutable Chem.Mol class instead. Perturbations are always executed in place by calling the mpt.ExecutePerturbation() free function. If you want to create a copy you will have to do so manually.
RDKit::RWMol rwmol (...);
RDKit::ROMol romol (...);
std::shared_ptr<MolecularPerturbation> perturbation (...);
perturbation(rwmol); // Perturb in place.
RDKit::RWMol perturbed_romol = perturbation(romol); // Copy and then perturb.molecule = Chem.Mol(...)
molecule_copy = Chem.Mol(molecule)
perturbation = mpt.MolecularPerturbation(...)
mpt.ExecutePerturbation(perturbation, molecule)
mpt.ExecutePerturbation(perturbation, molecule_copy)You can characterize a perturbation by retrieving its type or calculating its ID (i.e. hash).
MolecularPerturbation::Type perturbation_type = perturbation->GetType();
std::size_t perturbation_id = perturbation->ID();perturbation_type = perturbation.type
perturbation_id = perturbation.ID()Compared to standard functions, the added value of MolecularPerturbations is:
- Deferred execution
- Deduplication
- Constraint enforcement
In the RDKit specific atoms and bonds can be referenced through their pointers (RDKit::Atom* and RDKit::Bond*) or indices. Since atoms and bonds are stored in contiguous memory both their addresses and indices may be invalidated when you modify the topology of a molecule. This makes keeping track of atoms and bonds in ever-changing molecules challenging.
To keep track of an atom or bond throughout a molecule's lifespan we use tags. Tags are integers attached to atoms and bonds. They remain associated with the atom or bond regardless of how we modify the molecule. Tags shouldn't be re-used within the same molecule.
Before constructing perturbations you must tag your molecules! Perturbations that insert atoms or bonds will automatically tag the new atoms and bonds. You only need to tag the initial molecule.
RDKit::ROMol molecule (...);
TagMolecule(molecule);molecule = Chem.Mol(...)
mpt.TagMolecule(molecule)To get the tag of an atom or bond:
const RDKit::Atom* atom = molecule.getAtomWithIdx(0);
const RDKit::Bond* bond = molecule.getBondWithIdx(0);
std::size_t atom_tag = GetTag(atom);
std::size_t bond_tag = GetTag(bond);atom = molecule.GetAtomWithIdx(0)
bond = molecule.GetBondWithIdx(0)
atom_tag = mpt.GetAtomTag(atom)
bond_tag = mpt.GetBondTag(bond)You can construct perturbations by manually calling their constructor. Multiple constructor overloads are exposed. Since Python isn't strongly typed overload selection on the Python side is a bit finnicky. You may have to specify all arguments for overload resolution to work properly.
AtomicNumberChange perturbation1 (
0, // Atom index of the affected atom
8 // New atomic number
)
AtomDeletion perturbation2 (
molecule, // Subject molecule
0, // Atom index of the atom to delete
1 // Atom index of the reconnection atom that takes its place
)perturbation1 = mpt.AtomicNumberChange(
atom_idx=0,
atomic_number=8)
perturbation2 = mpt.AtomDeletion(
molecule=molecule,
atom_idx=0,
reconnection_atom_idx=1,
preserve_bond_types=True, # Default
default_bond_type=Chem.BondType.SINGLE # Default
)Molpert's real power rests in its MolecularPerturbation factory: MoleculePerturber.
MoleculePerturber perturber;perturber = mpt.MoleculePerturber()MoleculePerturber has a plethora settings to regulate which perturbations are generated. Reasonable default values are set for you. If you don't like them you can change them by accessing the corresponding public member variables. Broadly speaking settings can be classified into two categories:
- Perturbation specific settings
- Sampling values and weights
Each perturbation type has perturbation specific settings. Setting names are prefaced with the perturbation type's name. Most of these settings are unsigned integers or booleans.
perturber.atom_insertion_max_n_neighbors = 4; // Default is 3
perturber.bond_deletion_allow_reroutes = false; // Default is trueperturber.atom_insertion_max_n_neighbors = 3
perturber.bond_deletion_allow_reroutes = FalseSampling values are vectors (std::vector) from which atom and bond properties are sampled. Each value array has a corresponding size-matched vector of sampling weights. The default values and weights are derived from ChEMBL. If you manually modify the value vector make sure you modify the weights vector accordingly. If you are a Python user you'll have to use the corresponding setter instead. In the following example you would set a 60% probability of sampling carbons and a 20% probability of sampling nitrogens and oxygens. Atomic numbers in particular are sampled during AtomicNumberChanges and AtomInsertions.
perturber.atomic_numbers = {6, 7, 8};
perturber.atomic_numbers_weights = {0.6, 0.2, 0.2};perturber.SetAtomicNumbers(
atomic_numbers=[6, 7, 8],
weights=[0.6, 0.2, 0.2])You can conveniently modify some of these arrays by passing arguments to the MoleculePerturber constructor. This is less error prone than doing it manually.
MoleculePerturber perturber (
false, // Sample uniformly instead of using the ChEMBL sampling weights.
true, // Include aromatic bonds in the eligible bond types.
true // Allow sampling of aromatic bond type for acyclic bonds.
)perturber = MoleculePerturber(
use_chembl_distribution=False,
use_aromatic_bonds=True,
acyclic_can_be_aromatic=True)You can also adjust which perturbation types are sampled and with which probabilities. The main difference with the above is that perturbation sampling settings are fixed-size variables (std::bitset and std::array). It's recommended to modify them using operator[] instead of operator=. In Python you must use a setter function instead. By default all perturbations may be sampled, with sampling weights being derived from ChEMBL.
perturber.perturbation_types.reset() // Reset all perturbation types to false.
perturber.perturbation_types[MolecularPerturbation::Type::AtomicNumberChange] = true;
perturber.perturbation_types[MolecularPerturbation::Type::BondTypeChange] = true;
perturber.perturbation_types_weights[MolecularPerturbation::Type::AtomicNumberChange] = 0.5;
perturber.perturbation_types_weights[MolecularPerturbation::Type::BondTypeChange] = 0.5;perturber.SetPerturbationTypes(
perturbation_types=[
mpt.MolecularPerturbation.Type.AtomicNumberChange,
mpt.MolecularPerturbation.Type.BondTypeChange],
weights=[0.5, 0.5],
reset=True # Reset everything to 0 before setting the specified values (default)
)MoleculePerturber has member functions to randomly generate applicable perturbations. These functions rely on a random number generator which must be passed as argument to the perturbation generators.
std::random_device rd;
std::mt19937 prng (rd()); // Pseudo-random number generatorprng = mpt.PRNG()If you don't mind the type of the generated perturbations you can call MoleculePerturber::operator(). In Python the equivalent member function is MoleculePerturber.Perturbation().
// Perturb the molecule in some random way
std::shared_ptr<MolecularPerturbation> perturbation = perturber(
molecule, prng);# Perturb the molecule in some random way
perturbation = perturber.Perturbation(
molecule=molecule,
prng=prng)To limit the generation to a specific type of perturbation call the appropiate member function. Stochastic perturbation generators are named using the perturbation type in imperative tense:
// Insert an atom anywhere in molecule
std::shared_ptr<AtomInsertion> atom_insertion = perturber.InsertAtom(
molecule, prng);
// Delete any bond of molecule
std::shared_ptr<BondDeletion> bond_deletion = perturber.DeleteBond(
molecule, prng);# Insert an atom anywhere in molecule
atom_insertion = perturber.InsertAtom(
molecule=molecule,
prng=prng)
# Delete any bond of molecule
bond_deletion = perturber.DeleteBond(
molecule=molecule,
prng=prng)Generators have multiple overloads allowing you to specify, among other things, which part of the molecule (i.e. which atom or bond) to target.
// Insert an atom adjacent to atom with index 5
std::shared_ptr<AtomInsertion> atom_insertion = perturber.InsertAtom(
molecule,
5, // Central atom index
prng);
// Delete the bond with index 3
std::shared_ptr<BondDeletion> bond_deletion = perturber.DeleteBond(
molecule,
3, // Deleted bond index
prng);# Insert an atom adjacent to atom with index 5
atom_insertion = perturber.InsertAtom(
molecule=molecule,
central_atom_idx=5,
prng=prng)
# Delete the bond with index 3
bond_deletion = perturber.DeleteBond(
molecule=molecule,
bond_idx=5,
prng=prng)MoleculePerturber has member functions to enumerate all perturbations that are applicable to a molecule. Enumerated perturbations are stored in a queue, which must be passed as argument to the perturbation generators.
MolecularPerturbationQueue queue;queue = mpt.MolecularPerturbationQueue()MolecularPerturbationQueue automatically deduplicates enumerated perturbations, ensuring that the same perturbation can't be stored more than once. Beware that a perturbation is considered a duplicate if it ever entered the queue, regardless of whether it's in it at present time. If you don't like this behaviour call MolecularPerturbationQueue::clear() to erase the record of previously queued perturbations.
If you'd like to enumerate all perturbations of all kinds that can be applied to a molecule call MoleculePerturber::operator(). The equivalent function in Python is MoleculePerturber.Perturbations().
// Enumerate all perturbations that can be applied to molecule
perturber(queue, molecule);# Enumerate all perturbations that can be applied to molecule
perturber.Perturbations(
queue=queue,
molecule=molecule)To limit the generation to a specific type of perturbation call the corresponding member function. Deterministic perturbation generators are named after the perturbation type in plural:
// Enumerate all atom insertions that can be applied to molecule
perturber.AtomInsertions(queue, molecule);
// Enumerate all bond deletions that can be applied to molecule
perturber.BondDeletions(queue, molecule);# Enumerate all atom insertions that can be applied to molecule
perturber.AtomInsertions(queue=queue, molecule=molecule)
# Enumerate all bond deletions that can be applied to molecule
perturber.BondDeletions(queue=queue, molecule=molecule)Generators have multiple overloads allowing you to specify, among other things, which part of the molecule (i.e. which atom or bond) to target.
// Enumerate all atom insertions that insert an atom adjacent to atom 5
perturber.AtomInsertions(
queue,
molecule,
5); // Central atom index
// Enumerate all bond deletions that deleted bond 3
perturber.BondDeletions(
queue,
molecule,
3); // Deleted bond index# Enumerate all atom insertions that insert an atom adjacent to atom 5
perturber.AtomInsertions(
queue=queue,
molecule=molecule,
central_atom_idx=5)
# Enumerate all bond deletions that deleted bond 3
perturber.BondDeletions(
queue=queue,
molecule=molecule,
bond_idx=3)MoleculePerturbers perturbation generation functions accept MolecularConstraints as optional arguments. Constraints allow you to specify arbitrary requirements that molecules resulting from a perturbation must fulfill.
A constraint is a callback function that takes as input an atom, bond or molecule and returns a boolean, evaluating to true if the corresponding atom, bond or molecule is allowed, and false otherwise. Atom and bond constraints apply to a specific atom or bond. Molecule constraints apply to the whole molecule.
The C++ signatures of constraints are:
typedef std::function<bool(const RDKit::Atom*)> AtomConstraint;
typedef std::function<bool(const RDKit::Bond*)> BondConstraint;
typedef std::function<bool(const RDKit::ROMol*)> MoleculeConstraint;Let us define some constraints. Constraints can be free functions or functors:
bool HasPositiveCharge(const RDKit::Atom* atom) {
return atom->getFormalCharge() > 0;
};
bool IsHeteroatom(const RDKit::Atom* atom) {
int z = atom->getAtomicNum();
return z != 1 && z != 6;
};
bool NotHeteroBond(const RDKit::Bond* bond) {
return !(IsHeteroatom(bond->getBeginAtom()) &&
IsHeteroatom(bond->getEndAtom()));
};
struct IsRigid {
unsigned max_n_rotatable_bonds = 0;
IsRigid(unsigned max_n_rotatable_bonds) :
max_n_rotatable_bonds(max_n_rotatable_bonds) {};
bool operator()(const RDKit::ROMol* molecule) const {
unsigned n_rotatable_bonds = RDKit::Descriptors::calcNumRotatableBonds(
*molecule);
return n_rotatable_bonds <= max_n_rotatable_bonds;
};
};def HasPositiveCharge(atom: Chem.Atom) -> bool:
return atom.GetFormalCharge() > 0
def IsHeteroatom(atom: Chem.Atom) -> bool:
z = atom.GetAtomicNum()
return z != 1 and z != 6
def NotHeteroBond(bond: Chem.Bond) -> bool:
return not (IsHeteroatom(bond.GetBeginAtom()) and
IsHeteroatom(bond.GetEndAtom()))
class IsRigid:
def __init__(self, max_n_rotatable_bonds: int):
self.max_n_rotatable_bonds = max_n_rotatable_bonds
def __call__(self, molecule: Chem.Mol) -> bool:
n_rotatable_bonds = Chem.Descriptors.NumRotatableBonds(molecule)
return n_rotatable_bonds <= self.max_n_rotatable_bondsYou can store these constraints in a MolecularConstraints instance.
MolecularConstraints constraints;
// Atom with tag 3 should have a positive charge.
constraints.SetAtomConstraint(3, HasPositiveCharge);
// Bond with tag 2 shouldn't bond two heteroatoms.
constraints.SetBondConstraint(2, NotHeteroBond);
// Molecule may not have more than 4 rotatable bonds.
constraints.SetMoleculeConstraint(IsRigid(4));constraints = mpt.MolecularConstraints()
# Atom with tag 3 should have a positive charge.
constraints.SetAtomConstraint(
atom_tag=3,
atom_constraint=HasPositiveCharge)
# Bond with tag 2 shouldn't bond two heteroatoms.
constraints.SetBondConstraint(
bond_tag=2,
bond_constraint=NotHeteroBond)
# Molecule may not have more than 4 rotatable bonds.
constraints.SetMoleculeConstraint(
molecule_constraint=IsRigid(4))Pass the MolecularConstraints to the perturbation generator as the (optional) last argument:
// Stochastic perturbation generation
std::shared_ptr<AtomInsertion> perturbation = perturber.InsertAtom(
molecule, prng, &constraints);
// Deterministic perturbation generation
perturber.AtomInsertions(queue, molecule, &constraints);# Stochastic perturbation generation
perturbation = perturber.InsertAtom(
molecule=molecule,
prng=prng,
constraints=constraints)
# Deterministic perturbation generation
perturber.AtomInsertions(
queue=queue,
molecule=molecule,
constraints=constraints)You can store as many MoleculeConstraints as you please, but only up to one AtomConstraint or BondConstraint per atom or bond. If you need to check for multiple conditions merge them into a single function.
constraints.SetAtomConstraint(3, HasPositiveCharge); // OK
constraints.SetAtomConstraint(3, IsHeteroatom); // Careful! Overwrites constraint.If you need all of your atoms and/or bonds to fulfill some condition you have two options:
- Option A (recommended): Use a
MoleculeConstraintthat iterates over all atoms and/or bonds and checks each one for the condition. - Option B: Define constraint generators and construct
MolecularConstraintswith them. A constraint generator is a function that takes as input a pair of atoms or bonds (prior and posterior), compares both and generates or updates the constraint for the atom or bond whenever a molecule is perturbed. Use constraint generators only if your constraints depend on both the prior and posterior states. The signature of constraint generators is as follows:
typedef std::function<AtomConstraint(const RDKit::Atom*, const RDKit::Atom*)
> AtomConstraintGenerator;
typedef std::function<BondConstraint(const RDKit::Bond*, const RDKit::Bond*)
> BondConstraintGenerator;During topological perturbation generation the molecular graph may have to be traversed to determine if a perturbation is applicable or not. If you plan on perturbing a molecule just once it's advisable to perform these graph traversals lazily (default).
perturber.assess_connectivity_with_sssr = false;
perturber.assess_distances_with_distance_matrix = false;If you will be enumerating many neighbors of a molecule it's worthwhile to spend some extra time calculating some topological information once to be used in subsequent perturbation generation:
perturber.assess_connectivity_with_sssr = true;
perturber.assess_distances_with_distance_matrix = true;If you'd like to remove an upper limit for a setting set it to std::numeric_limits<unsigned>::max(). On most systems the Python equivalent of this is 0xFFFFFFFF. The code will check against this value and skip unnecessary calculations. DO NOT set it to an arbitrarily high value, as the graph traversals will still be executed.
perturber.bond_insertion_max_distance_partner = std::numeric_limits<unsigned>::max(); // OK
perturber.bond_insertion_max_distance_partner = perturber.max_unsigned; // OK. Same as above.
perturber.bond_insertion_max_distance_partner = 9999; // BAD. Will trigger graph traversal.perturber.bond_insertion_max_distance_partner = 0xFFFFFFFF # OKBe conservative in your use of constraints. It may be possible to achieve what you want without them by changing the MoleculePerturber settings. Even if changing settings doesn't get you all the way there it can greatly reduce the number of generated perturbations and with it the number of times your constraints are evaluated. For example, if you'd like to avoid generating macrocycles consider using the below settings. They won't guarantee it doesn't happen, but they will reduce the number of macrocycles that are generated in a computationally more efficient way.
// 5 consecutive bonds span 6 atoms.
// Bonding the peripheral atoms results in a 6-membered ring.
perturber.bond_insertion_max_distance_partner = 5;
// 4 consecutive bonds span 5 atoms.
// Inserting an atom between the peripheral atoms results in a 6-membered ring.
perturber.atom_insertion_max_distance_neighbor = 4;
// 6 consecutive bonds span 7 atoms. However, we are removing one of them.
// Bonding the peripheral atoms results in a 6-membered ring.
perturber.atom_deletion_max_distance_reconnection_atom = 6;