"""
Tools for creating graph inputs from molecule data
"""
import itertools
import os
import sys
from collections import deque
from functools import partial
from multiprocessing import Pool
from typing import Dict, Union, List
import numpy as np
from pymatgen.core import Molecule, Element
from pymatgen.analysis.local_env import NearNeighbors
from pymatgen.io.babel import BabelMolAdaptor
from megnet.data.graph import StructureGraph, GaussianDistance, BaseGraphBatchGenerator, GraphBatchGenerator, Converter
from megnet.utils.general import fast_label_binarize
from .qm9 import ring_to_vector
try:
import pybel # type: ignore
except ImportError:
try:
from openbabel import pybel
except ImportError:
pybel = None
try:
from rdkit import Chem # type: ignore
except ImportError:
Chem = None
__date__ = "12/01/2018"
# List of features to use by default for each atom
_ATOM_FEATURES = [
"element",
"chirality",
"formal_charge",
"ring_sizes",
"hybridization",
"donor",
"acceptor",
"aromatic",
]
# List of features to use by default for each bond
_BOND_FEATURES = ["bond_type", "same_ring", "spatial_distance", "graph_distance"]
# List of elements in library to use by default
_ELEMENTS = ["H", "C", "N", "O", "F"]
[docs]class SimpleMolGraph(StructureGraph):
"""
Default using all atom pairs as bonds. The distance between atoms are used
as bond features. By default the distance is expanded using a Gaussian
expansion with centers at np.linspace(0, 4, 20) and width of 0.5
"""
def __init__(
self,
nn_strategy: Union[str, NearNeighbors] = "AllAtomPairs",
atom_converter: Converter = None,
bond_converter: Converter = None,
):
"""
Args:
nn_strategy (str): NearNeighbor strategy
atom_converter (Converter): atomic features converter object
bond_converter (Converter): bond features converter object
"""
if bond_converter is None:
bond_converter = GaussianDistance(np.linspace(0, 4, 20), 0.5)
super().__init__(nn_strategy=nn_strategy, atom_converter=atom_converter, bond_converter=bond_converter)
[docs]class MolecularGraph(StructureGraph):
"""Class for generating the graph inputs from a molecule
Computes many different features for the atoms and bonds in a molecule, and prepares them
in a form compatible with MEGNet models. The :meth:`convert` method takes a OpenBabel molecule
and, besides computing features, also encodes them in a form compatible with machine learning.
Namely, the `convert` method one-hot encodes categorical variables and concatenates
the atomic features
## Atomic Features
This class can compute the following features for each atom
- `atomic_num`: The atomic number
- `element`: (categorical) Element identity. (Unlike `atomic_num`, element is one-hot-encoded)
- `chirality`: (categorical) R, S, or not a Chiral center (one-hot encoded).
- `formal_charge`: Formal charge of the atom
- `ring_sizes`: For rings with 9 or fewer atoms, how many unique rings
of each size include this atom
- `hybridization`: (categorical) Hybridization of atom: sp, sp2, sp3, sq.
planer, trig, octahedral, or hydrogen
- `donor`: (boolean) Whether the atom is a hydrogen bond donor
- `acceptor`: (boolean) Whether the atom is a hydrogen bond acceptor
- `aromatic`: (boolean) Whether the atom is part of an aromatic system
## Atom Pair Features
The class also computes features for each pair of atoms
- `bond_type`: (categorical) Whether the pair are unbonded, or in a single, double, triple, or aromatic bond
- `same_ring`: (boolean) Whether the atoms are in the same aromatic ring
- `graph_distance`: Distance of shortest path between atoms on the bonding graph
- `spatial_distance`: Euclidean distance between the atoms. By default, this distance is expanded into
a vector of 20 different values computed using the `GaussianDistance` converter
"""
def __init__(
self,
atom_features: List[str] = None,
bond_features: List[str] = None,
distance_converter: Converter = None,
known_elements: List[str] = None,
max_ring_size: int = 9,
):
"""
Args:
atom_features ([str]): List of atom features to compute
bond_features ([str]): List of bond features to compute
distance_converter (DistanceCovertor): Tool used to expand distances
from a single scalar vector to an array of values
known_elements ([str]): List of elements expected to be in dataset. Used only if the
feature `element` is used to describe each atom
max_ring_size (int): Maximum number of atom in the ring
"""
# Check if openbabel and RDKit are installed
if Chem is None or pybel is None:
raise RuntimeError("RDKit and openbabel must be installed")
super().__init__()
if bond_features is None:
bond_features = _BOND_FEATURES
if atom_features is None:
atom_features = _ATOM_FEATURES
if distance_converter is None:
distance_converter = GaussianDistance(np.linspace(0, 4, 20), 0.5)
if known_elements is None:
known_elements = _ELEMENTS
# Check if all feature names are valid
if any(i not in _ATOM_FEATURES for i in atom_features):
bad_features = set(atom_features).difference(_ATOM_FEATURES)
raise ValueError(f"Unrecognized atom features: {', '.join(bad_features)}")
self.atom_features = atom_features
if any(i not in _BOND_FEATURES for i in bond_features):
bad_features = set(bond_features).difference(_BOND_FEATURES)
raise ValueError(f"Unrecognized bond features: {', '.join(bad_features)}")
self.bond_features = bond_features
self.known_elements = known_elements
self.distance_converter = distance_converter
self.max_ring_size = max_ring_size
[docs] def convert(self, mol, state_attributes: List = None, full_pair_matrix: bool = True) -> Dict: # type: ignore
"""
Compute the representation for a molecule
Args：
mol (pybel.Molecule): Molecule to generate features for
state_attributes (list): State attributes. Uses average mass and number of bonds per atom as default
full_pair_matrix (bool): Whether to generate info for all atom pairs, not just bonded ones
Returns:
(dict): Dictionary of features
"""
# Get the features features for all atoms and bonds
atom_features = []
atom_pairs: List[Dict] = []
for idx, atom in enumerate(mol.atoms):
f = self.get_atom_feature(mol, atom)
atom_features.append(f)
atom_features = sorted(atom_features, key=lambda x: x["coordid"])
num_atoms = mol.OBMol.NumAtoms()
for i, j in itertools.combinations(range(0, num_atoms), 2):
bond_feature = self.get_pair_feature(mol, i, j, full_pair_matrix)
if bond_feature:
atom_pairs.append(bond_feature)
else:
continue
# Compute the graph distance, if desired
if "graph_distance" in self.bond_features:
graph_dist = self._dijkstra_distance(atom_pairs)
for pair in atom_pairs:
d: Dict = {"graph_distance": graph_dist[pair["a_idx"], pair["b_idx"]]}
pair.update(d)
# Generate the state attributes (that describe the whole network)
state_attributes = state_attributes or [
[mol.molwt / num_atoms, len([i for i in atom_pairs if i["bond_type"] > 0]) / num_atoms]
]
# Get the atom features in the order they are requested by the user as a 2D array
atoms = []
for atom in atom_features:
atoms.append(self._create_atom_feature_vector(atom))
# Get the bond features in the order request by the user
bonds = []
index1_temp = []
index2_temp = []
for bond in atom_pairs:
# Store the index of each bond
index1_temp.append(bond.pop("a_idx"))
index2_temp.append(bond.pop("b_idx"))
# Get the desired bond features
bonds.append(self._create_pair_feature_vector(bond))
# Given the bonds (i,j), make it so (i,j) == (j, i)
index1 = index1_temp + index2_temp
index2 = index2_temp + index1_temp
bonds = bonds + bonds
# Sort the arrays by the beginning index
sorted_arg = np.argsort(index1)
index1 = np.array(index1)[sorted_arg].tolist()
index2 = np.array(index2)[sorted_arg].tolist()
bonds = np.array(bonds)[sorted_arg].tolist()
return {"atom": atoms, "bond": bonds, "state": state_attributes, "index1": index1, "index2": index2}
def _create_pair_feature_vector(self, bond: Dict) -> List[int]:
"""Generate the feature vector from the bond feature dictionary
Handles the binarization of categorical variables, and performing the distance conversion
Args:
bond (dict): Features for a certain pair of atoms
Returns:
([float]) Values converted to a vector
"""
bond_temp: List[int] = []
for i in self.bond_features:
# Some features require conversion (e.g., binarization)
if i in bond:
if i == "bond_type":
bond_temp.extend(fast_label_binarize(bond[i], [0, 1, 2, 3, 4]))
elif i == "same_ring":
bond_temp.append(int(bond[i]))
elif i == "spatial_distance":
expanded = self.distance_converter.convert([bond[i]])[0]
if isinstance(expanded, np.ndarray):
# If we use a distance expansion
bond_temp.extend(expanded.tolist())
else:
# If not
bond_temp.append(expanded)
else:
bond_temp.append(bond[i])
return bond_temp
def _create_atom_feature_vector(self, atom: dict) -> List[int]:
"""Generate the feature vector from the atomic feature dictionary
Handles the binarization of categorical variables, and transforming the ring_sizes to a list
Args:
atom (dict): Dictionary of atomic features
Returns:
([int]): Atomic feature vector
"""
atom_temp = []
for i in self.atom_features:
if i == "chirality":
atom_temp.extend(fast_label_binarize(atom[i], [0, 1, 2]))
elif i == "element":
atom_temp.extend(fast_label_binarize(atom[i], self.known_elements))
elif i in ["aromatic", "donor", "acceptor"]:
atom_temp.append(int(atom[i]))
elif i == "hybridization":
atom_temp.extend(fast_label_binarize(atom[i], [1, 2, 3, 4, 5, 6]))
elif i == "ring_sizes":
atom_temp.extend(ring_to_vector(atom[i], self.max_ring_size))
else: # It is a scalar
atom_temp.append(atom[i])
return atom_temp
@staticmethod
def _dijkstra_distance(pairs: List[Dict]) -> np.ndarray:
"""
Compute the graph distance between each pair of atoms,
using the network defined by the bonded atoms.
Args:
pairs ([dict]): List of bond information
Returns:
([int]) Distance for each pair of bonds
"""
bonds = []
for p in pairs:
if p["bond_type"] > 0:
bonds.append([p["a_idx"], p["b_idx"]])
return dijkstra_distance(bonds)
[docs] def get_atom_feature(
self, mol, atom # type: ignore
) -> Dict: # type: ignore
"""
Generate all features of a particular atom
Args:
mol (pybel.Molecule): Molecule being evaluated
atom (pybel.Atom): Specific atom being evaluated
Return:
(dict): All features for that atom
"""
# Get the link to the OpenBabel representation of the atom
obatom = atom.OBAtom
atom_idx = atom.idx - 1 # (pybel atoms indices start from 1)
# Get the element
element = Element.from_Z(obatom.GetAtomicNum()).symbol
# Get the fast-to-compute properties
output = {
"element": element,
"atomic_num": obatom.GetAtomicNum(),
"formal_charge": obatom.GetFormalCharge(),
"hybridization": 6 if element == "H" else obatom.GetHyb(),
"acceptor": obatom.IsHbondAcceptor(),
"donor": obatom.IsHbondDonorH() if atom.type == "H" else obatom.IsHbondDonor(),
"aromatic": obatom.IsAromatic(),
"coordid": atom.coordidx,
}
# Get the chirality, if desired
if "chirality" in self.atom_features:
# Determine whether the molecule has chiral centers
chiral_cc = self._get_chiral_centers(mol)
if atom_idx not in chiral_cc:
output["chirality"] = 0
else:
# 1 --> 'R', 2 --> 'S'
output["chirality"] = 1 if chiral_cc[atom_idx] == "R" else 2
# Find the rings, if desired
if "ring_sizes" in self.atom_features:
rings = mol.OBMol.GetSSSR() # OpenBabel caches ring computation internally, no need to cache ourselves
output["ring_sizes"] = [r.Size() for r in rings if r.IsInRing(atom.idx)]
return output
[docs] @staticmethod
def create_bond_feature(mol, bid: int, eid: int) -> Dict:
"""
Create information for a bond for a pair of atoms that are not actually bonded
Args:
mol (pybel.Molecule): Molecule being featurized
bid (int): Index of atom beginning of the bond
eid (int): Index of atom at the end of the bond
"""
a1 = mol.OBMol.GetAtom(bid + 1)
a2 = mol.OBMol.GetAtom(eid + 1)
same_ring = mol.OBMol.AreInSameRing(a1, a2)
return {
"a_idx": bid,
"b_idx": eid,
"bond_type": 0,
"same_ring": bool(same_ring),
"spatial_distance": a1.GetDistance(a2),
}
[docs] def get_pair_feature(self, mol, bid: int, eid: int, full_pair_matrix: bool) -> Union[Dict, None]:
"""
Get the features for a certain bond
Args:
mol (pybel.Molecule): Molecule being featurized
bid (int): Index of atom beginning of the bond
eid (int): Index of atom at the end of the bond
full_pair_matrix (bool): Whether to compute the matrix for every atom - even those that
are not actually bonded
"""
# Find the bonded pair of atoms
bond = mol.OBMol.GetBond(bid + 1, eid + 1)
if not bond: # If the bond is ordered in the other direction
bond = mol.OBMol.GetBond(eid + 1, bid + 1)
# If the atoms are not bonded
if not bond:
if full_pair_matrix:
return self.create_bond_feature(mol, bid, eid)
return None
# Compute bond features
a1 = mol.OBMol.GetAtom(bid + 1)
a2 = mol.OBMol.GetAtom(eid + 1)
same_ring = mol.OBMol.AreInSameRing(a1, a2)
return {
"a_idx": bid,
"b_idx": eid,
"bond_type": 4 if bond.IsAromatic() else bond.GetBondOrder(),
"same_ring": bool(same_ring),
"spatial_distance": a1.GetDistance(a2),
}
@staticmethod
def _get_rdk_mol(mol, format: str = "smiles"):
"""
Return: RDKit Mol (w/o H)
"""
if format == "pdb":
return Chem.rdmolfiles.MolFromPDBBlock(mol.write("pdb"))
if format == "smiles":
return Chem.rdmolfiles.MolFromSmiles(mol.write("smiles"))
return None
def _get_chiral_centers(self, mol):
"""
Use RDKit to find the chiral centers with CIP(R/S) label
This provides the absolute stereochemistry. The chiral label obtained
from pybabel and rdkit.mol.getchiraltag is relative positions of the bonds as provided
Args:
mol (Molecule): Molecule to asses
Return:
(dict): Keys are the atom index and values are the CIP label
"""
mol_rdk = self._get_rdk_mol(mol, "smiles")
if mol_rdk is None:
# Conversion to RDKit has failed
return {}
chiral_cc = Chem.FindMolChiralCenters(mol_rdk)
return dict(chiral_cc)
[docs]def dijkstra_distance(bonds: List[List[int]]) -> np.ndarray:
"""
Compute the graph distance based on the dijkstra algorithm
Args:
bonds: (list of list), for example [[0, 1], [1, 2]] means two bonds formed by atom 0, 1 and atom 1, 2
Returns:
full graph distance matrix
"""
nb_atom = max(itertools.chain(*bonds)) + 1
graph_dist = np.ones((nb_atom, nb_atom), dtype=np.int32) * np.infty
for bond in bonds:
graph_dist[bond[0], bond[1]] = 1
graph_dist[bond[1], bond[0]] = 1
queue: deque = deque() # Queue used in all loops
visited: set = set() # Used in all loops
for i in range(nb_atom):
graph_dist[i, i] = 0
visited.clear()
queue.append(i)
while queue:
s = queue.pop()
visited.add(s)
for k in np.where(graph_dist[s, :] == 1)[0]:
if k not in visited:
queue.append(k)
graph_dist[i, k] = min(graph_dist[i, k], graph_dist[i, s] + 1)
graph_dist[k, i] = graph_dist[i, k]
return graph_dist
[docs]def mol_from_smiles(smiles: str):
"""
load molecule object from smiles string
Args:
smiles (string): smiles string
Returns:
openbabel molecule
"""
mol = pybel.readstring(format="smi", string=smiles)
mol.make3D()
return mol
[docs]def mol_from_pymatgen(mol: Molecule):
"""
Args:
mol(Molecule)
"""
mol = pybel.Molecule(BabelMolAdaptor(mol).openbabel_mol)
mol.make3D()
return mol
[docs]def mol_from_file(file_path: str, file_format: str = "xyz"):
"""
Args:
file_path(str)
file_format(str): allow formats that open babel supports
"""
mol = list(pybel.readfile(format=file_format, filename=file_path))[0]
return mol
def _convert_mol(mol: str, molecule_format: str, converter: MolecularGraph) -> Dict:
"""Convert a molecule from string to its graph features
Utility function used in the graph generator.
The parse and convert operations are both in this function due to Pybel objects
not being serializable. By not using the Pybel representation of each molecule
as an input to this function, we can use multiprocessing to parallelize conversion
over molecules as strings can be passed as pickle objects to the worker threads but
but Pybel objects cannot.
Args:
mol (str): String representation of a molecule
molecule_format (str): Format of the string representation
converter (MolecularGraph): Tool used to generate graph representation
Returns:
(dict): Graph representation of the molecule
"""
# Convert molecule into pybel format
if molecule_format == "smiles":
mol = mol_from_smiles(mol) # Used to generate 3D coordinates/H atoms
else:
mol = pybel.readstring(molecule_format, mol)
return converter.convert(mol)
[docs]class MolecularGraphBatchGenerator(BaseGraphBatchGenerator):
"""Generator that creates batches of molecular data by computing graph properties on demand
If your dataset is small enough that the descriptions of the whole dataset fit in memory,
we recommend using :class:`megnet.data.graph.GraphBatchGenerator` instead to avoid
the computational cost of dynamically computing graphs."""
def __init__(
self,
mols: List[str],
targets: List[np.ndarray] = None,
converter: MolecularGraph = None,
molecule_format: str = "xyz",
batch_size: int = 128,
shuffle: bool = True,
n_jobs: int = 1,
):
"""
Args:
mols ([str]): List of the string reprensetations of each molecule
targets ([ndarray]): Properties of each molecule to be predicted
converter (MolecularGraph): Converter used to generate graph features
molecule_format (str): Format of each of the string representations in `mols`
batch_size (int): Target size for each batch
shuffle (bool): Whether to shuffle the training data after each epoch
n_jobs (int): Number of worker threads (None to use all threads).
"""
super().__init__(len(mols), targets, batch_size, shuffle)
self.mols = np.array(mols)
if converter is None:
converter = MolecularGraph()
self.converter = converter
self.molecule_format = molecule_format
self.n_jobs = n_jobs
def mute():
with open(os.devnull, "w") as f:
sys.stdout = f
sys.stderr = f
self.pool = Pool(self.n_jobs, initializer=mute) if self.n_jobs != 1 else None
def __del__(self):
if self.pool is not None:
self.pool.close() # Kill thread pool if generator is deleted
def _generate_inputs(self, batch_index: list) -> np.ndarray:
# Get the molecules for this batch
mols = self.mols[batch_index]
# Generate the graphs
graphs = self._generate_graphs(mols)
# Return them as flattened into array format
return self.converter.get_flat_data(graphs)
def _generate_graphs(self, mols: List[str]) -> List[Dict]:
"""Generate graphs for a certain collection of molecules
Args:
mols ([string]): Molecules to process
Returns:
([dict]): Graphs for all of the molecules
"""
if self.pool is None:
graphs = [_convert_mol(m, self.molecule_format, self.converter) for m in mols]
else:
func = partial(_convert_mol, molecule_format=self.molecule_format, converter=self.converter)
graphs = self.pool.map(func, mols)
return graphs
[docs] def create_cached_generator(self) -> GraphBatchGenerator:
"""Generates features for all of the molecules and stores them in memory
Returns:
(GraphBatchGenerator) Graph genereator that relies on having the graphs in memory
"""
# Make all the graphs
graphs = self._generate_graphs(self.mols)
# Turn them into a fat array
atom_features, bond_features, state_features, index1_list, index2_list, targets = self.converter.get_flat_data(
graphs, self.targets
) # type: ignore
return GraphBatchGenerator(
atom_features=atom_features,
bond_features=bond_features,
state_features=state_features,
index1_list=index1_list,
index2_list=index2_list,
targets=targets,
is_shuffle=self.is_shuffle,
batch_size=self.batch_size,
)