def get_changed_bonds(rxn_smi):
reactants = Chem.MolFromSmiles(rxn_smi.split('>')[0])
products = Chem.MolFromSmiles(rxn_smi.split('>')[2])
conserved_maps = [a.GetProp('molAtomMapNumber') for a in products.GetAtoms() if a.HasProp('molAtomMapNumber')]
bond_changes = set() # keep track of bond changes
# Look at changed bonds
bonds_prev = {}
for bond in reactants.GetBonds():
nums = sorted(
[bond.GetBeginAtom().GetProp('molAtomMapNumber'),
bond.GetEndAtom().GetProp('molAtomMapNumber')])
if (nums[0] not in conserved_maps) and (nums[1] not in conserved_maps): continue
bonds_prev['{}~{}'.format(nums[0], nums[1])] = bond.GetBondTypeAsDouble()
bonds_new = {}
for bond in products.GetBonds():
nums = sorted(
[bond.GetBeginAtom().GetProp('molAtomMapNumber'),
bond.GetEndAtom().GetProp('molAtomMapNumber')])
bonds_new['{}~{}'.format(nums[0], nums[1])] = bond.GetBondTypeAsDouble()
for bond in bonds_prev:
if bond not in bonds_new:
bond_changes.add((bond.split('~')[0], bond.split('~')[1], 0.0)) # lost bond
else:
if bonds_prev[bond] != bonds_new[bond]:
bond_changes.add((bond.split('~')[0], bond.split('~')[1], bonds_new[bond])) # changed bond
for bond in bonds_new:
if bond not in bonds_prev:
bond_changes.add((bond.split('~')[0], bond.split('~')[1], bonds_new[bond])) # new bond
return bond_changes
def edit_mol(rmol, edits, tatoms):
new_mol = Chem.RWMol(rmol)
[a.SetNumExplicitHs(0) for a in new_mol.GetAtoms()]
amap = {}
for atom in rmol.GetAtoms():
amap[atom.GetAtomMapNum() - 1] = atom.GetIdx()
for x, y, t, v in edits:
bond = new_mol.GetBondBetweenAtoms(amap[x], amap[y])
# a1 = new_mol.GetAtomWithIdx(amap[x])
# a2 = new_mol.GetAtomWithIdx(amap[y])
if bond is not None:
new_mol.RemoveBond(amap[x], amap[y])
if t > 0:
new_mol.AddBond(amap[x], amap[y], BOND_FLOAT_TO_TYPE[t])
pred_mol = new_mol.GetMol()
pred_smiles = Chem.MolToSmiles(pred_mol)
pred_list = pred_smiles.split('.')
pred_mols = []
for pred_smiles in pred_list:
mol = Chem.MolFromSmiles(pred_smiles)
if mol is None: continue
atom_set = set([atom.GetAtomMapNum() - 1 for atom in mol.GetAtoms()])
if len(atom_set & tatoms) == 0:
continue
for atom in mol.GetAtoms():
atom.SetAtomMapNum(0)
pred_mols.append(mol)
return '.'.join(
sorted([Chem.MolToSmiles(pred_mol) for pred_mol in pred_mols]))
def __init__(self, smiles=None, rdk=None, conv_enabled=False):
"""Constructor
Keyword Arguments:
smiles {str} -- SMILES representation of a molecule (default: {None})
rdk {rdkit Mol} -- molecule as an RDKit object (default: {None})
conv_enabled {bool} -- whether to set both smiles and graph
arguments here or lazily defer until called
(default: {False})
Raises:
ValueError -- if neither a correct smiles string
or a rdkit mol are provided
"""
if conv_enabled:
if isinstance(smiles, str):
# also checks if smiles can be parsed
rdk = Chem.MolFromSmiles(smiles)
assert rdk is not None
elif rdk is not None:
smiles = Chem.MolToSmiles(rdk)
else:
raise ValueError("Invalid arguments")
self.smiles = smiles
self.rdk = rdk
self.graph = None # should be obtained from rdk when needed
self.synthesis_path = [] # list of Reactions
self.begin_flag = True
def smiles2graph(smiles, idxfunc=lambda x: x.GetIdx()):
mol = Chem.MolFromSmiles(smiles)
if not mol:
raise ValueError("Could not parse smiles string:", smiles)
n_atoms = mol.GetNumAtoms()
n_bonds = max(mol.GetNumBonds(), 1)
fatoms = np.zeros((n_atoms, atom_fdim))
fbonds = np.zeros((n_bonds, bond_fdim))
atom_nb = np.zeros((n_atoms, max_nb), dtype=np.int32)
bond_nb = np.zeros((n_atoms, max_nb), dtype=np.int32)
num_nbs = np.zeros((n_atoms, ), dtype=np.int32)
for atom in mol.GetAtoms():
idx = idxfunc(atom)
if idx >= n_atoms:
raise Exception(smiles)
fatoms[idx] = atom_features(atom)
for bond in mol.GetBonds():
a1 = idxfunc(bond.GetBeginAtom())
a2 = idxfunc(bond.GetEndAtom())
idx = bond.GetIdx()
if num_nbs[a1] == max_nb or num_nbs[a2] == max_nb:
raise Exception(smiles)
atom_nb[a1, num_nbs[a1]] = a2
atom_nb[a2, num_nbs[a2]] = a1
bond_nb[a1, num_nbs[a1]] = idx
bond_nb[a2, num_nbs[a2]] = idx
num_nbs[a1] += 1
num_nbs[a2] += 1
fbonds[idx] = bond_features(bond)
return fatoms, fbonds, atom_nb, bond_nb, num_nbs
def get_bond_label(r, edits, max_natoms):
rmol = Chem.MolFromSmiles(r)
n_atoms = rmol.GetNumAtoms()
rmap = np.zeros((max_natoms, max_natoms, nbos))
for s in edits.split(';'):
a1, a2, bo = s.split('-')
x = min(int(a1) - 1, int(a2) - 1)
y = max(int(a1) - 1, int(a2) - 1)
z = bo_to_index[float(bo)]
rmap[x, y, z] = rmap[y, x, z] = 1
labels = []
sp_labels = []
for i in range(max_natoms):
for j in range(max_natoms):
for k in range(len(bo_to_index)):
if i == j or i >= n_atoms or j >= n_atoms:
labels.append(INVALID_BOND) # mask
else:
labels.append(rmap[i, j, k])
if rmap[i, j, k] == 1:
sp_labels.append(i * max_natoms * nbos + j * nbos + k)
# TODO: check if this is consistent with how TF does flattening
return np.array(labels), sp_labels
def get_feature_batch(r_list):
max_natoms = 0
for r in r_list:
rmol = Chem.MolFromSmiles(r)
if rmol.GetNumAtoms() > max_natoms:
max_natoms = rmol.GetNumAtoms()
features = []
for r in r_list:
features.append(get_bin_feature(r, max_natoms))
return np.array(features)
def get_bin_feature(r, max_natoms):
'''
This function is used to generate descriptions of atom-atom relationships, including
the bond type between the atoms (if any) and whether they belong to the same molecule.
It is used in the global attention mechanism.
'''
comp = {}
for i, s in enumerate(r.split('.')):
mol = Chem.MolFromSmiles(s)
for atom in mol.GetAtoms():
comp[atom.GetIntProp('molAtomMapNumber') - 1] = i
n_comp = len(r.split('.'))
rmol = Chem.MolFromSmiles(r)
n_atoms = rmol.GetNumAtoms()
bond_map = {}
for bond in rmol.GetBonds():
a1 = bond.GetBeginAtom().GetIntProp('molAtomMapNumber') - 1
a2 = bond.GetEndAtom().GetIntProp('molAtomMapNumber') - 1
bond_map[(a1, a2)] = bond_map[(a2, a1)] = bond
features = []
for i in range(max_natoms):
for j in range(max_natoms):
f = np.zeros((binary_fdim, ))
if i >= n_atoms or j >= n_atoms or i == j:
features.append(f)
continue
if (i, j) in bond_map:
bond = bond_map[(i, j)]
f[1:1 + bond_fdim] = bond_features(bond)
else:
f[0] = 1.0
f[-4] = 1.0 if comp[i] != comp[j] else 0.0
f[-3] = 1.0 if comp[i] == comp[j] else 0.0
f[-2] = 1.0 if n_comp == 1 else 0.0
f[-1] = 1.0 if n_comp > 1 else 0.0
features.append(f)
return np.vstack(features).reshape((max_natoms, max_natoms, binary_fdim))
def sanitize_smiles(smi, largest_fragment=False):
mol = Chem.MolFromSmiles(smi)
if mol is None:
return smi
try:
mol = standardizer.standardize(
mol) # standardize functional group reps
if largest_fragment:
mol = standardizer.largest_fragment(
mol) # remove product counterions/salts/etc.
mol = standardizer.uncharge(
mol) # neutralize, e.g., carboxylic acids
except Exception:
pass
return Chem.MolToSmiles(mol)
def _test_sas(self):
sas_func = lambda mol: calculateSAScore(Chem.MolFromSmiles(mol.smiles))
print(sas_func(Molecule("CC")))
test_pool = ["CC", "O=C=O", "C#N", "CCN(CC)CC", "CC(=O)O", "C1CCCCC1", "c1ccccc1"]
test_pool = [Molecule(smiles) for smiles in test_pool]
exp = RandomExplorer(sas_func, initial_pool=test_pool)
print("Starting SA score optimization")
t0 = time()
exp.run(10)
#check
print("Completed SA score optimization, time elapsed: %.3fs" % (time()-t0))
print(exp.pool)
top = exp.get_best(1)[0]
print(top.get_synthesis_path())
def get_all_batch(re_list):
mol_list = []
max_natoms = 0
for r, e in re_list:
rmol = Chem.MolFromSmiles(r)
mol_list.append((r, e))
if rmol.GetNumAtoms() > max_natoms:
max_natoms = rmol.GetNumAtoms()
labels = []
features = []
sp_labels = []
for r, e in mol_list:
l, sl = get_bond_label(r, e, max_natoms)
features.append(get_bin_feature(r, max_natoms))
labels.append(l)
sp_labels.append(sl)
return np.array(features), np.array(labels), sp_labels
def to_rdkit(self):
"""
Converter to rdkit library format, which is
used for computation of molecular properties
and for synthesis. Performs a validity check.
Returns:
rdkit.Mol -- molecule in RDKit format
Raises:
ValueError -- if SMILES cannot be decoded
into a chemically valid molecule.
"""
if self.rdk is None:
self.rdk = Chem.MolFromSmiles(self.smiles)
if self.rdk is None:
raise ValueError(f"Molecule {self.smiles} is not valid.")
return self.rdk
def predict(self,
react,
top_cand_bonds,
top_cand_scores=[],
scores=True,
top_n=100):
'''react: atom mapped reactant smiles
top_cand_bonds: list of strings "ai-aj-bo"'''
cand_bonds = []
if not top_cand_scores:
top_cand_scores = [0.0 for b in top_cand_bonds]
for i, b in enumerate(top_cand_bonds):
x, y, t = b.split('-')
x, y, t = int(float(x)) - 1, int(float(y)) - 1, float(t)
cand_bonds.append((x, y, t, float(top_cand_scores[i])))
while True:
src_tuple, conf = smiles2graph(react,
None,
cand_bonds,
None,
core_size=core_size,
cutoff=MAX_NCAND,
testing=True)
if len(conf) <= MAX_NCAND:
break
ncore -= 1
feed_map = {x: y for x, y in zip(self.src_holder, src_tuple)}
cur_scores, cur_probs, candidates = self.session.run(
self.predict_vars, feed_dict=feed_map)
idxfunc = lambda a: a.GetAtomMapNum()
bond_types = [
Chem.rdchem.BondType.SINGLE, Chem.rdchem.BondType.DOUBLE,
Chem.rdchem.BondType.TRIPLE, Chem.rdchem.BondType.AROMATIC
]
bond_types_as_double = {0.0: 0, 1.0: 1, 2.0: 2, 3.0: 3, 1.5: 4}
# Don't waste predictions on bond changes that aren't actually changes
rmol = Chem.MolFromSmiles(react)
rbonds = {}
for bond in rmol.GetBonds():
a1 = idxfunc(bond.GetBeginAtom())
a2 = idxfunc(bond.GetEndAtom())
t = bond_types.index(bond.GetBondType()) + 1
a1, a2 = min(a1, a2), max(a1, a2)
rbonds[(a1, a2)] = t
cand_smiles = []
cand_scores = []
cand_probs = []
for idx in candidates:
cbonds = []
# Define edits from prediction
for x, y, t, v in conf[idx]:
x, y = x + 1, y + 1
if ((x, y) not in rbonds and t > 0) or (
(x, y) in rbonds and rbonds[(x, y)] != t):
cbonds.append((x, y, bond_types_as_double[t]))
pred_smiles = edit_mol(rmol, cbonds)
cand_smiles.append(pred_smiles)
cand_scores.append(cur_scores[idx])
cand_probs.append(cur_probs[idx])
outcomes = []
if scores:
for i in range(min(len(cand_smiles), top_n)):
outcomes.append({
'rank': i + 1,
'smiles': cand_smiles[i],
'score': cand_scores[i],
'prob': cand_probs[i],
})
else:
for i in range(min(len(cand_smiles), top_n)):
outcomes.append({
'rank': i + 1,
'smiles': cand_smiles[i],
})
return outcomes
restore_path = tf.train.latest_checkpoint(opts.model_path)
saver.restore(session, restore_path)
sys.stderr.write('restored')
sys.stderr.flush()
total = 0.0
idxfunc = lambda x: x.GetIntProp('molAtomMapNumber')
try:
while not coord.should_stop():
total += 1
r, conf = queue.get(timeout=30)
if r is None: # reached end of data set
break
cur_pred = session.run(pred_topk)
rmol = Chem.MolFromSmiles(r)
rbonds = {}
for bond in rmol.GetBonds():
a1 = idxfunc(bond.GetBeginAtom())
a2 = idxfunc(bond.GetEndAtom())
t = bond_types.index(bond.GetBondType()) + 1
a1, a2 = min(a1, a2), max(a1, a2)
rbonds[(a1, a2)] = t
if opts.verbose:
for idx in cur_pred:
# record the bond changes for this candidate
for x, y, t, v in conf[idx]:
# convert ids to atom map numbers
x, y = x + 1, y + 1
# make sure this bond change is really a _change_
def edit_mol(rmol, edits):
new_mol = Chem.RWMol(rmol)
# Keep track of aromatic nitrogens, might cause explicit hydrogen issues
aromatic_nitrogen_idx = set()
aromatic_carbonyl_adj_to_aromatic_nH = {}
aromatic_carbondeg3_adj_to_aromatic_nH0 = {}
for a in new_mol.GetAtoms():
if a.GetIsAromatic() and a.GetSymbol() == 'N':
aromatic_nitrogen_idx.add(a.GetIdx())
for nbr in a.GetNeighbors():
if a.GetNumExplicitHs() == 1 and nbr.GetSymbol(
) == 'C' and nbr.GetIsAromatic() and any(
b.GetBondTypeAsDouble() == 2 for b in nbr.GetBonds()):
aromatic_carbonyl_adj_to_aromatic_nH[
nbr.GetIdx()] = a.GetIdx()
elif a.GetNumExplicitHs() == 0 and nbr.GetSymbol(
) == 'C' and nbr.GetIsAromatic() and len(nbr.GetBonds()) == 3:
aromatic_carbondeg3_adj_to_aromatic_nH0[
nbr.GetIdx()] = a.GetIdx()
else:
a.SetNumExplicitHs(0)
new_mol.UpdatePropertyCache()
amap = {}
for atom in rmol.GetAtoms():
amap[atom.GetIntProp('molAtomMapNumber')] = atom.GetIdx()
# Apply the edits as predicted
for x, y, t in edits:
bond = new_mol.GetBondBetweenAtoms(amap[x], amap[y])
a1 = new_mol.GetAtomWithIdx(amap[x])
a2 = new_mol.GetAtomWithIdx(amap[y])
if bond is not None:
new_mol.RemoveBond(amap[x], amap[y])
# Are we losing a bond on an aromatic nitrogen?
if bond.GetBondTypeAsDouble() == 1.0:
if amap[x] in aromatic_nitrogen_idx:
if a1.GetTotalNumHs() == 0:
a1.SetNumExplicitHs(1)
elif a1.GetFormalCharge() == 1:
a1.SetFormalCharge(0)
elif amap[y] in aromatic_nitrogen_idx:
if a2.GetTotalNumHs() == 0:
a2.SetNumExplicitHs(1)
elif a2.GetFormalCharge() == 1:
a2.SetFormalCharge(0)
# Are we losing a c=O bond on an aromatic ring? If so, remove H from adjacent nH if appropriate
if bond.GetBondTypeAsDouble() == 2.0:
if amap[x] in aromatic_carbonyl_adj_to_aromatic_nH:
new_mol.GetAtomWithIdx(
aromatic_carbonyl_adj_to_aromatic_nH[
amap[x]]).SetNumExplicitHs(0)
elif amap[y] in aromatic_carbonyl_adj_to_aromatic_nH:
new_mol.GetAtomWithIdx(
aromatic_carbonyl_adj_to_aromatic_nH[
amap[y]]).SetNumExplicitHs(0)
if t > 0:
new_mol.AddBond(amap[x], amap[y], BOND_TYPE[t])
# Special alkylation case?
if t == 1:
if amap[x] in aromatic_nitrogen_idx:
if a1.GetTotalNumHs() == 1:
a1.SetNumExplicitHs(0)
else:
a1.SetFormalCharge(1)
elif amap[y] in aromatic_nitrogen_idx:
if a2.GetTotalNumHs() == 1:
a2.SetNumExplicitHs(0)
else:
a2.SetFormalCharge(1)
# Are we getting a c=O bond on an aromatic ring? If so, add H to adjacent nH0 if appropriate
if t == 2:
if amap[x] in aromatic_carbondeg3_adj_to_aromatic_nH0:
new_mol.GetAtomWithIdx(
aromatic_carbondeg3_adj_to_aromatic_nH0[
amap[x]]).SetNumExplicitHs(1)
elif amap[y] in aromatic_carbondeg3_adj_to_aromatic_nH0:
new_mol.GetAtomWithIdx(
aromatic_carbondeg3_adj_to_aromatic_nH0[
amap[y]]).SetNumExplicitHs(1)
# Tried:
# bonds_to_remove.sort(key=lambda x: x[0], reverse=True)
# for (idx, bond) in bonds_to_remove:
# start = bond.GetBeginAtomIdx()
# end = bond.GetEndAtomIdx()
# new_mol.RemoveBond(start, end)
# pred_mol = new_mol.GetMol()
pred_mol = new_mol.GetMol()
# Clear formal charges to make molecules valid
# Note: because S and P (among others) can change valence, be more flexible
for atom in pred_mol.GetAtoms():
atom.ClearProp('molAtomMapNumber')
if atom.GetSymbol() == 'N' and atom.GetFormalCharge(
) == 1: # exclude negatively-charged azide
bond_vals = sum(
[bond.GetBondTypeAsDouble() for bond in atom.GetBonds()])
if bond_vals <= 3:
atom.SetFormalCharge(0)
elif atom.GetSymbol() == 'N' and atom.GetFormalCharge(
) == -1: # handle negatively-charged azide addition
bond_vals = sum(
[bond.GetBondTypeAsDouble() for bond in atom.GetBonds()])
if bond_vals == 3 and any(
[nbr.GetSymbol() == 'N' for nbr in atom.GetNeighbors()]):
atom.SetFormalCharge(0)
elif atom.GetSymbol() == 'N':
bond_vals = sum(
[bond.GetBondTypeAsDouble() for bond in atom.GetBonds()])
if bond_vals == 4 and not atom.GetIsAromatic(
): # and atom.IsInRingSize(5)):
atom.SetFormalCharge(1)
elif atom.GetSymbol() == 'C' and atom.GetFormalCharge() != 0:
atom.SetFormalCharge(0)
elif atom.GetSymbol() == 'O' and atom.GetFormalCharge() != 0:
bond_vals = sum(
[bond.GetBondTypeAsDouble()
for bond in atom.GetBonds()]) + atom.GetNumExplicitHs()
if bond_vals == 2:
atom.SetFormalCharge(0)
elif atom.GetSymbol() in ['Cl', 'Br', 'I', 'F'
] and atom.GetFormalCharge() != 0:
bond_vals = sum(
[bond.GetBondTypeAsDouble() for bond in atom.GetBonds()])
if bond_vals == 1:
atom.SetFormalCharge(0)
elif atom.GetSymbol() == 'S' and atom.GetFormalCharge() != 0:
bond_vals = sum(
[bond.GetBondTypeAsDouble() for bond in atom.GetBonds()])
if bond_vals in [2, 4, 6]:
atom.SetFormalCharge(0)
elif atom.GetSymbol(
) == 'P': # quartenary phosphorous should be pos. charge with 0 H
bond_vals = [
bond.GetBondTypeAsDouble() for bond in atom.GetBonds()
]
if sum(bond_vals) == 4 and len(bond_vals) == 4:
atom.SetFormalCharge(1)
atom.SetNumExplicitHs(0)
elif sum(bond_vals) == 3 and len(
bond_vals) == 3: # make sure neutral
atom.SetFormalCharge(0)
elif atom.GetSymbol(
) == 'B': # quartenary boron should be neg. charge with 0 H
bond_vals = [
bond.GetBondTypeAsDouble() for bond in atom.GetBonds()
]
if sum(bond_vals) == 4 and len(bond_vals) == 4:
atom.SetFormalCharge(-1)
atom.SetNumExplicitHs(0)
elif atom.GetSymbol() in ['Mg', 'Zn']:
bond_vals = [
bond.GetBondTypeAsDouble() for bond in atom.GetBonds()
]
if sum(bond_vals) == 1 and len(bond_vals) == 1:
atom.SetFormalCharge(1)
elif atom.GetSymbol() == 'Si':
bond_vals = [
bond.GetBondTypeAsDouble() for bond in atom.GetBonds()
]
if sum(bond_vals) == len(bond_vals):
atom.SetNumExplicitHs(max(0, 4 - len(bond_vals)))
# Bounce to/from SMILES to try to sanitize
pred_smiles = Chem.MolToSmiles(pred_mol) # <--- TODO: error occurs here
pred_list = pred_smiles.split('.')
pred_mols = [Chem.MolFromSmiles(pred_smiles) for pred_smiles in pred_list]
for i, mol in enumerate(pred_mols):
# Check if we failed/succeeded in previous step
if mol is None:
logging.debug('##### Unparseable mol: {}'.format(pred_list[i]))
continue
# Else, try post-sanitiztion fixes in structure
mol = Chem.MolFromSmiles(Chem.MolToSmiles(mol))
if mol is None:
continue
for rxn in clean_rxns_postsani:
out = rxn.RunReactants((mol, ))
if out:
try:
Chem.SanitizeMol(out[0][0])
pred_mols[i] = Chem.MolFromSmiles(
Chem.MolToSmiles(out[0][0]))
except Exception as e:
print(e)
print('Could not sanitize postsani reaction product: {}'.
format(Chem.MolToSmiles(out[0][0])))
print('Original molecule was: {}'.format(
Chem.MolToSmiles(mol)))
pred_smiles = [
Chem.MolToSmiles(pred_mol) for pred_mol in pred_mols
if pred_mol is not None
]
return pred_smiles
def smiles2graph(rsmiles,
psmiles,
core_bonds,
gold_bonds,
cutoff=500,
idxfunc=lambda x: x.GetIntProp('molAtomMapNumber') - 1,
core_size=20,
kmax=5,
return_found=False,
testing=False):
'''This is the function that takes reactants, a true product (when defined), and the candidate bonds
to generate all of the candidate products according to some bounds on the enumeration'''
mol = Chem.MolFromSmiles(rsmiles)
if not mol:
raise ValueError("Could not parse smiles string:", rsmiles)
if not testing:
pmol = Chem.MolFromSmiles(psmiles)
if not pmol:
raise ValueError("Could not parse smiles string:", psmiles)
n_atoms = mol.GetNumAtoms()
n_bonds = max(mol.GetNumBonds(), 1)
fatoms = np.zeros((n_atoms, atom_fdim))
fbonds = np.zeros((n_bonds, bond_fdim))
atom_nb = np.zeros((n_atoms, max_nb), dtype=np.int32)
bond_nb = np.zeros((n_atoms, max_nb), dtype=np.int32)
num_nbs = np.zeros((n_atoms, ), dtype=np.int32)
raw_atom_nb = np.zeros((n_atoms, max_nb), dtype=np.int32)
raw_bond_nb = np.zeros((n_atoms, max_nb), dtype=np.int32)
raw_num_nbs = np.zeros((n_atoms, ), dtype=np.int32)
free_vals = np.zeros((n_atoms, ))
pfree_vals = np.zeros((n_atoms, ))
is_c2_of_pyridine = np.zeros((n_atoms, ), dtype=bool)
is_c = np.zeros((n_atoms, ), dtype=bool)
is_p = np.zeros((n_atoms, ), dtype=bool)
is_s = np.zeros((n_atoms, ), dtype=bool)
is_o = np.zeros((n_atoms, ), dtype=bool)
is_n = np.zeros((n_atoms, ), dtype=bool)
# gbonds = {(x,y):0 for x,y in core_bonds}
#Feature Extraction
for atom in mol.GetAtoms():
idx = idxfunc(atom)
fatoms[idx] = atom_features(atom)
free_vals[idx] += atom.GetTotalNumHs() + abs(atom.GetFormalCharge())
# TODO: review these rules
# Aromatic carbon next to an aromatic nitrogen can get a carbonyl b/c stupid bookkeeping of hydroxypyridines
if atom.GetAtomicNum() == 6:
is_c[idx] = True
if atom.GetIsAromatic():
for nbr in atom.GetNeighbors():
if nbr.GetAtomicNum() == 7 and nbr.GetDegree() == 2:
is_c2_of_pyridine[idx] = True
break
# Nitrogen should be allowed to become positively charged
elif atom.GetAtomicNum() == 7:
free_vals[idx] += 1 - atom.GetFormalCharge()
is_n[idx] = True
# Phosphorous can form a phosphonium
elif atom.GetAtomicNum() == 15:
free_vals[idx] += 1 - atom.GetFormalCharge()
is_p[idx] = True
elif atom.GetAtomicNum() == 8:
is_o[idx] = True
elif atom.GetAtomicNum() == 16:
is_s[idx] = True
# special information needed for valence filtering
if not testing:
tatoms = set()
#Calculate free slots for each atom in product
for bond in pmol.GetBonds():
a1 = idxfunc(bond.GetBeginAtom())
a2 = idxfunc(bond.GetEndAtom())
t = bond_types.index(bond.GetBondType()) + 1
a1, a2 = min(a1, a2), max(a1, a2)
tatoms.add(a1)
tatoms.add(a2)
if (a1, a2) in core_bonds:
# gbonds[(a1,a2)] = t
tval = t if t < 4 else 1.5
pfree_vals[a1] += tval
pfree_vals[a2] += tval
rbonds = {}
rbond_vals = {} # bond orders
ring_bonds = set()
#Calculate free slots for each atom in reactant
for bond in mol.GetBonds():
idx = bond.GetIdx()
a1 = idxfunc(bond.GetBeginAtom())
a2 = idxfunc(bond.GetEndAtom())
t = bond_types.index(bond.GetBondType())
a1, a2 = min(a1, a2), max(a1, a2)
tval = t + 1 if t < 3 else 1.5
rbonds[(a1, a2)] = t + 1
rbond_vals[(a1, a2)] = tval
if (a1, a2) in core_bonds:
free_vals[a1] += tval
free_vals[a2] += tval
if bond.IsInRing():
ring_bonds.add((a1, a2))
# Get all possible core configurations - NEW IN DIRECT VERSION
from itertools import combinations
core_configs = [
] # will be list of lists of (x, y, t, v) tuples, where t is the bond order and v is CoreFinder score
# print('rbond_vals:')
# print(rbond_vals)
# Filter out core bonds that exactly match reactants
prev_len = len(core_bonds)
core_bonds = [(x, y, t, v) for (x, y, t, v) in core_bonds
if ((x, y) not in rbond_vals) or (rbond_vals[(x, y)] != t)]
# print('{}/{} core bonds kept after filtering existing bonds'.format(prev_len, len(core_bonds)))
# Pare down to top-core_size only
core_bonds = core_bonds[:core_size]
# Helper function to check if a combination is connected - this helps the number of valid combinations
core_bonds_adj = np.eye(len(core_bonds), dtype=bool)
for i in range(len(core_bonds)):
a1, b1, t1, v1 = core_bonds[i]
for j in range(i, len(core_bonds)):
a2, b2, t2, v2 = core_bonds[j]
if a1 == a2 or a1 == b2 or b1 == a2 or b1 == b2:
core_bonds_adj[i, j] = core_bonds_adj[j, i] = True
# print(core_bonds)
# print('Calculated core bonds adj matrix: {}'.format(core_bonds_adj * 1.0))
def check_if_connected(combo_i):
'''Checks if a set of candidate edits (by indeces) are all connected'''
if len(combo_i) == 1:
return True # only one change, always connected
temp_adj_pow = np.linalg.matrix_power(
core_bonds_adj[combo_i, :][:, combo_i],
len(combo_i) - 1)
return np.all(temp_adj_pow)
# Helper function to check if a combiation is valid
def check_if_valid(bond_change_combo):
force_even_parity = np.zeros((n_atoms, ), dtype=bool)
force_odd_parity = np.zeros((n_atoms, ), dtype=bool)
seen = defaultdict(lambda: False)
free_vals_temp = free_vals.copy()
for x, y, t, v in bond_change_combo:
x, y = tuple(sorted([x, y]))
if seen[(x, y)]:
# print('already seen this bond in the list of cand changes')
return False # can't have two distinct bond change types in same combo
seen[(x, y)] = True
# TODO: review these valence rules
# Special rules:
# - if phosphorous or sulfur, don't count formation of =O toward valence but require odd/even
# - if c2 carbon in a pyridine ring, let it get a =O
tx = ty = t
if t == 2:
if is_o[x]:
if is_c2_of_pyridine[y]:
ty = 1. # pretend it's just a hydroxylation for the sake of valence
elif is_p[y]:
ty = 0. # don't count toward valence
force_odd_parity[
y] = True # but require odd valence parity
elif is_s[y]:
ty = 0.
force_even_parity[y] = True
elif is_o[y]:
if is_c2_of_pyridine[x]:
tx = 1.
elif is_p[x]:
tx = 0.
force_odd_parity[x] = True
elif is_s[x]:
tx = 0.
force_even_parity[x] = True
elif is_n[x] and is_p[y]:
ty = 0.
force_odd_parity[y] = True
elif is_n[y] and is_p[x]:
tx = 0.
force_odd_parity[x] = True
elif is_p[x] and is_c[y]:
tx = 0.
force_odd_parity[x] = True
elif is_p[y] and is_c[x]:
ty = 0.
force_odd_parity[y] = True
if (x, y) in rbond_vals:
free_vals_temp[x] += rbond_vals[(x, y)] - tx
free_vals_temp[y] += rbond_vals[(x, y)] - ty
else:
free_vals_temp[x] += -tx
free_vals_temp[y] += -ty
# too many connections? sulfur valence not even? phosphorous valence not odd?
if any(free_vals_temp < 0) \
or any(aval % 2 != 0 for aval in free_vals_temp[force_even_parity]) \
or any(aval % 2 != 1 for aval in free_vals_temp[force_odd_parity]):
# print('invalid valence?')
# print(free_vals_temp)
return False
return True
# N choose k combinatorics
# up to 4 bond changes at once - only 0.19% of train examples have 5 bonds changed, we can take the hit...
core_bonds_i = range(len(core_bonds))
for k in range(1, kmax + 1):
for bond_change_combo_i in combinations(core_bonds_i, k):
# Check if connected
if not check_if_connected(bond_change_combo_i):
# print('This combination is not connected!')
continue
bond_change_combo = [core_bonds[i] for i in bond_change_combo_i]
if check_if_valid(bond_change_combo):
core_configs.append(bond_change_combo)
# print('Found a total of {} core configs that seem valid'.format(len(core_configs)))
if not testing:
random.shuffle(core_configs)
idx = -1
for i, cand_bonds in enumerate(core_configs):
if set([(x, y, t) for (x, y, t, v) in cand_bonds]) == gold_bonds:
idx = i
break
# If we are training and did not find the true outcome, make sure it is the first entry
if idx == -1:
# print('Did not find true outcome')
found_true = False
core_configs = [[(x, y, t, 0.0)
for (x, y, t) in gold_bonds]] + core_configs
else:
# print('Found true outcome')
found_true = True
core_configs[0], core_configs[idx] = core_configs[
idx], core_configs[0] # swap order so true is first
else:
found_true = False
if not testing:
# If it is possible to recover the true smiles from the set of bonds using the edit_mol method,
# remove duplicates from the list by converting each candidate into a smiles string
# note: get_product_smiles is HIGHLY imperfect, but that's not a huge deal. training tries to pick the
# right bonds. The evaluation script has a more robust function to get product_smiles
smiles0 = get_product_smiles(mol, core_configs[0], tatoms)
if len(smiles0) > 0: #
cand_smiles = set([smiles0])
new_core_configs = [core_configs[0]]
for core_conf in core_configs[1:]:
smiles = get_product_smiles(mol, core_conf, tatoms)
# print('candidate smiles: {}'.format(smiles))
if smiles in cand_smiles or len(smiles) == 0:
continue
cand_smiles.add(smiles)
new_core_configs.append(core_conf)
core_configs = new_core_configs
else:
print('\nwarning! could not recover true smiles from gbonds: {}'.
format(psmiles))
print('{} {}'.format(rsmiles, gold_bonds))
# print('After removing duplicates, {} core configs'.format(len(core_configs)))
core_configs = core_configs[:cutoff]
n_batch = len(core_configs) + 1
if not testing:
labels = np.zeros((n_batch - 1, ))
labels[0] = 1
# Calculate information that is the same for all candidates; do small updates based on specific changes later
pending_reactant_neighbors = [
] # reactant neighbors that *might* be over-ridden
core_bonds_noScore = [(x, y, t) for (x, y, t, z) in core_bonds]
for bond in mol.GetBonds():
idx = bond.GetIdx()
a1 = idxfunc(bond.GetBeginAtom())
a2 = idxfunc(bond.GetEndAtom())
a1, a2 = min(a1, a2), max(a1, a2)
if (
a1, a2, 0.0
) not in core_bonds_noScore: # are a1 and a2 guaranteed to be neighbors?
raw_atom_nb[a1, raw_num_nbs[a1]] = a2
raw_atom_nb[a2, raw_num_nbs[a2]] = a1
raw_bond_nb[a1, raw_num_nbs[a1]] = idx
raw_bond_nb[a2, raw_num_nbs[a2]] = idx
raw_num_nbs[a1] += 1
raw_num_nbs[a2] += 1
else:
pending_reactant_neighbors.append(
(a1, a2, bond.GetBondTypeAsDouble()))
# Reactants have this bond...
atom_nb[a1, num_nbs[a1]] = a2
atom_nb[a2, num_nbs[a2]] = a1
bond_nb[a1, num_nbs[a1]] = idx
bond_nb[a2, num_nbs[a2]] = idx
num_nbs[a1] += 1
num_nbs[a2] += 1
fbonds[idx] = bond_features(bond)
# print('What is core_bonds here?: {}'.format(core_bonds))
if not testing:
num_newbonds = max(
len(gold_bonds), len(core_bonds)
) * 2 + 1 # CC fixed in case where core_bonds isn't large enough
else:
num_newbonds = len(core_bonds) * 2 + 1
new_fbonds = np.zeros(
(n_bonds + num_newbonds + len(pending_reactant_neighbors),
bond_fdim)) # CC added + len(pending_reactant_neighbors)
new_fbonds[:n_bonds, :] = fbonds
fbonds = new_fbonds
batch_fbonds, batch_anb, batch_bnb, batch_nbs = [fbonds], [atom_nb], [
bond_nb
], [num_nbs] # first entry is reactants
batch_corebias = []
for core_bonds in core_configs:
atom_nb2 = np.copy(raw_atom_nb)
bond_nb2 = np.copy(raw_bond_nb)
num_nbs2 = np.copy(raw_num_nbs)
fbonds2 = np.copy(fbonds)
n_bonds2 = n_bonds + 1
# Add back reactant bonds?
core_bonds_nobo = [(x, y) for (x, y, t, v) in core_bonds]
for (x, y, t) in pending_reactant_neighbors:
if (x, y) not in core_bonds_nobo:
core_bonds.append((x, y, t, 0.0))
for x, y, t, v in core_bonds: # add new bond features to the "default" reactant ones
if t == 0: continue
atom_nb2[x, num_nbs2[x]] = y
atom_nb2[y, num_nbs2[y]] = x
bond_nb2[x, num_nbs2[x]] = n_bonds2
bond_nb2[y, num_nbs2[y]] = n_bonds2
num_nbs2[x] += 1
num_nbs2[y] += 1
fbonds2[n_bonds2] = onek_encoding_unk(t, [1.0, 2.0, 3.0, 1.5, -1])
if (x, y) in ring_bonds:
fbonds2[n_bonds2][4] = 1
n_bonds2 += 1
batch_fbonds.append(fbonds2)
batch_anb.append(atom_nb2)
batch_bnb.append(bond_nb2)
batch_nbs.append(num_nbs2)
batch_corebias.append(sum([v for (x, y, t, v) in core_bonds]))
# TODO: change atom features for each candidate? Maybe update degree at least
if return_found:
return (np.array([fatoms] * n_batch), np.array(batch_fbonds),
packnb(batch_anb), packnb(batch_bnb), np.array(batch_nbs),
np.array(batch_corebias), labels), core_configs, found_true
if not testing:
return (np.array([fatoms] * n_batch), np.array(batch_fbonds),
packnb(batch_anb), packnb(batch_bnb), np.array(batch_nbs),
np.array(batch_corebias), labels), core_configs
return (np.array([fatoms] * n_batch), np.array(batch_fbonds),
packnb(batch_anb), packnb(batch_bnb), np.array(batch_nbs),
np.array(batch_corebias)), core_configs
a = np.zeros((len(arr_list), N))
for i, arr in enumerate(arr_list):
for j in range(arr.shape[0]):
a[i][j] = 1
return a
def smiles2graph_list(smiles_list, idxfunc=lambda x: x.GetIdx()):
res = list(map(lambda x: smiles2graph(x, idxfunc), smiles_list))
fatom_list, fbond_list, gatom_list, gbond_list, nb_list = zip(*res)
return pack2D(fatom_list), pack2D(fbond_list), pack2D_withidx(
gatom_list), pack2D_withidx(gbond_list), pack1D(nb_list), get_mask(
fatom_list)
m = Chem.MolFromSmiles('CC')
assignProperties(m)
atom = m.GetAtoms()[0]
bond = m.GetBonds()[0]
atom_fdim = len(atom_features(atom))
bond_fdim = len(bond_features(bond))
if __name__ == "__main__":
np.set_printoptions(threshold='nan')
a, b, c, d, e, f = smiles2graph_list(["c1cccnc1", 'c1nccc2n1ccc2'])
print(a)
print(b)
print(c)
print(d)
print(e)
print(f)