def predict_outcome(self, reaction, k=1):
"""
Using a predictor, produce top-k most likely reactions
Params:
reaction {Reaction}
k {int} - how many top predictions to set and return
Returns:
{list[Molecule]} - list of products of reaction
"""
react = reaction.get_input_str()
try:
(react, bond_preds, bond_scores, cur_att_score) = self.directcorefinder.predict(react)
outcomes = self.directcandranker.predict(react, bond_preds, bond_scores)
except RuntimeError as e:
logging.error(f"Error occured in DirectCandRanker.predict: {e}")
raise e
res = []
for out in outcomes:
if out["smiles"]: # may be empty for some reason?
smiles = out["smiles"][0]
mol = Molecule(smiles)
mol.set_synthesis(reaction.inputs)
res.append(mol)
else:
continue
res = res[:k]
# setting predicted products, if not already set:
reaction.set_products(res)
return res
def _draw_node(self, node: Molecule):
import os
self._node_counter += 1
if self._draw_mode == "smiles":
self._dot.node(name=node.to_smiles(), label=node.to_smiles())
elif self._draw_mode == "formula":
self._dot.node(name=node.to_smiles(), label=node.to_formula())
elif self._draw_mode == "plot":
mol_img_path = os.path.join(self._sub_dir,
str(self._node_counter) + ".png")
visualize_mol(node, path=mol_img_path)
self._dot.node(name=node.to_smiles(),
label="",
image=mol_img_path,
shape="plaintext")
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_chembl(n_mols=None, as_mols=True, option='', max_size=1000):
"""
Return list of SMILES.
NOTE: this function should be located
in the same directory as data files.
"""
path = os.path.join(__location__, "ChEMBL.txt")
with open(path, "r") as f:
if n_mols is None:
res = [line.strip() for line in f]
else:
res = [f.readline().strip() for _ in range(n_mols)]
mols = [Molecule(smile) for smile in res]
if len(mols) < max_size:
return mols
gen = np.random.RandomState(42)
mols = list(gen.choice(mols, max_size, replace=False))
if option == '':
return mols
elif option == 'small_qed':
qed_func = get_objective_by_name("qed")
return [mol for mol in mols if qed_func(mol) < 0.6]
elif option == 'large_qed':
qed_func = get_objective_by_name("qed")
return [mol for mol in mols if qed_func(mol) >= 0.6]
else:
raise ValueError(f"Dataset filter {option} not supported.")
def test_chembl(self):
"""
Problem with fixed-prop testing:
Almost all of the results (<10% for init_pool of 50) seem to be outside of the database,
and even less for smaller pool. Hence cannot get its score for testing;
setting them to zero leads to slow exploration.
"""
pool_all, dd = get_chembl_prop()
# loading with mol conversions takes 8 minutes
# pool_all = [Molecule(smiles, conv_enabled=True) for smiles in tqdm(pool_all[:10000])]
pool_all = [Molecule(smiles, conv_enabled=False) for smiles in pool_all]
start_pool = list(np.random.choice(pool_all, size=100, replace=False))
def print_props(pool):
props = [dd[mol.smiles] for mol in pool]
print("Props of pool", len(pool), np.min(props), np.mean(props), np.max(props))
print_props(pool_all)
print_props(start_pool)
func = lambda mol: dd[mol.smiles]
exp = RandomExplorer(lambda mol_list: func(mol_list[0]), initial_pool=start_pool)
print("Starting ChEMBL score 1 optimization")
t0 = time()
exp.run(30)
print("Completed ChEMBL score 1 optimization, time elapsed: %.3fs" % (time()-t0))
# print(exp.pool)
top = exp.get_best(1)[0]
print(top.get_synthesis_path())
print("Best achieved score: %.3f" % func(top))
props = [dd[mol.smiles] for mol in pool_all]
print("Best possible score: %.3f" % np.max(props))
def get_min_score(syn):
res = float('inf')
for mol, syn_graph in syn.items():
# if mol.begin_flag:
if isinstance(syn_graph, str):
return sa_score(Molecule(mol))
res = min(res, get_min_score(syn_graph))
return res
def get_graph_data_for_distance_computation(mol):
""" Returns graph representation for a molecule. """
if isinstance(mol, str):
from mols.molecule import Molecule
mol = Molecule(mol)
rdk_mol = mol.to_rdkit()
rdk_mol = Chem.AddHs(rdk_mol)
adj_matrix = Chem.rdmolops.GetAdjacencyMatrix(rdk_mol)
bonds = [(b.GetBeginAtomIdx(), b.GetEndAtomIdx())
for b in rdk_mol.GetBonds()]
bond_types = [
rdk_mol.GetBondBetweenAtoms(b[0], b[1]).GetBondType() for b in bonds
]
atom_idxs = list(range(len(rdk_mol.GetAtoms())))
atomic_numbers = [
rdk_mol.GetAtomWithIdx(idx).GetAtomicNum() for idx in atom_idxs
]
atomic_symbols = [
rdk_mol.GetAtomWithIdx(idx).GetSymbol() for idx in atom_idxs
]
atomic_masses = [
rdk_mol.GetAtomWithIdx(idx).GetMass() for idx in atom_idxs
]
num_atoms = len(atom_idxs)
bonds_of_each_atom = [
get_neighbors_and_bond_types(idx, bonds, atomic_symbols, bond_types)
for idx in range(num_atoms)
]
bond_type_counts_of_each_atom = [
get_bond_type_counts(bt) for bt in bonds_of_each_atom
]
# Return
graph_data = Namespace(
rdk_mol=rdk_mol,
adj_matrix=adj_matrix,
bonds=bonds,
bond_types=bond_types,
atom_idxs=atom_idxs,
atomic_numbers=atomic_numbers,
atomic_symbols=atomic_symbols,
atomic_masses=atomic_masses,
num_atoms=num_atoms,
bonds_of_each_atom=bonds_of_each_atom,
bond_type_counts_of_each_atom=bond_type_counts_of_each_atom,
)
return graph_data
def draw_molecule(mol: Molecule) -> PIL.Image.Image:
"""
Draw a single molecule `mol` (make it `PIL.Image.Image`)
:param mol: molecule to draw
:return: corresponding image to `mol`
"""
img = Draw.MolToImage(mol.to_rdkit())
return img
def compute_synthesizability(exp_path):
sas = get_objective_by_name("sascore")
mol = None
with open(os.path.join(exp_path, 'exp_log'), 'r') as f:
for line in f:
if 'Resulting molecule' in line:
mol = Molecule(smiles=line.split()[2])
if not mol: return
sa_score = sas(mol)
return sa_score
def _test_len(self):
dummy_func = lambda mol: len(mol.smiles)
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(dummy_func, initial_pool=test_pool)
print("Starting len of SMILES optimization")
exp.run(2)
#check
print(exp.pool)
def draw_synthesis_path(target_smiles: str, synth_path: str,
out_path: str) -> None:
"""Draw the synthesis path and save to provided location.
:param target_smiles: SMILES of the molecule being synthesized
:param synth_path: dictionary of format SMILES m -> synpath of m
:param out_path: where to save the resulting pdf
"""
with open("./mols/best_molecule.pkl", "rb") as f:
synpath = pickle.load(f)
synpath = smile_synpath_to_mols(Molecule(smiles=target_smiles), synpath)
drawer = SynPathDrawer(synpath, "plot")
drawer.render(out_path)
def get_chembl(option='', max_size=1000, as_mols=True):
"""
Return list of Molecules.
NOTE: this function should be located
in the same directory as data files.
Arguments:
option {str} -- either empty or of format '{small,large}_{objective name}'
max_size {int} -- number of molecules to sample, if None, returns all,
else randomly samples a subset. Attention: there is a randomly set random seed
that seeds this sampler now, so the subset will always be the same.
as_mols {bool} -- whether to wrap SMILES into the Molecule class
"""
path = os.path.join(__location__, "ChEMBL.txt")
with open(path, "r") as f:
mols = [line.strip() for line in f]
if as_mols:
mols = [Molecule(smile) for smile in mols]
if max_size == -1:
max_size = len(mols)
if len(mols) <= max_size:
return mols
# TODO: this logic is off, if filtering afterwards,
# we get less than max_size molecules in the end.
# Fix this if needed.
gen = np.random.RandomState(42)
mols = list(gen.choice(mols, max_size, replace=False))
if option == '':
return mols
elif option.startswith('small_'):
obj_name = option.split("_")[1]
obj_func = get_objective_by_name(obj_name)
small_thresh = get_threshold(obj_name, mode='low')
return [mol for mol in mols if obj_func(mol) < small_thresh]
elif option.startswith('large_'):
obj_name = option.split("_")[1]
obj_func = get_objective_by_name(obj_name)
large_thresh = get_threshold(obj_name, mode='high')
return [mol for mol in mols if obj_func(mol) >= large_thresh]
else:
raise ValueError(f"Dataset filter {option} not supported.")
def compute_min_sa_score(mol):
""" Compute sas scores along the synthesis path of molecule. """
sa_score = get_objective_by_name("sascore")
def get_min_score(syn):
res = float('inf')
for mol, syn_graph in syn.items():
# if mol.begin_flag:
if isinstance(syn_graph, str):
return sa_score(Molecule(mol))
res = min(res, get_min_score(syn_graph))
return res
synthesis_path = mol.get_synthesis_path()
if isinstance(synthesis_path, dict):
min_sa_score = get_min_score(synthesis_path)
else:
min_sa_score = sa_score(Molecule(synthesis_path))
return min_sa_score
def __init__(self, mol: Molecule, draw_mode: str):
"""
:param mol: the molecule to draw synthesis path for
:param draw_mode: "smiles" | "formula" | "plot" way of plotting each single molecule
Examples::
>>> drawer = SynPathDrawer(root_mol, "smiles") # or "formula" or "plot"
>>> drawer.render("some_output_dir/some_file_name") # please, no file extension
"""
assert draw_mode in ["smiles", "formula", "plot"]
from graphviz import Digraph
self._mol = mol
self._dot = Digraph(comment="Synthesis path for {}".format(
mol.to_smiles()),
format="pdf")
self._draw_mode = draw_mode
self._node_counter = 0
self._sub_dir = None
def get_zinc250(option='', max_size=1000):
path = os.path.join(__location__, "zinc250k.csv")
zinc_df = pd.read_csv(path)
list_of_smiles = list(map(lambda x: x.strip(), zinc_df.smiles.values))
# other columns are logP, qed, and sas
mols = [Molecule(smile) for smile in res]
if len(mols) < max_size:
return mols
gen = np.random.RandomState(42)
mols = list(gen.choice(mols, max_size, replace=False))
if option == '':
return mols
elif option == 'small_qed':
qed_func = get_objective_by_name("qed")
return [mol for mol in mols if qed_func(mol) < 0.6]
elif option == 'large_qed':
qed_func = get_objective_by_name("qed")
return [mol for mol in mols if qed_func(mol) >= 0.6]
else:
raise ValueError(f"Dataset filter {option} not supported.")
def get_zinc250(option='', max_size=1000, as_mols=True):
"""
Return list of Molecules.
NOTE: this function should be located
in the same directory as data files.
Arguments:
option {str} -- either empty or of format '{small,large}_{objective name}'
max_size {int} -- number of molecules to sample, if None, returns all,
else randomly samples a subset. Attention: there is a randomly set random seed
that seeds this sampler now, so the subset will always be the same.
as_mols {bool} -- whether to wrap SMILES into the Molecule class
"""
path = os.path.join(__location__, "zinc250k.csv")
zinc_df = pd.read_csv(path)
list_of_smiles = list(map(lambda x: x.strip(), zinc_df.smiles.values))
# other columns are logP, qed, and sas
mols = [Molecule(smile) for smile in list_of_smiles]
if max_size == -1:
max_size = len(mols)
if len(mols) <= max_size:
return mols
gen = np.random.RandomState(42)
mols = list(gen.choice(mols, max_size, replace=False))
if option == '':
return mols
elif option.startswith('small_'):
obj_func = get_objective_by_name(option.split("_")[1])
return [mol for mol in mols if obj_func(mol) < 0.6]
elif option.startswith('large_'):
obj_func = get_objective_by_name(option.split("_")[1])
return [mol for mol in mols if obj_func(mol) >= 0.6]
else:
raise ValueError(f"Dataset filter {option} not supported.")
except RuntimeError as e:
logging.error(f"Error occured in DirectCandRanker.predict: {e}")
raise e
res = []
for out in outcomes:
if out["smiles"]: # may be empty for some reason?
smiles = out["smiles"][0]
mol = Molecule(smiles)
mol.set_synthesis(reaction.inputs)
res.append(mol)
else:
continue
res = res[:k]
# setting predicted products, if not already set:
reaction.set_products(res)
return res
if __name__=="__main__":
list_of_mols = ["[CH3:26][c:27]1[cH:28][cH:29][cH:30][cH:31][cH:32]1",
"[Cl:18][C:19](=[O:20])[O:21][C:22]([Cl:23])([Cl:24])[Cl:25]",
"[NH2:1][c:2]1[cH:3][cH:4][c:5]([Br:17])[c:6]2[c:10]1[O:9][C:8]"+
"([CH3:11])([C:12](=[O:13])[O:14][CH2:15][CH3:16])[CH2:7]2"
]
list_of_mols = [Molecule(smiles) for smiles in list_of_mols]
t = RexgenForwardSynthesizer()
reaction = Reaction(list_of_mols)
t.predict_outcome(reaction)
def setUp(self):
S1, S2, S3 = "Cc1ccccc1", "C1OC1", "CCOC(=O)C1=C[C@@H](OC(CC)CC)[C@H](NC(C)=O)[C@@H](N)C1"
self.mols = [Molecule(S1), Molecule(S2), Molecule(S3)]
def members_are_equal(cls, point_1, point_2):
""" Technically, because SMILES are not unique,
this may sometimes give false negatives.
TODO: graph structure matching?
"""
return mol1.to_smiles() == mol2.to_smiles()
def __str__(self):
""" Returns a string representation. """
cc_attrs = ""
if hasattr(self, "constraint_checker") and self.constraint_checker is not None:
cc_attrs = {key:getattr(self.constraint_checker, key)
for key in self.constraint_checker.constraint_names}
return 'Mol(%s):%s'%(self.mol_type, cc_attrs)
# Different constraint checker functions(Molecule -> bool) --------------------
def has_carbon(mol):
rdk = mol.to_rdkit()
atomic_symbols = [rdk.GetAtomWithIdx(idx).GetSymbol() for idx in range(len(rdk.GetAtoms()))]
mol_has_carbon = ('C' in atomic_symbols)
print(atomic_symbols, mol_has_carbon)
return mol_has_carbon
if __name__ == "__main__":
mol = Molecule("C=C1NC(N(C)C)=NC12CCN(CC(C)c1ccccc1)CC2")
has_carbon(mol)
def setUp(self):
self.mol = Molecule("CC")
# def draw_synthesis_path(mol):
# def compute_depth(syn_path):
# depth = 1
# if not mol.begin_flag:
# for inp, inp_syn_path in syn_path:
# inp_depth = compute_depth(inp_syn_path)
# depth = max(depth, inp_depth)
# return depth
#
# syn_path = mol.get_syn_path()
# depth = compute_depth(syn_path) # number of rows to allocate for plotting
# imgs_per_row = []
# min_shape = None
#
# # traverse the synthesis path and append images to imgs_per_row
# # each row should be concatenated: see
# # https://stackoverflow.com/questions/30227466/combine-several-images-horizontally-with-python
#
# # TODO
#
# imgs_comb = np.vstack([np.asarray(img.resize(min_shape))
# for img in imgs_per_row])
# result_img = PIL.Image.fromarray(imgs_comb)
# return result_img
if __name__ == "__main__":
mol = Molecule("CCCC")
img = draw_molecule(mol)
img.save('./experiments/results/test.png')
def _draw_edge(self, tail: Molecule, head: Molecule):
self._dot.edge(tail_name=tail.to_smiles(), head_name=head.to_smiles())