def tanimoto_sml(queryism, targetism):
"""
Returns the Tanimoto similarity between the two molecules in SMILES format.
"""
# no Unicode here
querymol = MolFromSmiles(str(queryism))
targetmol = MolFromSmiles(str(targetism))
if querymol and targetmol:
queryfp = RDKFingerprint(querymol)
targetfp = RDKFingerprint(targetmol)
return TanimotoSimilarity(queryfp, targetfp)
def similarity_search(fps_db, smile):
fps_test = RDKFingerprint(MolFromSmiles(smile))
ts = []
for i, s_top in enumerate(fps_db):
ts.append(DataStructs.FingerprintSimilarity(s_top, fps_test))
ts = np.array(ts)
return ts.mean() # ts.max()
def _set_target_fps(self, pickaxe: Pickaxe):
for smiles in pickaxe.target_smiles:
mol = MolFromSmiles(smiles)
if self.fingerprint_method == "Morgan":
fp = AllChem.GetMorganFingerprintAsBitVect(mol, **self.fingerprint_args)
else:
fp = RDKFingerprint(mol)
self.target_fps.append(fp)
def fingerprint(self, m, fpsize=1024, bitsperhash=2, tgtDensity=0.3):
"""Compute bit fingerprint of molecule m"""
from rdkit.Chem import RDKFingerprint
#from rdkit import DataStructs
#fp=RDKFingerprint(m, minPath=1, maxPath=7,fpSize=1024, bitsPerHash=2, useHs=False, tgtDensity=0.3)
return RDKFingerprint(m,
minPath=1,
maxPath=7,
fpSize=fpsize,
nBitsPerHash=bitsperhash,
tgtDensity=tgtDensity,
minSize=fpsize)
def calculate_fp(mol, method='maccs', n_bits=2048): # mol = Chem molecule object # Function to calculate molecular fingerprints given the number of bits and the method if method == 'maccs': return MACCSkeys.GenMACCSKeys(mol) if method == 'ecfp4': return GetMorganFingerprintAsBitVect(mol, 2, nBits=n_bits, useFeatures=False) if method == 'ecfp6': return GetMorganFingerprintAsBitVect(mol, 3, nBits=n_bits, useFeatures=False) if method == 'torsion': return GetHashedTopologicalTorsionFingerprintAsBitVect(mol, nBits=n_bits) if method == 'rdk5': return RDKFingerprint(mol, maxPath=5, fpSize=1024, nBitsPerHash=2)
def transform(self):
super().transform()
fts = []
self.mol_names = []
for mol in self.structures:
fp = RDKFingerprint(mol)
arr = np.zeros((0, ), dtype=np.int8)
DataStructs.ConvertToNumpyArray(fp, arr)
fts.append(arr)
self.features = np.array(fts)
self.mol_names.append(mol.GetProp("_Name"))
self.columns = [str(i) for i in list(range(self.features.shape[1]))]
return self.features
def rdk_molstring(molecule, fptype):
"""
Method for make molstring for rdk fingerprint
:param molecule: molecule object
:param fptype: type, radius and size of fingerprint
:type fptype: dict
:return: molstring for rdk fingerprint
"""
arr = np.zeros((1, ), dtype=int)
DataStructs.ConvertToNumpyArray(
RDKFingerprint(molecule, fpSize=fptype['Size']), arr)
return arr
def ensemble_test(self, submission, data_list, reversed_token_map,
transform):
"""
ensemble test function
:param submission: submission file
:param data_list: list of test data path
:param reversed_token_map: converts prediction to readable format
:param transform: normalize function
"""
# load .yaml file that contains information about each model
with open('model/prediction_models.yaml') as f:
p_configs = yaml.load(f)
predictors = []
for conf in p_configs.values():
predictors.append(
Predict.remote(conf, self._device, self._gpu_non_block,
self._decode_length, self._model_load_path))
loop = asyncio.get_event_loop()
async def process_async_calculate_similarity(combination_of_smiles,
combination_index):
return {
idx: await self.async_fps(comb[0], comb[1])
for comb, idx in zip(combination_of_smiles, combination_index)
}
def ray_prediction(imgs):
return ray.get([p.decode.remote(imgs) for p in predictors])
conf_len = len(p_configs) # configure length == number of model to use
fault_counter = 0
sequence = None
model_contribution = np.zeros(conf_len)
for i, dat in enumerate(data_list):
imgs = Image.open(self._test_file_path + dat)
imgs = self.png_to_tensor(imgs)
imgs = transform(imgs).pin_memory().cuda()
# predict SMILES sequence form each predictors
preds_raw = ray_prediction(imgs)
preds = []
for p in preds_raw:
# predicted sequence token value
SMILES_predicted_sequence = list(
torch.argmax(p.detach().cpu(), -1).numpy())[0]
# converts prediction to readable format from sequence token value
decoded_sequences = decode_predicted_sequences(
SMILES_predicted_sequence, reversed_token_map)
preds.append(decoded_sequences)
del (preds_raw)
# fault check: whether the prediction satisfies the SMILES format or not
ms = {}
for idx, p in enumerate(preds):
m = MolFromSmiles(p)
if m != None:
ms.update({idx: m})
if len(
ms
) == 0: # there is no decoded sequence that matches to SMILES format
print('decode fail')
fault_counter += 1
sequence = preds[0]
elif len(
ms
) == 1: # there is only one decoded sequence that matches to SMILES format
sequence = preds[list(ms.keys())[0]]
else: # there is more than two decoded sequence that matches to SMILES format
# result ensemble
ms_to_fingerprint = [RDKFingerprint(x) for x in ms.values()]
combination_of_smiles = list(combinations(
ms_to_fingerprint, 2))
ms_to_index = [x for x in ms]
combination_index = list(combinations(ms_to_index, 2))
# calculate similarity score
smiles_dict = loop.run_until_complete(
process_async_calculate_similarity(combination_of_smiles,
combination_index))
# sort the pairs by similarity score
smiles_dict = sorted(smiles_dict.items(),
key=(lambda x: x[1]),
reverse=True)
if smiles_dict[0][
1] == 1.0: # if a similar score is 1 we assume to those predictions are correct.
sequence = preds[smiles_dict[0][0][0]]
else:
score_board = np.zeros(conf_len)
for i, (idx, value) in enumerate(smiles_dict):
score_board[list(idx)] += conf_len - i
pick = int(np.argmax(score_board)
) # choose the index that has the highest score
sequence = preds[pick] # pick the decoded sequence
model_contribution[pick] += 1 # logging witch model used
sequence = preds[np.argmax(score_board)]
print('{} sequence:, {}'.format(i, sequence))
# print('decode_time:', time.time() - start_time)
submission.loc[submission['file_name'] == dat, 'SMILES'] = sequence
del (preds)
loop.close()
print('total fault:', fault_counter)
print('model contribution:', model_contribution)
return submission
def rdk_fp(self):
"""
Receives the csv file which is used to generate rdk fingerprints (2048) and saves as numpy file
Parameter
---------
input smiles : str
Compouds in the form of smiles are used
return : np.array
Features are saved in the form of numpy files
"""
df = pd.read_csv(self.csv_path)
smiles_list = df['Smiles'].tolist()
fingerprints = []
not_found = []
for i in tqdm(range(len(smiles_list))):
try:
mol = Chem.MolFromSmiles(smiles_list[i])
fp = RDKFingerprint(mol, nBitsPerHash=1)
bits_array = (np.fromstring(fp.ToBitString(), 'u1') - ord('0'))
fingerprints.append(bits_array)
except:
fingerprints.append(np.nan)
not_found.append(i)
pass
df.drop(not_found, axis=0, inplace=True)
print('Number of FPs not found: {}'.format(len(not_found)))
df.reset_index(drop=True, inplace=True)
labelencoder = LabelEncoder()
Y = labelencoder.fit_transform(df['Label'].values)
Y = Y.reshape(Y.shape[0], 1)
print('Output shape: {}'.format(Y.shape))
fp_array = (np.asarray((fingerprints), dtype=object))
X = np.delete(fp_array, not_found, axis=0)
X = np.vstack(X).astype(np.float32)
print('Input shape: {}'.format(X.shape))
final_array = np.concatenate((X, Y), axis=1)
# Removing rows, from final_array, where duplicate FPs are present
final_array_slice = final_array[:, 0:(final_array.shape[1] - 1)]
_, unq_row_indices = np.unique(final_array_slice,
return_index=True,
axis=0)
final_array_unique = final_array[unq_row_indices]
print(
'Number of Duplicate FPs: {}'.format(final_array.shape[0] -
final_array_unique.shape[0]))
print('Final Numpy array shape: {}'.format(final_array_unique.shape))
print('Type of final array: {}'.format(type(final_array_unique)))
final_numpy_array = np.asarray((final_array_unique), dtype=np.float32)
return final_numpy_array
def fingerprint(fingerprint_method, keyword_dict, smi):
mol = AllChem.MolFromSmiles(smi)
if fingerprint_method == "Morgan":
return AllChem.GetMorganFingerprintAsBitVect(mol, **keyword_dict)
else:
return RDKFingerprint(mol)
def main_bo(vocab_path,
model_path,
save_dir,
descriptor_path,
sampling=60,
iterations=2,
epochs=2,
hidden_size=450,
latent_size=56,
depthT=20,
depthG=3,
random_seed=1,
pIC50_weight=0,
QED_weight=0,
logP_weight=1,
SA_weight=1,
cycle_weight=1,
sim_weight=0):
if os.path.isdir(save_dir) is False:
os.makedirs(save_dir)
vocab = [x.strip("\r\n ") for x in open(vocab_path)]
vocab = Vocab(vocab)
model = JTNNVAE(vocab, hidden_size, latent_size, depthT, depthG)
model.load_state_dict(torch.load(model_path))
model = model.cuda()
if sim_weight != 0:
df_100 = pd.read_csv(
'../data/covid/MPro_6wqf_A_ProteaseData_smiles_top100.csv')
ms_db = [MolFromSmiles(x) for x in df_100['SMILES'].tolist()]
fps_db = [RDKFingerprint(x) for x in ms_db]
# We load the random seed
np.random.seed(random_seed)
# Path of the files
latent_feature = os.path.join(descriptor_path, './latent_features.txt')
target = os.path.join(descriptor_path, './targets.txt')
logp_value = os.path.join(descriptor_path, './logP_values.txt')
QED_value = os.path.join(descriptor_path, './QED_values.txt')
pIC50_value = os.path.join(descriptor_path, './pIC50_values.txt')
sa_score = os.path.join(descriptor_path, './SA_scores.txt')
cycle_score = os.path.join(descriptor_path, './cycle_scores.txt')
sim_score = os.path.join(descriptor_path, './sim_values.txt')
# We load the data (y is minued!)
X = np.loadtxt(latent_feature)
y = -np.loadtxt(target)
y = y.reshape((-1, 1))
n = X.shape[0]
permutation = np.random.choice(n, n, replace=False)
X_train = X[permutation, :][0:np.int(np.round(0.9 * n)), :]
X_test = X[permutation, :][np.int(np.round(0.9 * n)):, :]
y_train = y[permutation][0:np.int(np.round(0.9 * n))]
y_test = y[permutation][np.int(np.round(0.9 * n)):]
np.random.seed(random_seed)
pIC50_values = np.loadtxt(pIC50_value)
QED_values = np.loadtxt(QED_value)
logP_values = np.loadtxt(logp_value)
SA_scores = np.loadtxt(sa_score)
cycle_scores = np.loadtxt(cycle_score)
sim_values = np.loadtxt(sim_score)
iteration = 0
while iteration < iterations:
# We fit the GP
np.random.seed(iteration * random_seed)
M = 500
sgp = SparseGP(X_train, 0 * X_train, y_train, M)
# TODO: test hyperparameters
sgp.train_via_ADAM(X_train,
0 * X_train,
y_train,
X_test,
X_test * 0,
y_test,
minibatch_size=10 * M,
max_iterations=5,
learning_rate=0.001)
pred, uncert = sgp.predict(X_test, 0 * X_test)
error = np.sqrt(np.mean((pred - y_test)**2))
testll = np.mean(sps.norm.logpdf(pred - y_test, scale=np.sqrt(uncert)))
print('Test RMSE: ', error)
print('Test ll: ', testll)
pred, uncert = sgp.predict(X_train, 0 * X_train)
error = np.sqrt(np.mean((pred - y_train)**2))
trainll = np.mean(
sps.norm.logpdf(pred - y_train, scale=np.sqrt(uncert)))
print('Train RMSE: ', error)
print('Train ll: ', trainll)
# We pick the next 60 inputs
next_inputs = sgp.batched_greedy_ei(sampling, np.min(X_train, 0),
np.max(X_train, 0))
# joblib.dump(next_inputs, './next_inputs.pkl')
# next_inputs = joblib.load('./next_inputs.pkl')
valid_smiles = []
new_features = []
for i in tqdm(range(sampling)):
all_vec = next_inputs[i].reshape((1, -1))
tree_vec, mol_vec = np.hsplit(all_vec, 2)
tree_vec = create_var(torch.from_numpy(tree_vec).float())
mol_vec = create_var(torch.from_numpy(mol_vec).float())
tree_vecs, _ = model.rsample(tree_vec, model.T_mean, model.T_var)
mol_vecs, _ = model.rsample(mol_vec, model.G_mean, model.G_var)
s = model.decode(tree_vecs, mol_vecs, prob_decode=False)
if s is not None:
valid_smiles.append(s)
new_features.append(all_vec)
print(len(valid_smiles), "molecules are found")
valid_smiles = valid_smiles
new_features = next_inputs
new_features = np.vstack(new_features)
save_object(
valid_smiles,
os.path.join(save_dir, "valid_smiles{}.pkl".format(iteration)))
scores = []
if pIC50_weight != 0:
current_pIC50 = calculate_pIC50(valid_smiles)
for i in range(len(valid_smiles)):
current_pIC50_normalized = (
current_pIC50[i] -
np.mean(pIC50_values)) / np.std(pIC50_values)
else:
current_pIC50_normalized = []
for i in range(len(valid_smiles)):
current_pIC50_normalized.append(0)
for i in range(len(valid_smiles)):
if sim_weight != 0:
current_sim_value = similarity_search(fps_db, valid_smiles[i])
current_sim_value_normalized = (
current_sim_value -
np.mean(sim_values)) / np.std(sim_values)
else:
current_sim_value_normalized = 0
current_QED_value = QED.qed(MolFromSmiles(valid_smiles[i]))
current_log_P_value = Descriptors.MolLogP(
MolFromSmiles(valid_smiles[i]))
current_SA_score = -sascorer.calculateScore(
MolFromSmiles(valid_smiles[i]))
cycle_list = nx.cycle_basis(
nx.Graph(
rdmolops.GetAdjacencyMatrix(MolFromSmiles(
valid_smiles[i]))))
if len(cycle_list) == 0:
cycle_length = 0
else:
cycle_length = max([len(j) for j in cycle_list])
if cycle_length <= 6:
cycle_length = 0
else:
cycle_length = cycle_length - 6
current_cycle_score = -cycle_length
current_SA_score_normalized = (
current_SA_score - np.mean(SA_scores)) / np.std(SA_scores)
current_QED_value_normalized = (
current_QED_value - np.mean(QED_values)) / np.std(QED_values)
current_log_P_value_normalized = (
current_log_P_value -
np.mean(logP_values)) / np.std(logP_values)
current_cycle_score_normalized = (
current_cycle_score -
np.mean(cycle_scores)) / np.std(cycle_scores)
score = (SA_weight * current_SA_score_normalized +
QED_weight * current_QED_value_normalized +
logP_weight * current_log_P_value_normalized +
cycle_weight * current_cycle_score_normalized +
pIC50_weight * current_pIC50_normalized[i] +
sim_weight * current_sim_value_normalized)
scores.append(-score) # target is always minused
print(valid_smiles)
print(scores)
save_object(scores,
os.path.join(save_dir, "scores{}.pkl".format(iteration)))
if len(new_features) > 0:
X_train = np.concatenate([X_train, new_features], 0)
y_train = np.concatenate([y_train, np.array(scores)[:, None]], 0)
iteration += 1
def rdkit_fingerprint(mol, **kwargs):
return list(RDKFingerprint(mol, **kwargs).GetOnBits())
def scorer(smiles, pIC50_weight, QED_weight, logP_weight, SA_weight, cycle_weight, sim_weight):
smiles_rdkit = []
for i in range(len(smiles)):
smiles_rdkit.append(
MolToSmiles(MolFromSmiles(smiles[i]), isomericSmiles=True))
# calculate IC50 of training set using MPNN
#IC50_scores=calculateScore(smiles_rdkit)
# read in IC50 of training set from database
IC50_scores = np.loadtxt('../data/covid/ic50-fulltrain.txt')
IC50_scores = [x for x in IC50_scores]
IC50_scores_normalized = (np.array(IC50_scores) - np.mean(IC50_scores)) / np.std(IC50_scores)
if sim_weight != 0:
# df_100 = list of molecules to match similarity
df_100 = pd.read_csv('../data/covid/MPro_6wqf_A_ProteaseData_smiles_top100.csv')
ms_db = [MolFromSmiles(x) for x in df_100['SMILES'].tolist()]
fps_db = [RDKFingerprint(x) for x in ms_db]
sim_values = []
for i in range(len(smiles)):
sim_values.append(
similarity_search(fps_db, smiles_rdkit[i]))
sim_values_normalized = (
np.array(sim_values) - np.mean(sim_values)) / np.std(sim_values)
else:
sim_values, sim_values_normalized = [], []
for i in range(len(smiles)):
sim_values.append(0)
sim_values_normalized.append(0)
sim_values_normalized=np.array(sim_values_normalized)
logP_values = []
for i in range(len(smiles)):
logP_values.append(
Descriptors.MolLogP(MolFromSmiles(smiles_rdkit[i])))
qed_values = []
for i in range(len(smiles)):
qed_values.append(
QED.qed(MolFromSmiles(smiles_rdkit[i])))
SA_scores = []
for i in range(len(smiles)):
SA_scores.append(
-sascorer.calculateScore(MolFromSmiles(smiles_rdkit[i])))
cycle_scores = []
for i in range(len(smiles)):
cycle_list = nx.cycle_basis(
nx.Graph(
rdmolops.GetAdjacencyMatrix(MolFromSmiles(smiles_rdkit[i]))))
if len(cycle_list) == 0:
cycle_length = 0
else:
cycle_length = max([len(j) for j in cycle_list])
if cycle_length <= 6:
cycle_length = 0
else:
cycle_length = cycle_length - 6
cycle_scores.append(-cycle_length)
SA_scores_normalized = (
np.array(SA_scores) - np.mean(SA_scores)) / np.std(SA_scores)
qed_values_normalized = (
np.array(qed_values) - np.mean(qed_values)) / np.std(qed_values)
cycle_scores_normalized = (
np.array(cycle_scores) - np.mean(cycle_scores)) / np.std(cycle_scores)
logP_values_normalized = (
np.array(logP_values) - np.mean(logP_values)) / np.std(logP_values)
targets = (pIC50_weight * IC50_scores_normalized +
logP_weight * logP_values_normalized +
SA_weight * SA_scores_normalized +
QED_weight * qed_values_normalized +
cycle_weight * cycle_scores_normalized +
sim_weight * sim_values_normalized)
return (IC50_scores, qed_values, logP_values, SA_scores, cycle_scores, sim_values, targets)