def encode_smiles(series):
print("Encoding smiles")
# parameter is a pd.series with ZINC_IDs as the indicies and smiles as the elements
encoded_smiles = DDModel.process_smiles(series.values, 100, fit_range=100, use_padding=True, normalize=True)
encoded_dict = dict(zip(series.keys(), encoded_smiles))
# returns a dict array of the smiles.
return encoded_dict
os.mkdir(file_path + '/iteration_' + str(it) + '/morgan_1024_predictions')
except OSError:
print(file_path + '/iteration_' + str(it) + '/morgan_1024_predictions',
"already exists")
thresholds = pd.read_csv(file_path + '/iteration_' + str(it) +
'/best_models/thresholds.txt',
header=None)
thresholds.columns = ['model_no', 'thresh', 'cutoff']
tr = []
models = []
for f in glob.glob(file_path + '/iteration_' + str(it) +
'/best_models/model_*'):
if "." not in f: # skipping over the .ddss & .csv files
mn = int(f.split('/')[-1].split('_')[1])
tr.append(thresholds[thresholds.model_no == mn].thresh.iloc[0])
models.append(
DDModel.load(file_path + '/iteration_' + str(it) +
'/best_models/model_' + str(mn)))
print("Number of models to predict:", len(models))
t = time.time()
returned = prediction_morgan(fn, models, tr)
print(time.time() - t)
with open(
file_path + '/iteration_' + str(it) +
'/morgan_1024_predictions/passed_file_ct.txt', 'a') as ref:
ref.write(fn + ',' + str(returned) + '\n')
if CONTINUOUS:
print('Using continuous labels...')
y_valid = valid_data.r_i_docking_score
y_test = test_data.r_i_docking_score
y_train = train_data.r_i_docking_score
if NORMALIZE:
print('Adding cutoff to be normalized')
cutoff_ser = pd.Series([cf], index=['cutoff'])
y_train = y_train.append(cutoff_ser)
print("Normalizing docking scores...")
# Normalize the docking scores
y_valid = DDModel.normalize(y_valid)
y_test = DDModel.normalize(y_test)
y_train = DDModel.normalize(y_train)
print('Extracting normalized cutoff...')
cf_norm = y_train['cutoff']
y_train.drop(labels=['cutoff'], inplace=True) # removing it from the dataset
cf_to_use = cf_norm
else:
cf_to_use = cf
# Getting all the ids of hits and non hits.
y_pos = y_train[y_train < cf_to_use]
y_neg = y_train[y_train >= cf_to_use]
all_sc = {}
path_to_model = SAVE_PATH+'/iteration_'+str(n_iteration)+'/all_models/'
print('Model_to_use_with_cf:', model_to_use_with_cf)
for i in range(len(model_to_use_with_cf)):
cf = model_to_use_with_cf[i][0]
# y_train<cf returns a bool array for this condition on each element
y_test_cf = y_test<cf
y_valid_cf = y_valid<cf
models = []
# loading the models matching the cutoff and appending them to the models list
for mn in model_to_use_with_cf[i][-1]:
print('\tLoading model:', path_to_model + '/model_'+str(mn))
models.append(DDModel.load(path_to_model+'/model_'+str(mn)))
print('num models:', len(models))
prediction_valid = []
scc = []
for model in models:
print('using valid as validation')
model_pred = model.predict(X_valid)
if model.output_activation == 'linear':
# Converting back to binary values to get stats
model_pred = model_pred < cf
prediction_valid.append(model_pred)
precision, recall, thresholds = precision_recall_curve(y_valid_cf, model_pred)
scc.append([precision, recall, thresholds])
tr = []