def predict_interaction(smiles,
smiles_b,
model='mlp',
feature='ECFP',
directory=''):
'''Use model to predict interactions
Args :
smiles (str): First SMILES string
smiles_b (str): Second SMILES string
model (str): Name of model used to train
feature (str): Name of feature embedding used
directory (str): Path to directory containing model
Returns :
prediction (numpy.ndarray): Array containing prediction from model
'''
model_path = os.path.join(directory, model + '_' + feature + '.h5')
model = tf.keras.models.load_model(model_path)
vec_a = smiles_to_ECFP(smiles)
vec_b = smiles_to_ECFP(smiles_b)
test = np.concatenate((vec_a, vec_b)).reshape((1, -1))
prediction = model.predict(test)
return prediction
def predict_from_files(candidates_file: object, drugs_file: object,
target_file: object, model_file: object) -> object:
'''Use model to predict interaction between candidates and drugs
Args :
candidates_file (str): Path to txt file with candidate SMILES strings
drugs_file (str): Path to txt file with drug SMILES strings
target_file (str): Path to csv file to write results to
model (object): Pre-trained model to use to predict interactions from
Returns :
None
'''
candidates_list = []
with open(candidates_file) as file:
for line in file:
candidates_list.append(line)
print('Loaded drug candidates.')
drugs_list = []
with open(drugs_file) as file:
for line in file:
drugs_list.append(line)
print('Loaded existing drugs')
print('********************')
print(os.path)
label_lookup = read_dict_from_csv(os.path.join('label_lookup.csv'))
interactions_df = pd.DataFrame(columns = ['Candidate SMILES', 'Drug SMILES', \
'Interaction 1', 'Probability 1', 'Interaction 2', 'Probability 2', \
'Interaction 3', 'Probability 3'])
model = tf.keras.models.load_model(model_file)
print('Predicting interactions ...')
for candidate in tqdm(candidates_list, desc='Candidates : '):
vec_a = smiles_to_ECFP(candidate)
if vec_a is not None:
for drug in tqdm(drugs_list, desc='Drugs : '):
vec_b = smiles_to_ECFP(drug)
if vec_b is not None:
test = np.concatenate((vec_a, vec_b)).reshape((1, -1))
prediction = model.predict(test)
top_labels, top_probs = get_top_n(prediction, 3)
top_labels = list(
map(lambda x: label_lookup[str(x)], top_labels))
interactions_df = interactions_df.append({
'Candidate SMILES':candidate, 'Drug SMILES':drug, \
'Interaction 1':top_labels[0], 'Probability 1':top_probs[0], \
'Interaction 2':top_labels[1], 'Probability 2':top_probs[1], \
'Interaction 3':top_labels[2], 'Probability 3':top_probs[2]}, \
ignore_index = True)
interactions_df.to_csv(target_file, index=False)
def predict():
global SMILES_DICT
smiles = request.args.get('SMILES').strip()
drug_name = request.args.get('drug_name').strip()
if smiles == '':
return render_template('error.html', error='no_smiles')
if drug_name == '':
return render_template('error.html', error='no_drug')
fp = smiles_to_ECFP(smiles)
if fp is None:
return render_template('error.html', error='invalid_smiles')
if drug_name not in SMILES_DICT:
return render_template('error.html', error='drug_not_found')
prediction = predict_interaction(smiles,
SMILES_DICT[drug_name],
directory='flaskapp')
top_labels, top_probs = get_top_n(prediction, 5)
label_lookup = read_dict_from_csv('flaskapp/label_lookup.csv')
top_labels = list(map(lambda x: label_lookup[str(x)], top_labels))
return render_template('result.html', top_labels = top_labels, top_probs = top_probs, \
smiles = smiles, drug_name = drug_name)