CCGNet (Co-Crystal Graph Network) is a deep learning framework for virtual screening of binary organic cocrystals, which integrates the priori knowledge into the feature learning on molecular graph and achieves a great improvement of performance for data-driven cocrystal screening. https://doi.org/10.1038/s41467-021-26226-7
Requirements:
- Python 3.7
- Tensorflow (1.6=<Version<2.0)
- RDKit
- Openbabel 2.4.1
- CCDC Python API
- Scikit-learn
Note: CCDC Python API is a commercial module, which is included in CSD softwares.
The ccgnet module could be used to build CCGNet models. In ccgnet, the message passing is implement by Graph-CNN, where the code of this part we borrow from https://github.com/fps7806/Graph-CNN. The Featurize module represents the combination of two molecules as the input of CCGNet models.
We integrated the ensemble model as pipeline that can provide the high throughput screening for the defined compounds pairs and generate a report form automatically. predict.py can run the ensemble model that were composed by trained 10 models, whose mean balanced accuracy achieve 98.60% in 10-fold cross validation.
usage: predict.py [-h] [-table TABLE] [-mol_dir MOL_DIR] [-out OUT]
[-fmt {sdf,mol,mol2}] [-model_path MODEL_PATH]
optional arguments:
-h, --help show this help message and exit
-table TABLE The table of coformer pairs
-mol_dir MOL_DIR path of molecular file, default is "./"
-out OUT name of predictive file, default is "Result.xlsx"
-fmt {sdf,mol,mol2} format of molecular file，which support sdf, mol, mol2.
default is sdf
-model_path MODEL_PATH
The path where the models are stored, default is
./snapshot/CCGNet_block/CC_Dataset/
First, you need to prepare the structure files of each coformer, whose format can be 'sdf', 'mol', 'mol2'. Then, a table that reprensents the pairs of coformers should be generated. Each item in line is separated by "\t". like this:
The value of the last column is arbitrary, as long as it is an integer.
Here, we use the testing samples in our work as an example
python predict.py -table ./Test/Test_Table.tab -mol_dir ./Test/coformers -out cc_test.xlsx -fmt sdf -model_path ./inference_model/CC_Model/
All pairs were sorted by the predictve score from high to low. The higher score, the higher possibility to form cocrystal.
Finally, the predictive result is written in a .xlsx file, eg. cc_test.xlsx.
In CCGNet, Coformer object is the basic data unit, for each compound:
from Featurize import Coformer
c = Coformer('./Test/coformers/1983.sdf')
A = c.AdjacentTensor.OnlyCovalentBond(with_coo=False)
A.shape
(20, 4, 20)If with_coo=True, it will return the adjacent tensor with COO format, which is easy to feed other GNN framework, such as pytorch-geometric. The method 'OnlyCovalentBond' only consider the four type of covalent bonds, such as single, double, triple and aromatic bonds. If you want to add bond lenth:
A = c.AdjacentTensor.WithBondLenth(with_coo=False)
A.shape
#Out: (20, 5, 20)The extra channel in axis=1 denotes the bond lenth. Each element either 0 or the bond lenth. There are some other methods to compute adjacent tensor, please see Featurize/AdjacentTensor.py .
The atom features and global_state can be computed by:
V = c.VertexMatrix.feature_matrix()
global_state = c.descriptors()
print(V.shape, global_state.shape)
#Out: (20, 34) (12,)Two Coformer objects can be transformed to a Cocrystal object. The features can be calculated by the same way:
from Featurize import Coformer, Cocrystal
c1 = Coformer('./Test/coformers/1983.sdf')
c2 = Coformer('./Test/coformers/1110.sdf')
cc = Cocrystal(c1, c2)
A_cc = cc.AdjacentTensor.OnlyCovalentBond(with_coo=False)
V_cc = cc.VertexMatrix.feature_matrix()
global_state_cc = cc.descriptors()
print(A_cc.shape, V_cc.shape, global_state_cc.shape)
#Out: (34, 4, 34) (34, 34) (24,)Here, we also defined the possible intermolecular interaction. To reprensent potential H-bonds, you can add edge between potential H-bond donnors and accptors
A_hb = cc.CCGraphTensor(t_type='OnlyCovalentBond', hbond=True)
A_hb.shape
#Out: (34, 5, 34) If you want to add possible π-π stack:
A_hb_pp = cc.CCGraphTensor(t_type='OnlyCovalentBond', hbond=True, pipi_stack=True)
A_hb_pp.shape
#Out: (34, 6, 34)
#You can add edge between potential aromatic atoms, but the complexity of cc graph will increase.Furthermore, if you want to add possible weak H-bond interaction, such as C-H···O, C-H···N :
A_hb_pp_c = cc.CCGraphTensor(t_type='OnlyCovalentBond', hbond=True, pipi_stack=True, contact=True)
A_hb_pp_c.shape
#Out: (34, 7, 34)If you want to make your own dataset, ccgnet also provide Dataset class. First, you need prepare two files like 'CC_Table.tab' and 'Mol_Blocks.dir' in './Samples/'. 'CC_Table.tab' reprensents the pairs of coformers. 'Mol_Blocks.dir' is a text file with python dictionary format, whose key is the coformer name in 'CC_Table.tab' and value is string with 3D 'sdf' format.
Note that the input 'sdf' string should preferably be an optimized 3D structure.
mol_block = eval(open('./Samples/Mol_Blocks.dir').read())
print(mol_block['1983'])
'''
1983
-OEChem-03111921513D
20 20 0 0 0 0 0 0 0999 V2000
3.8509 0.4516 0.0012 O 0 0 0 0 0 0 0 0 0 0 0 0
-2.5999 1.4041 -0.0018 O 0 0 0 0 0 0 0 0 0 0 0 0
-1.5705 -0.7171 0.0001 N 0 0 0 0 0 0 0 0 0 0 0 0
-0.2066 -0.4231 -0.0002 C 0 0 0 0 0 0 0 0 0 0 0 0
0.2205 0.9047 0.0004 C 0 0 0 0 0 0 0 0 0 0 0 0
0.7298 -1.4570 -0.0007 C 0 0 0 0 0 0 0 0 0 0 0 0
1.5841 1.1986 0.0002 C 0 0 0 0 0 0 0 0 0 0 0 0
2.0933 -1.1629 -0.0007 C 0 0 0 0 0 0 0 0 0 0 0 0
2.5204 0.1648 -0.0003 C 0 0 0 0 0 0 0 0 0 0 0 0
-2.6485 0.1782 0.0009 C 0 0 0 0 0 0 0 0 0 0 0 0
-3.9735 -0.5420 0.0010 C 0 0 0 0 0 0 0 0 0 0 0 0
-0.4436 1.7577 0.0012 H 0 0 0 0 0 0 0 0 0 0 0 0
0.4113 -2.4963 -0.0010 H 0 0 0 0 0 0 0 0 0 0 0 0
-1.8010 -1.7086 0.0001 H 0 0 0 0 0 0 0 0 0 0 0 0
1.9053 2.2370 0.0009 H 0 0 0 0 0 0 0 0 0 0 0 0
2.8180 -1.9726 -0.0008 H 0 0 0 0 0 0 0 0 0 0 0 0
-4.0655 -1.1463 -0.9058 H 0 0 0 0 0 0 0 0 0 0 0 0
-4.7904 0.1844 0.0288 H 0 0 0 0 0 0 0 0 0 0 0 0
-4.0445 -1.1886 0.8802 H 0 0 0 0 0 0 0 0 0 0 0 0
3.9650 1.4176 0.0017 H 0 0 0 0 0 0 0 0 0 0 0 0
1 9 1 0 0 0 0
1 20 1 0 0 0 0
2 10 2 0 0 0 0
3 4 1 0 0 0 0
3 10 1 0 0 0 0
3 14 1 0 0 0 0
4 5 2 0 0 0 0
4 6 1 0 0 0 0
5 7 1 0 0 0 0
5 12 1 0 0 0 0
6 8 2 0 0 0 0
6 13 1 0 0 0 0
7 9 2 0 0 0 0
7 15 1 0 0 0 0
8 9 1 0 0 0 0
8 16 1 0 0 0 0
10 11 1 0 0 0 0
11 17 1 0 0 0 0
11 18 1 0 0 0 0
11 19 1 0 0 0 0
M END
$$$$
'''We can build the dataset based on these two file, such as:
from ccgnet.Dataset import Dataset, DataLoader
data1 = Dataset('data/CC_Table/ECC&CC_Table.tab', mol_blocks_dir='data/Mol_Blocks.dir')
data1.make_graph_dataset(Desc=1, A_type='OnlyCovalentBond', hbond=0, pipi_stack=0, contact=0, make_dataframe=True)
data2 = Dataset('data/CC_Table/CC_Table-DataAug.tab', mol_blocks_dir='data/Mol_Blocks.dir')
data2.make_graph_dataset(Desc=1, A_type='OnlyCovalentBond', hbond=0, pipi_stack=0, contact=0, make_dataframe=True)If save_name is not None, the dataset will be saved as a 'save_name'.npz file to your disk. If make_dataframe=True, object data will make a new property .dataframe. Each entry holds the data for the corresponding sample, like:
data.dataframe['UNEYOB'].keys()
#Out: dict_keys(['V', 'A', 'label', 'global_state', 'tag', 'mask', 'graph_size', 'subgraph_size'])To train and test the model, we firstly split out the test set.
nico = eval(open('data/Test/Test_Samples/Nicotinamide_Test.list').read())
carb = eval(open('data/Test/Test_Samples/Carbamazepine_Test.list').read())
indo = eval(open('data/Test/Test_Samples/Indomethacin_Test.list').read())
para = eval(open('data/Test/Test_Samples/Paracetamol_Test.list').read())
pyre = eval(open('data/Test/Test_Samples/Pyrene_Test.list').read())
apis = list(set(nico + indo + para + carb))
cl20 = eval(open('data/Test/Test/Test_Samples/CL-20_Test.list').read())
tnt = eval(open('data/Test/Test_Samples/TNT_Test.list').read())
# these samples that are selected as out-of-sample test are the same with our paper.
test = list(set(nico + carb + indo + para + pyre))Here, we perform a 10-fold cross-validation.
from sklearn.model_selection import KFold
cv = np.array([i for i in data.dataframe if i not in test])
np.random.shuffle(cv)
fold_10 = {}
fold_10['test'] = test
kf = KFold(n_splits=10, shuffle=True, random_state=1000)
n = 0
for train_idx, test_idx in kf.split(cv):
fold = 'fold-{}'.format(n)
fold_10[fold] = {}
fold_10[fold]['train'] = cv[train_idx]
fold_10[fold]['valid'] = cv[test_idx]
n += 1The model construction.
import tensorflow as tf
from ccgnet import experiment as exp
from ccgnet.finetune import *
from ccgnet import layers
from ccgnet.layers import *
import numpy as np
import time
from sklearn.metrics import balanced_accuracy_score
from ccgnet.Dataset import Dataset, DataLoader
from Featurize.Coformer import Coformer
from Featurize.Cocrystal import Cocrystal
class CCGNet(object):
def build_model(self, inputs, is_training, global_step):
V = inputs[0]
A = inputs[1]
labels = inputs[2]
mask = inputs[3]
graph_size = inputs[4]
tags = inputs[5]
global_state = inputs[6]
subgraph_size = inputs[7]
# message passing
V, global_state = CCGBlock(V, A, global_state, subgraph_size, no_filters=64, mask=mask, num_updates=global_step, is_training=is_training)
V, global_state = CCGBlock(V, A, global_state, subgraph_size, no_filters=16, mask=mask, num_updates=global_step, is_training=is_training)
V, global_state = CCGBlock(V, A, global_state, subgraph_size, no_filters=64, mask=mask, num_updates=global_step, is_training=is_training)
V, global_state = CCGBlock(V, A, global_state, subgraph_size, no_filters=16, mask=mask, num_updates=global_step, is_training=is_training)
# readout
V = ReadoutFunction(V, global_state, graph_size, num_head=2, is_training=is_training)
# predict
with tf.compat.v1.variable_scope('Predictive_FC_1') as scope:
V = layers.make_embedding_layer(V, 256)
V = layers.make_bn(V, is_training, mask=None, num_updates=global_step)
V = tf.nn.relu(V)
V = tf.compat.v1.layers.dropout(V, 0.457, training=is_training)
with tf.compat.v1.variable_scope('Predictive_FC_2') as scope:
V = layers.make_embedding_layer(V, 1024)
V = layers.make_bn(V, is_training, mask=None, num_updates=global_step)
V = tf.nn.relu(V)
V = tf.compat.v1.layers.dropout(V, 0.457, training=is_training)
with tf.compat.v1.variable_scope('Predictive_FC_3') as scope:
V = layers.make_embedding_layer(V, 256)
V = layers.make_bn(V, is_training, mask=None, num_updates=global_step)
V = tf.nn.relu(V)
V = tf.compat.v1.layers.dropout(V, 0.457, training=is_training)
out = layers.make_embedding_layer(V, 2, name='final')
return out, labelsThen, we will fit model. In this case, model and log will be saved at './snapshot/CCGNet_block/CC_Dataset/time_*'.
We call data.split to split train_data, valid_data and test_data. The test_data is optional. In model training, ccgnet will save the models that hold best 5 performance (default). You can increase or decrease the number of saving models by changing the value of 'max_to_keep'. When higher performance is achieved in valid set, ccgnet will save the weights and make inference for test set (if with_test=True). Also, you can use the saved model to perform inference on the test set by using predict.py.
start = time.time()
snapshot_path = './snapshot/'
model_name = 'CCGNet'
dataset_name = 'CC_Dataset'
for fold in ['fold-{}'.format(i) for i in range(10)]:
print('\n################ {} ################'.format(fold))
train_data1, valid_data1, test_data1 = data1.split(train_samples=fold_10[fold]['train'], valid_samples=fold_10[fold]['valid'], with_test=True, test_samples=fold_10['test'])
train_data2, valid_data2 = data2.split(train_samples=fold_10[fold]['train'], valid_samples=fold_10[fold]['valid'], with_test=False)
train_data = []
for ix, i in enumerate(train_data1):
train_data.append(np.concatenate([i, train_data2[ix]]))
del train_data2
del train_data1
tf.reset_default_graph()
model = CCGNet()
model = exp.Model(model, train_data, valid_data1, with_test=True, test_data=test_data1, snapshot_path=snapshot_path, use_subgraph=True,
model_name=model_name, dataset_name=dataset_name+'/time_{}'.format(fold[-1]))
history = model.fit(num_epoch=100, save_info=True, save_att=True, silence=0, train_batch_size=128,
metric='acc')
end = time.time()
time_gap = end-start
h = time_gap//3600
h_m = time_gap%3600
m = h_m//60
s = h_m%60
print('{}h {}m {}s'.format(int(h),int(m),round(s,2)))You can use the Featurize.MetricsReport module to assess model performance in 10-fold CV.
from Featurize.MetricsReport import model_metrics_report
model_metrics_report('{}/{}/'.format(snapshot_path, model_name))