-
Notifications
You must be signed in to change notification settings - Fork 16
/
AMES_aux.py
209 lines (183 loc) · 8.01 KB
/
AMES_aux.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
import pandas as pd
import numpy as np
from rdkit import Chem
from rdkit.Chem import Draw
from itertools import product
import matplotlib.pyplot as plt
from PIL import Image, ImageOps
from keras.models import load_model
import keras.backend as K
from keras.models import Model
from keras.layers import Input, Conv1D, AveragePooling1D
from keras.layers import Dense, Dropout, ZeroPadding1D, Dot, Reshape, Concatenate
from keras import optimizers
from sklearn.metrics import roc_auc_score
def Generator(molsList, one_hot, pad_len=1200, xdtype=np.float32):
# this function transform rdkit molecules in the needed format for the
# graph convolutional neural networks
N = len(molsList)
if isinstance(molsList[0], str):
molsList = [Chem.MolFromSmiles(x) for x in molsList]
x = []
idx_map = []
for i in range(0,N):
try:
smiEnc, per = get_mol_and_pos(molsList[i], one_hot, pad_len=pad_len)
except:
smiEnc = np.zeros((int(pad_len/8),len(one_hot)+4))
per = np.zeros((pad_len,int(pad_len/8)+1))
# print('problematic smiles: '+Chem.MolToSmiles(molsList[i])) if molsList[i] != None else print('None')
x.append(smiEnc)
idx_map.append(per)
return np.asarray(x,dtype=xdtype)/(len(one_hot)+4), np.asarray(idx_map,dtype=np.int8)
def get_mol_and_pos(mol, one_hot, pad_len):
# this function converts each molecule in two arrays
# the first array indicates the atom types of the molecule with an one hot encoding
# plus a bond encoding
# the second array indicates the atom neighborhood
atoms = mol.GetAtoms()
mol_grid = np.zeros((int(pad_len/8),len(one_hot)+4))
pos_grid = np.zeros((pad_len,int(pad_len/8)+1))
k = 1
i = 0
for a in range(min(len(atoms), int(pad_len/8))):
atom = atoms[a]
c_sym = atom.GetSymbol()
center = get_one_hot(atom, atom, one_hot)
mol_grid[a]=center
center_idx = atom.GetIdx()+1
for neighbor in atom.GetNeighbors()[:4]: # at max 4 neighbors are considered
idx = neighbor.GetIdx()+1
n_sym = atom.GetSymbol()
pos_grid[i, idx] = 1
pos_grid[i+1, center_idx] = 1
i+=2
i = (a+1)*8*k
return(mol_grid, pos_grid)
def get_one_hot(a, bond, one_hot):
# this function transforms a rdkit-atom into a one hot vector plus a bond encoding
sym = a.GetSymbol()
arom = a.GetIsAromatic()
if arom:
sym = sym.lower()
else:
add = str(a.GetNumImplicitHs()) #add number of implicit Hs for the one hot encoding
sym = "".join((sym, add))
if sym not in one_hot:
sym = sym[:-1]
if sym not in one_hot:
sym = 'X'
vec = np.zeros(len(one_hot))
bonds = ['SINGLE', 'DOUBLE', 'TRIPLE', 'AROMATIC']
bond_vec = [str(b.GetBondType()) for b in a.GetBonds()]
bond_vec = [int(bond_vec.count(bond)>1) for bond in bonds]
vec[one_hot.index(sym)] = 1
vec = np.concatenate([vec, bond_vec])
return(vec)
def create_model(input_shape, output_shape, filters, dense_layers, opt):
# this function creates a keras network for graph convolutions
inputx = Input(shape=input_shape)
x = inputx
idx_input = Input(shape=(int(input_shape[0]*8), input_shape[0]+1))
nr_atoms = int(input_shape[0])
f=0
p=0
for f in range(len(filters)):
x = ZeroPadding1D(padding=(1,0))(x)
x = Dot(axes=[2,1])([idx_input, x])
x = Conv1D(kernel_size=2, filters=filters[f], strides=2, kernel_initializer='he_normal',
activation='relu', padding='valid', use_bias=False)(x)
x = AveragePooling1D(pool_size=4, padding='valid')(x)
x = AveragePooling1D(pool_size=nr_atoms)(x)
x = Reshape((filters[f],))(x)
for i in range(len(dense_layers)):
x = Dense(dense_layers[i], activation='relu', kernel_initializer='he_normal', bias_initializer='zeros')(x)
x = Dropout(0.5)(x)
outputs = Dense(output_shape, activation='sigmoid',kernel_initializer='lecun_normal',bias_initializer='zeros')(x)
model = Model(inputs=[inputx, idx_input], outputs=outputs)
model.compile(loss=K.binary_crossentropy, optimizer=opt, metrics=['accuracy'])
return model
def run_model(model, inputTrain, yTrain, inputVal, yVal, inputTest, yTest, n=100, batch_size=32, model_file=None):
# this function trains the network and tracks the process
tr = np.zeros((n, yTrain.shape[1]))
val = np.zeros((n, yTrain.shape[1]))
tst = np.zeros((n, yTrain.shape[1]))
best_val = 0
not_improv = 0
for i in range(n):
model.fit(x=inputTrain, y=yTrain, batch_size=batch_size, epochs=1)
P_train = model.predict(inputTrain)
P_val = model.predict(inputVal)
P_tst = model.predict(inputTest)
tr[i] = roc_auc_score(yTrain, P_train)
val[i] = roc_auc_score(yVal, P_val)
tst[i] = roc_auc_score(yTest, P_tst)
if model_file!=None:
if val[i]>best_val:
best_val = val[i]
model.save(model_file+'.h5')
not_improv=0
else:
not_improv+=1
print(i, tr[i], val[i], tst[i])
if not_improv == 50 or np.logical_and(i>50,best_val<0.63):
return model, tr, val, tst
return model, tr, val, tst
def get_activations(model, x, p, layers, filters, pad_len=800, bs = 56):
# this function calculates the hidden representations in a batch mode
atoms = int(pad_len/8)
dim = x.shape[0]
mol_shape = filters[-1]
act_atom = np.zeros((dim*atoms, np.sum(filters)))
act_mol = np.zeros((dim, mol_shape))
act_last0 = K.function([model.input[0], model.input[1]],[model.layers[layers[0]].output])
act_last1 = K.function([model.input[0], model.input[1]], [model.layers[layers[1]].output])
act_last2 = K.function([model.input[0], model.input[1]],[model.layers[layers[2]].output])
mol_act = K.function([model.input[0], model.input[1]],[model.layers[layers[3]].output])
for b in range(int(dim/bs)+int(dim%bs>0)):
train0 = act_last0([x[bs*b:bs*(b+1)], p[bs*b:bs*(b+1)],False])[0].reshape(-1,filters[0])
train1 = act_last1([x[bs*b:bs*(b+1)], p[bs*b:bs*(b+1)],False])[0].reshape(-1,filters[1])
train2 = act_last2([x[bs*b:bs*(b+1)], p[bs*b:bs*(b+1)],False])[0].reshape(-1,filters[2])
mol_layer = mol_act([x[bs*b:bs*(b+1)], p[bs*b:bs*(b+1)],False])[0].reshape(-1,mol_shape)
act_atom[bs*b*atoms:bs*(b+1)*atoms] = np.concatenate([train0, train1, train2], axis=1)
act_mol[bs*b:bs*(b+1)] = mol_layer
return(act_atom, act_mol)
def reduce_act(act, mols, nr_atoms=100):
# this function reduces the atom activations to unique representations
bool_idx = np.zeros(len(mols)*nr_atoms).astype(bool)
atoms_per_mol = []
atoms = []
for i in range(len(mols)):
mol = mols[i]
if mol != None:
nr = np.min([len(mol.GetAtoms()),nr_atoms])
atomsx = mol.GetAtoms()
atoms.extend(np.array(atomsx)[:nr])
else:
nr = 1
nr = np.min([nr_atoms,nr])
atoms_per_mol.append(nr)
bool_idx[i*nr_atoms:(i+1)*nr_atoms] = np.concatenate([np.repeat(True,nr), np.repeat(False, nr_atoms-nr)])
reduced = act[bool_idx]
reduced_df = pd.DataFrame(reduced)
unique_mask = np.logical_not(reduced_df.duplicated())
unique_reduced = reduced_df[unique_mask].values
rep_idx = np.repeat(np.arange(x.shape[0]), atoms_per_mol)
aa = []
idx = []
for i in range(np.sum(bool_idx)):
atom = atoms[i].GetIdx() if atoms[i]!= 'x' else 'x'
if unique_mask[i]:
aa.append(atom)
idx.append(rep_idx[i])
return unique_reduced, idx, aa
def get_substruct(mol, atom_idx, radius=1):
# this function creates submolecules
for r in range(radius)[::-1]:
env = Chem.FindAtomEnvironmentOfRadiusN(mol, r, atom_idx)
amap={}
submol=Chem.PathToSubmol(mol,env,atomMap=amap)
smi = Chem.MolToSmiles(submol)
if smi!="":
break
return submol