Spaces:
Running
Running
Upload 7 files
Browse files- inference.py +61 -37
- layers.py +2 -2
- loss.py +39 -15
- models.py +56 -6
- smiles_cor.py +1291 -0
- utils.py +18 -7
inference.py
CHANGED
@@ -4,19 +4,20 @@ import pickle
|
|
4 |
import random
|
5 |
from tqdm import tqdm
|
6 |
import argparse
|
7 |
-
|
8 |
import torch
|
9 |
from torch_geometric.loader import DataLoader
|
10 |
import torch.utils.data
|
11 |
from rdkit import RDLogger
|
12 |
torch.set_num_threads(5)
|
13 |
RDLogger.DisableLog('rdApp.*')
|
14 |
-
|
15 |
from utils import *
|
16 |
from models import Generator
|
17 |
from new_dataloader import DruggenDataset
|
18 |
from loss import generator_loss
|
19 |
from training_data import load_molecules
|
|
|
20 |
|
21 |
|
22 |
class Inference(object):
|
@@ -43,6 +44,7 @@ class Inference(object):
|
|
43 |
|
44 |
self.inference_model = config.inference_model
|
45 |
self.sample_num = config.sample_num
|
|
|
46 |
|
47 |
# Data loader.
|
48 |
self.inf_raw_file = config.inf_raw_file # SMILES containing text file for first dataset.
|
@@ -103,8 +105,7 @@ class Inference(object):
|
|
103 |
dim=self.dim,
|
104 |
depth=self.depth,
|
105 |
heads=self.heads,
|
106 |
-
mlp_ratio=self.mlp_ratio
|
107 |
-
submodel = self.submodel)
|
108 |
|
109 |
self.print_network(self.G, 'G')
|
110 |
|
@@ -113,7 +114,7 @@ class Inference(object):
|
|
113 |
|
114 |
def decoder_load(self, dictionary_name):
|
115 |
''' Loading the atom and bond decoders'''
|
116 |
-
with open("data/decoders/" + dictionary_name + "_" + self.dataset_name + '.pkl', 'rb') as f:
|
117 |
return pickle.load(f)
|
118 |
|
119 |
|
@@ -139,18 +140,25 @@ class Inference(object):
|
|
139 |
self.restore_model(self.submodel, self.inference_model)
|
140 |
|
141 |
# smiles data for metrics calculation.
|
142 |
-
chembl_smiles = [line for line in open("data/chembl_train.smi", 'r').read().splitlines()]
|
143 |
-
chembl_test = [line for line in open("data/chembl_test.smi", 'r').read().splitlines()]
|
144 |
-
drug_smiles = [line for line in open("data/akt_inhibitors.smi", 'r').read().splitlines()]
|
145 |
drug_mols = [Chem.MolFromSmiles(smi) for smi in drug_smiles]
|
146 |
drug_vecs = [AllChem.GetMorganFingerprintAsBitVect(x, 2, nBits=1024) for x in drug_mols if x is not None]
|
147 |
|
148 |
|
149 |
# Make directories if not exist.
|
150 |
-
if not os.path.exists("experiments/inference/{}".format(self.submodel)):
|
151 |
-
os.makedirs("experiments/inference/{}".format(self.submodel))
|
152 |
-
|
153 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
154 |
self.G.eval()
|
155 |
|
156 |
start_time = time.time()
|
@@ -158,7 +166,9 @@ class Inference(object):
|
|
158 |
uniqueness_calc = []
|
159 |
real_smiles_snn = []
|
160 |
nodes_sample = torch.Tensor(size=[1,45,1]).to(self.device)
|
161 |
-
|
|
|
|
|
162 |
val_counter = 0
|
163 |
none_counter = 0
|
164 |
# Inference mode
|
@@ -182,7 +192,7 @@ class Inference(object):
|
|
182 |
g_edges_hat_sample = torch.max(edge_sample, -1)[1]
|
183 |
g_nodes_hat_sample = torch.max(node_sample, -1)[1]
|
184 |
|
185 |
-
fake_mol_g = [self.inf_dataset.matrices2mol_drugs(n_.data.cpu().numpy(), e_.data.cpu().numpy(), strict=
|
186 |
for e_, n_ in zip(g_edges_hat_sample, g_nodes_hat_sample)]
|
187 |
|
188 |
a_tensor_sample = torch.max(a_tensor, -1)[1]
|
@@ -197,34 +207,47 @@ class Inference(object):
|
|
197 |
if molecules is None:
|
198 |
none_counter += 1
|
199 |
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
metric_calc_dr.append(molecules)
|
210 |
-
|
211 |
|
|
|
212 |
generation_number = len([x for x in metric_calc_dr if x is not None])
|
213 |
if generation_number == self.sample_num or none_counter == self.sample_num:
|
214 |
break
|
215 |
-
|
216 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
217 |
et = time.time() - start_time
|
218 |
-
gen_vecs = [AllChem.GetMorganFingerprintAsBitVect(Chem.MolFromSmiles(x), 2, nBits=1024) for x in uniqueness_calc if Chem.MolFromSmiles(x) is not None]
|
219 |
-
real_vecs = [AllChem.GetMorganFingerprintAsBitVect(x, 2, nBits=1024) for x in real_smiles_snn if x is not None]
|
220 |
-
print("Inference mode is lasted for {:.2f} seconds".format(et))
|
221 |
|
222 |
-
|
223 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
224 |
|
225 |
return{
|
226 |
"Runtime (seconds)": f"{et:.2f}",
|
227 |
-
"Validity":
|
228 |
"Uniqueness": f"{fraction_unique(uniqueness_calc):.2f}",
|
229 |
"Novelty (Train)": f"{novelty(metric_calc_dr, chembl_smiles):.2f}",
|
230 |
"Novelty (Inference)": f"{novelty(metric_calc_dr, chembl_test):.2f}",
|
@@ -237,13 +260,14 @@ if __name__=="__main__":
|
|
237 |
# Inference configuration.
|
238 |
parser.add_argument('--submodel', type=str, default="DrugGEN", help="Chose model subtype: DrugGEN, NoTarget", choices=['DrugGEN', 'NoTarget'])
|
239 |
parser.add_argument('--inference_model', type=str, help="Path to the model for inference")
|
240 |
-
parser.add_argument('--sample_num', type=int, default=
|
241 |
-
|
|
|
242 |
# Data configuration.
|
243 |
parser.add_argument('--inf_dataset_file', type=str, default='chembl45_test.pt')
|
244 |
-
parser.add_argument('--inf_raw_file', type=str, default='data/chembl_test.smi')
|
245 |
parser.add_argument('--inf_batch_size', type=int, default=1, help='Batch size for inference')
|
246 |
-
parser.add_argument('--mol_data_dir', type=str, default='data')
|
247 |
parser.add_argument('--features', type=str2bool, default=False, help='features dimension for nodes')
|
248 |
|
249 |
# Model configuration.
|
|
|
4 |
import random
|
5 |
from tqdm import tqdm
|
6 |
import argparse
|
7 |
+
import pandas as pd
|
8 |
import torch
|
9 |
from torch_geometric.loader import DataLoader
|
10 |
import torch.utils.data
|
11 |
from rdkit import RDLogger
|
12 |
torch.set_num_threads(5)
|
13 |
RDLogger.DisableLog('rdApp.*')
|
14 |
+
from rdkit.Chem import QED
|
15 |
from utils import *
|
16 |
from models import Generator
|
17 |
from new_dataloader import DruggenDataset
|
18 |
from loss import generator_loss
|
19 |
from training_data import load_molecules
|
20 |
+
from smiles_cor import smi_correct
|
21 |
|
22 |
|
23 |
class Inference(object):
|
|
|
44 |
|
45 |
self.inference_model = config.inference_model
|
46 |
self.sample_num = config.sample_num
|
47 |
+
self.correct = config.correct
|
48 |
|
49 |
# Data loader.
|
50 |
self.inf_raw_file = config.inf_raw_file # SMILES containing text file for first dataset.
|
|
|
105 |
dim=self.dim,
|
106 |
depth=self.depth,
|
107 |
heads=self.heads,
|
108 |
+
mlp_ratio=self.mlp_ratio)
|
|
|
109 |
|
110 |
self.print_network(self.G, 'G')
|
111 |
|
|
|
114 |
|
115 |
def decoder_load(self, dictionary_name):
|
116 |
''' Loading the atom and bond decoders'''
|
117 |
+
with open("DrugGEN/data/decoders/" + dictionary_name + "_" + self.dataset_name + '.pkl', 'rb') as f:
|
118 |
return pickle.load(f)
|
119 |
|
120 |
|
|
|
140 |
self.restore_model(self.submodel, self.inference_model)
|
141 |
|
142 |
# smiles data for metrics calculation.
|
143 |
+
chembl_smiles = [line for line in open("DrugGEN/data/chembl_train.smi", 'r').read().splitlines()]
|
144 |
+
chembl_test = [line for line in open("DrugGEN/data/chembl_test.smi", 'r').read().splitlines()]
|
145 |
+
drug_smiles = [line for line in open("DrugGEN/data/akt_inhibitors.smi", 'r').read().splitlines()]
|
146 |
drug_mols = [Chem.MolFromSmiles(smi) for smi in drug_smiles]
|
147 |
drug_vecs = [AllChem.GetMorganFingerprintAsBitVect(x, 2, nBits=1024) for x in drug_mols if x is not None]
|
148 |
|
149 |
|
150 |
# Make directories if not exist.
|
151 |
+
if not os.path.exists("DrugGEN/experiments/inference/{}".format(self.submodel)):
|
152 |
+
os.makedirs("DrugGEN/experiments/inference/{}".format(self.submodel))
|
153 |
+
if self.correct:
|
154 |
+
correct = smi_correct(self.submodel, "DrugGEN_/experiments/inference/{}".format(self.submodel))
|
155 |
+
search_res = pd.DataFrame(columns=["submodel", "validity",
|
156 |
+
"uniqueness", "novelty",
|
157 |
+
"novelty_test", "AKT_novelty",
|
158 |
+
"max_len", "mean_atom_type",
|
159 |
+
"snn_chembl", "snn_akt", "IntDiv", "qed"])
|
160 |
+
|
161 |
+
|
162 |
self.G.eval()
|
163 |
|
164 |
start_time = time.time()
|
|
|
166 |
uniqueness_calc = []
|
167 |
real_smiles_snn = []
|
168 |
nodes_sample = torch.Tensor(size=[1,45,1]).to(self.device)
|
169 |
+
f = open("DrugGEN/experiments/inference/{}/inference_drugs.txt".format(self.submodel), "w")
|
170 |
+
f.write("SMILES")
|
171 |
+
f.write("\n")
|
172 |
val_counter = 0
|
173 |
none_counter = 0
|
174 |
# Inference mode
|
|
|
192 |
g_edges_hat_sample = torch.max(edge_sample, -1)[1]
|
193 |
g_nodes_hat_sample = torch.max(node_sample, -1)[1]
|
194 |
|
195 |
+
fake_mol_g = [self.inf_dataset.matrices2mol_drugs(n_.data.cpu().numpy(), e_.data.cpu().numpy(), strict=False, file_name=self.dataset_name)
|
196 |
for e_, n_ in zip(g_edges_hat_sample, g_nodes_hat_sample)]
|
197 |
|
198 |
a_tensor_sample = torch.max(a_tensor, -1)[1]
|
|
|
207 |
if molecules is None:
|
208 |
none_counter += 1
|
209 |
|
210 |
+
for molecules in inference_drugs:
|
211 |
+
if molecules is not None:
|
212 |
+
molecules = molecules.replace("*", "C")
|
213 |
+
f.write(molecules)
|
214 |
+
f.write("\n")
|
215 |
+
uniqueness_calc.append(molecules)
|
216 |
+
nodes_sample = torch.cat((nodes_sample, g_nodes_hat_sample.view(1,45,1)), 0)
|
217 |
+
pbar.update(1)
|
218 |
+
metric_calc_dr.append(molecules)
|
|
|
|
|
219 |
|
220 |
+
real_smiles_snn.append(real_mols[0])
|
221 |
generation_number = len([x for x in metric_calc_dr if x is not None])
|
222 |
if generation_number == self.sample_num or none_counter == self.sample_num:
|
223 |
break
|
224 |
+
|
225 |
+
|
226 |
+
f.close()
|
227 |
+
print("Inference completed, starting metrics calculation.")
|
228 |
+
if self.correct:
|
229 |
+
corrected = correct.correct("DrugGEN/experiments/inference/{}/inference_drugs.txt".format(self.submodel))
|
230 |
+
gen_smi = corrected["SMILES"].tolist()
|
231 |
+
|
232 |
+
else:
|
233 |
+
gen_smi = pd.read_csv("DrugGEN/experiments/inference/{}/inference_drugs.txt".format(self.submodel))["SMILES"].tolist()
|
234 |
+
|
235 |
+
|
236 |
et = time.time() - start_time
|
|
|
|
|
|
|
237 |
|
238 |
+
with open("DrugGEN/experiments/inference/{}/inference_drugs.txt".format(self.submodel), "w") as f:
|
239 |
+
for i in gen_smi:
|
240 |
+
f.write(i)
|
241 |
+
f.write("\n")
|
242 |
+
|
243 |
+
if self.correct:
|
244 |
+
val = round(len(gen_smi)/self.sample_num,3)
|
245 |
+
else:
|
246 |
+
val = round(fraction_valid(gen_smi),3)
|
247 |
|
248 |
return{
|
249 |
"Runtime (seconds)": f"{et:.2f}",
|
250 |
+
"Validity": str(val),
|
251 |
"Uniqueness": f"{fraction_unique(uniqueness_calc):.2f}",
|
252 |
"Novelty (Train)": f"{novelty(metric_calc_dr, chembl_smiles):.2f}",
|
253 |
"Novelty (Inference)": f"{novelty(metric_calc_dr, chembl_test):.2f}",
|
|
|
260 |
# Inference configuration.
|
261 |
parser.add_argument('--submodel', type=str, default="DrugGEN", help="Chose model subtype: DrugGEN, NoTarget", choices=['DrugGEN', 'NoTarget'])
|
262 |
parser.add_argument('--inference_model', type=str, help="Path to the model for inference")
|
263 |
+
parser.add_argument('--sample_num', type=int, default=100, help='inference samples')
|
264 |
+
parser.add_argument('--correct', type=str2bool, default=False, help='Correct smiles')
|
265 |
+
|
266 |
# Data configuration.
|
267 |
parser.add_argument('--inf_dataset_file', type=str, default='chembl45_test.pt')
|
268 |
+
parser.add_argument('--inf_raw_file', type=str, default='DrugGEN/data/chembl_test.smi')
|
269 |
parser.add_argument('--inf_batch_size', type=int, default=1, help='Batch size for inference')
|
270 |
+
parser.add_argument('--mol_data_dir', type=str, default='DrugGEN/data')
|
271 |
parser.add_argument('--features', type=str2bool, default=False, help='features dimension for nodes')
|
272 |
|
273 |
# Model configuration.
|
layers.py
CHANGED
@@ -82,7 +82,7 @@ class Encoder_Block(nn.Module):
|
|
82 |
|
83 |
def forward(self, x, y):
|
84 |
x1 = self.ln1(x)
|
85 |
-
x2,y1 = self.attn(x1,y)
|
86 |
x2 = x1 + x2
|
87 |
y2 = y1 + y
|
88 |
x2 = self.ln3(x2)
|
@@ -102,5 +102,5 @@ class TransformerEncoder(nn.Module):
|
|
102 |
|
103 |
def forward(self, x, y):
|
104 |
for Encoder_Block in self.Encoder_Blocks:
|
105 |
-
x, y = Encoder_Block(x,y)
|
106 |
return x, y
|
|
|
82 |
|
83 |
def forward(self, x, y):
|
84 |
x1 = self.ln1(x)
|
85 |
+
x2, y1 = self.attn(x1, y)
|
86 |
x2 = x1 + x2
|
87 |
y2 = y1 + y
|
88 |
x2 = self.ln3(x2)
|
|
|
102 |
|
103 |
def forward(self, x, y):
|
104 |
for Encoder_Block in self.Encoder_Blocks:
|
105 |
+
x, y = Encoder_Block(x, y)
|
106 |
return x, y
|
loss.py
CHANGED
@@ -1,36 +1,60 @@
|
|
1 |
import torch
|
2 |
|
3 |
|
4 |
-
def discriminator_loss(generator, discriminator,
|
5 |
# Compute loss with real molecules.
|
6 |
-
|
|
|
|
|
|
|
|
|
7 |
prediction_real = - torch.mean(logits_real_disc)
|
8 |
|
9 |
# Compute loss with fake molecules.
|
10 |
node, edge, node_sample, edge_sample = generator(z_edge, z_node)
|
11 |
-
|
12 |
-
|
|
|
|
|
|
|
|
|
13 |
prediction_fake = torch.mean(logits_fake_disc)
|
14 |
|
15 |
-
# Compute gradient
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
|
21 |
-
# Calculate total loss
|
22 |
-
d_loss = prediction_fake + prediction_real + d_loss_gp * lambda_gp
|
23 |
return node, edge, d_loss
|
24 |
|
25 |
|
26 |
-
def generator_loss(generator, discriminator, adj, annot, batch_size):
|
27 |
# Compute loss with fake molecules.
|
28 |
node, edge, node_sample, edge_sample = generator(adj, annot)
|
|
|
|
|
|
|
|
|
|
|
29 |
|
30 |
-
graph = torch.cat((node_sample.view(batch_size, -1), edge_sample.view(batch_size, -1)), dim=-1)
|
31 |
-
|
32 |
-
logits_fake_disc = discriminator(graph)
|
33 |
prediction_fake = - torch.mean(logits_fake_disc)
|
|
|
34 |
g_loss = prediction_fake
|
35 |
|
36 |
return g_loss, node, edge, node_sample, edge_sample
|
|
|
1 |
import torch
|
2 |
|
3 |
|
4 |
+
def discriminator_loss(generator, discriminator, drug_edge, drug_node, batch_size, device, grad_pen, lambda_gp, z_edge, z_node, submodel):
|
5 |
# Compute loss with real molecules.
|
6 |
+
if submodel == "DrugGEN":
|
7 |
+
logits_real_disc = discriminator(drug_edge, drug_node)
|
8 |
+
else:
|
9 |
+
logits_real_disc = discriminator(drug_node)
|
10 |
+
|
11 |
prediction_real = - torch.mean(logits_real_disc)
|
12 |
|
13 |
# Compute loss with fake molecules.
|
14 |
node, edge, node_sample, edge_sample = generator(z_edge, z_node)
|
15 |
+
if submodel == "DrugGEN":
|
16 |
+
logits_fake_disc = discriminator(edge_sample, node_sample)
|
17 |
+
else:
|
18 |
+
graph = torch.cat((node_sample.view(batch_size, -1), edge_sample.view(batch_size, -1)), dim=-1)
|
19 |
+
logits_fake_disc = discriminator(graph.detach())
|
20 |
+
|
21 |
prediction_fake = torch.mean(logits_fake_disc)
|
22 |
|
23 |
+
# Compute gradient penalty.
|
24 |
+
eps_edge = torch.rand(batch_size, 1, 1, 1, device=device) # Shape adapted for broadcasting with edges and nodes
|
25 |
+
eps_node = torch.rand(batch_size, 1, 1, device=device) # Shape adapted for broadcasting with edges and nodes
|
26 |
+
int_node = eps_node * drug_node + (1 - eps_node) * node_sample
|
27 |
+
int_edge = eps_edge * drug_edge + (1 - eps_edge) * edge_sample
|
28 |
+
int_node.requires_grad_(True)
|
29 |
+
int_edge.requires_grad_(True)
|
30 |
+
|
31 |
+
# Compute discriminator output for interpolated samples
|
32 |
+
if submodel == "DrugGEN":
|
33 |
+
logits_interpolated = discriminator(int_edge, int_node)
|
34 |
+
else:
|
35 |
+
graph = torch.cat((int_node.view(batch_size, -1), int_edge.view(batch_size, -1)), dim=-1)
|
36 |
+
logits_interpolated = discriminator(graph)
|
37 |
+
|
38 |
+
# Compute gradient penalty for nodes and edges
|
39 |
+
grad_penalty = grad_pen(logits_interpolated, int_node)
|
40 |
+
|
41 |
+
# Calculate total discriminator loss
|
42 |
+
d_loss = prediction_fake + prediction_real + lambda_gp * grad_penalty
|
43 |
|
|
|
|
|
44 |
return node, edge, d_loss
|
45 |
|
46 |
|
47 |
+
def generator_loss(generator, discriminator, adj, annot, batch_size, submodel):
|
48 |
# Compute loss with fake molecules.
|
49 |
node, edge, node_sample, edge_sample = generator(adj, annot)
|
50 |
+
if submodel == "DrugGEN":
|
51 |
+
logits_fake_disc = discriminator(edge_sample, node_sample)
|
52 |
+
else:
|
53 |
+
graph = torch.cat((node_sample.view(batch_size, -1), edge_sample.view(batch_size, -1)), dim=-1)
|
54 |
+
logits_fake_disc = discriminator(graph)
|
55 |
|
|
|
|
|
|
|
56 |
prediction_fake = - torch.mean(logits_fake_disc)
|
57 |
+
|
58 |
g_loss = prediction_fake
|
59 |
|
60 |
return g_loss, node, edge, node_sample, edge_sample
|
models.py
CHANGED
@@ -5,9 +5,8 @@ from layers import TransformerEncoder
|
|
5 |
class Generator(nn.Module):
|
6 |
"""Generator network."""
|
7 |
|
8 |
-
def __init__(self, act, vertexes, edges, nodes, dropout, dim, depth, heads, mlp_ratio
|
9 |
super(Generator, self).__init__()
|
10 |
-
self.submodel = submodel
|
11 |
self.vertexes = vertexes
|
12 |
self.edges = edges
|
13 |
self.nodes = nodes
|
@@ -30,8 +29,8 @@ class Generator(nn.Module):
|
|
30 |
self.transformer_dim = vertexes * vertexes * dim + vertexes * dim
|
31 |
self.pos_enc_dim = 5
|
32 |
|
33 |
-
self.node_layers = nn.Sequential(nn.Linear(nodes, 64), act, nn.Linear(64,dim), act, nn.Dropout(self.dropout))
|
34 |
-
self.edge_layers = nn.Sequential(nn.Linear(edges, 64), act, nn.Linear(64,dim), act, nn.Dropout(self.dropout))
|
35 |
self.TransformerEncoder = TransformerEncoder(dim=self.dim, depth=self.depth, heads=self.heads, act = act,
|
36 |
mlp_ratio=self.mlp_ratio, drop_rate=self.dropout)
|
37 |
|
@@ -63,12 +62,61 @@ class Generator(nn.Module):
|
|
63 |
edge = self.edge_layers(z_e)
|
64 |
edge = (edge + edge.permute(0, 2, 1, 3)) / 2
|
65 |
|
66 |
-
node, edge = self.TransformerEncoder(node,edge)
|
67 |
|
68 |
node_sample = self.readout_n(node)
|
69 |
edge_sample = self.readout_e(edge)
|
|
|
70 |
return node, edge, node_sample, edge_sample
|
71 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
|
73 |
class simple_disc(nn.Module):
|
74 |
def __init__(self, act, m_dim, vertexes, b_dim):
|
@@ -82,6 +130,8 @@ class simple_disc(nn.Module):
|
|
82 |
act = nn.Sigmoid()
|
83 |
elif act == "tanh":
|
84 |
act = nn.Tanh()
|
|
|
|
|
85 |
|
86 |
features = vertexes * m_dim + vertexes * vertexes * b_dim
|
87 |
self.predictor = nn.Sequential(nn.Linear(features,256), act, nn.Linear(256,128), act, nn.Linear(128,64), act,
|
@@ -90,4 +140,4 @@ class simple_disc(nn.Module):
|
|
90 |
|
91 |
def forward(self, x):
|
92 |
prediction = self.predictor(x)
|
93 |
-
return prediction
|
|
|
5 |
class Generator(nn.Module):
|
6 |
"""Generator network."""
|
7 |
|
8 |
+
def __init__(self, act, vertexes, edges, nodes, dropout, dim, depth, heads, mlp_ratio):
|
9 |
super(Generator, self).__init__()
|
|
|
10 |
self.vertexes = vertexes
|
11 |
self.edges = edges
|
12 |
self.nodes = nodes
|
|
|
29 |
self.transformer_dim = vertexes * vertexes * dim + vertexes * dim
|
30 |
self.pos_enc_dim = 5
|
31 |
|
32 |
+
self.node_layers = nn.Sequential(nn.Linear(nodes, 64), act, nn.Linear(64, dim), act, nn.Dropout(self.dropout))
|
33 |
+
self.edge_layers = nn.Sequential(nn.Linear(edges, 64), act, nn.Linear(64, dim), act, nn.Dropout(self.dropout))
|
34 |
self.TransformerEncoder = TransformerEncoder(dim=self.dim, depth=self.depth, heads=self.heads, act = act,
|
35 |
mlp_ratio=self.mlp_ratio, drop_rate=self.dropout)
|
36 |
|
|
|
62 |
edge = self.edge_layers(z_e)
|
63 |
edge = (edge + edge.permute(0, 2, 1, 3)) / 2
|
64 |
|
65 |
+
node, edge = self.TransformerEncoder(node, edge)
|
66 |
|
67 |
node_sample = self.readout_n(node)
|
68 |
edge_sample = self.readout_e(edge)
|
69 |
+
|
70 |
return node, edge, node_sample, edge_sample
|
71 |
|
72 |
+
class Discriminator(nn.Module):
|
73 |
+
|
74 |
+
def __init__(self, act, vertexes, edges, nodes, dropout, dim, depth, heads, mlp_ratio):
|
75 |
+
super(Discriminator, self).__init__()
|
76 |
+
self.vertexes = vertexes
|
77 |
+
self.edges = edges
|
78 |
+
self.nodes = nodes
|
79 |
+
self.depth = depth
|
80 |
+
self.dim = dim
|
81 |
+
self.heads = heads
|
82 |
+
self.mlp_ratio = mlp_ratio
|
83 |
+
self.dropout = dropout
|
84 |
+
|
85 |
+
if act == "relu":
|
86 |
+
act = nn.ReLU()
|
87 |
+
elif act == "leaky":
|
88 |
+
act = nn.LeakyReLU()
|
89 |
+
elif act == "sigmoid":
|
90 |
+
act = nn.Sigmoid()
|
91 |
+
elif act == "tanh":
|
92 |
+
act = nn.Tanh()
|
93 |
+
|
94 |
+
self.features = vertexes * vertexes * edges + vertexes * nodes
|
95 |
+
self.transformer_dim = vertexes * vertexes * dim + vertexes * dim
|
96 |
+
|
97 |
+
self.node_layers = nn.Sequential(nn.Linear(nodes, 64), act, nn.Linear(64, dim), act, nn.Dropout(self.dropout))
|
98 |
+
self.edge_layers = nn.Sequential(nn.Linear(edges, 64), act, nn.Linear(64, dim), act, nn.Dropout(self.dropout))
|
99 |
+
self.TransformerEncoder = TransformerEncoder(dim=self.dim, depth=self.depth, heads=self.heads, act = act,
|
100 |
+
mlp_ratio=self.mlp_ratio, drop_rate=self.dropout)
|
101 |
+
self.node_features = vertexes * dim
|
102 |
+
self.edge_features = vertexes * vertexes * dim
|
103 |
+
self.node_mlp = nn.Sequential(nn.Linear(self.node_features, 64), act, nn.Linear(64, 32), act, nn.Linear(32, 16), act, nn.Linear(16, 1))
|
104 |
+
|
105 |
+
def forward(self, z_e, z_n):
|
106 |
+
b, n, c = z_n.shape
|
107 |
+
_, _, _ , d = z_e.shape
|
108 |
+
|
109 |
+
node = self.node_layers(z_n)
|
110 |
+
edge = self.edge_layers(z_e)
|
111 |
+
edge = (edge + edge.permute(0, 2, 1, 3)) / 2
|
112 |
+
|
113 |
+
node, edge = self.TransformerEncoder(node, edge)
|
114 |
+
|
115 |
+
node = node.view(b, -1)
|
116 |
+
|
117 |
+
prediction = self.node_mlp(node)
|
118 |
+
|
119 |
+
return prediction
|
120 |
|
121 |
class simple_disc(nn.Module):
|
122 |
def __init__(self, act, m_dim, vertexes, b_dim):
|
|
|
130 |
act = nn.Sigmoid()
|
131 |
elif act == "tanh":
|
132 |
act = nn.Tanh()
|
133 |
+
else:
|
134 |
+
raise ValueError("Unsupported activation function: {}".format(act))
|
135 |
|
136 |
features = vertexes * m_dim + vertexes * vertexes * b_dim
|
137 |
self.predictor = nn.Sequential(nn.Linear(features,256), act, nn.Linear(256,128), act, nn.Linear(128,64), act,
|
|
|
140 |
|
141 |
def forward(self, x):
|
142 |
prediction = self.predictor(x)
|
143 |
+
return prediction
|
smiles_cor.py
ADDED
@@ -0,0 +1,1291 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import os
|
3 |
+
import pandas as pd
|
4 |
+
import random
|
5 |
+
from chembl_structure_pipeline import standardizer
|
6 |
+
from rdkit.Chem import MolStandardize
|
7 |
+
from rdkit import Chem
|
8 |
+
import time
|
9 |
+
import torch
|
10 |
+
import torch.nn as nn
|
11 |
+
from torchtext.data import TabularDataset, Field, BucketIterator, Iterator
|
12 |
+
import random
|
13 |
+
import os
|
14 |
+
import torch
|
15 |
+
import torch.nn as nn
|
16 |
+
from torch.utils.data import DataLoader
|
17 |
+
import random
|
18 |
+
from torch import optim
|
19 |
+
import numpy as np
|
20 |
+
import itertools
|
21 |
+
import time
|
22 |
+
import statistics
|
23 |
+
from rdkit.Chem import GraphDescriptors, Lipinski, AllChem
|
24 |
+
from rdkit.Chem.rdSLNParse import MolFromSLN
|
25 |
+
from rdkit.Chem.rdmolfiles import MolFromSmiles
|
26 |
+
import torch
|
27 |
+
import torch.nn as nn
|
28 |
+
import torch.optim as optim
|
29 |
+
import pandas as pd
|
30 |
+
import numpy as np
|
31 |
+
from rdkit import rdBase, Chem
|
32 |
+
import re
|
33 |
+
from rdkit import RDLogger
|
34 |
+
RDLogger.DisableLog('rdApp.*')
|
35 |
+
|
36 |
+
SEED = 42
|
37 |
+
random.seed(SEED)
|
38 |
+
torch.manual_seed(SEED)
|
39 |
+
torch.backends.cudnn.deterministic = True
|
40 |
+
|
41 |
+
##################################################################################################
|
42 |
+
##################################################################################################
|
43 |
+
# #
|
44 |
+
# THIS SCRIPT IS DIRECTLY ADAPTED FROM https://github.com/LindeSchoenmaker/SMILES-corrector #
|
45 |
+
# #
|
46 |
+
##################################################################################################
|
47 |
+
##################################################################################################
|
48 |
+
def is_smiles(array,
|
49 |
+
TRG,
|
50 |
+
reverse: bool,
|
51 |
+
return_output=False,
|
52 |
+
src=None,
|
53 |
+
src_field=None):
|
54 |
+
"""Turns predicted tokens within batch into smiles and evaluates their validity
|
55 |
+
Arguments:
|
56 |
+
array: Tensor with most probable token for each location for each sequence in batch
|
57 |
+
[trg len, batch size]
|
58 |
+
TRG: target field for getting tokens from vocab
|
59 |
+
reverse (bool): True if the target sequence is reversed
|
60 |
+
return_output (bool): True if output sequences and their validity should be saved
|
61 |
+
Returns:
|
62 |
+
df: dataframe with correct and incorrect sequences
|
63 |
+
valids: list with booleans that show if prediction was a valid SMILES (True) or invalid one (False)
|
64 |
+
smiless: list of the predicted smiles
|
65 |
+
"""
|
66 |
+
trg_field = TRG
|
67 |
+
valids = []
|
68 |
+
smiless = []
|
69 |
+
if return_output:
|
70 |
+
df = pd.DataFrame()
|
71 |
+
else:
|
72 |
+
df = None
|
73 |
+
batch_size = array.size(1)
|
74 |
+
# check if the first token should be removed, first token is zero because
|
75 |
+
# outputs initaliazed to all be zeros
|
76 |
+
if int((array[0, 0]).tolist()) == 0:
|
77 |
+
start = 1
|
78 |
+
else:
|
79 |
+
start = 0
|
80 |
+
# for each sequence in the batch
|
81 |
+
for i in range(0, batch_size):
|
82 |
+
# turns sequence from tensor to list skipps first row as this is not
|
83 |
+
# filled in in forward
|
84 |
+
sequence = (array[start:, i]).tolist()
|
85 |
+
# goes from embedded to tokens
|
86 |
+
trg_tokens = [trg_field.vocab.itos[int(t)] for t in sequence]
|
87 |
+
# print(trg_tokens)
|
88 |
+
# takes all tokens untill eos token, model would be faster if did this
|
89 |
+
# one step earlier, but then changes in vocab order would disrupt.
|
90 |
+
rev_tokens = list(
|
91 |
+
itertools.takewhile(lambda x: x != "<eos>", trg_tokens))
|
92 |
+
if reverse:
|
93 |
+
rev_tokens = rev_tokens[::-1]
|
94 |
+
smiles = "".join(rev_tokens)
|
95 |
+
# determine how many valid smiles are made
|
96 |
+
valid = True if MolFromSmiles(smiles) else False
|
97 |
+
valids.append(valid)
|
98 |
+
smiless.append(smiles)
|
99 |
+
if return_output:
|
100 |
+
if valid:
|
101 |
+
df.loc[i, "CORRECT"] = smiles
|
102 |
+
else:
|
103 |
+
df.loc[i, "INCORRECT"] = smiles
|
104 |
+
|
105 |
+
# add the original drugex outputs to the _de dataframe
|
106 |
+
if return_output and src is not None:
|
107 |
+
for i in range(0, batch_size):
|
108 |
+
# turns sequence from tensor to list skipps first row as this is
|
109 |
+
# <sos> for src
|
110 |
+
sequence = (src[1:, i]).tolist()
|
111 |
+
# goes from embedded to tokens
|
112 |
+
src_tokens = [src_field.vocab.itos[int(t)] for t in sequence]
|
113 |
+
# takes all tokens untill eos token, model would be faster if did
|
114 |
+
# this one step earlier, but then changes in vocab order would
|
115 |
+
# disrupt.
|
116 |
+
rev_tokens = list(
|
117 |
+
itertools.takewhile(lambda x: x != "<eos>", src_tokens))
|
118 |
+
smiles = "".join(rev_tokens)
|
119 |
+
df.loc[i, "ORIGINAL"] = smiles
|
120 |
+
|
121 |
+
return df, valids, smiless
|
122 |
+
|
123 |
+
|
124 |
+
def is_unchanged(array,
|
125 |
+
TRG,
|
126 |
+
reverse: bool,
|
127 |
+
return_output=False,
|
128 |
+
src=None,
|
129 |
+
src_field=None):
|
130 |
+
"""Checks is output is different from input
|
131 |
+
Arguments:
|
132 |
+
array: Tensor with most probable token for each location for each sequence in batch
|
133 |
+
[trg len, batch size]
|
134 |
+
TRG: target field for getting tokens from vocab
|
135 |
+
reverse (bool): True if the target sequence is reversed
|
136 |
+
return_output (bool): True if output sequences and their validity should be saved
|
137 |
+
Returns:
|
138 |
+
df: dataframe with correct and incorrect sequences
|
139 |
+
valids: list with booleans that show if prediction was a valid SMILES (True) or invalid one (False)
|
140 |
+
smiless: list of the predicted smiles
|
141 |
+
"""
|
142 |
+
trg_field = TRG
|
143 |
+
sources = []
|
144 |
+
batch_size = array.size(1)
|
145 |
+
unchanged = 0
|
146 |
+
|
147 |
+
# check if the first token should be removed, first token is zero because
|
148 |
+
# outputs initaliazed to all be zeros
|
149 |
+
if int((array[0, 0]).tolist()) == 0:
|
150 |
+
start = 1
|
151 |
+
else:
|
152 |
+
start = 0
|
153 |
+
|
154 |
+
for i in range(0, batch_size):
|
155 |
+
# turns sequence from tensor to list skipps first row as this is <sos>
|
156 |
+
# for src
|
157 |
+
sequence = (src[1:, i]).tolist()
|
158 |
+
# goes from embedded to tokens
|
159 |
+
src_tokens = [src_field.vocab.itos[int(t)] for t in sequence]
|
160 |
+
# takes all tokens untill eos token, model would be faster if did this
|
161 |
+
# one step earlier, but then changes in vocab order would disrupt.
|
162 |
+
rev_tokens = list(
|
163 |
+
itertools.takewhile(lambda x: x != "<eos>", src_tokens))
|
164 |
+
smiles = "".join(rev_tokens)
|
165 |
+
sources.append(smiles)
|
166 |
+
|
167 |
+
# for each sequence in the batch
|
168 |
+
for i in range(0, batch_size):
|
169 |
+
# turns sequence from tensor to list skipps first row as this is not
|
170 |
+
# filled in in forward
|
171 |
+
sequence = (array[start:, i]).tolist()
|
172 |
+
# goes from embedded to tokens
|
173 |
+
trg_tokens = [trg_field.vocab.itos[int(t)] for t in sequence]
|
174 |
+
# print(trg_tokens)
|
175 |
+
# takes all tokens untill eos token, model would be faster if did this
|
176 |
+
# one step earlier, but then changes in vocab order would disrupt.
|
177 |
+
rev_tokens = list(
|
178 |
+
itertools.takewhile(lambda x: x != "<eos>", trg_tokens))
|
179 |
+
if reverse:
|
180 |
+
rev_tokens = rev_tokens[::-1]
|
181 |
+
smiles = "".join(rev_tokens)
|
182 |
+
# determine how many valid smiles are made
|
183 |
+
valid = True if MolFromSmiles(smiles) else False
|
184 |
+
if not valid:
|
185 |
+
if smiles == sources[i]:
|
186 |
+
unchanged += 1
|
187 |
+
|
188 |
+
return unchanged
|
189 |
+
|
190 |
+
|
191 |
+
def molecule_reconstruction(array, TRG, reverse: bool, outputs):
|
192 |
+
"""Turns target tokens within batch into smiles and compares them to predicted output smiles
|
193 |
+
Arguments:
|
194 |
+
array: Tensor with target's token for each location for each sequence in batch
|
195 |
+
[trg len, batch size]
|
196 |
+
TRG: target field for getting tokens from vocab
|
197 |
+
reverse (bool): True if the target sequence is reversed
|
198 |
+
outputs: list of predicted SMILES sequences
|
199 |
+
Returns:
|
200 |
+
matches(int): number of total right molecules
|
201 |
+
"""
|
202 |
+
trg_field = TRG
|
203 |
+
matches = 0
|
204 |
+
targets = []
|
205 |
+
batch_size = array.size(1)
|
206 |
+
# for each sequence in the batch
|
207 |
+
for i in range(0, batch_size):
|
208 |
+
# turns sequence from tensor to list skipps first row as this is not
|
209 |
+
# filled in in forward
|
210 |
+
sequence = (array[1:, i]).tolist()
|
211 |
+
# goes from embedded to tokens
|
212 |
+
trg_tokens = [trg_field.vocab.itos[int(t)] for t in sequence]
|
213 |
+
# takes all tokens untill eos token, model would be faster if did this
|
214 |
+
# one step earlier, but then changes in vocab order would disrupt.
|
215 |
+
rev_tokens = list(
|
216 |
+
itertools.takewhile(lambda x: x != "<eos>", trg_tokens))
|
217 |
+
if reverse:
|
218 |
+
rev_tokens = rev_tokens[::-1]
|
219 |
+
smiles = "".join(rev_tokens)
|
220 |
+
targets.append(smiles)
|
221 |
+
for i in range(0, batch_size):
|
222 |
+
m = MolFromSmiles(targets[i])
|
223 |
+
p = MolFromSmiles(outputs[i])
|
224 |
+
if p is not None:
|
225 |
+
if m.HasSubstructMatch(p) and p.HasSubstructMatch(m):
|
226 |
+
matches += 1
|
227 |
+
return matches
|
228 |
+
|
229 |
+
|
230 |
+
def complexity_whitlock(mol: Chem.Mol, includeAllDescs=False):
|
231 |
+
"""
|
232 |
+
Complexity as defined in DOI:10.1021/jo9814546
|
233 |
+
S: complexity = 4*#rings + 2*#unsat + #hetatm + 2*#chiral
|
234 |
+
Other descriptors:
|
235 |
+
H: size = #bonds (Hydrogen atoms included)
|
236 |
+
G: S + H
|
237 |
+
Ratio: S / H
|
238 |
+
"""
|
239 |
+
mol_ = Chem.Mol(mol)
|
240 |
+
nrings = Lipinski.RingCount(mol_) - Lipinski.NumAromaticRings(mol_)
|
241 |
+
Chem.rdmolops.SetAromaticity(mol_)
|
242 |
+
unsat = sum(1 for bond in mol_.GetBonds()
|
243 |
+
if bond.GetBondTypeAsDouble() == 2)
|
244 |
+
hetatm = len(mol_.GetSubstructMatches(Chem.MolFromSmarts("[!#6]")))
|
245 |
+
AllChem.EmbedMolecule(mol_)
|
246 |
+
Chem.AssignAtomChiralTagsFromStructure(mol_)
|
247 |
+
chiral = len(Chem.FindMolChiralCenters(mol_))
|
248 |
+
S = 4 * nrings + 2 * unsat + hetatm + 2 * chiral
|
249 |
+
if not includeAllDescs:
|
250 |
+
return S
|
251 |
+
Chem.rdmolops.Kekulize(mol_)
|
252 |
+
mol_ = Chem.AddHs(mol_)
|
253 |
+
H = sum(bond.GetBondTypeAsDouble() for bond in mol_.GetBonds())
|
254 |
+
G = S + H
|
255 |
+
R = S / H
|
256 |
+
return {"WhitlockS": S, "WhitlockH": H, "WhitlockG": G, "WhitlockRatio": R}
|
257 |
+
|
258 |
+
|
259 |
+
def complexity_baronechanon(mol: Chem.Mol):
|
260 |
+
"""
|
261 |
+
Complexity as defined in DOI:10.1021/ci000145p
|
262 |
+
"""
|
263 |
+
mol_ = Chem.Mol(mol)
|
264 |
+
Chem.Kekulize(mol_)
|
265 |
+
Chem.RemoveStereochemistry(mol_)
|
266 |
+
mol_ = Chem.RemoveHs(mol_, updateExplicitCount=True)
|
267 |
+
degree, counts = 0, 0
|
268 |
+
for atom in mol_.GetAtoms():
|
269 |
+
degree += 3 * 2**(atom.GetExplicitValence() - atom.GetNumExplicitHs() -
|
270 |
+
1)
|
271 |
+
counts += 3 if atom.GetSymbol() == "C" else 6
|
272 |
+
ringterm = sum(map(lambda x: 6 * len(x), mol_.GetRingInfo().AtomRings()))
|
273 |
+
return degree + counts + ringterm
|
274 |
+
|
275 |
+
|
276 |
+
def calc_complexity(array,
|
277 |
+
TRG,
|
278 |
+
reverse,
|
279 |
+
valids,
|
280 |
+
complexity_function=GraphDescriptors.BertzCT):
|
281 |
+
"""Calculates the complexity of inputs that are not correct.
|
282 |
+
Arguments:
|
283 |
+
array: Tensor with target's token for each location for each sequence in batch
|
284 |
+
[trg len, batch size]
|
285 |
+
TRG: target field for getting tokens from vocab
|
286 |
+
reverse (bool): True if the target sequence is reversed
|
287 |
+
valids: list with booleans that show if prediction was a valid SMILES (True) or invalid one (False)
|
288 |
+
complexity_function: the type of complexity measure that will be used
|
289 |
+
GraphDescriptors.BertzCT
|
290 |
+
complexity_whitlock
|
291 |
+
complexity_baronechanon
|
292 |
+
Returns:
|
293 |
+
matches(int): mean of complexity values
|
294 |
+
"""
|
295 |
+
trg_field = TRG
|
296 |
+
sources = []
|
297 |
+
complexities = []
|
298 |
+
loc = torch.BoolTensor(valids)
|
299 |
+
# only keeps rows in batch size dimension where valid is false
|
300 |
+
array = array[:, loc == False]
|
301 |
+
# should check if this still works
|
302 |
+
# array = torch.transpose(array, 0, 1)
|
303 |
+
array_size = array.size(1)
|
304 |
+
for i in range(0, array_size):
|
305 |
+
# turns sequence from tensor to list skipps first row as this is not
|
306 |
+
# filled in in forward
|
307 |
+
sequence = (array[1:, i]).tolist()
|
308 |
+
# goes from embedded to tokens
|
309 |
+
trg_tokens = [trg_field.vocab.itos[int(t)] for t in sequence]
|
310 |
+
# takes all tokens untill eos token, model would be faster if did this
|
311 |
+
# one step earlier, but then changes in vocab order would disrupt.
|
312 |
+
rev_tokens = list(
|
313 |
+
itertools.takewhile(lambda x: x != "<eos>", trg_tokens))
|
314 |
+
if reverse:
|
315 |
+
rev_tokens = rev_tokens[::-1]
|
316 |
+
smiles = "".join(rev_tokens)
|
317 |
+
sources.append(smiles)
|
318 |
+
for source in sources:
|
319 |
+
try:
|
320 |
+
m = MolFromSmiles(source)
|
321 |
+
except BaseException:
|
322 |
+
m = MolFromSLN(source)
|
323 |
+
complexities.append(complexity_function(m))
|
324 |
+
if len(complexities) > 0:
|
325 |
+
mean = statistics.mean(complexities)
|
326 |
+
else:
|
327 |
+
mean = 0
|
328 |
+
return mean
|
329 |
+
|
330 |
+
|
331 |
+
def epoch_time(start_time, end_time):
|
332 |
+
elapsed_time = end_time - start_time
|
333 |
+
elapsed_mins = int(elapsed_time / 60)
|
334 |
+
elapsed_secs = int(elapsed_time - (elapsed_mins * 60))
|
335 |
+
return elapsed_mins, elapsed_secs
|
336 |
+
|
337 |
+
|
338 |
+
class Convo:
|
339 |
+
"""Class for training and evaluating transformer and convolutional neural network
|
340 |
+
|
341 |
+
Methods
|
342 |
+
-------
|
343 |
+
train_model()
|
344 |
+
train model for initialized number of epochs
|
345 |
+
evaluate(return_output)
|
346 |
+
use model with validation loader (& optionally drugex loader) to get test loss & other metrics
|
347 |
+
translate(loader)
|
348 |
+
translate inputs from loader (different from evaluate in that no target sequence is used)
|
349 |
+
"""
|
350 |
+
|
351 |
+
def train_model(self):
|
352 |
+
optimizer = optim.Adam(self.parameters(), lr=self.lr)
|
353 |
+
log = open(f"{self.out}.log", "a")
|
354 |
+
best_error = np.inf
|
355 |
+
for epoch in range(self.epochs):
|
356 |
+
self.train()
|
357 |
+
start_time = time.time()
|
358 |
+
loss_train = 0
|
359 |
+
for i, batch in enumerate(self.loader_train):
|
360 |
+
optimizer.zero_grad()
|
361 |
+
# changed src,trg call to match with bentrevett
|
362 |
+
# src, trg = batch['src'], batch['trg']
|
363 |
+
trg = batch.trg
|
364 |
+
src = batch.src
|
365 |
+
output, attention = self(src, trg[:, :-1])
|
366 |
+
# feed the source and target into def forward to get the output
|
367 |
+
# Xuhan uses forward for this, with istrain = true
|
368 |
+
output_dim = output.shape[-1]
|
369 |
+
# changed
|
370 |
+
output = output.contiguous().view(-1, output_dim)
|
371 |
+
trg = trg[:, 1:].contiguous().view(-1)
|
372 |
+
# output = output[:,:,0]#.view(-1)
|
373 |
+
# output = output[1:].view(-1, output.shape[-1])
|
374 |
+
# trg = trg[1:].view(-1)
|
375 |
+
loss = nn.CrossEntropyLoss(
|
376 |
+
ignore_index=self.TRG.vocab.stoi[self.TRG.pad_token])
|
377 |
+
a, b = output.view(-1), trg.to(self.device).view(-1)
|
378 |
+
# changed
|
379 |
+
# loss = loss(output.view(0), trg.view(0).to(device))
|
380 |
+
loss = loss(output, trg)
|
381 |
+
loss.backward()
|
382 |
+
torch.nn.utils.clip_grad_norm_(self.parameters(), self.clip)
|
383 |
+
optimizer.step()
|
384 |
+
loss_train += loss.item()
|
385 |
+
# turned off for now, as not using voc so won't work, output is a tensor
|
386 |
+
# output = [(trg len - 1) * batch size, output dim]
|
387 |
+
# smiles, valid = is_valid_smiles(output, reversed)
|
388 |
+
# if valid:
|
389 |
+
# valids += 1
|
390 |
+
# smiless.append(smiles)
|
391 |
+
# added .dataset becaue len(iterator) gives len(self.dataset) /
|
392 |
+
# self.batch_size)
|
393 |
+
loss_train /= len(self.loader_train)
|
394 |
+
info = f"Epoch: {epoch+1:02} step: {i} loss_train: {loss_train:.4g}"
|
395 |
+
# model is used to generate trg based on src from the validation set to assess performance
|
396 |
+
# similar to Xuhan, although he doesn't use the if loop
|
397 |
+
if self.loader_valid is not None:
|
398 |
+
return_output = False
|
399 |
+
if epoch + 1 == self.epochs:
|
400 |
+
return_output = True
|
401 |
+
(
|
402 |
+
valids,
|
403 |
+
loss_valid,
|
404 |
+
valids_de,
|
405 |
+
df_output,
|
406 |
+
df_output_de,
|
407 |
+
right_molecules,
|
408 |
+
complexity,
|
409 |
+
unchanged,
|
410 |
+
unchanged_de,
|
411 |
+
) = self.evaluate(return_output)
|
412 |
+
reconstruction_error = 1 - right_molecules / len(
|
413 |
+
self.loader_valid.dataset)
|
414 |
+
error = 1 - valids / len(self.loader_valid.dataset)
|
415 |
+
complexity = complexity / len(self.loader_valid)
|
416 |
+
unchan = unchanged / (len(self.loader_valid.dataset) - valids)
|
417 |
+
info += f" loss_valid: {loss_valid:.4g} error_rate: {error:.4g} molecule_reconstruction_error_rate: {reconstruction_error:.4g} unchanged: {unchan:.4g} invalid_target_complexity: {complexity:.4g}"
|
418 |
+
if self.loader_drugex is not None:
|
419 |
+
error_de = 1 - valids_de / len(self.loader_drugex.dataset)
|
420 |
+
unchan_de = unchanged_de / (
|
421 |
+
len(self.loader_drugex.dataset) - valids_de)
|
422 |
+
info += f" error_rate_drugex: {error_de:.4g} unchanged_drugex: {unchan_de:.4g}"
|
423 |
+
|
424 |
+
if reconstruction_error < best_error:
|
425 |
+
torch.save(self.state_dict(), f"{self.out}.pkg")
|
426 |
+
best_error = reconstruction_error
|
427 |
+
last_save = epoch
|
428 |
+
else:
|
429 |
+
if epoch - last_save >= 10 and best_error != 1:
|
430 |
+
torch.save(self.state_dict(), f"{self.out}_last.pkg")
|
431 |
+
(
|
432 |
+
valids,
|
433 |
+
loss_valid,
|
434 |
+
valids_de,
|
435 |
+
df_output,
|
436 |
+
df_output_de,
|
437 |
+
right_molecules,
|
438 |
+
complexity,
|
439 |
+
unchanged,
|
440 |
+
unchanged_de,
|
441 |
+
) = self.evaluate(True)
|
442 |
+
end_time = time.time()
|
443 |
+
epoch_mins, epoch_secs = epoch_time(
|
444 |
+
start_time, end_time)
|
445 |
+
info += f" Time: {epoch_mins}m {epoch_secs}s"
|
446 |
+
|
447 |
+
break
|
448 |
+
elif error < best_error:
|
449 |
+
torch.save(self.state_dict(), f"{self.out}.pkg")
|
450 |
+
best_error = error
|
451 |
+
end_time = time.time()
|
452 |
+
epoch_mins, epoch_secs = epoch_time(start_time, end_time)
|
453 |
+
info += f" Time: {epoch_mins}m {epoch_secs}s"
|
454 |
+
|
455 |
+
|
456 |
+
torch.save(self.state_dict(), f"{self.out}_last.pkg")
|
457 |
+
log.close()
|
458 |
+
self.load_state_dict(torch.load(f"{self.out}.pkg"))
|
459 |
+
df_output.to_csv(f"{self.out}.csv", index=False)
|
460 |
+
df_output_de.to_csv(f"{self.out}_de.csv", index=False)
|
461 |
+
|
462 |
+
def evaluate(self, return_output):
|
463 |
+
self.eval()
|
464 |
+
test_loss = 0
|
465 |
+
df_output = pd.DataFrame()
|
466 |
+
df_output_de = pd.DataFrame()
|
467 |
+
valids = 0
|
468 |
+
valids_de = 0
|
469 |
+
unchanged = 0
|
470 |
+
unchanged_de = 0
|
471 |
+
right_molecules = 0
|
472 |
+
complexity = 0
|
473 |
+
with torch.no_grad():
|
474 |
+
for _, batch in enumerate(self.loader_valid):
|
475 |
+
trg = batch.trg
|
476 |
+
src = batch.src
|
477 |
+
output, attention = self.forward(src, trg[:, :-1])
|
478 |
+
pred_token = output.argmax(2)
|
479 |
+
array = torch.transpose(pred_token, 0, 1)
|
480 |
+
trg_trans = torch.transpose(trg, 0, 1)
|
481 |
+
output_dim = output.shape[-1]
|
482 |
+
output = output.contiguous().view(-1, output_dim)
|
483 |
+
trg = trg[:, 1:].contiguous().view(-1)
|
484 |
+
src_trans = torch.transpose(src, 0, 1)
|
485 |
+
df_batch, valid, smiless = is_smiles(
|
486 |
+
array, self.TRG, reverse=True, return_output=return_output)
|
487 |
+
unchanged += is_unchanged(
|
488 |
+
array,
|
489 |
+
self.TRG,
|
490 |
+
reverse=True,
|
491 |
+
return_output=return_output,
|
492 |
+
src=src_trans,
|
493 |
+
src_field=self.SRC,
|
494 |
+
)
|
495 |
+
matches = molecule_reconstruction(trg_trans,
|
496 |
+
self.TRG,
|
497 |
+
reverse=True,
|
498 |
+
outputs=smiless)
|
499 |
+
complexity += calc_complexity(trg_trans,
|
500 |
+
self.TRG,
|
501 |
+
reverse=True,
|
502 |
+
valids=valid)
|
503 |
+
if df_batch is not None:
|
504 |
+
df_output = pd.concat([df_output, df_batch],
|
505 |
+
ignore_index=True)
|
506 |
+
right_molecules += matches
|
507 |
+
valids += sum(valid)
|
508 |
+
# trg = trg[1:].view(-1)
|
509 |
+
# output, trg = output[1:].view(-1, output.shape[-1]), trg[1:].view(-1)
|
510 |
+
loss = nn.CrossEntropyLoss(
|
511 |
+
ignore_index=self.TRG.vocab.stoi[self.TRG.pad_token])
|
512 |
+
loss = loss(output, trg)
|
513 |
+
test_loss += loss.item()
|
514 |
+
if self.loader_drugex is not None:
|
515 |
+
for _, batch in enumerate(self.loader_drugex):
|
516 |
+
src = batch.src
|
517 |
+
output = self.translate_sentence(src, self.TRG,
|
518 |
+
self.device)
|
519 |
+
# checks the number of valid smiles
|
520 |
+
pred_token = output.argmax(2)
|
521 |
+
array = torch.transpose(pred_token, 0, 1)
|
522 |
+
src_trans = torch.transpose(src, 0, 1)
|
523 |
+
df_batch, valid, smiless = is_smiles(
|
524 |
+
array,
|
525 |
+
self.TRG,
|
526 |
+
reverse=True,
|
527 |
+
return_output=return_output,
|
528 |
+
src=src_trans,
|
529 |
+
src_field=self.SRC,
|
530 |
+
)
|
531 |
+
unchanged_de += is_unchanged(
|
532 |
+
array,
|
533 |
+
self.TRG,
|
534 |
+
reverse=True,
|
535 |
+
return_output=return_output,
|
536 |
+
src=src_trans,
|
537 |
+
src_field=self.SRC,
|
538 |
+
)
|
539 |
+
if df_batch is not None:
|
540 |
+
df_output_de = pd.concat([df_output_de, df_batch],
|
541 |
+
ignore_index=True)
|
542 |
+
valids_de += sum(valid)
|
543 |
+
return (
|
544 |
+
valids,
|
545 |
+
test_loss / len(self.loader_valid),
|
546 |
+
valids_de,
|
547 |
+
df_output,
|
548 |
+
df_output_de,
|
549 |
+
right_molecules,
|
550 |
+
complexity,
|
551 |
+
unchanged,
|
552 |
+
unchanged_de,
|
553 |
+
)
|
554 |
+
|
555 |
+
def translate(self, loader):
|
556 |
+
self.eval()
|
557 |
+
df_output_de = pd.DataFrame()
|
558 |
+
valids_de = 0
|
559 |
+
with torch.no_grad():
|
560 |
+
for _, batch in enumerate(loader):
|
561 |
+
src = batch.src
|
562 |
+
output = self.translate_sentence(src, self.TRG, self.device)
|
563 |
+
# checks the number of valid smiles
|
564 |
+
pred_token = output.argmax(2)
|
565 |
+
array = torch.transpose(pred_token, 0, 1)
|
566 |
+
src_trans = torch.transpose(src, 0, 1)
|
567 |
+
df_batch, valid, smiless = is_smiles(
|
568 |
+
array,
|
569 |
+
self.TRG,
|
570 |
+
reverse=True,
|
571 |
+
return_output=True,
|
572 |
+
src=src_trans,
|
573 |
+
src_field=self.SRC,
|
574 |
+
)
|
575 |
+
if df_batch is not None:
|
576 |
+
df_output_de = pd.concat([df_output_de, df_batch],
|
577 |
+
ignore_index=True)
|
578 |
+
valids_de += sum(valid)
|
579 |
+
return valids_de, df_output_de
|
580 |
+
|
581 |
+
|
582 |
+
class Encoder(nn.Module):
|
583 |
+
|
584 |
+
def __init__(self, input_dim, hid_dim, n_layers, n_heads, pf_dim, dropout,
|
585 |
+
max_length, device):
|
586 |
+
super().__init__()
|
587 |
+
self.device = device
|
588 |
+
self.tok_embedding = nn.Embedding(input_dim, hid_dim)
|
589 |
+
self.pos_embedding = nn.Embedding(max_length, hid_dim)
|
590 |
+
self.layers = nn.ModuleList([
|
591 |
+
EncoderLayer(hid_dim, n_heads, pf_dim, dropout, device)
|
592 |
+
for _ in range(n_layers)
|
593 |
+
])
|
594 |
+
|
595 |
+
self.dropout = nn.Dropout(dropout)
|
596 |
+
self.scale = torch.sqrt(torch.FloatTensor([hid_dim])).to(device)
|
597 |
+
|
598 |
+
def forward(self, src, src_mask):
|
599 |
+
# src = [batch size, src len]
|
600 |
+
# src_mask = [batch size, src len]
|
601 |
+
batch_size = src.shape[0]
|
602 |
+
src_len = src.shape[1]
|
603 |
+
pos = (torch.arange(0, src_len).unsqueeze(0).repeat(batch_size,
|
604 |
+
1).to(self.device))
|
605 |
+
# pos = [batch size, src len]
|
606 |
+
src = self.dropout((self.tok_embedding(src) * self.scale) +
|
607 |
+
self.pos_embedding(pos))
|
608 |
+
# src = [batch size, src len, hid dim]
|
609 |
+
for layer in self.layers:
|
610 |
+
src = layer(src, src_mask)
|
611 |
+
# src = [batch size, src len, hid dim]
|
612 |
+
return src
|
613 |
+
|
614 |
+
|
615 |
+
class EncoderLayer(nn.Module):
|
616 |
+
|
617 |
+
def __init__(self, hid_dim, n_heads, pf_dim, dropout, device):
|
618 |
+
super().__init__()
|
619 |
+
|
620 |
+
self.self_attn_layer_norm = nn.LayerNorm(hid_dim)
|
621 |
+
self.ff_layer_norm = nn.LayerNorm(hid_dim)
|
622 |
+
self.self_attention = MultiHeadAttentionLayer(hid_dim, n_heads,
|
623 |
+
dropout, device)
|
624 |
+
self.positionwise_feedforward = PositionwiseFeedforwardLayer(
|
625 |
+
hid_dim, pf_dim, dropout)
|
626 |
+
self.dropout = nn.Dropout(dropout)
|
627 |
+
|
628 |
+
def forward(self, src, src_mask):
|
629 |
+
# src = [batch size, src len, hid dim]
|
630 |
+
# src_mask = [batch size, src len]
|
631 |
+
# self attention
|
632 |
+
_src, _ = self.self_attention(src, src, src, src_mask)
|
633 |
+
# dropout, residual connection and layer norm
|
634 |
+
src = self.self_attn_layer_norm(src + self.dropout(_src))
|
635 |
+
# src = [batch size, src len, hid dim]
|
636 |
+
# positionwise feedforward
|
637 |
+
_src = self.positionwise_feedforward(src)
|
638 |
+
# dropout, residual and layer norm
|
639 |
+
src = self.ff_layer_norm(src + self.dropout(_src))
|
640 |
+
# src = [batch size, src len, hid dim]
|
641 |
+
|
642 |
+
return src
|
643 |
+
|
644 |
+
|
645 |
+
class MultiHeadAttentionLayer(nn.Module):
|
646 |
+
|
647 |
+
def __init__(self, hid_dim, n_heads, dropout, device):
|
648 |
+
super().__init__()
|
649 |
+
assert hid_dim % n_heads == 0
|
650 |
+
self.hid_dim = hid_dim
|
651 |
+
self.n_heads = n_heads
|
652 |
+
self.head_dim = hid_dim // n_heads
|
653 |
+
self.fc_q = nn.Linear(hid_dim, hid_dim)
|
654 |
+
self.fc_k = nn.Linear(hid_dim, hid_dim)
|
655 |
+
self.fc_v = nn.Linear(hid_dim, hid_dim)
|
656 |
+
self.fc_o = nn.Linear(hid_dim, hid_dim)
|
657 |
+
self.dropout = nn.Dropout(dropout)
|
658 |
+
self.scale = torch.sqrt(torch.FloatTensor([self.head_dim])).to(device)
|
659 |
+
|
660 |
+
def forward(self, query, key, value, mask=None):
|
661 |
+
batch_size = query.shape[0]
|
662 |
+
# query = [batch size, query len, hid dim]
|
663 |
+
# key = [batch size, key len, hid dim]
|
664 |
+
# value = [batch size, value len, hid dim]
|
665 |
+
Q = self.fc_q(query)
|
666 |
+
K = self.fc_k(key)
|
667 |
+
V = self.fc_v(value)
|
668 |
+
# Q = [batch size, query len, hid dim]
|
669 |
+
# K = [batch size, key len, hid dim]
|
670 |
+
# V = [batch size, value len, hid dim]
|
671 |
+
Q = Q.view(batch_size, -1, self.n_heads,
|
672 |
+
self.head_dim).permute(0, 2, 1, 3)
|
673 |
+
K = K.view(batch_size, -1, self.n_heads,
|
674 |
+
self.head_dim).permute(0, 2, 1, 3)
|
675 |
+
V = V.view(batch_size, -1, self.n_heads,
|
676 |
+
self.head_dim).permute(0, 2, 1, 3)
|
677 |
+
# Q = [batch size, n heads, query len, head dim]
|
678 |
+
# K = [batch size, n heads, key len, head dim]
|
679 |
+
# V = [batch size, n heads, value len, head dim]
|
680 |
+
energy = torch.matmul(Q, K.permute(0, 1, 3, 2)) / self.scale
|
681 |
+
# energy = [batch size, n heads, query len, key len]
|
682 |
+
if mask is not None:
|
683 |
+
energy = energy.masked_fill(mask == 0, -1e10)
|
684 |
+
attention = torch.softmax(energy, dim=-1)
|
685 |
+
# attention = [batch size, n heads, query len, key len]
|
686 |
+
x = torch.matmul(self.dropout(attention), V)
|
687 |
+
# x = [batch size, n heads, query len, head dim]
|
688 |
+
x = x.permute(0, 2, 1, 3).contiguous()
|
689 |
+
# x = [batch size, query len, n heads, head dim]
|
690 |
+
x = x.view(batch_size, -1, self.hid_dim)
|
691 |
+
# x = [batch size, query len, hid dim]
|
692 |
+
x = self.fc_o(x)
|
693 |
+
# x = [batch size, query len, hid dim]
|
694 |
+
return x, attention
|
695 |
+
|
696 |
+
|
697 |
+
class PositionwiseFeedforwardLayer(nn.Module):
|
698 |
+
|
699 |
+
def __init__(self, hid_dim, pf_dim, dropout):
|
700 |
+
super().__init__()
|
701 |
+
self.fc_1 = nn.Linear(hid_dim, pf_dim)
|
702 |
+
self.fc_2 = nn.Linear(pf_dim, hid_dim)
|
703 |
+
self.dropout = nn.Dropout(dropout)
|
704 |
+
|
705 |
+
def forward(self, x):
|
706 |
+
# x = [batch size, seq len, hid dim]
|
707 |
+
x = self.dropout(torch.relu(self.fc_1(x)))
|
708 |
+
# x = [batch size, seq len, pf dim]
|
709 |
+
x = self.fc_2(x)
|
710 |
+
# x = [batch size, seq len, hid dim]
|
711 |
+
|
712 |
+
return x
|
713 |
+
|
714 |
+
|
715 |
+
class Decoder(nn.Module):
|
716 |
+
|
717 |
+
def __init__(
|
718 |
+
self,
|
719 |
+
output_dim,
|
720 |
+
hid_dim,
|
721 |
+
n_layers,
|
722 |
+
n_heads,
|
723 |
+
pf_dim,
|
724 |
+
dropout,
|
725 |
+
max_length,
|
726 |
+
device,
|
727 |
+
):
|
728 |
+
super().__init__()
|
729 |
+
self.device = device
|
730 |
+
self.tok_embedding = nn.Embedding(output_dim, hid_dim)
|
731 |
+
self.pos_embedding = nn.Embedding(max_length, hid_dim)
|
732 |
+
self.layers = nn.ModuleList([
|
733 |
+
DecoderLayer(hid_dim, n_heads, pf_dim, dropout, device)
|
734 |
+
for _ in range(n_layers)
|
735 |
+
])
|
736 |
+
self.fc_out = nn.Linear(hid_dim, output_dim)
|
737 |
+
self.dropout = nn.Dropout(dropout)
|
738 |
+
self.scale = torch.sqrt(torch.FloatTensor([hid_dim])).to(device)
|
739 |
+
|
740 |
+
def forward(self, trg, enc_src, trg_mask, src_mask):
|
741 |
+
# trg = [batch size, trg len]
|
742 |
+
# enc_src = [batch size, src len, hid dim]
|
743 |
+
# trg_mask = [batch size, trg len]
|
744 |
+
# src_mask = [batch size, src len]
|
745 |
+
batch_size = trg.shape[0]
|
746 |
+
trg_len = trg.shape[1]
|
747 |
+
pos = (torch.arange(0, trg_len).unsqueeze(0).repeat(batch_size,
|
748 |
+
1).to(self.device))
|
749 |
+
# pos = [batch size, trg len]
|
750 |
+
trg = self.dropout((self.tok_embedding(trg) * self.scale) +
|
751 |
+
self.pos_embedding(pos))
|
752 |
+
# trg = [batch size, trg len, hid dim]
|
753 |
+
for layer in self.layers:
|
754 |
+
trg, attention = layer(trg, enc_src, trg_mask, src_mask)
|
755 |
+
# trg = [batch size, trg len, hid dim]
|
756 |
+
# attention = [batch size, n heads, trg len, src len]
|
757 |
+
output = self.fc_out(trg)
|
758 |
+
# output = [batch size, trg len, output dim]
|
759 |
+
return output, attention
|
760 |
+
|
761 |
+
|
762 |
+
class DecoderLayer(nn.Module):
|
763 |
+
|
764 |
+
def __init__(self, hid_dim, n_heads, pf_dim, dropout, device):
|
765 |
+
super().__init__()
|
766 |
+
self.self_attn_layer_norm = nn.LayerNorm(hid_dim)
|
767 |
+
self.enc_attn_layer_norm = nn.LayerNorm(hid_dim)
|
768 |
+
self.ff_layer_norm = nn.LayerNorm(hid_dim)
|
769 |
+
self.self_attention = MultiHeadAttentionLayer(hid_dim, n_heads,
|
770 |
+
dropout, device)
|
771 |
+
self.encoder_attention = MultiHeadAttentionLayer(
|
772 |
+
hid_dim, n_heads, dropout, device)
|
773 |
+
self.positionwise_feedforward = PositionwiseFeedforwardLayer(
|
774 |
+
hid_dim, pf_dim, dropout)
|
775 |
+
self.dropout = nn.Dropout(dropout)
|
776 |
+
|
777 |
+
def forward(self, trg, enc_src, trg_mask, src_mask):
|
778 |
+
# trg = [batch size, trg len, hid dim]
|
779 |
+
# enc_src = [batch size, src len, hid dim]
|
780 |
+
# trg_mask = [batch size, trg len]
|
781 |
+
# src_mask = [batch size, src len]
|
782 |
+
# self attention
|
783 |
+
_trg, _ = self.self_attention(trg, trg, trg, trg_mask)
|
784 |
+
# dropout, residual connection and layer norm
|
785 |
+
trg = self.self_attn_layer_norm(trg + self.dropout(_trg))
|
786 |
+
# trg = [batch size, trg len, hid dim]
|
787 |
+
# encoder attention
|
788 |
+
_trg, attention = self.encoder_attention(trg, enc_src, enc_src,
|
789 |
+
src_mask)
|
790 |
+
# dropout, residual connection and layer norm
|
791 |
+
trg = self.enc_attn_layer_norm(trg + self.dropout(_trg))
|
792 |
+
# trg = [batch size, trg len, hid dim]
|
793 |
+
# positionwise feedforward
|
794 |
+
_trg = self.positionwise_feedforward(trg)
|
795 |
+
# dropout, residual and layer norm
|
796 |
+
trg = self.ff_layer_norm(trg + self.dropout(_trg))
|
797 |
+
# trg = [batch size, trg len, hid dim]
|
798 |
+
# attention = [batch size, n heads, trg len, src len]
|
799 |
+
return trg, attention
|
800 |
+
|
801 |
+
|
802 |
+
class Seq2Seq(nn.Module, Convo):
|
803 |
+
|
804 |
+
def __init__(
|
805 |
+
self,
|
806 |
+
encoder,
|
807 |
+
decoder,
|
808 |
+
src_pad_idx,
|
809 |
+
trg_pad_idx,
|
810 |
+
device,
|
811 |
+
loader_train: DataLoader,
|
812 |
+
out: str,
|
813 |
+
loader_valid=None,
|
814 |
+
loader_drugex=None,
|
815 |
+
epochs=100,
|
816 |
+
lr=0.0005,
|
817 |
+
clip=0.1,
|
818 |
+
reverse=True,
|
819 |
+
TRG=None,
|
820 |
+
SRC=None,
|
821 |
+
):
|
822 |
+
super().__init__()
|
823 |
+
self.encoder = encoder
|
824 |
+
self.decoder = decoder
|
825 |
+
self.src_pad_idx = src_pad_idx
|
826 |
+
self.trg_pad_idx = trg_pad_idx
|
827 |
+
self.device = device
|
828 |
+
self.loader_train = loader_train
|
829 |
+
self.out = out
|
830 |
+
self.loader_valid = loader_valid
|
831 |
+
self.loader_drugex = loader_drugex
|
832 |
+
self.epochs = epochs
|
833 |
+
self.lr = lr
|
834 |
+
self.clip = clip
|
835 |
+
self.reverse = reverse
|
836 |
+
self.TRG = TRG
|
837 |
+
self.SRC = SRC
|
838 |
+
|
839 |
+
def make_src_mask(self, src):
|
840 |
+
# src = [batch size, src len]
|
841 |
+
src_mask = (src != self.src_pad_idx).unsqueeze(1).unsqueeze(2)
|
842 |
+
# src_mask = [batch size, 1, 1, src len]
|
843 |
+
return src_mask
|
844 |
+
|
845 |
+
def make_trg_mask(self, trg):
|
846 |
+
# trg = [batch size, trg len]
|
847 |
+
trg_pad_mask = (trg != self.trg_pad_idx).unsqueeze(1).unsqueeze(2)
|
848 |
+
# trg_pad_mask = [batch size, 1, 1, trg len]
|
849 |
+
trg_len = trg.shape[1]
|
850 |
+
trg_sub_mask = torch.tril(
|
851 |
+
torch.ones((trg_len, trg_len), device=self.device)).bool()
|
852 |
+
# trg_sub_mask = [trg len, trg len]
|
853 |
+
trg_mask = trg_pad_mask & trg_sub_mask
|
854 |
+
# trg_mask = [batch size, 1, trg len, trg len]
|
855 |
+
return trg_mask
|
856 |
+
|
857 |
+
def forward(self, src, trg):
|
858 |
+
# src = [batch size, src len]
|
859 |
+
# trg = [batch size, trg len]
|
860 |
+
src_mask = self.make_src_mask(src)
|
861 |
+
trg_mask = self.make_trg_mask(trg)
|
862 |
+
# src_mask = [batch size, 1, 1, src len]
|
863 |
+
# trg_mask = [batch size, 1, trg len, trg len]
|
864 |
+
enc_src = self.encoder(src, src_mask)
|
865 |
+
# enc_src = [batch size, src len, hid dim]
|
866 |
+
output, attention = self.decoder(trg, enc_src, trg_mask, src_mask)
|
867 |
+
# output = [batch size, trg len, output dim]
|
868 |
+
# attention = [batch size, n heads, trg len, src len]
|
869 |
+
return output, attention
|
870 |
+
|
871 |
+
def translate_sentence(self, src, trg_field, device, max_len=202):
|
872 |
+
self.eval()
|
873 |
+
src_mask = self.make_src_mask(src)
|
874 |
+
with torch.no_grad():
|
875 |
+
enc_src = self.encoder(src, src_mask)
|
876 |
+
trg_indexes = [trg_field.vocab.stoi[trg_field.init_token]]
|
877 |
+
batch_size = src.shape[0]
|
878 |
+
trg = torch.LongTensor(trg_indexes).unsqueeze(0).to(device)
|
879 |
+
trg = trg.repeat(batch_size, 1)
|
880 |
+
for i in range(max_len):
|
881 |
+
# turned model into self.
|
882 |
+
trg_mask = self.make_trg_mask(trg)
|
883 |
+
with torch.no_grad():
|
884 |
+
output, attention = self.decoder(trg, enc_src, trg_mask,
|
885 |
+
src_mask)
|
886 |
+
pred_tokens = output.argmax(2)[:, -1].unsqueeze(1)
|
887 |
+
trg = torch.cat((trg, pred_tokens), 1)
|
888 |
+
|
889 |
+
return output
|
890 |
+
|
891 |
+
|
892 |
+
def remove_floats(df: pd.DataFrame, subset: str):
|
893 |
+
"""Preprocessing step to remove any entries that are not strings"""
|
894 |
+
df_subset = df[subset]
|
895 |
+
df[subset] = df[subset].astype(str)
|
896 |
+
# only keep entries that stayed the same after applying astype str
|
897 |
+
df = df[df[subset] == df_subset].copy()
|
898 |
+
|
899 |
+
return df
|
900 |
+
|
901 |
+
|
902 |
+
def smi_tokenizer(smi: str, reverse=False) -> list:
|
903 |
+
"""
|
904 |
+
Tokenize a SMILES molecule
|
905 |
+
"""
|
906 |
+
pattern = r"(\[[^\]]+]|Br?|Cl?|N|O|S|P|F|I|b|c|n|o|s|p|\(|\)|\.|=|#|-|\+|\\\\|\\|\/|:|~|@|\?|>|\*|\$|\%[0-9]{2}|[0-9])"
|
907 |
+
regex = re.compile(pattern)
|
908 |
+
# tokens = ['<sos>'] + [token for token in regex.findall(smi)] + ['<eos>']
|
909 |
+
tokens = [token for token in regex.findall(smi)]
|
910 |
+
# assert smi == ''.join(tokens[1:-1])
|
911 |
+
assert smi == "".join(tokens[:])
|
912 |
+
# try:
|
913 |
+
# assert smi == "".join(tokens[:])
|
914 |
+
# except:
|
915 |
+
# print(smi)
|
916 |
+
# print("".join(tokens[:]))
|
917 |
+
if reverse:
|
918 |
+
return tokens[::-1]
|
919 |
+
return tokens
|
920 |
+
|
921 |
+
|
922 |
+
def init_weights(m: nn.Module):
|
923 |
+
if hasattr(m, "weight") and m.weight.dim() > 1:
|
924 |
+
nn.init.xavier_uniform_(m.weight.data)
|
925 |
+
|
926 |
+
|
927 |
+
def count_parameters(model: nn.Module):
|
928 |
+
return sum(p.numel() for p in model.parameters() if p.requires_grad)
|
929 |
+
|
930 |
+
|
931 |
+
def epoch_time(start_time, end_time):
|
932 |
+
elapsed_time = end_time - start_time
|
933 |
+
elapsed_mins = int(elapsed_time / 60)
|
934 |
+
elapsed_secs = int(elapsed_time - (elapsed_mins * 60))
|
935 |
+
return elapsed_mins, elapsed_secs
|
936 |
+
|
937 |
+
|
938 |
+
def initialize_model(folder_out: str,
|
939 |
+
data_source: str,
|
940 |
+
error_source: str,
|
941 |
+
device: torch.device,
|
942 |
+
threshold: int,
|
943 |
+
epochs: int,
|
944 |
+
layers: int = 3,
|
945 |
+
batch_size: int = 16,
|
946 |
+
invalid_type: str = "all",
|
947 |
+
num_errors: int = 1,
|
948 |
+
validation_step=False):
|
949 |
+
"""Create encoder decoder models for specified model (currently only translator) & type of invalid SMILES
|
950 |
+
|
951 |
+
param data: collection of invalid, valid SMILES pairs
|
952 |
+
param invalid_smiles_path: path to previously generated invalid SMILES
|
953 |
+
param invalid_type: type of errors introduced into invalid SMILES
|
954 |
+
|
955 |
+
return:
|
956 |
+
|
957 |
+
"""
|
958 |
+
|
959 |
+
# set fields
|
960 |
+
SRC = Field(
|
961 |
+
tokenize=lambda x: smi_tokenizer(x),
|
962 |
+
init_token="<sos>",
|
963 |
+
eos_token="<eos>",
|
964 |
+
batch_first=True,
|
965 |
+
)
|
966 |
+
TRG = Field(
|
967 |
+
tokenize=lambda x: smi_tokenizer(x, reverse=True),
|
968 |
+
init_token="<sos>",
|
969 |
+
eos_token="<eos>",
|
970 |
+
batch_first=True,
|
971 |
+
)
|
972 |
+
|
973 |
+
if validation_step:
|
974 |
+
train, val = TabularDataset.splits(
|
975 |
+
path=f'{folder_out}errors/split/',
|
976 |
+
train=f"{data_source}_{invalid_type}_{num_errors}_errors_train.csv",
|
977 |
+
validation=
|
978 |
+
f"{data_source}_{invalid_type}_{num_errors}_errors_dev.csv",
|
979 |
+
format="CSV",
|
980 |
+
skip_header=False,
|
981 |
+
fields={
|
982 |
+
"ERROR": ("src", SRC),
|
983 |
+
"STD_SMILES": ("trg", TRG)
|
984 |
+
},
|
985 |
+
)
|
986 |
+
SRC.build_vocab(train, val, max_size=1000)
|
987 |
+
TRG.build_vocab(train, val, max_size=1000)
|
988 |
+
else:
|
989 |
+
train = TabularDataset(
|
990 |
+
path=
|
991 |
+
f'{folder_out}{data_source}_{invalid_type}_{num_errors}_errors.csv',
|
992 |
+
format="CSV",
|
993 |
+
skip_header=False,
|
994 |
+
fields={
|
995 |
+
"ERROR": ("src", SRC),
|
996 |
+
"STD_SMILES": ("trg", TRG)
|
997 |
+
},
|
998 |
+
)
|
999 |
+
SRC.build_vocab(train, max_size=1000)
|
1000 |
+
TRG.build_vocab(train, max_size=1000)
|
1001 |
+
|
1002 |
+
drugex = TabularDataset(
|
1003 |
+
path=error_source,
|
1004 |
+
format="csv",
|
1005 |
+
skip_header=False,
|
1006 |
+
fields={
|
1007 |
+
"SMILES": ("src", SRC),
|
1008 |
+
"SMILES_TARGET": ("trg", TRG)
|
1009 |
+
},
|
1010 |
+
)
|
1011 |
+
|
1012 |
+
|
1013 |
+
#SRC.vocab = torch.load('vocab_src.pth')
|
1014 |
+
#TRG.vocab = torch.load('vocab_trg.pth')
|
1015 |
+
|
1016 |
+
# model parameters
|
1017 |
+
EPOCHS = epochs
|
1018 |
+
BATCH_SIZE = batch_size
|
1019 |
+
INPUT_DIM = len(SRC.vocab)
|
1020 |
+
OUTPUT_DIM = len(TRG.vocab)
|
1021 |
+
HID_DIM = 256
|
1022 |
+
ENC_LAYERS = layers
|
1023 |
+
DEC_LAYERS = layers
|
1024 |
+
ENC_HEADS = 8
|
1025 |
+
DEC_HEADS = 8
|
1026 |
+
ENC_PF_DIM = 512
|
1027 |
+
DEC_PF_DIM = 512
|
1028 |
+
ENC_DROPOUT = 0.1
|
1029 |
+
DEC_DROPOUT = 0.1
|
1030 |
+
SRC_PAD_IDX = SRC.vocab.stoi[SRC.pad_token]
|
1031 |
+
TRG_PAD_IDX = TRG.vocab.stoi[TRG.pad_token]
|
1032 |
+
# add 2 to length for start and stop tokens
|
1033 |
+
MAX_LENGTH = threshold + 2
|
1034 |
+
|
1035 |
+
# model name
|
1036 |
+
MODEL_OUT_FOLDER = f"{folder_out}"
|
1037 |
+
|
1038 |
+
MODEL_NAME = "transformer_%s_%s_%s_%s_%s" % (
|
1039 |
+
invalid_type, num_errors, data_source, BATCH_SIZE, layers)
|
1040 |
+
if not os.path.exists(MODEL_OUT_FOLDER):
|
1041 |
+
os.mkdir(MODEL_OUT_FOLDER)
|
1042 |
+
|
1043 |
+
out = os.path.join(MODEL_OUT_FOLDER, MODEL_NAME)
|
1044 |
+
|
1045 |
+
torch.save(SRC.vocab, f'{out}_vocab_src.pth')
|
1046 |
+
torch.save(TRG.vocab, f'{out}_vocab_trg.pth')
|
1047 |
+
|
1048 |
+
# iterator is a dataloader
|
1049 |
+
# iterator to pass to the same length and create batches in which the
|
1050 |
+
# amount of padding is minimized
|
1051 |
+
if validation_step:
|
1052 |
+
train_iter, val_iter = BucketIterator.splits(
|
1053 |
+
(train, val),
|
1054 |
+
batch_sizes=(BATCH_SIZE, 256),
|
1055 |
+
sort_within_batch=True,
|
1056 |
+
shuffle=True,
|
1057 |
+
# the BucketIterator needs to be told what function it should use to
|
1058 |
+
# group the data.
|
1059 |
+
sort_key=lambda x: len(x.src),
|
1060 |
+
device=device,
|
1061 |
+
)
|
1062 |
+
else:
|
1063 |
+
train_iter = BucketIterator(
|
1064 |
+
train,
|
1065 |
+
batch_size=BATCH_SIZE,
|
1066 |
+
sort_within_batch=True,
|
1067 |
+
shuffle=True,
|
1068 |
+
# the BucketIterator needs to be told what function it should use to
|
1069 |
+
# group the data.
|
1070 |
+
sort_key=lambda x: len(x.src),
|
1071 |
+
device=device,
|
1072 |
+
)
|
1073 |
+
val_iter = None
|
1074 |
+
|
1075 |
+
drugex_iter = Iterator(
|
1076 |
+
drugex,
|
1077 |
+
batch_size=64,
|
1078 |
+
device=device,
|
1079 |
+
sort=False,
|
1080 |
+
sort_within_batch=True,
|
1081 |
+
sort_key=lambda x: len(x.src),
|
1082 |
+
repeat=False,
|
1083 |
+
)
|
1084 |
+
|
1085 |
+
|
1086 |
+
# model initialization
|
1087 |
+
|
1088 |
+
enc = Encoder(
|
1089 |
+
INPUT_DIM,
|
1090 |
+
HID_DIM,
|
1091 |
+
ENC_LAYERS,
|
1092 |
+
ENC_HEADS,
|
1093 |
+
ENC_PF_DIM,
|
1094 |
+
ENC_DROPOUT,
|
1095 |
+
MAX_LENGTH,
|
1096 |
+
device,
|
1097 |
+
)
|
1098 |
+
dec = Decoder(
|
1099 |
+
OUTPUT_DIM,
|
1100 |
+
HID_DIM,
|
1101 |
+
DEC_LAYERS,
|
1102 |
+
DEC_HEADS,
|
1103 |
+
DEC_PF_DIM,
|
1104 |
+
DEC_DROPOUT,
|
1105 |
+
MAX_LENGTH,
|
1106 |
+
device,
|
1107 |
+
)
|
1108 |
+
|
1109 |
+
model = Seq2Seq(
|
1110 |
+
enc,
|
1111 |
+
dec,
|
1112 |
+
SRC_PAD_IDX,
|
1113 |
+
TRG_PAD_IDX,
|
1114 |
+
device,
|
1115 |
+
train_iter,
|
1116 |
+
out=out,
|
1117 |
+
loader_valid=val_iter,
|
1118 |
+
loader_drugex=drugex_iter,
|
1119 |
+
epochs=EPOCHS,
|
1120 |
+
TRG=TRG,
|
1121 |
+
SRC=SRC,
|
1122 |
+
).to(device)
|
1123 |
+
|
1124 |
+
|
1125 |
+
|
1126 |
+
|
1127 |
+
return model, out, SRC
|
1128 |
+
|
1129 |
+
|
1130 |
+
def train_model(model, out, assess):
|
1131 |
+
"""Apply given weights (& assess performance or train further) or start training new model
|
1132 |
+
|
1133 |
+
Args:
|
1134 |
+
model: initialized model
|
1135 |
+
out: .pkg file with model parameters
|
1136 |
+
asses: bool
|
1137 |
+
|
1138 |
+
Returns:
|
1139 |
+
model with (new) weights
|
1140 |
+
"""
|
1141 |
+
|
1142 |
+
if os.path.exists(f"{out}.pkg") and assess:
|
1143 |
+
|
1144 |
+
|
1145 |
+
model.load_state_dict(torch.load(f=out + ".pkg"))
|
1146 |
+
(
|
1147 |
+
valids,
|
1148 |
+
loss_valid,
|
1149 |
+
valids_de,
|
1150 |
+
df_output,
|
1151 |
+
df_output_de,
|
1152 |
+
right_molecules,
|
1153 |
+
complexity,
|
1154 |
+
unchanged,
|
1155 |
+
unchanged_de,
|
1156 |
+
) = model.evaluate(True)
|
1157 |
+
|
1158 |
+
|
1159 |
+
# log = open('unchanged.log', 'a')
|
1160 |
+
# info = f'type: comb unchanged: {unchan:.4g} unchanged_drugex: {unchan_de:.4g}'
|
1161 |
+
# print(info, file=log, flush = True)
|
1162 |
+
# print(valids_de)
|
1163 |
+
# print(unchanged_de)
|
1164 |
+
|
1165 |
+
# print(unchan)
|
1166 |
+
# print(unchan_de)
|
1167 |
+
# df_output_de.to_csv(f'{out}_de_new.csv', index = False)
|
1168 |
+
|
1169 |
+
# error_de = 1 - valids_de / len(drugex_iter.dataset)
|
1170 |
+
# print(error_de)
|
1171 |
+
# df_output.to_csv(f'{out}_par.csv', index = False)
|
1172 |
+
|
1173 |
+
elif os.path.exists(f"{out}.pkg"):
|
1174 |
+
|
1175 |
+
# starts from the model after the last epoch, not the best epoch
|
1176 |
+
model.load_state_dict(torch.load(f=out + "_last.pkg"))
|
1177 |
+
# need to change how log file names epochs
|
1178 |
+
model.train_model()
|
1179 |
+
else:
|
1180 |
+
|
1181 |
+
model = model.apply(init_weights)
|
1182 |
+
model.train_model()
|
1183 |
+
|
1184 |
+
return model
|
1185 |
+
|
1186 |
+
|
1187 |
+
def correct_SMILES(model, out, error_source, device, SRC):
|
1188 |
+
"""Model that is given corrects SMILES and return number of correct ouputs and dataframe containing all outputs
|
1189 |
+
Args:
|
1190 |
+
model: initialized model
|
1191 |
+
out: .pkg file with model parameters
|
1192 |
+
asses: bool
|
1193 |
+
|
1194 |
+
Returns:
|
1195 |
+
valids: number of fixed outputs
|
1196 |
+
df_output: dataframe containing output (either correct or incorrect) & original input
|
1197 |
+
"""
|
1198 |
+
## account for tokens that are not yet in SRC without changing existing SRC token embeddings
|
1199 |
+
errors = TabularDataset(
|
1200 |
+
path=error_source,
|
1201 |
+
format="csv",
|
1202 |
+
skip_header=False,
|
1203 |
+
fields={"SMILES": ("src", SRC)},
|
1204 |
+
)
|
1205 |
+
|
1206 |
+
errors_loader = Iterator(
|
1207 |
+
errors,
|
1208 |
+
batch_size=64,
|
1209 |
+
device=device,
|
1210 |
+
sort=False,
|
1211 |
+
sort_within_batch=True,
|
1212 |
+
sort_key=lambda x: len(x.src),
|
1213 |
+
repeat=False,
|
1214 |
+
)
|
1215 |
+
model.load_state_dict(torch.load(f=out + ".pkg",map_location=torch.device('cpu')))
|
1216 |
+
# add option to use different iterator maybe?
|
1217 |
+
|
1218 |
+
valids, df_output = model.translate(errors_loader)
|
1219 |
+
#df_output.to_csv(f"{error_source}_fixed.csv", index=False)
|
1220 |
+
|
1221 |
+
|
1222 |
+
return valids, df_output
|
1223 |
+
|
1224 |
+
|
1225 |
+
|
1226 |
+
class smi_correct(object):
|
1227 |
+
def __init__(self, model_name, trans_file_path):
|
1228 |
+
# set random seed, used for error generation & initiation transformer
|
1229 |
+
|
1230 |
+
self.SEED = 42
|
1231 |
+
random.seed(self.SEED)
|
1232 |
+
self.model_name = model_name
|
1233 |
+
self.folder_out = "DrugGEN/data/"
|
1234 |
+
|
1235 |
+
self.trans_file_path = trans_file_path
|
1236 |
+
|
1237 |
+
if not os.path.exists(self.folder_out):
|
1238 |
+
os.makedirs(self.folder_out)
|
1239 |
+
|
1240 |
+
self.invalid_type = 'multiple'
|
1241 |
+
self.num_errors = 12
|
1242 |
+
self.threshold = 200
|
1243 |
+
self.data_source = f"PAPYRUS_{self.threshold}"
|
1244 |
+
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
|
1245 |
+
self.initialize_source = 'DrugGEN/data/papyrus_rnn_S.csv' # change this path
|
1246 |
+
|
1247 |
+
def standardization_pipeline(self, smile):
|
1248 |
+
desalter = MolStandardize.fragment.LargestFragmentChooser()
|
1249 |
+
std_smile = None
|
1250 |
+
if not isinstance(smile, str): return None
|
1251 |
+
m = Chem.MolFromSmiles(smile)
|
1252 |
+
# skips smiles for which no mol file could be generated
|
1253 |
+
if m is not None:
|
1254 |
+
# standardizes
|
1255 |
+
std_m = standardizer.standardize_mol(m)
|
1256 |
+
# strips salts
|
1257 |
+
std_m_p, exclude = standardizer.get_parent_mol(std_m)
|
1258 |
+
if not exclude:
|
1259 |
+
# choose largest fragment for rare cases where chembl structure
|
1260 |
+
# pipeline leaves 2 fragments
|
1261 |
+
std_m_p_d = desalter.choose(std_m_p)
|
1262 |
+
std_smile = Chem.MolToSmiles(std_m_p_d)
|
1263 |
+
return std_smile
|
1264 |
+
|
1265 |
+
def remove_smiles_duplicates(self, dataframe: pd.DataFrame,
|
1266 |
+
subset: str) -> pd.DataFrame:
|
1267 |
+
return dataframe.drop_duplicates(subset=subset)
|
1268 |
+
|
1269 |
+
def correct(self, smi):
|
1270 |
+
|
1271 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
1272 |
+
|
1273 |
+
model, out, SRC = initialize_model(self.folder_out,
|
1274 |
+
self.data_source,
|
1275 |
+
error_source=self.initialize_source,
|
1276 |
+
device=device,
|
1277 |
+
threshold=self.threshold,
|
1278 |
+
epochs=30,
|
1279 |
+
layers=3,
|
1280 |
+
batch_size=16,
|
1281 |
+
invalid_type=self.invalid_type,
|
1282 |
+
num_errors=self.num_errors)
|
1283 |
+
|
1284 |
+
valids, df_output = correct_SMILES(model, out, smi, device,
|
1285 |
+
SRC)
|
1286 |
+
|
1287 |
+
df_output["SMILES"] = df_output.apply(lambda row: self.standardization_pipeline(row["CORRECT"]), axis=1)
|
1288 |
+
|
1289 |
+
df_output = self.remove_smiles_duplicates(df_output, subset="SMILES").drop(columns=["CORRECT", "INCORRECT", "ORIGINAL"]).dropna()
|
1290 |
+
|
1291 |
+
return df_output
|
utils.py
CHANGED
@@ -42,7 +42,15 @@ class Metrics(object):
|
|
42 |
|
43 |
@staticmethod
|
44 |
def max_component(data, max_len):
|
|
|
|
|
45 |
|
|
|
|
|
|
|
|
|
|
|
|
|
46 |
return ((np.array(list(map(Metrics.mol_length, data)), dtype=np.float32)/max_len).mean())
|
47 |
|
48 |
@staticmethod
|
@@ -347,7 +355,7 @@ def canonic_smiles(smiles_or_mol):
|
|
347 |
if mol is None:
|
348 |
return None
|
349 |
return Chem.MolToSmiles(mol)
|
350 |
-
def fraction_unique(gen, k=None, n_jobs=1, check_validity=
|
351 |
"""
|
352 |
Computes a number of unique molecules
|
353 |
Parameters:
|
@@ -363,11 +371,13 @@ def fraction_unique(gen, k=None, n_jobs=1, check_validity=False):
|
|
363 |
"gen contains only {} molecules".format(len(gen))
|
364 |
)
|
365 |
gen = gen[:k]
|
366 |
-
|
367 |
-
|
368 |
-
|
369 |
-
|
370 |
-
|
|
|
|
|
371 |
|
372 |
def novelty(gen, train, n_jobs=1):
|
373 |
gen_smiles = mapper(n_jobs)(canonic_smiles, gen)
|
@@ -375,7 +385,8 @@ def novelty(gen, train, n_jobs=1):
|
|
375 |
train_set = set(train)
|
376 |
return 0 if len(gen_smiles_set) == 0 else len(gen_smiles_set - train_set) / len(gen_smiles_set)
|
377 |
|
378 |
-
|
|
|
379 |
|
380 |
def average_agg_tanimoto(stock_vecs, gen_vecs,
|
381 |
batch_size=5000, agg='max',
|
|
|
42 |
|
43 |
@staticmethod
|
44 |
def max_component(data, max_len):
|
45 |
+
|
46 |
+
# There will be a name change for this function to better reflect what it does
|
47 |
|
48 |
+
"""Returns the average length of the molecules in the dataset normalized by the maximum length.
|
49 |
+
|
50 |
+
Returns:
|
51 |
+
array: normalized average length of the molecules in the dataset
|
52 |
+
"""
|
53 |
+
|
54 |
return ((np.array(list(map(Metrics.mol_length, data)), dtype=np.float32)/max_len).mean())
|
55 |
|
56 |
@staticmethod
|
|
|
355 |
if mol is None:
|
356 |
return None
|
357 |
return Chem.MolToSmiles(mol)
|
358 |
+
def fraction_unique(gen, k=None, n_jobs=1, check_validity=True):
|
359 |
"""
|
360 |
Computes a number of unique molecules
|
361 |
Parameters:
|
|
|
371 |
"gen contains only {} molecules".format(len(gen))
|
372 |
)
|
373 |
gen = gen[:k]
|
374 |
+
if check_validity:
|
375 |
+
|
376 |
+
canonic = list(mapper(n_jobs)(canonic_smiles, gen))
|
377 |
+
canonic = [i for i in canonic if i is not None]
|
378 |
+
set_cannonic = set(canonic)
|
379 |
+
#raise ValueError("Invalid molecule passed to unique@k")
|
380 |
+
return 0 if len(canonic) == 0 else len(set_cannonic) / len(canonic)
|
381 |
|
382 |
def novelty(gen, train, n_jobs=1):
|
383 |
gen_smiles = mapper(n_jobs)(canonic_smiles, gen)
|
|
|
385 |
train_set = set(train)
|
386 |
return 0 if len(gen_smiles_set) == 0 else len(gen_smiles_set - train_set) / len(gen_smiles_set)
|
387 |
|
388 |
+
def internal_diversity(gen):
|
389 |
+
return 1 - average_agg_tanimoto(gen, gen, agg="mean")
|
390 |
|
391 |
def average_agg_tanimoto(stock_vecs, gen_vecs,
|
392 |
batch_size=5000, agg='max',
|