-
Notifications
You must be signed in to change notification settings - Fork 0
/
cgan_tagmol.py
228 lines (186 loc) · 8.95 KB
/
cgan_tagmol.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
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
###################################################################
# This is an official PyTorch implementation for Target-specific
# Generation of Molecules (TagMol)
# Author: Junde Li, The Pennsylvania State University
# Date: Mar 25, 2023
###################################################################
import os
import time
import argparse
from tqdm import tqdm
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import torch.autograd as autograd
import torch.utils.data as Data
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
import pickle
from rdkit import Chem
from dataloader import PDBbindPLDataset, Normalize, RandomRotateJitter, ToTensor
from model import PointNetEncoder, Generator, Discriminator, m_Discriminator
import itertools
from utils import *
from frechetdist import frdist
import csv
import warnings
warnings.filterwarnings("ignore")
# --------------------------
# Hyperparameters
# --------------------------
lr = 3e-5
batch_size = 64
max_epoch = 2000
num_workers = 2
ligand_size = 14
x_dim = 64 # standard GAN
z_dim = 8
n_pc_points = 4096
g_conv_dims = [64, 256, 1024]
d_conv_dim = [[128, 64], 128, [128, 64]]
node_dim = 64
n_atom_types = 7
n_bond_types = 5
dropout = 0.2
dataset = 'refined'
n_critic = 5
lambda_gp = 10
alpha_l2 = 1e-3
beta_le = 1e-5
gamma_lr = 1e-4
save_step = 500
atom_decoder = {0: 0, 1: 6, 2: 7, 3: 8, 4: 9, 5: 16, 6:17}
bond_decoder = {0: Chem.rdchem.BondType.ZERO,
1: Chem.rdchem.BondType.SINGLE,
2: Chem.rdchem.BondType.DOUBLE,
3: Chem.rdchem.BondType.TRIPLE,
4: Chem.rdchem.BondType.AROMATIC}
name = f"model/tagmol-ligsize/{ligand_size}"
log_dir = f"{name}"
models_dir = f"{name}/{dataset}_saved_models"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
cuda = True if torch.cuda.is_available() else False
Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor
os.makedirs(log_dir, exist_ok=True)
os.makedirs(models_dir, exist_ok=True)
proteins = pickle.load(open(f'data/pdbbind/{dataset}-set/proteins.p', "rb"))
atoms = pickle.load(open(f'data/pdbbind/{dataset}-set/atoms.p', "rb"))
bonds = pickle.load(open(f'data/pdbbind/{dataset}-set/bonds.p', "rb"))
train_dataset = Data.TensorDataset(proteins, atoms, bonds)
train_loader = DataLoader(train_dataset, batch_size=batch_size,
shuffle=True, drop_last=False)
# Initialize models and make optimizers.
encoder = PointNetEncoder(x_dim, channel=4, feature_transform=True)
generator = Generator(x_dim, z_dim, g_conv_dims, ligand_size, n_atom_types, n_bond_types)
discriminator = m_Discriminator(d_conv_dim, n_atom_types, n_bond_types, dropout)
# discriminator = Discriminator(c_in=n_atom_types, c_out=node_dim, c_hidden=32, n_relations=n_bond_types, n_layers=3)
opt_enc_gen = torch.optim.Adam(itertools.chain(encoder.parameters(), generator.parameters()), lr, (0.9, 0.999))
opt_disc = torch.optim.Adam(discriminator.parameters(), lr, (0.9, 0.999))
if cuda:
encoder.cuda()
generator.cuda()
discriminator.cuda()
start_epoch = 0
if start_epoch:
for model in [generator, discriminator]:
path = os.path.join(models_dir, f'{ligand_size}-{start_epoch}-{str(type(model)).split(".")[-1][:-2]}.ckpt')
model.load_state_dict(torch.load(path))
model.train()
print(f'Loaded model checkpoints from {models_dir}.')
def main():
# train loop
print('Start traning...')
for epoch in tqdm(range(start_epoch, max_epoch), desc='total progress'):
losses_D = []
losses_G = []
fd_scores = []
cur_time = time.strftime("%D %H:%M:%S", time.localtime())
epoch_log = f"{cur_time}\tepoch {epoch+1}\t"
for batch, sample_batched in enumerate(train_loader):
protein = sample_batched[0]
r_atoms, r_bonds = sample_batched[1:]
batch_log = f"{epoch+1}:{batch}\n"
bs = protein.size(0)
# -----------------------
# Train Discriminator
# -----------------------
opt_disc.zero_grad()
# Encode protein features using encoder.
x = encoder(protein.transpose(2, 1))
# Sample noise as generator input.
z = Variable(Tensor(np.random.normal(0, 1, (bs, z_dim))))
f_atoms, f_bonds = generator(z, x)
# Hard categorical sampling fake ligands from probabilistic distribution.
f_atoms = F.gumbel_softmax(f_atoms, tau=1, hard=True)
f_bonds = F.gumbel_softmax(f_bonds, tau=1, hard=True)
# Validity for real and fake samples.
r_validity = discriminator(r_bonds, None, r_atoms)
f_validity = discriminator(f_bonds, None, f_atoms)
# Calculate adient penalty.
gradient_penalty = compute_gradient_penalty(discriminator, r_atoms, r_bonds, f_atoms, f_bonds, True)
# Adversarial loss plus gradient penalty.
loss_D = -torch.mean(r_validity) + torch.mean(f_validity) + lambda_gp * gradient_penalty
losses_D.append(loss_D.item())
loss_D.backward()
opt_disc.step()
batch_log += f"loss_d: {-torch.mean(r_validity).item():.2f}\t" + \
f"{torch.mean(f_validity).item():.2f}\t{lambda_gp * gradient_penalty.item():.2f}\n"
# Train other networks every n_critic steps
if (batch+1) % n_critic == 0:
# -------------------------------
# Train Generator and Encoder
# -------------------------------
opt_enc_gen.zero_grad()
if x_dim:
x = encoder(protein.transpose(2, 1))
z = Variable(Tensor(np.random.normal(0, 1, (batch_size, z_dim))))
f_atoms, f_bonds = generator(z, x)
# Hard categorical sampling fake ligands from probabilistic distribution.
f_atoms = F.gumbel_softmax(f_atoms, tau=1, hard=True)
f_bonds = F.gumbel_softmax(f_bonds, tau=1, hard=True)
# Validity for fake samples.
f_validity = discriminator(f_bonds, None, f_atoms)
loss_G_fake = - torch.mean(f_validity)
loss_G = loss_G_fake
losses_G.append(loss_G.item())
loss_G.backward()
opt_enc_gen.step()
r_dist =[list(r_atoms[i].reshape(-1).cpu().detach().numpy()) \
+ list(r_bonds[i].reshape(-1).cpu().detach().numpy()) for i in range(batch_size)]
f_dist =[list(f_atoms[i].reshape(-1).cpu().detach().numpy()) \
+ list(f_bonds[i].reshape(-1).cpu().detach().numpy()) for i in range(batch_size)]
fd_bond_atom = frdist(r_dist, f_dist)
fd_scores.append(fd_bond_atom)
# Saving model checkpoint with lowest FD score
if "fd_bond_atom_min" not in locals():
fd_bond_atom_min = 30
if fd_bond_atom_min > fd_bond_atom:
if "lowest_ind" not in locals():
lowest_ind = 0
if lowest_ind:
for model in [encoder, generator, discriminator]:
os.remove(os.path.join(models_dir, \
f'{ligand_size}-{lowest_ind}-{str(type(model)).split(".")[-1][:-2]}.ckpt'))
lowest_ind = epoch+1
fd_bond_atom_min = fd_bond_atom
for model in [encoder, generator, discriminator]:
path = os.path.join(models_dir, \
f'{ligand_size}-{epoch+1}-{str(type(model)).split(".")[-1][:-2]}.ckpt')
torch.save(model.state_dict(), path)
with open(os.path.join(models_dir, 'lowest_indices.csv'), 'a') as file:
writer = csv.writer(file)
writer.writerow([epoch+1] + [fd_bond_atom])
batch_log += f"loss_g: {loss_G_fake.item():.2f}\n"
print_and_save(batch_log, f"{log_dir}/batch-log.txt")
epoch_log += f"loss_d:{np.mean(losses_D):.4f}\t loss_g:{np.mean(losses_G):.4f}\t fd:{np.mean(fd_scores):.4f}\t"
print_and_save(epoch_log, f"{log_dir}/epoch-log.txt")
# Save model checkpoints.
if (epoch+1) % save_step == 0:
for model in [encoder, generator, discriminator]:
path = os.path.join(models_dir, f'{ligand_size}-{epoch+1}-{str(type(model)).split(".")[-1][:-2]}.ckpt')
torch.save(model.state_dict(), path)
print(f'Saved model checkpoints into {models_dir}...')
if __name__ == '__main__':
main()