-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathinference_multimer.py
311 lines (239 loc) · 14 KB
/
inference_multimer.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
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
import os, sys, argparse, time
from multiprocessing import Pool
import numpy as np
import pandas as pd
from gate.tool.utils import *
from gate.feature.feature_generation import *
from gate.sample.sample import *
from gate.feature.config import *
from gate.model.build_graph import *
from torch.utils.data import Dataset, DataLoader
from gate.model.graph_transformer_v3 import Gate
def cal_average_score(dfs):
prev_df = None
for i in range(len(dfs)):
curr_df = dfs[i].add_suffix(f"{i + 1}")
curr_df['model'] = curr_df[f'model{i + 1}']
curr_df = curr_df.drop([f'model{i + 1}'], axis=1)
if prev_df is None:
prev_df = curr_df
else:
prev_df = prev_df.merge(curr_df, on=f'model', how="inner")
# print(prev_df)
avg_scores = []
for i in range(len(prev_df)):
sum_score = 0
for j in range(len(dfs)):
sum_score += prev_df.loc[i, f"score{j+1}"]
avg_scores += [sum_score/len(dfs)]
models = prev_df['model']
ensemble_df = pd.DataFrame({'model': models, 'score': avg_scores})
ensemble_df = ensemble_df.sort_values(by='score', ascending=False)
ensemble_df.reset_index(inplace=True, drop=True)
return ensemble_df
class DGLData(Dataset):
"""Data loader"""
def __init__(self, dgl_folder):
self.dgl_folder = dgl_folder
self.data = []
self.data_list = []
self._prepare()
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
return self.data[idx], self.data_list[idx]
def _prepare(self):
for dgl_file in os.listdir(self.dgl_folder):
g, tmp = dgl.data.utils.load_graphs(os.path.join(self.dgl_folder, dgl_file))
self.data.append(g[0])
self.data_list.append(os.path.join(self.dgl_folder, dgl_file))
def collate(samples):
"""Customer collate function"""
graphs, data_paths = zip(*samples)
batched_graphs = dgl.batch(graphs)
return batched_graphs, data_paths
def cli_main():
parser = argparse.ArgumentParser()
parser.add_argument('--fasta_path', type=str, required=True)
parser.add_argument('--input_model_dir', type=str, required=True)
parser.add_argument('--pkldir', type=str, default="", required=False)
parser.add_argument('--output_dir', type=str, required=True)
parser.add_argument('--use_af_feature', default=False, type=lambda x: (str(x).lower() == 'true'))
parser.add_argument('--sample_times', default=5, type=int)
args = parser.parse_args()
device = torch.device('cuda') # set cuda device
features_multimer = features_multimer_dict()
complex_feature_generation(fasta_path=args.fasta_path,
input_model_dir=args.input_model_dir,
output_dir=args.output_dir,
config=CONFIG,
use_alphafold_features=args.use_af_feature,
pkldir=args.pkldir,
features_multimer=features_multimer)
to_be_average_dfs = []
for i in range(args.sample_times):
sample_i_output_dir = os.path.join(args.output_dir, 'workdir' + str(i))
os.makedirs(sample_i_output_dir, exist_ok=True)
sample_dir = os.path.join(sample_i_output_dir, 'sample')
os.makedirs(sample_dir, exist_ok=True)
print("Start to sample subgraphs......")
if not os.path.exists(os.path.join(sample_i_output_dir, 'sample.done')):
model_to_cluster = sample_models_by_kmeans(pairwise_usalign_file=features_multimer.pairwise_usalign,
pairwise_mmalign_file=features_multimer.pairwise_mmalign,
pairwise_qsscore_file=features_multimer.pairwise_qsscore,
pairwise_dockq_wave_file=features_multimer.pairwise_dockq_wave,
pairwise_dockq_ave_file=features_multimer.pairwise_dockq_ave,
pairwise_cad_score_file=features_multimer.pairwise_cad_score,
sample_number_per_target=3000,
outdir=sample_dir)
with open(os.path.join(sample_i_output_dir, 'cluster.txt'), 'w') as fw:
for modelname in model_to_cluster:
fw.write(f"{modelname}\t{model_to_cluster[modelname]}\n")
os.system(f"touch {os.path.join(sample_i_output_dir, 'sample.done')}")
print("Generating dgl files for the subgraphs.....")
dgl_base_dir = os.path.join(sample_i_output_dir, 'dgl')
configs = {}
if args.use_af_feature:
configs = {
'v8': {'use_af_feature': args.use_af_feature,
'use_gcpnet_ema': True,
'use_interface_pairwise': True},
}
else:
configs = {
'casp15_official_v7_nogcpnet': {'use_af_feature': args.use_af_feature,
'use_gcpnet_ema': False,
'use_interface_pairwise': False},
'casp15_official_v7_gcpnet': {'use_af_feature': args.use_af_feature,
'use_gcpnet_ema': True,
'use_interface_pairwise': False},
'casp15_official_v8': {'use_af_feature': args.use_af_feature,
'use_gcpnet_ema': True,
'use_interface_pairwise': True},
}
for config in configs:
dgl_dir = os.path.join(dgl_base_dir, config)
os.makedirs(dgl_dir, exist_ok=True)
if os.path.exists(os.path.join(dgl_base_dir, config + '.done')):
continue
generate_multimer_dgls(sample_dir=sample_dir,
dgl_dir=dgl_dir,
features_multimer=features_multimer,
sim_threshold=0.5,
use_af_feature=configs[config]['use_af_feature'],
use_gcpnet_ema=configs[config]['use_gcpnet_ema'],
use_interface_pairwise=configs[config]['use_interface_pairwise'])
if len(os.listdir(dgl_dir)) == 3000:
os.system(f"touch {os.path.join(dgl_base_dir, config + '.done')}")
print("Generating predictions for the subgraphs.....")
prediction_base_dir = os.path.join(sample_i_output_dir, 'prediction')
gate_model_names, dgl_dirs = [], []
if args.use_af_feature:
# gate_model_names.append('casp15_inhouse_full_v7_nogcpnet')
# dgl_dirs.append(os.path.join(dgl_base_dir, "v7_nogcpnet"))
# gate_model_names.append('casp15_inhouse_full_v7_gcpnet')
# dgl_dirs.append(os.path.join(dgl_base_dir, "v7_gcpnet"))
# gate_model_names.append('casp15_inhouse_full_v8')
# dgl_dirs.append(os.path.join(dgl_base_dir, "v8"))
# gate_model_names.append('casp15_inhouse_top_v7_nogcpnet')
# dgl_dirs.append(os.path.join(dgl_base_dir, "v7_nogcpnet"))
# gate_model_names.append('casp15_inhouse_top_v7_gcpnet')
# dgl_dirs.append(os.path.join(dgl_base_dir, "v7_gcpnet"))
gate_model_names.append('casp15_inhouse_top_v8')
dgl_dirs.append(os.path.join(dgl_base_dir, "v8"))
else:
gate_model_names.append('casp15_official_v7_nogcpnet')
dgl_dirs.append(os.path.join(dgl_base_dir, "casp15_official_v7_nogcpnet"))
gate_model_names.append('casp15_official_v7_gcpnet')
dgl_dirs.append(os.path.join(dgl_base_dir, "casp15_official_v7_gcpnet"))
gate_model_names.append('casp15_official_v8')
dgl_dirs.append(os.path.join(dgl_base_dir, "casp15_official_v8"))
ensemble_dfs = []
for gate_model_name, dgl_dir in zip(gate_model_names, dgl_dirs):
print(dgl_dir)
test_data = DGLData(dgl_folder=dgl_dir)
prediction_dir = os.path.join(prediction_base_dir, gate_model_name)
os.makedirs(prediction_dir, exist_ok=True)
prediction_dfs = []
for fold in range(10):
if os.path.exists(os.path.join(prediction_dir, f'fold{fold}.csv')):
print(f"Prediction for fold{fold} has been generated!")
df = pd.read_csv(os.path.join(prediction_dir, f'fold{fold}.csv'))
prediction_dfs += [df]
continue
fold_model_config = GATE_MODELS[gate_model_name]['fold' + str(fold)]
test_loader = DataLoader(test_data,
#batch_size=fold_model_config.batch_size,
batch_size=128,
num_workers=4,
pin_memory=True,
collate_fn=collate,
shuffle=False)
model = Gate(node_input_dim=fold_model_config.node_input_dim,
edge_input_dim=fold_model_config.edge_input_dim,
num_heads=fold_model_config.num_heads,
num_layer=fold_model_config.num_layer,
dp_rate=fold_model_config.dp_rate,
layer_norm=fold_model_config.layer_norm,
batch_norm=not fold_model_config.layer_norm,
residual=True,
hidden_dim=fold_model_config.hidden_dim,
mlp_dp_rate=fold_model_config.mlp_dp_rate)
model = model.load_from_checkpoint(os.path.join(CKPTDIR, f"{gate_model_name}_fold{fold}.ckpt"))
model = model.to(device)
model.eval()
target_pred_subgraph_scores = {}
for idx, (batch_graphs, data_paths) in enumerate(test_loader):
batch_x = batch_graphs.ndata['f'].to(torch.float)
batch_e = batch_graphs.edata['f'].to(torch.float)
batch_graphs = batch_graphs.to(device)
batch_x = batch_x.to(device)
batch_e = batch_e.to(device)
batch_scores = model.forward(batch_graphs, batch_x, batch_e)
pred_scores = batch_scores.cpu().data.numpy().squeeze(1)
start_idx = 0
for subgraph_path in data_paths:
subgraph_filename = os.path.basename(subgraph_path)
subgraph_df = pd.read_csv(f"{sample_dir}/{subgraph_filename.replace('.dgl', '.csv')}", index_col=[0])
for i, modelname in enumerate(subgraph_df.columns):
if modelname not in target_pred_subgraph_scores:
target_pred_subgraph_scores[modelname] = []
target_pred_subgraph_scores[modelname] += [pred_scores[start_idx + i]]
start_idx += len(subgraph_df.columns)
ensemble_scores, ensemble_count, std, normalized_std = [], [], [], []
for modelname in target_pred_subgraph_scores:
mean_score = np.mean(np.array(target_pred_subgraph_scores[modelname]))
median_score = np.median(np.array(target_pred_subgraph_scores[modelname]))
if fold_model_config.ensemble_mode == "mean":
ensemble_scores += [mean_score]
else:
ensemble_scores += [median_score]
ensemble_count += [len(target_pred_subgraph_scores[modelname])]
std += [np.std(np.array(target_pred_subgraph_scores[modelname]))]
normalized_std += [np.std(np.array(target_pred_subgraph_scores[modelname])) / mean_score]
df = pd.DataFrame({'model': list(target_pred_subgraph_scores.keys()),
'score': ensemble_scores,
'sample_count': ensemble_count,
'std': std,
"std_norm": normalized_std})
df.to_csv(os.path.join(prediction_dir, f'fold{fold}.csv'))
prediction_dfs += [df]
ensemble_df = cal_average_score(prediction_dfs)
ensemble_df.to_csv(os.path.join(prediction_dir, 'ensemble.csv'))
ensemble_dfs += [ensemble_df]
if args.use_af_feature:
to_be_average_dfs += [ensemble_dfs[0]]
else:
non_af_ensemble_df = cal_average_score(ensemble_dfs)
non_af_ensemble_df.to_csv(os.path.join(sample_i_output_dir, 'ensemble.csv'))
to_be_average_dfs += [non_af_ensemble_df]
final_ensemble_df = cal_average_score(to_be_average_dfs)
resultfile = os.path.join(args.output_dir, 'ensemble_af.csv') if args.use_af_feature else os.path.join(args.output_dir, 'ensemble_nonaf.csv')
final_ensemble_df.to_csv(resultfile)
# create a summary csv for all the scores in GATE
summary_df = generate_feature_summary(workdir=os.path.join(args.output_dir, 'feature'),
gate_prediction_df=final_ensemble_df,
use_af_feature=args.use_af_feature)
summary_df.to_csv(os.path.join(args.output_dir, 'gate_af_summary.csv') if args.use_af_feature else os.path.join(args.output_dir, 'gate_nonaf_summary.csv'))
if __name__ == '__main__':
cli_main()