a96123155 commited on
Commit
dedea73
·
1 Parent(s): c79ecae
Files changed (2) hide show
  1. .DS_Store +0 -0
  2. app.py +0 -550
.DS_Store CHANGED
Binary files a/.DS_Store and b/.DS_Store differ
 
app.py DELETED
@@ -1,550 +0,0 @@
1
- import streamlit as st
2
- from io import StringIO
3
- from Bio import SeqIO
4
-
5
- st.title("IRES-LM prediction and mutation")
6
-
7
- # Input sequence
8
- st.subheader("Input sequence")
9
-
10
- seq = st.text_area("FASTA format only", value=">vir_CVB3_ires_00505.1\nTTAAAACAGCCTGTGGGTTGATCCCACCCACAGGCCCATTGGGCGCTAGCACTCTGGTATCACGGTACCTTTGTGCGCCTGTTTTATACCCCCTCCCCCAACTGTAACTTAGAAGTAACACACACCGATCAACAGTCAGCGTGGCACACCAGCCACGTTTTGATCAAGCACTTCTGTTACCCCGGACTGAGTATCAATAGACTGCTCACGCGGTTGAAGGAGAAAGCGTTCGTTATCCGGCCAACTACTTCGAAAAACCTAGTAACACCGTGGAAGTTGCAGAGTGTTTCGCTCAGCACTACCCCAGTGTAGATCAGGTCGATGAGTCACCGCATTCCCCACGGGCGACCGTGGCGGTGGCTGCGTTGGCGGCCTGCCCATGGGGAAACCCATGGGACGCTCTAATACAGACATGGTGCGAAGAGTCTATTGAGCTAGTTGGTAGTCCTCCGGCCCCTGAATGCGGCTAATCCTAACTGCGGAGCACACACCCTCAAGCCAGAGGGCAGTGTGTCGTAACGGGCAACTCTGCAGCGGAACCGACTACTTTGGGTGTCCGTGTTTCATTTTATTCCTATACTGGCTGCTTATGGTGACAATTGAGAGATCGTTACCATATAGCTATTGGATTGGCCATCCGGTGACTAATAGAGCTATTATATATCCCTTTGTTGGGTTTATACCACTTAGCTTGAAAGAGGTTAAAACATTACAATTCATTGTTAAGTTGAATACAGCAAA")
11
- st.subheader("Upload sequence file")
12
- uploaded = st.file_uploader("Sequence file in FASTA format")
13
-
14
- # augments
15
- global output_filename, start_nt_position, end_nt_position, mut_by_prob, transform_type, mlm_tok_num, n_mut, n_designs_ep, n_sampling_designs_ep, n_mlm_recovery_sampling, mutate2stronger
16
- output_filename = st.text_input("output a .csv file", value='IRES_LM_prediction_mutation')
17
- start_nt_position = st.number_input("The start position of the mutation of this sequence, the first position is defined as 0", value=0)
18
- end_nt_position = st.number_input("The last position of the mutation of this sequence, the last position is defined as length(sequence)-1 or -1", value=-1)
19
- mut_by_prob = st.checkbox("Mutated by predicted Probability or Transformed Probability of the sequence", value=True)
20
- transform_type = st.selectbox("Type of probability transformation",
21
- ['', 'sigmoid', 'logit', 'power_law', 'tanh'],
22
- index=2)
23
- mlm_tok_num = st.number_input("Number of masked tokens for each sequence per epoch", value=1)
24
- n_mut = st.number_input("Maximum number of mutations for each sequence", value=3)
25
- n_designs_ep = st.number_input("Number of mutations per epoch", value=10)
26
- n_sampling_designs_ep = st.number_input("Number of sampling mutations from n_designs_ep per epoch", value=5)
27
- n_mlm_recovery_sampling = st.number_input("Number of MLM recovery samplings (with AGCT recovery)", value=1)
28
- mutate2stronger = st.checkbox("Mutate to stronger IRES variant, otherwise mutate to weaker IRES", value=True)
29
-
30
- if not mut_by_prob and transform_type != '':
31
- st.write("--transform_type must be '' when --mut_by_prob is False")
32
- transform_type = ''
33
-
34
-
35
- # Import necessary libraries
36
- # import matplotlib
37
- # import matplotlib.pyplot as plt
38
- import numpy as np
39
- import os
40
- import pandas as pd
41
- # import pathlib
42
- import random
43
- # import scanpy as sc
44
- # import seaborn as sns
45
- import torch
46
- import torch.nn as nn
47
- import torch.nn.functional as F
48
- # from argparse import Namespace
49
- from collections import Counter, OrderedDict
50
- from copy import deepcopy
51
- from esm import Alphabet, FastaBatchedDataset, ProteinBertModel, pretrained, MSATransformer
52
- from esm.data import *
53
- from esm.model.esm2 import ESM2
54
- # from sklearn import preprocessing
55
- # from sklearn.metrics import (confusion_matrix, roc_auc_score, auc,
56
- # precision_recall_fscore_support,
57
- # precision_recall_curve, classification_report,
58
- # roc_auc_score, average_precision_score,
59
- # precision_score, recall_score, f1_score,
60
- # accuracy_score)
61
- # from sklearn.model_selection import StratifiedKFold
62
- # from sklearn.utils import class_weight
63
- # from scipy.stats import spearmanr, pearsonr
64
- from torch import nn
65
- from torch.nn import Linear
66
- from torch.nn.utils.rnn import pad_sequence
67
- from torch.utils.data import Dataset, DataLoader
68
- from tqdm import tqdm, trange
69
-
70
- # Set global variables
71
- # matplotlib.rcParams.update({'font.size': 7})
72
- seed = 19961231
73
- random.seed(seed)
74
- np.random.seed(seed)
75
- torch.manual_seed(seed)
76
- # torch.cuda.manual_seed(seed)
77
- # torch.backends.cudnn.deterministic = True
78
- # torch.backends.cudnn.benchmark = False
79
-
80
-
81
- global idx_to_tok, prefix, epochs, layers, heads, fc_node, dropout_prob, embed_dim, batch_toks, device, repr_layers, evaluation, include, truncate, return_contacts, return_representation, mask_toks_id, finetune
82
-
83
- epochs = 5
84
- layers = 6
85
- heads = 16
86
- embed_dim = 128
87
- batch_toks = 4096
88
- fc_node = 64
89
- dropout_prob = 0.5
90
- folds = 10
91
- repr_layers = [-1]
92
- include = ["mean"]
93
- truncate = True
94
- finetune = False
95
- return_contacts = False
96
- return_representation = False
97
-
98
- device = "cpu"
99
-
100
- global tok_to_idx, idx_to_tok, mask_toks_id
101
- alphabet = Alphabet(mask_prob = 0.15, standard_toks = 'AGCT')
102
- assert alphabet.tok_to_idx == {'<pad>': 0, '<eos>': 1, '<unk>': 2, 'A': 3, 'G': 4, 'C': 5, 'T': 6, '<cls>': 7, '<mask>': 8, '<sep>': 9}
103
-
104
- # tok_to_idx = {'<pad>': 0, '<eos>': 1, '<unk>': 2, 'A': 3, 'G': 4, 'C': 5, 'T': 6, '<cls>': 7, '<mask>': 8, '<sep>': 9}
105
- tok_to_idx = {'-': 0, '&': 1, '?': 2, 'A': 3, 'G': 4, 'C': 5, 'T': 6, '!': 7, '*': 8, '|': 9}
106
- idx_to_tok = {idx: tok for tok, idx in tok_to_idx.items()}
107
- # st.write(tok_to_idx)
108
- mask_toks_id = 8
109
-
110
- global w1, w2, w3
111
- w1, w2, w3 = 1, 1, 100
112
-
113
- class CNN_linear(nn.Module):
114
- def __init__(self):
115
- super(CNN_linear, self).__init__()
116
-
117
- self.esm2 = ESM2(num_layers = layers,
118
- embed_dim = embed_dim,
119
- attention_heads = heads,
120
- alphabet = alphabet)
121
-
122
- self.dropout = nn.Dropout(dropout_prob)
123
- self.relu = nn.ReLU()
124
- self.flatten = nn.Flatten()
125
- self.fc = nn.Linear(in_features = embed_dim, out_features = fc_node)
126
- self.output = nn.Linear(in_features = fc_node, out_features = 2)
127
-
128
- def predict(self, tokens):
129
-
130
- x = self.esm2(tokens, [layers], need_head_weights=False, return_contacts=False, return_representation = True)
131
- x_cls = x["representations"][layers][:, 0]
132
-
133
- o = self.fc(x_cls)
134
- o = self.relu(o)
135
- o = self.dropout(o)
136
- o = self.output(o)
137
-
138
- y_prob = torch.softmax(o, dim = 1)
139
- y_pred = torch.argmax(y_prob, dim = 1)
140
-
141
- if transform_type:
142
- y_prob_transformed = prob_transform(y_prob[:,1])
143
- return y_prob[:,1], y_pred, x['logits'], y_prob_transformed
144
- else:
145
- return y_prob[:,1], y_pred, x['logits'], o[:,1]
146
-
147
- def forward(self, x1, x2):
148
- logit_1, repr_1 = self.predict(x1)
149
- logit_2, repr_2 = self.predict(x2)
150
- return (logit_1, logit_2), (repr_1, repr_2)
151
-
152
- def prob_transform(prob, **kwargs): # Logits
153
- """
154
- Transforms probability values based on the specified method.
155
-
156
- :param prob: torch.Tensor, the input probabilities to be transformed
157
- :param transform_type: str, the type of transformation to be applied
158
- :param kwargs: additional parameters for transformations
159
- :return: torch.Tensor, transformed probabilities
160
- """
161
-
162
- if transform_type == 'sigmoid':
163
- x0 = kwget('x0', 0.5)
164
- k = kwget('k', 10.0)
165
- prob_transformed = 1 / (1 + torch.exp(-k * (prob - x0)))
166
-
167
- elif transform_type == 'logit':
168
- # Adding a small value to avoid log(0) and log(1)
169
- prob_transformed = torch.log(prob + 1e-6) - torch.log(1 - prob + 1e-6)
170
-
171
- elif transform_type == 'power_law':
172
- gamma = kwget('gamma', 2.0)
173
- prob_transformed = torch.pow(prob, gamma)
174
-
175
- elif transform_type == 'tanh':
176
- k = kwget('k', 2.0)
177
- prob_transformed = torch.tanh(k * prob)
178
-
179
- return prob_transformed
180
-
181
- def random_replace(sequence, continuous_replace=False):
182
- if end_nt_position == -1: end_nt_position = len(sequence)
183
- if start_nt_position < 0 or end_nt_position > len(sequence) or start_nt_position > end_nt_position:
184
- # raise ValueError("Invalid start/end positions")
185
- st.write("Invalid start/end positions")
186
- start_nt_position, end_nt_position = 0, -1
187
-
188
- # 将序列切片成三部分:替换区域前、替换区域、替换区域后
189
- pre_segment = sequence[:start_nt_position]
190
- target_segment = list(sequence[start_nt_position:end_nt_position + 1]) # +1因为Python的切片是右开区间
191
- post_segment = sequence[end_nt_position + 1:]
192
-
193
- if not continuous_replace:
194
- # 随机替换目标片段的mlm_tok_num个位置
195
- indices = random.sample(range(len(target_segment)), mlm_tok_num)
196
- for idx in indices:
197
- target_segment[idx] = '*'
198
- else:
199
- # 在目标片段连续替换mlm_tok_num个位置
200
- max_start_idx = len(target_segment) - mlm_tok_num # 确保从i开始的n_mut个元素不会超出目标片段的长度
201
- if max_start_idx < 1: # 如果目标片段长度小于mlm_tok_num,返回原始序列
202
- return target_segment
203
- start_idx = random.randint(0, max_start_idx)
204
- for idx in range(start_idx, start_idx + mlm_tok_num):
205
- target_segment[idx] = '*'
206
-
207
- # 合并并返回最终的序列
208
- return ''.join([pre_segment] + target_segment + [post_segment])
209
-
210
-
211
- def mlm_seq(seq):
212
- seq_token, masked_sequence_token = [7],[7]
213
- seq_token += [tok_to_idx[token] for token in seq]
214
-
215
- masked_seq = random_replace(seq, n_mut) # 随机替换n_mut个元素为'*'
216
- masked_seq_token += [tok_to_idx[token] for token in masked_seq]
217
-
218
- return seq, masked_seq, torch.LongTensor(seq_token), torch.LongTensor(masked_seq_token)
219
-
220
- def batch_mlm_seq(seq_list, continuous_replace = False):
221
- batch_seq = []
222
- batch_masked_seq = []
223
- batch_seq_token_list = []
224
- batch_masked_seq_token_list = []
225
-
226
- for i, seq in enumerate(seq_list):
227
- seq_token, masked_seq_token = [7], [7]
228
- seq_token += [tok_to_idx[token] for token in seq]
229
-
230
- masked_seq = random_replace(seq, continuous_replace) # 随机��换n_mut个元素为'*'
231
- masked_seq_token += [tok_to_idx[token] for token in masked_seq]
232
-
233
- batch_seq.append(seq)
234
- batch_masked_seq.append(masked_seq)
235
-
236
- batch_seq_token_list.append(seq_token)
237
- batch_masked_seq_token_list.append(masked_seq_token)
238
-
239
- return batch_seq, batch_masked_seq, torch.LongTensor(batch_seq_token_list), torch.LongTensor(batch_masked_seq_token_list)
240
-
241
- def recovered_mlm_tokens(masked_seqs, masked_toks, esm_logits, exclude_low_prob = False):
242
- # Only remain the AGCT logits
243
- esm_logits = esm_logits[:,:,3:7]
244
- # Get the predicted tokens using argmax
245
- predicted_toks = (esm_logits.argmax(dim=-1)+3).tolist()
246
-
247
- batch_size, seq_len, vocab_size = esm_logits.size()
248
- if exclude_low_prob: min_prob = 1 / vocab_size
249
- # Initialize an empty list to store the recovered sequences
250
- recovered_sequences, recovered_toks = [], []
251
-
252
- for i in range(batch_size):
253
- recovered_sequence_i, recovered_tok_i = [], []
254
- for j in range(seq_len):
255
- if masked_toks[i][j] == 8:
256
- st.write(i,j)
257
- ### Sample M recovery sequences using the logits
258
- recovery_probs = torch.softmax(esm_logits[i, j], dim=-1)
259
- recovery_probs[predicted_toks[i][j]-3] = 0 # Exclude the most probable token
260
- if exclude_low_prob: recovery_probs[recovery_probs < min_prob] = 0 # Exclude tokens with low probs < min_prob
261
- recovery_probs /= recovery_probs.sum() # Normalize the probabilities
262
-
263
- ### 有放回抽样
264
- max_retries = 5
265
- retries = 0
266
- success = False
267
-
268
- while retries < max_retries and not success:
269
- try:
270
- recovery_indices = list(np.random.choice(vocab_size, size=n_mlm_recovery_sampling, p=recovery_probs.cpu().detach().numpy(), replace=False))
271
- success = True # 设置成功标志
272
- except ValueError as e:
273
- retries += 1
274
- st.write(f"Attempt {retries} failed with error: {e}")
275
- if retries >= max_retries:
276
- st.write("Max retries reached. Skipping this iteration.")
277
-
278
- ### recovery to sequence
279
- if retries < max_retries:
280
- for idx in [predicted_toks[i][j]] + [3+i for i in recovery_indices]:
281
- recovery_seq = deepcopy(list(masked_seqs[i]))
282
- recovery_tok = deepcopy(masked_toks[i])
283
-
284
- recovery_tok[j] = idx
285
- recovery_seq[j-1] = idx_to_tok[idx]
286
-
287
- recovered_tok_i.append(recovery_tok)
288
- recovered_sequence_i.append(''.join(recovery_seq))
289
-
290
- recovered_sequences.extend(recovered_sequence_i)
291
- recovered_toks.extend(recovered_tok_i)
292
- return recovered_sequences, torch.LongTensor(torch.stack(recovered_toks))
293
-
294
- def recovered_mlm_multi_tokens(masked_seqs, masked_toks, esm_logits, exclude_low_prob = False):
295
- # Only remain the AGCT logits
296
- esm_logits = esm_logits[:,:,3:7]
297
- # Get the predicted tokens using argmax
298
- predicted_toks = (esm_logits.argmax(dim=-1)+3).tolist()
299
-
300
- batch_size, seq_len, vocab_size = esm_logits.size()
301
- if exclude_low_prob: min_prob = 1 / vocab_size
302
- # Initialize an empty list to store the recovered sequences
303
- recovered_sequences, recovered_toks = [], []
304
-
305
- for i in range(batch_size):
306
- recovered_sequence_i, recovered_tok_i = [], []
307
- recovered_masked_num = 0
308
- for j in range(seq_len):
309
- if masked_toks[i][j] == 8:
310
- ### Sample M recovery sequences using the logits
311
- recovery_probs = torch.softmax(esm_logits[i, j], dim=-1)
312
- recovery_probs[predicted_toks[i][j]-3] = 0 # Exclude the most probable token
313
- if exclude_low_prob: recovery_probs[recovery_probs < min_prob] = 0 # Exclude tokens with low probs < min_prob
314
- recovery_probs /= recovery_probs.sum() # Normalize the probabilities
315
-
316
- ### 有放回抽样
317
- max_retries = 5
318
- retries = 0
319
- success = False
320
-
321
- while retries < max_retries and not success:
322
- try:
323
- recovery_indices = list(np.random.choice(vocab_size, size=n_mlm_recovery_sampling, p=recovery_probs.cpu().detach().numpy(), replace=False))
324
- success = True # 设置成功标志
325
- except ValueError as e:
326
- retries += 1
327
- st.write(f"Attempt {retries} failed with error: {e}")
328
- if retries >= max_retries:
329
- st.write("Max retries reached. Skipping this iteration.")
330
-
331
- ### recovery to sequence
332
-
333
- if recovered_masked_num == 0:
334
- if retries < max_retries:
335
- for idx in [predicted_toks[i][j]] + [3+i for i in recovery_indices]:
336
- recovery_seq = deepcopy(list(masked_seqs[i]))
337
- recovery_tok = deepcopy(masked_toks[i])
338
-
339
- recovery_tok[j] = idx
340
- recovery_seq[j-1] = idx_to_tok[idx]
341
-
342
- recovered_tok_i.append(recovery_tok)
343
- recovered_sequence_i.append(''.join(recovery_seq))
344
-
345
- elif recovered_masked_num > 0:
346
- if retries < max_retries:
347
- for idx in [predicted_toks[i][j]] + [3+i for i in recovery_indices]:
348
- for recovery_seq, recovery_tok in zip(list(recovered_sequence_i), list(recovered_tok_i)): # 要在循环开始之前获取列表的副本来进行迭代。这样,在循环中即使我们修改了原始的列表,也不会影响迭代的行为。
349
-
350
- recovery_seq_temp = list(recovery_seq)
351
- recovery_tok[j] = idx
352
- recovery_seq_temp[j-1] = idx_to_tok[idx]
353
-
354
- recovered_tok_i.append(recovery_tok)
355
- recovered_sequence_i.append(''.join(recovery_seq_temp))
356
-
357
- recovered_masked_num += 1
358
- recovered_indices = [i for i, s in enumerate(recovered_sequence_i) if '*' not in s]
359
- recovered_tok_i = [recovered_tok_i[i] for i in recovered_indices]
360
- recovered_sequence_i = [recovered_sequence_i[i] for i in recovered_indices]
361
-
362
- recovered_sequences.extend(recovered_sequence_i)
363
- recovered_toks.extend(recovered_tok_i)
364
-
365
- recovered_sequences, recovered_toks = remove_duplicates_double(recovered_sequences, recovered_toks)
366
-
367
- return recovered_sequences, torch.LongTensor(torch.stack(recovered_toks))
368
-
369
- def mismatched_positions(s1, s2):
370
- # 这个函数假定两个字符串的长度相同。
371
- """Return the number of positions where two strings differ."""
372
-
373
- # The number of mismatches will be the sum of positions where characters are not the same
374
- return sum(1 for c1, c2 in zip(s1, s2) if c1 != c2)
375
-
376
- def remove_duplicates_triple(filtered_mut_seqs, filtered_mut_probs, filtered_mut_logits):
377
- seen = {}
378
- unique_seqs = []
379
- unique_probs = []
380
- unique_logits = []
381
-
382
- for seq, prob, logit in zip(filtered_mut_seqs, filtered_mut_probs, filtered_mut_logits):
383
- if seq not in seen:
384
- unique_seqs.append(seq)
385
- unique_probs.append(prob)
386
- unique_logits.append(logit)
387
- seen[seq] = True
388
-
389
- return unique_seqs, unique_probs, unique_logits
390
-
391
- def remove_duplicates_double(filtered_mut_seqs, filtered_mut_probs):
392
- seen = {}
393
- unique_seqs = []
394
- unique_probs = []
395
-
396
- for seq, prob in zip(filtered_mut_seqs, filtered_mut_probs):
397
- if seq not in seen:
398
- unique_seqs.append(seq)
399
- unique_probs.append(prob)
400
- seen[seq] = True
401
-
402
- return unique_seqs, unique_probs
403
-
404
- def mutated_seq(wt_seq, wt_label):
405
- wt_seq = '!'+ wt_seq
406
- wt_tok = torch.LongTensor([[tok_to_idx[token] for token in wt_seq]]).to(device)
407
- wt_prob, wt_pred, _, wt_logit = model.predict(wt_tok)
408
-
409
- st.write(f'Wild Type: Length = ', len(wt_seq), '\n', wt_seq)
410
- st.write(f'Wild Type: Label = {wt_label}, Y_pred = {wt_pred.item()}, Y_prob = {wt_prob.item():.2%}')
411
-
412
- # st.write(n_mut, mlm_tok_num, n_designs_ep, n_sampling_designs_ep, n_mlm_recovery_sampling, mutate2stronger)
413
- # pbar = tqdm(total=n_mut)
414
- mutated_seqs = []
415
- i = 1
416
- pbar = st.progress(i, text="mutated number of sequence")
417
- while i <= n_mut:
418
- if i == 1: seeds_ep = [wt_seq[1:]]
419
- seeds_next_ep, seeds_probs_next_ep, seeds_logits_next_ep = [], [], []
420
- for seed in seeds_ep:
421
- seed_seq, masked_seed_seq, seed_seq_token, masked_seed_seq_token = batch_mlm_seq([seed] * n_designs_ep, continuous_replace = True) ### mask seed with 1 site to "*"
422
-
423
- seed_prob, seed_pred, _, seed_logit = model.predict(seed_seq_token[0].unsqueeze_(0).to(device))
424
- _, _, seed_esm_logit, _ = model.predict(masked_seed_seq_token.to(device))
425
- mut_seqs, mut_toks = recovered_mlm_multi_tokens(masked_seed_seq, masked_seed_seq_token, seed_esm_logit)
426
- mut_probs, mut_preds, mut_esm_logits, mut_logits = model.predict(mut_toks.to(device))
427
-
428
- ### Filter mut_seqs that mut_prob < seed_prob and mut_prob < wild_prob
429
- filtered_mut_seqs = []
430
- filtered_mut_probs = []
431
- filtered_mut_logits = []
432
- if mut_by_prob:
433
- for z in range(len(mut_seqs)):
434
- if mutate2stronger:
435
- if mut_probs[z] >= seed_prob and mut_probs[z] >= wt_prob:
436
- filtered_mut_seqs.append(mut_seqs[z])
437
- filtered_mut_probs.append(mut_probs[z].cpu().detach().numpy())
438
- filtered_mut_logits.append(mut_logits[z].cpu().detach().numpy())
439
- else:
440
- if mut_probs[z] < seed_prob and mut_probs[z] < wt_prob:
441
- filtered_mut_seqs.append(mut_seqs[z])
442
- filtered_mut_probs.append(mut_probs[z].cpu().detach().numpy())
443
- filtered_mut_logits.append(mut_logits[z].cpu().detach().numpy())
444
- else:
445
- for z in range(len(mut_seqs)):
446
- if mutate2stronger:
447
- if mut_logits[z] >= seed_logit and mut_logits[z] >= wt_logit:
448
- filtered_mut_seqs.append(mut_seqs[z])
449
- filtered_mut_probs.append(mut_probs[z].cpu().detach().numpy())
450
- filtered_mut_logits.append(mut_logits[z].cpu().detach().numpy())
451
- else:
452
- if mut_logits[z] < seed_logit and mut_logits[z] < wt_logit:
453
- filtered_mut_seqs.append(mut_seqs[z])
454
- filtered_mut_probs.append(mut_probs[z].cpu().detach().numpy())
455
- filtered_mut_logits.append(mut_logits[z].cpu().detach().numpy())
456
-
457
-
458
-
459
- ### Save
460
- seeds_next_ep.extend(filtered_mut_seqs)
461
- seeds_probs_next_ep.extend(filtered_mut_probs)
462
- seeds_logits_next_ep.extend(filtered_mut_logits)
463
- seeds_next_ep, seeds_probs_next_ep, seeds_logits_next_ep = remove_duplicates_triple(seeds_next_ep, seeds_probs_next_ep, seeds_logits_next_ep)
464
-
465
- ### Sampling based on prob
466
- if len(seeds_next_ep) > n_sampling_designs_ep:
467
- seeds_probs_next_ep_norm = seeds_probs_next_ep / sum(seeds_probs_next_ep) # Normalize the probabilities
468
- seeds_index_next_ep = np.random.choice(len(seeds_next_ep), n_sampling_designs_ep, p = seeds_probs_next_ep_norm, replace = False)
469
-
470
- seeds_next_ep = np.array(seeds_next_ep)[seeds_index_next_ep]
471
- seeds_probs_next_ep = np.array(seeds_probs_next_ep)[seeds_index_next_ep]
472
- seeds_logits_next_ep = np.array(seeds_logits_next_ep)[seeds_index_next_ep]
473
- seeds_mutated_num_next_ep = [mismatched_positions(wt_seq[1:], s) for s in seeds_next_ep]
474
-
475
- mutated_seqs.extend(list(zip(seeds_next_ep, seeds_logits_next_ep, seeds_probs_next_ep, seeds_mutated_num_next_ep)))
476
-
477
- seeds_ep = seeds_next_ep
478
- i += 1
479
- # pbar.update(1)
480
- pbar.progress(i/n_mut, text="Mutating")
481
- # pbar.close()
482
- st.success('Done', icon="✅")
483
- mutated_seqs.extend([(wt_seq[1:], wt_logit.item(), wt_prob.item(), 0)])
484
- mutated_seqs = sorted(mutated_seqs, key=lambda x: x[2], reverse=True)
485
- mutated_seqs = pd.DataFrame(mutated_seqs, columns = ['mutated_seq', 'predicted_logit', 'predicted_probability', 'mutated_num']).drop_duplicates('mutated_seq')
486
- return mutated_seqs
487
-
488
- def read_raw(raw_input):
489
- ids = []
490
- sequences = []
491
-
492
- file = StringIO(raw_input)
493
- for record in SeqIO.parse(file, "fasta"):
494
-
495
- # 检查序列是否只包含A, G, C, T
496
- sequence = str(record.seq.back_transcribe()).upper()
497
- if not set(sequence).issubset(set("AGCT")):
498
- st.write(f"Record '{record.description}' was skipped for containing invalid characters. Only A, G, C, T(U) are allowed.")
499
- continue
500
-
501
- # 将符合条件的序列添加到列表中
502
- ids.append(record.id)
503
- sequences.append(sequence)
504
-
505
- return ids, sequences
506
-
507
- def predict_raw(raw_input):
508
- state_dict = torch.load('model.pt', map_location=torch.device(device))
509
- new_state_dict = OrderedDict()
510
-
511
- for k, v in state_dict.items():
512
- name = k.replace('module.','')
513
- new_state_dict[name] = v
514
-
515
- model = CNN_linear().to(device)
516
- model.load_state_dict(new_state_dict, strict = False)
517
- model.eval()
518
- st.write(model)
519
- # st.write('====Parse Input====')
520
- ids, seqs = read_raw(raw_input)
521
-
522
- # st.write('====Predict====')
523
- res_pd = pd.DataFrame()
524
- for wt_seq, wt_id in zip(seqs, ids):
525
- try:
526
- st.write(wt_id, wt_seq)
527
- res = mutated_seq(wt_seq, wt_id)
528
- st.write(res)
529
- res_pd.append(res)
530
- except:
531
- st.write('====Please Try Again this sequence: ', wt_id, wt_seq)
532
- # st.write(pred)
533
- return res_pd
534
-
535
- # Run
536
- if st.button("Predict and Mutate"):
537
- if uploaded:
538
- result = predict_raw(uploaded.getvalue().decode())
539
- else:
540
- result = predict_raw(seq)
541
-
542
- result_file = result.to_csv(index=False)
543
- st.download_button("Download", result_file, file_name=output_filename+".csv")
544
- st.dataframe(result)
545
-
546
-
547
-
548
-
549
-
550
-