-
Notifications
You must be signed in to change notification settings - Fork 13
/
Copy pathprottrans_models.py
208 lines (178 loc) · 8.39 KB
/
prottrans_models.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
"""
A module that works with ProtTrans models. Functions were
adapted from ProtTrans authors' Google Colab notebook
"""
from transformers import T5EncoderModel, T5Tokenizer
import torch
import os
import sys
from hashlib import sha256
def get_pretrained_model(model_path):
"""
Fetches the model accordingly to the model_path
model_path - STRING that identifies the model to fetch
Returns model.
"""
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
if(os.path.exists(model_path+'/pytorch_model.bin') and
os.path.exists(model_path+'/config.json')):
model = T5EncoderModel.from_pretrained(model_path+'/pytorch_model.bin',
config=model_path+'/config.json')
else:
model = T5EncoderModel.from_pretrained(model_path)
model = model.to(device)
model = model.eval()
return model
def get_tokenizer(model_path):
"""
Fetches the tokenizer accordingly to the model_path
model_path - STRING that identifies the model whose tokenizer
should be fetched
returns tokenizer
"""
tokenizer = T5Tokenizer.from_pretrained(model_path, do_lower_case=False)
return tokenizer
def load_model_and_tokenizer(pt_dir, pt_server_path):
"""
Load ProtTrans model and tokenizer.
pt_dir - STRING to determine the path to the directory with ProtTrans
"pytorch_model.bin" file
pt_server_path - STRING of the path to ProtTrans model in its server
returns (ProtT5-XL model, tokenizer)
"""
if(not os.path.exists(f"{pt_dir}/")):
os.system(f"mkdir -p {pt_dir}/")
if(os.path.isfile(f"{pt_dir}/pytorch_model.bin")):
# Only loading the model
model = get_pretrained_model(pt_dir)
else:
# Downloading and saving the model
model = get_pretrained_model(pt_server_path)
model.save_pretrained(pt_dir)
if(os.path.isfile(f"{pt_dir}/tokenizer_config.json")):
# Only loading the tokenizer
tokenizer = get_tokenizer(pt_dir)
else:
# Downloading and saving the tokenizer
tokenizer = get_tokenizer(pt_server_path)
tokenizer.save_pretrained(pt_dir)
return (model, tokenizer)
def process_FASTA(fasta_path, split_char="!", id_field=0):
"""
Reads in fasta file containing multiple sequences.
Split_char and id_field allow to control identifier extraction from header.
E.g.: set split_char="|" and id_field=1 for SwissProt/UniProt Headers.
Returns dictionary holding multiple sequences or only single
sequence, depending on input file.
"""
seqs = dict()
orig_seq_headers = dict()
orig_seqs = dict()
with open(fasta_path, 'r') as fasta_f:
for line in fasta_f:
if line.startswith('>'):
uniprot_id = line.replace('>', '').strip().split(split_char)[id_field]
uniprot_id = uniprot_id.replace("/", "_").replace(".", "_")
seqs[uniprot_id] = ''
orig_seq_headers[uniprot_id] = line.replace('>', '').strip()
orig_seqs[uniprot_id] = ''
else:
orig_seqs[uniprot_id] += line.strip()
seq = ''.join(line.split()).upper().replace("-", "")
seq = seq.replace('U','X').replace('Z', 'X').replace('O', 'X')
seqs[uniprot_id] += seq
example_id = next(iter(seqs))
return (seqs, orig_seq_headers, orig_seqs)
def get_embeddings(model, tokenizer, seqs, per_residue, per_protein,
max_residues=4000, # number of cumulative residues per batch
max_seq_len=2000, # max length after which we switch to single-sequence processing to avoid OOM
max_batch=100 # max number of sequences per single batch
):
"""
Generation of embeddings via batch-processing.
per_residue indicates that embeddings for each residue in a protein
should be returned.
per_protein indicates that embeddings for a whole protein should be
returned (average-pooling).
returns results depending on the option in the input.
"""
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
results = {'per_res_representations': dict(),
'mean_representations': dict()}
seq_dict = sorted(seqs.items(), key=lambda kv: len(seqs[kv[0]]),
reverse=True)
batch = list()
for seq_idx, (pdb_id, seq) in enumerate(seq_dict, 1):
seq_len = len(seq)
seq = ' '.join(list(seq))
batch.append((pdb_id, seq, seq_len))
n_res_batch = sum([s_len for _, _, s_len in batch]) + seq_len
if (len(batch) >= max_batch) or (n_res_batch >= max_residues) or (seq_idx == len(seq_dict)) or (seq_len > max_seq_len):
pdb_ids, seqs, seq_lens = zip(*batch)
batch = list()
token_encoding = tokenizer.batch_encode_plus(seqs,
add_special_tokens=True, padding='longest')
input_ids = torch.tensor(token_encoding['input_ids']).to(device)
attention_mask = torch.tensor(
token_encoding['attention_mask']).to(device)
try:
with torch.no_grad():
embedding_repr = model(input_ids,
attention_mask=attention_mask)
except RuntimeError:
print(f"{sys.argv[0]}: runtime error generating embedding for {pdb_id} (L={seq_len}). "+\
f"Try lowering batch size. If single sequence processing does not work, you need "+\
f"more vRAM to process your protein.", file=sys.stderr)
continue
for batch_idx, identifier in enumerate(pdb_ids):
s_len = seq_lens[batch_idx]
emb = embedding_repr.last_hidden_state[batch_idx,:s_len]
if per_residue:
results["per_res_representations"][identifier] = \
emb.detach().cpu().numpy().squeeze()
if per_protein:
protein_emb = emb.mean(dim=0)
results["mean_representations"][identifier] = \
protein_emb.detach().cpu().numpy().squeeze()
return results
def save_embeddings(sequences, embeddings, embeddings_directory, embedding_type="mean"):
"""
Saving embeddings to PT files for later use.
sequences - DICT with sequence ids as keys and sequences themselves
as values
embeddings - DICT with generated embeddings for each sequence
embeddings_directory - STRING that determines the path to directory for embeddings
embedding_type - STRING that determines, which type of embeddings to save
"""
embedding_type_key = embedding_type+"_representations"
for seq_id in list(sequences.keys()):
seq_data = {"label": seq_id}
if(seq_id in embeddings[embedding_type_key].keys()):
seq_data["sequence"] = sequences[seq_id]
seq_data[embedding_type_key] = torch.from_numpy(
embeddings[embedding_type_key][seq_id])
seq_code = sha256(sequences[seq_id].encode('utf-8')).hexdigest()
torch.save(seq_data, f"{embeddings_directory}/{embedding_type}_{seq_code}.pt")
def print_embeddings_generation_stats(iteration, portion_size, embeddings,
seqs_wo_emb, start_time, end_time):
"""
Printing embeddings generation statistics.
iteration - INT index of the iteration (indexing from zero)
portion_size - INT size of the processed portion
embeddings - DICT with keys 'mean_representations' and 'per_res_representations'
seqs_wo_emb - DICT with a portion of sequence ids as keys and amino acid sequences as values
start_time - datetime object of the beginning of embeddings' generation
end_time - datetime object of the end of embeddings' generation
"""
print(f"Portion {int(iteration/portion_size)+1}.", file=sys.stderr)
print(f"{len(embeddings['mean_representations'].keys())}/{len(list(seqs_wo_emb.keys()))}: "+\
"sequences with generated mean embeddings",
file=sys.stderr)
print(f"{len(embeddings['per_res_representations'].keys())}/{len(list(seqs_wo_emb.keys()))}: "+\
"sequences with generated per-residue embeddings",
file=sys.stderr)
print("%s: time to generate embeddings" % (end_time - start_time),
file=sys.stderr)
print("%s: time to generate embeddings per protein" % ((end_time - \
start_time)/len(embeddings["mean_representations"])),
file=sys.stderr)