MassimoGregorioTotaro commited on
Commit
634752b
1 Parent(s): cd5b5ac

add application file

Browse files
Files changed (1) hide show
  1. app.py +218 -0
app.py ADDED
@@ -0,0 +1,218 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from huggingface_hub import HfApi, ModelFilter
2
+ from transformers import AutoTokenizer, AutoModelForMaskedLM
3
+ import pandas as pd
4
+ import re
5
+ from tqdm import tqdm
6
+ import torch
7
+ import gradio as gr
8
+ import warnings
9
+ warnings.filterwarnings('ignore')
10
+
11
+ MODEL, MODEL_NAME, BATCH_CONVERTER, ALPHABET = None, None, None, None
12
+ OFFSET = 1
13
+ MODELS = [m.modelId for m in HfApi().list_models(filter=ModelFilter(author="facebook", model_name="esm", task="fill-mask"), sort="lastModified", direction=-1)]
14
+ SCORING = ["masked-marginals (more accurate)", "wt-marginals (faster)"]
15
+
16
+ def label_row(row, sequence, token_probs):
17
+ wt, idx, mt = row[0], int(row[1:-1]) - OFFSET, row[-1]
18
+ assert sequence[idx] == wt, "The listed wildtype does not match the provided sequence"
19
+
20
+ wt_encoded, mt_encoded = ALPHABET[wt], ALPHABET[mt]
21
+
22
+ score = token_probs[0, 1 + idx, mt_encoded] - token_probs[0, 1 + idx, wt_encoded]
23
+ return score.item()
24
+
25
+ def initialise_model(model_name):
26
+ global MODEL, MODEL_NAME, BATCH_CONVERTER, ALPHABET
27
+ MODEL_NAME = model_name
28
+ MODEL = AutoModelForMaskedLM.from_pretrained(model_name)
29
+ BATCH_CONVERTER = AutoTokenizer.from_pretrained(model_name)
30
+ ALPHABET = BATCH_CONVERTER.get_vocab()
31
+ if torch.cuda.is_available():
32
+ MODEL = MODEL.cuda()
33
+
34
+ def parse_input(seq, sub):
35
+ assert seq.isalpha(), "Sequence must be alphabetic"
36
+ substitutions, mode = list(), None
37
+
38
+ if len(sub.split()) == 1 and len(sub.split()[0]) == len(seq):
39
+ mode = 'seq vs seq'
40
+ for resi,(src,trg) in enumerate(zip(seq,sub), OFFSET):
41
+ if src != trg:
42
+ substitutions.append(f"{src}{resi}{trg}")
43
+ elif len(targets := sub.split()) > 1:
44
+ if all(re.match(r'\d+', x) for x in targets):
45
+ mode = 'deep mutational scan'
46
+ for resi in map(int, sub.split()):
47
+ src = seq[resi-OFFSET]
48
+ for trg in "ACDEFGHIKLMNPQRSTVWY".replace(src,''):
49
+ substitutions.append(f"{src}{resi}{trg}")
50
+ elif all(re.match(r'[A-Z]\d+[A-Z]', x) for x in targets):
51
+ mode = 'aa substitutions'
52
+ substitutions = targets
53
+
54
+ if not mode:
55
+ raise RuntimeError("Unrecognised running mode")
56
+
57
+ return mode, pd.DataFrame(substitutions, columns=['0'])
58
+
59
+ def run_model(sequence, substitutions, batch_tokens, scoring_strategy):
60
+ if scoring_strategy.startswith("wt-marginals"):
61
+ with torch.no_grad():
62
+ token_probs = torch.log_softmax(MODEL(batch_tokens)["logits"], dim=-1)
63
+ substitutions[MODEL_NAME] = substitutions.apply(
64
+ lambda row: label_row(
65
+ row['0'],
66
+ sequence,
67
+ token_probs,
68
+ ),
69
+ axis=1,
70
+ )
71
+ elif scoring_strategy.startswith("masked-marginals"):
72
+ all_token_probs = []
73
+ for i in tqdm(range(batch_tokens.size()[1])):
74
+ batch_tokens_masked = batch_tokens.clone()
75
+ batch_tokens_masked[0, i] = ALPHABET['<mask>']
76
+ with torch.no_grad():
77
+ token_probs = torch.log_softmax(
78
+ MODEL(batch_tokens_masked)["logits"], dim=-1
79
+ )
80
+ all_token_probs.append(token_probs[:, i])
81
+ token_probs = torch.cat(all_token_probs, dim=0).unsqueeze(0)
82
+ substitutions[MODEL_NAME] = substitutions.apply(
83
+ lambda row: label_row(
84
+ row['0'],
85
+ sequence,
86
+ token_probs,
87
+ ),
88
+ axis=1,
89
+ )
90
+
91
+ return substitutions
92
+
93
+ def parse_output(output, mode):
94
+ if mode == 'aa substitutions':
95
+ output = output.sort_values(MODEL_NAME, ascending=False)
96
+ elif mode == 'deep mutational scan':
97
+ output = pd.concat([(output.assign(resi=output['0'].str.extract(r'(\d+)', expand=False).astype(int))
98
+ .sort_values(['resi', MODEL_NAME], ascending=[True,False])
99
+ .groupby(['resi'])
100
+ .head(19)
101
+ .drop(['resi'], axis=1)).iloc[19*x:19*(x+1)].reset_index(drop=True) for x in range(output.shape[0]//19)]
102
+ , axis=1).set_axis(range(output.shape[0]//19*2), axis='columns')
103
+
104
+ return output.style.format(lambda x: f'{x:.2f}' if isinstance(x, float) else x).hide_index().hide_columns().background_gradient(cmap="RdYlGn", vmax=8, vmin=-8).to_html()
105
+
106
+
107
+ # mode = 'deep mutational scan' #@param ['seq vs seq', 'deep mutational scan', 'aa substitutions']
108
+ # sequence = "MVEQYLLEAIVRDARDGITISDCSRPDNPLVFVNDAFTRMTGYDAEEVIGKNCRFLQRGDINLSAVHTIKIAMLTHEPCLVTLKNYRKDGTIFWNELSLTPIINKNGLITHYLGIQKDVSAQVILNQTLHEENHLLKSNKEMLEYLVNIDALTGLHNRRFLEDQLVIQWKLASRHINTITIFMIDIDYFKAFNDTYGHTAGDEALRTIAKTLNNCFMRGSDFVARYGGEEFTILAIGMTELQAHEYSTKLVQKIENLNIHHKGSPLGHLTISLGYSQANPQYHNDQNLVIEQADRALYSAKVEGKNRAVAYREQ" #@param {type:"string"}
109
+ # target = "61 214 19 30 122 140" #@param {type:"string"}
110
+ # substitutions = list()
111
+ # scoring_strategy = "masked-marginals"
112
+
113
+ # if mode == 'seq vs seq':
114
+ # for resi,(seq,trg) in enumerate(zip(sequence,target), OFFSET):
115
+ # if seq != trg:
116
+ # substitutions.append(f"{seq}{resi}{trg}")
117
+ # elif mode == 'deep mutational scan':
118
+ # for resi in map(int, target.split()):
119
+ # seq = sequence[resi-OFFSET]
120
+ # for trg in "ACDEFGHIKLMNPQRSTVWY".replace(seq,''):
121
+ # substitutions.append(f"{seq}{resi}{trg}")
122
+ # elif mode == 'aa substitutions':
123
+ # substitutions = target.split()
124
+ # else:
125
+ # raise RuntimeError("Unrecognised running mode")
126
+
127
+ # df = pd.DataFrame(substitutions, columns=['0'])
128
+ # mutation_col = df.columns[0]
129
+
130
+ # batch_tokens = batch_converter(sequence, return_tensors='pt')['input_ids']
131
+
132
+ # if scoring_strategy == "wt-marginals":
133
+ # with torch.no_grad():
134
+ # token_probs = torch.log_softmax(model(batch_tokens)["logits"], dim=-1)
135
+ # df[model_name] = df.apply(
136
+ # lambda row: label_row(
137
+ # row[mutation_col],
138
+ # sequence,
139
+ # token_probs,
140
+ # alphabet,
141
+ # OFFSET,
142
+ # ),
143
+ # axis=1,
144
+ # )
145
+ # elif scoring_strategy == "masked-marginals":
146
+ # all_token_probs = []
147
+ # for i in tqdm(range(batch_tokens.size()[1])):
148
+ # batch_tokens_masked = batch_tokens.clone()
149
+ # batch_tokens_masked[0, i] = alphabet['<mask>']
150
+ # with torch.no_grad():
151
+ # token_probs = torch.log_softmax(
152
+ # model(batch_tokens_masked)["logits"], dim=-1
153
+ # )
154
+ # all_token_probs.append(token_probs[:, i]) # vocab size
155
+ # token_probs = torch.cat(all_token_probs, dim=0).unsqueeze(0)
156
+ # df[model_name] = df.apply(
157
+ # lambda row: label_row(
158
+ # row[mutation_col],
159
+ # sequence,
160
+ # token_probs,
161
+ # alphabet,
162
+ # OFFSET,
163
+ # ),
164
+ # axis=1,
165
+ # )
166
+
167
+ # if mode == 'aa substitutions':
168
+ # df = df.sort_values(model_name, ascending=False)
169
+ # elif mode == 'deep mutational scan':
170
+ # df = pd.concat([(df.assign(resi=df['0'].str.extract(f'(\d+)', expand=False).astype(int))
171
+ # .sort_values(['resi', model_name], ascending=[True,False])
172
+ # .groupby(['resi'])
173
+ # .head(19)
174
+ # .drop(['resi'], axis=1)).iloc[19*x:19*(x+1)].reset_index(drop=True) for x in range(df.shape[0]//19)]
175
+ # , axis=1).set_axis(range(df.shape[0]//19*2), axis='columns')
176
+
177
+ # df.style.hide_index().hide_columns().background_gradient(cmap="RdYlGn", vmax=8, vmin=-8)
178
+
179
+ def app(*argv):
180
+ seq, trg, model_name, scoring_strategy, *_ = argv
181
+
182
+ mode, substitutions = parse_input(seq, trg)
183
+
184
+ if model_name != MODEL_NAME:
185
+ initialise_model(model_name)
186
+
187
+ batch_tokens = BATCH_CONVERTER(seq, return_tensors='pt')['input_ids']
188
+
189
+ df = run_model(seq, substitutions, batch_tokens, scoring_strategy)
190
+
191
+ return parse_output(df, mode)
192
+
193
+ # demo = gr.Interface(
194
+ # theme=gr.themes.Base(),
195
+ # title="Protein Sequence Mutagenesis",
196
+ # description="Predict the effect of mutations on protein stability",
197
+ # fn=app,
198
+ # inputs=[gr.Textbox(lines=2, label="Sequence", placeholder="Sequence here...", required=True, value='MVEQYLLEAIVRDARDGITISDCSRPDNPLVFVNDAFTRMTGYDAEEVIGKNCRFLQRGDINLSAVHTIKIAMLTHEPCLVTLKNYRKDGTIFWNELSLTPIINKNGLITHYLGIQKDVSAQVILNQTLHEENHLLKSNKEMLEYLVNIDALTGLHNRRFLEDQLVIQWKLASRHINTITIFMIDIDYFKAFNDTYGHTAGDEALRTIAKTLNNCFMRGSDFVARYGGEEFTILAIGMTELQAHEYSTKLVQKIENLNIHHKGSPLGHLTISLGYSQANPQYHNDQNLVIEQADRALYSAKVEGKNRAVAYREQ'),
199
+ # gr.Textbox(lines=2, label="Substitutions", placeholder="Substitutions here...", required=True, value="61 214 19 30 122 140"),
200
+ # gr.Dropdown(MODELS, label="Model", value=MODELS[1]),
201
+ # gr.Dropdown(["masked-marginals (more accurate)", "wt-marginals (faster)"], label="Scoring strategy", value="wt-marginals (faster)"),
202
+ # ],
203
+ # outputs=gr.HTML(formatter="html", label="Output"),
204
+ # )
205
+
206
+ with gr.Blocks() as demo:
207
+ gr.Markdown("""Protein Sequence Mutagenesis""", name="title")
208
+ gr.Markdown("""Predict the effect of mutations on protein stability""", name="description")
209
+ seq = gr.Textbox(lines=2, label="Sequence", placeholder="Sequence here...", required=True, value='MVEQYLLEAIVRDARDGITISDCSRPDNPLVFVNDAFTRMTGYDAEEVIGKNCRFLQRGDINLSAVHTIKIAMLTHEPCLVTLKNYRKDGTIFWNELSLTPIINKNGLITHYLGIQKDVSAQVILNQTLHEENHLLKSNKEMLEYLVNIDALTGLHNRRFLEDQLVIQWKLASRHINTITIFMIDIDYFKAFNDTYGHTAGDEALRTIAKTLNNCFMRGSDFVARYGGEEFTILAIGMTELQAHEYSTKLVQKIENLNIHHKGSPLGHLTISLGYSQANPQYHNDQNLVIEQADRALYSAKVEGKNRAVAYREQ')
210
+ trg = gr.Textbox(lines=1, label="Substitutions", placeholder="Substitutions here...", required=True, value="61 214 19 30 122 140")
211
+ model_name = gr.Dropdown(MODELS, label="Model", value=MODELS[1])
212
+ scoring_strategy = gr.Dropdown(SCORING, label="Scoring strategy", value=SCORING[1])
213
+ btn = gr.Button(label="Submit", type="submit")
214
+ btn.click(fn=app, inputs=[seq, trg, model_name, scoring_strategy], outputs=[gr.HTML()])
215
+
216
+ if __name__ == '__main__':
217
+ demo.launch()
218
+ # demo.launch(share=True, server_name="0.0.0.0", server_port=7878)