Spaces:
Sleeping
Sleeping
matanninio
commited on
Commit
•
4fb0503
1
Parent(s):
f98cc68
save snapshot
Browse files- mammal_demo/tcr_task.py +196 -0
mammal_demo/tcr_task.py
ADDED
@@ -0,0 +1,196 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import torch
|
3 |
+
from fuse.data.tokenizers.modular_tokenizer.op import ModularTokenizerOp
|
4 |
+
from mammal.examples.dti_bindingdb_kd.task import DtiBindingdbKdTask
|
5 |
+
from mammal.keys import (
|
6 |
+
ENCODER_INPUTS_STR,
|
7 |
+
ENCODER_INPUTS_TOKENS,
|
8 |
+
ENCODER_INPUTS_ATTENTION_MASK,
|
9 |
+
CLS_PRED,
|
10 |
+
SCORES,
|
11 |
+
)
|
12 |
+
from mammal.model import Mammal
|
13 |
+
|
14 |
+
from mammal_demo.demo_framework import MammalObjectBroker, MammalTask
|
15 |
+
|
16 |
+
|
17 |
+
class TcrTask(MammalTask):
|
18 |
+
def __init__(self, model_dict):
|
19 |
+
super().__init__(name="T-cell receptors-peptide binding specificity", model_dict=model_dict)
|
20 |
+
self.description = "T-cell receptors-peptide binding specificity (TCR)"
|
21 |
+
self.examples = {
|
22 |
+
"tcr_beta_seq": "NAGVTQTPKFQVLKTGQSMTLQCAQDMNHEYMSWYRQDPGMGLRLIHYSVGAGITDQGEVPNGYNVSRSTTEDFPLRLLSAAPSQTSVYFCASSYSWDRVLEQYFGPGTRLTVT",
|
23 |
+
"epitope_seq": "LLQTGIHVRVSQPSL",
|
24 |
+
}
|
25 |
+
self.markup_text = """
|
26 |
+
# Mammal based T-cell receptors-peptide binding specificity demonstration
|
27 |
+
|
28 |
+
Given the TCR beta sequance and the epitope sequacne, estimate the binding specificity.
|
29 |
+
"""
|
30 |
+
|
31 |
+
|
32 |
+
|
33 |
+
|
34 |
+
def create_prompt(self,tcr_beta_seq, epitope_seq):
|
35 |
+
prompt = (
|
36 |
+
"<@TOKENIZER-TYPE=AA><BINDING_AFFINITY_CLASS><SENTINEL_ID_0>"+
|
37 |
+
f"<@TOKENIZER-TYPE=AA><MOLECULAR_ENTITY><MOLECULAR_ENTITY_TCR_BETA_VDJ><SEQUENCE_NATURAL_START>{tcr_beta_seq}<SEQUENCE_NATURAL_END>"+
|
38 |
+
f"<@TOKENIZER-TYPE=AA><MOLECULAR_ENTITY><MOLECULAR_ENTITY_EPITOPE><SEQUENCE_NATURAL_START>{epitope_seq}<SEQUENCE_NATURAL_END><EOS>"
|
39 |
+
)
|
40 |
+
|
41 |
+
return prompt
|
42 |
+
|
43 |
+
|
44 |
+
|
45 |
+
|
46 |
+
def crate_sample_dict(self, sample_inputs: dict, model_holder: MammalObjectBroker):
|
47 |
+
"""convert sample_inputs to sample_dict including creating a proper prompt
|
48 |
+
|
49 |
+
Args:
|
50 |
+
sample_inputs (dict): dictionary containing the inputs to the model
|
51 |
+
model_holder (MammalObjectBroker): model holder
|
52 |
+
Returns:
|
53 |
+
dict: sample_dict for feeding into model
|
54 |
+
"""
|
55 |
+
sample_dict= dict()
|
56 |
+
sample_dict[ENCODER_INPUTS_STR] = self.create_prompt(*sample_inputs)
|
57 |
+
tokenizer_op = model_holder.tokenizer_op
|
58 |
+
model = model_holder.model
|
59 |
+
tokenizer_op(
|
60 |
+
sample_dict=sample_dict,
|
61 |
+
key_in=ENCODER_INPUTS_STR,
|
62 |
+
key_out_tokens_ids=ENCODER_INPUTS_TOKENS,
|
63 |
+
key_out_attention_mask=ENCODER_INPUTS_ATTENTION_MASK,
|
64 |
+
)
|
65 |
+
sample_dict[ENCODER_INPUTS_TOKENS] = torch.tensor(
|
66 |
+
sample_dict[ENCODER_INPUTS_TOKENS], device=model.device
|
67 |
+
)
|
68 |
+
sample_dict[ENCODER_INPUTS_ATTENTION_MASK] = torch.tensor(
|
69 |
+
sample_dict[ENCODER_INPUTS_ATTENTION_MASK], device=model.device
|
70 |
+
)
|
71 |
+
|
72 |
+
return sample_dict
|
73 |
+
|
74 |
+
def run_model(self, sample_dict, model: Mammal):
|
75 |
+
# Generate Prediction
|
76 |
+
batch_dict = model.generate(
|
77 |
+
[sample_dict],
|
78 |
+
output_scores=True,
|
79 |
+
return_dict_in_generate=True,
|
80 |
+
max_new_tokens=5,
|
81 |
+
)
|
82 |
+
return batch_dict
|
83 |
+
|
84 |
+
@staticmethod
|
85 |
+
def positive_token_id(tokenizer_op: ModularTokenizerOp):
|
86 |
+
"""token for positive binding
|
87 |
+
|
88 |
+
Args:
|
89 |
+
model (MammalTrainedModel): model holding tokenizer
|
90 |
+
|
91 |
+
Returns:
|
92 |
+
int: id of positive binding token
|
93 |
+
"""
|
94 |
+
return tokenizer_op.get_token_id("<1>")
|
95 |
+
|
96 |
+
@staticmethod
|
97 |
+
def negative_token_id(tokenizer_op: ModularTokenizerOp):
|
98 |
+
"""token for negative binding
|
99 |
+
|
100 |
+
Args:
|
101 |
+
model (MammalTrainedModel): model holding tokenizer
|
102 |
+
|
103 |
+
Returns:
|
104 |
+
int: id of negative binding token
|
105 |
+
"""
|
106 |
+
return tokenizer_op.get_token_id("<0>")
|
107 |
+
|
108 |
+
def decode_output(self, batch_dict, tokenizer_op: ModularTokenizerOp)-> dict:
|
109 |
+
|
110 |
+
"""
|
111 |
+
Extract predicted class and scores
|
112 |
+
"""
|
113 |
+
|
114 |
+
# positive_token_id = self.positive_token_id(tokenizer_op)
|
115 |
+
# negative_token_id = self.negative_token_id(tokenizer_op)
|
116 |
+
|
117 |
+
negative_token_id = tokenizer_op.get_token_id("<0>")
|
118 |
+
positive_token_id = tokenizer_op.get_token_id("<1>")
|
119 |
+
|
120 |
+
label_id_to_int = {
|
121 |
+
negative_token_id: 0,
|
122 |
+
positive_token_id: 1,
|
123 |
+
}
|
124 |
+
classification_position = 1
|
125 |
+
|
126 |
+
decoder_output=batch_dict[CLS_PRED][0]
|
127 |
+
decoder_output_scores=batch_dict[SCORES][0]
|
128 |
+
|
129 |
+
|
130 |
+
if decoder_output_scores is not None:
|
131 |
+
scores = decoder_output_scores[classification_position,positive_token_id]
|
132 |
+
else:
|
133 |
+
scores=[None]
|
134 |
+
|
135 |
+
ans = dict(
|
136 |
+
pred=label_id_to_int.get(int(decoder_output[classification_position]), -1),
|
137 |
+
score=scores.item(),
|
138 |
+
)
|
139 |
+
return ans
|
140 |
+
|
141 |
+
|
142 |
+
|
143 |
+
def create_and_run_prompt(self, model_name, tcr_beta_seq, epitope_seq):
|
144 |
+
model_holder = self.model_dict[model_name]
|
145 |
+
inputs = {
|
146 |
+
"tcr_beta_seq": tcr_beta_seq,
|
147 |
+
"epitope_seq": epitope_seq,
|
148 |
+
}
|
149 |
+
sample_dict = self.crate_sample_dict(
|
150 |
+
sample_inputs=inputs, model_holder=model_holder
|
151 |
+
)
|
152 |
+
prompt = sample_dict[ENCODER_INPUTS_STR]
|
153 |
+
batch_dict = self.run_model(sample_dict=sample_dict, model=model_holder.model)
|
154 |
+
res = prompt, *self.decode_output(batch_dict, tokenizer_op=model_holder.tokenizer_op)
|
155 |
+
return res
|
156 |
+
|
157 |
+
|
158 |
+
|
159 |
+
def create_demo(self, model_name_widget):
|
160 |
+
|
161 |
+
|
162 |
+
with gr.Group() as demo:
|
163 |
+
gr.Markdown(self.markup_text)
|
164 |
+
with gr.Row():
|
165 |
+
tcr_textbox = gr.Textbox(
|
166 |
+
label="T-cell receptor beta sequence",
|
167 |
+
# info="standard",
|
168 |
+
interactive=True,
|
169 |
+
lines=3,
|
170 |
+
value=self.examples["tcr_beta_seq"],
|
171 |
+
)
|
172 |
+
epitope_textbox = gr.Textbox(
|
173 |
+
label="Epitope sequace",
|
174 |
+
# info="standard",
|
175 |
+
interactive=True,
|
176 |
+
lines=3,
|
177 |
+
value=self.examples["epitope_seq"],
|
178 |
+
)
|
179 |
+
with gr.Row():
|
180 |
+
run_mammal = gr.Button(
|
181 |
+
"Run Mammal prompt for TCL-Epitope Interaction",
|
182 |
+
variant="primary",
|
183 |
+
)
|
184 |
+
with gr.Row():
|
185 |
+
prompt_box = gr.Textbox(label="Mammal prompt", lines=5)
|
186 |
+
|
187 |
+
with gr.Row():
|
188 |
+
decoded = gr.Textbox(label="Mammal prediction")
|
189 |
+
binding_score = gr.Number(label="Binding score")
|
190 |
+
run_mammal.click(
|
191 |
+
fn=self.create_and_run_prompt,
|
192 |
+
inputs=[model_name_widget, tcr_textbox, epitope_textbox],
|
193 |
+
outputs=[prompt_box, decoded, binding_score],
|
194 |
+
)
|
195 |
+
demo.visible = False
|
196 |
+
return demo
|