AmelieSchreiber
commited on
Commit
•
2839395
1
Parent(s):
9a1d58d
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,53 @@
|
|
1 |
---
|
2 |
license: mit
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: mit
|
3 |
---
|
4 |
+
|
5 |
+
# ESM-2 for Predicting Binding Sites
|
6 |
+
|
7 |
+
## Using the Model
|
8 |
+
|
9 |
+
```python
|
10 |
+
from transformers import AutoModelForTokenClassification, AutoTokenizer
|
11 |
+
from peft import PeftModel
|
12 |
+
import torch
|
13 |
+
|
14 |
+
# Path to the saved LoRA model
|
15 |
+
model_path = "AmelieSchreiber/esm2_t33_650M_qlora_binding_12M"
|
16 |
+
# ESM2 base model
|
17 |
+
base_model_path = "facebook/esm2_t12_35M_UR50D"
|
18 |
+
|
19 |
+
# Load the model
|
20 |
+
base_model = AutoModelForTokenClassification.from_pretrained(base_model_path)
|
21 |
+
loaded_model = PeftModel.from_pretrained(base_model, model_path)
|
22 |
+
|
23 |
+
# Ensure the model is in evaluation mode
|
24 |
+
loaded_model.eval()
|
25 |
+
|
26 |
+
# Load the tokenizer
|
27 |
+
loaded_tokenizer = AutoTokenizer.from_pretrained(base_model_path)
|
28 |
+
|
29 |
+
# Protein sequence for inference
|
30 |
+
protein_sequence = "MAVPETRPNHTIYINNLNEKIKKDELKKSLHAIFSRFGQILDILVSRSLKMRGQAFVIFKEVSSATNALRSMQGFPFYDKPMRIQYAKTDSDIIAKMKGT" # Replace with your actual sequence
|
31 |
+
|
32 |
+
# Tokenize the sequence
|
33 |
+
inputs = loaded_tokenizer(protein_sequence, return_tensors="pt", truncation=True, max_length=1024, padding='max_length')
|
34 |
+
|
35 |
+
# Run the model
|
36 |
+
with torch.no_grad():
|
37 |
+
logits = loaded_model(**inputs).logits
|
38 |
+
|
39 |
+
# Get predictions
|
40 |
+
tokens = loaded_tokenizer.convert_ids_to_tokens(inputs["input_ids"][0]) # Convert input ids back to tokens
|
41 |
+
predictions = torch.argmax(logits, dim=2)
|
42 |
+
|
43 |
+
# Define labels
|
44 |
+
id2label = {
|
45 |
+
0: "No binding site",
|
46 |
+
1: "Binding site"
|
47 |
+
}
|
48 |
+
|
49 |
+
# Print the predicted labels for each token
|
50 |
+
for token, prediction in zip(tokens, predictions[0].numpy()):
|
51 |
+
if token not in ['<pad>', '<cls>', '<eos>']:
|
52 |
+
print((token, id2label[prediction]))
|
53 |
+
```
|