README.md

---
license: gpl-3.0
language:
- en
base_model:
- facebook/esm2_t30_150M_UR50D
tags:
- peptides
- neurotoxicity
- protein-classification
- therapeutic-peptides
- bioinformatics
- esm2
- transformer
---

# 🧠 NTxPred2: A large language model for predicting neurotoxic peptides and neurotoxins

NTxPred2 is a fine-tuned transformer model built on top of the [ESM2-t30_150M_UR50D](https://huggingface.co/facebook/esm2_t30_150M_UR50D) protein language model. It is specifically trained for **binary classification** of peptide sequences — predicting whether a peptide is **neurotoxic** or **non-toxic**.

🎯 **Use Case:** Accelerating the identification and design of safe peptide therapeutics by filtering out neurotoxic candidates early in the drug development pipeline.

---

### 🖼️ NTxPred2 Workflow

![NTxPred2 Workflow](https://github.com/raghavagps/NTxPred2/raw/main/NTxPred2_workflow.png)

---

## 🧬 Model Highlights

- **Base Model:** Facebook’s ESM2-t30 (150M parameters)
- **Fine-Tuning Task:** Neurotoxicity prediction (binary classification)
- **Input:** Short peptide sequences (7–50 amino acids)
- **Output:** Binary label → `1` (neurotoxic), `0` (non-toxic)
- **Architecture:** ESM2 encoder + linear classification head

---

## 🗂️ Files Included

- `config.json` – Contains configuration settings for the model architecture, hyperparameters, and training details.

- `model.safetensors` – This is the actual trained model weights saved in the SafeTensors format, which is safer and faster than the traditional .bin files.

- `special_tokens_map.json` – Stores mappings for special tokens, like [CLS], [SEP], or any custom tokens used in your tokenizer.

- `tokenizer_config.json` – Contains tokenizer-related settings (like vocabulary size, tokenization method).

- `vocab.txt` – Lists all tokens and their corresponding IDs; it's essential for text tokenization.

---

## 🚀 How to Use

### 🔧 Install Dependencies

```bash
pip install torch esm biopython huggingface_hub


### Loading the Model from Hugging Face

```python
import torch
import torch.nn as nn
import esm
import json
from huggingface_hub import hf_hub_download

# Define the classifier model (ESM encoder + linear head)
class ProteinClassifier(nn.Module):
    def __init__(self, esm_model, embedding_dim, num_classes):
        super(ProteinClassifier, self).__init__()
        self.esm_model = esm_model
        self.fc = nn.Linear(embedding_dim, num_classes)

    def forward(self, tokens):
        layer_index = len(self.esm_model.layers)  # Get number of layers
        results = self.esm_model(tokens, repr_layers=[layer_index])
        embeddings = results["representations"][layer_index].mean(1)
        return self.fc(embeddings)

# Device setup
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load config from your repo
config_path = hf_hub_download(repo_id="anandr88/NTxPred2", filename="config.json")
with open(config_path, 'r') as f:
    config = json.load(f)

# Load ESM2 model - UPDATED METHOD
model_name = "esm2_t30_150M_UR50D"
esm_model, alphabet = esm.pretrained.load_model_and_alphabet(model_name)
batch_converter = alphabet.get_batch_converter()

# Initialize a NEW classifier (with random weights)
classifier = ProteinClassifier(
    esm_model, 
    embedding_dim=config['embedding_dim'], 
    num_classes=config['num_classes']
)
classifier.to(device)
classifier.eval()

print("✅ Model loaded successfully!")
print(f"Using device: {device}")
print(f"Model architecture: {classifier}")
```

---

## 🧪 Example Usage (Optional)

---

```python
# Example Usage for Binary Classification
sequence = ("TEST_SEQUENCE", "ACDEFGHIKLMNPQRSTVWY")  # Your peptide sequence

# Convert to model input format
_, _, batch_tokens = batch_converter([sequence])
batch_tokens = batch_tokens.to(device)

# Predict
with torch.no_grad():
    logits = classifier(batch_tokens)
    probability = torch.sigmoid(logits).item()  # Sigmoid for binary classification

# Interpret results
threshold = 0.5  # Standard threshold (adjust if needed)
prediction = "Neurotoxic" if probability >= threshold else "Not-toxic"

print("\n" + "="*50)
print(f"🔬 Input Sequence: {sequence[1]}")
print(f"📊 Neurotoxicity Probability: {probability:.4f}")
print(f"🏷️ Prediction: {prediction} (threshold={threshold})")

```

---

## 📊 Applications

- **Neurotoxic peptide filtering** in therapeutic design  
- **Toxicity scanning** of synthetic peptides  
- **Dataset annotation** for bioactivity studies  
- **Educational use** in bioinformatics and deep learning for proteins  

---

## 🌐 Related Links

- 🔬 Project Web Server: [NTxPred2 Web Tool](http://webs.iiitd.edu.in/raghava/ntxpred2)
- 🧾 Documentation & Source: [GitHub – raghavagps/NTxPred2](https://github.com/raghavagps/NTxPred2)

---

## 🧠 Citation

📖 Rathore et al.  
_A Large Language Model for Predicting Neurotoxic Peptides and Neurotoxins._  
**#Coming Soon#**

---

👨‍🔬 Start using **NTxPred2** today to enhance your peptide screening pipeline with the power of **transformer-based intelligence**!