README.md

metadata

title: CRISPR-BERT Prediction API
emoji: 🚀
colorFrom: blue
colorTo: purple
sdk: docker
sdk_version: latest
app_file: app.py
pinned: false

CRISPR-BERT Prediction API

CRISPR off-target prediction API using hybrid CNN-BERT architecture deployed on Hugging Face Spaces.

🚀 API Endpoints

• GET / - API information
• GET /health - Health check (shows if model is loaded)
• POST /predict - Make a single prediction
• POST /batch_predict - Make batch predictions
• GET /model/info - Get model information

📝 Usage

Single Prediction

curl -X POST https://santu0032-crispr-bert-api.hf.space/predict \
  -H "Content-Type: application/json" \
  -d '{"sgRNA": "GGTGAGTGAGTGTGTGCGTGTGG", "DNA": "TGTGAGTGTGTGTGTGTGTGTGT"}'

Response

{
  "prediction": 0,
  "confidence": 0.9919,
  "probabilities": {
    "class_0": 0.9919,
    "class_1": 0.0081
  },
  "threshold_used": 0.65,
  "sgRNA": "GGTGAGTGAGTGTGTGCGTGTGG",
  "DNA": "TGTGAGTGTGTGTGTGTGTGTGT",
  "timestamp": "2025-10-31T..."
}

📋 Model Requirements

• sgRNA: Exactly 23 nucleotides (A, T, C, G, or - for indels)
• DNA: Exactly 23 nucleotides (A, T, C, G, or - for indels)

🔧 Files Structure

.
├── app.py                    # Main Flask application
├── requirements.txt          # Python dependencies
├── Dockerfile               # Docker configuration
├── sequence_encoder.py       # Sequence encoding utilities
├── data_loader.py           # Data loading utilities
├── final1/
│   └── weight/
│       ├── final_model.keras           # Trained model
│       ├── threshold_schedule.json    # Threshold config
│       └── bert_weight/                # BERT weights
└── README.md                # This file

🧬 About CRISPR-BERT

This API uses a hybrid CNN-BERT architecture to predict CRISPR off-target effects:

• CNN Branch: Multi-scale convolutions for sequence pattern recognition
• BERT Branch: Transformer attention for contextual understanding
• BiGRU Layers: Bidirectional GRU for sequence modeling
• Final Output: Binary classification (on-target vs off-target)

📊 Model Architecture

• Input: 23-nt sgRNA and DNA sequences
• CNN Encoding: 26x7 one-hot encoding
• BERT Encoding: 26 token IDs
• Output: Binary prediction with confidence scores

🔗 Integration

Update your backend to use this API:

MODEL_API_URL=https://santu0032-crispr-bert-api.hf.space

📝 License

MIT License

🙏 Credits

Built with TensorFlow, Flask, and deployed on Hugging Face Spaces.