██████╗ ██╗  ████████╗      ██████╗███╗   ██╗███╗   ██╗
██╔══██╗██║  ╚══██╔══╝     ██╔════╝████╗  ██║████╗  ██║
██████╔╝██║     ██║        ██║     ██╔██╗ ██║██╔██╗ ██║
██╔══██╗██║     ██║        ██║     ██║╚██╗██║██║╚██╗██║
██████╔╝███████╗██║        ╚██████╗██║ ╚████║██║ ╚████║
╚═════╝ ╚══════╝╚═╝         ╚═════╝╚═╝  ╚═══╝╚═╝  ╚═══╝

🧬 BLT-CNN: Antibiotic Resistance Predictor

When every second counts, your model shouldn't guess.

BLT-CNN is a novel hybrid deep learning architecture that predicts antibiotic resistance
from raw DNA sequences — combining entropy-guided patching, local motif detection, and
global sequence context in a single end-to-end trainable model.

---

🌍 Why This Matters

"Antimicrobial resistance is predicted to kill 10 million people per year by 2050 — more than cancer."
— UN Interagency Coordination Group on AMR

Traditional resistance testing takes 24–72 hours. Patients die waiting.
BLT-CNN predicts resistance from a DNA sequence in milliseconds.

---

⚡ Architecture — Three Engines, One Model

                    DNA Sequence (300 nucleotides)
                            │
                    ┌───────▼───────┐
                    │   Embedding   │  5-token vocab → 64-dim vectors
                    └───────┬───────┘
                            │
              ┌─────────────┼─────────────┐
              │                           │
    ┌─────────▼──────────┐    ┌──────────▼──────────┐
    │  BLT Entropy       │    │   1D CNN             │
    │  Patcher           │    │   Local Motifs       │
    │                    │    │                      │
    │  H(i) > 1.5 bits   │    │  Conv(256, k=9)      │
    │  → 3-nt patch      │    │  Conv(128, k=9)      │
    │  H(i) ≤ 1.5 bits   │    │  MaxPool(2)          │
    │  → 12-nt patch     │    │                      │
    │                    │    └──────────┬──────────┘
    │  Output: 64-dim    │               │
    └─────────┬──────────┘    ┌──────────▼──────────┐
              │               │  Transformer         │
              │               │  Encoder             │
              │               │                      │
              │               │  4 layers, 8 heads   │
              │               │  GELU, sinusoidal PE │
              │               │  Output: 128-dim     │
              │               └──────────┬──────────┘
              │                          │
              └──────────┬───────────────┘
                         │  Concatenate [128 + 64] = 192-dim
                ┌────────▼────────┐
                │  Fusion MLP     │  192 → 256 → 128
                │  LayerNorm      │  + Dropout
                └────────┬────────┘
                         │
                ┌────────▼────────┐
                │  Classifier     │  128 → 3
                │  Sigmoid        │  Multi-label output
                └────────┬────────┘
                         │
          ┌──────────────┼──────────────┐
          ▼              ▼              ▼
     Methicillin   Ciprofloxacin   Vancomycin
       99.8%           0.6%           0.0%
      RESISTANT     SUSCEPTIBLE    SUSCEPTIBLE

---

📊 Results

Metric	Score
🎯 Accuracy	93.34%
📈 AUC	98.55%
⚖️ F1 Score	94.18%
🔬 Precision	95.57%
📡 Recall	93.35%
🧮 Parameters	944,387

Per-Class Performance

Antibiotic	Samples	Precision	Recall	F1	Status
🔵 Methicillin	3,612	0.9761	0.9852	0.9806	✅ Excellent
🟡 Ciprofloxacin	283	0.8857	0.5849	0.7045	⚠️ Moderate
🔴 Vancomycin	171	0.6500	0.5417	0.5909	🔧 Improving

---

🧬 The Three Antibiotics

💉 Methicillin Beta-lactam antibiotic Targets S. aureus (MRSA) Resistance gene: mecA Most common hospital-acquired infection worldwide

💊 Ciprofloxacin Fluoroquinolone antibiotic Targets E. coli, Klebsiella Resistance genes: qnr, gyrA Most prescribed broad-spectrum antibiotic

⚠️ Vancomycin Glycopeptide antibiotic Targets Enterococcus (VRE) Resistance genes: vanA, vanB Last-resort antibiotic — critical priority

---

🚀 Quick Start

1. Clone & Install

git clone https://github.com/man-ra/BLT-CNN-Project.git
cd BLT-CNN-Project
pip install -r requirements.txt

2. Run the 3D Web UI

# Start the FastAPI backend
python app_api.py

# Then open blt_cnn_ui.html in your browser
# → Real predictions with 3D DNA helix animation

3. Run Streamlit App

python -m streamlit run app.py
# Opens at http://localhost:8501

4. Train from Scratch

# Process CARD database
python scripts/process_card.py

# Train BLT-CNN
python scripts/train_card.py

---

🗂️ Project Structure

BLT-CNN-Project/
│
├── 🧠 src/
│   ├── models/
│   │   ├── blt_patcher.py      ← Entropy-guided dynamic patching
│   │   ├── cnn1d.py            ← 1D CNN local motif detection
│   │   ├── transformer.py      ← Transformer encoder
│   │   └── blt_cnn.py          ← Hybrid fusion model ⭐
│   ├── data/
│   │   ├── dataset.py          ← CARD data loader
│   │   └── encoding.py         ← Nucleotide encoder
│   └── training/
│       ├── trainer.py          ← Training loop
│       ├── losses.py           ← Loss functions
│       └── metrics.py          ← Evaluation metrics
│
├── 🔬 scripts/
│   ├── process_card.py         ← CARD database processor
│   ├── train_card.py           ← Main training script
│   └── evaluate.py             ← Evaluation script
│
├── 🌐 Frontend/
│   ├── blt_cnn_ui.html         ← 3D animated web UI ✨
│   ├── app.py                  ← Streamlit app
│   └── app_api.py              ← FastAPI backend
│
├── 📊 results/
│   └── card_results.json       ← Training results
│
└── 🤖 models/
    └── blt_cnn_card_best.pth   ← Best trained model (epoch 18)

---

🔬 How BLT Entropy Patching Works

Unlike fixed tokenization, BLT dynamically adjusts resolution based on information content:

DNA:  A T G A A A C G T A T C G G A A T T C G ...
      │ │ │ │ │ │ │ │ │ │ │ │ │ │ │ │ │ │ │ │
Ent:  1.2 1.1 2.1 2.3 1.9 2.2 1.0 1.0 1.1 1.2
       ↓   ↓    ↓   ↓   ↓   ↓   ↓   ↓   ↓   ↓
      [  large  ][small][small][  large  ][large]
      12-nt patch 3-nt  3-nt  12-nt patch
      conserved   hotspot      conserved

• High entropy (>1.5 bits) → 3-nucleotide patches → captures mutation hotspots
• Low entropy (≤1.5 bits) → 12-nucleotide patches → efficiently compresses conserved regions

This mirrors biology: resistance mutations cluster in specific variable regions while flanking sequences remain conserved.

---

📦 Dependencies

torch>=2.0
numpy
scikit-learn
fastapi
uvicorn
streamlit
biopython
pandas
matplotlib

---

📄 Citation

If you use this work, please cite:

@misc{bltcnn2026,
  title     = {BLT-CNN: A Hybrid Deep Learning Architecture for
               Antibiotic Resistance Prediction from Genomic Sequences},
  author    = {Mantasha},
  year      = {2026},
  url       = {https://github.com/man-ra/BLT-CNN-Project},
  note      = {B.Tech Research Project}
}

---

🙏 References

• Pagnoni et al. (2024) — Byte Latent Transformer, Meta AI
• Alcock et al. (2023) — CARD 2023, Nucleic Acids Research
• Arango-Argoty et al. (2018) — DeepARG, Microbiome
• Vaswani et al. (2017) — Attention Is All You Need, NeurIPS

---

Built with 🧬 by Mantasha

B.Tech 3rd Year · Antibiotic Resistance Research

⭐ Star this repo if you found it useful

---

"The bacteria are evolving. So should our tools."