Genomic Quixer: Quantum Next-Token Prediction for Genomics

This project implements a Quantum Transformer (Quixer) for genomic sequence modeling (Next-Base Prediction). It compares the performance of a quantum-enhanced attention mechanism (LCU + QSVT) against a classical transformer baseline on genomic data.

🧬 Pipeline Overview

The pipeline focuses on K-mer Tokenization and Next-Token Prediction, treating DNA sequences like language.

graph TD
    A[Genomic FASTA] -->|preprocess.py| B(K-mer Tokenization)
    B --> C[Tokenized Dataset]
    C --> D{Model Training}
    D -->|model_type='quixer'| E[Quixer Quantum Transformer]
    D -->|model_type='classical'| F[Classical Transformer Baseline]
    E --> G[Evaluation Metrics]
    F --> G

Loading

📂 Project Structure

The codebase is organized as a modular Python package genomic_quixer:

.
├── genomic_quixer/             # Main Package
│   ├── data/
│   │   ├── dataset.py          # Memory-efficient Dataset class
│   │   └── preprocess.py       # K-mer tokenization logic
│   ├── models/
│   │   ├── quixer.py           # Quantum Transformer (Ansatz 14 + LCU/QSVT)
│   │   └── classical.py        # Classical Transformer Baseline
│   └── training/
│       └── trainer.py          # Training loops and visualization
├── train.py                    # Main entry point for training
├── preprocess.py               # Entry point for data generation
└── legacy_lambeq/              # Archive of previous Lambeq experiments

🚀 Quick Start

1. Preprocess Data

Convert raw FASTA sequences into K-mer tokens.

# Default: 4-mer tokenization (Vocab size = 256)
python preprocess.py --k_mer 4

2. Train Quixer (Quantum Model)

Train the quantum transformer model.

python train.py \
    --model_type quixer \
    --epochs 10 \
    --batch_size 256 \
    --window_size 128 \
    --qubits 6

3. Train Classical Baseline

Train a classical transformer with comparable parameters.

python train.py \
    --model_type classical \
    --epochs 10 \
    --batch_size 256 \
    --window_size 128

📊 Key Features

• Quixer Model: Uses TorchQuantum to simulate parameterized quantum circuits (Ansatz 14) with Linear Combination of Unitaries (LCU) and Quantum Singular Value Transformation (QSVT).
• Efficient Data Loading: Uses memory mapping (mmap) to handle large genomic datasets without loading everything into RAM.
• Direct Comparison: Unified training script allowing side-by-side comparison of Quantum vs Classical approaches on the same data splits.

🛠 Dependencies

• torch
• torchquantum
• numpy
• tqdm
• matplotlib

---

Note: This pipeline replaces the previous Lambeq-based approach. Old files can be found in legacy_lambeq/.