Breast Cancer Classification of DNA Sequences

📚 Overview

This project implements a hybrid machine learning approach for classifying breast cancer from DNA sequences using bidirectional embeddings generated by DNABERT. The study processes over 46 million high-quality DNA sequences to distinguish between cancerous and non-cancerous genomic material.

Key Features

• Bidirectional Analysis: Utilizes both forward and reverse DNA strand representations
• Hybrid Classification: Combines Random Forest (forward embeddings) and Deep Neural Networks (backward embeddings)
• High-Quality Data: Implements Q30 Phred score filtering for 99.9% base call accuracy
• Scalable Processing: Handles large genomic datasets using efficient batch processing

🎯 Research Objectives

• Apply DNA sequence analysis to distinguish between genetic material linked to breast cancer
• Develop a bidirectional DNA sequence embedding approach for improved classification
• Demonstrate the potential of genomic information for early, non-invasive cancer diagnosis

📖 Research Paper

📄 Main Publication: Breast_Cancer_Classification_of_DNA_Sequences3.pdf

Authors:

• Aakash Walavalkar (Michigan Technological University, USA)
• Anushka Kumar (NMIMS, Mumbai)
• Laavanya Mishra (NMIMS, Mumbai)

Keywords: Phred Score, Base Pair, Sequence Embeddings, Breast Cancer, DNA Sequencing, Classification

🗂️ Project Structure

├── 📄 Breast_Cancer_Classification_of_DNA_Sequences3.pdf  # Main research paper
├── 📊 Data Processing & Analysis
│   ├── 6. Generating Clean Readings for All Batches.pdf   # Quality control procedures
│   ├── cleaning_sequences.ipynb                          # Sequence cleaning implementation
│   ├── data_read_forward.ipynb                          # Forward strand processing
│   ├── data_read_backward.ipynb                         # Backward strand processing
│   └── rough.ipynb                                      # Utility functions
├── 🧬 Embedding Generation
│   ├── embeddings.py                                    # DNABERT embedding generation
│   └── embeddings.ipynb                                # Embedding analysis notebook
├── 🤖 Model Training
│   ├── forw_train_df.ipynb                             # Forward embeddings training
│   └── backw_train_df.ipynb                            # Backward embeddings training
└── 📈 Results & Analysis
    ├── UMAP visualizations                             # Dimensionality reduction plots
    └── Classification reports                          # Model performance metrics

🔬 Methodology

1. Data Acquisition

• Source: National Genomics Data Center (NGDC) and NCBI SRA
• Datasets:
  • Cancerous: Primary breast cancer sequences (SRR5177930)
  • Non-Cancerous: Normal breast tissue samples (SRR6269879)
• Format: FASTQ files converted to Parquet for efficient processing

2. Quality Control & Filtering

def is_quality_good(quality_scores):
    return np.min(quality_scores) >= 30

Filtering Criteria:

• Phred Score: ≥ Q30 (99.9% base call accuracy)
• Sequence Length: ≥ 100 base pairs
• Batch Size: 100,000 reads per batch for memory efficiency

3. Data Processing Pipeline

Phase 1: FASTQ to Parquet Conversion

• Parse FASTQ files using Biopython's SeqIO
• Extract sequence ID, nucleotide sequence, and quality scores
• Batch processing to handle 60M+ sequences efficiently

Phase 2: Quality Filtering

• Apply Q30 filtering to ensure high-quality sequences
• Remove sequences with any base having Phred score < 30
• Filter sequences shorter than 100 base pairs

Phase 3: Embedding Generation

# DNABERT-6 tokenization and embedding
model = AutoModel.from_pretrained("zhihan1996/DNA_bert_6")
embeddings = model.encode(sequences)  # 768-dimensional vectors

4. Model Architecture

Forward Embeddings: Random Forest Classifier

• Algorithm: RandomForestClassifier with Optuna hyperparameter optimization
• Best Parameters:
  • n_estimators: 161
  • max_depth: 21
  • max_features: sqrt
• Performance: AUC = 0.9753, Accuracy = 97.5%

Backward Embeddings: Deep Neural Network

• Architecture: Feedforward Neural Network
  • Input: 768-dimensional DNABERT embeddings
  • Layers: [4096, 2048, 1024, 512, 256, 128]
  • Dropout: 0.3
  • Optimizer: Adam
• Performance: AUC = 0.9493, Accuracy = 88.05%

📊 Results Summary

Model Type	Embedding Direction	Accuracy	Precision	Recall	F1-Score	AUC
Random Forest (Tuned)	Forward	97.50%	0.97-0.98	0.97-0.98	0.97-0.98	0.9753
Neural Network	Backward	88.05%	0.87-0.90	0.85-0.91	0.87-0.89	0.9493

🛠️ Installation & Setup

Prerequisites

• Python 3.9.21
• CUDA-compatible GPU (recommended for DNABERT processing)
• Azure VM or similar cloud compute environment

Required Libraries

pip install torch transformers
pip install biopython pandas pyarrow numpy scikit-learn
pip install optuna umap-learn
pip install duckdb  # For efficient Parquet querying

DNABERT Model

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("zhihan1996/DNA_bert_6", trust_remote_code=True)
model = AutoModel.from_pretrained("zhihan1996/DNA_bert_6", trust_remote_code=True)

🔧 Usage

1. Data Preprocessing

# Clean and filter sequences
jupyter notebook cleaning_sequences.ipynb

# Process forward and backward strands
jupyter notebook data_read_forward.ipynb
jupyter notebook data_read_backward.ipynb

2. Generate Embeddings

# Run DNABERT embedding generation
python embeddings.py --input_dir ./clean_sequences --output_dir ./embeddings

3. Train Models

# Train forward embeddings model
jupyter notebook forw_train_df.ipynb

# Train backward embeddings model  
jupyter notebook backw_train_df.ipynb

📁 Data Organization

Cleaned Datasets Structure

clean_forward_reads/     # Forward cancerous sequences (604 batches, 28M sequences)
clean_backward_reads/    # Backward cancerous sequences (604 batches, 9M sequences)  
clean_forward_noncan/    # Forward non-cancerous sequences (553 batches, 5.7M sequences)
clean_backward_noncan/   # Backward non-cancerous sequences (553 batches, 3.8M sequences)

Embedding Files Structure

embeddings/
├── forward_cancerous_embeddings.npy     # 768-dim vectors + metadata CSV
├── forward_noncancerous_embeddings.npy  # 768-dim vectors + metadata CSV
├── backward_cancerous_embeddings.npy    # 768-dim vectors + metadata CSV
└── backward_noncancerous_embeddings.npy # 768-dim vectors + metadata CSV

🔬 Technical Implementation Details

Infrastructure

• Platform: Azure Virtual Machine (cloud-hosted)
• Access: SSH with Visual Studio Code remote development
• Storage: Efficient Parquet format for large genomic datasets
• Monitoring: Weights & Biases for experiment tracking

Key Technologies

• DNABERT: Pre-trained transformer for genomic sequences with 6-mer tokenization
• Optuna: Hyperparameter optimization framework
• UMAP: Dimensionality reduction for visualization
• DuckDB: Lightweight querying for large Parquet files

Quality Metrics

• Total Sequences Processed: 46,968,954 high-quality sequences
• Quality Threshold: Phred Q30 (1 in 1,000 error rate)
• Batch Processing: 100,000 sequences per batch for memory efficiency

📈 Key Findings

1. Bidirectional Approach: Forward and backward embeddings show different separability patterns in UMAP visualization
2. Model Selection: Random Forest optimal for forward embeddings; Neural Networks better for backward embeddings
3. Quality Impact: Q30 filtering significantly reduces dataset size but improves classification accuracy
4. Performance: Both models achieve high accuracy (>88%) demonstrating feasibility of genomic-based cancer classification

Future Work

Future Directions

• Multi-class cancer type classification
• Integration with clinical data
• Real-time diagnostic applications
• Cross-population validation studies

🤝 Contributing

This is an academic research project. For collaboration opportunities or questions:

• Aakash Walavalkar: aakash.muskurahat@gmail.com
• Anushka Kumar: anushka.ayyanar@gmail.com
• Laavanya Mishra: mishralaavanya@gmail.com

📚 References

The complete reference list is available in the research paper. Key references include:

1. DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language
2. National Genomics Data Center (NGDC) datasets
3. Apache Parquet for efficient genomic data storage
4. Optuna for automated hyperparameter optimization

📄 License

This project is for academic research purposes. Please cite the paper if you use this work in your research.

---

Last Updated: September 2025
Version: 1.0
Status: Research Complete