Multi-Omics Cancer Classification using Foundation Models

This project implements a multi-omics approach to cancer classification, leveraging transfer learning from foundation models specialized in biological data. We integrate DNA Methylation and RNA-seq data from TCGA (The Cancer Genome Atlas) to improve diagnostic accuracy across multiple cancer types.

Motivation

Cancer is a complex disease requiring analysis across multiple molecular levels. Foundation models like MethylGPT (or CpGPT) for methylation and Geneformer for transcriptomics offer rich representations that capture intricate biological dependencies better than traditional methods.

Methodology

The project follows a multi-phase approach:

1. Parallel Preprocessing:
   • RNA-seq: Processed for Geneformer to capture gene interactions.
   • DNA Methylation: Processed for MethylGPT to extract epigenetic embeddings.
2. Fusion Strategies:
   • Early Fusion: Concatenating embeddings before classification.
   • Late Fusion: Ensemble of specialized models for each modality.
3. Classification: Final layer utilizing Neural Networks or XGBoost with k-fold cross-validation.

Data Acquisition

We target 5-6 primary cancer types from TCGA:

• Breast (TCGA-BRCA)
• Lung (TCGA-LUAD)
• Colon (TCGA-COAD)
• Kidney (TCGA-KIRC)
• Liver (TCGA-LIHC)
• Thyroid (TCGA-THCA)

Data is downloaded using the GDC API, specifically filtering for patients with matched RNA-seq and Methylation 450k data.

Initial Setup

Prerequisites

• Python 3.8+
• Disk Space: ~50-60GB for full dataset (currently using subset for testing)

Installation

pip install -r requirements.txt

Data Download

To download the initial test subset:

python scripts/download_tcga_data.py