Batch-Harmonized AI for Multi-Center Pancreatic Cancer RNA Data

A reproducible pipeline for cross-cohort RNA biomarker discovery using batch correction and machine learning

---

Project Overview

This project develops an AI-powered framework to harmonize RNA expression data across multiple cancer datasets (TCGA + GEO) and build robust machine learning classifiers for pancreatic cancer subtyping. The goal is to create a reproducible, clinically-translatable pipeline that addresses batch effects, a major barrier in multi-center biomarker studies.

Why Pancreatic Cancer?

• High unmet need: 5-year survival rate <10%
• Understudied: Less AI/ML work compared to breast/lung cancer
• Clear subtypes: Classical, Basal-like, with distinct prognosis
• Novel contribution: Limited prior work on cross-cohort harmonization

---

Research Objectives

1. Harmonize RNA data across TCGA and GEO datasets using multiple batch correction methods
2. Build ML classifiers (RandomForest, XGBoost) for cancer subtype prediction
3. Validate cross-cohort performance with true external validation
4. Deploy interactive tool via Shiny app for reproducible analysis

---

Project Structure

rna-harmonization-ai/
├── data/
│   ├── raw/              # Original TCGA & GEO downloads
│   └── processed/        # Cleaned, normalized, harmonized data
├── scripts/
│   ├── 00_setup_environment.R      # Install all packages
│   ├── 01_download_tcga_data.R     # Download TCGA-PAAD
│   ├── 02_download_geo_data.R      # Download GEO validation sets
│   ├── 03_eda_initial.R            # Exploratory data analysis
│   ├── 04_preprocessing.R          # Normalization & feature alignment
│   ├── 05_batch_correction.R       # ComBat, Harmony, MNN comparison
│   ├── 06_ml_models.R              # Train ML classifiers
│   └── 07_evaluation.R             # Cross-cohort validation
├── figures/              # All generated plots
├── results/              # Model outputs, performance metrics
├── shiny_app/            # Interactive web application
└── docs/                 # Manuscript drafts, supplementary materials

---

Quick Start

1. Setup Environment (First Time Only)

# In RStudio, run:
source("scripts/00_setup_environment.R")

This installs all required packages (~20-30 minutes).

2. Download Data

# Download TCGA pancreatic cancer data (~2-3 GB, 15-30 min)
source("scripts/01_download_tcga_data.R")

# Download GEO validation datasets (~5-10 min)
source("scripts/02_download_geo_data.R")

3. Run Initial EDA

# Explore data and identify batch effects
source("scripts/03_eda_initial.R")

# Check generated figures in figures/ folder

---

Datasets

Primary Dataset: TCGA-PAAD

• Project: The Cancer Genome Atlas - Pancreatic Adenocarcinoma
• Samples: 178 primary tumors
• Data type: RNA-seq (STAR aligned, raw counts)
• Platform: Illumina HiSeq
• Clinical data: Vital status, stage, demographics

Validation Dataset: GEO

• GSE71729: RNA-seq PDAC cohort (357 samples, Illumina)

Note: We focused on RNA-seq datasets only to maximize gene overlap (14,137 common genes vs. <100 with microarray inclusion). This provides better feature coverage for robust machine learning models.

---

Methods Pipeline

1. Preprocessing

• Normalize RNA counts (DESeq2 VST or log2 CPM)
• Filter low-count genes
• Align gene features across platforms
• Handle missing values

2. Batch Correction (Comparison Study)

• ComBat (empirical Bayes)
• Harmony (PCA-based)
• MNN (mutual nearest neighbors)
• Evaluate: PCA mixing, silhouette scores, kBET

3. Machine Learning

• RandomForest: Interpretable, robust to outliers
• XGBoost: High performance gradient boosting
• Neural Network (optional): Deep learning approach

4. Evaluation Strategy

• Training: TCGA-PAAD (in-distribution)
• Testing: GEO datasets (out-of-distribution)
• Metrics: Accuracy, F1-score, AUROC
• Interpretability: Feature importance, SHAP values

---

Expected Outcomes

Key Results Demonstrated

1. Gene coverage: 14,137 common genes between TCGA and GEO RNA-seq
2. Batch correction: ComBat successfully harmonizes datasets (PCA visualization)
3. Cross-cohort classification: Models trained on TCGA (178 samples) predict on GEO (357 samples)
4. Stable biomarkers: 5 top genes (LAMC2, DKK1, ITGB6, GPRC5A, MAL2) with biological relevance
5. Survival prediction: Alive vs Dead classification (85 vs 93 samples, balanced classes)

Figures for Manuscript

• PCA before/after batch correction
• Cross-cohort classification performance
• Feature importance and biological interpretation
• Shiny app screenshots

---

Technology Stack

• Language: R 4.0+
• Data Access: TCGAbiolinks, GEOquery
• Preprocessing: DESeq2, edgeR, limma
• Batch Correction: sva (ComBat), harmony, batchelor (MNN)
• Machine Learning: randomForest, xgboost, caret
• Visualization: ggplot2, pheatmap, ComplexHeatmap
• Deployment: Shiny, shinydashboard

---

Manuscript Roadmap

Target Journal

• bioRxiv (preprint) → Bioinformatics or Briefings in Bioinformatics

Estimated Timeline

• Weeks 1-2: Data download & EDA
• Weeks 3-4: Batch correction comparison
• Weeks 5-6: ML model development
• Weeks 7-8: Cross-cohort validation
• Weeks 9-10: Shiny app development
• Weeks 11-12: Manuscript writing

Novelty Statement

"We present the first open-source R pipeline that integrates batch correction (ComBat) with machine learning for robust, cross-cohort RNA biomarker discovery in pancreatic cancer. By focusing on RNA-seq platforms, we achieved high gene coverage (14,137 genes) enabling identification of 5 novel prognostic biomarkers. Our Shiny application enables reproducible analysis and accelerates clinical translation."

---

Contributing

This is a research project. For questions or collaborations, please open an issue or contact the author.

---

Key References

1. Johnson et al. (2007) - ComBat batch correction
2. Korsunsky et al. (2019) - Harmony algorithm
3. Haghverdi et al. (2018) - MNN method
4. TCGA Research Network - Pancreatic cancer characterization

---

License

MIT License - Open source for academic and commercial use.

---

Project Status: Active Development
Last Updated: October 2025
Author: Mark Barsoum Markarian