🧬 AI-Driven Multi-Omics Gastric Cancer Subtyping

A Biological Hybrid AI pipeline for molecular subtyping of gastric cancer integrating WES mutational signatures, DNA methylation, and clinical data

Results • Architecture • Install • Quick Start • Methodology

---

📋 Overview

An end-to-end multi-omics pipeline for classifying TCGA-STAD gastric cancer into 5 molecular subtypes — CIN, EBV, GS, MSI, and POLE.

Standard ML classifiers suffer from the accuracy paradox — a model can hit ~89% accuracy by favoring the majority CIN class while catastrophically failing on rare but clinically critical subtypes like POLE (1.9% of samples) and GS (12.3%). This pipeline solves that with the Biological Hybrid AI (BHAI) framework — deterministic biological rules for POLE and GS that supplement SMOTE-XGBoost where synthetic data alone isn't enough.

Data Layer	Source	Samples
Somatic Mutations	GDC Masked MAF files (WES)	431
DNA Methylation	Illumina 450K β-values	457
Clinical Labels	GDC + cBioPortal	375 labelled

Improved version of Machine-Learning-Mutational-Signatures
Original used MAF only → this adds DNA methylation, 20 driver gene features, and the Hybrid AI framework.

---

🏆 Key Results

Stepwise Performance

Strategy	Accuracy	Macro F1	MSI Recall	GS Recall	POLE Recall
SMOTE + XGBoost Baseline	89.3%	0.812	100%	41.3%	28.6%
+ Hybrid Biological Rules	90.5%	0.847	100%	58.7%	57.1%
+ Threshold Optimisation	91.2%	0.860	100%	63.0%	57.1%

Per-Class Breakdown (Final Model)

Subtype	Precision	Recall	F1-Score	Support
CIN	0.94	0.96	0.95	219
EBV	0.90	0.83	0.86	30
GS	0.72	0.63	0.67	46
MSI	0.99	1.00	0.99	73
POLE	0.80	0.57	0.67	7

Highlights

• POLE recall: 28.6% → 57.1% via single biological rule (SBS10 + TMB)
• GS recall: 41.3% → 63.0% via CDH1/RHOA mutation logic + threshold tuning
• MSI recall: 100% maintained across all strategies
• COSMIC reconstruction: 0.933 avg cosine similarity across 431 samples

Figures

---

🏗️ Pipeline Architecture

 ┌───────────────────────────────────────────────────────────────┐
 │                      INPUT DATA SOURCES                       │
 │  MAF Files (431)   │  Methylation (457)  │  Clinical (GDC)    │
 └───────┬────────────┴──────────┬──────────┴──────────┬─────────┘
         │                       │                     │
         ▼                       │                     │
 ┌─ Step 1 ─────────────┐       │                     │
 │ Process MAF Files     │       │                     │
 │ • 96-ch SBS Matrix    │       │                     │
 │ • 20 Driver Genes     │       │                     │
 └───────┬───────────────┘       │                     │
         ▼                       │                     │
 ┌─ Step 2 ─────────────┐       │                     │
 │ NMF Extraction (k=9)  │       │                     │
 └───────┬───────────────┘       │                     │
         ▼                       │                     │
 ┌─ Step 3 ─────────────┐       │                     │
 │ COSMIC v3.4 NNLS      │       │                     │
 │ 83 active signatures  │       │                     │
 └───────┬───────────────┘       │                     │
         │                       │                     │
         ▼                       ▼                     ▼
 ┌─ Step 4 ─────────────────────────────────────────────────────┐
 │ Multi-Omics Feature Matrix → 281 features                    │
 └───────┬──────────────────────────────────────────────────────┘
         │
         ▼
 ┌─ Step 5 ─────────────┐
 │ Methylation Processing│
 │ Two-pass, RAM-optimized│
 └───────┬───────────────┘
         ▼
 ┌─ Step 6 ─────────────────────────────────────────────────────┐
 │ Biological Hybrid AI Classification                          │
 │  POLE Rule (SBS10+TMB) │ GS Rule (CDH1/RHOA) │ SMOTE-XGB   │
 └───────┬──────────────────────────────────────────────────────┘
         ▼
 ┌─ Step 7 ─────────────┐
 │ Visualization (SHAP,  │
 │ t-SNE, UMAP, Survival)│
 └───────────────────────┘

---

📁 Repository Structure

├── 02_methylation_data/          # Processed methylation features
│   ├── methylation_features.csv
│   ├── methylation_pca.csv          (30 PCA components)
│   ├── methylation_targeted.csv     (MLH1, CDH1, CDKN2A, MGMT, BRCA1, RUNX3)
│   └── methylation_summary.csv      (CIMP score, global stats)
│
├── 03_pipeline_scripts/          # Core pipeline
│   ├── step1_process_maf.py         → SBS96 + gene features
│   ├── step2_extract_signatures.py  → NMF (k=9)
│   ├── step3_cosmic_assignment.py   → COSMIC v3.4 NNLS
│   ├── step4_build_features.py      → 281-feature matrix
│   ├── step5_process_methylation.py → Two-pass methylation
│   ├── step6_classify.py            → Hybrid AI classifier
│   ├── step7_visualize.py           → Publication figures
│   └── run_pipeline.sh              → Master runner
│
├── 07_figures/                   # Publication-ready figures
├── 08_clinical_data/             # Clinical + subtype labels
├── 09_documentation/             # Detailed docs
├── requirements.txt
├── LICENSE
└── README.md

---

⚙️ Installation

git clone https://github.com/Ahsansayz/AI-Driven-Multi-Omics-Gastric-Cancer-Subtyping.git
cd AI-Driven-Multi-Omics-Gastric-Cancer-Subtyping

python -m venv venv
source venv/bin/activate

pip install -r requirements.txt

Data Setup

1. MAF Files — Download TCGA-STAD masked somatic MAFs from GDC → 01_raw_data/maf_files/
2. COSMIC Reference — Download COSMIC v3.4 SBS signatures → 01_raw_data/cosmic_reference/
3. Methylation (optional) — Pre-processed features already included in 02_methylation_data/

---

🚀 Quick Start

# Full pipeline
bash 03_pipeline_scripts/run_pipeline.sh

# Individual steps
python 03_pipeline_scripts/step1_process_maf.py      # MAF → SBS96 + genes
python 03_pipeline_scripts/step2_extract_signatures.py # NMF extraction
python 03_pipeline_scripts/step3_cosmic_assignment.py  # COSMIC mapping
python 03_pipeline_scripts/step4_build_features.py     # Feature matrix
python 03_pipeline_scripts/step5_process_methylation.py # Methylation
python 03_pipeline_scripts/step6_classify.py           # Classification
python 03_pipeline_scripts/step7_visualize.py          # Figures

# Run specific steps only
bash 03_pipeline_scripts/run_pipeline.sh 6 7

---

🧠 Methodology

Biological Hybrid AI

Three strategies combined:

1. POLE Override — SBS10 total > threshold + TMB > 7.0 mut/Mb + MSI safeguard (SBS15 check)
2. GS Rescue — Evidence scoring: CDH1 truncating (3pt) + RHOA (2pt) + CDH1 missense (1pt) + TP53 wt (1pt). Override at score ≥ 3 when ML confidence < 0.50
3. SMOTE-XGBoost — 500 trees, stratified 5-fold CV, per-class threshold optimization

281 Feature Matrix

Block	Count	Source
COSMIC activities (abs + proportional)	166	NNLS fitting
Engineered + Clinical	15	TMB, MSI burden, age, etc.
Methylation (targeted + PCA + summary)	47	Two-pass processing
Gene mutations (20 genes × binary + count + CDH1 detail + variant stats)	53	MAF parsing
Total	281

Comparison with Original Pipeline

	Original	This Version
Data	MAF only	MAF + Methylation + Clinical
Features	~179	281
Gene Mutations	❌	✅ 20 driver genes
Methylation	❌	✅ CpGs + CIMP + PCA
Classification	XGBoost	Hybrid Biological AI
POLE Recall	~28%	57.1%
GS Recall	~41%	63.0%
Accuracy	~85%	91.2%

---

📄 License

MIT License — see LICENSE

---

_{Built for precision oncology research}