🩸 Early Cancer Detection: A Transcriptomic Fingerprint in a Single Blood Drop

Multi-cancer early detection from platelet RNA sequencing (RNA-seq). Platelets act like "molecular sponges," absorbing RNA released by tumors — effectively carrying the cancer's transcriptomic fingerprint throughout the bloodstream. This project builds a bioinformatics + machine learning pipeline that reads that fingerprint to distinguish cancer patients from healthy controls.

⚠️ Research and educational project. Not a medical device and not for clinical use.

---

📋 Table of Contents

• The Challenge
• The Approach
• Pipeline
• Step 1 — Data Preprocessing & Bioinformatics
• Step 2 — Machine Learning Pipeline
• Step 3 — Results
• Explainable AI (SHAP)
• Repository Structure
• Tech Stack
• Limitations

---

🎯 The Challenge

There is a need for non-invasive, early-stage cancer screening that works across multiple cancer types. This project explores platelet RNA-seq as that signal source.

• Key insight: platelets absorb tumor-released RNA, carrying a cancer transcriptomic fingerprint in the blood.
• Data overview: 2,351 RNA-seq samples (cancer patients vs. healthy controls).

---

🧠 The Approach

A two-stage pipeline: a bioinformatics stage (R) that cleans and normalizes raw RNA-seq counts and finds differentially expressed genes, followed by a machine learning stage (Python) that trains and stacks multiple models into a final ensemble.

---

🔗 Pipeline

Raw RNA-seq counts
      │
      ▼
Bioinformatics (ThromboSeq, R)  →  Normalization · Batch correction · DE analysis · Filtering
      │
      ▼
Feature engineering (Python)    →  Top 300 differentially expressed genes + SMOTE
      │
      ▼
Model training                  →  XGBoost · LightGBM · SVM · NN · Decision Tree
      │
      ▼
Stacking meta-model (Champion)  →  Prediction + SHAP explainability

---

🧬 Step 1 — Data Preprocessing & Bioinformatics

Raw RNA-seq count matrices from multiple centers were inconsistent and noisy, so the project uses the specialized ThromboSeq bioinformatics pipeline:

• Normalization — DESeq2 / edgeR for RNA-seq count data.
• Batch correction — RUVSeq to eliminate technical variation between sequencing centers.
• Differential expression analysis — identifying genes (cancer vs. control).
• Filtering — multiple testing correction (FDR < 0.005).

---

🤖 Step 2 — Machine Learning Pipeline

Feature engineering

• Selection of 300 differentially expressed genes as model input features, based on correlation factor.
• SMOTE applied to handle class imbalance (original distribution: 80.6% cancer / 19.4% control → balanced to 50/50).

Models

Group	Models	Role
🌳 Tree Masters	XGBoost, LightGBM	Efficient, high-performance
🔍 Pattern Hunters	SVM, Neural Network	Capture complex patterns
📏 Interpretable Baseline	Decision Tree	For comparison
🏆 The Champion	Ensemble Stacking Meta-Model	Combines all models' predictions

---

📊 Step 3 — Results

Overall highlight — Stacking Meta-Model: 95.8% AUC

Metric	Value
Accuracy	0.9481
Balanced accuracy	0.9240
Sensitivity	0.9633
Specificity	0.8846
Precision	0.9722
F1	0.9677
AUC	0.9575
Average precision	0.9854

Robustness across test settings

Test	Setting	AUC
Test 1	Asymptomatic controls (clean cancer signal)	95.8%
Test 3	Symptomatic patients (noise from inflammation, medication, other diseases)	86.8%

The real win — early-stage detection

Stage	Detection accuracy	Sample size
Stage I	100%	n = 19
Stage II	100%	n = 35
Later stages	~95%	—

Detecting Stage I and II cancers is the clinically hardest and most valuable case — this is where early, non-invasive screening matters most.

---

🔬 Explainable AI (SHAP)

• Goal — identify which transcripts drive the model's predictions.
• Biomarker gold — SHAP revealed genes such as ENSG00000158578 as key cancer signatures.
• Biological insight — these genes can be fed into Gene Set Enrichment (GSE) analysis to reveal the biological pathways they affect.

---

📁 Repository Structure

.
├── count-matrix-preprocessing.r        # Count matrix preprocessing (R)
├── thromboSeqTools_PreProcessing.R     # ThromboSeq preprocessing utilities (R)
├── thromboSeqTools_ANOVA.R             # ThromboSeq ANOVA / differential expression (R)
├── models-training.ipynb               # ML pipeline: training, evaluation, SHAP (notebook)
├── models-training.py                  # ML pipeline (script version)
└── .Rhistory                           # R session history

---

🛠️ Tech Stack

Bioinformatics (R): ThromboSeq · DESeq2 · edgeR · RUVSeq Machine Learning (Python): scikit-learn · XGBoost · LightGBM · imbalanced-learn (SMOTE) · SHAP

---

⚠️ Limitations

• Symptomatic-patient performance drops (AUC 86.8%) — real-world noise from inflammation, medication, and comorbidities reduces the signal.
• Small early-stage samples — Stage I (n = 19) and Stage II (n = 35) accuracies are promising but based on limited data; larger cohorts are needed to confirm.
• Results reflect this dataset and may not transfer directly to unseen populations or sequencing centers without re-validation.

---

Reading the cancer's fingerprint — from a single blood drop. 🩸