Machine Learning For Biomarker Discovery for Alzheimer's Disease

This repository contains a complete machine learning workflow for biomarker discovery, from raw gene expression data to identifying genes as potential biomarkers. The project demonstrates pre-processing, feature scaling, dimensionality reduction, hyperparameter tuning, cross-validation, and feature attribution (SHAP) using Python.

---

Project Overview

The goal of this project is to identify potential biomarkers from gene expression data. The workflow includes:

Clinical Relevance: Top biomarkers (APP, APOE, PSEN1) align with known AD mechanisms involving amyloid processing and lipid metabolism, validating model's biological interpretability.

Folder Structure

ML_alzheimer_gene_/
├── scripts/            # Analysis scripts (Python)
├── results/            # Evaluation performance data
├── plots/              # Distribution plots, PCA plots, confusion matrix
├── config.yml          # configuration file
└── README.md           # Project documentation

---

Data Overview

Samples: 206 (control vs. condition) Features: 19,297 genes (RNA-seq expression values)

Data Pre-processing

• Samples were balanced and did not need resampling

- Removed batch ID column (no batch effect detected) - Label encoding: Control → 0, Condition → 1 - Train-test split: 80/20 stratified by class

Methodology

Dimensionality Reduction: PCA vs. Raw Features

Compared model performance with/without PCA (95% variance = 213 components)

Experiment	XGBoost Accuracy	Random Forest	Logistic Regression
Raw Features	96.55%	93.10%	80.70%
PCA-performed	60.34%	60.33%	46.62%

• 36% reduced accuracy with PCA
• Raw features preserve biological interpretations Proceeded with raw features

Model Selection & Optimization

From different ML models, 3 models were selected based on accuracy - worst, mid and best Model selected: Logistic Regression, Random Forest, XGBoost Optimisation: RandomizedSearchCV with 5-fold CV (accuracy metric) *Final Evaluation (10-fold Stratified CV on entire dataset)

Model	Accuracy	Precision	F1 Score	ROC-AUC
XGBoost	94.43 ± 3.86%	98.75 ± 3.75%	94.10 ± 4.13%	0.995 ± 0.025
Random Forest	89.53 ± 4.37%	93.13 ± 5.63%	88.64 ± 5.04%	0.974 ± 0.014
Logistic Regression	80.70 ± 6.88%	83.23 ± 5.72%	82.00 ± 6.47%	0.867 ± 0.068

Best Model: XGBoost

• Best accuracy and stability (lowest variance)
• Compatible with SHAP for feature importance analysis

Biomarker Discovery: SHAP Feature Attribution

SHAP (SHapley Additive exPlanations) reveals:

• Individual gene contribution
• Regulation of gene; upregulated or downregulated
• Based on gene interaction where contribution are measured.

Conclusion

The project concludes XGBoost as the most robust classifier for this dataset. While PCA helps in reducing the number of features involved, much of the information with important discriminative power is lost. Key genes as potential biomarker's for Alzheimer's Disease was also identified. Thus, demonstrates that machine learning techniques can be used reveal biological insights.

---

Tools and Libraries

• Python 3.10+
• scikit-learn
• XGBoost
• SHAP
• pandas, numpy
• matplotlib, seaborn
• joblib