This project presents a machine learning workflow for classification of glioblastoma (GBM) and non-tumor brain samples using gene expression data.
The analysis was performed in Python using the scikit-learn machine learning framework.
The aim of the project was to investigate whether gene expression profiles can be used to distinguish glioblastoma samples from non-tumor brain samples.
The workflow includes gene expression preprocessing, probe annotation, gene-level aggregation, feature selection, machine learning classification, model evaluation, and cross-validation.
GSE4290 (GEO)
│
▼
Affymetrix expression matrix
54,613 probes
│
▼
Sample selection
GBM vs Non-Tumor samples
│
▼
GPL570 annotation
│
▼
Probe → Gene mapping
│
▼
Gene-level expression
22,189 genes
│
▼
Feature Selection
Top 100 genes
│
┌─────────┴─────────┐
▼ ▼
Logistic Regression Random Forest
│ │
└─────────┬─────────┘
▼
Model Evaluation
│
┌─────────────┴─────────────┐
▼ ▼
Test Set Evaluation 5-Fold Cross-Validation
│ │
└─────────────┬─────────────┘
▼
Performance Metrics
Accuracy / Precision / Recall
F1-score / ROC-AUC
Can gene expression profiles be used to distinguish glioblastoma (GBM) from non-tumor brain samples using machine learning?
| Information | Value |
|---|---|
| Database | Gene Expression Omnibus (GEO) |
| Accession | GSE4290 |
| Organism | Homo sapiens |
| Platform | Affymetrix Human Genome U133 Plus 2.0 Array |
| Platform ID | GPL570 |
| Data type | Gene expression microarray |
| Total samples | 100 |
| GBM samples | 77 |
| Non-tumor samples | 23 |
| Original probes | 54,613 |
| Genes after annotation | 22,189 |
| Selected features | 100 |
Dataset:
https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE4290
The original GEO expression data are not included in this repository.
- Download GSE4290 expression matrix from GEO
- Select GBM and non-tumor samples
- Inspect sample diagnostic information
- Load GPL570 annotation
- Map Affymetrix probes to gene symbols
- Aggregate multiple probes corresponding to the same gene
- Generate a gene-level expression matrix
- Select the 100 most informative genes
- Create a stratified training and test set
- Train Logistic Regression classifier
- Train Random Forest classifier
- Evaluate classification performance
- Generate confusion matrices
- Generate ROC curves
- Compare model performance
- Perform 5-fold stratified cross-validation
- Compare model stability across validation folds
The gene-level expression matrix contained 22,189 genes.
Because the number of genes was much larger than the number of samples, feature selection was performed using:
SelectKBest + ANOVA F-test (f_classif)
The 100 highest-ranking genes were selected as features for classification.
Feature selection was incorporated into the machine learning pipeline to reduce the risk of information leakage during cross-validation.
Logistic Regression was used as a linear classification model.
The selected gene expression features were standardized before model training.
The model was used to determine whether a linear relationship between gene expression patterns and the diagnostic groups could distinguish GBM from non-tumor samples.
Random Forest was used as a nonlinear classification model based on an ensemble of decision trees.
The model was used to determine whether more complex relationships between gene expression features could improve classification performance.
The dataset was divided using a stratified train/test split:
- Training set: 80 samples
- Test set: 20 samples
| Model | Accuracy | Precision | Recall | F1-score | ROC-AUC |
|---|---|---|---|---|---|
| Logistic Regression | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 |
| Random Forest | 0.95 | 1.00 | 0.93 | 0.97 | 1.00 |
Logistic Regression correctly classified all 20 samples in the test set.
Random Forest correctly classified 19 out of 20 test samples.
To assess model performance across multiple data partitions, 5-fold stratified cross-validation was performed.
| Model | Accuracy | Precision | Recall | F1-score | ROC-AUC |
|---|---|---|---|---|---|
| Logistic Regression | 0.950 ± 0.055 | 0.988 ± 0.024 | 0.947 ± 0.078 | 0.965 ± 0.041 | 0.997 ± 0.005 |
| Random Forest | 0.950 ± 0.055 | 0.988 ± 0.024 | 0.947 ± 0.078 | 0.965 ± 0.041 | 0.992 ± 0.016 |
Logistic Regression achieved a slightly higher and more stable ROC-AUC than Random Forest during cross-validation.
The ROC curve compares the classification performance of Logistic Regression and Random Forest across different classification thresholds.
The model comparison summarizes the main classification metrics obtained on the test set.
The confusion matrix shows the classification results obtained by Logistic Regression on the test set.
The confusion matrix shows the classification results obtained by Random Forest on the test set.
The cross-validation comparison shows the performance of both machine learning models across the five validation folds.
The ROC-AUC plot illustrates the stability of model performance across the five cross-validation folds.
The machine learning analysis demonstrated that gene expression profiles can effectively distinguish glioblastoma from non-tumor samples within the GSE4290 dataset.
Both Logistic Regression and Random Forest achieved high classification performance.
Logistic Regression achieved the strongest cross-validation performance, with a mean ROC-AUC of:
0.997 ± 0.005
Random Forest achieved:
0.992 ± 0.016
The results indicate that the selected gene expression features contain strong information associated with the distinction between GBM and non-tumor samples.
This project demonstrates practical experience with:
- Gene expression data analysis
- Microarray data preprocessing
- Affymetrix probe annotation
- Probe-to-gene mapping
- Gene-level feature construction
- Statistical feature selection
- Machine learning classification
- Logistic Regression
- Random Forest
- Train/test splitting
- Stratified cross-validation
- Model performance evaluation
- ROC-AUC analysis
- Confusion matrix visualization
- Scientific data visualization in Python
- Reproducible machine learning workflows
- Python programming
- Linux / WSL
- Git & GitHub project organization
- Python 3.11
- pandas
- NumPy
- scikit-learn
- matplotlib
- seaborn
- PyYAML
- NCBI GEO
- Affymetrix GPL570
Possible extensions of this project include:
- Validation using an independent GBM dataset
- Hyperparameter optimization
- Comparison with Support Vector Machines
- Gradient Boosting or XGBoost classification
- Recursive feature elimination
- Biological interpretation of selected genes
- Pathway enrichment analysis of machine learning features
- Survival analysis using independent clinical datasets
04-Machine-Learning-GBM
│
├── data
│ └── raw
│
├── figures
│ ├── confusion_matrix_logistic_regression.png
│ ├── confusion_matrix_random_forest.png
│ ├── ROC_curve_models.png
│ ├── model_performance_comparison.png
│ ├── cross_validation_model_comparison.png
│ └── cross_validation_ROC_AUC.png
│
├── results
│ ├── model_performance.csv
│ ├── cross_validation_summary.csv
│ ├── cross_validation_fold_results.csv
│ ├── selected_features.csv
│ ├── train_test_split.csv
│ └── test_predictions.csv
│
├── scripts
│ ├── 01_prepare_data.py
│ ├── 02_feature_selection.py
│ ├── 03_train_models.py
│ ├── 04_evaluate_models.py
│ └── 05_external_validation.py
│
├── .gitignore
├── LICENSE
└── README.md
The project was developed using Python 3.11 in a dedicated Conda environment under Linux/WSL.
Create the environment:
conda create -n ml-gbm python=3.11
conda activate ml-gbmInstall the required packages:
pip install pandas numpy scikit-learn matplotlib seaborn pyyamlDownload the GSE4290 expression matrix and GPL570 annotation from GEO and place them in:
data/raw/
Run the analysis scripts sequentially:
python scripts/01_prepare_data.py
python scripts/02_feature_selection.py
python scripts/03_train_models.py
python scripts/04_evaluate_models.py
python scripts/05_external_validation.pyThe analysis was performed on a relatively small dataset containing 100 samples.
The classes are also imbalanced, with more GBM samples than non-tumor samples.
The very high classification performance should therefore be interpreted in the context of this particular dataset.
The models are not clinically validated.
Independent external validation using a separate dataset would be required to determine whether the observed performance generalizes to other patient cohorts.
Katarzyna Zielińska
Bioinformatics Portfolio
2026
Created as part of a Bioinformatics Portfolio project focused on machine learning applications in cancer genomics.