Gene Expression–Based Breast Cancer Classifier

by mikeph_

A bioinformatics machine learning workflow using publicly available gene expression data from GEO.
The goal is to classify cancer vs normal samples based on transcriptomic profiles and to practice proper evaluation, validation, and interpretation.

---

Project Overview

• Dataset: GEO accession GSE45827
• Task: Binary classification (Cancer vs Normal)
• Data type: Microarray gene expression
• Samples: 155
• Genes: ~30,000
• Model: Logistic Regression
• Language: Python

---

Dataset Description

• Source: Gene Expression Omnibus (GEO)
• Platform: Microarray
• Labels were extracted from !Sample_characteristics_ch1 metadata
• Samples containing keywords such as "cancer" or "tumor" were labeled as cancer

---

Methods

Preprocessing

• Transposed expression matrix to samples × genes
• Converted all values to numeric
• Missing values filled using gene-wise means

Feature Selection

• Removed low-variance genes using VarianceThreshold
• Reduced dimensionality from ~30,000 to ~200 genes

Model

• Logistic Regression
• Evaluated using:
  • Train/test split
  • Stratified k-fold cross-validation
  • Permutation testing

---

Results

On Training Dataset:

Metric	Value
Accuracy	1.00
ROC-AUC	1.00
CV ROC-AUC	1.00 ± 0.00
Permutation Test p-value	~0.01

Interpretation:
The classifier perfectly separates cancer and normal samples, indicating a strong biological signal in gene expression data. Permutation testing confirms that performance is statistically significant and not due to chance.

---

Visualizations

• ROC curve demonstrates near-perfect classification performance

  

• PCA plot shows clear separation between cancer and normal samples

  

• Volcano plot visualizes the differential gene expression

  

---

⚠️ Notes on Evaluation

• ROC-AUC is not defined for Leave-One-Out CV in imbalanced datasets
• Accuracy was used instead for LOOCV
• This behavior is expected and reflects known statistical limitations

---

Key Outputs

• important_genes.csv: genes ranked by logistic regression coefficients
• expression_selected_genes.csv: reduced feature matrix
• Exported train/test splits for reproducibility

---

Results on external datasets

GSE42568:

GEO NCBI

Class	Precision	Recall	F1‑score	Support
0	0.61	0.82	0.70	17
1	0.97	0.91	0.94	104
Macro avg	0.79	0.87	0.82	121
Weighted avg	0.92	0.90	0.91	121

• ROC-AUC

  

• PCA

  

• PCA Training Data vs GSE42568

  

---

GSE61304:

GEO NCBI

• Characteristics

  Found characteristic rows: 11
  Selected diagnosis row index: 1
  Example diagnosis values: ['diagnosis: Adenocarcinoma of breast, ductal'
  'diagnosis: Adenocarcinoma of breast, ductal, mucinous'
  'diagnosis: Within normal limits'
  'diagnosis: Adenocarcinoma of breast, ductal'
  'diagnosis: Adenocarcinoma of breast, ductal']
  External labels (0=Normal, 1=Cancer): [ 4 58]
  Common genes: 200
  

  Class	Precision	Recall	F1‑score	Support
  0	0.00	0.00	0.00	4
  1	0.93	0.86	0.89	58
  Macro avg	0.46	0.43	0.45	62
  Weighted avg	0.87	0.81	0.84	62

• ROC-AUC

  

• PCA

  

• PCA Training Data vs GSE61304

  

---

GSE10810

GEO NCBI

• Characteristics

  Found characteristic rows: 8
  External labels (0=Normal, 1=Cancer): [27 31]
  Unique labels: [0 1]
  External data shape after aligning with training genes: (58, 200)
  Scaler expects features: 200
  External data successfully scaled.
  

  Class	Precision	Recall	F1-score	Support
  0	0.80	0.74	0.77	27
  1	0.79	0.84	0.81	31
  Accuracy			0.79	58
  Macro Avg	0.79	0.79	0.79	58
  Weighted Avg	0.79	0.79	0.79	58

• ROC-AUC

  

• PCA

  

• PCA Training Data vs GSE61304

  

---

GSE65194

GEO NCBI

• Characteristics

  Found characteristic rows: 4
  External labels (0=Normal, 1=Cancer): [ 11 153]
  Unique labels: [0 1]
  External data shape after aligning with training genes: (164, 200)
  Scaler expects features: 200
  External data successfully scaled.
  

Best threshold for F1: 0.9561

Class	Precision	Recall	F1-score	Support
0 (Normal)	1.00	0.64	0.78	11
1 (Cancer)	0.97	1.00	0.99	153
Accuracy			0.98	164
Macro avg	0.99	0.82	0.88	164
Weighted avg	0.98	0.98	0.97	164

• ROC-AUC

  

• PCA

  

• PCA Training Data vs GSE61304

  

---

How to Run

Install depedencies

pip install depedencies.txt

Download datasets

Because of the file limits of github, the datasets are available here.