Classification of Gene Expression Levels by Microarray Datasets

This repository implements an end-to-end classification pipeline for benchmark microarray gene-expression datasets. It focuses on selecting a compact set of informative genes and comparing multiple classifiers under cross-validation and held-out testing.

Datasets

Two datasets are included in ARFF format:

• SRBCT: Multi-class classification (4 tumor types). File: SRBCT.arff. 83 instances with 2308 features (genes) and class information.
  

• Colon Tumor: Binary classification (Tumor vs Normal). File: Colon.arff. 62 instances with 2000 features, and binary class information.
  

Dataset source: https://csse.szu.edu.cn/staff/zhuzx/Datasets.html

Reference paper: Zhu et al., 2007, “Markov blanket-embedded genetic algorithm for gene selection”.

Methodology

1) Data loading and label handling

• SRBCT is loaded using scipy.io.arff.loadarff and class labels are decoded and converted to integers.
• Colon Tumor is parsed manually because the ARFF header contains duplicate gene identifiers; duplicates are disambiguated by appending suffixes (e.g., R39465_1, R39465_2). Labels are mapped to Tumor = 1, Normal = 0.
  
  

2) Exploratory data analysis (EDA)

Standard microarray EDA visualizations are generated for each dataset, including class distribution, per-gene box/violin plots, histograms, summary-statistic heatmaps, and correlation heatmaps.

3) Train/test split

Data are split into 80% training and 20% test sets before feature selection and normalization to reduce data leakage risk.

4) Feature selection (you need to run one of them)

Feature selection is run on the training set and then applied to the test set:

• Mutual Information: selects the top 50 genes.
• RFE (Recursive Feature Elimination) with Logistic Regression: selects 50 genes.
• LASSO (LassoCV): ranks coefficients and keeps the top 50 genes by absolute magnitude.
  
  

5) Normalization

Selected gene-expression features are standardized with StandardScaler, fit only on the training set, and then applied to both training and test data.

6) Modeling and evaluation

The pipeline compares five model families:

• Random Forest (RF)
• Support Vector Machine (SVM)
• Linear Discriminant Analysis (LDA)
• XGBoost
• Multi-Layer Perceptron (MLP; Keras/TensorFlow)

For RF/SVM/LDA/XGBoost, 5-fold cross-validation is performed via GridSearchCV for hyperparameter selection. For the MLP, 5-fold StratifiedKFold is performed with early stopping and learning-rate scheduling, then a final model is trained on the full training split.

Evaluation includes accuracy, F1, precision, recall, and confusion matrices, reported on the held-out test set.

Results

The following results are reported as accuracy (%) under 5-fold cross-validation (CV) and on the held-out test set after selecting 50 genes. Values are taken from the performed experiments in this repository (the exact numbers can vary with different splits, seeds, and library versions).

SRBCT (4-class)

Feature selection	RF (CV/Test)	SVM (CV/Test)	LDA (CV/Test)	XGBoost (CV/Test)	MLP (CV/Test)
Mutual Information	100.0 / 100.0	100.0 / 100.0	100.0 / 100.0	95.6 / 100.0	100.0 / 100.0
RFE	100.0 / 100.0	100.0 / 100.0	100.0 / 100.0	95.6 / 100.0	98.6 / 100.0
LASSO	96.9 / 100.0	98.5 / 94.1	96.9 / 88.2	93.9 / 94.1	95.4 / 94.1

Key observation: for SRBCT, perfect test accuracy is achieved for all models under Mutual Information and RFE in the reported runs; LASSO yields lower test accuracy for several models.

Colon Tumor (binary)

Feature selection	RF (CV/Test)	SVM (CV/Test)	LDA (CV/Test)	XGBoost (CV/Test)	MLP (CV/Test)
Mutual Information	92.0 / 76.9	94.0 / 76.9	92.0 / 84.6	89.6 / 76.9	89.8 / 84.6
RFE	86.0 / 76.9	92.0 / 76.9	92.0 / 84.6	83.8 / 84.6	94.0 / 76.9
LASSO	83.6 / 76.9	91.8 / 69.2	87.6 / 84.6	87.8 / 69.2	91.8 / 76.9

Key observation: for Colon Tumor, the best reported test accuracy is 84.6%, most consistently reached by LDA across all three feature-selection approaches.

Outputs

• Figures and confusion matrices are generated during execution.
• Example figures and result screenshots are provided under analyses/ (EDA plots, feature-selection plots, and summary comparisons).
  
  

How to run

1. Ensure SRBCT.arff and Colon.arff are in the working directory.
2. Run gene_expression_levels.py cell-by-cell (it is organized with #%% sections for an IDE workflow similar to a notebook).
   
   

Dependencies

Core: numpy, pandas, scipy, matplotlib, seaborn

ML: scikit-learn, xgboost, tensorflow (Keras)

Notes

• The Colon dataset contains duplicate gene names in the ARFF header; duplicates are resolved by creating unique column names.
• Reported results can vary slightly with different train/test splits or random seeds.

📫 Contact me at memisoguzhants@gmail.com
I'm waiting for your suggestions and contributions.