ML-Based-Multi-Cancer-Classification-Using-Gene-Expression

Overview

This project is a Bioinformatics and Machine Learning application developed to classify cancer types using gene expression microarray data.

The system predicts whether a sample belongs to:

• Breast Cancer
• Liver Cancer
• Lung Cancer
• Skin Cancer
• No Cancer (among the above four cancer types)

An ensemble learning approach combining SVM, Random Forest, KNN, and XGBoost models is used to improve prediction performance.

---

Live Demo (deployed in render)

🔗 Live Application:

https://ml-based-cancer-classification-using-187j.onrender.com

Dataset Information

• Dataset Sources

The gene expression datasets used in this project were obtained from the NCBI Gene Expression Omnibus (GEO) database.

NCBI GEO Repository

https://www.ncbi.nlm.nih.gov/geo/

GEO Datasets Used

Cancer Type	GEO Accession	Dataset Link
Skin Cancer (Melanoma / Nevus)	GSE3189	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE3189
Liver Cancer (HCC)	GSE14520	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE14520
Breast Cancer	GSE15852	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE15852
Lung Cancer	GSE19188	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE19188
Colon Cancer (Exploratory Dataset)	GSE44076	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE44076

Final Dataset

After merging and preprocessing:

• Total Samples: 757

• Original Gene Features: 22,268

• Final Selected Features: 1,268

• Cancer Classes:

  • Breast Cancer
  • Liver Cancer
  • Lung Cancer
  • Skin Cancer
  • No Cancer

The final dataset was created by combining the GEO datasets, standardizing gene expression values, and applying feature selection using Variance Thresholding, Mutual Information, and LASSO.

Dataset Statistics

Item	Value
Total Samples	757
Original Gene Features	22,268
Metadata Columns	5
Final Selected Features	1,268

Cancer Categories

• Breast Cancer
• Liver Cancer (HCC)
• Lung Cancer (ADC, SCC, LCC)
• Skin Cancer (Melanoma, Nevus)
• Healthy / Normal Samples

---

Data Preprocessing Pipeline

1. Variance Thresholding

Removed low-variance genes that contribute little to classification.

Threshold	Features Remaining
1000	21,440
5000	19,192
10000	17,626

Selected Threshold = 5000

---

2. Standard Scaling

Gene expression values were normalized using StandardScaler:

• Mean = 0
• Standard Deviation = 1

This ensures all genes contribute equally during training.

---

3. Mutual Information Feature Selection

Selected the most informative genes.

Top Features Selected
5,000
10,000
15,000

Selected:

Top 10,000 genes

---

4. LASSO Feature Selection

Further reduced dimensionality using L1 Regularization.

C Value	Features Selected
0.1	60
0.3	129
0.5	189
1.0	457
2.0	1,268
5.0	4,124
10.0	6,405

Final Selection:

• C = 2.0
• Features = 1,268

---

Machine Learning Models

Support Vector Machine (SVM)

Best Parameters:

• Kernel = Linear
• C = 100
• Gamma = Scale

Results:

• Accuracy = 91.45%
• F1 Score = 91.81%

---

Random Forest

Best Parameters:

• n_estimators = 200
• max_depth = 20
• min_samples_split = 2
• min_samples_leaf = 1

Results:

• Accuracy = 92.76%
• F1 Score = 92.95%

---

K-Nearest Neighbors (KNN)

Best Parameters:

• n_neighbors = 9
• weights = distance
• p = 2

Results:

• Accuracy = 94.08%
• F1 Score = 94.13%

---

XGBoost

Best Parameters:

• n_estimators = 300
• max_depth = 6
• learning_rate = 0.3
• subsample = 0.8
• colsample_bytree = 0.8

Results:

• Accuracy = 93.42%
• F1 Score = 93.37%

---

Ensemble Learning

Soft Voting Ensemble combining:

• SVM
• Random Forest
• KNN
• XGBoost

Final Performance

Metric	Score
Accuracy	95.39%
Precision	95.55%
Recall	95.39%
F1 Score	95.44%

Best Model: Voting Ensemble

---

Web Application

A Flask-based web application was developed for deployment.

Features:

• Upload gene expression CSV file
• Automatic preprocessing
• Cancer type prediction
• Download sample input file
• Interactive prediction results page

---

Technologies Used

• Python
• Flask
• Pandas
• NumPy
• Scikit-Learn
• XGBoost
• Joblib
• Bootstrap 5
• HTML/CSS

---

Project Structure

Bio/
│
├── app.py
├── requirements.txt
├── merged_LASSO_c_2.csv
│
├── final_models/
│   ├── best_model.pkl
│   ├── scaler.pkl
│   ├── label_encoder.pkl
│
├── templates/
│   ├── index.html
│   └── result.html
│
├── sample_input.csv
└── README.md

Academic Project

Developed as part of a Bioinformatics Machine Learning project for cancer classification using gene expression data.