Smoker Status Classification in Lung Adenocarcinoma RNA-seq

This repository contains the code and data used for classifying smoker status in Lung Adenocarcinoma (LUAD) using RNA-seq data and machine learning techniques.

Project Overview

Lung cancer is the leading cause of cancer-related deaths worldwide, with smoking accounting for almost 85% of lung cancer cases. Classifying smoker status can aid in early diagnosis and improve the accuracy of lung cancer diagnosis. Gene expression can be altered due to smoking and is therefore a potential biomarker for lung cancer.

Objectives

1. Dataset Collection: Identify datasets of RNA-seq with adequate smoking annotation.
2. Algorithm Selection: Find the best performing machine learning algorithm for classification.
3. Model Development: Write a hierarchical multiclassifier.

Methods and Data

Dataset

• TCGA Lung Adenocarcinoma Dataset: Comprised of 522 patients.

Tools and Libraries

• Python Libraries:
  • Machine Learning: scikit-learn, LightGBM, pycaret, optuna
  • Feature Engineering: SHAP, decoupleR
• R Libraries:
  • Pathway Analysis: PROGENY

Data Preparation

• Data Transformation: Multiclass Y transformed to binary Y.
• Gene Filtration with SHAP: Estimated feature importance using SHAP values and removed the least important features iteratively.

Model Training

• Initial Model Assessment:

  • Performance estimation using pycaret on the training dataset (396 samples, 3000 genes).
  • Best model: LightGBM with Train F1 score ~ 0.75.

• Hyperparameter Tuning:

  • Hyper-tuned LightGBM using Optuna.
  • Train F1: 1.00, Test F1: 0.73.

• Pathway Activities:

  • Assessed pathway activities with PROGENY, resulting in 14 new features.
  • Train F1: 0.91, Test F1: 0.74.

Multiclassification

• Custom Hierarchical Classifier:
  • Represented multiclass Y as a tree hierarchy.
  • Implemented a Custom Hierarchical Classifier.
  • Upsampled classes with SMOTE.
  • Trained and hyper-tuned the LightGBM multiclassifier.

Future Plans

1. Identify additional RNA-seq LUAD datasets.
2. Explore deep learning algorithms.
3. Improve the Custom Hierarchical Multiclassifier.

How to Use

1. Clone the Repository:

   git clone https://github.com/michtrofimov/smoker_class.git
   cd smoker_class

2. Install Dependencies: Install with any environment manager from pyproject.toml

3. Run the Notebooks: Open the Jupyter notebooks in the notebooks/ directory to see data preprocessing steps, model training, and evaluation.

Contact

For any questions or issues, please contact Michil Trofimov.