MEGAFS: Multi-stage Explainable Gene-Aware Feature Selection

MEGAFS a multi-stage explainable feature selection framework that combines multivariate filtering with evolutionary optimization.

The library employs a robust two-stage architecture:

1. Search Space Pruning: It utilizes the mRMR (Minimum Redundancy Maximum Relevance) algorithm to drastically reduce the initial feature space, filtering out redundant variables.
2. Fine-tuning Optimization: It initializes a Genetic Algorithm (GA) on the reduced set to select an optimal, stable, and minimal feature subset.

This approach is validated through a case study on HIV-1 seroconversion susceptibility, benchmarking against a wide spectrum of state-of-the-art extraction and selection techniques.

🚀 Key Features

• Hybrid Efficiency: Combines the speed of filter methods (mRMR) with the precision of wrapper methods (Genetic Algorithms).
• Scikit-learn Compatible: Works natively with any sklearn classifier (RandomForest, SVM, XGBoost, etc.).
• Automated Workflow: Handles data loading, validation, scaling, feature selection, and cross-validation in a single function call.
• Stability: Performs multiple independent runs to ensure the selected features are robust and not artifacts of randomness.
• Excel Integration: Directly accepts .xlsx datasets and generates Excel reports.

🛠 Installation

To use MEGAFS, you need to install the following dependencies:

pip install pandas numpy scikit-learn mrmr-selection sklearn-genetic xgboost openpyxl

(Note: Once the package is published to PyPI, you will be able to install it simply via pip install MEGAFS)

📊 Data Format

The input dataset must be an Excel file (.xlsx) structured as follows:

1. Rows: Represent individual samples (patients, observations).
2. Columns: Represent features (genes, biomarkers, variables).
3. Target: A specific column named class containing the target labels (0/1, integers, etc.).

📖 Usage / Quick Start

The main entry point is the megafs function. Here is a complete example of how to run an experiment:

from megafs import megafs
from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier

# 1. Define the path to your dataset
dataset_path = 'data/my_biomarkers.xlsx'

# 2. (Optional) Define a custom dictionary of classifiers
# If you pass None, MEGAFS uses a default set of robust classifiers (SVC, GBC, etc.)
my_classifiers = {
    'RFC': RandomForestClassifier(n_estimators=100, random_state=42),
    'GBC': GradientBoostingClassifier(random_state=42)
}

# 3. Run the MEGAFS pipeline
megafs(
    dataset=dataset_path,
    model_list=my_classifiers,
    metric='roc_auc',       # Metric to optimize
    n_splits=5,             # 5-fold Cross-Validation
    n_runs=3,               # 3 Independent runs per model for stability
    output_file='Experiment_Results'
)

⚙️ API Parameters

The megafs function accepts the following arguments:

📂 Output Files

The execution generates two files in your working directory based on the output_file name:

1. Log File (.txt)

Contains detailed execution logs:

• Global parameters used.
• Per-run details: Max features allowed, number of selected features, Train/Test scores.
• Selected Features: A list of the specific feature names selected in each run.
• Genetic Evolution: A list showing the fitness score evolution across generations.

2. Summary File (.xlsx)

A spreadsheet summarizing the results:

• Mean Metric: Average performance (e.g., ROC-AUC) across all runs.
• Mean Features: Average number of features selected.
• Aggregated rows for every model tested.

⚖️ License

This project is licensed under the MIT License. See the LICENSE file for more details.