featsel

Feature selection pipeline for high-dimensional data with a focus on genomics and bioinformatics.

The Problem

Modern datasets often contain thousands of features but relatively few samples. This is especially common in:

• Genomics: Gene expression profiles with 20,000+ genes per patient
• Text analysis: Document classification with large vocabularies
• Sensor data: IoT and industrial monitoring systems

Training models on such data leads to overfitting, long computation times, and poor interpretability. Feature selection addresses this by identifying the most predictive variables while discarding noise.

What This Project Does

This project implements a feature selection pipeline that:

• Applies multiple feature selection methods: filter (variance, ANOVA F with a fixed count, a percentile, or a false discovery or family-wise error rate, mutual information, correlation), embedded (Lasso, random forest importance) and wrapper (recursive feature elimination, with or without cross-validation, and sequential selection), plus a random-selection control. Higher Criticism is planned.
• Compares them under one protocol: stratified Monte Carlo splits, with every selector fit on the training part only
• Runs the splits in parallel
• Records accuracy metrics, selection stability, selection time and memory

Every selector is a scikit-learn estimator, so it also works on its own in a scikit-learn Pipeline, cross_val_score or GridSearchCV.

The primary use case is predicting breast cancer molecular subtypes from gene expression data, but the pipeline generalizes to other high-dimensional classification problems.

Project Structure

├── featsel/            # The Python package
│   └── selectors/      # One subpackage per selector, with its tests and a tutorial notebook
├── configs/            # Dataset and experiment configuration files (YAML)
├── datasets/           # Input data (one subfolder per dataset, not in git)
├── results/            # Experiment output CSVs (not in git)
├── figures/            # Generated plots
├── notebooks/          # Analysis and usage notebooks
├── scripts/            # One-off runs and examples
├── tests/              # Package-level tests
├── docs/               # Design notes; report chapters in docs/report/
└── references/         # Project proposal and papers

Installation

# Clone the repository
git clone https://github.com/drormeir/featsel.git
cd featsel

# Create virtual environment
python -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Future PyPI installation (not yet available):

pip install featsel

Usage

Run a comparison from an experiment config. It names a dataset config, the selectors, the models and the splits, and writes one CSV row per execution:

python -m featsel.run --config configs/experiment_classifiers.yaml --n-jobs -1
python -m featsel.run --config configs/experiment_classifiers.yaml --resume   # continue a killed run

The same from Python, on data already in memory:

from featsel.run import run

config = {
    'selectors': [{'name': 'RandomSelector', 'k': [10, 50]},
                  {'name': 'ANOVAFSelector', 'k': [10, 50]}],
    'models': [{'name': 'logistic_regression', 'params': {'max_iter': 1000}}],
    'n_splits': 20,
    'output': 'results/my_run.csv',
}
run(X, y, config)

A single selector, as a scikit-learn step:

from featsel.selectors import ANOVAFSelector

X_selected = ANOVAFSelector(n_features=50).fit_transform(X, y)

See notebooks/03_usage_demo.ipynb for an end-to-end example.

To use your own dataset, add a data folder in datasets/, a dataset config (see configs/template.yaml) and an experiment config that names it (see configs/experiment_classifiers.yaml). Paths in a config are relative to the config file.

Datasets

The pipeline is dataset-agnostic. Each dataset needs:

• A subfolder in datasets/ with features.csv and metadata.csv
• A dataset config file in configs/

SCAN-B Breast Cancer (included config)

• 3,069 patients, 9,259 gene expression features after cleaning
• PAM50 molecular subtype labels (Basal, LumA, LumB, HER2, Normal)
• Clinical metadata (ER status, survival data)

Note: Data files are not included due to size. Download from [TBD] and place in datasets/scanb_small/.

Status

This project is an M.Sc. final project at Reichman University, supervised by Dr. Ben Galili.