Wrapshap a high dimension feature selection framework to explore biological systems through machine learning...

Area of Science:

• Machine Learning
• Bioinformatics
• Computational Biology

Background:

• Wrapper methods in machine learning (ML) offer effective feature selection (FS) but are computationally expensive and prone to overfitting, limiting their use in high-dimensional biological data.
• Existing FS methods struggle with redundancy, multicollinearity, and interaction effects common in biological datasets.

Purpose of the Study:

• To introduce the Wrapshap framework, a novel FS method designed for high-dimensional biological data.
• To address the computational cost and overfitting issues of traditional wrapper methods.
• To enhance the interpretability and usability of ML in biomedical research.

Main Methods:

• Wrapshap leverages the local accuracy property of SHapley Additive exPlanations (SHAP) to translate complex biological data into interpretable insights.
• The framework integrates a trained ML model, enabling explanations to be visualized as an exploratory data space.
• It requires only a single training of the ML model, mitigating computational inefficiencies.

Main Results:

• Benchmarks on 84 regression and 106 classification datasets show Wrapshap outperforms state-of-the-art methods like Recursive Feature Elimination in computational speed, feature ranking quality, and predictive performance.
• The method demonstrates adaptability across various ML models, explanation methods, and tasks.
• Wrapshap enables performance monitoring and opens possibilities for real-time, interactive FS.

Conclusions:

• Wrapshap offers a computationally efficient and effective solution for feature selection in high-dimensional biological data.
• The framework enhances the interpretability of ML models, facilitating informed biological hypothesis generation.
• It represents a significant advancement in applying ML to biomedical research.