Biomarker identification by interpretable maximum mean discrepancy

Area of Science:

• Biomedical data analysis
• Bioinformatics
• Machine learning

Background:

• Biomedical research often involves comparing paired sample groups (e.g., treated vs. control) using high-dimensional omics data.
• Identifying discriminating features, or biomarkers, is crucial for understanding biological differences.
• Current methods often separate two-sample testing and feature selection, limiting integrated analysis.

Purpose of the Study:

• To develop a unified statistical framework for simultaneous two-sample testing and feature selection.
• To introduce a sparse, interpretable, and optimized Maximum Mean Discrepancy (MMD) test (SpInOpt-MMD).
• To demonstrate the versatility and effectiveness of SpInOpt-MMD across diverse data types.

Main Methods:

• Developed the SpInOpt-MMD algorithm, integrating multivariate two-sample testing with sparse feature selection.
• Applied SpInOpt-MMD to synthetic and real-world datasets, including images, gene expression, and text data.
• Compared SpInOpt-MMD performance against established feature selection techniques like SHapley Additive exPlanations and univariate association analysis.

Main Results:

• SpInOpt-MMD successfully performs simultaneous two-sample testing and feature selection.
• The method demonstrates effectiveness in identifying relevant features, even with small sample sizes.
• SpInOpt-MMD outperformed alternative feature selection methods in several experimental comparisons.

Conclusions:

• SpInOpt-MMD offers a powerful and versatile approach for biomarker discovery in high-dimensional biomedical data.
• The integrated method provides both statistical significance testing and interpretable feature identification.
• The developed method enhances the analysis of complex biological datasets, particularly in scenarios with limited samples.