Heterogeneity-Preserving Discriminative Feature Selection for Disease-Specific Subtype Discovery

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Disease heterogeneity necessitates identifying specific subtypes for personalized medicine and targeted therapies.
• High-throughput molecular data (e.g., single-cell RNA-seq) aids subtype discovery but presents high-dimensionality challenges.
• Existing feature selection methods often fail to preserve data heterogeneity while identifying subtypes.

Purpose of the Study:

• To develop a novel feature selection method that preserves biological heterogeneity and effectively identifies disease or cell subtypes.
• To address the limitations of current methods in discovering subtype signatures from high-dimensional molecular datasets.

Main Methods:

• Utilized deep metric learning for feature embedding to explore heterogeneity-preserving statistical properties.
• Developed PHet (Preserving Heterogeneity), a statistical method using iterative subsampling, differential interquartile range (IQR) analysis, and Fisher's method.
• Applied PHet to single-cell RNA-seq and microarray datasets for validation.

Main Results:

• Identified features with significant interquartile range (IQR) differences between classes as crucial for subtype information.
• PHet successfully preserved sample heterogeneity and distinguished known disease/cell states, outperforming existing methods.
• Discovered two basal cell subtypes in mouse tracheal epithelial cells differentiating towards a luminal secretory phenotype.

Conclusions:

• PHet is an effective method for identifying disease and cell subtypes by preserving heterogeneity.
• The method enhances subtype clustering quality and aids in understanding complex biological systems.
• PHet contributes to precision medicine by improving disease mechanism insights and cell differentiation studies.