EMD-HVG: a normalization-independent method for highly variable gene selection based on Earth mover's distance.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Identifying highly variable genes (HVGs) is crucial for single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (ST).
• Existing methods often rely on normalization, which can obscure biological variability and struggle with sparse, noisy scRNA-seq/ST data.
• Current distributional assumptions in HVG identification may not accurately reflect the nature of this data.

Purpose of the Study:

• To develop a normalization-independent and nonparametric method for robust HVG identification.
• To overcome the limitations of conventional normalization-based approaches in scRNA-seq and ST data analysis.

Main Methods:

• Proposed the Earth Mover's Distance based Highly Variable Gene identification method (EMD-HVG).
• EMD-HVG utilizes a mixture distribution model for gene expression patterns, preserving biological heterogeneity.
• Employs Earth Mover's Distance (EMD), a nonparametric metric, to assess gene expression variability without distributional assumptions.

Main Results:

• EMD-HVG demonstrated superior performance compared to existing HVG detection methods across numerous evaluation scenarios.
• Achieved top performance in 26 out of 30 tested scenarios based on various accuracy metrics.
• The identified HVGs significantly improved the quality of downstream clustering and cell-type delineation.

Conclusions:

• EMD-HVG is a robust and broadly applicable method for HVG identification in both scRNA-seq and ST.
• The method enhances downstream analysis precision, facilitating more accurate cell-type identification.
• EMD-HVG offers a significant advancement for transcriptomic data analysis by avoiding normalization pitfalls.