DUBStepR is a scalable correlation-based feature selection method for accurately clustering single-cell data.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Feature selection is crucial for single-cell RNA sequencing (scRNA-seq) data clustering.
• Current methods often yield inconsistent results and ignore gene-gene correlations.

Purpose of the Study:

• To introduce DUBStepR, a novel feature selection algorithm for enhanced single-cell data clustering.
• To leverage gene-gene correlations and a new Density Index (DI) for improved accuracy.

Main Methods:

• DUBStepR (Determining the Underlying Basis using Stepwise Regression) utilizes gene-gene correlations and the Density Index (DI).
• The algorithm was benchmarked against existing feature selection methods across diverse scRNA-seq datasets.

Main Results:

• DUBStepR significantly outperformed existing methods in clustering accuracy, despite selecting fewer genes.
• It successfully deconvolved T and NK cell heterogeneity in rheumatoid arthritis patient PBMCs, identifying disease-associated cell types and subtypes.
• The method demonstrated scalability to over a million cells and applicability to single-cell ATAC-seq data.

Conclusions:

• DUBStepR offers a robust and scalable solution for feature selection in single-cell data analysis.
• It enhances clustering accuracy and facilitates the identification of cellular heterogeneity and disease-specific subtypes.
• The algorithm is proposed as a general-purpose tool for scRNA-seq and other single-cell data types.