Statistically principled feature selection for single cell transcriptomics.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell transcriptomics (scRNAseq) generates high-dimensional data, necessitating feature selection for downstream analyses like cell clustering.
• Evaluating feature selection methods is challenging as their performance varies significantly depending on the specific analytical task.
• Subtle cell type differences require careful consideration of both the number and selection strategy of features.

Purpose of the Study:

• To develop and present a novel, model-grounded feature selection method for scRNAseq data.
• To enable interpretable selection of an optimal number and subset of genes without arbitrary parameters.
• To facilitate the identification of biologically meaningful rare cell types.

Main Methods:

• Development of a feature selection method based on an analytical model.
• Comparison of the proposed method against default feature selection in Scanpy and Seurat, and SCTransform.
• Evaluation of method performance across different tasks, including routine and subtle cell type identification.

Main Results:

• The performance of feature selection methods is highly task-dependent.
• Randomly selected features can suffice for basic cell type identification.
• Subtle cell type distinctions necessitate strategic feature selection, where both quantity and method are critical.
• The proposed analytical method offers interpretable guidance on feature selection.
• The new method achieves higher accuracy with fewer features compared to existing approaches.

Conclusions:

• A new, interpretable feature selection method for scRNAseq data has been developed.
• This method enhances the accuracy of identifying subtle cell type differences and rare cell populations.
• The approach provides a more robust alternative to default methods in popular scRNAseq analysis packages.