Sphetcher: Spherical Thresholding Improves Sketching of Single-Cell Transcriptomic Heterogeneity

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Massive single-cell RNA sequencing (scRNA-seq) datasets pose computational challenges for routine analyses like cell type detection.
• Geometric sketching offers an alternative to uniform subsampling, selecting representative cells to accelerate analysis and identify rare cell types.

Purpose of the Study:

• To introduce Sphetcher, a novel algorithm for efficient geometric sketching of scRNA-seq data.
• To demonstrate Sphetcher's ability to generate more accurate transcriptomic landscape representations using spherical regions.
• To showcase the utility of Sphetcher's fairness-aware optimization for biological inference, such as trajectory analysis.

Main Methods:

• Sphetcher employs a thresholding technique to select representative cells within spherical regions covering the transcriptomic space.
• The algorithm optimizes cell selection to ensure even coverage and allows for the incorporation of prior biological knowledge through fairness constraints.
• The performance of Sphetcher was evaluated by comparing its spherical sketches to existing methods and by applying it to infer differentiation trajectories.

Main Results:

• Spherical sketches generated by Sphetcher provide a more accurate representation of the original transcriptomic landscape compared to traditional methods.
• Sphetcher's fairness-aware sampling successfully informed the inference of human skeletal muscle myoblast differentiation trajectories.
• The algorithm efficiently selects representative cells, accelerating downstream analyses of large scRNA-seq datasets.

Conclusions:

• Sphetcher offers an effective and accurate method for sketching large scRNA-seq datasets, overcoming computational limitations.
• The algorithm's ability to incorporate fairness enhances its utility for biologically relevant downstream applications, including trajectory inference.
• Sphetcher represents a significant advancement in the computational analysis of single-cell genomics data.