Hubness reduction improves clustering and trajectory inference in single-cell transcriptomic data.

Area of Science:

• Computational Biology
• Bioinformatics
• Data Science

Background:

• Single-cell RNA sequencing (scRNAseq) data presents high dimensionality, leading to the "dimensionality curse" in analyses.
• Hubness, the presence of data points with excessive neighbors, is a manifestation of this curse in scRNAseq neighborhood graphs.
• Current methods often involve aggressive dimensionality reduction, potentially losing valuable information.

Purpose of the Study:

• To investigate the phenomenon of hubness in scRNAseq datasets.
• To evaluate methods for reducing hubness as an alternative to dimensionality reduction.
• To assess the impact of hubness reduction on common scRNAseq analysis tasks.

Main Methods:

• Investigated hubness in scRNAseq datasets.
• Applied and evaluated various hubness reduction techniques.
• Assessed the performance of hubness reduction on clustering, trajectory inference, and visualization tasks.
• Compared hubness reduction against state-of-the-art neighborhood graph improvement methods.

Main Results:

• Hub cells in scRNAseq data do not appear to be linked to technical or biological biases.
• Hubness reduction enhances neighborhood graph properties for machine learning applications.
• Hubness reduction outperforms existing methods for improving neighborhood graphs.
• Improved performance in clustering, trajectory inference, and visualization, particularly for datasets with high intrinsic dimensionality.

Conclusions:

• Hubness is a significant factor in scRNAseq data analysis, impacting neighborhood graph construction.
• Reducing hubness can serve as a viable alternative to drastic dimensionality reduction.
• Hubness reduction enhances the utility of scRNAseq data for machine learning-based analyses, especially in high-dimensional scenarios.