Supervised dimensionality reduction for exploration of single-cell data by HSS-LDA.

Area of Science:

• Single-cell biology
• Bioinformatics
• Computational biology

Background:

• Single-cell technologies produce large, high-dimensional omics datasets.
• Dimensionality reduction techniques are crucial for visualizing and understanding data structure and heterogeneity.
• Current unsupervised methods often ignore valuable biological labels like cell type or time.

Purpose of the Study:

• To repurpose the linear discriminant analysis (LDA) classification algorithm for supervised dimensionality reduction of single-cell data.
• To develop a computationally efficient feature selection method (Hybrid Subset Selection - HSS) for LDA.
• To demonstrate the utility of HSS-LDA for exploring specific aspects of cellular heterogeneity.

Main Methods:

• Implemented Hybrid Subset Selection (HSS) for feature selection.
• Applied linear discriminant analysis (LDA) for supervised dimensionality reduction.
• Benchmarked HSS-LDA against popular unsupervised dimensionality reduction algorithms.

Main Results:

• HSS-LDA generates non-stochastic, interpretable axes for dimensionality reduction.
• The method effectively visualizes biological processes like cell differentiation and cell cycle.
• Demonstrated utility across single-cell mass cytometry, transcriptomics, and chromatin accessibility data.

Conclusions:

• HSS-LDA offers a powerful supervised approach for analyzing single-cell data.
• This method enhances the interpretability of cellular heterogeneity and biological processes.
• HSS-LDA is a versatile tool applicable to diverse single-cell omics datasets.