Structure-Aware Principal Component Analysis for Single-Cell RNA-seq Data.

Area of Science:

• Single-cell transcriptomics
• Computational biology
• Bioinformatics

Background:

• Droplet-based technologies enable high-throughput single-cell RNA sequencing, generating large datasets for cellular heterogeneity studies.
• Traditional Principal Component Analysis (PCA) for dimensionality reduction in single-cell data is often biased by dominant cell populations, hindering the detection of rare cell types.
• Identifying minority cell types is crucial for comprehensive biological knowledge discovery.

Purpose of the Study:

• To develop a novel dimensionality reduction method that accurately identifies minority cell populations in large single-cell transcriptomic datasets.
• To overcome the limitations of traditional PCA in detecting rare cell types due to biased principal directions.

Main Methods:

• Proposed Locality-Sensitive PCA (LSPCA), a scalable PCA variant incorporating structure-aware data sampling.
• Structure-aware sampling ensures a neutral data spread, mitigating bias from redundant samples.
• Benchmarked LSPCA against traditional PCA and PCA with random sampling using ten public single-cell expression datasets.

Main Results:

• LSPCA demonstrated superior performance in detecting minority cell populations compared to traditional PCA and PCA with random sampling.
• Clustering analyses on annotated datasets confirmed LSPCA's higher accuracy in identifying rare cell types.
• The proposed method effectively reduces bias in principal directions, facilitating more balanced data representation.

Conclusions:

• LSPCA is an effective and scalable method for dimensionality reduction in single-cell transcriptomics, particularly for identifying rare cell populations.
• The structure-aware sampling approach in LSPCA enhances the discovery of cellular heterogeneity by mitigating biases inherent in traditional PCA.