Random Projection Methods Outperform Principal Component Analysis for Dimensionality Reduction in Single Cell RNA-Seq

Area of Science:

• Computational Biology
• Bioinformatics
• Data Science

Background:

• Principal Component Analysis (PCA) is widely used for dimensionality reduction in high-dimensional datasets like single-cell RNA sequencing (scRNA-seq).
• PCA's performance degrades with increasing dataset size, and it is sensitive to outliers and assumes linearity.
• Random Projection (RP) methods present a promising alternative to overcome PCA's limitations.

Purpose of the Study:

• To systematically evaluate and compare the performance of PCA and various RP methods on scRNA-seq datasets.
• To introduce and assess a novel Matching Sparsity Random Projection algorithm for improved computational scalability and effectiveness.
• To provide guidance on selecting optimal dimensionality reduction strategies for scRNA-seq data analysis.

Main Methods:

• Evaluated PCA, Singular Value Decomposition (SVD), and multiple RP methods (sparse, Gaussian, adaptive sparsity) on public scRNA-seq datasets.
• Assessed clustering performance using Hierarchical Clustering and Spherical K-Means.
• Quantified performance using metrics like Hungarian algorithm accuracy, Mutual Information, Dunn Index, Gap Statistic, and Within-Cluster Sum of Squares.

Main Results:

• RP methods demonstrated substantial computational speed improvements over PCA.
• RP methods, including the adaptive sparsity approach, outperformed PCA in preserving locality and several other evaluated metrics.
• RP methods rivaled, and in some cases exceeded, PCA in data variability preservation and clustering quality.

Conclusions:

• Random projection methods are a computationally efficient and effective alternative to PCA for scRNA-seq data.
• The adaptive sparsity RP algorithm shows promise for handling data sparsity patterns effectively.
• This study offers critical guidance for choosing dimensionality reduction techniques that balance computational efficiency, scalability, and analytical performance.