A Benchmarking Study of Random Projections and Principal Components for Dimensionality Reduction Strategies in Single...

Area of Science:

• Computational biology
• Bioinformatics
• Data science

Background:

• Principal Component Analysis (PCA) is a standard for dimensionality reduction in high-dimensional data, such as single-cell RNA sequencing (scRNA-seq).
• PCA's limitations include performance degradation with large datasets, sensitivity to outliers, and an assumption of linearity.
• Random Projection (RP) methods present emerging alternatives addressing some of PCA's shortcomings.

Purpose of the Study:

• To systematically evaluate and compare the performance of PCA and various RP methods.
• To assess computational efficiency and effectiveness in downstream analyses for scRNA-seq data.
• To provide insights for selecting optimal dimensionality reduction techniques.

Main Methods:

• Benchmarking PCA (including SVD and randomized SVD) against RP algorithms (Sparse and Gaussian RP).
• Utilizing multiple scRNA-seq datasets (labeled and unlabeled) for performance evaluation.
• Assessing downstream clustering quality using Hierarchical Clustering and Spherical K-Means.
• Measuring clustering accuracy (Hungarian algorithm, Mutual Information) and separation (Dunn Index, Gap Statistic).
• Evaluating data variability (Within-Cluster Sum of Squares - WCSS) and locality preservation.

Main Results:

• RP methods demonstrated superior computational speed compared to PCA.
• RP methods rivaled or exceeded PCA in preserving data variability and downstream clustering quality.
• Locality preservation was notably better with RP in several evaluated metrics.
• RP methods showed improved performance across various metrics on both labeled and unlabeled scRNA-seq datasets.

Conclusions:

• Random Projection methods are a computationally efficient and effective alternative to PCA for scRNA-seq dimensionality reduction.
• RP techniques offer advantages in scalability and downstream analysis performance.
• This study provides a comprehensive benchmark to guide the selection of dimensionality reduction methods based on specific analytical needs.