A Comparison for Dimensionality Reduction Methods of Single-Cell RNA-seq Data.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) provides insights into cellular heterogeneity.
• scRNA-seq data is characterized by high dimensionality, noise, and sparsity.
• Dimension reduction is essential for analyzing scRNA-seq data.

Purpose of the Study:

• To evaluate the stability, accuracy, and computational cost of 10 dimensionality reduction methods.
• To investigate the sensitivity of these methods to hyperparameter tuning.
• To provide recommendations for dimensionality reduction in scRNA-seq analysis.

Main Methods:

• Developed an evaluation strategy using 30 simulated and 5 real scRNA-seq datasets.
• Assessed 10 different dimensionality reduction techniques.
• Analyzed method performance concerning stability, accuracy, and computational expense.

Main Results:

• t-distributed Stochastic Neighbor Embedding (t-SNE) demonstrated the highest accuracy but also the highest computational cost.
• Uniform Manifold Approximation and Projection (UMAP) exhibited superior stability and preserved cell population structures effectively.
• UMAP offered a balance of moderate accuracy and the second-highest computational cost.

Conclusions:

• UMAP is recommended for its stability and ability to maintain cell population integrity in scRNA-seq data.
• Hyperparameter tuning is critical for non-linear and neural network-based dimensionality reduction methods.
• Method selection should consider the trade-offs between accuracy, stability, and computational resources.