Dirichlet process mixture models for single-cell RNA-seq clustering.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables gene expression analysis at the cellular level.
• Clustering scRNA-seq data is essential for identifying cell types and states.
• Determining the optimal number of clusters remains a significant challenge in scRNA-seq analysis.

Purpose of the Study:

• To compare the performance of Bayesian latent Dirichlet allocation (LDA) and hierarchical Dirichlet process (HDP) for scRNA-seq data clustering.
• To evaluate the impact of the number of clusters on clustering accuracy.
• To provide insights into selecting appropriate clustering methods and resolutions for scRNA-seq data.

Main Methods:

• Comparative analysis of LDA and HDP algorithms on four diverse scRNA-seq datasets.
• Utilized intrinsic (DB-index) and extrinsic (ARI) metrics to assess clustering quality.
• Systematically varied the number of clusters for LDA to compare against HDP's non-parametric approach.

Main Results:

• Clustering performance of LDA and HDP is dataset-specific.
• HDP sometimes yields more appropriate clustering without requiring pre-specified cluster numbers.
• LDA, with optimized cluster numbers, can outperform HDP by better capturing biological features, while HDP may over-cluster.

Conclusions:

• Neither LDA nor HDP universally outperforms the other for scRNA-seq clustering.
• The choice of method and the number of clusters significantly impact biological interpretation.
• Emphasizes the need for careful evaluation of clustering results and number of clusters in scRNA-seq data analysis.