scDVAE:Single-Cell Data Clustering Based on Variational Autoencoder With Disentangled Latent Representations.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) reveals cellular heterogeneity but faces challenges like high dimensionality and dropout events.
• Cell clustering is vital for analyzing scRNA-seq data and identifying distinct cell populations.
• Existing clustering methods struggle with the inherent complexities of scRNA-seq data.

Purpose of the Study:

• To develop a novel deep generative model, scDVAE, for improved single-cell RNA sequencing data clustering.
• To address challenges in scRNA-seq data analysis, including high dimensionality, sparsity, and dropout events.
• To enhance the identification of cellular heterogeneity through robust clustering.

Main Methods:

• scDVAE utilizes a variational autoencoder with disentangled latent representations.
• Latent representations are separated into clustering and generative features for task-specific optimization.
• A Student's t-mixture model is employed as the prior distribution for clustering features to improve robustness.
• A hybrid data augmentation strategy is implemented to increase dataset diversity and reduce noise.

Main Results:

• scDVAE demonstrated significantly improved clustering performance across 10 real-world datasets.
• The disentangled latent space effectively separated clustering and generative information.
• The method showed enhanced robustness against dropout events compared to existing approaches.
• Experimental results confirmed the superiority of scDVAE over state-of-the-art clustering methods.

Conclusions:

• scDVAE offers a powerful new approach for clustering single-cell RNA sequencing data.
• The model effectively handles the complexities and noise inherent in scRNA-seq datasets.
• This method advances the analysis of cellular heterogeneity in complex biological systems and diseases.