Compact and informative representation learning for scRNA-seq data clustering with masked information bottleneck.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables cellular heterogeneity characterization.
• scRNA-seq data suffers from sparsity, noise, and redundancy, hindering accurate cell clustering.
• Common dimensionality reduction methods using highly variable genes may retain noisy or redundant information.

Purpose of the Study:

• To develop a robust representation learning framework for scRNA-seq data.
• To improve cell clustering accuracy and robustness in the presence of data noise and sparsity.
• To mitigate the impact of noise and redundancy on biological signal extraction.

Main Methods:

• Proposed scMIB, a masked information bottleneck framework.
• Implemented a masking-based denoising strategy to perturb and recover gene expression patterns.
• Integrated an information bottleneck objective for signal compression and relevant information preservation.
• Employed mask consistency learning to capture stable gene-level patterns.

Main Results:

• scMIB demonstrated consistent improvements in clustering accuracy and robustness across multiple scRNA-seq datasets.
• The framework effectively mitigated the influence of noise and sparsity in gene expression data.
• Masking-based perturbation combined with information bottleneck learning proved effective for extracting informative representations.

Conclusions:

• The proposed scMIB framework offers a robust solution for scRNA-seq data analysis and clustering.
• Effective denoising and signal compression strategies enhance the identification of cellular heterogeneity.
• This approach facilitates more reliable biological discoveries from complex single-cell transcriptomic data.