scCompressSA: dual-channel self-attention based deep autoencoder model for single-cell clustering by compressing...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for understanding cell differentiation and diseases.
• Accurate cell type detection from stochastic scRNA-seq data remains challenging, particularly for human samples.
• Deep neural networks show promise in identifying cell-specific patterns, outperforming traditional statistical methods.

Purpose of the Study:

• To develop a novel method for improving cell type detection in scRNA-seq data.
• To leverage gene-gene interactions and topological patterns for enhanced clustering accuracy.
• To address the limitations of existing methods in capturing complex biological features.

Main Methods:

• Proposes the scCompressSA method, integrating topological patterns from scRNA-seq data.
• Utilizes a self-attention (SA) based coefficient compression (CC) block to capture gene-gene interactions.
• Employs cross-correlation to extract static gene-gene interactions as temporal features.

Main Results:

• scCompressSA enhances clustering accuracy across multiple benchmark scRNA-seq datasets.
• The method effectively integrates both topological and temporal features for improved analysis.
• Demonstrates superior performance compared to approaches relying solely on temporal patterns.

Conclusions:

• Static gene-gene interactions, when extracted as temporal features, significantly boost single-cell clustering performance.
• The dual-channel SA-based CC block in scCompressSA effectively integrates topological features.
• scCompressSA exhibits superior detection accuracy in single-cell clustering compared to prior methods.