scGCRC: Graph and Contrastive-Based Representation Learning for Single-Cell RNA-Seq Data Clustering.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) enables cellular-level biological analysis.
• Cell clustering is crucial for identifying cell types in scRNA-seq data.
• Existing deep learning methods often use a difficult two-stage learning process.

Purpose of the Study:

• To develop a novel, more efficient method for scRNA-seq clustering.
• To improve cell representation learning for better clustering accuracy.
• To overcome the limitations of traditional two-stage deep learning approaches.

Main Methods:

• A local self-attention network aggregates cell information via a cell relationship graph.
• Dual contrastive learning optimizes cell representations at both cell and cluster levels.
• The Leiden algorithm performs clustering on the learned representations.

Main Results:

• The proposed method directly learns low-dimensional cell representations without autoencoder pretraining.
• It preserves local structural relationships between cells effectively.
• Demonstrated superior performance across diverse datasets compared to baseline methods.

Conclusions:

• The novel approach simplifies scRNA-seq clustering by integrating representation learning and contrastive learning.
• This method offers a more efficient and effective solution for cell type identification.
• The findings highlight the potential of local self-attention and contrastive learning in single-cell data analysis.