scDGCL: A Dual-Level and Graph-Constrained Contrastive Learning Method for Single-Cell RNA Sequencing Data...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Single-cell RNA sequencing (scRNA-seq) is vital for life science, but its high dimensionality and sparsity challenge data analysis.
• Clustering is a fundamental step in scRNA-seq analysis, yet existing methods struggle with suboptimal data representations, limiting performance.

Purpose of the Study:

• To develop an advanced clustering method for scRNA-seq data that overcomes limitations of existing approaches.
• To improve the accuracy and biological relevance of cell clustering in scRNA-seq data analysis.

Main Methods:

• Propose scDGCL, a novel dual-level and graph-constrained contrastive learning framework.
• Implement Dual-level Contrastive Learning (DCL) to optimize cell representations at cell and cluster levels.
• Integrate Graph-constrained Contrastive Learning (GCL) to align representations with graph priors, enhancing biological insights.

Main Results:

• scDGCL demonstrates superior performance in scRNA-seq data clustering across 12 real and 8 simulated datasets.
• Comparative analysis against 17 methods confirms scDGCL's effectiveness.
• Ablation and hyperparameter studies validate the robustness and component efficacy of scDGCL.

Conclusions:

• scDGCL significantly advances scRNA-seq data clustering by improving cell representation.
• The method's biological plausibility is confirmed through marker gene expression and cell trajectory inference.
• scDGCL offers a robust and effective tool for analyzing complex single-cell transcriptomic data.