AGTformer: Synergistic global transformer and adaptive graph gating for Accurate scRNA-seq clustering

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) offers high-resolution transcriptomic data but faces challenges in cell subpopulation identification due to data complexity.
• Existing deep learning methods often focus on local cell connections, potentially missing crucial long-range dependencies.

Purpose of the Study:

• To introduce AGTformer, an unsupervised clustering framework designed to enhance the accuracy of cell subpopulation identification in scRNA-seq data.
• To address limitations in current deep learning clustering approaches by integrating global context with local topology.

Main Methods:

• AGTformer utilizes an Adaptive Adjacency Gating mechanism for dynamic reweighting of cell-cell graph edges, stabilizing topology-aware learning.
• A Global Transformer module is incorporated to capture long-range cell-cell dependencies, complementing local graph propagation.
• The framework synergistically combines local and global information to learn discriminative latent representations for clustering.

Main Results:

• AGTformer demonstrated superior clustering performance compared to existing methods across ten public scRNA-seq datasets.
• Component effectiveness was validated through visualization, sensitivity analysis, and ablation studies, confirming improved representation quality.
• The framework successfully learned discriminative latent representations crucial for accurate clustering.

Conclusions:

• AGTformer provides a robust framework for unsupervised clustering of scRNA-seq data, effectively characterizing cellular heterogeneity.
• The integration of adaptive graph gating and global transformer modeling represents a significant advancement in scRNA-seq data analysis.