Specificity-driven cell-gene graph learning identifies rare cell states in single-cell and spatial transcriptomic...

Area of Science:

• Single-cell transcriptomics
• Computational biology
• Genomics

Background:

• Detecting rare cell populations is vital for understanding biological processes and disease.
• Existing single-cell transcriptomic analysis methods struggle to identify rare cells due to reliance on cell-cell similarity, which can obscure subtle transcriptional signatures.
• Multi-sample analyses exacerbate this challenge, as batch correction can dilute rare-cell signals.

Purpose of the Study:

• To develop a sensitive and robust framework for rare-cell discovery from single-cell transcriptomic data.
• To address the limitations of existing methods in resolving rare cell populations, especially in multi-sample datasets.
• To provide a unified approach for identifying biologically meaningful rare cell populations while mitigating batch effects.

Main Methods:

• Introduction of scFormer, a heterogeneous graph transformer (HGT) framework.
• Construction of a Z-score-guided cell-gene heterogeneous graph utilizing highly specific marker genes as informational bridges.
• Integrated optimization strategy for representation learning, clustering, and optional batch correction.

Main Results:

• scFormer demonstrated competitive or superior performance across 125 simulated and 18 real datasets compared to existing approaches.
• Successfully recovered known but weakly represented cell populations in diverse multi-sample single-cell and spatial transcriptomics datasets.
• Revealed previously obscured cell states, including airway club cells, intestinal revival stem cells, and rare embryonic cell states.

Conclusions:

• scFormer offers a novel and effective framework for sensitive rare-cell discovery.
• The Z-score-guided graph construction effectively embeds rare-cell features, overcoming limitations of neighbor-based similarity.
• scFormer provides a unified solution for identifying rare cell populations and mitigating batch effects in complex biological datasets.