Soft graph clustering for single-cell RNA sequencing data.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Clustering is crucial for single-cell RNA sequencing (scRNA-seq) to reveal cellular diversity.
• Graph neural networks (GNNs) improve scRNA-seq clustering but struggle with hard graph constructions that lose information and introduce errors.
• Hard graphs simplify cell relationships, losing continuous similarity data and causing issues in GNNs.

Purpose of the Study:

• To develop a novel soft graph clustering method (scSGC) for scRNA-seq data.
• To overcome limitations of hard graph constructions in existing GNN-based clustering approaches.
• To improve the characterization of continuous intercellular similarities and enhance clustering accuracy.

Main Methods:

• scSGC utilizes a zero-inflated negative binomial (ZINB)-based feature autoencoder to handle scRNA-seq data sparsity and dropouts.
• A dual-channel cut-informed soft graph embedding module captures continuous cell similarities and preserves data structure.
• An optimal transport-based clustering optimization module ensures biologically relevant cell population delineation.

Main Results:

• scSGC effectively characterizes continuous similarities among cells using non-binary edge weights.
• The method mitigates information loss and erroneous message propagation common in hard graph methods.
• Experiments show scSGC outperforms 13 state-of-the-art models in clustering accuracy and efficiency across ten datasets.

Conclusions:

• scSGC integrates advanced techniques to overcome hard graph construction challenges in GNNs for scRNA-seq.
• The method demonstrates superior performance in clustering accuracy, cell type annotation, and computational efficiency.
• scSGC holds significant potential for advancing scRNA-seq data analysis and understanding cellular heterogeneity.