scNAME: neighborhood contrastive clustering with ancillary mask estimation for scRNA-seq data

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables the study of cellular heterogeneity.
• Cell clustering is crucial for scRNA-seq analysis but faces challenges due to data noise, high dimensionality, and scale.
• Existing deep learning methods struggle with rare cell type identification and fully utilizing gene dependencies or cell similarity.

Purpose of the Study:

• To develop a novel scRNA-seq clustering algorithm, scNAME, to address limitations in current methods.
• To improve the accuracy and robustness of cell clustering, especially for rare cell types.
• To effectively leverage gene dependencies and cell similarity for better biological structure detection.

Main Methods:

• scNAME incorporates a mask estimation task for gene pertinence mining.
• A neighborhood contrastive learning framework is employed for cell intrinsic structure exploitation.
• An offline memory bank is utilized for global cellular similarity representation.

Main Results:

• scNAME effectively reveals uncorrupted data structure and denoises scRNA-seq data.
• The algorithm demonstrates improved robustness and data capacity through augmented data in contrastive learning.
• Experimental results confirm scNAME's accuracy, robustness, and scalability in simulations and real data.
• scNAME facilitates rare cell type detection and achieves intra-cluster compactness with inter-cluster separation.

Conclusions:

• scNAME offers a significant advancement in scRNA-seq clustering analysis.
• The method's unique combination of mask estimation and contrastive learning enhances rare cell type identification.
• scNAME provides a powerful tool for accurate and scalable cell type structure discovery in complex single-cell datasets.