scNAME

scNAME implements a clustering algorithm for single-cell RNA sequencing (scRNA-seq) data to improve cell-type identification accuracy and robustness by combining mask estimation and neighborhood contrastive learning.

Key Features:

• Mask Estimation for Gene Pertinence Mining: Performs mask estimation to identify pertinent genes and denoise original single-cell expression data to reveal uncorrupted data structure.
• Neighborhood Contrastive Learning Framework: Exploits intrinsic cell structures using neighborhood contrastive learning and generates augmented data randomly to increase sample variety and robustness.
• Global Memory Bank for Enhanced Feature Representation: Employs an offline/global memory bank within the neighborhood contrastive paradigm to store representations and promote intra-cluster compactness and inter-cluster separation.
• Improved Rare Cell Type Detection: Integrates mask estimation, neighborhood contrastive learning, and the global memory bank to enhance detection of rare cell types.

Scientific Applications:

• Biological Analysis: Supports marker gene identification, gene ontology, and pathway enrichment analyses to validate biological significance of clusters.
• Scalability and Robustness: Demonstrates accuracy, robustness, and scalability across simulations and real scRNA-seq datasets.
• Gene Relationship Exploration and Global Similarity Repositories: Facilitates exploration of gene relationships and construction of global cellular similarity repositories for studying cellular heterogeneity.

Methodology:

Mask estimation task, random data augmentation, neighborhood contrastive learning, an offline/global memory bank, and denoising of single-cell expression data.