MscNMF

MscNMF applies non-negative matrix factorization to decompose scRNA-seq data and learn multi-subspace cell similarities for identifying cellular subpopulations and marker genes.

Key Features:

• Non-Negative Matrix Factorization Framework: Employs non-negative matrix factorization (NMF) to decompose high-dimensional scRNA-seq datasets into interpretable components.
• Multi-Subspace Cell Similarity Learning: Learns cell similarity across multiple low-dimensional subspaces to enhance detection of gene and cell features.
• Data Decomposition, Similarity Learning, and Fusion: Integrates data decomposition, similarity learning, and similarity fusion to combine insights from different subspaces.
• Reduction of Redundancy and Noise: Reduces redundancy and noise within each low-dimensional feature space to improve signal quality.
• Gene Weight Information Analysis: Analyzes gene weight information derived from NMF to estimate the optimal number of subpopulations.
• Clustering Performance and Genetic Marker Extraction: Demonstrated strong clustering performance and effective extraction of genetic markers across eight real scRNA-seq datasets.

Scientific Applications:

• Cellular heterogeneity analysis: Identifies and characterizes distinct cell subpopulations within scRNA-seq datasets.
• Marker gene identification: Extracts genetic markers associated with identified cell subpopulations.
• Developmental biology studies: Supports analysis of cell population dynamics relevant to development.
• Cancer research: Facilitates detection of tumor cell subpopulations and associated marker genes.
• Immunology: Enables characterization of immune cell subsets and their marker genes.
• Disease mechanism and therapeutic target discovery: Provides insights into disease mechanisms and potential therapeutic targets through subpopulation and marker analysis.

Methodology:

Uses non-negative matrix factorization to decompose scRNA-seq matrices, performs multi-subspace cell similarity learning and similarity fusion, reduces redundancy and noise within low-dimensional feature spaces, analyzes gene weight information to determine optimal subpopulation number, and performs clustering and marker extraction; evaluated on eight real scRNA-seq datasets.