Optimal Feature Non-negative Matrix Factorization (OFNMF) for Enhanced Single-Cell RNA Sequencing Analysis

Introduction: Single-cell RNA sequencing (scRNA-seq) technology enables the parallel processing of millions of genetic data points, providing deeper insights into species evolution, biological metabolism, and disease mechanisms. However, the parallel processing capability of scRNA-seq introduces challenges in data analysis, including the presence of redundant features and technical noise. To address these issues, we propose a model called Optimal Feature Non-negative Matrix Factorization (OFNMF), which aims to effectively process and analyze scRNA-seq data. Methods: The OFNMF model consists of three main components: Non-negative Matrix Factorization (NMF), similarity learning, and spectral clustering. NMF reduces high-dimensional features of the scRNA-seq dataset by filtering out noise, outliers, and irrelevant data while mapping the data to a lower-dimensional subspace. OFNMF utilizes cell information differences to assess the adequacy of learned features and determine the optimal number of features. To capture the heterogeneity between cells, OFNMF introduces a cell similarity learning framework. Finally, spectral clustering is employed to refine cell clustering for more accurate results. Results: The OFNMF model demonstrates improved performance in data processing by effectively reducing noise, revealing more meaningful features, and efficiently identifying relevant features and clustering cells based on their similarities. Discussion: The model's robustness is validated through cell visualizations and its application to gene marker extraction, confirming its superiority over traditional methods. This demonstrates that OFNMF offers a promising approach for improving single-cell RNA sequencing data analysis. Conclusion: The method holds significant potential for enhancing downstream applications, such as gene marker extraction, and furthering our understanding of cellular heterogeneity.