Single-cell RNA-seq data clustering by deep information fusion.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell transcriptomics is crucial for biological discovery but faces computational challenges like high dropout rates, sparsity, and dimensionality.
• Existing deep learning methods struggle to effectively integrate gene attribute and cell topology information for robust cell clustering and data imputation.

Purpose of the Study:

• To develop a novel deep learning framework, scDeepFC, for accurate cell clustering and data imputation from single-cell transcriptomics data.
• To leverage both gene attribute and cell-cell topological information for enhanced downstream analysis.

Main Methods:

• scDeepFC employs a deep auto-encoder (DAE) with integrated zero-inflated negative binomial (ZINB) loss to handle dropout events and model gene expression data.
• A deep graph convolution network (DGCN) captures high-order cell-cell topological information.
• A deep information fusion network integrates representations from DAE and DGCN to create a comprehensive consensus representation.

Main Results:

• scDeepFC demonstrates superior performance in cell clustering and data imputation compared to existing single-cell analysis methods on real datasets.
• The integration of gene attribute and cell topology information significantly improves the accuracy and robustness of cell clustering.

Conclusions:

• scDeepFC provides an effective deep information fusion strategy for single-cell data analysis, addressing key computational challenges.
• The method's ability to utilize diverse data features offers a more comprehensive understanding of cellular heterogeneity and improves data quality.