scDFN: enhancing single-cell RNA-seq clustering with deep fusion networks.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables high-resolution transcriptome analysis.
• Interpreting heterogeneous scRNA-seq data requires robust cell clustering methods.
• Existing methods struggle with inherent data heterogeneity and limited gene expression.

Purpose of the Study:

• To introduce scDFN, a novel deep learning algorithm for improved scRNA-seq data clustering.
• To enhance the deciphering of transcriptomic diversity and cell behavior patterns.
• To provide an effective tool for nuanced single-cell transcriptomics analysis.

Main Methods:

• scDFN utilizes a fusion network strategy combining an autoencoder and an improved graph autoencoder.
• A cross-network information fusion mechanism integrates attribute and topological information.
• Triple self-supervision and four distinct loss functions optimize the clustering process.

Main Results:

• scDFN significantly outperforms five leading scRNA-seq clustering methods based on NMI and ARI metrics.
• The algorithm demonstrates robust performance on multi-cluster datasets and resilience to batch effects.
• Ablation studies confirm the importance of its core components and loss functions.

Conclusions:

• scDFN establishes a new benchmark for single-cell clustering accuracy and robustness.
• The algorithm effectively addresses challenges in analyzing complex scRNA-seq data.
• scDFN offers a powerful tool for advancing single-cell transcriptomics research.