A model-based constrained deep learning clustering approach for spatially resolved single-cell data.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Spatially resolved single-cell RNA sequencing (sp-scRNA-seq) offers insights into gene expression within tissue context.
• Computational methods for sp-scRNA-seq data analysis are crucial but underdeveloped.
• Existing methods struggle to fully leverage spatial information for accurate cell clustering.

Purpose of the Study:

• To develop an advanced deep learning approach for clustering sp-scRNA-seq data.
• To integrate spatial coordinates and marker gene expression patterns into a unified clustering framework.
• To enhance the accuracy and robustness of cell type identification in spatial transcriptomics.

Main Methods:

• Developed Deep Spatially constrained Single-cell Clustering (DSSC), a deep learning model.
• Utilized a graphical neural network to encode spatial cell information.
• Incorporated cell-to-cell constraints based on marker gene spatial expression.
• Applied deep embedding clustering using autoencoders and Kullback-Leibler divergence.

Main Results:

• DSSC significantly improves clustering performance compared to state-of-the-art methods.
• Demonstrated robust performance across diverse simulated and real sp-scRNA-seq datasets.
• Successfully integrated spatial coordinates and marker gene data for enhanced clustering.
• Showcased effectiveness across various tissue organizations and spatial dependencies.

Conclusions:

• DSSC is a pioneering model leveraging both spatial coordinates and marker genes for cell/spot clustering.
• The method offers a promising advancement for analyzing sp-scRNA-seq data.
• DSSC enhances the potential of spatial transcriptomics for biological discovery.