Assembling spatial clustering framework for heterogeneous spatial transcriptomics data with GRAPHDeep.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics enables high-resolution analysis of tissue microenvironments.
• Spatial clustering is critical for understanding cellular organization and function.
• Graph neural networks (GNNs) show promise for integrating spatial and gene expression data.

Purpose of the Study:

• To develop a framework, GRAPHDeep, for selecting optimal GNNs for spatial clustering in spatial transcriptomics.
• To investigate the impact of different graph deep learning modules and GNN architectures.
• To provide guidance for choosing appropriate GNNs for spatial omics data analysis.

Main Methods:

• Developed GRAPHDeep, a framework integrating 2 graph deep learning modules and 20 GNNs.
• Evaluated GNN performance on heterogeneous spatial transcriptomics datasets.
• Compared GRAPHDeep's spatial clustering with state-of-the-art methods.

Main Results:

• The number of genes/proteins significantly impacts spatial clustering performance.
• Variational graph autoencoder outperformed deep graph infomax for this task.
• Recommended GNNs include UniMP, SAGE, SuperGAT, GATv2, GCN, and TAG.
• Existing spatial clustering frameworks may not use the optimal GNN.

Conclusions:

• GRAPHDeep provides an effective approach for spatial clustering in spatial transcriptomics.
• The study offers crucial insights into selecting appropriate GNNs for spatial omics data.
• This work guides researchers in optimizing spatial clustering analysis for biological discovery.