DiffuScope: A diffusion-regularized autoencoder for spatial transcriptomic clustering.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Spatial transcriptomics technologies generate large, diverse datasets with significant heterogeneity.
• Existing clustering algorithms struggle to adapt to this diversity, necessitating generalizable solutions.

Purpose of the Study:

• To develop a novel, generalizable clustering framework for spatial transcriptomics data.
• To improve the accuracy and cross-dataset generalization of spatial data analysis.

Main Methods:

• Proposed DiffuScope, a framework utilizing Graph Convolutional Variational Autoencoders (GC-VAE).
• Employed self-supervised learning for latent representation extraction.
• Incorporated Reconstruction Loss and Diffusion Consistency Loss for enhanced feature learning.

Main Results:

• Systematically benchmarked DiffuScope against state-of-the-art methods on 12 diverse datasets.
• Demonstrated robust performance, even with introduced noise in spatial expression data.
• Successfully applied DiffuScope to breast and gastric cancer datasets for domain delineation.

Conclusions:

• DiffuScope offers a robust and generalizable clustering solution for heterogeneous spatial transcriptomics data.
• The framework effectively identifies spatial domains and reveals biologically relevant tissue structures.
• DiffuScope advances the analysis of complex biological spatial patterns.