Define and visualize pathological architectures of human tissues from spatially resolved transcriptomics using deep...

Area of Science:

• Computational biology
• Genomics
• Neuroscience

Background:

• Spatially resolved transcriptomics offers insights into disease pathogenesis but faces limitations in associating spatial information for detailed architectural analysis.
• Existing computational frameworks struggle to fully reveal and localize detailed spatial architectures and functional zonation.

Purpose of the Study:

• To introduce RESEPT, a novel deep-learning framework for characterizing and visualizing tissue architecture from spatial transcriptomics data.
• To overcome limitations in current methods for associating spatial information and revealing detailed tissue organization.

Main Methods:

• RESEPT utilizes a spatial retained graph neural network to learn a 3D embedding from spatial transcriptomics data (gene expression or RNA velocity).
• The learned embedding is visualized by mapping it into color channels of an RGB image.
• A supervised convolutional neural network model is employed for segmentation of the visualized tissue architecture.

Main Results:

• RESEPT accurately infers and visualizes tissue architecture on 10x Genomics Visium datasets of human and mouse cortex.
• The framework successfully localized cortex layers and cell types in Alzheimer's disease samples, identifying amyloid-beta plaques.
• RESEPT distinguished distinct regions in a glioblastoma sample, including tumor-enriched, non-tumor, and infiltrating tumor cells.

Conclusions:

• RESEPT provides a powerful tool for analyzing tissue architecture from spatial transcriptomics, enhancing understanding of complex diseases.
• The framework demonstrates potential for clinical and prognostic applications in oncology and neurodegenerative diseases.
• RESEPT's ability to localize specific features like plaques and tumor regions highlights its utility in pathological studies.