Inferring single-cell spatial gene expression with tissue morphology via explainable deep learning

Area of Science:

• Computational Biology
• Genomics
• Biotechnology

Background:

• Cellular spatial arrangement is crucial for development and organogenesis.
• Deep learning with spatial omics data reveals complex biological patterns and disease insights.
• Histological images and computational methods analyze cellular heterogeneity and spatial data.

Purpose of the Study:

• To develop a vision transformer (ViT) framework, SPiRiT, for mapping histological signatures to spatial single-cell transcriptomic signatures.
• To enhance the framework with cross-validation and model interpretation for hyper-parameter tuning.
• To predict single-cell spatial gene expression from histopathological images in human and mouse models.

Main Methods:

• Developed SPiRiT, a vision transformer (ViT) framework integrating cross-validation and model interpretation.
• Applied SPiRiT to predict spatial gene expression from H&E stained histological images.
• Evaluated SPiRiT using Xenium and Visium (10x Genomics) datasets for human breast cancer and whole mouse pup.

Main Results:

• SPiRiT accurately predicts single-cell spatial gene expression from tissue morphology.
• Model interpretation identified high-resolution, high attention areas (HAR) linked to specific cell types and marker genes (FASN, POSTN, IL7R).
• SPiRiT demonstrated a 40% improvement in predictive accuracy over ST-Net, with high consistency between predicted gene expression and tumor region annotation.

Conclusions:

• SPiRiT enables inference of spatial single-cell gene expression from tissue morphology across multiple species and organs.
• The framework's integration of model interpretation and ViT offers a general-purpose tool for spatial transcriptomics.
• This approach accelerates scientific discovery and enhances precision in medical diagnostics and treatments.