Novae: a graph-based foundation model for spatial transcriptomics data.

Area of Science:

• Molecular biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics offers high-resolution gene expression data within tissue microenvironments.
• Understanding spatial organization is crucial for tissue function and disease research.
• Current models face limitations with multi-slide analyses and batch effect correction.

Purpose of the Study:

• To develop a novel graph-based foundation model, Novae, for spatial transcriptomics.
• To overcome limitations of existing models in handling multiple slides and batch effects.
• To enable robust zero-shot domain inference across diverse datasets.

Main Methods:

• Designed Novae, a graph-based foundation model for extracting cell representations in spatial contexts.
• Trained Novae on a large dataset of nearly 30 million cells across 18 tissues.
• Implemented native batch effect correction and construction of a nested hierarchy of spatial domains.

Main Results:

• Novae achieves zero-shot domain inference across multiple gene panels, tissues, and technologies.
• The model natively corrects batch effects, improving data consistency.
• Novae successfully constructs a nested hierarchy of spatial domains.
• Supported downstream analyses including spatially variable gene/pathway analysis and trajectory analysis.

Conclusions:

• Novae is a robust and versatile tool for advancing spatial transcriptomics.
• The model facilitates deeper understanding of tissue microenvironments and disease mechanisms.
• Novae enhances biomedical research by providing powerful spatial gene expression analysis capabilities.