Data-driven fine-grained region discovery in the mouse brain with transformers.

Area of Science:

• Computational biology
• Neuroscience
• Genomics

Background:

• Spatial transcriptomics enables detailed analysis of tissue organization, but scaling to organ-level datasets presents computational challenges.
• Accurate detection of spatial domains is crucial for understanding tissue architecture and function.

Purpose of the Study:

• To establish a scalable workflow for self-supervised spatial domain detection in large-scale spatial transcriptomic data.
• To develop a method capable of identifying known and novel tissue domains with high spatial coherence.

Main Methods:

• Utilized an encoder-decoder architecture (CellTransformer) for hierarchical feature learning from cellular and molecular data.
• Integrated representation learning with GPU-accelerated clustering for scalability to millions of cells.
• Applied the workflow to MERFISH and Slide-seqV2 datasets from mouse brain tissue.

Main Results:

• CellTransformer successfully scaled to multi-million cell datasets, outperforming existing methods.
• Identified spatial domains consistent with existing brain atlases (e.g., Allen Mouse Brain CCF) and discovered hundreds of novel regions.
• Demonstrated high consistency across multiple tissue sections and animals, enabling complex multi-animal analyses.

Conclusions:

• CellTransformer provides a performant and scalable solution for fine-grained spatial domain detection in organ-scale spatial transcriptomics.
• The workflow facilitates the integration of cells across tissue sections and aids in the discovery of uncataloged anatomical areas.
• This advancement supports more comprehensive neuroanatomical studies and the exploration of tissue organization.