TissueFormer: a neural network for labeling tissue from grouped single-cell RNA profiles

Area of Science:

• Computational Biology
• Genomics
• Neuroscience

Background:

• Single-cell RNA sequencing (scRNA-seq) offers deep gene expression insights but current analysis often misses crucial population-level signals.
• Interpreting scRNA-seq data typically focuses on individual cells, neglecting the importance of cellular composition for inferring sample phenotypes.
• Tissue identity, disease state, and other sample-level characteristics are often dictated by the mix of cells within a sample.

Purpose of the Study:

• To introduce TissueFormer, a Transformer-based neural network designed to analyze groups of single-cell RNA profiles.
• To enable the inference of population-level labels from cellular composition while maintaining single-cell resolution.
• To provide a computational framework for predicting sample-level phenotypes influenced by cellular diversity and tissue organization.

Main Methods:

• Development of TissueFormer, a Transformer-based neural network architecture.
• Application of TissueFormer to spatial transcriptomic data from mouse brains.
• Comparison of TissueFormer's performance against single-cell foundation models and traditional machine learning methods using pseudobulk and cell type composition data.

Main Results:

• TissueFormer successfully predicted cortical areas from groups of cells in spatial transcriptomic data.
• The model outperformed existing single-cell foundation models and machine learning approaches.
• Automated construction of high-resolution brain region maps in individual mice was enabled.

Conclusions:

• TissueFormer effectively leverages cellular composition for accurate prediction of population-level phenotypes.
• The framework advances the analysis of spatial transcriptomic data for high-resolution mapping.
• TissueFormer offers a versatile tool for diverse applications in biological and clinical research where cellular diversity is key.