SpatialFormer: universal spatial representation learning from subcellular molecular to multicellular landscapes

Area of Science:

• Genomics
• Systems Biology
• Computational Biology

Background:

• Understanding gene spatial expression and multicellular organization is crucial for disease diagnosis and biological studies.
• Current models face challenges in integrating gene expression data with cellular spatial information.

Purpose of the Study:

• Introduce SpatialFormer, a novel hybrid framework to effectively integrate multimodal and multiscale single-cell information.
• To leverage convolutional networks and transformers for learning gene expression and spatial distribution within cellular niches.

Main Methods:

• Developed SpatialFormer, a hybrid framework combining convolutional networks and transformers.
• Pretrained the model on a large dataset of 700 million cell pairs from 17 million spatially resolved single cells across 71 Xenium slides.
• Employed a pairwise training strategy to merge gene spatial expression profiles with cell niche information.

Main Results:

• SpatialFormer successfully distills biological signals for various tasks, including single-cell batch correction, cell-type annotation, and co-localization detection.
• Perturbation analysis identified key gene pairs involved in immune cell-cell communication in pulmonary fibrosis.
• Detected epithelial-myoepithelial co-localization and tumor transition signals in breast cancer.

Conclusions:

• SpatialFormer significantly advances the integration of gene expression and spatial data in single-cell analysis.
• The framework enhances the understanding of cellular dynamics and provides new avenues for biomedical research applications.
• This approach offers improved capabilities for disease diagnosis and fundamental biological process studies.