Nicheformer: a foundation model for single-cell and spatial omics

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Tissue structure and function are dictated by the cellular microenvironment.
• Spatial single-cell genomics offers a powerful method for studying cell interactions.
• Understanding spatial context is crucial for interpreting single-cell data.

Purpose of the Study:

• To introduce Nicheformer, a transformer-based foundation model for spatial single-cell analysis.
• To develop a model capable of learning spatial context from diverse transcriptomic data.
• To enable the prediction of spatial microenvironments for dissociated cells.

Main Methods:

• Training Nicheformer on a large dataset (SpatialCorpus-110M) of human and mouse single-cell and spatial transcriptomics data.
• Utilizing a transformer architecture to capture spatial relationships.
• Evaluating model performance on downstream tasks like spatial composition and label prediction.

Main Results:

• Nicheformer successfully learns cell representations that incorporate spatial context.
• The model excels in predicting spatial composition and labels.
• Demonstrated that models trained solely on dissociated data cannot fully capture spatial microenvironment complexity.
• Nicheformer can transfer spatial information to single-cell RNA sequencing datasets.

Conclusions:

• Nicheformer represents a significant advancement in machine learning for spatial single-cell analysis.
• Integrating multiscale data is essential for comprehensive understanding of spatial biology.
• The model facilitates the prediction of spatial context for dissociated cells, enhancing scRNA-seq data interpretation.