This paper details the introduction and evaluation of Nicheformer, a novel transformer-based foundation model designed for single-cell and spatial omics data analysis. This model is pretrained on SpatialCorpus-110M, a massive collection of over 110 million cells encompassing both dissociated single-cell and spatially resolved transcriptomics data from human and mouse tissues. Nicheformer learns cell representations that encode spatial context, enabling it to perform effectively on spatial-specific downstream tasks, such as spatial composition prediction and spatial label prediction, where models trained only on dissociated data fail. The research highlights Nicheformer's ability to transfer spatially aware annotations from spatial to dissociated data and demonstrates its superior performance against existing baselines like scVI and PCA in predicting cellular niche structures and densities. Overall, the work positions Nicheformer as an advance toward creating a generalizable, multiscale model for single-cell and spatial biology.

References:

• Schaar A C, Tejada-Lapuerta A, Palla G, et al. Nicheformer: a foundation model for single-cell and spatial omics[J]. bioRxiv, 2024: 2024.04. 15.589472.