Nicheformer: Foundation Model for Single-Cell and Spatial Omics

Nicheformer is a transformer-based foundation model published in Nature Methods (2025, vol. 22, pp. 2525–2538), pretrained on SpatialCorpus-110M — a curated corpus of over 57 million dissociated and 53 million spatially resolved cells across 73 tissues from human and mouse. The model enables prediction of spatial context for dissociated scRNA-seq cells, spatial composition prediction, and label transfer via linear probing and fine-tuning. Key finding: models trained only on dissociated data fail to recover spatial microenvironment complexity, underscoring the need for joint single-cell and spatial pretraining.