Nicheformer (spatial single-cell)

Nicheformer is a single-cell transformer from Helmholtz Munich that adds spatial context tokens, so it can learn how a cell's neighbourhood shapes its state rather than treating cells as isolated. The 2024 bioRxiv preprint pretrained it on 110M cells, 57M dissociated and 53M spatial, and applied it to niche and spatial-context tasks in tumours, such as predicting a cell's tissue neighbourhood from its expression. It is intended for researchers working with spatial transcriptomics of the tumour microenvironment. The limitation is that spatial data remain sparse compared with dissociated data, so the spatial half of the training set is thin and predictions for rare tissue contexts are uncertain. For a newcomer: Nicheformer learns that where a cell sits changes what it does, which matters for understanding how tumours organise themselves.