Nicheformer is a transformer foundation model for single-cell and spatial omics from Fabian Theis's lab — the flagship scientific FM of the TUM/Helmholtz single-cell program (joint→Helmholtz Munich ICB). It is pretrained on SpatialCorpus-110M, a curated corpus of ~110 million cells (57M dissociated + 53M spatially resolved) spanning 73 tissues across human and mouse.

By jointly modeling dissociated single-cell data and targeted spatial transcriptomics, Nicheformer transfers spatial context to dissociated data and excels at downstream spatial tasks — notably niche/spatial-composition and spatial-label prediction — under both linear-probing and fine-tuning. Published in Nature Methods (2025, vol. 22, pp. 2525–2538); weights are released on HuggingFace under theislab/Nicheformer.

Model Details

Architecture DENSE
sciencebiologyfoundation-modelresearch