What am I looking at?
Each point is one single cell from the CELLxGENE Census: its learned-identity sample summary token, reduced to two dimensions. Cells that the model represents similarly sit close together.
This map contains the exported exp8 cells, balanced across labels and studies; it is not the whole training corpus. Cell types and CELLxGENE datasets are annotations added after training. Nearby cells may share biology, technical effects, or both.
Hovera point to read its metadata.
Clickto pin it and see its ten nearest neighbours in the model’s space.
Legendclick a class to isolate it, then “Show datasets” to colour its CELLxGENE datasets.
Searcha cell type, tissue, donor or assay to light up matching cells.
