Overview
- Developed by UCSF and the Allen Institute, CellTransformer uses a transformer architecture to learn spatial relationships among cells, similar in principle to large language models.
- The data-driven parcellation shows strong agreement with the Allen Institute’s Common Coordinate Framework, providing external validation of the inferred anatomical boundaries.
- Beyond recapitulating known structures, the model proposes previously uncataloged subregions in poorly annotated areas such as the midbrain reticular nucleus, with further computational and experimental checks anticipated.
- The workflow scales to multimillion-cell datasets and demonstrated nearly perfect cross-animal consistency for up to 100 spatial domains across four mice comprising about nine million cells.
- Authors report the method is tissue-agnostic, positioning it for mapping other organs and cancer tissues where large spatial transcriptomics datasets are available.