Elucidating Spatial Cell Composition of Neuroblastoma (Group Project)

Three-month MRes group computing project with three other students, in collaboration with Dr Ximena Montano at King’s College London, who provided the spatial data.

The question

Neuroblastoma is a solid tumour whose cellular makeup varies substantially between patients. Single-cell RNA-seq describes which cell types are present but discards where they sit; single-cell spatial transcriptomics keeps the positions but measures a targeted gene panel with far less depth per cell. The project was a pilot combining the two, on NanoString CosMx spatial transcriptomics from 13 neuroblastoma tumours spanning stages 1 to 4.

Approach

Building a reference atlas. Public single-cell neuroblastoma studies were meta-analysed into a single reference spanning tumour cell diversity. This was the phase I worked on, and it faces the same integration problem as any multi-study atlas: batch effects between studies have to be removed without flattening the biological differences that make the atlas worth building.

Transferring labels to the spatial data. Cell populations in the CosMx data were annotated by transfer learning from that reference, with label-transfer methods benchmarked against each other rather than one chosen in advance.

Multimodal clustering. We adapted SiGra, a graph deep-learning model that combines multichannel immunohistochemistry images, a spatial neighbour graph and gene expression, so that clustering uses tissue morphology and physical neighbourhood alongside expression.

Outcome

Presented at Spatial Biology Research at King’s in July 2024, with a manuscript in preparation.

The CosMx data is Dr Montano’s and unpublished.