The problem. Single-cell RNA-seq gives deep, whole-transcriptome cell-type detail but throws away location. Spatial methods keep location but are either coarse (capture) or targeted (imaging). Each has what the other lacks. Can you combine them — use rich single-cell data to annotate spatial data, or use spatial data to give single cells a position?

The idea. Tangram learns an alignment: it optimises a mapping of single-cell profiles onto spatial locations so that predicted spatial expression matches the measured spatial data. Once learned, that mapping transfers whatever the single-cell data has — fine cell types, genes outside a targeted panel, annotations — onto the tissue coordinates, and conversely gives single cells spatial context. It’s a deep-learning optimisation over the correspondence between the two modalities.

Why it matters. This is the integration capstone for the spatial day: it explicitly unites the single-cell world (yesterday) with the spatial world (today), which is precisely the analytical core of STU-style work — deconvolving spots, imputing unmeasured genes, annotating spatial data from a single-cell reference. It’s the same “transfer from reference to target” theme (scANVI, SingleR, Seurat) now spanning modalities rather than batches.

Verdict. Foundational for single-cell–to–spatial integration and widely used for deconvolution and annotation transfer. Its mappings are inferred, so they warrant validation against known markers. Read it as the method that makes single-cell and spatial data one analysis instead of two.