In spatial transcriptomics, every experiment forces a choice between resolution and coverage. In my systematic re-analysis of 128 kidney spatial datasets, that trade-off became starkly evident.
On the single-cell imaging side (Xenium and CosMx), we can analyze ~1.9 million cells. Because the kidney's architecture is tightly organized, scoring 18 nephron functional modules yielded a 90.1% confident cell-type assignment rate across sections without any supervised reference training.
On the whole-transcriptome spot side (Visium and Visium HD), we can gain coverage of ~18,000 genes, but each 55 µm spot blends 5 to 30 cells.
To resolve these mixtures, researchers rely on deconvolution algorithms. I compared two leading tools: RCTD and Tangram, across the cohort. The finding worth paying attention to: deconvolution outputs are hypersensitive to the single-cell reference prior used. Changing the reference skewed inferred cell fractions by more than 2-fold in adjacent cortical spots.
Deconvolution gives an estimated neighborhood composition, not an experimental measurement of single-cell state.
Find out more from my case study here: https://zqzneptune.github.io/cases/renalspatial/04_spot_bin_lens.html
How do you benchmark spot deconvolution when orthogonal spatial single-cell data is unavailable?
Spot deconvolution algorithms are hypersensitive to the single-cell reference prior used—changing the reference skews inferred cell fractions by more than 2-fold in adjacent cortical spots. Deconvolution estimates spot neighborhood mixture, not true single-cell states.