Spatial transcriptomics analysis, explained without the marketing

Spatial data is sold as "single-cell plus coordinates". It is not, and the difference determines which analysis is valid. The first question is always whether your measurement units are cells or spots.

Two families, two different problems

Spot-based platforms (Visium, Visium HD at coarser binning, Stereo-seq at larger bin sizes) capture whole transcriptomes under a region that usually contains several cells. Your unit is a mixture, so the central problem is deconvolution: estimating what proportion of each spot comes from which cell type.

Imaging-based platforms (Xenium, MERFISH, CosMx) measure a targeted gene panel at single-molecule resolution and assign transcripts to segmented cells. Your unit is closer to a cell, so deconvolution is unnecessary — but segmentation error becomes the dominant technical artefact, and the panel limits what you can discover to what someone chose to include.

Deconvolution in practice

Every method (RCTD, cell2location, SPOTlight, Tangram and others) needs a single-cell reference of the same tissue. The reference is the analysis: if a cell type is missing from it, its signal is redistributed to whatever is closest, and you will confidently report a population that is not there. Matching the reference to your tissue, condition and species matters more than the choice of algorithm.

Sanity check the output against the histology. If the pathologist's annotation says epithelium and your deconvolution says immune, the deconvolution is wrong, not the pathologist.

Spatially variable genes

These are genes whose expression depends on position rather than being randomly distributed — the honest starting point for "what is spatially organised here". Two cautions. First, tissue architecture alone produces significance: a gene expressed only in a structure that occupies one region of the section will be spatially variable without that being a discovery. Second, these methods return long lists; interpret them in the context of structures you can name, not as a ranked table of hits.

Niche and neighbourhood analysis

Once cell types or proportions are assigned, you can ask which types co-occur more than expected. This is where the biology usually is — an immune population at a tumour boundary, a stromal niche around a vessel. Two things to control: the null model (co-occurrence must be tested against a randomisation that preserves tissue shape, not a uniform field), and section-to-section variability (one section is one observation; conclusions need several).

Where spatial analyses usually go wrong

• Treating spots as cells. Clustering Visium spots and calling the clusters cell types is the most frequent error in the field. They are regions, and should be described as such.
• One section per condition. With n = 1 per group, every difference is confounded with the section. Reviewers know this.
• Ignoring the tissue image. The H&E is data. An analysis that never compares against it has skipped the cheapest available validation.
• Over-interpreting low-count spots. Edge spots and folded tissue produce artefacts that look like biology in a UMAP and disappear when you look at the slide.

What a complete spatial deliverable contains

Per-section QC, clustering with spatial overlays, deconvolution proportions with the reference stated, spatially variable genes tied to identifiable structures, neighbourhood statistics with the null model described, and figures overlaid on the tissue image rather than floating in UMAP space. If the report never shows the tissue, it is not a spatial analysis.