Is that spatial domain real? Telling a tissue niche from an artefact

Spatial clustering never fails. Ask for ten domains and you get ten, coloured neatly across the tissue, and the figure looks like biology whether or not any of it is. The question that matters is which of those domains would survive being questioned.

The short version: look at what each domain's marker genes actually are before you interpret the map. In our own public Visium run, one of ten domains — 248 spots — is defined entirely by mitochondrial genes. It is a tissue-quality region, not a niche, and no amount of spatial statistics would have revealed that. Reading the marker list did, in about ten seconds.

Check 1 · Read the marker genes before you read the map

This is the cheapest and most decisive check, and it is skipped constantly. On a public human lymph node section (4,025 spots after QC, clustered into 10 domains), the top markers per domain tell you immediately which are real:

• Domain 0 — IGHG1, IGKC, IGHG4, IGHG3: immunoglobulin-producing plasma cell regions. Real, and exactly what a lymph node should contain.
• Domain 6 — CXCL13, MS4A1: B-cell follicle and germinal centre. Real.
• Domain 2 — TRBC1, TRAC, IL7R, CCL21: the T-cell zone. Real, and its position relative to the follicles is interpretable.
• Domain 3 — MT-ND4, MT-ATP6, MT-ND3, MT-CO3, MT-CO2, MT-ND1. Every one of the top markers is mitochondrial. This is not a cell niche; it is the part of the section where tissue quality is worst.

Domain 3 would have been perfectly presentable in a figure. It has a spatial position, a coherent shape, a cluster number and a colour. Interpreted as biology it would have generated a story about a metabolically distinct region — a story with no support whatsoever.

Check 2 · Ask whether the domain is explained by a technical gradient

For every domain, compare its median UMI count, median genes per spot and mitochondrial percentage against the rest of the section. If one domain sits at the extreme of any of those, the burden of proof shifts: it needs biological markers that are not explained by depth or damage. Tissue edges, folds and regions under a bubble all produce convincing-looking domains.

Check 3 · Confirm the genes are actually spatially structured

Clustering can produce domains from genes that vary randomly across spots. Moran's I measures whether a gene's expression is spatially organised at all. In the same section the top spatially variable genes were IGKC (0.878), IGHG4 (0.853), FDCSP (0.748) and CCL21 (0.681) — follicular and T-zone genes with strong spatial structure, which is what a lymph node should give. A domain whose defining genes all have near-zero Moran's I is a clustering artefact regardless of how tidy the map looks.

Check 4 · Change the resolution and see what survives

Domain counts are a parameter, not a discovery. Re-run the clustering at a coarser and a finer resolution: real structures merge and split predictably, while artefacts appear and vanish. Any claim about a specific domain should hold across a reasonable range, and the report should say what range was tested.

What to ask for in a spatial deliverable

1. Marker genes per domain — the full list, not a curated selection.
2. QC metrics per domain, so quality-driven domains are visible rather than hidden.
3. Moran's I (or equivalent) for the genes the interpretation rests on.
4. A note on which domains were stable across clustering resolutions.

Our worked example, with the code and every number, is on the spatial sample page. If you have a section and want the interpretation to survive review, tell us what you have.