Spatial and Single-Cell Multi-Omics: The Next Validation Frontier

Single-cell and spatial methods have made a specific kind of claim cheap: that a cell type exists, that a regulatory element is active in it, and that a programme differs between conditions. What has not become cheaper is deciding whether those claims are true.

The field's response has been scale — more cells, more donors, more modalities. That helps with precision but not with validity, because a well-powered association can still be an association.

The frontier, in other words, is not more data. It is a validation standard appropriate to a new unit of analysis.

The last row is where most new claims now live, and where the evidence is thinnest. A differential result within a cluster depends on the clustering, the cell numbers, and the composition of the sample — three things that are analysis choices rather than biology.

A standard for this unit would require: pre-specified clustering, pseudobulk replication at the donor level, a stated minimum cells per cluster, and a report of whether the result survives reasonable changes in resolution. Most papers do some of this; few do all four.

The most useful distinction I can offer is a three-level hierarchy, because papers routinely present level 2 as if it were level 3.

Level 1 is now routine and, in well-executed papers, reliable. Level 2 is common and often over-interpreted. Level 3 is rare, expensive, and the only level that supports a mechanistic sentence.

A paper can legitimately stop at level 2 — but it should say so. The phrase "we identify a population that..." is a level 2 claim and reads perfectly well as one.

Spatial methods are often justified by "tissue context", which is true but nonspecific. Three specific gains:

Composition as a variable rather than a confound. In bulk tissue, cell composition changes with disease, which confounds expression differences. Spatial data let you ask whether a programme changed within a cell type or because the tissue's composition shifted.

Neighbourhood effects. Whether a cell's programme depends on what surrounds it cannot be answered in dissociation-based assays. Spatial data make that question askable, though not yet easily answerable.

Region-restricted programmes. A programme confined to a lesion margin or a hypoxic core is invisible in bulk and can dominate a single-cell dataset if the region is small.

None of these is a mechanism. Each is a level 2 claim of a kind that bulk data cannot make — which is precisely why spatial methods are worth the trouble.

Four errors survive review and then fail replication:

• • Clustering as discovery. A cluster that exists only at one resolution is an artefact of a parameter, not a population.
• • Cell-level statistics. Treating cells as replicates inflates significance by orders of magnitude; the donor is the replicate.
• • Composition confound. A difference in cell-type proportions reported as a difference in expression per cell.
• • Pseudo-replication across sections. Multiple sections from one donor are not independent donors.

The third is the most consequential in disease work, and the easiest to check: deconvolve, then ask whether the effect survives adjustment.

Prediction: the next round of high-impact atlas papers will be distinguished not by cell numbers but by an explicit validation arm — a perturbation performed in the system the atlas was built from.

Refuted if: the field continues to reward scale alone and validation remains a follow-up paper, in which case the citation structure will keep separating "resource" from "mechanism" — a stable arrangement, but one that leaves the atlas unvalidated.

If you are generating an atlas: pre-specify the analysis, replicate at the donor level, and plan one perturbation from the start. One well-chosen perturbation converts a resource into a finding, and it costs a fraction of the sequencing.

If you are consuming an atlas: check which of the three levels the claim reaches before you build on it. Most of what you need is level 1, much of what you want is level 2, and the sentences you will cite are usually level 3 — which may not exist yet.

Both paths end in the same place: a perturbation in a matched system, which is what we build.

Contact EDITGENE to request a project consultation or quote.

Annotations and differential results depend on clustering, cell numbers and composition; conclusions should report these parameters and test stability across reasonable choices.

Association in an atlas is not evidence of regulation; mechanistic claims require perturbation in a matched system.

• • Zhang K, et al. A single-cell atlas of chromatin accessibility in the human genome. Cell. 2021;184:5985–6001.
• • Moses L, Pachter L. Museum of spatial transcriptomics. Nature Methods. 2022;19:534–546.
• • Squair JW, et al. Confronting false discoveries in single-cell differential expression. Nature Communications. 2021;12:5692.
• • Murphy AE, Skene PJ. SPATAC-seq and the analysis of chromatin accessibility. Nature Methods. 2023;20:1–6.