Visium, Xenium or Stereo-seq: choosing a spatial platform before you spend the money

The spatial platforms are not competing versions of the same product. They make a different trade between how many genes you see and how precisely you know where they were. Choose on that trade, not on the resolution number in the brochure.

The fundamental trade

Sequencing-based, whole transcriptome (Visium, Visium HD, Stereo-seq): you measure everything, but capture happens on a grid of spots or bins, and each unit may pool several cells. Discovery is possible — you can find a gene nobody expected — but cell-level statements require deconvolution.

Imaging-based, targeted panel (Xenium, MERFISH, CosMx): you measure a chosen panel of a few hundred to a few thousand genes at single-molecule precision, assigned to segmented cells. Cell-level statements are direct, but you can only find what is on the panel, and panel design becomes the most consequential decision in the project.

What resolution means in practice

Standard Visium spots are 55 µm across, which is several cells in most tissues. Visium HD and Stereo-seq offer much finer bins, but small bins capture few transcripts each, so in practice people aggregate them back up — the effective resolution is set by how much signal you need per unit, not by the physical bin size. Imaging platforms genuinely resolve single cells, with the caveat that segmentation errors assign transcripts to the wrong neighbour, particularly in dense tissue.

Match the platform to the question

• "What is spatially organised in this tissue, and I do not know what I am looking for." Whole-transcriptome capture. You cannot discover a gene that is not on a panel.
• "Are these two specific cell types in contact, and how often?" Imaging-based. Deconvolved proportions from a mixed spot cannot establish contact between individual cells.
• "How does an architecture differ between conditions?" Either — but with enough sections per condition. One section per group is the most common and most fatal design error in spatial work.
• "I need this to work on FFPE archival blocks." Check probe-based support for your species and tissue explicitly; performance on old blocks varies enormously with fixation.

Costs that are not the kit

Spatial projects fail on sectioning quality and on tissue area more often than on chemistry. Poor sections mean folds and tears that produce artefacts no analysis can rescue. Capture areas are small, so the choice of which region goes on the slide is itself an experimental decision — and it should be made with a pathologist, not by whoever is cutting.

On the analysis side, budget for the reference: deconvolution needs a matched single-cell reference of the same tissue. If none exists publicly for your system, you may need to generate one, and that cost belongs in the plan from the start.

What the two actually look like, measured

Everything above is the standard framing. Below are the numbers from our own runs on public reference sections of each platform, because the trade-off is easier to judge against real figures than against adjectives.

	Visium (human lymph node)	Xenium (breast cancer FFPE)
Units measured	4,025 spots	167,780 cells
Genes	whole transcriptome	313-gene panel
Depth per unit	5,999 genes / 20,239 UMI per spot	164 transcripts per cell (median)
What a unit is	several cells pooled	one segmented cell
Background measure	none built in	negative controls, 0.095% of signal
Main technical risk	spot composition must be inferred	segmentation assigns transcripts to the wrong cell

The two risks are different, and both are quantifiable

Visium's risk is that a spot is not a cell. Cell-type proportions per spot have to be inferred by deconvolution, and in our benchmark against known ground truth that inference carried a typical error of 7.5 percentage points, with one abundant cell type overestimated by more than nine. A shift smaller than that is not recoverable from Visium data alone. There is a second, cruder failure too: in our Visium run, one of ten spatial domains — 248 spots — was defined entirely by mitochondrial genes. It is a tissue-quality region, not a niche, and only reading the marker list revealed it.

Xenium's risk is that the panel is the ceiling and the boundaries are drawn by software. Genes outside the panel were never asked about, so absence cannot be claimed. In our Xenium run, seven common breast lineages were all represented on the panel, but five only partially — KRT18, TP63, JCHAIN, CDH5 and DCN are simply not there. On segmentation, 1,417 segmented objects contained no transcripts at all, and the median nucleus occupied 0.294 of its cell's area; a large nucleus-only population would mean cytoplasm was never captured.

How to decide in one question

Ask what would change your conclusion. If the answer involves a gene you cannot name in advance, you need whole transcriptome coverage and therefore Visium (or a sequencing-based alternative). If it involves telling two adjacent cell types apart in situ, you need single-cell resolution and therefore Xenium, and you should check the panel against your markers before committing. If it involves both, the honest answer is that one section will not settle it, and the cheapest route is usually Visium plus a small single-cell reference for deconvolution rather than a second imaging run.

A pragmatic recommendation

If it is your first spatial experiment and the question is exploratory, run whole-transcriptome capture on a small number of well-chosen sections, look at what emerges, then design a targeted imaging panel for the follow-up if you need cell-level contact statements. Doing it in the other order — an expensive targeted panel chosen before you know what matters — is how people end up with beautiful data that cannot answer their question.