What is spot deconvolution in spatial transcriptomics?

It's the computational step that estimates what fraction of each cell type is present in a spatial transcriptomics spot, since sequencing-based platforms like 10x Visium capture multiple cells per spot rather than single-cell resolution. It typically uses a matched scRNA-seq reference to translate a spot's mixed expression profile into per-cell-type proportions.

Cell2location vs RCTD: which should I use?

Cell2location tends to score higher on accuracy and is more robust when sequencing depth varies across spots, based on a mouse brain benchmark (PCC = 0.9837, RMSE = 0.0373). RCTD is faster and simpler to run, a reasonable tradeoff if your dataset has consistent depth and you're processing many samples.

Do I need deconvolution for Visium data?

If you plan to claim which cell types occupy which regions, yes, a raw Visium spot is a mixture, not a cell, and reading marker genes directly off a spot without deconvolving risks misassigning identity. If you only need broad tissue domains or expression patterns, you can sometimes skip it, but any per-cell-type claim needs deconvolution first.

What reference data does spot deconvolution require?

Most methods, including RCTD and CARD, require a single-cell RNA-seq reference from the same or a closely matched tissue with cell types already annotated. Some newer methods are reference-free, but the field's most benchmarked tools (RCTD, CARD, cell2location) all depend on a good-quality scRNA-seq reference.

Does spot deconvolution fix RNA spillover?

No. Deconvolution assumes a spot's mixed signal reflects genuine cell-type composition, not technical bleed-through from neighboring spots during sectioning. Spillover can masquerade as a mixed-cell-type spot, so check spot mixtures against histology before trusting deconvolution output as biology.

Glossary · Single-Cell and Spatial

Spot deconvolution

Turning a Visium spot's blended signal into per-cell-type proportions, and why the method you pick changes the biology you report.

By Ming "Tommy" Tang, Director of Bioinformatics in Big Pharma · Reviewed September 2026 · 3 min read

Also: cell2location, RCTD

Definition

Spot deconvolution is the computational estimation of cell-type composition within each spot of sequencing-based spatial transcriptomics data, such as 10x Visium, because each spot spans roughly 50-100 µm and usually captures a blended transcriptional signal from multiple cells rather than one. It combines a spot's gene expression profile with a matched single-cell RNA-seq reference to infer the proportion of each cell type present in that spot. Methods span several algorithmic families: probabilistic (RCTD, CARD, cell2location), non-negative matrix factorization (SPOTlight), graph-based (DSTG), optimal transport (SpaOTsc), and deep learning (Tangram).

You run into spot deconvolution the moment you move from single-cell data to sequencing-based spatial transcriptomics like 10x Visium and try to say "this region is macrophage-rich" instead of just "this region has high CD68." A Visium spot is roughly 50-100 µm across and can hold zero to several cells, so its expression profile is a composite of whatever cells happen to sit under it. Deconvolution is the step that pulls that composite apart into estimated proportions of each cell type.

This decides whether your spatial map of "T cell zone" or "fibroblast niche" reflects real tissue architecture or an artifact of spot size. It's also the step where you commit to a reference scRNA-seq dataset and a specific algorithm, both of which shape the answer more than most people expect.

Why it matters

Get this wrong and you can invent cell types that don't exist. A spot with mixed fibroblast and T cell signal is not automatically a hybrid cell: it can be ordinary spillover, where transcripts bleed into a neighboring spot during sectioning, before deconvolution gets a chance to explain the mixture biologically. Treating every mixed spot as genuine composition without checking for spillover is a fast way to report a cell type that is really a technical artifact.

The method you pick also moves your numbers directly. In a mouse brain benchmark, cell2location scored the best overall accuracy (PCC = 0.9837, RMSE = 0.0373) and, along with spatialDWLS, stayed more robust than RCTD when sequencing depth varied across spots. RCTD trades some of that robustness for speed and a simpler workflow, which matters when you're deconvolving many capture areas without GPU time to spare. Choosing RCTD for a depth-uneven dataset purely because it's fastest can quietly bias which cell types you report as dominant in low-depth regions.

Where people get it wrong

The most common mistake is treating a Visium spot as a single cell, then reading a cluster of spots on a spatial plot as literal cell identities instead of a mixture that still needs deconvolving. A close second: feeding normalized, log-transformed counts into a method like CARD, which automatically normalizes cell-type proportions to sum to 1.0 per spot but expects raw, unnormalized counts as input, normalize twice and the pipeline runs fine while the proportions become meaningless.

The other trap is conflating a genuine mixed-cell-type spot with RNA spillover. Both look the same on paper: a spot expressing markers from two unrelated lineages. Only one is biology. Deconvolution methods generally assume the mixture is real composition and don't distinguish it from spillover, so a spot called "50% T cell, 50% fibroblast" deserves a second look before you trust it as a genuine mixed niche rather than a boundary artifact.

A concrete example

You have a Visium slide of a tumor section and a matched scRNA-seq reference from the same tissue type. Before deconvolving, confirm the Visium counts are raw (not normalized) and that the scRNA-seq reference covers the cell types you actually expect, since deconvolution can only assign proportions to types present in the reference. With RCTD (from the spacexr package), you choose whether a spot can hold many cell types or just two, then sanity-check the output against the histology image rather than trusting the proportions at face value.

library(spacexr)
myRCTD <- create.RCTD(spatialRNA, reference, max_cores = 4)
# 'full' mode: arbitrary number of cell types per spot
# 'doublet' mode: caps each spot at two cell types, faster and simpler
myRCTD <- run.RCTD(myRCTD, doublet_mode = "full")

Related terms

Questions people ask

What is spot deconvolution in spatial transcriptomics?: It's the computational step that estimates what fraction of each cell type is present in a spatial transcriptomics spot, since sequencing-based platforms like 10x Visium capture multiple cells per spot rather than single-cell resolution. It typically uses a matched scRNA-seq reference to translate a spot's mixed expression profile into per-cell-type proportions.
Cell2location vs RCTD: which should I use?: Cell2location tends to score higher on accuracy and is more robust when sequencing depth varies across spots, based on a mouse brain benchmark (PCC = 0.9837, RMSE = 0.0373). RCTD is faster and simpler to run, a reasonable tradeoff if your dataset has consistent depth and you're processing many samples.
Do I need deconvolution for Visium data?: If you plan to claim which cell types occupy which regions, yes, a raw Visium spot is a mixture, not a cell, and reading marker genes directly off a spot without deconvolving risks misassigning identity. If you only need broad tissue domains or expression patterns, you can sometimes skip it, but any per-cell-type claim needs deconvolution first.
What reference data does spot deconvolution require?: Most methods, including RCTD and CARD, require a single-cell RNA-seq reference from the same or a closely matched tissue with cell types already annotated. Some newer methods are reference-free, but the field's most benchmarked tools (RCTD, CARD, cell2location) all depend on a good-quality scRNA-seq reference.
Does spot deconvolution fix RNA spillover?: No. Deconvolution assumes a spot's mixed signal reflects genuine cell-type composition, not technical bleed-through from neighboring spots during sectioning. Spillover can masquerade as a mixed-cell-type spot, so check spot mixtures against histology before trusting deconvolution output as biology.

Sources

Deconvolution chapter in Orchestrating Spatial Transcriptomics Analysis with Bioconductor — Definitions, algorithm categories, RCTD and CARD parameter behavior
Benchmarking and integration of methods for deconvoluting spatial transcriptomic data — cell2location vs RCTD accuracy and robustness benchmarks
SpatialPrompt: spatially aware scalable and accurate tool for spot deconvolution and domain identification in spatial transcriptomics — Spatial-aware deconvolution approach and runtime figures