Is FOS/JUN expression in my spatial data a dissociation artifact if I never dissociated the tissue?

Visium and Xenium sections aren't dissociated, so the stress induction from warm collagenase digestion doesn't happen directly to your spots. But if the scRNA-seq reference or database you're using to name clusters was built from dissociated cells digested warm, its stress signature can still leak into your spatial labels through label transfer or deconvolution.

How many marker genes do I need before I trust a cell-type label?

At least two concordant markers that also pass a specificity filter, show spatial coherence on the tissue image, and match a curated database like CellMarker or PanglaoDB for the tissue you're working with. A single significant gene from FindMarkers is a hypothesis, not confirmation.

Does FindAllMarkers with only.pos=TRUE fix the ribosomal/stress gene problem?

No. only.pos=TRUE just removes downregulated genes from the output for readability; ribosomal and housekeeping genes that are upregulated by fold-change alone will still pass. You still need a specificity filter (like filter_rank_genes_groups' thresholds) on top of it.

Does marker validation differ between Visium spots and Xenium cells?

Yes. Visium spots contain one to ten cells, so a cluster is a mixture by construction and needs deconvolution (RCTD, Cell2location, and similar) before you can trust a single cell-type label per spot. Xenium, MERFISH, and CosMx approach single-cell resolution, but their targeted gene panels of a few hundred to a few thousand genes mean a marker you'd expect from a whole-transcriptome reference may simply not be on the panel.

Sanity check · Spatial Transcriptomics

How to Sanity-Check Marker Genes and Cell Type Labels in Spatial Transcriptomics

FindMarkers and rank_genes_groups will happily hand you ribosomal genes and dissociation stress as your top "cell type marker", here is the checklist that catches it before it reaches a figure legend.

By Ming "Tommy" Tang, Director of Bioinformatics in Big Pharma · Reviewed September 2026 · 5 min read

You clustered your Visium or Xenium data, ran FindAllMarkers() or sc.tl.rank_genes_groups(), and now you have a table of top genes per cluster sitting on your screen. The next step feels obvious: eyeball the top few genes, match them to a cell type you recognize, and write "Fibroblast" or "Tumor epithelium" in the cluster label. That step is where most spatial marker calls quietly go wrong.

FindMarkers and rank_genes_groups are purely statistical. They return whatever genes differ most between one group and the rest, with no concept of biology attached. In dissociated single-cell data that bias already produces ribosomal, mitochondrial and stress-response genes at the top of marker lists often enough to be a known failure mode. In spatial data it gets worse, because a Visium spot is not a cell, it is a 55-micron patch of tissue holding one to ten cells, and because the reference atlases and companion scRNA-seq datasets used to name spatial clusters carry their own processing artifacts into your labels.

This page gives you an ordered set of checks to run on a marker table before you trust it, cheapest first, plus the fixes for the failure modes you'll actually hit. Budget the next hour to run checks 1 through 6 on your current cluster set; save the database cross-check and deconvolution step for the clusters that survive.

What it looks like when it's happening

Top of a cluster's marker table is dominated by RPS/RPL genes or MT-/mt- genes instead of anything you could name a cell type after
FOS, JUN, or heat-shock genes (HSPA1A, DNAJB1) sit near the top of every cluster's marker list, not just one, suggesting a shared artifact rather than distinct biology
VlnPlot for your candidate marker shows most cells or spots at zero with a long right tail from a handful of outliers driving the mean
The gene that supposedly defines a cluster looks like random salt-and-pepper noise, not a coherent region, when plotted back onto the tissue image with spatial_scatter or SpatialFeaturePlot
Two clusters that look histologically identical on the H&E image get different top markers, and the difference tracks ribosomal or mitochondrial fraction rather than anything you can name
A cluster's top 'marker' is MALAT1, which is a nuclear lncRNA QC signal for spot/cell quality, not a cell-identity gene
Your marker list returns no hits, or only generic hits, when checked against CellMarker or PanglaoDB for the tissue you're working with
P-values from FindMarkers or rank_genes_groups look absurdly small, like 1e-200, for nearly every gene in a cluster

Why it happens

A Visium spot is a mixture, not a cell. It covers one to ten cells, so the expression profile clustering algorithms compare is already an average across whatever cell types happen to co-occur in that patch of tissue. When FindMarkers/rank_genes_groups looks for genes that separate one cluster's average profile from another's, the genes that win are often whichever cell type is more abundant in the mix, or genes that are simply the most highly expressed transcripts in nearly every cell type present. Ribosomal protein genes and mitochondrial genes fall into that second category: they carry large absolute counts everywhere, so small proportional differences between clusters translate into large fold changes and tiny p-values under a Wilcoxon or t-test run over thousands of cells or spots, even though the difference is a depth or composition artifact, not a marker of identity.

Dissociation-stress genes get into spatial analysis by a less direct route. Visium and Xenium sections are not dissociated, so FOS, JUN and heat-shock genes are not being directly induced in your spots the way they are in scRNA-seq prepared with warm collagenase digestion at 37°C. But the reference data you use to name spatial clusters usually was dissociated: a public scRNA-seq atlas you're pulling markers from, or a matched scRNA-seq run on the same specimen that you're using for label transfer or deconvolution. If that reference was digested warm, its own marker lists carry the stress signature, and that signature rides straight into your spatial cluster names the moment you borrow its markers or run RCTD/Cell2location against it.

There's a statistical mechanism underneath both problems. Cells or spots within a cluster are not independent observations, they come from the same tissue section, the same patient, often the same local neighborhood, so a cell-level Wilcoxon or t-test massively understates the true variance and inflates significance. This is why marker p-values from single-cell-resolution tests routinely look implausible; pseudobulking to the sample level before running a proper DE tool like PyDESeq2 gives a much more honest picture, at the cost of needing multiple biological replicates.

Finally, remember that any RNA-seq measurement, spatial or not, is a single snapshot of a process running on a molecular clock: transcription runs at roughly 10-100 nucleotides per second and translation at roughly 10 amino acids per second. A gene that lights up as a top marker may reflect a transient state, stress from the exact moment of tissue processing, a cell-cycle phase, a local hypoxia gradient, captured at the instant of the assay rather than a stable identity you can build a cell-type label on.

The checks

Run them in order. Each one tells you what healthy looks like and what the problem looks like.

0/8 checked · saved in this browser

01Grep the marker table for known junk gene families
Take the top 15-20 genes per cluster from your FindMarkers/rank_genes_groups output and flag anything starting with RPS/RPL (ribosomal), MT-/mt- (mitochondrial), or matching FOS, JUN, HSPA1A, HSPA1B, DNAJB1 (dissociation stress). This takes under a minute and catches the most common failure before you invest in anything downstream.
python
```
junk = adata.uns['rank_genes_groups']
import pandas as pd
df = pd.DataFrame(junk['names']).head(20)
for cl in df.columns:
    flagged = [g for g in df[cl] if g.startswith(('RPS','RPL','MT-','mt-')) or g in ('FOS','JUN','HSPA1A','HSPA1B','DNAJB1')]
    print(cl, flagged)
```
Healthy
Top 15-20 genes per cluster are receptor, enzyme, transcription factor or structural genes tied to known cell biology (e.g. CD3E for T cells, KRT8 for epithelium), with at most one or two incidental ribosomal/mito hits.
Red flag
Three or more of the top 10 genes for a cluster are RPS/RPL, MT-/mt-, or stress-response genes.
02Filter the marker list for specificity, not just significance
Run scanpy's filter_rank_genes_groups (or the Seurat equivalent of manually filtering on pct.1/pct.2) to require genes be expressed in a minimum fraction of the target cluster, have a minimum fold change, and stay below a maximum expression fraction in the rest of the data.
python
```
sc.tl.filter_rank_genes_groups(
    adata,
    min_in_group_fraction=0.25,
    min_fold_change=1,
    max_out_group_fraction=0.5,
)
```
Healthy
The surviving marker list per cluster is noticeably shorter and cleaner, dominated by genes with obvious cell-type relevance.
Red flag
Almost nothing survives the filter for a given cluster (no gene clears a 25% in-group / 50% out-group specificity bar, meaning the cluster likely isn't a real, distinct population), or the list is unchanged from the raw output, meaning ribosomal/housekeeping genes were already passing on fold-change alone.
03Check expression is spread across the cluster, not driven by outliers
Violin-plot each candidate marker split by cluster and look at the shape of the distribution within the cluster itself, not just the difference in means.
r
```
VlnPlot(seurat_obj, features = c("CD3E", "CD8A"), group.by = "seurat_clusters")
```
Healthy
Most cells or spots in the cluster sit clearly above background for the marker, and the marker stays near zero in other clusters.
Red flag
A bimodal violin where most of the cluster sits at zero and a handful of cells/spots with very high expression are pulling the mean up: the gene is 'significant' because of outliers, not because it defines the cluster.
04Confirm the marker is spatially coherent on the tissue image
Plot the candidate marker's expression directly onto the tissue coordinates and compare against the H&E histology or your annotated regions.
python
```
sq.pl.spatial_scatter(adata, color="GENE_NAME")
```
Healthy
Expression concentrates in an anatomically sensible region that matches the histology, e.g. an epithelial layer, a stromal band, or an immune aggregate.
Red flag
Expression looks like random salt-and-pepper noise across the section, tracks tissue edges or fold artifacts, or is diffusely uniform everywhere, none of which is consistent with a spatially organized cell-identity marker.
05Check the marker isn't just tracking sequencing depth
Correlate the candidate marker's expression per spot or cell against total UMI counts (or total genes detected) for that spot/cell.
Healthy
Weak or no correlation between the marker and total counts across the dataset; a true identity marker should decouple from sequencing depth.
Red flag
The marker's expression pattern across the tissue matches the total-count or total-gene heatmap almost exactly, you're looking at a depth artifact, not biology.
06Quantify ribosomal fraction and MALAT1 as cluster-level QC
Compute ribosomal gene fraction and MALAT1 expression per spot/cell and compare distributions across clusters rather than just eyeballing single genes in the marker table.
python
```
adata.obs['ribosomal_fraction'] = (
    adata[:, adata.var_names.str.startswith(('RPS', 'RPL'))].X.sum(axis=1)
    / adata.X.sum(axis=1)
)
```
Healthy
Ribosomal fraction is roughly comparable across clusters (allowing for genuinely protein-synthesis-heavy cell types), and MALAT1 is uniformly high (above roughly 1% of the median) in spots/cells you intend to call real biology on.
Red flag
A 'cluster' turns out to be defined almost entirely by high mitochondrial fraction or low MALAT1, that's a damaged-cell or empty-spot QC bucket, not a cell type.
07Cross-check candidate markers against a curated database
For each marker that survives the checks above, look it up in CellMarker (about 13,600 human markers across 467 cell types) or PanglaoDB (over 6,000 gene-cell-type associations) for the tissue you're working with, and require at least two independent markers to agree on the same identity.
Healthy
Markers are found in the database, tied to a cell type that fits the tissue, and at least two markers per cluster point to the same identity.
Red flag
The marker is absent from the database, or is listed as expressed broadly across many unrelated cell types (i.e. it's not specific enough to be a real identity marker).
08For spot-based platforms, validate with deconvolution before trusting the cluster label
For Visium or Slide-seq data, run a reference-based deconvolution method (RCTD, Cell2location, SPOTlight, Stereoscope or DestVI) against a matched scRNA-seq reference, and compare the predicted per-spot cell-type proportions to what your cluster marker calls implied.
Healthy
Spots assigned to a given cluster show one dominant deconvolved cell type that matches the marker-based call.
Red flag
Deconvolution shows the 'cluster' is actually a roughly even mixture of two or three cell types, you clustered a mixture region and gave it a single-cell-type name.

What to do about it

Filter marker output computationally before eyeballing it

When: Top markers for a cluster are visibly dominated by ribosomal, mitochondrial, or stress genes and you want a clean shortlist without hand-curating every cluster.

Run scanpy's filter_rank_genes_groups with min_in_group_fraction, min_fold_change, and max_out_group_fraction thresholds (0.25/1/0.5 are the documented defaults) on the raw rank_genes_groups output before you look at the table.

Caveat: Default thresholds are somewhat arbitrary. A real but sparsely detected marker (common with spot dropout or a shallow targeted panel) can get filtered out at a 25% in-group fraction; loosen it for sparse data rather than trusting the defaults blindly.

Require three lines of evidence before naming a cluster

When: A marker clears the statistical test but the violin plot is bimodal, the spatial plot is speckled, or the database check comes back generic.

Only accept a marker-based label once it clears expression-distribution (VlnPlot), spatial coherence (spatial_scatter/SpatialFeaturePlot), and database confirmation together, not any single one in isolation.

Caveat: Slower and manual. Build a small script that runs all three checks per cluster automatically once you have more than a handful of clusters to review.

Deconvolve spot data instead of clustering-then-labeling

When: Working with Visium or Slide-seq, where each spot genuinely contains one to ten cells and a cluster is by construction a mixture.

Run RCTD or Cell2location against a matched scRNA-seq reference and report continuous per-spot cell-type proportions rather than forcing a single hard label per cluster.

Caveat: Needs a good matched reference. A reference from the wrong tissue, wrong disease state, or the wrong dissociation protocol produces confident-looking but wrong proportions.

Rebuild the scRNA-seq reference with cold dissociation if you control it

When: You're generating your own scRNA-seq reference for label transfer or deconvolution, and stress genes (FOS, JUN, heat-shock genes) dominate that reference's own cluster markers.

Switch to cold active protease dissociation at around 6°C rather than warm collagenase at 37°C, which is documented to substantially reduce the conserved stress signature.

Caveat: Not possible for archival or public reference atlases you don't control. If you're stuck with a warm-dissociated public reference, at minimum flag and downweight its stress genes before using it for labeling.

Pseudobulk before trusting cluster-level p-values

When: P-values from FindMarkers or rank_genes_groups look implausibly small (near-zero) for most genes in a cluster, a sign that cell-level non-independence is inflating significance.

Aggregate raw counts to the sample or section level per cluster, then run a proper bulk DE tool such as PyDESeq2 instead of testing at the cell/spot level.

Caveat: Requires enough biological replicates (multiple sections or patients per condition); doesn't work with a single tissue section, where you have no independent replication to pseudobulk over.

When not to "fix" it

Not every ribosomal, mitochondrial, or stress-gene hit is an artifact to strip out. Plasma cells, hepatocytes, and other genuinely protein-synthesis-heavy cell types legitimately carry high ribosomal gene expression as real biology, not contamination, filtering it out just because it's a ribosomal gene will erase a correct marker. Similarly, a tumor's necrotic or hypoxic core, or tissue from an ischemic or treated specimen, can legitimately show elevated mitochondrial and stress transcripts because the tissue itself is under stress, not because of how it was processed; in that case the "artifact" is the finding. And if you're doing spatial domain or niche analysis rather than strict cell typing, a marker that is spatially coherent but fails a cell-type specificity or database check may still be exactly the signal you want, cell-type-specific spatially variable genes identified through dedicated spatial methods routinely surface context-specific markers that general-purpose databases like CellMarker or PanglaoDB simply don't contain.

Five things experienced analysts do here

Read every marker table with the gene-family flags on. Before trusting the top-N list for a cluster, check it against RP[SL], MT-/mt-, and the FOS/JUN/HSP family first, it takes thirty seconds and catches most naive labeling mistakes before they cost you an hour of database searching.
Never call a cell type off one gene. Require at least two concordant markers plus a database hit before a label goes in a figure legend; a single gene, however significant, is a hypothesis, not an identification.
Treat a p-value from FindMarkers/rank_genes_groups as a starting point, not a conclusion. Always run the VlnPlot and spatial plot before you write a cell-type name anywhere a reviewer or pathologist will see it.
Know your platform's actual resolution before you use cell-type language. Visium and Slide-seq spots are mixtures of one to ten cells, call them regions or domains until you've deconvolved them; Xenium, MERFISH, and CosMx approach single-cell resolution but still depend on segmentation quality you should check separately.
Track the dissociation protocol behind any reference you borrow markers or labels from. A public scRNA-seq atlas or companion dataset built with warm collagenase digestion will hand you stress genes baked into its own cluster markers, and those ride straight into your spatial labels via label transfer or deconvolution.

From the book

Ch. 9 · Adapting from scRNA-seq to Spatial Transcriptomics — Spatial marker validation is scRNA-seq validation plus spatial coordinates and image context, so getting the core matrix thinking right first is what lets you adapt to Visium, Xenium or whatever comes next.
Ch. 9 · MALAT1 and NEAT1 as scRNA-seq QC Markers — MALAT1 showing up as a cluster 'marker' is a tell that you're looking at a QC artifact, not a cell type, and this lesson gives the threshold to catch it.
Ch. 10 · Mitochondrial and Ribosomal Gene Naming Conventions — Knowing exactly which prefixes mark mitochondrial and ribosomal genes across species is what makes the thirty-second grep-the-marker-table check possible in the first place.

These lessons are of 290 in What They Don't Teach You in Bioinformatics.

Questions people ask

Why are ribosomal genes (RPS, RPL) showing up as top markers in my spatial clusters?: Ribosomal protein genes carry large absolute counts in nearly every cell type, so small proportional differences between clusters translate into large fold changes and very small p-values under a Wilcoxon or t-test run over thousands of cells or spots. Filter with scanpy's filter_rank_genes_groups (min_in_group_fraction, min_fold_change, max_out_group_fraction), or check ribosomal fraction per cluster directly before trusting the raw marker table.
Is FOS/JUN expression in my spatial data a dissociation artifact if I never dissociated the tissue?: Visium and Xenium sections aren't dissociated, so the stress induction from warm collagenase digestion doesn't happen directly to your spots. But if the scRNA-seq reference or database you're using to name clusters was built from dissociated cells digested warm, its stress signature can still leak into your spatial labels through label transfer or deconvolution.
How many marker genes do I need before I trust a cell-type label?: At least two concordant markers that also pass a specificity filter, show spatial coherence on the tissue image, and match a curated database like CellMarker or PanglaoDB for the tissue you're working with. A single significant gene from FindMarkers is a hypothesis, not confirmation.
Does FindAllMarkers with only.pos=TRUE fix the ribosomal/stress gene problem?: No. only.pos=TRUE just removes downregulated genes from the output for readability; ribosomal and housekeeping genes that are upregulated by fold-change alone will still pass. You still need a specificity filter (like filter_rank_genes_groups' thresholds) on top of it.
Does marker validation differ between Visium spots and Xenium cells?: Yes. Visium spots contain one to ten cells, so a cluster is a mixture by construction and needs deconvolution (RCTD, Cell2location, and similar) before you can trust a single cell-type label per spot. Xenium, MERFISH, and CosMx approach single-cell resolution, but their targeted gene panels of a few hundred to a few thousand genes mean a marker you'd expect from a whole-transcriptome reference may simply not be on the panel.

Guide · How to Tell If You Overclustered in Spatial Transcriptomics
Guide · How to Detect Integration Over-Correction in Spatial Transcriptomics
Guide · How to Log-Transform Counts Without Fooling Yourself in Single-Cell RNA-seq
Compare · Squidpy vs Seurat: Which One Should You Use?
Glossary · Pseudobulk
Glossary · Spatial transcriptomics

Sources

Exploring Spatial Transcriptomics A Dive into Visium Data Analysis in Python — Spot-mixture reasoning and spatial coherence validation via spatial_scatter
Scanpy filter_rank_genes_groups documentation — Specificity filter thresholds (min_in_group_fraction, min_fold_change, max_out_group_fraction)
Scanpy rank_genes_groups documentation — Statistical test options and the case for pseudobulking with PyDESeq2
Seurat Guided Clustering Tutorial — FindAllMarkers, VlnPlot/FeaturePlot/DoHeatmap validation, and ROC classification power
Dissociation of solid tumor tissues with cold active protease for single-cell RNA-seq minimizes conserved collagenase-associated stress responses — Cold active protease dissociation fix for dissociation-stress gene artifacts
Fully-automated and ultra-fast cell-type identification using specific marker combinations from single-cell transcriptomic data — CellMarker and PanglaoDB as curated marker reference databases
Identifying cell-type-specific spatially variable genes with ctSVG — Context-specific spatial markers missed by general-purpose databases
Gut mucosa dissociation protocols influence cell type proportions and single-cell gene expression levels — Epithelial cells' sensitivity to dissociation artifacts

Part of the Marker gene sanity series.