Your GeoMx Data Failed QC — Now What?

You’ve spent months on experimental design, tissue preparation, and ROI selection. You’ve run your GeoMx DSP experiment. The data comes back — and the QC report has red flags everywhere.

This is more common than you’d think — and it doesn’t necessarily mean your experiment failed.

Here’s what to do.

Understanding GeoMx QC Metrics

Before panicking, understand what the QC metrics actually measure and which ones matter most.

Sequencing Saturation

What it measures: The proportion of unique reads relative to total reads. As you sequence deeper, you encounter more duplicate reads and saturation increases.

Healthy range: ~50% or higher is generally acceptable for most experiments. Below 50% means you may benefit from additional sequencing.

What to do if it’s low: This is actually the easiest problem to fix — you can often re-sequence the same library to increase depth without repeating the experiment. Contact NanoString or your sequencing core about top-up sequencing.

Area and Nuclei Count

What it measures: The physical size of each ROI and the number of nuclei detected within it.

Why it matters: Very small ROIs (few nuclei) tend to have lower signal-to-noise and higher variability — though quality is also influenced by tissue quality, staining, segmentation, and assay type.

Red flag threshold: ROIs with fewer than ~100 nuclei often show increased variability. Below 50, results typically become unreliable for most downstream analyses.

What to do: Consider merging small adjacent ROIs if biologically appropriate, or exclude very small ROIs and note the exclusion in your methods.

Limit of Quantification (LOQ)

What it measures: The number of genes detected above background noise in each ROI. LOQ essentially defines the threshold above which a gene’s signal is distinguishable from background noise. Calculated from negative control probes — a gene is “detected” if its expression exceeds the LOQ derived from the negative probes.

Healthy range: This depends heavily on your probe panel, tissue type, and ROI size. Well-performing WTA samples often detect several thousand genes per ROI, commonly in the 5,000-10,000+ range depending on tissue type and ROI size.

Red flag: If some ROIs detect 1,000 genes while others in the same experiment detect 8,000, investigate. The low-detection ROIs may have tissue damage, poor hybridization, or insufficient material.

Background-to-Signal Ratio

What it measures: How much of the signal in each ROI comes from non-specific binding versus true gene expression.

Red flag threshold: As a general guideline, ratios approaching 1.0 indicate that signal is dominated by background noise — at that point, distinguishing real biology from noise becomes extremely difficult. Ratios below 0.3 are typically healthy.

What to do: High background can sometimes be addressed computationally (background subtraction), but if the ratio is too high, the data may be unusable for that ROI.

The Triage Decision

Not all QC failures are equal — and not all require the same response. Here’s how to categorize:

Salvageable (fix computationally)

Low sequencing saturation → re-sequence
A few ROIs with low nuclei counts → exclude those ROIs
Moderate background → apply background correction and compare results with/without
Uneven normalization → try alternative normalization methods (Q3, CPM, TMM)

Partially salvageable (reduced scope)

Subset of ROIs fail while others pass → analyze the passing subset, note the reduced sample size
One compartment fails while others pass → restrict analysis to passing compartments
One tissue type performs well, another doesn’t → analyze separately

Not salvageable (re-run)

Uniform high background across all ROIs → hybridization failure, tissue degradation, or protocol issue
Very low gene detection across all ROIs → RNA quality issue (check DV200 or other FFPE-appropriate RNA quality metrics if available; RIN can be less informative for FFPE tissue)
Systematic batch effects that confound your experimental design → design problem, not data problem

Common Causes of GeoMx QC Failure

Most QC failures can be traced back to three areas: tissue quality, ROI selection, and protocol execution.

Tissue Quality

FFPE tissue varies enormously in RNA preservation. Factors that degrade RNA quality:

Extended fixation time (over-fixation)
Old blocks (RNA degrades over years in paraffin)
Inadequate fixation (autolysis before fixation)
Tissue necrosis

If you suspect tissue quality, check whether the QC failures correlate with specific tissue blocks or patients. If one block fails and others pass, the tissue — not the protocol — is the issue.

ROI Selection

ROIs that include necrotic regions, tissue folds, or edge artifacts will produce poor data. During ROI selection:

Avoid tissue edges and folds
Exclude obviously necrotic regions
Ensure each ROI contains a homogeneous tissue compartment
Verify fluorescent marker staining quality before committing to ROI placement

Panel and Protocol

Expired reagents can reduce hybridization efficiency
Incorrect probe concentrations affect sensitivity
Protocol deviations (hybridization temperature, wash stringency) impact background levels

What Computational Analysis Can (and Can’t) Fix

Background Correction

GeoMx experiments include negative control probes that estimate non-specific binding. You can subtract this background from your expression values to improve signal-to-noise.

When it helps: Moderate background levels where real signal is still present above the noise floor.

When it hurts: If background approaches signal strength, subtraction can remove real biological signal along with the noise. In our experience, aggressive background correction on noisy data can produce worse results than no correction at all.

Alternative Normalization

Q3 (upper quartile) normalization is a standard GeoMx recommendation. But if your data has issues, alternatives may perform better depending on the study design and batch structure:

TMM normalization: Often performs well as an alternative to Q3, particularly for datasets with composition differences across groups
Quantile normalization: Can reduce technical variability, though assumptions should be checked
Housekeeping-based normalization: Uses a stable gene set as reference, useful when global normalization assumptions are violated
RUV-based correction: Packages like standR provide remove-unwanted-variation workflows designed specifically for spatial transcriptomics data

Try multiple approaches and compare. If your biological conclusions are consistent across normalization methods, you can be more confident in the findings.

Outlier Exclusion

Removing failing ROIs is legitimate — but document it rigorously. Report:

How many ROIs were excluded
What criteria were used for exclusion
Whether the exclusion changed the balance of your experimental groups
Whether key conclusions hold with and without the excluded ROIs

The key is not to treat QC as pass/fail, but as a guide to what your data can reliably support.

When to Call for Help

GeoMx QC troubleshooting requires a combination of computational skills and biological knowledge. If you’re unsure whether your data is salvageable, or if you’ve tried standard corrections and the results still don’t make sense, bringing in an experienced bioinformatics partner can save months of circular troubleshooting.

At Cytogence, GeoMx data rescue is one of our most common consulting engagements. We’ve worked with datasets ranging from “slightly noisy” to “nearly unusable” and have developed systematic approaches for determining what can be salvaged, what requires re-analysis with different parameters, and what genuinely needs to be re-run.

The earlier you involve bioinformatics expertise, the more options you have.

Cytogence provides expert GeoMx DSP analysis and troubleshooting. If your spatial data needs help, we can help you determine what’s salvageable — and what’s not.