Want expert eyes on your spatial transcriptomics pipeline? From raw images to final figures, we help researchers catch silent failures before they become costly. Request a free consultation →
The Problem
There is an irresistible temptation in spatial analysis to interpret visually coherent clusters as biologically meaningful. Tools like BayesSpace and SpaGCN encourage this by incorporating spatial proximity into the clustering process.
Why It Happens
The resulting plots often look beautiful - smooth domains, sharp boundaries - but visual elegance does not equal biological truth. We’ve encountered several projects where “spatial domains” perfectly aligned to histology, yet collapsed when marker gene expression or cell type composition was examined. In some cases, the spatial coherence was purely driven by coordinate regularity and not expression similarity.
This happens because spatial clustering algorithms enforce locality as a prior. If expression is noisy or uninformative, the model still segments based on position. The output is a map - but not necessarily a meaningful one.
How We Address It
In our workflow, we always test whether spatial domains correlate with known cell types, known tissue zones, or functionally enriched pathways. We also compare to non-spatial clustering as a control. If spatial-only models add no interpretive value, we discard them. Clustering must serve the biology - not the other way around.
The Problem
Deconvolving cell types in spot-based data is powerful - but dangerous. Tools like stereoscope, Cell2location, and Tangram can infer cell type proportions across spots, but they depend heavily on the quality and match of the scRNA-seq reference.
Why It Happens
The most common failure we see is transferring labels from poorly matched references - different species, different developmental stages, or different tissue processing. The result: T cells in fibroblast zones, neurons in stromal tissue, or glial signatures in muscle. Worse, since the output looks quantitative, it is often accepted without questioning.
We also observe errors due to underpowered reference datasets, which fail to capture the diversity of cell types actually present in the tissue. These lead to “unknown” zones or false homogenization.
How We Address It
Our approach includes careful curation of the reference, including removal of low-quality or ambiguous cells, and thorough matching of preprocessing steps. We also compare multiple deconvolution tools and cross-check with spatial marker gene expression. If the spatial map contradicts biology, we discard or rework the deconvolution - even if it took days to run.
Using Xenium or CosMx for spatial analysis? We help research teams design, interpret, and avoid overclaims when working with panel-based spatial transcriptomics platforms. Request a free consultation →
The Problem
A common practice in spatial transcriptomics is to plot known marker genes across the tissue and interpret their localization. While this seems intuitive, it is also one of the most misleading forms of evidence.
Why It Happens
Gene expression levels in spatial platforms are noisy, dropout-prone, and confounded by local background and spot overlap. A smooth gradient in UMI counts may be an artifact of local capture efficiency. A hotspot of a marker gene might reflect only one outlier spot.
We have reviewed figures where the color scale was manipulated to exaggerate spatial signal, or where a gene’s presence in two spots was interpreted as a pattern. Worse, these visualizations often drive conclusions - about zonation, cell migration, or treatment response - without statistical validation.
How We Address It
Our standard practice is to combine marker gene maps with robust differential analysis (e.g., SPARK, SpatialDE), and to plot expression alongside spatial autocorrelation metrics. We also advise showing raw counts, not just normalized or log-scaled versions. Marker genes are only evidence if supported by statistics and matched histology.
The Problem
Detecting spatially variable genes (SVGs) is one of the most popular applications - but also one of the easiest to misuse.
Why It Happens
Tools like SPARK, SpatialDE, and trendsceek are powerful, but highly sensitive to underlying assumptions about spatial continuity and variance. Without careful pre-filtering and quality control, the outputs can be misleading. We have seen reports claiming hundreds of SVGs where a majority were mitochondrial genes or ribosomal proteins - artifacts of spatial noise, tissue folds, or section thickness, not true biological gradients.
Some groups misinterpret spatial variability itself as inherently meaningful - when in fact, it may arise from technical heterogeneity, slicing bias, or simply diffusion from necrotic regions.
How We Address It
We start by validating spot-level quality and applying spatial normalization methods that correct for tissue depth, RNA content, and slide-specific variance. Only then do we perform SVG analysis, typically with multiple tools and overlapping thresholds. We stratify by expression intensity, filter by spot density, and verify spatial patterns against known histological features. Most importantly, we test if the SVGs enrich expected pathways, or align with known cell type distributions. A list of spatially variable genes without context is not a result - it’s a hypothesis.
The Problem
High-resolution platforms like Xenium and CosMx offer subcellular spatial resolution - but with targeted gene panels. Many users mistakenly treat them like whole-transcriptome data.
Why It Happens
Workflows optimized for Visium are often ported directly to CosMx or Xenium data. But a 100–500 gene panel cannot support the same kinds of clustering, pathway inference, or deconvolution - especially when gene coverage is uneven or biologically unbalanced. We’ve seen projects where inflammatory signaling was declared absent simply because the key pathway genes weren’t included in the panel.
How We Address It
We begin every panel-based analysis by carefully matching panel content with biological questions. If the panel is not appropriate - we say so, and often stop before wasting resources. In suitable cases, we emphasize local co-expression, spatial enrichment, and segmentation-based analysis. We avoid overclustering, and discourage overstated conclusions from underpowered panels. CosMx and Xenium are powerful tools - when treated with respect for their design limits.
Asked to revise your spatial analysis for peer review? We've helped researchers rebuild reproducible pipelines, respond to reviewer demands, and regain confidence in their results. Request a free consultation →
The Problem
The most painful failures are not analytical - but procedural. Reviewers may request re-clustering, alternative normalization, or revised annotation - and the project team realizes too late that their pipeline is fragile or undocumented.
Why It Happens
Many spatial analyses are exploratory in nature, with ad hoc steps coded into notebooks or done interactively in GUI-based tools. When the time comes to regenerate figures or revise based on feedback, nothing is reproducible. Sometimes the data structure has changed. Sometimes the analyst is gone. Either way, the project stalls.
How We Address It
We treat all spatial analyses as reproducibility challenges from day one. Our workflows are scripted, modular, and version-controlled - typically using Snakemake or Nextflow. We document every image preprocessing step, spot filtering rule, and parameter value in code, with outputs logged and timestamped. For every figure, we can say exactly where it came from and how to regenerate it. Reanalysis doesn’t need to be painful - if the system was built for it.
Spatial transcriptomics is not a shortcut to pretty figures. It is a complex integration of histology, molecular biology, image processing, and statistical modeling - and each step can fail silently. Many failures arise not from lack of skill, but from false assumptions, uncritical workflow reuse, or overconfidence in tools.
At AccuraScience, we have seen firsthand how a single overlooked distortion or a poorly matched reference can ruin weeks of work. But we have also seen how disciplined, critical analysis - with full awareness of platform limitations - can unlock real spatial insight, even from noisy data.
If you take one thing from this series, let it be this: do not confuse automation with understanding. The best spatial bioinformaticians are those who question every default, validate every assumption, and respect the complexity of tissues, not just the convenience of software.
We are proud to help researchers uncover spatial biology that stands up to scrutiny - and survives peer review.
If you’ve read this far, you know how easily spatial transcriptomics analyses can go wrong - often without obvious warning signs.
We've encountered these failure modes in many forms: image misalignments mistaken for gradients, marker maps driving unsupported claims, and deconvolution results that collapse under scrutiny. Fixing these issues takes more than technical skill - it takes experience, pattern recognition, and a willingness to question the defaults.
Our senior bioinformaticians work closely with researchers to design, troubleshoot, and refine spatial transcriptomics pipelines that hold up under peer review. Whether you're preparing a new study or reanalyzing a challenging dataset, we’re here to help you get it right.
Request a free consultation →