KCNMA1 in fibroblast niches and the PDAC tumor-stroma interface

This repository is a focused continuation of the cell-resolved Xenium analysis of GSE313662. The first project reconstructed and explored the treated PDAC atlas. Here, the analysis is narrowed to one point: quantifying how fibroblast KCNMA1 expression relates to the tumor-stroma interface.

I carried out this work during my six-month Master 1 bioinformatics internship with the Regulations of Ion Channel in Cancer team at the Institut de Biologie Valrose, within Universite Cote d'Azur, CNRS UMR 7277 and Inserm U 1091.

The data come from six Xenium Prime 5K sections in GSE313662. Five sections contain an interpretable malignant compartment and are used for the interface analyses.

From Visium to Xenium

The ym22000/pdac-ion-channels-m1 project placed KCNMA1 on the stromal side of PDAC and associated it with activated or myofibroblastic CAF programs. Visium also suggested that KCNMA1 could occur at the tumor-stroma interface in some sections. This was an interesting starting point, but each Visium spot contains several cells. The apparent interface signal could therefore reflect a mixture of tumor cells, fibroblasts and other stromal cells.

Xenium makes the question testable at cell resolution. It allows us to identify the expressing cells, preserve their coordinates and measure their distance from the tumor compartment. The analysis therefore separates cell identity, local spatial organization and distance to tumor rather than trying to answer everything with one correlation.

Building the spatial approach

A positive control: WNT5A

Before testing KCNMA1, the workflow was evaluated with WNT5A. This gene is involved in tumor-CAF communication in PDAC and is expected to mark fibroblast territories close to tumor cells (Nat Commun. 2023). Its Xenium distribution also provides a clear visual reference for an interface-associated signal.

Limits of pooled Spearman correlation

A first analysis correlated gene expression with an interface signature across all cells. This standard Spearman test summarizes a global monotonic relationship, but it does not use the spatial coordinates. Cells next to one another and cells on opposite sides of a section are treated in the same way.

This is problematic for a localized interface. A gene may be concentrated in small tumor-proximal niches while remaining absent from most cells. Pooling all cells dilutes this pattern, and spatially neighboring cells are not independent observations. Spearman is therefore retained as a descriptive screen, not as the main spatial test.

Step 1: Moran's spatial autocorrelation

Fibroblasts were grouped into 50 micrometer microdomains containing at least five fibroblasts. Moran's I then tests whether microdomains with similar expression values occur near one another.

• A positive Moran's I indicates spatial clustering.
• A value close to zero is compatible with a weak or random spatial pattern.
• A negative value indicates neighboring dissimilar values.

The test used 499 permutations and FDR correction. Radii of 35, 50 and 75 micrometers were also examined to check whether the result was stable across spatial scales.

For WNT5A, Moran's I was positive and significant in the five tumor-bearing sections. The signal was therefore not randomly scattered among fibroblasts: it formed reproducible local domains. Pt01 lacked an exploitable tumor compartment and was not used to support an interface conclusion.

Step 2: distance from the tumor

Moran's I establishes spatial organization, but it does not say whether a domain is close to the tumor. A KD-tree was therefore used to calculate, for each fibroblast, the distance to the nearest annotated tumor cell. The fibroblasts were then summarized in the same 50 micrometer microdomains.

For each microdomain, the analysis retained the median tumor distance, mean gene expression and detection rate. Correlations were calculated separately for each section before looking for a recurrent direction across patients. A negative rho means that the fibroblast signal decreases as distance from the tumor increases.

WNT5A showed a negative gradient in all five tumor-bearing sections. The median rho was -0.237 for mean expression and -0.230 for detection. The section-level direction was significant after correction (FDR = 0.0469). Together with Moran's I, this confirms that the workflow can recover an organized fibroblast program preferentially located near tumor cells.

Analysis overview

1. Reconstruct cell states from DEG-supported Xenium clusters.
2. Identify tumor epithelial cells and seven fibroblast states.
3. Compare KCNMA1 across fibroblast states with patient-aware edgeR.
4. Test whether KCNMA1 forms spatially autocorrelated fibroblast microdomains with Moran's I.
5. Compare KCNMA1 with a myCAF program by local neighborhoods and bivariate spatial statistics.
6. Construct a signed malignant/non-malignant boundary.
7. Measure the distance from each fibroblast microdomain to the nearest tumor cell.
8. Use WNT5A and ANO1 as positive spatial controls for near-tumor fibroblast gradients.
9. Evaluate consistency at the section level rather than treating every cell as an independent patient.

Main findings

KCNMA1 is not a universal tumor-stroma boundary marker.

The data support a more specific result:

• KCNMA1 forms significant fibroblast spatial structure in four of the five tumor-bearing sections at 50 micrometers;
• activated and myCAF-like fibroblasts contain more KCNMA1 than several other fibroblast states;
• local KCNMA1-myCAF association is strongest in Pt09 and Pt11;
• KCNMA1 does not show a consistent decrease with increasing tumor distance across the five sections;
• the custom signed-boundary test is also inconsistent between patients;
• WNT5A and ANO1, used as controls, show reproducible negative tumor-distance gradients.

A cautious interpretation is:

KCNMA1 marks a subset of spatially organized activated/myCAF-like fibroblast niches in treated PDAC, but these niches are not consistently restricted to the immediate tumor-stroma interface.

This is an association at RNA level. It does not prove BK channel activity, direct tumor-fibroblast signaling or a causal fibroblast phenotype.

KCNMA1 and ANO1

ANO1 provided a second control. Like WNT5A, it formed organized fibroblast domains and showed a negative tumor-distance gradient in all five tumor-bearing sections. Its median expression rho was -0.223, with a section-level FDR of 0.0469. The two controls show that the workflow can detect a recurrent tumor-proximal fibroblast pattern when it is present.

KCNMA1 gave a different and less uniform result. Its expression was spatially clustered in four of the five tumor-bearing sections and was enriched along an activated/myCAF-like fibroblast axis. However, only two sections showed a negative tumor-distance gradient. Across sections, the median rho was +0.041 and the cohort-level FDR was 0.688.

This does not make KCNMA1 uninteresting. It indicates that the signal is not a pure or universal boundary pattern. Some KCNMA1+ fibroblast niches occur near tumor cells, especially in sections such as Pt11, whereas other niches occur in deeper activated or myCAF-rich stroma. The exploratory deep-stroma analysis was positive in all five sections, although it did not remain significant after correction across the tested metrics.

The result leaves an open biological question: what determines whether a KCNMA1-positive fibroblast state is positioned at the tumor-stroma interface or maintained in deeper stroma?

Possible determinants include the local tumor program, extracellular-matrix organization, vascular or pericyte-associated niches and treatment-induced remodeling. These remain hypotheses. The present data support spatial heterogeneity, not a causal mechanism.

Documentation

• Xenium KCNMA1 interface notes
• Biological question
• Chronological methods
• Results
• How to interpret the statistics
• Critical assessment
• Data availability
• Software and parameters

Outputs

• KCNMA1 Moran maps
• KCNMA1/myCAF spatial statistics
• Tumor-distance gradients
• Signed boundary analysis
• Fibroblast differential expression and UMAP
• Fibroblast-state continuum

Repository layout

.
|-- assets/        README illustration
|-- config/        sample and cell-state maps
|-- docs/          biological, statistical and critical documentation
|-- environment/   software versions
|-- results/       selected final figures and aggregate tables
|-- scripts/       core reconstruction and interface analyses
|-- requirements.txt
`-- run_pipeline.ps1

Raw matrices, full per-cell tables and large RDS objects remain outside the repository and can be rebuilt from GEO.

Reproduce

Rscript .\environment\install_r_packages.R
.\run_pipeline.ps1

If the reconstructed annotation objects already exist locally, use -SkipReconstruction. A full run is computationally heavy because it rebuilds the cell-state atlas before performing the interface tests.

The script order and outputs are indexed in scripts/README.md.