My goal: Which cell states inferred in 10x Visium co-occur in the same regions, and are those co-occurrences supported by real cell-cell neighborhoods in 10x Xenium?

This is a toy project designed to help me work with mostly with Xenium, but alsso Visium and scFFPE-seq datasets together and explore ways to integrate them. You can find my code in the notebooks folder.

I explore a single FFPE tissue of breast cancer block (Stage II-B, ER + /PR − /HER2 +) analyzed with a trio of complementary technologies scFFPE-seq, Visium and Xenium by Janesick et al Nat. Commun. 2023. As far as I know, its the first round of such analysis being done at that time. Datasets were obtained from the GEO repository under id GSE243280 and Xenium FFPE Human Breast with Custom Add-on Panel

Overview of the platforms:

• scFFPE-seq technology designed to perform high-throughput single-nucleus RNA sequencing on archival, formalin-fixed paraffin-embedded (FFPE) tissue samples from 10x Genomics. It is well suited for building a reference atlas from archived FFPE samples.
  • Readout: next-generation sequencing
  • Coverage: targeted fixed RNA panel
  • Resolution: single-cell
  • Output matrix: cell × gene
• Visium is a sequencing-based spatial transcriptomics technology from 10x Genomics. It is well suited for capturing global tissue architecture and gives whole-transcriptome spatial context but at spot level (mixed cells per spot).
  • Readout: next-generation sequencing
  • Coverage: whole transcriptome
  • Resolution: ~55 µm spots (typically containing 2-10 cells)
  • Output matrix: spot × gene
• Xenium is an imaging-based spatial transcriptomics platform from 10x Genomics. It is well suited for high-resolution, cell-level spatial analysis and gives single-cell spatial positions but usually a targeted panel (fewer genes).
  • Readout: microscopy imaging
  • Coverage: targeted gene panel (typically 100-5000 genes)
  • Resolution: single-cell to subcellular
  • Output matrix: cell × gene

How to combine them? A practical strategy is to use scFFPE-seq as the reference atlas (to to learn what cell types/states look like molecularly), use that reference to deconvolve mixed Visium spots into likely cell-type proportions and then use Xenium to validate whether those inferred co-occurrences are truly adjacent at single-cell resolution.

In practice:

1. Build a high-quality annotated scFFPE-seq reference.
2. Deconvolve Visium spots using the reference (for example, with cell2location).
3. Test whether co-occurring Visium cell states are supported by true Xenium neighborhoods.

Some useful links:

• intro to models and methods for spatial transcriptomics by Ben Raphael (CGSI 2023) https://www.youtube.com/watch?v=CRuSrd8JWI0
• https://spatialdata.scverse.org - I find it a bit better in teh context of spatial transcirptomics to the sc-best-practices tutorial (https://www.sc-best-practices.org/spatial/introduction.html)

Core single-cell & spatial analysis:

🧬 scanpy • anndata • squidpy

Probabilistic modeling & deconvolution:

🧬 scvi-tools • cell2location

Spatial data infrastructure:

🧬 spatialdata • spatialdata-io

General analysis & plotting:

🧬 pandas • numpy • matplotlib

visium-xenium-reference-mapping/
├── env/
├── data/
│   ├── reference/
│   ├── visium/
│   └── xenium/
├── notebooks/
│   ├── 01_reference_qc_annotation.ipynb
│   ├── 02_visium_qc_exploration.ipynb
│   ├── 03_visium_cell2location.ipynb
│   ├── 04_xenium_loading_qc.ipynb
│   ├── 05_xenium_annotation_neighborhoods.ipynb
│   └── 06_cross_platform_validation.ipynb
└── src/

My steps:

• Get the reference into a solid annotated AnnData.
• Run Visium + cell2location.
• Load Xenium and transfer the same labels and markers.
• Run cross-platform validation in Notebook 06:
  • matched cell types between Visium and Xenium
  • abundance/fraction concordance plots
  • pairwise Visium A-B co-occurrence vs Xenium A-B neighborhood z comparison

Results are reasonable, but the cross-platform concordance is quite moderate. This is my first-pass integration and validation analysis on this dataset. Additional QC (at the cell/spot level), more careful label transfer/lifting across platforms, and tighter harmonization of cell-state definitions could improve agreement in future iterations.