Visium HD data analysis: from 2 µm bins to cells, domains and niches
How to analyze Visium HD: QC and tissue masking, 8 µm bins vs nucleus segmentation, RCTD or cell2location deconvolution, domains and niches.
Key takeaways
- A 2 µm Visium HD bin is smaller than a cell and holds only a few UMIs. Do not cluster raw 2 µm bins.
- Use 8 µm bins for robust domains and gradients; use nucleus segmentation and bin-to-cell assignment for cell-level questions.
- Deconvolution and label transfer need a reference that contains every cell type in the tissue.
- Mask folds, edges and off-tissue bins before analysis; they otherwise form their own clusters.
Visium HD captures whole-transcriptome expression on a continuous grid of 2 × 2 µm barcoded squares. That is close to single-cell resolution, but it is not single-cell data. A large part of a good Visium HD analysis is deciding how to turn bins into biologically meaningful units. This guide covers the workflow our spatial team uses.
Understand the data structure
Space Ranger outputs expression at 2 µm resolution and, by default, binned to 8 µm and 16 µm. Recent versions can also produce segmentation-based cell outputs. Key facts:
- A 2 µm bin covers a fraction of a typical cell and often contains only a handful of UMIs.
- 8 µm bins approach the size of a small cell and hold more signal per unit, but often mix neighboring cells.
- Tissue morphology (H&E or immunofluorescence) is registered to the grid. That image is what makes cell segmentation possible.
Step 1: Processing and registration
Re-run Space Ranger when needed with the correct probe set, slide serial and image registration, especially CytAssist alignment. Poor registration shifts every bin relative to the morphology, and every downstream cell assignment inherits the error. We check registration visually on tissue landmarks.
Step 2: QC and tissue masking
- Plot UMIs and genes per bin across the tissue at 8 and 16 µm.
- Mask folds, tears, bubbles and edges. These regions have abnormal counts and form clusters that look like tissue compartments.
- Assess lateral diffusion: transcripts detected just outside the tissue boundary show how far signal spreads. This sets a realistic limit on boundary sharpness.
- For FFPE, check RNA quality metrics and probe detection.
Step 3: Choose the resolution for your question
| Question | Recommended unit |
|---|---|
| Tissue domains and architecture | 8 µm (or 16 µm) bins |
| Expression gradients (distance to tumor border, crypt axis) | 8 µm bins |
| Cell-type neighborhoods and niches | Segmented cells |
| Cell-level differential expression | Segmented cells, then pseudobulk per sample |
| Low-expression genes | 16 µm bins, or aggregation by domain |
We usually deliver both 8 µm-bin and segmented-cell analyses and explain where each is used.
What our experts avoid: clustering raw 2 µm bins. Sparse bins cluster by UMI depth and tissue artifacts, not by biology. In one rescued project this produced 38 depth-driven clusters; moving to segmented cells gave 17 interpretable cell types.
Step 4: Segment nuclei and assign bins to cells
- Segment nuclei on the H&E (StarDist's H&E model works well) or on a DAPI/IF channel (Cellpose, StarDist).
- Expand nuclei by a tissue-appropriate distance to capture cytoplasm, without merging into neighbors.
- Assign 2 µm bins to cells (the bin2cell approach) and sum counts per cell.
- Filter cells with too few counts, and keep bins outside cells for background estimation.
Step 5: Cell identity
- Deconvolution of bins (RCTD, cell2location) estimates cell-type mixtures in each 8 µm bin.
- Label transfer to segmented cells (RCTD single-cell mode, scVI-based mapping, Tangram) assigns identities with confidence scores.
- Reference quality matters more than the algorithm. If the reference lacks a cell type, its transcripts are forced onto others. We use matched scRNA-seq or curated atlases and report signal that the reference cannot explain.
- A biologist checks annotations against morphology. Hepatocytes should look like hepatocytes on the H&E.
Step 6: Tissue architecture
- Spatial domains: BANKSY, SpaGCN, GraphST or CellCharter group bins or cells into coherent regions using both expression and neighborhood.
- Spatially variable genes: Moran's I, SpatialDE or nnSVG, with multiple-testing correction and filtering of domain-marker redundancy.
- Neighborhoods and niches: neighborhood enrichment, co-occurrence and niche clustering by local cell-type composition (Squidpy).
- Gradients: expression as a function of distance to annotated structures such as tumor borders, vessels or epithelium.
Step 7: Compare samples and conditions
Treat each tissue section from an independent donor as the replicate. Aggregate by domain, niche or cell type per sample (pseudobulk), then test between conditions with DESeq2 or limma. Several sections from one block are technical replicates.
Step 8: Spatial signaling
Ligand-receptor inference is far more meaningful in space, because you can restrict interactions to neighboring cells. We use spatially aware methods (COMMOT, LIANA+ spatial metrics, NicheNet with niche-defined senders and receivers), and compare signaling between conditions rather than listing every pair.
Deliverables
A complete Visium HD analysis should include masked QC maps, 8 µm and cell-level objects (SpatialData/AnnData or Seurat), annotation overlays on the H&E at publication resolution, domain and niche maps, statistical comparisons, code and methods. That is our standard Visium HD analysis service. Not sure Visium HD is the right platform? Read our platform comparison.
Frequently asked questions
What bin size should I use for Visium HD?
Use 8 µm bins for tissue domains, spatial gradients and robust multi-sample comparisons. Use segmented cells, created by assigning 2 µm bins to nuclei segmented on H&E or IF images, for cell-type neighborhoods and cell-level statistics. 16 µm bins are useful for quick exploration and low-expression genes. Raw 2 µm bins are too sparse to cluster directly.
What is bin2cell?
bin2cell is an open approach that segments nuclei on the H&E image (and optionally on gene expression), then aggregates 2 µm Visium HD bins into cells, with expansion around nuclei. It produces cell-level expression profiles for annotation and neighborhood analysis.
Work with OmicsQuest. Related services:
References
- Cable DM, Murray E, Zou LS, et al. Robust decomposition of cell type mixtures in spatial transcriptomics. Nat Biotechnol. 2022;40:517-526.
- Kleshchevnikov V, Shmatko A, Dann E, et al. Cell2location maps fine-grained cell types in spatial transcriptomics. Nat Biotechnol. 2022;40:661-671.
- Biancalani T, Scalia G, Buffoni L, et al. Deep learning and alignment of spatially resolved single-cell transcriptomes with Tangram. Nat Methods. 2021;18:1352-1362.
- Palla G, Spitzer H, Klein M, et al. Squidpy: a scalable framework for spatial omics analysis. Nat Methods. 2022;19:171-178.
- Svensson V, Teichmann SA, Stegle O. SpatialDE: identification of spatially variable genes. Nat Methods. 2018;15:343-346.
- Stringer C, Wang T, Michaelos M, Pachitariu M. Cellpose: a generalist algorithm for cellular segmentation. Nat Methods. 2021;18:100-106.
- Moses L, Pachter L. Museum of spatial transcriptomics. Nat Methods. 2022;19:534-546.