Visium HD: when 2 µm bins were analyzed as cells

The problem

The team expected Visium HD to show where immunosuppressive macrophages met tumor cells at the invasive margin. Their first analysis clustered 2 µm bins directly. It produced 38 clusters that tracked UMI count and tissue edges rather than biology, and deconvolution results conflicted with pathology.

What we found

• Median UMIs per 2 µm bin were in the single digits. Clusters were dominated by sparsity and depth, not cell identity.
• Folds and edge artifacts in two sections formed their own clusters.
• The deconvolution reference lacked cancer-associated fibroblast subtypes, so fibroblast signal was forced onto other cell types.

How we rescued it

1. Masked folds and off-tissue regions and re-ran Space Ranger with correct image registration.
2. Segmented nuclei on the H&E with StarDist and aggregated 2 µm bins into cells with tissue-tuned expansion.
3. Built a matched reference from a public colorectal cancer scRNA-seq atlas with complete fibroblast and myeloid subtypes, and annotated cells with RCTD.
4. Defined niches with neighborhood composition and identified a SPP1-positive macrophage and FAP-positive fibroblast niche at the invasive front, present in 5 of 6 tumors.
5. Ran spatially restricted ligand-receptor analysis and pseudobulk comparisons by niche across tumors.

The result

The niche signature matched pathologist annotations of the invasive margin and became the basis for a multiplex IHC validation panel. The team now uses our pipeline as their standard for Visium HD.