Where this lecture fits

• Previous: Lec 13 — WGCNA
• You are here: Lec 16 — measuring gene expression while keeping the tissue context
• Next: Lec 14 — Trajectory Inference & Cell–Cell Communication
• Companion tutorial: Tutorial 16 — Spatial Transcriptomics with Visium
• Companion book chapter: Chapter 16 — Spatial Transcriptomics
• Parallel reading: scNotebooks Module 10

Goals of this lecture

• Understand what spatial transcriptomics measures that scRNA-seq cannot
• Compare the major technologies — Visium, Visium HD, Xenium, MERFISH, CosMx — by resolution / throughput / coverage
• Read a Visium experiment as a spot-level dataset (≠ single-cell)
• Apply the standard analysis pipeline: QC → SCTransform → cluster → spatially variable features → integration
• Know when to use deconvolution (RCTD, CARD, cell2location) vs spatial clustering (BANKSY, STAGATE)
• Recognize the limits — bin size vs cell size, edge artifacts, depth heterogeneity

The dissociation lie

scRNA-seq starts with dissociation: tissue → cell suspension. That gives you each cell’s transcriptome but loses everywhere it ever was in the tissue. For developmental biology, immunology, and disease microenvironments, position is the answer to half your questions.

• Where in the tumor are the exhausted T cells? Adjacent to which tumor sub-clones?
• In the developing kidney, which transcriptional state precedes which?
• Are immune cells in the ductal periphery or the lobular core? Granular / scattered or concentrated in foci?
• Which genes are expressed only at tissue boundaries, where dissociation would shred the spatial signal?

Spatial transcriptomics at a glance

Spatial preview — what gets measured, where, and how

• Counts plus coordinates — every observation has a (x, y) location in the tissue
• Different platforms make different trade-offs between resolution, gene panel size, and field-of-view
• The downstream toolchain reuses scRNA-seq machinery but adds spatial-aware steps (Moran’s I, deconvolution, region detection)

What spatial measures

The left panel shows Hpca (a hippocampal marker) measured on a Visium tissue section — the gene’s expression traces the hippocampus. The right panel shows the same cells re-shuffled (as scRNA-seq would after dissociation) — same expression, no anatomy.

The landscape

Visium 101 — the workhorse

The key constraint: a Visium spot is not a single cell. Your “marker gene” expression at one spot is a mixture of whatever cells are within it. This shapes everything downstream — clustering recovers tissue regions (not cell types), DE recovers region-vs-region differences, and “deconvolution” is its own analytical step.

Visium HD — sub-cellular bins

In 2023 10x released Visium HD with 2 µm × 2 µm capture squares (much smaller than a cell). Output bins at 8 µm (~1 cell) or 16 µm (multi-cell) provide the choice between resolution and depth. Same chemistry as classic Visium but ~1000× more bins per section.

For atlas-scale work in 2026, Visium HD has largely replaced classic Visium. The analysis pipeline is the same; the file sizes are bigger.

Spot-level QC — three metrics, plus a fourth

Per-spot QC metrics generalize from scRNA-seq (nCount_Spatial, nFeature_Spatial, percent.mt) but you should always plot them on the tissue — spatial QC is the fourth metric:

• A localized region of low UMI counts? White matter, edge artifacts, bubbles. Don’t filter blindly.
• A localized region of high percent.mt? Could be necrotic tissue (real biology) or storage damage (technical).
• Correlated drops in both? The spot is on tissue but in poor condition.

Tutorial dataset — stxBrain (matched mouse brain sections)

Why SCTransform on Visium

Visium spots have a wider dynamic range than single cells (each spot is 1–10 cells). LogNormalize works but SCTransform is preferred — its regularized negative binomial models the depth-vs-variance relationship more cleanly at the spot level.

library(Seurat)
brain <- LoadData("stxBrain", type = "anterior1")
brain <- SCTransform(brain, assay = "Spatial", verbose = FALSE)

Clustering — but the clusters are regions, not cell types

Spatially variable features — Moran’s I

Some genes vary spot-to-spot with spatial structure (clear pattern on the tissue); some vary spot-to-spot but spatially randomly. The first kind is what you want.

Moran’s I is the standard test:

\[I = \frac{N}{W} \cdot \frac{\sum_i \sum_j w_{ij} (x_i - \bar x)(x_j - \bar x)}{\sum_i (x_i - \bar x)^2}\]

where \(w_{ij}\) is the spatial weight (large for spatially close spots, zero for far). High Moran’s I means high spatial autocorrelation — adjacent spots have similar expression.

The mixture problem

Each Visium spot ≈ 1–10 cells. If a spot’s cluster is “putative neuron-rich region”, which kinds of neurons? In what proportions?

Deconvolution decomposes each spot’s mixture using a single-cell reference of the same tissue:

Tools

Beyond standard SNN+Leiden

Standard clustering treats spots as iid points in expression space — but adjacent spots are spatially correlated. A spatial-aware clustering uses the tissue layout as an additional signal:

• BANKSY — augments each spot’s expression with its neighbors’ average expression before clustering
• STAGATE — graph attention auto-encoder on the spatial graph; learns smoother spatial domains
• SpaGCN — graph convolutional network combining expression + histology image
• GraphST — contrastive self-supervised graph learning

These typically yield smoother, more biologically interpretable spatial domains than standard Seurat clustering. Worth using for any “what are the regions?” question.

Six things that bite

1. A spot is not a cell. Any “marker gene” plot you draw is a mixture readout. For cell-type questions, deconvolve.
2. Edge effects. The outermost spots only partially cover tissue and have lower UMI counts. Most pipelines treat them like the rest. Either filter (SpatialFeaturePlot(... pt.size.factor = 1.0, alpha = c(0.0, 1))) or annotate them.
3. Cross-section integration. Two sections from the same tissue still need integration — chip-to-chip variation is real. Same Harmony / SCT-anchor methods as scRNA-seq.
4. Reference mismatch in deconvolution. A reference that’s missing your tissue’s main cell type will distribute that cell’s transcripts across the cell types it does have. Always use a same-tissue reference.
5. Over-interpreting the histology image. The H&E underneath is a real image but is registered, not a ground truth — slight misalignment between spots and image is expected.
6. Visium HD data volume. A single Visium HD section is ~10 GB; loading naively into Seurat will OOM. Use the bin-aggregated outputs (8 µm, 16 µm) unless you really need 2 µm.

Reading spatial outputs — what to look for

• SpatialFeaturePlot: color on the H&E image should trace anatomy — Hpca in hippocampus, Plp1 in white matter; a uniform smear signals ambient RNA or failed normalization
• SpatialDimPlot: clusters that form contiguous anatomical regions are real; clusters that scatter randomly across the tissue are probably technical artifacts
• Moran’s I top genes: expect distinct patches or stripes matching known structures; a gene with high variance but low Moran’s I is spotty, not spatial
• Integration UMAP (slice coloring): spots from both sections should interleave after integration — remaining separation by slice means the technical batch was not corrected

What to remember from Lecture 16

• Spatial transcriptomics measures gene expression while keeping the position — answers “where” questions scRNA-seq throws away
• Visium is the workhorse (whole-tx, low resolution, easy); Xenium / MERFISH / CosMx trade coverage for resolution
• Visium spots are mixtures (1–10 cells) — clusters recover tissue regions, deconvolution recovers cell-type fractions
• SCTransform is the standard normalization; Moran’s I finds spatially variable genes
• Spatial-aware clustering (BANKSY, STAGATE) gives cleaner regions than vanilla SNN+Leiden
• Always plot QC metrics on the tissue — spatial QC is itself a sanity check

Coming up next

• Lec 14 — Trajectory Inference & Cell–Cell Communication
• Hands-on: Tutorial 16 — Spatial Transcriptomics

Further reading

• Moses & Pachter 2022, Nat Methods — comprehensive spatial-methods landscape, doi:10.1038/s41592-022-01409-2
• Cable et al. 2022 — RCTD / spacexr, doi:10.1038/s41587-021-00830-w
• Kleshchevnikov et al. 2022 — cell2location, doi:10.1038/s41587-021-01139-4
• Singhal et al. 2024 — BANKSY, doi:10.1038/s41588-024-01664-3
• scNotebooks Module 10 — Spatial Transcriptomics