When the map's pixel size flips the biology

In Short. On a Visium HD map of mouse intestine, Paneth and Goblet cells look weakly segregated at 8 µm bins (r = −0.12) and strongly co-localized at 64 µm (r = +0.80). That sign flip is geometry: the resolution you choose is part of the biological claim.

---

Tissue RNA maps promise a picture of who sits next to whom. A slide is cut into a grid of capture bins. Visium HD’s native bins are 2 µm on a side; at that grain each bin holds too little RNA for a stable cell-type call (median 17 UMIs in this intestine dataset), so the analysis that follows starts at 8 µm aggregations and steps up through 16, 32, 64, and 128 µm. Each bin yields a gene-count inventory. Software then estimates which cell types could have produced that inventory. The neighborhood story, the tumor-edge story, the niche story all ride on those estimates.

The catch is that the grid is a dial. Yang, Chen, and Zhang show, on Visium HD data from mouse small intestine, that turning that dial can reverse what the map says about cell-cell relationships (Fig. 6f). Paneth cells live at crypt bases. Goblet cells spread along the crypt-villus axis. At 8 µm bins, their proportions across 351,817 bins correlate at r = −0.12: weak mutual exclusion at cellular scale. At 64 µm, the same pair peaks at r = +0.80. Adjacent but distinct niches collapse into the same measurement unit and look like partners.

Researchers reading conventional Visium-scale bins near 55 µm would see the strong positive number and could invent a shared microenvironment that the fine map never supported. The paper names the older spatial-statistics label for this failure: the modifiable areal unit problem. Here it is a tissue-scale instance with a qualitative consequence, a sign change, not a soft blur.

A horizon, not a preference

The same intestine series marks where purity collapses. At 8 µm, 61.5% of bins are dominated by a single cell type (proportion above 80%). At 16 µm that share falls to 13.3%. Normalized Shannon entropy of mixing rises from 0.23 to 0.48 across that step (Fig. 6c). The authors call the inflection a resolution horizon and note that it sits near the characteristic diameter of intestinal epithelial cells, about 10-15 µm. Below the horizon, bins mostly hold one cell’s worth of identity. Above it, averages erase that identity.

Brain cortex, run the same way, does not show an intestinal-style cliff. Purity declines gradually, and twelve of twenty-eight lineage pairs still flip correlation sign, including astrocyte-microglia (r = −0.12 → +0.51). The horizon moves with tissue architecture. It is a checkable property of assay plus tissue, not a universal software default.

Checking the mixing model

The mixing software is an obvious suspect. The authors check that path with Xenium in situ data from mouse colon, where cell positions are known without deconvolution. The ground-truth aggregates themselves flip: negative correlation at 8 µm (r = −0.06) and strong positive correlation at 128 µm (r = +0.68). Geometry writes the reversal before any deconvolution step runs.

Crypt-villus boundary sharpness tells the same story in anatomical language. Paths from Paneth-rich crypt cores to enterocyte-rich villi lose 77% of their mean enterocyte gradient sharpness between 8 µm and 16 µm. At 32 µm and coarser, crypt cores meeting their criteria disappear. The information loss tracks anatomy blurring into the bin.

Why a million-bin sweep is now cheap

FlashDeconv estimates each bin’s cell-type mix by matching its RNA counts to a single-cell reference while sampling genes so rare markers stay in the compressed math, which is why million-bin Visium HD maps finish on ordinary hardware fast enough to repeat the analysis at many bin sizes.

Speed is the enabler, not the lesson. On the intestine series the tool processed 351,817 bins at 8 µm in 12 seconds. That throughput makes a five-resolution sweep (8, 16, 32, 64, 128 µm) cheap enough to treat resolution as an experiment rather than a one-shot setting. Once you can afford that sweep, the horizon and the sign flips become visible.

At 8 µm the same maps recover Tuft-cell niches with 16.2-fold enrichment for intestinal stem cells, a chemosensory neighborhood that washes out as bins coarsen (maximum Tuft proportion falls from 61% at 8 µm to 4% at 128 µm). In a human colorectal Visium HD cohort of about 1.6 million bins, neutrophil inflammatory microdomains at the tumor-stroma interface are sharpest at 8 µm and decay toward 64 µm; discrete-label workflows that withhold mixed bins label only 2.3% of those hotspot bins as neutrophil singlets. Those findings are what the fine bins buy: niches that a coarser grid, or a discrete label that withholds mixed bins, will not report.

How to read the next neighborhood figure

Any figure that plots two cell types as neighbors on a capture-based spatial map inherits a bin geometry. Ask for that geometry in micrometers. Ask whether the co-localization survives a resolution sweep across the cell-size scale of that tissue. Ask whether an imaging ground truth without deconvolution shows the same sign. If the answer is missing, the neighborhood may be a property of the grid.

Visium HD and Xenium appear here as instruments under study. The commercial names are coordinates on a measurement trade. Coarser maps still answer some questions. They answer different ones. The honest practice is to say which resolution produced the claim, and whether the claim’s sign holds when the bins shrink toward the cells.

Yang C., Chen J., Zhang X. FlashDeconv reveals resolution horizons in atlas-scale spatial transcriptomics. bioRxiv (v5, 2026-09-11). Preprint, not peer-reviewed.