Multiscale harmonization, semantic integration, and immersive exploration of single-cell data in support of novel biological insights

Key results

1. ShownEight scenes, each from a different Data Provider, are reached through a Multiscale Elevator System. Each floor shrinks the user by one order of magnitude, from a whole-body scene down to 100-micron cell groups, and each scene marks where its dataset came from in the body using its HRA registration.

   Levels: 10^0 (1 meter), 10^-2 (1 centimeter), 10^-3 (1 millimeter), 10^-4 (100 microns). Organs: lymph node, brain, large and small intestine, liver. Assays: CODEX, Visium, volumetric cyclic immunofluorescence (v-CyCIF), Xenium, serial block-face scanning electron microscopy (SBF-SEM). Start scene: 81 HRA v2.5 organs, 2,762 tissue blocks, 13,344 datasets, 1,021 donors, 54 providers. · Introduction; Insights from VR; hra-whole_body-hra-1_meter-10_0; Extraction Site Component

2. ShownMaking the data run on standalone headsets took mesh decimation and data subsampling. The large intestine v-CyCIF scene was cut from about 6 million to about 3 million polygons. It still ran without significant frame-rate drops, although it stayed above Meta's polygon budget.

   Meta recommends 400-600 draw calls and 1.3-1.8 million polygons in view (Quest 3/3S). A decimation factor of 0.05 kept ~5% of each confocal mesh. Xenium polyp: model cut ~41% from 1.6 million to 942,436 polygons. All 29 layers would total 2.32 million cells, so only layer 16 (18,285 cells, 20,919 RNA molecules) is shown, rendered within 300 (cells) and 100 microns (RNA) of the user. · Mesh optimization and performance considerations; Performance; Visualization; Performance metrics for Meta Quest VR headsets

3. SuggestedIn a light-sheet v-CyCIF crop of human large intestine, most of ~200 CD8+ conventional T cells appear to touch TUBB3+ enteric nerves directly, and some seem to migrate toward the lumen. In VR, users can pull cells apart and color them by properties such as volume or sphericity.

   This is a qualitative observation from one crop, with no counts, distances or statistics. A Blender Python script matched each 3D cell to its quantification-table row: it first aligned on a shared reference cell, then required position within 3.0 scene units and volume within 20%. The Data Provider validated the mapping. · hubmap-large_intestine-wong-100_microns-10_4; hubmap-large_intestine-wong-100_microns-10_4

4. SuggestedA CODEX small-intestine region is shown as 4 stacked layers (cell type, neighborhood, community, tissue unit), linked by a selection cylinder and an alluvial diagram. The tissue-unit layer matches known intestinal histology from mucosa to muscularis externa.

   Dataset HBM573.SWGH.988: 3,252 cells shown at 4 levels. Above the cell-type level, cells are grouped by neighbor composition over increasingly broad ranges. The selection cylinder counts selected cells per layer, and the alluvial diagram links the layers. CD4+ and CD8+ T cells in distinct multicellular settings are described without quantification. · hubmap-small_intestine-miao-100_microns-10_4; hubmap-small_intestine-miao-100_microns-10_4

5. ShownA 3D spike plot shows 8 senescence hallmark scores as lines above or below Visium spots from the dorsolateral prefrontal cortex of a 74-year-old female donor. It is presented as a way to generate hypotheses about senescence microenvironments.

   4,992 spots (55-micron diameter, 100-micron spacing) carry 39,936 hallmark values. A mean/variance filter reduces the plot to 1,576 lines for 197 spots. A gallery of 10 plotly 2D plots gives scatter graphs that each show one hallmark, plus 2 auxiliary graphs showing which spots get lines and where they sit on the Visium grid. The paper reports no specific spatial pattern. · sennet-brain-phatnani-1_millimeter-10_3; sennet-brain-phatnani-1_millimeter-10_3; 3D spike plot; 2D plots