GPTomics/bio-spatial-transcriptomics-spatial-data-io

Version Compatibility

Reference examples tested with: spatialdata 0.2+, spatialdata-io 0.1.5+, squidpy 1.4+, scanpy 1.10+, anndata 0.10+

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Spatial Data I/O

"Load my spatial data" -> Parse a platform's output bundle into one coordinate frame holding the expression matrix, coordinates, images, and (for imaging) the molecule table and segmentation shapes.

• Imaging/in-situ (Xenium, MERSCOPE/MERFISH, CosMx, seqFISH): spatialdata_io.{xenium, merscope, cosmx} -> SpatialData with a per-transcript points table AND a derived per-cell tables matrix.
• Sequencing/capture (Visium, Visium HD, Slide-seq/Curio, Stereo-seq): squidpy.read.visium or spatialdata_io.{visium, visium_hd, curio, stereoseq} -> spot/bin matrix + coordinates; NO molecule table.

Governing Principle

The single most consequential I/O fact is that the two platform classes emit different primary objects, and one class emits two of them that are easy to confuse.

Imaging/in-situ platforms emit TWO physically distinct primary objects. The first is a per-TRANSCRIPT molecule table -- one row per decoded molecule with x, y (and often z), gene, a decoding-quality value, and a cell-assignment-or-unassigned. The second is a per-CELL expression matrix, genes-by-cells, DERIVED by overlaying a segmentation and counting the molecules that fall inside each boundary. The matrix looks exactly like scRNA-seq and is therefore wrongly trusted as ground truth, but it is a downstream product: it inherits every segmentation error and is usually quality-filtered (Xenium keeps Q>=20 in the matrix while the transcript table keeps everything). The molecule table is the source of truth and the ONLY object that lets the analyst re-segment, recover unassigned molecules, or do subcellular work. A loader that returns only the cell matrix has silently discarded the re-segmentable layer.

Sequencing/capture platforms (Visium, Slide-seq, Stereo-seq) have NO molecule table -- a spot/bead/bin is mini-bulk over the cells beneath it, captured as a single barcoded profile. Do not go looking for a transcript table that does not exist; the only objects are the barcode-by-gene matrix, the coordinates, and the tissue image.

A second trap is coordinate frames. Images, spot/cell coordinates, and molecule points each live in their own intrinsic pixel or array axes; overlaying transcripts on histology, or building a neighbor graph with a micron radius, requires the right transform (Visium scalefactors; imaging micron-to-pixel matrices). SpatialData makes the frames explicit (intrinsic vs a shared extrinsic "global" system); the legacy AnnData layout hides them in uns['spatial'][library_id]['scalefactors']. Mixing frames silently places points off the image or builds a graph at the wrong scale.

The Platform-Class Fork

The first question of any spatial dataset is which side of the fork it sits on, because it decides what objects exist and what the rest of the pipeline must do.

| Class | Platforms | Primary objects | Molecule table? | Cell unit | Downstream | |---|---|---|---|---|---| | Imaging / in-situ | Xenium, MERSCOPE/MERFISH, CosMx, seqFISH | molecule table + segmentation-derived cell matrix + images + shapes | YES (source of truth) | from segmentation (a hypothesis) | segment, then label-transfer typing | | Sequencing / capture | Visium, Visium HD, Slide-seq/Curio, Stereo-seq, GeoMx | barcode/bin-by-gene matrix + coordinates + image | NO | spot/bin = 1-10-cell MIXTURE (GeoMx ROI = many-cell region; sub-cell bins = fraction of a cell) | deconvolution (or bin/segment-up for sub-cell bins) |

Object-Model Landscape

Each toolkit is a strategy for co-storing four things in one frame: an expression matrix, geometry (spot circles, cell/nucleus polygons, centroids), raster images (H&E, DAPI, multiplex IF -- often gigapixel), and (imaging only) the molecule point cloud. They differ in how separately they keep these and which is the forward path.

| Framework (language) | Core object | Molecule table | Per-cell matrix | Segmentation geometry | Images | Best when | |---|---|---|---|---|---|---| | SpatialData / scverse (Python) | SpatialData of elements | points (dask -> Parquet) | tables (AnnData) | labels (masks) + shapes (geopandas polygons) | images (xarray, OME-NGFF/Zarr, lazy) | Multimodal, larger-than-memory, multiple platforms in one store; re-segmentation | | Squidpy + AnnData (Python) | AnnData | via SpatialData/readers | adata.X | in obs/external | uns['spatial'] (legacy) | Standard AnnData spatial graph stats on spot or cell matrices | | Seurat v5 (R) | Seurat, geometry in @images | FOV@molecules | Assay5 counts | FOV@boundaries (Centroids + Segmentation) | platform SpatialImage | R single-cell users; v5 integration | | SpatialExperiment / SFE-Voyager (Bioc, R) | SpatialExperiment / SpatialFeatureExperiment | rowGeometries (sf points, SFE only) | SCE assay | colGeometries (cellSeg/nucSeg/spotPoly) | imgData() | Bioconductor scran/scater; geospatial ESDA (Moran's I) | | Giotto Suite (R) | giotto (multi-scale) | subcellular molecule layer | aggregated cell layer | polygon/cell layers | image layers | One technology-agnostic object spanning molecule -> cell -> region |

SpatialData is the forward-path Python standard because it is the only framework that natively keeps the molecule table, multiscale OME-NGFF images, and segmentation shapes as first-class, frame-aware elements. Because a SpatialData table IS an AnnData, all of squidpy.gr.* runs on it unchanged.

Platform-to-Reader Map

| Platform | spatialdata-io reader | squidpy.read | Key I/O fact | |---|---|---|---| | Visium | visium | visium | spot, no molecule table; tissue_positions.csv gained a header at Space Ranger v2.0 (readers handle both) | | Visium HD | visium_hd | -- | bins (2/8/16um); tissue_positions.parquet (PARQUET, not CSV); 8um bin still spans ~2 cells | | Xenium | xenium | -- (none) | molecule table transcripts.parquet (all Q) + Q>=20 cell matrix; needs experiment.xenium manifest | | MERSCOPE / MERFISH | merscope | vizgen | there is NO merfish reader -- merscope handles both; boundaries went hdf5-folder -> single cell_boundaries.parquet at instrument SW v232 | | CosMx | cosmx | nanostring | flat CSVs with a run/slide prefix; tx_file (molecule table) absent for protein-only panels | | Slide-seq / Curio | curio | -- (none) | bead, no molecule table | | Stereo-seq | stereoseq | -- | DNB sub-cellular; binned up to cells |

squidpy.read provides ONLY visium, vizgen, and nanostring -- it has no xenium or slideseq reader. For Xenium, Slide-seq/Curio, Stereo-seq, and Visium HD, use the spatialdata_io reader. scanpy.read_visium is deprecated as of scanpy 1.11 -- prefer squidpy.read.visium (identical obsm/uns layout) or spatialdata_io.visium.

Load Visium (Spot/Capture)

Goal: Read a Space Ranger bundle into an AnnData with coordinates, image, and scalefactors, without reaching for the deprecated scanpy reader.

Approach: Use squidpy.read.visium; coordinates land in obsm['spatial'] (pixels of the full-res image), image and scalefactors in uns['spatial'][library_id].

import squidpy as sq

adata = sq.read.visium('spaceranger_out/')         # filtered matrix + spatial/; NOT a cell -- each spot is a 1-10-cell mixture
library_id = list(adata.uns['spatial'].keys())[0]
scalef = adata.uns['spatial'][library_id]['scalefactors']
# obsm['spatial'] is in FULL-RES pixels; multiply by tissue_hires_scalef to index the hires image
print(adata.n_obs, 'spots', adata.n_vars, 'genes', '| spot diameter (px):', scalef['spot_diameter_fullres'])

Load Imaging Data and Keep the Molecule Table

Goal: Load Xenium (or MERSCOPE/CosMx) so BOTH the per-transcript molecule table and the derived cell matrix are available, not just the matrix.

Approach: Use the spatialdata_io reader, which returns a SpatialData object; the molecule table lives in sdata.points, the cell matrix in sdata.tables, segmentation polygons in sdata.shapes, images in sdata.images. Inspect element names with print(sdata) -- they vary by platform and reader version.

import spatialdata_io as sdio

sdata = sdio.xenium('xenium_out/')                 # needs experiment.xenium manifest
print(sdata)                                       # lists points/tables/shapes/images element names

transcripts = sdata.points['transcripts']          # dask DataFrame: x, y, z, feature_name, qv, cell_id -- ALL Q-scores
adata = sdata.tables['table']                      # AnnData cell matrix -- Q>=20 filtered, inherits segmentation error
# the matrix is a DERIVED product; the molecule table is the re-segmentable source of truth
print('molecules:', transcripts.shape[0].compute(), '| cells in matrix:', adata.n_obs)

For MERSCOPE substitute sdio.merscope('merscope_out/'); for CosMx sdio.cosmx('cosmx_out/'). As an AnnData-only alternative for MERSCOPE, sq.read.vizgen(path, counts_file='cell_by_gene.csv', meta_file='cell_metadata.csv') returns the cell matrix but discards the molecule table.

Load Other Capture Platforms

Goal: Load Visium HD bins, Slide-seq/Curio beads, or Stereo-seq into a SpatialData object.

Approach: Use the matching spatialdata_io reader; none of these has a molecule table, and Visium HD / Stereo-seq bins are smaller than a cell (the inverse-of-deconvolution regime -- see spatial-deconvolution).

import spatialdata_io as sdio

sdata_hd = sdio.visium_hd('visium_hd_out/')        # tissue_positions are PARQUET; pick a bin (8um default still ~2 cells)
sdata_ss = sdio.stereoseq('stereoseq_out/')        # DNB sub-cellular spots, binned up to cells
sdata_bead = sdio.curio('slideseq_out/')           # Slide-seq/Curio beads; ~1 cell but ~1/3 carry >=2 types

Inspect and Register Coordinate Frames

Goal: Confirm whether coordinates are in pixels or microns before building a graph or overlaying on histology, so a neighbor radius or a plotted point lands at the right scale.

Approach: In SpatialData read the element transformations (intrinsic vs the shared "global" extrinsic frame); in the AnnData layout read the scalefactors. Never assume obsm['spatial'] units -- Visium is full-res pixels, most imaging readers place a micron "global" frame.

from spatialdata.transformations import get_transformation

# SpatialData: every element carries transforms into shared coordinate systems
print(sdata.coordinate_systems)                    # e.g. ['global']
print(get_transformation(sdata['transcripts'], get_all=True))   # intrinsic -> global (often micron scaling)

Convert SpatialData to AnnData

Goal: Extract the cell/spot matrix as a plain AnnData for tools that expect one, while keeping coordinates.

Approach: Copy the table, set obsm['spatial'] from the matching shapes/centroids. Persist via sdata.write(...) ONLY to a scratch path -- a .zarr store is a DIRECTORY, not a file, so it must never be committed.

adata = sdata.tables['table'].copy()
# write a zarr STORE (a directory) to scratch, never the repo; rm -rf when done
# sdata.write('/tmp/scratch/store.zarr')

Common Errors

| Symptom | Cause | Fix | |---|---|---| | AttributeError: module 'squidpy.read' has no attribute 'xenium' (or slideseq) | squidpy.read only has visium, vizgen, nanostring | Use spatialdata_io.xenium / spatialdata_io.curio for those platforms | | spatialdata_io has no merfish reader | The reader is named for the instrument, not the chemistry | Use spatialdata_io.merscope (handles MERFISH and MERSCOPE) | | DeprecationWarning / future removal on scanpy.read_visium | Deprecated as of scanpy 1.11 | Use squidpy.read.visium or spatialdata_io.visium (same layout) | | Cell matrix has far fewer transcripts than the molecule table | Imaging cell matrix is Q>=20 filtered and segmentation-derived | Treat the matrix as provisional; use sdata.points (all Q) to re-segment or audit | | Trusting the cell matrix as ground truth; weird co-expression | The matrix inherits all segmentation/spillover error | Validate against the molecule table; re-segment (see image-analysis) | | Looking for a transcript table in Visium/Slide-seq and finding none | Capture platforms have no molecule table | Stop -- a spot is mini-bulk; there is nothing to re-segment | | Visium HD reader fails reading tissue_positions.csv | Visium HD positions are PARQUET (tissue_positions.parquet) | Use spatialdata_io.visium_hd, which expects the parquet bundle | | Older Visium positions parse with a shifted header | tissue_positions.csv gained a header at Space Ranger v2.0 | Current readers handle both; upgrade spatialdata-io/squidpy if parsing legacy files | | MERSCOPE boundaries not found | hdf5-folder boundaries became single cell_boundaries.parquet at SW v232 | Match reader version to instrument software; point at the parquet if present | | Seurat @coordinates slot missing (R interop) | Seurat 5.1 VisiumV2 has no coordinates slot (V1 did) | Use GetTissueCoordinates(); there is no V1->V2 converter | | Transcripts plot off the image | Points and image in different frames/units (pixel vs micron) | Apply the reader's transform / scalefactor before overlaying | | A stray .zarr directory left in the repo after writing | sdata.write makes a directory store; git status hides untracked dirs | Write to scratch; rm -rf the store; run git status --porcelain | grep '/$' |

Related Skills

• spatial-preprocessing - QC floors and normalization that differ by platform class after loading
• image-analysis - re-segment the molecule table; the cell matrix is a segmentation hypothesis
• spatial-deconvolution - recover cell-type proportions from spot mixtures that have no molecule table
• high-resolution-binning - bin/segment-up sub-cellular Visium HD and Stereo-seq captures
• spatial-visualization - plot spots vs imaging FOVs with the correct coordinate frame
• single-cell/data-io - non-spatial scRNA-seq loading for the deconvolution/label-transfer reference

References

• Marconato L, Palla G, Yamauchi KA, et al. (2025) SpatialData: an open and universal data framework for spatial omics. Nature Methods 22(1):58-62. DOI 10.1038/s41592-024-02212-x
• Palla G, Spitzer H, Klein M, et al. (2022) Squidpy: a scalable framework for spatial omics analysis. Nature Methods 19(2):171-178. DOI 10.1038/s41592-021-01358-2
• Wolf FA, Angerer P, Theis FJ (2018) SCANPY: large-scale single-cell gene expression data analysis. Genome Biology 19:15. DOI 10.1186/s13059-017-1382-0
• Virshup I, Rybakov S, Theis FJ, Angerer P, Wolf FA (2024) anndata: Access and store annotated data matrices. Journal of Open Source Software 9(101):4371. DOI 10.21105/joss.04371
• Hao Y, Stuart T, Kowalski MH, et al. (2024) Dictionary learning for integrative, multimodal and scalable single-cell analysis. Nature Biotechnology 42(2):293-304. DOI 10.1038/s41587-023-01767-y
• Righelli D, Weber LM, Crowell HL, et al. (2022) SpatialExperiment: infrastructure for spatially-resolved transcriptomics data in R using Bioconductor. Bioinformatics 38(11):3128-3131. DOI 10.1093/bioinformatics/btac299
• Moore J, Allan C, Besson S, et al. (2021) OME-NGFF: a next-generation file format for expanding bioimaging data-access strategies. Nature Methods 18:1496-1498. DOI 10.1038/s41592-021-01326-w
• Janesick A, Shelansky R, Gottscho AD, et al. (2023) High resolution mapping of the tumor microenvironment using integrated single-cell, spatial and in situ analysis (Xenium). Nature Communications 14:8353. DOI 10.1038/s41467-023-43458-x