henhao
Sign in
BioinformaticsNature Methods
Peer-reviewed journal article

SpatialData: an open and universal data framework for spatial omics

Luca MarconatoiD, Giovanni PallaiD, Kevin A. YamauchiiD, Isaac VirshupiD, Elyas HeidariiD, Tim TreisiD, Wouter-Michiel VierdagiD, Marcella TothiD, Sonja StockhausiD, Rahul Shrestha, Benjamin RombautiD, Lotte PollarisiD, Laurens LehneriD, Harald S. VöhringeriD, Ilia KatsiD, Yvan SaeysiD, Sinem K. SakaiD, Wolfgang K. HuberiD, Moritz GerstungiD, Josh MooreiD, Fabian Joachim TheisiD, Oliver StegleiD

Abstract

Spatially resolved omics technologies are transforming our understanding of biological tissues. However, handling uni- and multi-modal spatial omics datasets remains a challenge owing to large volumes of data, heterogeneous data types and the lack of unified spatially-aware data structures. Here, we introduce SpatialData, a framework that establishes a unified and extensible multi-platform file-format, lazy representation of larger-than-memory data, transformations, and alignment to common coordinate systems. SpatialData facilitates spatial annotations and cross-modal aggregation and analysis, the utility of which is illustrated via multiple vignettes, including integrative analysis on a multi-modal Xenium and Visium breast cancer study.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
231
FWCI
29.84
Field percentile
top 1%
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (8 May 2023) · CC BY 4.0

SpatialData, a Zarr/OME-NGFF-based format plus Python library, unifies images, labels, points, shapes and tables across spatial omics platforms with coordinate alignment, lazy loading and cross-modal aggregation.

Spatial omics data mix very large images, transcript point clouds, polygons and count tables in vendor-specific formats, often at different resolutions and from different tissue regions. Existing tools lack joint support for large images, multimodal alignment, cross-modality aggregation and interactive annotation, and SpatialData offers one language-independent standard and API covering these four gaps.

Key results

  1. ShownTwo Xenium replicates and one Visium CytAssist section were aligned to a common coordinate system using landmarks and affine transforms. This allowed Xenium cell-type fractions to be transferred onto Visium capture spots, where they agreed with the other replicate and with deconvolution estimates.

    3 landmarks per image, with Xenium replicate 1 as the reference. Cell-type fractions at Visium locations: median Pearson's R=0.75 between Xenium replicates; median R=0.60 between Xenium-derived fractions and cell2location (v0.1.3) deconvolution of Visium restricted to the 313 Xenium-panel genes. · Application of SpatialData to a multi-modal and…; Alignment to create common coordinate systems; Cell type interpolation for Visium; Cell type deconvolution using cell2location

  2. ShownSumming Xenium single-molecule transcripts inside Visium spot geometries allowed gene-by-gene comparison across technologies. Agreement was higher between the two Xenium replicates than between Xenium and Visium, and it fell for lowly expressed genes.

    Across 313 shared genes: median Pearson's R=0.62 between aggregated Xenium replicates vs median R=0.48 between Xenium and Visium counts. Per-gene correlation rose with overall transcript abundance; aggregated FOXA1 and UCP1 levels at Visium spots are also visualized. · Application of SpatialData to a multi-modal and…; Transcript aggregations; Figure legends

  3. SuggestedThe same aggregation interface computed cell-type composition over four hand-drawn histological regions of interest (ROIs) and over Visium-derived tumor subclones. Compositions were consistent between the Xenium replicates and with Visium deconvolution.

    CopyKat (v1.1.0) followed by hierarchical clustering gave 3 major clusters among aneuploid spots, used as subclones. ROIs were drawn on the H&E image in napari-spatialdata. Agreement is shown only graphically; no correlation statistic is reported for the ROI or clone comparisons. · Application of SpatialData to a multi-modal and…; ROI selection with napari-spatialdata; Clone detection on Visium using CopyKat; ROI cell type fractions

  4. ShownThe storage model uses five element types (Images, Labels, Points, Shapes, Tables) saved as OME-NGFF-compliant Zarr, with coordinate transformations stored alongside them. Technology readers and a curated set of pre-converted example datasets are distributed for reuse.

    Pre-formatted SpatialData objects from over 40 datasets across 8 technologies (Table S3), downloadable as .zip or readable from public S3. Readers for vendor formats are listed in Table S2. Points are stored as Apache Parquet rather than as Zarr arrays. · SpatialData concept and implementation; Application to further datasets and illustration…; SpatialData framework; Figure legends

  5. SuggestedLandmark alignment was extended to a multi-slide prostate cancer study, and the coordinate framework supplied a PyTorch Dataset for training an image-based cell-type classifier. Both are infrastructure demonstrations rather than evaluated analyses.

    12 Visium slides mapped onto a large prostate section (the Supplementary Figure 9 legend shows 15 fields-of-view, about 580 megapixels each, viewed in napari-spatialdata). 32×32 pixel tiles around Xenium cells were fed to a MONAI DenseNet encoder trained for only a few epochs, with suboptimal accuracy. · SpatialData concept and implementation; Application to further datasets and illustration…; Figure legends

Methods, data and reproducibility

Defines an OME-NGFF/Zarr storage specification with five element types and stored coordinate transformations, plus a Python library (spatialdata-io readers, napari-spatialdata annotation, spatialdata-plot) offering lazy Dask loading, composable affine transforms, spatial queries and aggregation; demonstrated on a Xenium/Visium breast cancer study.

  • 10x Genomics Xenium in situ sequencing replicates from consecutive breast tumor sections: DAPI images, cell/nucleus segmentations, transcripts, count tables · human · n = two replicates; 313-gene panel
  • 10x Visium CytAssist section with H&E image from a consecutive section of the same breast tumor · human · n = one section
  • Publicly available breast cancer scRNA-seq atlas used as reference for ingest label transfer and cell2location deconvolution · n = 9 major cell types, 29 subtypes
  • Prostate cancer Visium and Spatial Transcriptomics datasets (Erikson et al.) aligned to whole-tissue layout images · n = 12 Visium slides; 15 fields-of-view shown for one tissue
  • NanoString CosMx dataset used to demonstrate lasso-based interactive cell annotation
  • MERFISH mouse brain dataset (Allen Institute prototype pipeline) with polygonal anatomical ROIs · mouse
  • Collection of pre-formatted SpatialData sample datasets spanning multiple technologies · n = over 40 datasets across 8 technologies
Methods & tools
spatialdata (core Python library), spatialdata-io (technology readers), napari-spatialdata (napari plugin), spatialdata-plot, OME-NGFF / OME-Zarr specification, Dask (lazy loading), Scanpy v1.9 ingest (label transfer), cell2location v0.1.3, CopyKat v1.1.0, Squidpy, PyTorch Dataset with MONAI DenseNet encoder, Apache Parquet (Points storage)
Code
https://github.com/scverse/spatialdata-notebooks/tree/main/notebooks/paper_reproducibility
Data & accessions
https://www.10xgenomics.com/products/xenium-in-situ/preview-dataset-human-breasthttps://spatialdata.scverse.org/en/latest/tutorials/notebooks/datasets/README.html
Software & environment
Python 3.9 and above; pip-installable, BSD 3-Clause licensed packages with unit tests and pre-commit CI; Scanpy v1.9, cell2location v0.1.3 (default tutorial parameters), CopyKat v1.1.0 run in R.
Assay conditions & compute
Exploring the aligned breast cancer data was reported to work on a standard low-memory laptop thanks to Dask lazy loading and Zarr multiscale storage; no runtime or memory benchmarks are given.
Reproducibility gaps
CopyKat analysis needed an .h5ad export to R; scRNA-seq atlas and prostate data cited without accessions; parts of the format (Points, Shapes, tables, transforms) are not yet ratified by the OME-NGFF community.
Limitations (5)
  • Coordinate systems, transforms and tables are still under OME-NGFF community review, and Points and Shapes have not yet been submitted. Points storage in Parquet may switch to a Zarr alternative. · stated by the authors
  • No direct R/Bioconductor interoperability yet: the CopyKat step required round-tripping AnnData through .h5ad. R support, multiscale point and polygon representations, and cloud access via Vitessce are listed as future work. · stated by the authors
  • The deep learning example was trained for only a few epochs without hyperparameter tuning, giving suboptimal cell-type predictions. It shows infrastructure, not model performance. · stated by the authors
  • No quantitative benchmark of load time, memory use or scalability against existing tools is reported. Table S1 compares features only, and laptop-level performance is described qualitatively. · noted in this summary
  • Alignment uses 3 manually placed landmarks with affine transforms and reports no registration error. Validation rests on one breast tumor, and the prostate example's 12 slides (text) vs 15 fields-of-view (legend) are not reconciled. · noted in this summary

Who should read:Read this if you need to store, align and jointly query images, transcripts, segmentations and count tables from several spatial omics platforms (e.g., Xenium plus Visium) in Python.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISLangseth '26Baker '23Chang '26Cai '25Chen '23Venturelli '26
73% semantic matchNeuroscience
Christoffer Mattsson Langseth, Bastien Hervé, Hanna P. Piechaczyk+3 more·bioRxiv·Mar 12, 2026