henhaobio
Sign in

Nature CommunicationsCell BiologyPeer-reviewed journal article

MuSpAn: A Toolbox for Multiscale Spatial Analysis

Joshua A. BulliD, Joshua W. MooreiD, Shania M. CorryiD, Muyang Lin, Hayley L. Belnoue-DavisiD, Eoghan J. Mulholland-IllingworthiD, Simon J. LeedhamiD, Helen Mary ByrneiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Advances in multiplex imaging and spatial omics have revolutionised spatial data generation in biology, revealing complex tissue organisation across multiple scales. However, methods for analysing these data have lagged behind, with fragmented, study-specific pipelines and limited guidance for tool selection. To address this, we introduce MuSpAn, a Multiscale Spatial Analysis package offering intuitive, flexible access to a wide range of mathematical tools - including spatial statistics, topological data analysis, geometry, and networks - within a unified framework. MuSpAn supports efficient data querying, is agnostic to imaging modality, and provides extensive documentation and community support. It enables users to create custom pipelines or conduct unbiased exploratory analyses. We demonstrate MuSpAns capacity to interrogate cross-compartmental cell interactions at multiple length scales in both normal and neoplastic tissue using mouse intestinal spatial transcriptomic datasets. Applied to a CMS4-like murine intestinal cancer model, MuSpAn identifies a continuum of fibroblastic functional phenotypes associated with discrete and coordinated fibroblast-immune interactions, highlighting its utility as a discovery tool across diverse biological contexts.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 4 (24 Sept 2025) · CC BY 4.0

MuSpAn, a Python library of over 100 spatial methods, quantifies multiscale tissue organisation and maps cancer-associated fibroblast subtypes to distinct immune niches in mouse colorectal tumours.

Spatial omics and multiplex imaging are typically analysed with fixed, modality- or question-specific pipelines (e.g., SPIAT, SpottedPy), and few platforms (SquidPy being a notable one) let users systematically compare and integrate diverse methods. MuSpAn offers a modality-agnostic object-and-query framework in which geometric, topological, network, statistical and density-based methods can be applied and cross-checked at any scale, from transcripts to whole sections.

Key results

  1. SuggestedIn AKPT mouse colorectal tumours, four transcriptional cancer-associated fibroblast (CAF) subtypes occupy distinct locations and immune niches: CAF1 sits inside tumours alongside Spp1+ macrophages, whereas CAF4 lies outside the tumour periphery with Tregs, B cells and dendritic cells.

    n = 4 mice, bespoke 480-gene Xenium panel. APT on ≤1.5µm contact networks: CAF1 with MAC2 (Spp1+), CAF3 with Cxcl2+ cells, CAF4 with B and dendritic cells. 3-hop neighbourhoods (k-means, k = 4): MAC2-dense around 51% of CAF1, Treg-type around 67% of CAF4. Cross-PCFs (to 500µm): CAF1 correlated with MAC2 and excluded from Tregs; CAF4 the reverse. · Multiscale spatial profiling of…; Multiscale spatial profiling of…; Cell-cell contact analysis; Neighbourhood analysis

  2. SuggestedCombining contact, neighbourhood, local correlation and cross-PCF metrics into one multiscale vector per fibroblast separates the CAF subtypes in spatial feature space. A minimum spanning tree over subtype medians traces a fibroblast-immune continuum that matches the transcriptional CAF1-to-CAF4 trajectory.

    87-dimensional z-scored vector per fibroblast: contact counts (7), 3-hop counts (7), TCM at r = 30µm (7), cross-PCF contributions at 0, 100, 200, 300, 400µm (55) and cumulative cross-PCF ratio to R = 150µm (11); weighted minimum spanning tree on 1-norm distances between subtype medians projected into UMAP. Segregation is shown visually, without a separation statistic. · Multiscale spatial profiling of…; Multiscale spatial profiling of…; Trajectory inference; Multiscale spatial feature vectors

  3. ShownIn the public healthy mouse colon Xenium section, nearest-neighbour and quadrat-correlation analyses both recover the expected SMC-STR1 colocalisation and PROG-colonocyte separation, while contact networks confirm frequent SMC-STR1 contacts in the lamina propria. This checks the toolbox against known biology.

    ROI ≈ 0.38 mm^2: nearest-neighbour distance 10.9µm for SMC-STR1 vs 106.0µm for PROG-COL, Average Nearest Neighbour Index z-scores p < 10^−3; QCM (75µm quadrats, K = 1000 shuffles) z = 2.87 for SMC-STR1 and z = −1.69 for PROG-COL; SMCs show low contact-network degree but frequent STR1 neighbours. · Quantification of cell proximity and interaction…; D.2.1 Nearest neighbour metrics; D.2.2 Quadrat correlation matrix; Figure legends

  4. ShownIn the same tissue, MuSpAn separates SMC and progenitor cell shapes by circularity and orientation. Boundary-corrected cross-PCFs give the length scales of PROG clustering and PROG-SMC exclusion, and persistent homology picks up crypts as loops.

    SMC vs PROG (0.13mm^2 ROI, Wilcoxon): area p = 0.26, circularity p < 10^−3, principal angle p < 10^−3. Cross-PCF (≈ 1.13 mm^2 ROI): PROG-PROG g > 1 for r < 150µm, PROG-SMC g < 1 for r < 80µm. Vietoris-Rips H1 loops of radius ≈ 25µm in PROG cells correspond to transected crypts. · Cell geometry and subcellular transcript…; Assessment of multiscale spatial patterns; D.3.2 Vietoris-Rips filtration; Figure legends

  5. ShownRebuilding three published pipelines (stacked IHC, imaging mass cytometry, CODEX) reproduces and extends them: compartment-restricted statistics show stronger macrophage clustering in tumour than stroma, and neighbourhood contacts differ between inflammation groups.

    Head and neck IHC: at 20µm, tumour macrophages around 2.75 times and stromal around 1.75 times more likely than random to pair. CODEX (Schürch et al.): APT effect sizes for Neighbourhood 4-8 contacts are higher in DII than CLR TMAs (p ≈ 0.0013, unequal variance t-test). IMC lung: TCM resolves conflicting QCM and APT signals for one pair. · A.1 Macrophage localisation in head and neck…; A.2 COVID-19 infected lung (Imaging Mass…; A.3 CODEX - Schürch et al. (2020); Figure legends

Methods, data and reproducibility

Python library that stores points and polygons (transcripts, centroids, cell boundaries, annotations) with labels in a domain, filtered via composable queries, and applies over 100 methods from six modules (distribution, geometry, networks, region-based, spatial statistics, topology) at any scale, with parent-child links propagating features between scales.

  • Public 10x Xenium fresh-frozen healthy mouse colon section (Mouse Tissue Atlas Panel, 379 genes; 19 of 25 Leiden clusters annotated) · mouse (8-week-old male C57)
  • AKPT (Villin-CreERT2; Apc; Kras G12D; Trp53; Tgfbr1) colorectal tumours, FFPE Swiss rolls profiled with a custom 480-gene Xenium panel · mouse · n = 4 mice (n = 2 Apc fl/fl, n = 2 Apc fl/+) · 10.5281/zenodo.17176282 (MuSpAn domains)
  • Resected head and neck cancer imaged as registered stacked IHC (PanCK, CAIX, pimonidazole, CD68), reanalysing a published PCF/J-function study · human · n = four consecutive slices; 1.5mm × 1.5mm ROI
  • COVID-19 infected lung (organising pneumonia) imaged by imaging mass cytometry, reanalysing a QCM/cross-PCF/TCM pipeline · human · n = 2mm × 2mm ROI
  • Colorectal cancer CODEX tissue microarrays from Schürch et al. (2020), Crohn's-like reaction vs diffuse inflammatory infiltration groups · n = 68 group 1 (CLR) TMAs shown
  • Synthetic-Points-Architecture point dataset bundled with MuSpAn (ring of two cell types amid randomly placed points)
Methods & tools
MuSpAn v1.2.0 (Python), 10x Genomics Xenium In Situ and Xenium Explorer (v3), Seurat v5.2.0 in R v4.3.3, ComplexHeatmap v2.18.0, networkx v3.4.2 (weighted minimum spanning tree), UMAP (Manhattan metric), Cross pair correlation function (cross-PCF) and cumulative PCF ratio, Adjacency permutation test and k-hop neighbourhood k-means clustering, Topographical correlation map (TCM) and Local Getis-Ord, Persistent homology (Vietoris-Rips filtration), Quadrat correlation matrix and Average Nearest Neighbour Index, Gaussian KDE, KL divergence and sliced 2-Wasserstein distance
Software & environment
MuSpAn v1.2.0 (Python); Seurat v5.2.0 in R v4.3.3; ComplexHeatmap v2.18.0; networkx v3.4.2; Xenium Explorer (v3); Xenium v1 chemistry.
Reproducibility gaps
MuSpAn is free only for non-commercial use under an Academic Use Licence (commercial users license via Oxford University Innovation); no deposition of raw AKPT Xenium outputs is stated; only derived MuSpAn domains are on Zenodo.
Limitations (5)
  • Because MuSpAn is a flexible library rather than a fixed pipeline, the authors caution that users are responsible for assembling analyses suited to their biological question (guidance in Appendix B). · stated by the authors
  • Single statistics can disagree or mislead when used alone: in the COVID-19 lung IMC example, QCM, APT and cross-PCF gave conflicting signals for one cell pair, so the authors recommend combining methods from several modules. · stated by the authors
  • MuSpAn performs no image preprocessing; segmentation, region annotation and phenotyping must be done beforehand in external tools (e.g., Xenium Explorer, QuPath, Halo, Deepcell). · stated by the authors
  • The CAF-immune niche conclusions rest on four tumours from a single genetically engineered model, compared with Mann-Whitney tests (p < 0.05), with no orthogonal validation (e.g., protein staining or functional perturbation) of the proposed interactions. · noted in this summary
  • No runtime, memory or scaling benchmarks against SquidPy or other toolkits are reported, and the method count is stated inconsistently (over 100 methods vs 40+ methods in v1.2.0), which makes coverage claims hard to assess. · noted in this summary

Read this if you analyse imaging-based spatial omics (Xenium, CODEX, IMC, IHC) and want one Python toolkit to compare spatial statistics, networks and topology across scales, or study CAF-immune niches in CMS4-like colorectal cancer.

Citations
17
FWCI
18.89
Field percentile
top 1%

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
THISBlampey '24Liu '25Li '25Khawatmi '24Xiao '24Feng '26
64% semantic matchBioinformatics
Quentin Blampey, Kevin Mulder, Margaux Gardet+5 more·Nature Communications·Jun 11, 2024· 59 citations