henhaobio
Sign in

bioRxivBioinformaticsPreprint

KSTITCH links cellular morphology and gene expression in spatial transcriptomics

Shashwat KumariD, Yingxiao ShiiD, Tuulia ValliusiD, Chi‐Ping DayiD, P.-A. Absil, Anuj SrivastavaiD, Sridhar S. HannenhalliiD, Vishaka DattaiD

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

Abstract

In situ spatial (ISS) sequencing can uncover co-variation between cellular morphology and gene expression in vivo. However, a principled and interpretable mathematical representation of morphology has not yet been applied in this context. In particular, current deep learning-based representations of cell images confound a cell's shape with its size. We present an interpretable representation of cellular boundary contours, based on tangent principal component analysis (TPCA) in a Kendall shape manifold, that captures size-independent contour shape features. This approach successfully recovers shape-perturbing genes in an RNAi screen than a previous metric geometry-based approach. We build on TPCA to develop KSTITCH (Kendall Shape-TranscriptomIc Correlation and Harmonization), an approach to reveal covariation between cell morphology with gene expression in ISS datasets. In a Xenium dataset, KSTITCH recovers known morphology-transcriptomic relationships in keratinocytes, macrophages and endothelial cells. Across samples in a melanoma CosMx dataset, KSTITCH reproducibly associates elongated and triangular fibroblasts with proximity to malignant cells and myofibroblast-like transcriptional program. Finally, KSTITCH independently recovers a known link between mesenchymal-like malignant cell states and increased cell area in two melanoma cohorts. KSTITCH can thus yield interpretable morphology-transcriptome relationships across cell types, patients, and spatial transcriptomics platforms. KSTITCH is available at https://github.com/vishakagopalan/kstitch .

In-depth summary

Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 2 (26 Aug 2026) · CC0 1.0

STITCH pairs size-independent tangent-PCA contour shape modes with expression factors by canonical correlation, recovering keratinocyte, fibroblast and melanoma shape-expression links in Xenium and CosMx data.

CosMx and Xenium pair each cell's contour with its transcriptome, but prior pipelines (MUSE, using Inception v3 image embeddings; CAJAL, using Gromov-Wasserstein distances) give no interpretable shape axes and conflate size with shape, although size can reflect field-of-view edge effects or magnification. STITCH separates area from shape, models shape by tangent PCA on a Kendall manifold, and links both to expression factors by canonical correlation analysis.

Key results

  1. ShownTangent PCA (TPCA) distances flagged nuclear-shape genes in a published siRNA screen with higher odds-ratio enrichment and far fewer calls than CAJAL, whose recovery collapsed once nuclear size was removed.

    TPCA (10 PCs) called 21 shape genes, 7 in the 34-gene ground truth (odds ratio 6.84; enrichment of area under the precision-recall curve, auPRC, over baseline 3.10). CAJAL called 277 including all ground-truth genes (odds ratio 2.51, auPRC enrichment 3.46) but recovered none once nuclei were size-normalized. Dropping PC1 or PC1-2 still gave 17 and 16 calls (odds ratios 8.45, 7.69). · Tangent Principal Component Analysis generates…; TPCA and CAJAL analysis of CRL1474 fibroblast…

  2. ShownMalignant-cell area tracked a mesenchymal-like (MES) expression axis (CEP1, first canonical expression projection) in two CosMx melanoma cohorts analysed separately; projected onto bulk RNA-seq of melanoma cell lines, CEP1 scored MES lines (largest measured area) above melanocytic (MEL) and Transitory lines.

    Love et al, 10 samples: canonical correlation 0.1 to 0.25; area loaded positively on the shape projection CSP1; inter-sample loading cosine similarity 0.9 (CSP1), 0.37 (CEP1), p < 0.001; high-CEP1 factors overlapped Andrews et al MES and Tsoi et al undifferentiated-neural crest-like genes. Dong et al, 25 tumors: similarities 0.312 and 0.146 (p < 0.001). Both cohorts' loadings ranked MES lines highest. · STITCH links mesenchymal-like melanoma cell…; STITCH analysis of malignant cells in Love et al…

  3. SuggestedIn the Love et al CosMx melanoma cohort, fibroblasts nearer malignant cells were smaller in area, more eccentric and triangular, and expressed factors overlapping myofibroblast signatures; this axis recurred across the ten samples.

    In Metastasis 1, CEP1-CSP1 R = 0.51 and CEP1 rose with distance from malignant cells; low-CEP1 factors 2 and 5 overlapped Gao et al c20-CD74 and c04-LRRC15 myofibroblast signatures. Across 10 samples, the first canonical correlation was 0.1 to 0.5, CEP1 correlated positively with malignant-cell distance in each, and loadings aligned (mean cosine similarity 0.78 for CEP1, 0.53 for CSP1, p < 0.001). · STITCH reproducibly recovers shape changes…; STITCH analysis of fibroblasts from Love et al…

  4. SuggestedIn one Xenium Prime melanoma skin section, clustering keratinocytes on STITCH's joint shape-expression projections recovered the keratinocyte layers and their nuclear eccentricity gradient, which MUSE did not.

    2,331 keratinocytes: each of six STITCH clusters was enriched (adjusted p < 0.05) for one layer only, and basal- or granular-enriched clusters had more eccentric nuclei than spinous ones (cluster 0 excepted); most of MUSE's eleven clusters mixed layers, with no significant eccentricity differences. Normalized mutual information 0.202 vs 0.002; spatial coherence 0.486 vs 0.447; CC1-3 r = 0.21, 0.16, 0.1. · STITCH recovers known associations between…; Comparison of MUSE and STITCH in keratinocytes

  5. ShownEuclidean distances on the first ten tangent PCs closely approximated geodesic shape distances, and pooled shape embeddings showed no noticeable batch structure across replicates or CosMx slides.

    Rho = 0.981 between first-ten-PC and geodesic distances for nuclei sampled from the Schibler et al data (N = 1000); 10 PCs held 90% of variance, PC1 46% (tracking eccentricity and roundness). Fibroblast shape and expression-factor embeddings pooled over three CosMx slides gave average kBET p-value > 0.9. · Tangent Principal Component Analysis generates…; STITCH reproducibly recovers shape changes…; Figure legends

Methods, data and reproducibility

Cell or nuclear contours (50 landmarks) are stripped of position, scale, rotation, reflection and starting point on a Kendall shape manifold and summarized by tangent principal component analysis (TPCA). Per cell type, area-augmented shape PCs and LIGER integrative non-negative matrix factorization factors are residualized for technical covariates and linked by canonical correlation analysis (CCA).

  • Schibler et al nuclear morphology siRNA screen in immortalized CRL1474 fibroblasts; DAPI images re-segmented with Cellpose · n = 2,011,829 nuclei; 861 genes (867 proteins in Methods), 3 siRNAs each
  • 10x Xenium Prime formalin-fixed paraffin-embedded (FFPE) primary dermal melanoma section, 5,006-gene panel · human (10x human skin dataset) · n = 112,551 cells; 2,331 keratinocytes analysed
  • Love et al CosMx melanoma cohort (nevi, primary and metastatic melanoma; 968-gene panel; three slides) · human (melanoma patients) · n = 10 patients · doi:10.5061/dryad.ksn02v7b1
  • Dong et al CosMx melanoma cohort, clinical stages I-IV; malignant cells only · human (melanoma tumors) · n = 25 tumors
  • De Blander et al melanoma cell line bulk RNA-seq labelled MES, MEL or Transitory, with measured mean cell area (NCL lines excluded) · GSE134432
  • Published signatures: Gao et al fibroblast states; Tsoi et al and Andrews et al melanoma differentiation states · n = 20 Gao et al fibroblast signatures (filtered subset analysed)
Methods & tools
Cellpose v3.0.5, OpenCV v4.12.0, LIGER (iNMF), Seurat v5.3.0, SpaceTrooper, kBET, CAJAL, MUSE, Google Inception v3, PhenoGraph, Human Protein Atlas, aricode
Reagents & environment
Cellpose v3.0.5 (pretrained nuclear model), OpenCV v4.12.0, Seurat v5.3.0; covariate regression with the R glm function; no container or requirements file stated.
Reproducibility gaps
TPCA pseudocode is in an appendix (Algorithms 1-4) but no code repository is linked; NCL melanoma cell line RNA-seq was not publicly available, so those lines were excluded; no source is given for the Schibler et al raw images.
Limitations (4)
  • Imperfect segmentation can assign transcripts to the wrong neighbouring cell, creating spurious shape-expression covariation or diluting real covariation; the Human Protein Atlas gene filter guards only against mis-assignment from neighbours of another type, not the same type or subtly different states. · stated by the authors
  • Canonical correlation analysis is linear and may miss switch-like links between shape and expression; nonlinear substitutes such as Deep CCA or kernel CCA would sacrifice interpretability. · stated by the authors
  • The MUSE comparison rests on one Xenium section (2,331 keratinocytes), and layer ground truth came from marker-gene expression, so agreement may partly reflect the expression half of the STITCH embedding rather than nuclear shape. · noted in this summary
  • TPCA's edge over CAJAL depends on the metric (CAJAL had the higher auPRC enrichment, 3.46 vs 3.10); shape-expression correlations were modest (malignant-cell first canonical correlation 0.1 to 0.25); and the cell-line CEP1 comparison is reported in the text without a test statistic. · noted in this summary

Read this if you work with CosMx or Xenium data and want interpretable, size-independent cell or nuclear shape features linked to expression programs instead of image-embedding distances.

Citations
0

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
THISChen '26Hou '26Xiao '26Shan '22Tan '19Pham '23
58% semantic matchBioinformatics
Po-Wei Chen, Guanao Yan, Shan Yu+1 more·bioRxiv·Nov 9, 2026