Multiscale topology classifies cells in subcellular spatial transcriptomics

Abstract

Spatial transcriptomics measures in situ gene expression at millions of locations within a tissue¹, hitherto with some trade-off between transcriptome depth, spatial resolution and sample size². Although integration of image-based segmentation has enabled impactful work in this context, it is limited by imaging quality and tissue heterogeneity. By contrast, recent array-based technologies offer the ability to measure the entire transcriptome at subcellular resolution across large samples^3-6. Presently, there exist no approaches for cell type identification that directly leverage this information to annotate individual cells. Here we propose a multiscale approach to automatically classify cell types at this subcellular level, using both transcriptomic information and spatial context. We showcase this on both targeted and whole-transcriptome spatial platforms, improving cell classification and morphology for human kidney tissue and pinpointing individual sparsely distributed renal mouse immune cells without reliance on image data. By integrating these predictions into a topological pipeline based on multiparameter persistent homology^7-9, we identify cell spatial relationships characteristic of a mouse model of lupus nephritis, which we validate experimentally by immunofluorescence. The proposed framework readily generalizes to new platforms, providing a comprehensive pipeline bridging different levels of biological organization from genes through to tissues.

Fig. 1. Method overview.

a, Spatial transcriptomics measurements can be grouped according to their relative scale: subcellular, single-cell and multicellular. Existing methods decompose multicellular readings into cell type proportions. With subcellular data, it is necessary to aggregate data to reach the single-cell level. b, Automatic spatial cell type identification requires integration of single-cell and spatial transcriptomics. c, A common approach is to aggregate expression into a fixed-window and then run a standard classifier; however, this fails to resolve individual cells and can lose information about sparsely dispersed cells caught between bin boundaries. d, TopACT minimizes information loss by taking a flexible topological approach. Each spot is classified independently using local neighbourhoods at several scales, accommodating heterogeneous cell sizes and varying per-spot transcriptional abundance. This flexibility allows detection of finer structural information, including individual sparsely dispersed cells. e, A radius–codensity bifiltration defined on a two-dimensional point cloud. At each radius–codensity pair, a ball of that radius is drawn on top of all points with at most the given codensity (that is, sparseness). The hue indicates the underlying point codensity. The radius parameter therefore changes the scale of interaction between points and the codensity parameter controls the level of noise reduction. f, Archetypal spot assignment patterns (top) and their corresponding MPH landscapes (bottom). (i) The large loop structure activates the landscape at high radius values. (ii) The small loop structure activates the landscape at low radius values. (iii) A saturated loop structure with central clusters does not activate the landscape. (iv) A point cloud with no underlying loop structure does not activate the landscape. (v) The codensity parameter ensures that the landscape is still activated even in the presence of outliers and misclassifications. Panels a–d were created with BioRender (https://BioRender.com).

Fig. 2. Benchmarking TopACT with synthetic data.

a, Synthetic data generation schematic. b, Sample output of cell type identification algorithms on synthetic data. From left to right: ground truth, TopACT, RCTD. Colours indicate cell types. c, Box plots of per-iteration accuracy of cell type classification methods on synthetic data (n = 100 iterations). Centre line shows median; box limits show interquartile range (IQR); whiskers show full range. Modal is the optimal bin 20 classification assigning to each bin its most common ground truth cell type. d, TopACT performance on rare cell types. Top, number of cells of each type detected per iteration. Coloured regions denote cells detected by TopACT, red regions denote cells not detected by TopACT. Bottom, box plots showing recall of sparsely dispersed cells (n = 100 iterations). Centre lines show median; box limits show interquartile range; whiskers show full range of non-outlier points. Outliers are points more than 1.5× IQR from the upper or lower quartiles. Full distributions overlaid. e, Accuracy of methods under simulated molecular diffusion. Methods run for n = 10 iterations each on mean diffusion magnitudes of λ μm for λ = 0,1,2,…,7. Lines show mean, bands show s.e.m. Vertical dashed lines refer to previous diffusion estimates in literature^,. CCD, cortical collecting duct; DCT, distal convoluted tubule; ECS, endothelial cells; PT, proximal tubule; TAL, thick ascending limb of the loop of Henle; VSM, vascular smooth muscle. Source Data

Fig. 3. TopACT predicts previously unidentified PVM cells in adult mouse brain.

a, Spatial locations of TopACT-predicted PVM cells (black crosses). Background heatmap shows smoothed transcript count. Black dashed line shows convex hull of high-density regions, to which analysis is restricted. Scale bar, 0.5 mm. b, Violin plots for expression of common markers of PVM cells, for TopACT-predicted PVM cells (blue, n = 66 cells) and randomly sampled background cells (green, n = 66 cells), across the entire mouse brain sample. Each plot corresponds to the expression counts of a single given marker gene in cells labelled with the given cell type. Violins show kernel density estimate of data distribution. Inner box-and-whisker shows summary statistics as follows: white centre line shows median; box limits show IQR; whiskers show full range. A log scale is used. Source Data

Fig. 4. TopACT cell segmentation of human IgAN kidney profiled on the Xenium platform.

a, H&E staining of an exemplar tissue region including tubular structures. b, Supervised tubular cell type annotation on imaging-based cell segmentation. Cell types profiled are proximal tubule (pale blue), thick ascending limb (dark blue) and principal cell (PC, turquoise). c, TopACT tubular cell type annotations. d, Comparison of H&E (left) and TopACT-annotated (right) tubular region at higher magnification. e, H&E staining of an exemplar tissue region showing glomeruli, marked with red circles. f, Supervised podocyte (Pod, purple) and endothelial (green) cell type annotation on imaging-based cell segmentation. Cells annotated as podocytes outside glomerular regions are identified by red circles. g, TopACT podocyte and endothelial cell type annotations. A single pixel (1 μm) outside the glomerular regions, marked with a red circle, is identified as containing podocyte markers but does not meet size criteria to be defined as a cell. h, Comparison of H&E (left) and TopACT-annotated (right) glomerular region at higher magnification. All annotations overlaid on DAPI fluorescent imaging showing nuclei distribution. All images are representative regions of a single Xenium experiment. Scale bars, 100 μm.

Fig. 5. TopACT predicts immune cell ring structure in mouse glomeruli.

a–e, Analysis of Stereo-seq kidney sections (four control, six treated). a, Example TopACT-predicted immune cells. Background, transcript density. Scale bars, 0.2 mm. b, Example glomerular (blue) and non-glomerular (orange) patch distribution. Scale bar, 0.2 mm. c, Mean TopACT-predicted immune count per patch, by condition and patch type. Error bars, s.e.m. Increased immune cell numbers (P = 4.2 × 10⁻⁵) observed in glomerular (n = 161) versus non-glomerular (n = 180) patches in treated samples. d, Histogram of distances between immune cells and nearest podocyte. e, MPH analysis. (i) Glomerular patches. Scale bars, 0.2 mm. (ii) TopACT spot-level immune annotations. Scale bars, 20 μm. (iii) Single-patch MPH landscapes. (iv) Average MPH landscapes over all patches. Treated average indicates large peripheral loop structures (compare Fig. 1f(i) and (v)). f–h, Multiplex immunofluorescence analysis (three control, three treated kidneys). f, Representative renal cortex immunofluorescence. Scale bars, 100 μm. g, Mean immune intensities in glomerular versus non-glomerular regions, n = 75 regions per region type and condition. Box plots, centre line shows mean; box limits show IQR; whiskers show full range. Increase (P = 5.6 × 10⁻¹²) in glomerular regions of treated samples consistent with TopACT. h, Ratio of mean intensity in outer to central glomerular region, by condition, for immune subtypes. Increased ratio (P = 7.3 × 10⁻⁵, n = 90 glomeruli per condition) over all immune types in treated samples consistent with MPH prediction. Individual increases for T cells (P = 2.7 × 10⁻⁶, n = 45 glomeruli per condition) and myeloids (P = 2.3 × 10⁻³, n = 75 glomeruli per condition). All statistical tests are one-sided Welch’s t-tests. NS (not significant), P ≥ 0.06, *P ≤ 0.05, **P ≤ 0.01, ***P ≤ 0.001, ****P ≤ 0.0001. Source Data

Extended Data Fig. 1. Reference snRNA-seq data from four healthy human kidney biopsies.

a, UMAP embedding with labelled cell clusters. b, Dot plot showing expression of key marker genes for each cell type. PCT = proximal convoluted tubule; TAL = Thick ascending limb; LOH = loop of Henle; DCT = distal convoluted tubule; PC = principal cell; IC-A = α intercalated cell, IC-B = β intercalated cell; ENDO = endothelial cell; MES = mesangial cell; POD = podocyte; PEC = parietal epithelial cell; LEU = Leucocyte.

Extended Data Fig. 2. Reference snRNA-seq data for control and IMQ-treated mouse kidney.

Top: UMAP embeddings with labelled cell type clusters. Bottom: Violin plots displaying highly differentially expressed genes per cell type. TAL = thick ascending limb; ENDO = endothelium; PODS = podocyte; VSM = vascular smooth muscle; PT = proximal tubule; DCT = distal convoluted tubule; CNT = connecting tubule; CCD = cortical collecting duct; CCD_aIC = cortical collecting duct α intercalated cell; CCD_bIC = cortical collecting duct β intercalated cell.

Extended Data Fig. 3. Concordance of TopACT predicted cells with ssDNA-based cell segmentation in a representative IMQ-treated mouse kidney sample.

a, ssDNA-based cell segmentation annotated with TopACT-predicted immune (orange) and podocyte (purple) cells. Grey regions indicate ssDNA cell bins without a corresponding TopACT annotation. Right: The whole sample. Left: A magnified representative patch. Crosses denote TopACT-predicted cell centres. Note that each cell bin in the representative patch contains at most one TopACT prediction. b, Magnification of the three ssDNA-bins with more than one assigned TopACT cell. In each case the bin is assigned two TopACT cells of the same type.

Extended Data Fig. 4. Mouse kidney patch distributions.

Squares show glomerular (blue) and non-glomerular (orange) patches defined on each sample. Background heatmaps show smoothed transcript count. Scale bars: 0.2 mm.

Extended Data Fig. 5. Analysis of spatial immune subpopulations in mouse kidney samples.

a, Proportions of T-cells, macrophages, dendritic cells (DC) and B-cells across all samples. b, Proportions of T-cells, macrophages, dendritic cells (DC) and B-cells, split according to glomerular proximity. Glom = within glomerulus. Non-glom = not within glomerulus. c, Expression of immune subtype marker genes across immune subpopulations, separated by IMQ-treated and control. Macrophage markers: Apoe, C1qa, Lyz2, activated macrophage marker: Ly6a, T-cell marker: Sugct. B-cell markers: Ighg2b, Jchain. DC markers: Tnfrsf21, Cfb.

Extended Data Fig. 6. Spatial distribution of immune subpopulations relative to glomerulus loci.

Top row: Macrophage in glomerulus (pink), macrophage outside of glomerulus (dark blue), glomerulus (pale blue). Middle row: T-cell in glomerulus (pink), macrophage outside of glomerulus (dark blue), glomerulus (light blue). Bottom row: B-cell in glomerulus (pink), B-cell outside of glomerulus (dark blue), glomerulus (light blue).