Spatial atlas of diabetic kidney disease reveals a B cell-rich subgroup

Abstract

Diabetic kidney disease (DKD), the leading cause of kidney failure, is marked by clinical and molecular heterogeneity, making therapeutic development exceedingly difficult¹. Here we used Xenium and CosMx single-cell spatial transcriptomics, integrated with single-nucleus RNA sequencing, to build a cross-platform kidney atlas that makes tissue architecture computable for prognosis, non-invasive detection and patient selection. Using this atlas, we defined reproducible tissue niches and injury-linked microenvironments and uncovered a profibrotic context that expands with disease and tracks with worse kidney function. Within this architecture, we identified a B cell-predominant, tertiary lymphoid structure-like immune microenvironment that defines a distinct DKD subset with accelerated progression to renal end-points. We developed tissue biomarkers and a matched plasma protein panel that capture this biology, stratify patients in a population biobank and improve risk prediction beyond clinical models—supporting their potential for biomarker-guided selection in future B cell-targeted DKD trials.

A single-cell spatial atlas identifies a B cell-predominant microenvironment within the profibrotic tubular niche that marks a subset of patients with diabetic kidney disease with rapid progression.

Main

The human kidney is a highly complex organ whose three-dimensional architecture enables essential functions such as fluid and electrolyte balance, waste excretion and blood pressure regulation^2–4. Chronic kidney disease (CKD) is characterized by a persistent reduction in glomerular filtration rate (GFR), accumulation of toxins, salt and water retention, and abnormalities in electrolyte and hormonal balance⁵. If untreated, kidney disease can lead to end-stage renal failure, requiring life-sustaining dialysis or transplantation. DKD is the leading cause of CKD and develops in approximately 40% of patients with diabetes^6,7. Recent therapeutic advances—including sodium–glucose cotransporter 2 inhibitors, non-steroidal mineralocorticoid receptor antagonists and selective endothelin receptor antagonists—have demonstrated efficacy in slowing disease progression^8–12. Despite these advances, end-stage renal failure remains common, reflecting the heterogeneity of DKD, where disease progression ranges from rapid decline to long-term stability¹. With the increasing global burden of diabetes, there is an urgent need for precise, mechanism-based approaches to stratify and treat DKD. Clinical factors such as age, sex, estimated GFR (eGFR) and albuminuria predict outcomes but reveal little about the underlying biology, and genetic risk scores offer limited value for risk or treatment guidance^13,14. Circulating tumour necrosis factor (TNF) receptors 1 and 2 strongly associate with renal decline and end-stage renal failure^15,16, yet the cellular and spatial context of kidney inflammation in DKD remains unclear, limiting translation into targeted therapies. Next-generation single-cell spatial transcriptomics enables near-genome-wide gene profiling at true cellular resolution within intact tissue, preserving spatial context and revealing disease-relevant tissue organization and distinct pathological subtypes directly from clinical biopsies.

A spatial atlas at cellular resolution

We generated a single-cell spatial atlas from 64 formalin-fixed paraffin-embedded (FFPE) kidney samples from 58 patients (26% female, with a mean age of 57.3 years) spanning the full range of GFR and diverse demographics (Fig. 1b–e and Supplementary Table 1). Samples were profiled using CosMx (1,000 genes, n = 48) and Xenium (5,100 genes, n = 16) platforms, with 656 overlapping transcripts. After filtering low-quality cells, we integrated each spatial dataset with a reference single-nucleus RNA-sequencing (snRNA-seq) atlas from 150 patients¹⁷ (Supplementary Table 2) using scVI¹⁸ and scANVI¹⁹. Integration performance assessed by the single-cell integration benchmark²⁰ showed that scANVI performed best (Extended Data Fig. 1b). We retained 3,386,822 CosMx cells and 951,041 Xenium cells after filtering and integrated them with scANVI¹⁹ into a unified spatial atlas of more than 5 million cells from more than 200 patients (Fig. 1f–i and Extended Data Fig. 2; quality control metrics are shown in Supplementary Figs. 1–8). Unbiased Leiden clustering followed by marker-based annotation identified 20 cell types, all present across samples except 3 that lacked glomeruli (Fig. 1f, Extended Data Fig. 1a and Supplementary Tables 3 and 4). The integrated spatial dataset revealed distinct endothelial subtypes, including glomerular, peritubular and descending vasa recta cells, and the higher resolution of Xenium enabled detection of lymphatic endothelium (Fig. 1f). Stromal cells are difficult to distinguish owing to shared markers and spatial overlap, but we identified fibroblasts, vascular smooth muscle cells (VSMCs) and mesangial cells based on gene expression and location. Mesangial cells showed higher IGFBP5 (CosMx), PDGFRB and IL1RL1 (Xenium) expression (Extended Data Fig. 1c,d). Spatial resolution also refined annotation of parietal epithelial cells and the loop of Henle (LOH), distinguishing thick ascending (TAL) from thin descending (DTL) and ascending (ATL) segments (Extended Data Fig. 1c,d and Supplementary Tables 3 and 4). We identified injured proximal tubule (iPT), proximal tubule (PT), injured thick ascending limb (iTAL) and thick ascending limb (TAL) cells by injury-associated gene signatures. Differential expression between injured and healthy segments revealed conserved markers across platforms, including SOX4, VCAM1 and HAVCR1 in iPT cells, and CLU, SLPI and ITGB6 in iTAL cells (Extended Data Fig. 2a–d). Because CosMx and Xenium are not genome wide, we imputed 3,000 highly variable genes from the snRNA-seq reference using the scANVI latent space. Imputed expression patterns aligned with known cell identities, as canonical markers localized to expected populations (Supplementary Fig. 9, Supplementary Table 5). We combined Periodic acid–Schiff (PAS)-stained CosMx and haematoxylin and eosin (H&E)-stained Xenium data for paired histological and spatial analyses (Fig. 1j,k). DKD tissue showed marked proximal tubule disruption with abundant injured epithelial cells. High-resolution views revealed key features: a healthy interlobular artery (with VSMCs, endothelial cells and fibroblasts; Fig. 1l), an inflamed tubulointerstitial region (with immune cells, injured epithelium and fibroblasts; Fig. 1m) and a collecting duct (CD) with principal and intercalated cells (Fig. 1n).

Fig. 1. A spatial atlas at cellular resolution.

a, We generated a spatial atlas from 64 human kidney samples to define kidney niches and characterize fibrosis through injured tubular microenvironments. We built a spatial immune atlas to study renal inflammation, uncovering a B cell-predominant immune microenvironment in DKD that correlated with disease severity and kidney function decline. Finally, bulk RNA-seq and machine learning identified tissue and circulating biomarkers of this disease-associated immune microenvironment, validated in the UK Biobank. b–e, Sex (b), age (c), race classification (d) and GFR (e) of participants in this study. snRNA-seq samples in blue; spatial samples in pink. B, Black; O, other; U, unspecified; W, white. f, UMAP of the integrated spatial single-cell atlas combining CosMx 1k, Xenium 5k and snRNA-seq data using scVI¹⁸ and scANVI¹⁹. Each dot represents one cell. CNT, connecting tubule; DCT, distal convoluted tubule; DVR, distal vasa recta; EC, endothelial cell; Fibro, fibroblast; glom, glomerular; IC, intercalated cell; Lymph, lymphatic; MC1, mesangial cell; PC, principal cell; PEC, parietal epithelial cell; Peritub, peritubular; Podo, podocyte. g, CosMx 1k spatial data from 48 samples and 3,386,822 cells. h, snRNA-seq data from 150 samples and 720,924 cells. i, Xenium 5k spatial data from 16 samples and 951,041 cells. j, PAS staining and CosMx 1k cell-type maps of a healthy paediatric kidney (top) and a DKD kidney (bottom). Cells are shown by spatial location with segmentation masks (Nanostring pipeline). Scale bars, 300 µm. Images are representative regions from the indicated samples. k, H&E staining and Xenium 5k cell-type maps of a healthy kidney (top) and a DKD kidney (bottom). Cells are shown by spatial location with their segmentation masks (10x Genomics Xenium Explorer 3.2.0). Scale bars, 300 µm. Images are representative regions from the indicated samples. l, High-magnification CosMx view of an interlobular artery region from j. Scale bar, 100 µm. m, High-magnification Xenium view of tubulointerstitial inflammation from k. Scale bar, 100 µm. n, High-magnification Xenium view of a collecting duct region from k. Scale bar, 100 µm. Inj. T., injured tubular.

Extended Data Fig. 1. a Cell type composition across samples in CosMx spatial transcriptomics data.

Stacked bar plot showing the relative proportions of cell type categories across individual samples. Each bar represents a sample, and the height of each colored segment corresponds to the proportion of that cell type within the total sample. Bars were normalized by the total number of cells per sample. Colors correspond to cell type identities. b: Benchmarking dataset integration: Methods are sorted by overall score. Overall score is a weighted mean of the batch correction score and the bio conservation score, which in turn are a mean of the individual metrics within the category. Metrics have previously been described (Luecken et al.). Briefly, these metrics were grouped into two broad categories: conservation of biological variance (Bio conversation) and removal of batch effects. All scores are normalized between 0 and 1, where higher values indicate better performance. c: Expression of selected marker genes across annotated cell types in the CosMx dataset. For each gene, expression was averaged per cell type and scaled to a 0–1 range using MinMax normalization. The color intensity represents the relative expression level of each gene across cell types. Visualization was performed using the Seaborn heatmap function. d: Expression of selected marker genes across annotated cell types in the Xenium dataset. For each gene, expression was averaged per cell type and scaled to a 0–1 range using MinMax normalization. Visualization was performed using the Seaborn heatmap function.

Extended Data Fig. 2. Differential gene expression analysis of injured versus non-injured epithelial subtypes in CosMx and Xenium datasets.

Volcano plots displaying differentially expressed genes between injured and non-injured epithelial subtypes in spatial transcriptomic data from CosMx and Xenium platforms. a: iPT vs PT (CosMx), b: iPT vs PT (Xenium), c: iTAL vs TAL (CosMx), and d: iTAL vs TAL (Xenium). Differential expression analysis was performed using the two-sided Wilcoxon rank-sum test after normalization and log-transformation of raw counts. Each point represents a gene, with the x-axis indicating log₂ fold change and the y-axis representing the –log₁₀ adjusted p-value. Genes with significant differential expression (adjusted p-value < 0.05 and |log₂FC| > 1) were highlighted, and selected markers were annotated. Horizontal and vertical dashed lines denote significance and fold-change thresholds, respectively.

Niches decode tissue architecture

The complex three-dimensional architecture of the kidney underlies its diverse functions, from glomerular filtration to medullary urine concentration. Although key structures are well characterized, we hypothesized that spatial transcriptomics could reveal additional patterns of cellular organization, which we refer to as ‘niches’ throughout this study.

Without a defined spatial scale for kidney niches, we aggregated cell–cell relationships within 20–80 µm radii to capture multiscale context and used machine learning to identify 11 recurring cellular neighbourhoods (niches) that were observed consistently across Xenium and CosMx samples (Fig. 2a,b). This unbiased analysis aligned with known kidney architecture: the glomerular niche contained glomerular podocytes, endothelial, parietal and mesangial cells and was enriched for genes linked to nephrotic syndrome and focal segmental glomerulosclerosis (Fig. 2b,e–g,n). The vascular niche consisting of endothelial and VSMCs showed enrichment for muscle contraction and calcium regulation pathways (Fig. 2b,e–g,n). Tubular regions separated into PT, DCT, CNT, LOH, CD and injury-associated iPT and injured loop of Henle (iLOH) niches. The PT niche featured transport and lipid metabolism genes, whereas the iPT niche was enriched for apoptosis and STAT3 signalling (Fig. 2b,e–m). The full list of niche-specific differentially expressed genes (DEGs) is presented in Supplementary Tables 6–8.

Fig. 2. Niches decode tissue architecture.

a, Schematic of spatial niche identification based on neighbourhood fractions across 20 cell types. Each dot represents a single cell, coloured by cell type. Large circles represent distinct spatial niches, and the smaller circles within each niche depict the cell types that colocalize and define that niche. b, Sample niche composition ordered by GFR. c, Relative niche fraction changes in DKD versus controls. d, Correlation of niche fractions with GFR (two-sided Pearson test; DCT: P = 7.4 × 10⁻⁷, immune: P = 1.4 × 10⁻⁶, PT: P = 1.7 × 10⁻⁵, iPT: P = 4.1 × 10⁻⁶). Shaded bands indicate 95% confidence intervals. e, Heat map of niche cell-type composition. f,g, Spatial niche labels overlaid on segmentation masks (representative healthy region) (g) and PAS staining of adjacent section (f). Scale bar, 250 μm. h–m, Spatial transcriptomic assignment of cell types in representative PT (h), iPT (i), CD (j), LOH (k), iLOH (l) and CNT (m) niches. Colours denote cell types (Xenium Explorer v3.2.0). Scale bars, 100 μm. n, Differentially expressed imputed genes between niches identified by two-sided Wilcoxon test and analysed by Gene Set Enrichment Analysis (GSEA; WikiPathways 2023; GSEApy); normalized enrichment score (NES) is shown; adjusted P < 0.05. ECM, extracellular matrix; FSGS, focal segmental glomerulosclerosis; RCC, renal cell carcinoma. o, Imputed DEGs of vascular niches (DKD versus control) analysed by over-representation analysis (GSEApy Enrichr; one-sided Fisher’s exact test with Benjamini–Hochberg correction); adjusted P < 0.05. p,q, Mesangial (p) and immune (q) cell fraction changes across niches (n = 44; two-sided Welch’s t-test, mesangial: P = 0.047, immune: P = 0.029). Violin plots display distributions with embedded box plots (median, interquartile range (IQR), 1.5× IQR whiskers). r, Ligand–receptor interactions between glomerular endothelial cells and podocytes (CellPhoneDB v5, imputed expression) correlated with GFR (two-sided Pearson test; NRG1–ERBB3: P = 0.005, PLAU–PLAUR: 0.014); shaded bands denote 95% confidence intervals. Schematic created in BioRender; Dumoulin, B. https://biorender.com/rwud0db (2026).

Source data

To validate niche patterns, we applied the COVET framework, which models the local gene–gene covariance for each cell²¹. Pairwise niche distances visualized by uniform manifold approximation and projection (UMAP) showed clear separation by niche labels, supporting the biological relevance of our assignments (Extended Data Fig. 3m–o). We also used NicheCompass²², a deep learning model integrating gene expression and spatial context, which identified eight niches consistent with our annotations and revealed niche-specific programmes such as APOE in PT and IL-32 in iPT (Extended Data Fig. 3p–r and Supplementary Fig. 11). Unbiased spatial mapping of cellular neighbourhoods (niches) revealed disease-associated changes in tissue organization. Niche frequencies differed markedly between healthy and disease samples (Fig. 2c) and DCT (r = 0.61) and PT (r = 0.52) correlated positively with eGFR, whereas immune (r = –0.59) and iPT (r = –0.55) niches correlated negatively with eGFR (Fig. 2d and Extended Data Fig. 3a–k). These compositional shifts reflect loss of epithelial integrity and inflammation, suggesting potential utility as quantitative indicators of kidney function and disease severity.

Extended Data Fig. 3. Correlation between glomerular filtration rate (GFR) and niche-specific cell fractions across spatial niches.

Scatter plots showing Pearson correlations between estimated glomerular filtration rate (GFR) and the relative abundance of individual cell niches across all samples in the spatial transcriptomics dataset. Each plot represents a distinct spatial niche (a: CNT niche, b: DCT niche, c: Fibroblast niche, d: Glomerular niche, e: iLOH niche, f: Immune niche, g: iPT niche, h: LOH niche, i: PC niche, j: PT niche, k: Vascular niche) and shows niche fraction vs. GFR per sample. Cell fractions were calculated by normalizing the number of niche-specific cells to the total cell count within each sample. A linear regression line is overlaid, shaded band represents 95% confidence interval of the linear regression. Pearson correlation coefficients and corresponding two-sided p-values are indicated. l: Scatter plot showing the relationship between the proportion of immune cells and mesangial cells within the glomerular niche across individual patient samples. Only samples from control and DKD patients were included in the analysis (n = 44). Cell fractions were calculated by normalizing the number of immune or mesangial cells to the total number of glomerular niche cells per sample. A linear regression line is overlaid in red, shaded band represents 95% confidence interval of the linear regression. Pearson correlation coefficients and corresponding two-sided p-values are indicated. Orthogonal niche definition and visualization using COVET: We applied the recently published COVET method to define cellular niches based on spatial covariance of gene expression. Inputs to the COVET matrix included: (1) spatially resolved gene expression, (2) 64 highly variable genes (selected via Scanpy’s highly_variable_genes), and (3) the number of spatial neighbors (k = 8. For each cell in the CosMx dataset with at least 50 neighbors within 200 μm, we computed an adjusted covariance matrix across the 64 genes to capture niche-specific co-expression patterns. Pairwise niche distances were then derived using Approximate Optimal Transport (AOT). Based on this distance matrix, we calculated UMAP (Uniform Manifold Approximation and Projection) for visualization after PCA which help to better examine the relationships within the COVET niche representation. Panels show UMAP embeddings of the resulting COVET adjacency matrices, with each point representing a single cell and colored by our annotation based niche annotations. m: UMAP for all patient samples combined. n: UAMP restricted to control patients. o: UMAP restricted to patients with diabetic kidney disease (DKD). Orthogonal niche definition and visualization using NicheCompass analysis. p: UMAP of integrated NicheCompass embeddings colored by celltypes. q: Uniform Manifold Approximation and Projection (UMAP) of integrated NicheCompass embeddings, colored by annotated NicheCompass niches. Niches were defined based on characteristic gene programs inferred from the spatial context. r: UMAP of integrated NicheCompass embeddings colored by annotation based identification of niches.

We hypothesized that the niche framework could uncover DKD-associated transcriptional changes. Indeed, we found niche-specific alterations linked to extracellular matrix remodelling, adhesion, inflammation and neovascularization within the vascular niche (Fig. 2o and Supplementary Table 9). Using Hotspot²³ on niche-specific COVET embeddings, we identified conserved spatial gene–gene correlations across control and DKD samples (Extended Data Figs. 4–6). Differential correlation analysis revealed strengthened links such as COL1A1–SERPINE1, CXCL12–MHCI and JUNB–CD74 in DKD, consistent with the observed pathway enrichments in the vascular niche (Extended Data Fig. 4b).

Extended Data Fig. 4. Spatially correlated transcripts across kidney niches.

Hotspot was applied to niche-specific latent representations from the COVET analysis to identify spatially correlated genes. For each niche (Vascular, PT, CD, and CNT), heatmaps show pairwise gene–gene correlations (deeper red indicates stronger correlation). Gene pairs conserved across conditions are highlighted (a, c, e, g). Differential correlation analysis identifies interactions significantly increased in DKD (b, d, f, h).

Extended Data Fig. 6. Spatially correlated transcripts across kidney niches.

Hotspot was applied to niche-specific latent representations from the COVET analysis to identify spatially correlated genes. For each niche (iPT, Glomerular, DCT), heatmaps show pairwise gene–gene correlations (deeper red indicates stronger correlation). Gene pairs conserved across conditions are highlighted (a, c, e). Differential correlation analysis identifies interactions significantly increased in DKD (b, d, f). g: overview of two glomeruli in the glomerular niche, podocytes (Podo) indicated in green and glomerular endothelial cells (Glom EC). Scale bar, 250 µm. h-i: Ligand receptor expression in the glomerular niche. Transcripts are colored as indicated; one circle represents one transcripts (10x Genomics Xenium Explorer 3.2.0). Scale bar, 50 µm.

As DKD primarily affects the glomerulus, we investigated changes in cellular composition within the glomerular niche. Mesangial (P < 0.05; Fig. 2p) and immune cells (P < 0.05; Fig. 2q) were modestly increased in DKD, consistent with hallmark features of disease pathology²⁴. Notably, these changes were correlated (r = 0.32, P < 0.05), suggesting coordinated expansion of mesangial and immune populations (Extended Data Fig. 2l). Differential correlation analysis further revealed increased mesangial–immune interactions, including COL3A1–ITGB8, CHI3L1–JUN and IGFBP5–CD24 (Extended Data Fig. 6d). The scale of our dataset enabled analysis of niche-level cell–cell interactions across disease severity. Using CellPhoneDB, we found that NRG1–ERBB3 and CXCL9–DPP4 interactions correlated positively with kidney function, whereas PLAU–PLAUR and IL-6–IL-6R signalling between glomerular endothelial cells and podocytes were enriched in severe DKD (Fig. 2r and Extended Data Fig. 6g–l), highlighting candidate pathways for future mechanistic studies.

In conclusion, spatial transcriptomics revealed kidney niches through complementary annotation- and gene expression-based approaches, recapitulating known structures and uncovering new functional domains. This niche-based framework offers a molecular lens to define disease-associated changes in composition and signalling, and potential therapeutic targets.

The profibrotic microenvironment in DKD

Beyond identifying spatial niches, spatial transcriptomics captures how a cell’s state and gene expression vary with its immediate surroundings—the ‘microenvironment’ (Fig. 3a). Unlike multicellular niches, microenvironments reflect local context-dependent changes within a specific cell type. We focused on injured tubular (iPT and iLOH) and immune cell microenvironments to understand how local interactions contribute to disease progression.

Fig. 3. The profibrotic microenvironment in DKD.

a, Schematic of injured tubular epithelial cell (iTEC; includes iTAL and iPT cells)-centred microenvironment (ME) identification based on neighbourhood fractions across 20 cell types. Each dot represents a single cell, coloured by cell type, with iTEC (yellow) at the centre. Clustering reflects the cellular neighbourhoods surrounding iTECs, defining distinct microenvironments. b, Injured tubular microenvironment fractions across samples, ordered by eGFR. c, Association between profibrotic niche fraction and eGFR (two-sided Pearson test, P = 3 × 10⁻⁷); the black line is the linear regression with 95% confidence interval. d, Injured tubular microenvironment fraction in control versus DKD kidney. e,f, Heat maps of key DEGs in iPT (e) and iTAL (f) centre cells across microenvironment fractions (left, CosMx; right, Xenium); gene label colours indicate platform specificity. g–k, Spatial transcriptomic analysis of representative periglomerular (g), perivascular (h), proximal (i), profibrotic (j) and distal (k) microenvironments (Xenium Explorer v3.2.0). Scale bars, 100 μm. l, Centre cell (pie) and local tissue composition (doughnut) of injured tubular microenvironments. m, imputed DEGs between microenvironments identified by two-sided Wilcoxon test and analysed by pathway over-representation analysis (GSEApy Enrichr; one-sided Fisher’s exact test with Benjamini–Hochberg correction); adjusted P < 0.05. FDR, false discovery rate. n, Representative profibrotic microenvironment region (Xenium 5k map) with adjacent Sirius Red staining. Scale bar, 250 μm. o, Ligand–receptor interactions between iPT cells (sender) and fibroblasts (receiver) inferred using CellPhoneDB v5 (imputed expression). Interaction scores were correlated with HAVCR1 expression in iPT (HBEGF→EGFR: P = 4.7 × 10⁻¹⁸, TNFSF10→TNFRSF10B: P = 6.36 × 10⁻¹⁷) and COL1A1 expression in fibroblasts (JAG2→NOTCH2: P = 5.9 × 10⁻¹¹, PDGFβ→PDGFRβ: 8.2 × 10⁻¹¹) using two-sided Pearson tests; shaded bands denote 95% confidence intervals. Schematic created in BioRender; Dumoulin, B. https://biorender.com/w4x4t61 (2026).

Source data

Using neighbourhood composition, we identified five distinct injured tubular microenvironments, each defined by unique neighbouring cell types and transcriptional states (Fig. 3g–l, Extended Data Fig. 7a–e and Supplementary Tables 12–15). The proximal tubular microenvironment, enriched for PT and DCT neighbours, showed relatively healthy iPT and iTAL centre cells expressing CUBN, APOE, GPX3, EGF and COL6A1. The distal microenvironment, enriched for TAL, DTL/ATL, PC, CNT and IC neighbours, featured UMOD⁺/EGF⁺ iTAL and KRT19⁺/SLC39A8⁺ iPT cells. The periglomerular microenvironment, adjacent to glomerular cells, showed SPOCK2⁺/DUSP1⁺/IGFBP7⁺ iPT and CCN2⁺/BCAM⁺ iTAL cells at the centre, suggesting early stress and matrix remodelling. The perivascular microenvironment, enriched for EC_DVR and VSMC neighbours, displayed SERPINA1⁺ and PFKFB3⁺ iPT and iTAL centre cells, consistent with hypoxia-driven metabolic reprogramming. Finally, the profibrotic microenvironment, rich in immune and fibroblast neighbours, featured MMP7⁺/HAVCR1⁺/HIF1A⁺ iPT and MMP7⁺/ANXA2⁺/CLU⁺/ITGB6⁺ iTAL centre cells, indicating proinflammatory and profibrotic activation consistent with previous reports. Notably, ITGB6 has previously been reported as upregulated in iTAL cells²⁵, and MMP7 has been proposed as both a biomarker of kidney disease and an indicator of renal injury^26,27. Moreover, these findings are consistent with recent reports of HAVCR1⁺ iPT cells localizing near fibrotic regions^28–30.

Extended Data Fig. 7. a–e Differential cell-type abundance across injured tubular microenvironments.

We quantified the relative abundance of selected cell types within injured tubular microenvironments (MEs), excluding cells outside these regions. For each sample, we calculated cell-type fractions across MEs and performed pairwise comparisons using two-sided Mann–Whitney U tests, with Benjamini–Hochberg correction for multiple testing. Inner boxplots indicate the median (center line) and interquartile range (box), with whiskers extending to 1.5x IQR. Points represent individual samples. a: TAL cells were significantly enriched in the Distal ME compared to the Perivascular ME (adjusted p = 2.6 × 10^–15), Profibrotic ME (adjusted p = 3.57 × 10^–20), Proximal ME (adjusted p = 1.05 × 10^–19), and Periglomerular ME (adjusted p = 2.2 × 10^–17). b: PT cells were significantly enriched in the Proximal ME compared to the Perivascular ME (adjusted p = 7.88 × 10^–17), Profibrotic ME (adjusted p = 4.14 × 10^–17), Distal ME (adjusted p = 4.14 × 10^–17), and Periglomerular ME (adjusted p = 4.14 × 10^–17). c: Glomerular cells were significantly enriched in the Periglomerular ME compared to the Perivascular ME (adjusted p = 1.09 × 10^–15), Profibrotic ME (adjusted p = 2.3 × 10^–14), Proximal ME (adjusted p = 7.32 × 10^–15), and Distal ME (adjusted p = 1.11 × 10^–14). d: Fibroblasts were significantly enriched in the Profibrotic ME compared to the Perivascular ME (adjusted p = 0.013), Distal ME (adjusted p = 2.8 × 10^–13), Proximal ME (adjusted p = 7.4 × 10^–19), and Periglomerular ME (adjusted p = 4.3 × 10^–13). e: Vascular cells (EC_Peritub, VSMC, EC_DVR) were significantly enriched in the Perivascular ME compared to the Distal ME (adjusted p = 6.57 × 10^–20), Profibrotic ME (adjusted p = 9.68 × 10^–20), Proximal ME (adjusted p = 9.68 × 10^–20), and Periglomerular ME (adjusted p = 2.32 × 10^–19). COL1A1 expression in fibroblasts across tubular microenvironments. f: Dot plot showing the normalized expression of COL1A1 across tubular microenvironments for fibroblasts profiled with the CosMx 1k platform. Dot size represents the proportion of cells expressing the gene, while color intensity reflects the average log-normalized expression level within each microenvironment. g: Dot plot showing the normalized expression of COL1A1 across tubular microenvironments for fibroblasts profiled with the Xenium 5k platform. Dot size represents the proportion of cells expressing the gene, while color intensity reflects the average log-normalized expression level within each microenvironment. Bulk transcriptomic correlations of receptor–ligand pairs with markers of tubular injury and fibrosis. Scatter plots showing Pearson correlations between select ligand and receptor genes and markers of tubular injury (HAVCR1) and fibrosis (COL1A1) in human kidney tubule bulk RNA-seq data (n = 843). Expression values were log₂-transformed, centered, and min–max scaled. TNFSF10 (h), TNFRSF10B (i), HBEGF (j), and EGFR (k) expression correlate with HAVCR1. PDGFBRB (l), PDGFB (m), NOTCH2 (n), and JAG2 (o) expression correlate with COL1A1. A linear regression line is overlaid, shaded band represents 95% confidence interval of the linear regression. Pearson correlation coefficients and corresponding two-sided p-values are indicated.

Source data

The profibrotic microenvironment was markedly expanded in DKD (Fig. 3d) and negatively correlated with eGFR (r = –0.6, P < 0.001; Fig. 3c). Differential expression analysis of aggregated microenvironment profiles using cell states mapped from snRNA-seq through the scANVI latent space revealed enrichment of inflammatory and fibrotic pathways—including, for example, focal adhesion or IL-7 signalling (Fig. 3m and Supplementary Tables 16–18). Spatial validation using Sirius Red staining on an adjacent section confirmed colocalization with COL1A1 deposition. Consistent with this, fibroblasts in this microenvironment showed higher COL1A1 expression than in other injured tubular microenvironments (Extended Data Fig. 7f,g).

We next analysed ligand–receptor interactions between iPT cells and fibroblasts (Supplementary Table 19) using HAVCR1 as a marker of injury. HBEGF–EGFR and TNFSF10–TNFRSF10B interactions correlated positively with HAVCR1 expression (Fig. 3p and Extended Data Fig. 8a,c), suggesting that stromal cells promote epithelial injury through EGFR and death receptor signalling³¹. In relation to COL1A1, a fibrosis marker, JAG2–NOTCH and PDGFβ–PDGFRβ interactions were linked to matrix production via Notch and PDGF pathways³² (Fig. 3o and Extended Data Fig. 8a,c). Bulk tubulointerstitial RNA-sequencing (RNA-seq) data from 843 kidney samples confirmed positive correlations between these pairs and HAVCR1 or COL1A1, supporting these findings (Extended Data Fig. 7n,o). Together, these results identify a disease-associated profibrotic injured tubular microenvironment that is enriched in fibroblasts and immune cells and marked by proinflammatory and profibrotic signalling.

Extended Data Fig. 8. a Profibrotic injured tubular microenvironment and ligand–receptor interactions.

Overview of profibrotic injured tubular microenvironment, cells are colored in red. Scale bar, 500 µm. b-c: Ligand receptor expression in the profibrotic injured tubular microenvironment niche. Transcripts are colored as indicated; one circle represents one transcripts (10x Genomics Xenium Explorer 3.2.0). Scale bar, 25 µm.

Spatial mapping of the immune landscape

Building on our identification of injured tubular cell states shaped by their microenvironment, we next focused on immune cells—whose functional diversity, spatial organization and transcriptional heterogeneity have historically been difficult to resolve owing to the limited spatial resolution of earlier technologies.

For this purpose, we generated a comprehensive spatially resolved human kidney immune cell atlas. Unbiased clustering identified 11 immune populations, including T cell, natural killer (NK) cell, dendritic cell, B cell, plasma, macrophage, basophil/mast and neutrophil subsets (Fig. 4a, Supplementary Table 3 and Extended Data Fig. 9a,b). Immune cells were increased in DKD, with plasma and B cells showing 3.1-fold and 2.4-fold higher abundance, respectively (Fig. 4b and Extended Data Fig. 9c). Spatial mapping in DKD revealed that whereas macrophages and plasma cells were broadly distributed, CD4⁺ and CD8⁺ T cells, and especially B cells, showed localized clustering; B cells also had the highest proportion of immune neighbours, indicating enrichment within dense immune microenvironments (Fig. 4c and Extended Data Fig. 10c).

Fig. 4. Spatial mapping of the immune landscape.

a, Human kidney immune atlas (UMAP of integrated data coloured by cell type). cDC, conventional dendritic cell; mDC, myeloid dendritic cell; Neutro, neutrophil; pDC, plasmacytoid dendritic cell. b, Increased immune cell populations in DKD versus controls. c, Spatial immune organization in DKD (Xenium Explorer v3.2.0). Scale bars, 500 μm. d, Schematic of immune cell-centred microenvironment identification based on neighbourhood fractions of coarse celltypes. Each dot represents a single cell, coloured by cell type, with immune cells (IC, yellow) at the centre. Clustering reflects the cellular neighbourhoods surrounding ICs, defining distinct microenvironments. e, Immune microenvironment fractions per sample. f–l, Representative fibroblast (f), residential (g), T-predominant (h), vascular (i), glomerular (j), injured tubular (k) and B-predominant (l) immune microenvironments (Xenium Explorer v3.2.0). Scale bars: 100 μm (f–h,j,k), 200 μm (i,l). m–s, Immune and local tissue composition of fibroblast (m), residential (n), T-predominant (o), vascular (p), glomerular (q), injured tubular (r) and B-predominant (s) immune microenvironments. Pies indicate immune fractions, and doughnuts denote neighbouring cell types. t, Immune subtype fractions in glomerular microenvironments (control versus DKD), normalized per sample and compared by two-sided Welch’s t-test (n = 41; samples with fewer than 50 glomerular immune microenvironments were excluded). Box plots show median and interquartile range. u, Ligand–receptor interactions inferred using CellPhoneDB v5 (imputed expression) and correlated with eGFR using two-sided Pearson tests (FN1→ITGA4B7: P = 0.007, BMP4→BMPR2: P = 0.008); shaded bands indicate 95% confidence intervals. MC, mesangial cell. v, Immune subtype fractions in injured tubular microenvironments (control versus DKD), compared by two-sided Welch’s t-test (n = 44). Box plots show median and interquartile range. w, Ligand–receptor interactions (CellPhoneDB v5) correlated with eGFR using two-sided Pearson tests (CXCL12→CXCR4: P = 0.017, ANXA1→FPR3: P = 0.007, TNFSF13→TNFRSF13B: P = 0.003); shaded bands denote 95% confidence intervals. Schematics in u,w created in BioRender; Dumoulin, B. https://biorender.com/yz4gxcl (2026).

Source data

Extended Data Fig. 9. a Immune cell marker genes and residential immune microenvironment composition.

Expression of selected marker genes across annotated immune cell types in the immune cell atlas (CosMx dataset). For each gene, expression was averaged per cell type and scaled to a 0–1 range using MinMax normalization. The color intensity represents the relative expression level of each gene across cell types. Visualization was performed using the Seaborn heatmap function. b: Expression of selected marker genes across annotated immune cell types in the immune cell atlas (Xenium dataset). For each gene, expression was averaged per cell type and scaled to a 0–1 range using MinMax normalization. Visualization was performed using the Seaborn heatmap function. c: Bar plot showing the fraction of immune cells relative to all immune cells in each sample. d: Bar plot showing the fraction of immune cells classified as part of the residential immune microenvironment relative to all immune cells in each sample. e: Stacked bar plot showing the relative proportions of immune cell types within the residential immune microenvironment across all spatial transcriptomic samples (CosMx and Xenium). Each bar represents one sample, and cell type proportions were normalized by the total number of immune cells assigned to the residential microenvironment in that sample. Sample names are abbreviated in the end based on the platform they were run on (C = CosMx, X = Xenium).

Extended Data Fig. 10. a Spatial Immune Composition and gene expression and across Immune Microenvironments.

Visualizations of expressed genes in macrophages between the residential immune microenvironment macrophages and macrophages outside of this microenvironment in the CosMx dataset and b: in the Xenium dataset. Dot size represents the fraction of cells expressing the respective gene, and color intensity indicates the mean normalized expression level. c: Heatmap showing the mean fraction of neighboring immune cells for each annotated immune cell type across control and DKD samples. For each cell, the fraction of neighboring immune cells was computed across four spatial scales (20 µm, 40 µm, 60 µm, and 80 µm), and the mean was calculated per immune subtype. d: Dotplot showing the expression of PDCD1 and CXCR5 in CD4 + T cells of the Xenium dataset across immune microenvironments. Dot size represents the fraction of cells expressing the respective gene, and color intensity indicates the mean normalized expression level. Bulk expression correlations between plasma cell markers, and associated ligand receptor interactions. TPM expression values from bulk RNA-seq from 843 patients were log₂-transformed, centered and min-max scaled. e: Scatter plots shows Pearson correlations between XBP1 as plasma cell markers and interstitial fibrosis as scored by a nephropathologist. f: Scatter plots shows Pearson correlations between XBP1 as plasma cell markers and TNFRSF13B. g: Scatter plots shows Pearson correlations between CXCR4 as plasma cell markers and XBP1. h: Scatter plots shows Pearson correlations between XBP1 as plasma cell markers and CXCL12. i: Scatter plots shows Pearson correlations between XBP1 as plasma cell markers and TNFSF13. j: Scatter plots shows Pearson correlations between XBP1 as plasma cell markers and ANXA1. k: Scatter plots shows Pearson correlations between XBP1 as plasma cell markers and FPR3. l: Scatter plots shows Pearson correlations between the interaction partners TNFSF13 and TNFRSF13B. m: Scatter plots shows Pearson correlations between XBP1 as plasma cell markers and FPR3. n: Scatter plots shows Pearson correlations between the interaction partners CXCR4 and CXCL12. A linear regression line is overlaid, shaded band represents 95% confidence interval of the linear regression. Pearson correlation coefficients and corresponding two-sided p-values are indicated.

Source data

To define the spatial organization of immune cells, we clustered them by their local cellular context (Fig. 4d). This revealed microenvironments enriched in fibroblast, injured tubular, vascular or glomerular neighbours—termed fibro, injured tubular, vascular and glomerular immune microenvironments (Fig. 4f,i–k,m,p–r). We also identified two immune-dense microenvironments: a T cell-predominant niche composed mainly of CD4⁺ and CD8⁺ T cells, and a B cell-predominant niche enriched in B cells with some CD4⁺ T cells (Fig. 4h,l,o,s).

One identified immune microenvironment contained few neighbouring immune or parenchymal cells and was termed the residential kidney immune microenvironment (Fig. 4e,g). It consisted mainly of macrophages with some CD8⁺ T cells. Because tissue-resident immune cells are difficult to identify spatially, previous studies proposed that immune cells in histologically normal regions are likely to represent resident populations³³. Consistent with this idea, in two healthy paediatric kidneys, most immune cells localized to this compartment, which remained proportionally stable and compositionally preserved in DKD (Extended Data Fig. 9d,e and Supplementary Fig. 12a–h).

We next examined macrophages, the main cell type in the residential immune microenvironment (Extended Data Fig. 9e). In both CosMx and Xenium datasets, residential macrophages showed tissue-resident and anti-inflammatory signatures, including MRC1, CD163, ZFP36 (CosMx), MRC1, GPR34, STAB1, ALDOB and RNASE1 (Xenium). By contrast, non-residential macrophages expressed higher levels of inflammatory and antigen-presentation genes such as LYZ, S100A4, S100A6 (CosMx), ITGAX and STAT1 (Xenium) (Extended Data Fig. 10a,b).

Unbiased analysis revealed increased immune infiltration within the glomerular niche in DKD (Fig. 2q). Even healthy kidneys contained a small population of macrophages (Fig. 4t), but in DKD both macrophages and CD8⁺ T cells were significantly increased in number (P < 0.05). Leveraging our large cohort, we analysed ligand–receptor interactions across disease severity (Fig. 4u). The mesangial FN1–macrophage ITGA4 or ITGB7 interaction negatively correlated with eGFR, consistent with fibrotic remodelling and potential homing of macrophages³⁴, whereas the interaction of podocyte BMP4 with BMPR2 on CD8⁺ T cells correlated positively with eGFR, suggesting a potential anti-inflammatory role³⁵.

We next examined the injured tubular immune microenvironment to better understand inflammation linked to tubular injury. Even healthy kidneys contained some injured tubules (Figs. 3b and 4e) consistent with previous observations³⁶, but in DKD this microenvironment showed increased numbers of plasma cells (P = 0.008; Fig. 4v). Ligand–receptor analysis revealed reduced ANXA1–FPR3 signalling, suggesting impaired inflammation resolution, and stronger CXCL12–CXCR4 and TNFSF13–TNFRSF13B interactions, consistent with plasma cell recruitment and survival (Fig. 4w). Bulk RNA-seq from 843 kidneys confirmed positive correlations between these interactions and the plasma cell marker XBP1 (Extended Data Fig. 10e–n).

In summary, we generated a comprehensive, spatially resolved immune cell atlas of the human kidney, integrating cell type identity, gene expression and spatial context across health and disease. This identified a stable tissue-resident immune cell organization dominated by anti-inflammatory macrophages. In DKD, we observed marked increases in B cells and plasma cells forming a highly organized B cell-predominant microenvironment that can be revealed by spatial transcriptomics.

B cell microenvironment marks DKD subset

In DKD, B cell numbers were markedly increased and exhibited the highest degree of spatial organization among immune cell types. In a subset of patients, these B cells further assembled into a distinct, DKD-associated B cell-predominant immune microenvironment.

The B cell-predominant immune microenvironment was embedded within the profibrotic tubular niche (Figs. 3 and 5c), linking B cell accumulation to regions of injury and fibrosis. This region showed strong spatial co-expression of CXCL13–CXCR5 (Fig. 5o), a chemokine axis that organizes B and T follicular helper (T_FH) cells in secondary lymphoid organs, suggesting the presence of tertiary lymphoid structures (TLSs). Supporting this, CD4⁺ T cells expressed CXCR5 and PDCD1, consistent with a T_FH-like phenotype³⁷ (Extended Data Fig. 10d), and TNFSF13B–TNFRSF13C signalling indicated active B cell survival and differentiation (Fig. 5o), implying a TLS-like niche sustaining local B cell responses in DKD.

Fig. 5. B cell microenvironment marks DKD subset.

a–g, Xenium plots showing immune and profibrotic tubular microenvironments (a), including CD8⁺ T cells (b), CD4⁺ T cells (c), B cells (d), plasma cells (e) and B cells expressing IRF4 (f) or AICDA (g). Scale bars, 200 μm. h–n, IMC of CD45 (h), CD8 (i), CD4 (j,n), CD20 (k,m), CD38 (l), Ki-67 (m,n), AQP1 and DAPI. Arrowheads indicate colocalization of Ki-67 and CD20. Scale bars, 200 μm. o, Xenium visualization of CXCL13 and CXCR5 (left) and TNFSF13B and TNFRSF13C over H&E staining (dots represent transcripts). Scale bars, 200 µm. p,q, Integration of single-cell B cell atlas with Xenium using scVI and scANVI. UMAP of integrated B and plasma cells across all patients (p). UMAP restricted to Xenium-derived B and plasma cells from this study (q). Act, activated B cells; Atm, atypical memory B cells; Bm, memory B cells; Bnai, naive B cells; Gcb, germinal centre B cells. r, A B-predominant microenvironment with B cell subsets. Scale bar, 100 µm. s–u, IMC of CD20 with CD27 (s), CD11c (t) and BCL6 (u), plus AQP1 and DAPI. Scale bars, 200 µm. v, The TRIDENT cohort (n = 248) clustered into B− (n = 231) and B+ (n = 17) using a 10-gene B cell signature (Ward’s method, Euclidean distance). Kaplan–Meier analysis of renal end-point (dialysis, transplantation or eGFR decline ≥40%); two-sided log-rank test (P = 0.02). w, Differential expression between B− and B+ DKD analysed with limma. Heat map shows significant DEGs (z-scored). x, IgA immunoprecipitation–mass spectrometry (n = 9 control; n = 9 with B+ gene signature from the Susztaklab Biobank); differential and covariation analysis identifies IgA-interacting proteins. FC, fold change. y, Elastic-net logistic model (SMOTE, fivefold cross-validation) identifies plasma proteins that distinguish B+ (n = 13) and B− (n = 182) DKD in the TRIDENT cohort. Box plots show median and IQR. z, Proteins with non-zero elastic-net coefficients were entered into a Cox model for renal survival; performance is compared to clinical-only models using 1,000 bootstraps in the TRIDENT cohort.

Source data

To validate these findings at the protein level, we performed imaging mass cytometry (IMC) on adjacent tissue, confirming the presence of abundant CD20⁺ B cells and CD4⁺ T cells within this microenvironment, whereas CD8⁺ T cells and plasma cells were rare (Fig. 5a–n). Integration with a single-cell RNA-seq atlas³⁸ identified naive, memory and atypical memory B cells (Fig. 5p–r), with a marked enrichment of IgD⁺ naive B cells (P = 2.38 × 10⁻¹¹; Extended Data Fig. 12). IMC for CD27, CD11c and BCL6 supported this classification (Fig. 5s–u), and Ki-67 staining showed proliferating CD20⁺ B cells but not CD4⁺ T cells (Fig. 5m,n). Finally, Xenium detected expression of AICDA (Fig. 5g; required for class-switch recombination³⁹) and IRF4 (Fig. 5f; essential for plasma cell differentiation⁴⁰), indicating local B cell activation and differentiation into antibody-producing cells within the kidney.

Extended Data Fig. 12. Antibody isotype distribution within spatial immune microenvironments.

B cells were classified into four categories—IgD+, IgM+, IgA/G+, or negative—based on normalized and scaled expression of IGHD, IGHM, IGHA1, IGHG1, and IGHG2, assigning each cell to the dominant isotype expressed. a: Fraction of IgD+ B cells across immune microenvironments. The B cell–predominant immune microenvironment showed a significantly higher proportion of IgD+ B cells compared to all other microenvironments (two-sided Mann–Whitney U test, p = 2.38 × 10^–11). b: Fraction of IgM+ B cells across immune microenvironments. A significantly higher proportion of IgM+ B cells was also observed in the B cell–predominant immune microenvironment (two-sided Mann–Whitney U test, p = 1.06 × 10^–4). c: Fraction of IgA/G+ B cells across immune microenvironments. No significant difference was observed between the B cell–predominant microenvironment and other immune microenvironments (two-sided Mann–Whitney U test).

To assess the effect of intrarenal B cell aggregation and immune complex formation, we derived a spatial gene signature for the B cell-predominant niche and applied it to bulk RNA-seq data from the Susztaklab Biobank to identify B cell-rich samples. IgA immunoprecipitation followed by tandem mass tag-based mass spectrometry revealed strong enrichment of immune effector proteins, including granzymes and complement components (C3, C4A and C8B), which are known contributors to tissue damage in inflamed organs⁴¹ (Fig. 5x).

We next assessed the link between B cell-predominant inflammation and kidney disease severity. Applying our spatial B cell signature to bulk RNA-seq data from 843 kidney samples in the Susztaklab Biobank revealed a strong positive correlation with interstitial fibrosis (Extended Data Fig. 11m). Using the TRIDENT cohort, we stratified patients into B+ DKD (8.4%) and B− DKD (91.6%) groups based on this signature (Supplementary Table 24). Kaplan–Meier analysis showed that patients with B+ DKD reached kidney failure end-points significantly faster over three years (Fig. 5v).

Extended Data Fig. 11. Expression of niche-specific gene signatures across kidney cell types in CosMx spatial transcriptomics data.

Heatmap showing the scaled expression of top differentially expressed genes from 11 spatially defined kidney niches across annotated cell types in CosMx data. For each niche, the top 10 upregulated genes were selected. Gene expression was log-transformed and averaged across cell types, followed by MinMax scaling (0–1) per gene. Correlation of niche signatures with interstitial fibrosis. TPM expression values from bulk RNA-seq from 843 patients were log₂-transformed, centered and min-max scaled. Scatter plots shows Pearson correlations between interstitial fibrosis as scored by a nephropathologist and the 10 most significant genes of the spatial gene signature of the b: CNT niche, c: DCT niche, d: Fibroblast niche, e: Glomerular niche, f: iLOH niche, g: Immune niche, h: iPT niche, i: LOH niche, j: PC niche, k: PT niche, l: Vascular niche. m: B cell predominant microenvironment. A linear regression line is overlaid, shaded band represents 95% confidence interval of the linear regression. Pearson correlation coefficients and corresponding two-sided p-values are indicated.

To identify biomarkers of the B cell-predominant inflammatory subset, we compared bulk RNA-seq profiles of B+ and B− DKD patients. B+ DKD resulted in higher expression of B cell markers (CD19, CD79A and PAX5), TLS-associated chemokine (CXCL13, LTA and LTB) and plasma cell (MZB1, CD38 and PRDM1) genes. We also found upregulation of BAFF/APRIL receptor genes (TNFRSF13C and TNFRSF13B) and T_FH markers (CXCR5, PDCD1 and CD40LG), consistent with the spatial enrichment of these genes in the B cell-predominant microenvironment.

To identify circulating biomarkers that distinguish B+ from B− DKD, we trained an elastic-net logistic regression model on paired plasma proteomic data from TRIDENT, using SMOTE for class balance and fivefold cross-validation. The model selected 14 proteins (Fig. 5y), termed the B+ DKD protein set, which improved kidney outcome prediction beyond clinical features in Cox models (Fig. 5z, Supplementary Table 26 and Supplementary Fig. 13). External validation in 3,309 UK Biobank participants with diabetes confirmed superior performance over the kidney failure risk equation (area under the curve (AUC) = 0.70 versus 0.63; Supplementary Fig. 13a). Patients in the top quartile of the B+ DKD protein score had a substantially higher risk of progression (Supplementary Fig. 14b), suggesting potential clinical utility for identifying candidates for B cell-targeted therapy.

Discussion

In this study, we used Xenium and CosMx to build a single-cell spatial atlas of healthy and DKD human kidneys, enabling spatial mapping of cell types and defining computationally derived, biologically validated kidney niches. In DKD, we identified disease-associated niches enriched in injured tubular and immune cells. High-resolution analysis revealed a profibrotic injured tubular microenvironment with distinct cell states and signalling networks, within which a B cell-predominant microenvironment emerged in a subset of patients. A spatial gene signature of the B cell-predominant microenvironment was associated with rapid disease progression and linked to candidate tissue and plasma biomarkers.

The complex spatial organization of the kidney demands high-resolution mapping, and here we present the most comprehensive single-cell spatial dataset of the human kidney to date⁴². Whereas previous studies were limited by spot-level resolution or small sample size^{25,28,43–47}, our work integrates CosMx 1k and Xenium 5k spatial transcriptomics with single-nucleus RNA-seq, enabling cross-platform, cross-modality analysis. This integration provides a harmonized, multimodal map of human kidney cell types at unprecedented resolution.

We present a roadmap for next-generation molecular cartography using the human kidney as a model. High-resolution single-cell spatial transcriptomics, combined with complementary computational approaches, defined 11 distinct kidney niches—including both canonical and previously underappreciated compartments—consistent across platforms and samples. These niches reveal how disease reshapes tissue architecture by expanding injured and immune compartments while depleting homeostatic ones such as the DCT niche. Anchoring disease-associated signalling networks in spatial context, this framework lays the foundation for a standardized digital kidney atlas to advance cross-study comparison and precision medicine.

We introduced the concept of microenvironments—clusters of a given cell type defined by their local cellular neighbourhoods—to capture how spatial context shapes cellular states. Applying this framework to injured tubular epithelial cells revealed that tubular injury in DKD occurs within five distinct microenvironments, each with unique neighbouring cell types and gene expression profiles. The profibrotic microenvironment emerged as a key disease-associated niche, enriched in fibroblast and immune neighbours, expanded in DKD, and negatively correlated with kidney function. It contained HAVCR1⁺, MMP7⁺ injured PT and ITGB6⁺ LOH cells, both previously implicated in kidney disease^{25,36,48–51} with fibroblasts expressing high levels of COL1A1. Ligand–receptor analysis implicated epithelial–stromal crosstalk via HBEGF–EGFR, TNFSF10–TNFRSF10B, JAG2–NOTCH and PDGFβ–PDGFRβ pathways, identifying this microenvironment as a focal unit of fibrosis and disease progression.

The creation of a high-resolution spatial atlas of human kidney immune cells marks a major advance in renal immunology⁵². Despite recognition of immune involvement in DKD^53–55, a comprehensive map of immune cell types, states and spatial organization has been lacking. Our atlas fills this gap by capturing the full spectrum of immune populations and their spatial context across healthy and diseased kidneys. Integrating spatial transcriptomics with multiplex imaging and bulk multiomics, it provides an unprecedented framework for dissecting renal inflammation at cellular resolution.

One of the most notable findings from our atlas is the marked expansion of B cells and plasma cells in DKD. Although immune dysregulation in kidney disease is well known, B cell organization had not been fully appreciated^56–60. In DKD, plasma cells localized near injured tubules, whereas B cells formed a distinct TLS-like immune microenvironment^60–62 embedded in the profibrotic niche, featuring Ki-67⁺ B cells with AICDA and IRF4 expression, strong CXCL13–CXCR5 and BAFF–BAFFR signalling and CXCR5⁺/PDCD1⁺ T_FH cells. This structure, typical of autoimmune or transplant rejection settings^63,64, contained predominantly memory and IgD⁺ naive B cells. IgA pull-down mass spectrometry from B cell-rich tissue revealed granzymes and complement proteins within immune complexes, suggesting that local antibody production contributes to inflammation and tissue injury. Notably, this B cell-predominant microenvironment defines a new therapeutic axis in DKD. Its spatial gene signature correlated with disease severity, identified a rapidly progressive DKD subtype, and yielded distinct tissue and plasma biomarkers. Although B cell-targeted therapies are established in autoimmunity and cancer, they remain unexplored in DKD. Our findings highlight the need for clinical trials testing such therapies, with biomarker-based stratification to identify patients that are most likely to benefit.

In summary, spatial single-cell transcriptomics transformed kidney anatomy into a computable disease map, revealing how DKD reshapes tissue into profibrotic microenvironments and a B cell-predominant TLS-like niche marking a progressive subset. By compressing these patterns into a tissue signature and plasma panel, we improved detection and risk stratification beyond clinical models. These results offer a blueprint for translating spatial biology into practice, defining actionable microenvironments and biomarkers to guide targeted therapy.

Methods

Sample acquisition and ethics approval

The University of Pennsylvania institutional review board (IRB) approved the collection of human kidney tissue for this study. Left-over kidney samples were irreversibly de-identified, and no personal identifiers were gathered. Therefore, they were exempt from IRB review (category 4). We engaged an external, honest broker who was responsible for clinical data collection without disclosing personally identifiable information. Participants were not compensated. Some samples were collected as part of the Transformative Research in Diabetic Nephropathy (TRIDENT) study, a multicentre observational cohort study that enrolls, follows, and performs multiomics characterization of individuals with DKD. This study was approved by the University of Pennsylvania IRB and by the TRIDENT Steering Committee. Informed consent was obtained from each participant. The present research using the UKBB Resource was approved under Application Number 273810. Participants from the UKBB provided written informed consent allowing the use of their samples and data for medical research purposes.

CosMx sample preparation and data preprocessing

Tissue sections were cut at 5 µm thickness and prepared according to the manufacturer specifications (NanoString Technologies). We used the human universal cell characterization RNA probes, and a 50 additional custom probes for the following genes: ESRRB, SLC12A1, UMOD, CD247, SLC8A1, SNTG1, SLC12A3, TRPM6, ACSL4, SCN2A, SATB2, STOX2, EMCN, MEIS2, SEMA3A, PLVAP, NEGR1, SERPINE1, CSMD1, SLC26A7, SLC22A7, SLC4A9, SLC26A4, CREB5, HAVCR1, REN, AP1S3, LAMA3, NOS1, PAPPA2, SYNPO2, RET, LHX1, SIX2, CITED1, WNT9B, AQP2, SCNN1G, ALDH1A2, CFH, NTRK3, WT1, NPHS2, PTPRQ, CUBN, LRP2, SLC13A3, ACSM2B, SLC4A4, PARD3, XIST and UTY. We used the DAPI, CD298/B2M, CK8/18 and PanCK/CD45 for additional staining per Nanostring protocol. Imaging was performed using configuration A. After imaging was completed, the flowcell was incubated in 100% xylene overnight, the coverslip was removed from the slide with a razor blade—the slide was then stained with haematoxylin and eosin. The expression matrix and metadata from each CosMx run was exported from the AtoMx platform and was converted to a Python object using squidpy. All samples were merged and preprocessed and analysed together using scanpy. Cells with fewer than 30 counts were filtered out.

Xenium sample preparation and data preprocessing

Tissue sections were cut at 5 µm thickness and cut onto a Xenium slide. according to the manufacturer specifications (10X Genomics). We used the human Xenium Prime 5K Human Pan Tissue and Pathways Panel with 100 additional custom probes for the following genes: TPM1, ESRRB, COL6A3, AGR2, SLC26A7, ATP1B1, SLC8A1, ATP6AP2, TAGLN, SPP1, SAT1, MYL9, LDB2, DEFB1, COL1A2, ACTA2, ST6GALNAC3, SLC13A3, SLC12A3, SLC12A1, MGP, IGHG1, FN1, C7, ACSM2B, AIF1, APOE, AQP3, AZGP1, C1QA, C1QB, C1QC, CAV1, PPIA, CD74, CHI3L1, COL1A1, COL6A1, CRYAB, CXCL14, ENO1, HLA-DPA1, HLA-DRA, IFI27, IGHA1, IL32, KLF2, LGALS3, LUM, MMP7, PIGR, S100A2, SLC4A4, SLPI, SOD2, SPINK1, SOX4, SPOCK2, TACSTD2, TM4SF1, TPM2, VIM, ZFP36, AQP2, RNASE1, ALDOB, PCGF6, RHOB, CD81, ASS1, MYL6, COX8A, CTSB, GATM, MT1G, TMSB10, COL3A1, MIF, TPT1, COL6A2, BST2, CLU, APOC1, APOD, PHKG2, RGCC, HLA-DQA2, CORO1A, HSPB1, ADIRF, CKB, HLA-DQB1, COX5B, MT1H, RAMP3, TYROBP, LAMTOR5, ITM2B, UBB and CTSD. Additionally, the same sections were stained according to the Xenium Cell Segmentation workflow for automated morphology-based cell segmentation, and subsequently loaded onto the Xenium Analyzer for in situ transcriptomic analysis. Xenium raw output files were processed using the spatialdata framework (v0.0.14) with the spatialdata-io Xenium plugin. Xenium transcript and segmentation data were loaded from the manufacturer’s output directory using the xenium() function, which parses transcript tables, cell segmentation boundaries and spatial metadata into a structured SpatialData object. The gene expression table was extracted as an AnnData object for downstream single-cell analysis. Cells with fewer than 30 counts were filtered out.

Integration of the single-cell spatial kidney atlas

To integrate snRNA-seq with spatial transcriptomic data, we applied a deep generative modelling framework using scVI and scANVI, as implemented in the scvi-tools Python package (v1.0+). Two parallel integrations were performed: one with CosMx and one with Xenium data. Raw count matrices were used, and cells were annotated with metadata, including technology, project, and sample identifiers. The SCVI.setup_anndata() function was used to register these batch and covariate annotations. The scVI model was initialized with a negative binomial likelihood and trained using 4 hidden layers and a 30-dimensional latent space. We applied early stopping based on validation loss and used a batch size of 2,048. We refined this by training a scANVI model based using the scVI-trained model as a prior and using the annotations from the annotated snRNA-seq reference atlas. scANVI was trained with early stopping, The mean distance between a given CosMx cells and 10 snRNA-seq cells was then calculated in the resulting latent space using k-nearest neighbours. Cells with a high distance were discarded and reintegration was performed. To create a unified reference map across technologies, we then concatenated the filtered CosMx and Xenium objects. scVI was trained using the raw counts as input, with tech as batch key, and proj, orig_ident, nCount_RNA, and percent_mt as additional covariates. The model was trained using 30 latent dimensions, 4 hidden layers, and a negative binomial likelihood, with dropout and early stopping enabled. scANVI was subsequently initialized using the trained scVI model and fine-tuned using coarse annotations. The model was trained for 50 epochs with a learning rate of 0.001. UMAP coordinates were computed based on the scANVI latent space, and Leiden clustering was performed for annotation. Based on the final scanVI latent space, the distance from each spatial cell was calculated to the nearest snRNA-seq cell using a k-nearest neighbour graph based on Euclidean distance in the latent space. The expression of 3,000 highly variable genes of the snRNA-seq dataset was then inferred for each spatial cell by averaging the corresponding gene expression values of its nearest neighbours, with contributions weighted by the squared inverse of their latent space distances, such that closer neighbours contributed more strongly.

Annotation-based niche generation

To quantify cell-type-specific neighbourhood compositions across multiple spatial scales, we computed the number of neighbouring cells of each cell type within 20, 40, 60 and 80 μm radii for each cell in the combined CosMx and Xenium dataset. The input contained spatial connectivity matrices at each distance threshold. For each sample, we extracted subsets of the data and iteratively computed, for every cell type, the total number of neighbours per cell using the respective spatial connectivity matrix. This process yielded four neighbourhood matrices, each representing a different spatial scale. To capture neighbours newly included at increasing distances, we computed differences between consecutive shells (for example, 40 μm minus 20 μm). Cells with zero neighbours across all types were excluded, and the bottom 5% of sparse neighbourhoods (based on total neighbour count at 80 µm) were filtered out to mitigate edge effect. Cell-type-specific neighbour compositions across all four scales were then used as input features to define spatially resolved kidney niches. These features were normalized using StandardScaler, and dimensionality reduction was performed using principal component analysis (PCA). The optimal number of clusters was determined using the elbow method, followed by MiniBatchKMeans clustering to assign niche labels. To assess disease-associated shifts in kidney niche composition, we quantified the relative abundance of annotated niche labels across healthy and DKD samples. For each condition and niche, we calculated the number of contributing cells and normalized this count by the total number of cells from the respective condition to correct for differences in sampling depth. This yielded a normalized abundance matrix for each niche within control and DKD groups. To compare niche representation between groups, we computed the relative increase or depletion in DKD by calculating the ratio of normalized counts (DKD/control), assigning negative values when niches were more abundant in controls. These relative changes were visualized as a heat map.

COVET and Hotspot analysis

We used the covariance environment (COVET) methodology²¹ on all cells in the CosMx dataset with more than 50 neighbours within a 200 µm radius. This approach incorporates the gene–gene covariate structure with spatial location considerations. Inputs to the COVET matrix included: (1) spatially resolved gene expression; (2) 64 highly variable genes (selected via Scanpy’s highly_variable_genes); and (3) the number of spatial neighbours (k = 8). For each cell in the CosMx dataset with at least 50 neighbours within 200 μm, we computed an adjusted covariance matrix across the 64 genes to capture niche-specific co-expression patterns. Pairwise niche distances were then derived using approximate optimal transport (AOT). Based on this distance matrix, we calculated UMAP for visualization after PCA which help to better examine the relationships within the COVET niche representation. For analysis of spatially correlated transcripts, we used Hotspot²³ for each niche, based on the niche-specific latent representations calculated by COVET analysis. Specifically, we first construct a cell–cell similarity graph in which nodes represent cells and edges encode similarity in the COVET latent space. Gene–gene colocalization is then quantified using a local correlation statistic that evaluates whether two genes tend to be expressed in the same regions of this similarity graph. For a pair of genes $\mathit{x}$ and $y$, the local correlation is defined as:

$$H_{\mathit{xy}}=\sum _{(i,j)\in E}{w}_{\mathit{ij}}\left({\tilde{x}}_i{\tilde{y}}_j,|,{\tilde{y}}_i{\tilde{x}}_j\right),$$

where ${\tilde{x}}_i$ and ${\tilde{y}}_i$ denote standardized expression values of genes $x$ and $\mathit{y}$ in cell $i$, respectively, and the sum is taken over all neighbouring cell pairs $\left(i\right|j)$ in the similarity graph with edge weights $w_{\mathit{ij}}$. The difference in gene–gene correlations between DKD and control was calculated by subtracting the control matrix from the DKD matrix. To normalize values across datasets, each correlation matrix (control, DKD and the difference matrix) was scaled to the range [–1, 1] using min–max scaling. Hierarchical clustering was applied to the scaled control matrix to define a consistent gene ordering, which was subsequently used to reorder all matrices for visualization. Clustered heat maps were generated using Seaborn’s clustermap and heat map functions, with selected gene pairs highlighted to visualize conserved and differential co-regulation.

We visualized gene pairs with absolute correlation values above 0.99 in both control and DKD, identifying robust co-expressed gene modules. Additionally, we computed the top 30 gene pairs with the greatest correlation differences between DKD and control and visualized these.

NicheCompass analysis

To infer spatially defined gene programmes and cellular niches, we trained the NicheCompass model on all cells in the CosMx dataset with more than 50 neighbours within a 200 µm radius. The model was configured with raw count data, sample-specific categorical covariates, and a spatial adjacency matrix for edge-aware graph neural network learning. The encoder used a graph attention convolutional layer (gatv2conv), and categorical covariates were injected into the decoder layer. Gene programme activity thresholds were determined using a relative cutoff (active_gp_thresh_ratio=0.01). The model was trained for 50 epochs and 25 additional epochs across all gene programmes, using the following optimization parameters: learning rate (lr = 0.001), edge reconstruction penalty (λ_edge_recon = 5 × 10⁵), gene expression reconstruction penalty (λ_gene_expr_recon = 300), and L1 regularization for novel gene programmes (λ_L1_addon = 30). The batch size for edge sampling was set to 1,900. We computed a neighbourhood graph and UMAP embedding based on the latent representation and performed clustering using the Leiden algorithm. Clusters were manually annotated to define discrete cellular. To identify gene programmes enriched in specific niches, we performed differential gene programme testing across latent clusters using a log Bayes factor threshold of 2. The resulting active gene programmes were visualized via heat maps after min–max normalization, highlighting niche-specific transcriptional signatures.

Injured tubular microenvironment generation

To characterize injured tubular microenvironments, we subset the CosMx and Xenium datasets to include injured proximal and thick ascending limb cells (iPT, iTAL). For each of these cells, we quantified the surrounding neighbourhood composition across four spatial shells (20 μm, 40 μm, 60 μm and 80 μm) and aggregated related cell types (for example, glomerular, collecting duct and endothelial) to reduce dimensionality. The resulting neighbourhood matrices were normalized using StandardScaler, and dimensionality reduction was performed using PCA. We applied MiniBatchKMeans clustering to identify niche subtypes. Clusters were manually annotated based on dominant neighbourhood composition and labelled as profibrotic, perivascular, periglomerular, proximal or distal microenvironments. We then examined the cell type composition of each niche and visualized sample-specific niche distributions. Differential gene expression analysis was performed within each injured tubular subtype (iPT and iTAL), stratified by microenvironment, across both CosMx and Xenium platforms. To quantify disease-associated shifts in injured tubular microenvironments, we compared the frequency of iPT/iTAL microenvironment labels between DKD and control samples. For each sample, the number of cells assigned to each microenvironment was calculated and normalized to the total number of PT, iPT, TAL, or iTAL to correct for sampling differences. We compared normalized microenvironment frequencies between DKD and control samples, computing relative changes as fold increases (DKD/control) or depletions (−control/DKD). The resulting values were visualized in a heat map.

Immune microenvironment generation

To construct spatial neighbourhood profiles centred around immune cells, we used the immune cell type annotation from the kidney immune cell atlas. For each sample in the integrated Xenium and CosMx dataset, we computed cell-type-resolved neighbourhood matrices at four spatial scales (20 µm, 40 µm, 60 µm, 80 µm) using precomputed spatial connectivity graphs. For each radius, we quantified the number of neighbouring cells belonging to each cell type per index cell, resulting in four DataFrames capturing neighbourhood composition. To isolate newly added neighbours at increasing spatial distances, we calculated the difference between consecutive shells (e.g., 40 µm minus 20 µm). Cells with zero neighbours across all types were excluded, and the bottom 5% of sparse neighbourhoods (based on total neighbour count at 80 µm) were filtered out to ensure robust representation. Each DataFrame was then normalized by the total number of neighbours per cell at that distance to yield cell-type-specific neighbour fractions, and column names were suffixed with the corresponding spatial scale. Finally, cells shared across all four scales were retained and the resulting matrices were concatenated into a single comprehensive neighbourhood profile matrix. This matrix served as the input for downstream immune microenvironment classification. Using these precomputed neighbour fraction matrices at 20–80 µm radii, we aggregated cell types into broader tissue compartments, such as glomerular cells, injured tubules, collecting ducts, loop of Henle, endothelium, and healthy tubules. The resulting DataFrame was standardized and subjected to PCA. MiniBatchKMeans clustering was then performed on the PCA-reduced data. The optimal number of clusters was determined using the elbow method. Cells were assigned cluster labels corresponding to immune microenvironment types. To visualize immune cell composition within each immune microenvironment, we grouped cells by immune microenvironment and immune cell type, and counted the number of occurrences to obtain cell type frequencies per microenvironment. To simplify visualization, rarer immune cell types (basophils/mast cells, conventional and plasmacytoid dendritic cells, NK cells, neutrophils) were aggregated into a single “other” category. We then generated pie charts representing the immune composition of each microenvironment using Matplotlib. To visualize the full cellular composition of the 20 µm immune microenvironments, we grouped cells in the 20 µm surrounding each immune microenvironment and cell type, then counted occurrences to calculate the local composition. To reduce complexity and highlight key structures, related cell types were collapsed into broader categories: PT and TAL into “PT_TAL”, iPT and iTAL into “iPT_iTAL”, glomerular cells (Podo, MC1, PEC, EC_glom) into “glom”, and vascular populations (EC_Peritub, EC_DVR, VSMC) into “ec”. Remaining cell types (e.g., DCT, CNT, intercalated cells) were grouped into an “other” category. We then generated pie charts representing the cellular composition of each microenvironment using Matplotlib.

To assess changes in immune cell composition within the glomerular microenvironment, we calculated the frequency of immune cell subtypes by counting the number of cells per immune subtype and normalizing to the total number of cells in the glomerular immune microenvironment in control and DKD samples. Samples with fewer than 50 glomerular cells were excluded. Fractions were centred by subtracting the median of control samples for each immune subtype. Group-wise comparisons between DKD and control were performed using Welch’s t-test. Results were visualized using box plots of centred immune cell fractions. To analyse immune cell composition in the injured tubular immune microenvironment, we calculated the frequency of immune cell subtypes by counting the number of cells per immune subtype and normalizing to the total number of cells in the injured tubular immune microenvironment in control and DKD samples. Immune cell subtype counts were computed per sample and normalized by the total number of cells in the corresponding niche to obtain cell-type fractions. Samples with fewer than 50 total cells in the injured tubular immune niche were excluded. Immune cell fractions were centred by subtracting the median fraction of control samples. Statistical comparisons between DKD and control samples were performed using Welch’s t-test. Centred fractions were visualized using violin and box plots to assess relative enrichment or depletion of specific immune cell types within the tubular immune microenvironment in DKD.

Human kidney immune cell atlas generation

For the integration of the human kidney immune cell atlas, we integrated immune cells from spatial transcriptomic data with immune cells from snRNA-seq, we used a deep generative modelling framework based on scVI and scANVI, implemented in the scvi-tools Python package (v1.0+). Immune cells were first subset separately from the CosMx and Xenium spatial datasets. Each subset was integrated independently with immune cells from the snRNA-seq reference atlas. Raw count matrices were used and additional covariates included technology, project, and sample identifiers were used. The SCVI.setup_anndata() function was used to register batch and covariate annotations. scVI was trained with a negative binomial likelihood, four hidden layers, and a 30-dimensional latent space and early stopping. After training, scANVI was initialized using the scVI model and refined with annotations from the snRNA-seq reference. scANVI was trained with a learning rate of 0.001. We then trained a scANVI model using known annotations and predicted cell identities for Xenium/CosMx cells labelled as ‘Unknown’. To refine predictions, we iteratively reassigned Xenium/CosMx cell types based on their 10 nearest neighbours in the scANVI latent space, using majority voting if their mean distance fell below a defined percentile threshold (initially 50%, increased to 70% and 90% in subsequent iterations). After each round, updated labels were used to retrain the scANVI model, leading to progressively improved annotation accuracy. Following independent integration of CosMx and Xenium immune cells, the resulting AnnData objects were concatenated. A final scVI model was then trained on the combined dataset using raw counts, with tech as the batch key and proj, orig_ident, nCount_RNA, and percent_mt as covariates. scANVI was reinitialized on the combined latent space and refined using coarse annotations. UMAP coordinates were computed from the final scANVI latent space, and Leiden clustering was applied. Expression of immune cell genes was then inferred for each spatial cell by averaging the corresponding gene expression values of its nearest neighbours, with contributions weighted by the squared inverse of their latent space distances, such that closer neighbours contributed more strongly.

B cell atlas generation

Raw count matrices for B cells and plasma cells were provided in Matrix Market Exchange (MTX) format from a recently published single-cell B cell atlas³⁸, along with matching cell-level metadata (CSV) and gene name annotations. The count matrices were loaded using scipy.io.mmread and transposed to ensure cells corresponded to rows and genes to columns. Each count matrix was then combined with its corresponding metadata and gene names to construct a sparse AnnData object using Scanpy (v1.9+). To construct a unified B cell atlas, we first excluded a curated list of blacklisted genes prior to downstream modelling. The dataset was normalized using total count normalization followed by log-transformation. We identified 3,000 highly variable genes using the Seurat flavour in Scanpy with dataid as the batch key. Cells with fewer than 200 counts were removed. The resulting dataset was used as input for the scVI model. We initialized a deep generative model using scVI (v1.0+, scvi-tools) with the raw count layer as input, dataid as batch, and additional covariates including orig_ident, nCount_RNA, and percent_mt. The model used a negative binomial likelihood, three hidden layers, and a 30-dimensional latent space. Training was performed with a learning rate of 0.001, early stopping. For cell type refinement, we used scANVI and coarse labels from the original annotation. The scANVI model was initialized from the trained scVI model with celltype_coarse as the label key and Unknown as the unlabelled category. It was trained for 100 epochs using a learning rate of 0.001, batch size of 256, and n_samples_per_label = 50, with early stopping enabled. Leiden clustering was applied to the scANVI latent space and cell types assigned to known B cell states including BIe, atypical memory B, age-associated B cells, germinal centre B cells and plasma cells. To enable integrated analysis of single-cell and spatial transcriptomic B cell data, we concatenated annotated single-cell RNA-seq data with B and plasma cells extracted from Xenium datasets. To transfer annotations from single-cell RNA-seq to Xenium spatial transcriptomics data, we trained using scVI and scANVI. Annotated single-cell and unannotated Xenium B and plasma cells were concatenated and filtered to include only samples with at least 50 cells. The scVI model was trained using tech as the batch key and proj and orig_ident as categorical covariates, with nCount_RNA and percent_mt as continuous covariates. We then trained a scANVI model using known annotations and predicted cell identities for Xenium cells labelled as Unknown. To refine predictions, we iteratively reassigned Xenium cell types based on their 10 nearest neighbours in the scANVI latent space, using majority voting if their mean distance fell below a defined percentile threshold (initially 50%, increased to 70% and 90% in subsequent iterations). After each round, updated labels were used to retrain the scANVI model, leading to progressively improved annotation accuracy. Following scVI/scANVI integration, we computed UMAP embeddings based on the scANVI latent space, and Leiden clustering was applied. Based on the clustering results and expression profiles, cluster identities were manually curated and mapped to B cell subtypes, including naive B cells, memory B cells, activated B cells, germinal centre B cells, atypical memory and plasma cells.

Imaging mass spectrometry

FFPE tissue sections were first deparaffinized and rehydrated following standard protocols. Antigen retrieval was carried out using citrate buffer (pH 6.0), followed by blocking with 5% BSA. Primary antibodies targeting proteins of interest were conjugated to cadmium (Cd) metal isotopes using the Maxpar Antibody Labeling Kit (Fluidigm) and incubated with the tissue sections overnight at 4 °C. After washing, metal-conjugated secondary antibodies were applied to enhance signal amplification. Nuclear staining was performed using an iridium intercalator. IMC was conducted on a Hyperion Imaging System (Fluidigm) equipped with a 20× objective. Tissue ablation was performed using a laser with 1 μm resolution, and metal isotope signals were quantified using a time-of-flight mass spectrometer.

Differentially expressed genes

Raw counts were normalized, and log-transformed DEGs were calculated using Wilcoxon rank-sum test with the built in scanpy function. DEGs between patient subcluster in the bulk RNA-seq of the TRIDENT cohort were calculated using Limma⁶⁵.

Pathway enrichment analysis using Enrichr

To identify pathways enriched in niche-specific upregulated genes, we performed enrichment analysis using the gseapy Python package (v1.0.6)⁶⁶. Differential gene expression results for each niche were loaded from precomputed.csv files, and low-confidence genes with adjusted P values of zero were replaced with a small non-zero constant (10⁻³¹⁰). Only genes with positive log fold change and an adjusted P value < 0.05 were retained for enrichment. The background gene set was defined as the union of all tested genes across niches. Gene set enrichment was performed using the enrichr function with the WikiPathways 2023 Human gene set collection (min size = 20). For each niche, significantly enriched pathways (P < 0.05) were extracted from the Enrichr results. Enrichment results were visualized using dot plots or stored for downstream comparison.

Pseudobulk differential expression analysis and spatial gene signature derivation

To identify niche-specific gene signatures, we performed pseudobulk differential expression analysis using the CosMx dataset. We aggregated raw gene expression counts by summing expression values across all cells from the same sample and annotated niche, resulting in a pseudobulk expression matrix (cells grouped by sample and niche). Corresponding metadata containing sample and niche labels was generated. The resulting count matrix and metadata were exported and imported into R for differential gene expression analysis. We then used the DESeq2 package to perform differential expression for each annotated niche versus all other niches. For each contrast, we created a binary group label (‘niche’ versus ‘non-niche’) and ran the DESeq2 pipeline with estimateSizeFactors (type = “poscounts”) and DESeq. Results were extracted using the results() function. DEGs were defined as those with an adjusted P value < 0.05 and absolute log₂ fold change >0.5. To quantify significance and ranking, a π-score was computed as the signed −log₁₀ of the adjusted P value. Genes were filtered for adjusted P value < 0.05 and log₂ fold change >0.5. To rank genes by both significance and magnitude of change, we computed a π-score for each gene, defined as the product of the sign of the log₂ fold change and the negative log₁₀ of the adjusted P value. This composite metric prioritizes genes with strong upregulation and high statistical confidence. The top 10 upregulated genes per niche were selected as spatial gene signatures.

Ligand–receptor interactions

To identify microenvironment-specific cell–cell communication signals, we performed the statistical ligand–receptor interaction analysis using CellPhoneDB v5 (ref. ⁶⁷. An AnnData object containing imputed gene expression was used as input, along with metadata specifying cell-type annotations and microenvironment labels. Cells were first filtered to exclude minor cell types comprising less than 1% of cells within each niche. For downstream compatibility with CellPhoneDB, we annotated each cell with its corresponding microenvironment and defined composite labels combining microenvironment and cell type. These metadata were formatted into two CSV files: one mapping barcodes to cell types and another associating each cell with a microenvironment. We used the CellPhoneDB statistical framework with 1,000 iterations and a threshold of 1% gene expression to determine significant ligand–receptor pairs. Interactions were scored, and significance was determined using a P value cutoff of 0.05. The HGNC gene symbol was used for annotation, and the analysis was performed using 8 computational threads. Results were saved for downstream interpretation and visualization. For calculation of ligand–receptor interactions, we used the imputed gene expression. To identify disease-relevant cell–cell interactions associated with kidney function, we used CellPhoneDB statistical output tables. Interaction scores were indexed by ligand–receptor pairs and filtered to include only interactions between specific sender and receiver cell types within the glomerular niche. For each sender–receiver pair (for example, EC_glom to Podo), we calculated Pearson correlations between interaction scores and GFR across control and DKD samples. Only interactions expressed in more than eight samples were retained for robust correlation estimates.

Bulk RNA-seq correlation

Processing was performed as previously described²⁸. In summary, quality control of the sequencing data was assessed using FASTQC. Adaptors and low-quality bases were removed with TrimGalore (v0.4.5). The trimmed FASTQ files were then aligned to the human genome (hg19/GRCh37) using STAR (v2.7.3a) based on Genecode v19 annotations. Gene expression levels were quantified with RSEM by calculating uniquely mapped reads as transcripts per million (TPM). Count matrices were log-transformed, and the indicated genes extracted. Gene expression values were centred and min–max normalized (range: –1 to 1). For each gene, Pearson correlation was computed. Scatterplots with regression lines and correlation coefficients were generated using Seaborn. For each spatial gene signature, the corresponding 10 genes were extracted. Each gene was scaled to a range from −1 to 1, and the scaled values of the 10 genes within each spatial gene signature were summed. The resulting score was then correlated with the fibrosis score of the respective sample using Pearson correlation.

Hierarchical clustering of bulk RNA kidney samples

Hierarchical clustering was conducted on the scaled TPM matrices using 10 genes from the B predominant immune microenvironment gene list as a reference (Supplementary Table 23). The datasets were clustered using Ward’s method with Euclidean distances.

TRIDENT plasma proteomics analysis

We trained an elastic-net logistic regression classifier using tidymodels⁶⁸. Because B cell classes were imbalanced, we applied the synthetic minority oversampling technique (SMOTE) to the training data. Model complexity (α, λ) was tuned via fivefold cross-validation to maximize the area under the receiver operating characteristic curve (AUROC). Predictors with non-zero coefficients in the refit model were then used as covariates in a Cox proportional hazards model of kidney survival, defined as a composite outcome of death, dialysis, or ≥40% decline in eGFR. We compared the C-statistic of a model including clinical characteristics plus these proteins with a clinical-only model. To account for potential overfitting, we performed 1,000 bootstrap resamples and reported the mean C-statistic for each model.

Validation in UK Biobank cohort

The UK Biobank (UKBB) is a large-scale, community-based cohort comprising over 500,000 participants aged 40–69 years at recruitment (2006–2010) from 22 assessment centres across the UK. For validation purposes, we included 3,309 participants who underwent plasma proteomic profiling and were diagnosed with diabetes mellitus¹. This subset is representative of the overall UKBB population. Primary outcomes in the UKBB included end-stage kidney disease (ESKD) (identified via ICD-10 codes N185 and N180), dialysis or kidney transplantation. ESKD diagnoses and dates were determined using first-occurrence records from multiple data sources. eGFR was calculated using the CKD-EPI formula based on serum creatinine and cystatin C measurements obtained from baseline and follow-up records. (Study Application 273810). The predictive performance of the developed models was evaluated using time-dependent receiver operating characteristic curve (tAUC) were calculated to assess the model’s performance at different time points⁶⁹.

Immune complex pull-down, LC–MS acquisition and data analysis

Kidney tissues from nine controls and nine samples with B cell-predominant inflammation from the Susztaklab Biobank were lysed under native conditions to preserve protein–protein interactions. Immunoglobulin and associated proteins were captured using a biotin-conjugated anti-IgA antibody (ThermoFisher, 7102882100) and pulled down with Dynabeads MyOne Streptavidin C1 (ThermoFisher, 65001). The bead-bound complexes were prepared for liquid chromatography–mass spectrometry (LC-MS) analysis following an on-bead digestion workflow: samples were reduced with tris(2-carboxyethyl)phosphine (TCEP), alkylated with iodoacetamide (IAA) and digested overnight with a Trypsin/LysC mix (Promega, V5072). The resulting peptides were then labelled with tandem mass tags (TMTpro; ThermoFisher, A44520 and A52045), multiplexed, and cleaned prior to high pH reversed-phase peptide fractionation to achieve deep proteome coverage. The resulting fractions were analysed on an Orbitrap Eclipse mass spectrometer equipped with FAIMS Duo and real-time search for in-depth quantitative profiling. Raw data were processed using the GFY core pipeline and statistically analysed with the msTrawler framework⁷⁰. Differential expression and covariation analyses were performed to identify patterns of protein interactions, and IgA correlation metrics were combined to generate a ranked list of IgA-interacting proteins.

Statistics and reproducibility

Statistical analyses are described in the Methods and figure legends. All tests were two-sided unless stated otherwise. Differential expression was assessed using Wilcoxon or limma with Benjamini–Hochberg correction. Group comparisons used Welch’s t-test, survival differences were evaluated by log-rank test, and associations were assessed by Pearson correlation with 95% confidence intervals. Pathway enrichment was performed using GSEApy (GSEA or over-representation analysis with Fisher’s exact test and Benjamini–Hochberg correction). No statistical method was used to predetermine sample size. The spatial transcriptomics discovery cohort comprised 64 human kidney samples (48 CosMx, 16 Xenium), and the snRNA-seq cohort included 150 samples. Sample size was determined by tissue availability and feasibility of high-resolution spatial profiling. The study was observational and not randomized. Samples were assigned to groups based on established clinical diagnoses. Investigators were not blinded during analysis; however, spatial transcriptomics data processing were conducted using automated computational pipelines, minimizing subjective bias. Representative histological and spatial images reflect patterns observed in the indicated samples. Biological replicates correspond to patient samples; in cases where samples were profiled twice, measurements were treated as independent technical datapoints. Sample-level comparisons were performed on aggregated per sample values unless otherwise stated. Samples and cells were excluded based on predefined quality control criteria (see Methods). Robustness was supported by cross-platform replication (CosMx and Xenium), integration with snRNA-seq data, external validation cohorts and cross-validation or bootstrap resampling where applicable.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41586-026-10363-4.

Extended data figures and tables

Extended Data Fig. 5. Spatially correlated transcripts across kidney niches.

Hotspot was applied to niche-specific latent representations from the COVET analysis to identify spatially correlated genes. For each niche (Fibroblast, iLOH, Immune, and LOH), heatmaps show pairwise gene–gene correlations (deeper red indicates stronger correlation). Gene pairs conserved across conditions are highlighted (a, c, e, g). Differential correlation analysis identifies interactions significantly increased in DKD (b, d, f, h).

Author contributions

B.D., J.L., A.B., T.N. and D.T., performed experiments. B.D., J.L., K.K., C.L., S.M., L.M., D.H., T.N. and L.P. performed computational analysis. E.H., V.F., V.G.B., H.P., M.A.S., A.A., M.J.M., K.H.K., M.R., F.E.M., A.A.H., M.L., E.J.W., C.A.H. and M.P.C., and the TRIDENT consortium offered experimental and analytical suggestions. K.S. was responsible for the overall design and oversight of the experiments. M.P. performed pathological scoring. K.S. supervised the experiments. B.D. and K.S. wrote the original draft. All authors contributed to and approved the final version of the manuscript. Members of the TRIDENT Consortium contributed to study design discussions, provided experimental and analytical guidance, facilitated cohort oversight and sample acquisition, and contributed to data interpretation.

Peer review

Peer review information

Nature thanks Carlos Talavera-López and the other, anonymous, reviewers for their contribution to the peer review of this work. Peer review reports are available.

Data availability

Raw data, processed data and metadata from snRNA-seq have been deposited previously (https://zenodo.org/records/15007208 (ref. ⁷¹)), as well as bulk sequencing (GSE115098). The single-cell B cell sequencing atlas has been published previously (https://zenodo.org/records/11468564 (ref. ⁷²)). The full spatial data will be made available upon publication (10.5281/zenodo.18674907 (ref. ⁷³)) and the data can be viewed on the http://www.susztaklab.com browser. Source data are provided with this paper.

Code availability

All the codes used for the analysis were deposited on Github: https://github.com/Woopsydaisy/Spatial-Human-Kidney-Map.

Competing interests

The Susztak laboratory is supported by Boehringer Ingelheim, Regeneron, Bayer, GSK, Gilead, Novartis, KKC, Gilead, Genentech, ONO and Novo Nordisk for work that is unrelated to the current manuscript. E.J.W. is a member of the Parker Institute for Cancer Immunotherapy. E.J.W. is an advisor for Arpelos Bio, Arsenal Biosciences, Coherus, Danger Bio, IpiNovyx, New Limit, Marengo, Pluto Immunotherapeutics, Related Sciences, Santa Ana Bio and Synthekine. E.J.W. is a founder of Arpelos Bio, Arsenal Biosciences and Danger Bio, and holds stock in Coherus. The other authors declare no competing interests.

Extended data

is available for this paper at 10.1038/s41586-026-10363-4.

Supplementary information

The online version contains supplementary material available at 10.1038/s41586-026-10363-4.