Spatial analysis reveals cellular microenvironments and mechanisms of inflammation and injury in acute interstitial nephritis

Abstract

Acute interstitial nephritis (AIN) is present in 15-20% of all acute kidney injury cases but lacks proven effective therapies beyond corticosteroids. Using high-resolution imaging mass cytometry and single-cell spatial transcriptomics to analyze human kidney biopsies with AIN, acute tubular injury (ATI), and reference tissue, we show that interferon-γ-induced CXCL9-CXCR3 signaling is markedly increased in AIN, with significantly more predicted interactions relative to ATI, dominated by homotypic T cell-T cell amplification networks concentrated in lymphoid aggregates. Tubular cells, predominantly VCAM1⁺ injured proximal tubules, markedly upregulate complement components, with local enrichment of C3AR1⁺ immune cells, linking tubular injury to immune activation in AIN. Nicotinamide phosphoribosyltransferase (NAMPT) is the strongest spatial correlate of VCAM1⁺ tubular microenvironments, with expression by both injured tubules and surrounding immune cells coordinating metabolic-inflammatory niches. These findings reveal distinct molecular circuits underlying AIN pathogenesis and nominate potential therapeutic targets for improving clinical management and preventing progression to chronic kidney disease.

Spatial analysis of human kidney biopsies reveals that acute interstitial nephritis differs from nonimmune kidney injury, identifying complement-producing injured tubules, T cell signaling and NAMPT as markers of sustained inflammatory niches.

Introduction

Acute kidney injury (AKI) affects 10–15% of hospitalized patients and up to 50% of critically ill patients, with mortality reaching 50% in severe cases requiring dialysis^1–4. Among the various causes of AKI, acute interstitial nephritis (AIN) is present in 15–20% of biopsy-proven cases and carries significant morbidity, with high rates of progression to chronic kidney disease (CKD) despite treatment^5–10. The clinical burden of AIN is increasing, particularly due to the broadening range of medications identified as triggers, most notably immune checkpoint inhibitors^11–13. Treatment remains limited to discontinuing the offending agent and empirical corticosteroid therapy, which fails in 20–40% of patients^14,15. Despite well-characterized histopathology^12,13, the molecular circuits underlying inflammation in AIN remain undefined, limiting targeted therapy development. The therapeutic challenge is compounded by the narrow 7–10 day window for steroid efficacy¹⁰ and significant steroid-associated complications, particularly in cancer patients receiving immune checkpoint inhibitors.

While the immune cell infiltrate in AIN has been carefully described^7,12,16–18, understanding AIN pathogenesis and outcomes requires detailed characterization of the bidirectional interactions between immune cells and tubular epithelium. Upon injury, proximal tubule (PT) cells upregulate KIM-1 (kidney injury molecule-1, encoded by HAVCR1) and, in some cases, VCAM1 (vascular cell adhesion molecule-1). VCAM1 mediates cellular interactions with VLA-4⁺ leukocytes and often marks a failed-repair tubular state that is prominent in CKD biopsies^19–22. Elucidating how distinct tubular injury states participate in the orchestration of immune responses, and how these immune responses accelerate tubular injury, could reveal targetable pathways for interrupting this pathogenic cycle. Spatial proteomics and transcriptomics enable comprehensive tissue mapping at single-cell resolution^22–24, offering opportunities to identify targetable pathways in AIN.

In this work, we apply imaging mass cytometry (IMC, a spatial proteomics technique) and single-cell spatial transcriptomics (ST) to kidney tissues from patients with pathologist-adjudicated AIN, acute tubular injury (ATI, a non-immune-mediated AKI comparator), and reference tissue. We show that VCAM1⁺ injured proximal tubule cells exist within an NAMPT-enriched metabolic-inflammatory niche where they produce complement components that recruit and activate C3AR1⁺ innate immune cells. We further show that an interferon-γ-driven CXCL9–CXCR3 axis forms self-amplifying CD8⁺ T cell networks that distinguish AIN from ATI and concentrate within lymphoid aggregates lacking normal checkpoint regulation. Together, these findings define AIN as a spatially organized, interferon-driven disease and nominate interferon-γ signaling, complement, and NAMPT as candidate therapeutic targets.

Results

Cohort design and single-cell spatial profiling of AIN, ATI, and reference kidney tissue

We performed imaging mass cytometry (IMC) on 106 kidney tissues from patients with pathologist-adjudicated AIN (n = 37), ATI (n = 37), and reference tissue (n = 32) across discovery (Yale, n = 65) and validation (Johns Hopkins, n = 41) cohorts, with single-cell spatial transcriptomics on 20 discovery cohort tissues (eight AIN, seven ATI, five reference) (Supplementary Fig. 1a, b, and Supplementary Data 1). Most patients with AIN received corticosteroid treatment after biopsy (77% in the discovery cohort). Clinical parameters including baseline eGFR, nadir eGFR, and 6-month eGFR change were otherwise comparable between AIN and ATI in both cohorts (Supplementary Fig. 1a–c). Notably, reference tissues from the validation cohort had lower baseline eGFR and higher injury scores compared to discovery reference tissues (Supplementary Data 2 and 3), likely reflecting unrecognized kidney injury.

IMC using 31 antibodies (Supplementary Data 4) generated single-cell spatial proteomic profiles from 2,565,136 cells across all 106 kidney tissues (Fig. 1a and Supplementary Fig. 2a). Unsupervised clustering defined 50 clusters simplified to 18 core cell populations and two mixed cell populations (Fig. 1c, Supplementary Fig. 2b, and Supplementary Data 5), that were reproduced in the validation cohort (Supplementary Fig. 1d). Single-cell ST was performed on 20 tissues from the IMC discovery cohort (eight AIN, seven ATI, five reference), identifying 321,318 cells with transcriptional designation of 26 core cell types and one mixed cell type (Fig. 1b, d, Supplementary Fig. 3a, b, and Supplementary Data 6).

Fig. 1. Multi-modal spatial profiling reveals conserved cellular architecture across imaging and transcriptomic platforms in acute kidney injury.

a Imaging mass cytometry (IMC) workflow. Left to right: H&E of AIN biopsy; IMC segmentation mask (white outline) on pseudo-colored DNA intercalator (blue); pseudo-colored cellular markers (DNA, blue; megalin and aquaporin 1, red; beta-catenin, green; nestin, yellow); final cell-type annotations (colors as in c). Scale bar, 50 µm. Representative of the 106 IMC tissues. b Spatial transcriptomics (ST) workflow, matched AIN tissue from the same patient as in (a). Left to right: H&E; cell segmentation method (interior, pink; boundary, yellow; nuclear, red) on H&E; raw transcript expression on H&E; final cell-type annotations (colors as in d) on H&E. Scale bar, 50 µm. Representative of the 20 ST tissues. c UMAP embedding of 1,331,664 cells from the IMC discovery cohort (n = 65 patients). Numbered populations: (1) PT, proximal tubule; (2) tDL, thin descending limb; (3) TAL, thick ascending limb; (4) DCT, distal convoluted tubule; (5) CNT, connecting tubule; (6) CD, collecting duct; (7) ddTE, dedifferentiated tubular epithelium; (8) EC, endothelial cells; (9) LEC, lymphatic endothelial cells; (10) vSMC, vascular smooth muscle cells; (11) stromal; (12) parietal; (13) glomerular stromal; (14) podocytes; (15–20) immune populations. d UMAP embedding of 321,318 cells from the ST cohort (n = 20 patients), showing concordant cell-type identification across platforms. Annotations as in (c), plus (11) pericyte, (12) fibroblasts, (13) myofibroblasts, (15) JGA, (19) plasma cells, (22) dendritic cells, (23) APC, antigen presenting cells.

Proximal tubule injury states in AIN associate with distinct inflammatory programs

Unsupervised PhenoGraph clustering of 31-marker IMC protein expression data identified four major distinct proximal tubule (PT) cell states, annotated based on marker expression profiles: healthy (PT-H; megalin^highAQP1^high), dedifferentiated (PT-dd; megalin^lowAQP1^lowKIM-1^–VCAM1^–), PT-injured state 1 (PT-INJ1; megalin^lowAQP1^lowKIM-1⁺VCAM1^–) and PT injured state 2 (PT-INJ2; megalin^lowAQP1^lowKIM-1^±VCAM1⁺) (Fig. 2a). Examination of adjacent H&E-stained sections from regions enriched for specific PT subtypes confirmed that PT-H cells appeared normal, whereas PT-INJ1 cells had loss of brush border, cytoplasmic vacuolization, and nuclear drop-out and PT-INJ2 cells exhibited epithelial flattening in dilated tubules with peritubular scarring (Fig. 2b). Composite injury scores by a pathologist blinded to the protein expression data increased progressively across PT-H, PT-dd, and PT-INJ1 (Jonckheere-Terpstra trend test p = 0.009; n = 9 PT-H, n = 6 PT-dd, n = 3 PT-INJ1 regions; Supplementary Fig. 4a and Supplementary Data 7), consistent with these three clusters representing positions along a continuum from healthy to injured. PT-INJ2 cells and associated tubules were identified as morphologically atrophic.

Fig. 2. Tubule injury states exhibit distinct cellular neighborhoods and complement expression patterns.

a Protein marker expression across IMC-defined PT subtypes in cortex (n = 63; AIN = 21, ATI = 21, Ref = 21). PT-Ki67+ excluded for insufficient cell numbers (Supplementary Fig. 2b and Supplementary Data 5). Dot size indicates percentage expressing the marker above the 75th percentile of all tubular epithelium; color, log-transformed min–max-scaled expression. b Morphological correlation of IMC-defined PT states: segmentation masks with PT subtypes (red) and matched H&E (right). Scale bars, 50 µm. Representative of the 106 IMC tissues. c PT proportional shifts across disease states in cortex; height, proportion of total PT cells (densities in d). n = 63 cortex biopsies. d PT subtype densities (cells/mm²): reference (green, n = 21), AIN (red, n = 21), ATI (orange, n = 21) cortex. Kruskal–Wallis with post hoc two-sided Wilcoxon rank-sum tests and Bonferroni correction; *P < 0.05, **P < 0.01, and ***P < 0.001 (exact P in Source Data). Box plots: median (center line), 25th–75th percentiles (box bounds), whiskers to 1.5× IQR; all individual patient data points overlaid (showing minima and maxima). e Cellular composition within 20 µm of PT subtypes in cortex (n = 63). Dot size, percentage of neighbors; color, z-score enrichment versus the mean across subtypes. f Disease-specific PT subtype neighbor enrichment in cortex (n = 63). Only significant interactions shown (q < 0.05, mixed models); dot size, fold change; color, comparison direction. g PT subtype neighbor changes within AIN cortex (PT-INJ1 and PT-INJ2 versus PT-H; mixed models). Red, enriched; blue, reduced; gray, non-significant (q ≥ 0.05). n = 21 AIN. h Top 10 DEGs in PT cells from AIN cortex (n = 7) versus reference (left; n = 2, descriptive only) and versus ATI (right; n = 6; adjusted P < 0.05, Wilcoxon with Bonferroni). i Percent of PT and thick ascending limb (TAL) cell subtypes expressing C3 in cortex (n = 15; AIN = 7, ATI = 6, Ref = 2). Color, mean normalized expression (log1p). TAL stratified by uromodulin: TAL-H (>median UMOD), TAL-dd (≤median). j Percent of PT cells expressing C3 stratified by HAVCR1/VCAM1 expression (n = 15). Color, mean normalized expression (log1p). PT proximal tubule, MP mononuclear phagocyte, EC blood vascular endothelial cell, LEC lymphatic endothelial cell. Source data are provided as a Source Data file.

While overall PT cell density was reduced in AIN, VCAM1⁺ PT-INJ2 cells were specifically enriched (28.9% of all PT cells vs 5.1% in ATI and 4.6% in reference) (Fig. 2c and Supplementary Fig. 4b), translating to 6.5-fold higher PT-INJ2 cells/mm² in AIN compared to ATI (median 267 vs 41 cells/mm², 95% bootstrap CI for the ratio of medians 1.8–11.7; adj. p = 0.013). PT-INJ1 density showed no significant difference between AIN and ATI (Fig. 2d and Supplementary Fig. 4c). The overall reduction in PT density in AIN was primarily driven by a significant reduction in PT-H cells (median 458 vs 1768 cells/mm² in reference, adj. p < 0.001). These relationships were maintained in the validation cohort except for higher numbers of PT-INJ1 cells in reference samples (Supplementary Fig. 4d–f), concordant with the higher baseline injury scores in validation cohort reference tissues (Supplementary Data 2 and 3) and consistent with designation of PT-INJ1 as an injured state of the PT. Together, these findings position PT-INJ1 as a less injured state shared across AKI etiologies and found even in the absence of clinical AKI, with the AIN-versus-ATI difference concentrated in the VCAM1⁺ PT-INJ2 compartment.

IMC-based characterization of immediately adjacent cell neighbors revealed unique neighborhood compositions around distinct PT states. Across all three conditions, PT-H cells had endothelial cells as their most common non-tubular neighbor, whereas PT-dd had more stromal cell neighbors (Fig. 2e). Analysis of the nearest neighbors of injured PT cells in both AIN and ATI showed that PT-INJ2 cells had enrichment of T cell nearest neighbors in AIN, whereas PT-INJ1 cells had mononuclear phagocytes (MPs) as their most prominent neighbors in both conditions, albeit with distinct polarization states (Fig. 2f and Supplementary Fig. 5a). The predilection for T cells as enriched neighbors of PT-INJ2 and MPs as enriched neighbors of PT-INJ1 persisted when PT cells were analyzed within AIN cortex alone (Fig. 2g and Supplementary Fig. 5b).

Given the established roles of KIM1 and VCAM1 in defining injured and failed-repair PT states in prior single-cell studies of human and murine kidney injury^20–22, and our IMC data supporting distinct immune cell associations based on their protein expression, we next analyzed PT cell transcriptional signatures across the three conditions. Unsupervised clustering of pooled PT cells from ST data identified 21 distinct PT cell clusters that could be distilled down to healthy (PT-H, 7 clusters; LRP2^highCUBN^high), dedifferentiated (PT-dd, 4 clusters; LRP2^lowCUBN^low, without HAVCR1, VCAM1, or LCN2 expression), and injured (PT-INJ, 10 clusters; HAVCR1, VCAM1, and/or LCN2 expression) states (Supplementary Data 6 and 8). HAVCR1 and VCAM1 positivity were concentrated in the PT-INJ population (20.4% HAVCR1⁺VCAM1^–; 28.7% HAVCR1^–VCAM1⁺; and 6.7% HAVCR1⁺VCAM1⁺; n = 9650 cells) compared to PT-H (4.2%, 5.2%, and 0.5%; n = 1241 cells) and PT-dd (2.6%, 5.1%, and 0.3%; n = 1952 cells). Within AIN cortex tissues, HAVCR1⁺VCAM1^– cells (representing PT-INJ1 in the IMC data) exhibited 157 differentially expressed genes (DEGs) compared to HAVCR^–VCAM1^– (PT^neg) cells, including upregulation of other injury markers (VIM, C3, and ITGB3) and downregulation of transport genes (Supplementary Fig. 6). VCAM1⁺ PT cells (representing PT-INJ2 in the IMC data) showed 232 DEGs, with prominent upregulation of complement genes (C3, C4A, C4B, and CFB), MHC class II-related gene CD74, and survival pathways (THY1, SERPINA3)(Supplementary Fig. 7). Independent validation of this stepwise injury progression in the KPMP snRNA-seq atlas²² confirmed concordant marker gradients across an independent cohort and technology (Supplementary Fig. 8a, b).

Comparison across disease contexts revealed a robust AIN-enriched inflammatory signature in both HAVCR1⁺VCAM1^– and VCAM1⁺ PT cells, with upregulation of SERPINA3, THY1, C3, ITGB6, and STAT1 relative to matched states in ATI (all log2FC > 1.3, p-adj. < 0.001; Fig. 2h and Supplementary Fig. 6). Notably, this inflammatory program extended into HAVCR1^–VCAM1^– (PT^neg) cells as well, with tubular C3 positivity in PT^neg cells rising from 2% in reference tissue to 30% in AIN (compared to 14% in ATI; Fig. 2j), indicating that PT reprogramming in AIN extends beyond what canonical injury markers alone capture. Beyond this shared inflammatory signature, VCAM1⁺ cells in AIN additionally upregulated genes implicating a metabolic-inflammatory niche, including NAMPT, SOD2, HMOX1, HAVCR2, CXCL12, and CYP24A1 relative to VCAM1⁺ cells in ATI (Supplementary Fig. 6), consistent with oxidative stress adaptation and cytokine-driven reprogramming unique to the AIN VCAM1⁺ compartment. C3 upregulation was a consistent feature across PT injury states in AIN, suggesting tubular cell-expressed complement might help orchestrate immune recruitment.

Local complement expression aligns with disease-specific immune recruitment

Given C3’s prominence as a defining AIN feature, we quantified its cellular sources. Tubular epithelial cells contributed 63.7% of total C3 mRNA in AIN cortex, with immune cells contributing 21.9% (Supplementary Fig. 9), aligning with histologic observations of C3 protein deposition along tubules in AIN^25–28. The proximal tubule was the dominant tubular source of C3 mRNA, accounting for 78.2% of tubular C3 signal across all cortical tissues. AIN samples showed significantly higher C3 expression than ATI or reference across all PT cells (mean normalized expression 0.441 (AIN) vs 0.160 (ATI), Wilcoxon test, p < 0.0001). Within AIN, C3 expression progressively increased from PT^neg to HAVCR1⁺VCAM1^– to VCAM1⁺ cells (Fig. 2i, j), a pattern paralleled by C4A, C4B, and CFB, with C4A and C4B reaching highest expression in VCAM1⁺ cells (both p. adj. < 0.005 vs HAVCR1⁺VCAM1^–; Supplementary Figs. 6 and 7). A similar injury-associated complement signature was evident in the thick ascending limb (TAL). Consistent with recent observations in AKI and CKD, AKI samples showed increased numbers of dedifferentiated TAL cells (TAL-dd) with diminished UMOD expression^22,29,30. In AIN cortex, TAL-dd comprised 87.1% of TAL (versus 7.6% in reference), with C3 expression 23-fold higher than healthy TAL and 3.6-fold higher than TAL-dd in ATI.

At the tissue level, tubular C3 expression correlated with density of immune cells expressing the C3a receptor, C3AR1 (ρ = 0.86, p = 0.014, Fig. 3a). AIN macrophages expressed higher C3AR1 mRNA than those found in ATI samples (mean normalized expression 0.086 vs 0.060, Wilcoxon test, p < 0.0001). Using a 20 μm distance threshold, predicted paracrine C3–C3AR1 interactions were 4.8-fold higher in AIN than reference and 1.5-fold higher than ATI (Fig. 3b), with tubular, immune, and stromal cells as C3⁺ sources and C3AR1⁺ macrophages as the primary targets (Fig. 3c). C3AR1⁺ macrophages adjacent to C3⁺ PT were predominantly CD74⁺LIPA-high pro-inflammatory macrophages (51% of PT-macrophage C3–C3AR1 interactions; Supplementary Fig. 10), consistent with the association of CD206^– MPs with PT-INJ1 in our IMC analysis and suggesting direct complement-mediated recruitment of this subset to tubule interfaces.

Fig. 3. Tubular complement production drives spatially organized macrophage recruitment in AIN.

a Tubular C3 expression versus C3AR1⁺ immune infiltration in AIN cortex (n = 7). All immune cells (red): two-sided Spearman ρ = 0.86, P = 0.014; macrophages (cyan): ρ = 0.79, P = 0.036. Line, linear least-squares fit; band, 95% CI; no multiple-comparison adjustment. b Predicted C3–C3AR1 paracrine interaction density (interactions/mm²) in cortex (cell pairs co-expressing C3 and C3AR1 within 20 µm; autocrine excluded). Mean ± 95% CI with patient values overlaid: reference (green, n = 2, descriptive only), ATI (orange, n = 6), AIN (red, n = 7). c Chord diagram of predicted C3–C3AR1 interactions (≤20 µm) in AIN cortex (n = 7; 2628 interactions). Width indicates interaction frequency; arrows connect C3⁺ senders (right) to C3AR1⁺ receivers (left). PT–macrophage interactions include CD74⁺LIPA-high (50%). d IMC mononuclear phagocyte (MP) densities in discovery cortex tissues: reference (n = 21), AIN (n = 21), ATI (n = 21). Kruskal–Wallis with post hoc two-sided Wilcoxon rank-sum tests and Bonferroni correction; *P < 0.05, **P < 0.01, and ***P < 0.001 (exact P in Source Data). Box plots: median (center line), 25th–75th percentiles (box bounds), whiskers to 1.5× IQR; all individual patient data points overlaid (showing minima and maxima). e Transcriptomic macrophage polarization states in cortex: AIN (red, n = 7), ATI (blue, n = 6), reference (green, n = 2, descriptive only). Y-axis log_₁₀. AIN versus ATI by two-sided Wilcoxon rank-sum test, unadjusted; *P < 0.05, **P < 0.01 (exact P in Source Data). Box plots as in (d). f Representative macrophage spatial distribution in AIN cortex: proximal tubules (black), other tubular cells (dark gray), CXCL9/10-signature pro-inflammatory (red), CD74⁺LIPA-high (blue), other macrophages (yellow), lymphocytes (green), and other cells (light gray). Scale bar, 500 µm. CXCL9/10-signature macrophages were not enriched near tubules (n = 7, paired Wilcoxon V = 13, P = 0.93). g CD206⁻ MP density versus percent recovery of baseline eGFR at 6 months after AKI in AIN cortex (n = 11 steroid-treated AIN patients with baseline, nadir, and 6-month eGFR). Two-sided Spearman ρ = 0.682, adjusted P = 0.0255 (Benjamini–Hochberg). eGFR recovery, % of baseline restored at 6 months. Line, linear fit; band, 95% CI. ST spatial transcriptomics, IMC imaging mass cytometry, Ref reference tissue, PT proximal tubule, MP mononuclear phagocyte. Source data are provided as a Source Data file.

In addition to the CD74⁺LIPA-high proinflammatory MPs enriched adjacent to injured PT cells, a second subset of proinflammatory MPs with a CXCL9/10-expression signature was enriched in AIN (Fig. 3e and Supplementary Fig. 10). Despite having a 9.9-fold higher overall density in AIN versus ATI, these CXCL9/10-signature pro-inflammatory macrophages participated in only 7.9% of predicted PT-macrophage C3–C3AR1 interactions (versus 50% for CD74⁺LIPA-high; Fig. 3c) and showed no preferential accumulation at tubular interfaces (median 7.2 vs 4.4 CXCL9/10-signature macrophages per 1,000 non-PT cortical cells in peri-tubular versus distant zones; n = 7 AIN cortex tissues, paired Wilcoxon V = 13, p = 0.93; Fig. 3f), indicating that CXCL9/10-signature macrophages spatially organize in response to signals that are distinct from the predicted tubular complement:CD74⁺LIPA-high MP interactions.

These transcriptional analyses of macrophage polarization and spatial distribution aligned with mononuclear phagocyte (MP) accumulation and activation states identified in the larger IMC cohort, with tissue MP density of 120 cells/mm² in AIN versus 63 cells/mm² in ATI and 26 cells/mm² in reference tissues (p < 0.05 and p < 0.001, respectively; Fig. 3d and Supplementary Fig. 4c), findings confirmed in the validation cohort (Supplementary Fig. 4g). The disease-level pro-inflammatory macrophage skew seen in ST was reinforced at the tubular interface: PT-INJ1 cells in AIN exhibited a 3.5-fold enrichment of MP neighbors lacking the CD206 alternative-activation marker, while PT-INJ1 cells in ATI showed CD206⁺ MP neighbor enrichment (Fig. 2f, g, and Supplementary Fig. 5a, b). Within AIN, PT-INJ2 (VCAM1⁺) cells were enriched for mixed MP and T cell neighbors rather than isolated macrophages, reflecting tightly apposed macrophages and T cell infiltrates around VCAM1⁺ tubules that precluded phenotypic resolution of individual macrophages by IMC for this injury state. While overall MP density in AIN did not correlate with eGFR recovery at 6 months (n = 15, ρ = 0.443, adj. p = 0.40, Supplementary Fig. 11b), stratification by CD206 revealed that higher CD206^– MP density specifically predicted improved recovery in AIN patients treated with steroids (n = 11, ρ = 0.682, adj. p = 0.0255, Fig. 3g). Notably, the MP–recovery correlation was directionally opposite in AIN versus ATI (AIN ρ = 0.443; ATI ρ = − 0.564; Fisher’s z = 2.83, p = 0.0046, adj. p = 0.02; Cohen’s q = 1.11, Supplementary Fig. 11b).

Layered inflammatory and metabolic microenvironments surround HAVCR1⁺ and VCAM1⁺ proximal tubule cells in AIN

To characterize microenvironments associated with PT injury, we applied SORBET³¹ to identify transcripts within a two-cell radius that were most predictive of PT HAVCR1 and VCAM1 status, complemented by DEG analysis within 50 μm (Fig. 4a and Supplementary Figs. 12–16). In AIN cortex, HAVCR1⁺VCAM1^– PT microenvironments were best predicted by transcripts coordinating acute inflammation: IL7R (SORBET association score 1.00, 1.6-fold upregulation versus PT^neg microenvironment), SOCS3 (score 0.93, 1.3-fold), and CXCL9 (score 0.89, 1.3-fold). VCAM1⁺ microenvironments additionally showed metabolic and stromal remodeling features, with upregulation of NAMPT (score 1.00, 1.9-fold) alongside the matrix and vascular remodeling genes COL4A1 (score 0.60, 1.6-fold), LUM (score 0.40, 1.1-fold), and ENG (score 0.46). Consistent with the IMC finding that VCAM1⁺ cells are closely associated with lymphocyte populations in AIN (Fig. 2f, g), MS4A1 (CD20) and CD8A expression were also highly enriched in and predictive of the VCAM1⁺ microenvironment (Fig. 4a), reflecting the B and CD8⁺ T cell infiltrates in these regions.

Fig. 4. SORBET machine learning-guided spatial analysis reveals hierarchical inflammatory programs in kidney injury microenvironments in AIN cortex tissues.

a PT injury microenvironments in AIN cortex (n = 7). Left: SORBET spatial association scores (≥0.40 shown) and fold-change for genes predictive of PT states (HAVCR1⁻VCAM1⁻ (PT^neg), HAVCR1⁺VCAM1⁻, VCAM1⁺); Middle: gene names; Right: log₂ fold-change from differential expression of neighboring cells within 50 µm of each PT subtype. PT transcripts were excluded from both analyses. DEG significance is shown by formatting of log₂ fold-change (Bonferroni-adjusted): bold with asterisk, P < 0.001; bold, P < 0.01; plain, P < 0.05; not shown if non-significant. Genes ranked by combined SORBET score between PT injury states, with genes scoring highest for (VCAM1⁺ PT-associated genes on top, followed by HAVCR1⁺VCAM1^– and PT^neg). Right-panel dot size, percentage of each grouped cell type expressing the gene; color, normalized log10-scaled expression. b NAMPT expression in cortical PT cells grouped by HAVCR1/VCAM1 expression in AIN (n = 7). NAMPT positivity significantly rises from PT^neg to HAVCR1⁺VCAM1⁻ to VCAM1⁺ PT cells (two-sided Cochran–Armitage trend test, χ² = 177.2, df = 1, exact P = 2.0 × 10⁻⁴⁰; n = 12,843 PT cells). c Local paracrine NAMPT regulation in AIN cortex (n = 7): non-PT cells show increasing NAMPT expression with increasing VCAM1⁺ PT cell neighbors within 50 µm. Dot size indicates the number of non-PT cells per category. Two-sided Cochran–Armitage trend test, χ² = 18.4, df = 1, exact P = 1.8 × 10⁻⁵; n = 3777 non-PT cells across 7 tissues. d Percentage of NAMPT⁺ cells >200 µm from any VCAM1⁺ PT versus whole-tissue VCAM1⁺ PT density (two-sided Pearson r = 0.78, df = 4, P = 0.065, 95% CI −0.09 to 0.98; n = 6). One AIN tissue excluded for <100 cells >200 µm from VCAM1⁺ PT. Line, linear fit; band, 95% CI. PT proximal tubule, DEG differentially expressed gene, ECM extracellular matrix, ncRNA non-coding RNA, vSMC vascular smooth muscle cells, SCT SCTransform. Source data are provided as a Source Data file.

Expression of the NAD+ biosynthetic enzyme and extracellular inflammatory cytokine NAMPT was the strongest predictor of VCAM1⁺ PT cell microenvironments in AIN (Fig. 4a), with stepwise upregulation along the PT injury gradient in virtually all neighbor cell types (microenvironment DEG 1.5 for HAVCR1⁺VCAM1^–, 1.9 for VCAM1⁺). PT cells also demonstrated stepwise NAMPT upregulation rising from 8.6% of PT^neg cells to 14.6% in HAVCR1⁺VCAM1^– and 17.3% in VCAM1⁺ cells (p < 0.0001, Fig. 4b). NAMPT expression correlated with VCAM1⁺ PT cell density at both local (Fig. 4c) and regional (Fig. 4d) scales, suggesting tissue-level field effects beyond the immediate PT microenvironment. CXCL9 emerged as a consistent inflammatory mediator across lymphocyte, myeloid, and stromal populations in AIN, with progressive upregulation in microenvironments along the PT injury gradient (1.3-fold near HAVCR1⁺VCAM1^– PT and 1.6-fold near VCAM1⁺ PT; Fig. 4a). The differential enrichment of T cell signaling transcripts (IL7R, CXCL9) and lymphocyte markers (MS4A1, CD8A) across PT microenvironments led us to examine lymphocyte subsets populating these niches.

Interferon-activated T cells dominate CXCL9–CXCR3 inflammatory networks in AIN

IMC analysis showed that T and B cell populations were elevated in AIN versus ATI and reference tissues (Fig. 5a, Supplementary Figs. 4c, g, and 17, 18). These cells localized around VCAM1⁺ tubules, with PT-INJ2 cells in AIN showing the greatest lymphocyte recruitment: mixed B and T cell aggregates (5.7-fold increased vs PT-INJ2 in ATI, p = 0.001) and CD8⁺ T cells (2.7-fold increased vs PT-INJ2 in ATI, p = 0.002) as immediate neighbors (Fig. 2f, g, and Supplementary Fig. 5a, b). PT-INJ1 cells showed no significant lymphocyte enrichment, suggesting PT VCAM1 expression specifically aligns with lymphocyte accumulation.

Fig. 5. Spatial profiling reveals CXCL9–CXCR3 inflammatory networks and immune cell dynamics in acute kidney injury.

a IMC cortical immune-cell densities (cells/mm²) across 63 biopsies in discovery cohort for CD4⁺, CD8⁺ and CD4^low/⁻CD8^low/⁻ T cells, B/T aggregates, MP/T infiltrates and mast cells: reference (green, n = 21), AIN (red, n = 21), ATI (blue, n = 21). Kruskal–Wallis with post hoc Wilcoxon rank-sum tests and Bonferroni correction; *P < 0.05, **P < 0.01, and ***P < 0.001 (exact P in Source Data). Box plots: median (center), 25th–75th percentiles (box), 1.5× IQR whiskers; all data points overlaid (minima and maxima shown). b ST T-cell populations across 15 cortex tissues (n = 2 Ref, 7 AIN, 6 ATI). Left, total T-cell subset densities (5.6-fold higher in AIN vs ATI); middle, CD4⁺ T-cell subtypes; right, CD8⁺ T-cell subtypes. Y-axis log₁₀. AIN vs ATI by two-sided Wilcoxon rank-sum test, unadjusted; *P < 0.05, **P < 0.01 (exact P in Source Data); reference descriptive only. Box plots as in (a). c Tissue CXCL9 expression versus CXCR3⁺ cell density across all categories (n = 20; two-sided Spearman ρ = 0.73, P = 2.4 × 10⁻⁴). Line, linear fit; band, 95% CI. d Predicted CXCL9–CXCR3 paracrine interaction density (interactions/mm²) in cortex (co-expression within 20 µm; autocrine excluded). Mean ± 95% CI, patient values overlaid: reference (green, n = 2, descriptive only), ATI (orange, n = 6), AIN (red, n = 7). AIN vs ATI by two-sided Wilcoxon rank-sum test: W = 39, P = 0.012, rank-biserial r = 0.86; reference excluded from testing. e Chord diagram of predicted CXCL9–CXCR3 interactions (within 20 µm) in AIN cortex (n = 7; 2335 interactions; 44-fold higher vs ATI). Width indicates interaction frequency; arrows connect CXCL9⁺ senders (right) to CXCR3⁺ receivers (left). f Urinary CXCL9 versus cortex CXCL9⁺ cell density (two-sided Pearson r = 0.637, df = 11, P = 0.019, 95% CI 0.13–0.88) in paired AIN (red, n = 7) and ATI (blue, n = 6); reference excluded. Line, linear fit; band, 95% CI. g Predicted CXCL9–CXCR3 interaction density versus checkpoint-high CD8⁺ TRM presence (n = 20): reference (green, n = 5), AIN (n = 8), ATI (n = 7). Checkpoint-high CD8⁺ TRM absent in all patients exceeding 50 interactions/mm². LOESS fit; x-axis log. MP mononuclear phagocyte, Ref reference tissue, IFN interferon, TRM tissue-resident memory, Tfc follicular cytotoxic T cell. Source data are provided as a Source Data file.

ST analysis confirmed T cells as the dominant immune cell in AIN, with 5.6-fold elevation in median cell density versus ATI (1271 vs 227 cells/mm²) (Fig. 5b and Supplementary Fig. 11a). CD8⁺ T cells comprised 62% of the T cell infiltrate (median 465 cells/mm²). Reporting per-patient median proportions across the seven AIN cortex ST samples, a resting subset predominated (54% of CD8⁺ T cells, Supplementary Fig. 3d), followed by tissue-resident memory (TRM, 9.3%), memory (8.8%), interferon (IFN)-activated (7.7%), and follicular cytotoxic (Tfc, 6.5%) CD8⁺ T cells. Interferon-activated T cell populations showed the greatest density enrichment in AIN versus ATI (12.3-fold for CD4⁺ and 8.1-fold for CD8⁺, p < 0.01). FOXP3⁺ regulatory T cells were also enriched in AIN (8.7-fold higher cell density than ATI, p < 0.01) (Supplementary Fig. 3c). Despite this enrichment in AIN, CD4⁺ regulatory T cells comprised only 3% of total T cells, with TIGIT expression³² detected in only 25% of these, indicating a small regulatory compartment with limited inhibitory receptor expression relative to the large effector T cell infiltrate.

Since urinary CXCL9 is an AIN-specific biomarker³³, we examined cellular CXCL9 sources and their proximity to CXCL9 receptor (CXCR3)-expressing cells. There was a significant correlation between tissue CXCL9 expression and CXCR3⁺ cell density (ρ = 0.732, p = 0.00024, Fig. 5c), with CD8⁺ T cells and macrophages contributing equally to CXCL9 expression. Among CD4⁺ subsets, IFN-activated T cells had the highest CXCR3 positivity (11.9% of cells), while TRM cells were the top CXCR3⁺ CD8⁺ subset (7.5%). Spatially constrained (20 μm) paracrine ligand–receptor analysis revealed a 44-fold enrichment in predicted CXCL9–CXCR3 interactions in AIN versus ATI (median 52.1 vs 1.2 interactions/mm², Wilcoxon p = 0.012, 95% bootstrap CI 1.8– 506.8, Fig. 5d). CD8⁺ T cell-to-CD8⁺ T cell interactions comprising 41.2% of all predicted CXCL9–CXCR3 interactions (Fig. 5e). These were dominated by homotypic amplification loops, with IFN-activated CD8⁺ T cell-to-CD8⁺ T cell interactions representing 49.2% of CD8⁺-CD8⁺ pairs, and follicular cytotoxic (Tfc) pairings representing 19.5%. Predicted CXCL9–CXCR3 interaction density per tissue correlated with tissue HAVCR1⁺VCAM1^– PT density (r = 0.803, p = 0.030), linking tubular injury to the CXCL9–CXCR3 signaling network. Urinary CXCL9 was positively associated with overall CXCL9⁺ cell density (r = 0.637, p = 0.019) rather than interaction density (r = 0.113, p = 0.714, Fig. 5f).

Notably, these homotypic CXCL9–CXCR3 amplification networks occurred in the setting of a profound reduction of checkpoint-high CD8⁺ tissue resident memory (TRM) cells. This subset, characterized by co-expression of multiple inhibitory receptors (LAG3, TIGIT, and PDCD1) with retained cytotoxic capacity (PRF1⁺) but minimal cytotoxic activity (GZMB^–), represents a memory population under multilayered inhibitory control (Supplementary Fig. 3d). These cells were present in all five reference tissues (median 17.0 cells/mm²) but absent in 88% (7/8) of AIN samples (Fisher’s exact test, p = 0.005), and their absence correlated with increased inflammatory signaling intensity (Fig. 5g). Samples lacking checkpoint-high CD8⁺ TRM cells (n = 11: 7 AIN, 4 ATI) showed significantly more predicted CXCL9–CXCR3 signaling interactions (median 10.8 interactions/mm²) compared to samples retaining these cells (median 0 interactions/mm², n = 9: 1 AIN, 3 ATI, 5 Ref, Wilcoxon rank-sum p = 0.05). All patients exceeding 50 CXCL9–CXCR3 predicted interactions/mm² showed complete absence of checkpoint-high CD8⁺ TRM cells (4/4 patients, all AIN). Together with the limited TIGIT expression among the few regulatory T cells present, these observations suggest that AIN inflammation proceeds without the cellular brakes typically restraining T cell-driven signaling. T cell expansion, IFN-activated subset enrichment, CXCL9–CXCR3 interaction density, and checkpoint-high CD8⁺ TRM depletion findings were robust to leave-one-out exclusion of any single AIN cortex patient, with direction and statistical significance (p < 0.05) preserved across all seven iterations of each comparison (Supplementary Data 9). This unopposed interferon-driven inflammatory network, together with the CXCL9/10-signature macrophages whose functional contribution to AIN remained unaccounted for in our PT neighborhood analysis, raised the question of where in the tissue these inflammatory cells and signaling loops spatially concentrate.

Lymphoid aggregates are hubs for CXCL9–CXCR3 signaling in AIN

In light of the predicted spatial predominance of T cell-T cell interactions rather than tubular cell-immune cell interactions underlying CXCL9–CXCR3 signaling, we used criteria adapted from Liu et al.³⁴ for identifying tertiary lymphoid structures (TLS) in our ST samples. Twenty-eight lymphoid aggregate regions were identified across all tissues, with disease-specific enrichment: 21 aggregates in 7/8 AIN patients versus 7 aggregates in 2/7 ATI patients (33.3%), and no aggregates in reference tissues (Fig. 6a). These aggregates contained B cells, T cells, dendritic cells, and plasma cells, with universal expression of TLS-organizing chemokines CCL19 and CCL21, and 57% also expressing the B cell-recruiting chemokine CXCL13. In contrast, checkpoint-high CD8⁺ TRM cells were absent from 27 of 28 lymphoid aggregates (96.4%), while CD4⁺ regulatory T cells were detected in 17 of 28 aggregates (60.7%).

Fig. 6. Lymphoid aggregates spatially compartmentalize inflammatory signaling in acute interstitial nephritis.

a Tissue-wide mapping of lymphoid aggregates across 20 ST biopsies. Reference (green header, n = 5), no aggregates; AIN (red, n = 8), 21 aggregate regions in 7/8 patients (87.5%); ATI (blue, n = 7), 7 aggregates in 2/7 (28.6%). Colored circles, aggregate regions (gradient, summed CXCL9/CXCR3 score; yellow = low, red = high); gray, parenchyma. Numbers, unique patients; M/C, medullary/cortical (REF-2 and ATI-5 had both). b Representative H&E-stained AIN cortex with cell-type segmentation overlay: macrophages (yellow), T cells (blue), B/plasma cells (green), tubular cells (black). Lymphoid aggregate boundary outlined (yellow). Scale bar, 100 µm. c High-magnification predicted ligand–receptor interactions (within 20 µm) within and around a representative lymphoid aggregate (yellow). Symbols: CXCL9⁺ interacting macrophages (dark blue crosses), CXCL9⁺ interacting CD8⁺ T cells (red crosses), CXCR3⁺ interacting CD8⁺ T cells (red circles), C3⁺ interacting tubular cells (cyan circles), C3AR1⁺ interacting immune cells (magenta circles); non-interacting cells, gray. Scale bar, 100 µm. d Predicted interaction densities (interactions/mm²) within versus outside lymphoid aggregates in AIN cortex (n = 7, per patient per compartment). C3–C3AR1 interactions (median 189 outside aggregates vs 166 within); CXCL9–CXCR3 interactions 3.6-fold higher within aggregates (median 121 vs 434). e Fold change in predicted interaction density (aggregate vs non-aggregate) in AIN cortex (n = 7): CXCL9–CXCR3 interactions: 3.6-fold enrichment; C3–C3AR1 interactions: 0.9-fold. Source data are provided as a Source Data file.

Spatial mapping of ligand–receptor proximity (Fig. 6b, c) and quantitative analysis of predicted ligand–receptor interactions revealed that predicted CXCL9–CXCR3 interactions were 3.6-fold enriched within lymphoid aggregates compared to non-aggregate cortex (median 433.8 vs 121.0 interactions/mm², Fig. 6d, e), while predicted C3–C3AR1 interactions showed no enrichment (median 116.0 vs 189.0/mm²). CXCR3⁺ cells were enriched within aggregate regions in all seven AIN cortex patients (median 1.14% vs 0.47% non-aggregate, patient-level paired Wilcoxon p = 0.022), consistent with spatial concentration of CXCR3⁺ T cells at sites of predicted CXCL9–CXCR3 signaling. Parallel enrichment was observed on the ligand side: CXCL9⁺ T cells comprised 34% of CXCL9⁺ cells within aggregates versus 21% outside, while CXCL9⁺ macrophage contribution was stable across compartments (14–17%), indicating that T cells rather than macrophages underlie the aggregate-specific amplification of CXCL9 production. CD8⁺ IFN-activated and follicular cytotoxic T cells dominated predicted CXCL9–CXCR3 interactions within aggregates. Among macrophages, the CXCL9/10-signature pro-inflammatory subset, which was underrepresented in tubular C3–C3AR1 interactions, accounted for 69% of macrophage-driven CXCL9–CXCR3 interactions within aggregates despite comprising only 15% of AIN cortex macrophages overall. Thus, lymphoid aggregates and injured tubular cells emerge as distinct signaling compartments: aggregates as niches for CXCL9–CXCR3-mediated T cell signaling, and tubular interfaces as sites of C3–C3AR1-mediated innate recruitment.

Discussion

Our spatial analysis of 106 kidney biopsies reveals AIN as a spatially organized inflammatory disease driven by distinct cellular circuits that distinguish it from non-immune-mediated kidney injury. Four major discoveries emerge: first, VCAM1⁺ PT cells are central to inflammatory microenvironments by providing local complement production to recruit innate immune cells and by supporting metabolic-inflammatory coupling via NAMPT. Second, interferon-induced programs pervade AIN at multiple levels and are predicted to drive CXCL9 expression, STAT1-associated tubular reprogramming, and preferential expansion of IFN-activated T cell subsets, creating an AIN-defining inflammatory axis absent in ATI. Third, the CXCL9–CXCR3 axis generates homotypic T cell amplification networks dominated by CD8⁺ T cells. Fourth, lymphoid aggregates compartmentalize these CXCL9–CXCR3 circuits within discrete interstitial foci that lack normal checkpoint regulatory elements. These discoveries provide potential therapeutic targets needed to improve patient outcomes given the increasing incidence of drug-induced AIN³⁵.

Our identification of 6.5-fold enrichment of VCAM1⁺ PT cells specifically in AIN extends prior identification of “failed repair” PT cells as drivers of kidney disease progression^20,36. In AIN, VCAM1⁺ PT cells express protective markers (SERPINA3, THY1)³⁷ alongside complement components promoting inflammatory recruitment, contrasting with profibrotic signatures in ATI. Unsupervised subclustering confirmed that HAVCR1 and VCAM1 define independent axes of variation within injured PT cells (Supplementary Fig. 8d), and independent validation in the KPMP snRNA-seq atlas²² reproduced the stepwise injury progression across distinct platforms (Supplementary Fig. 8a–c). The concurrent upregulation of STAT1 alongside other interferon-stimulated transcripts (GBP1, GBP5, IRF1, and CXCL9) in AIN-exposed PT cells is consistent with a tubular interferon response program rather than passive bystander status. STAT1 expression is itself interferon-inducible through a well-characterized autoregulatory feedback loop³⁸.

NAMPT emerged as the strongest discriminator of VCAM1⁺ PT microenvironments in AIN, representing the first demonstration of its role in immune-mediated kidney injury. Both PT and surrounding cells showed stepwise NAMPT upregulation along the injury gradient from PT^neg to HAVCR1⁺VCAM1^– to VCAM1⁺ tubules. NAMPT functions as an intracellular rate-limiting NAD+ biosynthetic enzyme and extracellular inflammatory cytokine (eNAMPT)²⁹. While NAD+ biosynthetic pathway alterations are established features of AKI^39–41, SORBET analysis indicates that enriched NAMPT expression specifically defines the injury-to-atrophy microenvironment, potentially reflecting both tubular metabolic dysfunction with loss of fatty acid oxidation^42,43 and high metabolic demands of infiltrating immune cells.

VCAM1⁺ PT cells in AIN specifically associate with lymphocytes through unique spatial relationships. Within AIN tissues themselves, only VCAM1⁺ PT cells showed enrichment of adjacent CD8⁺ T cell and mixed B and T cell aggregates, whereas KIM-1⁺VCAM1^– (PT-INJ1) cells and healthy tubules showed no lymphocyte enrichment in the same biopsies. This selective lymphocyte tropism suggests that VCAM1⁺ PT cells may specifically express lymphocyte chemotactic or adhesive signals. Melchinger et al. showed VCAM1 overexpression on cultured PT cells increased lymphocyte adhesion while CRISPR-based knockout abolished TNF-α-induced adhesion⁴⁴, providing one mechanism for this T cell interaction effect.

C3 and other complement factor expression by injured tubular cells extends work demonstrating local kidney complement synthesis^45,46 but reveals disease specificity: AIN shows 2-3-fold higher C3 expression across all segments versus ATI, with VCAM1⁺ PT cells showing the highest expression. This hierarchy creates complement gradients orchestrating spatial immune organization, with a strong correlation between tubular C3 and immune infiltration consistent with complement-mediated chemotaxis^47,48. That CD74⁺LIPA-high macrophages are predicted to interact with C3-expressing tubules, while the equally abundant CXCL9/10 macrophage subset are not, suggests that tubular complement does not broadly recruit pro-inflammatory macrophages but rather engages a specific subset at the tubular interface, partitioning innate complement sensing from interferon-driven chemokine production into distinct inflammatory microdomains that together sustain direct tubular damage and broader tissue inflammation.

The marked enrichment of CXCL9–CXCR3 interactions in AIN provides the strongest distinction between primary immune-mediated injury versus secondary immune responses in ATI. Because CXCL9 is a canonical interferon-γ-induced chemokine⁴⁹, the 44-fold enrichment of CXCL9–CXCR3 predicted interactions likely reflects an underlying interferon-γ response program in AIN, consistent with the marked expansion of IFN-activated CD8⁺ and CD4⁺ T cells, CXCL9/10-signature pro-inflammatory macrophages, and interferon-stimulated transcripts (STAT1, GBP1, and GBP5) in injured PT microenvironments. We show that CD8⁺ T cells match macrophages as CXCL9 producers, expanding the prior macrophage-centric paradigm of CXCL9 expression^50,51. CXCL9 signaling in AIN operated across two complementary spatial compartments. SORBET identified CXCL9 among the top transcripts predictive of injured PT microenvironments, establishing a peri-tubular component tied directly to the injured epithelium. At the same time, predicted CXCL9–CXCR3 interaction density was 3.6-fold higher within lymphoid aggregate regions outside the tubular microenvironment, concentrating the CXCL9–CXCR3 axis in interstitial amplification centers. These aggregates, forming in 87.5% of AIN versus 28.6% of ATI cases, may sustain inflammatory responses independent of secondary lymphoid organs, while the peri-tubular CXCL9 component maintains a direct link to the epithelial injury that initiates recruitment.

These lymphoid aggregates share features with tertiary lymphoid structures (TLS), including spatial B and T cell enrichment, canonical organizing chemokines (CCL19/CCL21), and a transcriptional signature consistent with local antibody production. However, they lack classical stromal infrastructure (high endothelial venules, follicular dendritic cells, and fibroblastic reticular cells), with CXCL13 expression instead coming from immune cells themselves, suggesting self-organizing inflammatory circuits. Similar structures have been characterized in other acute inflammatory conditions^52–54, reflecting a continuum of lymphoid organization where structures lacking full stromal specialization may still coordinate pathogenic immune responses. Whether this absence of specialized stromal infrastructure influences aggregate resolution or persistence, and whether specific organizational features predict treatment response or progression to CKD, represents an important area for future investigation.

Multiple regulatory checkpoints that normally suppress inflammation fail in AIN. Despite 8.7-fold higher CD4⁺ regulatory T cells versus ATI, their median proportion among total T cells was only 3%, suggesting inadequate compensation for the massive lymphocyte infiltration. Checkpoint-high CD8⁺ TRM cells were present in all reference kidneys, but nearly absent in AIN. The loss of this checkpoint-high population may remove a brake on tissue inflammation, as exhausted T cells can occupy tissue niches while maintaining functionally restrained states, and in their absence, more active memory populations may contribute to persistent inflammation through unchecked effector responses. The IMC panel did not include LAG3 or TIGIT to confirm this depletion at the protein level, and activation-induced downregulation of checkpoint markers or niche exclusion from the inflamed cortex could contribute to the apparent absence rather than population loss alone. None of the AIN patients with spatial transcriptomic data had ICI-induced AIN (Supplementary Data 10), so this finding is observed in conventional drug-induced AIN rather than reflecting checkpoint inhibitor pharmacology.

Finally, the apparent paradox that macrophage infiltration predicts functional recovery in steroid-treated AIN but progression in ATI correlates with fundamentally different innate programs in response to primary (AIN) versus secondary (ATI) immune injury. Samples were obtained at similar timepoints after AKI initiation (6 days for AIN, 8 for ATI), yet the pro-inflammatory-to-alternatively-activated macrophage ratio was heavily skewed toward alternatively-activated subsets in ATI. This alternatively activated macrophage predominance in ATI is well described in mouse models^21,55 and correlates with eventual matrix deposition and fibrosis that steroids cannot reverse^56–58. Critically, the association between CD206^– macrophage abundance and better functional recovery was observed only in steroid-treated AIN, suggesting that pro-inflammatory macrophage infiltration marks a steroid-responsive state where tubular survival pathways (e.g., THY1, autophagy) remain intact, whereas the alternatively-activated macrophage response in ATI represents a trajectory toward fibrosis, consistent with failure of steroids to improve ATI outcomes. Notably, CD206^– MP density did not differ significantly between disease groups, indicating that the prognostic value of these cells in steroid-treated AIN reflects within-group variation in inflammatory state rather than a disease-specific abundance pattern.

Strengths of our study include a large cohort size (106 tissues), use of two complementary spatial technologies, a validation cohort from a separate institution, and comparison of AIN with both ATI and reference tissues. However, important limitations merit consideration. The cross-sectional design prevents determination of causality; we cannot establish whether VCAM1⁺ PT cells drive inflammation or result from it, nor can we track evolution of cellular states. The 77% steroid exposure in our AIN cohort confounds outcome associations that are further limited by modest sample size (n = 11). Limited reference cortex samples for ST (n = 2) preclude formal disease-versus-reference expression comparisons, although AIN-versus-ATI comparisons are unaffected. Heterogeneous etiologies of AIN cases (drug-induced, checkpoint inhibitor-associated, idiopathic) allow analysis of shared pathophysiology but prevent evaluation of subtype-specific mechanisms. Our ligand–receptor analyses identify spatially co-localized cells expressing the relevant transcripts and predict probable signaling interactions, but detailed analyses of protein expression, activation, and cleavage states will be required to prove that these interactions actually occur.

Despite these limitations, multiple pharmacologically targetable pathways that may be clinically relevant in AIN emerge from our findings. These include JAK-STAT inhibition (via agents such as baricitinib or ruxolitinib that are already in clinical use for other IFN-driven conditions^59–62), direct IFN-γ neutralization (emapalumab, approved for hemophagocytic lymphohistiocytosis⁶³), complement inhibition^64,65, NAD+ pathway augmentation (recently shown safe and beneficial with nicotinamide for AKI prevention in cardiac surgery^42,66), and eNAMPT inhibition⁶⁷. Because AIN encompasses drug-induced, checkpoint inhibitor-associated, and idiopathic forms with distinct antigenic triggers and effector profiles, the relative contribution of these pathways is likely to differ across subtypes, and subtype-stratified evaluation will be required before clinical translation. Future priorities include direct functional experiments to validate one or more of these candidate therapeutic pathways, determining how distinct AIN triggers impact these pathogenic mechanisms, longitudinal sampling to capture disease evolution and identify how and when reversible inflammation transitions to autonomous injury, and ultimately clinical trials stratifying patients by pathway biomarkers rather than clinical AIN diagnosis alone.

Methods

Study approval and participants

This research complies with all relevant ethical regulations. For the discovery cohort, we included participants of the Yale kidney biobank enrolled between 2015 and 2018, which prospectively enrolled participants who underwent clinically indicated kidney biopsies at Yale New Haven Hospital and St. Raphael’s Hospital (both in New Haven, Connecticut, USA). This cohort has been described previously^33,68. The study excluded participants undergoing biopsy for evaluation of renal mass and allograft biopsies. All participants provided written informed consent, and this study was approved by the Yale Human Investigation Committee (Protocol 11110009286).

The validation cohort of this study was approved by the Johns Hopkins Institutional Review Board under two protocols: IRB00221958 for the overall project and IRB00090103 for the use of FFPE tissue samples. These were archived, de-identified clinical pathology FFPE specimens, for which the Board granted a waiver of informed consent. Identical definitions and criteria were used to obtain clinical data. All data are reported in de-identified form, and no individual participant is identifiable; patient ages are presented in banded ranges (Supplementary Data 11).

Establishing diagnosis and clinical cohorts

In the discovery cohort, three renal pathologists independently evaluated biopsy slides or whole slide scans from the Yale Kidney Biobank to establish AIN diagnosis for this study. The pathologists were blinded to clinical history and official biopsy report. For this study, we selected all cases that were unanimously adjudicated as AIN (n = 32) or ATI (n = 34), of which 23 participants with AIN and 23 with ATI had FFPE archival tissue available for analysis and were included in the discovery cohort.

For the validation cohort from Johns Hopkins University School of Medicine, cases with top-line diagnoses of AIN (n = 15) and ATI (n = 16) were identified from archival tissues spanning 2012–2024 and screened for both tissue adequacy and clinical diagnosis of AKI. These samples were analyzed using identical pathologist adjudication criteria. Identical definitions and criteria were used to obtain clinical data.

Reference tissues

To accurately characterize the importance of findings in injured kidneys and as a quality control assessment of expected signals of resident cell antibodies for each batch, we sought to identify histopathologically normal kidney tissues for inclusion on each slide⁶⁹. Reference tissues consisted of previously prospectively accrued living donor samples⁷⁰, obtained by back-bench biopsy at the time of organ donation, as well as highly selected prospectively obtained and banked samples of nephrectomy tissues taken from tumor-remote regions. Each kidney biopsy used as reference tissue was scored by a pathologist at ×100 magnification using a raster grid consisting of a matrix of 10 rows and 10 columns of equal-sized squares (100 total squares per grid). Squares overlying the renal cortex were counted for histological features of interstitial fibrosis and tubular atrophy (IFTA), interstitial cell infiltrate (ICI), and acute tubular injury (ATI). Tubular injury was defined by the presence of three features: loss of brush border, tubular lumen dilatation, and apical blebbing. Squares containing each type of injury were divided by the total number of squares reviewed and converted into a scoring scheme as follows: none (<5% of area involved), mild (6–25%), moderate (26–50%), and severe (>50%). The overwhelming majority of reference samples used scored as “less than five percent” in all categories and were therefore included in our analysis (Supplementary Data 2).

For the validation cohort, reference tissues were obtained from tumor-remote regions of nephrectomy specimens performed in 2024 and underwent the same histopathological scoring protocol used for the Yale cohort to ensure minimal baseline pathology (Supplementary Data 3).

Clinical data collection

In the discovery cohort, clinical data were collected in the parent protocol through review of individual patient chart review using the Epic electronic health record system and stored in secured web-based data entry system (Supplementary Fig. 1a, b). We collected over 200 variables including demographics, medical history, medications, vital signs, and laboratory data. Sex and age for all participants in both the discovery and validation cohorts are provided in Supplementary Data 11. Baseline serum creatinine values were defined based on two physician chart review of serum creatinine trends over the 24-month period prior to onset of the AKI event prompting kidney biopsy. Where only one value was available, this value was used as baseline. Baseline serum creatinine values were unable to be established in 8 of the 43 patients with AKI in the discovery cohort and 2 of the 31 patients with AKI in the validation cohort due to a lack of serum creatinine measurements within the defined period prior to AKI.

Onset of AKI was defined as an absolute increase in serum creatinine of 0.3 mg/dL or more (within 48 h if occurred during an inpatient stay, no time limitation if outpatient), or a 50% or greater increase in serum creatinine (within 7 days if occurred during an inpatient stay, no time limitation if outpatient). Follow-up serum creatinine measurements at 6 months were taken from values obtained from 4 to 40 weeks after the date of biopsy, with the value closest to the 6-month time point being the value that was used. There was no significant difference between the length of time from the biopsy to follow-up serum creatinine measurement between the two groups (discovery cohort mean 173 days for AIN, 162 days for ATI, p = 0.74; validation cohort mean 163 days for AIN, mean 155 days for ATI, p = 0.67). Recovery of kidney function was calculated as the percentage of baseline estimated glomerular filtration rate restored at 6 months using the CKD-Epi equation. Serologic and urinary data associated with individual biopsies were available through the Yale Kidney Biobank for the discovery cohort.

Urinary CXCL9 measurement

Urinary CXCL9 levels were measured in AIN participants from the Yale Kidney Biobank using the Olink Explore HT proximity extension assay platform (Olink Proteomics, Uppsala, Sweden), which quantifies approximately 5400 proteins by DNA-tagged antibody pairs. Urine samples were obtained at the time of biopsy through the Yale Kidney Biobank and processed according to standard Olink protocols. CXCL9 values are reported as normalized protein expression (NPX) units on a log₂ scale as provided by the Olink platform.

Sample size determination and blinding

Our power calculation indicated that we needed 29 cases with AIN to detect a 1.75-fold difference in cell numbers between AIN and ATI with at least 90% power with alpha of 0.05. Sample sizes were determined based on tissue availability and quality requirements for spatial analysis, with power calculations indicating >80% power to detect 2-fold differences in cell densities between groups. No statistical methods were used to predetermine sample size for the discovery cohort. Three samples in the discovery cohort (1 AIN, 2 ATI) were not included in the analysis due to missing data from technical disruptions in CyTOF software causing loss of data (Supplementary Data 2), one sample was excluded in the validation cohort due to accidental duplication of patient sample and tissue region (Supplementary Data 3), and three reference tissues and one ATI tissue are missing data from ST analysis due to inadequate tissue fixation to the slide and loss of tissue during Xenium processing (Supplementary Data 1). Analysts performing antibody hybridization, IMC analysis, and initial clustering were blinded to diagnosis of AIN vs. ATI until after completion of cell phenotyping. For spatial transcriptomics analysis, investigators were not blinded, but cell segmentation and initial clustering took place on the entire cell population aggregate, without respect to tissue type.

Imaging mass cytometry

Antibody selection, validation, and conjugation

The antibody panel used included metal-conjugated antibodies previously validated for IMC use in kidney tissue using our protocol^69,70, as well as antibodies against new targets (Supplementary Data 4). New targets were selected against protein products of cell-type-specific differentially expressed genes and of signaling, state, and injury pathways from published single-cell RNA sequencing datasets. Only BSA-free and carrier-free commercial antibodies against selected targets were suitable for inclusion to be compatible with heavy metal conjugation, and monoclonal antibodies with supplier knock-out validation were given preference in our testing.

Antibody validation and IMC immunolabeling were performed as previously described⁶⁹. Antibodies were validated using tissues in which target antigen would be expected to be found. For example, antibodies against immune cells were validated in a combination of human tonsil and chronic severe interstitial nephritis tissues. Antibodies against kidney-resident cells were validated in normal kidney tissues. Antibodies in which positive staining was detected in expected cells with a maximal signal-to-noise ratio >3 and in which significant ectopic or background staining was not observed in immunofluorescence staining underwent heavy metal conjugation. These antibodies were conjugated using the Maxpar conjugation kit (Standard BioTools) following the manufacturer’s protocol, scaled for 100 μg of carrier-free antibody. Metals were selected based on expected strength and to minimize overlap due to isotopic impurity. Newly conjugated antibodies were validated using IMC along with the full panel of previously validated antibodies, to ensure both that binding specificity had not changed through the conjugation process, and that the addition to the panel did not interfere with the signal of previously validated antibodies. The final panel consisted of 31 antibodies targeting structural markers, immune cells, injury markers, and functional states.

Slide generation and antibody hybridization

Each slide contained one AIN, one ATI, one reference healthy kidney tissue, and one lymph node, which collectively constituted one “batch” (Supplementary Data 2,3). While including one healthy reference kidney tissue on each slide allowed for a quality assessment of antibody targets of resident kidney cells, inclusion of a lymph node on each slide allowed for a quality assessment of immune cell markers.

For each biopsy, the first 3-μm thick section was stained for H&E, and an adjacent 5-μm thick unstained section was used for IMC. H&E slides underwent Aperio scanning for quality control of each biopsy, and the images were analyzed on Aperio ImageScope (v12.4.6). Unstained tissue slides allocated for IMC were labeled using a cocktail of metal-conjugated antibodies within 2 weeks of tissue fixation to slide to minimize tissue oxidation. For IMC, the unstained tissue sections were deparaffinized, rehydrated, antigen retrieved at pH 9, and then hybridized with a cocktail of metal-conjugated antibodies. Staining was performed within 2 weeks of sectioning to minimize tissue oxidation. By design, all tissues on a given slide underwent antibody hybridization and IMC ablation at the same time. After antibody hybridization, tissue slides were then ablated and imaged using the ND:YAG 213 nm laser in the Hyperion™ Imaging System (Standard BioTools) with 1-μm pixel size, with each aerosolized plume analyzed in a Standard BioTools CyTOF mass cytometer, within 4 days of antibody hybridization to minimize deterioration of antibody signal and tissue oxidation.

For the validation cohort, tissue sections from FFPE tissue blocks were processed at John Hopkins University (under supervision of author AR), and shipped overnight to our site for further IMC analysis as described and underwent antibody hybridization and IMC using the same hybridized antibodies, identical protocols and tools, and by the same author (MLB) prior to undergoing IMC at the CyTOF Core at Yale.

Image acquisition

The stained sections were ablated using the ND:YAG 213 nm laser in the Hyperion™ Imaging System (Standard BioTools) with 1 μm² resolution, and the released metals were analyzed in a tandem CyTOF mass cytometer (Standard BioTools). Regions of interest for ablation were selected to span the largest tissue areas possible and to ablate cortex and medulla separately, covering the entire AIN and ATI tissue biopsy. Larger regions of interest were ablated for all reference tissues where additional tissue was available. For tumor-remote nephrectomy reference samples reconstructed into tissue microarrays and for lymph node samples, a region of interest of 1500 μm by 700 μm was selected.

Image quality control and generation

Images were initially viewed, and antibody hybridization was validated using MCD Viewer (Standard BioTools, v1.0.560.6). For all images shown, thresholding was performed in MCD Viewer with gamma set to 1, minimum color range set to 0, and maximum color range set to 100. The threshold minimum was set within the range of 0–2 to remove background signal, as assessed morphologically. For all comparative images, signal thresholds (minimum and maximum) were set identically for each marker between samples to allow for unbiased comparison. Analyses of cortex and medulla were performed separately, with the corticomedullary junction being analyzed with cortical tissues. Cortex and medulla were defined morphologically.

Pre-processing

IMC images were pre-processed using a Python-based algorithm, “Steinbock” as described⁷¹. Pre-processing included configuring the antibody panel and a hot pixel-filtering step for raw image files. Single-cell segmentation was performed using custom-optimized Mesmer⁷² as described below. Generated segmentation masks were used within Steinbock to measure cell intensities, cell morphology, cell neighbors, and ultimately export single-cell data for downstream analysis.

Object segmentation

Cell segmentation was augmented by implementing the Maxpar® IMC™ Cell Segmentation kit in the antibody panel and pixel expansion method on Mesmer⁷², a deep-learning-enabled segmentation algorithm. While trained on a dataset containing more than 1 million manually labeled cells, the Mesmer algorithm default parameters were modified for optimal cell segmentation of kidney tissue. Specifically, the modifications included resolution from 1 to 0.8, small object threshold modified from 15 to 53, maxima threshold modified from 0.1 to 0.3, and interior threshold modified from 0.2 to 0.47. Cell segmentation was performed by the expansion method by expanding by 1 µm around the nucleus. Performance of the optimized cell segmentation method was validated using “the-segmentation-game” plugin on napari, a fast, interactive viewer for multi-dimensional images⁷³. F1-score⁷⁴ was determined on 15 different tissue samples for which segmentation masks were generated using custom modified Mesmer parameters: 3 chronic kidney disease samples, 3 acute kidney injury samples, 3 diabetic kidney disease samples, 3 tumor remote nephrectomy reference samples, and 3 living donor biopsy reference samples. Across all these samples, the mean F1-score was 0.838 with a standard deviation of 0.0274 (Supplementary Data 12).

Signal spillover correction

While the use of mass spectroscopy of rare earth metals conjugated to antibodies to distinguish unique antibody binding greatly decreases the extent of crosstalk between channels relative to dye-based methods, spillover still exists in mass cytometry data. We employed a previously described bead-based compensation workflow and R-based software⁷⁵ to estimate and compensate for interference between channels.

Image and cell-level quality control and batch effect correction

All downstream analysis was performed within the imcRtools workflow⁷¹. We performed image-level quality control through calculating the signal-to-noise ratio for individual channels and found that each of our channels had an acceptable ratio >3 using the Otsu thresholding approach. Batch effect correction was performed between different batches (slides) using the fast mutual nearest neighbors function within the Batchelor package⁷⁶.

Cell phenotyping

The resultant low-dimensional embedding of integrated cells after batch effect correction was used for unsupervised cell phenotyping. We employed the RPhenoGraph clustering approach, which includes calculation of a graph by detecting the K nearest neighbors based on Euclidean distance in expression space, weighting of edges between nodes by overlap in nearest neighbor sets using the Jaccard index, and Louvain modularity optimization to detect connected communities and partition the graph into clusters of cells⁷⁷. Initial unsupervised clustering results were analyzed through mapping back onto original tissues to assess cell locations, as well as analysis of marker expression using violin plots of raw expression patterns. Subclustering was attempted on all unsupervised clusters using all markers appearing within the cluster to be expressed more highly than the average across all unsupervised clusters. Ultimate cell annotations were validated through marker expression patterns and spatial localization, identifying 50 detailed cell populations (including mixed populations, Supplementary Data 5), which were distilled to 18 core cell populations and two core mixed cell populations (Fig. 1c). All core cell populations identifiable through our RPhenoGraph clustering and subclustering approach were reproduced in the validation cohort (Supplementary Fig. 1d).

Pathologist H&E scoring of PT cluster-enriched regions

To histologically validate the IMC-defined PT subtypes, a renal pathologist blinded to cluster identity scored adjacent H&E sections at tissue regions enriched for a single PT subtype. For each PT cell cluster, regions were selected from the IMC image where multiple PT cells across multiple tubular cross-sections were algorithmically assigned to that cluster; the corresponding field was then cropped from the adjacent 3-μm H&E-stained section at ×40 magnification. A total of 21 cluster-enriched regions were scored, with each PT subtype represented by at least three regions where tissue permitted. The pathologist scored each region on six features of acute tubular injury (brush border shedding, apical blebbing, nuclear dropout, tubular dilation, epithelial flattening, epithelial shedding; 0 = absent, 1 = present), with a composite injury score defined as the sum of these features (maximum = 6). Individual scores are provided in Supplementary Data 7. Because PT-H, PT-dd, and PT-INJ1 represent positions along an ordered injury continuum, scores across these three subtypes were compared using a Jonckheere-Terpstra trend test; a pairwise comparison of PT-INJ1 vs PT-H was performed using a one-sided Wilcoxon rank-sum test. Feature-level positivity across the three subtypes was compared using the Cochran–Armitage trend test. Two of three PT-INJ2 regions were independently characterized by the pathologist as chronic tubular atrophy with peritubular fibrosis and were excluded from acute injury scoring.

Spatial neighborhood analysis

Cell–cell interactions allow cells to exist in communities, perform collective functions, and respond to tissue-level stimuli in a coordinated fashion. Using a k-nearest neighbors-based approach calibrated to capture only the direct, immediate neighboring cells of each cell across different kidney tissues, we were able to define the frequency at which defined interstitial cell types were in immediate spatial proximity to defined tubular cell types. We were interested in capturing each cell’s immediately adjacent neighboring cells, the cells most likely to be directly interacting with a given cell. Using two living donor biopsy tissues, one AKI biopsy tissue, and one CKD biopsy tissue, we manually visually counted the number of immediately adjacent neighboring cells for a subset of 100 random cells per tissue. Across 400 cells and three tissue types, we found that individual cells in our kidney tissues have five immediately adjacent neighboring cells, on average, with a mean centroid-to-centroid distance of 16 µm. Using a k-nearest neighbors approach and a threshold of 5, we generated a spatial interaction graph which was used to determine the identity of each cell’s immediately adjacent neighbors across the entire project. For tubular-interstitial interaction analysis, we used tubular epithelial clusters as index cell types and immune, endothelial, vascular, and stromal clusters as neighboring cell types.

Statistical analysis of cell densities and spatial neighbors for IMC data

Cell type densities (cells/mm²) in IMC data were calculated per patient by dividing total cell count by tissue area, yielding one density value per cell type per patient. Tissue area was measured manually using Fiji/ImageJ. For Supplementary Figs. 17 and 18 (immune cell fractions), patient-level proportions were compared across AIN, ATI, and reference groups using pairwise Wilcoxon rank-sum tests without multiple-testing correction, as these analyses were presented as descriptive companions to the density-based comparisons in Fig. 5a and Supplementary Fig. 4c, g.

For analysis testing the association of diagnosis with the number of index and surrounding cell pairs, we allowed each type of cell to be surrounded by up to 5 nearest neighbors. To test the association of each diagnosis with index cell-surrounding cell pair, we fit a generalized linear mixed model with the outcome as the number of index cell surrounding cell pairs and the predictor as diagnosis using a log link. We clustered the analysis at the level of participant. We present fold difference in the number of cells between various diagnoses (β coefficients) as well as multiple comparison-adjusted q-values using the Benjamini–Hochberg procedure (Fig. 2f and Supplementary Fig. 5a).

To examine whether different PT injury states within AIN tissues exhibited distinct patterns of immune cell recruitment, we analyzed the immediate cell neighbors of PT-H, PT-INJ1, and PT-INJ2 cells exclusively within AIN cortex samples (Fig. 2g and Supplementary Fig. 5b). For each PT cell, we identified up to 5 nearest neighbors. We fit a generalized linear mixed model with a log link for each surrounding cell type, with the number of index-surrounding cell pairs as the outcome and PT cell type as the predictor, using PT-H as the reference category. These analyses were conducted in Stata Statistical Software: Release 18, College Station, TX: StataCorp LLC.

Spatial transcriptomics

Tissue cohort and gene panel

Single-cell spatial transcriptomic analysis was performed on 20 patients (eight with AIN, seven with ATI, five with reference tissue) included in the Yale cohort IMC analysis. With individual patient tissues including cortex, medulla, or both, our cohort included 15 unique cortex tissue samples (seven AIN, six ATI, two reference) and seven unique medulla tissue samples (one AIN, two ATI, four reference) (Supplementary Data 1). 10× Genomics Xenium Analyzer was chosen as the platform to perform spatial transcriptomic analysis because of its single-cell resolution, relatively high sensitivity without sacrificing specificity in FFPE tissues^78,79, and multimodal cell segmentation algorithm. The analysis employed the Xenium Prime 5K Human Pan Tissue & Pathways Panel (5001 genes) plus a custom 100-gene panel (Supplementary Data 13) targeting kidney-resident cells, immune cells, and cell state-specific markers. The custom panel included genes targeting kidney tubular segment markers (SLC12A1, SLC12A3, AQP2, SLC34A1, and SLC22A6), immune-related markers (PTPRC, CD74, CD69, and CCL2), injury markers (LCN2, SPP1), proliferation markers (PCLAF, PRC1), and fibrosis markers (COL1A1, COL1A2, VIM, and TAGLN).

Cell segmentation

Cell segmentation was performed using Xenium’s multimodal algorithm, which employs deep learning to analyze nuclear, cell interior, and cell boundary staining patterns. To validate segmentation accuracy and spatial fidelity, we compared the ST results with post-workflow H&E staining and cell segmentation target images of the same tissue sections. Visual inspection using Xenium Explorer confirmed accurate transcript localization for cell-specific markers (e.g., LRP2 in PT cells) and proper cell boundary identification across all samples.

Data integration and clustering

Single-cell spatial transcriptomics data were processed using Seurat v5. Following quality control to remove cells expressing fewer than 10 unique genes, we created a unified Seurat object containing patient identifiers, region, and category information. After quality control, our dataset had a median of 139 (IQR: 53–374) transcripts per cell and a median cell diameter of 8.33 μm. Data layers were split by patient groups to enable sample-specific processing before integration. Expression matrices underwent SCTransform normalization, which uses regularized negative binomial regression to simultaneously normalize the data and identify variable features while accounting for technical variation in sequencing depth. Gene expression values were scaled to standardize the mean expression and variance across cells. To address batch effects while preserving biological variation, we applied Harmony integration through Seurat’s IntegrateLayers function, creating a shared embedding space for cross-patient comparison. Shared nearest neighbor graphs were constructed using the first 30 Harmony dimensions, followed by Louvain clustering. Optimal clustering resolution was determined by evaluating silhouette scores across multiple parameters (0.4-1.2), with resolution 0.8 providing the highest score and optimal biological separation. Cell type annotation combined differential gene expression analysis with spatial localization patterns and histological correlation, particularly for thick ascending limb and parietal cells, resulting in identification of 26 core cell types and one mixed population (Supplementary Data 6).

Proximal tubule subclustering

Given the centrality of PT injury states to the study, PT cells underwent dedicated subclustering. All cells annotated as PT in the primary clustering step were extracted, re-normalized using SCTransform, batch-corrected using Harmony, and reclustered using the shared nearest-neighbor graph approach with Louvain community detection, identical to the parameters used for primary clustering. This yielded 21 PT subclusters that were collapsed into three major states based on canonical marker expression and established literature on kidney tubular injury: healthy (PT-H, 7 subclusters; LRP2^highCUBN^high, preserved apical transport machinery, minimal injury marker expression), dedifferentiated (PT-dd, 4 subclusters; LRP2^lowCUBN^low with low expression of both canonical PT transport genes and injury markers HAVCR1, VCAM1, and LCN2), and injured (PT-INJ, 10 subclusters; HAVCR1, VCAM1, and/or LCN2 expression; Supplementary Data 8). To parallel the IMC-based PT state definitions (which, in part, used KIM-1 and VCAM1 protein expression to define PT-INJ1 and PT-INJ2), cortical PT cells from AIN biopsies were further stratified by HAVCR1 and VCAM1 mRNA expression status (expression >0 defining positivity), yielding four cross-classified groups (PT^neg, HAVCR1⁺VCAM1⁻, HAVCR1⁻VCAM1⁺, and HAVCR1⁺VCAM1⁺). This marker-based stratification was used alongside the cluster-based classification because cluster assignments and marker status describe overlapping but non-identical facets of injury: within each unsupervised subcluster, cells showed variable HAVCR1 and VCAM1 expression, and marker-positive cells were also present at low frequency within PT-H and PT-dd clusters.

Immune cell subclustering

To capture cellular heterogeneity, we performed targeted subclustering of T cells (Supplementary Fig. 3c, d) using lineage markers (CD4, CD8A, FOXP3, CCR7, and CD69), inhibitory receptors (LAG3, TIGIT, and PDCD1), cytotoxic effectors (PRF1, GZMB), and interferon-response transcripts (STAT1, GBP5); and macrophages (Supplementary Fig. 10) using pro-inflammatory/interferon-response markers (STAT1, IRF1, CXCL9, CXCL10, CXCL11, GBP1, and GBP5), antiviral interferon-stimulated genes (MX1, IFIT1, IFIT3, and OAS1), antigen processing and catabolic markers (CD74, LIPA), lipid-associated macrophage markers (SPP1, APOE, and TREM2), tissue-resident markers (LYVE1, FOLR2, MAF, CD163, MRC1, LILRB5, and CD209), matrix and fibrosis-associated transcripts (COL4A1, COL1A1, COL1A2, LUM, IGFBP7, CXCL12, THBS1, HIF1A, LDHA, and CSF1R), and proliferation markers (MKI67, PCNA, TOP2A, and CCNB1). Subset annotations were based on marker co-expression patterns.

KPMP snRNA-seq validation of PT injury states

To independently validate the PT injury progression observed in our Xenium data, we analyzed single-nucleus RNA sequencing data from the Kidney Precision Medicine Project (KPMP) human kidney atlas Lake et al.²². The full annotated dataset was obtained from the CZ CELLxGENE Discover collection for this study (collection ID bcb61471-2a44-4d00-a0af-ff085512674c; https://cellxgene.cziscience.com/collections/bcb61471-2a44-4d00-a0af-ff085512674c) as an AnnData (.h5ad) object and processed in R using the anndata and reticulate packages. Analyses were restricted to cells annotated as PT (75,222 cells total) across four Lake et al. annotation categories: PT-S1/2 (healthy proximal convoluted/straight tubule), PT-S3 (healthy PT), aPT (adaptive PT, representing cells with injury-associated transcriptional changes), and dPT (degenerative PT, representing cells with a dedifferentiated injury profile).

We performed two complementary analyses using this dataset. First, we quantified the percentage of PT cells positive for HAVCR1 and VCAM1 (expression >0 defining positivity) within each of Lake et al.’s annotation categories to test whether these canonical injury markers track with the authors’ independently defined injury spectrum (Supplementary Fig. 8c). Second, to parallel the stratification scheme applied to our Xenium data, we cross-classified KPMP PT cells along two axes: (i) the Lake et al. annotation label, collapsed into “healthy” (PT-S1/2 + PT-S3) or “injured” (aPT + dPT); and (ii) HAVCR1/VCAM1 expression status using the same thresholds as for Xenium (>0 for positivity). This yielded four mutually exclusive groups paralleling the Xenium stratification: healthy-annotated HAVCR1⁻VCAM1⁻ (n = 19,820), healthy-annotated HAVCR1⁺VCAM1⁻ (n = 1402), injured-annotated HAVCR1⁺VCAM1⁻ (n = 4593), and injured-annotated HAVCR1⁺VCAM1⁺ (n = 3546).

Stepwise injury marker analysis (“staircase” plots; Supplementary Fig. 8a, b)

Within each of the four cross-classified groups (in both Xenium and KPMP datasets), we calculated the percentage of cells expressing each of a panel of injury-associated transcripts (C3, VIM, STAT1, NAMPT, THY1, and CD74) and differentiation markers (CUBN, LRP2), with a cell scored as positive for a given gene if its normalized expression exceeded zero. For Xenium, the four groups comprised: PT-H HAVCR1⁻VCAM1⁻ (n = 1130); PT-H HAVCR1⁺VCAM1⁻ (n = 46); PT-INJ HAVCR1⁺VCAM1⁻ (n = 1315); and PT-INJ HAVCR1⁺VCAM1⁺ (n = 650). This dual-axis design evaluates whether injury markers escalate both across independently annotated PT categories (PT-H → PT-INJ, or PT-S1/2 + S3 → aPT + dPT) and within them as cells acquire HAVCR1 and VCAM1 positivity.

Independent axes analysis (“dumbbell” plot; Supplementary Fig. 8d)

To test whether HAVCR1 and VCAM1 define independent biological axes within the injured PT compartment of our Xenium data, we computed Spearman rank correlations between the percent-positivity of each marker gene and the percent-positivity of HAVCR1 or VCAM1 across the 10 unsupervised PT-INJ-dominant Xenium subclusters (Supplementary Data 8). Separate correlation coefficients were calculated for HAVCR1 and VCAM1, with significance assessed at α = 0.05. Markers tracking preferentially with one marker but not the other (e.g., VIM, THY1, ITGB3 with VCAM1; CUBN, ICAM1 with HAVCR1) indicated dissociable biological programs along each axis.

Statistical analysis of cell densities and spatial neighbors for spatial transcriptomic data

Cell type densities (cells/mm²) in spatial transcriptomic data were calculated per tissue section by dividing total cell count by tissue area. Tissue area was measured using the shoelace formula after boundary delineation in R Shiny v1.7.4 with Plotly v4.10.1, with cell inclusion determined by a ray-casting algorithm. Each density value represented one tissue section from one patient. T cell subset proportions were calculated as the percentage of each subtype within CD4⁺ or CD8⁺ T cells for each patient separately, and the per-patient median was reported across AIN cortex ST samples (n = 7). Cell type annotations were validated by comparing densities between ST and IMC data from 20 overlapping patient samples (Supplementary Fig. 19). Equivalent cell types were mapped between platforms, encompassing tubular epithelial, endothelial, stromal, and immune cell populations. Cell type proportions were consistent between IMC and ST platforms.

Predicted ligand–receptor interaction analysis

To validate molecular communication mechanisms identified by IMC, we performed spatially constrained ligand–receptor interaction analysis on ST data. We analyzed two key ligand–receptor pairs based on differential expression results: CXCL9–CXCR3 and C3–C3AR1. Cells were classified as ligand- or receptor-positive based on detectable expression (>0 normalized counts) of respective genes. We quantified spatial interactions by calculating Euclidean distances between all ligand-expressing and receptor-expressing cells using transformed spatial coordinates. Cell pairs within 20 μm were considered interacting, consistent with our k-nearest neighbor analysis showing mean intercellular distances of 16 μm of immediately adjacent neighbors. Possible autocrine interactions, or expression of receptor and ligand within the same cell, were not included in these analyses.

Interaction densities (interactions/mm²) were calculated by normalizing total interactions to tissue area. For each ligand–receptor pair and disease category, we characterized interaction composition by identifying participating cell types, classified as “ligand_cell_type → receptor_cell_type.” Mean minimum distances were calculated as the average distance from each ligand-positive cell to its nearest receptor-positive cell, regardless of whether that nearest cell was within the 20 μm interaction threshold, providing a measure of overall spatial clustering between cell types. Statistical comparisons used Poisson tests for interaction counts with patient-level clustering.

Lymphoid aggregate region identification

Following established criteria³⁴, we identified organized lymphoid aggregates through a two-step process. First, we detected B cell aggregates (≥10 B cells within 100 μm radius) using spatial clustering with the DBSCAN algorithm. Second, we refined aggregate centers using density-based analysis to identify the highest concentration of B cells and plasma cells within a 250 μm search radius. For each refined center, we evaluated cellular composition within a 150 μm radius, classifying regions as organized lymphoid aggregates if they contained: (1) B cells (≥10), (2) T cells (≥5), (3) dendritic cells (≥2) or plasma cells (≥5), and (4) expression of lymphoid organizing chemokines (CXCL13, CCL19, or CCL21). This density-based refinement ensured accurate capture of lymphoid aggregate cores while maintaining standardized criteria for comparison across samples.

To characterize the organizational features of these aggregates, we assessed markers of specialized lymphoid stromal cells found in classical tertiary lymphoid structures. Specifically, we evaluated: high endothelial venule markers (GLYCAM1, MADCAM1), follicular dendritic cell markers (CR2), and fibroblastic reticular cell markers (co-expression of PDPN and CCL19). We did not identify cells expressing these specialized markers within the aggregates. Additionally, we quantified cellular sources of CXCL13 expression and found it was predominantly derived from immune cells (B cells: 289/793, T cells: 257/793 CXCL13⁺ cells) with minimal stromal contribution (fibroblasts: 13/793). Despite lacking these specialized stromal components typically associated with mature tertiary lymphoid structures, these aggregates met organizational criteria used in recent studies, including Liu et al.³⁴, who similarly characterized structures lacking classical FDCs as TLS in nasopharyngeal carcinoma.

Proportion comparisons across spatial compartments

To test whether specific cell populations were enriched within defined spatial compartments, we computed per-patient densities in paired compartments and compared them using patient-level paired Wilcoxon signed-rank tests. Two spatial compartment definitions were used. For peri-tubular analyses (e.g., CXCL9/10-signature macrophages, Fig. 3f), a peri-tubular compartment was defined as all non-PT cells within 20 μm of any PT cell, using the same 20 μm Euclidean-distance threshold applied in ligand–receptor analyses. Distance from each non-PT cell to its nearest PT cell was computed by k-nearest-neighbor lookup on transformed spatial coordinates (R package FNN, v.1.1.4). The complementary “distant” compartment comprised all remaining non-PT cortical cells. Cell type densities were expressed as cells per 1000 local non-PT cortical cells, computed separately per patient per compartment. For lymphoid aggregate analyses (e.g., CXCR3⁺ cells, Fig. 6d), aggregate regions were defined by the 150 μm-radius circles identified through the aggregate identification pipeline described above; the non-aggregate compartment comprised all remaining cortical tissue, and cell type proportions were expressed as percentages of local cortical cells per patient per compartment. Paired patient-level values were compared using Wilcoxon signed-rank tests. Only patients with at least one cell of the relevant cell type in both compartments contributed to the analysis.

SORBET analysis

We adapted Spatial Omics Framework for Biomarker Extraction and Tissue classification (SORBET)³¹—a supervised, geometric deep learning model—to predict PT cell state from ST data. SORBET requires two inputs: tissue samples profiled with spatial ‘omics technologies and a label (PT class) associated with each subregion (subgraph). Analyses were restricted to the renal cortex and only considered the AIN and ATI pathologies. AIN and ATI were modeled separately.

PT cells were classified as PT^neg (HAVCR1^–VCAM1^–), HAVCR1⁺VCAM1^– PT, or VCAM1⁺ PT. A cell was considered positive for HAVCR1 or VCAM1 if its expression of the specified marker exceeded zero.

Each tissue sample was modeled as a graph $G_K=\left(V_K,E_K\right)$, where $V_K$ is the set of cells and $E_K$ a set of edges encoding spatial adjacency. A single distance threshold was set for each graph to identify edges. Briefly, let ${\mathrm{d}}_{10}\left(c_i\right)$ denote the distance from $c_i$ to its 10th nearest neighbor. Then the edge $\left(c_i,c_j\right)\in E_K$ if $\mathrm{d}\left(c_i,c_j\right)\le \mathrm{Median}\left(\left\{{\mathrm{d}}_{10}\left(c_i\right)\mid c_i\in V_k\right\}\right)$.

Next, cell niches were extracted around each PT cell; the niche was labeled according to that PT cell’s classification. SORBET’s automated subgraph extraction algorithm was to extract niches with the parameters $k=2$ (niche radius/hop-distance) and $\mathrm{MS}=5$ (minimum subgraph size). To avoid confounding, all PT cells and their incident edges were excluded from extracted cell niches.

SORBET’s graph convolutional network (GCN)-based model was modified to account for multi-class prediction. Rather than a sigmoid function, $\sigma \left(z\right)=\frac{1}{1+e^{-z}}$, the final layer is replaced with a soft-max function, ${\sigma }_i\left(\mathbf{z}\right)=\frac{e^{-{\mathbf{z}}_i}}{{\sum }_{j=1}{e}^{-{\mathbf{z}}_j}}$, over three outputs, $\mathbf{z}\in {\mathbb{R}}^3$, corresponding to the three potential PT classes. The output loss function is modified to compute the (multi-class) cross-entropy loss,

$$\mathcal{L}\left(y,\widehat{y_j}\right)=-{\sum }_{j=1}^3{1}_j\left(y\right)\log \left(\widehat{y_j}\right)$$ 1

where $1_j\left(y\right)$, is the indicator that the true label is j and $\widehat{y_j}$ is the predicted likelihood that the sample is of class j. Models were otherwise trained using the same methods as described in ref. ³¹; hyperparameters were optimized to maximize the F1-score.

Models of PT cell state in the AIN and ATI populations were interpreted using cell phenotype similarity (CPS) score and canonical correlation analysis (CCA)-based analyses to identify cell- and marker-level patterns, as described in ref. ³¹. These methods were originally designed for interpreting binary classification tasks. Each disease (AIN and ATI) was analyzed in three separate comparisons: PT^neg vs HAVCR1⁺VCAM1^– PT, PT^neg vs VCAM1⁺ PT, and HAVCR1⁺VCAM1^– PT vs VCAM1⁺ PT.

For any of the above pairwise comparisons, the CPS method determines which cells are most relevant for pairwise discrimination between the corresponding types of PT cells in their respective neighborhood. Briefly, let ${\mathbf{z}}_i$ represent the model’s representation of a cell’s spatial context and ${\mathbf{s}}_j$ the representation of the j-th PT cell’s niche. Moreover, let $k\in \left\{{\mathrm{PT}}^{\mathrm{neg}},{HAVCR1}^{+}{VCAM1}^{-}\mathrm{PT},{VCAM1}^{+}\mathrm{PT}\right\}$. Then, the score for cell $c_i$, with model representation, ${\mathbf{z}}_i$, is

$${\mathrm{w}}_k\left({\mathbf{z}}_i\right)=\frac{{\sum }_{j=1}^J\mathrm{d}{\left({\mathbf{z}}_i,{\mathbf{s}}_j\right)}^{-p}{\mathrm{f}}_k\left({\mathbf{s}}_j\right)}{{\sum }_{j=1}^J\mathrm{d}{\left({\mathbf{z}}_i,{\mathbf{s}}_j\right)}^{-p}}$$ 2

where J denotes the number of PT cell niches and ${\mathrm{f}}_k\left({\mathbf{s}}_j\right)$ is a function that is 1 if the j-th PT cell is in state k and −1 if it is not in the k state.

Similarly, we fit CCA models to the input data, ${\mathbf{c}}_i$, and the representation of the PT cell’s niche, ${\mathbf{s}}_j$. Let $C_k=\left[{\mathbf{c}}_i\right]$ denote a matrix of the cells input data ($C\in {\mathbb{R}}^{N\times d}$, where N is the number of cells in cell niches and $d=5101$, the number of markers profiled). Then $S_k=\left[{\mathbf{s}}_j\right]$ is a matrix of niche representations ($S\in {\mathbb{R}}^{N\times h}$, where h is the size of the PT cell niche representation), where cell i is in PT niche j. Here we modify the CCA analysis in SORBET, which analyzes each hop distance. Instead, the CCA analysis is computed between each cell and the inferred subgraph representations.

Differential expression analyses

Differential expression analysis was performed using Seurat’s FindMarkers function with Wilcoxon rank-sum test. Three pairwise comparisons were conducted: AIN-cortex versus Ref-cortex, ATI-cortex versus Ref-cortex, and AIN-cortex versus ATI-cortex. Testing was restricted to cell types containing at least 3 cells per condition group. Parameters included a minimum gene detection threshold of 10% in either group, a minimum average log2 fold-change of 0.25, and Bonferroni correction for multiple testing. Genes were considered significantly differentially expressed at adjusted p-value < 0.05. For visualization, volcano plots were generated with the EnhancedVolcano package showing -log10(adjusted p-value) versus log2 fold-change.

NAMPT spatial expression analyses

Local and regional spatial regulation of NAMPT expression around VCAM1⁺ PT cells was evaluated in AIN cortex tissues. For local analysis (Fig. 4c), up to 10,000 non-PT cells were randomly sampled from AIN cortex, and for each sampled cell the Euclidean distance to every VCAM1⁺ PT cell in the same tissue was computed. Cells within 50 μm of at least one VCAM1⁺ PT cell were stratified by the number of VCAM1⁺ PT neighbors within 50 μm (1, 2, 3, or ≥4), and the percentage of NAMPT-positive cells (SCT-normalized expression >0) was calculated per stratum. Statistical significance of the dose–response relationship was assessed using the Cochran–Armitage trend test. For regional analysis (Fig. 4d), VCAM1⁺ PT density was calculated per patient as the number of VCAM1⁺ PT cells divided by tissue area (cells/mm²). For each patient, 200 non-PT cells located >200 μm from any VCAM1⁺ PT cell were randomly sampled, and the percentage of NAMPT-positive cells among these sampled cells was computed. The patient-level correlation between tissue-wide VCAM1⁺ PT density and the percentage of NAMPT⁺ cells in distant regions was assessed by Pearson correlation (n = 6 patients; one AIN cortex tissue was excluded due to insufficient (<100) cells >200 μm from any VCAM1⁺ PT due to high tissue VCAM1⁺ PT burden).

Statistics and reproducibility

General approach. All statistical analyses were performed in R (v.4.3.0) unless otherwise specified. Patient-level summaries (cell densities in cells/mm², interaction densities in interactions/mm², tissue-level expression averages, and within-compartment proportions computed per patient) were the unit of analysis for group comparisons. Cell-level analyses were used only when patient clustering was modeled explicitly (generalized linear mixed models for IMC spatial neighbors; Poisson models with patient-level clustering for ST ligand–receptor interactions).

Group comparisons. Continuous variables were compared between disease groups by Kruskal–Wallis tests with post-hoc pairwise Wilcoxon rank-sum tests for three-group comparisons (AIN vs ATI vs Ref), or Wilcoxon rank-sum tests for two-group comparisons (AIN vs ATI). Patient-level paired comparisons across spatial compartments (within vs outside lymphoid aggregates; peri-tubular vs distant) used Wilcoxon signed-rank tests. Categorical variables were compared using Fisher’s exact tests.

Correlations. Continuous variables were correlated using Spearman’s method, except for Fig. 5f, where Pearson correlation was applied to urinary and tissue CXCL9 values (displayed on log₁₀-scaled axes). For recovery analyses where each patient contributed a value derived from a varying number of index cells (e.g., per-cell expression averaged within patient), weighted Spearman correlation was used with weights equal to the number of contributing cells per patient. Fisher’s z-transformation compared independent correlation coefficients: z = 0.5 × ln((1 + r)/(1 − r)), SE = 1/√(n − 3).

Trend tests. Ordered group comparisons used Cochran–Armitage trend tests (proportions) or Jonckheere-Terpstra trend tests (continuous outcomes), as one-sided tests by design.

Multiple testing correction. Bonferroni adjustment was applied to predefined hypothesis-driven comparisons (cell densities across disease groups, spatial interaction densities, differential expression). Benjamini–Hochberg false discovery rate was applied to exploratory analyses (correlation networks, generalized linear mixed model q-values).

Bootstrap confidence intervals. For the CXCL9–CXCR3 predicted interaction density and PT-INJ2 cell density comparisons between AIN and ATI, 95% confidence intervals on the ratio of medians and ratio of means were computed by nonparametric percentile bootstrap with 10,000 iterations, resampling patients with replacement within each disease group. Bootstrap iterations in which the resampled ATI median was zero (yielding an undefined ratio of medians) were excluded from the ratio-of-medians CI calculation; this affected 1024 of 10,000 iterations for CXCL9–CXCR3 and 0 of 10,000 iterations for PT-INJ2.

Leave-one-out sensitivity analyses. For each headline AIN-versus-ATI (Blocks 1–6) and AIN-versus-reference (Block 7) comparison in the Xenium cortex cohort, the comparison was re-run with each of the seven AIN cortex patients sequentially excluded. Per-iteration sample sizes, medians, ratios, test statistics, and p-values are reported in Supplementary Data 9.

PT-INJ2 demographic stratification. Because VCAM1⁺ PT-INJ2 enrichment is a central finding of this study, PT-INJ2 cell density was examined across patient sex strata (Supplementary Data 11) within AIN and ATI groups of the IMC discovery cohort; no significant differences were observed across strata.

Reference tissue. Reference tissue values for ST analyses (n = 2 cortex tissues) are shown descriptively without formal statistical testing. ST statistical comparisons were limited to AIN vs ATI within cortex (n = 7 vs n = 6) or across cortex+medulla (n = 8 vs n = 7 vs n = 5).

Effect sizes. Spearman ρ or Pearson r for correlations; Cohen’s q for differences between independent correlations; rank-biserial correlation r for Wilcoxon rank-sum comparisons.

Test directionality and significance. Unless otherwise specified, all tests were two-sided with significance threshold P < 0.05 after adjustment. Trend tests (Jonckheere-Terpstra, Cochran–Armitage) and the pre-specified pathologist scoring contrast (PT-INJ1 vs PT-H, one-sided Wilcoxon rank-sum) used one-sided alternatives.

Software. R v.4.3.0; Seurat v5.0; Harmony v0.1.1; batchelor v1.12.3; imcRtools v1.14.0; FNN v1.1.4; Python v3.9 with scanpy v1.9.3.

Author contributions

M.L.B., D.G.M., and L.G.C. conceived and designed the study. M.L.B., V.R.K., and T.B. performed the imaging mass cytometry experiments. M.L.B., V.R.K., T.B., and M.W. developed the imaging mass cytometry data analysis pipeline. M.L.B. performed and coordinated the single-cell spatial transcriptomics experiments. G.M. performed the pathological diagnosis adjudication and injury scoring. G.L. performed the histopathologic neighborhood scoring. J.M.C. and Y.K. developed the SORBET computational method and, with input from M.L.B., adapted it to address a specific question in this study; J.M.C. and Y.K. applied the adapted method to the data, and M.L.B. interpreted and integrated the results. M.L.B. analyzed the data. D.G.M., S.S., C.R.P., and L.G.C. assembled the clinical data. A.Z.R. and C.R.P. provided and characterized the validation cohort. M.L.B., V.R.K., J.M.C., Y.K., C.R.P., D.G.M., and L.G.C. acquired funding. L.G.C. and D.G.M. supervised the study. M.L.B. wrote the original draft, and L.G.C., D.G.M., and M.W. reviewed and edited the manuscript. All authors reviewed and approved the final manuscript.

Peer review

Peer review information

Nature Communications thanks Stefan Bonn and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This work was supported by NIH grants R01DK126815 (L.G.C.), R01DK128087 (D.G.M.), T32DK007276 (M.L.B.), U01DK133768 (V.R.K. and L.G.C.), P30DK045735 (Yale Diabetes Research Center Pilot and Feasibility Grant, M.L.B. and L.G.C.), and U54DK137331 (C.R.P.); American Diabetes Association Postdoctoral Fellowship 11-23-PDF-63 (M.L.B.). Development of SORBET was supported by NIH grants R01DA063148, U01DA051410, U54AG076043, U54AG079759, and P50CA121974 (J.M.C. and Y.K.). The Kidney Precision Medicine Project (KPMP) is supported by the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) through the following grants: U01DK133081, U01DK133091, U01DK133092, U01DK133093, U01DK133095, U01DK133097, U01DK114866, U01DK114908, U01DK133090, U01DK133113, U01DK133766, U01DK133768, U01DK114907, U01DK114920, U01DK114923, U01DK114933, U24DK114886, UH3DK114926, UH3DK114861, UH3DK114915, and UH3DK114937. We gratefully acknowledge the essential contributions of our patient participants and the support of the American public through their tax dollars. NIH disclaimer: the content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Data availability

Processed data objects supporting the findings of this study (annotated SpatialExperiment and Seurat objects with cell-type annotations, embeddings, and per-patient clinical metadata) are deposited at Zenodo⁸⁰ and accessible at https://doi.org/10.5281/zenodo.19899468. The raw imaging mass cytometry data generated in this study (multichannel images, segmentation masks, and single-cell intensities for the Discovery cohort, and raw Hyperion text exports for the Validation cohort) have been deposited in the BioImage Archive (EMBL-EBI) under accession code S-BIAD3694⁸¹. The raw Xenium spatial transcriptomics data generated in this study (morphology images, decoded transcripts, cell-feature matrices, and cell and nuclear segmentation) have been deposited in the Gene Expression Omnibus (GEO) under accession code GSE336890⁸². The KPMP human kidney atlas (Lake et al., Nature 2023) used for cross-platform validation of proximal tubule injury states is publicly available at the CZ CELLxGENE Discover collection (collection ID bcb61471-2a44-4d00-a0af-ff085512674c; https://cellxgene.cziscience.com/collections/bcb61471-2a44-4d00-a0af-ff085512674c). De-identified clinical metadata accompanying the deposited objects is provided as part of the Zenodo deposit. Additional patient-level clinical data are available from the corresponding authors under restricted access owing to patient-privacy constraints; access requires institutional review board approval (Johns Hopkins IRB00221958, IRB00090103) and compliance with HIPAA and the Yale Kidney Biobank data use agreement (ClinicalTrials.gov NCT04343417). Requests will be answered within 60 days. Source data are provided with this paper.

Code availability

All custom analysis code used in this study is available at the GitHub repository https://github.com/megleila/AIN-spatial-analysis and archived at Zenodo⁸³: https://doi.org/10.5281/zenodo.19900536. The repository documents the imaging mass cytometry workflow (signal spillover correction, fastMNN batch correction, Rphenograph clustering, k-nearest-neighbor spatial graph construction, density and neighborhood analyses) and the Xenium spatial transcriptomics workflow (SCTransform normalization, Harmony integration, Louvain clustering, proximal tubule subclustering, and KPMP atlas cross-validation). Companion methods that are not custom-developed in this work include Steinbock for IMC pre-processing, Mesmer for cell segmentation (with custom kidney-tuned parameters described in Methods), and SORBET for spatial PT injury microenvironment classification.

Competing interests

The authors declare the following competing interests: L.G.C. is a consultant for DropShot Therapeutics, Novo Nordisk, GondolaBio, and Pfizer in areas not related directly to this research. M.L.B. is a consultant for DropShot Therapeutics in areas not related directly to this research. D.G.M. and C.R.P. are co-inventors and the applicants on a pending patent titled “Methods and Systems for Diagnosis of Acute Interstitial Nephritis (Patent number US20200072847A1).” No aspect of the manuscript is covered in the patent, which relates to urine biomarkers for diagnosis of acute interstitial nephritis. D.G.M. and C.R.P. are also co-founders of Predict AIN, LLC, a diagnostics company that has entered into a technology license and development agreement with Machaon, Inc., another diagnostics firm. D.G.M. serves as a consultant for BioHaven, Inc., and on the editorial boards for MDCalc and Kidney360. All other authors (V.R.K., T.B., M.W., J.M.C., G.L., G.M., A.Z.R., and Y.K.) declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77210-y.