Integration of spatial protein imaging and transcriptomics in the human kidney tracks the regenerative potential of proximal tubules

Abstract

The organizational principles of nephronal segments are based on anatomical and physiological attributes that are linked to the homeostatic functions of the kidney. Recent molecular approaches have uncovered layers of deeper signatures and states in tubular cells that arise at various time points on the disease trajectory. Here, we introduce an analytical pipeline of multiplexed spatial protein imaging integrated with RNA expression to characterize proximal tubular subpopulations and neighborhoods in human kidney tissue. We demonstrate that, in reference tissue, a large proportion of S1 proximal tubular epithelial cells expresses thymus antigen 1 (THY1), a mesenchymal stromal and stem cell marker that regulates differentiation. Kidney disease is associated with loss of THY1 and transition toward expression of prominin 1 (PROM1), another stem cell marker recently linked to failed repair. Our data support a model in which the interplay between THY1 and PROM1 expression in proximal tubules associates with their regenerative potential and marks the timeline of disease progression.

The interplay in expression of the stem cell markers THY1 and PROM1 in kidney proximal tubules marks the timeline of disease.

INTRODUCTION

The kidney maintains body homeostasis through the integrated functions of nephronal segments and numerous cell populations. High-resolution molecular interrogation techniques such as single-cell RNA sequencing (RNA-seq) have underscored the complexity of the kidney, which is composed of at least 70 interacting cell types (1–4). Our understanding of these specialized cells has increased in granularity because we now recognize the presence of subpopulations in what was thought to be uniform cell types. Furthermore, within a given population, recent depictions of various cell states that correlate with disease trajectory or outcomes usher previously unidentified opportunities to uncover key pathways in the pathogenesis of kidney disease and precise biomarkers that clock the disease timeline (1, 2). Some altered cell states (typically defined by the expression profile of specific stress- or mesenchymal-associated genes or proteins) are associated with maladaptive or failed repair following injury and may predict the progression to chronic kidney disease (1). The power of single-cell methodologies can be greatly expanded when analyzed in a “spatial” context. The interaction of various cells in spatially defined microenvironments within the kidney likely governs many of these pathological processes (1, 2, 5–7). Recently described large-scale multiplexed spatial protein imaging techniques such as CO-Detection by indEXing (CODEX) capture cell types and their distribution within the complex landscape of kidney tissue (7). This evolving technology allows the detection of an increasing number of target markers (up to 100) covering structural, immune, stromal, and various injury states including posttranslational protein modifications. Applying such a spatially anchored analytical pipeline can then provide “single-cell data” with spatial protein expression features. The power of such spatial protein-based analysis can be further augmented when integrated with spatial gene expression platforms such as spatial transcriptomics (ST), performed on a sequential section from the same specimen (4, 8). This spatially anchored integrative analysis bridging protein to RNA at the single-cell level will greatly improve our understanding of the pathophysiology of kidney disease and the drivers of its progression.

As the most upstream tubular segment of the nephron, the proximal tubule (PT) is central to the functioning of the kidney and orchestrates tight communication with more distal segments (9, 10). It is also a frequent target of disease, and its pathology rapidly spreads to other cell populations. Many PT cell subpopulations have been described along the spectrum from health to disease (1, 11–15). While some are permanent components of the PT, others have a more fleeting existence, appearing or vanishing at specific points in the timeline of disease. Of particular interest are subtypes of PTs that undergo maladaptive repair, ushering the progression toward fibrosis (11–13). These cells are defined by the persistent expression of several injury markers such as HAVCR1 [Kidney injury Molecule-1 (KIM1)] or VCAM1 (13, 14, 16). These PT subtypes have been well characterized in both experimental models of kidney injury and human kidney tissue from patients with kidney disease (1, 11–13). In addition, genes that typically mark stem cells and are involved in repair, such as PROM1 (CD133) and SOX9, are up-regulated postinjury (17–19). However, persistent activation of these genes is frequently associated with maladaptive repair and progressive fibrosis (1, 18, 20).

Thymus antigen 1 (THY1, also known as CD90) is a small glycoprotein that was originally described as a T cell marker in the thymus but later found to be expressed in stem cells and in a variety of neuronal, mesenchymal, stromal, and endothelial cells (21). THY1 interacts with various ligands to regulate signaling pathways related to cell growth and regeneration. For example, THY1 is thought to play a key regulatory role in the differentiation of mesenchymal stromal cells (21, 22). A recent report by Wu et al. (23) showed that THY1 is expressed in human PTs. The function of THY1 in the kidney is unclear, and its expression in various PT subpopulations has not been investigated. Because of its association with stem cells and regeneration, there has been a growing interest in its potential role in kidney development and kidney repair after injury (21, 23). THY1 also exists in a soluble form, and an inverse correlation between soluble THY1 levels and kidney function has been described (21, 23). Despite reports suggesting THY1 expression in mouse kidneys (21), previous data (24) and various murine single-cell expression databases (2, 13, 25) have consistently demonstrated that THY1 is not expressed in mouse tubular cells. THY1 expression in the mouse kidneys is likely to originate from extratubular sources such as infiltrating immune cells. Whether THY1 expression in the kidney tubules is unique to humans remains unknown.

In this work, we aimed to leverage a CODEX analytical pipeline integrated with spatial and single-cell transcriptomics to establish the presence and distribution of distinct populations of THY1-positive and prominin 1 (PROM1)–positive PT cells. We also define the molecular trajectory and significance of these cells in human reference and disease kidney tissue. Our findings highlight the importance of PT cell states in understanding the timeline of kidney disease progression and suggest that THY1 loss and a shift toward high PROM1 states are associated with kidney disease.

RESULTS

Building a multiplexed imaging and analytical pipeline to decipher cell types in kidney health and disease

To delineate the various cell types and cell states using spatial protein expression information, we first defined the identity of each cell in a set of healthy and diseased kidney tissue samples using a CODEX marker panel with structural, immune, and injury markers (tables S1 and S2 and fig. S1). Summary demographics and diagnosis of reference and disease kidney tissue specimens are presented in table S1. The initial core marker panel applied to all specimens (23 markers) was subsequently extended to include more specialized markers (total of 38) (see tables S1 and S2 for details by specimen). CODEX imaging was performed at a large scale, scanning the entire tissue sections (7). Quality control steps and technical and biological reproducibility were described in Materials and Methods and shown in fig. S2. Applying our imaging and analysis pipeline within a single analytical space using the core markers, we classified a total of 169,802 cells (Fig. 1, A to C). This approach was successful in identifying all major expected cell types in the kidney based on the protein expression profile including glomerular, PT, distal tubular, endothelial, and immune cells. The cell classes were spatially mapped to the images to validate the cell labels (Fig. 1, D to K).

Fig. 1. Cell classification based on CODEX imaging in reference and disease specimens.

Reference and disease specimens were stained and imaged using CODEX. (A) Heatmap showing the z-scaled mean fluorescence intensity for CODEX markers for all the cells segmented from the images of reference and disease kidney tissue sections. The specimen identification number and diagnosis are indicated on the left. IgA, immunoglobulin A; Ref, reference. (B) Violin plots showing the mean fluorescence intensity distribution per cell of the markers indicated in each of the cell clusters. (C) Uniform manifold approximation and projection (UMAP) plot of the Louvain clusters identified from the reference and disease specimens. The UMAP axis labels X1 and X2 represent abstract dimensions derived from the original data. Cell annotation was based on the marker expression profile (B) and mapping back of the clusters to the images (D to K). Distal convoluted tubules, connecting segments, and collecting duct cells could not be individually resolved on the basis of the markers used and are labeled collectively as distal nephron. [(D) to (G)] Representative CODEX images of select markers in reference tissue [(D) and (F)] and the corresponding cell clusters mapped back as nuclear overlays [(E) and (G)]. Scale bars, 500 μm. [(H) to (K)] Representative CODEX images of select markers in a kidney biopsy with chronic kidney disease (CKD) [(H) and (J)] and the corresponding cell clusters mapped back as nuclear overlays [(I) and (K)]. CDH1, Cadherin-1; Endo, endothelial cells; DCT, distal convoluted tubule.

Identifying PT cell subtypes and states using protein expression analysis with CODEX

To increase the resolution of PT cell subtypes, we reclustered PT cells in a new analytical space (Fig. 2). Clustering was predominantly based on the expression of low density lipoprotein receptor-related protein 2 (LRP2) and aquaporin 1 (AQP1) and the presence or absence of CD68, THY1, PROM1, and insulin-like growth factor-binding protein 7 (IGFBP7) (Fig. 2, A and B). We uncovered two PT subtypes that uniquely expressed THY1 or PROM1 (Fig. 2, A to D). As discussed, THY1 (also known as CD90) is a stem cell marker implicated in cell differentiation and regeneration (21). PROM1 is a marker of progenitor cells that is also expressed in cells undergoing adaptive or maladaptive repair (1, 19). The PT identity of these subtypes was confirmed morphologically by backmapping the cell clusters onto the images (Fig. 2, E to J). An extended CODEX panel applied to a disease tissue was also used to validate the profile of the PT subtypes, showing a protein expression signature consistent with injury [high KIM-1, phosphorylated c-JUN proto-oncogene protein (p-cJUN), and phosphorylated mixed lineage kinase domain-like protein (pMLKL)] for PROM1-positive PTs as compared to THY1-positive or other PTs (fig. S3).

Fig. 2. Subclustering of PT cells identifies groups expressing injury and regenerative state markers.

Post-CODEX and –cell-type classification, PT cells were reclustered with a reduced set of molecular markers. (A) Violin plots showing the mean intensity distribution per cell of the markers relevant for each of the PT cell clusters. (B) UMAP plot visualizing the various PT clusters, highlighting the unique separation of the THY1-positive and PROM1-positive clusters. (C and D) Feature plots of THY1 and PROM1 expression. (E to G) Representative images from reference tissue showing staining of 4′,6-diamidino-2-phenylindole (DAPI), LRP2, THY1, CD68, and PROM1 [(E) and (F)] and mapping of the corresponding cell clusters as nuclear overlays (G). (H to J) Representative images from disease tissue with similar staining [(H) and (I)] and mapping of the corresponding cell clusters (J). Scale bar, 200 μm

Bridging THY1 expression from protein to RNA by integrating CODEX and ST

PROM1 expression has been described in a subpopulation of PTs, but the expression of THY1 in the human kidney and within PT cells is less well characterized. Here, we aimed to spatially define the relationship between THY1 protein and RNA expression. To determine the concordance of THY1 protein and mRNA expression, we used spatially coregistered sequential sections of human kidney cortex that underwent CODEX and ST (Fig. 3, A to C). ST spots overlaying THY1-positive PTs (defined as THY1 positive and LRP2 positive by CODEX) were identified (Fig. 3, D to F) and found to have high THY1 RNA by differential gene expression as compared to spots overlaying THY1-negative PT cells as defined by CODEX (Fig. 3G). Furthermore, THY1 RNA (from ST) and protein (from CODEX) showed significant correlation in these spots overlaying PTs identified as LRP2 positive by CODEX (fig. S4). Additional genes were differentially up-regulated in THY1-positive PT spots, including MT1G, GATM, GPX3, APOE, PDZK1IP1, and ALDOB (Fig. 3G). To validate and extend the findings at the single-cell level, we used the single nuclear RNA-seq (snRNA-seq) Kidney Precision Medicine Project (KPMP) atlas to subset PTs (Fig. 3H). We identified a PT subcluster (cluster 5; arrow in Fig. 3I) with a high level of THY1 expression, showing enrichment with similar genes to those uncovered by ST in CODEX-defined THY1-positive PTs (Fig. 3, G and I). This was also independently confirmed by comparing THY1-positive PT gene expression to all other PTs in the snRNA-seq atlas (Fig. 3, J and K). Further analysis with additional markers showed that THY1-positive PTs are enriched in genes corresponding to the S1 segment (fig. S5A). Pathway analysis based on the differentially up-regulated genes in THY1-positive PTs revealed pathways related to cell growth, metabolism, proliferation, and differentiation (Fig. 3L and fig. S6), which are consistent with proposed functions of THY1 and its potential link to a regenerative capability (26). We also confirmed distinct PT clusters with high PROM1 expression (clusters 1 and 8 in Fig. 3I), analogous to the findings with CODEX. The gene expression profile of PROM1⁺ PTs has been reported previously (27).

Fig. 3. Bridging THY1 expression at the cell level from protein to RNA.

(A) Representative CODEX image of a reference tissue showing seven markers. (B) LRP2, PODXL, and THY1 expression registered on the sequential ST section stained with hematoxylin and eosin (H&E). THY1 colabels a subset of PT (LRP2-positive). (C) THY1 gene expression by ST registered with protein expression in CODEX. Clear/blue spots = minimal gene expression. Red/orange = higher gene expression. The black rectangle in (C) corresponds to (D) to (F). (D) H&E image from ST. (E) CODEX with LRP2, PODXL, and THY1 registered on the ST H&E image. (F) Yellow circles correspond to ST spots overlaying regions that are positive for LRP2 and THY1 proteins. Cyan circles correspond to ST spots overlaying regions that are LRP2-positive and THY1-negative protein. (G) Volcano plot of differentially expressed genes (DEGs) comparing the two sets of spots defined in (F) from combined three ST sections, [yellow versus cyan, statistical test: negative binomial generalized linear model, adjusted P < 0.05, |log₂FC (fold change)| > 0.25, n = 3]. (H) UMAP of PT class from KPMP snRNA-seq kidney atlas of health and disease (KPMP snRNA-seq atlas) clustered on the basis of highly variable gene expression. (I) Gene expression profile of the PT clusters from (H) displaying the genes uncovered in ST spots in (G) in addition to PROM1. Arrow indicates a subcluster with high expression of THY1. (J) Volcano plot comparing gene expression between THY1-positive PTs and all other PTs in the KPMP snRNA-seq atlas (statistical test: negative binomial generalized linear model, adjusted P < 0.05, |log₂FC| > 0.25, n = 29). (K) Violin plots comparing the expression of PROM1 and a subset of up-regulated genes from (J) in THY1-positive versus THY1-negative PTs. (L) Pathway analysis of differentially up-regulated genes in THY1-positive versus THY1-negative PTs. rRNA, ribosomal RNA; GTP, guanosine 5′-triphosphate; MAPK, mitogen-activated protein kinase; MHC, major histocompatibility complex; BRAF, V-Raf Murine Sarcoma Viral Oncogene Homolog B; EIF2AK4 or GCN2, Eukaryotic Translation Initiation Factor 2 Alpha Kinase 4; SLIT, Slit glycoprotein; ROBO, Roundabout receptor. Scale bars, 500 μm [(A) and (C)] and 200 μm (D).

Changes of THY1 and PROM1 expression in PTs during kidney disease

We compared the changes in PT cell types identified by CODEX between reference and disease (Fig. 4, A to C). THY1-expressing PT cells were significantly decreased in disease (Fig. 4D). To validate this finding and further increase the robustness of our results, we examined THY1 and PROM1 expression in PTs in a larger validation cohort (n = 49 separate specimens with 2.35 million cells in total) using biopsies from the KPMP and healthy nephrectomy tissue from the Human Biomolecular Atlas Program (HuBMAP) (table S3). These tissue specimens were imaged with the newer PhenoCycler system. Similar analysis was performed to uncover subtypes of PTs (n = 441,509 cells) from the total cell clusters (Fig. 4, E to H, and figs. S7 to S9). Because of the larger sample size of this validation cohort, disease biopsies were separated by diagnosis: reference, acute kidney injury (AKI), or chronic kidney disease (CKD). Similar to the CODEX discovery cohort, we observed that THY1-positive PTs were highly abundant in reference tissue, comprising ~40% of PT cells, but they decreased significantly in AKI and CKD (Fig. 4, H to N). In healthy reference tissue sections, the expression of THY1 (RNA and protein) was periglomerular and absent from S3 PTs in the medullary rays (Fig. 4, I and L, and fig. S5B), consistent with the snRNA-seq suggesting that the expression of THY1 at baseline is primarily localized to S1 tubules. Unlike THY1, the proportion of PROM1-expressing PT cells was trending higher in the CODEX data with disease (Fig. 4D), was significantly increased in the PhenoCycler data in AKI, and was also trending higher, albeit not significantly, in CKD (Fig. 4, H, J, K, M, and N). No major differences were observed in other PTs (Fig. 4H and fig. S10).

Fig. 4. Changes in THY1 expression in disease.

(A) Integrated UMAP of PT cells from the combined CODEX dataset, separated into reference (n = 5) or disease (n = 4) groups (B and C, respectively). (D) Quantitative analysis of cell proportions in reference and disease CODEX tissue sections [two-way analysis of variance (ANOVA) test, adjusted P < 0.05]. (E to G) Integrated UMAPs of PT cells from the combined PhenoCycler dataset visualized on the basis of condition: combined, reference, or CKD. (H) Quantitative analysis of cell proportions in reference (n = 13) and disease conditions (AKI, n = 15; CKD, n = 21) in the PhenoCycler dataset (two-way ANOVA test, adjusted P < 0.05). (I to K) Representative large-scale images of the PhenoCycler data showing THY1, PROM1, and markers for endothelium (CD31), PTs (LRP2), collecting ducts (AQP2), and thick ascending limbs (TAL) (UMOD). Scale bars, 500 μm. (L to N) Insets of the boxed areas from (I) to (K). Scale bars, 200 μm.

CODEX-based cell trajectory analysis relative to THY1 and PROM1 protein expression

To understand whether the THY1 and PROM1 states of the PT represent steps along a spectrum or unique divergent cell processes, we next studied cell trajectories of PTs in the CODEX data, which highlighted the dynamics of THY1 and PROM1 expression (Fig. 5). The trajectories of PTs (high LRP2) in reference specimens show distinct paths toward THY1 and PROM1 expression [Fig. 5, A to C; trajectories (t) t₁ and t₂ for THY1 and t₃ and t₅ for PROM1]. This is consistent with distinct expression profiles (such as injury markers) of these two subpopulations outlined in fig. S3. Some PROM1-positive PTs in reference tissue had a low level of THY1 expression (Fig. 5, B and C, and fig. S11). This low level of THY1 is transient and lost with increasing PROM1 expression along trajectory 3 (t₃ in Fig. 5, D and E). In disease specimens, the expression of THY1 is markedly reduced (Fig. 5, F and G). The trajectories toward THY1 and PROM1 expression (trajectories t₄ and t₁) remain distinct. We also performed trajectory analysis on the integrated dataset of reference and disease, showing similar results (fig. S12). Furthermore, a population of PROM1-positive cells that also express THY1 was absent in disease in the CODEX data, likely because of small sample size (Fig. 5, B and C versus G and H), but was observed in the larger phenocycler data (Fig. 4). This THY1⁺ PROM1⁺ PT population also had the highest expression of vascular cell adhesion molecule–1 (VCAM1), a marker associated with injury and repair (Fig. 6, A to F). Trajectory analysis on this population showed multiple paths toward healthy PTs, PROM1⁺ PTs, and THY1⁺ PTs, suggesting that THY1⁺ PROM1⁺ VCAM1⁺ PTs are in an injured transitional state (Fig. 6B). VCAM1 was also expressed in PROM1⁺ and IGFBP7⁺ (but not THY1⁺) cells, albeit at a lower level (Fig. 6F), which could be consistent with a path toward failed repair. Conversely, THY1⁺ cells at baseline do not have high expression of VCAM1, and the trajectory analysis in Fig. 6 (B and G to I; trajectory 4) suggests that any observed low expression of VCAM1 in THY1⁺ cells likely indicates a prior state of injury. In fig. S13, we also performed trajectory analysis starting from THY1⁺ cells in CODEX data (unintegrated dataset), also supporting that THY1⁺ PTs predominantly transition into healthy PT and, to a lesser extent, to a PROM1⁺ state, indicating failed repair.

Fig. 5. Trajectory analysis of PT cells accounting for the dynamics of THY1 and PROM1 expression in CODEX data.

Reference (n = 5) (A to E) or disease (n = 4) (F to J) samples were separated and reclustered and projected into UMAP space. (A) PT cell clusters from CODEX data in reference tissue with trajectory analysis (t₁ to t₅) starting from clusters of PT with high LRP2 expression. [(B) and (C)] Feature plots for THY1 and PROM1 expression in reference tissue samples. [(D) and (E)] Pseudotime spectrum of PT cells based on THY1 and PROM1 expression, in reference samples. (F) PT cell clusters from CODEX data in disease tissue with trajectory analysis (t₁ to t₄) starting from clusters of PT with high LRP2 expression. [(G) and (H)] Feature plots for THY1 and PROM1 expression in disease tissue samples. [(I) and (J)] Pseudotime spectrum of PT cells based on THY1 and PROM1 expression in disease samples.

Fig. 6. Trajectory analysis of PT cells expressing high VCAM1 from the integrated phenocycler datasets of reference and disease.

(A) Integrated UMAP showing various PT cell clusters from Fig. 4E. (B) Trajectory analysis and pseudotime starting from the THY1⁺ PROM1⁺ clusters, which are also the cluster with high VCAM1 expression. (C to E) Feature plots for THY1, PROM1, and VCAM1. (F) Violin plots showing the distribution of THY1, PROM1, and VCAM1 expression among the various PT clusters. (G to I) Pseudotime spectrum of PT cell trajectories based on THY1, PROM1, and VCAM1 expression, respectively.

snRNA-seq–based cell trajectory analysis relative to THY1 and PROM1 expression

To determine whether similar cellular trajectories are found in transcriptomics data, we studied the dynamics of the various populations of PT cells in the snRNA-seq dataset from the KPMP atlas organized by clinical cohort: reference, AKI, and CKD (Fig. 7 and fig. S14). In reference tissue, there was a distinct population of THY1-positive PTs (cluster 3 in Fig. 7A and fig. S14A) and three subpopulations of PROM1-positive PTs grouped together (clusters 4, 6, and 8). PROM1-positive subpopulations express varying degrees of injury markers (consistent with CODEX data; fig. S3) such as HAVCR1 (KIM1) and VCAM1 (clusters 6 > 8 > 4; fig. S14A). In contrast to the phenocycler data, the THY1⁺ cluster had a weak VCAM1 signal. Notably, trajectories from noninjured LRP2⁺ tubules toward THY1-positive and PROM1-positive PT cells are also distinct, similar to our CODEX data in reference tissue (Fig. 5). In further concordance with our CODEX data, we observed a low level of expression of THY1 in the PROM1-positive PT subpopulation (cluster 4) that expresses the least amount of injury markers (Fig. 7A and fig. S14A; trajectory t₃ in Fig. 7, D and E) compared to the other PROM1-positive PT clusters that do not express THY1.

Fig. 7. Trajectory analysis of PT cells accounting for the dynamics of THY1 and PROM1 expression in snRNA-seq data.

(A) PT cell clusters from the KPMP snRNA-seq data in reference tissues (n = 8 samples with 18,178 nuclei) with trajectory (t) analysis starting with a cluster of PTs without any injury markers and expressing genes known to be present in healthy differentiated PT cells. (B and C) Feature plots for THY1 and PROM1 expression in reference PT cells. (D and E) Pseudotime spectrum of PT cells based on THY1 and PROM1 expression for each of the trajectories shown in (A). (F to J) The same analysis performed in (A) to (E) was done on cells from AKI tissue specimens (n = 10 samples with 16,067 nuclei). (K to O) Similar analysis performed on cells from CKD tissue specimens (n = 11 samples with 13,953 nuclei).

In AKI, the main population of THY1-expressing PTs (cluster 5; Fig. 7, F and G, and fig. S14B) remains distinct from PROM1-expressing PTs (clusters 0 and 7; Fig. 7, F and H, and fig. S14B), with separate trajectories from uninjured LRP2⁺ PTs toward those with high THY1 or PROM1 expression (trajectories t₃ and t₅ versus trajectories t₁, t₂, and t₄, respectively; Fig. 7, F, I, and J). High-THY1 PTs were characterized by near-absent HAVCR1 expression but high VCAM1 expression. HAVCR1 was closely linked to PROM1-expressing cells (clusters 0 and 7). There were also several subpopulations of PTs with intermediate levels of PROM1 and THY1 expression that also expressed VCAM1 (fig. S14B), particularly on trajectories t₁, t₂, and t₄ (Fig. 7, I and J) from uninjured LRP2⁺ PTs to high PROM1–expressing cells. This suggests an expression spectrum leading to an injury state characterized by low THY1 and high PROM1, VCAM1, and HAVCR1 expression.

In CKD, the expression of THY1 is markedly reduced and only observed at low levels in two subpopulations (clusters 3 and 4; Fig. 7, K and L, and fig. S14C) that have reduced HAVCR1 or VCAM1 expression. Conversely, the expression of PROM1 remained substantial and distinct from THY1 expression (clusters 2 and 6; Fig. 7, K and M, and fig. S14C), with distinct trajectories from uninjured LRP2⁺ PTs toward PTs with high THY1 or PROM1 expression (trajectory t₃ versus trajectories t₂ and t₄, respectively; Fig. 7, K, M, and O). PTs expressing PROM1 also had high HAVCR1 and VCAM1 expression, suggesting a shift toward a maladaptive repair phenotype (11, 13, 16). Similar analyses were done on integrated and unintegrated datasets starting from THY1⁺ or LRP2^high THY^– PTs, showing consistent results (figs. S15 and S16).

Cumulatively, the cell expression and trajectory data from CODEX and snRNA-seq are, for the most part, concordant and suggest that THY1 and PROM1 are likely on an opposite spectrum of disease progression with overlapping intermediary states during injury. Specifically, the loss of THY1 and shift toward high PROM1 states are associated with kidney disease. The complementary information from RNA and protein highlights the need to study the concurrent expression of these various markers to better understand the evolution of disease states.

Neighborhood analysis from the CODEX imaging data

To determine the cellular microenvironments for the distinct THY1 and PROM1 PT subpopulations, we performed a cell-centric neighborhood analysis (7, 28), which uncovered 19 cell neighborhood clusters or niches where specific types of cells are significantly associated in the same neighborhoods (Fig. 8). When considering the predominant cell types present within each niche and their relationships to other niches in a dimensionally reduced space (Fig. 8, A and B), we deconstructed the niche landscape into nephronal, immune, endothelial, and stromal microenvironments. For example, we can specify niches enriched in glomerular cells (N15 > N5, likely glomerular or periglomerular environments), PTs (N2, N3, N4, N6, and N8), immune cells (myeloid: N5 > N18 > N7 > N17; T cells: N18 > N7 > N17), and fibroblasts (N7). THY1-positive and PROM1-positive PTs account for the largest proportion of cells in the N3 and N17 niches, respectively. Notably, the abundance of immune cells in N3 is in the lowest tertile compared to the N17 niche, which is in the highest tertile of immune cell abundance within all the niches (Fig. 8B), suggesting an association between PROM1-positive PTs and immune cell activity. Comparing the distribution of niches in disease versus reference, we observed an increase in fibroblast (N7), immune (N5 and N18), and PROM1-positive PT (N17) niches and a decrease in the THY1-positive PT (N3) niche with disease (Fig. 8C). Pairwise analysis further confirmed the positive association of PROM1-positive PTs and negative association of THY1-positive PTs with immune cells (Fig. 8D). These data are consistent with the observed changes in PT cell composition and trajectories in health and disease detected in both the CODEX and snRNA-seq analysis.

Fig. 8. Neighborhood analysis in CODEX data highlights epithelial immune cell interactions.

(A) t-SNE plot showing neighborhood clusters or niches (N1 to N19; each dot is a niche) based on the average distribution of cell types in each niche. The major cell type or tubular type in the underlying niches is indicated. (B) Distribution of specific cell types in all neighborhood clusters. (C) (Top) Distribution of cell types in neighborhoods and the propensity of niches to be found in reference or disease as assessed by odds ratio (OR) [P < 0.05, confidence interval (CI): 95%], and (C) (bottom) the average distribution of cell types in each niche. Undef., undefined. (D) Interactions of PT, immune cells, and fibroblasts in niches. Top: Chord plot to visualize the pairwise cell-cell interactions in all niches. Bottom: Pairwise correlation between cell types in all neighborhoods by Pearson’s coefficient (P < 0.05). Red boxes highlight negative correlation of THY1-positive PTs with immune cells, and green boxes highlight the positive correlation between the immune cells and PROM1-positive PTs.

Cell-cell interaction analysis

To further characterize the functional significance of THY1 expression in PTs, we performed ligand-receptor analysis between subtypes of PTs and all cells in the KPMP snRNA-seq atlas (Fig. 9 and fig. S17). Our results demonstrate that some PT populations are prone to interact with epithelial, endothelial, and immune cells using basement membrane ligands such as collagen IV subunits that interact with various integrins (Fig. 9, A to C, and fig. S17, A to C). These interactions are key for attachment to extracellular matrix components, cell growth, and differentiation (29, 30). These ligand-receptor interactions are most pronounced for PROM1-positive PTs, which also have other interactions that are consistent with potential immune activation [CD226 activation (31)] and angiogenesis [thrombospondin 1 (THBS1) signaling (32)], particularly in disease. Conversely, these interactions are markedly absent in THY1-positive PT cells (Fig. 9, A to C, and fig. S17, A to D). THY1-positive PTs only retain Secreted Phosphoprotein 1 (SPP1, also known as osteopontin)-linked cell interactions, predominantly with nonimmune cells. These interactions of SPP1 were with integrins and CD44, which may be important in cell migration (33). Cumulatively, these findings suggest that THY1-positive PTs are associated with molecular interactions characteristic of differentiation and transmigration, which are key features of cells that participate in regeneration.

Fig. 9. Ligand-receptor analysis for PT cells from snRNA-seq data.

(A to C) Bubble plots from reference (n = 8 samples with 51,732 nuclei for all cells), AKI (n = 10 samples with 50,442 nuclei for all cells), and CKD (n = 11 samples with 40,450 nuclei for all cells) tissues showing the interactions of PROM1⁺ and THY1⁺ PTs with all other cells (x axis) and the ligand-receptor pairs used in such interactions (y axis) (one-sided permutation test, P < 0.05). See fig. S17 for all PT types and additional analysis. Commum.prob, communication probability.

DISCUSSION

In this work, we characterized subpopulations of PTs in the human kidney expressing THY1 and PROM1, markers linked with repair and regeneration. We localized these subpopulations with spatial gene and protein expression and further identified a THY1-positive PT subcluster within an snRNA-seq atlas, characterized by regenerative pathways and distinct from the PROM1-positive PTs. We then quantified changes in THY1-positive and PROM1-positive PT cells and charted their trajectories in health and disease. Broadly, THY1 and PROM1 are considered markers of stem cells and thought to modulate tissue response to injury (19, 21). Whereas persistent expression of PROM1 in renal epithelial cells has been associated with maladaptive repair and progression toward fibrosis (1), the role of THY1 and changes in its expression in PTs is less well understood.

In nonepithelial cells such as fibroblasts, leukocytes, neurons, and endothelial cells, THY1 has been implicated in modulating processes involved in tissue regeneration such as cell migration, proliferation, and matrix remodeling (21, 22, 34–36). It is known that THY1 is not usually expressed in murine kidney epithelium (24), and this has been further confirmed by the publicly available single-cell transcriptomic databases (37). Therefore, the low expression reported in some mouse models (38) likely originates from other cell types such as immune, endothelial, or stromal cells. Induced expression of human THY1 in murine kidney causes proliferative abnormalities of PTs (24), thereby supporting the posited role for THY1 in tissue regeneration. Contrary to the mouse, and consistent with this work, Wu et al. (23) recently reported constitutive THY1 protein and RNA expression in the healthy human kidney tubulointerstitium, which were significantly decreased in diabetic kidneys. Single-cell RNA-seq databases of human kidney confirmed the expression of THY1 in PTs along with other cells such as fibroblasts and immune cells (23). Unexpectedly, increased levels of soluble urinary THY1 were also associated with kidney disease, which was attributed to possible increased shedding of THY1.

Our data demonstrate that in reference kidney tissue, ~20 to 50% of PTs (based on the imaging platform used) express THY1, predominantly in S1 segments. Although we cannot completely rule out that some filtered and reabsorbed THY1 accounts, in part, for that signal, our combined transcriptomics and spatial protein data support actual expression of THY1 in that specific tubular segment. For unclear reasons, the ubiquity of THY1 expression was underrepresented in the snRNA-seq atlas and underscores the important complementary roles of spatial protein imaging and transcriptomics. A much smaller proportion of PTs expresses PROM1 in healthy states. Furthermore, THY1 loss in PTs and a shift toward higher PROM1 expression are associated with kidney disease. Although the increase in PROM1-expressing PTs was relatively small in these temporally cross-sectional biopsies, it could represent a substantial shift when considered over the life span of the disease. THY1-positive and PROM1-positive PTs are likely on opposite ends of a spectrum of disease progression with some overlapping intermediary states during acute injury. These intermediary states can be further stratified on the basis of the expression of additional markers such as VCAM1. The latter has been associated with states of injury and adaptive and maladaptive repair (3, 13). Our data support that the concurrent expression of THY1, PROM1, and VCAM1 marks a central transition state that is still not fully committed to one specific outcome. Our molecular cell-cell interaction analysis is consistent with a regenerative potential in THY1-positive PT cells and pro-inflammatory features for PROM1-positive PT cells. This was further supported in the neighborhood analysis where niches enriched with PROM1-positive PTs had high abundance of immune cells. Our data support a model in which THY1 loss and the shift toward PROM1 expression in PTs could be a valuable indicator of kidney disease progression. Therefore, we propose that dual THY1 and PROM1 assessment in kidney biopsies could serve as a staging tool and a prognostic biomarker for kidney disease progression. This will require further validation studies in larger cohorts.

However, the complex role that THY1 and PROM1 play in various stages of kidney disease can still be confounded by the overlap between AKI and CKD that is frequently observed, both at the clinical level and the actual histopathology. Our data presented in this work also suggest substantial overlap between AKI and CKD. Most AKI encountered in clinical practice frequently occurs in the setting of existing CKD (39). This may explain why THY1 expression is reduced both in AKI and CKD samples. Therefore, correlation of THY1 and PROM1 expression with rigorous grading of histological injury and chronicity within kidney biopsies is needed to fully extract their powerful potential as staging biomarkers. Consortia examining the detailed histopathological changes in kidney biopsies and their link to molecular signatures, such as the KPMP, will likely be the best venue to support such ongoing efforts.

There are several unique approaches and findings in this work that are worth highlighting. In a robust pipeline, we integrated large-scale CODEX multiplexed imaging data from multiple specimens in a single analytical space. Although a spatial protein imaging–based analysis was recently described in diabetic kidneys (40), our approach benefited from cross-validation with orthogonal techniques and datasets. We identified unique PT clusters that express either THY1 or PROM1 with high confidence, amidst other PT subtypes that were classified on the basis of their protein expression profile. To increase the robustness and generalizability of our findings in the CODEX dataset, we used a large PhenoCycler (newer iteration of CODEX) validation cohort. In this larger cohort, we validated the findings from the initial CODEX discovery cohort, showing a significant decrease in THY1-positive PTs and an increase in PROM1-positive PT population in disease.

A strong agreement and alignment between protein and RNA expression data were demonstrated by the integrated analysis based on coregistration of CODEX and ST on consecutive tissue sections. Not only did this analysis confirm the concordance of THY1 protein and RNA expression in PTs, but it also allowed us to investigate protein-defined THY1-positive PTs at full transcriptome depth. This integrative analysis could be adopted for other molecules or even other tissues when consecutive sections are available for interrogation using spatial protein imaging and spatial RNA expression. The concordance of the ST-derived gene expression profile of THY1-positive PT cells in the CODEX data with the transcriptomics snRNA-seq data from the KPMP atlas, which were generated independently, increases the robustness of the findings.

This work also demonstrates the feasibility of charting the trajectories of PTs based on spatial protein imaging. These trajectory and pseudotime analyses in CODEX and snRNA-seq yielded similar results. These congruent findings by independent technologies from two separate datasets not only enhance the robustness of the results but also support the usefulness of trajectory analysis in the CODEX data, especially when validated by snRNA-seq, which is considered a gold standard for such analyses (41).

Integrated neighborhood analysis on the CODEX data showed a negative association between THY1 and immune cells, whereas PROM1-positive PTs were enriched within neighborhoods that have notable immune activity. These findings could be consistent with a failed repair phenotype of PROM1-positive PTs, which are associated with immune activation and expression of injury markers such as KIM1 and VCAM1 (11–14, 16). The cell chat analysis using the snRNA-seq data was also consistent with these observations and further extended our understanding of potential interactions of THY1-positive and PROM1-positive PTs with other cell types. THY1-positive PT cells have reduced receptor-ligand interactions and were mostly centered around SPP1 (osteopontin). The latter predominantly interacts with integrins and CD44 on nonimmune cells. These interactions have been linked to differentiation and transmigration (33), which could be part of the regenerative potential associated with THY1. The reduced ligand-receptor interactions associated with THY1⁺ PTs could also suggest that these cells are predominantly in a quiescent state. Conversely, PROM1-positive PTs exhibited extensive ligand-receptor interactions with epithelial, endothelial, and immune cells using collagen IV subunits that could engage various integrins. These findings support a proinjury and inflammatory profile of PROM1-positive cells. The consistency of findings from the snRNA-seq data cell-cell interaction with the CODEX neighborhood and trajectory analyses (from CODEX and snRNA-seq) is quite notable and provides cross-validation across various assay platforms and datasets.

Our studies have a few limitations that are worth discussing. The CODEX discovery cohort has a relatively small sample size, which could be limiting in terms of sampling bias and generalizability. We mitigated this limitation by including a larger validation cohort, which reproduced and further extended the results. This larger validation cohort is currently being further analyzed and will be the subject of future work. Furthermore, the concordance of the CODEX data with the transcriptomics data validates the findings and the robustness of the results. On the basis of our data, we cannot infer causality on the role of THY1 or PROM1 PTs in repair or regeneration, but the existing experimental data from other groups for both THY1 and PROM1 strongly support our conclusions (19, 24). Last, although the cell chat analyses provide important insights into the biology, these data are associative and require further validation through interventional studies.

In summary, we show that protein imaging modalities such as CODEX offer a robust platform for data analysis that could complement existing RNA-based platforms. Our data point to THY1 and PROM1 as important players in PT biology and could serve as useful biomarkers for disease staging and progression and potential therapeutic targets.

METHODS

Tissue sources and study approvals

Ethical compliance: We have complied with all ethical regulations related to this study. All experiments on human samples followed all relevant guidelines and regulations. The relevant oversight information is given on the basis of tissue sources (detailed below).

CODEX discovery cohort

Detailed analyses were done using CODEX imaging on a cohort of biopsies or reference kidney tissue consisting predominantly of cortex. Human kidney tissue biopsies and nephrectomies were preserved in optimal cutting temperature (OCT) medium and were acquired from the Biopsy Biobank Cohort of Indiana, under a waiver of informed consent as approved by the Indiana University Institutional Review Board (IRB; IRB #1906572234). Reference nephrectomy tissues were also obtained from the KPMP central biorepository. Three reference kidney tissues, also preserved in OCT, were obtained via percutaneous nephrolithotomy carried out in patients with stone disease, as part of an ongoing study (Indiana University IRB protocol #1010002261). For relevant demographic information about each tissue donor and the corresponding tissue source, please see table S1.

PhenoCycler validation cohort

Kidney biopsies imaged with Akoya Biosciences PhenoCycler-Fusion 2.0 from the KPMP atlas were used for validation of the changes of THY1 and PROM1 staining in PTs. Imaging data are publicly available in the KPMP atlas. Human samples collected as part of the KPMP consortium were obtained with informed consent and approved under a protocol by the KPMP single IRB of the University of Washington (IRB #20190213). We also used reference tissue from HuBMAP (https://hubmapconsortium.org/hubmap-data/). Samples as part of the HuBMAP consortium were collected using informed consent by the Kidney Translational Research Center under a protocol approved by the Washington University IRB (IRB #201102312). Detailed analysis of this cohort (table S3) will be published elsewhere as part of a Spatial Atlas output from KPMP.

snRNA-seq sample source

A combined snRNA-seq atlas from HuBMAP (https://hubmapconsortium.org/hubmap-data/) and KPMP (www.kpmp.org) datasets (200,338 nuclei) with 41 samples from 36 subjects was used (1). The medulla and papilla were excluded, and the analysis was limited to the cortex and corticomedullary junction because kidney biopsies predominantly recover these regions. This dataset was previously published and is publicly available (GSE183277).

ST sample source

Two reference tumor nephrectomy tissues were also obtained from the KPMP central biorepository (table S1).

Antibody selection, conjugation, and validation

CODEX

Following a careful selection of structural, immune, and injury markers, based on available literature, public data domains, and previous work, we composed an initial panel of 23 markers that was applied to all specimens. This panel was subsequently extended to include more specialized markers (total of 38) to label cell types and cell states (see table S1 for the marker panel used for each specimen and table S2 for details about the panels).

Twenty-six of the antibodies from the final panel were conjugated in house using the protocol outlined by Akoya Biosciences and Black et al. (42). Commercially purchased antibodies first underwent a reduction step using a Reduction Master Mix (Akoya Biosciences). Lyophilized barcodes were then resuspended using Molecular Biology Grade Water and Conjugation Solution. The barcode solution was then added to the reduced antibody solution and incubated for 2 hours at room temperature. After incubation, the newly conjugated antibody barcode was purified in a three-step wash/spin process and stored at 4°C. Successful conjugation was validated via gel electrophoresis, immunofluorescent staining, and confocal imaging.

PhenoCycler

The antibody panel used in CODEX was further expanded to 42 markers. Forty-one markers used in this analysis (nestin was excluded) have been validated and published as part of Organ Mapping Antibody Panels (43) (https://doi.org/10.48539/HBM542.NNZP.924) and are shown in table S4.

Tissue preparation and imaging

CODEX

Ten-micrometer-thick human kidney tissue sections were cut from OCT blocks onto poly–l-lysine–coated square coverslips, which were subsequently processed using the protocol outlined by Akoya Biosciences and as described previously by Goltsev et al. (44) and Melo Ferreira et al. (8). Antigen retrieval was conducted with a three-step hydration process, followed by fixation with 1.6% paraformaldehyde after initial fixation, and an antibody cocktail of the markers listed in table S2 was dispensed among the coverslips. The antibody solution was left on the coverslips overnight at 4°C. The following day, the staining solution was washed from the tissues, and a second fixation was performed. Oligonucleotide probe staining and fluid handling were performed with the CODEX system from Akoya Biosciences. Automated tile-scan imaging of the tissue between probe staining rounds was performed on a Keyence BZ-X810 microscope fitted with a 20× objective. The resulting images were processed using the CODEX Processor (Akoya Biosciences) and visualized using Fiji/ImageJ (45).

PhenoCycler

Tissue processing for the PhenoCycler follows the same protocol as described above for the CODEX system, but the poly–l-lysine–coated coverslips are replaced by SuperFrost Gold+ charged slides. After staining, slides were covered with an Akoya Biosciences FlowCell. Cover-slipped slides are incubated in 1× CODEX Buffer with Buffer Additive (Akoya Biosciences) for 10 min to ensure proper adhesion of the FlowCell to the slide. Slides were either kept in storage buffer until ready for imaging or imaged immediately after the incubation. Imaging of the tissues was conducted with a 20× objective fitted on the Akoya Biosciences PhenoCycler 2.0 microscope and fluidics handler. Image stitching and processing were also performed with the PhenoCycler 2.0.

Image analysis

CODEX

CODEX images were brought into ImageJ/Fiji and analyzed using a custom-made plug-in: Volumetric Tissue Cytometry and Exploration (VTEA) (7, 46). Nuclei stained with 4′,6-diamidino-2-phenylindole (DAPI) were segmented with VTEA. A morphological dilation around each segmented nucleus was used as a proxy segmentation of cytosol. Using these nuclear or cytosol segmentations, the mean intensity in aligned images from additional CODEX rounds was calculated to detect protein markers in the putative nuclei and cytosol of individual cells. The quality of individual protein markers was assessed on a channel-by-channel basis. Channels were excluded from the dataset with aberrations including (i) no detectable signal, (ii) signal correlated with background images, or (iii) imaging artifacts (e.g., exposure inconsistency across image tiles). Furthermore, the semiunsupervised clustering of segmented cells with measured protein markers was used to help identify aberrations (e.g., cell clusters that correlated with inconsistent exposures across tile scanning). For this, cells were projected into t-distributed Stochastic Neighbor Embedding (t-SNE) space with a learning rate of 100 and a perplexity of 40, and clustering was conducted using Ward hierarchical clusters based on the same feature space, with a maximum of 15 clusters and mapped back to the image to identify correlations with artifacts.

After tissues were processed individually and artifacts were removed, measurements from segmented cells were imported into R and combined into a single dataset following z-scaling. Combined datasets were clustered using the Louvain methods and a feature space including the mean intensity of CD31, CD3, CD45R0, CD4, CD45, AQP1, CD11c, THY1, α–smooth muscle actin (α-SMA), E-cadherin, PROM1, podocalyxin (PODXL), LRP2, CD68, Uromodulin (UMOD), IGFBP7, and cytokeratin 8. After labeling the major clusters, the PT clusters were extracted as subsets and relabeled using a tubular epithelium–specific set of features that included the mean intensities of AQP1, THY1, α-SMA, E-cadherin, PROM1, LRP2, CD68, UMOD, IGFBP7, and cytokeratin 8.

PhenoCycler

Final “qptiffs” images generated by PhenoCycler-Fusion (Akoya Biosciences) were imported into ImageJ/Fiji for cropping biopsies into individual image files. Segmentation of nuclei based on DAPI staining was done using VTEA, as with the CODEX discovery cohort. Cortical and medullary regions were defined for each tissue to be used as additional features in downstream analysis. The resulting segmentation from all biopsies was imported into R, and data were z scaled. Forty-one markers and the region feature (cortex or medulla) were used to cluster the scaled data of 49 samples using FastPG (https://github.com/sararselitsky/FastPG), with k = 100. After clustering, uniform manifold approximation and projection (UMAP) embeddings were calculated, and data were plotted along with violin plots of both scaled data and raw intensity values to identify clusters. Cluster identities were further validated by backmapping the clusters onto representative tissue sections. Similar to the CODEX analytical pipeline, PT clusters (based on LRP2 and AQP1 expression) were extracted into subsets and reclustered. The feature space for this subclustering was the same as the one used in the CODEX discovery cohort in addition to the region (cortex or medulla). Clustering was done using FastPG with k = 500, which still yielded a high number of clusters because of the larger sample size. Redundant clusters based on the expression of THY1, PROM1, and IGFBP7 were further grouped together.

Cell neighborhood analysis in CODEX data

The PT subclusters were imported back into VTEA for neighborhood generation and analysis (7). The “spatial by cell” method in VTEA was used, with a neighborhood diameter of 55 μm to match the spot size used in Visium ST. VTEA generates the counts (“class sums”) and the fraction of the total of each cell type for each neighborhood, which were imported into R for further analysis. To determine which cell types were spatially associated in pairs, correlation matrices were calculated for all samples (corrplot v. 0.92). Chord plots (circlize v. 0.4.16) were also generated for cell types of interest, informed by the correlation matrices. For neighborhood analysis, the class sums were clustered with Louvain (igraph v. 1.5.1) and t-SNE (Rtsne v. 0.16) embedding, and neighborhood clusters or niches were identified. Stacked bar plots were used to visualize the distribution of cell types within each niche, and the odds ratios were calculated to assess the prevalence of a given niche in the reference versus disease cohorts.

Spatial transcriptomics

ST was done on three 10-μm sections obtained from two human reference kidney tissue specimens. The preparation and imaging of tissues were done using the Visium Spatial Gene Expression protocol for polyadenylate capture in fresh frozen tissue (10x Genomics, CG000240 protocol, Visium Tissue Preparation Guide) (8).

Using a Keyence BZ-X810 microscope equipped with a Nikon 10× CFI Plan Fluor objective, hematoxylin and eosin (H&E)–stained bright-field mosaics were collected and stitched before mRNA capture. mRNA was extracted from tissue after 12 min of tissue permeabilization. Isolated mRNA was captured by oligonucleotides in the fiducial capture areas and reverse transcribed. Libraries were prepared on the basis of the Visium 1.0 protocol (10x Genomics, CG000239 protocol), which included second-strand synthesis, denaturation, cDNA amplification, and cleanup using SPRIselect (Beckman Coulter). Multiplexed sequencing was performed on a NovaSeq 6000 (Illumina). Sequencing data were demultiplexed and mapped to the reference genome GRCh38 3.0.0 and counted in Space Ranger (10x Genomics, v1.0.0). Data processing was performed in Seurat (v.4.4) (47).

The snRNA-seq kidney atlas data, generated by KPMP and HuBMAP consortia (1), were mapped to the ST. SCTransform was used for read normalization (47, 48). In addition, dimensionality reduction was done with Seurat’s implementation of UMAP, and the expression signatures of spots for cell types were calculated and mapped using a transfer score system based on the Seurat anchor method (47). This score shows the association of the transcriptomics profile of each spot with a specific cell type. A higher transfer score implies that a greater proportion of a specific cell type is mapped to a particular ST spot.

The expression datasets, obtained from the mappings of each section, included the gene counts and sample metadata. Three datasets from each section were merged, and a UMAP was regenerated. Batch correction was accomplished by normalizing the raw counts of the merged object with SCTransform (47, 48). In an additional limited analysis aimed at identifying regions of high THY1 expression in an ST sample, spots were highlighted using the upper 0.5 quantile threshold for THY1 expression.

Coregistration and integrated analysis with CODEX

To coregister ST and CODEX, three consecutive ST and CODEX sections from two separate reference tissue specimens were used. In the CODEX data, LRP2, THY1, and PODXL staining were used to outline PTs, THY1-positive PTs, and glomeruli, respectively.

ST-CODEX alignment was done manually using Photoshop (Photoshop v.23.3.1). CODEX images composed of LRP2, THY1, and PODXL were aligned with the H&E images of the ST slide using linear transformations to align landmark structures, such as glomeruli and large tubules. Accordingly, other structures were overlayed precisely. Using this alignment, the spots in ST samples were categorized on the basis of expression of THY1 and LRP2 in CODEX. Using Loupe Browser (v.6), four groups of spots were defined as LRP2 positive, THY1 negative; LRP2 positive, THY1 positive; the edge layer of tissues; and the rest of the cells. The last two groups were excluded from the analysis.

To find the differentially expressed genes (DEGs) in the two groups of LRP2-positive THY1-negative and LRP2-positive THY1-positive spots, the “FindMarkers” Seurat function was used with the P value of 0.05 and average log2 fold change > 0.25.

Correlation analysis of THY1 expression between ST and CODEX

Correlation between THY1 protein and RNA expression was done using the three consecutive ST and CODEX sections described above. To perform this analysis, a semiautomated registration was used to align the CODEX and ST consecutive sections. The CODEX sections were transformed to align with the consecutive ST section using the BigWarp plugin in Fiji (ImageJ 1.52p). Regions of interest (ROIs) overlaying each ST spot were identified, linked, and numbered using ImageJ. First, we used THY1 protein fluorescence from ROIs matching the LRP2-positive ST spots defined above that were manually selected and used in the analysis in Fig. 3. We also used an unbiased methodology to define LRP2-positive ST spots based on Otsu thresholding of LRP2 fluorescence intensity in all the ROIs. We used a 10% Otsu threshold of LRP2 expression, which was visually validated on the tissue to ensure that mostly PTs and no other structures were identified in the ROIs. In both analyses, THY1 expression values from the ST spots and the corresponding THY1 mean fluorescence intensities from the CODEX ROIs were log₁P transformed and then used to generate correlation plots using the ggplot2 package in R. The Pearson correlation method was implemented to evaluate the strength, direction, and statistical significance of the correlations.

snRNA-seq analysis of PTs from KPMP atlas

The cortical subset of PTs including 29 specimens was taken from the snRNA-seq atlas (1). To compensate for batch effects, the Seurat anchor method was applied to integrate samples in this cortical PT subset. PTs were clustered using the Louvain algorithm with a resolution of 0.5. On the basis of THY1 differential expression, cluster 5 was selected as a THY1-positive subclass of PT cells. Expression of various genes of interest (uncovered by ST analysis or related to PT subsegment identity or cell state) was visualized in a Dot Plot. Using the FindMarkers function, the DEGs were found in this cluster compared to other 13 clusters. Pathway enrichment was done for the top 33 up-regulated genes. The R package pathfindR (49) was used for pathway analysis using the Kyoto Encyclopedia of Genes and Genomes (KEGG) and Reactome databases.

In addition, reference, AKI, and CKD samples in PTs were analyzed independently (unintegrated analysis) using the Seurat anchor method. PT subtypes within each cohort were identified with a clustering resolution of 0.5. Expression of PT markers, injury-related genes, and other genes of interest across reference, AKI, and CKD clusters was presented in a combined PT dataset using Dot Plot.

Trajectory analysis

Trajectory and pseudotime analyses of CODEX, snRNA-seq, and PhenoCycler data were performed in R using the slingshot (50) (v.2.6.0) and tradeSeq (41)(v.1.12.0-1.18.0) packages.

CODEX

Trajectory and pseudotime analyses were implemented on unintegrated reference and disease samples, as well as an integrated dataset combining both using the average expression of AQP1, THY1, α-SMA, E-cadherin, PROM-1, LRP2, CD68, UMOD, IGFBP7, and cytokeratin 8. Two different clusters were selected as the start point of trajectories including clusters with the highest average expression of LRP2 and the THY1-positive PT cluster. Feature plots showing the expression of THY1 and PROM1 were generated.

PhenoCycler

Trajectory and pseudotime analyses were conducted on the integrated dataset combining reference and disease samples starting from VCAM1-positive PT cells coexpressing THY1 and PROM1. Expression of VCAM1, THY1, and PROM1 was visualized using feature plots.

Single nuclear RNA-seq

Trajectory and pseudotime analyses were performed on unintegrated datasets of PTs from reference, AKI, and CKD samples, as well as on an integrated dataset combining all three conditions initiating from two separate clusters: one starting from a THY1-positive PT cluster and the other from a cluster with high LRP2 and low injury markers.

Cell-cell communication

We evaluated the interactions of cells in THY1-positive, PROM1-positive, S1, or S2 PTs with other kidney cell types in the snRNA-seq atlas using the CellChatDB package (v 1.6.1) (51). This approach defines significant interactions between ligands on the source cells with receptors on target cells.

Statistics

Unless otherwise specified, values are reported as mean ± SDs. Bar graph scatter plots were generated using GraphPad Prism (10.0.2). Identification of DEGs between groups was performed with a negative binomial generalized linear model. Statistically significant gene changes were assigned using an adjusted P value of <0.05 with Bonferroni correction. A two-way analysis of variance (ANOVA) model with Tukey’s test for multiple corrections was used to compare the different cell proportions between conditions in the CODEX and PhenoCycler data using adjusted P < 0.05. For cell communication analysis, significance was derived from a one-sided permutation test (the CellChat package default). Association of disease conditions with neighborhoods in CODEX was determined using the odds ratio.

Supplementary Materials

The PDF file includes:

List of the Kidney Precision Medicine Project collaborators and affiliations

Figs. S1 to S17

Legends for tables S1 to S4

Other Supplementary Material for this manuscript includes the following:

Tables S1 to S4