Single-cell spatial mapping of human kidney development implicates the microenvironment in guiding cell fate decisions

Abstract

Evolution has used cell–cell communication as a strategy to coordinate organ development, enabling the reproducible generation of intricate structures. Classically, these interactions have been studied one at a time in model organisms, limiting our understanding of how cellular interplay coordinates human development. We investigated human kidney development using single-cell RNA sequencing and spatial transcriptomics, analyzing over 700,000 cells. By mapping gene expression and differentiation trajectories in space, we define the spatial organization of kidney development. Our analysis revealed unrecognized plasticity, showing that cell fate established during early patterning can be later revised. This plasticity provides a potential mechanism for how cell fate is robustly established in complex patterned tissues. Additionally, through a genome-wide, spatially aware cell–cell interaction analysis, we link localized ligand signals to cell fate decisions. We also define biologically meaningful cellular neighborhoods based on aggregated extracellular cues, providing a blueprint to understand the coordination of human development at scale.

The spatial organization of human kidney development is explored using single-cell RNA sequencing and spatial transcriptomics, highlighting the crucial role of microenvironmental signals in guiding cell fate.

Main

The kidney is a spatially complex organ with over 30 specialized cell types¹. Its functional unit is the nephron. The number of nephrons that healthy individuals are born with varies between 200,000 and 2,000,000, with low nephron endowment correlating with increased risk of developing hypertension and chronic kidney disease². Grossly abnormal kidney development causes congenital anomalies of the kidney and urinary tract, which are the leading cause of chronic kidney disease in children³. However, research on human kidney development has technical limitations⁴.

Nephron formation is spatially choreographed. Nephrons are formed from cells derived from two separate progenitor populations located in the outermost region of the developing kidney. The proximal part of the mature nephron—including the renal corpuscle and a large portion of the tubule—is formed from nephron progenitor cells (NPCs), while the ureteric bud (UB) gives rise to portions of the connecting tubule and collecting system⁵. NPCs form a cap of mesenchyme atop a UB tip. NPCs then move adjacent to the UB tip and begin to epithelialize into primitive aggregates. These aggregates then adopt increasingly complex patterned, yet reproducible, spatial characteristics, forming renal vesicles, comma-shaped bodies and S-shaped bodies⁴, before forming the mature nephron^1,6. The NPC-derived and UB-derived segments must connect to form a functional conduit for the filtered plasma⁷.

Mouse studies have identified key cell–cell and ligand–receptor interactions critical for kidney development, with deletion of specific cell–cell junction proteins and ligands causing major developmental defects in the kidney^8–11. However, despite years of careful study, the full extent of these mechanisms is unclear in mouse kidney development⁸, and poorly understood in human kidney development. While developmental pathways are thought to be largely conserved between mouse and human, key biological differences have been noted^6,12.

Recent studies have used single-cell RNA sequencing (scRNA-seq) to understand and describe human kidney development, using concepts like pseudotime for trajectory analysis^13–16. Studies performed in mice indicated that NPC-derived intermediates adopt renal corpuscle and tubule fates, further differentiating into podocytes and tubule cells (proximal tubule (PT), loop of Henle (LOH), distal convoluted tubule (DCT)). In contrast, human studies have suggested that intermediates take on proximal versus distal fates, with proximal cells giving rise to podocytes and PT cells^17–19.

By integrating classic histological analysis with single-cell transcriptomics and single-cell spatial transcriptomics, we defined genetic programs associated with NPC aging and changes during differentiation and renewal. We found that there is an early decision between pretubular (PT, DCT and LOH lineages) and prerenal corpuscle (parietal epithelial cell (PEC), and podocyte) fates. Notably, we identified developmental plasticity, showing evidence that cells with a mature PT transcriptional state convert into PECs. We mapped developmental trajectories in space, identifying locations where trajectories diverge and their relationship with ligand signals, thereby mapping cell fate decisions.

Results

scRNA atlas of the developing human kidney

We performed scRNA-seq using five human fetal kidneys with gestational ages ranging from 12.5 to 20.5 weeks (Supplementary Table 1 and Supplementary Fig. 1). We then generated a new atlas containing 65,289 single cells that contained cell types from five distinct parent lineages—nephron, UB, stromal, endothelial and immune cells (Fig. 1a). We further classified nephron cell subtypes into NPCs, podocytes, PECs, PT, LOH cells and DCT, and intermediates (Int; Fig. 1b,c). Cell-type-specific markers (Fig. 1b,c, Supplementary Table 2 and Supplementary Fig. 2a,b) were consistent with the literature^12,18–22.

Fig. 1. scRNA atlas of the developing human kidney.

a, UMAP of scRNA-seq data of 65,289 cells after filtering and quality control shows populations identified within human fetal kidneys. X axis and y axis denote UMAP dimensions 1 and 2, respectively. Cell types are colored and labeled. b, Feature plots showing the number of raw counts for selected marker gene, as indicated by color scale. The following genes are shown: CITED1, LHX1, RET, CUBN, WNK1, UMOD, NPHS2, CFH and AQP2. c, Dot plot expression of representative cell-type markers in each cell population. The dot size indicates the fraction of cells expressing the gene. The color of the dot shows mean log counts. d, Pathway analysis of time-dependent genes for NPC and PT cell types. Percentage of genes within pathway is shown by dot size, with color indicating negative log FDR. Combined score is indicated by x axis. Int, intermediate nephron cells; UE, ureteric epithelium; UMAP, Uniform Manifold Approximation and Projection; FDR, false discovery rate; CITED1, carboxy-terminal domain 1; RET, ret proto-oncogene; CUBN, cubilin; WNK1, WNK lysine-deficient protein kinase 1; UMOD, uromodulin; CFH, complement factor H; AQP2, aquaporin 2.

NPC renewal and differentiation are key determinants of nephron number. The NPC population is exhausted in humans by late gestation and in mice perinatally^23,24. Gestational age-associated expression changes in NPCs have functional importance in mice²⁵, and here we identified 467 genes with significant age-dependent changes in human NPCs (Supplementary Table 3). Pathway analysis of NPC age-dependent genes showed higher bone morphogenetic protein (BMP)/WNT pathway activity, forkhead box protein O1 (FOXO1) and insulin signaling with advancing gestational age (Fig. 1d). BMP and WNT signaling are known to have a crucial role in NPC epithelialization and early nephron differentiation, indicating the correlation between gestational age and differentiation⁸. While FOXO1 has not been shown to have a role in NPCs²⁶, FOXO1 has an important role in insulin signaling, metabolism and WNT signaling^27,28.

We calculated gene expression changes associated with gestational age for each cell population (Supplementary Fig. 2c and Supplementary Tables 9–14). Pathway enrichment analysis of these genes showed that in maturing PT cells, PT-specific solute transport, peroxisome proliferator-activated receptor signaling and carbon metabolism pathways increased with gestational age (Fig. 1d and Supplementary Fig. 2d). We found that pathways associated with proliferation became less prominent with gestational age (Supplementary Fig. 2d).

Cellular differentiation trajectories of the fetal human kidney

We then estimated cellular differentiation trajectories using CellRank, which calculates the long-term fate probabilities for all cells²⁹. This contrasts with other methods, which only determine discrete branch points and therefore struggle to model plasticity^30,31. This is important as NPCs give rise to multiple terminal fates, which could interconvert^{5,14,19,20,32}.

NPC progeny pattern tubule and glomerular nephron segments, yet uncertainty exists over the contribution of UB-derived and NPC-derived progenitor sources to the patterning of distal tubule and connecting segment^5,33. Similar to recent studies, we observed a population of distal nephron and intermediate cells that created a continuous bridge between UB and NPC, making it difficult to precisely partition these lineages (Fig. 1a)^12,18,19,34. To partition these lineages, we used CellRank to calculate the likelihood that each cell was derived from the UB population or the NPC population, finding intercalated cells (IC) and some connecting tubule cells belonging to a UB lineage, while the remaining distal nephron cells were assigned to the nephron lineage (Supplementary Fig. 1f)^33,35.

We then examined cellular differentiation dynamics for the NPC lineage, calculating the probability for every nephron cell to transition into each mature cell type (podocyte, PT, LOH or DCT; Fig. 2a,b and Supplementary Fig. 3). NPCs transition to an early intermediate state and then take on either a renal corpuscle or a tubular fate, consistent with observations from mouse studies¹⁹. We examined the UB lineage observing trajectories for the CNT, IC and urothelium fates (Supplementary Fig. 4).

Fig. 2. Cellular differentiation trajectories of the developing kidney.

a, UMAP of nephron cell populations with RNA velocity. Colors indicate cell-type annotations, and arrows indicate RNA velocity direction and intensity (longer arrows indicate greater velocity). b, CellRank absorption probabilities for transition from NPCs to each differentiated cell state (Podo, PT, LOH and DCT) plotted on UMAP of nephron cells. Absorption probability value for each cell is indicated by color, with yellow color indicating higher absorption probability. c, Directed map of NPC differentiation showing direction and connectivity of Leiden clusters. Cells are colored by Leiden clusters and each cluster has a pie chart showing an aggregate of cell fate probability of all cells within the Leiden cluster, with each terminal state’s likelihood indicated by color within the pie chart. Arrows indicate likely transitions, with darker arrows being more frequent than thin arrows, which are more frequent than dotted lines. d, Absorption probabilities for all cells with potential terminal states of NPC and any differentiated state (aggregate of Podo, PT, LOH or DCT). Absorption probability indicated by color on a scale of orange (high NPC probability) to blue (high differentiated state) plotted on a UMAP of nephron cells. e, Volcano plot examining gAge-dependent expression in NPCs with renewal genes marked by orange color. Several example genes are labeled. P values are shown for each β coefficient determined using PyDESeq2, regressing gestational age against each gene expression. Multiple testing correction was performed for each gene examined. f, Volcano plot examining gAge-dependent expression in NPCs with differentiation genes marked by orange color. Several example genes are labeled. P values are shown for each β coefficient determined using PyDESeq2, regressing gestational age against each gene expression. Multiple testing correction was performed for each gene examined. g, UMAP of nephron cells shows a small number of annotated PECs located in the lower portion PT cluster (inset) at the terminus of RNA velocity vectors. h, CellRank absorption probabilities for all cells with potential terminal states of renal corpuscle (Podo) and any tubular state (aggregate of PT, LOH or DCT). Absorption probability indicated by color on a scale of orange (high tubule probability) to blue (high renal corpuscle probability), showing a small population of PT cells with a high probability of becoming an RC. Below is a UMAP showing macrostates, demonstrating a specific metastable cell state labeled PT2 that spans PT and PEC populations. i, Immunofluorescence of human fetal kidney tissue showing canonical PT marker ACE2 expression (purple), and PEC/Podo marker WT1 expression (green) and DAPI (white). Scale bar = 100 µm. This was repeated twice with this antibody set, with similar results, and additionally with other antibodies for ACE2/WT1 (Extended Data Fig. 1). gAge, gestational age; Podo, podocyte.

We then calculated the probability of each cell to transition to either a tubule fate (PT, LOH or DCT) or a renal corpuscle fate (podocyte). The analysis identified populations of immature cells—referred to as intermediate cells—that had already adopted prerenal corpuscle and pretubular characteristics (Fig. 2b,c). This analysis indicated a binary decision of NPCs becoming intermediates for the PT, LOH and DCT populations (pretubular), while the other cells were intermediates transitioning to podocytes and PECs (prerenal corpuscle; Fig. 2c,d and Supplementary Fig. 3), consistent with some previous models^12,19.

We then correlated the absorption probability for each trajectory with each gene’s expression to identify potential lineage drivers (Supplementary Tables 10 and 11). For the LOH trajectory, we identified Iroquois homeobox 1 (IRX1), which is thought to be an important driver of nephron morphogenesis³⁶. However, many of these genes were markers of the mature cell type and unlikely to be actively mediating cell fate, such as podocin (NPHS2) for podocytes.

NPC renewal and differentiation are critical for nephron formation. Therefore, we used CellRank to calculate the likelihood of each cell becoming an NPC (renewal) compared to the likelihood of that cell becoming any differentiated state (PT, podocyte, LOH or DCT; Fig. 2d), and genes associated with renewal (Supplementary Table 12 and Supplementary Fig. 2a). We found that genes associated with renewal had higher expression in earlier gestational ages (Fig. 2e and Supplementary Fig. 5). Conversely, expression of differentiation associated genes were higher at later gestational ages (Fig. 2f and Supplementary Fig. 5).

We then re-examined what seemed to be an inconsistency within our trajectory analysis. A small number of PECs were located near the PT cluster on the original UMAP (purple dots at the bottom of the PT cluster; Fig. 2a and Supplementary Fig. 3). This led us to suspect that our two-dimensional UMAP did not fully represent similarities between PTs and PECs. We generated a three-dimensional UMAP that showed the PEC population as a continuum linking PT and podocyte clusters, further suggesting that some PT cells may transition to PECs, and perhaps onto podocytes (Supplementary Data 1).

We then performed a random walk analysis, which showed direct conversions between PT and PECs (Supplementary Fig. 3b,c). RNA velocity indicated that this transition is directed from PT to the PEC fate, rather than from PEC to PT (Fig. 2g). Consistent with this, we identified a metastable macrostate of cells that overlapped the PT and PEC clusters (Fig. 2h). To examine these predictions, we performed immunofluorescence on human fetal kidney tissue and found that some renal corpuscles were lined with ACE2-positive cells—a marker of PT cell identity—that suggested that some morphologic PECs showed molecular characteristics of PT cells, supporting the predicted state transition (Fig. 2i and Supplementary Fig. 6). This ACE2+/PEC pattern was absent in adult human kidneys indicating that such double positive cells may represent a fetal-specific phenomenon (Extended Data Fig. 1).

Extended Data Fig. 1. Additional immunofluorescence images of WT1 and ACE2 in human adult, and fetal kidney.

WT1 (green), ACE2 (purple), and DAPI (white) were used to stain sections of human fetal, adult human. WT1: Wilms tumor 1, ACE2: angiotensin converting enzyme 2. Scale = 100 µm.

To investigate this further, we used spatial transcriptomics to determine whether PT genes are widely activated in morphological PECs. We analyzed three human fetal kidneys (gestational ages of 15–19 weeks) using the CosMx platform, as well as one human fetal and one pediatric kidney with the Xenium platform. We integrated these spatial datasets with our scRNA data (Fig. 3a, Supplementary Fig. 6 and Supplementary Table 1). These data show robust cell typing in space and on UMAP (Fig. 3a,b). To avoid the limitations of single-marker validation, we quantified PT identity using a composite PT score derived from canonical PT transcripts (HNF4A, CUBN, LRP2, ACE2) and measured simultaneously at single-cell resolution using RNA-based spatial transcriptomics. We observed renal corpuscles that housed morphologic PECs that demonstrated widespread expression of PT markers within the fetal sample (Fig. 3c and Extended Data Fig. 2). Notably, these PT markers were not detected in renal corpuscles in pediatric kidney tissue (Fig. 3d and Extended Data Fig. 2)³⁷, suggesting against nonspecific marker expression or technical spillover.

Fig. 3. Spatial transcriptomics demonstrates fetal-specific widespread PT molecular expression patterns within the renal corpuscle.

a, UMAP of human fetal kidney cell types from integrated scRNA-seq, CosMx and Xenium modalities. Each cell type is colored and labeled. Right, dot plot of these data, showing expression of marker genes within annotated populations. b, Human fetal kidney shown in space from fetal and pediatric kidney samples analyzed using Xenium spatial transcriptomics. Cell types are indicated by color. Scale bar = 100 µm. c, Gene expression of proximal tubular cell marker score shown in space (low, transparent; high, yellow) in human fetal tissue. Images also show staining of DAPI (gray), and cell-interior 18S RNA (white). Scale bar = 100 µm. d, Gene expression of proximal tubular cell marker score shown in space (low, transparent; high, yellow) in human pediatric tissue. Images also show staining for DAPI (gray) and cell-interior 18S RNA (white). Scale bar = 100 µm. e, Schematic of mouse construct for lineage tracing experiments. The ROSA26 allele drives expression of fluorophores that are determined by random recombination events allowing for labeling of individual clones. f, Schematic showing the timing strategy for induction and evaluation of clones for the mouse lineage tracing experiment. g, Representative image of traced clones in a mouse renal corpuscle. The image shows a glomerular anlage entirely composed of a single clone of cells colabeled in blue and red (in homozygous Confetti mice), demonstrating a common clonal origin for PECs, PT and Podo. Scale bar = 10 µm. Data were obtained from six independent experiments. In each experiment, one mother was induced, and kidneys from at least four pups were analyzed. h, Representative image of mouse clones showing a single yellow clone constituted of proximal tubular cells, PECs and some contiguous Podo demonstrating their shared origin from a single precursor. Scale bar = 25 µm. Data were obtained from six independent experiments. In each experiment, one mother was induced, and kidneys from at least four pups were analyzed. i, Representative image of mouse kidney clones stained with WT1 (white). The image shows a glomerular anlage entirely composed of a single clone of red-labeled cells, some strongly positive for WT1 (primordial Podo) and others weakly positive (PECs). Primordial tubular cells are visible within the developing glomerular tuft as labeled. Scale bar = 25 µm. Data were obtained from six independent experiments. In each experiment, one mother was induced, and kidneys from at least four pups were analyzed. j, Schematic showing previous models of nephron development, and our updated model that incorporates our identified plasticity. Prox PTA, proximal pretubular aggregate; dist PTA, distal pretubular aggregate; pre-RC, prerenal corpuscle cell; peds, pediatric.

Extended Data Fig. 2. Xenium spatial transcriptomics of proximal tubular markers in fetal and pediatric human kidney tissue.

a, Gene expression of composite (HNF4A/CUBN/LRP2/ACE2) is shown in space (low: transparent, high: yellow) for representative fetal renal corpuscles. b, Gene expression of composite PT score (HNF4A/CUBN/LRP2/ACE2) is shown in space (low: transparent, high: yellow) for representative pediatric renal corpuscles. c, PT marker gene expression shown individually in space for a fetal renal corpuscle. d, PT marker gene expression shown individually in space for a pediatric renal corpuscle. DAPI is colored gray and 18S RNA is shown in white to represent intracellular regions. Scale bar = 100 µm.

To determine whether this phenomenon is conserved in mice, we performed trajectory analysis on a mouse single-nucleus RNA/single-nucleus ATAC dataset³⁸. After identifying robust cell types (Supplementary Fig. 7a–c), we found that PECs were split between two clusters, one near the PT cluster and one near the podocyte cluster, with random walk analysis (Supplementary Fig. 7d), and increased PT absorption probability of the PECs that cluster near podocytes (Supplementary Fig. 7e) further supporting that these cell types interconvert.

We then performed clonal lineage tracing using the Pax2rtTA;TetO.Cre/R26.Confetti mouse model, previously used in developmental studies^39,40. The Pax2-rtTA;TetO.Cre/R26.Confetti transgenic mouse model labels renal progenitors during development with one of four fluorescence proteins (YFP/yellow, CFP/cyan, RFP/red or GFP/green) upon doxycycline induction, enabling single-cell and clonal tracking of Pax2+ cells (Fig. 3e). Cells were labeled starting at embryonic day 8 (E8) and clones were examined at postnatal day 6 (P6; Fig. 3f). We noted single clones encompassing morphologically defined PT, PEC and podocyte cell types within a single nephron (Fig. 3g–i). Given that each nephron contains multiple clones, these labeled clones do not reflect whole-nephron labeling, but rather clonal contributions to multiple epithelial segments. These results show that progeny of a single nephron progenitor can contribute to both renal corpuscle and tubular-fated cells, challenging the proposed model discussed in ref. ¹⁹. (Fig. 3j). Collectively, our mouse and human data provide strong evidence that some cells within the fetal renal corpuscle—although morphologically resembling PECs—express PT genes and resolve into a PEC identity over time, as alternative explanations like cell death or migration are unsupported by our genomic and histologic data. These findings necessitate a revision of the current nephron differentiation model.

Spatial molecular characterization of the developing human kidney

Cells interpret their spatial position within a tissue to guide fate decisions, ensuring proper tissue organization and function⁴¹. To examine how this may occur within the kidney, we integrated our CosMx and scRNA datasets using scVI and scANVI⁴², which enabled joint embedding, transcriptome-wide gene expression imputation and trajectory imputation for the spatial data (Extended Data Fig. 3 and Supplementary Fig. 8b).

Extended Data Fig. 3. Integration of scRNA-seq and CosMx data to allow for full transcriptome imputation.

a, Cartoon depicting integration strategy, showing merging of CosMx and single-cell raw count matrices using shared genes. scVI and scANVI were used to generate a shared embedding in latent space. Using the latent space for each cell, nearest neighbors with the opposite technology were identified and were used to calculate a weighted average for numeric data such as gene expression or pseudotime calculations. b, Shared embedding of scRNA-seq and CosMx data colored by modality and by cell annotation. c, UMAP showing expression of measured CosMx counts and imputed counts for the CosMx cells. Expression is indicated by color of each cell. d, UMAP of original scRNA-seq with cell annotations and absorption probability for tubular and renal corpuscle fates as indicated by color (above). Below is the UMAP embedding of the integrated object showing cell annotations and imputed CellRank absorption probability. e, Raw and imputed expression of WT1 show in space, followed by imputed CellRank absorption probabilities in space for the renal corpuscle and tubular trajectories. Values indicated by color. PEC: parietal epithelial cell, PT: proximal tubule, LOH: loop of Henle, DCT: distal convoluted tubule, UB_CT: ureteric bud and collecting duct, Int: intermediate, CD: collecting duct.

Using this atlas, we annotated cell types (Fig. 4a and Supplementary Figs. 9a,b and 10). Spatial locations of cell types were consistent with known tissue organization and histology (Fig. 4b, Extended Data Fig. 4 and Supplementary Fig. 10). We plotted the spatial locations of gene expression for all genes in the developing human kidney samples (Fig. 4c,d and Supplementary Fig. 10). Renal corpuscles were readily visible, marked by mature podocytes (NPHS2; Fig. 4c). We also identified mature PT markers such as apolipoprotein E (APOE; Fig. 4d). We found early NPC markers such as UNC homeobox (UNCX) localized to the most superficial region of the tissue (Fig. 4d). Deep to the NPCs we found higher expression of marker genes for intermediate cell states such as LIM homeobox 1 (LHX1) and jagged canonical notch ligand 1 (JAG1; Fig. 4d).

Fig. 4. Spatial molecular characterization of the developing human kidney.

a, Integrated UMAP of spatial (CosMx) and dissociated scRNA-seq data showing cell-type annotations indicated by color. b, H&E-stained human fetal kidney section adjacent to the corresponding cell-type annotation from our spatial transcriptomic data. Cell-type annotation is indicated by color. Performed for each CosMx run (n = 3), a representative image is shown. c, Expression of WT1 and NPHS2 (podocyte marker genes) in space within a representative region of fetal kidney sample. Gene expression quantity is indicated by color. d, Expression of MEIS2, UNCX, LHX1, APOE and JAG1 (yellow) in space within a representative region of fetal kidney sample. Gene expression quantity is indicated by color. e, Pseudotime of NPC lineages, colored from early (purple) to late (yellow); absorption probabilities for PT, DCT and LOH colored from low (purple) to high (yellow); renal corpuscle fate absorption probabilities colored from low (blue) to high (red), in space within a representative region of fetal kidney sample. RC, renal corpuscle cells; abs., absorption.

Extended Data Fig. 4. Imputed gene expression within space on CosMx samples.

a, Histology and CosMx cellular annotations are shown for the entire imaged region on the FK1 tissue sample (18 weeks, 5 days gestation) and on a zoomed in piece of tissue. b, Renal corpuscles were typically comprised of PEC and podocytes; however, for a subset of corpuscles, we noted that they were lined with PT cells. c, Imputed expression for marker genes as shown in space with lower expression indicated by purple coloring and higher expression indicated by yellow coloring. PT: proximal tubule, LOH: loop of Henle, DCT: distal convoluted tubule, UB: ureteric bud, Int: intermediate.

The fully integrated dataset enabled visualization of developmental and differentiation stages—represented as absorption probabilities—in space (Fig. 4e). This confirmed that the most developmentally immature nephron cells were located in the most superficial portion of the tissue, while more differentiated stages were located deeper⁶. We found podocyte and DCT specification in a relatively superficial region, indicating an early lineage specification consistent with our CellRank analysis (Fig. 4e). Renal corpuscle absorption probability showed that commitment of intermediate cells occurs superficially (Fig. 4e).

Molecular characterization of histologic neighborhoods

We then examined how gene expression and differentiation trajectories relate to classic microanatomic structures (for example, comma-shaped body). We had an expert pathologist annotate these structures (Fig. 5a,b and Extended Data Fig. 5a), and we examined gene expressions, cell-type fractions and trajectories for each neighborhood using our gene expression information.

Fig. 5. Molecular characterization of the histologic neighborhoods in the developing human kidney.

a, H&E image of the nephrogenic zone in a human fetal kidney with manual neighborhood annotation. Blastema is circled by blue lines, PES in red and green shows E. RC, with the lower boundary of the nephrogenic zone outlined in yellow. b, The H&E-identified neighborhoods were then plotted on the CosMx data as shown in a small section and across the entire tissue, with cells colored by neighborhood. c, Pseudotime trajectory characteristics of each neighborhood. For each histologic neighborhood, we graph aggregate trajectory enrichment (y axis) against pseudotime (x axis) for each lineage, as indicated by color. Remaining neighborhoods and cell types are presented in Extended Data Fig. 5. d, Neighbor characteristics of each histologic neighborhood. Circular plots show node cell, and neighbor frequency with increasing neighbor radius. Neighbor type is indicated by color. Remaining neighborhoods and cell types are presented in Extended Data Fig. 5. e, Aggregate trajectory and expression characteristics of each histologic neighborhood across lineages. Matrix plot showing gene expression or differentiation characteristics within each neighborhood as parsed by lineage. Each column shows a histologic neighborhood and each row a specific gene or trajectory. Value is indicated by color, with low being purple and high being yellow. f, UMAP of integrated scRNA-seq and CosMx, showing cell annotations. Stromal cells (circled) were subclustered and these subclusters were annotated as shown on adjacent UMAP of just stromal cells. Subclusters are marked by color. g, Aggregate expression characteristics of each stromal subcluster. Matrix plot showing gene expression quantity within each neighborhood as parsed by lineage. Each column shows a subcluster and each row a specific gene. Value is indicated by color with low being purple and high being yellow. h, Stromal subclusters in space. Stromal cells are plotted in space with annotation of each cell type indicated by color. Insets i and j are marked by a labeled white box. Gray cells indicate nonstromal cells. i, Stromal subclusters in space, inset of h, showing a region of nephrogenic zone. Stromal cells are plotted in space with annotation of each cell type indicated by color. j, Stromal subclusters in space, inset of h, showing a deeper region within the kidney. Stromal cells are plotted in space with annotation of each cell type indicated by color. E. RC, early renal corpuscle; PES, primitive epithelial structure; NZO, nephrogenic zone other.

Extended Data Fig. 5. Histologic neighborhood molecular characteristics.

a, Spatial locations of histologic annotations for entire tissue section. Specific neighborhood is indicated by dot color. b, Trajectory characteristics of each neighborhood. For each histologic neighborhood, we graph aggregate trajectory enrichment (y-axis) against pseudotime (x-axis) for each lineage, as indicated by color. Remaining neighborhoods and cell types are presented in Fig. 4. Also shown for renal corpuscle and tubular trajectories, as indicated by color (bottom). c, Neighbor characteristics of each histologic neighborhood. Circular plots show node cell, and neighbor frequency with increasing neighbor radius. Neighbor type is indicated by color. Remaining neighborhoods and cell types are presented in Fig. 4. d, Pathway analysis of differentially expressed genes in NPC lineage cell types within the blastema and PES, with dot size indicating number of genes within pathway and color indicating fold enrichment of the pathway genes. PT: proximal tubule, LOH: loop of Henle, DCT: distal convoluted tubule, PES: primitive epithelial structures. NZO: nephrogenic zone other. PEC: parietal epithelial cell.

The blastema had the earliest pseudotime and no enrichment for a particular cell fate, consistent with the notion that this neighborhood houses stem cells (Fig. 5c). Primitive epithelial structures demonstrated a later pseudotime, but no enrichment for a particular cell fate (Fig. 5c). Early renal corpuscles had an even later pseudotime and an enrichment for the podocyte lineage, demonstrating that a fate decision has been made within this neighborhood (Fig. 5c). Our analysis was corroborated by mature podocytes demonstrating full enrichment for the podocyte lineage and the latest pseudotime.

Recognizing that fate decisions are likely mediated by neighboring cells, we quantified the cell types present in each histologic neighborhood (Fig. 5d). We analyzed cells within 25–100 µm of the central cell. Cells within the blastema were almost exclusively surrounded by other undifferentiated nephron cells and stromal cells. Primitive epithelial structures had the highest fraction of UB cells, consistent with observations that NPC differentiation occurs at this stage near the UB. The endothelial cell fraction increased in early renal corpuscles and podocytes.

We then examined cell-type gene expressions within each neighborhood (Fig. 5e). The gene expression data indicated remarkable gene expression variability for the same cell type within different structures (Fig. 5e). For example, the early intermediate nephron cells expressed high levels of SIX1/SIX2 in the blastema, but gradually expressed markers of differentiated cells, podocytes, PT, LOH, etc., in different neighborhood contexts. Similarly, UB cells expressed RET in the blastema and AQP2/AQP3 in differentiated principal cells (PC). Endothelial cells showed important expression differences between the glomerulus and peritubular endothelial cells (Fig. 5f,g)^22,43.

We then subclustered stromal cells, identifying progenitor cells, intermediate cells, medullary fibroblasts, pericytes, mesangial cells, vascular smooth muscle cells and myofibroblasts, and examined marker gene expression as well as location (Fig. 5g–j and Extended Data Fig. 6), which was consistent with, and expands upon, previous efforts to categorize fetal kidney stromal heterogeneity^43,44.

Extended Data Fig. 6. Subclustering of stromal cells shows reproducible spatial distinct variability.

a, Spatial locations of stromal subclusters for each CosMx sample. Specific subcluster is indicated by dot color. b, Matrix plot of measured expression for each of the stromal subclusters. Each subcluster is indicated by a given row and gene by column. Relative expression is indicated by color with low expression being purple, and high expression being yellow.

Spatially resolved, unbiased cell–cell interactions

Traditional scRNA-seq ligand–receptor interaction analyses are limited by the lack of spatial information. Our new spatial gene expression data enabled a spatially aware ligand–receptor analysis, with full-transcriptome-estimated expression allowing widespread discovery.

First, we used CellPhoneDB⁴⁵, restricting interactions to those found within histologically identified neighborhoods (Supplementary Table 13). CellPhoneDB highlighted 569 spatially plausible interactions, including classic ligand–receptor interactions such as WNT (Wingless/Integrated), fibroblast growth factor (FGF) and BMP ligands, which are known to be critical for kidney development in mouse genetic experiments (Supplementary Table 13).

We then used CytoSignal⁴⁶, which leverages spatially resolved gene expression data to calculate signaling among single cells and is not limited to previously defined structures, unlike CellPhoneDB. CytoSignal accounts for ligand diffusion to identify ligand–receptor interactions and direct cell–cell interactions by requiring signaling cells to be in direct contact. In the developing human kidney, we observed 246 interactions conserved across all three samples (Supplementary Table 14). CytoSignal revealed key signaling nodes such as Wnt family member 4 (WNT4), which coordinates early epithelialization of nephron progenitors in mice.

We then calculated correlations between each sender signal (Fig. 6a; for example, ligands) in the CytoSignal database and each trajectory (Fig. 6b). This analysis identified cell–cell interactions associated with specific cell fate decisions. We found that insulin-like growth factor 2 (IGF2), Glial cell line-derived neurotrophic factor (GDNF) and R-spondin 3 (RSPO3) showed a high correlation with NPC renewal (Fig. 6b). Furthermore, the presence of IGF2 appeared to be associated with an early stem cell state across all NPC lineages (Fig. 6c). Cognate receptors for IGF2 were constitutively expressed in NPC lineage cells (Fig. 6d and Supplementary Fig. 11), while the ligand was produced by NPCs, stromal and UB cells. This led to the hypothesis that downstream IGF2 signaling might have a specific role in NPC function.

Fig. 6. Spatial transcriptomics highlights the importance of IGF2-mediated signaling in NPC renewal.

a, Spatial locations of CytoSignal ligand scores for IGF1 and IGF2 in the human kidney, with score indicated by color (low being purple and high being yellow). b, Histogram of correlations of microenvironment score with absorption probability for 878 microenvironment signals for the renewal trajectory. X axis shows correlations and y axis shows frequency (density estimate) of that correlation. Some individual genes of interest are labeled. c, Ligand score (y axis) is plotted against pseudotime (x axis) for IGF2. Each trajectory is shown by its discrete color. d, UMAP of scRNA-seq showing that IGF2 is most expressed in the NPC population, with cognate receptors IGF1R and INSR expressed in all cell populations. e, EpCAM immunofluorescence staining (white, ureteric tree) images and quantified number of UB tips (y axis) for mouse fetal explants incubated with ectopic IGF2 ligand. The total number of tips was normalized for average number of tips in the control samples for each replicate (control—1.00 ± 0.226, n = 10; IGF2—1.39 ± 0.266, n = 10; P = 0.002, two-sided t test). This was performed across three separate litters. Scale bar = 400 µm. f, EpCAM immunofluorescence staining (white, ureteric tree) images and quantified number of UB tips (y axis) for mouse fetal explants incubated with ectopic insulin ligand for 5 days. The total number of tips was normalized for average number of tips in the control samples for each replicate (control—1.00 ± 0.09, n = 8; insulin—0.63 ± 0.133, n = 10; P = 1.23 × 10⁻⁷, two-sided t test). This was performed across three separate litters. Scale bar = 400 µm. Each replicate is shown separately in Supplementary Fig. 12. g, SIX2 immunofluorescence staining (white, NPC marker) images and quantified size of SIX2 caps on mouse explant cultures for 5 days with IGF2 (500 ng ml⁻¹). Area was normalized to average area for control explants (control—1.00 ± 0.492, n = 153; IGF2—1.498 ± 0.679, n = 193; P = 2.5 × 10⁻¹², two-sided t test). This was performed across two separate litters. Each replicate is shown separately in Supplementary Fig. 12. Scale bar = 400 µm. h, SIX2 (purple) and EpCAM (white) immunofluorescence of explants treated with IGF1R inhibitor linsitinib at various concentrations for 5 days of culture. This was repeated in a separate litter and had similar results. Scale bar = 400 µm. i, Phase contrast images of human kidney organoids treated with IGF1R inhibitor linsitinib on days 10–12 at various concentrations, demonstrating loss of tubular structures with increasing linsitinib concentrations. This experiment was repeated with similar results in a separate organoid batch. Scale bar = 1,000 µm. j, Images of human kidney organoids treated with IGF1R inhibitor linsitinib on days 10–12 at various concentrations, followed by whole-mount immunofluorescence. These organoids demonstrate loss of both SIX2-positive (magenta) cells and JAG1 (cyan) tubular structures with increasing linsitinib concentrations. This experiment was repeated with similar results in a separate organoid batch. Scale bar = 1,000 µm for large image and 100 µm for inset. k, Images of human kidney organoids treated with vehicle or IGF2 (500 ng ml⁻¹) without basal insulin on days 10–12 of culture followed by whole-mount immunofluorescence on day 12 with SIX2 (magenta) and JAG1 (white). Scale bar = 1,000 µm. Images show equivalent intensity unless noted as being reduced LUT, in which case it is fivefold reduced from matched LUT. l, Inset images showing JAG1 aggregates and SIX2-positive nuclei from images in k. Scale bar = 100 µm. m, Quantified area for SIX2 and JAG1-positive cells for each organoid treated with IGF2 (500 ng ml⁻¹) or vehicle without basal insulin. This was completed across two batches (Extended Data Fig. 7; control—1.00 ± 0.252, n = 14; IGF2—2.07 ± 0.597343, n = 16; P = 9.37 × 10⁻⁷, two-sided t test). Box plot shown depicts mean (white bar), box depicts bounds of inner quartiles and whiskers depict outer quartiles. Violin depicts the kernel density estimate. n, Quantified area for SIX2 and JAG1-positive cells for each organoid treated with IGF2 (500 ng ml⁻¹) or vehicle with basal insulin. This was completed across two batches (Extended Data Fig. 7; control—1.00 ± 0.6277, n = 14; IGF2—6.88442 ± 4.266872, n = 16; P = 2.10 × 10⁻⁵, two-sided t test). Images of these organoids are shown in Extended Data Fig. 7. Box plot shown depicts mean (white bar), box depicts bounds of inner quartiles and whiskers depict outer quartiles. Violin depicts the kernel density estimate. Linsit, linsitinib; LUT, look-up table.

To investigate this, we cultured fetal mouse kidney explants from embryonic day 13 (E13) embryos. Explants were exposed to ectopic mouse IGF2, which resulted in more UB tips after 5 days of ex vivo culture compared to controls, indicating that IGF2 increases UB branching (Fig. 6e and Supplementary Fig. 12a). IGF2 is part of the insulin signaling pathway and canonically acts on the IGF1 receptor (IGF1R) but can also bind to the insulin receptor. Insulin can also bind IGF1R, especially at higher concentrations, as observed in clinical situations like maternal diabetes. Cultured explants treated with high-dose insulin showed poor growth and decreased UB branching relative to controls (Fig. 6f and Supplementary Fig. 12b). To specifically test the effect on NPCs, we performed immunofluorescence on IGF2-incubated explants and found that caps were larger in the IGF2-treated group (Fig. 6g and Supplementary Fig. 12c). Furthermore, incubation with an IGF1R inhibitor, linsitinib, led to the loss of a SIX2-positive population at high concentrations and smaller, disorganized caps at lower concentrations (Fig. 6h).

We examined NPC lineage human kidney organoids using a single-nucleus RNA-seq dataset capturing multiple time points⁴⁷. These data showed that IGF1R, IGF2R and INSR receptors were highly expressed, with early IGF2 expression present at day 12, but decreasing by day 26 (Supplementary Fig. 13), suggesting that organoids could be used for validation. We found that organoid treatment with linsitinib showed decreased SIX2-positive cells and fewer tubular aggregates at day 12 with increasing linsitinib concentrations (Fig. 6i,j). Conversely, adding IGF2 increased SIX2-positive and JAG1-positive populations at 12 days in both standard organoid media and in insulin-free media, indicating a positive effect on NPC renewal and early nephron fate (Fig. 6k–m and Extended Data Fig. 7).

Extended Data Fig. 7. IGF2 increases SIX2-positive cell population in human organoids.

a, Images of human kidney organoids treated with vehicle or IGF2 (500 ng ml⁻¹) with basal insulin on days 10–12 of culture followed by whole-mount immunofluorescence on day 12 with SIX2 (magenta) and JAG1 (white). Scale bar = 1000 µm. Images show equivalent intensity. b, Inset images showing JAG1 aggregates and SIX2 positive nuclei from images in panel ‘a’. Scale bar = 100 µm. c,d, Quantified area of SIX2 and JAG1 partitioned by batch (indicated by color) for data shown in Fig. 5. e, Single z-plane of SIX2 immunofluorescence (purple) demonstrating a nuclear pattern.

We then examined the ligands associated with the PT lineage. One of the highly spatially variable ligands we identified was arginine vasopressin (AVP). Interestingly, while AVP is produced by the hypothalamus, recent studies have shown that some local AVP production also occurs in the adult kidney, although its function is unclear⁴⁸. Within the fetal kidney, our data show that AVP is spatially correlated with the PT trajectory, which produces the ligand, but lacks the cognate receptor. It is also spatially associated with the DCT trajectory that does not produce the ligand but expresses the cognate receptors (Extended Data Fig. 8). This suggests that local AVP production in the kidney may act to synchronize differentiation across nephron segments. This is supported by previous work in organoids showing that AVP improved differentiation of IC and PC⁴⁹.

Extended Data Fig. 8. Local AVP as a potential mediator of tubular synchronization.

a, Histogram of correlations of microenvironment score with absorption probability for 878 microenvironment signals for the PT trajectory. X-axis shows correlations and y-axis shows frequency (density estimate) of that correlation. Some individual genes of interest are labeled. b, Spatial locations of CytoSignal ligand scores for AVP in the human kidney, with score indicated by color (low being purple and high being yellow). c, UMAP of scRNA-seq showing that AVP is most expressed in the PT cell population, with cognate receptors AVPR1A and AVPR2 expressed in the DCT cell population. d, Ligand score (y-axis) is plotted against pseudotime (x-axis) for AVP. DCT and PT trajectories are shown by their own discrete color.

We performed ligand association analysis for the other NPC trajectories, finding that FGF9 and Kininogen 1 (KNG1) were correlated with LOH trajectory, while ephrin A4 (EFNA4) and BMP3 were associated with the DCT (Extended Data Fig. 9). We also identified vascular endothelial growth factor A and ephrin B2 (EFNB1) for podocytes, which are consistent with their well-known roles in this compartment (Extended Data Fig. 9 and Supplementary Table 15).

Extended Data Fig. 9. Histogram of correlations of microenvironment score with absorption probability for 878 microenvironment signals for the DCT, podocyte and LOH trajectory.

X-axis shows correlations and y-axis shows frequency (density estimate) of that correlation. Some individual genes of interest are labeled. Bottom, the ligands correlated with the tubular and renal corpuscle fates within the PT (circled) population.

We also examined the ligands associated with general tubular cell fate (renal corpuscle or tubule), finding AGT associated with tubules, while tetherin (BST2) and JAGGED1 (JAG1) were associated with transition to a renal corpuscle fate (Extended Data Fig. 9). We investigated whether ligands appeared to be shared between the UB and NPC distal populations, finding that nephroblastoma overexpressed and midkine signals are associated with distal tubule/connecting tubule fates in both lineages (Supplementary Fig. 14 and Supplementary Table 18).

Unbiased spatial neighborhood-based microenvironments

Primitive developing nephron structures are heterogeneous and exist along a continuum. Therefore, we sought to develop an unbiased approach to characterize development based on aggregate microenvironmental effects, differing from our previous niche assignment that focuses on cell-type annotations within a fixed distance⁵⁰.

To do this, we used the ligand signal matrix, instead of the conventionally used gene expression, to generate a UMAP, which showed specific marker signals and cell fate choices localized to specific cell populations (Fig. 7a and Extended Data Fig. 10). In this analysis, each cluster was defined by ligand abundance (Supplementary Table 16). For example, we identified GDNF neighborhoods (blastema, NPC), vascular endothelial growth factor A neighborhood (glomeruli) and AVP ligand/receptor neighborhoods (collecting duct). These ligand/receptor neighborhoods also correlated with the probabilities of cell renewal and differentiation.

Fig. 7. Unbiased spatial neighborhood-based microenvironments in developing human kidney.

a, UMAPs of fetal kidney cells constructed based on ligand signals present at each cell location. Color of each cell indicates signal presence (top), or absorption probability (bottom left) or neighborhood (bottom right). For signal presence/absorption probability, low values are indicated as yellow and high being green. Neighborhood identity indicated by color—0, red; 1, orange; 2, yellow; 3, green; 4, blue; 5, purple; 6, pink; 7, brown; 8, black; 9, dark blue; 10, turquoise; 11, gold; 12, light green; 13, cyan; 14, light purple; 15, maroon; 16, dark purple; 17, dark pink; 18, dark yellow; 19, silver; 20, dark brown; 21, reddish orange; 22, dark orange (also shown in Extended Data Fig. 10c). b, Histologically defined unbiased neighborhoods in space (left) and microenvironment (ligand) defined neighborhoods in space (right). Neighborhoods are indicated by color. c, The heatmap of gene expression and trajectory characteristics of microenvironment-derived neighborhoods. Unbiased neighborhood spatial location is shown above and the columns under show representative gene expression of heatmap. Gene expression and trajectory characteristics are rows. Low values are indicated in purple and high values in yellow. d, Pseudotime trajectory characteristics of each microenvironment-derived neighborhood. For each microenvironment neighborhood, we graph aggregate trajectory enrichment (y axis) against pseudotime (x axis) for each lineage, as indicated by color.

Extended Data Fig. 10. Spatially defined neighborhoods.

a, UMAP of environment signal UMAP with color of each cell indicating absorption probability estimates for PT, podocytes, LOH and DCT lineages. b, UMAP of the ligand signal matrix with each sample indicated by color. c, Leiden clustering of the environment signal UMAP with individual Leiden clusters colored uniquely. d, UMAP of all CosMx cells colored by neighborhood cluster. e, Spatial locations of environment-based neighborhoods on the FK4 and the HK3524 fetal kidney samples, with each cell colored by Leiden cluster using color map from panel b. f, Renal corpuscle absorption of PT cells (y-axis) within microenvironment neighborhoods (x-axis). g, Dot plot showing PT marker genes in all cell populations, with PT population partitioned between PTs with those with >50% renal corpuscle absorption probability (PT_RC) and those with <50% absorption probability (PT_tubule). Int: intermediate nephron cells; NPC: nephron progenitor cells, PT_RC: proximal tubule with renal corpuscle absorption probability >50%, PT_tubule: proximal tubule with RC absorption probability <50%, PEC: parietal epithelial cell, LOH: loop of Henle, DCT: distal convoluted tubule, UB_CT: ureteric bud, collecting duct.

We found that the ligand-based neighborhoods recapitulated features of classic histologically defined neighborhoods, differentiating the blastema from other portions of the nephrogenic zone, the nephrogenic zone from the cortex and identifying medullary regions (Fig. 7b and Extended Data Fig. 10). This identified additional spatially defined developmental environments. We found that these neighborhoods predicted gene expression across cell types and trajectories (Fig. 7c). We found that neighborhood 10, which spatially resembles the blastema, corresponded with NPC renewal; neighborhood 7 with early NPC differentiation; neighborhood 0 with early tubular differentiation; neighborhood 1 with late tubular differentiation; and neighborhood 8 with renal corpuscle maturation (Fig. 7c,d). We noted that these neighborhoods correlated with PT fate (Extended Data Fig. 10).

Discussion

We present a comprehensive analysis of human fetal kidney development, driven by integration of spatially resolved single-cell gene expression, high-quality scRNA-seq and advanced probabilistic computational analyses. We substantially advance the interpretation of cellular differentiation by embedding fate trajectories within their native spatial context. This approach improves our understanding of cell differentiation trajectories during kidney development, supporting changes to the current model of nephrogenesis.

Previous work in mouse kidney development described a sequence of binary fate decisions, beginning with NPC renewal versus differentiation²⁴, followed by commitment to prerenal corpuscle or pretubular fates, culminating in proximal–distal segmentation and terminal differentiation into LOH or DCT cells¹⁹. In contrast, early human single-cell trajectory studies suggested a dominant proximal–distal decision in NPC-derived intermediates, in which proximal cells gave rise to both podocytes and PT cells¹⁸. Our analysis reveals an early renal corpuscle versus tubular fate decision at the intermediate stage. However, we find that this early fate decision is not fixed. Our data support the conclusion that some cells with PT identity contribute to renal corpuscle epithelium by adopting a PEC fate.

The implications of this finding extend beyond PEC biology. Cell fate and spatial position are tightly coupled in mature tissues, yet early developmental patterning relies on morphogen gradients that operate over micrometer scales and are therefore unlikely to encode the remarkable architectural precision observed in adult organs. The problem of reproducibly generating repetitive units with sharp fate boundaries and high architectural precision is not limited to the kidney, and it is unclear whether the developmental plasticity we observed exists in other organs.

A second major conceptual advance of this work is the explicit linkage of cell fate transitions to their surrounding signaling environment. Traditional ligand–receptor analyses applied to dissociated scRNA-seq data lack spatial proximity information and therefore cannot reliably infer which extracellular cues are available to individual cells in situ. By integrating spatial transcriptomics with trajectory inference, we generated microenvironment-aware, genome-wide ligand interaction maps and defined cellular neighborhoods based on aggregated extracellular signals rather than cell-type composition alone. This approach identified IGF2 as a locally acting factor associated with NPC renewal and early nephron differentiation, which we experimentally validated in mouse explants and human organoids. These results demonstrate that spatially resolved ligand information allows one to move beyond descriptive mapping and instead nominate and validate regulators of human development.

By embedding differentiation trajectories within a microenvironment-aware context, we were able to systematically identify extracellular cues that influence cell fate. This approach may be broadly applicable, beyond developmental biology, to disease contexts characterized by fate transitions, including regeneration, fibrosis and cancer.

Study limitations

Due to technical constraints of current spatial transcriptomic platforms, we inferred full-transcriptome profiles in space by mapping scRNA-seq data onto CosMx data rather than relying on measured counts. While this strategy mitigates transcript dropout, limited probe panels and segmentation errors, it relies on accurate integration, potentially excluding cells that lack close scRNA-seq counterparts. Our approach likely captures a subset of fate-associated interactions, as some ligands had low expression or were not assessed, such as mechanical forces. Furthermore, individual candidate ligands require further validation to prove their functionality.

With respect to the PT-to-PEC transition, direct lineage tracing in human fetal tissue is not feasible. Nonetheless, the convergence of spatial transcriptomics, probabilistic trajectory inference, orthogonal molecular markers and cross-species validation represents the current gold standard for inferring human developmental state transitions in situ.

In summary, this work provides a spatially resolved, single-cell framework for understanding human kidney development. We introduce two broad conceptual advances. First, we use a microenvironment-aware approach to cell fate determination in situ. Second, we identify late-stage cell fate tuning that refines previous fate decisions established during early patterning. Our findings establish a foundation for dissecting how spatial context shapes human development.

Methods

Fetal sample acquisition

Human fetal kidneys were obtained from donors undergoing elective abortion at the Gynecology Clinic, University of Pennsylvania and from the Sheba Medical Center. All donors provided written, informed consent. All experimental procedures were approved by the Institutional Review Boards at the University of Pennsylvania (832470) and at Sheba (approval 8993-21-SMC). Embryo ages were determined through ultrasonographic measurement of crown-rump length. Kidney tissue was dissected in RPMI-1640 medium (Gibco) using a dissection microscope.

Human sample acquisition

Human pediatric kidney tissue was obtained retrospectively from surgical resections at the Children’s Hospital of Philadelphia as part of the approved IRB protocol 22-020263, which included a waiver of consent, given the retrospective nature of the work and minimal risk for participants. Tissue was examined by an expert pathologist to ensure that it did not contain tumor cells. An honest broker de-identified the tissue and clinical information. Participants were not compensated.

Mouse explants

Embryos dissected from killed female CD1 mice (Charles River Production) were used at E13 developmental stages for explant experiments involving IGF2. For linsitinib experiments, C57BL/6 mice were used for availability purposes. Animal studies were approved by the University of Pennsylvania IACUC (807000).

scRNA-seq

Fresh human kidneys were collected in RPMI media and, using a razor blade, were cut into small pieces of a diameter of 2–4 mm. Tissue was further dissociated with gentleMACS C tube (Miltenyi, 130-110-201) using program Multi_B. Tissue was digested with enzyme D, enzyme R and enzyme A in RPMI at 37 °C for 30 min. Tissue was homogenized again with gentleMACS using the Multi_B program. The resultant solution was strained through a 70 µm filter, and then centrifuged at 600g for 7 min. The cell pellet was incubated in RBC lysis buffer for 3 min on ice, followed by the addition of 10 ml of PBS and centrifuged again for 5 min at 500g. Supernatant was removed, and the pellet was resuspended in PBS. Cell number and cell viability were examined using the Countess AutoCounter (Invitrogen, C10227) and manual examination. An estimated 20,000 cells, based on AutoCounter results, were loaded into the Chromium Controller (10x Genomics, PN-120223) in a Chromium Next GEM chip G single-cell kit (10x Genomics, PN-1000120). The cDNA and gene expression library were made using the Chromium Next GEM Single Cell 3′ GEM Kit (v3.1; 10x Genomics, PN-1000121) and Single Index Kit T Set A (10x Genomics, PN-120262) according to the manufacturer’s protocol. Quality control for the libraries was performed using Agilent Bioanalyzer High Sensitivity DNA kit (Agilent Technologies, 5067-4626). Libraries were sequenced on the Illumina NovaSeq 6000 system with 2 × 150 paired-end kits. Reads were then demultiplexed by sample.

CosMx sample preparation

Tissue sections were cut at 5-µm thickness and prepared according to the manufacturer’s specifications (NanoString Technologies). The human universal cell characterization RNA probes were used with 50 additional custom probes set for the following genes, which were manually selected: ESRRB, SLC12A1, UMOD, CD247, SLC8A1, SNTG1, SLC12A3, TRPM6, ACSL4, SCN2A, SATB2, STOX2, EMCN, MEIS2, SEMA3A, PLVAP, NEGR1, SERPINE1, CSMD1, SLC26A7, SLC22A7, SLC4A9, SLC26A4, CREB5, HAVCR1, REN, AP1S3, LAMA3, NOS1, PAPPA2, SYNPO2, RET, LHX1, SIX2, CITED1, WNT9B, AQP2, SCNN1G, ALDH1A2, CFH, NTRK3, WT1, NPHS2, PTPRQ, CUBN, LRP2, SLC13A3, ACSM2B, SLC4A4, PARD3, XIST and UTY. We used the DAPI, CD298/B2M, CK8/18 and PanCK/CD45 for additional staining per NanoString protocol. Imaging was performed using configuration A. After imaging was completed, the flowcell was incubated in 100% xylene overnight and the coverslip was removed from the slide with a razor blade. The slide was then stained with H&E and imaged. The adjacent section of each tissue piece was also stained with H&E and imaged as well.

Xenium sample preparation

Tissue sections were cut at 5 µm thickness and prepared according to the manufacturer’s specifications (10x Genomics). We used the Xenium Prime 5K human pan tissue & pathways panel. We used 100 additional custom probes for the following genes: TPM1, ESRRB, COL6A3, AGR2, SLC26A7, ATP1B1, SLC8A1, ATP6AP2, TAGLN, SPP1, SAT1, MYL9, LDB2, DEFB1, COL1A2, ACTA2, ST6GALNAC3, SLC13A3, SLC12A3, SLC12A1, MGP, IGHG1, FN1, C7, ACSM2B, AIF1, APOE, AQP3, AZGP1, C1QA, C1QB, C1QC, CAV1, PPIA, CD74, CHI3L1, COL1A1, COL6A1, CRYAB, CXCL14, ENO1, HLA-DPA1, HLA-DRA, IFI27, IGHA1, IL32, KLF2, LGALS3, LUM, MMP7, PIGR, S100A2, SLC4A4, SLPI, SOD2, SPINK1, SOX4, SPOCK2, TACSTD2, TM4SF1, TPM2, VIM, ZFP36, AQP2, RNASE1, ALDOB, PCGF6, RHOB, CD81, ASS1, MYL6, COX8A, CTSB, GATM, MT1G, TMSB10, COL3A1, MIF, TPT1, COL6A2, BST2, CLU, APOC1, APOD, PHKG2, RGCC, HLA-DQA2, CORO1A, HSPB1, ADIRF, CKB, HLA-DQB1, COX5B, MT1H, RAMP3, TYROBP, LAMTOR5, ITM2B, UBB and CTSD. Visualizations of Xenium data were made using Xenium Explorer 4.

Fetal explant culture

Fetal kidneys from E13 mice were dissected and placed on a transwell plate (CellTreat, 230615) and cultured in DMEM with penicillin/streptomycin with the addition of 500 ng ml⁻¹ IGF2 (R&D Systems, 792-MG), 70 µM Insulin (Novolog) or PBS. Media was changed daily, and, after 5 days of culture, the explants were fixed in 4% paraformaldehyde at room temperature for 10 min. Experimental comparisons were only made within the same litter of mice. Linsitinib (Selleckchem, OSI-906) experiments were performed the same way using 40, 4 and 0.4 µM concentrations or DMSO being added to the media and changed daily.

Human nephron organoid generation

NPC organoids were generated from the SV20 induced pluripotent stem cell (iPSC) line (iPSC core; University of Pennsylvania, PENN123i-SV20; RRID:CVCL_EL23) according to published protocols^51–53. iPSCs were maintained in standard tissue-culture-treated 6-well plates in stem cell maintenance as well as in supplement (mTeSR+ kit; STEMCELL Technologies, 100-0276). Cells were passaged using Accutase (STEMCELL Technologies, 07920) and plated at a density of 5,200 cells per cm². Differentiation was induced with TeSR-E6 medium as well as with supplements (TeSR-E6+; STEMCELL Technologies, 05946) and 7 µM CHIR99021 (Tocris, 4423) the following day, and media was refreshed every other day for 5 days. On day 5, wells were washed with DPBS and media was changed to TeSR-E6+ with 1 µg ml⁻¹ heparin (Sigma-Aldrich, H4784) and 200 ng ml⁻¹ FGF9 (R&D Systems, 273-F9-025). Media was refreshed every day for 5 days. On day 10, cells were lifted using Accutase and excess TeSR-E6+ was used to quench digestion. Cells were pelleted by centrifugation at 300g for 3min, then resuspended in TeSR-E6+ and counted. An appropriate volume of cells corresponding to 300,000 cells per organoid was then transferred to microcentrifuge tubes and centrifuged at 300g for 3 min. The pellet was resuspended in TeSR-E6+ with 7 µM CHIR99021 to create a dense cell slurry at 3 × 10⁵ cells per µl. One microliter aliquots of the slurry were then spotted onto 6-well 0.4 µm polyester transwell membranes (CellTreat, 230607) overlying TeSR-E6+ and 7 µM CHIR99021 ‘pulse’ media with five to eight organoids spotted per well. Organoids were ‘pulsed’ for 2 h. After the CHIR ‘pulse’, samples were washed with DPBS before changing to chase media to TeSR-E6+ with 1 µg ml⁻¹ heparin, 200 ng ml⁻¹ FGF9 and IGF2 (500 ng ml⁻¹) or equivolume DPBS for days 10–12, and, for linsitinib experiments, 0 µM, 40 nM, 400 nM, 4 µM or 40 µM linsitinib (Selleckchem, OSI-906) with equivolume DMSO (Fig. 5 and Extended Data Fig. 7) for days 10–12. Media was refreshed every day until D12 when samples were either fixed for immunofluorescence or washed with 1× DPBS before changing media to TeSR-E6+ with equivolume 1× DPBS, 500 ng ml⁻¹ IGF2. Media was refreshed every other day until experimental endpoint.

To further test IGF2, organoids were separately produced using media without insulin. iNPCs produced using the protocol above were previously frozen on D10 and stored in cryopreservation media made from a 50:40:10 percent volume of TeSR-E6 with supplements, KnockOut Serum Replacement (Gibco, 10828010) and DMSO (Sigma-Aldrich, 276855-100ML), as well as 10 µM Y27632 (STEMCELL Technologies, 72304). Upon thawing, samples were transferred to excess DMEM/F12 (Gibco, 11330-032) and centrifuged at 300g for 3 min. Organoids were spotted, ‘pulsed’ and handled postpulse as described above with the exception of no insulin differentiation media used as a replacement for TeSR-E6. No insulin differentiation media was prepared as follows: DMEM/F12 with 1% GlutaMAX (Invitrogen, 35050-079), 1% MEM nonessential amino acids (Invitrogen, 11140-050), 1% penicillin–streptomycin (Invitrogen, 15140122) and 110 μM 2-mercaptoethanol (Bio-Rad, 1610710).

Human nephron organoid immunofluorescence

At the experimental endpoint, the transwell membrane was separated from the transwell apparatus and placed in microcentrifuge tubes containing 4% paraformaldehyde for 15 min at room temperature, then washed twice with DPBS with 100 mM glycine, and finally exchanged with DPBS. Organoids were permeabilized with 0.5% Triton X-100 in DBPS for 30 min at room temperature, then blocked in 1× immunofluorescence wash (DPBS + 1 g l⁻¹ bovine serum albumin + 0.2% Triton X-100 + 0.04% Tween-20) with 5% donkey serum for 1 h at room temperature. Primary and secondary antibody solutions were diluted in blocking solution and samples were incubated in antibody solutions overnight at 4 °C. After primary or secondary antibody incubation, samples were washed with three exchanges of 1× immunofluorescence wash solution over 3 h at room temperature. After the last wash, samples were briefly exchanged with DPBS before being transferred to glass slides for imaging. Once on the slides, excess DPBS was wicked away and FocusClear (Cedarlane Labs, FC-101) solution was added to cover each organoid. Two 38 µm spacers (Precision Brand, 44115) were added to each slide before adding a 140 µm-thick coverslip (Globe Scientific, 1415-10), which was sealed to the slide with clear nail polish.

D12 organoids were stained with rabbit anti-SIX2 (1:400; Proteintech, 11562-1-AP; RRID:AB_2189084) and goat anti-JAG1 (1:150; R&D Systems, AF599; RRID:AB_2128257). Donkey secondary antibodies were used at 1:400 dilution as follows: donkey antigoat—405 (Invitrogen, A48257) and donkey antirabbit—647 (Invitrogen, A32795).

Samples were imaged using a Nikon Ti2-E microscope equipped with a CSU-W1 spinning disk (Yokogawa), a white light LED, laser illumination (100 mW 405, 488 and 561 nm lasers and a 75 mW 640 nm laser), a Prime 95B back-illuminated sCMOS camera (Photometrics), motorized stage, ×4/0.2 NA, ×10/0.25 NA and ×20/0.5 NA lenses (Nikon), and a stagetop environmental enclosure (OkoLabs). For 10× images of day 12 organoids, z projections were acquired by taking 5 µm steps spanning the entire height of each organoid with 2 × 2 fields of view stitched together at 1% overlap.

Human nephron organoid imaging analysis

Imaged organoids were analyzed in ImageJ. Jython scripts used to quantify the presence of SIX2, JAG1 for day 12 organoids, available at https://github.com/jlevins2010/Fetal_atlas. Briefly, area for each marker was calculated as being above a manually set intensity threshold, specific to each marker, using the ‘analyze particles’ command for each sample to calculate area, which was normalized to the control across each experimental batch. Images within an experiment are shown with consistent LUT unless otherwise noted. Images shown are maximum-projection stacks with 50-pixel rolling-ball background subtraction.

Whole-mount immunofluorescence

After fixation, explants were blocked in 5% donkey serum and then stained with the following primary antibodies before imaging: JAG1—goat (1:200; R&D Systems, AF599-SP), SIX2—mouse (1:100; Proteintech, 66347-1-Ig) and SIX2—rabbit (1:200; Proteintech, 11562-1). Secondary antibodies used were the following: donkey antimouse—555 (Invitrogen, A31570), donkey antirabbit—647 (Invitrogen, A32795), donkey antirabbit—488 (Invitrogen, A21206), conjugated EpCAM (anti-Mo) and FITC (Invitrogen, 11-5791-82). Explants were washed after incubation with primary and secondary antibodies in PBS with BSA, Triton and Tween. Before imaging, explants were cleared with one drop of FocusClear (CelExporer, FC-101). To quantify tips, EpCAM-positive regions of tissue were counted if at edge of tissue and bordered by SIX2-positive cells. Images were analyzed using ImageJ, and brightness of images was adjusted to visualize the ureteric tree. Antibodies are also listed in Supplementary Table 19.

Slide-based immunofluorescence

Sections of tissue were deparaffinized in xylene, rehydrated in ethanol baths and antigen retrieval was performed with incubation in sodium citrate (10 mM, pH 6.0) at 100 °C for 20 min. Samples were blocked in 10% BSA. Primary antibodies used were WT1 (1:200; Novus, NB110-60011, and Abcam, CAN-R9(IHC)-56-2) and ACE2 (Thermo Fisher Scientific, goat PA5-47488 and mouse MA5-32307). Slides were then washed with PBS, and then incubated with 1:400 secondary Alexa Fluor 555 goat antirabbit (Thermo Fisher Scientific, A-21428) and Alexa Fluor 488 donkey antigoat (Thermo Fisher Scientific, A-11055).

Confetti mouse lineage tracing

To trace Pax2⁺ cells during nephrogenesis, recombination of the Confetti reporter in Pax2.rtTA;TetO.Cre;R26.Confetti embryos was induced at E8 by administering 4 mg ml⁻¹ doxycycline in drinking water supplemented with 2.5% sucrose for 4 days through the mothers. Pups were killed on postnatal day 6. Confocal microscopy was performed on 10 μm renal tissue sections using a Leica confocal microscope. For Confetti reporter imaging, signal acquisition was set in the cyan, green, yellow and red channels, using excitation wavelengths of 470, 488, 514 and 543 nm, respectively.

Bioinformatic analysis: scRNA-seq

Alignment of samples was performed with cellRanger⁵⁴ (version 7.2.0) to generate the original expression matrix used for cell-type annotation and connectivity and STARSOLO⁵⁵ (version 2.7.9a) to generate spliced, unspliced and ambiguous gene x cell matrices. CellBender⁵⁶ (version 0.2.0) was used to remove ambient RNA, using the following parameters: expected cells 10,000, total droplets included 50,000, false-positive rate 0.01, epochs 150 and a default learning rate. Learning curves for each sample were manually inspected to ensure local training was appropriate.

Basic scRNA-seq processing was performed using SCANPY⁵⁷. Doublets were removed using Scrublet (version 0.2.3). HarmonyPy⁵⁸ (version 0.09) was used for batch correction of the scRNA-seq data using default parameters. Cells were filtered to remove those with ≥25% mitochondrial counts and <500 genes. Only genes expressed in at least three cells in all samples were used for further analysis. Variable genes were called using SCANPY (version 1.9.1) using min_mean = 0.0125, max_mean = 3, min_disp = 0.5, batch_key = ‘sample’. Cell cycle was regressed out of each PC, and neighbors—15 for each cell—were called using 40 PCs. Clustering was performed using the Leiden algorithm (leidenAlg; version 0.8.10), with resolution of 1.0. Leiden-clusters DEGs, called using the Wilcoxon test, were inspected to determine cell-type annotation for each cluster. UMAPs were used to visualize cells in dimension-reduced space (umap-learn; version 0.5.3).

RNA velocity was calculated using scVelo (version 0.2.5), using 30 neighbors and a dynamical model. CellRank²⁹ (version 1.5.1) was used for calculating absorption probabilities to calculate trajectories of the single-cell data, using a mixed kernel (0.5 weight connectivity and 0.5 weight velocity). Seven macrostates were calculated and 400 cells with the top score for each macrostate were called. Absorption probabilities were called for each terminal state (PT, LOH, podocyte, DCT), for renal corpuscle versus tubular (podocyte compared to aggregate PT, LOH and DCT), and renewal/differentiation (aggregate terminal states versus NPC).

Age-dependent DEGs were calculated using PyDESeq2 (ref. ⁵⁹ ; version 0.4.4) through pseudobulking each annotation by sample and plotted in matplotlib (version 3.6.3).

Bioinformatic analysis: spatial single-cell transcriptomics

The expression matrix and metadata from each CosMx run were exported from the AtoMx platform and were converted to a Python object using Squidpy⁶⁰. All samples were merged and preprocessed and analyzed together. Cells with <30 counts were filtered out. Counts from MALAT1 and hemoglobin genes were removed. Count matrices were normalized, and log transformed. Principal component analysis was performed on these matrices. Neighbors were calculated using 30 PCs and 15 nearest neighbors, which were used to generate an initial UMAP. Leiden clustering was performed using a resolution of 1 and initial cell-type annotation was performed. Raw counts of the initial object were used for scVI batch correction using five layers to generate 30 latent variables. This model was trained for a maximum of 1,000 epochs using a learning rate of 0.001. These latent variables were used to recalculate neighbors. Using 15 neighbors per cell Leiden clustering was again performed for repeat annotation. Using annotations from scRNA-seq cells and the CosMx cells with the 80% closest intratechnology distance, we used scANVI to recalculate latent variables, which were used for neighbor calling (using all 30 latent variables), and Leiden clustering of 15 nearest neighbors (based on the expression) was annotated for the final time. Metadata and expression were imputed for each remaining CosMx cell by averaging distance-squared weighted values of the most similar scRNA-seq cells. Spatial neighbors were called for 25-µm, 50-µm, 75-µm and 100-µm distances using Squidpy. Differential expression across cell types was performed using the rank genes function of SCANPY using a Wilcoxon test. Generalized additive models were constructed using pyGAM (version 0.8.0) and plotted using matplotlib (version 3.6.3).

Pathway analysis

GSEApy⁶¹ (version 1.1.2) was used through the EnrichR⁶² method for pathway analysis. Total expressed genes measured were used for the background list and WikiPathway_2023_Human pathway list, and input genes with an adjusted P value <0.05.

Histologic neighborhood annotation

Whole-slide imaging was performed on hematoxylin and eosin (H&E)-stained images of both original CosMx slides and sequential sections. These images were aligned using QuPath, and annotations were performed by a pathologist. Spatial locations from annotations were calculated using GeoPandas and merged into the SCANPY object for further analysis. Differential expression between neighborhoods was performed using the rank gene function, using a Wilcoxon test. CellPhoneDB was used to identify spatially plausible interactions, using each neighborhood as a distinct microenvironment. CytoSignal was also used for spatially aware ligand–receptor analysis, with a P value cutoff of 0.05, read threshold of 100 and a significance threshold of 100.

Ligand microenvironments were determined using the dge.lig matrix for each sample using CytoSignal, and ligand names were pulled from the dge.lig column names and were converted from UniProt to gene symbol. These were merged and imported into an anndata object for principal component analysis, neighbor calculation (15 neighbors, using 30 PCs) followed by Leiden clustering (resolution 0.3) and UMAP construction. Differential ligand expression between neighborhoods was performed using the rank genes function of SCANPY using a Wilcoxon test. Generalized additive models were constructed using pyGAM (version 0.8.0) and plotted using matplotlib (version 3.6.3).

Statistics and reproducibility

Independent sample two-sided t test was used to compare the likelihood that two groups were drawn from the same population using SciPy stats package. A P value <0.05 was considered significant. No specific statistical method was used to predetermine sample size, and replication number was chosen to be consistent with those typically used for these types of experiments. The number of times an experiment was replicated is indicated in Figs. 3 and 6 legends. The investigators were not blinded to treatment conditions during experiments and outcome assessment.

Reporting summary

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

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41588-026-02665-0.

Author contributions

J.L., S.G., B.D., A.A., C.C., A.H., R.L.-K., B.W., G.R. and A.M.B. performed experiments. J.L. and B.D. performed computational analysis. B.D., R.L.-K., E.H., K.A.K., P.R., J.P.A., O.P., M.L., P.T., A.J.H., N.Z., J.-S.P., P.B. and K. Sasaki offered experimental and analytical suggestions. K. Susztak was responsible for overall design and oversight of the experiments. L.S.F. performed pathological scorings. K. Susztak and A.J.H. supervised the experiments. J.L. and K. Susztak wrote the original draft of the manuscript. All authors contributed to writing and approved the final version of the manuscript.

Peer review

Peer review information

Nature Genetics thanks Zhongwei Li, Rajasree Menon and Yoshiharu Muto for their contribution to the peer review of this work.

Data availability

Raw data, processed data and metadata from the scRNA-seq and CosMx spatial transcriptomics have been deposited in Gene Expression Omnibus with the accession codes GSE277893 and GSE278614.

Code availability

Scripts have been shared at https://github.com/jlevins2010/Fetal_atlas, and are archived on Zenodo⁶³.

Competing interests

All authors declare no competing interests. P.T.’s research contributing to this manuscript was conducted at the University of Pennsylvania while serving as a faculty member (2017-2025). P.T. is currently an employee of Eli Lilly and Company; however, the research contributing to this manuscript, as well as the discussion and viewpoints expressed, are not affiliated with, nor endorsed by, Eli Lilly and Company. P.T. is acting on their own in the preparation and submission of this manuscript.

Extended data

is available for this paper at 10.1038/s41588-026-02665-0.

Supplementary information

The online version contains supplementary material available at 10.1038/s41588-026-02665-0.