Integrated Transcriptomic and Machine Learning Analyses Identify KCNN3 and TLR10 as Candidate Cell-Type-Associated Molecules in Idiopathic Membranous Nephropathy

Abstract

Purpose

Idiopathic membranous nephropathy (IMN) is a common immune-mediated glomerular disease, but the key molecules driving its progression remain unclear. This study integrated bulk RNA sequencing (RNA-seq), machine learning, and single-cell RNA sequencing (scRNA-seq) to screen and preliminarily validate key molecules, providing a basis for subsequent research.

Patients and Methods

We analyzed our published urinary bulk RNA-seq data, as well as IMN kidney datasets obtained from the GEO database. Differentially expressed genes were identified and subjected to enrichment analysis. Three machine learning algorithms screened key genes. ScRNA-seq data were used to identify the cell types expressing the key genes and to perform CellChat analysis. Immunohistochemistry validated key gene expression, and immunofluorescence explored their cellular localization. External GEO datasets validated expression differences and diagnostic performance.

Results

We identified 94 genes commonly upregulated in both urine and kidney tissues, enriched in transmembrane transport pathways. Two key genes, KCNN3 and TLR10, were identified by machine learning. Single-cell analysis suggested KCNN3 enrichment in podocytes and TLR10 in dendritic cells (DCs), and CellChat analysis predicted potential crosstalk between these cell types through the CXCL12-CXCR4 axis. Immunohistochemistry confirmed their elevated expression in IMN kidneys. Immunofluorescence showed spatial associations of KCNN3 with podocytes and TLR10 with DCs. External validation showed expression trends of KCNN3 and TLR10 were not entirely consistent across datasets, with combined AUCs of 0.705 and 0.748 in two validation sets, showing no significant improvement over single-gene models.

Conclusion

KCNN3 and TLR10 may serve as cell type-associated candidate molecules in IMN, warranting further investigation into their functions and underlying mechanisms.

Introduction

Nephrotic syndrome (NS) is a common glomerular disease characterized by massive proteinuria and hypoalbuminemia.^1–3 Membranous nephropathy (MN) is the most common pathological type of NS in adults.^4,5 Among them, cases without identifiable causes are classified as idiopathic membranous nephropathy (IMN), accounting for approximately 75% of MN cases and representing the predominant subtype. Early diagnosis of IMN is crucial to delay the progression of renal function impairment. Renal biopsy and serological biomarkers (eg, anti‑PLA2R autoantibodies) constitute the primary diagnostic modalities for IMN.⁶ Although patients positive for anti-PLA2R antibodies may potentially avoid renal biopsy, available non-invasive biomarkers remain limited. Therefore, identifying key molecules in IMN pathogenesis is important. This may help us understand disease mechanisms and provide clues for future research.

Urine collection is a low-cost and non-invasive detection method.⁷ Different from renal biopsy that only reflects pathological characteristics at a specific time point, urine samples have unique advantages in reflecting the dynamic changes of renal lesions.⁸ Moreover, renal biopsy only obtains tiny local renal tissue samples, and it remains uncertain whether these samples can represent the overall information of the kidney. Since all functional nephrons participate in the process of urine formation, urine can reflect the pathological state of the kidney more comprehensively than renal biopsy.⁹ Previous studies have identified early biomarkers of allograft injury through urinary transcriptome analysis.¹⁰ Notably, RNA sequencing (RNA-seq) results have shown that the diversity of immune cell types in urinary cells is higher than that in biopsy samples, suggesting that urinary specimens have superiorities in exploring the immune mechanisms of diseases.¹⁰ These findings indicate that urinary transcriptomic studies in IMN warrant further exploration.

Second-morning urine is considered an ideal sample type in urinary transcriptomic studies. A urinary metabolomics study found that, compared with first-morning urine, metabolite profiles in second-morning urine were minimally affected by dietary intake from the previous day. In addition, with respect to cortisol level measurement, the collection timing of second-morning urine avoids the pronounced fluctuations associated with the post-awakening period, leading to the recommendation of second-morning urine as the preferred sample for urinary metabolomic biomarker studies.¹¹ Furthermore, a scRNA-seq study of focal segmental glomerulosclerosis (FSGS) utilized second-morning urine as the sequencing sample and mapped the landscape of urinary immune cells and renal epithelial cells, providing experimental evidence for subsequent biomarker screening.¹² Another proteomic study on urine samples showed that first-morning urine had the highest protein concentration, while random urine exhibited the greatest coefficient of variation in protein concentration; second-morning urine achieved a relatively balanced combination of protein concentration and variability.¹³ Collectively, these findings support second-morning urine as a reliable sample type for urinary omics analyses.

Microarray and transcriptomic technologies have been widely applied to the screening and identification of MN-related genes.^14–16 Bulk RNA-seq can provide the average gene expression profile of tissue samples, whereas scRNA-seq allows for the resolution of cellular heterogeneity, although its cost and sample limitations make it unsuitable as a preferred strategy for large-scale screening.^17–19 Integrative transcriptomic analysis can complement the advantages of both approaches and has been increasingly applied in complex kidney disease research in recent years. Noel, Huang et al performed global screening via microarray or bulk RNA-seq, followed by cell localization and mechanistic exploration using scRNA-seq.^20,21 Lee, Liu et al screened novel cell subsets and candidate genes via scRNA-seq, and further explored the molecular regulation of these cell subsets as well as the clinical significance of key genes using microarray or bulk RNA-seq.^22,23 Research teams headed by Zhang, Chen and Dong combined bulk RNA-seq and scRNA-seq, performed differential expression gene (DEG) analysis separately and then extracted the overlapping genes, thereby improving the accuracy of key gene screening.^24–26 Different integrated methods have confirmed the effectiveness of this strategy. However, systematic studies that integrate urinary transcriptomics and kidney tissue transcriptomics for the screening of key molecules in IMN are currently lacking.

Our research group previously collected three types of urine samples from patients with IMN and performed bulk RNA sequencing on urinary exfoliated cells. Based on these data, we established an integrative analysis pipeline. Different from prior approaches that directly implemented machine‑learning algorithms across all three urine specimen groups, our work targeted differentially expressed transcripts uniquely elevated in second‑morning urine. We further intersected these candidates with up‑regulated genes derived from GEO renal tissue datasets to generate a gene pool for subsequent machine‑learning modelling.²⁷ This “urine–kidney” dual-locking strategy was designed to ensure that the selected genes were associated with renal pathology while remaining detectable in non-invasive urine samples. Subsequently, we employed publicly available kidney scRNA-seq data to further explore the expression patterns of key genes at the single-cell level, and used CellChat analysis to predict potential crosstalk between the two cell populations (podocytes and DCs) that highly expressed the key genes. In parallel, we performed immunohistochemistry and immunofluorescence to preliminarily observe the expression and cellular localization of the key genes at the protein level. Validation of gene‑expression signatures and assessment of diagnostic capacity were performed using independent external cohorts. Through the above analyses, KCNN3 and TLR10 were identified as cell type-associated candidate molecules in IMN, warranting further investigation into their functions and underlying mechanisms.

Material and Methods

Collection and Processing of Urine Specimens

All samples were procured from the Second Affiliated Hospital of Guangxi Medical University. Urine samples were collected from 17 patients with biopsy-proven PLA2R-positive IMN.

Inclusion criteria were as follows: (1) Patients admitted within one week of hospitalization; (2) Patients hospitalized primarily for nephropathy. Exclusion criteria included: (1) Anuria, receiving dialysis or critically ill patients; (2) Patients with a history of renal biopsy conducted in the prior three months; (3) Subjects unable or unwilling to cooperate with the research for any reason. Data from 17 healthy controls were acquired from the physical examination department, with age and sex matched to IMN patients. Every participant was fully informed of the study protocol and signed relevant consent paperwork.

Urine specimens were collected as follows: random urine was collected from each patient on the first night, followed by the first morning urine after getting up the next day, and then the second urine sample. All samples were processed within 30 minutes after collection. Each urine sample was transferred into a 50 mL centrifuge tube and centrifuged at 490 g for 10 minutes. After centrifugation, the supernatant was discarded. The tube wall was rinsed twice with DPBS to collect exfoliated urinary cells, and the resultant suspension was transferred into a 1.5 mL EP tube before centrifugation at 2000 rpm for 5 minutes. Following supernatant removal, 1 mL DPBS was added for resuspension, and the centrifugation step was repeated. After removing the supernatant completely with a 10 μL pipette tip to eliminate residual droplets, the urinary cells were stored at −80 °C.

Bulk RNA-Seq of Urinary Cells

The urinary cells were taken out from the −80 °C refrigerator and thawed on ice to 4 °C, followed by lysis with 30 μL AccuraCode Lysis Mix. The ONE Step reaction system was prepared using the AccuraCode® HTP OneStep RNAseq Kit (Singleron, Nanjing, China), and 18 μL of the mixture was dispensed into each well of a high-throughput PCR plate. Afterwards, 2 μL of lysed urinary cells were added to each well and mixed thoroughly by pipetting. The high-throughput PCR plate was placed in a thermal cycler for reverse transcription to obtain amplified products, which were subsequently purified and subjected to quality control.

Library construction consisted of fragmentation reaction setup, adapter ligation, purification after adapter ligation, PCR enrichment and transcriptome library size selection. The constructed libraries were quantified and assessed for fragment size via quality control. Bulk RNA libraries were sequenced on the Illumina NovaSeq system following the paired-end sequencing protocol.

Dataset Collection

The NCBI GEO database was utilized to screen publicly available high-throughput sequencing datasets from patients diagnosed with IMN. One bulk renal RNA-seq dataset (accession number: GSE216841) containing 20 samples was obtained, including 12 IMN patients and 8 healthy controls. Three transcriptomic datasets generated by scRNA-seq were acquired via the GEO repository, detailed as follows: GSE241302 contains 4 samples (3 IMN, 1 control), GSE171458 contains 8 samples (6 IMN, 2 controls), while GSE131685 includes 3 healthy control samples exclusively. Two microarray datasets, GSE108109 (44 IMN, 6 controls) and GSE200828 (51 IMN, 6 controls), and one RNA-seq dataset, GSE197307 (59 IMN, 8 controls), were used as external validation sets. Detailed information for each dataset is provided in Supplementary Table 1.

Analysis of Bulk RNA-Seq Data from Urine Samples and Kidney Tissue

Raw sequencing data from five bulk RNA-seq plates were processed by Singleron Biotechnologies using standard pipelines to generate gene expression matrices. After extracting the common gene set across the five plates, morning urine, second-morning urine, and random urine samples were separately selected from each plate. Matrices with the same sample type were merged to obtain three independent matrices. These three matrices were then combined into a single complete matrix, and genes with expression levels > 1 in fewer than two samples were filtered out to remove lowly expressed genes. Batch effects across plates were corrected using the ComBat_seq function from the R package sva (version 3.52.0), with sequencing plate as the batch covariate and urine sample type as the biological covariate. Principal component analysis (PCA) was performed before and after batch correction to ensure effective removal of batch effects. After correction, the matrix was split back into three separate matrices by sample type (morning urine, second-morning urine, and random urine) for subsequent analyses.

Differential expression analysis between IMN patients and healthy controls was performed separately for the morning urine, second-morning urine, and random urine matrices using the edgeR package (version 3.42.4). Differentially expressed genes (DEGs) were identified with thresholds of |log₂FC| > 1 and FDR < 0.05, and multiple testing was corrected using the Benjamini–Hochberg method. Venn diagram analysis was conducted using the R package VennDiagram (version 1.8.2) to identify genes that were upregulated exclusively in second-morning urine across the three urine sample types. The kidney tissue RNA-seq dataset GSE216841 (12 IMN and 8 controls) downloaded from the GEO database was also analyzed using the edgeR package (version 3.42.4), and DEGs were identified with thresholds of |log₂FC| > 1 and P < 0.05. FDR‑based multiple‑testing correction was not applied for this cohort.

Genes upregulated exclusively in second-morning urine were intersected with genes upregulated in IMN kidney tissues to obtain the commonly upregulated genes, which served as the initial candidate gene set for subsequent machine learning analysis.

Enrichment Analysis of DEGs

Gene Ontology (GO) enrichment and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses were performed on the overlapping genes using the R package clusterProfiler (version 4.10.0). Multiple testing was corrected using the Benjamini–Hochberg method, and an adjusted p-value (FDR) < 0.05 was considered statistically significant. GO enrichment analysis covered three ontologies: biological process (BP), cellular component (CC), and molecular function (MF). The enrichment results were visualized as bubble plots using the R package ggplot2.

Screening of Hub Genes via Machine Learning Algorithms

To screen for key genes, we selected the Least Absolute Shrinkage and Selection Operator (LASSO), Random Forest (RF), and XGBoost algorithms. The three models were constructed as follows: (1) The LASSO model was fitted using the glmnet package (family = “binomial”, alpha = 1). First, five-fold cross-validation was performed using the default lambda sequence to determine the optimal regularization parameter λ_min based on the mean squared error (MSE). To directly obtain the regression coefficients at this λ value, the model was fitted to the full data using a fine grid of 1000 lambda values, and genes with non-zero coefficients at λ_min were extracted. (2) Feature selection was performed using the Random Forest algorithm combined with Recursive Feature Elimination (RFE) using the R package caret. The original candidate gene matrix was used as the input independent variables, with the binary grouping as the dependent variable. Model performance was evaluated using Leave-One-Out Cross-Validation (LOOCV), and accuracy was used as the metric for feature subset evaluation. The gene subset corresponding to the highest accuracy was selected as the final result. (3) The XGBoost algorithm was implemented using the R packages xgboost and caret. The original candidate gene matrix was used as the input independent variables, with the binary grouping as the dependent variable. A 10-fold cross-validation combined with grid search was performed to tune hyperparameters, including the number of trees, maximum depth, learning rate, minimum loss reduction, subsampling ratios of features and samples, and minimum child weight. The optimal parameter combination was selected based on the minimization of mean logarithmic loss (logloss). The final model was retrained on the full dataset using the optimal parameters. Feature importance was evaluated using the Gain metric provided by XGBoost, and a feature importance ranking plot was generated. The discriminative performance of the three models was internally evaluated using cross-validation (five-fold cross-validation for LASSO, leave-one-out cross-validation for RF, and ten-fold cross-validation for XGBoost). ROC curves were plotted based on out-of-fold predicted probabilities to calculate the area under the curve (AUC) and its 95% confidence interval. The optimal cutoff value was determined using the Youden index, and the corresponding sensitivity and specificity were calculated. Detailed performance metrics for each model are provided in Supplementary Tables 2 and 3.

To reduce false positive bias that may be introduced by a single machine learning algorithm, we performed Venn diagram analysis to intersect the feature genes selected by LASSO, Random Forest (RF), and XGBoost, thereby identifying the key genes for IMN. The underlying principles of these algorithms are complementary: LASSO performs sparse feature selection via a penalty mechanism, RF evaluates feature importance through ensemble learning, and XGBoost optimizes classification performance via gradient boosting. Taking the intersection of the three gene sets as the final result may help balance the inherent biases of each algorithm and enhance the robustness of the identified candidate genes.

Single-Cell RNA-Seq Data Analysis

Datasets downloaded from the GEO database were processed using the Seurat R package (version 5.1.0). Cells with gene counts between 200 and 2,500 and mitochondrial gene percentages less than 30% were selected. To remove batch effects introduced by different donor sources, batch correction was performed using the Harmony package. Principal component analysis (PCA) was used for linear dimensionality reduction, and the number of principal components was determined based on the ElbowPlot for UMAP/tSNE visualization and clustering analysis. Major cell types were identified based on known cell-type marker genes. Further subclustering analysis was performed on the target cell populations, followed by visualization.

Cell-Cell Interaction Analysis

CellChat analysis was performed to predict potential cell–cell interactions among different cell types in the single-cell transcriptomic data. The integrated Seurat object containing pooled cells from all 9 IMN donors was used as the input for CellChat analysis. The predictions were based on the expression levels of receptors and ligands, with a focus on ligand–receptor pairs in the CXCL signaling pathway that exhibited differential interaction strength.

Immumohistochemical Staining

We cut paraffin-embedded renal tissues from IMN patients and normal controls into 3 μm thick slices (n = 6 per group), which were then processed for deparaffinization and rehydration. KCNN3 antigen retrieval was implemented in Tris-EDTA buffer (Sevier, China): the solution was boiled at medium-high heat and then simmered at low heat for 15 min. For TLR10, enzymatic antigen retrieval was conducted with trypsin incubation at 37 °C for 35 min, and the digestion was terminated by rinsing sections with PBS buffer. We blocked endogenous peroxidase activity in 3% H₂O₂ solution at ambient temperature for 15 min, and non-specific binding was blocked with primary antibody diluent for 30 min. Subsequently, sections were incubated with primary antibodies against KCNN3 (1:200, TD13233, Abmart, China) and TLR10 (1:100, PAB992Hu01, Cloud-Clone Corp., China). The anti-KCNN3 antibody was incubated at room temperature for 1 h, while the anti-TLR10 antibody was incubated overnight at 4 °C. After thorough washing to remove unbound primary antibodies, sections were incubated with secondary antibodies at room temperature for 30 min. DAB chromogenic reaction was applied, followed by hematoxylin counterstaining for 2 min and bluing in TBST buffer. Finally, all sections were dehydrated and mounted with neutral balsam. Microscopic observation was performed, and the percentage of positive staining area (Area%) in IMN and normal renal tissues was quantified using Fiji software. For the KCNN3 staining, three high-power fields (×400) containing glomeruli were randomly selected from each section, avoiding tissue edges and damaged areas. The percentage of positive area was measured in each field using Fiji software, and the mean value of the three fields was calculated as the final value for that sample. For the TLR10 staining, the renal interstitium was first identified at low-power magnification, and then three non-overlapping high-power fields (×400) were randomly selected from the interstitial regions. The percentage of positive area was similarly measured using Fiji software, and the mean of the three fields was used as the final value. Image acquisition and analysis were performed by a single independent observer who was blinded to the group assignment of the samples. Comparisons between two groups were performed using unpaired Student’s t-test with GraphPad Prism software, and a P-value < 0.05 was considered statistically significant.

Immunofluorescence Staining

Paraffin sections were deparaffinized and rehydrated, followed by antigen retrieval. For KCNN3 and WT1, antigen retrieval was performed using Tris-EDTA buffer (pH 9.0) with heat treatment for 15 min. For TLR10 and CD11c, Tris-EDTA buffer (pH 8.0) was used for heat-induced antigen retrieval for 15 min. After cooling, sections were blocked with 3% BSA at room temperature for 30 min. For the heterologous antibody combination (TLR10/CD11c), the two primary antibodies were incubated together at 4 °C overnight, followed by incubation with matched fluorophore‑conjugated secondary antibodies for 50 min. For the homologous antibody pair (KCNN3/WT‑1), the first primary antibody was applied, then incubated with the matching polymer secondary antibody to enable TSA‑tyramide fluorescent signal deposition. Bound antibodies were thoroughly removed using immunostaining elution buffer, and the above‑mentioned workflow was repeated to accomplish staining for the second primary antibody. Finally, nuclei were counterstained with DAPI, and sections were subjected to autofluorescence quenching before mounting. Images were acquired using a 3DHISTECH Pannoramic MIDI automated digital slide scanner. For TLR10/CD11c immunofluorescence staining, because CD11c-positive cells were sparsely distributed in the renal interstitium, each section was first scanned at low magnification to locate positive regions, followed by high-magnification image acquisition within these regions. For KCNN3/WT1 staining, three glomeruli were randomly selected from each section for observation. Comprehensive details regarding the primary and secondary antibodies applied in immunofluorescence assays are summarized in Supplementary Table 4. Clinical information for tissue sections used for immunohistochemistry and immunofluorescence staining is provided in Supplementary Table 5.

External Validation of Key Genes

We selected two microarray datasets (GSE108109 and GSE200828) and one RNA-seq dataset (GSE197307) from the publicly available GEO database, all derived from the Nephrotic Syndrome Study Network (NEPTUNE) cohort. NEPTUNE is a multicenter prospective cohort study that enrolled patients with primary nephrotic syndrome. From this cohort, we included samples with membranous nephropathy for analysis. Because the control sample sizes in GSE108109 and GSE200828 were relatively small, we merged these two datasets after removing batch effects using the R package sva (version 3.58.0). The expression levels of the key genes in the validation sets were examined using the Wilcoxon rank-sum test, and the diagnostic performance of individual key genes as well as their combination was evaluated using receiver operating characteristic (ROC) curves.

Results

Identification and Functional Enrichment Analysis of DEGs

A total of three categories of urinary specimens were harvested from 17 patients with IMN and 17 healthy volunteers. Urinary cells were isolated from urine samples, and transcriptome sequencing was conducted after reverse transcription and library preparation. The corresponding raw expression matrix was constructed and publicly deposited by our research team in a previous study (DOI: 10.3389/fimmu.2025.1671891). Differential expression analysis was performed to identify upregulated DEGs among the three urine sample groups (Figure 1A–C). As presented in Figure 1D, 289 upregulated DEGs were identified in first morning urine, 1834 in second morning urine, and 988 in random urine. Subsequently, upregulated DEGs from renal tissues of patients with IMN were intersected with those derived from second morning urine, yielding 94 overlapping upregulated DEGs (Figure 1E).

Figure 1.

Identification of DEGs and gene functional enrichment analysis. (A–C) Volcano plots showing upregulated or downregulated DEGs in the first morning urine, second morning urine, and random urine of patients with IMN compared with healthy controls. (D) Intersection of upregulated DEGs in three urine samples from patients with IMN and healthy controls. This panel is derived from the same dataset reported in our group’s previous publication (DOI: 10.3389/fimmu.2025.1671891), and is reproduced with approval from the original authors. The original work was published under a CC‑BY‑4.0 license. (E) Venn diagram showing the intersection of commonly upregulated DEGs in renal tissue and second morning urine of patients with IMN compared with healthy controls. (F) Results of GO enrichment analysis. Panels displayed from top to bottom illustrate enriched terms for BP, CC, and MF, respectively. (G) Results of KEGG enrichment analysis.

Intersecting upregulated DEGs underwent GO and KEGG enrichment analysis to uncover their biological functions and relevant pathways, and bubble plots were generated for result visualization. Analysis of GO enrichment found that the above genes were chiefly concentrated in biological processes (BPs) including organic hydroxy compound transport, regulation of membrane potential, cholesterol transport, and sterol transport. For cellular components (CCs), enrichment was mainly observed in transmembrane transporter complex, transporter complex, and monoatomic ion channel complex. In terms of molecular functions (MFs), the genes were enriched in metal ion transmembrane transporter activity, monoatomic ion channel activity, channel activity, and passive transmembrane transporter activity (Figure 1F). KEGG enrichment analysis demonstrated significant enrichment in nicotine addiction, cholesterol metabolism, the MAPK signaling pathway, and neuroactive ligand–receptor interaction (Figure 1G).

Identification of Hub Genes

Three machine learning algorithms were applied to screen the 94 commonly upregulated differentially expressed genes and further identify key genes. LASSO regression identified 17 significant genes at the optimal λ value (Figure 2A), with a five-fold cross-validated AUC of 0.962 (95% CI: 0.91–1.00). Random Forest combined with Recursive Feature Elimination (RFE) was used to evaluate the effect of gene number on model accuracy, which reached its highest value when using only the five most important genes (Figure 2B), with a leave-one-out cross-validated AUC of 0.841 (95% CI: 0.693–0.989). The XGBoost model provided the relative importance ranking of all genes (Figure 2C), with a ten-fold cross-validated AUC of 0.803 (95% CI: 0.646–0.959). Comprehensive performance indicators across all models are presented in Figure S1. By intersecting the genes identified by the three algorithms, two core genes for IMN were obtained: KCNN3 and TLR10 (Figure 2D).

Figure 2.

Screening of hub genes for IMN using machine learning. (A) LASSO regression for feature selection. Left panel: coefficient path plot. Right panel: cross-validation curve. The LASSO model built with the optimal λ showed a five-fold cross-validated AUC of 0.962 (95% CI: 0.91–1.00). (B) Feature selection results using RF combined with RFE. The corresponding leave‑one‑out cross‑validated AUC was 0.841 (95% CI: 0.693–0.989). (C) Feature importance ranking plot of the XGBoost model. The x-axis represents the feature importance score (Gain), and the y-axis represents the genes. The model showed a ten‑fold cross‑validated AUC of 0.803 (95% CI: 0.646–0.959). (D) Venn diagram showing two hub genes (KCNN3 and TLR10) identified by the above three algorithms.

Immunohistochemical Validation of KCNN3 and TLR10 Expression

We conducted immunohistochemical staining targeting KCNN3 and TLR10 on renal biopsy slices from IMN patients, as well as normal peritumoral renal tissues adjacent to renal carcinoma, to validate the protein abundance of the two pivotal genes. Immunohistochemical staining showed that KCNN3 expression was upregulated in IMN tissues compared with controls (Figure 3A). Similarly, TLR10 expression was elevated in IMN samples, whereas control tissues displayed lower protein levels (Figure 3B). According to quantitative measurements, the IMN group exhibited a KCNN3‑immunopositive area fraction of 16.64% ± 2.35%. This value was markedly higher versus controls (0.77% ± 0.34%; n = 6, unpaired t‑test, P = 0.0010; Figure 3C). TLR10 was also significantly upregulated in IMN tissues (7.518% ± 0.6595% vs 0.2388% ± 0.06814%; n = 6, unpaired t-test, P < 0.0001; Figure 3D).

Figure 3.

KCNN3 and TLR10 protein expression is upregulated in IMN patients. (A) Representative immunohistochemical staining images of KCNN3 in kidney tissues from the IMN (n = 6) and control (n = 6) groups. Left panels: 10× magnification, scale bars = 200 μm; right panels: 20× magnification, scale bars = 100 μm. (B) Representative immunohistochemical staining images of TLR10 in kidney tissues from the IMN (n = 6) and control (n = 6) groups. Left panels: 10× magnification, scale bars = 200 μm; right panels: 20× magnification, scale bars = 100 μm. (C) Quantitative assessment of KCNN3‑positive area proportions. (D) Quantitative assessment of TLR10‑positive area proportions. Statistical analysis was performed using unpaired t‑test. ***P < 0.001, ****P < 0.0001. Image quantification was performed using Fiji software to calculate the percentage of positive‑staining area from randomly selected fields of view.

ScRNA-Seq Analysis Shows the Distribution of Key Genes Across Cell Subpopulations

Raw scRNA-seq datasets (GSE241302, GSE171458, GSE131685) were retrieved from the GEO database. After integration and quality control, the Harmony algorithm was applied to correct batch effects from different donors. Following correction, cells from different donors were uniformly distributed in the tSNE dimensionality reduction space, with no obvious donor-specific clustering observed (Figure S2). Relying on canonical marker genes sourced from the CellMarker 2.0 database, all cells were grouped into 29 clusters (Figure 4A) and annotated into 13 cell types: proximal tubular epithelial cells, loop of Henle cells, distal tubular cells, intercalated cells of collecting duct, principal cells of collecting duct, endothelial cells, mesangial cells, podocytes, parietal epithelial cells, smooth muscle cells, macrophages, dendritic cells (DCs), and NKT cells (Figure 4C). Feature genes for each cluster and cell type are shown in Figure 4B and D. Next, the cellular localization of KCNN3 and TLR10 was explored. KCNN3 was enriched in podocytes, whereas TLR10 was predominantly expressed in DCs (Figure 4E). Density-colored t‑SNE visualization confirmed the enrichment of KCNN3 within podocytes and TLR10 within DCs (Figure 4F and G).

Figure 4.

Single-cell transcriptomic analysis showed cell clusters and marker gene expression patterns. (A) A total of 113,216 cells from publicly available scRNA-seq datasets were clustered into 29 clusters. (B) Bubble plot showing marker genes of each cluster. (C) t-SNE plot showing the annotation of 13 cell types. (D) Bubble plot showing marker genes of the 13 cell types. (E) Bubble plot showing that KCNN3 and TLR10 were enriched in podocytes and DCs, respectively. (F) Density‑colored t‑SNE plot illustrating prominent high expression of KCNN3 predominantly in podocytes. (G) Density‑colored t‑SNE plot showing that TLR10 is highly enriched in DCs.

KCNN3 is Highly Expressed in Podocyte Subclusters

To further identify the specific podocyte subpopulation with high expression of the key genes, we re-clustered podocytes and compared the transcriptomes of Cluster 0 and Cluster 1. A total of 998 DEGs were identified, with 499 upregulated in Cluster 0 and 499 in Cluster 1 (Figure 5A). A heatmap was generated to illustrate the top 10 most markedly differentially expressed genes across individual clusters (Figure 5B). Bubble plots show that Cluster 0 exhibits high expression of KCNN3, whereas Cluster 1 exhibits high expression of PLA2R1 (Figure 5C). This expression pattern is also visible in the t‑SNE plots (Figures 5D–F).

Figure 5.

Podocyte subclustering. (A) Volcano plot showing the distribution of differentially expressed genes between Cluster 0 and Cluster 1 of podocytes. (B) Heatmap showing distinct expression patterns of DEGs in the two clusters. (C) Density plot showing that KCNN3 is mainly expressed in Cluster 0, whereas PLA2R1 is mainly expressed in Cluster 1. (D) t‑SNE plot showing the expression distribution of KCNN3 and PLA2R1 in podocytes, with KCNN3⁺ and PLA2R1⁺ cell populations indicated. (E) t‑SNE plot showing the KCNN3⁺ podocyte population. (F) t‑SNE plot showing the PLA2R1⁺ podocyte population.

TLR10 is Highly Expressed in DC Subclusters

Similarly, to dissect the transcriptional heterogeneity within dendritic cells, we extracted dendritic cells for re-clustering. The dendritic cell population was divided into Cluster 0 and Cluster 1, and the DEGs between them were analyzed (Figure 6A). A heatmap showed the top 10 most significantly differentially expressed genes in each cluster (Figure 6B). t‑SNE plots showed that IL1‑B was mainly expressed in Cluster 0, whereas TLR10 was mainly expressed in Cluster 1 (Figure 6C and D). Density t‑SNE plots further indicated that both genes were predominantly distributed within their respective clusters (Figure 6E and F). A bubble plot displayed the expression differences of IL1‑B and TLR10 between Cluster 0 and Cluster 1 (Figure 6G).

Figure 6.

DC subclustering. (A) Volcano plot showing the distribution of upregulated DEGs between DC Cluster 0 and Cluster 1. (B) Heatmap visualizing the expression patterns of these upregulated DEGs in the two clusters. (C) and (D) t‑SNE plots showing DC subclustering. Plot (C) displays DC Cluster 0 and Cluster 1 obtained via unsupervised clustering, while plot (D) highlights IL‑1B⁺ and TLR10⁺ subpopulations. (E) and (F) Density t-SNE plots depicting the IL-1B-expressing and TLR10-expressing DC groups, respectively. (G) Bubble plot showing that IL-1B is mainly expressed in Cluster 0 and TLR10 in Cluster 1.

Predicted Cell–Cell Communication Between KCNN3⁺ Podocytes and TLR10⁺ DCs

To investigate the intercellular communication between KCNN3⁺ podocytes and TLR10⁺ DCs in IMN, we performed CellChat analysis. Figure 7A shows the extensive communication network among different cell types. Previous studies have shown that the CXCL signaling pathway plays an important role in kidney injury and fibrosi; however, its involvement in membranous nephropathy has been less frequently reported.^28,29 We therefore focused on the CXCL signaling pathway. Chord diagrams indicated communication links between KCNN3⁺ podocytes and TLR10⁺ DCs within the CXCL signaling pathway (Figure 7B), with podocyte KCNN3⁺ cells acting as signal senders and DC TLR10⁺ cells as signal receivers (Figure 7C). Figure 7D shows that CXCL12–CXCR4 was the most contributing ligand–receptor pair in the communication between KCNN3⁺ podocytes and TLR10⁺ DCs (highlighted by the red dashed box). Violin plots indicated that CXCL12 was predominantly expressed in KCNN3⁺ podocytes, whereas CXCR4 showed the highest expression in TLR10⁺ DCs (highlighted by the red dashed box; Figure 7E). Collectively, these results suggest a potential CXCL12–CXCR4-mediated communication link between KCNN3⁺ podocytes and TLR10⁺ DCs in IMN.

Figure 7.

CellChat analysis of intercellular interactions. (A) Overview of intercellular interactions. (B) Overview of the CXCL signaling pathway. (C) In the CXCL signaling pathway, KCNN3⁺ podocyte subcluster acts as the primary signal sender, and TLR10⁺ DC subcluster serves as the main signal receiver. (D) Dot plot suggests that the Podo‑KCNN3 subcluster and the DC‑TLR10 subcluster may interact predominantly via the CXCL12‑CXCR4 signaling axis. (E) Violin plot showing the expression of CXCL12 and CXCR4 signaling molecules. Red dotted rectangles in panels (D) and (E) highlight the key CXCL12‑CXCR4 signal‑related regions.

Key Gene Expression Validation and Diagnostic Performance Evaluation

In the training set (17 IMN and 17 controls), both KCNN3 and TLR10 were significantly upregulated in the IMN group compared with the control group (Figure 8A). KCNN3 showed an AUC of 0.796 (95% CI: 0.662–0.930) (Figure 8B), and TLR10 showed an AUC of 0.657 (95% CI: 0.528–0.787) (Figure 8C). The dual‑gene signature achieved an AUC of 0.832 (95% CI: 0.705–0.959), with better diagnostic performance than single‑gene models (Figure 8D). To externally validate the two key genes, we utilized two GEO microarray datasets (GSE108109 and GSE200828; 95 IMN and 12 controls) and one GEO RNA‑seq dataset (GSE197307; 59 IMN and 8 controls). Datasets GSE108109 and GSE200828 were integrated and analyzed following batch‑effect correction. PCA profiles before and after batch‑effect correction are presented in Figure S3. In the merged dataset, KCNN3 expression was significantly upregulated in IMN samples compared with controls, whereas TLR10 showed no significant difference (Figure 9A). KCNN3 achieved an AUC of 0.726 (95% CI: 0.541–0.912), indicating moderate diagnostic capability (Figure 9B). TLR10 showed an AUC of 0.579 (95% CI: 0.429–0.729), suggesting limited diagnostic value (Figure 9C). The combination of KCNN3 and TLR10 yielded an AUC of 0.705 (95% CI: 0.575–0.836) (Figure 9D). In the GSE197307 dataset, TLR10 expression was significantly upregulated in IMN samples compared with controls, whereas KCNN3 showed no significant difference (Figure 9E). KCNN3 yielded an AUC of 0.667 (95% CI: 0.492–0.843), indicating limited diagnostic value (Figure 9F). TLR10 achieved an AUC of 0.724 (95% CI: 0.581–0.866), showing moderate diagnostic capability (Figure 9G). The combination of KCNN3 and TLR10 produced an AUC of 0.748 (95% CI: 0.587–0.909), suggesting improved diagnostic performance over either gene alone (Figure 9H). Although the combined model achieved AUCs above 0.7 in both validation sets, the gain over single-gene models was relatively limited.

Figure 8.

Expression and diagnostic performance of key genes in the training set. (A) Boxplots showing the expression levels of KCNN3 and TLR10 in the IMN and control groups. (B) ROC curve of KCNN3. (C) ROC curve of TLR10. (D) ROC curve of the combination of KCNN3 and TLR10.

Figure 9.

Expression and diagnostic performance of key genes in the external validation sets. (A) Boxplots showing the expression of KCNN3 and TLR10 in the merged GSE108109 and GSE200828 dataset (batch-corrected). (B) ROC curve of KCNN3. (C) ROC curve of TLR10. (D) ROC curve of the combination of KCNN3 and TLR10. (E) Boxplots showing the expression of KCNN3 and TLR10 in the GSE197307 dataset. (F) ROC curve of KCNN3. (G) ROC curve of TLR10. (H) ROC curve of the combination of KCNN3 and TLR10.

Immunofluorescence Localization of Key Genes

Double immunofluorescence staining was performed using WT1 (a podocyte marker) and KCNN3. The results showed that KCNN3 exhibited a cytoplasmic distribution in IMN kidney tissues, with some signals closely adjacent to WT1‑positive podocyte nuclei (yellow arrows). Non‑malignant peritumoural kidney specimens obtained from patients with renal carcinoma exhibited limited KCNN3‑immunopositive staining in glomeruli (Figure 10A). These observations provide morphological clues for the association of KCNN3 with podocyte‑related regions in IMN glomeruli. In addition, double immunofluorescence staining was performed using CD11c (a widely used immune cell marker for identifying DCs and mononuclear phagocytes) and TLR10. In CD11c‑positive cells within the renal interstitium, CD11c staining was predominantly localized to the cell membrane. In the IMN group, TLR10 signals were observed in the cytoplasm of CD11c‑positive cells, whereas such signals were relatively scarce in the control group (Figure 10B). These findings suggest a spatial association between TLR10 and CD11c‑positive immune cells in the IMN renal interstitium.

Figure 10.

Representative double immunofluorescence staining images of renal tissues from IMN patients and control subjects (adjacent normal tissues from renal cancer patients). (A) Staining of KCNN3 (red) and WT1 (green, podocyte marker). In IMN glomeruli, KCNN3 signals were predominantly distributed in the cytoplasm, with some signals localized in the perinuclear regions of WT1‑positive podocytes (indicated by yellow arrows). No obvious KCNN3‑positive signals were detected in glomeruli from the control group. Scale bar = 50 μm. Three independent biological replicates (n = 3) were used for each group. (B) Staining of TLR10 (red) and CD11c (green, a marker for myeloid immune cells). In the IMN group, CD11c‑positive cells showed cytoplasmic TLR10 signals, whereas the control group exhibited such signals only infrequently. Scale bar = 50 μm. Three independent biological replicates (n = 3) were used for each group.

Discussion

MN remains one of the leading causes of adult nephrotic syndrome. Its natural clinical course conforms to the one-third rule: one-third of patients attain spontaneous remission, one-third develop persistent proteinuria, and the remaining one-third progress to end-stage kidney disease (ESKD).³⁰ The pathogenesis of IMN is closely associated with immune dysregulation, in which podocytes are damaged by autoantibody attack, leading to proteinuria.³¹ Current research on immune cells in MN has mainly focused on B cells and T cells. Existing evidence implies dendritic‑cell participation in IMN pathogenic processes; nevertheless, available research on this topic remains comparatively sparse.³² In this study, we identified two key candidate molecules, KCNN3 and TLR10, by integrating transcriptomic analyses of IMN urine and kidney tissue. Single‑cell RNA sequencing suggested that they were enriched in podocytes and dendritic cells, respectively, and CellChat analysis further predicted potential intercellular communication between these two cell populations. Immunohistochemistry supported the protein expression of these molecules, and immunofluorescence results suggested spatial associations of KCNN3 with podocytes and TLR10 with CD11c‑positive immune cells. These findings provide candidate molecular clues for future studies on IMN.

Potassium calcium-activated channel subfamily N member 3 (KCNN3, also termed SK3) encodes small-conductance Ca²⁺-activated K⁺ channels, which participate in the regulation of cellular ionic homeostasis, maintenance of electrophysiological stability, and intracellular signal transduction.³³ The functions of KCNN3 have been extensively studied in the nervous system and cardiovascular system, particularly in cardiac arrhythmias.^34,35 Studies have shown that KCNN genes play a role in polysubstance addiction (alcohol, nicotine, and opioids), which is consistent with our KEGG enrichment results.^36,37 KCNN3 belongs to the calcium‑activated potassium channel family, which includes large‑conductance (BK), intermediate‑conductance (IK), and small‑conductance (SK1‑3) calcium‑activated potassium channels.³⁸ In the kidney, the role of BK channels in regulating K⁺ secretion in the collecting duct has been well established.³⁹ In addition, studies have also suggested that BK channels are closely associated with podocyte injury. In IMN, podocyte injury is a critical component of the disease pathogenesis.^40–42 Autoantibodies attack podocytes, leading to podocyte injury, decreased glomerular filtration rate, and loss of renal function.⁴³ Activation of BK channels on podocytes can trigger intracellular K⁺ loss and Ca²⁺ overload, which in turn activate downstream damage signaling pathways and induce podocyte apoptosis.^44,45 Of note, SK3 channels can dynamically couple with TRPV4 to activate BK channels and regulate K⁺ secretion in the renal collecting duct.⁴⁶ However, the role of SK3 (KCNN3) in podocyte injury has not yet been reported. In the present study, single‑cell analysis showed that KCNN3 was highly expressed in podocytes, and immunofluorescence indicated that some KCNN3 signals were adjacent to WT1‑positive podocyte nuclei, suggesting a potential spatial association with podocytes. This finding is consistent with the characterization of IMN as a podocytopathy.⁴⁷ Combined with the aforementioned literature, we speculate that the high expression of KCNN3 in podocytes may affect BK channel activity via SK3 channels, thereby contributing to podocyte injury; however, this hypothesis requires experimental validation.

Toll-like receptor 10 (TLR10), a transmembrane protein and a member of the Toll-like receptor (TLR) family, is predominantly involved in innate immune responses in humans.^48–50 TLR10 is highly expressed in B cells, with low expression also detected in human plasmacytoid dendritic cells (pDCs), regulatory T cells, and monocytes/macrophages.⁵¹ Distinct from other TLR members, TLR10 is the sole family member with anti-inflammatory properties.^52,53 It has been studied to some extent in rheumatic and autoimmune diseases.^54,55 Findings from Dudkova’s group indicated marked TLR10 downregulation in individuals suffering from lupus nephritis. Such molecular changes might participate in SLE pathogenesis by dampening anti‑inflammatory responses.⁵⁶ Another study suggested that TLR10 may be involved in the progression of SLE by regulating unswitched memory B cells.⁵⁷ However, studies on TLR10 in kidney diseases remain limited, and its role in chronic kidney disease has not yet been reported. In the present study, TLR10 was found to be upregulated in IMN, single‑cell analysis showed its enriched expression in DCs, and immunofluorescence indicated that TLR10 signals were observed in the cytoplasm of CD11c‑positive cells, suggesting a potential spatial association between TLR10 and immune cells in IMN. Studies have shown that activation of TLR10 can suppress pDCs.⁵⁸ pDCs play immunomodulatory roles in various kidney diseases. In acute kidney injury, renal transplant rejection, and IgA nephropathy, pDCs contribute to disease progression by producing proinflammatory cytokines.^58–61 In contrast, in doxorubicin‑induced chronic kidney disease, pDCs pretreated with LPS were converted to a protective phenotype, exhibiting anti‑inflammatory effects and alleviating kidney injury.⁶² In addition, in renal cell carcinoma, the infiltration proportion of pDCs was higher in early‑stage tumors than in advanced tumors, suggesting that pDCs may participate in immune regulation during the early phase of tumor development.⁶³ In recent years, therapeutic strategies targeting DCs have been explored in chronic kidney diseases.^64,65 However, the role of pDCs in IMN remains unclear. Given the enriched expression of TLR10 in DCs and its known immunomodulatory properties, we speculate that the upregulation of TLR10 in IMN may contribute to disease progression by modulating DC function. Nevertheless, the precise regulatory mechanisms warrant further functional investigation.

CXCL12 (C-X-C motif chemokine ligand 12), also known as stromal cell-derived factor-1 (SDF-1), exerts biological functions primarily via binding to its receptor CXCR4.⁶⁶ CXCL12 has been reported to be associated with multiple renal disorders. Existing data indicate that podocytes constitutively express CXCL12.^66,67 Sha et al reported that CXCL12 was significantly increased in the serum and urine of MN patients, which may contribute to disease progression.⁶⁸ Our single‑cell data also showed that CXCL12 was expressed in the Podo‑KCNN3 population, consistent with the podocyte origin of CXCL12 reported in that study. Studies have reported that blocking this pathway in lupus nephritis reduces immune cell infiltration and podocyte injury.^69,70 In MN, this signaling acts within podocytes, and targeting it reduces podocyte apoptosis.⁶⁸ Similarly, in diabetic kidney disease, targeting this pathway also reduces podocyte apoptosis and fibrosis.⁷¹ Furthermore, the CXCL12–CXCR4 signaling pathway is also involved in communication between parenchymal cells and immune cells in kidney diseases. In immunoglobulin A nephropathy (IgAN), mesangial cells and macrophages interact through CXCL12–CXCR4, and blockade of this pathway alleviates inflammation and renal damage.⁷² In ANCA‑associated glomerulonephritis (AAGN), endothelial cells release CXCL12 to recruit CXCR4⁺ monocytes, promoting macrophage polarization and crescent formation, whereas blockade of this pathway attenuates renal fibrosis.⁷³ Our single‑cell data observed a potential CXCL12–CXCR4‑mediated communication link between Podo‑KCNN3 and DC‑TLR10 populations, providing preliminary clues for the potential role of this signaling axis in IMN. However, its specific role in IMN remains to be validated.

However, this study has several limitations. First, the sample size for machine learning training was relatively small (17 IMN and 17 controls), which may increase the risk of overfitting. Although we employed cross‑validation to mitigate this issue, the relatively high AUC in the training set and the decreased AUC in cross‑validation still suggest the possibility of overfitting. This leads to overestimation of the model’s classification performance and limits its generalizability. Second, we reused urinary transcriptomic datasets for the current investigation from a previously reported cohort generated by our research group. Although our screening strategy differed from that of previous studies, these data were not derived from a completely independent new cohort, which remains a limitation of this study. Third, this study integrated urinary bulk RNA‑seq and public kidney tissue RNA‑seq data. Although this integrative approach offered advantages in screening key molecules, the inherent differences between sample types and across different cohorts still constitute potential data heterogeneity, which may affect the robustness of the findings. Fourth, owing to the limited sample size of the public renal‑tissue bulk RNA‑seq dataset (12 IMN patients versus 8 controls), stringent FDR‑based multiple‑testing correction would substantially increase false‑negative rates under small‑sample conditions and risk discarding biologically meaningful differential molecules. We therefore adopted raw P‑values rather than FDR correction for DEG detection in this cohort. Nevertheless, candidate genes were required to be concurrently upregulated in both this renal dataset and our in‑house urine transcriptomic dataset, where rigorous Benjamini‑Hochberg FDR correction was implemented. This dual‑cohort overlapping strategy partially mitigates the false‑positive risk originating from uncorrected DEG calls in the public renal cohort. Even so, the small sample size of this public dataset remains an inherent methodological limitation, and future validation in larger independent cohorts is warranted. Fifth, in our external validation, the expression and diagnostic performance of KCNN3 and TLR10 showed heterogeneity across different validation datasets. GSE108109 and GSE200828 are microarray data, whereas GSE197307 is bulk RNA‑seq data. This heterogeneity may be related to platform differences and variations in patient characteristics across different enrollment periods, suggesting that the expression patterns of KCNN3 and TLR10 may differ across IMN patient populations. Current validation datasets lack full clinical metadata. Future work will demand larger cohorts carrying comprehensive clinical records to evaluate the expression signatures of these two genes across various IMN subgroups. Sixth, cell‑cell communication results derived from pooled IMN cells represent cohort‑level observational trends, and further validation in larger single‑cell cohorts is warranted. In addition, we attempted double‑immunofluorescence staining to localize key molecules with relevant cell populations. Constrained by the subcellular signal patterns of these targets, we could not obtain results suitable for robust quantitative co‑localization analysis. Finally, the specific roles of KCNN3 and TLR10 in IMN, as well as whether genuine crosstalk exists between podocytes and DCs, require further functional validation. Notwithstanding these noted caveats, outputs generated from our screening workflow could yield tentative pointers for follow‑up mechanistic studies into IMN.

Our group’s earlier work based on urinary‑cell transcriptomic data from IMN patients aimed to screen urine‑origin biomarkers and investigate the pro‑inflammatory and pro‑fibrotic roles of tubular‑cell‑derived SPP1 in IMN. In the present study, we identified two core candidate molecules of IMN, namely KCNN3 and TLR10, by leveraging both urinary and renal‑tissue transcriptomic profiles. Combined with single‑cell sequencing data and protein‑level assays, we preliminarily characterized their expression patterns within renal cell populations (podocytes, DCs). We further computationally inferred potential cell‑to‑cell crosstalk between these cell subsets. In comparison with our previous work centered on tubular cells, these observations offer one research avenue worth further exploration for understanding IMN from the podocyte‑immune‑cell perspective. Nevertheless, the exact biological functions and underlying mechanisms of KCNN3 and TLR10 still require further validation.

Conclusion

In this study, by integrating transcriptomic analysis and machine learning, we identified KCNN3 and TLR10 as cell type‑associated candidate molecules in IMN. These two molecules may be associated with podocytes and DCs, respectively. However, their functions and mechanisms warrant further investigation.

Funding Statement

This study was financially supported by the Joint Project on Regional High-Incidence Diseases Research of Guangxi Natural Science Foundation (Grant No. 2024GXNSFAA010316), together with the Guangxi Medical and health key discipline construction project.

Ethics Statement

The protocol of the present investigation received ethical authorization from the Medical Ethics Committee, Second Affiliated Hospital of Guangxi Medical University (No. 2024-KYL (030)). All research procedures strictly adhered to the principles outlined in the Declaration of Helsinki. Every study participant signed documented informed consent prior to enrollment.

Disclosure

The authors declare that there are no conflicts of interest regarding this study.