Defining PI3K Signaling as a Key Regulator of Molecular Heterogeneity and Inflammation in Glomerular Lupus Nephritis

Abstract

Lupus nephritis (LN) is a critical manifestation in systemic lupus erythematosus. However, there has been little progress in treatment outcomes, and the discovery of robust and clinically relevant molecular subtypes would be informative for guiding stratified therapy for LN. Glomerular transcriptomic datasets of LN were obtained from publicly repositories. We investigate the molecular heterogeneity of LN by analyzing renal glomerular transcriptomic data from 135 patients with LN using integrated optimal algorithms and propose a refined two-molecular subtype scheme for LN (designated GC1 and GC2). GC1 represents an active form of LN enriched by IFN signaling pathways, Fc receptor-mediated signals, and neutrophil activation-associated processes. In contrast, GC2 is a subset characterized by class A/1 (rhodopsin-like) and B cell (BC) receptor signaling pathways. The communication probability by receptor-ligand pairs is highly scored between BCs, CD4+ T cells and macrophages/monocytes. Proteinuria and renal insufficiency were more significant in GC1. We identified PIK3R2 as a hidden regulator in the disease module, and inhibition of PI3K subunits showed promising efficacy in counteracting altered molecular signatures in an active state of LN. The proposed two-molecular subtype scheme offers novel insights into the divergent mechanistic features of LN and represents a critical step toward precision medicine for LN.

Graphical Abstract

INTRODUCTION

Systemic lupus erythematosus (SLE) is an autoimmune disease of unknown cause involving loss of immune tolerance to endogenous nuclear material, leading to systemic autoimmunity that can damage various tissues and organs. Lupus nephritis (LN) is a form of glomerulonephritis (GN) and one of the most common and severe organ manifestations of SLE, affecting 20%–40% of patients. It is characterized by the deposition of immune complexes, complement activation, and a wide variety of immune cell infiltration, leading to irreversible glomerular injury (1,2,3,4). Despite the availability of treatment guidelines, the renal survival rate has remained stationary over a decade of evaluation, and 10%–30% of patients with LN progress to end-stage kidney disease (1,4). This discouraging outcome poses a major challenge and insights into LN’s current pathophysiology and classification system are needed.

Kidney biopsy is the gold standard for diagnosing LN and defining the pathology underlying the observed clinical phenotype. However, there is a discrepancy between histological findings and clinical response (5,6,7). Early clinical and histologic outcomes are discordant in proliferative LN, and neither correlates with long-term renal outcome. The kidney accrues chronic damage despite clinical response. Furthermore, urinary biomarkers did not consistently correlate with histological and clinical indices across studies (8,9). These results justify an elaborate investigation into the cellular and molecular heterogeneity and complexity behind the histological phenotypes.

Although a previous study using transcriptomics demonstrated concordant alterations in metabolic pathways in various forms of GN, including LN (10), alterations specific to LN need to be explored in a comprehensive way. Two recent studies identified class-oriented enriched pathways and treatment response-oriented molecular signatures, but these results were derived solely from an immunology panel that included a limited number of immune-related genes (11,12). Integrative systematic analysis of high-throughput omics datasets and the identification of disease-oriented modular networks that enlighten the molecular and cellular interactions of disease subtypes have revolutionized the understanding of the shared and divergent characteristics of complex diseases (13). Compared to individual studies, integrative analysis of multiple datasets has the merit of low sampling bias and a high chance of detecting minor but important characteristics of a complex disease.

In the present study, we systematically searched public data repositories and biomedical libraries for kidney transcriptomic datasets from patients with LN and specifically focused on glomeruli samples. We conducted unsupervised consensus clustering on the gene expression profiles of glomeruli and systematically dissected their molecular and cellular characteristics. We applied a network-based Bayesian inference method to define the key drivers of the disease subtypes. Finally, we evaluated the network-based perturbing effect of investigational drugs on each subtype.

MATERIALS AND METHODS

Systematic search and data collection

We used the keywords ‘systemic lupus erythematosus,’ ‘lupus nephritis,’ ‘glomerulonephritis,’ ‘kidney,’ ‘renal,’ ‘transcriptomics, microarray, or RNA-sequencing,’ ‘dataset’ in Google Scholar and PubMed (https://pubmed.ncbi.nlm.nih.gov/) to find relevant publications to the topic of LN signatures of patients with SLE. We retrieved all publications that were accompanied by high-throughput datasets and finally selected 13 datasets in total with ArrayExpress (https://www.ebi.ac.uk/arrayexpress/) and GEO (https://www.ncbi.nlm.nih.gov/geo/) series IDs: GSE32591, GSE60860, GSE69438, GSE99325, GSE99339, GSE104948, GSE104954, GSE112943, GSE113342, GSE127797, GSE157293, GSE175759, GSE200306 (Supplementary Tables 1 and 2). Four datasets of glomerulus origin (GSE32591, GSE99339, GSE104948, GSE127797) were finally selected and the combined datasets included 135 LN patient’s samples and 17 normal controls (NCs), and covered the 11,560 genes in common. Two datasets (GSE113342, GSE200306) were sequenced only for immunology panel of 606 genes, and GSE200306 was selected for validation because it has a larger number of samples and clinical information. For an analysis of lupus-prone animal model, we imported a microarray dataset of renal glomeruli tissues of NZM2328 mice (GSE206806). NZM2328 lupus-prone mouse strain as a model of human LN, with severe immune-complex (IC)-mediated nephritis and early mortality predominantly affecting female mice (14,15). Single cell RNA sequencing data of LN was imported from the ImmPort repository (accession SDY997) (16), which was processed using Seurat package according to the standard protocol (17) and annotated by singleR package (18). The summary of the data processing and analysis steps was illustrated in Supplementary Fig. 1.

Data normalization and removal of batch effects

For one-channel arrays, the robust multi-array average method was applied on the image data for a set of replicates for background correction, normalization, probe-set summarization using R package affy (19). Preprocessing of Illumina gene expression array data was performed using the lumi package and normalized using robust spline normalization of log-transformed raw data (20). Residual technical batch effects arising due to heterogeneous data integration were corrected using the ComBat function (21,22). Quality assurance and distribution bias was evaluated by principal component analysis. After preprocessing, the gene expression profiles have a significant reduction of systematic, dataset-specific bias in comparison with the same dataset before normalization and batch correction (Supplementary Fig. 2).

Unsupervised consensus clustering of gene expression profiles

In order to assess clustering patterns of gene expression profiles for significance, we used the Monte Carlo Reference-based Consensus Clustering (M3C) package, which uses a Monte Carlo simulation to generate null distributions of stability scores along the range of K which, by comparing with the real stability scores, are used to decide the optimal K and reject the null hypothesis K=1 (23,24,25). M3C was run on 1000 Monte Carlo iterations and 100 inner replications. The within-cluster consensus score, ranging between 0 and 1, is defined as the average consensus value for all pairs of individuals belonging to the same cluster. A value closer to one indicates better cluster stability. The cumulative distribution function (CDF), adjusted p-values, entropy, and proportion of ambiguous clustering (PAC) were applied to find the optimal K (26). The CDF curves show a flat middle segment only for the true K, suggesting that very few sample pairs are ambiguous when K is correctly inferred (27,28). As the lower the p-value and entropy are, the K is more significant (29). The PAC is defined as the fraction of sample pairs with consensus index values falling in the intermediate sub-interval (x₁, x₂) ∈ [0, 1]. A low value of PAC indicates a flat middle segment, allowing inference of the optimal K by the lowest PAC. Cluster stability was further validated using the Jaccard similarities index (30).

Classication using a gene expression-driven classier

The discovery datasets included 135 samples from four datasets (GSE32591, GSE99339, GSE104948, GSE127797) and the validating datasets include 34 samples from a dataset (GSE200306). We constructed a classier using a K-nearest neighbor (KNN) learning algorithm (31). Key predictors were sorted out using R package Boruta, a wrapper around a Random Forest classification algorithm for feature selection (32). For training the classier, we used the 65 featured genes and subgroup labels of the result of the M3C clustering process for the training samples. The classifier was further optimized by 10-fold cross validation during fitting the model. K parameter was evaluated from 1 to 20, with the highest accuracy of 0.9857 achieved at K=16. The final model was optimized at K=16, utilizing the Euclidean distance as the distance metric. Finally, we applied the fitted KNN classier to assign subgroups to the validation samples. The classifier reliably categorized LN samples in the testing set into two subtypes and a classification performance was estimated by area under the curve (AUC)-receiver operating characteristic (ROC).

Filtering of differentially expressed genes (DEGs)

In order to identify the DEGs, we used limma package, a software designed for the analysis of gene expression involving comparisons between many gene targets simultaneously (33). The p-values were adjusted by the Benjamin-Hochberg method, and the adjusted p-value cutoff of 0.05 and the absolute value of fold change >1.5 was applied to obtain the DEGs.

Pathway- and cell subset-driven enrichment analysis

We performed functional enrichment analysis focusing on the list of upregulated DEGs using the Enrichr software (34). The adjusted p-value was made using the Benjamini-Hochberg method for correction for multiple hypotheses testing and the terms were considered as significant if the adjusted p-value was less than 0.05. For the biological processes or signaling pathways, gene-set enrichment analysis (GSEA) was carried out using the GSEA software from the Broad Institute to assess their overrepresentation (35). Normalized enrichment score (NES) was used to compare analysis results across gene set by accounting for differences in gene set size and in correlations between gene sets and the expression dataset. Gene-set information on signaling pathways or biological processes was obtained from Kyoto Encyclopedia of Genes and Genomes (KEGG), Gene Ontology (GO), and the Reactome database (36,37,38) and the terms were regarded as significant if the false discovery rate was lower than 0.25, in accordance with the interpretation guideline (35). Specifically, gene sets for Th17 cell differentiation (hsa04659) and the IL-17 signaling pathway (hsa04657) were sourced from the KEGG pathway database. To test for gene enrichment in individual samples, we used a gene-set variation analysis (GSVA), which is a non-parametric, unsupervised method for estimating variation of gene set enrichment through the samples of an expression data set (39). GSVA performs a change in coordinate systems, transforming the data from a gene by sample matrix to a gene-set by sample matrix, thereby allowing the evaluation of pathway enrichment for each sample. To estimate the cellular infiltration into the bulk tissue sample, we used an algorithm called xCell (40), a powerful machine learning framework trained on the profiles of 64 immune and stroma cell datasets, for generating cell type enrichment scores and adjusting them to cell type proportions.

Cell-to-cell communication (CCC) analysis

To investigate potential interactions across different cell types in the renal glomerulus, CCC analysis was performed using CellChat according to a standard pipeline (41). CellChat is a tool that is able to quantitatively infer and analyze intercellular communication networks from single-cell RNA-sequencing data and predicts major signaling inputs and outputs for cells and how those cells and signals coordinate for functions using network analysis and pattern recognition approaches.

IFN-stimulated genes (ISGs) score

The gene sets of ISGs were extracted from the molecular signatures database (35) and an ISG scores were calculated using AUCell package previously described (42). According to the AUC value, gene-expression rankings were built for each sample. The AUC estimates the proportion of genes in the gene set that are highly expressed in each sample.

Construction of protein-protein interaction (PPI) network

To assess the interconnectivity of DEGs in the LN renal glomerulus tissue samples, we constructed a PPI network based on the human interactome database (43,44). In the network, nodes and edges represent genes and functional or physical relationships between them, respectively and includes 243,603 PPIs connecting 16,677 unique proteins.

Data-driven network-based Bayesian inference of drivers (NetBID) analysis for identification of putative drivers

We performed NetBID, version 2 (NetBID2) analysis to identify hidden drivers of LN using transcriptomic data of discovery dataset (45,46). NetBID2 reverse-engineers context-specic interactomes and integrates network activity inferred from large-scale multi-omics data, empowering the identication of hidden drivers using an improved version of SJARACNe (47), an information theory-based algorithm for regulatory network inference. Potential hidden driver genes were prioritized according to the driver activity score, which was expressed as a p-value format.

Network-based diffusion scoring

Genes perform the biological functions cooperatively with other genes, proteins and chemical compounds, and genes in close PPI network proximity to the established disease genes may share biological functions and therefore have a more chance to regulate disease modules (13). To quantify the leverage, we employed label propagation and network diffusion algorithm, which models heat flow from the seed genes along interaction in a PPI network. The diffuStats, an R package provides a collection of graph kernels and diffusion score (48). The regularized Laplacian kernel was employed as the default, given its standard application in biological network analysis. Although the standard Monte Carlo method typically scores nodes based on an empirical p-value computed through permutation, we selected the z-score method. The z-score is a parametric alternative that provides a clearer and more distinct numerical representation of scores. Seed genes were defined based on LN-associated pathways and biological processes, which were compiled from the KEGG, GO, and Reactome databases.

Molecular signature-based drug screening

The Library of Integrated Network-based Cellular Signatures (LINCS) is an NIH Common Fund program that catalogs a diverse and extensive reference library of cell-based perturbation-response signatures, along with novel data analytics tools to improve our understanding of human diseases at the systems level (49,50). This database has been exploited for drug repurposing through drug connectivity mapping (51). We employed metaLINCS R package (52), which take the output and perform a meta-level GSEA by combining them by compound, based on the ranked correlation scores for each experiment. This also perform an additional meta-level GSEA computation step to combine compounds into mode of actions or target genes, to highlight overarching themes in the data.

Statistical analysis

For continuous distributed data, between-group comparisons were performed using the one-way ANOVA, unpaired t-test or Mann-Whitney U test. Categorical or dichotomous variables were compared using the χ² test or Fisher’s exact test. Correlation analysis was performed using Pearson’s method. The Pearson correlation coefficient (γ) was calculated to assess the strength and direction of the linear relationship between two quantitative variables, with values ranging from −1 to +1. All analyses were conducted in R (https://www.r-project.org, version 4.1.0; The R Project for Statistical Computing, Vienna, Austria).

Ethics approval and consent to participate

All raw data used in this study are available in a public, open-access repository and only summary-level datasets without participants were utilized.

RESULTS

Discovery of gene expression-driven subgroups of LN glomeruli

We identified two distinct subgroups of LN glomeruli from the four merged gene expression profiles representing 135 patients with LN using the Monte Carlo reference-based consensus clustering method (Fig. 1A and B). Each subgroup was designated as GC1 (n=100) and GC2 (n=35), respectively. The segregation of these two clusters was verified using t-stochastic neighbor embedding and principal component analysis (Fig. 1C and D). The mean Jaccard scores for GC1 and GC2 were 0.9581 and 0.9032, respectively, indicating high stability within the identified clusters.

Figure 1. Unsupervised consensus clustering. The integrated dataset, constructed from four individual datasets (GSE32591, GSE99339, GSE104948, and GSE127797), was utilized. (A) Consensus heatmap of the two groups, GC1 and GC2. The color scale on the heatmaps ranges from 0 (blue) to 1 (red). (B) Identification of the optimum number of clustering K. Few sample pairs were ambiguous, and a flat middle segment appear on the CDF curve if the correct K was inferred. The PAC score was used to quantify the stability and ambiguity of cluster assignments, calculated based on the CDF of the consensus matrix. A low PAC value, reflecting a flatter middle segment of the CDF, indicates greater robustness in the clustering solution. The optimal number of clusters (K) was selected by prioritizing the solution that yielded the lowest p-value and minimal entropy. (C) PC analysis. (D) t-Distributed SNE analysis.

PC, principal component; SNE, stochastic neighbor embedding.

Characterization of the enriched pathways in LN subtypes

DEGs were obtained by comparing the gene expression profiles between each subtype and 17 NC samples (696 genes in GC1 and 1,731 genes in GC2), and 215 genes were shared between the two subtypes (Supplementary Fig. 3, Supplementary Table 3). We identified the enriched processes based on the DEGs (Supplementary Fig. 4). In GC1, IFN signaling pathways, neutrophil activation, and macrophage activation in the GO–biological processes panel and IFN signaling pathways, TLR signaling pathway, neutrophil degranulation, and signals by class A receptors in the Reactome panel were identified as the main processes. In GC2, B cell (BC) receptor, G protein-coupled receptor (GPCR), and class A/1 (rhodopsin-like) receptor signaling processes were prioritized, although the overall enrichment scores were lower than in GC1. To delineate the distinct immunopathogenic processes specific to LN, we gathered information on LN-associated biological processes and pathogenic pathways (1,2,3,53,54), and compared their enrichment scores using GSEA (Fig. 2A). IFN signaling pathways, Fc receptor-mediated signals, and neutrophil activation-associated processes were highly enriched in GC1, whereas the class A/1 receptor signaling process, signaling by GPCR, and a calcium signaling pathway were found in GC2. In GC2, CCL19, CCL7, and CX3CR1 were the top three leading-edge genes of the class A/1 receptor signaling process by GSEA (Supplementary Fig. 5). Overall, LN-associated immune-inflammatory signatures were much stronger in GC1 than in GC2. Differential activation in LN-associated pathways between the two subtypes was confirmed at a single sample level by GSVA (Supplementary Fig. 6).

Figure 2. Molecular signatures of the LN subtypes. The integrated dataset, constructed from four individual datasets (GSE32591, GSE99339, GSE104948, and GSE127797), was utilized. (A) LN-relevant biological pathways and processes by Gene set enrichment analysis. NES indicated the relative degree of overrepresentation across the gene sets. (B) ISGs scores. An unpaired t-test determined the p-value. (C) Gene expression values of chemokines and chemokine receptors in the class A/1 (rhodopsin-like receptor) gene set. ANOVA determined the p-values. The dashed vertical line was marked at a p-value of 0.01 (10⁻²). (D) Correlation between IL-17 signaling, IFN signaling pathways, and Th17 cell differentiation. Gene sets for Th17 cell differentiation (hsa04659) and the IL-17 signaling pathway (hsa04657) were sourced from the KEGG pathway database. The association was assessed using Pearson correlation analysis. (E) Gene expression values of the investigating LN biomarkers. The dashed vertical line was marked at a p-value of 0.01 (10⁻²). ANOVA determined the p-values.

ISG scores were also significantly higher in GC1 compared with GC2 (Fig. 2B). Class A/1 receptors and ligands were differentially expressed between the two subtypes. In GC2, CCL1, CCL7, CCL13, CCL17, CCL19, CCL22, CCL25, CCL27, CXCL5, CCRL2, CCR3, CCR4, CCR5, CCR6, CCR7, CCR8, CCR9, CXCR1, CXCR2, CXCR3, CXCR5, CXCR6, and CX3CR1 were consistently expressed. In contrast, their expression varied across the samples in GC1 (Fig. 2C).

Urinary IL-17 is a potential biomarker for LN (40,55). Its concentration is correlated with several key clinical indicators, including the severity of renal injury, overall disease activity, and the presence of proteinuria in patients diagnosed with LN. This analysis revealed a statistically significant, although modest, correlation between the IL-17 signaling pathway and IFN signaling pathways (Fig. 2D). For type I IFN, a significant correlation was observed in GC1 (γ=0.5250, p=2.045×10⁻⁸), whereas no significant correlation was found in GC2 (γ=0.1305, p=0.4549).

Urinary biomarkers are considered to be released from inflamed glomerulus or immune cells and have been investigated in terms of relapse or treatment response (8,9,56). ADIPOQ (adiponectin), LCN2 (NGAL), ICAM1, IL6, NAGLU (NAG), TGFB1, IL17A, FOXP3, and TF (transferrin), IL16, and COLEC12 were consistently expressed in GC2, whereas signals of PLG (angiostatin), TNFSF12 (TWEAK), TNFRSF1A, CCL2 (MCP-1), CXCL8, CXCL10 (IP-10), STAT1, TNFRSF11B (OPG), VCAM, CP and CD14 were more pronounced in certain samples of GC1 (Fig. 2E).

Proportion of immune cell subsets and their functional contributions

We then investigated the abundance of immune cell subsets in the two subtypes using xCell, a robust cell-type enrichment method (Fig. 3A). The main immune cell populations that infiltrate the glomeruli include monocytes (MN), macrophages (Mph), CD4+ T cells, CD8+ T cells, memory BCs, and activated dendritic cells (aDCs). In particular, MN and CD4+ memory T cells were more predominant in GC1, whereas CD4+ central memory T cells (Tcm), M2 Mph, and class-switched memory BCs were more prevalent in GC2.

Figure 3. Immune cell subsets and their enriched processes and cell-to-cell communication. (A) Heatmap of the estimated abundance of immune cells using xCell. An unpaired t-test determined the p-value. The integrated dataset, constructed from four individual datasets (GSE32591, GSE99339, GSE104948, and GSE127797), was utilized. (B) An illustrative UMAP plot, where cells form clusters based on similarities and differences in gene expression. Different colors designate clusters, differentiating between cell types and states. The SingleR package annotated clustered cell subsets. (C) DEG-driven pathway enrichment profiles by immune cell subsets. Enrichment was measured using the enriched package. (D) Heatmap of interaction strengths between pairs of cells based on cumulative ligand–receptor interaction scores using CellChat. (E) Bubble plot showing the selected ligand–receptor interactions between BC and Mph/MN. The p-value and communication probability are indicated by circle size and color, respectively. The contribution score was computed from pattern recognition analysis, and a higher contribution score implies the signaling pathway is more enriched in the corresponding cell group.

CD4T, CD4+ T cells; CD8Tcm, CD8+ Tcm; CD8Tem, CD8+ effector memory T cells; Msg, mesangial cells; NKCD56b, NK CD56^bright cells; NKCD56d, NK CD56^dim cells; PC, plasma cells.

Overall, GC1 was characterized by pronounced immune-inflammatory signatures driven by IFN signaling pathways and an abundance of MN and CD4+ memory T cells. Conversely, GC2 exhibited distinct class A/1 receptor signatures and was marked by a higher prevalence of CD4+ Tcm, aDCs, and M2 macrophage populations.

To sort through the cell subset-specific pathological processes, we imported single-cell RNA sequencing data (2,838 cells) obtained from kidney samples of LN patients, which were deposited in the ImmPort repository (accession code SDY997) (16). We analyzed these data using the methods implemented in the R toolkit Seurat. Eleven clusters were identified and annotated for cell types using R package singleR (Fig. 3B), and DEG-driven enriched processes were determined (Fig. 3C). Activation of IFN signaling pathways was observed in MN, Mph, and BC. The class A/1 receptor signaling process was enriched in CD8+ effector memory T cells, CD8+ Tcm, and NK CD56^dim cells.

CCC plays a critical role in maintaining tissue homeostasis. To dissect the complex interactions between immune cell subsets, we inferred intercellular CCC by analyzing the expression of ligand–receptor pairs using CellChat. A strong interaction occurred between BCs and CD4+ T cells, MN or Mph (Fig. 3D, Supplementary Fig. 7). BCs mainly act as the sender, whereas and Mph act as both sender and receiver. The ligand–receptor pairs, including HLAs–CD4, CD22–PTPRC, CD44–LGALS9, and CD45–LGALS9, played significant and contributing roles in mediating interactions between BCs and CD4+ T cells, Mph or MN (Fig. 3E).

Association of LN subtypes with clinical attributes

A transcriptomic dataset, GSE200306, included samples from the 34 LN patients before treatment. It was annotated with clinical information and sequenced for an immunology panel of 606 genes. We arranged the discovery dataset of 135 LN samples by the same genes as the immunology panel. A key gene-driven classifier was constructed using a KNN algorithm to validate the LN subtyping scheme. An independent transcriptome dataset, GSE200306, was put into the classifier. The KNN classifier categorized the samples into 27 GC1 and 7 GC2 subtypes, achieving an AUC of 0.985. All patients underwent a tapering course of prednisolone, starting at 1 mg/kg/day. They were treated with either mycophenolate (MMF) mofetil or intravenous cyclophosphamide (CYC), depending on their histopathological status, adherence, or personal preference.

The proportion of class IV LN was higher in GC1 than in GC2, although it did not reach statistical significance because of the low number of samples (Fig. 4A). Also, in GSE127797, a dataset with the largest sample size (n=41) in the discovery datasets, there was no difference in the proportion of the LN class between GC1 and GC2 (Supplementary Fig. 8). There was no difference in treatment outcome between the subgroups (Fig. 4B). In GC1, a rate of complete response (CR) in the MMF-treated subgroup was higher compared with that of the CYC-treated group (72% vs. 50%, p=0.434) (Fig. 4C). In GC2, all three patients treated with MMF was under CR, but only one of four patients treated with CYC was CR (100% vs. 25%, p=0.2253). Mean serum creatinine levels were slightly higher in GC1 compared with GC2 (0.8851 vs. 0.7142 mg/dl). However, the difference was not statistically significant (p=0.142), and renal insufficiency (defined by serum creatinine levels >1.0 mg/dl) was exclusively identified in five GC1 samples (Fig. 4D). The mean proteinuria was significantly greater in GC1 compared with GC2 (2.8118 vs. 1.7571, p=0.040) (Fig. 4E). There was no significant difference in activity and chronicity index score between GC1 and GC2 (Fig. 4F, Supplementary Fig. 9).

Figure 4. Clinical characteristics of LN molecular subtypes at baseline. A GSE2000306 dataset was assigned clusters by the fitted KNN model and was subsequently used for clinical characterization. (A) Proportion of LN classification. (B) Proportion of treatment response. (C) Frequency of treatment response by drugs. (D) Serum creatinine levels. (E) Urinary protein levels. (F) Activity and chronicity index score. The p-values were calculated by χ² test or unpaired t-test.

CR, complete responder; PR, partial responder; NR, nonresponder.

Molecular signatures of the renal glomerulus in in the NZM2328 lupus-prone mouse model

Previous studies established the NZM2328 lupus-prone mouse strain as a model of human LN, exhibiting severe IC-mediated nephritis and increased early mortality predominantly affecting female mice (14,15). We exploited GSE206806, a transcriptome microarray dataset of renal glomeruli from NZM2328 mice. This dataset included 14 samples with acute GN, five with transitional GN, and six with chronic GN, classified according to stages of GN as previously described (14,57). Compared with 10 non-GN controls, we identified 126,961 and 653 DEGs in the acute, transitional, and chronic stages of GN, respectively. In DEG-driven pathway enrichment analysis, the class A/1 receptor and TNF signaling pathways emerged as the most prominent processes in the acute GN stage. In contrast, major immuno-inflammatory processes, including IFN signaling pathways, were markedly activated in the transitional GN stage (Supplementary Figs. 10A and 11). ISG scores were also significantly higher in transitional GN than in acute and chronic GN (Supplementary Fig. 10B). MN and CD4+ memory T cells predominated in the transitional GN stage, whereas M2 Mph were more prevalent in the acute GN stage (Supplementary Fig. 10C).

Identification of hidden drivers in the disease module

A gene or molecule participates in specific biological or pathological processes through intimate interactions with neighboring components, forming a network module. A disease module comprises a collection of network components that collectively contribute to cellular function, and the disruption of this module is associated with a specific disease phenotype (13). Hidden driver genes or molecules operating the disease module are probably druggable, making them ideal therapeutic targets (58). We applied the NetBID algorithm, a data-driven network-based Bayesian inference of drivers, by integrating data from the transcriptomics of each LN subtype (45,46). Significant drivers involving LN-associated pathways were sorted from the top 50 driver by driver activity at each cluster. In GC1, 15 drivers were involved in neutrophil degranulation, followed by IFN signaling pathways (14 drivers; Fig. 5A). PIK3R2, also known as p85, was identified as the seventh highest priority driver and was associated with the broadest range of LN-related pathways (Fig. 5A, Supplementary Fig. 12A). PIK3R2 is the regulatory subunit binding to three catalytic subunits of p110 (PIK3CA, PIK3CB, and PIK3CD) (59). In GC2, 14 and 11 drivers, including CXCR3 and CXCR6, were involved in the GPCR and the class A/1 receptor signaling process, respectively (Supplementary Figs. 12A and 13A). MAPK13, one of the four p38 MAPKs, covered the 11 LN-associated pathways but did not participate in the class A/1 receptor signaling process (Supplementary Fig. 13A).

Figure 5. Identification of key driver genes and their perturbing activity in disease module network of GC1. (A) NetBID analysis for the identification of putative drivers in GC1. Driver candidates involved in LN-associated pathways and their activity score. Whether each driver was an element of the LN-associated pathways is marked in the upper panel. A red fill in the heatmap indicates that the gene product on the x-axis is a member of the corresponding pathway shown on the y-axis. Driver activity of genes was calculated using a z-scaled method in the lower panel. (B) Molecular targets and diffusion scores of investigational and candidate drugs of interest for LN. The z-scaled Monte-Carlo method installed in the R package diffuStats was used to compute the diffusion score of the nodes based on the genes across the pathway network of interest. (C) Kernel-based diffusion scores for the selected chemokines/chemokine receptors and PI3K subunits. (D) Meta-level analysis of LINCS L1000 drug signatures using stratified connectivity mapping for GC1. The bar plot shows the most enriched compound MoAs. On the vertical axis, the GSEA NES of the MoA class is plotted. A negative NES indicates that a compound could counteract a gene expression profile.

SIG, signaling factor; TF, transcription factor; MoA, mode of action.

To estimate the perturbation effect of each molecule and investigational drugs on LN in the disease module, we computed the diffusion scores of the genes using a kernel-based network diffusion algorithm (Fig. 5B) (48). JAK inhibitors such as baricitinib and brepocitinib highly scored for perturbing a broad scope of LN-associated processes, including IFNs and JAK-STAT signaling pathways. Voclosporin had high marks for both B and T cell receptor signaling pathways. In contrast, several investigational drugs only displayed favorable diffusion scores for a limited number of LN-associated pathways without a broad impact on the spectrum of pathological processes.

PIK3R2 was the key priority driver in GC1, and its binding subunits, PIK3CA, PIK3CB, and PIK3CD, were the major targets of PI3K inhibitors. CCL7, CCL19, CX3CR1, CXCR3, and CXCR6 were the leading-edge genes or key drivers implicated in the class A/1 receptor signaling process in GC2. All PI3K catalytic subunits had a significant diffusion score for several LN-associated pathways except IFNs signaling pathways and complement cascade, and CCL7, CCL19, CX3CR1, CXCR3, and CXCR6 showed strong leverage over the class A/1 receptor, GPCR, and chemokine signaling pathways by diffusion score (Fig. 5C).

To further validate the therapeutic effect of potential drugs or drug targets, we employed a novel approach to evaluate the correlation between LN gene expression profiles and perturbation signatures from the Library of Integrated Network-Based Cellular Signatures (LINCS) L1000 database using the metaLINCS package (49). A negative NES indicates that a compound could counteract a gene expression profile. In GC1, the mTOR inhibitor, PI3K inhibitor, and MEK inhibitor were the significant drugs that could counteract the GC1 phenotype (adjusted p<0.05; Fig. 5D), and MTOR, PIK3CD and MAP2K1 were the most enriched target genes with a significant perturbation effect (adjusted p<0.05; Supplementary Fig. 14). Protein synthesis inhibitor was the unique compound with signatures negatively correlated with GC2 gene signatures (adjusted p<0.05; Supplementary Fig. 13B). Gene signatures by CC chemokine receptor antagonist did not have a correlative association with GC2 signatures.

DISCUSSION

We obtained glomerular transcriptomic datasets of LN from publicly repositories and built the comprehensive gene expression profiles of the renal glomerulus from patients with LN and identified molecular patterns that result in subsets of similar characteristics using an advanced unsupervised clustering method (graphical abstract at Supplementary Fig. 15). We delineated two molecular subtypes of LN and characterized them for predominant molecular and cellular signatures. The two subtypes were further examined regarding clinical attributes, and particularly, GC1 was associated with proteinuria and renal insufficiency, the important predictors for treatment response or renal outcome (60,61,62,63). We sorted out the hidden pivotal drivers of the subtypes using the network-based Bayesian inference method and identified molecules that have high therapeutic potentials for each LN subtype.

Although histological findings of kidney biopsy are essential for diagnosing LN and assessing active and chronic inflammation of the pathological phenotypes, they were found to be inconsistent with early and long-term clinical outcomes (5,6). About one-third of the patients successfully treated in their initial episode later had relapse of nephritis (64,65). Furthermore, one-third of patients accrued progressive renal injury and developed renal insufficiency despite an apparent response to standard-of-care therapy (65). Current criteria to assess treatment response are based solely on clinical indices (66). The predictive ability of baseline clinical variables for the response to treatment, measured by the area under the ROC curves, ranged from 0.5 to 0.6, which only improved to approximately 0.7 when combined with urinary biomarkers (60,63). These observations emphasize the need for a deeper understanding and stratification of LN in terms of mechanical molecular signatures. Refined molecular subtyping of malignancies such as prostate, bladder, breast, and BC lymphomas has provided a framework for illustrating the functional mechanisms, assessing unrecognized heterogeneity, and creating a novel opportunity for precision treatment (67,68,69,70).

We identified two molecular subtypes of LN glomerulus. Subtype GC1 is an active form of LN, marked by significant enrichment of IFN signaling pathways, Fc receptor-mediated signaling, and processes associated with neutrophil activation. Type I IFN gene signatures are detectable in the kidney of patients with LN, and they become even more pronounced in active LN and are associated with both active kidney disease and treatment response (11,71,72,73). The important predictors for treatment response or renal outcome such as proteinuria and renal insufficiency (60,61,62,63) were more significant in GC1. A type I IFN-driven gene expression signature is a hallmark of SLE neutrophil (74,75) and it was reported that blood neutrophil signature was strongly associated with LN, particularly the severity of LN (76). Monocyte infiltration and upregulated CCL2 (also known as monocyte chemotactic protein-1) in LN was also associated with disease activity and tissue damage (77). LN-associated MN-derived macrophage subsets showed marked induction of FcγR response genes, avidly internalised circulating ICs, and presented IC-opsonised Ag (78). This is in line with the result that MN showed enriched molecular signatures regarding NET formation, FcγR phagocytosis, and IFN signaling pathways (Fig. 3C).

GC2 is an intriguing subset of LN characterized by class A/1 receptors and B-cell receptor signalling pathways. M2 Mph, CD4+ Tcm, and class-switched memory BC were the predominant immune cell subsets infiltrated into the glomerulus and the class A/1 receptor signaling pathway of GC2 was led mainly by CCL19, CCL7, CX3CR1, and CXCR6 (Supplementary Fig. 5). Chemokine expression precedes inflammatory cell infiltration during the initiation of murine LN (79). Activated kidney Mph are markers of disease onset and disease remission in LN (80,81). CX3CR1 is highly expressed in myeloid phagocytes, which are attracted to the renal endothelium and tubular epithelium chiefly producing CX3CL1 (82). Inhibition of CX3CL1/CX3CR1 interaction or macrophage depletion significantly ameliorated renal injury (83,84). Monocyte mobilisation into kidney and macrophage activation by chemokines and neighbouring immune cells might play a pivotal role in initiation of LN (79).

Two novel molecular subtypes of LN were applied to an independent dataset (n=34) with clinical information. The baseline proportions of ISN/RPS classification and the response to treatment between the two subtypes did not differ. This result needs reappraisal in a large-sized, long-term prospective study because the dataset was small and lacking in class II and V samples, and the treatment modality was not randomized. However, it seems certain that GC1 is a more inflammatory subtype, resulting in critical renal damage because samples with renal insufficiency were observed exclusively in GC1, and proteinuria was also more severe. This is in good agreement with the pattern of molecular signatures. Compared with the nephritis stage of the NZM2328 lupus-prone mouse, it is speculated that GC1 corresponds to transitional GN, the most active inflammatory stage, and GC2 is similar to acute GN, the early stage of development. However, it needs to be clarified whether they represent one of the stages of evolving disease or different disease subtypes.

DEGs often include coexpressed bystander or neighbor genes that may not meaningfully contribute to disease pathogenesis, and the analysis of DEGs at a single time point may not capture critical regulators with short half-lives or regulators that act by changing their functional activity rather than their expression levels. A comprehensive data-driven, network-based identification of the hidden drivers successfully operated in cancers or immune cell analyses (45,46). We identified PIK3R2 in GC1, which was linked to the largest LN-associated pathways among the significant hidden drivers. The product of PIK3R2, p85, directly interacts with p110α (PIK3CA), p110β (PIK3CB), and p110δ (PIK3CD), which were the real major targets of PI3K inhibitors (59) and showed promising potential to perturb the main LN-associated pathways effectively. LINCS is an NIH Common Fund program that catalogs how human cells respond globally to chemical, genetic, and disease perturbations (49). In evaluating the correlation between LN gene expression profiles and perturbation signatures from the LINCS L1000 database, the PI3K inhibitor showed a significant potential to reverse the gene expression profile of the GC1 subtype along with the mTOR inhibitor and MEK inhibitor. In a series of well-designed preclinical experiments using mouse models prone to lupus, PI3K inhibition improved LN by reducing cellular infiltration into the kidney, proteinuria, and tissue damage, finally leading to a prolonged life span (85,86,87,88,89). JAK inhibitors also showed a good diffusion score for many LN-associated pathways, and treatment with JAK inhibitors was reported to result in decreased proteinuria and amelioration of nephritis (90). However, JAK inhibitors failed to significantly counteract gene expression of LN GC1 in LINCS L1000 analysis. The true clinical efficacy of the drugs should be evaluated in clinical trials, and the clinical efficacy and resolution at the histological or molecular level should be separated because one-third of the patients who achieved complete clinical response after induction had persistently high histological activity (5,6). Clinical application of PI3K inhibitors should be titrated given their adverse effects and mechanistic differences in cancer and autoimmune disease (59,91).

Increased PI3Kδ but not PI3Kγ activity was observed in SLE patients, which correlated with resistance of activated and memory T cells to activation-induced cell death and increased memory T cells (92). PI3Kδ is pivotal in mediating BC receptor (BCR)-mediated response including BC proliferation, differentiation, homing, Ag presentation, and Ab and cytokine production (93). Sustained PI3Kδ activity is required for IFN-γ production and also involved in the other cytokines in T cells (94), while many studies showed that strong PI3K/AKT signaling activity negatively affects regulatory T cells (95). Balanced control of PI3K might be beneficial for alleviating the glomerular inflammation and maintaining the immune homeostasis in SLE. Cellular interaction between cells and Mph/MN played a significant role in maintaining the LN pathology. HLAs-CD4, CD22-PTPRC, CD44-LGALS9, and CD45-LGALS9 were the leading ligand–receptor pairs. It is worth noting that engagement of CD22, CD44, CD45, PTPRC, or LGALS9 transduces signals by recruiting or activating the PI3K signaling pathway (96,97,98,99). PIK3R2 and PIK3Cs are mediators regulating Ag receptor signaling (BCR, CD19, and TLRs), cytokine signaling (BAFF, IL-4, and TNFα), and chemokine signaling (CXCL12, CXCL13, and CCL19/21) (100). These signaling pathways were also enriched processes in GC1 (Fig. 5A). PI3K inhibitors may play a role in treatment of LN, similar to JAK inhibitors in rheumatoid arthritis although they do not block the IFN signaling pathway.

Although many promising drugs against LN are undergoing investigation or clinical trials, they do not adequately address the molecular alteration of LN compared with PI3K inhibitors. Most drugs, except for C5 inhibitor and prednisolone, did not significantly perturb the class A/1 receptor signaling pathway. In the GC2 subtype, the class A/1 receptor signaling pathway was the most prominent process, but it does not appear to take the lead in the molecular signatures of GC2 because CC chemokine receptor antagonist did not have a counteracting effect on gene expression of GC2. Therefore, targeting BCs or Mph, two major infiltrating immune cells, would be more reasonable in GC2.

There are several limitations to be addressed. First, some genes, albeit minor, were inevitably lost during the integration of multiple datasets from different sequencing platforms, and the batch effect correction might not be ideal. Second, because of a lack of information, we did not fully determine the correlation between molecular subtypes and clinical factors such as complement levels, and anti-dsDNA titer. Third, some gene signatures of specific cell subset that have significant roles in the development of LN might have not been detected because the gene expression profiles were obtained in the bulk kidney tissues. Fourth, analysis of tubulointerstitial compartments was not included because only 66 tubulointerstitial samples were paired with glomerular samples. Fifth, small sample size prevented us from securing statistical power on the treatment response by two molecular subtypes.

Therapeutic advances in SLE and LN lag behind those of other autoimmune rheumatic diseases, such as rheumatoid arthritis and ankylosing spondylitis. This is likely due to the inherent complexity and heterogeneity of the disease phenotypes and pathogenesis. In this study, we identified the two molecular subtypes of LN. We dissected their cellular and pathogenic properties and clinical features, thereby providing novel information on divergent and shared features of LN. Identifying key cellular components and driver genes could guide the development of rational agents and lead to a subtyping schema to prioritize treatment options. In particular, PIK3R2 was identified as a hidden regulator in the disease module, and inhibition against PI3K subunits showed promising efficacy in counteracting altered molecular signatures in an active state of LN. Further investigation is required to comprehensively characterize the molecular characteristics of the subtypes and assess their treatment response to specific regimens based on molecular signatures and clinical indices.

Glossary

aDC

activated dendritic cell

AUC

area under the curve

BC

B cell

BCR

BC receptor

CCC

cell-to-cell communication

CDF

cumulative distribution function

CR

complete response

CYC

cyclophosphamide

DEG

differentially expressed gene

GN

glomerulonephritis

GO

Gene Ontology

GPCR

G protein-coupled receptor

GSEA

gene-set enrichment analysis

GSVA

gene-set variation analysis

IC

immune-complex

ISG

IFN-stimulated genes

KEGG

Kyoto Encyclopedia of Genes and Genomes

KNN

K-nearest neighbor

LINCS

Library of Integrated Network-based Cellular Signatures

LN

lupus nephritis

M3C

Monte Carlo Reference-based Consensus Clustering

MMF

mycophenolate

MN

monocytes

NC

normal control

NES

normalized enrichment score

NetBID

data-driven network-based Bayesian inference of drivers

NetBID2

NetBID, version 2

PAC

proportion of ambiguous clustering

PPI

protein-protein interaction

ROC

receiver operating characteristic

SLE

systemic lupus erythematosus

Tcm

central memory T cells

SUPPLEMENTARY MATERIALS

Supplementary Table 1

A summary of the datasets used in the study

Supplementary Table 2

Summary of the demographic and clinical characteristics of the datasets used in the study

Supplementary Table 3

Differentially expressed genes in glomerular molecular subgroups

Supplementary Figure 1

Overview of data processing and analysis steps.

Supplementary Figure 2

Principal component analysis on the merged datasets of renal glomerulus transcriptomics before (A) and after (B) normalization and batch correction.

Supplementary Figure 3

Venn diagram of DEGs.

Supplementary Figure 4

Functional enrichment profiles in terms of GO–biological process (BP) and reactome pathways by the DEGs of GC1 (A) and GC2 (B).

Supplementary Figure 5

Gene-set enrichment analysis (GSEA) of the class A/1 (rhodopsin-like receptor) in GC2.

Supplementary Figure 6

Heatmap of pathway enrichment for each sample using GSVA.

Supplementary Figure 7

Cell-to-cell communications.

Supplementary Figure 8

Proportion of LN classification in the dataset GSE127797 by molecular subtypes.

Supplementary Figure 9

Histologic index scores.

Supplementary Figure 10

Molecular signatures of the NZM2328 lupus-prone mouse model.

Supplementary Figure 11

Functional enrichment profiles in terms of GO–biological process (BP) and reactome pathways by the DEGs of acute, transitional and chronic GN stages.

Supplementary Figure 12

Identification of hidden drivers inferred by NetBID2.

Supplementary Figure 13

Identification of key driver genes and their perturbing activity in disease module network of GC2.

Supplementary Figure 14

Barplot displaying the most enriched target genes in GC1.

Supplementary Figure 15

A graphical abstract.