Integrative transcriptomic analysis reveals KLHL2 downregulation as a potential driver of microglial inflammation in Alzheimer’s disease via the WNK3/JNK/c-Jun axis Authors: Yao Yang, Conghui Dai, Zongtang Xu, Linpei Zhuo, Kun Xiang, Xiaoou Lin, Jing Sun, Jiaming Liu, Qiuling Tong Published: 2026 ABSTRACT Alzheimer’s disease (AD) is a progressive neurodegenerative disorder characterized by prominent neuroinflammation. Microglia, the resident immune cells of the central nervous system, play a key role in AD-associated neuroinflammation, yet the molecular mechanisms regulating their inflammatory activation remain incompletely understood. In this study, we integrated human single-nucleus RNA sequencing with computational analyses and experimental validation to identify potential regulators of inflammatory microglial states in AD. Microglial subpopulations were characterized, and pseudotime analysis was performed to infer transcriptional state transitions across neuropathological stages. High-dimensional weighted gene co-expression network analysis (hdWGCNA) and machine learning were used to prioritize candidate genes for functional validation. KLHL2 was identified as a candidate regulator, and its expression progressively decreased along the inferred pseudotime trajectory, with concordant decreases in Klhl2 mRNA and KLHL2 protein levels observed in APP/PS1 mice and Aβ42-stimulated BV2 cells. In contrast, WNK3 expression level increased without corresponding changes in Wnk3 mRNA. In BV2 cells, Klhl2 overexpression reduced WNK3 expression level, JNK and c-Jun phosphorylation, and pro-inflammatory mediator expression, whereas Klhl2 knockdown produced the opposite effects. Collectively, these findings support KLHL2 as a negative regulator of microglial inflammatory activation and suggest the involvement of WNK3/JNK/c-Jun signaling in this process. Introduction Alzheimer’s disease (AD) is the most common neurodegenerative disease and the leading cause of dementia worldwide. Its major pathological features include amyloid-β (Aβ) deposition, aggregation of hyperphosphorylated tau, and chronic neuroinflammation [1]. For decades, AD research and therapeutic development have largely focused on Aβ and tau pathology. However, the modest clinical benefits achieved by therapies targeting Aβ or tau suggest that additional pathological mechanisms contribute to AD progression [2,3]. Accumulating evidence indicates that neuroinflammation is not merely a secondary response to cellular injury but an active contributor to AD pathogenesis and progression [[4], [5], [6]]. Therefore, elucidating the molecular mechanisms underlying AD-associated neuroinflammation may deepen our understanding of AD pathogenesis and provide a rationale for identifying novel therapeutic targets. Microglia are resident innate immune myeloid cells of the central nervous system (CNS) and play pivotal roles in the initiation, amplification, and resolution of neuroinflammatory responses [4]. They are highly sensitive to changes in the local microenvironment and exhibit substantial functional plasticity. Appropriately regulated microglial responses can facilitate the clearance of pathological material and restrain excessive inflammatory signaling, whereas persistent or dysregulated responses can promote the release of inflammatory mediators and amplify neuroinflammatory cascades [7,8]. In AD, sustained pathological insults, including abnormal protein aggregation, neuronal injury, and extracellular danger signals, can drive marked remodeling of microglial functional states and associated gene-expression programs [4,8,9]. This functional plasticity underscores the complexity of microglial involvement in AD-associated neuroinflammation. These diverse microglial states depend on tightly regulated intracellular signaling pathways, whose activity is partly determined by the levels and stability of key signaling proteins [10,11]. Kelch-like protein 2 (KLHL2), also known as Mayven, is a member of the BTB-BACK-Kelch protein family and functions as a substrate adaptor for CUL3-RING E3 ubiquitin ligase complexes [12,13]. Previous studies have shown that KLHL2 regulates WNK3 level [13,14]. However, whether KLHL2 modulates WNK3 level and downstream inflammatory signaling in microglia under AD-related pathological conditions remains unclear. In this study, we aimed to systematically investigate the potential role of the KLHL2 gene in the pathogenesis of AD by integrating bioinformatics analyses and experimental validation. We first leveraged single-nucleus RNA sequencing data of AD samples obtained from public databases to dissect the heterogeneity of microglial subpopulations and to characterize the dynamic transcriptional trajectories associated with inflammatory activation. By combining high-dimensional weighted gene co-expression network analysis (hdWGCNA) with a random forest (RF) algorithm, we sought to identify hub genes associated with inflammatory microglial state transitions. Finally, we performed in vivo expression validation and in vitro gain- and loss-of-function experiments to investigate the role of the prioritized candidate gene in Aβ42-associated microglial inflammatory activation. Materials and methods Data acquisition Single-nucleus RNA sequencing (snRNA-seq) data from the inferior temporal gyrus (ITG) in Alzheimer’s disease (AD) were obtained from the Gene Expression Omnibus (GEO) database under accession number GSE268599, which was generated using the Illumina NovaSeq 6000 platform (Homo sapiens) [15]. The dataset includes samples from individuals representing different Braak stages of AD pathology. Bulk transcriptome datasets were retrieved from the GEO for machine learning model construction and external validation. The dataset GSE36980 [16] was used as the training cohort, while GSE109887 [17], GSE118553 [18], GSE125583 [19], and GSE173954 [20] were employed as external validation cohorts. Single-nucleus RNA sequencing data processing Raw snRNA-seq data were processed using the Seurat package (version 5.3.0) in R [21]. Nuclei with <200 or >8,000 detected genes, UMI counts <1,000, mitochondrial transcript content >20%, or hemoglobin transcript content >3% were excluded. Data were normalized using LogNormalize with a scale factor of 10,000, and the top 3,000 highly variable genes were identified using FindVariableFeatures with the “vst” method. After scaling and principal component analysis (PCA), batch effects were corrected using Harmony (version 1.2.3) with sample identity (orig.ident) as the grouping variable [22]. Clustering was performed using FindNeighbors and FindClusters based on Harmony embeddings (dims = 1–10, resolution = 0.05), and UMAP was used for visualization. Detailed quality control statistics for each sample are provided in Fig. S1. Identification of microglial subpopulations To further characterize microglial heterogeneity, cells annotated as microglia were extracted and subjected to a second round of analysis following the same preprocessing workflow described above. For reclustering, we applied the same Harmony embeddings as in the initial clustering but used a higher resolution of 0.8 to identify finer microglial subpopulations. After reclustering, the FindAllMarkers function was used to identify differentially expressed genes (DEGs) for each microglial cluster. Microglial subclusters were annotated according to cluster-specific marker genes and previously reported microglial transcriptional states [9]. CytoTRACE analysis To estimate the developmental potential of inflammatory microglia, CytoTRACE 2 (version 1.1.0) was applied [23]. Higher CytoTRACE 2 scores indicate less differentiated cellular states. Scores were calculated for individual inflammatory microglial cells and compared across neuropathological stages. Pseudotime trajectory analysis Pseudotime trajectory analysis of inflammatory microglia was performed using Monocle2 (version 2.37.0) [24]. Highly variable genes identified by Seurat were used as ordering genes, with mean_expression ≥ 0.05 and dispersion_empirical ≥ 1.0 × dispersion_fit. Dimensionality reduction was performed using DDRTree with max_components = 2, num_dim = 6, and ncenter = 30, followed by cell ordering using the orderCells function. The trajectory state enriched in healthy control cells was designated as the root state. Genes significantly associated with pseudotime were identified using the differentialGeneTest function. High-dimensional weighted gene co-expression network analysis High-dimensional weighted gene co-expression network analysis (hdWGCNA) was performed to identify gene co-expression modules associated with inflammatory microglial states (version 0.4.07) [25]. The soft-thresholding power was selected according to scale-free topology criteria, followed by hierarchical clustering and module eigengene calculation. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were performed using the clusterProfiler package (version 4.17.0) [26] for functional annotation. Modules associated with inflammatory microglial changes were further analyzed to identify candidate hub genes. Machine learning model construction Machine-learning models were constructed using genes from the key hdWGCNA modules. Twelve machine-learning algorithms were systematically combined to generate 113 candidate models using 10-fold cross-validation in the training dataset GSE36980. These algorithms included least absolute shrinkage and selection operator (Lasso), Ridge regression, stepwise generalized linear model (StepGLM), extreme gradient boosting (XGBoost), random forest (RF), Elastic Net (Enet), partial least squares regression for generalized linear models (plsRglm), generalized boosted regression modeling (GBM), Naive Bayes, linear discriminant analysis (LDA), generalized linear model boosting (glmBoost), and support vector machine (SVM). Model performance was assessed by the area under the receiver operating characteristic curve (AUC), and the optimal model was externally validated in GSE109887, GSE118553, GSE125583, and GSE173954. Calibration curves and decision curve analysis (DCA) were used to further evaluate model performance. Animals and tissue collection Twelve-month-old APP/PS1 transgenic mice and age-matched C57BL/6J wild-type (WT) male mice were used as the AD and control groups, respectively. The adult mice were obtained from the Experimental Animal Center of Wenzhou Medical University and housed under a 12-h light/dark cycle with ad libitum access to food and water. Brain tissues were collected for subsequent analyses. Cell culture BV2 microglial cells were obtained from the Shanghai Institutes for Biological Sciences (China). BV2 cells were cultured in DMEM (10-013-CV, Corning, USA) supplemented with 10% (v/v) FBS (F8318, Sigma, Australia) at 37 °C in a humidified incubator containing 5% CO2. Klhl2 overexpression in BV2 microglial cells To generate stable Klhl2-overexpressing BV2 cells, cells were transduced at a multiplicity of infection (MOI) of 150 with recombinant lentiviral particles carrying a construct encoding FLAG-tagged KLHL2 or the corresponding empty control vector (viral titer, 1 × 108 TU/mL; GenePharma, China) in the presence of 10 μg/mL polybrene. After 24 h of lentiviral transduction, the cells were passaged and cultured in fresh DMEM supplemented with 10% (v/v) FBS and 5 μg/mL puromycin for 24 h. The surviving cells were subsequently maintained under selection in medium containing 1 μg/mL puromycin. Expression of FLAG-tagged KLHL2 was confirmed by Western blotting using an anti-FLAG antibody. The cells were then treated with 5 μM Aβ42 (A834109, Macklin, China) for 24 h before subsequent analyses [27]. Klhl2 knockdown in BV2 microglial cells BV2 microglial cells were transfected with small interfering RNA targeting Klhl2 (si-Klhl2) or the corresponding negative control siRNA (si-NC) using a transfection reagent (GenePharma, China) according to the manufacturer’s instructions. Klhl2 knockdown efficiency was confirmed by Western blotting. Cells were maintained for 48 h after transfection and then treated with 5 μM Aβ42 for an additional 24 h before subsequent analyses. Quantitative real-time PCR Quantitative real-time polymerase chain reaction (qPCR) was performed using the ChamQ Universal SYBR qPCR Master Mix kit (Q711-03, Vazyme, China) in accordance with the manufacturer’s standard protocols. The relative expression fold of target genes against the reference gene was calculated via the 2−ΔΔCt comparative method, and primer sequences are listed in Table S1. Western blotting Mouse brain tissues (n = 3 per genotype) and cultured cells (three independent preparations) were lysed in RIPA buffer (P0013B, Beyotime, China) supplemented with PMSF (ST506, Beyotime, China) and phosphatase inhibitors (P1260, Solarbio, China). Lysates were centrifuged at 12,000 × g for 15 min at 4 °C, and protein concentrations were determined using a BCA assay (P0011, Beyotime, China). Protein samples were mixed with loading buffer (AR0131, BOSTER, China), boiled at 95 °C for 5 min, separated by SDS-PAGE, and transferred to PVDF membranes. After blocking with 5% BSA for 1 h at room temperature, membranes were incubated overnight at 4 °C with primary antibodies against KLHL2 (1: 1,000, MG625073S, Abmart, China), WNK3 (1: 1,000, CSB-PA832503, CUSABIO, China), p-JNK (1: 1,000, BS4322, Bioworld, USA), JNK (1: 1,000, BS45380, Bioworld, USA), p-c-Jun (1: 1,000, BS43042, Bioworld, USA), c-Jun (1: 1,000, BS79651, Bioworld, USA), IL-1β (1: 1,000, BS3506, Bioworld, USA), TNF-α (1: 500, BS1857, Bioworld, USA), IL-6 (1: 500, BS6419, Bioworld, USA), iNOS (1: 500, T55993, Abmart, China), β-actin (1: 2,000, AC038, ABclonal, China), and FLAG (1: 2,000, AE063, ABclonal, China). Membranes were then incubated with HRP-conjugated secondary antibody (1: 1,000, BS22357, Bioworld, China) for 1 h at room temperature. Protein bands were visualized and quantified by densitometric analysis. All experiments were independently repeated at least three times. Statistical analysis Bioinformatics and machine learning analyses were conducted using R software (version 4.5.0). Where multiple testing was involved in the bioinformatics analyses, P values were adjusted using the default multiple-testing correction procedures implemented in the corresponding R packages, and adjusted P values (or q values, where applicable) were reported. Experimental data were statistically analyzed using SPSS (version 26) (Chicago, USA) and visualized with GraphPad Prism (version 10.1.1) (La Jolla, CA, USA). All data were first tested for normality. For multi-group comparisons, normally distributed data were analyzed using one-way analysis of variance (ANOVA) followed by Tukey’s or Dunnett’s post hoc test, whereas non-normally distributed data were assessed using the Kruskal–Wallis test. For two-group comparisons, an unpaired two-tailed Student’s t-test was used for normally distributed data, and the Mann-Whitney U test was used for non-normally distributed data. Data are presented as mean ± standard error of the mean (SEM). A P value < 0.05 was considered statistically significant. Results Single-nucleus RNA sequencing reveals microglial heterogeneity in AD Microglia accomplish their diverse roles in the central nervous system by adopting distinct transcriptional and functional states [28]. To systematically characterize the microglial heterogeneity across different neuropathological stages of AD, we analyzed the snRNA-seq dataset GSE268599, which includes 1 healthy control (HC) sample (Braak stage 0), 4 early-stage AD (EAD) samples (Braak stage I–II), 8 middle-stage AD (MAD) samples (Braak stage III–IV), and 18 late-stage AD (LAD) samples (Braak stage V–VI). These samples were strictly stratified according to the Braak staging system, which reflects the neuropathological progression of AD. As illustrated in Fig. 1A, the analytical workflow included quality control, unsupervised clustering, and cell-type annotation based on established marker genes. Specifically, neurons were identified by GRIA1 and GRIA2, oligodendrocytes by PLP1 and MBP, astrocytes by ADGRV1 and AQP4, endothelial cells by ABCB1 and CLDN5, and microglia by CSF1R and CD74 (Fig. 1B and C) [29,30]. The resulting annotated dataset comprised approximately 20,000–60,000 high-quality nuclei per sample and was used for subsequent analyses (Fig. 1D).Fig. 1Single-nucleus RNA sequencing reveals microglial heterogeneity in AD. (A) Schematic overview of the analytical workflow. Single-nucleus RNA sequencing (snRNA-seq) data from dataset GSE268599 were analyzed across four groups: healthy controls (HC), early-stage Alzheimer’s disease (EAD), middle-stage Alzheimer’s disease (MAD), and late-stage Alzheimer’s disease (LAD). (B) Uniform Manifold Approximation and Projection (UMAP) visualization of the snRNA-seq data showing the major brain cell types. Each dot represents a nucleus and is colored according to its annotated cell type. (C) Violin plots showing the expression of representative marker genes across major cell types. (D) Stacked bar plots showing the cell numbers (left) and relative proportions (right) of major cell types across individual samples. (E) UMAP visualization of reclustered microglia showing seven transcriptionally distinct microglial subpopulations. (F) Dot plot showing the expression of representative marker genes across microglial subpopulations. (G) Pie and doughnut charts showing the cell numbers (top) and relative proportions (bottom) of the seven microglial subpopulations. (H) Transcriptional characterization of microglial subpopulations. The left panel shows a heatmap of scaled expression of cluster-specific marker genes, and the right panel shows the enrichment of previously reported microglial transcriptional signatures across the identified subpopulations. Bubble color represents the negative base-10 logarithm of the Benjamini–Hochberg-adjusted P value (Padj), and bubble size represents the normalized enrichment score (NES). Highlighted bubbles indicate statistically significant enrichment (Padj < 0.05). Single-nucleus RNA sequencing reveals microglial heterogeneity in AD. (A) Schematic overview of the analytical workflow. Single-nucleus RNA sequencing (snRNA-seq) data from dataset GSE268599 were analyzed across four groups: healthy controls (HC), early-stage Alzheimer’s disease (EAD), middle-stage Alzheimer’s disease (MAD), and late-stage Alzheimer’s disease (LAD). (B) Uniform Manifold Approximation and Projection (UMAP) visualization of the snRNA-seq data showing the major brain cell types. Each dot represents a nucleus and is colored according to its annotated cell type. (C) Violin plots showing the expression of representative marker genes across major cell types. (D) Stacked bar plots showing the cell numbers (left) and relative proportions (right) of major cell types across individual samples. (E) UMAP visualization of reclustered microglia showing seven transcriptionally distinct microglial subpopulations. (F) Dot plot showing the expression of representative marker genes across microglial subpopulations. (G) Pie and doughnut charts showing the cell numbers (top) and relative proportions (bottom) of the seven microglial subpopulations. (H) Transcriptional characterization of microglial subpopulations. The left panel shows a heatmap of scaled expression of cluster-specific marker genes, and the right panel shows the enrichment of previously reported microglial transcriptional signatures across the identified subpopulations. Bubble color represents the negative base-10 logarithm of the Benjamini–Hochberg-adjusted P value (Padj), and bubble size represents the normalized enrichment score (NES). Highlighted bubbles indicate statistically significant enrichment (Padj < 0.05). To further characterize microglial heterogeneity, we extracted and reclustered the annotated microglial population, identifying seven transcriptionally distinct microglial subpopulations (Fig. 1E). Differential expression analysis revealed cluster-specific gene signatures for each subpopulation (Table S2). Based on cluster-specific marker genes and previously reported microglial transcriptional states, including the human microglial atlas described by Sun et al. [9], the seven clusters were annotated according to their representative marker genes and inferred functional characteristics: Mic0 (homeostatic, markers: P2RY12, CX3CR1), Mic1 (ribosome biogenesis, markers: FTL, FTH1), Mic2 (phagocytic, markers: CD163, F13A1), Mic3 (inflammatory, markers: CPEB4, TMEM163), Mic4 (lipid processing, markers: MYO1E, PTPRG), Mic5 (stress signature, markers: HSP90AA1, HSPH1), and Mic6 (cycling, markers: EZH2, BRIP1) (Fig. 1E and F). The absolute numbers and relative proportions of these microglial subpopulations are shown in Fig. 1G. Following the analytical pipeline described by Martins-Ferreira et al. [31], we characterized and validated the transcriptional identities of microglial subpopulations. We compared each cluster with the remaining clusters by differential expression analysis, visualized representative cluster-specific marker genes, and assessed the enrichment of established microglial gene signatures among these markers (Fig. 1H; Table S3). The reference gene sets included canonical human microglial signatures identified by comparing purified microglia with bulk cortical tissue [32,33] and the original disease-associated microglia (DAM) signature [34]. We also examined gene sets representing distinct DAM-related populations, including bona fide DAM, youth-associated microglia (YAM), and disease-inflammatory macrophages (DIMs), the latter representing microglia-like cells originating from infiltrating monocytes [35]. Marker sets reported in previous single-cell studies of human microglia were also included for comparison. These included signatures derived from datasets incorporated into the Human Microglia Atlas (HuMicA), including those from Mathys et al., Zhou et al., Schirmer et al., and Olah et al. [[36], [37], [38], [39]], together with signatures independently reported by Sun et al. and Gerrits et al. [9,40]. Overall, the identified microglial clusters showed preferential enrichment of previously reported transcriptomic signatures, indicating consistency with microglial populations described in independent datasets. These results support the biological plausibility and consistency of the cluster annotations used in the present study. Pseudotime analysis reveals stage-associated transcriptional changes in inflammatory microglia during AD progression Previous single-nucleus transcriptomic studies have shown that the proportions and gene-expression profiles of human microglial states vary across AD neuropathological stages [9,31]. Building upon the microglial subpopulations identified above, we next examined how these populations change during disease progression and focused on the transcriptional characteristics of inflammatory microglia. We first quantified the relative proportions of each microglial subtype across different AD stages. As shown in Fig. 2A, the composition of microglial subpopulations exhibited complex and dynamic alterations, reflecting the intricate heterogeneity of microglial states during AD pathogenesis. Given that inflammation plays a pivotal role in the pathogenesis of AD, we next used CytoTRACE 2 to estimate their transcriptionally inferred developmental potential across different neuropathological stages. As shown in Fig. 2B, the CytoTRACE 2 scores of inflammatory microglia remained consistently below 0.4 across all stages, indicating that these cells are largely terminally differentiated [23]. Although no statistically significant difference was detected between healthy controls and the early disease stage, the developmental potential of inflammatory microglia exhibited an overall downward trend with disease progression. These findings raise the possibility that inflammatory microglia exhibit reduced developmental plasticity during the middle and late stages of AD. Pseudotime analysis was performed using Monocle2 to infer transcriptional state relationships among inflammatory microglia [24]. The analysis revealed that inflammatory microglial states undergo a continuous and progressive transition along the inferred developmental timeline (Fig. 2C). Stage-stratified visualization showed that HC cells were predominantly located at earlier pseudotime positions, whereas EAD cells were broadly distributed throughout the trajectory. MAD and LAD cells were more frequently distributed along the upper branch and at later pseudotime positions (Fig. 2D). Consistent with these observations, visualization of cells colored by pseudotime further illustrated the gradual transition of inflammatory microglia along the trajectory, reflecting the continuous evolution of transcriptional states during disease progression (Fig. 2E).Fig. 2Pseudotime analysis reveals stage-associated transcriptional changes in inflammatory microglia during AD progression. (A) Stacked bar plot showing the relative proportions of microglial subpopulations across different disease stages. (B) Box plots showing CytoTRACE 2 scores of inflammatory microglia across disease stages, reflecting transcriptionally inferred developmental potential. (C) Pseudotime trajectory of inflammatory microglia showing an inferred transcriptional trajectory. (D) Distribution of inflammatory microglia from different disease stages along the inferred pseudotime trajectory. (E) Pseudotime map of inflammatory microglia colored according to pseudotime values. (F) Heatmap showing the expression patterns of genes significantly associated with pseudotime (q < 0.05) during inflammatory microglial state transitions. ns, not significant; ∗∗Padj < 0.01. Pseudotime analysis reveals stage-associated transcriptional changes in inflammatory microglia during AD progression. (A) Stacked bar plot showing the relative proportions of microglial subpopulations across different disease stages. (B) Box plots showing CytoTRACE 2 scores of inflammatory microglia across disease stages, reflecting transcriptionally inferred developmental potential. (C) Pseudotime trajectory of inflammatory microglia showing an inferred transcriptional trajectory. (D) Distribution of inflammatory microglia from different disease stages along the inferred pseudotime trajectory. (E) Pseudotime map of inflammatory microglia colored according to pseudotime values. (F) Heatmap showing the expression patterns of genes significantly associated with pseudotime (q < 0.05) during inflammatory microglial state transitions. ns, not significant; ∗∗Padj < 0.01. Finally, we analyzed gene expression dynamics along the pseudotime trajectory. Using the differentialGeneTest function, we identified 3600 genes (q < 0.05) whose expression was significantly associated with pseudotime. Clustering of these genes revealed distinct pseudotime-associated expression patterns among inflammatory microglia (Fig. 2F). hdWGCNA identification of core modules associated with inflammatory microglial changes Previous transcriptomic studies have shown that microglial gene expression dynamically changes across pathological stages of AD, forming stage-specific gene co-expression modules during disease progression [41]. We performed hdWGCNA to identify co-expression programs associated with inflammatory microglial changes during disease progression. Based on scale-free topology and connectivity criteria, a soft-thresholding power of 4 was selected for network construction (Fig. 3A). Hierarchical clustering identified seven gene co-expression modules, each represented by a distinct color (Fig. 3B).Fig. 3hdWGCNA identification of core modules associated with inflammatory microglial changes. (A) Selection of the soft-thresholding power for network construction. (B) Hierarchical clustering dendrogram showing seven gene co-expression modules identified by hdWGCNA. Different colors represent distinct modules. (C) Pseudotime-associated expression patterns and functional enrichment of the identified modules. The left panel shows module eigengene patterns along the inferred pseudotime trajectory, and the right panel shows representative GO biological processes and KEGG pathways enriched in each module. (D) M2 module scores across disease stages. (E) Distribution of module membership scores and representative hub genes within the M2 module. (F) Protein-protein interaction (PPI) network of genes within the M2 module, showing connectivity among representative hub genes. ∗∗Padj < 0.01. hdWGCNA identification of core modules associated with inflammatory microglial changes. (A) Selection of the soft-thresholding power for network construction. (B) Hierarchical clustering dendrogram showing seven gene co-expression modules identified by hdWGCNA. Different colors represent distinct modules. (C) Pseudotime-associated expression patterns and functional enrichment of the identified modules. The left panel shows module eigengene patterns along the inferred pseudotime trajectory, and the right panel shows representative GO biological processes and KEGG pathways enriched in each module. (D) M2 module scores across disease stages. (E) Distribution of module membership scores and representative hub genes within the M2 module. (F) Protein-protein interaction (PPI) network of genes within the M2 module, showing connectivity among representative hub genes. ∗∗Padj < 0.01. To characterize these gene modules, we examined their eigengene-expression patterns along the inferred pseudotime trajectory and performed GO and KEGG enrichment analyses (Fig. 3C). The left panel shows the pseudotime-associated expression patterns of the module eigengenes, whereas the right panel summarizes biological processes and signaling pathways enriched in each module. These analyses revealed distinct functional and transcriptional characteristics across the seven co-expression modules. The M2 module was predominantly enriched in pathways related to inflammation and immune activation, including Th17 cell differentiation, negative regulation of the inflammatory response, and the JAK-STAT signaling pathway, suggesting a potential inflammation-restraining function of this module. The M2 module eigengene progressively decreased along the inferred pseudotime trajectory and reached its lowest level at the terminal region (Fig. 3C). This suggests that the M2-represented inflammation-regulatory program gradually weakened, possibly indicating diminished microglial capacity to constrain inflammatory responses. We next evaluated the expression scores of the M2 module across different disease groups. The results showed statistically significant differences in M2 module scores among the groups: relatively higher scores were observed in the HC and EAD groups, whereas significantly lower scores were found in the MAD and LAD groups (Fig. 3D). To further explore the regulatory architecture of the M2 module, we analyzed the distribution of module membership scores and identified representative hub genes (Fig. 3E). In addition, protein-protein interaction (PPI) network analysis revealed strong connectivity among M2 module genes, highlighting key hub genes that may drive inflammatory microglial functional changes during AD progression (Fig. 3F). Machine learning prioritizes candidate genes associated with inflammatory microglial changes Given the critical role of the M2 module in inflammatory microglial dynamics, we applied machine learning approaches to prioritize candidate genes associated with inflammatory microglial changes. Twelve machine learning algorithms were integrated with a 10-fold cross-validation framework to construct predictive models based on M2 module genes [42]. Model performance was evaluated in the training dataset and four independent validation cohorts (Fig. 4A, Table S4).Fig. 4Machine learning prioritizes candidate genes associated with inflammatory microglial changes. (A) Heatmap showing AUC values of different machine-learning models in the training dataset and external validation cohorts. (B, C) ROC curves of the RF model in the training dataset and the GSE118553 validation cohort, respectively. (D, E) Calibration curves of the RF model in the training dataset and GSE118553 cohort, respectively. (F, G) Decision curve analyses of the RF model in the training dataset and GSE118553 cohort, respectively. (H) ROC curve of the final multigene model. (I) ROC curves of individual candidate genes. (J) Feature importance ranking of genes in the RF model. (K) Venn diagram showing the overlap among differentially expressed genes (DEGs) identified in inflammatory microglia, M2 module genes, and RF-selected genes. (L) Pseudotime-associated expression patterns of the three overlapping candidate genes. Machine learning prioritizes candidate genes associated with inflammatory microglial changes. (A) Heatmap showing AUC values of different machine-learning models in the training dataset and external validation cohorts. (B, C) ROC curves of the RF model in the training dataset and the GSE118553 validation cohort, respectively. (D, E) Calibration curves of the RF model in the training dataset and GSE118553 cohort, respectively. (F, G) Decision curve analyses of the RF model in the training dataset and GSE118553 cohort, respectively. (H) ROC curve of the final multigene model. (I) ROC curves of individual candidate genes. (J) Feature importance ranking of genes in the RF model. (K) Venn diagram showing the overlap among differentially expressed genes (DEGs) identified in inflammatory microglia, M2 module genes, and RF-selected genes. (L) Pseudotime-associated expression patterns of the three overlapping candidate genes. Among all tested models, the RF algorithm showed the best predictive performance. The RF model achieved an AUC of 0.963 (95% CI: 0.924–0.989) in the training dataset and maintained acceptable predictive ability in the external validation cohort GSE118553 with an AUC of 0.724 (95% CI: 0.615–0.819) (Fig. 4B and C). Calibration curves showed reasonable agreement between predicted and observed probabilities in both the training and validation datasets (Fig. 4D and E). Consistently, DCA further suggested a positive net benefit of the RF model across a range of threshold probabilities compared with the “treat-all” and “treat-none” strategies (Fig. 4F and G). The final multigene model also exhibited strong discriminative ability (Fig. 4H), while ROC analysis of individual genes further supported their discriminative performance (Fig. 4I). Feature importance analysis of the RF model identified eight key predictors: ACSS3, KLHL2, LGMN, KCNQ3, MTCH1, KPNB1, RANBP9, and TMEM63B (Fig. 4J). To further prioritize candidate genes associated with inflammatory microglial changes, we intersected DEGs identified in inflammatory microglia, hdWGCNA module genes, and RF-selected genes. Three overlapping genes (KCNQ3, LGMN, and KLHL2) were identified as candidate hub genes (Fig. 4K). Visualization of their expression patterns along pseudotime revealed dynamic transcriptional changes during inflammatory microglial state transitions (Fig. 4L), suggesting that these genes may play key roles in microglial functional remodeling during AD progression. Among these three genes, the expression pattern of KLHL2 was highly consistent with the M2 module, showing progressive downregulation along the inferred pseudotime. It was also ranked second in RF feature importance, suggesting its potential role as a key gene within the M2 module. Functionally, KLHL2 has been implicated in the regulation of WNK3 level, while WNK3 has been linked to downstream JNK/c-Jun signaling [13,14,43]. We therefore investigated whether KLHL2 regulates WNK3 level and downstream JNK/c-Jun signaling in microglia under AD-related conditions. KLHL2 expression level is downregulated and WNK3 level is post-translationally elevated in AD-related models Given the previously reported regulatory role of KLHL2 on WNK3, we first examined Klhl2 expression in vivo using brain tissues from 12-month-old male WT and APP/PS1 mice. qPCR analysis revealed a significant decrease in Klhl2 mRNA expression in APP/PS1 mice compared with WT controls (Fig. 5A), whereas no significant difference was observed in Wnk3 mRNA levels (Fig. 5B). Western blot analysis confirmed that KLHL2 level was reduced in APP/PS1 mice, while WNK3 level was significantly upregulated (Fig. 5C–E).Fig. 5The expression levels of KLHL2 and WNK3 in AD models were assessed. (A, B) Relative mRNA expression levels of Klhl2 and Wnk3 in brain tissues from WT and APP/PS1 mice (n = 3). (C) Representative Western blots of KLHL2 and WNK3 in brain tissues from WT and APP/PS1 mice. (D, E) Quantification of KLHL2 and WNK3 levels in the two groups (n = 3). (F, G) Relative mRNA expression levels of Klhl2 and Wnk3 in control (Con) and Aβ42-stimulated BV2 microglia (n = 3). (H) Representative Western blots of KLHL2 and WNK3 in Con and Aβ42-stimulated BV2 microglia. (I, J) Quantification of KLHL2 and WNK3 levels in the two groups (n = 3). Data are presented as mean ± SEM. ns, not significant; ∗P < 0.05; ∗∗P < 0.01. The expression levels of KLHL2 and WNK3 in AD models were assessed. (A, B) Relative mRNA expression levels of Klhl2 and Wnk3 in brain tissues from WT and APP/PS1 mice (n = 3). (C) Representative Western blots of KLHL2 and WNK3 in brain tissues from WT and APP/PS1 mice. (D, E) Quantification of KLHL2 and WNK3 levels in the two groups (n = 3). (F, G) Relative mRNA expression levels of Klhl2 and Wnk3 in control (Con) and Aβ42-stimulated BV2 microglia (n = 3). (H) Representative Western blots of KLHL2 and WNK3 in Con and Aβ42-stimulated BV2 microglia. (I, J) Quantification of KLHL2 and WNK3 levels in the two groups (n = 3). Data are presented as mean ± SEM. ns, not significant; ∗P < 0.05; ∗∗P < 0.01. We further validated these findings in an in vitro AD cell model using Aβ42-treated BV2 microglial cells. Consistent with the in vivo results, Aβ42 treatment induced a marked downregulation of Klhl2 mRNA expression (Fig. 5F), with no significant change in Wnk3 mRNA levels (Fig. 5G). Western blot analysis further confirmed that KLHL2 level was decreased in Aβ42-treated BV2 cells, whereas WNK3 level was increased (Fig. 5H–J). Collectively, these in vivo and in vitro results showed that KLHL2 expression was downregulated in AD-related models, whereas WNK3 level was significantly increased without corresponding changes in Wnk3 mRNA. These findings suggest a regulatory relationship between KLHL2 and WNK3. KLHL2 overexpression reduces WNK3/JNK/c-Jun signaling and inflammatory activation in Aβ42-treated BV2 microglia Previous studies have shown that WNK3 can regulate cellular functions through the downstream JNK/c-Jun signaling pathway [43]. To investigate the role of KLHL2 in regulating microglial inflammatory responses, we generated BV2 microglial cells stably overexpressing FLAG-tagged KLHL2 through lentiviral transduction. Successful overexpression was confirmed by Western blot for FLAG (Fig. 6A and B).Fig. 6KLHL2 overexpression reduces WNK3/JNK/c-Jun signaling and inflammatory activation in Aβ42-stimulated BV2 microglia. (A) Representative Western blots showing FLAG-tagged KLHL2 and WNK3 levels in Vector + Aβ42 and OE-KLHL2 + Aβ42 BV2 microglia. (B, C) Quantification of FLAG and WNK3 levels in the two groups (n = 3). (D) Representative Western blots of p-JNK, JNK, p-c-Jun, and c-Jun in Vector + Aβ42 and OE-KLHL2 + Aβ42 BV2 microglia. (E, F) Quantification of p-JNK/JNK and p-c-Jun/c-Jun ratios in the two groups (n = 3). (G) Representative Western blots of IL-6, TNF-α, IL-1β, and iNOS in Vector + Aβ42 and OE-KLHL2 + Aβ42 BV2 microglia. (H) Quantification of IL-6, TNF-α, IL-1β, and iNOS levels in the two groups (n = 3). Data are presented as mean ± SEM. ∗P < 0.05; ∗∗P < 0.01. KLHL2 overexpression reduces WNK3/JNK/c-Jun signaling and inflammatory activation in Aβ42-stimulated BV2 microglia. (A) Representative Western blots showing FLAG-tagged KLHL2 and WNK3 levels in Vector + Aβ42 and OE-KLHL2 + Aβ42 BV2 microglia. (B, C) Quantification of FLAG and WNK3 levels in the two groups (n = 3). (D) Representative Western blots of p-JNK, JNK, p-c-Jun, and c-Jun in Vector + Aβ42 and OE-KLHL2 + Aβ42 BV2 microglia. (E, F) Quantification of p-JNK/JNK and p-c-Jun/c-Jun ratios in the two groups (n = 3). (G) Representative Western blots of IL-6, TNF-α, IL-1β, and iNOS in Vector + Aβ42 and OE-KLHL2 + Aβ42 BV2 microglia. (H) Quantification of IL-6, TNF-α, IL-1β, and iNOS levels in the two groups (n = 3). Data are presented as mean ± SEM. ∗P < 0.05; ∗∗P < 0.01. Following Aβ42 stimulation, WNK3 levels were markedly reduced in Klhl2-overexpressing cells compared with control cells (Fig. 6A and C). Consistent with this, the downstream JNK/c-Jun signaling pathway was suppressed in Klhl2-overexpressing cells after Aβ42 treatment, as evidenced by decreased phosphorylation of JNK and c-Jun, while JNK and c-Jun levels remained unchanged (Fig. 6D–F). We next examined inflammatory responses in BV2 cells. Klhl2 overexpression significantly reduced the levels of the pro-inflammatory mediators IL-6, TNF-α, IL-1β and iNOS (Fig. 6G and H). To further validate these findings, we performed loss-of-function experiments using siRNA-mediated Klhl2 knockdown (Fig. S2A–B). Among the three siRNAs tested, si-KLHL2-2 showed the highest knockdown efficiency and was therefore selected for subsequent experiments. Under this condition, siRNA-mediated Klhl2 knockdown increased WNK3 level (Fig. S2C–D), enhanced JNK and c-Jun phosphorylation (Fig. S2E–G), and increased IL-6, TNF-α, IL-1β and iNOS expression levels (Fig. S2H–I). Collectively, these complementary gain- and loss-of-function findings support a negative regulatory role of KLHL2 in Aβ42-induced microglial inflammatory activation and suggest the involvement of WNK3/JNK/c-Jun signaling. Discussion In this study, we investigated the role of KLHL2 in inflammatory microglia during AD progression using an integrated single-nucleus transcriptomic and experimental approach. Pseudotime and CytoTRACE 2 analyses revealed stage-associated transcriptional changes in inflammatory microglia, while hdWGCNA and machine learning prioritized KLHL2 as a candidate regulator associated with these changes. KLHL2 expression progressively decreased along the inferred pseudotime trajectory, with concordant decreases in Klhl2 mRNA and KLHL2 levels also observed in APP/PS1 mice and Aβ42-treated BV2 cells. In contrast, WNK3 level was increased without corresponding changes in Wnk3 mRNA. In BV2 cells, Klhl2 overexpression reduced WNK3 level, JNK and c-Jun phosphorylation, and pro-inflammatory mediator expression, whereas Klhl2 knockdown produced the opposite effects. Collectively, these findings support a negative regulatory role of KLHL2 in microglial inflammatory responses and suggest the involvement of WNK3/JNK/c-Jun signaling. Advances in single-cell and single-nucleus transcriptomic technologies have profoundly expanded our understanding of microglial heterogeneity in AD. Consistent with previous studies [36,39], we identified multiple disease-associated microglial subpopulations with distinct transcriptional and functional features. These findings further support the view that microglial responses in AD are highly heterogeneous and cannot be represented by a single activation state. Sun et al. identified several inflammation-associated microglial states in the human brain, including the MG2 state characterized by CPEB4 and TMEM163, which showed a relatively sustained transcriptional response to inflammatory stimulation [9]. Based on this framework, we annotated the CPEB4/TMEM163-expressing population as inflammatory microglia. Although its relative level tended to decrease with AD progression, the difference was not statistically significant. This pattern may reflect redistribution among partially overlapping microglial transcriptional states, which are shaped by the evolving pathological microenvironment and dynamically remodeled during AD progression [[44], [45], [46]]. More importantly, the M2 co-expression module identified in inflammatory microglia progressively decreased along the inferred disease trajectory. Because this module was enriched in processes related to negative regulation of inflammatory responses, its reduced activity may indicate a progressive weakening of transcriptional programs that restrain excessive inflammation. Thus, inflammatory dysregulation in AD may involve not only changes in microglial subpopulation composition, but also impaired inflammatory regulation within specific microglial states. Such impairment may coexist with broader microglial dysfunction in AD. For example, terminally inflammatory microglia show both inflammatory and cellular stress features together with impaired Aβ-related processing [47]. Overall, AD progression may involve extensive remodeling of microglial states accompanied by a progressive loss of transcriptional programs that maintain inflammatory homeostasis. Within the M2 module, KLHL2 emerged as a key candidate associated with the progressive loss of inflammation-regulatory programs in inflammatory microglia. KLHL family proteins commonly function as substrate adaptors for Cullin-3 (CUL3) E3 ubiquitin ligase complexes and play important roles in maintaining cellular protein homeostasis [12]. Several KLHL family members have also been implicated in AD-related molecular and pathological processes [[48], [49], [50]]. In our study, KLHL2 expression progressively decreased along the pseudotime trajectory of inflammatory microglia and was consistently reduced in both in vivo and in vitro AD-related models. Together with the gain- and loss-of-function findings, these observations suggest that reduced KLHL2 expression may contribute to the weakening of inflammation-restraining mechanisms in microglia. Previous studies of KLHL2 have mainly focused on cancer and vascular regulation [14,51,52], whereas its role in microglial inflammation remains poorly understood. Our findings highlight a previously underappreciated link between KLHL2 and the dysregulation of inflammatory homeostasis in AD microglia. KLHL2 has been implicated in the regulation of WNK3 level [13,14]. In the present study, the inverse changes in KLHL2 and WNK3 levels, together with the gain- and loss-of-function results, suggest that KLHL2 negatively regulates WNK3 level. WNK3 has been implicated in inflammatory and secondary injury responses following intracerebral hemorrhage and ischemic stroke [53,54], and has also been linked to downstream JNK/c-Jun signaling [43]. Aberrant activation of the JNK/c-Jun pathway is closely associated with AD-related neuroinflammation [55,56]. Consistent with these observations, modulation of Klhl2 expression produced opposing changes in WNK3 level, JNK and c-Jun phosphorylation, and pro-inflammatory mediator expression. These findings suggest that reduced KLHL2 expression during AD progression may weaken its negative regulation of WNK3, thereby favoring enhanced downstream inflammatory signaling. Taken together, our results support KLHL2 as a negative regulator of Aβ42-associated microglial inflammatory activation, suggest the involvement of WNK3/JNK/c-Jun signaling in this regulatory process, and raise the possibility that preserving KLHL2 expression may help limit Aβ-driven neuroinflammation. This study has several limitations. First, the single-nucleus analyses were based on postmortem human brain tissues and therefore provide only cross-sectional snapshots of disease progression. Second, most functional experiments were performed in BV2 cells, which do not fully recapitulate primary human microglia or the in vivo microenvironment. Third, although the findings suggest that KLHL2 negatively regulates WNK3 level, the underlying molecular mechanism remains unclear. Finally, the effects of microglial KLHL2 modulation on Aβ pathology, neuroinflammation, and cognitive outcomes were not evaluated in vivo, and the efficacy and safety of KLHL2-targeted interventions therefore remain to be established. Future studies should employ AD mouse models with microglia-specific Klhl2 knockout or overexpression to determine the causal effects of microglial KLHL2 on neuroinflammation and cognition and to directly assess its influence on Aβ pathology in vivo. Such studies will also be necessary to evaluate the efficacy and safety of KLHL2-targeted interventions. Further functional validation in AD patient–derived iPSC microglia or primary human microglia would strengthen the physiological and translational relevance of these findings. In addition, protein stability assays, interactome profiling, and unbiased analyses of potential post-translational regulatory mechanisms will be needed to determine how KLHL2 regulates WNK3 level. In summary, our study identifies KLHL2 as a potential regulator of inflammatory microglial dysfunction in AD and suggests its involvement in WNK3/JNK/c-Jun signaling. These findings provide new insight into the molecular basis of microglial inflammatory dysregulation and support further investigation of KLHL2 in AD. Consent for publication Not applicable. Data availability The publicly available datasets analyzed in this study were obtained from the Gene Expression Omnibus (GEO). Additional data generated or analyzed during this study are available from the corresponding author upon reasonable request. Author contributions Yao Yang: Writing – original draft, Software, Methodology. Conghui Dai: Writing – review & editing. Zongtang Xu: Methodology, Formal analysis. Linpei Zhuo: Methodology, Formal analysis. Kun Xiang: Methodology, Formal analysis. Xiaoou Lin: Methodology, Formal analysis. Jing Sun: Project administration, Funding acquisition, Formal analysis. Jiaming Liu: Writing – review & editing, Project administration, Methodology, Funding acquisition, Formal analysis, Conceptualization. Qiuling Tong: Writing – review & editing, Project administration, Methodology, Formal analysis, Conceptualization. All authors read and approved the final manuscript. Funding This work was supported by the 10.13039/501100004731Natural Science Foundation of Zhejiang Province (No. LZ25H090004). Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.