Targeting the microglial phosphatidylethanolamine synthesis pathway promotes GABARAP-associated phagocytosis and Aβ clearance in Alzheimer’s disease

Abstract

Alzheimer’s disease (AD) is a major cause of dementia and a prevalent age-related neurodegenerative disorder characterized by progressive cognitive impairment and memory loss. Although metabolic activation or dysfunction of microglia is implicated in AD pathogenesis, the phospholipid metabolism–associated signaling mechanisms within microglia remain poorly defined. In this study, we demonstrate that quinolinic acid (QA), a byproduct of tryptophan catabolism via the kynurenine pathway, activates the microglial Kennedy pathway—responsible for de novo phosphatidylethanolamine (PE) biosynthesis—by upregulating the enzymes EPT1 and ETNK1. This activation markedly enhances the synthesis of PE species enriched in polyunsaturated fatty acids. Concurrently, QA significantly increases the expression of gamma-aminobutyric acid receptor–associated protein (GABARAP), promotes its lipidation, and facilitates the GABARAP-associated phagocytosis (GAP) of Aβ oligomers by microglia. Knockdown of EPT1 and ETNK1 attenuated QA-induced PE synthesis and impaired the GAP of Aβ oligomers, whereas inhibition of GABARAP lipidation via STBD1 deconjugase substantially reduced QA-mediated GAP. QA administration upregulated microglial Gabarap expression and decreased the Aβ plaque burden in the hippocampus of AD (5xFAD) mice, whereas Gabarap knockdown abrogated QA-induced microglial clearance of Aβ. Collectively, these findings reveal a paradoxically beneficial role of QA in activating a microglia-specific signaling cascade that promotes PE biosynthesis and GAP, thereby enhancing Aβ clearance and mitigating AD pathology. Targeting the microglial PE synthesis pathway and GAP may represent a promising therapeutic strategy to ameliorate Aβ accumulation and slow AD progression.

Introduction

Alzheimer’s disease (AD) is an age-associated neurodegenerative disorder characterized by a progressive decline in cognitive and memory functions.^1,2 The major pathological features of the AD brain include the formation of insoluble senile plaques composed of amyloid-beta (Aβ) (1-42) and the accumulation of neurofibrillary tangles formed by phosphorylated tau. Beyond these classical hallmarks, it is now widely accepted that chronic neuroinflammation plays a pivotal role in the initiation and propagation of neurodegeneration. Aβ and senile plaques trigger chronic, localized inflammation in the AD brain, which further amplifies neuronal damage and disease progression.^3,4

Microglia are a type of glial cell and the primary immune cells of the brain. They perform various functions, including surveilling for pathogens, monitoring neuronal synaptic activity, and phagocytosing dead cells.^5,6 These immune surveillance cells are responsible for the constant maintenance of the brain microenvironment and the preservation of neuronal integrity. Microglia respond to Aβ, brain injury, chemokines, cytokines, and other toxic molecules. Microglial activation is a pathological hallmark of presymptomatic AD and occurs before Aβ- and tau-induced neuronal dysfunction in the aging brain.^7–9 In AD, microglia undergo dynamic morphological and functional changes. As the disease advances, these cells transition from a homeostatic state to various disease-associated phenotypes, often losing their protective capacity. However, the precise mechanisms by which microglia become neurotoxic and contribute to memory loss remain poorly understood. Microglia rely on glycolysis to produce ATP and regulate diverse cellular functions. Recent advances in immunometabolism have highlighted that these functional transitions are governed by profound metabolic shifts within the cell. Not only do they have the ability to utilize multiple energy sources, but they also undergo metabolic reprogramming during AD pathogenesis, a process increasingly recognized for its contribution to brain aging and neurodegeneration.^10,11 This reprogramming is not merely a reactive change but a fundamental determinant of microglial survival and clearance efficiency in the pathological milieu.

One notable metabolic alteration in microglia is the abnormal increase in quinolinic acid (QA), a byproduct of the tryptophan-kynurenine metabolic pathway, observed under AD-related stress conditions. The kynurenine pathway, which accounts for the vast majority of central tryptophan catabolism, is frequently dysregulated in aging and chronic neurodegenerative disorders. QA is known to act as an endogenous neurotoxin.^12–15 An increasing body of evidence suggests that QA contributes to AD pathogenesis by promoting excitotoxicity and neuroinflammation.^15,16 Microglia-derived QA acts as an N-methyl-D-aspartate (NMDA) receptor agonist, disrupting calcium (Ca²⁺) homeostasis and mitochondrial function.^16–18 QA also induces tau phosphorylation and oxidative stress, ultimately leading to neuronal cell death. Additionally, QA stimulates astrocytes, increasing the release of chemokines that enhance inflammatory reactions in the brain.¹⁹ Notably, QA levels are elevated in the cerebrospinal fluid and brains of AD patients, indicating that it serves as a pathology-associated metabolic biomarker of AD, a condition deeply rooted in the aging process. While QA has long been characterized primarily by its neurodestructive properties, its role as a potential signaling molecule within glial cells under controlled physiological or acute stress conditions is beginning to emerge as a complex area of study. Moreover, increased QA levels correlate with the severity of cognitive impairment, and QA administration exacerbates cognitive deficits and tau hyperphosphorylation in animal models of AD.^20–25 This systemic-central crosstalk through the kynurenine pathway suggests that brain QA levels are sensitive to global metabolic status during aging.

Although the effects of QA on neurons and astrocytes have been reported, its impact on microglial phagocytic function in the context of AD pathology remains unexplored.^13–25 Despite the established link between lipid metabolism and immune function, the specific involvement of phospholipid synthesis in microglial adaptation to AD-related stress remains poorly defined. In our preliminary studies, we revealed that QA enhances microglial phagocytosis under AD-related stress in a microglia-specific manner. On the basis of our preliminary observations and prior reports characterizing QA as a signaling molecule in the brain,^26–29 we hypothesized that QA modulates microglial phagocytic function under AD-related stress. To test this hypothesis, in this study, we comprehensively investigated the cellular and molecular mechanisms by which QA modulates microglial phagocytosis within the pathological microenvironment of AD. We employed lipidomic profiling and single-cell transcriptomic analyses to determine how QA promotes de novo phospholipid synthesis and to identify QA-responsive gene networks contributing to Aβ clearance. Furthermore, we performed complementary in vitro and in vivo experiments to validate whether QA-induced phospholipid metabolism facilitates microglial adaptation to the AD milieu and enhances Aβ clearance. This GABARAP-mediated pathway represents a specialized branch of non-canonical autophagy that is distinctly tuned to handle extracellular protein aggregates. Collectively, the results of our study reveal a previously unrecognized signal transduction mechanism—Kennedy pathway activation coupled with GABARAP-associated phagocytosis (GAP)—that mediates the paradoxically beneficial effect of QA on the microglial clearance of Aβ in AD. These findings provide new insights into how microglial metabolic adaptation can be harnessed as a novel therapeutic strategy to mitigate neuropathology and maintain cellular homeostasis in the aging brain.

Results

Quinolinic acid (QA) increases PE synthesis via the Kennedy pathway in microglia

In light of evidence that QA is an endogenous metabolite of the kynurenine pathway that accumulates in the aging and AD brain and is known to modulate neuroinflammatory signaling, we sought to determine whether QA directly regulates microglial phagocytic function, a critical process for amyloid-β (Aβ) clearance in AD.^1–4 To this end, we initially examined the direct effects of QA on microglial behavior using human microglial (HMC3) cells, a well-established in vitro model for studying human microglial activation and phagocytic responses.^5,6,30 We first confirmed that the concentrations of QA that affected the metabolic activity of neuronal SH-SY5Y cells did not affect microglial viability, as determined by XTT, calcein, annexin V, and CTG assays (Supplementary Fig. 1a–p). To further evaluate the effect of QA on microglial functional activation, we assessed its effect on the phagocytosis of myelin, a key physiological substrate for microglial clearance. Quantitative analysis of pHrodo-labeled myelin revealed that QA treatment significantly increased myelin phagocytosis in HMC3 cells (Supplementary Fig. 2a–c). Under these microglia-sparing conditions, QA treatment significantly enhanced both microglial motility (Supplementary Fig. 2d–f) and the phagocytosis of Aβ oligomers (Supplementary Fig. 2g, h). Specifically, dual-label live-cell imaging was performed using pHrodo–Aβ (red) to monitor Aβ uptake into acidified compartments and LysoTracker (green) to visualize active lysosomes. Because microglia in the AD brain encounter both extracellular Aβ and metabolic signals, we examined whether QA modulates Aβ processing under Aβ-containing conditions. QA treatment led to a significant increase in both the red and green signals over time, and merged images showed enhanced colocalization between pHrodo–Aβ and LysoTracker. Quantitative analysis confirmed a significant increase in the total pHrodo–Aβ intensity in QA-treated microglia (Supplementary Fig. 2g, h). These findings indicate that QA enhances microglial uptake and lysosomal processing of Aβ, suggesting that QA acts additively with Aβ to promote microglial Aβ clearance. To investigate whether QA promotes Aβ signaling primarily through enhanced uptake and accelerated intracellular degradation, we performed a load-and-washout assay and monitored the decay of intracellular pHrodo-Aβ over time (Supplementary Fig. 2i). HMC3 microglia were preloaded with pHrodo-Aβ, washed to remove extracellular Aβ, and subsequently imaged under control or QA conditions (Supplementary Fig. 2j). In the control groups, the intracellular pHrodo-Aβ signal gradually decreased during time-lapse imaging. Notably, compared with the control groups, the QA-treated groups presented a reduction in the intracellular pHrodo-Aβ signal after washout (Supplementary Fig. 2k), resulting in a significantly greater Aβ degradation rate (Supplementary Fig. 2l). These results indicate that QA not only increases Aβ uptake but also accelerates the lysosomal degradation of internalized Aβ. Because phagocytosis requires rapid membrane remodeling, curvature generation, and membrane fusion, we reasoned that QA-induced phagocytic activation depends on lipid biosynthetic pathways that support these processes. Phosphatidylethanolamine (PE) is a key membrane lipid that promotes membrane curvature and phagosomal closure.^31–34 We therefore hypothesized that QA-induced microglial phagocytosis is mediated by enhanced PE biosynthesis through activation of the Kennedy pathway, the principal de novo route for PE synthesis. Importantly, PE is also required for the lipidation of ATG8 family proteins, which are essential for autophagy-related phagocytic processes. In this context, we examined whether QA-induced microglial phagocytosis is mediated by PE biosynthesis through activation of the Kennedy pathway. To assess the effect of QA on the expression of Kennedy pathway-associated enzymes, we measured the mRNA levels of ETNK1, PCYT2, and EPT1 in microglia treated with QA over a 24-h time course (0.2, 0.5, 1, 3, 6, 9, 12, and 24 h) (Fig. 1a, b). Quantitative PCR (qPCR) analysis revealed a significant increase in ETNK1 and EPT1 mRNA levels following QA treatment, with a time-dependent bell-shaped expression curve. In contrast, PCYT2 mRNA levels remained unchanged throughout the 24-h QA exposure (Fig. 1c–e). Having confirmed that QA selectively upregulates key enzymes in the Kennedy pathway, we next investigated whether QA affects de novo PE synthesis in microglia. Untargeted lipidomic analyses using UPLC‒Orbitrap‒MS/MS were performed (Fig. 1f, g). Principal component analysis (PCA) of lipidomic profiles revealed distinct clustering between control and QA-treated microglia, indicating significant alterations in lipid composition (Fig. 1h). An overview of the total lipid species composition in HMC3 cells is provided in Supplementary Fig. 1l, m. Notably, this separation between groups was more pronounced in the negative ionization mode, which is more sensitive for lipid species with acidic or anionic properties, including PE. These findings provide further support for our hypothesis that the Kennedy pathway and subsequent PE metabolism are primary targets of QA-mediated metabolic reprogramming.

Fig. 1.

QA increases PE synthesis via the Kennedy pathway in microglia. a A schematic representation of the Kennedy pathway related to PE biosynthesis and its associated enzymes. b Experimental workflow for qPCR analysis of QA-treated microglia at various time points (0.2, 0.5, 1, 3, 6, 9, 12, and 24 h). c ETNK1 mRNA levels significantly increased in a time-dependent manner following QA treatment (n = 3 independent experiments per group). d PCYT2 mRNA levels show little to no change over time following QA treatment (three independent experiments per group). e EPT1 mRNA levels significantly increased in a time-dependent manner following QA treatment. (three independent experiments per group). f Experimental workflow for the transfection of shETNK1 and shEPT1, followed by QA treatment of microglia for lipidomic analysis. g Workflow of lipidomics analysis in QA-treated microglia using LC‒MS/MS for lipid identification. h Principal component analysis (PCA) plot showing the lipidomic profiles of each group (Control and QA-treated microglia). i Heatmap of lipidomic results showing changes in various PEs across the control, QA, shETNK1 + QA, shEPT1 + QA, shETNK1 and shEPT1 groups (5 samples per group). j Extracted ion chromatograms (EICs) of specific PE species following the QA treatment. k PE 20:1/22:6, LPE O-16:1, PE 24:0/20:4, LPE 20:4, PE O-18:1/20:3, PE 17:1/18:1, PE 18:1/20:3, and PE 18:0/22:6 were significantly elevated in QA-treated microglia. ETNK1 knockdown or EPT1 knockdown significantly inhibited the increase in PE biosynthesis induced by QA treatment (5 samples per group). l RPKM values of ETNK1 and EPT1 mRNA were significantly greater in microglia from AD patients than in those from normal controls, as determined on the basis of data from the GSE125050 dataset (n = 9 patients per group). m Representative scanning electron microscopy (SEM) image of an HMC3 cell treated with saline or QA. Scale bars (yellow): 2.5 μm. n 3D surface plot analysis of SEM images visualizing changes in cell surface complexity. o Surface roughness was significantly greater in QA-treated microglia than in control microglia. A total of 90 ROIs per group were analyzed (15 ROIs per cell; 6 cells per group). The data are presented as the mean ± SEM. Statistical significance was evaluated by unpaired two-tailed Student’s t tests and one-way ANOVA with Tukey’s post hoc multiple-comparison tests. Significant differences at *p < 0.05, **p < 0.01

To verify the role of ETNK1 and EPT1 in this process, we utilized shRNA-mediated knockdown of these enzymes (Supplementary Fig. 3). Heatmap analysis of the lipidomic profiles revealed that QA-induced alterations in PE species were effectively reversed upon the knockdown of these enzymes (Fig. 1i). A representative extracted ion chromatogram (EIC) for LPE O 16:1 further demonstrated these shifts in PE biosynthesis (Fig. 1j). Subsequent quantitative analysis confirmed that QA treatment significantly increased specific PE species, including PE 20:1/22:6, LPE O-16:1, PE 24:0/20:4, LPE 20:4, PE O-18:1/20:3, PE 17:1/18:1, PE 18:1/20:3, and PE 18:0/22:6, whereas the knockdown of either ETNK1 or EPT1 significantly attenuated these increases (Fig. 1k). To further explore the relevance of Kennedy pathway-associated enzymes in AD, we analyzed our publicly available RNA sequencing dataset (GSE125050) from the postmortem brains of control and AD patients.³⁵ ETNK1 and EPT1 mRNA levels were significantly elevated in microglia from AD patients compared with those from control subjects (Fig. 1l). Immunohistochemical analysis of cortical sections revealed colocalization of EPT1 and the microglial marker IBA1 in both AD and control brains. Quantitative analysis confirmed a significant increase in EPT1 expression in IBA1-positive microglia from AD patients compared with controls (Supplementary Fig. 1n, o). Given that PE is a crucial phospholipid known to influence membrane dynamics, we investigated whether the QA-induced increase in PE levels alters microglial membrane surfaces. To explore this, we examined the surface of HMC3 cells. Representative scanning electron microscopy (SEM) images were used to visualize changes in cell surface morphology (Fig. 1m). Subsequent 3D surface plot analysis also revealed altered surface complexity (Fig. 1n), and quantitative analysis confirmed a statistically significant increase in cell surface roughness in QA-treated cells (Fig. 1o). Further ultrastructural analysis using transmission electron microscopy (TEM) revealed that QA treatment led to an increase in the number of intracellular vesicles and altered the morphology of the plasma membrane in HMC3 cells (Supplementary Fig. 1p).

Modulation of the Kennedy pathway regulates QA-induced phagocytosis of Aβ oligomers in microglia

Next, we determined whether the QA-induced increase in PE synthesis is directly associated with alterations in microglial phagocytosis. We employed both genetic and pharmacological approaches to modulate the activity of the Kennedy pathway and assess its effect on microglial phagocytic activity.

To determine whether the increase in PE via ETNK1 and EPT1 facilitates Aβ phagocytosis in QA-treated microglia, we knocked down ETNK1 and EPT1, performed time-lapse live-cell imaging, and analyzed microglial morphology and phagocytic function using AI-based single-cell tracking (Fig. 2a, b). Time-lapse microscopy captured the dynamic uptake of Aβ particles (visualized with pHrodo-Aβ (red)) in control, QA-treated, EPT1 knockdown, and EPT1 + QA knockdown microglia over a 6-h period (Fig. 2c and Supplementary Fig. 2d–l). Quantitative analysis revealed that compared with that in control microglia, the pHrodo-Aβ intensity in QA-treated microglia significantly increased. Notably, EPT1 knockdown attenuated this increase (Fig. 2d). Moreover, QA treatment increased the cumulative frequency of larger pHrodo-Aβ particles, indicating enhanced phagocytosis, whereas EPT1 knockdown effectively blocked this effect (Fig. 2e). EPT1 knockdown significantly reduced both the average intensity of pHrodo-Aβ uptake and the average size of QA-treated microglia (Fig. 2f, g). Similarly, ETNK1 knockdown attenuated both the uptake intensity and particle size (Supplementary Fig. 4a–e). To further validate the causal link between PE accumulation and enhanced phagocytosis, we focused on the PEMT pathway, which serves as a major consumer of PE by converting it to PC. We employed bezafibrate, a pharmacological inhibitor of phosphatidylethanolamine N-methyltransferase (PEMT),^36–39 to preserve the intracellular PE pool (Supplementary Fig. 4f). Lipidomic profiling confirmed that bezafibrate treatment led to a significant increase in specific PE species, most notably PE 18:1/20:3 and PE 18:0/22:6 (Supplementary Fig. 4g–i). As expected, the bezafibrate-mediated expansion of the PE pool was sufficient to increase zymosan particle phagocytosis in microglia, and this effect was further potentiated when the PE pool was combined with QA treatment (Supplementary Fig. 4j–n). Additionally, we examined the effects of meclizine, a PCYT2 enzyme inhibitor,⁴⁰ on PE synthesis and QA-induced phagocytosis (Fig. 2h, I). Meclizine suppressed specific PE species that were elevated by QA (Fig. 2j), significantly blocking QA-induced PE biosynthesis in microglia (Fig. 2k, l). Furthermore, meclizine significantly decreased QA-induced pH-induced zymosan uptake by microglia (Fig. 2m–o). Finally, ultrastructural analysis via transmission electron microscopy (TEM) further revealed that QA treatment induced various vesicular structures in microglia characterized by both single-membrane phagosomal compartments and double-membrane autophagosomes. Notably, meclizine treatment significantly reduced the presence of these QA-induced structures (Fig. 2p).

Fig. 2.

Modulation of the PE synthesis signaling pathway regulates microglial phagocytosis in response to QA. a A scheme of AI-based analysis for characterizing Aβ particle phagocytosis by microglia using time-lapse microscopy imaging. b Experimental workflow for the transfection of shETNK1 and shEPT1, followed by QA treatment of microglia for live-cell imaging. c Illustration of Aβ particles through the pHrodo-Aβ signal (red) in four groups of microglia, namely, the control, QA, shEPT1, and shEPT1 + QA groups, at different time points (0, 2, 4, and 6 h). Scale bars (black): 15 μm. (Control: 771 particles from 20 cells; QA: 699 particles from 20 cells; shEPT1: 556 particles from 15 cells; shEPT1 + QA: 739 particles from 15 cells.) d Cumulative frequency distribution of pHrodo-Aβ oligomer intensity in the four groups of microglia. (Control: 771 particles from 20 cells; QA: 699 particles from 20 cells; shEPT1: 556 particles from 15 cells; shEPT1 + QA: 739 particles from 15 cells.) e Representation of the cumulative frequency of the size of the pHrodo-Aβ particles in all groups of microglia. (Control: 771 particles from 20 cells; QA: 699 particles from 20 cells; shEPT1: 556 particles from 15 cells; shEPT1 + QA: 739 particles from 15 cells.). f Violin plots illustrating the average intensity of pHrodo-Aβ oligomers in all groups of microglia. (Control: 771 particles from 20 cells; QA: 699 particles from 20 cells; shEPT1: 556 particles from 15 cells; shEPT1 + QA: 739 particles from 15 cells.). g Violin plots illustrating a comparative analysis of the average size of pHrodo-Aβ oligomers in all groups of microglia. (Control: 771 particles from 20 cells; QA: 699 particles from 20 cells; shEPT1: 556 particles from 15 cells; shEPT1 + QA: 739 particles from 15 cells.). h Schematic representation of the Kennedy pathway related to PE biosynthesis and its PE synthesis inhibitor. i Experimental workflow for TEM, lipidomic analysis, and live imaging analysis of Meclizine- and QA-treated microglia. j Heatmap of lipidomics results showing changes in various PEs across the control, QA, meclizine + QA, and meclizine groups (5 samples per group). k, l, PE O-18:1/20:3 and PE 18:0/20:6 were significantly elevated in QA-treated microglia. Meclizine significantly inhibited the increase in lipid biosynthesis induced by QA treatment (5 samples per group). m Representative images of zymosan particles through the pHrodo-Zymosan signal (red) in four groups of microglia: Control, QA, Meclizine + QA, and Meclizine at different time points (0, 2, 4, and 6 h). Scale bars (white): 15 μm. n Representation of the cumulative frequency of the pH-induced increase in the intensity of Zymosan particles in all groups of microglia. (Control: 144 particles from 7 cells; QA: 180 particles from 8 cells; Meclizine+QA: 360 particles from 12 cells; Meclizine: 360 particles from 12 cells.). o Violin plots illustrating a comparative analysis of the average intensity of the pH-induced distribution of the rod-shaped Zymosan particles in all groups of microglia. (Control: 144 particles from 7 cells; QA: 180 particles from 8 cells; Meclizine+QA: 360 particles from 12 cells; Meclizine: 360 particles from 12 cells.). p Transmission electron microscopy (TEM) images showing the ultrastructures of microglia in the control, QA, meclizine + QA, and meclizine groups. Red arrowheads indicate diverse vesicular structures, including predominantly single-membrane compartments and some double-membrane vesicles. Scale bars (yellow): 1 μm. The data are presented as the mean ± SEM. Statistical significance was evaluated by one-way ANOVA with Tukey’s post hoc multiple-comparison tests. Significant differences at *p < 0.05 and **p < 0.01 are indicated

QA-mediated Kennedy pathway activation provides neuroprotection and reverses AD-like pathology in vivo

Building on our in vitro mechanistic findings, we next assessed the in vivo effects of QA administration in the 5xFAD mouse model. We first confirmed the time-dependent reduction in the number of Aβ plaques following stereotactic QA injection. To rigorously validate the Aβ clearance effect, we performed a blinded requantification of the IHC data, which confirmed a significant (~55%) reduction in Aβ plaque volume at 1 day post-injection (Supplementary Fig. 5a). Subsequent time-course analysis revealed that this clearance effect was initiated within 1 day, peaked at 3 days, and was partially sustained through 30 days post-injection (Supplementary Fig. 5b, c). Parallel to the reduction in amyloid pathology, we investigated the temporal dynamics of tau phosphorylation at multiple sites, including S199, AT8 (S202/T205), T217, and S396, over a 30-day period (Supplementary Fig. 6a–h). Quantitative analysis revealed that QA administration induced a significant, albeit transient, reduction in AT8 immunoreactivity at 1 d post-injection, which was followed by a progressive increase at 3 d and 30 d (Supplementary Fig. 6c, d). A similar late-stage increase was observed at the S396 site at 3 d and 30 d, although the initial decrease at 1 d did not reach statistical significance (Supplementary Fig. 6g, h). In contrast, p-Tau levels at S199 and T217 did not significantly change throughout the 30-day period (Supplementary Fig. 6a, b, e, f).

To evaluate the transcriptomic impact of QA in the 5xFAD mouse hippocampus, bulk RNA sequencing was conducted 24 h postinjection (Supplementary Fig. 7a). PCA revealed that the disease-associated gene expression of the QA-treated 5xFAD samples shifted toward the WT profile, indicating a partial normalization of disease-associated gene expression (Supplementary Fig. 7b). Differential expression analysis confirmed that QA substantially reversed 5xFAD-associated transcriptional alterations, downregulating genes whose expression was elevated in the disease state while restoring genes whose expression was suppressed (Supplementary Fig. 7c–e). GO enrichment analysis revealed that the QA-downregulated genes were primarily associated with immune and inflammatory pathways (cytokine signaling and leukocyte migration), whereas the QA-upregulated genes were enriched in neuronal and membrane-associated functions (Supplementary Fig. 7f). To assess cross-species conservation, we compared QA-responsive genes from human microglial scRNA-seq with those from mouse hippocampal bulk RNA-seq (Supplementary Fig. 7g). These results revealed that genes associated with shared functional programs were significantly enriched in cytoskeletal organization and membrane raft assembly (Supplementary Fig. 7h). Specifically, QA modulated the expression of actin dynamics regulators (Cnn2, Tpm4, Tpm2, Cfl2, Tpm1, Tagln2, and Tmsb10) and membrane trafficking genes (Anxa2 and S100a10), linking these conserved transcriptional changes to enhanced phagocytic cup formation and maturation. Pathway-level enrichment further supported the modulation of gene sets related to cytoskeletal and immune regulation across species (Supplementary Fig. 7i). Together, these results showed that QA induces conserved transcriptional programs associated with suppressed neuroinflammation and enhanced phagocytic remodeling across species. Consistent with these transcriptomic shifts, QA treatment significantly suppressed the expression of proinflammatory cytokines, including IL-1beta and TNF-alpha, in both microglia and the 5xFAD hippocampus (Supplementary Fig. 8a–g). Notably, QA administration further upregulated the expression of the Kennedy pathway enzymes ETNK1 and EPT1, specifically within IBA1-positive microglia in the 5xFAD mouse brain, mirroring our in vitro findings (Supplementary Fig. 8h–k).

These pathological improvements were accompanied by the restoration of PSD-95 synaptic density and significantly enhanced cognitive performance in terms of the NOR and NOPR recognition indices (Supplementary Fig. 9), establishing the net neuroprotective effect of QA. Having confirmed the protective role of QA, we sought to determine whether these effects are mediated through the Kennedy pathway in vivo. We stereotactically injected meclizine with or without QA into the dorsal hippocampus of 5xFAD mice (Fig. 3a). Compared with vehicle treatment, QA treatment significantly reduced the Abeta plaque volume by approximately 55% (Fig. 3b, c). Furthermore, QA treatment significantly increased the number of microglial branch points, an effect that was effectively attenuated by coadministration of meclizine (Fig. 3d). Meclizine treatment alone did not significantly affect plaque volume or microglial morphology. Finally, ultrastructural evaluation using TEM confirmed that meclizine blocked the QA-induced changes in Abeta and microglial morphology, providing cellular-level evidence for the involvement of this pathway (Fig. 3e). Together, these in vitro and in vivo pharmacological data indicate that the Kennedy pathway plays a crucial role in QA-induced microglial phagocytosis and Aβ clearance in AD models.

Fig. 3.

The inhibition of the PE synthesis pathway attenuated QA-induced microglial activation and Aβ clearance in AD mice. a Experimental workflow for the stereotactic injection of QA into the cortex and dorsal hippocampus of 5xFAD mice. b Representative images of IBA1 (microglial marker, green) and Aβ (red) immunofluorescence staining in the hippocampus of 5xFAD mice injected with saline, QA, meclizine + QA or meclizine. Scale bars (white): 40 μm. c The Aβ plaque volume was significantly lower in the QA group than in the saline group, but the addition of meclizine (meclizine + QA group) or meclizine alone did not significantly alter the Aβ plaque volume compared with that in the saline-injected 5xFAD group. A total of 75 plaques were analyzed (15 plaques per mouse; N = 5 mice per group). d The number of microglial branch points significantly increased in the QA treatment group compared with the saline group, but the addition of meclizine or meclizine alone significantly decreased the number of branch points compared with that in the QA-only group. A total of 50 cells were analyzed (10 cells per mouse; N = 5 mice per group). e TEM images showing the Aβ and microglial ultrastructures of the hippocampus in the saline, QA, meclizine + QA and meclizine-injected 5xFAD groups. Scale bars (black): 5 μm, (white): 500 nm. The data are presented as the mean ± SEM. Statistical significance was evaluated by one-way ANOVA with Tukey’s post hoc multiple-comparison tests. Significant differences at *p < 0.05, **p < 0.01

Single-cell transcriptome analysis reveals overlapping gene signatures between in vitro QA-exposed microglia and AD postmortem brain-derived microglia

Following our observation that QA enhances Aβ oligomer phagocytosis in microglia, we performed single-cell RNA sequencing (scRNA-seq) to investigate downstream molecular pathways involved in QA-dependent microglial phagocytosis (Fig. 4a and Supplementary Fig. 10a). Uniform manifold approximation and projection (UMAP) analysis of control versus QA-treated microglia revealed distinct clustering patterns, highlighting dynamic changes in microglial populations (Fig. 4b). We further generated a heatmap to compare the proportion of cells within each cluster (Fig. 4c and Supplementary Fig. 10b), identifying cluster ‘g’ as the most significantly altered population in response to QA. Functional annotation using KEGG pathway enrichment highlighted genes that were differentially expressed across the nine major microglial clusters, with strong enrichment for gene signatures associated with phagocytosis and neurodegeneration-related pathways (Fig. 4d). Gene Ontology (GO) analysis of the QA-responsive ‘g’ cluster revealed upregulated genes linked to exocytosis, autophagy, phagosome acidification, and synaptic vesicle endocytosis (Fig. 4e and Supplementary Fig. 10c).

Fig. 4.

Single-cell transcriptomics reveals that phagocytic gene signatures overlap between QA-stimulated and AD postmortem brain-derived microglia. a An experimental workflow for single-cell transcriptomics in QA-treated microglia. b UMAP (Uniform Manifold Approximation and Projection) analysis showing the distribution of microglial clusters from QA-treated microglia. UMAP was performed using all the cells in the dataset. c Heatmap showing the percentage of cells within each UMAP cluster from QA-treated microglia. d Heatmaps displaying the log2-fold change in the top differentially expressed genes in each cluster of QA-treated microglia. The functional annotation of clusters using KEGG pathways highlights the top 50 DEGs. e Gene Ontology (GO) analysis of the significant biological processes between control and QA-treated microglia, showing the enrichment of pathways related to phagocytosis, autophagy, and neurodegeneration. f Venn diagram showing the overlap of upregulated genes between QA-treated microglia and AD postmortem brain-derived microglia (from Olah et al.‘s dataset). g Canonical correlation analysis (CCA) comparing clusters from QA-activated microglia and an independent human microglial single-cell RNA-seq dataset. Significant correlations are color-coded on the basis of the −log10-transformed p values from the hypergeometric test of overlap between upregulated gene sets in each cluster. h Gene Ontology (GO) analysis of biological processes showing overlap of upregulated genes in QA-treated microglia and AD-derived microglia. i Scatter plots illustrating the enrichment of brain-related disease gene sets (including AD, HD, PD, etc.) from the disease ontology database in specific microglial clusters from QA-treated and AD-derived microglia. j Heatmaps of the top DEGs between QA-activated microglia and AD-derived microglia, with KEGG pathway analysis showing enrichment in phagocytosis, endocytosis, and lipid metabolism pathways. k Heatmaps showing the log2-fold change in the expression of the top differentially expressed genes in both QA-treated and AD-derived microglia, highlighting shared pathways such as phagocytosis, autophagy, and lipid metabolism

To explore whether our in vitro QA-activated microglial gene profile overlaps with that of AD patient-derived microglia, we compared our dataset to an independent human single-cell RNA-seq dataset from AD postmortem brains (reported by Olah et al.). Venn diagram analysis revealed a substantial overlap of upregulated gene signatures between the two datasets (Fig. 4f). Canonical correlation analysis (CCA) demonstrated that 300 upregulated genes from nine QA-treated microglial clusters significantly overlapped with genes from 14 clusters in AD-derived microglia, suggesting potential shared molecular mechanisms (Fig. 4g). Correlations were visualized by color-coded −log10 p values derived from a hypergeometric test. GO analysis of the 300 shared upregulated genes revealed enrichment in biological processes such as autophagy, lipid biosynthesis, and cytoskeleton organization (Fig. 4h). We specifically examined genes involved in phospholipid biosynthesis, including those in the Kennedy pathway, using our scRNA-seq data (Supplementary Fig. 10d). Heatmap analysis revealed that the expression of ETNK1 and EPT1, key enzymes in the Kennedy pathway, was significantly upregulated in QA-treated microglia, whereas the expression of enzymes related to phospholipid interconversion pathways changed little (Supplementary Fig. 10e). In support of this, analysis of publicly available scRNA-seq data (GSE125050) revealed that the expression of enzymes associated with phosphatidylethanolamine (PE) and phosphatidylserine (PS) biosynthesis was significantly altered in microglia from AD patients compared with those from controls (Supplementary Fig. 10f). Additionally, scatter plots revealed that gene sets related to autophagy and neurodegenerative diseases (AD, HD, PD, etc.) from the Disease Ontology database were highly enriched in the QA-induced cluster ‘g’ and in clusters #2 and #5 from AD-derived microglia (Fig. 4i). Heatmaps of the top upregulated genes in QA-activated and AD patient-derived microglia confirmed enrichment in KEGG pathways, including phagocytosis, endocytosis, and lipid metabolism (Fig. 4j). To prioritize candidates, we systematically evaluated these ‘top 15’ disease-relevant genes (Fig. 4k) for mechanistic relevance to our lipidomic findings (Fig. 1). This analysis revealed that GABARAP is uniquely positioned as a core ATG8 protein whose function is directly mediated by lipidation with phosphatidylethanolamine (PE), providing the most direct molecular bridge between our key observations. Notably, other ATG8 family members, including MAP1LC3B (LC3B), were absent from this list, indicating that GABARAP was the strongest candidate for subsequent functional validation.

GABARAP lipidation is required for the QA-induced phagocytosis of Aβ particles in microglia

After confirming that QA increases PE synthesis (Figs. 1 and 2) and identifying GABARAP as a key upregulated, disease-relevant target through our scRNA-seq analysis (Fig. 4), we next sought to determine the functional significance of the upregulation of this gene. We hypothesized that GABARAP is functionally linked to microglial phagocytosis and that this process is modulated by its PE lipidation. To test this hypothesis and further elucidate the dynamics of GABARAP- and LC3-associated phagocytosis, we conducted time-lapse live-cell imaging in microglia transfected with GABARAP-mCherry and LC3B-GFP plasmids. To selectively block GABARAP or LC3 lipidation, we used STBD1 fused with the catalytic domain of RavZ (STBD1-deconjugase) and FYCO1 fused with RavZ (FYCO1-deconjugase), respectively (Fig. 5a and Supplementary Fig. 11a).⁴¹ Under baseline conditions, GABARAP (red) and LC3B (green) were diffusely distributed throughout the cytoplasm. Upon QA stimulation, GABARAP signals prominently accumulated at the peripheral membrane, whereas the distribution of LC3B remained largely unchanged (Fig. 5b). To quantify these changes, we performed contour heatmap analysis using representative images from all four experimental groups (Fig. 5c, d). This, along with time-point analysis, revealed distinct intracellular localization patterns of GABARAP and LC3B (Fig. 5e, f). QA treatment significantly increased the number of GABARAP particles at the cell periphery. The expression of STBD1-deconjugase alone did not markedly alter the distribution of GABARAP. However, the coexpression of STBD1-deconjugase with QA prevented the QA-induced peripheral redistribution of GABARAP, maintaining its localization within the nucleus and cytoplasm, similar to the control and STBD1-deconjugase-only groups. In contrast, the LC3B distribution was less responsive to QA. Although there was a slight increase in peripheral LC3B in response to QA, it was not as prominent as the redistribution of GABARAP. STBD1 deconjugase had no significant effect on LC3B localization, with or without QA treatment.

Fig. 5.

QA signaling drives GABARAP relocalization through PE lipidation in microglial phagocytosis. a A scheme of the experimental flow for the transfection of GABARAP-mCherry, LC3B-GFP, and GABARAP-PE lipidation cutting enzyme (STBD1-deconjugase) plasmids in microglia and time-lapse live-cell imaging. b Representative live-cell images of GABARAP-mCherry particles (red) and LC3B-GFP particles (green) in microglia from four groups (Control, QA, STBD1-deconjugase, and STBD1-deconjugase + QA) at 0, 2, 4, and 6 h. Scale bars (black): 10 μm. c Data processing workflow for contour heatmap generation. d Contour heatmaps showing the distribution of GABARAP (red) and LC3B (green) particles in live-cell images of microglia across four groups (Control, QA, STBD1-deconjugase, and STBD1-deconjugase + QA). The heatmaps display the normalized particle distribution across all the time points. e Distribution of GABARAP particles across all groups of microglia at 0, 2, 4, and 6 h. The localization of GABARAP particles is shown for control (green) and QA-induced microglia (magenta). (Control: 523 particles from 16 cells; QA: 1062 particles from 10 cells; STBD1: 119 particles from 10 cells; STBD1 + QA: 697 particles from 10 cells.). f LC3B particle distribution across all groups of microglia at 0, 2, 4, and 6 h. The localization of LC3B particles is shown for control (green) and QA-induced microglia (magenta). (Control: 102 particles from 12 cells; QA: 158 particles from 10 cells; STBD1: 177 particles from 10 cells; STBD1 + QA: 195 particles from 10 cells.). g Comparative analysis of GABARAP particle speed across all groups of microglia at 0, 2, 4, and 6 h. The speed of GABARAP particles was compared between control (green) and QA-induced microglia (magenta). (Control: 75 particles from 16 cells; QA: 75 particles from 10 cells; STBD1: 107 particles from 10 cells; STBD1 + QA: 107 particles from 10 cells.). h Comparative analysis of LC3B particle speed across all groups of microglia at 0, 2, 4, and 6 h. The speed of LC3B particles was compared between control (green) and QA-induced microglia (magenta). (Control: 72 particles from 12 cells; QA: 72 particles from 10 cells; STBD1: 117 particles from 10 cells; STBD1 + QA: 115 particles from 10 cells.). i A scheme illustrating QA-induced GABARAP-associated phagocytosis (GAP) versus canonical LC3-associated phagocytosis (LAP). Schematic illustration was created with BioRender. The data are presented as the mean ± SEM. Statistical significance was evaluated by one-way ANOVA with Tukey’s post hoc multiple-comparison tests. Significant differences at *p < 0.05, **p < 0.01

We next tracked the movement of GABARAP and LC3B particles over time in all four experimental groups (Fig. 5g, h). GABARAP particle velocity was significantly greater in QA-treated microglia than in control microglia. STBD1-deconjugase alone did not alter this velocity, but its coexpression with QA significantly reduced the QA-induced velocity increase, restoring movement levels to those of control cells. Although QA also increased the LC3B particle velocity, the effect was modest, and STBD1-deconjugase had no significant effect on LC3 dynamics. To further validate the requirement of GABARAP lipidation in QA-induced phagocytosis, we employed FYCO1-deconjugase, another LIR binding protein with affinity for LC3 and GABARAP (Supplementary Fig. 11a–e). FYCO1-deconjugase similarly blocked QA-induced GABARAP translocation to the cell periphery and reduced its velocity, confirming the findings observed with STBD1-deconjugase. Western blot analysis further demonstrated that QA treatment significantly increased the levels of both lipidated GABARAP (GABARAP-PE) and LC3B (LC3B-PE). The expression of either STBD1-deconjugase or FYCO1-deconjugase effectively and significantly attenuated these QA-induced increases in ATG8 lipidation (Supplementary Fig. 11f–k). To determine whether QA-induced phagocytosis relies on canonical autophagy, we examined the effects of pharmacological inhibitors targeting different stages of the pathway (Supplementary Fig. 12a, b). While the blockade of autophagy initiation with 3-methyladenine (3-MA) or the inhibition of lysosomal degradation with chloroquine (CQ) or E64D/pepstatin A partially attenuated the increase in QA activity, microglial phagocytic activity remained significantly elevated compared with that of the control (Supplementary Fig. 12c, d). These results indicate that although canonical autophagy may contribute to basal microglial clearance, the predominant portion of QA-induced phagocytosis occurs through a mechanism distinct from the canonical machinery, primarily requiring GABARAP lipidation and redistribution (Fig. 5i).

QA upregulates GABARAP expression and lipidation in microglia and AD models, promoting Aβ oligomer phagocytosis

To further confirm whether GABARAP is involved in the QA-induced phagocytosis of Aβ oligomers in microglia, we investigated which ATG8 family genes (GABARAP or LC3) are preferentially regulated by QA in vitro. qPCR analysis revealed that QA significantly increased GABARAP mRNA levels in microglia at early time points (30 min) and maintained this elevation over a 12-h period. In contrast, QA increased MAP1LC3B/LC3B mRNA expression only between 6 and 12 h (Fig. 6a). Next, we performed Western blot analysis to quantify GABARAP and LC3B protein levels in microglia with or without QA treatment (Fig. 6b). Notably, the level of GABARAP protein was significantly increased in the soluble fraction of QA-treated microglia. Furthermore, the lipidated form of GABARAP (GABARAP-PE) was significantly elevated following QA treatment. Treatment with meclizine, an inhibitor of the Kennedy pathway, reduced GABARAP-PE levels, indicating the involvement of this pathway. Consistent with the mRNA data, LC3B protein levels also significantly increased in the soluble fraction after QA treatment (Fig. 6c–f). To determine whether GABARAP expression is also upregulated in the AD brain, we performed immunostaining of postmortem frontal cortex tissue from AD patients (Fig. 6g). As expected, GABARAP immunoreactivity was elevated in microglia from AD brains compared with those from control subjects (Fig. 6h, i). Since compared with other inflammatory stimuli, QA induced the most significant increase in GABARAP expression (Supplementary Fig. 13a–c), we further validated its effect in vivo using an AD (5xFAD) mouse model. GABARAP levels were elevated in microglia from AD (5xFAD) mice compared with those from wild-type (WT) controls. Notably, GABARAP immunoreactivity was significantly greater in QA-injected microglia than in vehicle-treated control microglia (Fig. 6j, k). To further validate the role of the Kennedy pathway in QA-induced GABARAP-associated phagocytosis, we examined the effects of meclizine on GABARAP localization and intensity in microglia from AD (5xFAD) mice. As expected, meclizine reduced GABARAP immunoreactivity and altered its cellular distribution in microglia (Fig. 6j, k and Supplementary Fig. 13d, e). These findings were further corroborated in primary microglial cultures, in which QA increased the number of GABARAP puncta and promoted their peripheral redistribution in a meclizine-dependent manner (Supplementary Fig. 13f–i). Collectively, these in vitro and in vivo findings indicate that the Kennedy pathway directly contributes to QA-induced, GABARAP-mediated phagocytosis in microglia.

Fig. 6.

GABARAP expression is modulated by QA signaling in microglia in vitro and is elevated in the postmortem brains of AD patients and AD transgenic (5xFAD) mice. a GABARAP mRNA expression significantly increased in a time-dependent manner following QA treatment (n = 8 independent experiments per group). b LC3B mRNA expression significantly increased in a time-dependent manner following QA treatment (eight independent experiments per group). c A scheme of experimental flow for Western blot analysis following QA treatment in cells. d Representative Western blot images showing the protein levels of GABARAP, GABARAP-PE, LC3B and LC3B-PE after treatment with QA in the presence or absence of meclizine. Short and Long indicate short- and long-exposure images of the GABARAP immunoblot. e The protein level of the GABARAP-PE lipid form significantly increased following QA treatment and was effectively blocked by cotreatment with meclizine (3 independent experiments per group). f The protein level of the LC3B-PE lipid form significantly increased following QA treatment and was effectively blocked by cotreatment with meclizine (3 independent experiments per group). g Representative low-magnification images of GABARAP immunoreactivity in hippocampal and entorhinal cortex regions from normal subjects and AD patients. Dashed lines indicate the hippocampal subfields and adjacent regions, including CA1, CA2, CA3, CA4, DG, and Ent Cx. h Representative images of IBA1 (microglial marker, red) and GABARAP (green) immunoreactivity in the cortex of AD patients and normal subjects. Scale bars (white): 50 μm, 5 μm. i Compared with that in normal subjects, the intensity of GABARAP in IBA1-positive microglia in the brains of AD patients was significantly greater. A total of 30 cells were analyzed (10 cells per case; N = 3 cases per group). j Representative images of IBA1 (microglial marker, green) and GABARAP (red) immunofluorescence staining in the hippocampus of 5xFAD mice injected with saline, QA, QA + Mec, or Mec. Scale bars (white): 5 μm. k Compared with that in the saline group, the GABARAP intensity in the QA-injected microglia in the 5xFAD group significantly increased, but the addition of meclizine (QA + Mec group) or meclizine alone (Mec group) significantly decreased the GABARAP intensity. A total of 50 cells were analyzed (10 cells per mouse; N = 5 mice per group). The data are presented as the mean ± SEM. Statistical significance was evaluated by unpaired two-tailed Student’s t tests and one-way ANOVA with Tukey’s post hoc multiple-comparison tests. Significant differences at *p < 0.05 and **p < 0.01 are indicated

QA-induced microglial activation reduces Aβ plaque formation via GAP in AD mice

To verify the critical role of GABARAP in the clearance of Aβ by microglia, we investigated the effect of GABARAP knockdown on Aβ oligomer phagocytosis in vitro using time-lapse live-cell imaging (Fig. 7a and Supplementary Fig. 14a–g). Phagocytosis of Aβ oligomers was visualized using the pHrodo-Aβ signal (red) in still-frame images from control, QA-treated control, shGABARAP-only, and QA-treated shGABARAP microglia at various time points (0 h, 2 h, 4 h, and 6 h) (Fig. 7b). Quantitative analyses, including cumulative frequency distribution and average intensity measurements, revealed that GABARAP knockdown not only reduced the baseline phagocytic activity (as indicated by the pHrodo-Aβ signal) but also markedly diminished the QA-induced increase in phagocytosis in microglia (Fig. 7c). Overall, the average intensity of the pHrodo-Aβ signal was significantly reduced by GABARAP knockdown, both in the presence and absence of QA (Fig. 7d). These in vitro findings support the conclusion that QA-induced microglial phagocytosis is mediated by GABARAP. Accordingly, this process is termed GABARAP-associated phagocytosis (GAP). To determine whether GAP contributes to Aβ clearance in vivo, we delivered AAV-shControl and AAV-shGabarap into the cortex and dorsal hippocampus of AD (APP/PS1) mice and evaluated pathological outcomes (Fig. 7e). To validate GABARAP knockdown, we examined GABARAP expression in human (HMC3) and mouse (BV2) microglial cell lines. qPCR analysis revealed that compared with control shRNA, GABARAP shRNA significantly reduced GABARAP mRNA levels in both cell types (Supplementary Fig. 14b, c). Western blot analysis confirmed a corresponding reduction in GABARAP protein levels in HMC3 cells (Supplementary Fig. 14e). The knockdown efficiency was further verified by a decrease in the mCherry signal in GABARAP–mCherry–expressing cells after shGABARAP transfection. In vivo, AAV-shGABARAP injection into the hippocampus of 5xFAD mice markedly reduced GABARAP expression, as shown by immunohistochemical analysis (Supplementary Fig. 14f, g), confirming efficient knockdown both in vitro and in vivo. As expected, compared with control virus, GABARAP knockdown resulted in increased Aβ plaque size in both the cortex and the dorsal hippocampus. Additionally, GABARAP knockdown altered microglial morphology, leading to pronounced shrinkage of microglia in both brain regions of AD (APP/PS1) mice (Fig. 7f–i and Supplementary Fig. 14h–k). Finally, to assess whether GABARAP is essential for the beneficial effects of QA on Aβ pathology in vivo, we coinjected QA with either AAV-shControl or AAV-shGABARAP into the hippocampus of AD (APP/PS1) mice and examined pathological changes, particularly alterations in Aβ plaques (Fig. 7j). Notably, confocal microscopy and quantitative image analysis revealed that compared with control virus, GABARAP knockdown significantly increased Aβ plaque size (Fig. 7k). These findings strongly support the conclusion that GABARAP-mediated microglial phagocytosis is essential for QA-induced Aβ clearance in AD models (Fig. 7l).

Fig. 7.

GABARAP knockdown attenuates QA-induced Aβ clearance in microglia and AD mice. a A scheme of the experimental flow for the transfection of shGABARAP for live-cell imaging. b Representative images of control microglia, QA-induced microglia, control microglia with GABARAP knockdown, and QA-induced microglia with GABARAP knockdown at different time points (0, 2, 4, and 6 h). Scale bars (black): 10 μm. c Cumulative frequency distribution of pHrodo-Aβ particle intensity in the four groups of microglia. (Control: 360 particles from 10 cells; QA: 323 particles from 10 cells; shGABARAP: 288 particles from 10 cells; shGABARAP + QA: 288 particles from 10 cells. d Violin plots illustrating the average intensity of pHrodo-Aβ particles in the four groups of microglia. (Control: 360 particles from 10 cells; QA: 323 particles from 10 cells; shGABARAP: 288 particles from 10 cells; shGABARAP + QA: 288 particles from 10 cells). e Experimental workflow for the stereotactic injection of AAV-shControl and AAV-shGabarap viruses into the dorsal hippocampus of APP/PS1 mice. f Representative immunofluorescence images showing Aβ plaques and microglia in the hippocampi of APP/PS1 mice with shControl or shGabarap knockdown. Scale bars (white): 20 μm. g Compared with that in the AAV-shControl-injected APP/PS1 group, the Aβ plaque volume in the AAV-shGabarap-injected APP/PS1 group significantly increased. A total of 75 plaques were analyzed (15 plaques per mouse; N = 5 mice per group). h Microglial branch points were significantly lower in the AAV-shGabarap-injected APP/PS1 group than in the AAV-shControl-injected APP/PS1 group. A total of 50 cells were analyzed (10 cells per mouse; N = 5 mice per group). i Compared with that in the AAV-shControl-injected APP/PS1 group, the perimeter of microglial bodies in the AAV-shGabarap-injected APP/PS1 group significantly decreased. A total of 50 cells were analyzed (10 cells per mouse; N = 5 mice per group). j Representative immunofluorescence images showing Aβ plaques in the hippocampi of APP/PS1 mice injected with QA along with either AAV-shControl or AAV-shGabarap virus. Scale bars (white): 100 μm. k Aβ plaque size was significantly greater in APP/PS1 mice injected with AAV-shGabarap + QA than in those injected with AAV-shControl + QA. A total of 75 plaques were analyzed (15 plaques per mouse; N = 5 mice per group). l Graphical summary illustrating a novel signal transduction pathway in which QA triggers the PE synthesis pathway and GABARAP lipidation and facilitates GAP for Aβ clearance in AD. Schematic illustration was created with BioRender. The data are presented as the mean ± SEM. Statistical significance was evaluated by unpaired two-tailed Student’s t tests and one-way ANOVA with Tukey’s post hoc multiple-comparison tests. Significant differences at *p < 0.05 and **p < 0.01 are indicated

Discussion

QA is an endogenous metabolite of the kynurenine pathway that has been recognized as an excitotoxin contributing to neuronal function in AD.^13–16 However, the potential role of QA in modulating microglial function, particularly at the cellular and molecular levels, remains incompletely understood. Understanding microglial metabolic responses in the context of brain aging and neurodegeneration is crucial.³⁶ Given the highly dynamic nature of microglial responses, metabolic signals such as QA may act as rapid regulators of microglial functional states in the AD brain. Previous studies have focused primarily on the neurotoxic effects of QA, including NMDA receptor–mediated excitotoxicity, oxidative stress, and tau pathology, reinforcing a largely neuron-centric view in which elevated QA levels are considered uniformly detrimental.^13–16 In contrast, emerging evidence suggests that endogenous metabolites may exert cell type–specific effects within the brain microenvironment.⁴² Consistent with this perspective, our findings reveal a previously unrecognized adaptive role for QA in microglia.⁴³ Specifically, we identify a novel mechanism by which QA modulates de novo PE synthesis via the Kennedy pathway and enhances the GAP of Aβ oligomers. While PE is known to play an essential role in membrane dynamics and phagocytosis, its regulatory mechanisms during AD pathogenesis are still poorly defined.^31–34 Our work addresses this gap by identifying QA as an upstream metabolic signal that selectively activates PE synthesis in microglia, thereby linking extracellular metabolic cues to membrane remodeling and phagocytic function.

PE is synthesized through the Kennedy pathway.^31,44–50 Importantly, we found that QA enhances PE production by upregulating ETNK1 and EPT1 expression in microglia. ETNK1 initiates this pathway, whereas EPT1 catalyzes the final step by converting CDP-PE to PE.^44–47 The resulting enrichment of PEs with long-chain unsaturated fatty acids promotes membrane curvature and facilitates the formation of the hexagonal II (HII) phase, thereby supporting membrane fusion and phagosomal closure.^31,48–50 In contrast, alternative PE-generating routes, including those mediated by PISD or PTDSS, were not affected, highlighting the specificity of the Kennedy pathway.^51–54 Consistent with this, elevated expression of Kennedy pathway–associated enzymes was also observed in microglia from postmortem AD brains.³⁵ Pharmacological interventions further corroborated these findings: inhibition of PCYT2 by meclizine reduced PE biosynthesis and phagocytic activity, whereas PEMT inhibition by bezafibrate increased PE levels and enhanced microglial phagocytosis.^36–40 In vivo, acute QA administration in 5xFAD mice promoted microglial engagement with Aβ plaques and was accompanied by a reduction in plaque burden. Together, these results indicate that metabolic cues such as QA can rapidly modulate microglial–plaque interactions via the Kennedy pathway–GABARAP axis.^55–57

In the current study, we show that QA upregulates GABARAP expression and facilitates the GAP of Aβ oligomers in microglia. Notably, QA enhanced microglial uptake and lysosomal processing of Aβ oligomers under Aβ-containing conditions, indicating an additive effect alongside Aβ-induced responses. Although ATG8 family proteins have been studied in the context of canonical autophagy and LC3-associated phagocytosis (LAP),^58–60 their specialized roles in microglia remain unclear.⁶¹ GABARAP lipidation with PE is essential for its association with autophagic membranes.^43,61–64 Using the GABARAP-PE lipidation-specific fusion enzyme STBD1-deconjugase, we show that GABARAP lipidation is required for QA-induced phagosome dynamics and GAP. Interestingly, the expression of deconjugases led to compensatory upregulation of the ATG conjugation machinery (e.g., ATG7 and ATG3) in HMC3 cells to maintain proteostatic homeostasis.⁴¹ On the basis of our findings, we propose that GABARAP lipidation contributes to early phagophore structures that subsequently mature into autophagosomes through fusion with LC3-positive compartments.^65,66 Our data further reveal a clear spatial and functional distinction between GABARAP- and LC3B-associated phagosomes in microglia.⁶⁶ GABARAP-positive structures preferentially localize to the plasma membrane and exhibit dynamic movement in response to QA, whereas LC3B-associated phagosomes are largely perinuclear and relatively static. These findings suggest that QA preferentially engages a GABARAP-associated phagocytosis (GAP) pathway rather than classical LAP. Consistently, the redistribution of GABARAP and the relative insensitivity of the QA-induced response to 3-MA indicate that GAP is the primary mechanism driving enhanced amyloid clearance in this system.

Conventional wisdom holds that elevated QA levels under inflammatory conditions contribute to neuronal dysfunction through excitotoxic mechanisms.^13–16 Our findings refine this view by demonstrating that QA exerts cell type–specific and context-dependent effects. While chronic or high concentrations are neurotoxic, acute, low-dose QA acts as a signaling metabolite that reprograms microglial metabolism, enhancing Aβ clearance and reducing pathological p-Tau levels. These beneficial effects are supported by preserved synaptic integrity and improved behavior in AD mice and are consistent with the concept of hormesis, in which physiological concentrations of a stress-associated metabolite elicit adaptive responses.⁶⁷ Our data further suggest that previously reported neurotoxic outcomes may largely reflect differences in dosage and exposure duration rather than an intrinsically deleterious property of QA. In vivo, QA induced amoeboid microglial morphology and reduced the plaque burden, effects that were abolished by GABARAP depletion.²⁸ These observations indicate that selectively targeting downstream pathways—such as the Kennedy pathway–GABARAP axis—may enable the beneficial aspects of QA-driven Aβ clearance while minimizing the risk of excitotoxicity. Although our experimental models capture key features of AD, they do not fully recapitulate the chronic and multifactorial nature of human disease, including progressive tau pathology or neuronal loss. Future studies will therefore be needed to evaluate GABARAP lipidation dynamics and autophagic flux under prolonged neurodegenerative conditions.

In conclusion, we identified a previously unrecognized microglia-specific signaling axis in which QA activates Kennedy pathway–dependent PE biosynthesis, enhances GABARAP lipidation, and promotes efficient Aβ phagocytosis. This work reveals a direct mechanistic link between metabolic signaling and the phagocytic machinery in microglia, redefining QA as a context-dependent regulator of microglial function in the AD microenvironment. This metabolic adaptation may play important roles in maintaining cellular homeostasis and mitigating AD-related neuropathology. Our findings suggest that targeting microglial lipid metabolism and GABARAP-dependent phagocytic pathways represents a promising therapeutic strategy for AD.

Methods

Postmortem human brain samples

Neuropathological examination of postmortem brain samples from normal subjects and severe AD patients was performed using procedures previously established by the Boston University Alzheimer’s Disease Center (BUADC).^68,69 The next of kin provided informed consent for participation and brain donation. This study was reviewed by the Institutional Review Board of the Boston University School of Medicine (Protocol H 28974) and was approved for exemption because it only included tissues collected from post-mortem subjects not classified as human subjects. The study was performed in accordance with institutional regulatory guidelines and principles of human subject protection in the Declaration of Helsinki. The sample information is listed in Supplementary Table 1.

Animals

Wild-type APP/PS1 mice on a C57BL/6 J background and 5xFAD mice on a C57BL/6 x SJL background were used as animal models of AD. APP/PS1 (stock no. 034832) and 5xFAD (stock no. 005359) mice were obtained from Jackson Laboratory. For the behavioral experiments, 6–10-month-old male mice were used. All the animals were maintained in a vivarium with a light/dark cycle (8:00 AM ~ 8:00 PM). Animal care and handling were performed according to the directives of the Animal Care and Use Committee and institutional guidelines of KIST (Seoul, Korea). Animals were randomly used for experiments.

Human cell lines

HMC3 cells, a human microglial cell line,³⁰ were purchased from ATCC (CRL-3304; ATCC, Manassas, VA, USA). In accordance with the ATCC guidelines, HMC3 is a validated model that retains key characteristics of primary human microglia, making it an appropriate system for analyzing neuroinflammation and biochemical functions. Cells were cultured at 37 °C in Dulbecco’s modified Eagle’s medium (DMEM; Gibco, Grand Island, NY, USA) supplemented with 10% fetal bovine serum (FBS; HyClone, Logan, UT, USA) and 1% P/S (100 μg/ml penicillin plus 100 μg/ml streptomycin; Gibco, Grand Island, NY, USA). The human neuroblastoma cell line SH-SY5Y (ATCC® CRL-2266™) was purchased from ATCC (Manassas, VA, USA). The cells were cultured at 37 °C in DMEM/F12 (1:1) (Gibco) supplemented with 10% fetal bovine serum (FBS; HyClone) and 1% P/S (Gibco).

Primary microglial culture

Primary microglia were isolated from the cerebral cortex of neonatal C57BL/6 mice (P0–P2). Briefly, the meninges were removed, and the cortical tissues were enzymatically dissociated using Trypsin-based manual dissociation. Dissociated cells were cultured in T75 flasks in DMEM/F12 supplemented with 10% FBS, 1% P/S, and 10 ng/mL recombinant mouse GM-CSF. After 10–14 days, the flasks were shaken at 200 rpm for 3 h to detach the microglia. The collected microglia were seeded onto plates at a density of 1 × 10⁵ cells/cm² and maintained in microglia-specific medium. The purity of the microglial cultures was confirmed to be >95% by IBA1 immunostaining.

Preparation of QA solutions

Quinolinic acid (QA) (Sigma‒Aldrich, St. Louis, MO, USA; purity ≥98%) was dissolved in phosphate-buffered saline (PBS, pH 7.4) to prepare a stock solution of 10 mM. This stock solution was stored in aliquots at −20 °C. Working solutions of QA were freshly prepared on the day of each experiment by diluting the stock solution in the appropriate experimental medium (cell culture medium for in vitro studies or saline for in vivo injections) to the final desired concentrations, typically 1 µM. The vehicle controls consisted of the experimental medium containing an equivalent concentration of PBS.

Aβ (1-42) oligomer preparation

Human synthetic Aβ(1-42) peptides (GenicBio, Shanghai, China; Cat. No. A-42-T-10) were dissolved in 100% 1,1,1,3,3,3-hexafluoro-2-propanol (HFIP) to monomerize the peptide, which was aliquoted and concentrated under vacuum. The dried peptide films were resuspended in anhydrous dimethyl sulfoxide (DMSO) to a concentration of 5 mM and diluted in sterile PBS to a final concentration of 100 µM. The solution was incubated at 4 °C for 24 h to promote oligomer formation. Oligomerization was validated by TEM, which revealed globular Aβ aggegates with diameters of 5–20 nm (Supplementary Fig. 15). The repared oligomers were used immediately or stored at −80 °C for short-term use. The oligomer preparation protocol and validation approach used here are consistent with widely adopted methods for generating biologically active Aβ oligomers.⁷⁰

Preparation of pHrodo Red–labeled Aβ

Human synthetic Aβ(1-42) peptide (GenicBio, Shanghai, China; Cat. No. A-42-T-10) was aggregated and labeled with pHrodo Red succinimidyl ester (pHrodo Red SE) (Invitrogen, Thermo Fisher Scientific, Waltham, MA, USA; Cat. No. P36600) according to the protocol provided by the vendor (Fujifilm Cellular Dynamics, Inc., Madison, WI, USA). Lyophilized Aβ was dissolved in 10 mM NaOH and diluted with HPLC-grade water, followed by neutralization with 10× TBS (pH 7.4) to obtain a final concentration of 100 μM. The peptide solution was incubated at 37 °C overnight to induce aggregation. For pHrodo Red SE labeling, aggregated Aβ was collected by centrifugation (16,000 × g, 1 min), washed with Hank’s balanced salt solution (HBSS), and resuspended in 0.1 M sodium bicarbonate buffer. pHrodo Red SE dissolved in DMSO was added to the suspension, and the reaction mixture was incubated for 2 h at room temperature in the dark to allow pHrodo labeling. Following labeling, the aggregates were pelleted by centrifugation and washed with methanol, followed by repeated washes with HBSS to remove excess dye. The labeled aggregates were then resuspended in HBSS and briefly sonicated. The pHrodo Red–labeled Aβ aggregates were used immediately or stored at −80 °C until further use.

Plasmids and pharmacological treatments

To selectively inhibit GABARAP or LC3 lipidation, cells were transfected with plasmids encoding STBD1-RavZ (GABARAP-specific deconjugase) or FYCO1-RavZ (GABARAP and LC3 deconjugase) using Lipofectamine 3000 (Invitrogen) according to the manufacturer’s instructions. These deconjugase constructs, which fuse the catalytic domain of the bacterial protease RavZ to the LIR domains of STBD1 or FYCO1, were utilized as previously described by Park et al.⁴¹ to specifically cleave ATG8 family proteins from the membrane. To pharmacologically modulate autophagic and metabolic pathways, various inhibitors have been applied at specific stages. Specifically, 3-methyladenine (3-MA, 5 mM; Sigma‒Aldrich) was added to the culture medium 1 h prior to QA stimulation to inhibit class III PI3K-dependent canonical autophagy. To assess autophagic flux and lysosomal function, cells were treated with chloroquine (CQ, 20 μM; Sigma‒Aldrich) to inhibit lysosomal acidification or a combination of pepstatin A and E64d (PEPA/E64d, 10 μg/mL each; Sigma‒Aldrich) to block lysosomal acid proteases. Additionally, meclizine (5 μM; Sigma‒Aldrich) and bezafibrate (100 μM; Sigma‒Aldrich) were used to modulate the Kennedy and PEMT pathways, respectively.

shRNA transfection

HMC3 cells were seeded at 70% confluency and transfected with shRNA plasmids (shETNK1, shEPT1, or shControl; 2.5 µg) using Lipofectamine 3000 (Thermo Fisher Scientific) according to the manufacturer’s protocol. Cells were incubated for 48 h before the QA treatment

AAV construction and injection

To generate the shGabarap virus, an AAV packaging plasmid expressing Rep2 and the AAV-MG1.1 capsid was used. AAV vectors were generated using a standard triple-plasmid transfection system in HEK293T cells. The packaging plasmid expressing Rep2 and the AAV-MG1.1 capsid was obtained from Addgene (Watertown, MA, USA), together with the pAAV transfer vector and adenoviral helper plasmid. shRNA expression was driven by a human U6 promoter. These AAV-MG variants are highly efficient at delivering genetic payloads into microglia, facilitating in vivo transgene delivery without inducing microglial immune activation.⁷¹ Viral titers were determined by qPCR analysis of the viral genome and were typically 2 × 10¹² vg/ml. For gene delivery, shGabarap and shControl viruses were stereotactically injected into the dentate gyrus (DG) and cortex of the dorsal hippocampus of mice (coordinates: AP − 1.8 mm, ML ± 1.5 mm, DV − 1.7 mm) using a stereotaxic apparatus. A total of 1–2 μL of virus was injected at a rate of 0.1 μL/min, and the animals were allowed to recover for 2 weeks to allow sufficient time for gene knockdown. Additionally, QA was injected into the same regions to assess its effect on microglial function.

Lipidomic sample preparation

The modified Folch method was used to extract lipids from cells.⁷² Briefly, cell pellets were obtained after the cells were washed with cold PBS. Before extraction, 10 µL of the PBS suspension was removed for DNA measurement. Afterward, 500 µL of ice-cold chloroform:methanol (2:1, v/v) extraction solvent containing 4 µg/mL arachidonic acid-d8 and 4 µg/mL 16:0-d31-18:1 phosphocholine as an internal standard (IS) was added to the pellet, which was subsequently vortexed for 10 min at room temperature. The sample was subsequently washed with 160 µL of distilled water and vortexed for a few seconds. The lower (organic) phase was obtained by centrifugation at 800 × g for 5 min at 4 °C. After the lower phase was transferred to a new centrifuge tube, the upper phase was re-extracted twice. Finally, the obtained lower phase was centrifuged at 2000 × g for 5 min at 4 °C. Three hundred microliters of the extracted sample was evaporated under nitrogen at 37 °C and reconstituted with mobile phase B. Finally, 5 µL of each sample was injected. Pooled QC samples were prepared to ensure accuracy and performance.

LC‒MS/MS lipidomics analysis

Lipids were analyzed on a Vanquish Flex UHPLC system coupled to an Orbitrap Exploris™ 120 mass spectrometer (Thermo Fisher Scientific, San Jose, CA, USA). Lipid metabolites were separated on an ACQUITY UPLCⓇ BEH C18 column (2.1 × 100 mm, 1.7 µm; Waters, Milford, MA, USA) at 45 °C. Mobile phases A and B consisted of 40% acetonitrile and isopropanol:acetonitrile (9:1, v/v), respectively. Each mobile phase contained 2 mM ammonium formate and 0.1% formic acid (v/v). The flow rate was 0.35 mL/min. The elution gradient was maintained at 90% of mobile phase A from 0 to 1 min. It was gradually decreased from 90 to 30% over 5 min. Then, mobile phase A was reduced to 10% over 6 min. After it reached 0% over 1 min, it was returned to the initial condition (90% mobile phase A) for 0.5 min, and the conditions were maintained for 2.5 min to re-equilibrate. Thus, the total running time was 16 min. Mass spectrometry data were obtained via data-dependent acquisition (DDA). The ionization of metabolites was achieved using an electrospray ionization (ESI) source in positive and negative modes. The mass range was analyzed as 80–1200 m/z in both ionization modes. MS2 fragmentation was performed using high-energy collisional dissociation (HCD). In addition, the detailed mass spectrometry parameters were as follows: sheath gas flow, 50 arb; auxiliary gas flow, 10 arb; sweep gas flow, 1 arb; ion transfer tube temperature, 325 °C; vaporizer temperature, 350 °C; and S-lens RF level, 70%.

Lipid identification and statistical analysis

For lipidomic data analysis, chromatograms were acquired using X-calibur 4.3 and analyzed with Compound Discoverer 3.3 (Thermo Fisher Scientific). SIMCA 18.0 (Sartorius AG, Göttingen, Germany) was used for multivariate data analysis. LC‒MS/MS data were normalized to the DNA concentration and internal standard (IS) area. Normalized data were used to generate lipid profiling figures, including principal component analysis (PCA) plots. Lipids were identified using our in-house library and mzVault database, including LipidBlast and MassBank of North America (MONA). The lipid identification data for shEPT1, shETNK1, meclizine, and bezafibrate are listed in Supplementary Tables 2, 3, and 4, respectively. Statistical analysis of the data was performed using Student’s t test. The fold change was calculated as the mean value of the normalized data for each lipid.

XTT assay

Cell viability was assessed using the Cell Proliferation Kit II (XTT) (Cat# 11465015001, MERK), which measures mitochondrial metabolic activity on the basis of the reduction of XTT tetrazolium salt to a water-soluble formazan dye. Cells were seeded into 96-well plates at a density of 1 × 10⁴ cells per well in 100 μL of medium and incubated overnight to allow stabilization. The cells were then treated with 0, 0.125, 0.25, 0.5, 1, or 2 μM QA for 6 h or 12 h. Following treatment, 50 μL of activated XTT reagent (XTT labeling reagent mixed with electron coupling reagent) was added to each well, following the manufacturer’s instructions. The plates were incubated for 4 h at 37 °C in the dark, after which the absorbance was measured at 450 nm using a microplate reader. Cell viability was calculated by normalizing the absorbance of the treated wells to that of the control group. Concentration‒response curves were generated using GraphPad Prism 8.0 (GraphPad Software). Each condition was measured in four biological replicates.

Calcein-AM assay

The cell viability of human microglia was assessed using a calcein-AM assay (Cat# C3099; Thermo Fisher Scientific), which measures esterase activity and membrane integrity through the enzymatic conversion of nonfluorescent calcein-AM to the fluorescent dye calcein in live cells. Cells were seeded into 96-well black-walled, clear-bottom plates at a density of 1 × 10⁴ cells per well in 100 μL of complete medium and allowed to stabilize overnight. After 50 μL of the culture medium from each well was removed, 50 μL of treatment solution containing QA was added at varying concentrations and incubated for different durations, according to the experimental design. After 24 h of incubation at 37 °C in a humidified incubator with 5% CO₂, the medium was removed, and the cells were gently washed once with phosphate-buffered saline (PBS). Afterward, 100 μL of 2 μM calcein-AM working solution was added to each well, and the plates were incubated for 30 min at 37 °C in the dark. The fluorescence was measured using a microplate reader (excitation/emission: 485/530 nm). Time- and concentration-dependent viability curves were generated using GraphPad Prism 8.0 (GraphPad Software). Each condition was tested in six technical replicates and repeated in three independent biological experiments.

Annexin-V assay

Apoptotic cell death was assessed using the Annexin V-CF488A conjugate from the Apoptotic and Necrotic Cell Quantification Kit (#30065; Biotium, USA). Human microglia (HMC3) were seeded in 8-well chamber slides and treated with increasing concentrations of QA for 12 h. The cells were stained with Annexin V working solution for 15 min at room temperature in the dark, following the manufacturer’s instructions. DAPI was used to visualize the nuclei, and the total number of cells was quantified. Live-cell imaging was conducted using the Image ExFluorer system (Live Cell Instrument, Korea), and the number of Annexin V–positive and DAPI-stained nuclei was manually counted using ImageJ software (Fiji; NIH, USA), and the percentage of apoptotic cells was calculated as the number of Annexin V–positive cells relative to the total number of cells per field of view.

CellTiter-Glo (CTG) assay

The cell viability of human microglia and human neurons was assessed using the CellTiter-Glo (CTG) Luminescent Cell Viability Assay (Cat# G7572; Promega), which quantifies intracellular ATP as an indicator of metabolically active cells through a luciferase-catalyzed luminescent reaction. Cells were seeded into 96-well plates at a density of 1 × 10⁴ or 5 × 10⁴ cells per well in 100 μL of complete medium and allowed to stabilize overnight. After 50 μL of the culture medium from each well was removed, 50 μL of treatment solution containing QA was added at varying concentrations and incubated for different durations, according to the experimental design. After 12 h of incubation at 37 °C in a humidified incubator with 5% CO₂, 80 μL of CTG reagent was added directly to each well. The plates were gently agitated for 2 min to induce cell lysis and incubated for 10 min at room temperature to allow the luminescent signal to stabilize. Luminescence was measured using a microplate reader. Time‒response curves were generated using GraphPad Prism 8.0 (GraphPad Software). Each condition was tested in six technical replicates and repeated in three independent biological experiments.

Live cell sample preparation

After the cells were treated with quinolinic acid and Aβ (1-42), they were recorded for 24 h at 15 min intervals using a live-cell time-lapse imaging system with a fully automated live-cell instrument and a heated incubator to 37 °C and 5% CO₂ (10× magnification; Live Cell Instrument, Avon, Nowon, Korea). To achieve accurate cell tracking, the maximum temporal resolution of acquisition was set to 1 frame per second.

AI-based microglial motility and GABARAP trafficking

To perform quantitative analysis of microglial dynamics under different experimental conditions, high-resolution live-cell images were recorded at 40x magnification using live microscopy. For whole-cell motility analysis, time-lapse images of HMC3 cells were acquired every 10 min for 24 h. The centroid of individual, nonoverlapping cells was tracked over time using the MTrackJ plugin in ImageJ (NIH). The average migration speed was calculated as the total path length divided by the observation time. In addition, we captured multiple images (>45) of microglia at a fixed time interval of 10 min. We recorded videos of microglia for more than 7 h at a rate of 6 images per h. In each image, different numbers of microglia were recorded. Some of them were either overlapping or touching neighboring microglia, and others were at significant distances. Hence, we categorized cell types into two classes: overlapping cells and nonoverlapping cells. It was challenging to ensure automatically that the subcellular vesicle was either from cells under observation or from neighboring cells. Overlapping cells were analyzed through IMARIS software, and nonoverlapping cells were analyzed through a Python script. To ensure the accuracy and reliability of the automated tracking, manual particle tracking was performed using the MTrackJ plugin in ImageJ (NIH).

In the current study, four subcellular vesicles were considered, i.e., GABARAP, LC3, Aβ, and zymosan. We calculated two parameters, i.e., the intensity and size of Aβ and zymosan in microglia under different experimental conditions. However, in the case of GABARAP and LC3, these two parameters were not appropriate for comparative analysis of microglia under different experimental conditions. Therefore, we determined the velocity and localization of GABARAP and LC3 in microglia. These two parameters indicate differences in the motility of GABARAP and LC3 under different conditions.

We manually cropped the region of interest (ROI) around a single cell prior to analyzing vesicle activity. This step reduced the processing time and chances of false positives as well. Afterward, we tracked the subcellular vesicles in overlapping microglia through IMARIS and manually removed the subcellular vesicles of neighboring microglia. For nonoverlapping microglia, we developed an algorithm in Python using a built-in library, PyTracking, for the automatic tracking of subcellular vesicles. The tracking results from the Python-based automated system were cross-validated with manual measurements obtained via MTrackJ to ensure consistency in the velocity calculations. Afterward, we recorded the intensity, size, and coordinates of all the tracked vesicles. The intensity and size of the Aβ signal were recorded to determine the aggregation and uptake of Aβ and zymosan in microglia. The coordinates of GABARAP and LC3 were recorded to calculate the velocity and distribution of particles in microglia. We designed a contour map to represent the distribution of subcellular vesicles in microglia under different experimental conditions. For this purpose, each cell was divided into four regions. The outermost region represents the plasma membrane, and the central region represents the nucleus of microglia. We calculated the percentage of subcellular vesicles in each region, which are presented in contour maps. Moreover, we also represented the distribution of subcellular vesicles in each region at different time frames through 3D line graphs. Similarly, the velocities of the subcellular vesicles at different time frames are presented through 3D line graphs.

DAB staining and light microscopy

The brain tissue sections were blocked using 5% BSA supplemented with 0.5% Triton X-100 in PBS for 1 h at room temperature. The brain sections were incubated with primary antibodies at 4 °C overnight. After being washed three times, the slides were incubated in 1% H₂O₂ in PBS for 15 min to quench endogenous peroxidase activity. Following quenching, the slides were processed with a Vector ABC Kit (Vector Laboratories, Inc., Burlingame, CA, USA). The immunoreactive signals were developed with DAB chromogen (Thermo Fisher Scientific, Meridian, Rockford, IL, USA) and analyzed under a bright field microscope.

Immunofluorescence staining and confocal microscopy

The animals were deeply anesthetized using 2% avertin and perfused with 0.1 M PBS, followed by 4% paraformaldehyde. The brains were postfixed in 4% paraformaldehyde at 4 °C for 24 h and 30% sucrose at 4 °C for 48 h. The brains were then cut into 30-µm coronal cryosections. The sections were blocked in 0.1 M PBS containing 0.3% Triton X-100 (Sigma) and 2% donkey serum (GeneTex) for 30 min at room temperature. The primary antibodies used were as follows: rabbit anti-Iba1 (1:1000; 019-19741; Wako), mouse anti-Iba1 (1:1000; 016-26721; Wako), mouse anti-beta amyloid (1:200; 803001; BioLegend), mouse anti-EPT1 (1:200; NBP3-17536; Novus), and mouse anti-GABARAP (1:200; M135-3; MBL). The brain samples were incubated overnight at 4 °C. Afterward, the sections were washed three times in 0.1 M PBS and incubated with the appropriate secondary antibodies from the Jackson Laboratory for 1.5 h. After three rinses in 0.1 M PBS and DAPI staining at 1:3000 (PIERCE), the sections were mounted on polysine microscopic glass (Thermo Scientific). Images were acquired using a Nikon A1R confocal microscope. We repeated each experiment at least three times.

Microglial morphology analysis

To quantify microglial morphology, confocal z-stacks of Iba1-positive cells were acquired. The images were binarized and skeletonized. A Sholl analysis was then performed using the ImageJ plugin, starting from the center of the cell soma. The number of intersections (branch points) at concentric circles of increasing radii was counted to assess morphological complexity.

Sample preparation for BSE-mode scanning electron microscopy – Brain tissue

Brain tissues were fixed overnight in 2.5% glutaraldehyde (GA) at 4 °C. The fixed tissues were coronally cut into 1-mm-thick slices using a mouse matrix (TED Pella). Brain slices were washed twice with phosphate-buffered saline (PBS) and once with 0.1 M sodium cacodylate buffer for 20 min each. Postfixation was performed with 2% osmium tetroxide (OsO₄) for 3 h, followed by two washes with 0.1 M sodium cacodylate buffer and one wash with distilled water for 20 min each.

For en bloc staining, the slices were incubated overnight at 4 °C in 2% aqueous uranyl acetate. The samples were subsequently washed three times with distilled water for 20 min each and subjected to dehydration with graded ethanol (30%, 50%, 70%, 80%, 90%, and 100%) for 15 min at each step. Further dehydration was performed using a 1:1 mixture of ethanol and acetone, followed by 100% acetone for 20 min each. The samples were then infiltrated with Spurr’s resin using stepwise ratios of acetone to resin (3:1, 1:1, 1:3) for 1 h each. After overnight infiltration in fresh resin, the samples were embedded and polymerized at 60 °C for 24–48 h.

SEM observations – Brain tissue

Resin-embedded brain samples were sliced at 50–100-nm thickness using an ultramicrotome (MT-XL, RMC, USA) installed in KIST BIO-Lab. Ultrathin sections were mounted onto silicon wafers using a loop and sequentially stained with 2% aqueous uranyl acetate for 10 min and Reynolds’ lead citrate for 5 min at room temperature. Imaging was performed with a scanning electron microscope (SEM; TeneoVS, FEI, USA) using a T1 backscattered electron (BSE) detector. Tile images were acquired at dimensions of 3072 × 2048 pixels or 6144 × 4096 pixels, with a working distance of approximately 6.5 mm, an acceleration voltage of 1.5 kV, and an emission current of 50 pA. Image acquisition and stitching were conducted using MAPS software.

Sample preparation for transmission electron microscopy – Cells

The cell pellets were fixed overnight in 2.5% glutaraldehyde prepared in PBS at 4 °C, followed by two washes with PBS and one wash with 0.1 M sodium cacodylate buffer for 15 min each. Postfixation was performed in 2% osmium tetroxide (OsO₄) for 2 h, followed by two washes with 0.1 M sodium cacodylate buffer and a final wash with distilled water for 15 min each. The samples were then embedded in low-melting-point agarose and allowed to solidify. After being washed with distilled water, the samples were subjected to en bloc staining in 2% aqueous uranyl acetate overnight at 4 °C. The samples were subsequently washed three times with distilled water for 15 min each and dehydrated through a graded ethanol series (30%, 50%, 70%, 80%, 90%, and 100%) for 15 min per step. Further dehydration was performed using a 1:1 mixture of ethanol and acetone, followed by 100% acetone for 15 min each. The samples were infiltrated stepwise with Spurr’s resin (acetone:resin = 3:1, 1:1, 1:3) for 1 h at each ratio. Following overnight infiltration in fresh resin, the samples were embedded and polymerized at 60 °C for 72 h.

TEM observations – cells

Resin-embedded cell samples were sliced at ~50-nm thickness using an ultramicrotome (MT-XL, RMC, USA) and collected onto Formvar-coated copper grids. The sections were stained with 2% aqueous uranyl acetate for 10 min, followed by Reynolds’ lead citrate for 5 min at room temperature. Imaging was performed using a transmission electron microscope (TEM; Tecnai-G2, FEI, USA).

Sample preparation for SEM and SEM observation – cells

To observe the surface of the cells, the cells on the coverslips were fixed in 2.5% glutaraldehyde prepared in PBS for 90 min, followed by two washes with PBS and one wash with 0.1 M sodium cacodylate buffer for 15 min each. Postfixation was performed in 2% osmium tetroxide (OsO₄) for 1 h, followed by two washes with 0.1 M sodium cacodylate buffer and a final wash with distilled water for 10 min each. The samples were subsequently washed three times with distilled water for 10 min each and dehydrated through a graded ethanol series (30%, 50%, 70%, 80%, 90%, and 100%) for 10 min per step and then with a hexamethyl-disilazane (HMDS) series (3:1, 1:1, 1:3 EtOH:HMDS), and 100% HMDS. After drying in air, the samples were mounted on SEM stubs, followed by a 5-nm-thick chromium coating with a sputter coater (ACE600, Leica). Surface imaging was carried out using SE mode at 2 kV and 0.2 nA with an ETD detector in SEM (TeneoVS, FEI, USA). To objectively quantify microglial surface roughness, SE-mode SEM images were analyzed using the 3D Surface Plot function in ImageJ (NIH). This tool converted the pixel intensity of the cell surface into topographic Z-height information. Surface roughness was then determined by calculating the standard deviation of these Z values from 15 ROIs (5 × 5 µm each) per cell, with a total of 90 ROIs analyzed per experimental group (n = 6 cells). This variance in surface height served as a quantitative measure of topographical complexity, reflecting the structural changes in the microglial membrane following QA treatment.

Library preparation for single-cell RNA sequencing using “gel bead-in-emulsion” technology

Microglial samples were prepared in two groups, namely, cells with no treatment (CONT) and cells treated with quinolinic acid for 6 h (QA). The cells were resuspended in 1X PBS with 0.04% BSA to obtain 1000 cells/µl. Each sample was used as an input for the library preparation process using the Chromium NEXT GEM Single-Cell 5’ Kit v2 (PN-1000265; 10x Genomics) according to the manufacturer’s instructions, which targeted 10,000 cells for each sample to be sequenced. Briefly, 16.5 µl of each sample and 22.2 µl of nuclease-free water were mixed with a 31.3 µl aliquot of master mix and loaded into the first row of the chip. The second row was filled with 50 µl of gel beads, and 45 µl of partitioning oil was loaded onto the third row. Any unused wells were filled with 50% glycerol in the same amount as the other wells in the corresponding row at the beginning. After loading, each GEM containing a single cell was generated by running the chip on a chromium controller. GEM was visually inspected and incubated in a thermocycler to obtain barcoded first-strand cDNA immediately after the run was complete. The barcoded cDNA of each sample was recovered from the cell pellet and reagents, followed by amplification in a thermocycler. The cDNA concentration was determined by a Qubit (Thermo Fisher Scientific). The libraries were subjected to paired-end sequencing on an Illumina NextSeq 550 (Illumina) to obtain at least 20 GB of FASTQ files per library.

Data preprocessing

FASTQ files were aligned to a prebuilt human reference transcriptome (GRCh38, downloaded from 10x Genomics), and a gene feature–cell barcode matrix was generated with Cell Ranger 6.1.2 (10x Genomics) with default settings. To identify true cell barcodes from background noise, the barcodes were sorted on the basis of the total unique molecular IDs, and the 99th percentile of the top 3000 cells was determined as the standard number of unique molecular IDs. Any cell barcode that had fewer unique molecular IDs than one-tenth of the standard number was discarded from the Cell Ranger pipeline. Additionally, the gene expression profiles of the barcodes with the lowest counts of unique molecular IDs were modeled as a multinomial distribution to further filter out barcodes that were unlikely to contain any cells. After filtering, 44.9% of the total reads (274,333,161) in the CONT group and 44.4% of the total reads (293,701,659) in the QA group were assigned to cells.

The standard workflow of the public R package for handling single-cell RNA-seq data (Seurat) was used for quality checks and clustering of the data. Metrics for quality control were checked in terms of the number of detected genes (min 500, max 7800), the number of total genes (min 600, max 61,000), and the percentage of mitochondrial genes (min 0, max 35). Because all empty droplets were filtered out in the first round of quality control, no further filtering was applied. After quality control, 6741 cells in the control group and 14,203 cells in the QA group were retained. The genes included 19,832 genes in the control group and 19,867 genes in the QA group, with a total of 21,590 genes detected across all cells. The gene counts of each cell were normalized to the total gene count and log transformed for further analysis.

Data integration, dimensionality reduction, and clustering

Integration of single-cell RNA sequencing data from all experimental groups was conducted by following the “Introduction to scRNA-seq integration” vignette of Seurat, which identified cells in different groups but showed similar gene expression patterns to correct for technical differences among the groups. To identify mutual nearest cells from each group, 2000 genes with the greatest variance in expression across all groups were chosen to calculate the distance between any pair of cells.⁷³

Cell-type identification by dimensionality reduction. For integrative analysis, we followed the workflow described in the Seurat-guided analysis. We first log-normalize the filtered matrixes and identify highly variable features for each sample using the FindVariableFeatures function with the parameters selection.method = vst and nfeatures = 1000. To integrate all 21 samples, we identified features for anchoring the samples using the FindIntegrationAnchors function with the parameter dims = 1:20 and used the identified anchors to integrate the dataset using the IntegrateData function with the parameter dims = 1:20. We subsequently scaled the integrated matrix and performed linear dimensional reduction using the RunPCA function with the parameter npcs = 50. We visualized the P value distribution of each principal component using the JackStrawPlot function and opted to use the first 20 principal components for graph-based clustering. We performed K-nearest neighbor clustering using the FindClusters function with a parameter resolution = 1 and UMAP clustering using the RunUMAP function with a parameter dims = 1:20, which initially yielded 43 cell clusters. We identified the DEGs in each cell cluster by the Wilcoxon rank-sum test using the FindAllMarkers function with the parameters logfc.threshold = 0.25 and test.use = wilcox. We then assigned a cell-type identity to each cell cluster according to the expression of known cell-type markers and identified additional cell-type-specific marker genes by the Wilcoxon rank-sum test using the FindAllMarkers function with the parameters logfc.threshold = 0.25 and test.use = wilcox. For cell-type markers, the level of statistical significance was set at an adjusted P < 0.1 and a log2-fold change ≥0.1 or ≤ −0.1. Subcluster analysis. For subcluster analysis, we first isolated individual cell types from the original Seurat dataset using the subset function. We subsequently reclustered each cell type using an approach similar to that used for our initial cell type clustering. We performed K-nearest neighbor clustering using the FindClusters function with a resolution of 0.2, 0.2, 0.3, and 0.2 for astrocytes, endothelial cells, microglia, and oligodendrocytes, respectively, as well as UMAP clustering using the RunUMAP function with a parameter dims = 1:20. To perform unbiased identification of the subpopulations of cells that contribute to AD-associated transcriptomic changes, we calculated the enrichment scores of AD-associated DEGs in each subpopulation by averaging the z scores of all AD-associated DEGs. A subpopulation with an enrichment score >0.5 was designated an AD-associated subpopulation. We identified the transcriptomic signatures of AD-associated subpopulations by comparing up- and downregulated subpopulations using the Wilcoxon rank-sum test with the FindMarkers function and the parameters logfc.threshold = 0 and test.use = wilcox. The level of statistical significance was set at an adjusted P < 0.1 and a log2-fold change ≥0.1 or ≤ −0.1. Data validation by comparison with previous studies. For data validation, we obtained datasets from published studies and compared them with our findings. We obtained DEG analysis results from a recent snRNA-seq study.⁷⁴ We selected a list of cell type-specific DEGs between healthy and disease samples with an adjusted P < 0.01 (two-sided Wilcoxon rank-sum test) and a log2-fold change ≥0.25 or ≤ −0.25. However, the proportion of endothelial cells in the previous dataset was low, rendering it unsuitable for validating our findings in endothelial cells. We also obtained the dataset from a study by Narayanan et al. (Gene Expression Omnibus [GEO] accession no. GSE33000), who performed bulk transcriptome microarray analysis of prefrontal cortical tissues in a large cohort (AD: n = 310; NC: n = 157).⁷⁵ We first filtered the samples according to disease status and kept only “Alzheimer’s disease” and “nondemented” samples for subsequent analysis. We removed genes that failed to be mapped to Entrez gene IDs. For genes mapped by multiple probes, we used the median value. To scale the variance, we performed log2 transformation and quantile normalization using the R limma package.⁷⁶ For differential expression analysis, we fitted the gene expression profiles by linear regression, adjusting for age and sex. We used the empirical Bayes method provided in limma to calculate t statistics and log-fold changes in differential expression. We adjusted the P values using the Benjamini–Hochberg procedure. We also used the same pipeline to process the microarray dataset of AD temporal cortical samples from Webster et al. (GEO accession no. GSE15222).⁷⁷ We obtained transcriptome data for mouse models of amyloid-beta deposition and Tau hyperphosphorylation from MOUSEAC.⁷⁸ Data Visualization. We visualized the data using Morpheus, Seurat’s DoHeatmap or DotPlot function, or Cytoscape (version 3.7.0) where appropriate. DEGs were functionally annotated according to Gene Ontology (GO) analysis (geneontology.org), ingenuity pathway analysis (QIAGEN), and STRING analysis (https://string-db.org). Data availability. Anonymized snRNA-seq sequencing data have been deposited in GEO (accession no. GSE157827).⁷⁰ All study data are included in the article and supporting information.

Consistency with other public data

AD-related microglial single-nucleus RNA-seq data were obtained^74,79. A list of genes related to phagocytosis-related microglia was obtained⁸⁰. Genes differentially expressed in amoeboid microglia were identified in⁸¹. Damage-associated microglial markers were obtained⁸². Differentially expressed genes among microglia, macrophages, and monocytes were obtained⁸³. Developmental microglial stage markers were obtained⁸⁴. LPS-treated HMC3 bulk RNA-seq data were obtained in ref. ⁸⁵.

Bulk RNA sequencing and pathway analysis of mouse brain tissue

Total RNA was extracted from the hippocampal tissues of saline- or QA-injected 5xFAD mice using TRIzol reagent (Invitrogen) according to the manufacturer’s instructions. RNA integrity and concentration were assessed using an Agilent 2100 Bioanalyzer. RNA-seq libraries were constructed using the TruSeq Stranded mRNA Library Prep Kit (Illumina). Sequencing was performed on an Illumina NovaSeq 6000 platform to generate 150-bp paired-end reads. The raw sequencing reads were quality controlled using FastQC, and the adapter sequences were trimmed. The processed reads were aligned to the mouse reference genome (mm¹⁰) using the STAR aligner. Differential gene expression analysis was performed using the DESeq2 R package, with a threshold of adjusted P < 0.05 and |log2-fold change| ≥ 1. Functional enrichment analysis, including Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses, was conducted to identify biological processes significantly modulated by QA treatment. To assess whether the QA-responsive transcriptional programs identified in HMC3 cells were recapitulated in the mouse hippocampus, genes upregulated by QA treatment in HMC3 cells were obtained from single-cell RNA-seq analysis as described above. These gene sets were compared with transcriptional responses in hippocampal bulk RNA-seq from QA-injected 5xFAD mice at the pathway level. Functional concordance between datasets was evaluated by testing whether biological processes enriched in HMC3 were also significantly enriched among QA-responsive genes in the mouse dataset using Gene Ontology (GO) enrichment analysis implemented in the clusterProfiler R package and the DAVID functional annotation tool. In addition, gene set enrichment analysis (GSEA) was performed using ranked gene lists derived from the 5xFAD bulk RNA-seq comparison (QA-injected vs. saline-injected) to assess enrichment of gene sets previously identified in single-cell analysis. Statistical significance was determined on the basis of p < 0.05.

Comparison of mouse brain RNA-seq with HMC3 SCRNA-seq

To assess whether the QA-responsive transcriptional programs identified in HMC3 cells were recapitulated in the mouse hippocampus, genes upregulated by QA treatment in HMC3 cells were obtained from single-cell RNA-seq analysis as described above. These gene sets were then compared at the pathway level with transcriptional responses in hippocampal bulk RNA-seq from QA-injected 5xFAD mice. Functional concordance between datasets was evaluated by determining whether biological processes enriched in QA-treated HMC3 cells were also significantly enriched among QA-responsive genes in the mouse dataset using Gene Ontology (GO) enrichment analysis implemented in the clusterProfiler R package and the DAVID functional annotation tool. In addition, gene set enrichment analysis (GSEA) was performed using ranked gene lists derived from the bulk RNA-seq comparison (QA-injected vs. saline-injected 5xFAD mice) to assess enrichment of gene sets previously identified in the scRNA-seq analysis. Statistical significance was defined as a p < 0.05.

RNA isolation and quantitative real-time PCR (qPCR)

qPCR was performed with a LightCycler 96 system instrument and software (Roche, Rotkreuz, Switzerland) according to the manufacturer’s recommended conditions. Total RNA was isolated from treated cells (TRIzol reagent; Invitrogen, CA), reverse transcribed (Superscript III; Invitrogen, CA), and subjected to quantitative PCR analysis using SYBR green master mix (Invitrogen, CA). The comparative threshold cycle (Ct) method was used to calculate the amplification factor, and the relative amount of each target was normalized to the GAPDH level in parallel reactions.

Western blot analysis

Western blot analyses were performed as described previously.^86–88 To prepare cell lysates, cell pellets were lysed in ice-cold RIPA buffer (50 mM Tris-HCl (pH 7.4), 150 mM NaCl, 1% Triton X-100, 1% sodium deoxycholate, and 0.1% SDS) supplemented with a cocktail of protease and phosphatase inhibitors. Following incubation on ice for 30 min, the lysates were centrifuged at 13,000 × g for 15 min at 4 °C. The resulting supernatant, containing both cytosolic and detergent-solubilized membrane proteins (including lipidated ATG8 family members), was collected and designated the ‘Triton X-100 soluble lysate’. Protein concentrations were determined using a BCA protein assay kit. Equal amounts of protein (20–30 μg) were separated by SDS‒PAGE (12% or 15% gels) and transferred to PVDF membranes. The membranes were blocked with 5% nonfat milk in Tris-buffered saline containing 0.05% Tween 20 (TBS-T) for 1 h at room temperature. The transferred blots were then incubated with the following primary antibodies at 4°C for 24 h: mouse anti-GABARAP (1:1000, MBL, USA), rabbit anti-LC3B (1:1000, MBL, USA), and mouse anti-ACTB/beta-actin (1:2000, Santa Cruz, USA). After being washed three times with TBS-T, the blots were incubated with appropriate secondary antibodies conjugated to horseradish peroxidase (HRP), such as anti-rabbit HRP or anti-mouse HRP (Amersham Pharmacia, USA), for 2 h at room temperature. The protein bands were developed using Immobilon Western ECL solution (Merck Millipore, USA) and visualized using an ImageQuant LAS 4000 system (GE Healthcare, USA). ACTB was utilized as the loading control for normalization.

Statistical analysis

Statistical analyses were performed using Prism 10. Differences between two different groups were analyzed with the two-tailed Student’s unpaired t test or Mann‒Whitney test (when the data were not normally distributed). To assess the change in a group in response to a specific intervention, the significance of the data was assessed by two-tailed Student’s paired t test. For comparisons of multiple groups, one-way analysis of variance (ANOVA) with Tukey’s or Dunnett’s multiple comparison test or two-way ANOVA with Bonferroni’s multiple comparison test was used. Data from multiple independent experiments were assumed to have normal variance. p < 0.05 was considered to indicate statistical significance throughout the study. The significance level is represented as an asterisk (*p < 0.05, **p < 0.01; ns, not significant). Unless otherwise specified, all the data are presented as the mean ± SEM. Sample sizes were determined empirically on the basis of our previous experience or a review of similar experiments in the literature. The numbers of animals used are described in the corresponding Fig. legends or on each graph. The experimental groups were balanced in terms of animal age, sex and weight. The animals were genotyped before the experiments, and they were all caged together and treated in the same way. The animals were randomly and evenly allocated to each experimental group. To perform the group allocation in a blinded manner during data collection, animal preparation and experiments were performed by different investigators. Unless otherwise specified, no data points were excluded.

Author contributions

S.J.H., J.L., and H.R. designed the research; S.J.H., S.C.K., J.C., Y(Yeonseo)K, U.P., P.T.T.N., S(Sojung)K, Y(Yeyun)K, and S(Suhyun)K performed the in vitro and in vivo experiments; H.L. and Y(Yeonseo)K performed the LC‒MS analysis for lipidomics; K.E.L. and A.E. performed the EM imaging; S.J.H., Z.T., and Y(Yeonseo)K analyzed the data; S.L., E.M.H., S.E.L., J.W., J(Jin-A)L, J(Junghee)L, and T.D.S. provided resources; and S.J.H., Z.T., H.L., J(Junghee)L, and H.R. wrote the manuscript. All the authors have read and approved the article.

Data availability

The sequencing data generated in this study have been deposited in the NCBI Gene Expression Omnibus (GEO) database under accession number GSE327618. Uncropped images used to generate the figures throughout the manuscript can be found within the Supplementary Materials.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41392-026-02789-z.