Aberrant excitatory neuronal ERBB4 promotes Alzheimer’s disease pathology

Synapse loss is evident across multiple brain regions from early stages of AD^5,6. A prevailing model proposes that amyloid-β (Aβ) oligomers induce reactive microglial states and activate complement-dependent phagocytic programs, leading to excessive synapse elimination^1,2,3,4. Previously, we revealed that astrocytes have a major role in the continuous elimination of both excitatory and inhibitory synapses in the normal adult hippocampus⁷. This astrocytic elimination of adult hippocampal synapses is highly dependent on hippocampal activity and has a critical role in maintaining circuit homeostasis and memory function⁷. As most research on the mechanisms of synapse loss in AD has focused on microglia and excitatory synapses, we initially set out to determine the contribution of astrocytes to the elimination of both excitatory and inhibitory synapses during AD progression.

Here we found that altered glial synapse elimination arises as a result of gene expression and functional changes in AD excitatory neurons. Using single-nucleus RNA-sequencing (snRNA-seq), gene set enrichment analysis (GSEA), weighted gene correlation network analysis (WGCNA) and candidate gene approach, we identified a single responsible pathway, an ectopically expressed ERBB4 tyrosine kinase receptor in excitatory neurons, as a key disease-promoting factor orchestrating broad aspects of AD pathophysiology. CRISPR-based Erbb4 knockout in excitatory neurons of the 5×FAD mouse model substantially normalized excitatory neuronal hyperactivity, inhibitory neuronal hypoactivity, abnormal synapse elimination, reactive gliosis and even Aβ accumulation, leading to significant recovery of cognitive functions. Conversely, restricted Erbb4 overexpression in wild-type (WT) excitatory neurons reproduced major AD-like phenotypes in the absence of Aβ plaques through downstream mTOR signalling. Human AD tissue and transcriptomic analyses further linked excitatory neuronal ERBB4 to amyloid and tau pathology and cognitive decline. These findings support a model in which early excitatory neuronal dysfunction, mediated by aberrant Erbb4 expression, precipitates maladaptive glial responses and synaptic imbalance, promoting cognitive decline and neurodegeneration.

We monitored glial synapse engulfment using ExPre and InhiPre reporters, in which acid-labile eGFP undergoes denaturation and loses fluorescence in acidic phagolysosomes, whereas acid-resistant mCherry remains fluorescent, thereby distinguishing intact from engulfed pre-synapses^7,8 (Extended Data Fig. 1a). ExPre was expressed in hippocampal CA3 excitatory projections and analysed in the stratum radiatum (SR) of the hippocampal CA1, where Schaffer collateral synapses predominate. InhiPre was expressed in CA1 inhibitory neurons and analysed in the stratum lacunosum moleculare (SLM), which is enriched in somatostatin (SST)-positive inhibitory synapses⁹. After reporter delivery to WT and APP/PS1 mice, we quantified astrocytic and microglial engulfment from 3 to 18 months to compare ageing- and AD-associated changes (Extended Data Fig. 1b–e).

Across APP/PS1 progression, excitatory pre-synapse elimination by both astrocytes and microglia progressively increased, whereas inhibitory pre-synapse elimination decreased (Extended Data Fig. 1b–e). Astrocytic changes were evident by 6 months and preceded comparable microglial changes. A similar opposing pattern emerged in the more rapidly progressing 5×FAD model at 3 months, but not at 2 months, coinciding with the appearance of hippocampal Aβ plaques (Fig. 1a–d and Extended Data Fig. 1f–i). Astrocytes engulfed more synapses than microglia across synapse types, ages and mouse models (Extended Data Fig. 1j,k), suggesting that they are major early responders to synaptic perturbation. No sex-dependent differences in phagocytosis were detected (Extended Data Fig. 2a–d).

Synapse numbers were unchanged in the 2- and 3-month-old cohorts but, by 4 months, 5×FAD mice had fewer excitatory synapses in the SR and more inhibitory synapses in the SLM compared with the WT controls (Fig. 1e–i and Extended Data Fig. 2e–l). These findings, in which changes in synapse numbers are negatively correlated with the extent of synapse engulfment by glial cells, suggest that glial phagocytosis may regulate excitatory as well as inhibitory synapse numbers in AD brains. Furthermore, the selective elimination of excitatory synapses, coupled with reduced elimination of inhibitory synapses, raises the possibility that intrinsic neuronal states may guide glial phagocytic behaviour and that indiscriminate neuroinflammation is unlikely to be the sole driver of synapse elimination in AD.

Notably, excitatory neurons become hyperactive, whereas inhibitory neurons become hypoactive, early in AD^10,11,12. As glial synapse elimination is neuronal activity dependent^7,13,14, we hypothesized that early alterations in neuronal activity in AD brains may misguide glial phagocytic processes, leading to aberrant synapse elimination. To modulate intrinsic neuronal activity, human G_q- or G_i-coupled muscarinic designer receptors (hM3Dq or hM4Di) were expressed selectively in excitatory or inhibitory neurons, followed by four daily clozapine-N-oxide (CNO) injections and confirmation of activity changes by FOS immunoreactivity (Extended Data Fig. 3). Importantly, glial synapse phagocytosis was increased with heightened neuronal activity and decreased when neuronal activity was suppressed in both excitatory and inhibitory neurons (Extended Data Fig. 3). These data suggest that imbalanced neuronal activity, characterized by excitatory neuronal hyperactivity and inhibitory neuronal hypoactivity, may shape abnormal patterns of glial synapse elimination in AD brains.

To test whether astrocytic phagocytosis directly contributes to excitatory synapse loss, we knocked down the astrocyte-enriched phagocytic receptor MEGF10 in CA1 of 5×FAD mice using two AAV-encoded shRNAs⁷ (Extended Data Fig. 4a–c). Megf10 knockdown increased the number of excitatory synapses without affecting inhibitory synapses (Extended Data Fig. 4d–g), consistent with a selective requirement for MEGF10 in excitatory synapse elimination⁶.

To determine the molecular mechanisms underlying early alterations in neuronal activity and corresponding synapse elimination in AD brains, we performed hippocampal snRNA-seq analysis in 3-month-old WT, 2-month-old 5×FAD and 3-month-old 5×FAD mice. After SysVI batch integration¹⁵, excitatory neurons showed the most prominent disease-associated shift on uniform manifold approximation and projection (UMAP) plots (Fig. 2a). Subclustering separated excitatory neurons into ten clusters (Fig. 2b and Supplementary Fig. 1a,b) and identified clusters 5–9 as selectively enriched in 3-month-old 5×FAD mice (Extended Data Fig. 5a). Their differentially expressed genes (DEGs) were enriched for Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways associated with AD, Huntington’s disease (HD), Parkinson’s disease and amyotrophic lateral sclerosis (ALS) (Fig. 2c). As these transcriptional alterations appeared early and were more pronounced than changes in astrocytes or microglia (Extended Data Fig. 5b), we designated these clusters early-responsive excitatory neurons (ERENs).

Next, we performed WGCNA^16,17 of excitatory neuronal genes and identified ten distinct gene modules (Extended Data Fig. 5c,d). Among these, ERENs exhibited a significantly elevated module score only for the turquoise module (Fig. 2d and Extended Data Fig. 5c). For further analysis, we selected the top 25% of turquoise module genes that were differentially expressed between ERENs and other excitatory neuronal clusters, ranked by adjusted P value and expression level (Fig. 2e and Supplementary Fig. 1c). Among these top-ranked genes, Erbb4 was the only gene that overlapped with Reactome pathway terms enriched by GSEA, prompting us to prioritize ERBB4 signalling for subsequent analyses (Fig. 2f). Independent analysis of cortical neurons from 7-month-old 5×FAD mice¹⁸ similarly identified an expanded Erbb4^high excitatory subset and firebrick module enriched in this subset (Extended Data Fig. 5e,f). Reactome GSEA again identified signalling by ERBB4 as the only significantly enriched pathway among firebrick-module genes (Extended Data Fig. 5g,h), supporting ERBB4 signalling as a conserved feature of EREN-like states.

While Erbb4 has been known to be exclusively expressed in parvalbumin (PV)⁺ inhibitory neurons in the central nervous system, where it contributes to synaptogenesis through the mTOR pathway¹⁹, its expression and function in excitatory neurons remain unclear. To validate our findings, we performed immunohistochemistry and observed a significant upregulation of ERBB4 in the CA1, but not in the CA3, of 3-month- and 4-month-old 5×FAD mice compared with WT mice (Fig. 2g,h and Extended Data Fig. 5i). Consistently, fluorescence in situ hybridization (FISH) analysis revealed a significant increase in nuclear Erbb4 mRNA in pyramidal neurons within the CA1 hippocampus of 5×FAD mice (Extended Data Fig. 5j,k).

EREN abundance was unchanged in Trem2-deficient 5×FAD mice¹⁸ (Extended Data Fig. 5l), arguing against an essential role for TREM2-dependent disease-associated microglia (DAM) in their generation. Conversely, CA1 injection of Aβ oligomers induced excitatory neuronal ERBB4 within 2 days (Extended Data Fig. 5m,n). These findings suggest that Aβ can promote ectopic Erbb4 expression in excitatory neurons, at least partly independently of TREM2-mediated microglial activation.

To determine the functional role of ERBB4 in AD pathology, we selectively deleted Erbb4 expression in excitatory neurons by delivering an AAV encoding a CaMKIIα promoter-driven Staphylococcus aureus Cas9 (SaCas9) along with a U6 promoter-driven single guide RNA (sgRNA) targeting Erbb4 into CA1 of 4-month-old 5×FAD mice (Fig. 3a and Extended Data Fig. 6a). This approach reduced ERBB4 in pyramidal neurons without altering its expression in PV interneurons (Fig. 3b,c and Extended Data Fig. 5j,k). Importantly, Erbb4 deletion reduced excitatory pre-synapse elimination and increased inhibitory pre-synapse elimination by both astrocytes and microglia in the AAV-injected 5×FAD mice, restoring values towards those in WT mice (Fig. 3d–g). It also rescued the loss of excitatory pre-synapses and post-synapses and the increase in inhibitory pre-synapses (Extended Data Fig. 6l–o).

Consistent with previous reports²⁰, assessment of neuronal activity through FOS immunoreactivity in the CA1 revealed an increased the number of FOS⁺ excitatory neurons in 5×FAD mice compared with in WT mice (Extended Data Fig. 6d,f). Notably, Erbb4 deletion in excitatory neurons strongly normalized the hyperactivity of excitatory neurons in 5×FAD mice (Fig. 3h,j). Conversely, the number of FOS⁺ SST neurons, which was significantly lower in 5×FAD mice than in WT mice (Extended Data Fig. 6e,g), was also strongly restored to WT levels after excitatory neuronal Erbb4 deletion (Fig. 3i,k). The number of FOS⁺ PV neurons remained unchanged in 4-month-old 5×FAD mice, regardless of Erbb4 manipulation (Extended Data Fig. 6h–k), indicating subtype-specific FOS changes in AD interneurons. These findings place aberrant excitatory neuronal ERBB4 upstream of excitatory hyperactivity, SST hypoactivity and the corresponding changes in synapse-selective glial phagocytosis. The restoration of SST activity despite Erbb4 deletion being confined to excitatory neurons further indicated that ERBB4-dependent dysfunction propagates through local circuit interactions rather than remaining cell-autonomous.

Given the association between abnormal neuronal activity and neuroinflammation in AD²¹, we next evaluated reactive gliosis by quantifying S100β, GFAP and IBA1 levels in 5×FAD mice with or without CA1 excitatory neuronal Erbb4 deletion. Notably, the 5×FAD mice with excitatory neuronal Erbb4 deletion exhibited significantly reduced levels of all three reactive gliosis markers (Fig. 3l,m), along with a reduction in AXL⁺ microglia, indicative of a diminished DAM population (Extended Data Fig. 6p,q). Moreover, we found that amyloid plaque burden, measured by both plaque area and number, was strongly reduced in 4-month-old 5×FAD mice after excitatory neuronal Erbb4 deletion (Fig. 3n,o). These results suggest that aberrant Erbb4 expression in ERENs not only disrupts synaptic homeostasis but also drives gliosis and plaque deposition. Importantly, the same AAV-mediated deletion of Erbb4 in WT CA1 excitatory neurons did not elicit detectable changes in synapse phagocytosis or reactive gliosis (Extended Data Fig. 6r–w), thereby highlighting excitatory neuronal ERBB4 as a promising therapeutic target for AD pathophysiology.

To determine whether normalization of excitatory activity was sufficient to reproduce the effects of Erbb4 deletion, we expressed hM4Di in CA1 excitatory neurons of 5×FAD mice. After four consecutive days of CNO injections, we found that activating hM4Di in AD excitatory neurons not only reduced the number of FOS⁺ excitatory neurons, but also increased the number of FOS⁺ SST neurons in 5×FAD mice (Extended Data Fig. 7a–e), supporting abnormal excitatory control over SST neurons in AD brains. Excitatory neuronal Erbb4 deletion also reduced the number of FOS⁺ vasoactive intestinal peptide (VIP) interneurons (Extended Data Fig. 6b,c). As VIP interneurons inhibit SST neurons²², these results are consistent with a potential excitatory neuron–VIP–SST disinhibitory circuit linking ERBB4-dependent excitatory hyperactivity to SST hypoactivity. Suppression of excitatory neuronal activity also decreased excitatory synapse phagocytosis by astrocytes and increased inhibitory synapse phagocytosis by both astrocytes and microglia (Extended Data Fig. 7f–i). However, despite normalization of neuronal activity and synapse phagocytosis, acute hM4Di activation did not reduce gliosis or amyloid plaques (Extended Data Fig. 7j–m), in contrast to Erbb4 deletion. These data suggest that ERBB4-expressing excitatory neurons initiate multiple aspects of AD pathophysiology by disrupting neuronal network activity. However, reactive gliosis may arise through ERBB4-dependent but neuronal-activity-independent mechanisms that have yet to be identified.

Finally, we tested durability and therapeutic timing. Erbb4 deletion at 3 months reduced reactive gliosis, AXL⁺ or TREM2⁺ DAM, and amyloid deposition when mice were analysed at 7 months (Extended Data Fig. 8a–h). Delivery at 8 months likewise reduced these pathological features at 10 months (Extended Data Fig. 8i–p). Thus, reducing excitatory neuronal Erbb4 produced sustained benefits and remained effective when initiated after substantial disease progression.

We next examined whether ectopic Erbb4 expression in WT excitatory neurons was sufficient to induce AD-like changes. A CaMKIIα-driven Erbb4 AAV was delivered to the CA1 of 2-month-old WT mice, with the viral dosage adjusted so that 5–15% of pyramidal neurons expressed ERBB4, matching the proportion observed in 5×FAD mice (Fig. 4a–c). Importantly, ERBB4 overexpression increased excitatory and decreased inhibitory synapse engulfment by both astrocytes and microglia (Fig. 4d–g), increased FOS⁺ excitatory neurons, reduced FOS⁺ SST neurons and left PV activity unchanged (Fig. 4h–k and Extended Data Fig. 9a,b). Along with these changes, the number of excitatory synapses was decreased while the number of inhibitory synapses was increased (Extended Data Fig. 9c–f), recapitulating the excitatory–inhibitory imbalance observed in 5×FAD mice. ERBB4 overexpression also elevated GFAP, S100β, IBA1 and AXL⁺ microglia without detectable cleaved caspase-3⁺ apoptosis (Fig. 4l–n and Extended Data Fig. 9g,h). By contrast, a kinase-dead ERBB4 mutant (K751M) did not alter synapse phagocytosis or gliosis (Extended Data Fig. 9i–q). Thus, ERBB4 kinase activity in excitatory neurons is sufficient to drive major AD-like neuronal, synaptic and glial phenotypes.

Erbb4 expression in cortical excitatory neurons similarly increased neuronal FOS activity and S100β, GFAP and IBA1 levels (Extended Data Fig. 9r–u). Together with the cortical EREN-like population identified in 7-month-old 5×FAD mice, these findings indicate that ectopic ERBB4 can promote AD-like neuronal and glial changes beyond the hippocampus.

As aberrant ERBB4 signalling altered both excitatory and inhibitory synapse numbers, we next examined whether these structural changes altered synaptic input onto CA1 pyramidal neurons. Whole-cell recordings were performed at 4 months, when ERBB4-dependent synapse changes first became evident in 5×FAD mice. Analysis of miniature excitatory postsynaptic currents (mEPSCs) revealed a reduced event frequency in 5×FAD mice compared with in the WT controls, and this reduction was restored to WT levels by excitatory-neuron-specific Erbb4 deletion (Extended Data Fig. 7n–p). Conversely, Erbb4 overexpression in WT excitatory neurons reduced the mEPSC frequency to levels comparable to those observed in 5×FAD mice (Extended Data Fig. 7n–p). By contrast, the mEPSC amplitude remained unchanged across groups (Extended Data Fig. 7o). Together, these data indicate that ERBB4 primarily decreases the number of functional excitatory inputs rather than changing postsynaptic strength at individual synapses.

Despite increased inhibitory presynaptic density in the SLM, both the frequency and amplitude of miniature inhibitory postsynaptic currents (mIPSCs) recorded from 5×FAD CA1 pyramidal neurons were reduced compared with WT controls (Extended Data Fig. 7q–s). Moreover, Erbb4 modulation in 5×FAD mice did not produce statistically significant changes in either mIPSC frequency or amplitude (Extended Data Fig. 7q–s). However, Erbb4 overexpression in WT excitatory neurons showed a trend toward increased mIPSC frequency, consistent with the increased inhibitory synapse number observed under this condition (P = 0.0504, unpaired Student’s t-test). Thus, the lack of a corresponding increase in mIPSC frequency despite the increased inhibitory presynaptic density suggests that these additional inhibitory contacts in 5×FAD mice may be immature, silent or otherwise functionally ineffective. Moreover, broader AD-associated dysfunction of PV or other interneuron populations could also contribute to the reduced inhibitory transmission^23,24.

Although the ERBB4-dependent reduction in mEPSC frequency is consistent with excitatory synapse loss, this synaptic input phenotype alone cannot readily explain the increased FOS activity in excitatory neurons (Extended Data Fig. 6d,f). We therefore examined whether aberrant ERBB4 expression alters the intrinsic neuronal excitability of CA1 pyramidal neurons after Erbb4 modulation. Input resistance was unchanged, but 5×FAD and Erbb4-overexpressing WT pyramidal neurons fired many more action potentials during depolarizing current injection, whereas Erbb4 deletion restored 5×FAD firing to WT levels (Extended Data Fig. 7t–v). These findings indicate that aberrant ERBB4 directly increases intrinsic excitability, providing a mechanism for excitatory hyperactivity. Hyperactivity may then promote glial synapse elimination and subsequent synapse loss, consistent with chemogenetic experiments showing that G_q-driven activation enhances, whereas G_i-driven suppression reduces, synapse phagocytosis (Extended Data Figs. 3 and 7a–i).

ERBB4 activates several downstream pathways, including the mTOR^19,25 and mitogen-activated protein kinase (MAPK) signalling cascades^26,27. Notably, neuronal mTOR hyperactivation is associated with seizures, excitatory–inhibitory imbalance, Down syndrome and AD^28,29, prompting us to test whether mTOR signalling acts downstream of ERBB4 to mediate AD-related pathological effects. We deleted the gene encoding regulatory-associated protein of mTOR (Rptor), an essential component of the mTOR complex 1 (mTORC1), in CA1 excitatory neurons of 4-month-old 5×FAD mice using CaMKIIα-SaCas9 and sgRptor (Extended Data Fig. 10a). Rptor deletion reduced phosphorylated ribosomal protein S6 (p-S6), confirming suppression of mTORC1 activity (Extended Data Fig. 10b,c). Similar to Erbb4 deletion, Rptor deletion normalized excitatory and inhibitory synapse elimination, restored corresponding synapse numbers, reduced excitatory activity and increased SST activity (Extended Data Fig. 10d–o). Moreover, Rptor deletion significantly attenuated gliosis, reduced both the AXL⁺ microglia population and amyloid plaque burden (Extended Data Fig. 10p–u). The close correspondence between Rptor and Erbb4 deletion phenotypes placed mTORC1 at a major convergence point linking excitatory neuronal signalling to synaptic, glial and amyloid-related outcomes.

We next examined whether excitatory neuronal mTOR signalling is required for the pathological effects induced by ectopic Erbb4 overexpression. To this end, we co-expressed Erbb4 with sgRptor in CA1 excitatory neurons of 2-month-old WT mice (Extended Data Fig. 11a). Despite robust ERBB4 expression, Rptor deletion suppressed p-S6 and prevented excitatory hyperactivity, SST hypoactivity, gliosis, AXL⁺ microglia, and changes in excitatory and inhibitory synapse numbers (Extended Data Fig. 11b–o). Together, these findings demonstrate that mTOR signalling is a critical downstream mediator of aberrant ERBB4 signalling and is necessary for the induction of multiple AD-like pathophysiological features.

Consistent with an ERBB4–mTOR signalling axis, we found that ERENs and excitatory neurons from Erbb4-overexpressing WT mice showed transcriptional changes that resembled those seen in Tsc2-deficient neurons³⁰ (Extended Data Fig. 12a). Notably, these overlapping changes involved genes related to neuronal excitability, including those encoding voltage-gated K⁺ channels, inwardly rectifying K⁺ channels, γ-aminobutyric acid type A (GABAA) receptors and GABAB receptors. By contrast, excitatory neurons from Erbb4-deficient 5×FAD mice showed opposing changes in these gene sets (Extended Data Fig. 12a). Given that these ion channels and inhibitory receptors are critical for regulating neuronal excitability^31,32,33,34, our findings suggest that aberrant ERBB4–mTOR signalling may promote excitatory neuronal hyperactivity, at least in part, by disrupting transcriptional programs that normally constrain neuronal excitability.

To better understand how Erbb4 contributes to multiple aspects of AD pathophysiology, we performed CA1 snRNA-seq analysis in 4-month-old 5×FAD mice with or without Erbb4 deletion and in WT mice with or without Erbb4 overexpression. First, in 5×FAD mice, UMAP data readily showed Erbb4 loss within the excitatory neuronal cluster after introduction of a CRISPR AAV targeting Erbb4 (Fig. 5a and Supplementary Fig. 1a,b). Moreover, deleting Erbb4 in 5×FAD excitatory neurons substantially altered the UMAP distribution of astrocytes and microglia (Fig. 5b,c). Importantly, DAM-associated genes, including Cst7, Lpl, Lyz2 and Apoe, and reactive astrocyte-associated genes, including Vim, Gfap, Serpina3n and Cp, were reduced. Conversely, homeostatic microglial markers such as P2ry12, Cx3cr1 and Tmem119 and astrocytic genes including Gpc5, Luzp2 and Lsamp were increased^35,36,37 (Fig. 5d–h).

In parallel, excitatory neuronal Erbb4 overexpression in WT mice generated a clearly separated Erbb4-expressing excitatory subcluster, analogous to that in 5×FAD mice (Figs. 2a and 5i and Supplementary Fig. 1a,b). Similarly, excitatory neuronal Erbb4 overexpression in WT mice was sufficient to shift microglia and astrocytes towards a more-reactive state (Fig. 5j,k). Microglia in Erbb4-overexpressing WT mice upregulated DAM-associated genes and concomitantly downregulated homeostatic genes (Fig. 5l,n,o). Likewise, astrocytes in this group upregulated inflammatory reactive-astrocyte-associated genes and reduced homeostatic gene expression (Fig. 5m,n,p). Notably, the expression levels of genes involved in the APP-processing pathway³⁸, including App, Psen1, Ncstn and Bace1, were significantly reduced in Erbb4-deleted 5×FAD excitatory neurons (Extended Data Fig. 12b), potentially providing a mechanistic explanation for the reduced Aβ plaque burden after excitatory neuronal Erbb4 deletion in 5×FAD mice.

Finally, we assessed hippocampus-dependent cognition using spontaneous alternation, novel-object location, novel-object recognition and Barnes maze tests^39,40,41. Notably, excitatory-neuron-specific Erbb4 deletion strongly rescued all behavioural deficits in 4-month-old 5×FAD mice, whereas Erbb4 overexpression impaired WT performance across these assays (Fig. 5q–t).

To test generalizability, Erbb4 was deleted in CA1 excitatory neurons of 7-month-old APP/PS1 mice, and behaviour was measured at 10 months. Control APP/PS1 mice were impaired relative to age-matched WT mice, whereas Erbb4 deletion improved performance in all behaviour tasks (Extended Data Fig. 12c,d). Taken together, these results underscore the pivotal role of aberrant excitatory neuronal ERBB4 in orchestrating gene expression changes not only in neurons but also in glial cells, ultimately leading to cognitive impairments through disrupting neural network activity and synapse homeostasis.

To assess the clinical relevance of our findings, we first assessed excitatory neuronal ERBB4 in human brain tissue. FISH analysis of human brain sections revealed increased ERBB4 signals in SLC17A7⁺ excitatory neurons from patients with AD compared with control individuals without dementia (Extended Data Fig. 12e,f). We next investigated whether ERBB4-expressing excitatory neurons in human patients with AD also show correlation with AD pathological metrics, including Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) score, Braak stage and Mini-Mental State Examination (MMSE). To address this, we analysed snRNA-seq data of CUX2⁺ and CUX2⁻ excitatory neurons from the dorsolateral prefrontal cortex of 446 participants in the Religious Orders Study and Rush Memory and Aging Project (ROSMAP) cohort^42,43,44. After integration¹⁵ with adjustment for confounding covariates such as age at death, sex and race, ERBB4^high clusters were identified within both CUX2⁺ and CUX2⁻ excitatory neuronal populations (Extended Data Fig. 12g,h). Notably, the abundance of these clusters was positively correlated with plaque severity as assessed by CERAD scores, and negatively correlated with cognitive performance measured by MMSE (Extended Data Fig. 12i,j). These correlations suggest that ectopic ERBB4 expression in excitatory neurons is linked to pathological burden and cognitive impairment in human AD.

We next used directed mediation analysis within the established amyloid-to-tau-to-cognitive-decline framework^45,46,47. Two non-mutually exclusive configurations were evaluated: one in which amyloid burden induces excitatory neuronal ERBB4, followed by tau pathology and cognitive decline; and another in which ERBB4 acts upstream of amyloid, tau and cognition. Both configurations showed a direct association between ERBB4 and cognitive decline, but the data provided stronger support for the upstream-ERBB4 model (Fig. 5u). In this configuration, ERBB4 was also indirectly associated with cognitive decline through sequential mediation by plaque severity and tau burden, whereas no plaque-independent direct association between ERBB4 and tau was detected.

On the basis of these results, we constructed a path model describing the relationships among excitatory neuronal ERBB4 expression and amyloid pathology, subsequent tau propagation and cognitive decline, and evaluated it using structural equation modelling. The model showed an excellent fit to the data (comparative fit index = 1.0, Tucker–Lewis index = 1.0; χ² test, P = 0.9863), supporting the plausibility of the proposed pathway (Fig. 5u). Importantly, replacing ERBB4 with ERBB2, another EGFR family member, did not yield an adequate model fit (comparative fit index = 0.9684, Tucker–Lewis index = 0.8895; χ² test, P = 0.0046), underscoring the specificity of ERBB4 in this framework. Collectively, these findings support a model in which aberrant ERBB4 expression in excitatory neurons forms a vicious cycle with amyloid pathology, acting as both a consequence and an amplifier of plaque accumulation, thereby promoting tau pathology and cognitive decline in human AD.

AD involves interconnected neuronal, glial and pathological changes, and therapies targeting a single component may therefore have limited efficacy. Although Aβ immunotherapies reduce brain amyloid⁴⁸, their cognitive benefits remain modest⁴⁹. Aβ accumulation may initiate multiple self-sustaining cascades that are not readily reversed by later plaque reduction^50,51,52. These findings highlight the need to identify core disease drivers downstream of Aβ accumulation.

Here we propose that the emergence of Erbb4-expressing excitatory neurons from the early stages of AD represents such a mechanism. Functional studies using CRISPR-mediated Erbb4 deletion revealed that excitatory neuronal ERBB4 is integral not only to synaptic regulation but also to excitatory hyperactivity, inhibitory hypoactivity, reactive gliosis, DAM induction and amyloid plaque deposition. Moreover, overexpression of Erbb4 in WT excitatory neurons recapitulated these pathological features, even in the absence of amyloid plaques. We further identified mTOR signalling as a key downstream effector of excitatory neuronal ERBB4.

Although we believe that our findings provide key mechanistic insight and a long-sought answer to the core drivers of early AD pathophysiology, several questions remain unresolved. First, while our data indicate that Aβ may act as a primary inducer of aberrant Erbb4 expression in excitatory neurons, the identity of the relevant ERBB4 ligands and the transcriptional regulation mechanisms governing Erbb4 remain to be elucidated. Second, our results suggest that inhibitory synaptic remodelling in 5×FAD mice may follow a more temporally or compartmentally complex trajectory than the excitatory phenotypes examined here. Analyses across interneuron subtypes, dendritic compartments and disease stages should clarify how it relates to excitatory dysfunction. Finally, given that our mouse study used amyloidogenic AD models, future studies should determine whether EREN-like populations arise in tauopathies and other neurodegenerative models and drive comparable pathology.

Notably, our transcriptomic analyses suggest that Erbb4-expressing excitatory neurons may not be unique to AD. Consistent with KEGG-GSEA results from mouse data (Fig. 2c), preliminary analyses of public human snRNA-seq datasets identified EREN-like excitatory neuronal clusters with elevated ERBB4 expression in the cortex of patients with progressive supranuclear palsy⁵³ (Extended Data Fig. 12k–m). Similarly, snRNA-seq data from patients with HD⁵⁴ revealed a linear correlation between excitatory neuronal ERBB4 and disease severity (Extended Data Fig. 12n). These observations suggest that ERENs may represent a convergent, disease-promoting neuronal state across diverse neurodegenerative conditions.

Mice

All mouse experiments were performed according to protocols approved by the Institutional Animal Care and Use Committees (IACUC) at Korea Advanced Institute of Science and Technology (KAIST) and the Institute for Basic Science (IBS), adhering strictly to ethical guidelines. B6.Cg-Tg(APPswe, PSEN1dE9)85Dbo/Mmjax (APP/PS1) mice and B6.Cg-Tg(APPSwFlLon, PSEN1*M146L*L286V)6799Vas/Mmjax (5×FAD) mice were produced in our laboratory and maintained by breeding with C57BL/6J mice. Mice were housed under a 12 h–12 h light–dark cycle. Both male and female mice were used unless otherwise specified in the figure legends. All mice were randomly assigned to experiments, which were performed by investigators blinded to experimental conditions. Gender was not considered unless explicitly noted in figure legends.

Human tissue sections

Human formalin-fixed paraffin-embedded (FFPE) sections were provided by the SNUH Brain Bank and use of these sections for this study was reviewed and determined to be exempt from review by the Public Institutional Review Board designated by the Ministry of Health and Welfare, Republic of Korea. All procedures involving human-derived materials were conducted in accordance with relevant regulations.

Cell lines

HEK293T cells (ATCC) were used solely for AAV production. Cells were maintained in our laboratory and were not independently authenticated for this study. HEK293T cells were confirmed to be free of mycoplasma contamination.

Antibodies and reagents

The antibodies used and their dilution factor were as follows: anti-S100β (Abcam, ab52642, Synaptic System, 287-004; Aves Labs, S100B-0020; 1:500), anti-IBA1 (Wako, 019-19741; Novus, NB100-1028; 1:500), anti-mCherry (Invitrogen, M11217; Aves Labs, mCherry-0020; 1:1,000), anti-vGLUT1 (Millipore, AB5905; 1:1,000), anti-PSD95 (Invitrogen, 51-6900; 1:500), anti-vGAT (Synaptic System, 131-004; 1:1,000), anti-gephyrin (Synaptic System, 147-008; 1:500), anti-FOS (Cell Signaling, 2250S; Synaptic System, 226-308; Synaptic System, 226-017; 1:500), anti-MEGF10 (Merck, ABC10; 1:500), anti-ERBB4 (Abcam, ab32375; Cell Signaling, 4795T; 1:500), anti-parvalbumin (Swant, GP72; 1:1,000), anti-beta amyloid (BioLegend, 803001; Cell Signaling, 2454S; 1:1,000), anti-GFAP (Abcam, ab4674; 1:1,000), anti-AXL (R&D systems, AF854; 1:100), anti-somatostatin (BMA Biomedicals, T-4103; 1:500), anti-VIP (Synaptic System, 443-005; 1:500), anti-NeuN (Sigma-Aldrich, ABN91; 1:500), anti-cleaved caspase-3 (Cell Signaling, 9661S; 1:500), anti-TREM2 (R&D systems, AF1729; 1:500), anti-HA (Cell Signaling, 3724S; 1:500) and anti-pS6 (Cell Signaling, 2211S; 1:500).

Secondary antibodies: donkey anti-goat IgG (H&L) Alexa Fluor 405 (Abcam, ab175665), donkey anti-goat IgG (H+L) Alexa Fluor 488 (Jackson Laboratory, 705-545-003), donkey anti-chicken DyLight 405-conjugated AffiniPure, donkey anti-chicken IgY (IgG) (H+L) (Jackson Laboratory, 703-475-155), donkey anti-chicken IgG (H+L) Alexa Fluor 488 (Jackson Laboratory, 703-545-155), donkey anti-chicken IgG (H+L) Alexa Fluor 594 (Jackson Laboratory, 709-585-155), donkey anti-rat IgG (H+L) Alexa Fluor 594 (Invitrogen, A-21209), donkey anti-rat IgG (H&L) Alexa Fluor 647 (Abcam, ab150155), donkey anti-rabbit IgG (H&L) Alexa Fluor 405 (Abcam, ab175649), donkey anti-rabbit IgG (H+L) Alexa Fluor 488 (Invitrogen, A-21206), donkey anti-rabbit IgG (H+L) Alexa Fluor 594 (Invitrogen, A-21207), donkey anti-guinea pig IgG (H+L) Alexa Fluor 488 (Jackson Laboratory, 706-545-148), donkey anti-guinea pig IgG (H+L) Alexa Fluor 594 (Jackson Laboratory, 706-585-148), donkey anti-guinea pig IgG (H+L) Alexa Fluor 647 (Jackson Laboratory, 706-605-148), donkey anti-sheep IgG H&L (Alexa Fluor 488) (Abcam, ab150177), donkey anti-mouse IgG (H+L) Highly Cross-Adsorbed Secondary Antibody, Alexa Fluor 594 (Invitrogen, A-21203). All Alexa-Fluor-488-conjugated secondary antibodies were used at a dilution of 1:1,000, whereas all other secondary antibodies were used at 1:500. The following reagents were used: HistoVT One (Nacalai, 06380-05), CNO (Sigma-Aldrich, C0832-5mg).

Python packages for data analysis

The following packages were used: Python v.3.10.0, numpy v.2.0.2, pandas v.2.2.3, matplotlib v.3.9.2, collections (built-in module in Python), scanpy v.1.10.4, scvi v.1.3.0, gseapy v.1.1.8, semopy v.2.3.11, pyWGCNA v.2.2.1, statsmodels v.0.14.5, scipy v.1.15.2, scikit-learn v.1.6.1 and magic-impute v.3.0.0.

snRNA-seq analysis

snRNA-seq data were analysed using Cell Ranger v.8.0.1. In brief, raw BCL files from Illumina HiSeq were demultiplexed to generate FASTQ files using ‘cellranger mkfastq’. FASTQ files were further processed with ‘cellranger count’, involving mapping to the mouse reference genome (mm10-2020-A), quantifying gene expression through unique molecular identifiers and cell barcodes, cell clustering and differential gene expression analysis. Multiple sequencing runs were combined using cellranger aggr.

Raw count matrices from 10x Genomics were imported into Scanpy v.1.10.4. Genes expressed in at least 3 cells and cells with 200–8,000 detected genes were used for downstream analysis. Cells with mitochondrial gene content over 5% were excluded. Doublets were removed by scanpy.pp.scrublet by default. After filtering, the selected count value was normalized by Pearson normalization, followed by highly-variable-gene identification with scanpy.experimental.pp.normalize_pearson_residuals and scanpy.experimental.pp.highly_variable_genes, respectively, in Scanpy. With Pearson-normalized counts and about 3,000 highly variable genes, we performed SysVI according to the package tutorial (https://docs.scvi-tools.org/en/stable/tutorials/notebooks/scrna/sysVI.html). Integrations with total cell clusters were performed with 100–200 epochs, whereas integrations with a single cell type such as excitatory neuron, astrocyte or microglia were performed with 5–30 epochs to obtain the minimum reconstruction loss between training set and validation set. Clustering and UMAP analysis were performed using scanpy.tl.leiden and scanpy.tl.umap by default. DEGs were obtained by scanpy.tl.rank_genes_groups using the Wilcoxon rank-sum method and then filtered by adjusted P value and log-transformed fold changes of their expressions. WGCNA was performed in Python with pyWGCNA v.2.2.1, as shown in the vignettes (https://github.com/mortazavilab/PyWGCNA/blob/main/tutorials/Quick_Start.ipynb). GSEA was performed with gseapy v.1.1.8 according to the tutorial (https://gseapy.readthedocs.io/en/latest/introduction.html). We used squared-rooted-normalized counts (numpy.sqrt(scanpy.pp.normalize_total(anndata, inplace=False)[‘X’])) for dot plots, UMAP plots and DEG analysis. We used magic-imputed counts⁵⁵ for visualization of violin plots, as shown in the vignettes (https://magic.readthedocs.io/en/stable/tutorial.html).

Analysed cell numbers

Cell numbers were as follows: Fig. 2a, 23,217 (3-month-old WT mice), 22,718 (2-month-old 5×FAD mice) and 33,623 (3-month-old 5×FAD mice) cells were analysed; Fig. 2b, 9,069 (3-month-old WT mice), 10,231 (2-month-old 5×FAD mice) and 14,044 (3-month-old 5×FAD mice) excitatory neurons were analysed. For each group, hippocampal samples were derived from two mice.

Figure 5a, 6,763 (sgControl) and 12,826 (sgErbb4) excitatory neurons; Fig. 5b, 3,199 (sgControl) and 2,840 (sgErbb4) microglia; Fig. 5c, 729 (sgControl) and 1,436 (sgErbb4) astrocytes; Fig. 5i, 8,570 (HA) and 10,442 (Erbb4) excitatory neurons; Fig. 5j, 1,037 (HA) and 1,450 (Erbb4) microglia; Fig. 5k, 1,857 (HA) and 1,267 (Erbb4) astrocytes. For each comparison, cells were derived from three mice per group.

Stereotaxic injection for adeno-associated virus transduction

All AAVs were produced in the laboratory as described previously⁷. In brief, we co-transfected pAAV9 capsid plasmid, helper plasmid for virus assembly and target plasmids into HEK293T (Korean Cell Line Bank) cells using a PEI-based (0.3 mg ml⁻¹) transfection method⁵⁶. HEK293T cells were maintained with FBS-containing (Gibco) Dulbecco’s modified Eagle’s medium (DMEM, Welgene), which was replaced with serum-free medium during transfection (6–12 h). Transfected cells were incubated in a 37 °C, 5% CO₂ conditioned cell incubator. Then, 72 h after medium replacement, culture medium was collected and fresh medium was added. Collected medium was stored at 4 °C. After 48 h, the culture medium and HEK293T cells were collected and purified using the polyethylene-glycol-mediated purification method⁵⁷. Purified AAVs were concentrated using a 100 kDa Amicon ultra centrifugal filter tube (Millipore) to 200 μl.

For sgRNA targeting Erbb4, the TTAGCGATATTCTTAAACTA sequence was cloned into the SaCas9 vector. The sequence GGTCGGGGCGTATGCGTCTA was cloned into the SaCas9 vector. For sgRNA targeting Rptor, the sequence TGCAGGTCGTATATGGACAG⁵⁸ was cloned into the SaCas9 vector. For control sgRNA, we used a non-targeting sgRNA sequence with no predicted target sites on the mouse genome, based on previously reported screening⁵⁹. For overexpression of Erbb4, the Erbb4 cDNA sequence was purchased from Sino Biological (MG51064) and cloned into an AAV target vector. For the kinase-dead mutant of Erbb4, Lys751 was mutated to methionine manually. For control shRNA, CATTGCTGGCACGAAGATTGAC and GTAGCAGAGCACCGTTTACATG were used. For shRNA targeting Megf10, TGAATCTTAAAAATGTGAATCC and GTTATTACAGAACCTAAGTGA were used.

Stereotaxic injection for delivering AAVs into the mouse brains was performed as described previously^7,37. In brief, mice were anaesthetized with isoflurane (Piramal) using a veterinary vaporizer (Surgivet). Heads were cleaned using 70% ethanol followed by hair removal and incision of the skin. For ExPre and chemogenetic approaches with hSyn promoter, AAVs were injected at CA3 (mediolateral (ML), −2.5 mm; anteroposterior (AP), −2.0 mm from bregma; dorsoventral (DV), −2.2 mm from the brain surface). For InhiPre, chemogenetic approaches with the Gad67 promoter, SaCas9 with sgRNAs, shRNAs, Erbb4 overexpression and oligomeric Aβ injections were performed at the CA1 (ML, −1.25 mm; AP, −2.0 mm from bregma; DV, −1.5 mm from the brain surface). We injected viruses bilaterally only for behavioural experiments. For oligomeric Aβ injection, mice were euthanized 2 days after injection. For Erbb4 overexpression experiments, mice were euthanized at 2 weeks after injection. For all other AAV-delivered experiments, mice were euthanized at 3 weeks after injection, or as otherwise reported in the figure legend. For chemogenetic approaches, CNO was injected at 0.5 mg per kg concentration for 4 consecutive days before euthanasia. We used same titre for each cohort. The incision was closed with Reflex 7 mm Wound Clips (ROBOZ) after injection.

Immunohistochemistry, FISH, antigen retrieval and image analysis

Mice were anaesthetized with avertin (20 μl g⁻¹) by intraperitoneal injection and perfused with 1× PBS (Welgene) followed by 4% PFA. Brains were post-fixed in 4% PFA at 4 °C overnight and transferred to 30% sucrose in 1× PBS for 48 h. Brains were embedded in OCT compound (Leica), sectioned into 30 μm coronal slices on cryo-stat microtomes (Leica). The sections were permeabilized by blocking buffer (4% BSA, 0.3% Triton X-100 in 1× PBS) for 1 h at room temperature followed by incubation with appropriate primary antibodies for 24 h at 4 °C. The sections were washed with PBST (0.1% Tween-20 in 1× PBS) and stained with the appropriate secondary antibodies conjugated with Alexa Fluor (Invitrogen, Abcam, Jackson ImmunoResearch) in PBST for 2 h at room temperature. The sections were washed and mounted onto slide glasses. TrueBlack (Biotium) diluted to 1/20 in 70% ethanol was applied to the sections for 2 min at room temperature to remove lipofuscin autofluorescence. The sections were washed with distilled water, and Vectashield with or without DAPI (Vector Lab) was used as the mounting medium. The samples were stored at 4 °C before imaging on a confocal laser-scanning microscope.

For immunohistochemistry analysis of ERBB4, we applied antigen retrieval before conventional immunohistochemistry. In brief, slices were placed in an Eppendorf tube filled with 500 μl 1× HistoVT One solution, and incubated at 70 °C for 20 min. After incubation, the slices were moved to a 24-well plate where conventional immunohistochemistry was performed.

For FISH, tissue sections were cut at a thickness of 10 μm and mounted onto SuperFrost Plus glass slides (Thermo Fisher Scientific). The sections were fixed in 4 °C PFA for 15 min, followed by dehydration through a graded ethanol series (50%, 70%, 100% and 100% ethanol; 5 min each) at room temperature. Human FFPE sections were deparaffinized before proceeding according to the same RNAscope workflow. After air-drying, FISH was performed using the RNAscope Multiplex Fluorescent Assay kits (Advanced Cell Diagnostics) according to the manufacturer’s instructions, with minor modifications. The slides were subjected to protease treatment for 30 min, followed by hybridization with target-specific RNAscope probes. All hybridization, amplification and wash steps were carried out manually at room temperature or at the temperatures specified by the manufacturer. Signal detection was achieved using Opal TSA fluorophores (Akoya Biosciences) diluted 1:1,000, depending on the probe and channel requirements. Nuclear counterstaining was performed using DAPI.

All confocal images from the brain sections were acquired using Zeiss LSM880 (×10 lens, ×20 lens or ×40 oil-immersion optical lens) for quantification as described below. Data indicate mice otherwise reported in figure legend.

To quantify Erbb4 mRNA levels in pyramidal neurons in mouse tissues, confocal single-plane images of Gad1 and Erbb4 within the pyramidal layer were isolated. Colocalization between Gad1 and Erbb4 was quantified using the DiAna plugin. To exclude Erbb4 mRNA originating from Gad1⁺ inhibitory neurons, the colocalized areas were subtracted from the total Erbb4⁺ areas.

To quantify ERBB4 mRNA levels in excitatory neurons in human tissue, confocal single-plane images of ERBB4, SLC17A7 and DAPI signals were analysed. SLC17A7⁺ excitatory neurons were identified by the presence of SLC17A7 puncta in DAPI⁺ nuclei. ERBB4 puncta colocalized with SLC17A7⁺DAPI⁺ signal were then measured using the DiAna⁶⁰ plugin.

To quantify synapse phagocytosis by glial cells, confocal single-plane images of GFP intensity subtracted from mCherry intensity (mCherry intensity − GFP intensity; mCherry alone) along with glial cells (S100β and IBA1) were separately isolated, and the areas of colocalization (mCherry alone⁺, glial cell⁺) were measured using the DiAna⁶⁰ plugin. To compensate for differences in the injected viruses, the colocalization areas were normalized to GFP areas.

To quantify synapse numbers, confocal single-plane images of excitatory synapses (vGLUT1 and PSD95) and inhibitory synapses (vGAT and gephyrin) were separately isolated, and the areas of themselves and their colocalization were measured using the Analyze Particle function in ImageJ and DiAna plugin, respectively.

To quantify the areas of GFAP, S100β, IBA1 and Ab, confocal single-plane images of these markers were isolated, and their areas were measured by Analyze Particle function.

To quantify the areas of pS6 in the pyramidal neurons and PV⁺ neurons, confocal single-plane images of pS6, PV and NeuN were isolated, and their colocalization (pS6⁺PV⁺NeuN⁺ or pS6⁺PV⁻NeuN⁺) was measured using the DiAna plugin.

To quantify the areas of DAM population (AXL⁺/TREM2⁺), confocal single-plane images of AXL, TREM2 and IBA1 were isolated, and their colocalization (AXL⁺IBA1⁺ or TREM2⁺IBA1⁺) was measured using the DiAna plugin.

To quantify MEGF10 in shRNA-applied samples, confocal single-plane images of MEGF10 and TagBFP were separately isolated, and the colocalization areas were measured using the DiAna plugin. To compensate for differences in the labelled astrocytes, the colocalization areas were normalized to TagBFP areas.

To quantify cleaved caspase-3⁺, FOS⁺ or ERBB4⁺ neurons, confocal z-stacked images of cleaved caspase-3, FOS or ERBB4 with neuronal markers (PV, SST or VIP with NeuN) were merged, and the numbers were manually counted.

Slice electrophysiology

Mice were anaesthetized with isoflurane and decapitated. The brains were quickly removed and transferred into ice-cold artificial cerebrospinal fluid (ACSF) continuously bubbled with 95% O₂ and 5% CO₂. The ACSF consisted of 125 mM NaCl, 2.5 mM KCl, 1.25 mM NaH₂PO₄, 25 mM NaHCO₃, 1 mM MgCl₂, 2 mM CaCl₂ and 15 mM glucose. Coronal slices were cut in ice-chilled ACSF using a vibratome (VT1200S, Leica). After slicing, the sections were allowed to recover in ACSF at 34 °C for 20 min, followed by incubation at room temperature for at least 40 min. For recordings, slices were placed into a submerged recording chamber and continuously perfused with oxygenated ACSF at a flow rate of 2–3 ml min⁻¹. The chamber temperature was maintained at 30–31 °C, and recordings were performed within 4–5 h after the recovery period. Neurons were visualized using infrared differential interference contrast microscopy with an upright microscope (BX51WI, Olympus). Whole-cell voltage-clamp recordings were obtained with borosilicate glass pipettes (2.5–3.5 MΩ) filled with a Cs⁺-based low-Cl⁻ internal solution containing 135 mM CsMeSO₃, 10 mM HEPES, 1 mM EGTA, 3.3 mM QX-314, 0.1 mM CaCl₂, 4 mM Mg-ATP, 0.3 mM Na₃-GTP, 8 mM Na₂-phosphocreatine (290–300 mOsm, pH 7.3 adjusted with CsOH). Recordings were included in the analysis only when the access resistance was between 10–20 MΩ and remained stable, with less than 20% change throughout the experiment. Whole-cell patch-clamp recordings were acquired using the Multiclamp 700B (Molecular Devices) system and signals were filtered at 2 kHz and digitized at 10 kHz (NI PCIe-6259, National Instruments). Data were monitored and acquired by WinWCP (Strathclyde Software), further analysed using Clampfit v.10.7 (Molecular Devices) and OriginPro 2017 (OriginLab). To measure the mEPSCs and mIPSCs, Nav channel blocker (TTX, 500 nM, Alomone Labs, T-550) and NMDA receptor antagonist (D-AP5, 25 μM, Tocris, 0106) were added to ACSF and applied throughout the recording session. Both mEPSCs and mIPSCs were recorded in the same cell. mEPSCs were obtained at −70 mV (reversal potential of chloride), and mIPSCs were subsequently obtained at 0 mV (reversal potential of ionotropic glutamate receptors). mEPSCs and mIPSCs were recorded for 3 min each. mEPSCs were analysed over the entire 3 min recording period, whereas mIPSCs were analysed during a 60 s segment from 60 s after the start of recording.

For intrinsic excitability recordings, coronal slices containing the dorsal hippocampal CA1 (300 µm) were cut in ice-cold oxygenated sucrose-based cutting solution containing 75 mM sucrose, 76 mM NaCl, 2.5 mM KCl, 25 mM NaHCO₃, 25 mM glucose, 1.25 mM NaH₂PO₄, 7 mM MgSO₄, 0.5 mM CaCl₂ with pH 7.3, equilibrated with 95% O₂ and 5% CO₂, using a vibratome (Leica, VT1200S), and then recovered in the same solution for 30 min at 32 °C. Slices were then transferred to an incubation chamber filled with oxygenated ACSF containing 124 mM NaCl, 2.5 mM KCl, 1.3 mM MgCl₂, 2.5 mM CaCl₂, 1.0 mM NaH₂PO₄, 26.2 mM NaHCO₃ and 20 mM glucose with pH 7.4 at room temperature and slices were kept for less than 6 h before recordings.

For patch-clamp recordings, slices were transferred to a recording chamber perfused with oxygenated ACSF at 30–32 °C controlled by a peristaltic pump. Patch microelectrodes were pulled from borosilicate glass (Harvard Apparatus, 30-0065) using a micropipette puller (Narishige, PC-10). Patch microelectrodes had a resistance of 5.0–7.0 MΩ. Signals were recorded using a patch-clamp amplifier (Multiclamp 700B, Molecular Devices) and digitized with Digidata 1550 digitizer (Molecular Devices) using Clampex software (Molecular Devices). Signals were amplified, filtered at 2 kHz and sampled at 10 kHz.

In current-clamp recordings, the membrane potential was held at −70 mV with intracellular solution: 135 mM K-gluconate, 7 mM NaCl, 10 mM HEPES, 0.5 mM EGTA, 2 mM Mg-ATP, 0.3 mM Na₂-GTP and 10 mM Na-phosphocreatine with pH 7.3 and 295 mOsm. Current-clamp experiments were recorded 5 min after obtaining whole-cell configuration. To evaluate intrinsic excitability, 500 ms depolarizing currents were injected from −200 to 500 pA with increments of 50 pA, and the mean firing rate was calculated based on the number of evoked action potentials in response to a depolarizing current injection. The input resistance (R_in) was estimated as the slope of the I–V relationship derived by measuring the difference between the baseline and the steady-state.

Behavioural test

Both novel-object recognition test and novel-object location test were performed in a square box (acrylic box, 30 cm × 30 cm × 28 cm, W × D × H, custom-made). One side of the box has a stripe to give information about location. Both tasks are composed of three steps: habituation, training and testing.

For the novel-object location test, mice were habituated to the testing box by freely moving in the open-field arena for 10 min. Then, 24 h after habituation, the mice were placed into the box, which contained two identical objects, for 10 min for training. Then, 24 h after training, the mice were again placed in the box, in which one object was moved to a new location for 10 min.

For the novel-object recognition test, mice were habituated under the same conditions as for the novel-object location test. At the training step, mice were exposed to two identical objects. For the testing step, one of the two objects was replaced with a novel object.

For spontaneous alternation, mice were placed on the centre of a Y-maze (acrylic box, 5 cm × 30 cm × 15 cm, W × D × H for one arm, custom-made) for 8 min.

To quantify performance in the novel-object location test and novel-object recognition test, the time spent interacting with the objects for both the old and novel object and location were measured. Then, the discrimination index was calculated according to the following equation:

For quantification of spontaneous alternation, arm entry for all arms was quantified and alternation percentages were calculated by the following equation:

where spontaneous alternation was defined as sequential arm entry for all arms.

For the Barnes maze, we followed the shortened protocol as described previously⁴¹. In brief, the maze consisted of a circular white platform (1 m diameter) elevated approximately 1 m above the floor, with 20 evenly spaced holes along the perimeter, one of which led to a dark escape cage. The task was performed in three phases: habituation, training and probe. On the habituation day, the mice were placed at the centre of the maze for 30 s and then gently guided to the target hole to allow familiarization with the escape cage. During the training phase, the mice were placed inside a transparent cylinder at the centre of the maze for 10 s, after which the cylinder was removed. Mice were allowed to explore the maze for up to 2 min per trial to locate the target hole. Training consisted of five trials over 2 days (three trials on day 1 and two trials on day 2), with an inter-trial interval of approximately 20 min. If a mouse failed to enter the escape cage within the allotted time, it was gently guided to the target hole. Primary latency, defined as the time from trial onset to the first nose poke at the target hole, was recorded as a measure of spatial learning. One day after final training session, a probe trial was conducted in which the escape cage was removed, and mice were allowed to explore the maze for 2 min. Time spent in each quadrant was recorded as a measure of spatial memory.

All behavioural tests were performed by researchers blinded to the groups assignments.

Illustrations

All illustrations were generated using a BioRender with a licence (https://www.biorender.com).

Software and statistical analysis

Zen (Zeiss) software acquisition system was used for acquisition of confocal images. For image analysis, ImageJ (NIH) and its plugin DiAna⁶⁰ were used.

No statistical methods were used to predetermine sample sizes. All statistical analyses were performed using GraphPad Prism 10 with 95% confidence. The Shapiro–Wilk test was first performed for all data to determine normality of the data distribution. If the data passed normality testing, comparisons between two groups were performed using two-tailed unpaired Student’s t-tests. For data that did not follow a normal distribution, the Mann–Whitney U-test was used. For comparisons of markers measured from different mouse brains, unpaired tests were applied, whereas for measurements obtained from the same mouse brain following bilateral injection of different AAVs, paired tests were used. For comparison of more than three groups with normal distribution, one-way ANOVA followed by Tukey’s multiple-comparison test was used. For comparison of more than two groups with two independent variables, two-way ANOVA followed by Tukey’s multiple-comparison test was used. For comparison of cell numbers in the excitatory neuronal cluster (Extended Data Fig. 5a,f), the generalized estimating equation test was performed using statsmodels v.0.14.5 in Python v.3.10. For comparison of excitability (Extended Data Fig. 7v), a linear mixed model was used and corrected using the ‘fdr_bh’ method in statsmodels and scipy v.1.15.2 in Python. For directed mediation analysis, statsmodels and semopy v.2.3.11 were used in Python. Linear regression was used by scikit-learn v.1.6.1 in Python. The statistical test used for each experiment is reported in the results.

Ethics statement

All animal experiments were performed in accordance with protocols approved by the Institutional Animal Care and Use Committees of the Korea Advanced Institute of Science and Technology and the Institute for Basic Science. Human FFPE hippocampal sections were provided by the SNUH Brain Bank. Their use in this study was reviewed and determined to be exempt from protocol review by the Public Institutional Review Board designated by the Ministry of Health and Welfare, Republic of Korea. All procedures involving human-derived materials were conducted in accordance with relevant regulations.

Reporting summary

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

1. Cameron, B. & Landreth, G. E. Inflammation, microglia, and Alzheimer’s disease. Neurobiol. Dis. 37, 503–509 (2010).

   Article  CAS  PubMed  Google Scholar 

2. Keren-Shaul, H. et al. A unique microglia type associated with restricting development of Alzheimer’s disease. Cell 169, 1276–1290 (2017).

   Article  ADS  CAS  PubMed  Google Scholar 

3. Hong, S. et al. Complement and microglia mediate early synapse loss in Alzheimer mouse models. Science 352, 712–716 (2016).

   Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

4. Crowley, G. et al. C1q and immunoglobulins mediate activity-dependent synapse loss in the adult brain. Science 393, eadv1219 (2026).

   Article  CAS  PubMed  Google Scholar 

5. Arendt, T. Synaptic degeneration in Alzheimer’s disease. Acta Neuropathol. 118, 167–179 (2009).

   Article  PubMed  Google Scholar 

6. Koffie, R. M., Hyman, B. T. & Spires-Jones, T. L. Alzheimer’s disease: synapses gone cold. Mol. Neurodegener. 6, 63 (2011).

   Article  PubMed  PubMed Central  Google Scholar 

7. Lee, J.-H. et al. Astrocytes phagocytose adult hippocampal synapses for circuit homeostasis. Nature 590, 612–617 (2021).

   Article  ADS  CAS  PubMed  Google Scholar 

8. Katayama, H. et al. Visualizing and modulating mitophagy for therapeutic studies of neurodegeneration. Cell 181, 1176–1187 (2020).

   Article  CAS  PubMed  Google Scholar 

9. Udakis, M., Pedrosa, V., Chamberlain, S. E. L., Clopath, C. & Mellor, J. R. Interneuron-specific plasticity at parvalbumin and somatostatin inhibitory synapses onto CA1 pyramidal neurons shapes hippocampal output. Nat. Commun. 11, 4395 (2020).

   Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

10. Busche, M. A. et al. Critical role of soluble amyloid-β for early hippocampal hyperactivity in a mouse model of Alzheimer’s disease. Proc. Natl Acad. Sci. USA 109, 8740–8745 (2012).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

11. Prince, S. M. et al. Alzheimer’s pathology causes impaired inhibitory connections and reactivation of spatial codes during spatial navigation. Cell Rep. 35, 109008 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

12. Xu, Y., Zhao, M., Han, Y. & Zhang, H. GABAergic inhibitory interneuron deficits in Alzheimer’s disease: implications for treatment. Front. Neurosci. 14, 660 (2020).

    Article  PubMed  PubMed Central  Google Scholar 

13. Chung, W.-S. et al. Astrocytes mediate synapse elimination through MEGF10 and MERTK pathways. Nature 504, 394–400 (2013).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

14. Schafer, D. P. et al. Microglia sculpt postnatal neural circuits in an activity and complement-dependent manner. Neuron 74, 691–705 (2012).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

15. Hrovatin, K. et al. Integrating single-cell RNA-seq datasets with substantial batch effects. BMC Genomics https://doi.org/10.1186/s12864-025-12126-3 (2025).

16. Langfelder, P. & Horvath, S. WGCNA: an R package for weighted correlation network analysis. BMC Bioinform. 9, 559 (2008).

    Article  Google Scholar 

17. Rezaie, N., Reese, F. & Mortazavi, A. PyWGCNA: a Python package for weighted gene co-expression network analysis. Bioinformatics 39, btad415 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

18. Zhou, Y. et al. Human and mouse single-nucleus transcriptomics reveal TREM2-dependent and TREM2-independent cellular responses in Alzheimer’s disease. Nat. Med. 26, 131–142 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

19. Bernard, C. et al. Cortical wiring by synapse type-specific control of local protein synthesis. Science 378, eabm7466 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

20. Zhang, P., Hirsch, E. C., Damier, P., Duyckaerts, C. & Javoy-Agid, F. c-fos protein-like immunoreactivity: distribution in the human brain and over-expression in the hippocampus of patients with Alzheimer’s disease. Neuroscience 46, 9–21 (1992).

    Article  CAS  PubMed  Google Scholar 

21. Vadisiute, A. et al. Glial cells undergo rapid changes following acute chemogenetic manipulation of cortical layer 5 projection neurons. Commun. Biol. 7, 1286 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

22. Tamboli, S. et al. Mouse hippocampal CA1 VIP interneurons detect novelty in the environment and support recognition memory. Cell Rep. 43, 114115 (2024).

    Article  CAS  PubMed  Google Scholar 

23. Hijazi, S. et al. Early restoration of parvalbumin interneuron activity prevents memory loss and network hyperexcitability in a mouse model of Alzheimer’s disease. Mol. Psychiatry 25, 3380–3398 (2020).

    Article  CAS  PubMed  Google Scholar 

24. Verret, L. et al. Inhibitory interneuron deficit links altered network activity and cognitive dysfunction in Alzheimer model. Cell 149, 708–721 (2012).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

25. Yang, D. et al. Modulation of mTOR and epigenetic pathways as therapeutics in gallbladder cancer. Mol. Ther. Oncolyt. 20, 59–70 (2021).

    Article  CAS  Google Scholar 

26. Cao, S. et al. ErbB4 regulate extracellular dopamine through the p38 MAPK signaling pathway. Neurosci. Lett. 751, 135830 (2021).

    Article  CAS  PubMed  Google Scholar 

27. Ramachandra, C. J. A., Mehta, A., Wong, P. & Shim, W. ErbB4 activated p38γ MAPK isoform mediates early cardiogenesis through NKx2.5 in human pluripotent stem cells. Stem Cells 34, 288–298 (2016).

    Article  CAS  PubMed  Google Scholar 

28. Kassai, H. et al. Selective activation of mTORC1 signaling recapitulates microcephaly, tuberous sclerosis, and neurodegenerative diseases. Cell Rep. 7, 1626–1639 (2014).

    Article  CAS  PubMed  Google Scholar 

29. Iyer, A. M. et al. mTOR hyperactivation in Down syndrome hippocampus appears early during development. J. Neuropathol. Exp. Neurol. 73, 671–683 (2014).

    Article  CAS  PubMed  Google Scholar 

30. Afshar-Saber, W. et al. Neuronal hyperactivity becomes mTORC1 independent due to transcriptional changes in tuberous sclerosis complex disease models. Cell Rep. 44, 116664 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

31. Gassmann, M. et al. Redistribution of GABA_B(1) protein and atypical GABA_B responses in GABA_B(2)-deficient mice. J. Neurosci. 24, 6086–6097 (2004).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

32. Peng, Z., Huang, C. S., Stell, B. M., Mody, I. & Houser, C. R. Altered expression of the δ subunit of the GABA_A receptor in a mouse model of temporal lobe epilepsy. J. Neurosci. 24, 8629–8639 (2004).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

33. Martínez-François, J. R. et al. BAD and KATP channels regulate neuron excitability and epileptiform activity. eLife 7, e32721 (2018).

    Article  PubMed  PubMed Central  Google Scholar 

34. Bond, C. T. Small conductance Ca2⁺-activated K⁺ channel knock-out mice reveal the identity of calcium-dependent afterhyperpolarization currents. J. Neurosci. 24, 5301–5306 (2004).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

35. Xu, J. et al. Multimodal single-cell/nucleus RNA sequencing data analysis uncovers molecular networks between disease-associated microglia and astrocytes with implications for drug repurposing in Alzheimer’s disease. Genome Res. 31, 1900–1912 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

36. Habib, N. et al. Disease-associated astrocytes in Alzheimer’s disease and aging. Nat. Neurosci. 23, 701–706 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

37. Jung, H. et al. Anti-inflammatory clearance of amyloid-β by a chimeric Gas6 fusion protein. Nat. Med. 28, 1802–1812 (2022).

    Article  CAS  PubMed  Google Scholar 

38. Hampel, H. et al. The amyloid-β pathway in Alzheimer’s disease. Mol. Psychiatry 26, 5481–5503 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

39. Creighton, S. D. et al. Dissociable cognitive impairments in two strains of transgenic Alzheimer’s disease mice revealed by a battery of object-based tests. Sci Rep. 9, 57 (2019).

    Article  ADS  PubMed  PubMed Central  Google Scholar 

40. Roy, E. R. et al. Concerted type I interferon signaling in microglia and neural cells promotes memory impairment associated with amyloid β plaques. Immunity 55, 879–894 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

41. Attar, A. et al. A shortened Barnes maze protocol reveals memory deficits at 4-months of age in the triple-transgenic mouse model of Alzheimer’s disease. PLoS ONE 8, e80355 (2013).

    Article  ADS  PubMed  PubMed Central  Google Scholar 

42. Bennett, D. A., Schneider, J. A., Arvanitakis, Z. & Wilson, R. S. Overview and findings from the Religious Orders Study. Curr. Alzheimer Res. 9, 628–645 (2012).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

43. Bennett, D. A. et al. Overview and findings from the rush memory and aging project. Curr. Alzheimer Res. 9, 646–663 (2012).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

44. Mathys, H. et al. Single-cell atlas reveals correlates of high cognitive function, dementia, and resilience to Alzheimer’s disease pathology. Cell 186, 4365–4385 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

45. Betthauser, T. J. et al. Amyloid and tau imaging biomarkers explain cognitive decline from late middle-age. Brain 143, 320–335 (2020).

    Article  PubMed  PubMed Central  Google Scholar 

46. Green, G. S. et al. Cellular communities reveal trajectories of brain ageing and Alzheimer’s disease. Nature 633, 634–645 (2024).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

47. Jack, C. R. et al. Longitudinal tau PET in ageing and Alzheimer’s disease. Brain 141, 1517–1528 (2018).

    Article  PubMed  PubMed Central  Google Scholar 

48. Sevigny, J. et al. The antibody aducanumab reduces Aβ plaques in Alzheimer’s disease. Nature 537, 50–56 (2016).

    Article  ADS  CAS  PubMed  Google Scholar 

49. Knopman, D. S., Jones, D. T. & Greicius, M. D. Failure to demonstrate efficacy of aducanumab: an analysis of the EMERGE and ENGAGE trials as reported by Biogen, December 2019. Alzheimers Dement. 17, 696–701 (2021).

    Article  PubMed  Google Scholar 

50. Kovbasyuk, Z., Tefera, E., Li, C., Baete, S. H. & Steriade, C. Imaging brain inflammation and blood brain barrier permeability in neurological and psychiatric diseases: a review. J. Neuroinflamm. 22, 275 (2025).

    Article  CAS  Google Scholar 

51. Li, S. et al. Soluble Aβ oligomers inhibit long-term potentiation through a mechanism involving excessive activation of extrasynaptic NR2B-containing NMDA receptors. J. Neurosci. 31, 6627–6638 (2011).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

52. Sondag, C. M., Dhawan, G. & Combs, C. K. Beta amyloid oligomers and fibrils stimulate differential activation of primary microglia. J. Neuroinflamm. 6, 1 (2009).

    Article  Google Scholar 

53. Rexach, J. E. et al. Cross-disorder and disease-specific pathways in dementia revealed by single-cell genomics. Cell 187, 5753–5774 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

54. Paryani, F. et al. Multi-omic analysis of Huntington’s disease reveals a compensatory astrocyte state. Nat. Commun. 15, 6742 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

55. Van Dijk, D. et al. Recovering gene interactions from single-cell data using data diffusion. Cell 174, 716–729 (2018).

    Article  PubMed  PubMed Central  Google Scholar 

56. Fischer, D., Bieber, T., Li, Y., Elsässer, H.-P. & Kissel, T. A novel non-viral vector for DNA delivery based on low molecular weight, branched polyethylenimine: effect of molecular weight on transfection efficiency and cytotoxicity. Pharm. Res. 16, 1273–1279 (1999).

    Article  CAS  PubMed  Google Scholar 

57. Challis, R. C. et al. Systemic AAV vectors for widespread and targeted gene delivery in rodents. Nat. Protoc. 14, 379–414 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

58. Long, L. et al. CRISPR screens unveil signal hubs for nutrient licensing of T cell immunity. Nature 600, 308–313 (2021).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

59. Morgens, D. W. et al. Genome-scale measurement of off-target activity using Cas9 toxicity in high-throughput screens. Nat. Commun. 8, 15178 (2017).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

60. Gilles, J.-F., Dos Santos, M., Boudier, T., Bolte, S. & Heck, N. DiAna, an ImageJ tool for object-based 3D co-localization and distance analysis. Methods 115, 55–64 (2017).

    Article  CAS  PubMed  Google Scholar 

Download references

Author notes

1. These authors contributed equally: Eunseok Park, Ha-Eun Lee

Authors and Affiliations

1. Center for Vascular Research, Institute for Basic Science (IBS), Daejeon, Republic of Korea

   Se Young Lee, Eunseok Park, Juwon Park, Young-Jin Choi & Won-Suk Chung

2. Department of Biological Sciences, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Republic of Korea

   Se Young Lee, Eunseok Park, Seongbin Kim, Yeji Yeo, Juwon Park, Young-Jin Choi, Kiheon Lee, Ki-Jun Yoon, Eunjoon Kim & Won-Suk Chung

3. Department of Biological Sciences, Ulsan National Institute of Science and Technology (UNIST), Ulsan, Republic of Korea

   Ha-Eun Lee & Jae-Ick Kim

4. Center for Synaptic Brain Dysfunctions, Institute for Basic Science (IBS), Daejeon, Republic of Korea

   Seongbin Kim, Yeji Yeo & Eunjoon Kim

5. Illimis Therapeutics, Seoul, Republic of Korea

   Sanghoon Park

Contributions

W.-S.C. and S.Y.L. designed projects and wrote the manuscript. H.-E.L. performed mEPSC and mIPSC recording. S.K. and Y.Y. performed intrinsic excitability recording. K.L., K.-J.Y., S.P., E.K. and J.-I.K. provided reagents, mice and discussion. Y.-J.C. designed shRNAs for Megf10. J.P. performed immunohistochemistry. E.P. performed behaviour experiments and FISH. S.Y.L. performed all other experiments including cloning, stereotaxic injection, scRNA-seq analysis and virus production.

Corresponding author

Correspondence to Won-Suk Chung.

Competing interests

S.Y.L. and W.-S.C. have filed a patent application based on this work. The other authors declare no competing interests.

Peer review information

Nature thanks Chun Chen, Sudeshna Das, Jorge Palop, Morgan Sheng and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions