Targeting notch signaling to restore neural development and behavior in mouse models of ASD

Abstract

Autism spectrum disorder (ASD) is a neurodevelopmental disorder with diverse genetic and environmental origins, yet whether these factors converge on common molecular pathways remains unclear. This study identifies dysregulation of the Notch signaling pathway as a shared mechanism in both hereditary and nonhereditary ASD models. Aberrant histone deacetylase 3-mediated epigenetic regulation of Notch signaling during embryonic forebrain development disrupts the specification of vasoactive intestinal peptide (VIP + ) GABAergic interneuron subtypes (VIP-INs), which originate in the caudal ganglionic eminence (CGE). CGE-specific ablation of Notch1/2 genes in ASD models restores the loss of VIP-INs, normalizes maladaptive excitatory and inhibitory balance, and selectively improves social behaviors. A single antenatal dose of a γ-secretase inhibitor ameliorates multiple ASD-associated neuronal, behavioral, and transcriptomic changes in adult models. The study indicates a strong convergence of ASD-related factors on Notch signaling dysregulation and establishes this pathway as a promising therapeutic target for developmental and behavioral deficits in ASD.

Here authors demonstrate embryonic disruption of Notch signaling impairs the development of specific inhibitory neuron subtypes, leading to autism-like behaviors. Modulating aberrant Notch activity restores circuit balance and behavior.

Introduction

Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder characterized by deficits in social communication and language, and repetitive behaviors. While approximately 60% of ASD cases are considered to have a genetic basis^1,2, the remaining cases arise from a complex interplay between genetic susceptibility and environmental influences, such as prenatal exposure to anticonvulsants, infections, or maternal nutritional deficiencies, which can significantly increase the incidence of ASD in offspring^3–6. Despite extensive genetic studies identifying de novo and inherited variants in approximately 15% of ASD cases^7,8, the underlying molecular mechanisms remain poorly understood.

Epigenetic modifications, such as histone acetylation and DNA methylation, are emerging as key regulators linking genetic and environmental risk factors of ASD during neurodevelopment^9–11. Mutations in chromatin regulators, such as chromodomain-helicase-DNA-binding protein 8 (CHD8) and methyl-CpG binding protein 2 (MeCP2), have been strongly linked to ASD^12–14, while environmental risk factors of ASD, such as prenatal exposure to valproic acid (VPA), affect chromatin structure, leading to widespread neurodevelopmental defects¹⁵. These findings underscore the role of epigenetic dysregulation in ASD; however, the specific downstream genes that are affected and how they contribute to ASD pathology remain unclear. Elucidating these pathways is a key to improving our understanding of ASD mechanisms.

A core neuropathological feature of ASD is an imbalance between excitatory and inhibitory (E/I) neural signaling¹⁶, which disrupts cortical circuit function and social behaviors. Many ASD models exhibit E/I imbalance^17–22. In humans, many individuals with ASD also have epilepsy, further confirming the link between ASD and hyperexcitation of neural circuits^23,24. A direct causal relationship between E/I imbalance and social deficits has been demonstrated in multiple ASD models^25,26. Studies have shown that restoring GABAergic function can alleviate some behavioral abnormalities in ASD models^18,19,27. Given the diverse genetic and environmental factors implicated in ASD, we hypothesize that these ASD-related factors would share molecular pathways that influence neuronal development and E/I balance. Elucidating such a common mechanism could provide therapeutic insights into ASD pathophysiology.

In this study, we uncover a convergent mechanism that links ASD heterogeneity, epigenetic dysregulation, and E/I imbalance. Our findings reveal a key pathogenic process involving dysregulation of the Notch signaling pathway—a key regulator of cell fate decisions during brain development^28–30. Enhanced Notch signaling, primarily driven by epigenetic dysregulation, disrupts the differentiation of caudal ganglionic eminence (CGE)-derived GABAergic neurons, including cortical vasoactive intestinal peptide (VIP⁺) interneurons (VIP-INs), which play a crucial role in maintaining E/I balance^31,32. Given the recent evidence that VIP-IN dysfunction is implicated in an ASD-related disorder, Dravet syndrome³³, our results suggest that notch-mediated VIP-IN deficits are a key driver of ASD pathophysiology. We demonstrate that genetic or pharmacological inhibition of Notch signaling rescues neural deficits and alleviates ASD-like behaviors, highlighting its potential as a therapeutic target across diverse ASD etiologies.

Results

Enhanced notch signaling pathway is a shared molecular signature across developing ASD models

To investigate the molecular pathogenesis underlying ASD, we conducted transcriptome analysis of the fetal brain from two nongenetic rodent models of ASD: in utero exposure to the environmental risk factors, VPA at E13 or polyinosinic:polycytidylic acid [poly(I:C)] at E12. Ribonucleic acid sequencing (RNA-seq) analysis identified significant gene expression changes, with 2734 altered genes (1520 upregulated, 1214 downregulated) in VPA-exposed brains and 1634 altered genes (906 upregulated, 728 downregulated) in poly(I:C)-exposed brains. Among these, 442 genes were commonly upregulated across both models, representing approximately 50% of the total upregulated genes in poly(I:C) mice (Fig. 1a).

Fig. 1. Notch signaling is hyperactivated in nongenetic and human genetic ASD models.

a Transcriptomic profiling of ASD models. Mouse embryos from nonhereditary ASD models were generated by in utero exposure to VPA (E13) or poly(I:C) (E12). RNA-sequencing (RNA-seq) was performed on forebrains (E14) to assess transcriptomic changes (n = 3, p < 0.05). b Gene ontology (GO) analysis by Metascape identified shared and distinct altered gene clusters in VPA- and poly(I:C)-exposed embryonic brains. c Volcano plots showing significant changes in genes associated with the Notch signaling pathway. a–c Two-sided unpaired t-test with Benjamini–Hochberg multiple testing correction. d Heatmaps depicting dysregulated Notch pathway genes in VPA and poly(I:C) models (n = 3, Control: C1–C3, VPA: V1–V3, poly(I:C): P1–P3). e Analysis of human ES-derived neurons with three ASD-related copy number variations (CNVs) (16p11 deletion, 1q21.1 deletion and 1q21.1 duplication) by pseudo-bulk analysis using single-cell RNA sequencing (scRNA-seq) datasets. Heatmaps of log2 fold change illustrate altered expression of Notch pathway genes in these models.

Gene ontology (GO) analysis of the upregulated genes in both VPA and poly(I:C) models highlighted significant enrichment in terms of epigenetic regulation and neural development. Chromatin-related terms such as “Chromatin organization,” “Chromatin remodeling,” and “Chromatin binding” emerged as the frequent common GO terms with high signal (Fig. 1b and Supplementary Fig. 1a, b), supporting the hypothesis that epigenetic dysregulation plays a central role in ASD pathogenesis. The extensive gene upregulation by environmental ASD-related factors may primarily disrupt epigenetic regulation, affecting downstream pathways critical for neural development. Notably, the term “Notch signaling pathway” was enriched in both ASD models (Fig. 1b and Supplementary Fig. 1a, b). As a conserved mechanism for cell fate decisions during development^28,29, Notch signaling appeared to be specifically activated in these ASD models. Indeed, upregulation of key components, including Notch1, Notch2, and the downstream transcription factor Hes5, was confirmed in both models (Fig. 1c and Supplementary Fig. 1c). Heatmap analysis exhibited a hyperactivation of multiple Notch pathway genes across both ASD models compared with controls (Fig. 1d). Notch signaling pathway had strong functional associations with chromatin regulation in VPA-exposed brains (Supplementary Fig. 2a).

To extend these findings to human ASD models, we examined the expression of the Notch pathway gene in transcriptomic data from neuronal cells derived from human embryonic stem cells (hESCs) carrying ASD-associated copy number variants (CNVs). Interestingly, expression of multiple Notch pathway genes was enhanced in the 16p11.2 deletion model³⁴, a most frequent CNV in hereditary ASD, but not in 1q21.1 deletion/duplication models³⁵ (Fig. 1e). Additionally, meta-analysis of ASD-associated human ESCs /iPSCs revealed that Notch pathway-associated genes were activated in MeCP2 R168X mutant³⁶ and tuberous sclerosis complex knockout (TSC KO)³⁷ cell lines (Supplementary Fig. 2b and Supplementary Data 1). These results suggest the possibility that Notch pathway dysregulation is a molecular hallmark in some cases of hereditary ASD. Our findings reveal that aberrant Notch signaling is a shared molecular signature in developing brains in both nonhereditary and certain hereditary ASD models, providing important insights into the common mechanisms underlying ASD pathogenesis.

Cortical VIP-IN deficits are a common pathology in nonhereditary ASD models

To investigate the postnatal consequences of prenatal Notch signaling dysregulation, we focused on its relationship with E/I imbalance—a hallmark phenotype in ASD. Despite its significance, the molecular basis of E/I imbalance remains poorly understood.

Using an established in vitro ASD model³⁸, where temporal VPA exposure at the initial stage of cortical culture induces E/I imbalance, we performed transcriptomic analysis to identify VPA-susceptible genes (Supplementary Fig. 3a). These genes were significantly enriched in ASD-associated genes listed in the Simons Foundation Autism Research Initiative (SFARI) Gene database (Supplementary Fig. 3b). Exon array analysis revealed significant changes in 126 genes (64 upregulated, 62 downregulated) (Supplementary Data 2). The transcript with the highest fold-change among downregulated genes was Vip, expressed in a cortical GABAergic interneuron subtype. Additionally, the top 10 downregulated genes included transcripts related to GABAergic interneuron subtypes, such as Neuron-derived neurotrophic factor (Ndnf), Reelin (Reln), and Somatostatin (Sst) (Fig. 2a and Supplementary Fig. 3c). Similarly, cell-type proportion analysis indicated the largest reduction of GABAergic interneurons (Fig. 2b). These drastic and selective effects on VIP-INs were confirmed by quantitative analysis of transcripts and immunostaining for markers of cortical interneuron subtypes (Supplementary Fig. 4a, b). Together with previously reported morphological and physiological evidence³⁸, these findings emphasize the significant impact of VPA exposure on E/I balance in vitro.

Fig. 2. ASD models exhibit selective alteration in GABAergic VIP-IN subtypes and their precursors in CGE.

a Transcriptomic analysis of the in vitro ASD model. Exon microarray analysis (n = 3) identified the top 20 altered genes, mapped to cell types using the Allen Brain Map. b Cell type-specific changes in the in vitro ASD models. Estimated percentages of altered gene expression were determined using Bisque⁶⁹ (n = 3). Box plots show the median (centre line), the 25th and 75th percentiles (bounds of the box). c Selective reduction in a GABAergic interneuron subtype in the VPA-exposed mouse ASD model. Representative coronal images of the somatosensory cortex (P21) and quantification of PV-INs (control = 21, VPA = 21, p = 0.7852), SST-INs (control = 16, VPA = 37, p = 0.5431), Reelin-INs (control = 21, VPA = 32, p = 0.2494), and VIP-INs (control = 21, VPA = 21, p < 0.0001). Scale bar: 100 µm. d Reduced cell density of VIP-INs and PV-INs in poly(I:C)-exposed mice. (control = 65, poly(I:C) = 86, p < 0.0001). Scale bar: 100 µm. e Illustration of developmental origins of cortical GABAergic interneurons. Cortical GABAergic interneurons arise from the medial ganglionic eminence (MGE) and caudal ganglionic eminence (CGE). Postmitotic CGE cells then migrate tangentially to the cortical zone. f, g Increased neural precursors in CGE of VPA-exposed embryonic brains. f Representative coronal images (E14) of CGE immunostaining with PROX1 and COUP-TFII antibodies. g Quantification of PROX1 (control = 45, VPA = 53, p = 0.9255), COUP-TFII (control = 34, VPA = 30, p = 0.0009), SOX6 (control = 17, VPA = 21, p = 0.2654), and LHX6 (control = 16, VPA = 22, p = 0.3403) expression levels in the CGE or MGE. Scale bar: 200 µm. h–i Increased proliferation in CGE of VPA and poly(I:C) models. h Representative coronal images (E14) of CGE EdU staining. i Quantification of EdU⁺ cell areas indicates significant increase in VPA (control = 52, VPA = 58, p < 0.0001) and poly(I:C) models (control=55, poly(I:C) = 62, p < 0.0001). Scale bar: 200 µm. j, k Reduced migration of CGE-derived cells in VPA-exposed mice. j Representative coronal images of the neocortex (P0) showing EdU⁺ cells. Arrowhead indicates EdU⁺/PROX1⁺, EdU⁺/COUP-TFII⁺, or EdU⁺/SOX6⁺ cells. k Quantification of EdU⁺/PROX1⁺ (control = 36, VPA = 31, p = 0.0027), EdU⁺/COUP-TFII⁺ (control = 25, VPA = 30, p = 0.0043), and EdU⁺/SOX6⁺ (control = 21, VPA = 19, p = 0.5643) neurons. Scale bar: 50 µm. c, d, g, i, k Data are presented as mean values +/− SEM, two-tailed unpaired t-test. Source data are provided as a Source Data file.

To validate these findings in vivo, we analyzed major interneuron subtypes in the mature cerebral cortex of VPA- and poly(I:C)-exposed mice. VIP-INs strikingly decreased in both ASD models by 38% and 49% relative to controls, respectively; meanwhile, other subtypes, such as Parvalbumin (PV)-, SST-, and RELN-INs, showed no significant changes, except for a modest reduction in PV-INs in poly(I:C)-exposed mice (Fig. 2c, d). VIP-INs decreased even after a single VPA exposure on E13 or E17 (Supplementary Fig. 5a). These results support the robust reduction in VIP-INs as a shared morphological pathology across these two nonhereditary ASD models.

Absolute reduction of cortical VIP-INs in ASD models is linked to impaired precursor-to-neuron specification in CGE

Cortical VIP-INs are major late-born GABAergic interneurons derived exclusively from the CGE, while PV-INs and SST-INs originate from the medial ganglionic eminence (MGE) during earlier stages^39–41. Therefore, we inquired whether other minor CGE-derived interneuron subtypes, not only VIP-INs, were preferentially impaired in the ASD models. We found that subsets of VIP-INs co-expressing calretinin (CR) and NDNF⁺ neurons modestly decreased to 60% and 70% of control levels in the VPA-exposed mice, respectively (Supplementary Fig. 5b, c). Collectively, these findings indicate that prenatal insults broadly disrupt the development of CGE-derived interneurons, not only the VIP⁺ subtype but also the NDNF⁺ subtype.

To explore the mechanism underlying the selective reduction of CGE-derived cell population in the ASD models, we focused on the developmental origins of cortical interneurons, as prenatal exposure to these nonhereditary ASD-risk factors is predicted to target interneuron progenitors. Transcriptional markers Coup-TFII (CGE-derived) and Sox6/Lhx6 (MGE-derived) were used for lineage tracing (Fig. 2e). In VPA-exposed embryonic brains, the COUP-TFII-expressing area in the CGE significantly expanded, whereas SOX6 and LHX6 expression areas in the MGE remained unchanged (Fig. 2f, g, and Supplementary Fig. 6a). These results suggest an inversely increased number of CGE-derived progenitors despite the reduced number of VIP-INs in postnatal ASD brains (Fig. 2c, d). This finding suggests that ASD-risk factors may disrupt the process of specification from mitotic progenitors to postmitotic neuronal progenitors in the CGE. To confirm this supposition, we labeled proliferating cells in the CGE with 5-ethynyl-2’-deoxyuridine (EdU) (Fig. 2h). EdU incorporation was observed in CGE progenitors co-expressing Ki-67 (Supplementary Fig. 6b). Both VPA and poly(I:C) exposure significantly increased the number of EdU⁺ cells in the CGE (Fig. 2h, i), indicating enhanced proliferation of VIP-IN progenitors in both ASD models.

During development, the CGE-derived cells tangentially migrate to the neocortex and are mainly integrated into layers II/III (Fig. 2e). Next, we chased the fate of CGE-derived cells in the neonatal cortex 6 days after EdU injection. EdU-labeled postnatal CGE markers (PROX1⁺/EdU⁺ and COUP-TFII⁺/EdU⁺) significantly reduced by VPA exposure (Fig. 2j, k and Supplementary Fig. 6c, d), consistent with the selective reduction of VIP-INs in mature ASD brains. By contrast, SOX6⁺/EdU⁺ cells and total EdU⁺ cells were unaffected (Fig. 2j, k and Supplementary Fig. 6e). These findings suggest that prenatal exposure to ASD-risk factors selectively disrupts the transition from proliferative progenitors to postmitotic VIP-IN precursors in the CGE, leading to a subsequent reduction in cortical VIP-INs.

Aberrant notch activity in the ganglionic eminence region disrupts cortical VIP-IN specification in ASD models

Given the crucial role of Notch signaling in cell fate determination during development^28–30, aberrant Notch activity is hypothesized to impair the transition from progenitors to neurons, leading to VIP-IN reduction. To determine whether aberrant Notch activation upon VPA and poly(I:C) exposure (Fig. 1) contributes to the altered VIP-IN progenitor pool and reduced VIP-INs observed in ASD models (Fig. 2), we manipulated Notch signaling in the in vitro ASD model using gain-of-function, loss-of-function, and rescue approaches.

We first confirmed that Notch expression was substantially higher in the ganglionic eminence (GE) than in the cortex (Fig. 3c and Supplementary Fig. 7a), indicating an abundance of Notch-positive neural progenitors in the GE during the critical developmental stages. These Notch-positive progenitors increased approximately twofold in the in vitro ASD model (Fig. 3b and Supplementary Fig. 7b). We also verified activation of Notch signaling at the protein level: the active (cleaved) Notch form was elevated in the in vitro ASD model, and this effect was attenuated by DAPT, a γ-secretase inhibitor that blocks Notch activation (Fig. 3c).

Fig. 3. Dysregulation of Notch signaling disturbs the specification of CGE precursors into GABAergic VIP-IN subtypes in ASD model.

a, b Flow cytometric analysis comparing Notch1 expression in the GE and the cortex (a), and in the transient VPA-treated forebrain primary culture (in vitro ASD model) and untreated culture (b). c Quantification of active NOTCH1 intracellular domain (ICN) protein levels in the in vitro ASD model, with or without pharmacological inhibition of Notch activity by DAPT treatment (Control= 7, VPA = 9, VPA + DAPT = 9, DAPT = 5). P = 0.0087 (Control vs. VPA), p < 0.0001 (VPA vs. VPA + DAPT). d Experimental scheme for genetic and pharmacological manipulation of Notch signaling in forebrain culture. Notch inhibition or activation was induced on DIV3–4, and maintained until DIV22–26. e, f Effects of Notch gain-of-function and loss-of-function on Vip transcript levels using the in vitro ASD models. e Overexpression of Notch1 ICN via retroviral infection reduced Vip transcript levels in forebrain cultures (Control = 9, ICN = 15). P = 0.0017 (Vip), p = 0.2867 (Sst). f Genome editing-mediated Notch1 gene disruption or DAPT treatment increased Vip transcript levels (Control = 5, sgN1 = 6; Control = 6, DAPT = 3). Single-guide RNAs of Notch1 (sgN1) was introduced with retrovirus vectors into forebrain cultures. The cultures used in the genome-editing experiments were prepared from embryos of Cas9-expressing mice. P = 0.0066 (Vip, sgN1), p = 0.6345 (Sst, sgN1), p < 0.0001 (Vip, DAPT), p = 0.0736 (Sst, DAPT). g Rescue of VPA-induced reduction in Vip transcript through Notch1 gene disruption (sgN1) or DAPT treatment (Control = 11, VPA = 21, VPA+sgN1 = 12; Control = 14, VPA = 15, VPA + DAPT = 4). P < 0.0001. h, i Representative images and quantification of VIP-IN density in forebrain cultures, showing significant reduction following overexpression of Notch1 ICN (Control = 9, ICN = 7). Arrowheads indicate VIP-stained cells. P = 0.0027. Scale bar: 50 µm. j, k Representative images and quantification of VIP-IN density showing restoration by gene disruption of Notch1 in VPA-exposed forebrain cultures prepared from Cas9-expressing embryos (Control = 7, VPA = 7, VPA+sgN1 = 15). P < 0.0001 (Control vs. VPA), p = 0.0003 (VPA vs. VPA+ sgN1). Scale bar: 50 µm. c, e, f, g, k, i Data are presented as mean values +/− SEM. c, g, k One-way ANOVA with Bonferroni’s multiple comparisons test. e, f, i Two-tailed unpaired t-test. Source data are provided as a Source Data file.

To examine the impact of Notch hyperactivation on neural progenitors, we introduced the intracellular domain of Notch (ICN), its active form, into forebrain neuronal cultures using a retroviral vector that specifically targets proliferative cells. ICN expression markedly reduced both the number of VIP-INs and the expression of Vip transcripts (Fig. 3d, e, h).

To evaluate the loss-of-function effect, we used CRISPR/Cas9-mediated removal of the Notch1 gene or pharmacological inhibition with DAPT. Both approaches significantly increased Vip transcripts, whereas Sst transcripts remained unaffected under all conditions (Fig. 3d,f). VIP-INs strongly reduced in the in vitro VPA model used for transcriptomic screening (Supplementary Fig. 4). Under this condition, both Notch1 gene removal and DAPT treatment significantly rescued the reduction in Vip transcripts and the number of VIP-INs in the VPA model (Fig. 3d,g,i). These findings reveal an inverse correlation between Notch activity and the differentiation of cortical VIP-INs, highlighting the pivotal role of precise regulation of the Notch signaling level during the late embryonic period. Therefore, the hyperactivation of Notch in the GE region at this developmental stage likely disrupts VIP-IN differentiation.

Aberrant histone deacetylase 3-mediated epigenetic regulation enhances notch signaling and impairs VIP-IN specification

We thus demonstrated that hyperactive Notch signaling disrupts the differentiation of CGE-derived VIP-INs (Fig.3a–e), directly contributing to the observed reduction of VIP-INs in ASD models. However, the upstream mechanisms driving this aberrant Notch activation remain unclear.

Given the widespread epigenetic dysregulation observed in the ASD model (Fig.1a) and the established role of class I histone deacetylase (HDAC) inhibitors such as VPA in altering gene expression during brain development^42–46, we examined the role of class I HDACs (HDAC1, HDAC2, HDAC3, and HDAC8) in regulating Notch signaling and VIP-IN specification. First, we screened the effects of selective inhibitors for these HDACs on cortical VIP-IN specification (Fig. 4a). Prenatal exposure to MS275 (a class I HDAC inhibitor), and RGFP966 (HDAC3 inhibitor), similar to VPA, significantly increased the number of COUPTFII⁺/EdU⁺ proliferating cells in the CGE (Fig. 4b–d) and decreased PROX1⁺ cells in the neonatal cortex (Supplementary Fig. 8a). Both inhibitors also reduced the number of cortical VIP-INs and Vip transcripts in both animals and forebrain cultures (Fig. 4e,f and Supplementary Fig. 8b,c), indicating disrupted VIP-IN specification. However, no such effects were observed with CAY10683 (HDAC2 inhibitor), highlighting HDAC3 as a key regulator (Fig. 4b–f). Further, RGFP966 exposure increased the active Notch proteins, suppressed by DAPT treatment (Supplementary Fig. 8d), and gene disruption of Notch1 restored RGFP966-induced reduction in Vip transcripts (Supplementary Fig. 8e), similar to VPA exposure (Fig. 3b,f). These findings suggest that HDAC3 plays a critical role in VIP-IN development. Given that HDAC3 is enriched in CGE and associates with the PROX1 (Supplementary Fig. 9a–c), we propose a model in which HDAC3 inhibition leads to aberrant Notch activation, thereby impairing the differentiation of CGE progenitors into VIP-INs (Supplementary Fig. 9d).

Fig. 4. HDAC3 is associated with Notch signaling-mediated specification of CGE precursors into GABAergic VIP-IN subtypes and social behavior.

a Table showing HDAC class I inhibitors and their IC₅₀. b–d Distinct effect of HDAC class I inhibitors on CGE progenitors. COUP-TFII expression and EdU⁺ cell areas are quantified in the CGE of each HDAC inhibitor-exposed mice. b Representative coronal images of EdU-labeled cells (green) in COUP-TFII⁺ (magenta) CGE progenitors (E14). Scale bar: 200 µm. c Relative expression levels of COUP-TFII (control = 46 images from 6 pups, VPA = 30 images from 5 pups, MS275 = 30 images from 4 pups, CAY10682 = 40 images from 5 pups, RGFP966 = 46 images from 5 pups). P = 0.0018 (Control vs. VPA), i = 0.0007 (Control vs. MS275), p > 0.9999 (Control vs. CAY10682), p = 0.0014 (Control vs. RGFP966). d Quantification of EdU⁺ area (control=82 images from 11 pups, VPA = 58 images from 5 pups, MS275 = 58 images from 8 pups, CAY10683 = 75 images from 7 pups, RGFP966 = 117 images from 12 pups). P < 0.0001 (Control vs. VPA), p = 0.0004 (Control vs. MS275), p = 0.0317 (Control vs. CAY10682), p < 0.0001 (Control vs. RGFP966). e, f Distinct effect of HDAC class I inhibitors on VIP-INs in the cerebral cortex (P21). e Representative coronal images of VIP-IN immunostaining. Scale bar: 100 µm. f VIP-IN density (control = 108 images from 9 pups, MS275 = 56 images from 4 pups, CAY10683 = 57 images from 5 pups, RGFP966 = 58 images from 6 pups). P < 0.0001 (Control vs. VPA), p = 0.0020 (Control vs. MS275), p > 0.9999 (Control vs. CAY10682), p < 0.0001 (Control vs. RGFP966). c, d, f One-way ANOVA with Dunnett’s multiple comparisons test. g–n The comparison of behavioral deficits among VPA, poly(I:C), or RGFP966-treated mice. g Duration of self-grooming (control = 60, VPA = 60, RG = 45, poly(I:C) = 50). P < 0.0001 (Control vs. VPA), p = 0.0124 (Control vs. RGFP966), p = 0.0092 (Control vs. poly(I:C)). h Number of ultrasonic vocalizations (control=60, VPA = 60, RG = 45, poly(I:C) = 50). P = 0.0285 (Control vs. VPA), p < 0.0001 (Control vs. RGFP966), p = 0.0291 (Control vs. poly(I:C)). i–n Three-chamber social interaction test (control = 21, VPA = 7, RG = 19). Representative traces of each test mouse and time spent in each chamber, on the sociability session (i and j) and the social novelty session (l, m). k Sociability index (SI). P = 0.2464 (Control vs. VPA), p = 0.7664 (Control vs. RGFP966). n Social novelty index (SNI). P = 0.0474 (Control vs. VPA), p = 0.0462 (Control vs. RGFP966). c, d, f, g, h, j, k, m, n Data are presented as mean values +/−SEM. g, h, k, n One-way ANOVA with Dunnett’s multiple comparisons test. (j, m) Two-tailed unpaired t-test. Source data are provided as a Source Data file.

Specific HDAC3 inhibition results in repetitive behaviors and social deficits linked to ASD

Based on the morphological abnormalities caused by HDAC3-specific inhibition (Fig. 4a–f), we next investigated whether prenatal inhibition of HDAC3 replicates the behavioral phenotypes characteristic of ASD, as seen in VPA- and poly(I:C)-exposed ASD models. Mice prenatally exposed to RGFP966 exhibited normal development and body weight comparable to control mice (data not shown). Nevertheless, these mice displayed significant increases in self-grooming behavior—a hallmark of repetitive behaviors often observed in ASD (Fig. 4g). Additionally, RGFP966-exposed mice showed social abnormalities, including reduced ultrasonic vocalization (USV) calls (Fig. 4h) and impaired preference for social novelty, but not sociability, in the three-chamber test (Fig. 4i–n), although general sociability remained partially intact (Supplementary Fig. 10). These findings indicate that HDAC3 inhibition alone is sufficient to recapitulate core behavioral phenotypes associated with ASD. Taken together, these data highlight the crucial role of HDAC3-mediated epigenetic regulation in normal social and behavioral development and its disruption as a potential mechanism underlying ASD pathology.

CGE region-targeted disruption of notch signaling rescues VIP-IN loss and social abnormalities in ASD models

To directly test whether the local hyperactivation of Notch signaling in developing brain underlies the morphological and behavioral abnormalities observed in ASD models, we generated CGE-specific Notch1/2 conditional knockout (cKO) mice (Fig. 5a). The deletion of Notch genes targets CGE progenitors around E18, based on the expression pattern of Htr3a⁴⁷. Both Notch1/2 cKO and control (WT) mice were prenatally exposed to VPA to evaluate the roles of Notch signaling on neural and behavioral ASD-like phenotypes (Fig. 5a).

Fig. 5. CGE-specific ablation of the Notch genes suppresses the reduction in the cortical VIP-IN subtype, normalizes E/I balance, and selectively restores social behavior in the mouse ASD model.

a Generation of CGE-specific Notch1/2 cKO mice. Htr3a-Cre mice were used to delete Notch1/2 genes in CGE progenitors exposed to VPA. b Restoration of VIP-IN density in VPA-exposed cKO mice. Representative coronal images of the somatosensory cortex and quantification of VIP-IN density (control=46, cKO=45, WT-VPA = 45, cKO-VPA = 46 images). P < 0.0001 (WT vs. WT-VPA, cKO vs. WT-VPA, WT-VPA vs. cKO-VPA). Scale bar: 100 µm. c–g Behavioral rescue in VPA-exposed cKO mice. c Duration of self-grooming (control=30, cKO=23, WT-VPA = 44, cKO-VPA = 36). P > 0.9999 (WT vs. cKO), p = 0.0018 (WT vs. WT-VPA), p = 0.0047 (WT vs. cKO-VPA), p > 0.9999 (WT-VPA vs. cKO-VPA). d Contact time in the reciprocal social interaction test (control=29, cKO=25, WT-VPA = 32, cKO-VPA = 26). P = 0.0006 (WT vs. WT-VPA), p < 0.0001 (cKO vs. WT-VPA), p = 0.0476 (WT-VPA vs. cKO-VPA). e–g Three-chamber social interaction test (control = 57, cKO=54, WT-VPA = 63, cKO-VPA = 51). e Time spent in each chamber. Increased time in the familiar chamber by VPA-exposure, p = 0.0447 (WT vs. WT-VPA), p = 0.0128 (cKO vs. WT-VPA), p = 0.0004 (WT-VPA vs. cKO-VPA). Decreased time in the centre chamber. p = 0.0154 (cKO vs. WT-VPA), p = 0.0252 (WT-VPA vs. cKO-VPA). f Representative heatmaps of each test mouse. g SNI in the social novelty session. P = 0.0378 (WT vs. WT-VPA), p = 0.0205 (cKO vs. WT-VPA), p = 0.0021 (WT-VPA vs. cKO-VPA). b, c, d, e, g Data are presented as mean values +/−SEM. b, c, d, g One-way ANOVA with Bonferroni’s multiple comparisons test. e Two-tailed unpaired t-test. h–m Transcriptomic changes analyzed by RNA-seq of the adult forebrain (n = 3) exhibited gene expression profiles between four groups (WT, WT-VPA, cKO, and cKO-VPA). h Number of differentially altered genes across genotypes and/or condition. i PCA showed distinct clustering of cKO-VPA mice from WT-VPA mice. j Gene enrichment analysis by STRING revealed increased glutamatergic synaptic transmission in WT-VPA, restored in cKO-VPA. k–m Volcano plots illustrating differential gene expression in comparisons: WT vs. cKO (k), WT vs. WT-VPA (l), and WT-VPA vs. cKO-VPA (m), highlighting the selective downregulation of Glutamatergic neuron-related genes in the cKO-VPA mice. j–m Two-sided unpaired t-test with Benjamini–Hochberg multiple testing correction. Source data are provided as a Source Data file.

In the absence of VPA exposure, there is no difference in both the number of cortical VIP-INs and PV-INs, and ASD-like behaviors between WT mice and Notch1/2 cKO mice (Fig. 5b–g and Supplementary Fig. 11a,b). However, while VPA-exposed WT mice exhibited a significant loss of cortical VIP-INs and pronounced social abnormalities in the same way as shown in Fig.4g–n and Supplementary Fig. 10, these impairments were largely prevented in the Notch1/2 cKO mice (Fig. 5b, d–g, and Supplementary Fig. 11c,d). This conditional deletion of Notch1/2 did not affect repetitive self-grooming behavior observed in VPA-exposed mice (Fig. 5c). These results demonstrate that CGE-specific Notch1/2 deletion reversed VIP-IN loss and social deficits in VPA-exposed mice, highlighting Notch dysregulation as a key driver of VIP-IN circuit abnormalities and consequent social impairments.

Developmental loss of CGE-derived interneuron circuits accompanies cortical E/I imbalance

Next, we investigated the molecular and circuit-level consequences of this prenatal genetic intervention by bulk RNA-seq analysis on the adult forebrain. The transcriptomic differences between WT and Notch1/2 cKO mice were modest, indicated by the small number of differentially expressed genes (DEGs) (Supplementary Data 3) and the close clustering profiles in principal component analysis (PCA) plots (WT vs. cKO, Fig. 5h,i). This is reasonable given that the Htr3a-Cre driver induces conditional Notch loss only in a small subset of CGE-derived interneurons. However, under VPA exposure, transcriptomic profiles diverged significantly between genotypes. VPA-exposed WT mice exhibited over 2101 DEGs (873 upregulated, 1228 downregulated) (WT vs. WT-VPA, Fig. 5h), reflecting widespread dysregulation of gene expression. Notably, gene expression in 49% of these DEGs was reversed in Notch1/2 cKO mice under VPA exposure (WT-VPA vs. cKO-VPA, Fig. 5h), and the aberrant transcriptomic profile observed in WT-VPA mice was partially normalized in the cKO-VPA condition (Fig. 5i).

Enriched GO terms of upregulated genes by VPA exposure (WT vs. WT-VPA) were “Synapse”-related categories with “Glutamatergic synapse” showing the highest enrichment (Fig. 5j and Supplementary Fig. 12). Importantly, the same GO terms were enriched for downregulated genes of Notch1/2 cKO mice under VPA exposure (WT-VPA vs. cKO-VPA, cKO vs. cKO -VPA or WT vs. cKO -VPA) (Supplementary Figs. 12, 13). Consistently, genes associated with glutamatergic synapse transmission including Grin2a, Grin2b, Scn2a, Nrxn1, and CamK2d were upregulated by VPA exposure, while these were normalized in Notch1/2 cKO-VPA mice (Fig. 5k–m and Supplementary Data 4). Transcriptomic changes in Grin2b and Grin2a, NMDA receptor subunits, were further supported by quantitative analysis (Supplementary Fig. 12c).

In the embryonic brain exposed to VPA or poly(I:C), RNA-seq analysis revealed the transcript level of Neurogenin2 (Neurog2) —a transcriptional activator promoting the differentiation of glutamatergic neuron—to be upregulated. However, the upregulation was restored in Notch1/2 cKO-VPA mice (Supplementary Fig. 12d), suggesting that Neurog2 elevation represents a secondary event of impaired inhibitory networks. Together, these results suggest that the developmental loss of CGE-derived interneurons may trigger maladaptive excitatory gene expression programs.

By contrast, the genes localized on medium spiny neurons (MSNs) in the striatum, such as Ppp1r1b, Pde10a/1b, Drd1/2, Penk, and Tac1, account for 11 out of the top 20 downregulated genes following VPA exposure, indicating a pronounced reduction of striatal MSNs, GABAergic neurons derived from the lateral ganglionic eminence (LGE) (Supplementary Fig. 13c,d). However, unlike glutamatergic regulators (Fig. 5m), these downregulated expressions were less normalized in the CGE-specific KO-VPA mice (Supplementary Fig. 13e).

These data suggest that the developmental loss of VIP-INs broadly impacts not only inhibitory circuits but also surrounding excitatory circuits, leading to maladaptive E/I balance and subsequent social deficits in ASD. Thus, Notch1/2 deletion demonstrates potential to restore both morphological and functional disruptions caused by exposure to ASD-related environmental factors.

A single dose of potent notch inhibitor prevents cortical and behavioral deficits in the ASD model

To explore the therapeutic potential of inhibiting Notch signaling against ASD, we examined whether a pharmacological approach could ameliorate the morphological and behavioral abnormalities in the ASD model. We used Ro4929097, a γ-secretase inhibitor that potently inhibits Notch activity and has previously been used for clinical purposes. This compound was originally developed as a candidate drug for Alzheimer’s disease and then advanced to Phase II clinical trials in 2011 for anticancer applications in melanoma, T-cell acute lymphoblastic leukemia, and other indications^48–50. A single dose of Ro4929097 was prenatally administrated to VPA-exposed mice (Fig. 6a). The potent inhibition effect of Notch activity by Ro4929097 was confirmed in forebrain neuronal cultures (Fig. 6b). Ro4929097 potently restored the reduction of cortical VIP-INs caused by VPA exposure, almost to normal levels (Fig. 6c,d). This result aligns with an earlier finding that hyperactive Notch signaling underlies the VIP-IN loss (Figs. 3, 5). Similarly, behavioral abnormalities in VPA-exposed ASD mice were restored by Ro4929097 (Fig. 6e–i). We noted hair loss—which appeared to be due to long-term self-grooming behavior ( > 13 weeks-old)—in 27% of VPA-exposed mice, but this was nearly absent in Ro4929097-treated mice (Fig. 6e). Ro4929097 also prevented repetitive self-grooming behavior as well as deficits in social behaviors (Fig. 6g–i), compared with the outcome achieved by CGE-specific Notch deletion, which did not prevent repetitive behavior (Fig. 5c).

Fig. 6. A single antenatal dose of a potent Notch inhibitor ameliorates multiple abnormalities in the mouse ASD model.

a Experimental scheme showing the in vivo rescue of Notch signaling by a single dose of γ-secretase inhibitor Ro4929097 (Ro) in the ASD-model mice exposure to VPA. b Normalized protein level of active Notch1 in the VPA-exposed mouse embryos by Ro treatment. c, d Rescue of VIP-IN density in Ro-treated VPA-exposed mice. c Representative coronal images of the somatosensory cortex. d Quantification of VIP-IN density showed a significant increase by Ro injection in VPA-exposed mice (control = 48, VPA = 68, VPA+Ro = 86 images). P < 0.0001 (Control vs. VPA, VPA vs. VPA+Ro). Scale bar: 100 µm. e–i Behavioral improvements in Ro-treated VPA-exposed mice. e Hair loss observed in 27% of VPA-exposed mice rescued by Ro exposure (control = 30, VPA = 34, VPA+Ro=29). P = 0.0024 (Control vs. VPA), p = 0.0414 (VPA vs. VPA+Ro). f Total grooming duration (control=70, VPA = 72, VPA+Ro=66). P < 0.0001 (Control vs. VPA), p = 0.0008 (VPA vs. VPA+Ro). g–i Reciprocal social interaction test (control=29, VPA = 41, VPA+Ro = 45). g Contact number, p = 0.0028 (Control vs. VPA), p = 0.0438 (VPA vs. VPA+Ro). h Contact time, p = 0.0034 (Control vs. VPA), p = 0.0375 (VPA vs. VPA+Ro). i Resting time, p < 0.0001 (Control vs. VPA, VPA vs. VPA+Ro). d, f, g–i Data are presented as mean values +/−SEM. One-way ANOVA with Bonferroni’s multiple comparisons test. e Two-sided Pearson’s chi-square test. j–o Single-cell RNA sequencing (scRNA-seq) analysis of whole brain (P2) from VPA-exposed mice revealed cellular and transcriptomic rescue by Ro. j UMAP plot of 20 clusters in whole-brain cells. k Restored perturbations in the whole cell compositions in VPA-exposed mouse with Ro exposure. l Heatmap of residuals and Pearson’s coefficients for cell-type distributions. Combined heatmap showing standardized residuals from chi-square tests and Pearson’s coefficients between groups (control, VPA, VPA+Ro). Ro treatment shifted the VPA-induced distribution pattern toward that of control mouse. m Sub-clustering of NSC/NP-enriched populations (clusters #6 and #7) and the normalized cellular composition of Htr3a⁺ CGE-enriched cell populations in VPA-exposed mouse after Ro treatment. n Normalized cell composition of oligodendrocyte-enriched clusters (clusters #5, #8, #12, #14, and #18) with Ro exposure. ovgy Cell composition of astrocyte-enriched clusters (clusters #2, #3, and #13) with Ro exposure. Source data are provided as a Source Data file.

Restoration of neural lineages in neonatal ASD models by pharmacological notch inhibition via single-cell transcriptomics

To elucidate the cellular mechanisms underlying robust rescue of ASD-like behaviors by the Notch inhibitor Ro4929097, we performed single-cell RNA seq (scRNA-seq) on neonatal whole brain and revealed 20 distinct cell clusters in the uniform manifold approximation and projection (UMAP) representation (Fig. 6j and Supplementary Fig. 14), with significant shifts in cluster composition in the ASD model compared with controls (Fig. 6k). Remarkably, the single dose of Ro4929097 almost restored this composition to control levels. To statistically validate these observations, we performed chi-square tests followed by residual analyses to compare cell-type proportions across the three groups. The resulting heatmap showed residual patterns consistent with the stacked bar plot in Fig. 6k (Fig. 6l). Pairwise chi-square tests revealed significant nonindependence between groups. Additionally, Pearson correlation coefficients indicated that administration of Ro4929097 shifted the ASD model profile toward that of controls (Fig. 6l).

Focusing on Notch1 and Notch2-positive populations, their expression was observed in neural progenitors (NSC/NPs) marked by Nestin (Supplementary Fig. 15). Their enrichment in proliferating clusters was also confirmed by the labeling of S-phase or G2/M-phase (Fig. 6m). We further re-clustered the NSC/NP population into four clusters and identified two clusters (clusters 0 and 3) that express the GABA precursor markers Dlx1 and Dlx5 (Fig. 6m and Supplementary Fig. 16). In ASD models, there was an expansion of a re-clustered progenitor subtype (cluster 3) enriched for CGE-selective markers (Dlx1/5, Htr3a, Prox1), consistent with the increased CGE populations seen in embryonic ASD brains (Fig. 2). Treatment with Ro4929097 reversed this expansion, restoring the cellular composition of CGE to control levels (Fig. 6m).

We extended this analysis to other Notch1/Notch2-positive cell types, revealing abnormalities in non-neuronal populations, including glial fibrillary acidic protein (GFAP)-positive astrocytes and Olig2-positive oligodendrocytes. Notably, while Ro4929097 significantly corrected abnormalities in oligodendrocyte-related populations (cluster numbers 5, 8, 12, 14, and 18) (Fig. 6n), it had limited effects on astrocytic populations (Fig. 6o). These findings underscore that the transient blockade of Notch activity is sufficient to restore ASD-like behavioral phenotypes and the disruption of developmental trajectories, including multiple neural and glial lineages in the ASD model.

Discussion

This study identifies dysregulated Notch signaling as a common pathological feature in both nonhereditary ASD models and human ESC-derived neurons mimicking hereditary ASD. Given the robust rescue effects of Notch inhibition on morphological and behavioral abnormalities in ASD animal models, our findings indicate that aberrant Notch activity plays a central role in ASD pathogenesis.

Recent studies have identified Notch genes as ASD risk factors^7,51,52, classified as high-priority genes in the SFARI database (Rank2)⁵³. Furthermore, previous studies on 1q21.1 duplication syndrome—a condition associated with developmental delays and ASD traits—have highlighted the role of Notch dysregulation, as overexpression of the human-specific Notch2NL gene delays neural differentiation through radial glial activation^54,55, aligning with our findings. We note that two cell lines used in Fig. 1e^34,35 do not carry a Notch2 gene copy at 1p21 and therefore may not be suitable for direct testing; future studies should employ models carrying the Notch2NL locus to fully assess its contribution.

Enhanced Notch signaling could originate from epigenetic dysregulation in embryonic brains. Epigenetic modifications, increasingly recognized as key mechanisms in ASD^9,10, were prominently detected in transcriptomic analyses of nonhereditary ASD models (Fig. 1). We found that disrupted histone modification mediated by HDAC3 increases Notch expression and related gene activity (Supplementary Fig. 8), suggesting that epigenetic control of the Notch signaling pathway may be particularly vulnerable to various hereditary and nonhereditary ASD-associated factors. However, the reason why Notch signaling is commonly enhanced by multiple ASD-risk factors, not only VPA, remains unclear. Future research should aim to elucidate the upstream mechanism causing dysregulation of Notch signaling through epigenome-oriented approaches.

Notch signaling governs cell fate decisions during neurodevelopment^28–30. Upregulation of Notch signaling by ASD-associated factors during a critical fetal time window disrupts the proper differentiation of CGE progenitor, leading to a drastic reduction in postnatal cortical VIP-INs (Figs. 2, 3), essential for maintaining cortical E/I balance. We identified this reduction in cortical VIP-INs as a common pathophysiological feature in nonhereditary ASD models—exposure to VPA or poly(I:C). These findings highlight the critical role of epigenetic Notch regulation in VIP-IN development and its direct implications for neurodevelopmental disorders including ASD.

Although Notch signaling is broadly active in neural progenitors, its disruption was selectively observed in CGE progenitor cells of embryonic ASD models. CGE-derived interneurons (VIP-INs, VIP/CR-INs and NDNF-INs) differentiate between E12.5 and E18.5. Around the same timing, in utero VPA or poly(I:C) exposure was performed to produce the ASD models. Consequently, VPA strongly affects the mitotic CGE progenitors, leading to the reduction in late-born CGE interneurons but not the early-born MGE interneurons (Fig. 2; Supplementary Figs. 5, 6). By contrast, VPA exposure after mitosis did not alter the number of GABAergic interneurons (Supplementary Fig. 4). Therefore, the timing and duration of VPA or poly(I:C) exposure restrict its effect on CGE progenitors. Similarly, VPA exposure may reduce striatal MSNs, originated from LGE during the late embryonic stage⁵⁶, in the ASD model (Supplementary Fig. 13).

Additionally, the restricted effect of Notch dysfunction may reflect the spatial inhibition of HDAC3, which interacts with PROX1—a key factor for precursor-to-neuron specification in VIP-IN populations—and suppresses Notch1 activity through this interaction⁵⁷. Accordingly, we observed strong HDAC3 expression in the CGE, colocalizing with PROX1 (Supplementary Fig. 9), along with an enrichment of Notch1-positive cells in the GE region, which increased further by VPA exposure (Fig. 3). These findings indicate that HDAC3 acts as a critical epigenetic regulator of Notch1 expression (Fig. 4)⁵⁷, suggesting that regional specificity in regulation by Notch signaling depends on HDAC3 and other cell-type-specific cofactors. We further observed that embryonic HDAC3 inhibition caused ASD-like social and repetitive behaviors in a similar way to VPA- or poly(I:C)-exposed mice (Fig. 4), supporting its contribution to dysregulation of Notch signaling in ASD models.

CGE-specific ablation of Notch1/2 in the VPA-exposed ASD model restored cortical VIP-INs and selectively rescued social behavior: deficits in social interaction and social novelty, but not repetitive behavior (Fig. 5). This results aligns with a recent finding that VIP-IN-specific gene disruption selectively affects social behaviors: a mouse model of Dravet syndrome (Scn1a KO) showed that the selective disruption of VIP-INs led to deficits in sociability, but not seizures or an increased risk of mortality³³. Thus, our findings, together with previous evidence, underscore the pivotal role of VIP-INs in social functions⁵⁸.

Consistent with their role in social behaviors, VIP-INs appear to be critical for broad cortical circuit formation and function. Our transcriptomic analysis of the adult forebrain in the ASD model suggests that the developmental loss of VIP-INs triggers maladaptive changes in gene expression (Fig. 5), potentially disrupting multiple cortical circuits. Surprisingly, we noted that either loss or rescue of VIP-INs has, respectively, an inverse correlation with the gene expression related to the maturation of the glutamatergic synaptic membrane, indicating E/I imbalance in the ASD model. Supporting this finding, we found that Neurog2—a driver of glutamatergic neuron differentiation—was elevated in poly(I:C)- or VPA-exposed mice, and rescued upon VPA exposure by CGE-specific ablation of Notch1/2 genes (Supplementary Fig. 12). Although VIP-INs represent only ~1% of all brain neurons (10% of cortical interneurons)⁵⁹, recent studies have highlighted their involvement in sensory processing⁶⁰, attention regulation⁶¹, and behavioral flexibility⁶², suggesting their crucial role in cortical circuit formation and activity. Our findings in the ASD models further emphasize the profound impact of this small GABAergic subpopulation on brain development and function. However, whether ASD-like social deficits arise directly from VIP-IN deficiency or secondarily via E/I imbalance remains unresolved.

A single antenatal administration of Ro4929097 to VPA-exposed pregnant females successfully ameliorated ASD-related behaviors and other abnormalities in their offspring (Fig. 6). Notably, repetitive behavior observed in the ASD mice were strongly rescued by the systemic Notch inhibition, whereas CGE-specific Notch1/2 cKO did not produce this rescue effect (Figs. 5, 6). This outcome highlights the therapeutic potential of targeting prenatal Notch signaling. The effect of Ro4929097 could reflect not only the rescue of VIP-IN deficits but also a broad restoration of disrupted cellular compositions. Single-cell transcriptomics indicated that prenatal enhancement of Notch signaling potentially affects the differentiation of multiple progenitor cells, contributing to the diverse postnatal abnormalities observed in ASD. Besides the loss of VIP-INs related to E/I balance and social function, this study highlighted other affected cell types, such as striatal MSNs that were not rescued by VIP-IN-specific Notch inhibition (Fig. 5), and oligodendrocytes that were rescued by systemic Notch inhibition (Fig. 6). Notably, reduction in MSNs may underlie repetitive behaviors, as mouse grooming behavior is correlated with dysfunction in the cortico-striatal-thalamo-cortical loop⁶³. Further cell type-specific rescue studies of Notch signaling could clarify which populations contribute to distinct ASD phenotype.

Finally, our findings suggest that dysregulated Notch signaling is relevant to some of nonhereditary and hereditary ASD models and may also play a role in some forms of hereditary human ASD (Fig. 1, Supplementary Fig. 2). Given the high etiological complexity of ASD, early-life intervention targeting Notch signaling holds great promise as a broadly applicable therapeutic strategy.

Methods

Generation of the mouse ASD models

C57BL6/J mice (Japan SLC, Inc., Shizuoka, Japan) and ICR (CD-1) mice (Charles River Laboratories, Inc., Wilmington, MA) were used in this study. Pregnancy was confirmed by the presence of a vaginal plug, designated as embryonic day 0 (E0). The dosages of drugs were adjusted according to the body weight of the dam on the day of injection.

The drugs were administered by intraperitoneal injections on the determined gestational days as follows: VPA (Sigma–Aldrich, St. Louis, MO) or histone deacetylase (HDAC) inhibitors were administered at E13, E14, E15, and/or E18, and poly(I:C) (Sigma–Aldrich) was administered at E12. ICR (CD-1) mice were used, except for the experiment using Htr3a-Cre; Notch1/2 conditional knockout mice. VPA, poly(I:C), and CAY10683 (Cayman Chemicals, Ann Arbor, MI) were suspended in 0.9% saline. RGFP966 and MS275 (Cayman Chemicals) were dissolved in dimethyl sulfoxide and further diluted in saline. The administered doses were as follows: VPA, 290–330 mg/kg; poly(I:C), 20 mg/kg; Ro4929097 (Cayman Chemicals), 1 mg/kg. For HDAC inhibitors, MS275 was 5–9.25 mg/kg for cortical analysis and 10–12.5 mg/kg for embryonic CGE/MGE analysis. RGFP966 was 6.5–10 mg/kg for cortical analysis and 7.5–10 mg/kg for embryonic analysis. CAY10683 was 35–40 µg/kg for cortical analysis and 85 µg/kg for embryonic analysis. Control dams received saline injections. After delivery, the dams were housed individually and allowed to raise their own litters. Both female and male offspring were used for the experiments up to 15 weeks of age.

All mice were housed at an animal facility of the Tokai University with constant temperature and air humidity, and a 12-h light/12-h dark cycle. All animal procedures complied with institutional and governmental regulations on animal experimentation, reviewed and approved by the Institutional Animal Care and Use Committee and Genetic Modification Safety Committee of Tokai University (nos. 244018 and 23-010-28). All surgeries were performed under sodium pentobarbital anesthesia, and all efforts were made to minimize animal suffering. Sex was not considered in the study design and analysis, since there were no significant sex-related differences in the behavioral, histological and biochemical analysis of this study.

Generation of Htr3a-Cre; Notch1/2 conditional knockout mice

Htr3a-Cre mice (MMRRC, ID: 37089, C57BL6/J background)⁶⁴ were crossed with Notch1 flox/flox (C57BL6/J background)⁶⁵ and Notch2 flox/flox (C57BL6/J background)⁶⁶ mice to generate Htr3a-Cre; Notch1 flox/flox; and Notch2 flox/flox conditional knockout (cKO) mice. Genotyping was performed using polymerase chain reaction (PCR) for the Cre transgene and floxed alleles as previously described^65,66.

All genetically modified mice were maintained on a C57BL/6 J background and were bred under specific pathogen-free (SPF) conditions.

Neuronal cell culture and in vitro ASD model

Primary neuronal cultures were prepared from the forebrains of embryonic day (E) 14.5-15.5 mouse embryos of ICR strain or Cas9 knockin mice (Jackson, Stock#026179)⁶⁷. The dissociation was performed using 0.05% trypsin in the presence of DNase I (Roche Applied Science) for 10 min at 37 °C. Trypsin digestion was inactivated using soybean trypsin inhibitor (Gibco; Thermo Fisher Scientific). Dissociated cells were plated onto six-well or 24-well culture dishes at a density of (1–2 × 10⁵ cells/cm²) and maintained in Neurobasal medium (Gibco) supplemented with 2% B27 (Gibco), 2 mM Glutamax (Gibco), and penicillin/streptomycin (Gibco) for up to 26 days.

For the in vitro ASD model, the cell line involved in E/I imbalance, neuronal cultures were treated with (1–2 mM) valproic acid (VPA) from DIV 0 to 6, following a protocol established previously³⁸. Cells were harvested at DIV 15 for biochemical analysis.

For the in vitro treatment with HDAC inhibitors shown in Fig. 4, neurons were exposed to 750 nM MS275, 1–2 µM RGFP966, or 1–10 nM CAY10683 for 4–6 days from the beginning of the culture, then maintained in a normal growth medium until the designated experimental time points.

To specifically modulate Notch1 signaling in proliferating cells, retroviral infection was performed at DIV3–4. For Notch1 intracellular domain (ICN) overexpression, a retrovirus carrying Notch1 ICN was introduced into neuronal cultures. For Notch1 gene disruption, neuronal cultures derived from Cas9-expressing transgenic mice were infected with retroviruses carrying sgRNAs targeting Notch1, following a previously described protocol⁶⁸.

Antibodies

The following commercially available primary antibodies were used:

• Rabbit anti-VIP (1:1000, 20077, ImmunoStar)

• Mouse anti-Reelin (1:500, D223-3 (CR-50), MBL)

• Rat anti-Somatostatin (1:300, MAB354 (YC7), Merck Millipore)

• Rabbit anti-Parvalbumin (1:1000, MSFR105210, Nittobo Medical)

• Rabbit anti-Calretinin (1:1000, MSFR100440, Nittobo Medical)

• Mouse anti-Calretinin (1:1000, MAB1568 (6B8.2), Merck Millipore)

• Mouse anti-GAD67 (1:1000, MAB5406 (1G10.2), Merck Millipore)

• Rabbit anti-PROX1 (1:1000, ab199359, Abcam)

• Mouse anti-COUP-TFII (1:1000, PP-H7147-00 (H7147), Perseus Proteomics)

• Rabbit anti-SOX6 (1:1000, ab30455, Abcam)

• Mouse anti-LHX6 (1:1000, H00026468-M02 (1B11), Abnova)

• Mouse anti-HDAC3 (1:1000, #3949 (7G6C5), Cell Signaling)

• Rabbit anti-cleaved NOTCH1 (1:1000, #4147 (Val1744) (D3D8), Cell Signaling,)

• Mouse anti-GAPDH (1:1000, Santa Cruz Biotechnology)

For immunostaining, Alexa Fluor 488- or 546-conjugated secondary antibodies (Molecular Probes) were used. For western blot analysis, horseradish peroxidase (HRP)-conjugated anti-mouse and anti-rabbit immunoglobulin G (IgG) secondary antibodies (Thermo Fisher Scientific) were used.

RNA isolation and analysis

Total RNA was isolated using RNAiso Plus (Takara, Kyoto, Japan), followed by DNase treatment with Turbo DNase (Ambion, Austin, TX, USA) to remove contaminating DNA. Two micrograms of total RNA were reverse transcribed using random hexamers and PrimeScript II Reverse Transcriptase (Takara).

Quantitative PCR (qPCR) was performed on a StepOnePlus qPCR system (Applied Biosystems, Foster City, CA, USA) using SYBR Green Master Mix (Applied Biosystems) and the comparative CT method. Custom primer sets (Supplementary Data 5) were designed with Primer3web 4.1.0, and mRNA levels were normalized to Gapdh mRNA.

Bulk RNA-sequence analysis

Total RNA was extracted from brain tissues and subjected to quality control procedures. mRNA was enriched using oligo(dT) beads, followed by rRNA depletion with the Ribo-Zero rRNA Removal Kit (Illumina, San Diego, CA, USA). The enriched mRNA was randomly fragmented using fragmentation buffer, and cDNA synthesis was carried out using the NEB Next Ultra RNA Library Kit (Illumina) according to the manufacturer’s protocol.

After library preparation, sequencing was performed on a NovaSeq6000 platform (Illumina), with pooling optimized based on effective concentration and anticipated data volume. The sequencing data underwent a filtering process to remove reads containing adapters, reads with undetermined bases (N > 10%), and reads with low-quality bases (Q score ≤ 5) that accounted for more than half of the total bases. Clean reads were then aligned to the mouse reference genome (mm10) using the HISAT2 program 47. Data analysis was performed using Metascape (https://metascape.org/gp/index.html) and STRING for gene ontology (GO) and protein interaction analysis. Using Metascape, the enriched GO terms were listed by the percentage of the genes on the hit list to the genes annotated with the GO term. Using STRING, the enriched GO terms were scored by the signal intensity defined as a weighted harmonic mean between the ratio of the annotated gene with the observed/expected annotated with a term and -log (false discovery rate; FDR). The enriched GO terms were sub-categorized into Biological Process, Cellular Component, Molecular Function, and/or Pathways, compared between experimental groups.

Single-cell RNA-sequence analysis

For analyzing VPA-exposed brains, single-cell RNA sequencing (scRNA-seq) was performed using the BD Rhapsody™ Whole Transcriptome Analysis (WTA) Amplification Kit (BD Biosciences). Briefly, viable cells were isolated from neural tissue using the Neural Tissue Dissociation Kit (P) (Miltenyi Biotec), following the manufacturer’s protocol. The resulting single-cell suspension was sorted using the BD Rhapsody™ Single-Cell Analysis System. Cell Sorter to enrich for viable single cells. A total of approximately 25,000 cells were sorted per animal group before proceeding to single-cell capture and cDNA synthesis. Library preparation was conducted according to the BD Rhapsody™ WTA Amplification Kit protocol. Sequencing was performed on an appropriate platform, generating approximately 5,000 high-quality single-cell transcriptomes per sample across three experimental groups, with an average sequencing depth of 19,164 reads per cell.

Data analysis was performed using Seurat (v5.0.3) in R4.3.0. After quality control, cells with >12% mitochondrial gene expression, fewer than 1,500 detected genes, or more than 5,000 genes were removed. Normalization was performed using Seurat’s “Normalize Data” function with the “Log Normalize” method, followed by feature selection and scaling. Three datasets were then integrated using Seurat’s integration anchors method. Dimensionality reduction was conducted using principal component analysis. Clustering was performed using the Louvain algorithm with a resolution parameter of 0.5, and visualization was performed using uniform manifold approximation and projection (UMAP). Differentially expressed genes were identified using the Wilcoxon rank-sum test with Bonferroni correction to characterize cluster-specific markers.

For the analysis of human ESC-derived neurons after differentiation, previously published scRNA-seq datasets were used (DDBJ accession IDs: DRX256644, DRX256645, DRA011369, and DRA010907)^34,35. Single-cell expression data were converted into pseudo-bulk profiles for downstream analysis. Pseudo-bulk analysis was performed using 10X Genomics Loupe Browser v8.1.2. To investigate Notch signaling dysregulation in ASD models, gene expression patterns were analyzed and visualized using heatmaps. The Log2 fold change values shown in the heatmaps represent the ratio of the normalized mean gene unique molecular identifier counts between conditions.

Mouse exon array and cell-type proportion analysis

Total RNA quality was assessed with an Agilent 2100 Bioanalyzer (Agilent Technologies, Santa Clara, CA, USA). One hundred nanograms of total RNA were labeled using the Low Input Quick Amp WT Labeling Kit, One-Color*3, and hybridized to the SurePrint G3 Mouse Exon Microarray 2×400 K (Agilent Technologies, Folsom, CA, USA) following the manufacturer’s protocol.

Hybridization signals were detected using a DNA Microarray Scanner G2600D (Agilent Technologies, Santa Clara, CA, USA). All scanned images were analyzed with Agilent Feature Extraction Software (v11.5.1.1), and raw data were processed using GeneSpring GX software (v13.0, Agilent Technologies). After 75th percentile shift normalization, differentially expressed genes in the VPA-treated groups were identified by comparing raw probe signal intensities, averaging across genotypes, and filtering out probes with signal intensities below 100. Scatter plots with Pearson’s correlation coefficients were generated, and a volcano plot was constructed using a moderated t-test with Benjamini–Hochberg multiple testing correction. The altered genes were aligned with the cell-types database of the Allen Brain Map (https://portal.brain-map.org/atlases-and-data/rnaseq).

For cell-type proportion analysis, significantly altered 126 genes (64 upregulated, 62 downregulated) in the mouse exon array were analyzed using Bisque⁶⁹, a computational tool for estimating cell-type composition from bulk RNA-seq data by leveraging single-cell RNA-seq references.

Flow cytometry analysis by fluorescence activated cell sorting

Cells were enzymatically dissociated using trypsin and resuspended in the neuronal culture medium described above. To block nonspecific Fc receptor binding, mouse Fc Blocker (BioLegend) was added to the cell suspension and incubated at 4 °C for 10 min. Cells were then incubated with a biotin-conjugated anti-NOTCH1 primary antibody (1:1000, HMN1-12, BioLegend) or the corresponding isotype control (Armenian Hamster IgG, 0.5 mg/mL, BioLegend) at 4 °C for 30 min. After washing with phosphate-buffered saline (PBS) supplemented with 0.5% bovine serum albumin (BSA), cells were stained with allophycocyanin (APC)-conjugated streptavidin (Thermo Fisher) for secondary detection. Propidium iodide (PI) was used for live/dead cell discrimination.

Flow cytometric data were acquired using a BD LSRFortessa™ cell analyzer (BD Biosciences) and analyzed with BD FACSDiva software v9.0 and FlowJo v10 (BD Biosciences). Dead cells were excluded based on PI staining, and debris was gated out using forward scatter (FSC) and side scatter (SSC) parameters. Doublets were excluded using FSC-H vs. FSC-W and SSC-H vs. SSC-W. HMN1-12–positive cells were gated using the isotype control as a reference.

Protein analysis

Brain tissues or cultured cells were homogenized in lysis buffer containing PBS, 10 mM ethylenediaminetetraacetic acid (EDTA), 1% Triton X-100, 0.2% sodium dodecyl sulfate (SDS), and a protease inhibitor cocktail (Roche Applied Science). For quantitative western blot analysis, 40 µg of total protein was separated by polyacrylamide gel electrophoresis and transferred onto a polyvinylidene fluoride membrane (Merck Millipore). Visualization was performed using an HRP-conjugated secondary antibody and enhanced chemiluminescent detection (Pierce; Thermo Fisher).

Signal acquisition was conducted with an image analyzer (LAS500, GE Healthcare, Chicago, IL), and quantification was performed using ImageJ (National Institutes of Health (NIH), Bethesda, MD). Signal intensities were normalized to GAPDH as an internal control. The uncropped blots are provided in the Source Data file.

Immunocytochemistry, image acquisition, and analysis

Cultured neurons were fixed in ice-cold 4% paraformaldehyde in 100 mM phosphate buffer (pH 7.4) for 20 min. After fixation, neurons were permeabilized with blocking solution containing 0.15% Triton X-100 and 5% normal donkey serum in PBS for at least 30 min at room temperature. Primary antibody incubation was performed overnight at 4 °C, followed by visualization using Alexa Fluor 546- or 488-conjugated secondary antibodies.

Images were acquired using a Zeiss LSM700 confocal microscope (Zeiss, Oberkochen, Germany) and processed with Adobe Photoshop and Illustrator (Adobe, San Jose, CA). For quantification, the positive cell of each GABAergic-INs was defined by the double stained cell with cell type-specific antibody and GAD67 antibody. The total number of the positive cells on coverslip was counted and expressed as a proportion of the total number of cells. Representative images were captured from randomly selected fields in each experimental condition.

Immunohistochemistry, image acquisition, and analysis

Postnatal week 3 − 14 mice were transcardially perfused with 4% paraformaldehyde in 100 mM phosphate buffer (pH 7.4). Brains were dissected and post-fixed for 60 min. Embryonic and neonatal brains (E14–P0) were collected and fixed in the fixative for 24 h. After washing with PBS, brains were cryoprotected in 25% sucrose/PBS until they sank. Sections were cut at 50 µm (P21-adult) or 14 µm (E14, P0) using a cryostat (Leica Microsystems, Wetzlar, Germany). For HDAC3 and PROX1 double staining, fixed E15 brains were embedded into paraffin blocks and sectioned at 3 μm. Sections were permeabilized with blocking solution containing 0.3% Triton X-100 and 10% normal donkey serum in PBS for 30 min at room temperature. Primary antibody incubation was performed overnight at 4 °C, followed by visualization using Alexa Fluor 546- or 488-conjugated secondary antibodies.

Images were acquired from the CGE/MGE of embryonic brains and the cerebral cortex (layers I–IV) of postnatal brains using a Zeiss LSM700 confocal microscope. Image analysis was performed using ImageJ (NIH). For quantification in postnatal brains, the number or density of marker-positive cells was measured per unit area (/mm²). A single threshold was set for each staining condition to ensure clear cell identification. For quantification in embryonic brains, the proportion of marker-positive and/or EdU-positive areas relative to the total CGE or MGE area was measured.

In utero EdU labeling

Pregnant female mice were exposed to VPA, poly(I:C), or HDAC inhibitors, simultaneously or prior to a single intraperitoneal injection of 8 mg/kg EdU (Click-iT EdU Imaging Kit, Thermo Fisher Scientific) on E13 or E14. The mice were sacrificed 1 day (Figs. 2h-i, 4b-d) or 6 days (Fig. 2j-k) after EdU injection. Embyonic and newborn brains were dissected and fixed in 4% paraformaldehyde for 24 h. EdU-labeled cells were detected according to the manufacturer’s protocol.

Behavioral analysis

Pregnant female mice were housed individually and raised their litters until weaning (P28). After weaning, offspring were housed in same-sex groups (2 − 5 mice/cage). Behavioral tests were performed in adult mice (8 − 13-weeks-old), except for the ultrasonic vocalization (USV) test (P7), as described previously⁷⁰. Both male and female mice were tested. Before each test, mice were habituated in the test room for at least 30 min. Between tests, equipment was cleaned with 70% ethanol and dried. Behavior was recorded using a digital camera placed on the side of (self-grooming) or above (three-chamber test, reciprocal social interaction) equipment and analyzed with Smart3.0 software.

USV (P7). Newborn pups emit USVs in response to maternal separation. USV calls were recorded for 5 min using an ultrasound microphone (Avisoft Bioacoustics, Berlin, Germany) placed 15 cm above the pup inside a soundproof chamber, and analyzed with Avisoft-SASLab Pro software. Body weight and axillary temperature were measured afterward.

Self-grooming (8 − 10 w). To assess repetitive behaviors, each test mouse was placed in a clean cage under white light (160 lux). The total duration of self-grooming was recorded over 25 min.

Three-chamber test (9 − 11 w). The test apparatus consisted of three chambers separated by Plexiglas walls with small openings (5 × 3 cm) for free movement. A small wire cage was placed in each side chamber. The test consisted of three 10-min sessions under white light (20 lux):

1. Habituation: The test mouse was placed in the centre chamber and allowed to explore all three chambers for 10 min.

2. Sociability: A novel mouse (stranger 1) was placed in one of the wire cages, and the test mouse was returned to the centre chamber to explore freely. Time spent in the stranger’s chamber was compared with that in the empty chamber.

3. Social Novelty: A second unfamiliar mouse (stranger 2) was placed in the opposite wire cage, whereas stranger 1 remained in the same location. The test mouse was again allowed to explore freely, and the time spent in the chamber with stranger 2 was compared with the time spent with stranger 1.

The sociability index (SI) and social novelty index (SNI) were calculated as follows:

• SI = (time in stranger chamber − time in empty chamber) / (time in stranger chamber + time in empty chamber)

• SNI = (time in stranger chamber − time in familiar chamber) / (time in stranger chamber + time in familiar chamber)

Reciprocal social interaction (11 − 13 w). Pairs of mice (matched for genotype, drug exposure, sex, age, and weight, with no prior contact) were placed in an open-field arena (45 × 45 cm) under white light (100 lux). Social interactions were recorded for 10 min, and the number of contacts, total contact duration, and resting time were measured.

Statistical analysis

GraphPad Prism 10 was used for statistical analyses. Pairwise comparisons were performed using the unpaired t-test or Mann–Whitney test. For multiple-group comparisons, analysis of variance (ANOVA) was followed by Bonferroni’s or Dunnett’s post-hoc tests. Data are presented as mean ± standard error of the mean (SEM). Statistical significance is indicated as follows: ****P < 0.0001, ***P < 0.001, **P < 0.01, *P < 0.05.

Reporting summary

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

Author contributions

YH and TI conceived the project; YH and TI designed the experiments; YH performed and analyzed experiments; YH and MN developed behavioral experiments; YH, AT, JN, Masami T, Masayuki T and TI analyzed the transcriptome data; YI performed flow cytometry: YH, HH, MI, KM, KH, GM, TT and TI contributed reagents/materials/analysis tools; YH and TI wrote the manuscript.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

The raw data of transcriptomic profiling generated in this study have been deposited in the NCBI’s Gene Expression Omnibus (GEO) database under accession code GSE293282 (Fig. 1), GSE133932 (Fig. 2), GSE293296 (Fig. 5) and GSE292235 (Fig. 6). Source data are provided with this paper.

Code availability

This study did not use custom code or proprietary algorithms. All computational analyses were conducted using publicly available packages.

Competing interests

All authors have seen and approved the manuscript, and the authors declare that they have no conflicts of interest with the contents of this article. K.M. has received personal fees from Shionogi & Co., Ltd., Sumitomo Pharma Co., Ltd., Takeda Pharmaceutical Co., Ltd., Lundbeck Japan K.K., EA Pharma Co., Ltd., and Otsuka Pharmaceutical Co., Ltd. These financial relationships are fully disclosed in accordance with the International Committee of Medical Journal Editors guidelines. These funding sources are unrelated to the present study and did not influence the design, execution, analysis, interpretation, or reporting of this research.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-70321-6.