Single-nucleus transcriptomic analysis reveals subfield-specific cell-to-cell synaptic reorganisation of the mouse hippocampus in focal temporal lobe epilepsy

Abstract

Background

Preclinical animal models are critical tools to study the pathophysiology of epilepsy. The intra-amygdala microinjection of kainic acid in mice recapitulates key features of drug-resistant temporal lobe epilepsy, including unilateral hippocampal pathology and broad resistance to anti-seizure medicines, and is increasingly used to discover novel targets for disease-modification. Defining the gene expression landscape at single-cell resolution may reveal how different cell types adapt to chronic seizures and help identify additional biomarkers and targets in this model.

Methods

We generated a single-nucleus RNA-seq dataset using ipsilateral hippocampi collected two weeks after status epilepticus in the mouse intra-amygdala kainic acid model. Samples from epileptic and control mice were processed using the 10X Genomics platform, with a total of 34,737 nuclei sequenced representing all major cellular subtypes in the hippocampus.

Results

Differential gene expression analysis revealed cell subtype- and subfield-specific dysregulation across the hippocampus in epileptic mice. Activity-regulated genes were increased in neurons in the dentate gyrus, CA1 and CA3 subfields. Dentate gyrus neurons displayed the greatest differential gene expression, enriched for synaptogenesis, plasticity and NMDA receptor signalling. Transcription profiles for CA3 glutamatergic neurons indicated mechanical and injury responses whereas CA1 signatures were enriched for neuroinflammation and wound healing. Astrocytes displayed the most differential gene expression among glial populations, characterised by mixed reactive pro- and anti-inflammatory responses and altered metabolic states. Cell-cell communication analysis predicted extensive remodelling of hippocampal signalling across cell subtypes, including for glutamate and latrophilin-mediated ligand-receptor pairing among dentate gyrus and CA2 neuron connections.

Conclusions

Our study provides an atlas of transcriptomic changes at the single cell level for the mouse intra-amygdala kainic acid model of temporal lobe epilepsy, highlighting network adaptations to hyperexcitability. The findings support a cell subtype basis for gene expression changes underlying pathophysiology and can guide novel therapeutic strategies to reduce or reverse chronic neuronal hyperexcitability in drug-resistant epilepsy.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12967-026-08978-2.

Introduction

Temporal lobe epilepsy (TLE) is a common and drug-resistant form of epilepsy in adults [1]. Frontline treatment is with anti-seizure medicines (ASMs) which principally work by dampening excitatory or enhancing inhibitory neurotransmission. Rates of seizure control have not changed for decades, however, despite the introduction of various new ASMs. In these patients, surgical resection of the hippocampus can yield significant reductions in seizure frequency and even seizure-freedom. Unfortunately, only a small proportion of eligible patients are suitable candidates for neurosurgery. Accordingly, there remains an urgent need for disease-modifying therapies for TLE and other epilepsies [2, 3].

Moving beyond symptomatic control of seizures requires an understanding of the cellular and molecular changes underlying the disease and their causal relationship. Analyses of resected hippocampi from TLE patients have revealed complex neuropathology comprising select neuron loss, gliosis, reorganisation of cell-to-cell contacts and neuroinflammation, among other changes [4, 5]. The recent availability of omics approaches has facilitated the cataloguing of the entire repertoire of molecular changes in TLE, revealing altered expression of hundreds of genes that may play roles in epileptogenesis, the network changes that persist in the chronic epileptic state, and drug resistance [6–9]. However, most studies have relied on bulk tissue-based approaches. By averaging, this obscures cell type-specific contributions to gene expression changes. Although deconvolution approaches can infer the cellular origin of the signal, there is an increasing adoption of single-cell approaches to brain disease. Recently, single-cell and -nucleus sequencing (snRNA-seq) approaches have been applied to human epilepsy. This has yielded insights into cell layer-specific vulnerabilities, neuron subtype-specific differential gene expression, glial signatures and immune-brain cell communication [10, 11].

The systemic or intracerebral administration of chemoconvulsants such as kainic acid (KA) is commonly used to model TLE in animals. An initial status epilepticus damages limbic structures such as the hippocampus and leads to an enduring state of hyperexcitability. The intra-amygdala microinjection of KA (IAKA) was originally described by Ben-Ari and colleagues in rats [12] and was later adapted for use in mouse strains [13–16]. The hallmark pathology of the mouse variant is degeneration of neurons within the ipsilateral CA3 subfield, with lesser injury to the CA1 and dentate gyrus (DG), accompanied by gliosis. The model also features enhanced neurogenesis, neuronal migration and cellular reorganisation including changes to dendritic structures within the DG [14, 17]. Spontaneous recurrent seizures (SRS) emerge within a few days in the mice, increasing in frequency over time, and these SRS are refractory to common ASMs [18]. The value of a given model depends in part on whether it reflects human pathophysiology. Gene expression analysis of the mouse IAKA model reveals extensive commonalities with the landscape in resected human TLE [19]. The model has been increasingly central to discovery and validation of novel therapeutic and disease-modifying targets and pathways including apoptosis [20], adenosine kinase [21], microRNAs [22], the TrkB receptor tyrosine kinase [23], neuroinflammation [24, 25], and an activity-regulated virally-delivered potassium channel therapy [26]. The model was recently adopted for the National Institute of Health (NIH) epilepsy therapy screening program to identify the next generation of treatments for epilepsy [18].

The availability of single-cell technologies provides opportunities to continue to align models and define the cellular source of gene expression differences and use this to inform optimal model choice for epilepsy research. Here, we used snRNA-seq to define gene expression changes across hippocampal cell types in the mouse IAKA model.

Methods

Study approval

All animal experiments were performed in accordance with the principles of the European Communities Council Directive (2010/63/EU). Procedures were reviewed and approved by the Research Ethics Committee of the Royal College of Surgeons in Ireland (RCSI) (REC 1587) and Health Products Regulatory Authority (HPRA) (AE19127/P083). All animals were treated according to European standards/regulations for animal experiments, and all efforts were made to minimize animal suffering and reduce the number of animals used in experiments.

Intra-amygdala kainic acid (IAKA) mouse model

Mice were anesthetized using isoflurane (5% induction, 1%–2% maintenance) and maintained normothermic by means of a feedback-controlled heat blanket (Harvard Apparatus). Once fully anesthetized, mice were placed in a stereotaxic frame, and a midline scalp incision was performed to expose the skull. A guide cannula (anteroposterior = 0.94 mm, lateral = 2.85 mm) was fixed in place with dental cement. Mice were also fitted with a single-channel EEG telemetry device (Data Sciences International, MN, USA), placed in the back and connected to surface skull screws overlying the hippocampus, as described [14]. Status epilepticus was induced by microinjection of 0.3 µg KA in 0.2 µL phosphate-buffered saline (PBS) into the right basolateral amygdala. Vehicle-injected control animals received 0.2 µL of sterile PBS. The anticonvulsant lorazepam (6 mg/kg; Wyetch) was delivered via intraperitoneal injection 40 min post-KA/vehicle. Mice were monitored for two weeks with EEG telemetry with video to confirm development of epilepsy. The number of SRS was determined by individually reviewing EEG records. Spontaneous seizures were defined as high-frequency (> 5 Hz), high-amplitude (more than twice baseline) polyspike discharges of ≥ 10 s duration. The number of daily seizures was quantified for each IAKA mouse.

Immunohistochemistry and microscopy

Brains were vibratome sectioned and embedded in 2% agarose and submerged in 200 mM glycine in PBS. Sections were made at 30 μm intervals and stored in PBS supplemented with 0.01% sodium azide at 4 °C until staining. Sections were retrieved from both IAKA and PBS animals and placed into wells as free-floating staining pairs. The tissue was washed with PBS, then quenched with 10 mM CuSO₄ in 50 mM NH₄Cl for 10 min. Sections were then washed first in ddH₂O, then PBS before blocking in 3% bovine serum albumin (BSA) in 0.3% PBS + Tween-20 (PBS + t) for 3 h at room temperature. The tissue was then incubated in 1% BSA in 0.1% TBS + t with GFAP; 1:1000 (Thermo Fisher, RRID: AB_10598206, Cat# 14-9892-82,) primary antibody overnight at 4 °C. The following day, tissue sections were rinsed in PBS and 0.1% PBS + t four times for 10 min each before being incubated in CF680-coupled goat anti-mouse; 1:500 (Sigma-Aldrich RRID: AB_3492091, Cat# No. SAB4600361,) secondary antibody and Hoechst; 1:10000 (Sigma-Aldrich, Cat# No. 14533-100MG) in 1% BSA in 0.1% PBS + t at room temperature for 2 h protected from light. The tissue was finally rinsed in 0.1% PBS + t four times for 10 min each and mounted on glass slides with ProLong Diamond Antifade Mountant (Thermo Fisher, REF No. P36961). Images were acquired on a Leica Stellaris laser-scanning confocal microscope with Power HyD S detectors, running Las X 4.8.1. Tile scans were performed with a 20X/0.75 dry objective maintaining zoom factor, laser intensities, and Z thickness with 24 1 μm step increments in 1024 × 1024 format. Tiles had 10% overlap and were statistically blended in the Las X software. Qualitative images were processed using ImageJ version 1.54 by Fiji [27].

Nuclei isolation

Frozen ipsilateral hippocampal tissue samples were placed in 2 ml of Nuclei Extraction Buffer (Miltenyi, 130-128-024) containing 0.2 U/µl RNAse inhibitor (Enzymatics, Y9240L) inside gentleMACS C Tubes (Miltenyi, 130-093-23) and then dissociated using a gentleMACS octo dissociator (Miltenyi, 130-096-427) with the preinstalled “Nuclei 4 C” program. The lysate was filtered through 100 μm MACS SmartStrainers (Miltenyi, 130-098-463) and centrifuged at 300 x g for 5 min at 4 °C. The resulting pellet was resuspended in 1 ml of Resuspension Buffer (0.5% BSA in PBS containing 0.2 U/µl RNAse inhibitor (Enzymatics, Y9240L)) and mixed with 300 µl Debris Removal Solution (Miltenyi, 130-109-398) in 2 ml microcentrifuge tubes. The samples were overlaid with 700 µl of Resuspension Buffer and centrifuged for 10 min at 3000 x g at 4 °C using a swinging bucket centrifuge. The resulting upper (i.e. debris-containing) layers of each sample were aspirated and discarded. The samples were washed twice by filling up their tubes with Resuspension buffer, inverting them 3 times, centrifuging them for 5 min at 300 x g at 4 °C and discarding the supernatant. The nuclei pellets were then resuspended in 200 µl of Resuspension Buffer and filtered through 40 μm flowmi strainers (SP Bel-Art, H13680-0040). To assess final nuclei concentration and quality, an aliquot of each sample was stained with Trypan Blue and counted using a haemocytometer.

Cell barcoding and library preparation

snRNAseq libraries were prepared using the Chromium Next GEM Single Cell 3’ Reagent Kit v3.1 (dual index) (10x Genomics) and the Chromium Controller (10x Genomics), according to the manufacturer’s instructions and targeting a final recovery of 3000 nuclei/sample. Libraries were quantified using a Qubit 4 fluorometer (Invitrogen, Q33238) along with the Qubit 1X dsDNA HS Assay Kit (Invitrogen, Q33231). Library quality control was performed using a Bioanalyzer 2100 (Agilent) along with a High Sensitivity DNA Kit (Agilent, 5067 − 4626).

Sequencing and 10X processing

Libraries were sequenced using a Novaseq system (Illumina) targeting a depth of 100 million, 150 bp paired-end reads per sample. The resulting FastQ files were aligned to the GRCm39 mouse genome and cell-specific gene expression was quantified using the count function of the CellRanger package (10x Genomics), using the option --include-introns=true.

snRNAseq data pre-processing and cell-subclass labelling

All analyses were conducted in R (v4.4.2). Count matrices from the different samples were merged with Seurat v5. Nuclei expressing less than 500 genes and with > 1% of mitochondrial DNA were flagged as low-quality cells and excluded from further analysis (Supplementary Fig. S1). Doublets were identified and filtered using the DoubletFinder package [28]. After filtering, nuclei were projected using UMAP clustering to examine potential batch effects. Cell-subtype identities were assigned using the Allen Brain Atlas MapMyCells tool (RRID: SCR_024672), which aligns user data to their whole–mouse brain multi-omic reference atlas and transfers corresponding labels across datasets [29].

Differential expression analysis

To investigate transcriptomic alterations, we followed a pseudobulk differential expression workflow. Gene counts were aggregated by biological replicate and cell type using AggregateExpression, followed by normalization and differential expression testing using DESeq2 (Wald test), with the statistical design accounting for the two conditions. Genes were considered significantly differentially expressed at a Benjamini-Hochberg adjusted p-value < 0.05. For each cellular subtype, a DESeqDataSet was constructed using DESeqDataSetFromMatrix, with a design formula ~ condition, where condition encoded Epi and Control samples. Library size normalization factors were estimated with estimateSizeFactors, dispersions were modeled using estimateDispersions with fitType = “local”, and Wald tests were performed via nbinomWaldTest. Differential expression between epileptic and control cells was then obtained using results with the contrast c(“condition”, “Epi”, “Control”) and alpha = 0.05. Genes with valid statistics and Benjamini–Hochberg–adjusted p-values < 0.05 were retained and exported in a Seurat-style results table (including p-value, adjusted p-value, log₂ fold change, Wald statistic, and base mean). Because pseudobulking may obscure changes in smaller subpopulations, we performed a complementary single-nucleus analysis using Seurat v5. After standard pre-processing including normalization, variable feature selection, and scaling; differential expression was assessed using the FindMarkers function with the MAST test. Results were filtered using the same significance threshold (adjusted p-value < 0.05).

Pathway enrichment analysis

Pathway enrichment analysis was performed using Ingenuity Pathway Analysis (IPA, Qiagen). For all analyses, input gene sets consisted of pseudobulk differentially expressed genes meeting the thresholds of adjusted p-value < 0.05 and absolute log₂ fold change > 0.3. IPA was used to identify enriched canonical pathways, upstream regulators, and functional interaction networks based on its curated mouse-specific knowledge base. In parallel, enrichment analyses were conducted in R using the clusterProfiler package [30], applying standard mouse organism annotations (org.Mm.eg.db) to perform Gene Ontology analysis (MF, CC and BP). Visualization of enriched pathways and gene sets was carried out using ggplot2.

Cell-cell communication analysis

Cell-cell communication changes in the model were analysed using CellChat v2.2.0 [31]. Seurat objects were subset to remove cell subtypes with < 20 nuclei, and raw UMI counts were normalized using CellChat’s internal procedure. Signalling inference was performed with CellChatDB.mouse, including protein, metabolic, and synaptic ligand–receptor categories. Signalling-relevant genes were selected with subsetData(), overexpressed genes and L-R pairs were identified, and communication probabilities were computed using the truncated mean method (trim = 0.1). Pathway-level communication scores and aggregated network statistics were then derived using standard CellChat functions. To compare conditions, CellChat objects were merged using mergeCellChat. High-confidence ligand–receptor interactions (p < 0.05, probability > 0.05, ≥ 5% expressing cells) and pathways (p < 0.05, probability > 0.02) were retained. For each interaction, the difference in communication probability (Epi − Control) was calculated and used to classify L-R pairs as gained, lost, changed, or unchanged across conditions. Visualization was carried out in R using ggplot2 and EnhancedVolcano.

qPCR validation

For RT-qPCR analysis of mRNA expression, brains were extracted from an independent cohort of mice (n = 5 per group) following transcardinal perfusion with PBS. The ipsilateral hippocampus was isolated, frozen with dry ice, and stored at -70 °C until use. Total RNA was extracted with Trizol (Invitrogen, Ireland, 15596018) using a tissue homogenizer and centrifuged at 12,000 g for 10 min at 4 °C. The supernatant was removed and incubated at room temperature for 5 min before adding chloroform (Sigma-Aldrich, Ireland, C2432), shaken and centrifuged at 13,000 rpm for 15 min at 4 °C. The resulting upper aqueous phase was transferred and isopropanol (Sigma-Aldrich, Ireland, I9516) added to it before incubation at -20˚C overnight. Samples were then centrifuged at 13,000 rpm for 30 min at 4 °C, the supernatant discarded and RNA pellet washed with and resuspended in cold 75% ethanol. The samples were further centrifuged at 12,000 g for 5 min at 4 °C, the wash steps repeated and pellet left to air dry completely. The pellet was finally resuspended in DNase/RNase-Free Water under agitation at 800 rpm for 10 min at 60 °C. The RNA concentration and quality was evaluated on a Nanodrop, only samples with 260:280 and 260:230 ratios of 1.8 to 2.2 were used.

For qPCR analysis, 500ng of RNA was treated with DNase for 15 min to remove any contaminating genomic DNA (Invitrogen, Ireland, 18068015). 25 mM EDTA (Invitrogen, Ireland, 2549307) was added before heating for 10 min at 65 °C. Samples were reverse transcribed on a thermocycler using random hexamers (Invitrogen, Ireland, 4368814) and reverse transcriptase Superscript III (Invitrogen, Ireland, 18080044). qPCR was performed on Applied Biosystems QuantStudio 5 Real-Time PCR System using ChamQ Universal SYBR qPCR Master Mix (CliniSciences, Ireland, NB-54-0173-0). 2-ΔCt normalisation was performed relative to Gapdh expression. Primers used for qPCR were: Gapdh, forward 5′-CATCACTGCCACCCAGAAGACTG-3′ and reverse 5′-ATGCCAGTGAGCTTCCCGTTCAG-3′; Abca1, forward 5′-GGAGCCTTTGTGGAACTCTTCC-3′ and reverse 5′-CGCTCTCTTCAGCCACTTTGAG-3′;Cyp51, forward 5′-ATCCAGAAGCGCAGGCTGTCAA-3′ and reverse 5′-CAGTCCGATGAGCATCCCTGAT-3′; Fdft1, forward 5′-GGATGTGACCTCCAAACAGGAC-3′ and reverse 5′-CAGACCCATTGAGTTGGCACAC-3′; Gpam, forward 5′-GCAAGCACTGTTACCAGCGATC-3′ and reverse 5′-TGCAATCAGCCTTCGTCGGAAG-3′; Hmgcs1, forward 5′-GGAAATGCCAGACCTACAGGTG-3′ and reverse 5′-TACTCGGAGAGCATGTCAGGCT-3′; Idi1, forward 5′-GCTCCTGTTACAGCAGAGATCAG-3′ and reverse 5′-CTCGCTTCACACCAATGGCGTT-3′; Adgrl3, forward 5′-CTCTTGCAGAGCCTATGTCCAG-3′ and reverse 5′-CAGTGTCAAGCAACATGGTGGC-3′; and Tenm4, forward 5′-GGAATCATCTCCACTTTGCTGGG-3′ and reverse 5′-GGTTGATGGCTAAGTCTGTGGG-3′.

Results

Single-nucleus RNA-seq profiling defines cell types in the IAKA model of epilepsy

To understand how distinct cell subtypes are affected and potentially contribute to the hyperexcitable state in the IAKA mouse model of TLE, we performed snRNA-seq using the droplet-based 10x Genomics platform (Fig. 1a). We focused on the hippocampus, a brain region that exhibits similar pathophysiological alterations in both TLE patients and IAKA mice [32, 33]. All animals were processed under identical experimental conditions, ensuring that observed differences reflect epilepsy-related effects rather than inter-individual variability or technical artifacts.

Fig. 1.

Overview of the single-nucleus RNA-seq dataset. (a) Schematic of the experimental and computational workflow used for snRNA-seq dataset generation and analysis. Mice received either intra-amygdala kainic acid (IAKA) or PBS on day 0 (D0) and brains were obtained two weeks later when mice had active epilepsy (d14). (b) Number of nuclei sequenced per condition (control or Epi (IAKA)) after quality control. (c) UMAP (uniform manifold approximation and projection) embedding of nuclei from all samples, coloured by sample of origin. (d) UMAP embedding of all nuclei labelled by annotated cell subtypes. Note the largest cell population are dentate gyrus (granule) neurons. (e) Cell type-specific expression of canonical marker genes used to identify and annotate major cell subtypes (MapMyCells)

Two weeks after intra-amygdala microinjections, ipsilateral hippocampal tissue was micro-dissected from eight mice (four epileptic/KA and four non-epileptic/PBS), followed by nuclei extraction and sequencing. All IAKA mice had multiple EEG-confirmed spontaneous seizures. The number of spontaneous seizures was within the expected range for the model [14, 16], although it varied at an individual animal level (Supplementary Fig. S2). Sequenced mice were not directly tested for ASM resistance in this study.

In total, we sequenced 34,737 nuclei, of which 29,260 passed quality control filters (Fig. 1b), with an average of 2,954 genes detected per sample. We used Seurat v5 [34] for data integration and downstream analysis. Dimensionality reduction confirmed successful integration across samples without a major batch effect, as confirmed by integration of all samples in each cluster (Fig. 1c).

Nuclei were annotated using the MapMyCells (MMC) framework from the Allen Brain Atlas to obtain reference-based annotations from a whole mouse brain dataset. Unbiased label transfer, transferring cell subtype labels to our cells from the MMC atlas, identified the expected hippocampal subtypes, indicating that the dataset captured the cellular diversity of the brain region (Fig. 1d). Hippocampal cell subtypes were consistently represented across all samples (Supplementary Fig. S3).

Initial annotation classified nuclei into 11 major cell clusters, demonstrating there is extensive cell type-specific transcriptional heterogeneity in the mouse hippocampus. We further divided these clusters due to biological importance in the hippocampus and relevance to TLE, refining the dataset into 21 subtypes based on marker genes defined in the MMC atlas (Fig. 1e). We additionally tested for region-specific gene markers to further confirm correct annotation (Supplementary Fig. S4-6). Cell subtypes represented by fewer than 20 nuclei per sample and that could not be integrated with other subtypes further up the annotation tree were discarded from downstream analyses.

Among excitatory neurons, which represent approximately 60% of the dataset, glutamatergic intratelencephalic–extratelencephalic-like (IT–ET) clusters were subdivided into subtypes corresponding to subiculum-prosubiculum glutamatergic (SUB-Pros Glut) neurons, CA1 Glut neurons, CA2 Glut neurons, CA3 Glut neurons and near-projecting cortex layer 6b glutamatergic (NP CT L6b Glut) neurons. Nuclei that could not be assigned to any of these subtypes were retained as IT-ET glutamatergic neurons. Dentate gyrus (DG) neurons and dentate gyrus-piriform (DG-PIR) immature neurons were also identified (Fig. 1d). Inhibitory neuronal subtypes represented approximately 7% of the neuronal population and included parvalbumin (Pvalb) GABA neurons, somatostatin (Sst) GABA neurons, vasointestinal peptide (Vip) GABA neurons, Lamp5 GABA neurons, and immature GABA neurons (Fig. 1d).

Glial populations represented approximately 25% of the detected cells, and comprised microglia, astrocytes, oligodendrocyte progenitor cells (OPCs) and oligodendrocytes. Additionally, we detected vascular cells and choroid plexus cells, which have been previously identified in studies on the hippocampus within mouse models [35].

The most abundant cell subtypes detected in our dataset were DG neurons (6534 nuclei) followed by CA1 Glut neurons (3645), oligodendrocytes (3566), and astrocytes (3226). Thus, our snRNA-seq dataset comprises a representative and sufficiently extensive collection of nuclei from the hippocampus that includes the major cell types known to undergo structural or functional changes in chronic epilepsy.

Cell subtype-specific transcriptional dysregulation highlights DG neurons in epilepsy

To identify the cell subtypes in which gene expression was altered, we performed differential gene expression (DEG) analysis between epileptic and control mice for each individual cell class. Uniform manifold approximation and projection (UMAP) clustering was first used to confirm the presence of samples from both conditions within each subtype (Fig. 2a). We also assessed if there were any changes in subtype proportion in the model. This revealed that cell numbers were largely similar between control and epileptic samples, although there were non-significant trends to higher numbers of microglia/astrocytes and lower numbers of CA1 Glut neurons, CA2 Glut neurons, SUB-ProS Glut neurons, and oligodendrocytes in epileptic mice (Fig. 2b, Supplementary Fig. S7).

Fig. 2.

Disease-related cell subtype-specific changes in gene expression in the epileptic hippocampus. (a) UMAP plot of the dataset with nuclei coloured by condition. (b) Barplot showing the number of cells in each subtype per condition (red: epileptic samples; blue: controls). (c) Barplot indicating the number of significant DEGs (p_adj < 0.05) across subtypes after pseudobulk analysis; scatter plots show the log₂ fold changes (log₂FC) of these genes. (d) Heatmap for the top 10 significant DEGs (p_adj < 0.05) across the 5 most dysregulated subtypes with pseudobulk analysis, with hierarchical clustering applied to both rows and columns. (e) Volcano plot of DEGs detected in Dentate Gyrus neurons. Dashed lines drawn at -log₁₀p = 1.3 and |log₂FC| = 0.5

To account for sample-level variability and reduce the risk of inflated p-values, we applied a pseudobulking approach prior to DEG analysis [36]. Using DESeq2 [37], we identified DEGs, filtering significant results with an adjusted p-value threshold of < 0.05 (Fig. 2c). DG neurons exhibited the most extensive transcriptional dysregulation with 941 DEGs (Fig. 2d), followed by astrocytes (174), CA1 Glut neurons (110), CA3 Glut neurons (26), and NP CT L6b Glut neurons (12) (Supplementary Table 1). As expected, markers of seizure activity in the brain were found to be elevated in certain cell types in the hippocampus from epileptic mice. This included Bdnf upregulation in CA1 Glut neurons and DG neurons, and Creb upregulation in DG neurons (Fig. 2e). We also detected several upregulated genes not previously linked with epilepsy, including higher levels of Yap1 and Slc29a4 in CA3 Glut neurons, (Fig. 2d, Supplementary Table 1).

Given the relatively low number of cells captured for some subtypes previously reported to be affected in epilepsy models, such as microglia and inhibitory neurons [38], we also analysed the data using Seurat’s FindMarkers function to support the identification of changes in smaller subtypes (data available in Supplementary Table 2). The non-pseudobulked analysis was able to identify DEGs in low abundance cell types, including genes such as Apoe overexpression in microglia, which has been linked to disease progression [39], and a reduction in the expression of Slc6a1 in Lamp5 GABA neurons, suggesting altered GABA dynamics in the model [40].

While this method increased our sensitivity to detect transcriptomic changes in subtypes that were under-detected by the stricter pseudobulk approach, results from non-pseudobulked approaches must be interpreted with caution, as there is a higher probability of false positives or Type I errors [36]. We continued the main analyses using the pseudobulked results. We additionally explored whether inter-animal variation in seizure burden was associated with cell subtype-specific transcriptional changes. Exploratory PCA-based analyses suggested that animals with greater seizure burden and/or more recent seizure activity may exhibit larger transcriptional shifts in some cell subtypes, although the small number of animals precluded statistical testing (Supplementary Fig. S8). Overall, these findings demonstrate large-scale cell subtype-specific dysregulation of gene expression that affects both neuronal and glial subtypes in the hippocampus of the mouse IAKA model.

Distinct cell subtype-specific signalling pathway alterations across the hippocampus

To explore the cellular pathways and processes associated with the observed gene changes in the model, we performed pathway enrichment analysis using Ingenuity Pathway Analysis (IPA) on DEGs across the various subtypes. Significantly enriched pathways were identified in five cell subtypes: four neuronal and one glial. We detected a strong cell subclass-specific pattern, where each subclass exhibited a largely unique set of enriched pathways (Fig. 3a), indicating distinct responses across subtypes and highlighting heterogeneous adaptations to epilepsy.

Fig. 3.

Pathway enrichment analysis of pseudobulked DEGs using ingenuity pathway analysis. (a) UpSet plot showing the overlap of significantly enriched pathways across major cell subtypes, based on IPA results from pseudobulked DEGs. b-d) Top 20 significantly enriched pathways in (b) Dentate gyrus (DG) neurons, (c) CA1 glutamatergic (Glut) neurons, and (d) CA3 Glut neurons. Bars are coloured by IPA activation Z-score, where positive values indicate predicted activation and negative values indicate predicted inhibition. Grey indicates there was not enough data for calculating the Z-score. Ratio indicates percentage of genes from the specific pathway that were present in our data. (e) Top 20 significantly enriched pathways in both astrocytes and DG neurons. Colouring and interpretation of Z-score as in b). (f) Heatmap with the top 20 predicted upstream regulators based on IPA results. Positive values indicate that activation of the regulator could generate the up- and down-regulated gene pattern observed in the dataset, while negative values indicate that the inactivation of the regulator would drive the changes

DG neurons exhibited the highest number of significantly enriched pathways (Fig. 3b, Supplementary Table 3), suggesting that these neurons undergo the most extensive transcriptional remodelling in the epileptic hippocampus. The top enriched pathways were associated with hyperexcitability phenotypes and excitatory neurotransmission, including Activation of NMDA receptors and postsynaptic events and Glutamatergic Receptor Signalling. Synaptic plasticity pathways were also upregulated, such as Synaptogenesis signalling pathway and Synaptic Long-Term Potentiation, consistent with synaptic remodelling in the region. While several enriched pathways are known hallmarks of epilepsy, we also identified pathways that have not been previously reported in DG neurons, such as the HEY1 signalling pathway (Supplementary Table 3), a regulator of neurogenesis and neurite growth [41, 42].

In contrast to DG neurons, the strongest enrichment in CA1 Glut neurons was in neuroimmune and inflammation-related pathways, including Neuroinflammation Signalling, Wound Healing Signalling Pathway, IL-4, IL-12 and IL-33 signalling (Fig. 3c, Supplementary Table 3). In addition, pathways associated with extracellular matrix (ECM) remodelling were enriched, such as Collagen Biosynthesis or Degradation of the extracellular matrix, suggestive of structural remodelling. Upregulation of Autophagy, TGF-β Signalling and p53 Signalling were also observed, consistent with cellular stress responses and inflammatory activation.

In CA3 Glut neurons, the hippocampal region displaying the main lesion site in the IAKA model [14], we observed enriched pathways such as HIPPO signalling (evidenced by alterations in genes Yap1, Pard3, and Tead1) and IL-3 signalling (due to alterations in Csf2rb2 and Prkcd expression) (Fig. 3d). These changes emphasize a mechanical and tissue injury response, which could reflect damage-induced structural remodelling in this region.

In the cortical NP CT L6b Glut neuronal subtype, enriched pathways were linked to neuronal damage related processes, such as the Wound Healing signalling Pathway, senescence-related pathways such as the Senescence-Associated Secretory Phenotype (SASP), and the Myelination Signalling Pathway. Alterations in this subtype were driven by changes in Fos, Lama2, and Dock10 expression and showed a similar enrichment profile to that observed in CA1 Glut neurons. Thus, although most altered pathways were unique to specific subtypes, we observed overlaps across cell populations, indicating common or coordinated responses. DG neurons and astrocytes shared the largest number of enriched pathways (n = 53; 25% of DG total enriched pathways and 49% in astrocytes), involving processes such as Synaptogenesis signalling Pathway, Synaptic Long-Term Potentiation, and CREB signalling in Neurons (Fig. 3e), pointing to coordinated astrocyte-neuron signalling alterations in the epileptic hippocampus. Together, the findings indicate that the hyper-excitable state features degenerative, inflammatory and resilience-based processes that differ in a cell subtype- and location-specific manner generated by network and localized microenvironments in the epileptic hippocampus.

Upstream regulators of transcriptional responses in the epileptic hippocampus

To identify potential transcriptional drivers of the observed gene expression changes, we performed IPA’s upstream regulator analysis on pseudobulked DEGs from the four cell subtypes with the largest transcriptional changes (Fig. 3f, Supplementary Table 3). This indicated a prominent inflammatory signature, driven by predicted activation of canonical pro-inflammatory cytokines including IL6, IL1B, IFNG and TNF. This pro-inflammatory profile was broadly shared across DG and CA1 glutamatergic neurons, as well as astrocytes. In addition, modulatory regulators such as TGFβ1 were also predicted to be active across multiple subtypes, indicating the coexistence of pro and anti-inflammatory signalling components within the same cellular populations rather than a strict subtype-specific segregation. Notably, our analysis also identified CREB1 as a key upstream regulator, but with different activity across subtypes. Predicted CREB1 activity was increased in DG and CA1 Glut neurons, but decreased in astrocytes, indicating divergent regulation across cell subtypes. CREB1 is a major regulator of activity-dependent transcription changes and has been linked to pathways involved in synaptic plasticity, including cAMP signalling and long-term potentiation. While CREB1 has previously been associated with epilepsy [19], this confirms its cell subtype-specific activity within this model.

Cellular communication analysis reveals bi-directional adjustment of cell contacts in the epileptic hippocampus

To investigate how cellular communication networks are altered in the epileptic hippocampus, we applied CellChat [43] to our snRNA-seq dataset. CellChat infers potential intercellular communication networks based on the expression of known ligand–receptor (L-R) pairs across cell subtypes allowing estimation of signalling probability, which allows us to investigate pathway-level communication. Inferred communication pathways were categorized based on whether signalling interactions were gained or lost in epilepsy or altered in strength (Δprob =probEpi - probControl).

CellChat identified 12,231 significant (padj < 0.05) L-R interactions across conditions (Fig. 4a, b). A total of 1,260 interactions were gained in epilepsy, 1,590 were lost, and 132 showed altered strength (Δprob > 0.05), indicating substantial rewiring of intercellular contacts and communication systems. We identified a major increase in L-R pairs detected in DG-PIR immature neurons and DG neurons when they act as signalling sources, likely reflecting neurogenesis-related processes. CA2 Glut neurons displayed the strongest loss of inferred L-R pairs in the same situation. DG neurons and DG-PIR immature neurons also presented the largest increase in differential interaction strength when serving as receivers. These results indicate extensive reshaping of the communication landscape among hippocampal cell subtypes in the IAKA model.

Fig. 4.

CellChat reveals altered cell-cell communication in the epileptic hippocampus. (a) Heatmap of inferred cell-cell communication changes. Cell subtypes in the x-axis are receivers of the information flow (expressing receptors) while those in the y-axis are senders (expressing ligands). Positive relative values (red shading) indicate a higher number of inferred communications between subtypes in the epileptic mice, while negative values indicate lower. (b) Bar plot ranking cell subtypes by the summed significant communication probabilities (p_adj < 0.05) across all signalling pathways. (c) Dot plot ranking signalling pathways by the summed significant communication probabilities (p_adj < 0.05) across all cell subtypes. (d) Dot plot showing glutamatergic signalling across all cell subtypes as the summed probability of L-R pairs. Positive values indicate increased signalling in epileptic mice. (e) Heatmap of CA2 Glut neurons mean interaction probability for glutamate-related ligand–receptor pairs grouped by receptor class. Incoming signalling denotes interactions where CA2 neurons express the receptor, whereas outgoing signalling denotes interactions where CA2 neurons express the ligand. Black outlines denote a number of L-R pairs higher than three for the communication

To identify the populations with the largest communication alterations, we ranked cell subtypes by the summed absolute difference in communication probability (|Δprob|) across all altered L-R pairs. The most altered subtypes were DG-PIR immature neurons, Vip GABA neurons, CA3 Glut neurons, CA2 Glut neurons, IT-ET Glut neurons and DG neurons (Fig. 4b). Neuronal subtypes generally exhibited larger differences in communication than glial subtypes and within glial subtypes, OPCs and astrocytes were predicted to be the most affected.

At the pathway level, 33% of signalling pathways were predicted differentially active between conditions with 464 gained in epilepsy, 309 lost, and 957 detected in both conditions and changed (Supplementary Table 4). Ranking signalling pathways across all subtypes by |Δprob| revealed that inter-cellular signalling in the epileptic hippocampus was dominated by changes in glutamate, NRXN (neurexin), ADGRL (latrophilin), GABA-A and PTPR (receptor-type protein tyrosine phosphatase) signalling (Fig. 4c, Supplementary Table 4). Glutamate pathways displayed both increased and decreased strength across subtypes. In contrast, GABA-A signalling showed an overall reduction, consistent with loss of inhibitory control of network activity in epilepsy.

We proceeded to examine the diverse responses in glutamate signalling across subtypes between conditions (Fig. 4d). We ranked subtypes by the summed differential probability of glutamate-related L-R pairs, which highlighted that DG-PIR immature neurons presented the strongest increase in glutamate-related signalling probability. Notably, CA2 Glut neurons were the neuronal subtype showing the greatest reduction in glutamate signalling. To better understand the reduced glutamate signalling in CA2 Glut neurons, we extracted all glutamate L-R pairs in this subtype that were statistically significant (padj < 0.05). We grouped the L-R pairs (400 L-R pairs in the Cellchat.mouse database) into four major glutamate receptor classes: AMPA (64/400 L-R pairs in database), kainate (144/400 pairs), NMDA (64/400 pairs) and mGluR (128/400 pairs). We identified an overall reduction in glutamate signalling across all receptors from both CA2 to other subtypes (outgoing), as well as in the opposite direction (incoming to CA2) (Fig. 4e). The receptor class with the highest number of altered L-R pairs across subtypes with CA2 Glut neurons was AMPA (206), followed by mGluR (81), kainate (41) and NMDA (25), indicating that epilepsy in the IAKA model is associated with major remodelling of AMPA-mediated glutamate signalling among CA2 Glut neurons. Overall, the findings indicate that neuronal subtypes are the main drivers of cell–cell communication changes in the epileptic hippocampus, with distinct subtype-specific responses consistent with substantial reorganisation of excitatory and inhibitory networks in the IAKA model.

Astrocyte metabolic dysfunction and altered neuron–glia signalling in the epileptic hippocampus

Glial dysfunction has been implicated in epilepsy [44, 45], and astrocytes are intimately involved in memory formation [46], a function compromised in the IAKA model of epilepsy. We confirmed the presence of astrogliosis in the hippocampus of IAKA mice by immunofluorescence microscopy (Fig. 5a). Astrocytes had the highest number of differentially expressed genes among glial subtypes in the snRNA-seq dataset (Fig. 2c). Pathway enrichment analysis of those DEGs highlighted that several cholesterol-related pathways were among the most significantly enriched, including cholesterol biosynthesis pathways, with an inferred reduction in pathway activity (Fig. 5b). In parallel, several neuronal-astrocyte crosstalk signalling pathways were predicted to be upregulated, including Calcium signalling, CREB signalling in neurons, and Synaptogenesis signalling pathway. Immune-related pathways such as IL-4 signalling, IL-10 signalling and Acute Phase Response signalling were also enriched, consistent with a reactive astrocyte profile (Fig. 5b). Pathway enrichment analysis also detected L1cam signalling, which was inferred to be activated in epilepsy, indicating altered astrocyte-mediated synaptogenesis and axon formation pathways. Consistent with our global cell–cell communication analysis, astrocyte-associated signalling pathways, including glutamate signalling and axon guidance-related families such as NRXN and ADGRL, showed altered signalling probabilities in epilepsy (Fig. 5c).

Fig. 5.

Dysregulated astrocyte function in the epileptic hippocampus. (a) Immunofluorescence images of the dentate gyrus from control (PBS) and epileptic (IAKA) mice. Images highlight the molecular layer (ML), granule cell layer (GCL), and hilus (H). Astrocytes are labelled with GFAP (yellow) and nuclei with Hoechst (blue). (b) Top 20 most significantly enriched pathways in astrocytes. Dots are coloured by IPA activation Z score, where positive values indicate predicted activation and negative values indicate predicted inhibition. Grey indicates there was not enough data for calculating the Z-score. Ratio indicates percentage of genes from the specific pathway that were present in our data. (c) Bubble plot with the summed mean difference of interaction strength for the altered cell-cell communication pathways found in astrocytes. The x-axis denotes mean |Sum. prob| (effect size across L-R pairs in each pathway). Colours indicate the direction of change (red for positive; blue for negative), and size reflects the number of altered L-R pairs contributing to each pathway. (d) Expression of astrocyte DEGs involved in the cholesterol biosynthesis pathway. Significant differences are marked with asterisks (p_adj < 0.05). (e) Expression of L1cam interaction DEGs in astrocytes. Significant differences are marked with asterisks (p_adj < 0.05)

We next investigated the genes driving these changes in astrocytes. Several key genes involved in cholesterol biosynthesis, including Cyp51, Gpam, Fdft1, Idi1, and Hmgcs1, were dysregulated in astrocytes (Fig. 5d). We additionally performed RT-qPCR for Abca1, Cyp51, Fdft1, Gpam, Hmgcs1, and Idi1 using an independent set of control and IAKA epileptic ipsilateral hippocampi. In contrast to the astrocyte-specific snRNA-seq signal, RT-qPCR on bulk tissue showed increased expression of all six genes at the whole-hippocampus level, with five reaching statistical significance and Abca1 showing a non-significant increase (Supplementary Fig. S9). We further compared these results with publicly available bulk RNA-seq datasets from the same model and tissue at different time points and observed a similar pattern of increased expression (Supplementary Fig. S9), suggesting divergence between single cell and bulk tissue findings for astrocyte-specific transcriptional responses.

We also examined the genes contributing to L1CAM signalling, a pathway involved in neuronal development and cell adhesion. This revealed increased expression of genes associated with adhesion (Itga5), guidance (Dnm3), and cytoskeletal remodelling (Ank3), suggesting that astrocytes may actively contribute to synaptic adhesion and circuit remodelling in the epileptic hippocampus (Fig. 5e). Altogether, our findings indicate that astrocytes adopt a distinct molecular program in the epileptic hippocampus, characterised by mixed inflammatory, disrupted cholesterol biosynthesis and altered synaptogenesis-related pathways.

Disrupted latrophilin signalling in epileptic hippocampal circuitry

After glutamatergic signalling, ADGRL (latrophilin) signalling was among the most prominently altered communication pathways in epileptic mice (Fig. 4c), suggesting that latrophilin signalling may represent a major synaptic communication axis affected in epilepsy. CellChat identified this based on the expression of the latrophilin receptors (genes Adgrl1-4) and their ligands, including teneurins and neurexins. Latrophilins represent compelling targets in epilepsy because they regulate synaptic assembly and organisation [47], processes disrupted in hyperexcitable networks that underlie epilepsy.

We first examined which L-R pairs within the ADGRL signalling network were altered across cell populations and identified Tenm4 - Adgrl3 as one of the most frequently altered interactions in the IAKA model (Fig. 6a), as well as the interaction with one of the largest changes in signalling strength in DG neurons (Fig. 6b). Differential expression analyses revealed that both Tenm4 (log₂FC = 0.88) and Adgrl3 (log₂FC = 1.406) were significantly upregulated in DG neurons (Fig. 6c, d). In addition, several known ligands of ADGRL3 showed altered expression in epilepsy (Fig. 6d), including Tenm1 (log₂FC=-0.5), Nrxn3 (log₂FC = 1.543) and Unc5c (log₂FC = 0.755) / Unc5d (log₂FC = 1.624). We additionally performed RT-qPCR for Adgrl3 and Tenm4 using samples from the independent batch of control and epileptic ipsilateral hippocampi. This confirmed statistically significant upregulation of both genes and the same direction of change as observed in our snRNA-seq dataset (Fig. 6e).

Fig. 6.

Altered latrophilin-associated signalling and gene expression in the epileptic dentate gyrus. (a) Heatmap showing cell subtypes ranked by the summed change in the number of inferred ligands receptor interactions within the ADGRL signalling pathway between conditions. (b) Heatmap showing cell subtypes ranked by the summed change in interaction strength (Δprob) of detected ligand–receptor pairs within the ADGRL signalling pathway. (c) Expression of latrophilin receptor gene family (ADGRL) in DG neurons (normalized counts). Significant differences are marked with asterisks (p_adj < 0.05). (d) Expression of known latrophilin ligand genes in DG neurons (normalized counts). Significant differences are marked with asterisks (p_adj < 0.05). The panel is split due to differences in the scale of the normalized counts. (e) RT-qPCR validation of Adgrl3 and Tenm4 expression in whole ipsilateral hippocampus from control and epileptic IAKA mice. Individual points represent biological replicates. (f) Heatmap of the log₂FC of latrophilin family genes and their ligands across different epilepsy bulk RNA-seq murine datasets and a human TLE single-nuclei dataset. Positive values indicate a higher expression of the gene in epilepsy. Significant differences are marked with asterisks (p_adj < 0.05). Grey indicates no data found in the differential expression analysis

To evaluate whether this pattern is present in other datasets, we examined publicly available bulk RNA-seq from several mouse epilepsy models, including the intrahippocampal kainic acid and pilocarpine models spanning time points encompassing both epileptogenesis and chronic stages of epilepsy [5, 19, 48, 49]. We observed consistent upregulation of Adgrl3 across several models and timepoints, accompanied by increased Bdnf expression, a downstream gene previously associated with ADGRL3-mediated excitatory synapse formation [47] (Fig. 6f, Supplementary Table 5). We also examined a single-nuclei dataset derived from cortex tissue resected from epilepsy patients [10], and identified upregulation of ADGRL3, particularly in L4_Rorb_Met neurons (Fig. 6f). Thus, latrophilin (Adgrl3)-mediated signalling and its gene expression profile may be a feature of the pathophysiology of epilepsy.

Discussion

Here, we report the single-nucleus transcriptional landscape of the ipsilateral hippocampus from the IAKA model of focal, temporal lobe epilepsy in mice. The dataset reinforces the transcriptional diversity of cell types and reveals distinct directional differences in neuronal and glial subtype transcript expression across hippocampal subfields and cellular compartments in mice with epilepsy. Moreover, we demonstrate a gradient of neuronal and glial subtypes undergoing predicted cell-to-cell communication shifts and identify potential underlying synaptic structure-associated ligand-receptor changes. Together, the data strengthen our understanding of transcriptional heterogeneity among cell types that define the epileptic state in the IAKA model and provide a resource for new biomarker and target discovery for drug-resistant epilepsy.

The present study is an important addition to our understanding of the molecular changes in experimental TLE [19]. Multi-omics techniques such as the 10X Genomics platform have enabled high-resolution mapping of gene activity in specific cell types and subclasses [50, 51]. This is increasingly being applied to yield a complete description of the pathophysiologic landscape of experimental and human epilepsy [10, 11, 52]. The IAKA model in mice has demonstrated validity in terms of reflecting the unilateral hippocampal pathology of human TLE, recurrent spontaneous seizures accompanied by clinical behaviour, a range of co-morbidities, multi-drug resistance profile and has been pivotal to the discovery and validation of anti-seizure and disease-modifying compounds [20–25]. The current mapping of cellular transcriptional heterogeneity represents an advance on earlier bulk (cell-averaged) tissue profiling of the model [19], enabling the identification of subtype-specific adaptations that would otherwise remain undetected and providing new insights into epilepsy-related circuit dysfunction. Notably, several of the processes identified here are consistent with findings in the hippocampus of TLE patients, including alterations in dentate gyrus neurons and cell-type specific changes linked to circuit dysfunction, supporting the translational relevance of the model [53]. Further, this preclinical resource enables researchers to generate testable hypotheses on cell type-contributions to disease mechanisms, biomarker sources and therapeutic targets.

The main lesion site and presumed site of ictogenesis in the IAKA model is the ipsilateral CA3 subfield [14, 21]. An unexpected finding here was that the most prominent transcriptional differences identified by snRNA-seq analysis were in DG and CA1 glutamatergic neurons. These cells are largely resilient hippocampal cell populations in the IAKA model. The transcriptional differences for the DG neurons included enhanced NMDA receptor and CREB signalling and the upregulation of genes such as Bdnf. The enhanced glutamatergic signalling may relate to mossy fibre sprouting, and there were signatures of structural remodelling (axon guidance molecules), possible compensatory neuroplastic changes, and neurogenesis (Wnt/β-catenin) [54, 55]. Thus, the transcriptional landscape of DG neurons seems to reflect multiple and ongoing processes linked to long-term circuit remodelling after epilepsy is established [14, 17, 22]. For CA1 neurons, we observed enrichment of immune and neuroinflammatory pathways, reflecting a distinct adaptive transcriptional response in this subfield. Thus, our snRNA-seq has provided cell-type resolution of these pathways in epilepsy and supports previous findings suggesting that neuroinflammatory signalling may have protective roles [24, 56].

In contrast, CA3 glutamatergic neurons and GABAergic neurons displayed only moderate differential gene expression. This could reflect the lower representation of these neuronal populations in the dataset and the consequently reduced statistical power to detect transcriptional changes. Alternatively, the findings may be due to seizures or injury causing suppression of transcription [57], selective survival of less vulnerable nuclei of these cell types, or the local cellular composition of the CA3 subfield buffering transcriptional responses of CA3 neurons to injury [58, 59]. Variance across neuron subtypes could also reflect uncoupling and non-linearity between transcriptional signatures and perceived cellular contributions to epilepsy. This has implications for how researchers appraise gene expression data, underscoring the need to integrate systems-level knowledge of a model, its pathology and electrophysiology, alongside gene expression data in reaching conclusions about disease mechanisms and potential drug targets. The present study also identifies the cellular origin of less well-characterised pathways relevant to epilepsy. These include senescence, recently reported in focal cortical dysplasia [60], and the mechano-sensitive YAP/TAZ pathway associated with Hippo signalling [61], which were dysregulated in CA3 glutamatergic neurons. These could be future therapeutic targets, and the findings reinforce that the chronic epileptic state is defined by transcriptional responses that may, individually or collectively, contribute to the recurrent seizures and drug resistance in the model.

Pathway and cell-cell analyses indicated there may be pronounced reductions in glutamatergic ligand-receptor interactions for CA2 neurons in the IAKA model. CA2 neurons are a morphologically, functionally and molecularly distinct population of pyramidal cells that lie between CA3 and CA1, receiving input from dentate granule and CA3 neurons as well as from outside the hippocampal formation [62]. The output from CA2 neurons supports hippocampal functions including contextual and social memory, and temporal encoding [62]. CA2 neurons have unique transcriptional signatures that may underlie their electrophysiologic properties as well as resilience to excitotoxic injury, including in TLE [62, 63]. Studies indicate that CA2 neurons become intrinsically more excitable in epilepsy [64–66]. Our findings offer a potential molecular explanation for these synaptic changes. Interestingly, cell communication analysis of single-cell data from other rodent models of epilepsy did not emphasize changes in CA2 [52, 67] although the intrahippocampal administration of KA causes increased connectivity to CA2 from mossy fibres [68]. This underscores the importance of single-cell data as a tool to appraise how well models reflect human-relevant pathophysiology. It is tempting to speculate that CA2 uncoupling is a homeostatic response to limit excitability within the epileptic hippocampus. Preclinical studies indicate that deactivation of CA2 neurons can suppress spontaneous seizures [65]. Decoupling of CA2 may also contribute to co-morbidities via deficits in hippocampus-related tasks [62]. CA2 changes could also represent altered cell state or changes in transcript detectability. Further studies including electrophysiologic assessment of anatomic coupling will be needed to assess the functional significance, if any, of the observed CA2 changes in the IAKA model and to determine whether these have utility in the design and delivery of anti-seizure or disease-modifying therapies.

Latrophilin signalling has recently emerged as an important regulator of synapse formation and neuronal connectivity [47]. Our analyses predict enhanced signalling involving the TENM4-ADGRL3 L-R interaction, driven by upregulation of the Adgrl3 receptor and its ligand Tenm4 in DG neurons. We confirmed the broad relevance of this pathway by independent validation of the snRNAseq data and by cross-referencing with available animal model and human TLE datasets. Dysregulation of this pathway may be a feature of epileptic hippocampal circuits. Indeed, given that teneurin-latrophilin interactions are required for synapse assembly and can elevate cAMP signalling to promote additional synaptogenesis [47, 69], this upregulation may reflect an activity-dependent increase in excitatory synapse formation on DG spines. Such changes could heighten excitatory drive, contribute to recurrent seizure activity, and further reinforce synaptogenic pathways [47, 69]. Dysregulation of this pathway in DG neurons may reflect integration of new-born hippocampal cells following neurogenesis, a process also recognized as a contributor to epileptogenesis and seizure frequency [70], and altered synaptic plasticity [38]. This is also consistent with reported alterations in human DG neurons in TLE, where increased dendritic thickness and spine density are associated with heightened excitability [53]. Functional experiments and analysis of human tissue will be needed to establish conservation within human TLE circuits and determine if ADGRL3 signalling represents a candidate for future therapeutic investigation.

Glial subtypes are increasingly recognized for their contributions to activity-dependent neuronal structure and function, and responses to injury that elicit repair or mal-adaptive states in epilepsy [44, 71]. Notably, these represent largely overlooked targets for anti-seizure medicines. Among glial subtypes, we found that astrocytes exhibited the strongest transcriptional reshaping in the IAKA model. This included both pro- and anti-inflammatory signatures. Lipid metabolism emerged as a major altered pathway in our data, consistent with previous studies where altered lipid metabolism in astrocytes contributes to hyperexcitability in both the intra-hippocampal kainate model and human TLE [72]. These changes suggest a coordinated metabolic and morphologic reprogramming that may contribute to compensatory effects or disrupt neuronal excitability. Consistent with this, cholesterol biosynthesis pathways were reduced, aligning with studies showing that impaired astrocytic cholesterol synthesis can impact neuronal function and contribute to disease progression [73].

Changes in cell number due to destruction of cell populations and structure within the hippocampus are a source of bias that has been reported in single cell studies for the intrahippocampal kainate and pilocarpine models [14, 38, 52, 74]. Our snRNA-seq dataset may be less impacted by under- or over-representation of specific cell types because of the moderate pathology in the model. Indeed, single-cell data from the intrahippocampal model revealed prominent downregulation of neuronal genes and upregulation of glial genes which is not apparent in the IAKA model dataset. This underscores the need to critically appraise the choice of models, and the clinical relevance of the IAKA model [19] which is part of the NIH epilepsy therapy screening program [75].

The present findings provide a strong rationale and may guide the design of cell type-specific therapies. We also found preliminary evidence that the extent of transcriptional changes in some cell types may track with seizure burden. There is an increasing shift in the field to genetic medicine and RNA therapies [76, 77], and it will be possible to steer therapies towards cell types. The results here and work by others [52, 74] are valuable in enabling this, for example through promoters or exploiting natural tropism, ensuring that therapies are expressed in the appropriate cell type.

While we detected transcriptional signatures consistent with cell type-specific inflammatory and immune-related signalling, we did not detect the type of immune cell infiltration recently highlighted in a single-cell analysis of resected tissue from epilepsy patients [11]. This may be due to the more limited neurodegeneration or blood-brain barrier disruption in the model. We cannot exclude, however, that very small numbers of critical cells may nevertheless influence outcomes. Regardless, since these cells may drive pathogenic or adaptive responses, the limited extent of immune cells here should be considered in the planning of discovery or evaluation of directed immune therapies in the IAKA model.

There are technical and other limitations to consider in the present study. Foremost, despite efforts to limit technical and batch-related effects, the number of biological replicates limits the strength of broad model-level conclusions, particularly for cell populations represented by lower numbers of nuclei. Sampling bias of surviving neuronal populations may also contribute to the observed differences, particularly for the transcriptional findings on DG neuron populations. Likewise, neuron loss, selective survival, lower sampling power and potential nuclei capture biases may contribute to the observed CA3 transcriptional profile. Tools such as IPA and CellChat provide predicted pathway activity and cell communication changes and are not definitive of functional activation or inhibition. Ever deeper sequencing and increasing sample sizes will yield further advances and may reveal signatures from very low abundance cell types. Whole cell sequencing would also complement nuclear sequencing and could be useful to infer effects of post-transcriptional factors such as microRNAs. Our analysis is restricted to mapping the heterogeneity of hippocampal cell types in already epileptic mice and we did not capture cell transcriptomes generated prior to spontaneous seizures which are relevant to epileptogenesis. Only male mice were used and the characterisation of the model in female mice to identify sex-specific differences in the single-cell landscape represents an important future goal. Likewise, the profiling of single-cell transcriptomes in other brain structures will add important context and be relevant to transcriptional cellular changes associated with sites of ictogenesis and seizure spread such as the contralateral hippocampus, amygdala and cortex. This could yield brain-wide mapping of the cellular heterogeneity of the transcriptional landscape in this model. Additional integration of common as well as distinct single-cell changes from other models and human data could yield a consensus atlas of epilepsy-associated genes at single cell level. Finally, integrating additional platforms could improve the spatial resolution of these findings. For example, combining 10x Genomics data with single-cell spatial technologies or Patch-seq approaches would help localize transcriptional changes across hippocampal substructures and laminae. Notably, recent studies have identified heterogeneous transcriptional profiles in dentate granule neurons, the cell subtype showing the greatest changes in the IAKA model, which are unevenly distributed along the axis of the gyrus in humans [78].

In conclusion, by using single-nucleus RNA-seq we have generated a comprehensive profiling of cell-type-specific transcriptional changes in the ipsilateral hippocampus of the IAKA model of epilepsy. Our study establishes a valuable resource to support our understanding of cell subtype-specific alterations in experimental TLE and provides a foundation for future biomarker validation and mechanistic and therapeutic investigations in drug-resistant epilepsy.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Original conception of the work: J.V.S, A.S, N.M.C and D.C.H. Contribution and design of the experiment: J.V.S, A.S, N.M.C and D.C.H. Generation of samples and collection of data: J.V.S, J.H, T.S, A.S, P.B, and O.M. Analysis and interpretation: J.V.S, N.M.C, and D.C.H. Drafting the article or revising it for important intellectual content: J.V.S, A.S, N.M.C and D.C.H. All authors approved the final version of the manuscript.

Funding

This research was funded by Research Ireland through the Research Ireland Centre for Research Training in Genomics Data Science under grant number 18/CRT/6214 and 21/RC/10294_P2 and co-funded under the European Regional Development Fund and by FutureNeuro industry partners. Images acquired in the RCSI Super resolution Imaging Consortium (SRIC) funded by Science Foundation Ireland (18/RI/5723). This work was supported by funding from Research Ireland/Irish Research Council (EpiGlymph, IRCLA/2022/3828; to JPH).

Data availability

Raw FASTQ files and processed expression matrices are available in GEO under accession code GSE343089. Supplementary files supporting the conclusions of this article are included within the article and its supplementary materials. Code to reproduce the analyses is available on GitHub (https://github.com/Javizuma/snIAKA_2026).

Declarations

Ethics approval

All procedures involving animals were performed in compliance with EU Directive 2010/63/EU on the protection of animals used for scientific purposes and procedures were approved by RCSI University of Medicine and Health Sciences’ Research Ethics Committee (REC 1587), and under license from the Ireland Health Products Regulatory Authority (AE19127/P083). Mice were housed at controlled conditions including a room temperature (20–25 °C), and humidity (40%–60%), on a 12 h dark-light cycle with ad libitum access to water and food.

Consent for publication

Not applicable.

Competing interests

Authors report no competing interests.