Cross-species transcriptomic analysis of rodent model fidelity to human mesial temporal lobe epilepsy Authors: Joonho Kim, Sangbo Lee, Bora Kim, Kyoung Hoon Jeong, Soojin Park, Soomi Cho, Hoon-Chul Kang, Won-Joo Kim, Sangwoo Kim Published: 2026 ABSTRACT AbstractMesial temporal lobe epilepsy (MTLE) is a common form of drug-resistant epilepsy, necessitating reliable rodent models for translational research. Although numerous rodent models reproduce hippocampal epileptogenesis, their molecular fidelity to human MTLE remains unclear. In this study, we conducted an integrative analysis of human MTLE transcriptomic data and evaluated rodent epilepsy models by comparing their transcriptomic profiles with those of human MTLE. By integrating 694 human samples (mean age 39.6 years, 40.3% female, 12 datasets) and 362 rodent samples (postnatal and adult mice and rats, 13 datasets), we identified synaptic dysfunction alongside elevated neuroinflammation and gliogenesis in chronic drug-resistant MTLE. Subsequent cross-species correlation analysis revealed that the intrahippocampal kainate-induced model best mimicked MTLE with hippocampal sclerosis (HS), characterized by neuroinflammation and gliogenesis. In contrast, the perforant path stimulation-induced model exhibited a high correlation with MTLE without HS, highlighting synaptic remodeling as distinct mechanism in this model. Overall, this cross-species analysis clarifies the transcriptomic landscape of human MTLE and provides a framework for selecting appropriate epilepsy models.Subject terms: Computational biology and bioinformatics, Computational neuroscience Introduction Epilepsy is a complex and debilitating neurological disorder characterized by recurrent seizures, affecting millions of individuals worldwide1. Despite advances in medical therapies, many patients develop drug resistance, demanding more effective treatment strategies2. Mesial temporal lobe epilepsy (MTLE) is a common form of drug-resistant epilepsy across different age groups3,4. In epilepsy research, the choice of an animal model that faithfully recapitulates the pathophysiology of human diseases is critical. Various rodent models have been employed to study epilepsy, reflecting substantial methodological variability5,6. In general, six key parameters shape model design and interpretation: species, induction method, administration route, lateralization, age, and disease phase. Mice and rats are the primary experimental models. Induction approaches include chemoconvulsants (for example, pilocarpine or kainate), electrical kindling via repeated subthreshold stimulation, and traumatic brain injury. Chemically induced models differ further depending on the route of administration, such as intrahippocampal, intra-amygdala, intraperitoneal or subcutaneous, each modulating seizure severity and spatial targeting. Lateralization is tailored to experimental objectives, with ipsilateral sampling capturing direct effects and contralateral sampling revealing compensatory changes. Variables, including animal age (postnatal and adult) and disease phase (from hyperacute to chronic), further influence phenotypic outcomes. Although rodent models establish an epileptic focus within the hippocampus, closely mirroring the pathophysiology of human MTLE6, the diversity and complexity of the available systems present challenges in selecting the most appropriate model. Optimizing clinical relevance requires systematic evaluation of how accurately each model recapitulates the molecular, cellular and network-level features of human epilepsy. In this study, we aimed to assess the clinical applicability of rodent epilepsy models by comparing their transcriptomic profiles with human MTLE across a comprehensive dataset consisting of 694 human samples (mean age 39.6 years, 40.3% female, 12 datasets) and 362 rodent samples (postnatal and adult mice and rats, 13 datasets) (Fig. 1a, b). First, we conducted an integrative analysis of human MTLE transcriptomic data to identify key epileptogenesis pathways. Second, we performed a cross-species correlation analysis to evaluate the degree of similarity between rodent models and human MTLE. Third, we characterized the transcriptomic landscapes of the representative models at the gene-set and individual gene levels, focusing on epileptogenesis-related pathways. This approach provides a framework for selecting appropriate epilepsy models for specific research objectives and enhances our understanding of MTLE pathophysiology. Study overview and landscape of human MTLE. a Overview of human and rodent samples. Human samples (694 samples, mean age 39.6 years, 40.3% female) were processed through sequential filtering steps, with the final dataset stratified by the presence or absence of HS. Rodent epilepsy models (362 samples, postnatal and adult mice and rats) were classified into 24 groups based on the SIALAP classification, including species, induction method, administration route, lateralization, age, and phase. b Analytical workflow outlining unified RNA processing and three main analysis steps: integrative analysis of human MTLE, cross-species correlation, and evaluation of the two most representative models. c Principal component analysis plot on TPM count illustrating the clustering of MTLE with HS, MTLE without HS, and NL, based on genes within the predefined epileptogenesis-related categories. d Box plots depicting TPM expression levels of GFAP, AQP4, NOTCH1, and ACTB in human samples. Color coding corresponds to MTLE with HS (green), MTLE without HS (blue), and NL (brown). Biological illustrations in a, b were created in BioRender. Kim, S. (2026) https://BioRender.com/2uqmryb. MTLE mesial temporal lobe epilepsy, HS hippocampal sclerosis, NL normal control, TPM transcript per million. Results Large-scale integration establishes a robust cross-species reference To overcome the small sample sizes and technical heterogeneity of individual studies, we integrated comprehensive public datasets into a unified processing pipeline, establishing a robust reference for cross-species analysis. We obtained bulk whole transcriptome data (FASTQ files) from brain tissue samples of human MTLE and rodent epilepsy models, all of which are publicly available in the Sequence Read Archive (SRA) (Fig. 1a). Datasets were systematically retrieved based on predefined criteria, including organism, disease keywords, and high-throughput sequencing technology, while microarray data were excluded due to technical incompatibility with our integration approach (see “Methods” and Supplementary Fig. 1). These datasets were integrated into a unified RNA processing pipeline, and technical variations across studies were corrected (Fig. 1b). Our study focused on genes within nine epileptogenesis-related categories: neurotransmission, ion channels, neuroinflammation, neuronal death, TNF signaling, neurogenesis, gliogenesis, mossy fiber sprouting, and synaptic plasticity (Supplementary Fig. 2 and Supplementary Data 1). These categories represent core processes replicated in rodent models7 and are defined based on established literature5,8 and Gene Ontology (GO) annotations. The workflow for human sample enrollment is detailed in the “Methods” section (Supplementary Fig. 3). In total, 694 samples from 12 datasets underwent the selection process, including RNA quantification and batch correction (Supplementary Data 2). After batch correction, the clustering pattern based on cohorts was removed successfully on raw and transcripts per million (TPM) normalized data, confirming batch effect minimization (Supplementary Fig. 4a). For downstream analyses, we selected 253 hippocampal samples from 11 datasets, including 108 MTLE cases with hippocampal sclerosis (HS), 18 MTLE cases without HS, and 127 post-mortem neurotypical controls (Supplementary Data 3). For rodent epilepsy models, we included 290 mouse samples from eight datasets and 72 rat samples from five datasets, including an intra-amygdala kainate-induced mouse model dataset generated specifically for this study (Supplementary Data 4). To systematically classify these models, we developed the SIALAP classification, which accounts for species, induction method, administration route, lateralization, age, and phase (Fig. 1a and Supplementary Fig. 5). Based on this classification, we categorized 24 distinct epilepsy models, comprising 18 chemically induced models, 5 electrically kindled models, and 1 traumatic brain injury model. All rodent datasets were processed using identical RNA processing steps, with additional batch correction to account for different sequencing platforms (Supplementary Fig. 6). Epileptogenesis-related genes reliably distinguish human MTLE from controls While MTLE and normal controls were difficult to distinguish at the whole-transcriptome level (Supplementary Fig. 4b), they became more distinct once the analysis was focused on predefined epileptogenesis-related genes (Fig. 1c). To further validate the difference between MTLE and normal controls, we examined the expression of GFAP, AQP4, and NOTCH1—key pathological markers of gliosis and neuroinflammation in the hippocampus9,10. All marker genes were significantly upregulated in MTLE, particularly in MTLE with HS, whereas the housekeeping gene ACTB remained consistently expressed across all groups (Fig. 1d). Additionally, the transcriptomic profile did not cluster according to the burden of anti-seizure medications in MTLE (Supplementary Fig. 4c). To ensure the human MTLE dataset served as a robust reference for evaluating rodent models, we applied outlier detection based on two criteria—principal component analysis (PCA) and marker gene expression, GFAP, AQP4 and NOTCH1 (see “Methods” and Supplementary Fig. 7). The first criterion was based on PC2, which provided a clearer distinction between MTLE and normal controls than PC1. The second criterion was the identification of samples with atypical expression patterns. Specifically, MTLE cases with markedly low GFAP, AQP4 and NOTCH1 expressions, as well as normal controls with abnormally high expression of these markers, were considered as outliers since they might not accurately represent their respective conditions. Thus, ten normal control samples met both exclusion criteria and were removed. The final Total Hippocampus Cohort included 243 human samples for the integrative analysis of human MTLE and cross-species comparisons (Supplementary Fig. 3). Synaptic dysfunction with concurrent neuroinflammation and gliogenesis characterizes human MTLE We conducted a human integrative analysis to characterize the transcriptomic profiles of human MTLE in Total Hippocampus Cohort. At the whole-transcriptome level, we observed an imbalance in gene expression, with MTLE-downregulated genes (5021) outnumbering MTLE-upregulated genes (1119) by nearly fivefold (Fig. 2a and Supplementary Data 5). However, this pattern was reversed within epileptogenesis-related categories, where upregulated genes (107) outnumbered those that were downregulated (72). To validate these findings, we compared our results with those of a previous analysis by Guelfi et al., which utilized microarray datasets from the temporal cortex of MTLE with HS11. Despite these technical (sequencing versus microarray) and anatomical (hippocampus versus cortex) differences, we observed substantial overlap: 15 (46.9%) of the top significant genes ( | fold change | > 1.5 and adjusted p value < 0.05) were also significant in our study. Furthermore, of all the significant DEGs reported by Guelfi et al. (adjusted p value < 0.05), 4077 genes (82.0%) showed an identical direction of effect across our entire transcriptomic profile (Supplementary Data 6). The high degree of concordance across different technologies and brain regions underscores the reproducibility of our integrative analysis. Integrative analysis of human MTLE. a Volcano plot illustrating DEGs in human MTLE compared to NL. Dots are color-coded according to the following nine epileptogenesis-related categories: neurotransmission (deep green), ion channel (vibrant green), neuroinflammation (deep indigo), neuronal death (light blue), TNF signaling (rich blue), neurogenesis (bold red), gliogenesis (distinct orange), mossy fiber sprouting (light yellow-orange), and synaptic plasticity (bright orange-yellow). The headers specify the number of DEGs associated with epileptogenesis categories, followed by the total whole-transcriptome DEGs in parentheses. b GSEA results for GO terms related to epileptogenesis are grouped into nine categories. NESs and nominal p values are shown for each category, with significant GO terms (p value < 0.05) indicated by gray borders around the dots. c Combined p values were calculated using Fisher’s method, separately for upregulated and downregulated GO terms within each epileptogenesis-related category. Upregulated GO terms in MTLE are positioned above y = 0, whereas downregulated GO terms are below y = 0. The shaded region represents non-significant findings (p value ≥ 0.05). d–f Distribution of DEGs across categories, including d neurotransmission and ion channel, e positive and negative regulation of the inflammatory response (GO:0050728, GO:0050729), and f positive and negative regulation of neurogenesis (GO:0050768, GO:0050769) and gliogenesis (GO:0014014, GO:0014015). DEGs differentially expressed genes, MTLE mesial temporal lobe epilepsy, NL normal control, GSEA gene set enrichment analysis, NES normalized enrichment score, GO gene ontology. GO term overrepresentation analysis (ORA) at the whole-transcriptome level revealed that most MTLE-upregulated genes were associated with immune-related pathways, whereas many other enriched GO terms had limited relevance to epileptogenesis (Supplementary Data 7). Conversely, MTLE-downregulated GO terms were primarily associated with the negative regulation of immune responses and chemical stimuli, indicating a persistent inflammatory state in MTLE. To further examine enrichment trends within epileptogenesis-related categories, we conducted Gene Set Enrichment Analysis (GSEA) between MTLE and normal controls (Fig. 2b). Most GO terms associated with neurotransmission, ion channels, and synaptic plasticity, except for the kainate receptor GO term, exhibited a negative trend in MTLE. In contrast, GO terms related to neuroinflammation, neuronal death, neurogenesis, and gliogenesis showed a positive trend in MTLE. When aggregating p values by enrichment direction, chronic-phase categories (neurogenesis and gliogenesis) showed the most prominent enrichment, followed by subacute-phase categories (neuroinflammation and neuronal death) (Fig. 2c). This trend also remained consistent in the Whole-Tissue Only Cohort, which excluded sub-dissected samples (Supplementary Fig. 8 and Supplementary Data 8). Although the smaller sample size in this sub-group led to a decrease in the total number of DEGs and overall statistical significance, the ratio of up- to downregulated genes and the directional enrichment trends were similar to those observed in the Total Hippocampus Cohort. To validate the gene-level changes observed in the Total Hippocampus Cohort, we utilized hippocampal single-nucleus transcriptomic (snRNA) data from external public cohorts and performed pseudobulk differential expression analysis (see “Methods” and Supplementary Fig. 9a–c). Within the ion channel category, the majority of DEGs were associated with potassium channels (Fig. 2d). Among these, only KCNQ1 and KCNH8 were consistently upregulated in MTLE across excitatory neurons, inhibitory neurons, and astrocytes (Supplementary Fig. 10). For neurotransmission-related genes, PTGS2 upregulation at the bulk level has been confirmed in both inhibitory neurons and astrocytes. While PTGS2 also exhibited a high effect size in excitatory neurons (log2 fold change = 2.97), this did not reach statistical significance, likely due to the limited sample size in the snRNA dataset (Supplementary Fig. 10a). In the neuroinflammation category, most MTLE-upregulated genes were associated with the positive regulation of neuroinflammation (GO:0050729), while MTLE-downregulated genes were linked to its negative regulation (GO:0050728) (Fig. 2e). The pro-inflammatory cytokine genes IL1B and OSM were elevated at the bulk level and consistently upregulated in MTLE across most cell types in single-cell data; however, this was not observed in microglia, as their baseline expression was already high in both MTLE and normal controls (Supplementary Fig. 10). Notably, SERPINE1 was the only gene in the neuroinflammation category to show significant upregulation across all cell types, including microglia. Regarding categories in the chronic phase, GO terms associated with the negative regulation of neurogenesis and positive regulation of gliogenesis were significantly upregulated in MTLE, whereas those associated with the positive regulation of axonogenesis and synaptic plasticity were downregulated, indicating impaired neural plasticity and accelerated gliogenesis in chronic MTLE (Fig. 2b). At the gene level, downregulation of SHANK3 further highlighted this loss of plasticity, although this finding was not replicated at the single-cell level (Fig. 2f and Supplementary Fig. 10). Most MTLE-upregulated genes in gliogenesis category were included in “positive regulation of gliogenesis” (GO:0014015), whereas MTLE-downregulated genes were included in “negative regulation of gliogenesis” (GO:0014014), consistent with the overall upregulation of gliogenesis-related GO terms. Notably, upregulation of EGR2 in MTLE was observed in both astrocytes and microglia with high effect sizes (log2 fold change > 4.0). Intracranial kainate and perforant path stimulation models best recapitulate human MTLE transcriptome We assessed whole-transcriptome similarity between human MTLE of the Total Hippocampus Cohort and rodent epilepsy models using one-to-one orthologs. The rodent models included in this study are well-documented for their phenotypic relevance to epilepsy, specifically the presentation of spontaneous recurrent seizures, and remain central to contemporary epilepsy research5. Among these models, the intrahippocampal kainate-induced (M-KAI-IH-IPSI-A-IM), intra-amygdala kainate-induced (M-KAI-IA-IPSI-A-IM), and perforant path stimulation-induced (M-PPS-A-DOFS) models demonstrated the highest correlation with human MTLE (Fig. 3a). Notably, the M-KAI-IH-IPSI-A-IM model shared the highest number of upregulated and downregulated genes with human MTLE. This suggests that the intrahippocampal kainate-induced model, utilizing adult mice with ipsilateral hippocampal tissue collected during the intermediate phase, closely replicates the transcriptomic landscape of the chronic drug-resistant human MTLE state. This trend was even more pronounced when focusing on epileptogenesis-related genes, which yielded higher correlation coefficients across most models than the whole-transcriptome analysis (Supplementary Fig. 11a). Cross-species transcriptomic correlation between human MTLE and rodent epilepsy models. a Correlation coefficients and the number of shared upregulated and downregulated genes between human MTLE and each rodent epilepsy model on whole-transcriptome. Correlation coefficients were calculated using Pearson’s correlation of log2 fold change vectors for shared one-to-one orthologs, comparing the differential expression signature of human MTLE against each rodent model. Correlation coefficients are presented as dots. Error bars represent 95% confidence intervals calculated via non-parametric bootstrapping. Statistical significance was determined using permutation test; models with non-significant results (p value ≥ 0.05) are highlighted in gray. The top annotation categorizes the 24 rodent epilepsy models based on the SIALAP classification (species, induction method, administration route, lateralization, age, and phase). b Correlation coefficients between human MTLE and rodent epilepsy models for nine epileptogenesis-related categories. Dot color represents the correlation coefficient, and dot size indicates statistical significance. MTLE mesial temporal lobe epilepsy. Among the chemical-induced epilepsy models, intracranial administration via intrahippocampal or intra-amygdala kainate injection yielded higher correlations with human MTLE than systemic administration of either kainate or pilocarpine (Fig. 3a). When analyzed according to the disease phase (hyperacute to chronic), the M-KAI-IH-IPSI-A models maintained the highest correlation with human MTLE across the acute, intermediate, and chronic phases, with the exception of the hyperacute phase (Supplementary Fig. 12). The intraperitoneal chemical injection models (M-PILO-IP-A-HA and M-KAI-IP-A-HA) showed stronger correlations than the intracranial chemical injection model (M-KAI-IH-IPSI-A-HA) during the hyperacute stage, while overall correlation levels in the hyperacute phase were lower than in subsequent phases. In electrical kindling models, the R-PPS-A-CR model exhibited a stronger correlation with human MTLE than the R-AMG-IPSI-A-CR model, exactly matched in all conditions except for induction and administration methods (Fig. 3a). Among the R-PPS-A models, samples collected on the “Day of First Seizure” phase showed the highest correlation, followed by those obtained in the intermediate and chronic phases. Cross-species correlation analysis across the nine epileptogenesis-related categories revealed distinctive transcriptomic similarities between human MTLE and specific rodent models (Fig. 3b). The intracranial chemical injection models, including M-KAI-IH-IPSI-A and M-KAI-IA-IPSI-A across the acute, intermediate, and chronic phases, exhibited stronger correlations within the ion channel category than the other models. In contrast, correlations in the synaptic plasticity category were notably higher in electrical kindling models, including R-PPS-A and R-AMG-IPSI-A across all phases, highlighting a divergence in the transcriptomic landscapes of chemical versus electrical induction methods. Correlations in neurogenesis and gliogenesis were consistently high across the M-KAI-IH-IPSI-A, M-KAI-IA-IPSI-A, and R-PPS-A models. We also examined cross-species correlations across different cell types. For cell deconvolution, snRNA and single-cell transcriptomic (scRNA) data from human and mouse hippocampi were used to generate cell type signatures, whereas scRNA data for the normal rat hippocampus was unavailable (see “Methods” and Supplementary Fig. 9). Once these signatures were established, we assessed the correlation between human and mouse cell types, confirming that each of the seven corresponding cell types was highly correlated (Supplementary Fig. 13). Notably, astrocytes in the intermediate and chronic phase of M-KAI-IH-IPSI-A and M-KAI-IA-IPSI-A models exhibited the strongest cross-species correlations compared to other cell types (Supplementary Fig. 11b, c). Although excitatory neuron correlations have been noted in some models, technical limitations in deconvolution preclude definitive conclusions. Neuroinflammatory and synaptic remodeling signatures characterize the two representative rodent models We performed a comparative analysis of the M-KAI-IH-IPSI-A and R-PPS-A models, which demonstrated the strongest transcriptomic similarity to human MTLE among the chemical-induced and electrical kindling models. ORA was first employed to assess model-specific expression changes. DEGs were aggregated across all four phases of each model, and genes unique to either the M-KAI-IH-IPSI-A or R-PPS-A models were identified. In the M-KAI-IH-IPSI-A model, the top 10 upregulated GO terms were primarily associated with neuroinflammation, neuronal death, and gliogenesis categories, whereas in the R-PPS-A model, the terms were mostly associated with ion channel and neurogenesis categories (Fig. 4a). When expanding the scope to the top 40 upregulated terms, the R-PPS-A model further exhibited significant enrichment in GO terms related to neurotransmission and synaptic plasticity, specifically “modulation of chemical synaptic transmission”, “regulation of trans-synaptic signaling”, and “regulation of synaptic plasticity” (Supplementary Data 9). The top 10 downregulated GO terms in the M-KAI-IH-IPSI-A model were mainly linked to neurotransmission and ion channels categories, whereas in the R-PPS-A model, they were associated with gliogenesis, neuronal death, and neurotransmission categories (Fig. 4b). Neurogenesis-related GO terms were present across the upregulated and downregulated genes in both models. When comparing human MTLE with these two models, the number of shared DEGs was considerably lower than that shared between rodent models; consequently, model-specific enrichment patterns were less discernible (Supplementary Fig. 14). Evaluation of the two representative rodent models at the gene-set and gene levels. a, b Overrepresentation analysis of uniquely a upregulated and b downregulated epileptogenesis-related genes in the M-KAI-IH-IPSI-A and R-PPS-A models. The top 10 GO terms for each model are displayed and categorized into nine epileptogenesis-related categories. c Connected dot plots illustrating the phase-specific trends of the two models using NESs from gene set enrichment analysis. The upper plots display neurotransmission and ion channel categories, whereas the lower plots represent neurogenesis and gliogenesis. The last group on the x axis in each plot represents NES in human MTLE. The color legend is provided next to the plots. d–f Forest plots comparing the log2 fold change of DEGs in human MTLE with their corresponding orthologs in the M-KAI-IH-IPSI-A-IM and R-PPS-A-DOFS models. Gene sets include d neurotransmission and ion channel, e positive and negative regulation of the inflammatory response (GO:0050728, GO:0050729), and f positive and negative regulation of neurogenesis (GO:0050768, GO:0050769) and gliogenesis (GO:0014014, GO:0014015), consistent with the human integrative analysis in Fig. 2d–f. Points are color-coded by group: human MTLE (brown), M-KAI-IH-IPSI-A-IM (blue), and R-PPS-A-DOFS (green). Points for rodent models are bordered with gray to denote statistical significance (p value 0.05), while all illustrated human MTLE genes are significant by default. MTLE mesial temporal lobe epilepsy, NES normalized enrichment score, GO gene ontology, DEG differentially expressed gene. Using GSEA, the enrichment trends across the four phases were assessed in the M-KAI-IH-IPSI-A and R-PPS-A models alongside human MTLE (Fig. 4c). In the M-KAI-IH-IPSI-A model, all GO terms related to neurotransmission and ion channels showed negative trends across all phases, mirroring the pattern observed in human MTLE, except for kainate receptor-related GO terms. In the neurogenesis and gliogenesis categories, GO terms associated with neuroblast proliferation and gliogenesis exhibited a positive trend after the hyperacute phase, whereas those related to axonogenesis and synaptic plasticity exhibited a negative trend. This pattern closely resembled human MTLE, with the exception of “negative regulation of axonogenesis.” In contrast, the R-PPS-A model displayed a positive trend in the chronic phase for all ion channel-related GO terms, as well as for ionotropic glutamate receptor-related GO terms, except for “NMDA Selective Glutamate Receptor Complex”. This trend was absent in the early phases but gradually emerged over time. However, broader GO terms, such as glutamatergic and GABAergic synaptic transmission, continued to show a negative trend. Additionally, the “regulation of short-term neuronal synaptic plasticity” GO term in the neurogenesis and gliogenesis categories was upregulated, in contrast to the patterns observed in human MTLE and the M-KAI-IH-IPSI-A model. These gene set-level findings from the ORA and GSEA indicate that the transcriptomic profile of the M-KAI-IH-IPSI-A model is characterized by a relative enrichment of neuroinflammatory and cell death-related signatures, whereas the R-PPS-A model shows a relative abundance of genes involved in ion channel activity and synaptic remodeling. For gene-level analysis, we examined the DEGs identified in human MTLE (Fig. 2d–f) across the following key categories: neurotransmission and ion channels, neuroinflammation (GO:0050728, GO:0050729), neurogenesis (GO:0050768, GO:0050769), and gliogenesis (GO:0014014, GO:0014015) (Fig. 4d–f). Analysis was focused on the intermediate phase of the M-KAI-IH-IPSI-A model and the “Day of First Seizure” phase of the R-PPS-A model, as these time points most closely replicated the human MTLE profile. We analyzed 59 shared orthologs for all three datasets among the 89 human MTLE DEGs. Within the neurotransmission and ion channel categories, the M-KAI-IH-IPSI-A-IM model exhibited an expression pattern that was more concordant with human MTLE than the R-PPS-A-DOFS model (Fig. 4d). Specifically, four human MTLE-upregulated potassium channel genes, including the snRNA-validated genes KCNQ1 and KCNH8, were consistently elevated across all phases of the M-KAI-IH-IPSI-A model, reinforcing the strong correlation between this model and human MTLE in the ion channel category (Supplementary Fig. 15). PTGS2, which was upregulated across all cell types in human MTLE, exhibited an initial increase in the early phases of both models, followed by a declining trend in the intermediate phase. Most genes in the neuroinflammation category were upregulated in both models, regardless of whether they were upregulated or downregulated in human MTLE, including OSM and IL1B, which remained elevated throughout most phases in both models (Fig. 4e and Supplementary Fig. 15). Notably, the snRNA-validated MTLE-upregulated gene SERPINE1 was consistently elevated across all phases in both models. Expression trends in the neurogenesis and gliogenesis categories generally showed closer alignment between human MTLE and rodent models, although model-specific differences were observed for certain key genes (Fig. 4f). For instance, EGR2, highly upregulated in astrocytes and microglia of human MTLE, exhibited sustained elevation across all phases in the R-PPS-A model, yet showed an initial increase followed by a decrease in the intermediate phase of the M-KAI-IH-IPSI-A model (Supplementary Fig. 15). Additionally, although SHANK3 was downregulated in human MTLE, no corresponding downregulation was observed in either rodent model. Hippocampal sclerosis-specific transcriptomic fidelity of the intrahippocampal kainate model We investigated the correlation between human MTLE subtypes and the two representative models, M-KAI-IH-IPSI-A-IM and R-PPS-A-DOFS models, to assess differences based on the presence of HS and history of initial precipitating injury (IPI). Clinical data on HS status were available across all datasets, whereas IPI history was available only for three datasets (Fig. 5a and Supplementary Data 3). IPI events were categorized as “seizure-related event,” “other event,” and “no event.” The M-KAI-IH-IPSI-A-IM model exhibited the strongest correlation with MTLE with HS, whereas its correlation with MTLE without HS was markedly low (Fig. 5b and Supplementary Fig. 16). In contrast, the R-PPS-A-DOFS model showed a comparable correlation with both MTLE with and without HS, suggesting a more generalized transcriptomic similarity. Regarding IPI history, the M-KAI-IH-IPSI-A-IM model demonstrated a low correlation with MTLE cases linked to seizure-related events, such as febrile seizures, or without IPI history. However, it showed a high correlation with MTLE cases associated with non-seizure-related events, such as head trauma. Compared to the M-KAI-IH-IPSI-A-IM model, the R-PPS-A-DOFS model maintained a moderate correlation even in MTLE cases with seizure-related IPI or without IPI history. However, these subtype-specific results should be interpreted with caution because a limited number of samples with IPI information were available. Cross-species correlation with human MTLE subtypes and transcriptomic profiles of the two representative rodent models. a Sankey plot illustrating the proportion of MTLE subtypes, categorized by the presence or absence of HS and history of IPI. b Correlation coefficients between human MTLE subtypes and the two most representative rodent models: M-KAI-IH-IPSI-A-IM and R-PPS-A-DOFS. Dot color represents the correlation coefficient, and dot size indicates statistical significance using permutation test. c Schematic representation of the model-specific transcriptomic profiles of the intrahippocampal kainate-induced and perforant path stimulation-induced models. The unique pathophysiological features associated with each transcriptomic profile, including neuronal loss, gliosis, and synaptic remodeling, contribute to MTLE-like hyperexcitability. Biological illustrations in c were created in BioRender. Kim, S. (2026) https://BioRender.com/2uqmryb. MTLE mesial temporal lobe epilepsy, HS hippocampal sclerosis, IPI initial precipitating injury. Discussion Our comprehensive transcriptomic analysis across human epilepsy and corresponding rodent models provides valuable insights into the epileptogenesis cascade in MTLE and its animal models. Additionally, this study evaluated the degree of similarity between rodent epilepsy models and human MTLE, offering a comparative framework for translational research. To achieve a robust human integrative analysis and cross-species comparison, we implemented a unified RNA processing pipeline with batch correction across all datasets, consisting of 694 human samples (12 datasets) and 362 rodent samples (13 datasets). Furthermore, we curated gene sets for nine epileptogenesis-related categories using GO terms, allowing a targeted investigation of epilepsy-associated transcriptomic changes. The transcriptomic landscape of nine epileptogenesis-related categories was first examined in drug-resistant chronic MTLE before evaluating rodent epilepsy models. In the neurotransmission and ion channel categories, most GO terms exhibited a negative trend in MTLE, likely reflecting extensive neuronal loss, particularly in MTLE with HS3. An exception was the kainate receptor GO term, which showed a slight positive trend driven by the increased expression of GRIK4 (log2 fold change = 0.568), one of the five kainate receptor subunit genes. Notably, GRIK4, which has a higher glutamate affinity than the other subunits, was also upregulated in single-cell transcriptomic studies of MTLE12,13. While these findings suggest a potential role for GRIK4 in drug-resistant MTLE, further investigation is needed, as most epilepsy studies have not specifically focused on kainate receptors. Our human integrative analysis revealed that PTGS2 was upregulated in the bulk and single-cell levels, which aligns with previous histological analyses confirming its elevation in the human epileptic hippocampus14. As the enzyme responsible for catalyzing the conversion of arachidonic acid to prostaglandins, PTGS2 likely promotes neuronal excitability through prostaglandin E2-mediated modulation of ion channels while simultaneously driving neuroinflammation via inflammatory prostaglandin signaling15,16. Its upregulation in the acute phase of the intrahippocampal kainate-induced and perforant path stimulation-induced models further supports their role in epileptogenesis. Given the decreased expression of a majority of receptors and ion channels in MTLE, the elevation of these specific genes across key cell types may be critical for sustaining hyperexcitability. Furthermore, four potassium channel genes upregulated in MTLE, including KCNQ1, KCNH8, KCNE5, and KCNK5, also showed sustained elevation across all phases in the intrahippocampal kainate-induced model. Considering the role of potassium channels in mediating hyperpolarization, their association with epileptogenesis requires further validation to determine whether this reflects a failed compensatory response or a contributor to maladaptive circuit remodeling17. Neuroinflammation was significantly increased in human MTLE at the gene-set and individual gene levels. However, the trends in neuronal death and TNF signaling were the opposite, as negative regulation showed a higher enrichment score than positive regulation. While neuronal death and TNF signaling play key roles in epileptogenesis5, their influence likely diminishes as epilepsy progresses to a chronic drug-resistant state. At the gene level, IL1B, encoding the pro-inflammatory cytokine IL-1β, has been extensively linked to MTLE in multidimensional studies18,19. Our integrative analysis confirmed this association, showing significant elevation in the bulk and single-cell levels. Similarly, OSM, a member of the leukemia inhibitory factor/oncostatin-M (LIF/OSM) cytokine family, promotes neuroinflammation by regulating cytokine production (e.g., IL-6)20. While the connection between OSM and epilepsy has previously been limited to pilocarpine-induced rodent models21, the consistent upregulation of both IL1B and OSM in intrahippocampal kainate-induced and perforant path stimulation-induced models suggests that these cytokines are fundamental drivers of sustained neuroinflammation throughout epileptogenesis. Beyond pro-inflammatory cytokines, SERPINE1, encoding an inhibitor of the tissue plasminogen activator (tPA)/plasmin system, was identified as a ubiquitous neuroinflammatory marker across all cell types in MTLE, as well as in intrahippocampal kainate-induced and perforant path stimulation-induced models22. tPA promotes hyperexcitability not only through its proteolytic activity but also by physically binding to and upregulating NR2B-containing NMDA receptors, thereby lowering the seizure threshold23. The widespread elevation of SERPINE1 likely represents a compensatory attempt to neutralize this excessive tPA activity and protect the brain from further protease-mediated insults24. Consequently, SERPINE1 may serve as a critical regulator of excitotoxic neuronal death in MTLE. We identified suppressed neurogenesis, deficient neural plasticity, and excessive gliogenesis in chronic MTLE, which align with the findings of recent immunofluorescence studies showing a dramatic decline in neuronal production and persistent astrogenesis in chronic MTLE25. The expression of SHANK3, a key regulator of synapse formation and dendritic spine maturation, was downregulated in MTLE. This finding contrasts with that of a previous study reporting elevated SHANK3 expression in the temporal cortex of temporal lobe epilepsy26. Given that SHANK3 deficiency is primarily linked to seizure development and epilepsy, its dysregulation may contribute to persistent seizure activity through impaired synaptic plasticity27,28. Additionally, a key driver of this pathological trend appears to be EGR2, which is markedly upregulated in both astrocytes and microglia with high effect sizes. While EGR2 is canonically recognized for its role in peripheral nerve myelination and hindbrain development, it is also known to modulate macrophage and microglial activation during chronic neuroinflammation29,30. Its dramatic elevation in astrocytes and microglia, combined with the significant upregulation of “positive regulation of gliogenesis” pathways, suggests that EGR2 may act as a transcription factor driving the reactive gliosis and persistent astrogenesis that characterize the chronic epileptic hippocampus. Further, we evaluated how well rodent models replicated the human epileptic transcriptome and how they manifested epileptic conditions. Among the chemical-induced epilepsy models, the intrahippocampal kainate-induced model significantly outperformed the systemic injection methods, closely mimicking the MTLE transcriptomic profile across the majority of epileptogenesis categories. Interestingly, this similarity increased over time, peaking during the intermediate phase (4–10 days post-treatment). This suggests that the intermediate phase may represent the optimal sampling window for capturing the transcriptomic environment of chronic drug-resistant MTLE. Meanwhile, the contralateral hippocampus, despite its exposure to recurrent seizures, failed to replicate the human MTLE environment as effectively as the ipsilateral focus. The predominant transcriptomic profile was characterized by neuroinflammation, neuronal death, and gliogenesis (Fig. 5c). Its higher correlation with MTLE with HS compared to MTLE without HS suggests that this profile likely contributes to the histopathological environment of HS. This pattern aligns with previous evidence that neuronal apoptosis plays a crucial role in the development of the kainate model31. Additionally, neuronal cell death followed by neuroinflammation has been shown to be sufficient to trigger spontaneous recurrent seizures32, reinforcing the relevance of this model in mimicking the pathophysiology of MTLE with HS. We additionally generated RNA-seq data for the intra-amygdala kainate-induced model. Although previous microarray studies suggested similarities between this model and human MTLE33, our analysis provides a more comprehensive transcriptomic comparison across multiple models. This model demonstrated performance comparable to that of the intrahippocampal kainate-induced model, specifically by reaching peak similarity to MTLE during the intermediate phase. Furthermore, the correlation patterns across the nine epileptogenesis categories closely mirrored those of the intrahippocampal model. These findings reconfirm the utility of the model in replicating the MTLE transcriptomic environment, as its effectiveness in producing high seizure frequencies makes it a compelling option for epilepsy modeling34. However, the higher mortality rate during immediate status epilepticus following intra-amygdala injection remains a significant practical challenge compared to the intrahippocampal approach. In the intraperitoneal pilocarpine-induced model, the correlation with human MTLE at the hyperacute phase was highest compared to other models within this phase. Moreover, the transcriptomic profile reached its peak correlation with human MTLE during the acute phase. This is consistent with the observation that pilocarpine induces more rapid neuronal damage than kainate35. Although these results suggest an earlier optimal sampling window for the pilocarpine model, this finding requires further validation under strictly controlled, head-to-head administration conditions to account for the potential variability between systemic and local induction. The perforant path stimulation model, which directly stimulates the hippocampus via the perforant path, demonstrated stronger performance than the amygdala stimulation model. Its unique transcriptomic profile was characterized by neurotransmission, ion channel activity, and synaptic plasticity compared to the intrahippocampal kainate-induced model (Fig. 5c). Over time, these distinct transcriptomic features became more pronounced, peaking during the chronic phase. This explains why the “Day of First Seizure” phase exhibited the highest correlation with human MTLE, whereas the correlation declined in the intermediate and chronic phases as expression patterns diverged from those of human MTLE. Notably, this model effectively recapitulated MTLE without HS, as well as MTLE with a seizure-related initial event, such as febrile seizures, making it a more suitable model for these subtypes than the intrahippocampal kainate-induced model. While neuroinflammation and gliogenesis remain integral processes to this model, the transcriptomic profile in this model was relatively more characterized by synaptic remodeling, representing a clear transcriptomic divergence from the intrahippocampal kainate-induced model. This finding aligns with those of previous studies, indicating that synaptic remodeling is the key mechanism underlying the long-term increase in seizure susceptibility following electrical kindling36. This study has some limitations. First, all human samples were derived from patients with drug-resistant chronic MTLE, limiting the evaluation of early epileptogenesis. However, chronic MTLE represents the clinical reality where drug resistance in epilepsy remains a therapeutic challenge, necessitating the development of therapeutic strategies. Therefore, assessing how well rodent models reflect this chronic epileptic condition remains a critical aspect of translational research. Second, while this study centers on the hippocampus as the primary epileptogenic focus, it does not account for other regions such as the temporal cortex. Future investigations incorporating the temporal cortex could provide complementary insights into the broader network-level mechanisms of the disease. Third, although we developed the SIALAP classification to provide detailed model specifications and minimize technical variability, certain variables, such as species and age, were not perfectly matched due to the inherent heterogeneity of multiple independent studies. Specifically, within the “Adult” classification (7–12 weeks), we could not fully evaluate the potential influence of age-related transcriptomic variations that may exist across that age range. Finally, while snRNA-seq provided cell-type-specific validation, its lower sensitivity for sparsely expressed genes and potential omission of rare cell types are recognized technical caveats. Therefore, these snRNA-seq-validated DEGs should be further verified using complementary modalities, such as immunohistochemistry, to confirm their biological significance at the protein level. In conclusion, our human integrative analysis identified DEGs associated with drug resistance in chronic MTLE that were not detected in individual studies. Additionally, we developed a cross-species evaluation framework to identify rodent epilepsy models that best replicate the human epileptic conditions. Our analysis revealed distinct transcriptomic signatures in the two representative models: intrahippocampal kainate-induced and perforant path stimulation-induced models. Our findings provide a transcriptomic basis for optimizing epilepsy rodent models and developing therapeutic strategies for drug-resistant MTLE. Methods Ethics statement This study was conducted in accordance with all relevant ethical regulations. Human brain tissue samples were obtained from publicly available repositories (SRA and Gene Expression Omnibus (GEO)), and all original studies received appropriate institutional ethics approval. All animal procedures for the intra-amygdala kainate-induced epilepsy model were approved by the Institutional Animal Care and Use Committee (IACUC) at the Yonsei University Health System (No. IACUC 2025-0326) and conducted in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals. Human samples Bulk whole-transcriptome RNA-seq data from brain tissues obtained from human patients with MTLE were procured from the SRA using the following search criteria in the GEO database: Homo sapiens (organism), temporal lobe epilepsy (keywords), and high-throughput sequencing (technology type) (Supplementary Fig. 1 and Supplementary Data 2). Microarray data were excluded because of their incompatibility with our integration approach. Nine hippocampal RNA-seq datasets from MTLE patients and normal controls14,37–44, along with three additional datasets from normal controls45–47, were downloaded as FASTQ files, totaling 694 samples with a mean age of 39.6 years and 40.3% female. During data processing, 3 and 10 samples were excluded owing to SRA download failure and the presence of fetal samples in PRJNA322318, respectively (Supplementary Fig. 3). Duplicate sequencing runs for 57 patients (n = 141 samples from PRJNA290212, PRJNA373909, PRJNA753504, and PRJNA773413) were merged at the FASTQ level. In instances where merged duplicates contained data from the dentate gyrus and cornu ammonis, the resulting aggregate was analyzed as a single whole-tissue hippocampal sample. Additionally, eight samples failed during alignment and quantification, leaving 589 samples for batch correction (MTLE, n = 161; normal controls, n = 179; focal cortical dysplasia, n = 50; tuberous sclerosis complex, n = 19; schizophrenia, n = 74; major depressive disorder, n = 29; bipolar disorder, n = 65; and alcohol use disorder, n = 12). After batch correction, only hippocampal samples from patients with MTLE and normal controls were retained for further analysis, comprising 11 datasets (excluding PRJNA787219) and 253 samples with a mean age of 42.9 years and 37.5% female (Fig. 1a and Supplementary Data 3). All MTLE samples were derived from drug-resistant patients who underwent surgical hippocampal resection, comprising 108 with HS and 18 without HS, with all tissues obtained from the ipsilateral side of the epileptic focus. In contrast, 127 hippocampal tissues from normal controls were obtained from post-mortem neurotypical individuals. Among these 253 samples, data on IPI were available for 35 samples in PRJNA290212, PRJNA373909, and PRJNA525671 (Supplementary Data 3). IPIs were manually categorized into three groups (“seizure-related event,” “other event,” and “no event”) using data extracted from the main tables and supplementary tables of each dataset. For PRJNA290212, patients with a history of febrile seizures were classified under “seizure-related event,” those with a history of head trauma under “other event,” and those without a documented medical history prior to epilepsy under “no event.” In PRJNA373909, we categorized patients with febrile seizures, generalized febrile seizures, or tonic-clonic postictal confusion under “seizure-related event,” those with a history of head trauma under “other event,” and the remaining patients, labeled as “ND” in the dataset, under “no event.” Since the dataset did not specify the meaning of “ND,” we presumed it indicated the absence of prior injury. In PRJNA525671, patients with a history of febrile seizures were classified as having a “seizure-related event,” those with stroke, head trauma, or meningitis under “other event,” and the remaining patients labeled as “Unknown” under “no event.” Additionally, one patient with “drug overdose” listed as the etiology was classified under “no event,” as this was not considered an appropriate IPI for MTLE. Data on the number of anti-seizure medication usage were available for 25 samples in PRJNA290212, PRJNA280563, and PRJNA373909 (Supplementary Data 3). Rodent samples Bulk whole-transcriptome RNA-seq data from the hippocampi of rodent models were obtained from the SRA using specific search criteria in the GEO database. The search included Mus musculus or Rattus norvegicus (organism); keywords such as epilepsy, pilocarpine, or kainic acid; and high-throughput sequencing (technology type). Seven datasets for mice7,48–53 and five datasets for rats54–58 were downloaded, totaling 588 FASTQ files across 326 samples (254 mice, 72 rats). For 88 samples from PRJNA564028, PRJNA880828, and PRJNA1006742, which had multiple FASTQ files per sample, we merged the files per sample as described in the original studies. All 326 samples, including epilepsy models and controls, were successfully processed for SRA download, RNA alignment, and quantification without errors. As RNA-seq data for the intra-amygdala kainate-induced epilepsy model were not publicly available, we generated these data in-house for comprehensive analysis. By incorporating these in-house datasets, we compiled a total of 13 rodent datasets encompassing 290 Mus musculus samples (C57BL/6 J, n = 109; C57BL/6 N, n = 24; C57BL/6 unspecified substrain, n = 157) and 72 Rattus norvegicus samples (Sprague-Dawley, n = 54; Wistar, n = 18), as detailed in Supplementary Data 4. One rat dataset involved rodents at postnatal week 1, while the remaining datasets utilized adult rodents ranging from 7 to 12 weeks of age. Most datasets consisted of male subjects only (n = 263), while two datasets included both males and females (n = 67; sex information at the individual level not specified by the original studies), and two datasets did not specify sex information (n = 32). Sex was not considered as a variable in the cross-species transcriptomic analysis due to the predominance of male-only datasets and the absence of individual-level sex information in a subset of datasets. SIALAP classification for rodent epilepsy models To systematically classify rodent epilepsy models, we developed the SIALAP classification, which categorizes models based on species, induction method, administration route, lateralization, age, and phase (Fig. 1a). The species included Mus musculus (mouse) and Rattus norvegicus (rat). Epilepsy was induced through three primary methods: (1) chemical induction using kainate or pilocarpine, (2) electrical kindling via perforant path or amygdala stimulation, and (3) TBI with lateral fluid percussion injury. Chemicals were administered via four routes: intrahippocampal, intra-amygdala, intraperitoneal, or subcutaneous injection. Lateralization in the SIALAP classification indicates the sampling location relative to the treatment site and is categorized as ipsilateral or contralateral. However, this does not apply to systemic treatments, such as intraperitoneal or subcutaneous injections, or the perforant path stimulation model, which involves bilateral stimulation. Rodent age was classified into two groups: “Postnatal” (1–2 weeks old) and “Adult” (7–12 weeks old). In datasets PRJNA215950, PRJNA304117, and PRJNA565693, the rodent age was unspecified but inferred to be “Adult” based on reported weight. The timing of sample collection relative to treatment was categorized into five phases: “Hyperacute” (within 24 hours after treatment), “Acute” (1–3 days after treatment), “Intermediate” (4–10 days after treatment), “Chronic” (more than 10 days after treatment), and “Day of First Seizure” (within 24 hours after first spontaneous seizure). The “Day of First Seizure” phase was applicable only to the perforant path model56. In total, 24 distinct epilepsy models were categorized using the SIALAP classification (Supplementary Data 4 and Supplementary Fig. 5). Each dataset included the corresponding control samples for differential expression analysis. Intra-amygdala kainate-induced epilepsy model To create the intra-amygdala kainate-induced epilepsy model, 8- to 10-week-old male C57BL/6 J mice (Orientbio, Gyeonggi, Korea) were used. Status epilepticus was triggered via a unilateral microinjection of kainic acid (0.3 μg in 0.15 μL phosphate-buffered saline [PBS]; Sigma-Aldrich, St. Louis, MO, USA) into the basolateral amygdala. Stereotaxic coordinates were defined as AP −1.31 mm, ML −3.0 mm, and DV −4.5 mm relative to Bregma. Injections were performed under isoflurane anesthesia, while control animals received an equivalent volume of vehicle (PBS). Successful seizure induction was confirmed, as all kainate-treated mice exhibited convulsive status epilepticus immediately after surgery. To manage seizure duration and reduce mortality, diazepam (10 mg/kg; Samjin Pharm, Seoul, Korea) was administered intraperitoneally 40 minutes after kainate or PBS injection. All mice in the chronic phase of the intra-amygdala kainate model exhibited spontaneous recurrent seizures ranging from Racine scale 3–5, as confirmed by 24-hour video monitoring. Hippocampus sampling and RNA sequencing Ipsilateral hippocampal tissues of the intra-amygdala kainate-induced model were harvested at 3, 7, and 14 days following the initial injection, corresponding to the acute, intermediate, and chronic phases. Six mice per group were used, totaling 36 male mice across all phases and control groups. Samples were preserved in RNAlater prior to total RNA extraction. The resulting RNA demonstrated high integrity, with RNA Integrity Number (RIN) values ranging between 8.0 and 9.5. Libraries were constructed using the TruSeq Stranded mRNA Sample Preparation Kit (Illumina, San Diego, CA, USA), following a poly(A) selection and strand-specific protocol. High-throughput sequencing was performed on the Illumina NovaSeq X Plus platform, generating 100 bp paired-end reads. RNA trimming and alignment The sequencing read layout (single-end or paired-end) and average read length varied across studies (Supplementary Data 2 and 4). To ensure data quality, FASTQ files were filtered and trimmed using FASTP (version 0.21.0), with parameters adjusted according to each read layout and length59. Reads shorter than 25 base pairs for single-end sequencing and 50 base pairs for paired-end sequencing were removed. Low-complexity reads with repetitive sequence patterns were filtered using the default “low_complexity_filter” option. Additionally, adapter sequences and overrepresented sequences were removed for paired-end reads. Filtered reads were aligned to the latest genome references—GRCh38.p14 for human, GRCm39 for mouse, and mRatBN7.2 for rat. Alignment was performed using STAR (version 2.7.10)60. Genome indices were generated separately for different read layouts (single-end and paired-end) and read lengths (35, 50, 75, 100, 125, and 150 bp). The aligned output was stored as BAM files, with annotation based on Ensembl release 112 GTF files for each species. RNA quantification and normalization Transcript quantification was performed using HTSeq (version 0.12.4) in intersection-strict mode61. For normalization, TPM values were calculated using GeoTcgaData (version 1.99.2)62. TPM values were used for expression comparison using Student’s t-test and as input for PCA clustering, GSEA, and the deconvolution process. Batch correction To mitigate batch effects across different cohorts, Combat-seq from the sva package (version 3.46.0) was applied63. A total of 589 human samples obtained from 12 datasets were included in the batch correction process. The dataset cohort variable was used to account for batch effects, while diagnosis was included as a biological covariate to preserve true biological signals. Eight diagnostic groups were considered, including MTLE, normal control, focal cortical dysplasia, tuberous sclerosis complex, schizophrenia, major depressive disorder, bipolar disorder, and alcohol use disorder. In contrast to the human MTLE samples distributed across multiple cohorts, the exclusivity of most rodent models to individual datasets rendered the use of the dataset cohort as a batch variable inappropriate, as it risked obscuring the true biological signals. Therefore, batch correction for rodent samples was performed using the sequencing platform as the batch variable to account for technical variability introduced by different sequencing methods. The platforms included GenomeAnalyzerIIx, GenomeAnalyzerII, NextSeq500, HiSeq1500, HiSeq2500, HiSeq4000, NovaSeq6000, and NovaSeq X Plus. Batch correction was applied separately to 290 mouse samples across eight datasets and 72 rat samples across five datasets. PCA clustering PCA clustering was conducted on both raw and TPM counts, before and after batch correction, using the prcomp function (from the stats R package, version 4.2.0). The assessment of the batch correction efficacy was conducted independently for human, mouse, and rat datasets to confirm the mitigation of platform-specific technical noise. For human outlier filtering process, TPM count was employed for PCA clustering. Human sample quality control and cohort stratification Outlier filtering was performed on 253 human samples to enhance data accuracy for cross-species evaluation (Supplementary Fig. 3). Two selection criteria were applied. First, PCA on TPM count with epileptogenesis-related genes was used; because the first principal component (PC1) did not effectively distinguish between clinical groups, the second principal component (PC2) was utilized to separate MTLE and normal samples (Supplementary Fig. 7a, b). Samples were generally clustered along the PC2 = 0 axis, with normal samples positioned below zero and MTLE samples above zero. However, a subset of samples deviated from this expected distribution. The second criterion involved TPM expression levels of GFAP, AQP4, and NOTCH1, which are key genes associated with gliosis and neuroinflammation in MTLE (Supplementary Fig. 7c, d). MTLE samples were considered outliers if the z-score of GFAP, AQP4, or NOTCH1 TPM expression was below −2.0, suggesting an absence of hallmark disease features. Conversely, normal samples were considered outliers if the z-score exceeded 2.0, indicating abnormally high gliosis or neuroinflammation. Samples meeting both outlier criteria were excluded, resulting in the removal of ten normal control samples (Supplementary Fig. 7e, f). The final Total Hippocampus Cohort included 243 human samples, which were used for the human integrative analysis of MTLE and cross-species comparison. Additionally, we established a Whole-Tissue Only Cohort (n = 143) to verify that the inclusion of subregion-specific samples within the Total Hippocampus Cohort did not introduce significant bias into the cross-species analysis (Supplementary Fig. 3). Epileptogenesis-related gene sets GO terms relevant to the acute, subacute, and chronic phases of the epileptogenic cascade were selected for downstream analysis (Supplementary Fig. 2 and Supplementary Data 1). GO terms associated with the acute phase included neurotransmission (GO:0051932, GO:0035249, GO:0008328, GO:0032281, GO:0017146, GO:0032983) and ion channel (GO:0005245, GO:0005249, GO:0005248). For the subacute phase, neuroinflammation (GO:0006954, GO:0150076, GO:0048143, GO:0001774, GO:0050729, GO:0050728), neuronal death (GO:0051402, GO:0110088, GO:0043525, GO:0043524), and TNF signaling (GO:0033209, GO:1903265, GO:0010804) were included. The chronic phase gene sets encompassed neurogenesis (GO:0050769, GO:0002052, GO:0050772, GO:0050768, GO:0007406, GO:0050771), gliogenesis (GO:0042063, GO:0014015, GO:0014014, GO:0048708, GO:0014004), mossy fiber sprouting (GO:0048668, GO:0098686), and synaptic plasticity (GO:0048168, GO:0048169, GO:0048172). A total of 1839 genes were included across nine epileptogenesis-related categories: neurotransmission, ion channel activity, neuroinflammation, neuronal death, TNF signaling, neurogenesis, gliogenesis, mossy fiber sprouting, and synaptic plasticity. Orthologous genes Orthologous genes shared between humans and rodents were identified using Ensembl BioMart (https://www.ensembl.org/biomart/martview/). To ensure accurate cross-species comparisons, only one-to-one orthologs were selected, while one-to-many and many-to-many orthologs were excluded. A total of 17,128 orthologous genes out of 26,360 genes (65.0%) were identified between humans and mice, while 16,211 orthologous genes out of 24,250 genes (66.8%) were identified between humans and rats. snRNA-seq data processing and cell type annotation Human snRNA-seq datasets from MTLE and normal hippocampal tissues64,65 were integrated to serve a dual purpose: validating DEGs identified at the bulk level and providing a reference framework for bulk RNA deconvolution. Human snRNA-seq data were obtained as processed matrix data (counts, genes, and metadata) from the GEO database. The MTLE dataset was obtained as a pre-processed object containing 131,325 cells and 17,180 features. This dataset had previously undergone quality control based on a gene count >300, transcript count <10,000 per cell, and mitochondrial gene ratio <5%. For the normal dataset, a total of 200,481 cells were selected from an initial 219,058 cells with 33,939 features, following the criteria including gene count between 200 and 5000 per cell, mitochondrial gene ratio below 10%, and inclusion of features expressed in at least 1% of cells. The two datasets were merged using Reciprocal PCA integration within Seurat (Supplementary Fig. 9a). Cells were subsequently annotated into seven major types: excitatory neurons, inhibitory neurons, astrocytes, microglia, oligodendrocytes, oligodendrocyte precursor cells (OPCs), and endothelial cells (Supplementary Fig. 9b and Supplementary Data 10). Immune cells and Cajal–Retzius cells were excluded, as they were not the focus of this study. All snRNA-seq data analyses were performed using the Seurat package (version 5.1.0)66. For the rodent study, we utilized a normal mouse hippocampal scRNA-seq dataset; however, no publicly available scRNA-seq data were identified for normal rat hippocampal tissue at the time of analysis. Mouse 10X scRNA-seq data were downloaded as an hdf5 file from the Allen Brain Map (https://portal.brain-map.org/)67. The data were pre-processed by the original authors to filter cells with fewer than 1500 detected genes, remove predicted doublets, and exclude low-quality or contaminated noise clusters. From the initial dataset of over one million cells, a subset of 82,535 hippocampal cells (7.1% of the total data) with 31,053 features was selected. Cells were categorized into the same seven major cell types used in the human analysis (Supplementary Data 10). While most annotations followed the original metadata, the “Oligo” cluster was manually sub-partitioned into mature oligodendrocytes and OPCs based on the expression of established lineage markers (Supplementary Fig. 9e). All mouse data analyses were conducted using Scanpy (v1.11.5) packages68. Expression matrix for cell-type signature generation To construct reference expression matrices for deconvolution, cell-type-specific marker genes were identified for each species. For the human dataset, markers were identified using Seurat’s FindAllMarkers function with a Wilcoxon Rank Sum test. The final signature matrix comprised the union of the top 200 DEGs per cell type, ensuring that the key markers for every major hippocampal population were included. Similarly, for the mouse dataset, cell-type-specific markers were identified using Scanpy’s rank_genes_groups function with a Wilcoxon Rank Sum test, following the same criteria for gene selection. Cell type signature extraction and deconvolution Cell type signatures and bulk RNA expression matrices for each cell type were generated using CIBERSORTx (https://cibersortx.stanford.edu/)69. The deconvolution process utilized TPM values from bulk RNA-seq data as the input mixture. First, cell-type-specific signatures were constructed for both human and mouse datasets using the CIBERSORTxFractions module (Supplementary Data 11). This utilized the snRNA-seq and scRNA-seq-derived expression matrix consisting of the union of the top 200 DEGs per cell type as a reference. Subsequently, we estimated the cell-type-specific gene expression profiles from the bulk data using CIBERSORTxHiRes. Because the HiRes module imposes a technical limit of 1000 genes per run, we implemented an iterative approach to cover all genes within our predefined epileptogenesis-related categories (Supplementary Data 1). This process was performed separately for each GO term across all cell types. Following these iterations, the results were merged into a comprehensive expression matrix. For genes appearing in multiple GO terms, expression values were collapsed by calculating the mean expression for each GO term per sample. Differential expression analysis Differential expression analysis was performed using DESeq2 (version 1.38.3)70. Batch-corrected raw counts were used to compute log2 fold change and adjusted p values. In humans, MTLE and its subtypes, classified based on HS and IPIs, were compared with normal controls. In rodents, each of the 24 epilepsy models was analyzed against its respective control group. Genes were considered significantly upregulated or downregulated if they met the following criteria: |log2 fold change| ≥ 1.0 and adjusted p value < 0.05. Pseudobulk differential expression analysis To validate cell-type-specific changes using integrated human snRNA-seq data, we performed pseudobulk differential expression analysis using the DESeq2 package (version 1.38.3). For each of the seven major cell types, raw counts were aggregated by summing the gene expression across all cells belonging to the same biological sample. Prior to the analysis, low-abundance genes with a total aggregate count of 10 or fewer were excluded. Significant DEGs were identified based on an adjusted p value < 0.05. Overrepresentation analysis ORA was performed using ToppFun (https://toppgene.cchmc.org/)71. Genes were considered upregulated or downregulated if they met the criteria of |log2 fold change| ≥ 1.0 and adjusted p value < 0.05. Statistical significance was assessed using false discovery rate-corrected p values following the Benjamini–Hochberg procedure. GO terms included in the “biological process” category were considered, whereas those with >1000 total genes were excluded. For rodent models, the analysis was restricted to genes within the epileptogenesis-related gene sets. The top 10 most significant GO terms were annotated based on their inclusion in nine epileptogenesis-related categories, determined by counting the number of associated genes within each category. Gene set enrichment analysis GSEA was conducted using GSEA software (version 4.3.3, https://www.gsea-msigdb.org/gsea/)72,73. Genes were ranked based on their TPM expression values, and enrichment analysis was performed using 1000 permutations with nine epileptogenesis-related gene sets (Supplementary Fig. 2 and Supplementary Data 1). The Ensembl gene ID chip version 2024 was used for each species, as provided by the GSEA software. Statistical significance was determined using nominal p values, as only selected GO terms of interest were investigated, rather than performing multiple hypothesis testing across several GO terms. For human samples, combined p values were calculated separately for the nine epileptogenesis-related categories using Fisher’s method, and upregulated and downregulated GO terms were analyzed independently within each category. Cross-species transcriptomic correlation analysis To evaluate the transcriptomic similarity between human MTLE and rodent epilepsy models, we performed gene-level Pearson correlation analysis. This approach compared differential expression signatures (model-level) rather than individual sample-to-sample expression. The input for each correlation was a vector of log2 fold change values derived from the respective DESeq2 analyses for all shared orthologs. Specifically, the log2 fold change vector of a human group (e.g., MTLE vs. control) was correlated with the log2 fold change vector of every epilepsy model group (e.g., M-KAI-IH-IPSI-A-IM vs. Control). To ensure the observed correlations represented global transcriptomic shifts rather than the influence of a few high-fold-change marker genes, we implemented two robust statistical measures using 1000 iterations each. First, non-parametric bootstrapping was used to calculate 95% confidence intervals (CIs) for correlation coefficients by resampling gene pairs with replacement. Second, permutation testing was conducted by randomly shuffling gene identities to generate an empirical null distribution. The p value was calculated by comparing the observed correlation coefficients with the empirical null distribution. All analyses were performed in R (version 4.2.0). This correlation framework was applied across three distinct scales. Initially, the analysis was performed on both the full set of one-to-one orthologous genes and a subset of genes within the nine epileptogenesis-related categories. Subsequently, correlation analysis was conducted for seven major cell types, including excitatory neurons, inhibitory neurons, astrocytes, microglia, oligodendrocytes, OPCs, and endothelial cells, using orthologous genes identified through the CIBERSORTxHiRes process. This cell-type-specific analysis was applied only to mouse models, as no corresponding cell-type-specific expression matrix was available for rats. Finally, we performed correlation analyses between specific MTLE subtypes, classified as HS or IPIs, and the most representative rodent models using the entire orthologous gene set. Statistics and reproducibility No statistical method was used to predetermine sample size. Instead, all publicly available RNA-seq datasets meeting the predefined selection criteria were incorporated to maximize statistical power. Data exclusions were applied during data processing and quality control steps, as described in “Human samples” and “Human sample quality control and cohort stratification” Methods subsections. The experiments were not randomized. The investigators were not blinded to allocation during experiments and outcome assessment. Statistical analyses, including differential expression analysis, gene set enrichment analysis, and cross-species correlation were performed as described in the respective Methods subsections. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data Reporting Summary Transparent Peer Review file Acknowledgements We would like to thank Editage (www.editage.co.kr) for English language editing. Author contributions J.K. designed the study. S.K. and W.J.K. supervised the project and secured funding. J.K. and B.K. produced the transcriptomic data for the intra-amygdala kainate model. J.K. and S.L. collected and curated transcriptomic datasets from human and rodent samples. J.K. processed raw RNA-seq data, performed alignment, quantification, and batch correction. J.K. conducted the human integrative analysis, cross-species correlation, and gene set enrichment analyses. S.L. and J.K. developed and maintained the bioinformatics pipelines and visualization scripts. K.H.J., S.P., S.C., H.C.K., and W.J.K. contributed to biological interpretation and contextualized findings within epileptogenesis mechanisms. J.K. and S.K. prepared all figures and tables and drafted the initial version of the manuscript. Peer review Peer review information Nature Communications thanks Konstantin Khodosevich, who co-reviewed with Dorottya Maria Ralbovszki, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Funding This work was supported by the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health and Welfare, Republic of Korea (RS-2024-00405260 to S.K.); the National Research Foundation of Korea (NRF), funded by the Ministry of Science and ICT (MSIT), Republic of Korea (RS-2025-23323132 to W.J.K. and RS-2023-00278314 to S.K.); the InnoCORE Program of the Ministry of Science and ICT (26-InnoCORE-02 to W.J.K.); and the Korean Neurological Association Neuro-Frontier Fellowship Award Grant (KNA-24-Neuro-Frontier Fellowship Award Grant-12 to J.K.). Data availability RNA-seq FASTQ data for the intra-amygdala kainate-induced model generated in this study have been deposited in SRA under accession code PRJNA1425071. RNA-seq FASTQ data for human samples are available at SRA under the following accession numbers: PRJNA280563, PRJNA290212, PRJNA322318, PRJNA373909, PRJNA525671, PRJNA556159, PRJNA600414, PRJNA753504, PRJNA773413, PRJNA787219, PRJNA1073977, and PRJNA1077986. RNA-seq FASTQ data for rodent samples are available at SRA under the following accession numbers: PRJNA215950, PRJNA293972, PRJNA304117, PRJNA388921, PRJNA564028, PRJNA565693, PRJNA600942, PRJNA622919, PRJNA815649, PRJNA844899, PRJNA880828, and PRJNA1006742. Processed snRNA data for human normal and MTLE hippocampal samples are available at GEO under the following accession numbers: GSE186538 and GSE160189. Processed 10X-seq scRNA data for the mouse sample is available at the Allen Brain Map website [https://portal.brain-map.org/]. DEGs of Guelfi et al., along with their associated p values and fold change estimates, were obtained from the supplementary materials. Code availability The code used in this study is available on GitHub (https://github.com/joonho345/Epilepsy_Rodent_Model_RNA) and has been archived on Zenodo74 (10.5281/zenodo.19640367). Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Joonho Kim, Sangbo Lee. Contributor Information Won-Joo Kim, Email: kzoo@yuhs.ac. Sangwoo Kim, Email: swkim@yuhs.ac. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-73796-5. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary Files Supplementary Data Reporting Summary Transparent Peer Review file Data Availability Statement RNA-seq FASTQ data for the intra-amygdala kainate-induced model generated in this study have been deposited in SRA under accession code PRJNA1425071. RNA-seq FASTQ data for human samples are available at SRA under the following accession numbers: PRJNA280563, PRJNA290212, PRJNA322318, PRJNA373909, PRJNA525671, PRJNA556159, PRJNA600414, PRJNA753504, PRJNA773413, PRJNA787219, PRJNA1073977, and PRJNA1077986. RNA-seq FASTQ data for rodent samples are available at SRA under the following accession numbers: PRJNA215950, PRJNA293972, PRJNA304117, PRJNA388921, PRJNA564028, PRJNA565693, PRJNA600942, PRJNA622919, PRJNA815649, PRJNA844899, PRJNA880828, and PRJNA1006742. Processed snRNA data for human normal and MTLE hippocampal samples are available at GEO under the following accession numbers: GSE186538 and GSE160189. Processed 10X-seq scRNA data for the mouse sample is available at the Allen Brain Map website [https://portal.brain-map.org/]. DEGs of Guelfi et al., along with their associated p values and fold change estimates, were obtained from the supplementary materials. The code used in this study is available on GitHub (https://github.com/joonho345/Epilepsy_Rodent_Model_RNA) and has been archived on Zenodo74 (10.5281/zenodo.19640367).