Skip to main content
NCBI home page
Wiley Open Access Collection logoLink to Wiley Open Access Collection
. 2026 Mar 13;67(6):3211–3225. doi: 10.1002/epi.70161

Decoding epilepsy's molecular blueprint: Machine learning unravels transcriptomic subtypes and regulatory networks

Yanping Weng 1, Yu Ma 2,3, Wanwan Hou 1, Haibo Li 4, Yuanfeng Zhou 5, Rui Zhao 6,7, Hao Li 7, Lian Chen 8, Yangyang Ma 8, Li Jin 1,9, Yi Wang 2,3,✉, Yu An 1,10,✉
PMCID: PMC13285232  PMID: 41823335

Abstract

Objective

Drug‐resistant epilepsy (DRE) affects approximately one‐third of patients with epilepsy. The molecular heterogeneity underlying DRE remains poorly defined, largely due to limited access to resected brain tissue and substantial genetic diversity. Current classifications rely primarily on clinical symptoms and histopathological features rather than molecular mechanisms, constraining mechanistic insight and the development of targeted therapies. This study aimed to develop a transcriptome‐based, machine learning–guided framework for molecular classification of DRE.

Methods

We performed comprehensive RNA sequencing on 153 surgically resected samples from 95 patients with DRE. Two transcriptomic subtypes were identified through unsupervised clustering. We also leveraged a weighted correlation network‐based framework and systematic transcriptional signature comparison and developed a classification model using machine learning algorithms.

Results

Unsupervised clustering revealed two molecular subtypes that diverged from traditional pathological classifications, indicating an alternative transcriptomic basis for epilepsy pathogenesis. A classification model was constructed based on four key differentially regulated pathways: (1) neuroactive ligand–receptor interaction, (2) cAMP signaling, (3) γ‐aminobutyric acid (GABA)ergic synapse, and (4) calcium signaling. Among the tested algorithms, the random forest model demonstrated superior performance, achieving 96% classification accuracy with an area under the curve (AUC) of .95.

Significance

These molecular subtypes and their pathways could serve as key molecular hallmarks of epilepsy, offering valuable insights for developing targeted therapies. Moreover, our findings introduce a novel framework for classifying epilepsy based on its molecular nature, potentially connecting the clinical symptoms with the underlying causes more effectively.

Keywords: classification model, epilepsy subtypes, gene network, machine learning, transcriptome sequencing

Key points.

  • Transcriptome‐based clustering reveals two molecular subgroups of drug‐resistant epilepsy independent of histopathological classification.

  • The two subgroups show distinct cellular compositions, characterized by high‐ and low‐neuron transcriptomic profiles.

  • Synaptic signaling pathways, including γ‐aminobutyric acid (GABA)ergic, calcium, cAMP, and neuroactive ligand–receptor pathways, underlie subtype differences.

  • Pathway‐level machine learning models robustly classify epilepsy subtypes and generalize to an external dataset.

1. BACKGROUND

Epilepsy is a common neurological disorder characterized by recurrent unprovoked seizures. 1 For carefully selected patients with drug‐resistant epilepsy (DRE), surgical intervention is an established and effective therapeutic option. Human brain tissues obtained during therapeutic resections, while collected primarily for clinical care, enable the investigation of the cellular and transcriptomic features of epileptogenic tissue, providing unique insights into disease mechanisms.

Recent transcriptomic studies, including bulk and single‐cell analyses, have revealed the gene expression patterns associated with epileptic tissues. For example, brain tissues from patients with focal cortical dysplasia (FCD) types I and III show greater inter‐individual transcriptomic heterogeneity than those with FCD type II or hemi‐megalencephaly. 2 Furthermore, the inflammatory response is believed to be related to the development of epilepsy. 3 Among the cellular mediators of this inflammatory response, astrocytes have been identified as both major sources and targets of pro‐epileptogenic inflammatory signaling, contributing to seizure generation and seizure‐associated neuronal injury. 4 , 5 , 6 Transcriptomic alterations involving ion channel splicing and mobile element activity have also been associated with epileptogenesis. 7 , 8 However, most efforts have been limited to a single subtype and lack integrative frameworks to capture the complex heterogeneity of epilepsy.

Despite these advances, epilepsy is currently classified according mainly to clinicopathological morphology, 9 with little consideration of the underlying molecular diversity. For instance, FCD, a major cause of DRE, is categorized into types I–III based on the morphology of neurons. 10 Mounting evidence suggests that these morphological differences may originate from uncharacterized molecular heterogeneity. 11 , 12 , 13 However, owing to the difficulty of collecting materials and high genetic heterogeneity, efforts to subtype the brain tissue of patients with DRE are relatively limited. 14 , 15

Machine learning provides a data‐driven approach to systematically capture molecular diversity. The types of machine learning are subclassified into supervised learning and unsupervised learning, and machine learning can be applied to develop disease identification models and redefine patient classifications. 16 One current application of machine learning in disease research is the integration of histopathological imaging with transcriptomic gene expression data via models such as convolutional neural networks (CNNs) for disease state classification. 17 , 18 The prediction of survival in patients with disease is another important application. 19 However, few studies have been conducted to improve epilepsy subtype classification.

To develop a molecular classification framework transcending conventional histopathological boundaries, we integrated RNA sequencing data from 153 surgically resected samples from patients with DRE. Probing unsupervised clusters, gene co‐expression network modules, and differentially expressed genes, we identified distinct molecular subtypes of epilepsy and further explored the underlying regulatory mechanisms. Finally, a machine learning–based classification model for epilepsy subtypes was constructed based on the key pathways.

2. MATERIALS AND METHODS

2.1. Study cohort

In this study, 95 patients with DRE had surgery to remove the seizure focus. For 58 of these patients, perilesional tissue adjacent to the lesions was collected to ensure complete removal of the epileptogenic zone. Postoperative histopathological examination confirmed that they showed no obvious structural abnormalities or gliosis. All resected specimens were obtained under the guidance of experienced neurosurgeons and neuropathologists. Preoperative evaluation included magnetic resonance imaging (MRI), video‐EEG (electroencephalography) monitoring, and [18F]‐FDG‐PET (fluorodeoxyglucose positron emission tomography) to localize the epileptogenic lesion, which was further confirmed by postoperative histopathological examination. For each case, the core lesion tissue identified intraoperatively was collected immediately after resection and snap‐frozen at −80°C for subsequent RNA extraction. Patient follow‐up was conducted through regular visits or telephone calls. Before the enrollment, each study participant was fully informed about the study and its associated risks and provided written consent. The study was approved by the ethics committee of the Children's Hospital of Fudan University (FE22264I).

2.2. RNA sample preparation and sequencing

Each resected tissue was ≈20 mg and was embedded in OCT (Optimal Cutting Temperature compound), sectioned at 10 μm thickness using a cryostat. A total of 200 sections were randomly collected for RNA extraction.

RNA was extracted using the miRNeasy Mini Kit (Qiagen), and RNA integrity was assessed with the Agilent Bioanalyzer. Libraries were prepared by rRNA depletion using the Ribo‐off rRNA Depletion Kit (Vazyme) and constructed with the VAHTS Universal V6 RNA‐seq Library Prep Kit. Library quality and concentration were evaluated using Qubit and the Bioanalyzer. Sequencing was performed on the Illumina NovaSeq platform (150 bp paired‐end).

Raw reads were quality‐checked with FastQC, trimmed with Trim‐galore, and aligned to GRCh38 using HISat2. 20 Gene counts were obtained with FeatureCount. Based on gene annotations from the Ensembl database, only protein‐coding messenger RNAs (mRNAs) were retained. Genes with low expression (counts <10) or zero expression in more than half of the samples were removed, resulting in a total of 17 389 genes.

2.3. Unsupervised clustering and cell composition analysis

The FPKM (Fragments Per Kilobase of transcript per Million mapped reads) values were normalized across samples. Euclidean distances between samples were then calculated and used to perform unsupervised clustering of the lesion samples. Cellular composition was estimated using CIBERSORT 21 (Cell‐type Identification By Estimating Relative Subsets Of RNA Transcripts) R package with the scRNA‐seq reference datasets from three studies (1000 permutations).

2.4. Network construction and functional analysis

To reduce noise from low‐variance genes, we selected the top 6000 genes ranked by coefficient of variation. Gene co‐expression network analysis was conducted using the R package WGCNA 22 (minimum module size = 30), and module eigengenes were correlated with clinical traits. Module preservation between lesion and perilesional tissue was assessed by permutation tests, and overlap between modules and 118 epilepsy‐related genes was evaluated using Fisher's exact test. Epilepsy‐related genes were collected from the epilepsy‐related databases, including the CarpeDB database, the epiGAD database, 23 and related reports on epilepsy‐causing genes. 24 , 25 The three measurement values were mapped to the range [0,1] using min–max normalization method and then summed as the final score.

Differentially expressed genes were identified with the R package DESeq2 26 (|FC| >1.5, p < .05). Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment and Gene Ontology (GO) enrichment analyses were performed using clusterProfiler 27 R package. A p value <.05 was considered statistically significant. Gene set enrichment analysis (GSEA) was performed using clusterProfiler R package to identify the biological pathways that were significantly enriched in GO and KEGG databases.

2.5. Machine learning models

The TensorFlow platform framework and Scikit‐learn library were used to construct a classification model for lesion subtype grouping. We used 70% of the data for training and validation, and the rest were used for testing. Fivefold cross‐validation was used to verify the model. In each round of cross‐validation, the accuracy, area under the curve (AUC), and F1 score were used to test the model's discriminative ability.

3. RESULTS

3.1. Diverse collection of patient brain samples and perilesional tissues

A total of 95 patients with epilepsy undergoing surgical resection were enrolled in this study. For transcriptomic profiling, we obtained tissue from the epileptogenic region as well as adjacent non‐lesional tissue, when available, resulting in 58 paired non‐lesional tissues.

During RNA extraction, 9 lesion and 13 perilesional samples with low RNA integrity (RIN <6) were excluded. In total, 86 lesion and 45 perilesional samples were retained for analysis. The patients had a mean age of 8.1 ± 7.4 years, and 68.6% were male. Lesions were most frequently located in the temporal lobe (46 cases), followed by the frontal (20), parietal (2), occipital (2), and insular lobes (4), with 12 patients showing involvement of multiple regions. Patients decreased their number of ASMs on average by 46.4% ± 24.7%. The details of the 86 lesions are listed in Table S1.

3.2. Epileptic lesions and perilesional tissues showed similar global transcriptional patterns

Principal component analysis (PCA) was performed on our data together with transcriptome data from 31 human organs in the Genotype‐Tissue Expression (GTEx) database, including 1152 brain transcriptome samples. In the resulting PCA plot, all samples from our study clustered within the GTEx brain tissue samples, indicating that our data exhibited transcriptional profiles consistent with brain tissue (Figure S1A). PCA of samples grouped by sequence batch, age at surgery (for both adults and young individuals), sex, RNA Integrity Number (RIN), pathology, and brain region showed no significant differences, suggesting that these variables do not contribute substantially to the observed gene expression patterns (Figure S1B–G).

To obtain a global understanding of the transcriptional landscape, we conducted PCA considering both lesion proximity (lesional vs perilesional) and brain region (Figure 1A). The PCA plot showed no clear separation between lesional and perilesional tissues or among different brain regions, indicating that the major sources of transcriptomic heterogeneity are not explained by these factors. This result is not unexpected because epilepsy is characterized by the abnormal activity in a relatively small subset of neurons and associated cell types. The same results were reported in another study involving single‐cell sequencing of the epileptic and non‐epileptic temporal cortex. 28

FIGURE 1.

FIGURE 1

Open in a new tab

Unsupervised clustering and group information. (A) Principal component analysis (PCA) plot of epilepsy brain tissues, grouped by disease status and patient identity. Each dot represents a sample, with the shape indicating the disease status (triangle for the lesion, circle for perilesional tissue), and the color denoting the patient number. (B) Euclidean distance clustering heatmap of brain tissue from lesions in the two clusters. The color indicates the Euclidean distance value. Red and blue indicate high and low distances, respectively. (C) Box plot showing the deconvolution results of two clusters based on the single‐cell sequencing data of the Allen Brain Atlas. Astro, Astrocytes; Oligo, Oligodendrocytes; Endo, Endothelial. **p < .05; ***p < .01; ****p < .001. Statistical analysis was performed using nonparametric two‐tailed Mann–Whitney test. (D, F) Comparison of the RNA quality number (RQN) and drug reduction rates across the two groups. The violin plots show the distribution, and the lines inside represent the means ± SDs. Medications reduction rate: Proportion of preoperative antiepileptic drugs (AEDs) discontinued after surgery. For instance, if a patient were taking three AEDs before surgery and continued one drug postoperatively, the reduction rate would be calculated as (3–1) / 3 = 66.7%. (E) Comparison of age, disease course, and starting age across the two groups. (G) Bar chart of the proportion of pathological information in the lesion clusters. The horizontal axis is the proportion of each pathology that appears in the group. The color represents the pathology.

3.3. Transcriptome‐based clustering defines subgroups with consistent cellular and pathological features

Hierarchical clustering of all epileptogenic focus samples revealed 11 outliers with large Euclidean distances, whereas the remaining samples can be divided into two distinct clusters. These two clusters comprised 56 and 19 epileptogenic focus samples, respectively (Figure 1B). We also performed clustering based on different clinical pathologies (FCD type I, type II, type III, and tumor). Of interest, clustering based on FCD subtypes did not align fully with the patterns observed in the overall transcriptomic profile (Figure S2A). To ensure that the observed separation was not driven merely by the pathological distinction of samples, we rigorously excluded all samples with a pathological diagnosis of hippocampal sclerosis and re‐performed the unsupervised clustering solely on the remaining cohort (comprising predominantly FCD cases). The persistence of the two distinct clusters in the absence of hippocampal sclerosis samples provides strong evidence that our molecular classification captures biological differences that are independent of the pathological diagnoses (Figure S3B).

Cell composition was then inferred by CIBERSORTx using the Allen Brain Atlas. 29 This analysis revealed marked differences in cellular composition between the two transcriptomic clusters. One cluster exhibited higher proportions of neuronal populations, including excitatory neuron subtypes such as L2/Pvalb, whereas the other cluster was relatively enriched for glial cell types, including astrocytes and oligodendrocytes (Figures 1C and S3D). Based on their cellular composition, particularly differences in neuronal proportions, we refer to them as the high‐neuron (HN) and low‐neuron (LN) hereafter. We also found that outlier samples contained only minimal neuron‐related cell types (Figure S3C).

We then explored the clinical differences between groups. No significant differences were observed between the clusters in RNA quality number (RQN), postoperative medication reduction rate, or age‐related clinical variables, including age at surgery, age at onset, and epilepsy duration (the number of years from seizure onset to surgery) (Figure 1D–F). Similarly, the clusters were well balanced in terms of sex distribution, hemisphere of origin, brain region, and Engel outcome (all p > .05; Figure S2C–F). Residual analysis further confirmed that clustering was driven by transcriptomic similarity rather than the influence of covariates (Figure S2B).

Pathological annotation revealed that the HN group was enriched for FCD (53.6%) and cortical structural abnormality (12.5%), whereas the LN group showed higher frequencies of hippocampal sclerosis (21.1%). Notably, vascular proliferation and neuronal abnormality were absent in the HN group (Figures 1G and S3A). Although this distribution was partly consistent with the known histopathological features of these lesions, a few samples in the HN group showed features of hippocampal sclerosis. Review of their pathological reports indicated that these cases represented mild hippocampal sclerosis, showing fewer transcriptional features of neuronal loss.

3.4. WGCNA identified neuron‐enriched modules with subtype‐specific associations

We constructed gene co‐expression networks for each group using WGCNA (Weighted correlation network analysis) algorithm. We identified 12 modules in the HN group and 13 in the LN group (Figure S4A,B). Several modules showed functional enrichment, including voltage‐gated ion channel pathways (HN‐turquoise, LN‐blue, LN‐turquoise) and MHC (Major Histocompatibility Complex)‐related pathways (HN‐tan, LN‐pink) (Figure S4D).

Module eigengenes (MEs) were correlated with disease status (lesion or perilesional tissue), revealing strong associations in the HN‐turquoise, HN‐yellow, HN‐brown, LN‐pink, LN‐tan, LN‐blue, and LN‐turquoise modules (Figure 2A). Among them, the HN‐turquoise, LN‐pink, and LN‐tan modules were positively correlated with disease, whereas the HN‐yellow, HN‐brown, and LN‐blue modules were negatively correlated.

FIGURE 2.

FIGURE 2

Open in a new tab

Overview of the modules of the gene network. (A) Comparison of the Pearson correlation coefficient between the disease status and module eigengene of each group. Gene modules were annotated by their biological function via gene enrichment analysis. The circle size indicates the number of genes in the module. (B) Circular heatmap showing the score of each group from three aspects. From inside to the outside: The Pearson correlation coefficient between disease status and module eigengenes (blue and red), the preservation rank of modules (cyan), the gene overlap ratio between modules and reported epilepsy‐related genes (purple), and the final score with scaling by the min‐max normalization method (green). The gene module names and their biological functions are shown on the outside.

In addition to considering module–disease correlations, we also calculated the preservation ranks of modules (Figure S4C) and the gene overlap ratios between modules and known epilepsy‐related genes (Figure S5A). Integrating these metrics, the HN‐turquoise and LN‐blue modules received the highest overall scores (Figure 2B).

Functional enrichment of these two key modules highlighted pathways related to membrane structures and ion signaling, with the neuroactive ligand–receptor interaction pathway being most significant (Figures 3A,B and S5B–G). Ten pathways were shared between the two modules (Figure 3C). Notably, the HN‐turquoise module was positively correlated with disease, whereas the LN‐blue module was negatively correlated, indicating that the mechanisms of the two groups might be opposite (Figure 2A).

FIGURE 3.

FIGURE 3

Open in a new tab

Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways of gene modules. (A, B) KEGG pathway enrichment analysis of the module with the highest score in each group. (C) Heatmap showing the gene ratios of pathways shared by both groups. (D) Heatmap showing the enrichment of genes within each cell type across modules. The enrichment scores are represented by color intensity, with asterisks (*) indicating significance. *p < .1; **p < .05; ***p < .01. Statistical analysis was performed using nonparametric two‐tailed Mann–Whitney test.

Cell‐type enrichment analysis further revealed that genes in the HN‐turquoise, LN‐blue, and LN‐turquoise modules were expressed mainly in neurons, especially excitatory neurons, consistent with functional enrichment results (Figure 3D).

3.5. Opposite regulation of shared pathways in subgroups

Differential expression analysis revealed that 686 genes were upregulated and 109 genes were downregulated in the HN group (Figure S6A), whereas 1219 genes were upregulated and 1245 genes were downregulated in the LN group (Figure S6C).

Functional enrichment (GO and KEGG) was performed on the up‐ and downregulated genes, respectively. Notably, in both subgroups, the most significant pathway identified in the KEGG analysis was neuroactive ligand–receptor interaction, which was upregulated in the HN group but downregulated in the LN group (Figure 4A,B). In addition, synaptic‐related pathways were downregulated and inflammation‐related pathways were upregulated in the LN group (Figure S6B,D). Seven shared pathways were upregulated in the HN group but downregulated in the LN group, whereas one showed the opposite pattern (Figure 4C). It is important to note that four of these dysregulated pathways, which were neuroactive ligand–receptor interaction, cAMP signaling, GABAergic synapse, and calcium signaling, were also identified as key components in the co‐expression network from the WGCNA analysis. The negative Pearson correlation of log2‐fold changes for genes within the four pathways between the two subgroups further corroborated the findings (Figure S7). In addition, GSEA validated the same opposing regulation pattern in these four pathways (Figure 4E,F).

FIGURE 4.

FIGURE 4

Open in a new tab

Differentially expressed pathways of each group. (A,B) Bar plot of the result of the Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis of the DEGs (Differential Expression Genes). The abscissa is the adjusted p‐value, and red color indicates the enrichment of upregulated genes, whereas blue indicates the downregulated pathways. The pathway demonstrated adjusted p‐value < .05. (C) The shared pathways of the two groups. Venn diagram showing the overlap between the up‐ and downregulated pathways in the high‐neuron (HN) group and the low‐neuron (LN) group. The color intensity corresponds to the number of overlapping pathways, with darker red indicating greater overlap. Heatmap showing the gene ratios of the overlapping pathways in the two groups. (D, E) GSEA (Gene Set Enrichment Analysis) of key pathways for the HN group (D) and LN group (E). Four pathways—neuroactive ligand–receptor interaction, cAMP signaling, γ‐aminobutyric acid (GABA)ergic synapse, and calcium signaling—were enriched in both subtypes, with opposite enrichment patterns. The x‐axis represents the rank of all genes ordered by differential expression between HN and LN, whereas the y‐axis indicates the running enrichment score reflecting the degree to which each pathway is overrepresented at the top or bottom of the ranked list.

To further validate the robustness and generalizability of our findings, we additionally collected 14 brain tissue samples from trauma patients without epilepsy as external controls. Following batch correction, differential expression analysis of both the HN and LN lesion tissues against these controls revealed that neuronal signaling–related pathways were consistently enriched in the KEGG pathway analysis (Figure S8).

3.6. Epilepsy classification model development using machine learning with identified key pathways

Based on the distinct pathway regulation patterns observed between the two subgroups, we used supervised machine learning algorithms to preoperatively predict the classification of epilepsy patients. Six machine learning algorithms were selected to build classification models, including the random forest, convolutional neural network (CNN), K‐nearest neighbor algorithm (KNN), support vector machine (SVM), least absolute shrinkage and selection operator (Lasso), and extreme gradient boosting (XGBoost). To avert the bias introduced by the unequal sample sizes between subgroups, we employed the synthetic minority over‐sampling technique (SMOTE) to supplement the data for groups with fewer samples. Then, six classification models were built based on the training dataset. The results showed that the random forest model performed best among the six classification models in terms of five‐fold cross validation (Figure 5A–C). The random forest model also performed best in the test set, with the lowest variance and the highest accuracy value of .96 and an AUC value of .9564 (Figure 5D). In the best‐performing random forest model, the gene with the highest contribution ranking was GNG3, which encodes a G‐protein gamma subunit involved in neuronal signaling (Figure 5E).

FIGURE 5.

FIGURE 5

Open in a new tab

Classification models constructed with machine learning approaches. (A–C) Box plots of the area under the curve (AUC) values (A), accuracy (B), and F1 scores (C) of the models in fivefold cross‐validation. Lines inside the box plots represent the 25th, 50th, and 75th quantiles. (D) Radar chart showing the performance of the classification models in the test set. Six machine learning models—SVM (support vector machine), KNN (K‐nearest neighbor algorithm), CNN (convolutional neural network), Lasso (least absolute shrinkage and selection operator), XGBoost (extreme gradient boosting), and random forest—were evaluated using three metrics: AUC, accuracy, and F1 score. Higher values toward the edges indicate better performance. (E) Contribution of genes in the random forest model. The figure shows the 10 genes that contributed the most to the model.

3.7. Validation of the subtyping framework in an independent cohort

Through the integrative analysis of epilepsy with network‐frame analysis, we identified two subgroups of epilepsy and emphasized their molecular hallmarks from several aspects. Nonetheless, the prevalence of these subgroups remains unknown in other datasets. To address this gap, we reviewed the epilepsy samples in the GSE256068. 30

We performed k‐means cluster analysis on the samples, and the clustering patterns were consistent with our findings rather than aligning with traditional pathological classification (Figure 6A).

FIGURE 6.

FIGURE 6

Open in a new tab

Subtyping epilepsy samples in the GSE256068. (A) Unsupervised clustering of epilepsy samples (TLE‐HS, FCD IIa, FCD IIb, and TSC) and control samples (cortex and hippocampus). (B) Prediction result of the samples using the random forest model. (C–E) Differences in the proportions of neurons (C), oligodendrocytes (D), and neoplastic cells (E) between the two groups. **p < .05; ***p < .01; ****p < .001. (F) Expression patterns of subtype markers across epilepsy subtypes (n = 135). The samples are arranged by subtype.

We predicted the subgroups of each sample with the random forest model, which achieved the best performance in the test set. Most samples were divided into the HN group, whereas four samples were divided into the LN group (Figure 6B). These four samples were also in the same cluster shown in Figure 6A.

We performed cell‐type deconvolution on each subgroup. Notably, the group classified as HN in the external data also presented a significantly greater neuronal proportion than the LN group did, mirroring the cell composition pattern observed in our original cohort (Figure 6C). In addition, we found that the LN group had significantly greater proportions of oligodendrocytes and endothelial cells, which was consistent with our initial findings (Figure 6D,E). We further explored the expression patterns of subgroup markers. The samples were not grouped according to the type of covariates, such as pathology and brain regions, but were related to the transcriptome expression profile (Figure 6F).

4. DISCUSSION

In this study, we identified transcriptome‐based molecular subtypes of drug‐resistant epilepsy (DRE) using data‐driven analytical framework. These subtypes capture reproducible molecular patterns that are largely independent of conventional histopathological classifications. Of note, this molecular stratification was observed consistently in an independent external cohort, supporting the robustness and generalizability of this classification scheme. Together, our findings suggest that transcriptomic profiling provides a complementary perspective on the molecular and cellular heterogeneity of DRE, extending beyond traditional pathology‐based approaches.

In our analysis, gene expression profiles did not show a clear separation between the lesion tissue and the putatively non‐lesional tissue. This observation is consistent with the notion that epilepsy does not uniformly affect all neuronal subtypes, 28 and that epileptogenic may retain a substantial proportion of transcriptionally intact cells. The coexistence of affected and relatively preserved cell populations may likely blur transcriptomic distinctions at the bulk tissue level. In line with this interpretation, neither our cohort nor the external validation dataset demonstrated a strong concordance between transcriptome‐based clusters and traditional pathological classifications. Growing evidence suggests that distinct pathological entities may share common molecular mechanisms. For instance, selective transcriptional impairment of parvalbumin‐positive GABAergic interneurons has been reported in FCD type I/III but not type II. 31 Another example is the dysregulation of the mTOR (mechanistic target of rapamycin) pathway has been implicated in both FCD and tuberous sclerosis complex. 32 Together these observations indicate pathology‐based classifications, which reflect primarily morphological features, may not effectively capture the underlying molecular heterogeneity underlying epilepsy.

Across molecular subtypes, we identified four critical pathways including neuroactive ligand‐receptor interaction, cAMP signaling, GABAergic synapse, and calcium signaling, which collectively define divergent regulatory states of synaptic excitability. These pathways were consistently upregulated in the HN group but downregulated in the LN group, indicating opposite regulatory patterns that may reflect distinct mechanisms of disease progression. Notably, the neuroactive ligand–receptor interaction pathway is one of the pathways able to distinguish groups, which has been implicated widely in epilepsy. 33 , 34 , 35 , 36 Both upregulation 37 and downregulation 38 , 39 of this pathway have been reported in different epilepsy contexts, including temporal lobe epilepsy and chemically induced mouse models. Similar epileptogenesis‐dependent roles have been described for cAMP 40 , 41 and calcium 42 signaling pathways, which are central to synaptic transmission and neuronal excitability. Alterations in the GABAergic synapse pathway further indicate disruption of inhibitory–excitatory balance, a hallmark of epileptic networks. In summary, these findings support the notion that the identified molecular subtypes may reflect divergent regulatory states of synaptic excitability.

Recent advances highlight the value of data‐driven subtyping in epilepsy, including imaging‐based classifications 43 and molecular subtypes associated with seizure recurrence. 14 Recent advances in DNA methylation–based classification have also emerged as a powerful tool for distinguishing epilepsy‐associated lesions. Studies by Jabari et al. 44 and Hoffmann et al. 45 have established comprehensive methylation reference maps that reliably distinguish FCD subtypes, hippocampal sclerosis, LEATs (long‐term epilepsy–associated tumors), and mixed pathology cases. Integrating transcriptomic and epigenomic information may therefore provide a more comprehensive framework for epilepsy classification. In the present study, we constructed epilepsy classification models based on pathway‐level features using multiple machine learning algorithms, among which the random forest model demonstrated the best performance. Of note, the interpretability of the random forest also enables identification of biologically meaningful contributors, such as the GNG3 gene, which is involved in G protein–coupled receptor signaling and has been implicated previously in epilepsy and neurodevelopmental disorders. 46 Compared with single‐gene markers, pathway‐level modeling reduces noise and enhances biological interpretability, supporting its utility for molecular stratification.

Cell‐type deconvolution revealed significant differences in neuronal composition between molecular subtypes. When our machine learning–based classification model was applied to an external dataset, subsequent cell‐type deconvolution demonstrated a consistent pattern: the HN group exhibited a greater neuronal proportion than did the LN group. Although the HN cluster was enriched for FCD cases and the LN cluster for hippocampal sclerosis, this correspondence was not absolute. Notably, cases of mild hippocampal sclerosis were present in the HN group, and a subset of FCD cases clustered within the LN group. These findings further support the ability of transcriptomic classification to capture molecular features that extend beyond conventional pathological categories.

Several limitations should be acknowledged. The sample size was relatively modest, which may limit generalizability and contribute to heterogeneity between cohorts. Despite our efforts to account for pathological subtype and brain region, subtle region‐specific effects cannot be completely excluded. In addition, although multiple machine learning models were evaluated, further optimization and validation in larger, independent datasets are warranted. Future studies integrating multi‐omics data and biological experiment verification will be necessary to enhance the biological significance and clinical relevance of this classification framework.

5. CONCLUSION

This study identified two transcriptomic subgroups (designated HN and LN) and developed a classification model. In summary, this network‐based system analysis revealed the dysregulated pathways in the different subtypes of epilepsy, which may indicate the different disease states. The network‐based system analysis method of this study provides a reference for related research. We suggest that anomalously regulated pathways could serve as targets for the development of new drugs and that the classification model of epilepsy lesions could provide a new basis for the classification of epilepsy.

AUTHOR CONTRIBUTIONS

Yu An, Yi Wang, and Li Jin conceived the study and designed the research. Yi Wang and Yuanfeng Zhou performed the clinical diagnosis. Rui Zhao and Hao Li conducted the surgical procedures. Lian Chen and Yangyang Ma carried out the clinical pathology assessments. Yu Ma collected and analyzed clinical data and prepared samples. Wanwan Hou and Haibo Li performed RNA library preparation and sequencing. Yaping Weng conducted the data analysis. Yanping Weng and Yu An developed the research methodology and drafted the manuscript. Yu An and Yi Wang revised the manuscript. All authors contributed to data interpretation, critically revised the manuscript for important intellectual content, approved the final version for submission, agreed on the journal to which the manuscript was submitted, and accept responsibility for all aspects of the work.

FUNDING INFORMATION

This study was supported by the National Key R&D Program of China (No. 2024YFC3406700, 2024YFC3406701), Ningbo Science and Technology project (2023Z178), and the Municipal Science and Technology Major Project (Grant No. 2017SHZDZX01).

CONFLICT OF INTEREST STATEMENT

The authors report no competing interests.

ETHICS STATEMENT

We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines. This research was conducted with the approval of the institutional review board. There was no harm to any participants.

Supporting information

Figure S1.

EPI-67-3211-s002.pdf (2.3MB, pdf)

Table S1.

EPI-67-3211-s001.docx (27.1KB, docx)

ACKNOWLEDGMENTS

This study was supported by the Human Phenome Data Center of Fudan University.

Contributor Information

Yi Wang, Email: yiwang@shmu.edu.cn.

Yu An, Email: anyu@fudan.edu.cn.

DATA AVAILABILITY STATEMENT

All datasets generated or analyzed during this study are included in the main manuscript or in the additional files. Raw RNA‐seq data have been deposited in The National Omics Data Encyclopedia (NODE; accession number: OEP00006719). Due to patient confidentiality and the restrictions of institutional review board (IRB) approval, the data are available under controlled access and can be obtained upon reasonable request through the NODE application process.

REFERENCES

  • 1. Fisher RS, Acevedo C, Arzimanoglou A, Bogacz A, Cross JH, Elger CE, et al. ILAE official report: a practical clinical definition of epilepsy. Epilepsia. 2014;55(4):475–482. [DOI] [PubMed] [Google Scholar]
  • 2. Chung C, Yang X, Bae T, Vong KI, Mittal S, Donkels C, et al. Comprehensive multi‐omic profiling of somatic mutations in malformations of cortical development. Nat Genet. 2023;55(2):209–220. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3. Tröscher AR, Mair KM, Verdú de Juan L, Köck U, Steinmaurer A, Baier H, et al. Temporal lobe epilepsy with GAD antibodies: neurons killed by T cells not by complement membrane attack complex. Brain. 2023;146(4):1436–1452. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4. Aronica E, Ravizza T, Zurolo E, Vezzani A. Astrocyte immune responses in epilepsy. Glia. 2012;60(8):1258–1268. [DOI] [PubMed] [Google Scholar]
  • 5. Pai B, Tome‐Garcia J, Cheng WS, Nudelman G, Beaumont KG, Ghatan S, et al. High‐resolution transcriptomics informs glial pathology in human temporal lobe epilepsy. Acta Neuropathol Commun. 2022;10(1):149. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6. Chen ZP, Wang S, Zhao X, Fang W, Wang Z, Ye H, et al. Lipid‐accumulated reactive astrocytes promote disease progression in epilepsy. Nat Neurosci. 2023;26(4):542–554. [DOI] [PubMed] [Google Scholar]
  • 7. Hu K, Liang P. Transcriptome analysis reveals higher levels of Mobile element‐associated abnormal gene transcripts in temporal lobe epilepsy patients. Front Genet. 2021;12:767341. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8. Heinzen EL, Yoon W, Weale ME, Sen A, Wood NW, Burke JR, et al. Alternative ion channel splicing in mesial temporal lobe epilepsy and Alzheimer's disease. Genome Biol. 2007;8(3):R32. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9. Blümcke I, Coras R, Miyata H, Özkara C. Defining clinico‐neuropathological subtypes of mesial temporal lobe epilepsy with hippocampal sclerosis. Brain Pathol. 2012;22(3):402–411. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10. Guerrini R, Duchowny M, Jayakar P, Krsek P, Kahane P, Tassi L, et al. Diagnostic methods and treatment options for focal cortical dysplasia. Epilepsia. 2015;56(11):1669–1686. [DOI] [PubMed] [Google Scholar]
  • 11. Galvão IC, Kandratavicius L, Messias LA, Athié MCP, Assis‐Mendonça GR, Alvim MKM, et al. Identifying cellular markers of focal cortical dysplasia type II with cell‐type deconvolution and single‐cell signatures. Sci Rep. 2023;13(1):13321. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12. Honke J, Hoffmann L, Coras R, Kobow K, Leu C, Pieper T, et al. Deep histopathology genotype‐phenotype analysis of focal cortical dysplasia type II differentiates between the GATOR1‐altered autophagocytic subtype IIa and MTOR‐altered migration deficient subtype IIb. Acta Neuropathol Commun. 2023;11(1):179. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13. Fang C, Zhang X, Yang L, Sun L, Lu Y, Liu Y, et al. Transcriptomic and morphologic vascular aberrations underlying FCDIIb etiology. Nat Commun. 2025;16(1):3320. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14. Hershberger CE, Louis S, Busch RM, Vegh D, Najm I, Bazeley P, et al. Molecular subtypes of epilepsy associated with post‐surgical seizure recurrence. Brain Commun. 2023;5(5):fcad251. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15. Stessman HAF, Turner TN, Eichler EE. Molecular subtyping and improved treatment of neurodevelopmental disease. Genome Med. 2016;8(1):22. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16. Deo RC. Machine learning in medicine. Circulation. 2015;132(20):1920–1930. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17. Wang Y, Kartasalo K, Weitz P, Ács B, Valkonen M, Larsson C, et al. Predicting molecular phenotypes from histopathology images: a transcriptome‐wide expression‐morphology analysis in breast cancer. Cancer Res. 2021;81(19):5115–5126. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18. Lombardo E, Hess J, Kurz C, Riboldi M, Marschner S, Baumeister P, et al. DeepClassPathway: molecular pathway aware classification using explainable deep learning. Eur J Cancer. 2022;176:41–49. [DOI] [PubMed] [Google Scholar]
  • 19. Chen X, Wang CC, Yin J, You ZH. Novel human miRNA‐disease association inference based on random Forest. Mol Ther Nucleic Acids. 2018;13:568–579. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20. Kim D, Paggi JM, Park C, Bennett C, Salzberg SL. Graph‐based genome alignment and genotyping with HISAT2 and HISAT‐genotype. Nat Biotechnol. 2019;37(8):907–915. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21. Newman AM, Liu CL, Green MR, Gentles AJ, Feng W, Xu Y, et al. Robust enumeration of cell subsets from tissue expression profiles. Nat Methods. 2015;12(5):453–457. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22. Langfelder P, Horvath S. WGCNA: an R package for weighted correlation network analysis. BMC Bioinformatics. 2008;9:559. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23. Tan NCK, Berkovic SF. The epilepsy genetic association database (epiGAD): analysis of 165 genetic association studies, 1996–2008. Epilepsia. 2010;51(4):686–689. [DOI] [PubMed] [Google Scholar]
  • 24. McTague A, Howell KB, Cross JH, Kurian MA, Scheffer IE. The genetic landscape of the epileptic encephalopathies of infancy and childhood. Lancet Neurol. 2016;15(3):304–316. [DOI] [PubMed] [Google Scholar]
  • 25. Wang J, Lin ZJ, Liu L, Xu HQ, Shi YW, Yi YH, et al. Epilepsy‐associated genes. Seizure. 2017;44:11–20. [DOI] [PubMed] [Google Scholar]
  • 26. Love MI, Huber W, Anders S. Moderated estimation of fold change and dispersion for RNA‐seq data with DESeq2. Genome Biol. 2014;15(12):550. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27. Yu G, Wang LG, Han Y, He QY. clusterProfiler: an R package for comparing biological themes among gene clusters. OMICS J Integr Biol. 2012;16(5):284–287. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28. Pfisterer U, Petukhov V, Demharter S, Meichsner J, Thompson JJ, Batiuk MY, et al. Identification of epilepsy‐associated neuronal subtypes and gene expression underlying epileptogenesis. Nat Commun. 2020;11(1):5038. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29. Ding SL, Royall JJ, Sunkin SM, Ng L, Facer BAC, Lesnar P, et al. Comprehensive cellular‐resolution atlas of the adult human brain. J Comp Neurol. 2016;524(16):3127–3481. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30. François L, Romagnolo A, Luinenburg MJ, Anink JJ, Godard P, Rajman M, et al. Identification of gene regulatory networks affected across drug‐resistant epilepsies. Nat Commun. 2024;15(1):2180. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31. Medici V, Rossini L, Deleo F, Tringali G, Tassi L, Cardinale F, et al. Different parvalbumin and GABA expression in human epileptogenic focal cortical dysplasia. Epilepsia. 2016;57(7):1109–1119. [DOI] [PubMed] [Google Scholar]
  • 32. Sumadewi KT, Harkitasari S, Tjandra DC. Biomolecular mechanisms of epileptic seizures and epilepsy: a review. Acta Epileptol. 2023;5(1):28. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33. Ran X, Li J, Shao Q, Chen H, Lin Z, Sun ZS, et al. EpilepsyGene: a genetic resource for genes and mutations related to epilepsy. Nucleic Acids Res. 2015;43(D1):D893–D899. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34. Lv YQ, Wang X, Jiao YZ, Wang YH, Wang N, Gao L, et al. Interactome overlap between risk genes of epilepsy and targets of anti‐epileptic drugs. PLoS One. 2022;17(8):e0272428. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35. Xu W, Zhang S, Feng Y, Zhang C, Xiao Y, Tian F. iTRAQ‐based proteomic analysis of the hippocampus of pentylenetetrazole‐kindled epileptic rats. Int J Dev Neurosci. 2021;81(2):125–141. [DOI] [PubMed] [Google Scholar]
  • 36. Wang XD, Liu S, Lu H, Guan Y, Wu H, Ji Y. Analysis of shared genetic regulatory networks for Alzheimer's disease and epilepsy. Biomed Res Int. 2021;2021:e6692974. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37. Li Y, Wang C, Wang P, Li X, Zhou L. Effects of febrile seizures in mesial temporal lobe epilepsy with hippocampal sclerosis on gene expression using bioinformatical analysis. Acta Epileptol. 2020;2(1):20. [Google Scholar]
  • 38. Xu J, Sun M, Wang Y, Xie A, Gao J. Identification of hub genes of Mesio temporal lobe epilepsy and prognostic biomarkers of brain low‐grade gliomas based on bioinformatics analysis. Cell Transplant. 2020;29:963689720978722. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39. Jang Y, Moon J, Lee ST, Jun JS, Kim TJ, Lim JA, et al. Dysregulated long non‐coding RNAs in the temporal lobe epilepsy mouse model. Seizure. 2018;58:110–119. [DOI] [PubMed] [Google Scholar]
  • 40. Mertz C, Krarup S, Jensen CD, Lindholm SEH, Kjær C, Pinborg LH, et al. Aspects of cAMP signaling in Epileptogenesis and seizures and its potential as drug target. Neurochem Res. 2020;45(6):1247–1255. [DOI] [PubMed] [Google Scholar]
  • 41. Chen TS, Huang TH, Lai MC, Huang CW. The role of glutamate receptors in epilepsy. Biomedicine. 2023;11(3):783. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 42. Taheri M, Pourtavakoli A, Eslami S, Ghafouri‐Fard S, Sayad A. Assessment of expression of calcium signaling related lncRNAs in epilepsy. Sci Rep. 2023;13(1):17993. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43. Xiao F, Caciagli L, Wandschneider B, Sone D, Young AL, Vos SB, et al. Identification of different MRI atrophy progression trajectories in epilepsy by subtype and stage inference. Brain. 2023;146(11):4702–4716. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44. Jabari S, Kobow K, Pieper T, Hartlieb T, Kudernatsch M, Polster T, et al. DNA methylation‐based classification of malformations of cortical development in the human brain. Acta Neuropathol. 2022;143(1):93–104. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 45. Hoffmann L, Coras R, Kobow K, López‐Rivera JA, Lal D, Leu C, et al. Ganglioglioma with adverse clinical outcome and atypical histopathological features were defined by alterations in PTPN11/KRAS/NF1 and other RAS‐/MAP‐kinase pathway genes. Acta Neuropathol. 2023;145(6):815–827. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46. Zhang YH, Zhao P, Gao HL, Zhong ML, Li JY. Screening targets and therapeutic drugs for Alzheimer's disease based on deep learning model and molecular docking. J Alzheimer's Dis. 2024;100(3):863–878. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Figure S1.

EPI-67-3211-s002.pdf (2.3MB, pdf)

Table S1.

EPI-67-3211-s001.docx (27.1KB, docx)

Data Availability Statement

All datasets generated or analyzed during this study are included in the main manuscript or in the additional files. Raw RNA‐seq data have been deposited in The National Omics Data Encyclopedia (NODE; accession number: OEP00006719). Due to patient confidentiality and the restrictions of institutional review board (IRB) approval, the data are available under controlled access and can be obtained upon reasonable request through the NODE application process.

Articles from Epilepsia are provided here courtesy of Wiley

RESOURCES