Transcriptional signatures and topological reorganization of morphometric similarity networks in temporal lobe epilepsy with unilateral hippocampal sclerosis

Abstract

Objective

To delineate morphometric similarity network (MSN) topological abnormalities and their underlying spatial transcriptomics in the normal‐appearing cortex of unilateral mesial temporal lobe epilepsy with hippocampal sclerosis (mTLE‐HS).

Methods

High‐resolution T1‐weighted magnetic resonance imaging (MRI) from 109 unilateral mTLE‐HS patients (64 left, 45 right) and 90 matched controls were analyzed to construct individual‐level MSNs by integrating five cortical morphometric features. Graph‐theoretical analysis quantified global and local network topology, and machine‐learning models assessed their values in patient identification and epileptogenic lateralization, with performance further evaluated in an independent validation dataset. Further connectome–transcriptome association analyses linked these macroscale topological abnormalities to the microscale substrates in specific gene expressions, biological pathways, cellular compositions, and neurodevelopmental windows.

Results

Patients exhibited global small‐worldness increase and local nodal reorganizations, with hyper‐connectivity in default mode and limbic networks and hypo‐connectivity in temporo‐occipital cortex. These MSN‐topology abnormalities enabled accurate patient classification (77.4%) and epileptogenic lateralization (83.2%), driven predominantly by features from limbic network (40.0% and 41.1%, respectively). These performances were reproducible in an independent validation dataset (classification accuracy = 75.4%, lateralization accuracy = 77.1%). Spatial transcriptomics mapped the MSN‐topology alterations to expression of genes enriched in RNA processing and mitochondrial energy metabolism, including key epilepsy risk genes such as DNM1 and PPFIA3. These genes showed enriched expression in excitatory neurons, astrocytes, and oligodendrocytes, peaking during neurodevelopment in early‐fetal striatum and neonatal‐to‐childhood cortex.

Significance

MSN topology delineates the pattern of cortical network reorganization in mTLE‐HS, aiding in patient identification and lateralization. The convergence of these macroscale connectomic alterations with microscale transcriptomic profiles points to an RNA–metabolic interplay that shapes cortical vulnerability of mTLE‐HS in a progressive way.

Key points.

• Morphometric similarity networks reveal cortical reorganization in mesial temporal lobe epilepsy with hippocampal sclerosis (mTLE‐HS), featuring limbic–default mode network (DMN) hyper‐connectivity and temporo‐occipital hypo‐connectivity.

• Morphometric similarity network (MSN) topological features enable high‐accuracy (>75%) individual‐level diagnosis and lateralization of the epileptogenic hemisphere, supporting epilepsy as a network disorder.

• Connectome–transcriptome mapping links MSN alterations to RNA processing and mitochondrial energy metabolism, implicating key epilepsy risk genes (e.g., DNM1 and PPFIA3).

• The spatiotemporal expression of MSN‐related genes peaks in early‐fetal striatum and neonatal‐to‐childhood cortex, suggesting a “two‐hit” model of vulnerability in mTLE‐HS.

1. INTRODUCTION

Mesial temporal lobe epilepsy with hippocampal sclerosis (mTLE‐HS) is the most common etiology of medically intractable epilepsy in adults and poses a substantial burden on health care systems. ¹ The neurobiological mechanisms underlying mTLE‐HS remain incompletely elucidated. ² Neuroimaging studies have sought to clarify its pathophysiological mechanisms from the perspective of brain networks, revealing widespread disturbances in functional and white‐matter structural connectivity that extend beyond the sclerotic hippocampus in patients with mTLE‐HS. ³ , ⁴ However, given that epilepsy is fundamentally characterized by synchronized neuronal discharges across cortical regions, research on cortical abnormalities in mTLE‐HS remains of primary importance. Currently, numerous studies employed single morphological features, such as volume, thickness, and surface area, to quantify cortical microstructural abnormalities in patients with mTLE‐HS, yet investigations into gray‐matter connectomic profiles at the network level remain scarce.

Morphometric similarity network (MSN) analysis integrates multiple morphological features to map cortical network organization, offering a promising method for recognizing neurological and psychological diseases as prototypical disorders of large‐scale network dysfunction. ⁵ Constructed at the individual level, MSN performs well in capturing subject‐specific variations, thereby preserving unique network signatures that are often obscured by conventional group‐level analyses. Li et al. observed a pattern of decreased fronto‐temporal MSN and increased insulo‐cingulato‐occipital MSN in patients with generalized epilepsy, ⁶ which was shown to have close associations with cortical lamination and cytoarchitecture. ⁵ , ⁷ In patients with TLE, Lu et al. reported increased MSN in the fronto‐parietal network and decreased MSN in the default mode network (DMN) ⁸ ; however, they neither quantified the contributive weights at the network level nor delineated abnormalities specific to HS, as their cohort included TLE patients with heterogeneous pathologies. In addition, their study did not include separate analyses of left and right TLE, thereby limiting assessment of hemispheric lateralization, especially given that voxel‐based morphometry evidence has shown more widespread abnormalities and greater contralateral involvement in left‐ compared to right‐sided mTLE‐HS patients. ⁹

Graph‐theoretical analysis models the brain as an interconnected network, characterized by high clustering within segregated neural communities and short paths between them. This architecture supports an optimal balance between network specialization and integration. Within this framework, neuroanatomic regions are represented as nodes and their interconnections as edges, enabling the computation of graph‐based metrics that quantitatively describe network topology and its pathological reorganization. In patients with mTLE‐HS, graph‐theoretical analyses based on probabilistic fiber density revealed increased limbic network clustering, as well as elevated nodal efficiency, degree, and clustering coefficient in the ipsilateral insula, superior temporal gyrus, and thalamus, indicating reorganization of the limbic system. ¹⁰ Another graph‐theoretical analysis derived from cortical thickness and subcortical volume showed increased clustering coefficients and path length in the orbitofrontal and temporal regions, suggesting a shift toward more regularized network topology in mTLE‐HS. ¹¹ However, graph‐theoretical analysis on MSN, which has recently demonstrated superior ability over those on single morphometric features in capturing individual‐level network reorganization in major depressive disorder, ¹² has not yet been applied to characterize the network topological reorganization in patients with mTLE‐HS.

Although neuroimaging offers a window into the large‐scale network reorganizations underlying mTLE‐HS, a critical next step is to explore its genetic and neurobiological substrates, particularly since HS is thought to have genetic underpinnings. ¹³ The Allen Human Brain Atlas (AHBA) helps bridge the gap between macro‐level connectome and micro‐level transcriptome by mapping abnormal network reorganization to spatial variations in gene expression and neuronal signatures based on postmortem whole‐brain sections. ¹⁴ , ¹⁵ It has been successfully used in schizophrenia, ⁷ major depressive disorder, ¹² and epilepsy, ¹¹ providing molecular insights into their pathophysiological mechanisms, and serving as a crucial complement for investigating brain regions beyond surgical resections. Cao et al. identified five distinct patterns of MSN alteration and their pattern‐specific associations with epilepsy‐related gene expression profiles in a cohort of heterogeneous drug‐resistant epilepsy. ¹⁶ In patients with TLE, MSN abnormalities have been linked to neuroinflammatory genes: PRR5, SMAD3, and IRF3. ⁸ Specific to mTLE‐HS, the genetic and cellular mechanisms underlying the vulnerable regions of MSN, particularly its topological properties, remain under investigation and represent a promising direction for elucidating brain network pathophysiology of mTLE‐HS.

In this study, we aimed to characterize cortical structural network disruptions in patients with unilateral mTLE‐HS by analyzing MSN topological properties and their genetic, transcriptomic, and neurodevelopmental underpinnings (Figure 1): first, to compare MSN topology among right‐sided mTLE‐HS, left‐sided mTLE‐HS, and healthy controls (HCs) for identifying syndrome‐specific disturbances; second, to develop an individual‐level machine‐learning classifier based on MSN topological features for patient identification and epileptogenic lateralization, and to quantify contributive weights of these features within the Yeo‐7 functional network; third, to integrate AHBA transcriptomic data for uncovering the spatial gene expression profiles associated with these topological reorganizations. We hypothesized that: (1) left‐ and right‐sided mTLE‐HS subtypes would exhibit distinct MSN topological signatures; (2) MSN topology‐derived machine‐learning models could accurately identify mTLE‐HS patients and lateralize their epileptogenic hemisphere, and pinpoint the most discriminative network at the individual level; and (3) spatial patterns of MSN topology reorganization would correlate with gene expressions enriched in specific biological process, cellular composition, and neurodevelopmental windows.

FIGURE 1.

Overview of the analysis pipeline. (A) Network topological analysis of the MSN. Individual MSNs were constructed based on five regional morphometric features (CT, SA, GMV, mean curvature, and Gaussian curvature) across 308 cortical regions defined by the Desikan–Killiany atlas. Structural networks were then established for each participant, with the 308 cortical regions serving as network nodes, and the morphometric correlations between each pair of regions serving as network edges. Network topological metrics were subsequently calculated for between‐group comparisons and individual‐level classification analysis, and further validated in an independent dataset. (B) Gene expression profiles. Gene expression data for 152 cortical regions in the left hemisphere were obtained from the Allen Human Brain Atlas (AHBA) dataset using the abagen toolbox. (C) Analysis of connectome–transcriptome association. Partial least squares (PLS) regression was conducted to examine spatial correlations between the MSN topological metrics and the transcriptional activity of all 15 633 genes. Enrichment analysis was then performed on genes associated with the first PLS component (PLS1), with absolute Z‐values exceeding 3.69. CT, cortical thickness; GMV, gray matter volume; HCs, healthy controls; L, left‐sided; MSN, morphometric similarity network; mTLE‐HS, mesial temporal lobe epilepsy with hippocampal sclerosis; R, right‐sided; SA, surface area.

2. MATERIALS AND METHODS

2.1. Subjects

This study was approved by the institutional review boards of West China Hospital of Sichuan University. Patients with unilateral mTLE‐HS were retrospectively recruited at our hospital from June 2017 to May 2023. Inclusion criteria: (1) diagnoses and lateralization of mTLE‐HS were established by a multidisciplinary epilepsy team, based on seizure semiology, EEG monitoring, and neuroimaging information; (2) preoperative high‐resolution 3.0T magnetic resonance imaging (MRI) scans were available. Exclusion criteria: (1) bilateral HS or presence of brain lesions other than HS on the MRI, (2) other serious neurological or psychiatric disorders, (3) history of alcohol or substance abuse, and (4) images with low quality and excessive head motion. The HCs matched for age and gender were recruited through advertisements.

2.2. MRI data acquisition and pre‐processing

MRI data were acquired using three 3.0T scanners (Siemens Trio Tim, Skyra and Prisma) with 32‐channel head coils. Detailed scanning parameters for high‐resolution T1‐weighted images are described in the Supplementary Materials. Data of each participant were pre‐processed using FreeSurfer (version 7.1.0, http://surfer.nmr.mgh.harvard.edu/). Individual cortical surface was reconstructed through skull stripping, brain tissue segmentation, separation of hemispheres and subcortical structures, and formation of the gray–white matter interfaces and pial surfaces. ¹⁷

2.3. Construction of MSN

First, a backtracking algorithm was employed to parcellate the individual reconstructed morphometric map into 308 spatially contiguous regions, based on the Desikan–Killiany atlas. ⁵ , ¹⁷ Regions were standardized to approximately equal sizes (~500 mm²) to minimize the potential influence of variability in parcel sizes. ⁵ Second, five regional morphometric features, including cortical thickness (CT), surface area (SA), gray matter volume (GMV), and mean and Gaussian curvatures, were extracted in native individual spaces. These morphometric feature vectors were Z‐score normalized across all cortical parcels within each participant. ⁷ , ¹⁷ Third, inter‐regional Pearson correlations were calculated between all cortical region pairs (308 × 308) based on the normalized morphometric features, generating individual‐level MSN matrices. The detailed procedures are illustrated in Figure 1A. Generally, larger MSN values indicate greater structural connectivity between two cortical regions, while values closer to zero indicate a weaker structural association.

2.4. Calculation of network topological metrics

Based on graph‐theoretic principles, the network was composed of 308 cortical regions as nodes and the morphometric correlations between each pair of regions as edges. GRETNA (version 2.1.0, https://www.nitrc.org/projects/gretna/) was used to calculate the network topological properties of MSN per participant, including: (1) three local metrics of nodes, that is, nodal betweenness, degree, and efficiency; and (2) five global metrics, that is, local efficiency, clustering coefficient, global efficiency, characteristic path length, and small‐world scalar. The range of sparsity (S) thresholds was set from .05 to .40 (interval = .01) to ensure that the thresholded networks meet small‐worldness criterion (σ > 1.0) and maintain an average node degree greater than 2 × log(N) (N = 308). Area under the curve (AUC) for each network metric was computed across this threshold range to provide a summarized description for the topological characteristics of brain networks.

2.5. Statistical analysis

Kolmogorov–Smirnov and Levene tests were used to assess the normality and homogeneity of variance of network metrics. Analysis of covariance (ANCOVA) was performed to examine group differences in MSN topological properties among right‐sided mTLE‐HS, left‐sided mTLE‐HS, and HCs, with age, gender, and scanning site and total brain volume (TBV) as covariates. Significant differences were followed by pairwise post hoc comparisons with Bonferroni correction (p < .05, two‐tailed). To assess the associations of MSN topological properties with clinical variables in mTLE‐HS patients, multiple linear regression models were employed, with age, sex, scanning site, and TBV included as covariates. The clinical variables included age at onset, epilepsy duration, early risk factors, seizure types, seizure frequency, seizure duration, and aura.

2.6. Individual‐level classification analysis

Two individual‐level classification tasks were performed to classify mTLE‐HS patients from HCs (Task 1), and to separate patients with right‐sided mTLE‐HS from patients with left‐sided mTLE‐HS (Task 2). Topological metrics derived from between‐group comparisons were used as features to train linear support vector machine (SVM) models. Model performance was evaluated using 10‐fold cross‐validation, with accuracy, sensitivity, specificity, and AUC. The absolute weights of each feature were extracted from the trained models and further mapped onto the Yeo 7 functional networks to assess their contributions at the network level. To validate the performance of Tasks 1 and 2, the trained SVM models were additionally applied to an independent dataset from the same center. Detailed descriptions of clinical characteristics and scanning parameters are provided in the Supplementary Materials.

2.7. Estimation of spatial gene expressions

The whole‐brain gene expression data were extracted from the AHBA, which provides postmortem transcriptomics information from six neurotypical adult donors. ¹⁵ Given the limited representation of the right hemisphere (two donors) in the AHBA relative to the left (six donors), we focused our analysis on the left hemisphere. Using the abagen toolbox (https://abagen.readthedocs.io/en/stable/), ⁵ the rigorous data processing pipeline included: probe re‐annotation, data filtering, probe selection, mapping sampling points to the brain regions, and normalization of expression values. Finally, a matrix (152 regions × 15 633 gene expression levels) of regional transcriptional level values was obtained.

2.8. Connectome–transcriptome association analyses

Partial least squares (PLS) regression was employed to assess the spatial correlation between the MSN topological connectome and the expression level of 15 633 genes. The t‐value maps across 152 cortical regions (derived from the mTLE‐HS vs HCs group comparisons) were set as the response variables and the normalized gene expression matrices as the predictor variables. PLS analyses generated latent components representing linear combinations of predictor variables that explained variance in the response variable. The first component (PLS1) was retained as it explained the greatest proportion of variance. ⁵ , ¹⁷ Permutation testing with 5000 rotations was performed for spatial autocorrelation correction, which tested the null hypothesis that the observed covariance between MSN topological metrics and gene expression patterns (explained by PLS1) was no greater than expected by chance. Further bootstrapping was conducted to estimate the variability of gene contributions to PLS1, with corrected weights calculated as each gene's raw weight divided by its bootstrap standard error. Gene rankings identified those with significant positive or negative contributions (PLS1±). Pearson correlations were also performed to confirm the regional alignment between PLS1 scores and t‐value maps.

2.9. Analysis of epilepsy‐related risk genes

To assess the contribution of epilepsy‐related risk genes to MSN topological reorganizations, we sourced a set of 119 candidate genes from previous genome‐wide association study (GWAS; meta−/mega‐analyses), whole‐exome sequencing studies, and systematic reviews. ⁶ The overlap between this gene set and 15 633 AHBA background genes was identified, and their Pearson correlations with the MSN topological alterations were then calculated with significance defined as false discovery rate (FDR) corrected p < .05. To account for spatial dependencies, significance was further validated using a spatial permutation test that preserves the intrinsic spatial autocorrelation of MSN metrics (5000 iterations) with the threshold at Pperm <5th or >95th centile. Subgroup analyses were conducted to assess potential lateralized effects. Specifically, MSN topological alteration maps derived from left‐ and right‐sided mTLE‐HS vs HCs were then correlated with epilepsy‐related risk gene expression using identical statistical procedures.

2.10. Enrichment analyses

Genes with absolute Z‐scores >3.69 or Z‐scores <−3.69 in PLS1 were analyzed using online bioinformatics platform (https://www.bioinformatics.com.cn) to identify their associated biological pathways. The significance of enrichments was determined by FDR correction (p < .05). To map significant genes to cell types, we compared PLS1 genes with cortical single‐cell RNA‐seq signatures (astrocytes, microglia, endothelial cells, oligodendrocyte precursor cells, oligodendrocytes, excitatory neurons, and inhibitory neurons) derived from five independent human samples, minimizing the potential biases of acquisition methods, analysis, or thresholding. ⁵ , ⁶ For each cell type, the p‐value of overlapping genes was calculated via permutation testing and compared against null distributions generated from 5000 random gene sets of matched size. Statistical significance was set as Pperm <5th or >95th centile, testing the null hypothesis that the obtained overlaps did not exceed chance expectations. To further investigate developmental trajectories, we cross‐referenced the PLS1 gene sets with the Brainspan developmental transcriptome atlas using the cell type–specific expression analysis (CSEA) tool (http://doughertytools.wustl.edu/CSEAtool.html). The CSEA evaluates the neurodevelopmental relevance of a gene set by examining its enrichment in the formation and differentiation of major brain regions (cortex, hippocampus, striatum, thalamus, amygdala, and cerebellum) across developmental stages from early fetal life to young adulthood. Statistical significance was determined using Fisher's exact test with FDR correction.

3. RESULTS

3.1. Demographic and clinical data

This study included 109 patients with unilateral mTLE‐HS. Among them, 64 cases were left‐sided mTLE‐HS, and 45 were right‐sided mTLE‐HS. The left‐sided mTLE‐HS group had a mean age of 28.4 years (standard deviation [SD] = 9.5), with 32 males (50%), and an average epilepsy duration of 10.5 years (SD = 9.2). The right‐sided mTLE‐HS group had a mean age of 30.6 years (SD = 11.1), with 23 males (51.1%), and an average epilepsy duration of 8.1 years (SD = 6.8). Ninety HCs were recruited, with a mean age of 31.4 years (SD = 9.3) and 48 males (53.3%). Age and sex did not differ significantly between the patient and HC groups, and clinical characteristics did not differ significantly between the left and right patient groups (Table 1).

TABLE 1.

Clinical and demographic characteristics of subjects.

	L‐mTLE‐HS (N = 64)	R‐mTLE‐HS (N = 45)	HCs (N = 90)	Statistics F, t, χ 2	p values
Sex (male/female)	32/32	23/22	48/42	.18	.91
Age, y	28.4 ± 9.5	30.6 ± 11.1	31.4 ± 9.3	1.84	.16
Age at onset, y	17.7 ± 11.4	21.6 ± 10.2		−1.79	.10
Epilepsy duration, y	10.5 ± 9.2	8.1 ± 6.8		1.53	.12
History of risk factors				4.71	.45
Head trauma	8 (13%)	3 (7%)
Febrile convulsions	11 (17%)	9 (20%)
Intracranial infection	9 (14%)	2 (4%)
Adverse perinatal events1	5 (8%)	3 (7%)
No risk factors	9 (14%)	10 (22%)
Missing	22 (34%)	18 (40%)
Type of seizures				1.91	.59
Focal2	22 (34%)	13 (29%)
FBTCS	12 (19%)	13 (29%)
BTCS/Unknown3	17 (27%)	12 (27%)
Missing	13 (20%)	7 (16%)
Seizure frequency, per mo				3.73	.16
≤1	15 (23%)	10 (22%)
>1	29 (45%)	26 (58%)
Missing	20 (31%)	9 (20%)
Average seizure duration, s				.66	.72
≤60	17 (27%)	15 (33%)
>60	24 (38%)	14 (31%)
Missing	23 (36%)	16 (36%)
Seizure aura				4.31	.12
Present	26 (41%)	17 (38%)
Absent	22 (34%)	22 (49%)
Missing	16 (25%)	6 (13%)

Note: Adverse perinatal events1: includes difficult labor, preterm birth, birth asphyxia, and an APGAR score <7 at birth; Focal2: includes both focal preserved aware and focal impaired awareness seizures; BTCS/Unknown3: unknown whether focal or generalized‐bilateral tonic–clonic seizure.

Abbreviations: BTCS, bilateral tonic–clonic seizures; FBTC, focal‐to‐bilateral tonic–clonic seizure; HCs, healthy controls; L, left‐sided; mo, month; mTLE‐HS, mesial temporal lobe epilepsy with hippocampal sclerosis; s, second; R, right‐sided; y, years.

3.2. Between‐group comparisons in MSN topological properties

3.2.1. Group differences in global topological properties

ANCOVA revealed significant differences in small‐world scalar and local efficiency among groups, but not in clustering coefficient, characteristic path length, and global efficiency (Table S2). Compared to HCs, small‐world scalar was significantly elevated in left‐sided mTLE‐HS patients, with an increasing trend in right‐sided mTLE‐HS patients that did not reach statistical significance. No significant difference in small‐world scalar was found in direct comparison between patient groups. Regarding local efficiency, left‐sided mTLE‐HS patients exhibited an increasing trend, whereas right‐sided mTLE‐HS patients presented a decreasing trend relative to HCs. Thus, direct comparison demonstrated significant elevation of local efficiency in left‐sided mTLE‐HS patients compared to right‐sided mTLE‐HS patients.

3.2.2. Group differences in local topological properties

ANCOVA revealed significant differences in all three local topological metrics—nodal betweenness, degree, and efficiency—among groups (Figure 2A, Tables S3–S5). Post hoc comparison revealed lateralized differences in nodal network properties between left‐ and right‐sided mTLE‐HS, primarily involving temporo‐occipital regions and the orbitofrontal cortex (OFC). Specifically, compared with right‐sided mTLE‐HS, left‐sided patients exhibited decreased nodal betweenness in the left middle temporal cortex, right cuneus, and right medial OFC, as well as reduced nodal degree and efficiency in the left lateral occipital cortex. In contrast, increased nodal efficiency was observed in the left superior temporal cortex and right lateral OFC.

FIGURE 2.

Topological properties of MSNs across L‐mTLE‐HS, R‐mTLE‐HS, and HCs groups. (A) Statistical parametric maps (F‐maps) showing brain regions with a significant main effect of group on nodal properties (betweenness, degree, and efficiency), based on ANCOVA. (B–D) Results of post hoc pairwise comparisons for the same nodal properties: (B) L‐mTLE‐HS vs R‐mTLE‐HS, (C) L‐mTLE‐HS vs HCs, and (D) R‐mTLE‐HS vs HCs. In B–D, green, orange, and purple spheres denote non‐significant regions, regions with a significant decrease, and regions with a significant increase, respectively. MSN, morphometric similarity network; ANCOVA, analysis of covariance; HCs, healthy controls; L, left‐sided; R, right‐sided; mTLE‐HS, medial temporal lobe epilepsy with hippocampal sclerosis.

When compared to HCs, both patient groups exhibited bilateral disturbances in nodal network properties. Nodal betweenness and degree showed more significant and widespread abnormalities in left‐sided patients, whereas nodal efficiency was characterized by an ipsilateral predominance (Figure 2C,D and Tables S3–S5). Specifically, left‐sided mTLE‐HS patients demonstrated (1) decreased nodal betweenness in the right medial OFC; (2) decreased nodal degree in the occipital cortex (left lateral occipital cortex and bilateral lingual gyri), alongside increases in limbic and DMN regions (left precuneus and right medial OFC) and the right postcentral gyrus; and (3) decreased nodal efficiency in the occipital cortex (bilateral lingual gyri and left lateral occipital cortex), concurrent with increases in limbic/DMN regions (left rostral anterior cingulate cortex [ACC] and left precuneus) and the right postcentral gyrus. Right‐sided mTLE‐HS patients demonstrated (1) no significant differences in nodal betweenness; (2) decreased nodal degree in the occipital cortex (right lingual gyrus) and increases in limbic/DMN regions (left rostral ACC and right medial OFC) as well as the right postcentral gyrus; and (3) decreased nodal efficiency in the right lateral OFC and the temporo‐occipital cortex (left superior temporal cortex, right middle temporal cortex, and right lingual gyrus), together with increases in the limbic/DMN regions (left rostral ACC and right medial OFC). No significant correlations were observed between these altered local topological metrics and clinical characteristics in both main and validation datasets (Tables S6 and S7).

3.3. Individual‐level classification performance of MSN topological properties

In establishing and evaluating the classification Task 1 for distinguishing patients with unilateral mTLE‐HS from HCs, we observed classification AUC of .86 with accuracy of 77.4%, sensitivity of 82.6%, and specificity of 71.1% (Figure 3A,B). For classification Task 2, which separated left‐sided from right‐sided mTLE‐HS patients, the AUC, accuracy, sensitivity, and specificity were .88, 83.2%, 77.8%, and 87.5%, respectively (Figure 3D,E). Feature importance analysis revealed the MSN topological properties within the limbic network as the top contributor to classification performance, accounting for 32.0% in Task 1 and 41.1% in Task 2 (Figure 3C,F). To validate these results, an independent dataset comprising 35 unilateral mTLE‐HS patients (18 left‐sided and 17 right‐sided) and 30 HCs was analyzed. In this dataset, Task 1 achieved an AUC of .81, with an accuracy of 75.4%, sensitivity of 80.0%, and specificity of 70.0% (Figure 3A,B). For Task 2, the AUC, accuracy, sensitivity, and specificity were .84, 77.1%, 64.7%, and 88.9%, respectively (Figure 3D,E).

FIGURE 3.

Classification performance for L‐mTLE‐HS, R‐mTLE‐HS, and HC groups. Classifier 1 discriminated patients with mTLE‐HS from HCs, whereas Classifier 2 discriminated L‐mTLE‐HS from R‐mTLE‐HS. (A, D) Confusion matrices for the main and validation analyses, (B, E) Receiver operating characteristic (ROC) curves with AUC for the main and validation analyses, (C, F) Spatial distribution of feature weights for each classifier, with radar plots summarizing the weight distributions across the seven canonical Yeo networks. AUC, area under the curve; DAN, dorsal attention network; DMN, default mode network; FPN, frontoparietal network; HCs, healthy controls; L, left‐sided; LIM, limbic network; mTLE‐HS, mesial temporal lobe epilepsy with hippocampal sclerosis; R, right‐sided; SMN, somatomotor network; VAN, ventral attention network; VIS, visual network.

3.4. Gene expression patterns co‐locate with MSN topological differences

The first PLS component accounted for 12.07%, 17.39%, and 17.22% of the variance in nodal betweenness, degree, and efficiency alterations, respectively, in patients with unilateral mTLE‐HS (Figure 4A). Spatial correlation analyses revealed significant associations between PLS1 score maps and case–control t‐maps of three MSN local topological properties: betweenness (r = .35, p < .001), degree (r = .42, p < .001), and efficiency (r = .41, p < .001) (Figure 4B). We found 449 genes with significantly positive PLS1 weights (Z > 3.69, PLS1+ gene sets) and 885 genes with significantly negative PLS1 weights (Z < −3.69, PLS1− gene sets) (all p < .05, FDR corrected).

FIGURE 4.

Co‐localization of gene expression patterns with MSN topological abnormalities in mTLE‐HS. (A) Spatial distribution of partial least squares (PLS) component scores (PLS1) projected onto brain maps. (B) Regional correlations between PLS1 scores and mTLE‐HS‐associated connectomic changes. Each data point represents a distinct brain region. (C) The spatial co‐localization between the expression patterns of epilepsy‐related risk genes and the topological alterations of MSNs in mTLE‐HS. mTLE‐HS, mesial temporal lobe epilepsy with hippocampal sclerosis. All p‐values were derived from permutation tests.

Among the 119 epilepsy‐related risk genes, 108 overlapped with the background genes from AHBA. Four genes showed spatial correlations with degree alterations: DNM1 (r = −.31, Pperm = .01), GOLIM4 (r = .31, Pperm = .007), GRIN1 (r = −.26, Pperm = .04), and PPFIA3 (r = −.26, Pperm = .01). Two genes showed spatial correlations with efficiency alterations: DNM1 (r = −.31, Pperm = .01) and PPFIA3 (r = −.26, Pperm = .01) (Figure 4C). Subgroup analyses showed largely consistent spatial associations in left‐ and right‐sided mTLE‐HS. Nodal degree alterations were correlated with DNM1, GRIN1, PPFIA3, and GOLIM4 in both subgroups, and nodal efficiency alterations were associated with DNM1. In addition, a significant association between nodal efficiency and PPFIA3 was observed in left‐sided mTLE‐HS but not in right‐sided mTLE‐HS (Figure S1).

3.5. Transcriptional enrichments for differences in topological properties

To clarify the biological significance of PLS1± gene sets, we conducted the functional enrichment analysis. The PLS1+ genes were significantly enriched in biological processes related to RNA processing, including RNA splicing and non‐coding RNA (ncRNA) metabolic process (Figure 5A). The PLS1− genes demonstrated significant pathway enrichment in mitochondrial energy metabolism, including oxidative phosphorylation, ATP synthesis coupled electron transport, respiratory electron transport chain, and cellular respiration (Figure 5B).

FIGURE 5.

Functional enrichment analysis and network visualization. (A, B) Biological processes enrichment for PLS1+ and PLS1− genes. The chord diagrams show the relationship between enriched pathways and genes. Node size and color represent gene count and pathway category, respectively. (C, D) Cellular specificity and developmental trajectories of both PLS1+ and PLS1− genes. Astro, astrocyte; Endo, endothelial cell; Micro, microglia; Neuro‐Ex, excitatory neuron; Neuro‐In, inhibitory neuron; Oligo, oligodendrocyte; OPC, oligodendrocyte precursor cell.

To further link regional MSN topological abnormalities to cellular composition, we performed cell‐type enrichment analysis based on gene distributions. The significant PLS1± genes were assigned to seven canonical cell classes; both PLS1+ and PLS1− genes were significantly related to astrocytes (PLS1+: 21 genes, Pperm = .002; PLS1−: 15 genes, Pperm = .004), oligodendrocytes (PLS1+: 26 genes, Pperm <.001; PLS1−: 16 genes, Pperm <.001), and excitatory neurons (PLS1+: 26 genes, Pperm <.001; PLS1−: 110 genes, Pperm <.001) (Figure 5C).

To uncover critical spatiotemporal windows of gene expression associated with MSN topological abnormalities, we performed developmental gene set enrichment analysis. Our results showed that PLS1+ genes were predominantly enriched in the striatum during the early fetal stage, whereas PLS1− genes displayed double‐peaked activation in the cortex across neonatal early infancy and mid‐to‐late childhood (Figure 5D).

4. DISCUSSION

This is the first study to incorporate multiple cortical morphological features to characterize the topological reorganizations of cytoarchitectonically similar networks in normal‐appearing cortex of patients with unilateral mTLE‐HS. We found global small‐worldness increase with regional nodal topological hyper‐connection in limbic‐DMN and hypo‐connection in temporo‐occipital cortex, showing more pronounced abnormalities in the ipsilateral hemisphere and in patients with left‐sided mTLE‐HS. At the individual level, MSN topological metrics could train machine‐learning models to distinguish unilateral mTLE‐HS patients from HCs, lateralize the epileptogenic hemisphere, and highlight the limbic system as their most critical network component. Transcriptomic analysis linked these MSN connectomic disturbances to gene expression patterns enriched in the biological pathways of RNA processing and mitochondrial energy metabolism, the cellular dysfunction of excitatory neuron and glial cells, and the neurodevelopment trajectory of early‐fetal striatum and neonatal‐to‐childhood cortex. Our findings offered novel integrative neurodevelopmental connectomic–transcriptomic insights into the neurobiological mechanisms underlying unilateral mTLE‐HS and its epileptogenic lateralization.

4.1. MSN topological reorganizations in patients with unilateral mTLE‐HS

Moving beyond isolated anatomic or morphometric measures like cortical thickness, curvature, or volume, MSNs synthesize diverse cortical features within each participant to capture multidimensional patterns of cortical architecture. ⁵ , ¹⁷ , ¹⁸ High morphological similarity between cortical regions is thought to mirror shared cytoarchitectonic organization, providing an anatomic basis for their increased likelihood of axonal connectivity. ¹⁷ When combined with graph‐theoretical analysis, network metrics of MSNs extend beyond pairwise similarity to capture higher‐order topological properties, thereby depicting network segregation and integration across both local and global scales. ¹⁹ Our analysis revealed significant global and local MSN topological reorganizations in patients with unilateral mTLE‐HS patients and between their subtypes, supporting the notion that mTLE‐HS involves impairments extending beyond the epileptogenic hippocampus to include large‐scale brain networks. ³ , ⁴ These aberrant MSN topological features showed no significant correlations with seizure‐related characteristics and early epilepsy risk factors, which suggests that these detected network abnormalities are largely independent of disease burden and may reflect inherent cortical reorganization underlying mTLE‐HS pathophysiology.

At the global level, we found an increased small‐world scalar of MSN in unilateral mTLE‐HS patients—a network alteration previously reported in other focal epilepsy cohort with good seizure control and implicated in non‐seizure symptoms such as cognitive impairment. ²⁰ This hyper‐connected topology was more pronounced in patients with left‐sided mTLE‐HS than in those with right‐sided mTLE‐HS, findings that aligned with previous reports of greater cortical thinning, ²¹ more widespread white matter impairment,²² and marker cognitive deficits ²³ in left‐sided mTLE‐HS patients, and further corroborated by our MSN nodal results of betweenness and degree. At the local level, reduced nodal degree and efficiency were observed in temporo‐occipital regions alongside with decreased betweenness in OFC. These regions have been reported to exhibit particular susceptibility to seizure‐related damage in mTLE, showing significant cortical thinning, ²⁴ reduced amplitude of low‐frequency fluctuations, ²⁵ and decreased effective EEG connectivity ²⁶ in previous studies. Our findings indicated reduced connection number, weakened ability to bridge information transfer, and inefficiency of information processing in these vulnerable regions, providing cortical topological substrates for cognitive deficits in patients with mTLE. On the contrary, hyper‐connectivity and increased transmission efficiency of MSN topology were found in cortical regions of DMN in our study. In line with it, a high‐density EEG study using graph‐theoretical analysis demonstrated significantly increased efficiency in the DMN during the interictal period among patients with unilateral mTLE‐HS. They proposed these hyper‐connectivities to represent epilepsy‐related hyperexcitable pathologic activity and an expansion of epileptic network, which may facilitate the seizure propagation. ²⁷ In addition, increased DMN efficiency may also reflect compensatory reorganization: hippocampal dysfunction in mTLE may drive reliance on some DMN structures, such as the precuneus—a key hub for episodic memory retrieval—to help sustain cognitive processes. ²⁸ McCormick et al. validated that elevated resting‐state functional connectivity of DMN indicated the capacity of episodic memory in mTLE‐HS patients and correlated with the postoperative memory retention. ²⁹ Our findings provide direct structural connectomic evidence for the cortical network reorganizations in patients with mTLE with unilateral HS; however, further studies are warranted to elucidate its implications for cognition and seizure propagation.

4.2. Individual discrimination and lateralization for patients with unilateral mTLE‐HS

The ultimate goal of advanced imaging technique is to achieve translation from group‐level differences to individual‐level prediction that informs clinical practice. Employing MSN topological features derived exclusively from cortex, our models achieved 77.4% accuracy in identifying patients with mTLE‐HS and 83.2% accuracy in lateralizing the epileptogenic hemisphere at the individual level. This performance was comparable to the classifiers trained by hippocampal features, which reported accuracies of 80.5% for identification and 88.0% for lateralization. ³⁰ , ³¹ , ³² With reporting accuracies of 64.5%–79.5%, another multimodal machine‐learning study corroborated that the normal‐appearing cortex, irrespective of the hippocampus, carried enough information to reliably predict both mTLE‐HS diagnosis and lateralization. ³³ A previous study has described MSN disturbances in patients with TLE, ⁸ and our findings extend this work by employing machine‐learning methods on an independent validation dataset to demonstrate the superior predictive ability of topological features derived from the MSN. It is consistent with a study on major depressive disorder that MSN topological features showed a greater contribution to identifying suicidal subtypes than MSN morphological features. ¹² Taken together, cortical MSN connectome might serve as a promising marker for mTLE patients' diagnosis and subtyping, further supporting the conceptualization of epilepsy as a network disorder and calling for more network‐level investigations.

Moreover, our machine‐learning models provided a data‐driven assessment of feature weights at the network level, with MSN topological abnormalities in the limbic network pinpointed as the most significant contributor to both patient identification and lateralization of epileptogenic hemisphere. It is reasonable that the limbic structures are anatomically and functionally proximal to the epileptogenic hippocampus and highly susceptible to epileptiform activity. ³⁴ Previous studies have also revealed limbic connectivity reorganization in patients with mTLE, especially those with HS, implying its involvement in the pathophysiology of seizure generation and maintenance. ¹⁰ , ³⁵ Graph‐theoretical analysis of functional connectivity revealed distinct, network‐specific alteration patterns between left and right TLE in the DMN, attention, limbic, and visual networks, where the limbic network yielded the highest accuracy for determining seizure laterality. ³⁵ Moreover, the limbic network plays a central role in memory regulation—functions that are often impaired in mTLE. ³⁶ Hemispheric lateralization of memory is well established, as patients with left HS typically exhibit more pronounced verbal memory deficits, whereas those with right HS tend to show greater impairment in nonverbal memory functions. ³⁷ A resting state functional MRI (fMRI) study has revealed different neural reorganization patterns of anterior and posterior hippocampal networks associated with verbal memory in left and right mTLE‐HS patients, ³⁸ implying a potential link for the hemispheric lateralization between their cognitive deficits and limbic network disturbance.

4.3. MSN connectome‐related gene expression and transcriptional enrichment

Spatial correlations between MSN topological reorganization and the expression of the genes (e.g., DNM1, PPFIA3, and GRIN1) indicate that distinct molecular mechanisms collectively shape cortical network remodeling in mTLE with unilateral HS. Among these genes, DNM1 and PPFIA3 serve as key regulators of synaptic vesicle trafficking and inhibitory signaling. ³⁹ , ⁴⁰ A large‐scale exome sequencing study (n = 1131) identified DNM1 as a novel epilepsy‐associated gene. ³⁹ DNM1 encodes dynamin‐1, a GTPase essential for synaptic vesicle endocytosis; pathogenic variants impair vesicle recycling and selectively weaken γ‐aminobutyric acid (GABA)ergic inhibitory transmission, driving network disinhibition and hyperexcitability. PPFIA3 variants, recently linked to neurodevelopmental disorders with epilepsy phenotype, disrupt inhibitory synapse formation and presynaptic organization, further exacerbating the imbalance between excitation and inhibition. ⁴⁰ GRIN1, encoding the obligatory GluN1 subunit of the N‐methyl‐d‐aspartate (NMDA) receptor, governs glutamatergic neurotransmission, and its dysregulation promotes excitotoxicity and neuronal hyperexcitability. ⁴¹ Collectively, these gene‐driven processes disrupt the excitation–inhibition balance, predisposing cortical networks to seizure generation and propagation.

Transcriptional analyses revealed that these genes were spatially enriched in biological processes related to RNA processing and mitochondrial energy metabolism. Dysregulation of RNA processing, particularly ncRNA metabolic pathways, has been implicated in TLE through the modulation of GABAergic and glutamatergic neurotransmission, thereby affecting synaptic transport and neuronal excitability. ⁴² Mitochondrial oxidative phosphorylation sustains vesicular trafficking and neurotransmitter release at excitatory synapses; its impairment disrupts synaptic energy balance and promotes hyperexcitability, as evidenced by reduced oxidative phosphorylation and impaired nicotinamide adenine dinucleotide (phosphate)‐related metabolic responses in hippocampal tissue from TLE patients with hippocampal sclerosis. ⁴³ , ⁴⁴ , ⁴⁵ In line with these findings, our cell‐type enrichment analysis further assigned these MSN topology–related genes to glutamatergic excitatory neurons, which are distributed widely across the hippocampus and cortex and are characteristically involved in mTLE‐HS, where hyper‐glutamatergic activity driven by altered kynurenine pathway metabolism has been demonstrated. ⁴⁶

At the level of cellular composition, astrocytes and oligodendrocytes showed significant enrichment in our study. Astrocytes are involved directly in multiple glutamate‐related processes, including glutamate synthesis, reuptake of released glutamate, and clearance of excess extracellular glutamate. Excessive glutamate contributes to neuronal hyperexcitability, which underlies epileptic seizure generation. ⁴⁷ Transcriptome‐wide network analyses have revealed dysregulations in neuronal support and myelination pathways in mechanistic target of rapamycin‐related epilepsies and mTLE‐HS, highlighting oligodendrocyte impairment as a molecular hallmark of drug‐resistant epilepsy and a potential avenue for upstream therapeutic targeting. ⁴⁸

Our trajectory analysis further revealed dynamic stage‐specific enrichment of the MSN topology–associated gene expression along neurodevelopment, supporting the mTLE‐HS as a progressive disorder. The fetal striatal peak of PLS1+ gene expression may represent the prenatal initiation of brain vulnerability to the inhibitory/excitatory imbalance and subsequent seizures. ⁴⁹ The cortical enrichment of PLS1− genes throughout postnatal development—spanning infancy to late childhood—overlaps with the typical age window for febrile seizures and epilepsy onset in patients with mTLE‐HS, ⁴⁵ implying its complex interaction with rapid brain maturational processes to reshape and consolidate the vulnerability of epileptic brain. These findings provide new evidence for the “two‐hit” hypothesis in TLE, ⁵⁰ with the prenatal striatal enrichment constituting the initial predisposition (“first hit”) and the postnatal cortical enrichment driving the subsequent pathology (“second hit”).

4.4. Limitations

First, we recruited patients from a tertiary epilepsy center and thus represent a cohort with heavier seizure burden and greater number of anti‐seizure medications. Caution must be taken in the generalizability of our findings to the broader unilateral mTLE‐HS population. Second, clinical data were extracted from historical medical records, with the assessment of seizure‐related variables confined to the 1 mo peri‐scan period. Due to the retrospective design of this study, several seizure‐related variables, including risk factor history, seizure type, seizure frequency, average seizure duration, and aura, were unavailable. To mitigate this limitation, correlation analyses were replicated in the independent validation cohort with complete clinical data; this procedure confirmed the initial findings. Third, our transcriptomic analyses were restricted to the left hemisphere due to the limited right‐hemisphere sampling (only two donors) in the AHBA. However, previous studies have well established that the dataset from the left hemisphere provides a sufficiently representative high‐resolution gene expression atlas for transcriptomic mapping. ⁵ , ⁶ Fourth, our genetic and transcriptomic analysis was based on a normative atlas (AHBA), rendering the spatial co‐localization between imaging phenotypes and molecular substrates theoretical and inferential. Further studies incorporating histopathological analysis of surgical brain specimens are necessary to clarify the actual underpinnings of the MSN topological reorganizations.

5. CONCLUSION

In conclusion, MSN topology delineates the pattern of cortical network reorganization in patients with unilateral mTLE‐HS. Machine‐learning models corroborate the ability of MSN topology to accurately identify patients and lateralize epileptogenic hemisphere at the individual level. These findings underscore the diagnostic and pathophysiological significance of cortical MSN topology beyond the sclerotic hippocampus, likely reflecting macro‐scale network remodeling underlying the epileptogenesis and cognitive comorbidity of unilateral mTLE‐HS. We further link the MSN topological disturbances to epilepsy risk genes across spatial co‐localization, biological pathway, cellular composition, and neurodevelopment trajectory. This multi‐scale convergence points to an RNA–metabolic interplay that shapes the cortical vulnerability of mTLE‐HS in a progressive way.

AUTHOR CONTRIBUTIONS

Fei Zhu: Conceptualization; formal analysis; investigation; resources; methodology; and writing‐review and editing. Shanshan He: Formal analysis and investigation. Bo Tao: Conceptualization; investigation; formal analysis; methodology; writing‐original draft; and writing‐review and editing. Yue Li: Investigation. Yi Liang: Resources; formal analysis; and investigation. Chunge Pan: Resources and investigation. Ziyang Gao: Resources and investigation. Menglian Wu: Investigation and methodology. Dong Zhou: Conceptualization and writing‐review and editing. Peilin Lu: Conceptualization and writing‐review and editing. Yingying Tang: Conceptualization; investigation; resources; and writing‐review and editing. Su Lui: Conceptualization; project administration; investigation; methodology; and writing‐review and editing.

FUNDING INFORMATION

This study was supported by the National Natural Science Foundation of China (Project Nos. 82471484, 82441007, 82120108014); the ST12030‐Major Projects (Project No. 2021ZD0201904); the Sichuan Science and Technology program (Project No. 2024NSFTD0043); the Natural Science Foundation of Sichuan Province (Project No. 2024NSFSC1649, 2025ZNSFSC1773); 1.3.5 project for disciplines of excellence, West China Hospital, Sichuan University (Project No. ZYGD23003); and Chengdu Science and Technology Office, major technology application demonstration project (Project No. 2022‐GH03‐00017‐HZ).

CONFLICT OF INTEREST STATEMENT

None of the authors has any conflict of interest to disclose.

ETHICS STATEMENT

This study was approved by the ethics committee on biomedical research of West China Hospital, Sichuan University. Written informed consent was obtained from all participants. 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.

Supporting information

Table S1.

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the corresponding author, upon reasonable request.