Neural Excitation–Inhibition Imbalance in Cervical Spondylotic Myelopathy

Abstract

Background

Cervical spondylotic myelopathy (CSM) is a chronic spinal cord compression disorder associated with degenerative changes in the cervical spine. Although morphological alterations in brain regions associated with CSM have been reported, the underlying molecular regulatory mechanisms remain incompletely understood. This study aims to reveal the structural changes among different brain regions in CSM and their complex associations with specific transcriptional expression patterns.

Methods

The study includes 170 CSM patients and 79 healthy controls(age range: 40-65 years). MRI is used to acquire cortical data for each participant, and the morphometric similarity network is then constructed for each individual. A general linear model is applied to analyze cortical structural differences between patients and healthy controls. Subsequently, partial least squares regression is used to link these cortical structural differences with gene expression profiles from the AHBA, in order to identify associated genes for enrichment analysis.

Results

We find that, compared with healthy controls, CSM patients show generally increased MSN strength in the cerebral cortex. Additionally,the differentially expressed genes are specifically enriched in both excitatory and inhibitory neurons.

Conclusions

Our study reveals that structural network hubs are more susceptible to damage under chronic spinal cord compression, and that disruption of the dynamic balance between excitatory and inhibitory neurons may serve as a candidate mechanism driving secondary cortical remodeling.

Plain Language Summary

Using brain morphometric network analysis, this study finds enhanced interregional connectivity in patients with cervical spondylotic myelopathy. By integrating human brain gene expression data, it further reveals that these structural changes are associated with specific gene expression, with these genes enriched in particular neuron types, suggesting an imbalance between excitatory and inhibitory neurons. This research provides new insights into the molecular mechanisms underlying brain alterations in CSM.

Wu, Zhu et al investigate enhanced interregional connectivity in patients with cervical spondylotic myelopathy (CSM) using brain morphometric network analysis. Findings reveal that these structural changes are associated with specific gene expression suggesting an imbalance between excitatory and inhibitory neurons in CSM.

Introduction

Cervical spondylotic myelopathy (CSM) is a chronic spinal cord compression disorder associated with degenerative changes in the cervical spine, and frequently leads to sensory, motor, and autonomic dysfunction^1,2. However, there are also patients with spinal cord compression who do not present with obvious symptoms in clinical practice³, suggesting that the functional impairment of CSM may originate not only from local lesions, but also involves remodeling and compensatory mechanisms within the central nervous system⁴. As the advanced center for the integration and regulation of peripheral signals⁵, the brain’s structural and functional connectivity responses to chronic spinal cord injury require further investigation. Importantly, chronic sensorimotor deafferentation can induce neural plasticity through multiple molecular and cellular pathways, including synaptic remodeling, neuroinflammation, and myelination-related processes^6,7. Histological evidence from spinal cord injury and related deafferentation models suggests that these micro-level changes can reshape cortical circuits via synaptic reorganization, disrupted excitation–inhibition balance, glial-mediated modulation, and altered myelin/axonal integrity^8,9. At the macro level, such adaptations may manifest as alterations in cortical morphology.

Previous studies have reported structural abnormalities in the brains of CSM patients; however, most of this work has focused on single morphological metrics, such as cortical volume¹⁰, and have failed to capture the joint alterations among multiple structural parameters, including cortical thickness, surface area, and sulcal depth¹¹.This one-sided perspective may lead to inconsistent results and makes it difficult to reveal the overall remodeling characteristics of complex brain functional modules¹².Morphometric similarity network (MSN), as an integrative approach for brain morphological analysis¹³, enables the integration of multiple structural parameters to construct similarity patterns of macroscopic organization between brain regions, and has been widely applied in the study of psychiatric and neurological disorders.^14–16.

On the other hand, the establishment of human brain gene expression atlases in recent years, such as the Allen Human Brain Atlas (AHBA), has made it possible to integrate neuroimaging data with transcriptomic data^17,18.Multiple studies have shown that macroscopic changes in cortical structure are highly correlated with their local gene expression profiles. This cross-scale integrative strategy offers extended perspectives for uncovering the molecular mechanisms underlying brain structural abnormalities¹⁴. The integration of neuroimaging multimodal data with gene transcription data has proven effective in uncovering the micro-level changes underlying various cortical structural anomalies. By linking morphometric alterations in the cortex of CSM patients to gene expression data, our aim is to identify the complex processes underlying structural changes and provide extended avenues for improving sensory and motor function recovery in these patients.We hypothesize that MSN alterations is associated with CSM, preferentially affecting structural network hubs, and that the spatial pattern of these changes is associated with baseline regional gene-expression gradients derived from the AHBA.

In this study, we find significant structural alterations in the cerebral cortex of patients with cervical spondylotic myelopathy. By integrating whole-brain transcriptomic data, we find genes related to brain structural remodeling in patients are associated with excitatory and inhibitory neurons.

Methods

Participants

A total of 170 patients with moderate CSM and 79 age-matched healthy controls (HCs) were included in this study. CSM patients were recruited from the First Affiliated Hospital of Anhui Medical University in Hefei, Anhui Province, China, between January 2022 and January 2025, while HCs were recruited from the local Shushan community. Inclusion criteria for CSM patients were as follows:①Age between 40 and 65 years;②presentation with typical clinical symptoms and signs of myelopathy;③a modified Japanese Orthopaedic Association (mJOA) score between 12 and 14¹⁹;④MRI-confirmed spinal cord compression;⑤No history of cervical spinal injury or surgery;⑥No evidence of intracranial organic lesions or cognitive impairment. Exclusion criteria included:①Alcohol or substance abuse, smoking;②Presence of severe systemic diseases, such as hypertension, diabetes, heart disease, or other chronic medical conditions;③Peripheral neuropathy;④MRI contraindications: claustrophobia, pregnancy, or presence of metal implants (e.g., cardiac pacemakers, metallic dentures, etc.).

HCs were recruited from the local community and underwent clinical neurological and imaging examinations to confirm the absence of myelopathy-related symptoms or signs, while also excluding individuals with spinal canal stenosis or spinal cord compression.All participants were assessed by at least two experienced clinicians.

Data Acquisition

Brain images of all participants were acquired using a 3.0-T MRI system (Discovery MR750w, General Electric, Milwaukee, WI, USA) equipped with a 24-channel head coil. To minimize scanning noise and head motion artifacts, participants wore earplugs and foam pads during the scanning process. They were instructed to keep their eyes closed and remain awake throughout the procedure. High-resolution 3D T1-weighted structural images were obtained with the following parameters:echo time (TE) = 3.2 ms;repetition time (TR) = 8.5 ms;inversion time (TI) = 450 ms;flip angle = 12°;field of view (FOV) = 256 × 256 mm;matrix = 256 × 256;slice thickness = 1 mm with no gap; acquisition time = 4 min 56 s;188 sagittal slices.

Image Preprocessing

T1-weighted structural images were processed using FreeSurfer(version 6.0, http://surfer.nmr.mgh.harvard.edu/). The preprocessing steps included skull stripping²⁰, tissue segmentation²¹, parcellation of the hemispheres and subcortical structures²², and reconstruction of the gray–white matter interface and cortical surfaces²³. Participants with head motion exceeding 3° were excluded from the study. In addition, the Euler number and total intracranial volume (TIV) were calculated for each T1-weighted image.

Construction of Morphometric Similarity Networks

The cortex was parcellated into 68 regions according to the Desikan-Killiany (D-K) atlas²⁴, and five structural metrics were extracted from each region: surface area, cortical thickness, gray matter volume, Gaussian curvature, and mean curvature. All metrics were z-standardized, and pairwise Pearson correlation analyses were performed to calculate inter-regional correlations, generating a 68 × 68 morphometric similarity network (MSN) matrix for each subject. The mean MSN strength of each region was then computed. For each of the 68 cortical regions, group differences in MSN strength were assessed using a general linear model (GLM) with age, sex, and total intracranial volume (TIV) as covariates, and the resulting p-values were corrected for multiple comparisons across regions using the Benjamini–Hochberg false discovery rate (FDR) procedure at q = 0.05.To control for global effects, regional MSN strength values were linearly regressed on the global mean MSN strength, and the resulting residuals were used as region‑specific MSN measures with global effects removed.

To enable network-level analyses, We performed edge-wise GLM analyses on the MSN matrices, using the MSN strength of each edge as the dependent variable and group as the independent variable, to test group differences in connectivity between each pair of regions and obtain the corresponding test statistics. Nonparametric permutation testing (5000 permutations) was then applied across all edges to obtain corrected p-values. Subsequently, we used publicly available software from GitHub(https://github.com/younghoo/brain-templates-atlases) to map each brain region to the Yeo 7-network parcellation²⁵.

Acquisition of Regional Gene Expression Profiles

The Allen Human Brain Atlas (AHBA, http://www.brain-map.org) is a comprehensive postmortem human brain gene expression database, providing detailed spatial annotations of gene expression based on whole-brain microarray analyses from 3,702 spatial locations across six adult donors. This dataset contains 58,692 probes designed to target 20,737 genes, with each probe assigned to a specific spatial location in the brain.We preprocessed the AHBA dataset using the Abagen toolbox (https://github.com/rmarkello/abagen) as described previously²⁶. The preprocessing steps were as follows: First, microarray probes were mapped to gene symbols, and probes with intensity values below background in ≥50% of samples were discarded. For genes targeted by multiple probes, the probe with the highest regional differential stability was selected, where differential stability was calculated as follows:

$${\mathrm{\Delta }}_{\mathrm{s}}\left(\mathrm{p}\right)=\frac{1}{\left(\genfrac{}{}{0ex}{}{\mathrm{N}}{2}\right)}{\sum }_{\mathrm{i}=1}^{\mathrm{N}-1}{\sum }_{\mathrm{j}=\mathrm{i}+1}^{\mathrm{N}}\mathrm{\rho }\left[{\mathrm{B}}_{\mathrm{i}}\left(\mathrm{p}\right),{\mathrm{B}}_{\mathrm{j}}\left(\mathrm{p}\right)\right]$$

Here, ρ represents the Spearman correlation of the expression values of probe pp across brain regions between each pair of donor brains Bi and Bj, and N is the total number of donor brains. This procedure retained 16,013 probes, each corresponding to a unique gene.Next, using corrected MNI coordinates (https://github.com/chrisfilo/alleninf), tissue samples were assigned to brain regions by selecting the nearest region within a 2-mm radius. Gene expression values were then robustly normalized across regions for each donor using a sigmoid function, resulting in gene expression data scaled to the [0, 1] range.Finally, a transcriptional matrix was generated comprising 68 brain regions and expression levels for 16,013 genes.

To investigate gene expression changes associated with cervical spondylotic myelopathy (CSM), we performed data mining and analysis in the Gene Expression Omnibus (GEO) database and identified one dataset (GSE47681) based on the GPL1261 platform. The gene expression matrix for this dataset was downloaded and subjected to differential expression analysis using the limma package. Genes with|log2(fold change, FC) | > 1 and P < 0.05 were considered significantly differentially expressed.

Potential Association Between Regional Transcriptomic Variation and MSN Strength

Partial least squares (PLS) regression is applied to explore the spatial relationship between the expression levels of all 16,013 genes (predictors) and the regional t-value map of MSN strength (response variable) across the cortex. Fifteen components are set, and the first PLS component (PLS1) shows the highest correlation with the t-statistic map. We perform 10,000 permutation tests to assess whether the covariance between the PLS1 transcriptomic scores and the t-value map exceeded that expected by chance. In addition, we use bootstrapping (10,000 iterations) to evaluate the variability of the weight assigned to each gene in the PLS1 component. Z-scores are defined as the ratio of the gene loading to its bootstrap-estimated standard error, and all genes are ranked according to their PLS1 weights. Genes are classified into two groups based on their PLS1 weight values: PLS1+ genes (Z > 1) and PLS1− genes (Z < − 1). The correlation between regional PLS1 scores and corresponding t-values was assessed using Spearman correlation and further evaluated for spatial dependence.

Functional Enrichment Analysis of PLS1-Positive and PLS1-Negative Genes

Differentially expressed genes identified from the GEO dataset were intersected with the gene lists of PLS1+ (Z > 1) and PLS1− (Z <−1) genes, respectively. The resulting gene sets were then submitted to the Metascape platform for functional enrichment analysis, including Gene Ontology (GO) Biological Process (BP) evaluation and KEGG pathway enrichment analysis. Enriched pathways were subject to FDR correction, and a 5% significance threshold (FDR < 0.05) was used to filter significant results.

Investigating Regional Cellular Changes by Mapping EOS-Related Genes to Cell Types

To investigate the enrichment of PLS1 genes in specific brain cell types, we collected five single-cell transcriptomic datasets from human cortical samples^27–30. Three of these datasets were excluded from further analyses: two did not subdivide neuronal populations^30,31, and one lacked annotation for the “Per” (pericyte) cell type²⁸, which has also been justified in previous studies³².

Next, the single-sample gene set enrichment analysis (ssGSEA) algorithm was applied, using gene sets representing seven major cortical cell types. The “gsva” function from the GSVA R package was used with the “Gaussian” method to calculate cell-type abundance scores (ssGSEA scores) for each region³³. Gene overlaps between PLS1− and PLS1+ gene sets and the gene sets of the seven cell types were evaluated using the GeneOverlap R package (version 1.26.0). Statistical significance of enrichment was assessed by Fisher’s exact test. The degree of overlap was quantified using odds ratio and Jaccard index.Finally, PLS1+ and PLS1− genes were intersected with each of the seven cell-type gene sets, and GO and KEGG pathway enrichment analyses were performed using the Metascape platform. All resulting p-values were corrected for multiple comparisons using FDR.

Statistics and Reproducibility

All statistical analyses were performed using GraphPad Prism 9.5 software (https://www.graphpad.com/). Group comparisons of age and brain volume were conducted using independent samples t-tests, while sex ratio differences were assessed using chi-square tests. Group comparisons of MSN were performed using multiple linear regression models, with age, sex, and TIV as covariates. False discovery rate (FDR) correction was applied to the results of all brain regions to control for multiple comparisons.

Ethical Statement

This study was approved by the Research Ethics Committee of the First Affiliated Hospital of Anhui Medical University (Approval No. PJ2022-01-12). All participants provided written informed consent in accordance with the ethical principles of the Declaration of Helsinki³⁴. In addition, participant data were de-identified to prevent any potential risk of reidentification.The ethics and informed consent statement of AHBA can be found on the Allen Institute’s Policies and Compliance page: https://alleninstitute.org/policies-and-compliance/.

Results

The overall research process consisted of four main stages: construction of the MSN, Partial Least Squares (PLS) analysis, and gene and cell enrichment analyses(Fig. 1). Demographic Data:

Fig. 1. The workflow of this study.

a For each participant, a morphological similarity network (MSN) matrix was constructed based on surface area, cortical thickness, gray matter volume, Gaussian curvature, and mean curvature. Brain regions showing significant differences between patients with cervical spondylotic myelopathy (CSM) and healthy controls were identified. Partial least squares (PLS) regression was then used to explore imaging–transcriptomic associations. b Enrichment analysis was performed on the overlapping genes between the PLS1 gene list and disease-related genes. c Single-sample gene set enrichment analysis (ssGSEA) was conducted to examine the enrichment of the PLS1 gene list in seven major brain cell types. GC:gaussian curvature;GM:gray matter volume;MC:mean curvature;SA:surface area; CT:cortical thickness;PLS:Partial least squares

A total of 170 patients with CSM (86 males and 84 females; mean age: 51.09 ± 8.10 years) and 79 healthy community-recruited controls (35 males and 44 females; mean age: 49.04 ± 10.78 years) are included in this study. Significant differences are observed between patients and controls in terms of age (p = 0.021) and individual brain volume (p = 0.038), whereas no significant difference is found in sex distribution (χ²=0.853, p = 0.356).

Alterations in MSN Strength in CSM Patients

Regional MSN strength changes in the CSM group are generally opposite to those in the HC group, with a global increase observed in CSM patients (Fig. 2a–d).Fig. 2c(Supplementary Data 2) shows the distribution of MSN strength, adjusted for covariates, in both healthy controls and CSM patients. MSN strength exhibits a normal distribution in both groups.Notably, significant increases in MSN strength are found in regions such as the caudal anterior cingulate, isthmus of the cingulate and so on. In contrast, significant decreases are observed in regions such as the precentral gyrus and superior frontal gyrus(Fig. 2b,Supplementary Data 3 and 4).To test whether the above regional differences were primarily driven by a global increase in MSN, we include each participant’s global mean MSN strength as a covariate. After FDR correction, significant region-specific differences remained(43 of 68) (Fig. 2b,Supplementary Data 5); a residualization approach yields the same result, with 43 of 68 regions significant after FDR correction (Supplementary Data 5).To impose more stringent control for multiple comparisons, we perform a Freedman–Lane permutation test and apply maxT family-wise error rate (FWER) correction. In total, 27 of 68 regions remain significant at p <0.05, indicating a set of regions with stable differences even under stringent correction(Supplementary Data 6).

Fig. 2. Group differences in MSN between CSM and HC.

a Mapping of the average residuals of the HC and CSM groups on the Desikan-Killiany (D-K) cortical atlas. b Cortical mapping of case–control t-values of MSN strength in CSM patients, t-values of the significantly different regions projected onto the cortical surface and case–control t-values of MSN strength after regressing out the global mean MSN strength. Increased MSN strength indicates greater morphometric similarity between a given region and the rest of the cortex, suggesting CSM-related compensatory or pathological cortical remodeling. Decreased MSN strength indicates reduced morphometric similarity between a given region and the rest of the cortex, suggesting lower concordance with the brain-wide morphometric pattern. c Histogram showing the relative frequency of brain regions within each range of residual values; the x-axis represents the mean residual MSN strength per region (after regressing out age, sex, and TIV), and the y-axis indicates the relative frequency. d Group comparison of overall MSN strength between healthy controls and CSM patients; each point represents the mean residual MSN strength of a cortical region in either the HC or CSM group.We perform independent paired t-tests for each brain region, and apply BH-FDR correction across all regions to control the false discovery rate. e Scatter plot showing the association between MSN strength in healthy controls and case–control t-values across brain regions, using pearson correlation analysis with BH-FDR correction.

There is a negative spatial correlation between the case–control t-value of CSM and the distribution of MSN strength in healthy controls (r = –0.789, p < 0.05). Specifically, 69.1% of cortical regions with positive MSN strength in healthy controls exhibit “decoupling”, while the remaining 30.9% of regions with negative MSN strength exhibit “dedifferentiation” (Fig. 2e).

To investigate between-network and within-network effects, the DK68 atlas is projected onto the Yeo 7-network parcellation for analysis. The results shows that within-network alterations are mainly concentrated in the visual network and the default mode network, both exhibiting increased MSN; at the between-network level, CSM patients also show significant MSN differences in the connections between multiple functional networks (Fig. 3,Supplementary Data 7) .

Fig. 3. Within-network and between-network effects.

R represents the left hemisphere, and L represents the right hemisphere.The abbreviations for each brain region are as follows: baSTS:bankssts; cACC:caudalanteriorcingulate; cMFG:caudalmiddlefrontal; Cun:cuneus; ENT:entorhinal; FuG:fusiform; IPL:inferiorparietal; ITG:inferiortemporal; iCC:isthmuscingulate; LO:lateraloccipital; LOF:lateralorbitofrontal; LinG:lingual; MOF:medialorbitofrontal; MTG:middletemporal; PHG:parahippocampal; PaC:paracentral;POp:parsopercularis;PTr:parstriangularis;PCAL:pericalcarine;PoCG:postcentral;PCu:precentral;SFG:superiorfrontal;SPL:superiorparietal;STG:superiortemporal; SMG:supramarginal; TP:temporalpole; TTG:transversetemporal.

Transcriptional Patterns Associated with Regional Changes in MSN Strength

Partial least squares regression identifies 15 components, with PLS1 explaining over 20% of the variance in regional structural differences, markedly higher than the other components (Fig. 4c; Supplementary Data 8). Spatial autocorrelation analysis further confirms a significant spatial correlation between the t-values of the altered brain regions and PLS1 scores (r = 0.482, p = 0.0027). Subsequently, a one-sample Z-test is applied to identify genes with the largest weights along the PLS1 component, resulting in 510 highly loaded genes for PLS1+ (Z > 1) and 606 lowly loaded genes for PLS1– (Z<–1) (Fig. 4a, b; Supplementary Data 10). The selected PLS1+ and PLS1– gene sets represent genes with the largest positive and negative weights along the PLS1 component, respectively.

Fig. 4. PLS analysis linking cortical MSN alterations to gene expression.

a Volcano plot showing the relationship between gene weights (Z-scores) in the PLS1 component and their statistical significance (−log₁₀p). Vertical dashed lines indicate thresholds for PLS1+ and PLS1− genes (Z > 1 and Z <−1), and the horizontal dashed line denotes the nominal significance threshold (p = 0.05). Red and blue dots represent PLS1+ and PLS1− genes, respectively.This analysis uses PLS to extract the first component, PLS1, and evaluate brain-region scores. Permutation test (n_permutations=10,000, two-sided) yields empirical p-values, with BH-FDR correction. Bootstrap (n_bootstrap=10,000) estimates SE and computes Z-values b Genes ranked by their PLS1 loading values. c Percentage of variance in the response variable explained by each of the 15 PLS components. d Surface mapping of regional t-values and PLS1 scores across the Desikan–Killiany cortical atlas. e Scatter plot of PLS1 scores vs. case–control t-values across 45 regions.We perform a univariate linear regression of t_value on PLS1_Score. The regression test is two-sided. No adjustment for multiple comparisons was applied for this single test. Each yellow dot denotes a region. The black line is the regression fit; the center line is the regression prediction y_hat, with the shaded area as the 95% CI for the mean (or 95% PI if specified). Marginal histograms show the distributions.

Fig. 4d illustrates the PLS scores for each brain region: regions such as the precentral gyrus and superior/inferior parietal lobules exhibit high negative loadings on PLS1, whereas regions including the temporal pole, parahippocampal cortex, and entorhinal cortex show high positive loadings (Supplementary Data 9). Moreover, PLS1 loadings are positively correlated with the case–control t-value (Fig. 4e).

Enrichment Analysis Related to MSN Changes

Based on the related genes from the GEO database (1,256 genes, see Supplementary Data 11), PLS1+ and PLS1− genes are screened, identifying 33 meaningful overlapping genes in PLS1+ and 31 in PLS1− (Fig. 5). Enrichment analysis (see Supplementary Data 12) shows that PLS1+ genes are significantly enriched in multiple GO biological processes. The top 12 enriched pathways are illustrated in Fig. 5, including “regulation of MAPK cascade,” “positive regulation of response to external stimulus,” “negative regulation of cellular response to growth factor stimulus,” and “response to wounding.” In addition, one GO cellular component is identified: “collagen-containing extracellular matrix.” Three significant KEGG pathways are also found: “platelet activation,” “Apelin signaling pathway,” and “Kaposi sarcoma–associated herpesvirus infection.”

Fig. 5. Enrichment analysis of common genes.

Shared genes between PLS1− and GEO disease-related genes (n = 31), and between PLS1+ and GEO disease-related genes (n = 33), were identified. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were performed for both sets of overlapping genes.

Genes related to PLS1– are mainly enriched in three GO biological processes: negative regulation of cell development, lipid localization, and inhibitory synapse assembly, as well as in one GO cellular component: “primary lysosome.” No significant KEGG pathway enrichment is found for PLS1– genes (Fig. 5).

Cellular characteristics of CSM-related genes

ssGSEA for cell-type genes across brain regions reveals that genes specific to excitatory and inhibitory neurons are highly expressed in the differential regions (Fig. 6a, Supplementary Data 13). GeneOverlap analysis shows that PLS1-negatively weighted genes are mainly enriched in excitatory neurons (number = 61, FDR-corrected) and inhibitory neurons (number = 59, FDR-corrected). Although PLS1+positively weighted genes are expressed in oligodendrocyte precursor cells and microglia, the enrichment does not survive FDR correction (Fig. 6b, Supplementary Data 14). Further enrichment analysis of the overlapping genes indicates that those related to excitatory neurons are mainly enriched in inorganic ion transmembrane transport and purine nucleotide metabolic processes. Genes related to inhibitory neurons are primarily enriched in synaptic signaling and involved in the composition of the presynaptic membrane, postsynaptic membrane, and transmembrane transporter complexes (Fig. 6c).

Fig. 6. Cell type–specific enrichment of PLS1 associated genes.

a Heatmap showing enrichment scores of gene sets corresponding to seven major brain cell types—astrocytes, endothelial cells, oligodendrocytes, oligodendrocyte precursor cells (OPCs), microglia, excitatory neurons, and inhibitory neurons—across the 45 differentially affected brain regions. b Enrichment of PLS1+ and PLS1− gene sets in the seven canonical brain cell types. c Bubble plots showing GO and KEGG enrichment results of marker genes enriched in excitatory and inhibitory neurons within the PLS1− gene set.

Discussion

In this study, we integrated the macroscopic structural alterations of the MSN with the region-specific transcriptional expression patterns of brain genes to explore the relationship between structural changes and gene expression. Previous research on the MSN has provided a theoretical foundation for this approach¹: Cortical regions with similar morphology are expected to share a high level of gene co-expression²; Combining multiple MRI-derived morphometric measures enables a more accurate inference of the structural properties of the human cortex; and³ There is a high likelihood of axonal connectivity between histologically similar cortical areas¹³. Additionally, the symptoms of CSM patients are highly associated with abnormal functional connectivity among brain regions^35,36. Integrating macroscopic MSN structural alterations with disease-specific transcriptional expression profiles thus offers a promising direction for future mechanistic studies and provides deeper insights into how disease-related changes at the microscopic level may drive the macroscopic brain abnormalities observed in CSM patients.

Our findings suggest that the impact of CSM on the brain is accompanied by more widespread cortical morphological alterations. The global increase in MSN may reflect long-term abnormalities in sensorimotor input caused by chronic spinal cord compression, leading multiple regions to undergo similar structural adaptations through plasticity mechanisms, thereby making cortical areas more consistent in their morphological features^37,38. Alternatively, it may indicate a degree of “compensatory hyper-coupling”. Brain regions that typically act in coordination may, in order to maintain core functions, exhibit strengthened inter-regional structural coupling, with different networks increasing cooperation to share the demands of the same function^39,40. Importantly, an increase in global MSN does not necessarily imply functional improvement; it may reflect both adaptive compensation and maladaptive remodeling.To adapt to prolonged pain in patients with chronic pain, regions such as the cingulate cortex and entorhinal cortex may undergo coordinated cross-regional morphological remodeling, reflected by shared similarity of structural features across regions. This is consistent with our finding of global increase in MSN^41,42.

To exclude the influence of the global effect on regional differences in MSN, we performed a sensitivity analysis by including each participant’s global mean MSN as a covariate in the model and applied a more stringent correction for multiple comparisons based on this adjusted model, after which 27 regions still showed significant between-group differences. These findings indicate that, even after controlling for global effects and applying strict multiple-comparison correction, regional MSN differences persist, suggesting that they are not solely driven by an overall increase in MSN but more likely reflect region-specific alterations such as the precentral gyrus and superior frontal gyrus in CSM patients compared to HC. The precentral gyrus, which serves as the primary motor cortex, is the origin site of the corticospinal tract (CST) and is responsible for the voluntary control of the limbs and trunk⁴³. Previous studies have confirmed that patients with CSM exhibit higher intrinsic activation in the sensorimotor cortex compared to healthy individuals, which can return to normal levels following decompression surgery⁴⁴. Although the superior frontal gyrus does not directly project CST fibers, it is involved in the planning and fine regulation of movement. Damage to the superior frontal gyrus has been associated with impairments in movement initiation and coordination^45,46. Both regions are highly related to the corticospinal tract and work together to control voluntary movement.Chronic spinal cord compression leads to impaired neuronal signaling, clinically manifesting as sensory and motor dysfunction. Centrally, this presents as reduced input of sensory and motor signals, decreased cortical activity, use-dependent atrophy, and reduced neuronal connectivity. Therefore, rehabilitation for CSM patients should not only focus on restoring the function of the motor cortex itself but also on enhancing external sensory input^47,48. Previous studies have demonstrated that transcranial electrical stimulation, when used as an adjunct to task training, can significantly improve hand function in patients with cervical spinal cord injury⁴⁹.

In addition to global and regional MSN alterations, we further examined within‑network and between‑network effects. The results showed that, compared with healthy controls, patients with CSM exhibited widespread coordinated remodeling across multiple large‑scale brain networks, characterized overall by a generalized increase in between‑network MSN. This pattern is consistent with our aforementioned hypothesis of “compensatory hyper‑coupling.” For example, the elevated MSN between the Somatomotor network and the Visual network may reflect increased reliance on visual information under gait instability and impaired proprioception, whereby visual input partially compensates for deficits in posture and gait control.The within-network changes were predominantly localized to the Visual network and the default mode network(DMN), suggesting a more pronounced morphological reorganization in this network. Given that the DMN is closely associated with the regulation of emotion and pain, the increased within-network MSN may indicate corresponding alterations in systems related to memory, attention, and emotional processing.^50,51.

Our findings also revealed a negative correlation between MSN strength in brain regions of healthy individuals and the case-control t-values. This indicates that regions more tightly connected with surrounding areas exhibit greater group differences, and may even serve as highly integrated hubs¹³.These results suggest that CSM-induced changes in the central nervous system are selective, with structural network hubs being more susceptible to damage than relatively independent areas.

Using PLS, we identified genes whose normative regional expression shows a spatial correspondence with MSN changes. Notably, PLS1 loadings were positively correlated with the case-control t-values, suggesting that MSN alterations co-localize with an AHBA-derived regional gene-expression pattern across regions. Importantly, this reflects a spatial association rather than disease-specific differential expression. The PLS1+ genes were primarily enriched in regulation of the MAPK cascade and positive regulation of responses to external stimuli, which may reflect regional molecular features conferring either vulnerability to pathological remodeling or capacity for adaptive plasticity, depending on the specific pathway branch and activation dynamics. The MAPK signaling pathway is a highly conserved transduction cascade with distinct branches that may exert different, or even opposite, effects on neural plasticity and functional recovery. For example, activation of the NF-κB pathway can induce microglial activation, leading to the expression of inflammatory mediators such as IL-6, TNF-α, or ROS⁵². The p38 family comprises at least four subtypes (p38α, β, γ, δ), among which p38α is a key regulator of CNS inflammation^53,54. Overactivation of these pathways generally inhibits neural plasticity and impedes functional recovery. In contrast, transient or moderate activation of the ERK pathway can promote protective autophagy and is involved in synaptic plasticity, thereby playing a crucial role in neural adaptation⁵⁵. Positive regulation of response to external stimuli” similarly captures enhanced responsiveness to environmental cues via changes in neural signaling, synaptic plasticity, and inflammatory responses, and therefore may index both susceptibility and adaptive potential.

The PLS1− genes were mainly enriched in lipid localization and inhibitory synapse assembly. Lipid localization is key for maintaining neuronal structure and function, as it regulates the distribution of lipids in the pre- and post-synaptic membranes, influences neurotransmitter release, and is also essential for mitochondrial function⁵⁶. Recent research has shown that cardiolipin levels significantly decrease after SCI, and maintaining mitochondrial cardiolipin homeostasis can reduce cell death and promote recovery from spinal cord injury⁵⁷. In this context, lipid localization may indicate regions with higher mitochondrial dependence and thus greater vulnerability to injury-related remodeling, while also representing a substrate that could support recovery if homeostasis is preserved. The enrichment of inhibitory synapse assembly suggests a potential disruption of the excitation-inhibition balance in CSM patients.Enrichment in inhibitory synapse assembly may reflect either a baseline circuit feature that increases vulnerability to excitation–inhibition imbalance under disease stress, and/or a capacity for activity-dependent inhibitory circuit reorganization. This is consistent with our cellular-level findings, where both excitatory and inhibitory neuron genes were highly enriched. Overall, our findings suggest that PLS1 + /PLS1− enrichments may reflect baseline regional identity with implications for vulnerability and plasticity; however, distinguishing these interpretations will require convergent evidence.

Previous studies have demonstrated that cellular gene expression profiles can explain intergroup differences in cortical structural changes across various psychiatric disorders, including ASD, SCZ, BD, and MDD⁵⁸. Although our study focused on patients with CSM, these individuals also showed cortical structural alterations, suggesting that CSM may similarly involve transcriptional abnormalities in specific cell types that impact brain structure and function. In our analysis, inhibitory and excitatory neurons constituted the largest proportion of PLS1- weighted genes. The balance between excitatory and inhibitory neuronal activity plays a pivotal role in functional recovery following spinal cord injury^59,60. Notably, research on spinal cord plasticity has shown that adult spinal cord injury leads to a neurochemical shift in excitatory interneurons toward an inhibitory phenotype, reducing the excitability of motor neurons. In contrast, neonatal spinal cord injury preserves an excitatory phenotype in interneurons, which facilitates synaptic sprouting⁶¹.

Our GO enrichment analysis further found that shared genes between PLS1- and both excitatory and inhibitory neurons are primarily involved in synaptic signaling processes: transmembrane transport of inorganic ions affects action potentials, while purine nucleotide metabolic processes regulate neurotransmitter synthesis. Genes shared specifically with inhibitory neurons are directly involved in synaptic assembly. These results suggest that cortical structural changes in CSM patients are likely attributable to impaired synaptic transmission; some neurons may undergo atrophy due to sustained low-input conditions, while neurons that typically play a supporting role under normal physiological circumstances may be strengthened to maintain overall function^62,63.

There are several limitations to our study. First, due to the lack of CSM-specific gene sets, we use gene sets associated with spinal cord injury as a substitute. SCI and CSM differ in etiology and temporal dynamics. Accordingly, using the SCI transcriptomic dataset to filter genes may bias the retained genes toward acute injury–related transcriptional changes (e.g., inflammatory responses), potentially leading to an overestimation of inflammation/stress-related pathways and an underestimation of chronic degenerative and long-term remodeling processes in downstream functional enrichment analyses. Second, gene enrichment analysis can only provide broad directions for mechanistic exploration; further basic research is needed to validate these findings.Third, the study was deliberately designed to probe the supraspinal adaptive and maladaptive mechanisms, using the brain as the readout so we didn’t collect spinal cord morphometrics metrics.

Conclusion

In summary, compared with healthy controls, patients with CSM exhibit distinct macroscopic structural changes in the cerebral cortex, with structural network hubs being more susceptible than brain regions with simpler structures. Abnormalities in synaptic composition and signal transmission may lead to an imbalance between excitatory and inhibitory neuronal activity, which can be considered a possible hypothesis to explain these cortical structural changes. In the future, we will validate the imbalance of excitatory-inhibitory activity through experiments such as animal models to simulate E/I imbalance, multi-channel electrophysiological techniques to record local electrical activity and inhibitory activities, and neurotransmitter level analysis.

Supplementary information

Supplementary Data Descriptions ^{(17.5KB, docx)}

Supplementary Data 1-14 ^{(428.9KB, xlsx)}

Author contributions

H.Q.W.: Data collection, data processing, formal analysis, drafting the original manuscript. K.Z.: Data collection, data processing, methodology. H.X.W.: Visualization, review and editing. Q.G.: Software, data processing. Z.C.L.: GEO database analysis. S.C.B.: Data collection,Data management. B.Y.X.: review and editing. H.Y.N.: Software. Y.Y.W.: Data management, methodology. Y.Q.Y.: Validation, supervision. F.L.D.: Review and editing, visualization, supervision, resources, project administration, funding acquisition.

Peer review

Peer review information

Communications Medicine thanks Hosni Cherif who co-reviewed with Shirin Sadighparvar and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Funding

This study was funded by the following funding sources: the First Affiliated Hospital of Anhui Medical University Clinical Research Project (Grant No. LCYJ2021YB018), University Natural Science Research Major Project of Anhui Province (2025AHGXZK20147), University Natural Science Research Key Project of Anhui Province (Grant No. 2022AH051158), Horizontal Research Project of Anhui Medical University (9021755201), and Health Research Project of Anhui Province (Grant No. AHWJ2023A20125).

Data availability

The source data for Fig. 2 is in Supplementary Data 2 to 5; the source data for Fig. 3 is in Supplementary Data 7; the source data for Fig. 4 is in Supplementary Data 8 to 10; the source data for Fig. 5 is in Supplementary Data 11 and 12; the source data for Fig. 6 is in Supplementary Data 13 and 14. Region-based gene expression maps are available from the Allen Brain Atlas (https://human.brain-map.org/static/download). The genes related to spinal cord injury are stored in a GEO dataset (GSE47681) at the GEO database (https://www.ncbi.nlm.nih.gov/geo/). To protect participant privacy, the data generated in this study are available under controlled access. Access to the data can be requested by contacting dongfulongtg@sina.com. Requests will typically be responded to within 8 weeks. Any restrictions on the use of the data will be detailed in the response.

Code availability

The code developed to perform the analyses described is available at https://github.com/zhukun990427/CSM-MSN-Code.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s43856-026-01661-z.