Development of an integrated predictive model for postoperative glioma-related epilepsy using gene-signature and clinical data

Data from 208 patients who underwent craniotomy and were diagnosed with primary DHGG at Beijing Tiantan Hospital, China, were retrospectively reviewed. Patients who previously underwent surgery for other intracranial diseases or had any other concomitant malignant diseases were excluded from the study. All patients had received the standard treatment, including maximal safe resection and adjuvant radiotherapy and/or chemotherapy, depending on their tumor grade. Demographic data (age and gender) and the main complaints from patients were retrieved from an electronic medical record system. Tumor samples collected during craniotomy were flash frozen in liquid nitrogen and stored at -80 °C. Hematoxylin and eosin staining was performed, and tumor tissues containing > 80% tumor cells were selected for further analysis. Molecular pathological characteristics, including IDH1/2 mutation, chromosome 1p19q co-deletion, and MGMT promoter methylation status were assessed through pyrosequencing or fluorescence, as previously described [12, 13]. The pathological diagnosis of each patient was re-evaluated by integrating histopathology and molecular biomarkers according to the 2016 WHO classification of tumors of the central nervous system[14]. Transcriptomic data of samples collected during surgery were generated on the Agilent Whole Human Genome Array platform, as previously described [15]. RNA expression data were converted into fragments per kilobase million matrix and used in subsequent analysis.

Preoperative GRE was defined as at least one unprovoked epileptic seizure prior to surgery based on the patients' main complaint and history of illness. All patients received prophylactic AEDs after surgery, including phenobarbital injections administered within 3 d and valproate or levetiracetam tablets administered for at least 3 months. Seizure outcome was followed up via telephone interviews conducted 12 months after surgery, and postoperative seizure occurrence was defined as at least one unprovoked epileptic seizure occurring between the day of discharge and the day of follow-up. Patients with overall survival time of less than 12 months were excluded from the study. All data analyzed here, except those related to seizures, have been uploaded to the CGGA portal (http://​www.​cgga.​org.​cn/​) for open access [16].

Patients were randomly assigned to either a training or validation cohort in a 4:1 ratio. The prediction model was established using the training cohort and validated using the validation cohort. The R package “Limma” was used to screen for differentially expressed genes (DEGs) between patients with or without preoperative GRE under the criteria of adjusted p-value < 0.05 (Benjamini and Hochberg Method) and |log2(fold change)|≥ 0.8. Gene Ontology (GO) enrichment analysis, Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis[17‐19], and Gene Set Enrichment Analysis (GSEA) were conducted using the R package “clusterProfiler.” Adjusted p < 0.05 was used as the cut-off threshold. Enrichment analyses used the Molecular Signatures Database (MSigDB) as a reference.

Immune cell infiltration in tumor samples was analyzed using the “CIBERSORT” algorithm (https://​cibersort.​stanford.​edu/​), which was used to evaluate the cell composition of tumor samples based on gene expression profiles. The proportions of 22 types of tumor-infiltrating immune cells were evaluated according to the LM22 signature genes file, with 1000 permutations. Tumor infiltration levels in patients with and without preoperative GRE were compared.

Risk score = \(\sum_{{\varvec{i}}=1}^{{\varvec{n}}}{{\varvec{\beta}}}_{{\varvec{i}}}{{\varvec{x}}}_{{\varvec{i}}}\)(1)

The performance of the risk score was evaluated using receiver operating characteristic (ROC) analysis and the area under the curve (AUC), implemented using the R package “pROC.” Patients in the training and validation cohorts were then divided into high-risk and low-risk groups based on the median risk score value.

Weighted correlation network analysis (WGCNA), implemented using the R package “WGCNA,” was conducted to identify key modules and genes with the highest correlation with preoperative GRE history, postoperative seizure occurrence, and the obtained gene-signature. Age, WHO grade, IDH1/2 mutation status, chromosome 1p/19q codeletion status, preoperative GRE history, postoperative seizure occurrence, and risk scores were included as variables in the analysis. Risk scores and patient age (with a cut-off of 45 years) were included as dichotomous categorical variables. The top 5000 genes (based on variance) were screened, and those with β = 5 (R² ≥ 0.9) were selected and used to construct an unsigned scale-free co-expression gene network. The smallest module contained at least 30 genes, and analogous modules were fused at a height cut-off of 0.25. Genes in the obtained key modules were extracted for further analysis.

WE generated a new set of DEGs based on risk scores (RDEGs) by identifying genes that were differentially expressed between high-risk and low-risk groups and selecting genes with potential effect on both the occurrence and outcome of GRE. Genes overlapping upregulated or downregulated RDEGs and genes extracted from WGCNA were selected. Protein–protein interaction (PPI) analysis was performed using the R package “STRINGdb” to explore the interactions among these genes. Genes showing high numbers of PPI interactions (those interacting with more than five other nodes) were identified as hub genes. The expression levels of hub genes in the validation cohort were also evaluated.

Statistical analysis was performed using R software version 4.0.5 (R Foundation for Statistical Computing, Vienna, Austria, www.​r-project.​org). Figures were generated using R packages, including “ggolot,” “pheatmap,” and “ggraph.” Chi-square and Fisher’s exact tests were used to evaluate the statistical significance of categorical variables. Patients with missing data were excluded from multivariate analyses. Statistical significance was set at p < 0.05, unless otherwise stated.

The 208 patients enrolled into the study were randomly assigned into a training cohort (n = 166) and a validation cohort (n = 42). Detailed information on the patients are presented in Table 1. Of the patients, 77 had a history of preoperative GRE (62 in the training cohort and 15 in the validation cohort), while postoperative seizures occurred in 37 patients (30 in the training cohort and 7 in the validation cohort) 12 months after surgery. We further compared the baseline characteristics between the enrolled patients and excluded patients, who died within 12 months, and found many variables including age at diagnosis, WHO grade, tumor pathology, incidences of pre- and postoperative GRE were significantly different.

Based on the “Limma” algorithm, 623 DEGs were associated with preoperative GRE in the training cohort, including 381 upregulated and 242 downregulated genes (Fig. 1A). Functional enrichment analysis revealed that upregulated DEGs were mainly associated with ion channel regulator activity and structural constituent of myelin sheaths (Fig. 1B, Supplementary Fig. 1A, GO terms), and neuroactive ligand-receptor interaction and cAMP signaling pathway (Supplementary Fig. 1B, KEGG). Downregulated DEGs were predominantly involved in extracellular matrix organization and structure (Fig. 1C, Supplementary Fig. 1C, GO terms), and ECM-receptor interaction and focal adhesion (Supplementary Fig. 1D, KEGG). GSEA showed that ion channel activity (GO terms) and GABAergic synapse and glutamatergic synapse pathway (KEGG) were activated, while lymphocyte and adaptive immunity (GO terms) and ECM-receptor interaction and TNF signaling pathway (KEGG) were suppressed in patients with preoperative GRE (Fig. 1D-F). In addition, the relative proportions of 22 types of tumor-infiltrating immune cells were evaluated using the “CIBERSORT” algorithm. The landscape of immune cell infiltration is shown in Fig. 1G. The results showed that the levels of monocytes were significantly increased, while the levels of M0 macrophages were significantly lower in patients with preoperative GRE (Fig. 1H).

LASSO logistic analysis of DEGs was used to identify genes with the highest ability to predict postoperative seizure occurrence. Seven genes: LUZP2, GPNMB, MMD2, GALNT13, IKBKGP1, SLC1A4, and RP1 were identified and subsequently used to construct a predictive gene signature (Fig. 2A-B). Detailed information on these seven genes is given in Table 2. The risk score for each individual was calculated using the following formula:

Risk score = 0.000646822 * Exp_LUZP2 + 0.002348291 * Exp_GPNMB + 0.003071414 * Exp_MMD2 + 0.005716866 * Exp_GALNT13 + 0.014817826 * Exp_IKBKGP1 + 0.01969116 * Exp_SLC1A4 + 0.019902233 * Exp_RP1(2).

ROC analysis was used to test the predictive capacity of the gene-signature in the training and validation cohorts, and AUC values of 0.842 and 0.75 were obtained from the respective cohorts (Fig. 2C-D).

WGCNA was performed to identify hub genes associated with preoperative GRE history, postoperative seizure occurrence, and the obtained gene-signature. Previously reported risk factors for GRE, including age, WHO grade, IDH1/2 mutation status, and chromosome 1p/19q codeletion status were also included in the analysis (Fig. 3A). WGCNA was conducted using the top 5000 genes with the largest variance, and the soft-threshold power was set to 5 with scale free R² ≥ 0.9 (Fig. 3B). A total of eight co-expression modules were identified (Fig. 3C). The topological overlap matrix constructed from the top 1000 genes revealed a strong co-expression of genes in these modules (Fig. 3D). Correlations between the eight modules and the included variables were analyzed (Fig. 3E). The MEpink module had the most significant positive correlation with higher gene-signature-derived risk score (r = 0.64, p = 2e-20), preoperative GRE (r = 0.33, p = 2e-05), and postoperative seizure occurrence (r = 0.25, p = 0.001) (Fig. 3E). This module was also correlated with IDH1/2 mutation (r = 0.35, p = 4e-06) and higher WHO grade (WHO grade 4, r = -0.31, p = 5e-05). Genes in the MEpink module (62 genes) showed similar correlations (Fig. 3F-H).

Subsequent analysis of DEGs between patients with high and low gene-signature-derived risk scores was conducted to generate a new set of RDEGs. A total of 1112 RDEGs were upregulated in the high-risk group, with 44 of them overlapping with genes in the pink module (Fig. 4A). PPI network analysis identified a densely connected module consisted of six genes (Fig. 4B, Table 3). The expression of these genes was also upregulated in patients with gene-signature-derived high risk scores both in the training and validation cohorts (Fig. 4C-D).

Multivariate logistic regression analysis was used to identify variables for constructing the postoperative seizure occurrence prediction model. Age (Odds Ratio [OR] = 0.944, p = 0.046), temporal lobe involvement (OR = 5.414, p = 0.011), preoperative GRE (OR = 4.120, p = 0.047), and higher risk scores (OR = 17.766, p < 0.001) were independently associated with postoperative GRE (Table 4). A predictive nomogram for postoperative seizure occurrence was generated using these four variables (Fig. 5A). The nomogram showed the weight of each variable and showed that the gene-signature-derived risk score was critical in the prediction of postoperative seizure occurrence. To further determine the role of gene-signature-derived risk scores in the model, ROC analysis was used to calculate the predictive performance of prediction models generated using different combinations of the four variables. The AUC was 0.75 when the model was generated using the three clinical variables, but increased to 0.88 when gene-signature-derived risk scores were added to the model (Fig. 5B-C). The calibration curve also showed good agreement between the predicted (red line) and observed (black dot line) probabilities (Fig. 5D). Finally, we evaluated the predictive capacity of our model using the validation cohort and obtained an AUC of 0.84 (Fig. 5E-H).

Postoperative GRE is a disturbing complication to both patients and treating physicians. Most neurosurgeons routinely prescribe postoperative AED prophylaxis for patients with glioma, even though their ability to reduce seizures is not clear [20, 21] and current research evidence and official guidelines do not support their administration [22, 23]. One potential root cause for this contradiction is the differences of seizure susceptibility among patients, for those with a high possibility of postoperative seizures, more aggressive AED prophylaxis maybe beneficial rather than current routine prophylaxis. Thus, studies on risk factors and predictors of postoperative seizure are important for personalized management of patients with glioma. Significant progress has been made in the study of patients with low-grade gliomas. Many clinical characteristics, including younger age, longer history of preoperative seizure, presenting as focal epilepsy, and sub-total resection, have been identified as risk factors for poor seizure control [24‐26]. We previously developed an accurate RNA-seq-based model for predicting postoperative seizure outcomes in patients with diffuse low-grade gliomas [10].

This is the first study to generate a predictive gene-signature for postoperative seizure occurrence in patients with DHGG. We further developed an integrated prediction model using both gene expression and clinical data. Age, temporal lobe involvement, preoperative GRE history, and gene-signature-derived risk scores were identified as independent risk factors for postoperative seizure occurrence and included in the model. Of these, gene-signature-derived risk score was the most significant. Age, temporal lobe involvement, and preoperative GRE have also been confirmed as predictors of postoperative seizures by previous studies [9, 25‐28]. Overall, the integrated prediction model showed good accuracy and stability. Given that patients who were excluded from this study differed from included patients in many aspects, the current integrated prediction model was statistically mainly applicable in DHGGs patients with relative longer survival. And the model can promote the prevention of seizures in these patients, improving their quality of life. However, limited by the conditions of enrolled patients, our model cannot provide sufficient information to avoid AEDs abuse in patients who may not need anti-epileptic treatment at all. Further prospective studies are needed to prompt rational use of AEDs.

GRE and glioma are comorbidities and share common pathogenic mechanisms. In this study, we explored the mechanisms underlying GRE and discovered that ion channel and immune system dysfunctions may play a major role in the epileptogenesis of GRE.

Epilepsy is correlated with changes in neuronal excitation/inhibition balance. Ion channel malfunctions can influence the extracellular concentrations of ions and neurotransmitters, inducing neuronal hyperexcitability that lead to seizures [29]. Altered ion channels are also associated with the proliferation and invasion of glioma cells [30]. In this study, functional enrichment analysis revealed that ion channel activities (especially potassium ion channels) were overactivated in patients with GRE. This may decrease interneuron activity and lead to disinhibition, resulting in seizures [31]. Moreover, pathway analysis showed that activities of the glutamatergic synapse and cAMP signaling pathways were enhanced in patients with GRE. Both pathways have been positively correlated with increased neuronal excitability [32, 33].

The correlation between immune dysfunction and glioma is widely accepted. Moreover, the relationship between immune dysfunction and epilepsy has been studied intensively, and the 2017 International League Against Epilepsy seizure classification identified immune group as a major etiological category for epilepsy [34, 35]. In our study, analysis of the immune cell infiltration landscape revealed an increase in the relative proportion of monocytes and a decrease in the relative proportion of M0 macrophages in patients with GRE. Both changes have been associated with better prognosis, representing lower tumor malignancy [36]. Thus, we inferred that GRE and tumor malignancy may be negatively correlated. Berendsen et al. found that epilepsy is an independent prognostic factor for longer survival in patients with glioblastoma. They proposed that the prognostic effect cannot be explained solely by the early diagnosis, but that the distinct biological features and alterations in gene expression may constitute the underlying mechanism [37]. The negative correlation can also be explained from the perspective of the epilepsy network. An integrated brain network is necessary for seizure generation and propagation, but tumors with higher malignancy can lead to the destruction of the brain network, blocking the propagation of seizures [38].

Seven genes were selected from the 623 DEGs and used to generate a predictive gene-signature for postoperative seizure occurrence. Of these, SLC1A4 was essential in the gene-signature (β = 0.019691). SLC1A4 is a member of the Solute Carrier Family 1, which has a high-affinity for glutamate and is involved in its transport in the central nervous system [39]. It encodes ASCT1, an effective transporter of glutamate under low pH, and is active in acidic tumor microenvironments [40]. We postulated that the mechanism underlying the correlation between SLC1A4 overexpression and GRE involved upregulated SLC1A4 resulting in increased concentration of glutamate in the peritumoral microenvironment, leading to overexcitement of the surrounding neurons and induction of seizures. SLC1A4 may be a therapeutic target for GRE.

Additionally, WGCNA, RDEG identification, and PPI analysis were performed to identify additional potential therapeutic targets for GRE. Of the final six selected genes, the expression of KCNJ10, which encodes the inward-rectifier potassium channel Kir4.1, was correlated with the epileptogenesis of GRE. It has been reported that gain-of-function mutations in KCNJ10 may lead to hyperexcitability through multiple unconfirmed mechanisms [31]. However, an animal model study suggested that reduced Kir4.1 activity could increase susceptibility to seizures by reducing potassium and glutamate buffering [41]. We inferred that upregulated KCNJ10 expression in patients with GRE was more likely a compensatory measure for cell hyperexcitability rather than the cause of GRE [42].