Integrating machine learning and retrospective cohort to explore the effects of AEDs on the prognosis of LGG and identify related molecular targets Authors: Ming Zhou, Qinhong Huang, Hui Liang, Hanjie Yang, Min Li Published: 2026 ABSTRACT Antiepileptic drugs (AEDs) were frequently employed in glioma patients, especially those with low-grade glioma (LGG), in which epilepsy manifested in roughly 70–90% of cases. It has been reported that some AEDs can improve the survival of glioma patients. However, the molecular mechanisms of AEDs through which affect LGG prognosis remained unclear. Therefore, this study integrated 105 targets of 10 AEDs, by using machine learning, molecular docking, a retrospective clinical cohort, and in vitro experiments to clarify the biological mechanisms through which AEDs affect LGG prognosis. Our study established a reliable 13-gene AEDs-related LGG prognostic model. Carbamazepine, Oxcarbazepine and Lacosamide were supposed to improve the OS by inhibiting the hazard factor SCN9A. Phenobarbital was supposed to restrict the OS by inhibiting the identified protectors GRIN2C and GRIN3A. Molecular docking visualized the strong affinities between the above drugs and targets. Retrospective cohort further verified our speculation about the effect of the above drugs on the prognosis of LGG. In vitro experiments demonstrated that inhibiting SCN9A, a common targets for most AEDs, could suppress LGG cells malignant behaviors; while the suppression of GRIN2C and GRIN3A enhanced the malignant behaviors of LGG cells. This study provided guidance for individualized AEDs selection for LGG patients, and provide new insights into the potential biological functions and molecular mechanisms of AEDs affecting LGG.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40360-026-01167-3. Introduction Glioma was the most common primary malignant tumor in the central nervous system, characterized by genetic heterogeneity, high lethality and therapeutic resistance [1–3]. According to the latest diagnosis and treatment guidelines, low-grade gliomas (LGG, usually described as WHO grade 2–3) accounted for 31.8% of all gliomas, with a median overall survival (OS) of 5 to 13 years [4–6]. Although the prognosis of LGG patients was much better than glioblastoma patients, whose median OS was only about 15 months, most LGG patients eventually suffered from relapses, resulting in poor prognosis [7–9]. Secondary epilepsy caused by glioma was defined as glioma-related epilepsy (GRE) [10]. As a common initial symptom or concomitant symptom of glioma, GRE seriously affected the quality of life of glioma patients and brought a serious burden to patients and their families [11, 12]. It was reported that GRE was much frequent in LGG patients with an incidence of about 80% [13–15]. Considering the high risk of epilepsy in glioma patients, the use of AEDs was quite common in clinical practice regardless therapy or prevention [16, 17]. Currently, the commonly used AEDs in clinical mainly included: Valproic acid (VPA), Carbamazepine (CBZ), Phenobarbital (PB), Phenytoin (PHT), Ethosuximide (ESX), Oxcarbazepine (OXC), Lamotrigine (LTG), Topiramate (TPM), Levetiracetam (LEV) and Lacosamide [18]. Clinically, many LGG patients needed to take AEDs for a long time, but clinical selection of AEDs for LGG patients lacked specific guidance. Studies have revealed that the long-term use of AEDs may impact the survival of glioma patients [19, 20]. A study involving 825 patients demonstrated that the median survival time of patients using LEV was longer than that of non-users (17 months vs. 15 months, p < 0.05) [21]. Besides, the usage of VPA was found to prolong the survival time of glioma patients (Mantel-Cox log-rank test 17.506, p < 0.001) [22]. Other studies have suggested that AEDs can exert a potential anti-tumor effect by targeting signaling pathways related to cell proliferation, apoptosis or invasion/migration [23, 24], or by increasing the sensitivity of glioma cells to chemotherapy and radiotherapy, achieving anti-tumor effects [25, 26]. However, lacking studies explored the specific targets and the molecular mechanisms of these AEDs affecting the prognosis of glioma patients. Therefore, this study constructed a comprehensive multidrug comparative framework, and integrated the drug-target-prognosis relationships, aiming to clarify the potential targets of AEDs affecting the survival of LGG patients, and provide theory support for the clinical selection of AEDs. Our results exhibited the different effects of AEDs on the prognosis of LGG, provided potential guidance for clinical rational selection of AEDs, and revealed the potential targets and biological mechanisms of AEDs affecting the survival of LGG patients. Materials and methods Data sources From the DrugBank (https://go.drugbank.com/), we obtained ten commonly prescribed AEDs along with their known protein targets (Supplemental Table 1). The targets of each AEDs were identified and confirmed by the previous published studies. RNA sequencing data and clinical information of LGG were retrieved from The Cancer Genome Atlas (TCGA, https://portal.gdc.cancer.gov), and the Chinese Glioma Genome Atlas (CGGA, http://www.cgga.org.cn/). Specifically, the TCGA dataset contributed 479 LGG samples, the CGGA mRNAseq-693 dataset (CGGA-1) provided 431 LGG samples, and the CGGA mRNAseq-325 dataset (CGGA-2) supplied 172 LGG samples. Normalization of the mRNA sequencing data was performed with the R package “limma”. The sequencing data were transformed into transcripts per kilobase million (TPM) values, and a log2 transformation (value+1) was applied before downstream analyses. No interpolation processing was performed on all the data. Clinical cohort characteristics A retrospective study cohort was assembled from 3,627 patients who underwent surgery and chemoradiotherapy at Zhujiang Hospital of Southern Medical University. We followed up on clinical information for 199 individuals diagnosed with LGG (Supplemental Table 2). The cohort was divided into five groups according to the medication history: Carbamazepine group (who only used Carbamazepine up to 3 months, n = 25), lacosamide group (who only used lacosamide up to 3 months, n = 47), Oxcarbazepine group (who only used Oxcarbazepine up to 3 months, n = 43). Phenobarbital group (who only used Phenobarbital, n = 20) and Control group (who did not use AEDs before and after surgery, or only used AEDs less than 1 month for prevention after surgery, n = 63). Patients with a history of AED use had to meet the following inclusion criteria [1]: a diagnosis of LGG confirmed by preoperative imaging and postoperative pathology [2]; clear documentation of preoperative epilepsy and clinical signs [3]; Electroencephalogram (EEG) findings showing epileptic discharge patterns characteristic of LGG [4]; complete medical records allowing retrieval of antiepileptic drug usage [5]; receipt of postoperative chemoradiotherapy; and [6] use of a single antiepileptic drug regimen. The main exclusion criteria comprised a diagnosis of glioblastoma, absence of epileptic symptoms during the disease course, or unknown medication status. From the electronic medical record system, we extracted patients’ primary symptoms, basic clinical information, drug use, and molecular pathological features. Tumor grading followed the World Health Organization (WHO) 2021 Classification of Central Nervous System Tumors. Functional analysis Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment and Gene Ontology (GO) analyses were carried out on 105 AED target genes using the R package “clusterProfiler”. Pathways with a false discovery rate (FDR) below 0.25 and a p-value under 0.05 were deemed statistically significant and were visualized with the ‘ggplot2’ package (Supplemental Table 3). A preliminary protein-protein interaction (PPI) network was built via the STRING database (https://cn.string-db.org/). Cox proportional hazard regression analysis, decision curve analysis and calibration plots Univariate Cox proportional hazard regression analysis was used to screening potential prognostic AEDs targeted genes for subsequent model construction. Multivariate Cox proportional hazard regression analysis was then performed to assess the independent prognostic value of the variates. Results were presented as hazard ratio (HR), 95% confidence interval (95%CI) and corresponding p-value. A p-value less than 0.05 was considered statistically significant. All survival analyses were conducted using the R package “survival”. The decision curve analysis (DCA) was performed using R package “rmda v1.6”. Calibration plots (1-, 3-, and 5-year) were conducted to ensure the accuracy of the nomogram using the R package “pec v2.5.4”. The establishment of AEDs-related prognostic model Leveraging10 excellent machine learning algorithms, which included “Lasso, generalized boosted modeling regression (GBM), Ridge, elastic network (Enet), stepwise Cox, CoxBoost. Random survival forest (RSF), supervised principal component (SuperPC), methods Partial least squares regression for Cox (plsRcox), survival support vector machine (survival-SVM)”, we developed over 100 distinct models. The machine learning process were conducted according to the guidance of the representive article [27]. The original code was obtained from https://github.com/l-magnificence/Mime. The TCGA cohort served as the training set, whereas the CGGA-1 and CGGA-2 cohorts were used for validation. The validation cohorts were independent and not used for iterative model selection. To avoid overfitting, feature selection employed a 10-fold cross-validation method for training (cv.folds = 10). For each cohort, Harrell’s concordance index (C-index) was computed; the “Mean C-index” reflected the average C-index across the two independent validation cohorts. The model achieving the highest Mean C-index was selected as the optional one. The sum of the product of the importance of each gene and its expression level was the Risk score (RS) of patients. RS = Importance (Gene1) * Expression (Gene1) + Importance (Gene2) * Expression (Gene2) + … + Importance (GeneX) * Expression (GeneX). The importance of the gene larger than 0 indicated the gene may be a hazard factor of the prognosis, on the contrary, the the importance of the gene smaller than 0 indicated that the gene may be a protector of the prognosis. Kaplan-Meier (K-M) analysis and Receiver Operating Characteristic (ROC) analysis Differences in prognosis among subgroups were evaluated using K-M analysis (R package ‘survival’), with statistical significance assessed by the log-rank test. Time-dependent receiver operating characteristic (ROC) analysis was employed to evaluate the predictive effects of RS on 1-year, 3-year and 5-year survival using “timeROC” R package. The visualization of results was generated using the “ggplot2” R package. Molecular docking Molecular docking was used to visualize the specific interaction sites between each drug and its protein target. The 3D protein structures corresponding to SCNA9 (pdb_000021tp), GRIN2C (pdb_00007yfg) and GRIN3A (pdb_00002rc9) were downloaded from the RCSB database (https://www.rcsb.org/) in PDB format. Molecular structure files for Carbamazepine (Compound CID: 2554), Lacosamide (Compound CID: 219,078), Oxcarbazepine (Compound CID: 34,312) and Phenobarbital (Compound CID: 4763) were downloaded from the PubChem database (https://pubchem.ncbi.nlm.nih.gov/) in SDF format. Molecular docking analyses were performed via the CB-Dock2 website (https://cadd.labshare.cn/cb-dock2/php/index.php). Binding energies were calculated and expressed as Vina scores (kcal/mol). Cell culture, RNA extraction, cDNA synthesis, and quantitative real-time PCR LGG cell lines SW1783 and SW1088, and normal human astrocyte cell line HA were used to verify the relative expression of SCN9A. All three lines were provided by Procell (Wuhan, China) and cultured in Dulbecco’s modified Eagle medium (DMEM; Gibco) supplemented with 10% fetal bovine serum (FBS; Procell) and 10 µL/mL penicillin-streptomycin solution (100×; Biosharp). Cells were maintained at 37 °C in 5% CO₂incubator. Total RNA was extracted from cell lines using AG RNAex Pro reagent (Accurate Biology, #cat:AG21102), strictly following the manufacturer’s instructions. SYBR Green SupTaq HS premix qPCR kit (Accurate Biology, AG11762) was used for gDNA removal, cDNA synthesis and PCR amplification. After preparing the reaction mixture, amplification was run on an ABI QuantStudio 1 Plus real‑time PCR system. Primer sequences for qRT-PCR were used: GAPDH, F: 5’-GGTGGTCTCCTCTGACTTCAAC-3’, R: 5’-GTTGCTGTAGCCAAATTCGTTGT-3’; SCN9A, F: 5’-TTCATGACCTTGAGCAACCC-3’, R: 5’-TCTCTTCGAGTTCCTTCCTG-3’. GRIN2C, F: 5’-GGGATCTGCCATAACGAGAAG-3’, R: 5’-GCACTGAGTGTCGAAGTTTCCA-3’. GRIN3A, F:5’-TTTTGCTCTTCCCACACACGCCCCCAGGT-3’. R: 5’-TTTTACCTGGGGGGCGTGTGTGGGAAGAGC-3’. Gene expression levels were calculated using the 2−ΔΔCt method. Lentiviral transfection Lentiviruses (GeneChem, Shanghai, China) with the following sequences were used: sh1: GCCCTCATTGAACAACGCATT, sh2: GCTGATTTGATTGAAACGTAT. Following optimization according to the manufacturer’s protocol, a multiplicity of infection (MOI) of 5 was chosen for transducing SW1088 and SW1783 cells. Cells were cultured under standard conditions, and those in good growth phase were selected for lentiviral transduction. After adding the virus‑containing medium, cells were incubated at 37 °C with 5% CO₂. After 24 hours, the virus-free medium was replaced to continue incubation. Puromycin (2 μg/mL) was used to screen the stable strains for three weeks. Knockdown efficiency was finally assessed by qRT-PCR, and the cell lines with stable knockdown were selected for subsequent experiments. siRNA transient transfection The siRNA transfection kits for the GRIN2C and GRIN3A genes were purchased from GenePharma (China, Shanghai). The gene sequences were as follows: siGRIN2C#1: 5’- CCAGGCCAAUGUGCUGAAGAUTT-3’siGRIN2C#2: 5’- CGTAGACTTCTCTGTACCCTT-3’; siGRIN2C#NC: 5’-CCTAAGGTTAAGTCGCCCTCG-3’,siGRIN3A#1:5’-CGCCAACATATCCGAGCTAAT-3’,siGRIN3A#2: 5’-CCTCCTCCTTACCCAGAATAA-3’; siGRIN3A#NC: 5’-CCTAAGGTTAAGTCGCCCTCG-3’. The siRNAs were transfected into SW1088 and SW1783 glioma cells using Lipofectamine 3000 according to the manufacturer’s instructions. Meanwhile, cells transfected with siRNA without homology to the target gene sequence were used as negative controls. SW1088 and SW1783 cells were transfected with the plasmid at a density of 1 × 10^5 cells per well in a 6-well plate. After 72 hours, cellular RNA was extracted for qRT-PCR verification of the target gene knockdown efficiency, with experiments repeated at least three times. Subsequent functional assays were conducted only after confirming that the target gene knockdown efficiency exceeded 50%. CCK-8 proliferation Assay For CCK-8 proliferation experiment, cells in the logarithmic growth phase were trypsinized and resuspended into single-cell suspensions on day 1. Cells were seeded into 96-well plates at 2,000 cells per well (five replicates per group). After the cells were adherent, 100 uL of complete medium containing 10% CCK-8 reagent (APExBIO, cat: K1018-5) was added to each well. Following a 2-hour incubation, absorbance was measured with a microplate reader (BioTek Synergy HT, USA). The assay was repeated every 48 hours for 7 days.The experimental data were visualized using Prism9, with significant differences *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. * Indicated statistical significance. EdU incorporation assay The 5-ethynyl-20-deoxyuridine (EdU) incorporation assay was performed with an EdU detection kit (Beyotime, cat: C0078S). Briefly, LGG cells were seeded in 96-well plates at 3000 cells per well and cultured for 24 hours. Then 10 μM EdU reagent was added to each well and incubated at 37 °C for 2 hours. Subsequently, after fixation, permeabilization, and fluorescence staining procedures, cell images were captured using a laser scanning confocal microscope (Olympus). Statistical analysis was performed using ImageJ software, with data visualization presented through Prism 9. Wound healing assay Cell migration capacity was evaluated using an in vitro wound healing assay. Briefly, SW1088 and SW1783 were seeded in 6-well plates (Corning, Cat# 3516) at 5 × 105 cells/well and cultured until 100% confluency. To eliminate proliferation effects, cells were pretreated with 10 μg/mL mitomycin (Sigma, Cat# M4287) for 2 hours. Uniform wounds (approximately 1000 μm wide) were created with a sterile 200 μL pipette tip, followed by three washes with PBS.Cells were maintained in Free-serum medium. Wound images at identical locations were captured at 0, 12, and 24 h using an inverted microscope (Nikon Eclipse Ts 2, ×10). Image processing was performed using Photoshop. Transwell invasion assay Migratory and invasive capacities were assessed using Transwell chambers (Corning, 8.0 μm pore size, Cat# 353,097). For invasion assays, the upper chamber was pre-coated with 50 μL Matrigel (BD Biosciences, Cat 354,234; diluted 1:8 in serum-free DMEM) and incubated at 37 °C for 2 hours. SW1088, SW1783cells were serum-starved for 24 h, harvested, and resuspended in serum-free medium. Cells (5 × 104 for migration; 1 × 105 for invasion) in 200 μL serum-free medium were seeded into the upper chamber, while 600 μL medium containing 10% FBS (chemoattractant) was added to the lower chamber. After 16–24 hours of incubation, non-migrated cells in the upper chamber were removed. The migrated/invaded cells were fixed with 4% paraformaldehyde and then stained with 0.1% crystal violet. Excess stain was washed away, and images were taken using a Nikon Eclipse Ti2 microscope. Five random fields of view per chamber were counted with ImageJ software. Data analysis All statistical analyses were performed with GraphPad Prism 9.0 (San Diego, CA). Normality and homogeneity of variance were confirmed by Shapiro-Wilk and Levene’s tests, respectively. Between-group differences were analyzed by one-way ANOVA with Sidak’s multiple comparisons test. Survival data were compared using the log-rank test and presented as Kaplan-Meier curves. Correlations were assessed via Pearson’s coefficient (*r*). Categorical variables were examined with the χ2 test and Yates’ continuity correction. Data are expressed as mean ± standard deviation (SD) unless otherwise indicated. Significance levels: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. Result Identification and functional analysis of AEDs-related targets In order to explore the potential targets and biological mechanisms of AEDs, 10 most commonly used AEDs were selected, including five traditional AEDs: Valproate (VPA), Carbamazepine (CBZ), Phenobarbital (PB), Phenytoin (PHT), Ethosuximide (ESX), and five new AEDs: Oxcarbazepine (OXC), Lamotrigine (LTG). Topiramate (TPM), Levetiracetam (LEV), Lacosamide. The proven functional targets of the above AEDs were obtained from DrugBank, and a total of 105 targets were obtained (Supplementary Table 1). KEGG analysis displayed that these 105 targets were mainly involved in Neuroactive ligand-receptor interaction, Calcium signaling pathway, cAMP signaling pathway, MAPK signaling pathway and other signaling pathways (Fig. 1A). As for Gene Ontology,GO-BP mainly involved chemical synaptic transmission, chloride transmembrane transport, sodium ion transmembrane transport; GO-CC included plasma membrane, postsynaptic membrane and transmembrane transporter complex; GO-MF included extracellular ligand-gated monoatomic ion channel activity, GABA-A receptor activity and other signaling receptor activity (Fig. 1B). PPI network provided the landscape of drug-target interactions (Fig. 1C).Fig. 1Functional annotation of AEDs-related targets. (A): KEGG analysis of 105 targets, the complete results were detailed in supplement Table 2, and the objects of interest and p < 0.05, FDR < 0.25 were selected for display. (B): GO analysis of 105 targets, the complete results were detailed in supplement Table 2, and the objects of interest and p < 0.05, FDR < 0.25 were selected for display. (C): The PPI network visualized the interactions of 10 AEDs and their 105 targets Functional annotation of AEDs-related targets. (A): KEGG analysis of 105 targets, the complete results were detailed in supplement Table 2, and the objects of interest and p < 0.05, FDR < 0.25 were selected for display. (B): GO analysis of 105 targets, the complete results were detailed in supplement Table 2, and the objects of interest and p < 0.05, FDR < 0.25 were selected for display. (C): The PPI network visualized the interactions of 10 AEDs and their 105 targets Construction and validation of AEDs-related prognostic model To explore the effect of the above AEDs targets on the prognosis of LGG, we performed univariate COX analysis on these 105 targets, and 59 AEDs targets were identified with potential prognostic effects on LGG (Fig. 2A). Then, these 59 potential LGG prognostic AEDs targets genes were included for machine learning algorithms to establish a AEDs-related prognostic model (Fig. 2B). The results exhibited that 117 combined models were established and “CoxBoost + SuperPC” reached the highest Mean C index (Mean C-index = 0.75695). To further enhance the reliability of model selection, we included the models with the top ten C-index into the Decision Curve Analysis (DCA) analysis, and the result suggested that “CoxBoost + SuperPC” reaching the highest decision-making efficiency (Supplemental Figure 1A). Besides, calibration plots further demonstrated the relative high degree of matching between the predicted prognosis and the real prognosis of patients of the model (Supplemental Figure 1B). These results consistently supported “CoxBoost + SuperPC” to be the optional model.Fig. 2The construction and validation of AEDs-related prognostic models. (A): Results of the univariate COX analysis on the 105 targets; (B): The summary of 117 combinations of 10 machine learning algorithms for the AEDs-related prognostic models. The C-index of each model was calculated across TCGA training cohort and CGGA validation cohorts. The Mean C-index of 2 CGGA validation cohorts was also calculated. (C) The importance of the 13 features including in the AEDs-related prognostic model based on the combination of CoxBoost + SuperPC. (D-F) Kaplan-Meier curves of OS between LGG patients in high- and low-RS subgroups in the TCGA, CGGA-1 and CGGA-2 cohorts, respectively. (G-I) The model prediction values of 1-, 3- and 5-year OS in the TGGA, CGGA-1 and CGGA-2 cohorts, respectively The construction and validation of AEDs-related prognostic models. (A): Results of the univariate COX analysis on the 105 targets; (B): The summary of 117 combinations of 10 machine learning algorithms for the AEDs-related prognostic models. The C-index of each model was calculated across TCGA training cohort and CGGA validation cohorts. The Mean C-index of 2 CGGA validation cohorts was also calculated. (C) The importance of the 13 features including in the AEDs-related prognostic model based on the combination of CoxBoost + SuperPC. (D-F) Kaplan-Meier curves of OS between LGG patients in high- and low-RS subgroups in the TCGA, CGGA-1 and CGGA-2 cohorts, respectively. (G-I) The model prediction values of 1-, 3- and 5-year OS in the TGGA, CGGA-1 and CGGA-2 cohorts, respectively The model “CoxBoost + SuperPC” included CA2, OGDH, HDAC1, CA3, CHRNA1, CD274, HGF, SCN9A, ALDH5A1, GABRD, GRIK4, GRIN2C and GRIN3A. The prognostic risk scoring (RS) formula of the model was: RS = 3.64*CA2 + 3.13*OGDH − 3.87*ALDH5A1 - 3.03*GABRD + 3.53*HDAC1 - 2.92*GRIK4 - 1.98*GRIN2C + 4.4*CA3 - 3.05*GRIN3A + 4.41*CHRNA1 + 3.81*CD274 + 3.68*HGF + 1.98*SCN9A (Fig. 2B, C). Then, according to the median RS, LGG patients were divided into high-RS and low-RS subgroups, K-M analysis revealed that the AEDs-related prognostic model can well stratify patients with different prognosis in all the cohorts. Patients in high-RS subgroup generally suffered from poorer prognosis that those in low-RS subgroup (Fig. 2D–F). Time-dependent ROC analysis was performed to validate the prediction value of the AEDs-related prognostic model. The TCGA cohort exhibited AUC values of 0.87 for 1-year OS, 0.87 for 3-year OS and 0.80 for 5-year OS (Fig. 2G). In validation cohorts, CGGA1 cohort generated AUC values of 0.71 for 1-year OS, 0.79 for 3-year OS and 0.80 for 5-year OS (Fig. 2H); while CGGA2 generated AUC values of 0.82 for 1-year OS, 0.86 for 3-year OS and 0.88 for 5-year OS (Fig. 3I). To further adjust the influence from the clinical variables such as age, tumor grade, 1p/19q status, MGMT status and IDH status, multivariate Cox regression analyses was used and displayed that RS were independent prognostic factor for LGG patients (Supplemental Figure 1C). Moreover, DCA suggested that the net benefit of the predictive strategy based on the RS significantly exceeded than that of the predictive strategy based on the above clinical variables(Supplemental Figure 1D). All these results supported that the excellent performance of AEDs-related prognostic model in survival prediction.Fig. 3Survival analysis on the 13 AED-related prognostic features in TCGA dataset. (A-H): K-M curves of OS between LGG patients with different expression levels of CA2, OGDH, HDAC1, CA3, CHRNA1, CD274, HGF, and SCN9A, respectively. High expressions of these targets were associated with poorer prognosis. (I-M): KM curves of OS between LGG patients with different expression levels of ALDH5A1, GABRD, GRIK4, GRIN2C and GRIN3A, respectively. High expression of these targets were associated with better prognosis Survival analysis on the 13 AED-related prognostic features in TCGA dataset. (A-H): K-M curves of OS between LGG patients with different expression levels of CA2, OGDH, HDAC1, CA3, CHRNA1, CD274, HGF, and SCN9A, respectively. High expressions of these targets were associated with poorer prognosis. (I-M): KM curves of OS between LGG patients with different expression levels of ALDH5A1, GABRD, GRIK4, GRIN2C and GRIN3A, respectively. High expression of these targets were associated with better prognosis The prognosis effect of the 13 AEDs-related prognostic targets To further evaluate the effect of 13 targets on the prognosis of LGG patients, we performed K-M analysis. In TCGA cohort, the high expressions of CA2, OGDH, HDAC1, CA3, CHRNA1, CD274, HGF and SCN9A were found to be associated with reduced survival time, indicating that these targets could be hazard factors for the survival of LGG patients (Fig. 3A–H); while the high expressions of ALDH5A1, GABRD, GRIK4, GRIN2C and GRIN3A were found to be associated with increased survival time, indicating that these targets could be protectors for the survival of LGG patients (Fig. 3I–M). Similar results were also obtained from the CGGA1 and CGGA2 validation cohorts (Supplementary Figure 1 and Supplementary Figure 2). These results were consistent with the coefficient of the RS formula in AEDs-related prognostic model, verifying the reliability of the model. Besides, due to the different prognostic effects of these AEDs targets, the choice of AEDs for LGG patients needed to be reconsidered with caution. Evaluation of the potential effects of AEDs on LGG survival Because different AEDs had different targets, and our above analysis suggested the identified prognostic AEDs targets had different prognostic effects on LGG, we conducted sankey map to visualize the “drug - target - prognostic effect” axis (Fig. 4A). It can be found that only Valproate, Carbamazepine, Phenobarbital, Oxcarbazepine, Lamotrigine. Topiramate and Lacosamide acted on the identified prognostic targets. Among them, Carbamazepine, Oxcarbazepine and Lacosamide were inhibitors of SCN9A, an identified prognostic hazard factor, which may be beneficial to the survival of LGG patients. However, Phenobarbital were inhibitor of GRIN2C and GRIN3A, the identified prognostic protectors, which may do harm to the survival of LGG patients. Valproate, Topiramate and Lamotrigine acted on both the prognostic hazard factors and protectors, and therefore their prognostic effects cannot be inferred in our analysis.Fig. 4Molecular docking analysis. (A): Sankey diagram revealed the pharmacological mechanism between 7 AEDs and the 13 prognostic targets. (B, C): Molecular docking result of GRIN2C/GRIN3A and Phenobarbital with the lowest binding energy. (D-F): Molecular docking result of SCN9A and Carbamazepine/Oxcarbazepine/Lacosamide with the lowest binding energy Molecular docking analysis. (A): Sankey diagram revealed the pharmacological mechanism between 7 AEDs and the 13 prognostic targets. (B, C): Molecular docking result of GRIN2C/GRIN3A and Phenobarbital with the lowest binding energy. (D-F): Molecular docking result of SCN9A and Carbamazepine/Oxcarbazepine/Lacosamide with the lowest binding energy Further, we chose Carbamazepine, Oxcarbazepine, Lacosamide, Phenobarbital and their targeted prognostic genes for molecular docking to verify their interactions. The Vina Scores were −6.9 kcal/mol for Phenobarbital-GRIN2C, −7.3 kcal/mol for Phenobarbital-GRIN3A, −9.8 kcal/mol for Carbamazepine-SCN9A, −9.5 kcal/mol for Oxcarbazepine-SCN9A and −7.0 kcal/mol for Lacosamide-SCN9A (Fig. 4B–F). All the Vina Scores were lower than −5 kcal/mol, indicating that these 4 AEDs can spontaneously bind to the targets with strong affinity. Validation of the effects of 4 AEDs on LGG survival via retrospective cohort To further verify the above speculation on the effects of the above 4 AEDs (Carbamazepine, Oxcarbazepine, Lacosamide, Phenobarbital) on the prognosis of LGG, we collected a retrospective cohort consisted of 199 LGG patients with detailed survival information, AEDs usage history and treatment information (Supplemental Table 2). According to the AEDs usage history, patients were divided into five groups: Carbamazepine group (n = 25), Lacosamide group (n = 47), Oxcarbazepine group (n = 44), Phenobarbital group (n = 20) and Control group (n = 63) (Fig. 5A). K-M analysis illustrated that, when compared to Control group, the HRs were 0.48 for Oxcarbazepine group (p = 0.01. Fig. 5B), 0.48 for Carbamazepine group (p=0.05. Fig. 5C) and 0.25 for Lacosamide group (p = 1.5e-4. Fig. 5D), indicating that the usage of Carbamazepine, Lacosamide or Oxcarbazepine was associated with the prolonged OS of LGG patients; while the HR for Phenobarbital group was 2.59 (p=0.02) when compared to Control group (Fig. 5E), suggesting that the usage of Phenobarbital was associated with the decreased OS of LGG patients.Fig. 5The different OS of LGG patients under different medication groups. (A): K-M curve of OS between Control group, Carbamazepine group, lacosamide group Oxcarbazepine group and Phenobarbital group. (B-E): The K-M curves of OS between Control group and Carbamazepine group/Lacosamide group/Oxcarbazepine group/Phenobarbital group, respectively. (F): Multivariate Cox regression analyses to adjust the inflence of the usage of AEDs on OS The different OS of LGG patients under different medication groups. (A): K-M curve of OS between Control group, Carbamazepine group, lacosamide group Oxcarbazepine group and Phenobarbital group. (B-E): The K-M curves of OS between Control group and Carbamazepine group/Lacosamide group/Oxcarbazepine group/Phenobarbital group, respectively. (F): Multivariate Cox regression analyses to adjust the inflence of the usage of AEDs on OS Further, multivariate Cox regression analyses was conducted to control the inflence from confounding factors, such as tumor grade and the known prognostic molecular status. Result display that the HRs were 0.44 for Oxcarbazepine group (p = 8.3e-3), 0.51 for Carbamazepine group (p = 0.08) and 0.26 for Lacosamide group (p = 8.4e-4) while the HR for Phenobarbital group was 2.62 (p = 0.03). These results indicated that the usage of Lacosamide, Oxcarbazepine and Phenobarbital could be independent prognostic factor for LGG patients, and the prognostic effect of Carbamazepine may be slightly affected by some clinical factors (Fig. 5F). Identification of SCN9A as promising therapy target for LGG To further evaluate the independent prognostic effects of 13 targets, we performed a multivariate Cox analysis. Specially, it was found that SCN9A, as target for most AEDs (Valproic acid, Carbamazepine, Topiramate, Lacosamide, Lamotrigine and Oxcarbazepine), has independently prognostic effect on the survival of LGG (Fig. 6A). Considering these AEDs all exerted inhibitory effects on SCN9A to perform biological functions, we constructed stable SCN9A knockdown LGG cell lines (SW1088 and SW1783) to explore the its biological functions. qRT-PCR results verified the knockdown efficiency in SW1088 and SW1783 (Fig. 6B, C). The CKK-8 assay and EDU incorporation assay consistently exhibited that the proliferation of LGG cell lines was significantly impeded when down-regulating SCN9A (Fig. 6D, G). Transwell assays suggested that the down-regulation of SCN9A impaired both the migration and invasion ability of LGG cells (Fig. 6H, I). Wound healing assay further verified the impact of SCN9A on LGG cells migration (Fig. 6J, K). These results suggested that AEDs could inhibit the malignant biological behaviors of LGG via SCN9A, improving the prognosis of LGG patients.Fig. 6Targeting SCN9A inhibits the malignant biological behavior of LGG cells. (A): Multivariate COX regression analysis on 13 prognostic targets. (B, C): qRT-PCR showed the knockdown efficiency of SCN9A in SW1088 and SW1783 (n = 3). (D, E): Results of CCK8 assays by down-regulating the SCN9A. (F, G): EdU incorporation assay, EDU was shown in red, DAPI in blue (scale bar = 400 μm). (H, I): Transwell assay showed the effect of SCN9A on LGG cells migration and invasion (scale bar = 400 μm). (J, K): Wound healing assay to evaluate the migration ability of LGG cells after inhibiting SCN9A (scale bar = 400 μm) Targeting SCN9A inhibits the malignant biological behavior of LGG cells. (A): Multivariate COX regression analysis on 13 prognostic targets. (B, C): qRT-PCR showed the knockdown efficiency of SCN9A in SW1088 and SW1783 (n = 3). (D, E): Results of CCK8 assays by down-regulating the SCN9A. (F, G): EdU incorporation assay, EDU was shown in red, DAPI in blue (scale bar = 400 μm). (H, I): Transwell assay showed the effect of SCN9A on LGG cells migration and invasion (scale bar = 400 μm). (J, K): Wound healing assay to evaluate the migration ability of LGG cells after inhibiting SCN9A (scale bar = 400 μm) Validation the protective effect of GRIN2C and GRIN3A on LGG Considering Phenobarbital acted on both GRIN2C and GRIN3A, and our analysis and previous studies consistently suggested that GRIN2C and GRIN3A were potential protector for glioma [28–30], we conducted vitro assays for further validation. GRIN2C or GRIN3A knockdown LGG cell lines (SW1088 and SW1783) were constructed. qRT-PCR results verified the knockdown efficiency in SW1088 and SW1783 (Fig. 7A). Transwell assays suggested that the down-regulation of GRIN2C or GRIN3A impaired both the migration and invasion ability of LGG cells (Fig. 7B–C). The EDU incorporation assay exhibited that the proliferation of LGG cell lines was significantly impeded when down-regulating GRIN2C or GRIN3A (Fig. 7D–E). These results suggested that the neagative association between the usage of Phenobarbital and OS of LGG patients may be due to the inhibition of GRIN2C and GRIN3A.Fig. 7Potential anti-glioma function of GRIN2C and GRIN3A in vitro. (A): qRT-PCR showed the knockdown efficiency of GRIN2C and GRIN3A in SW1088 and SW1783, respectively (n = 3). (B): Transwell assay showed the effect of GRIN2C and GRIN3A on the migration and invasion of SW1088 (scale bar = 400 μm). (C) Transwell assay showed the effect of GRIN2C and GRIN3A on migration and invasion of SW1783 (scale bar = 400 μm). (D) EdU incorporation assay showed the effect of GRIN2C and GRIN3A on the proliferation of SW1088 (scale bar = 400 μm). EDU was shown in red, DAPI in blue. (E) EdU incorporation assay showed the effect of GRIN2C and GRIN3A on the proliferation of SW1783 (scale bar = 400 μm). EDU was shown in red, DAPI in blue Potential anti-glioma function of GRIN2C and GRIN3A in vitro. (A): qRT-PCR showed the knockdown efficiency of GRIN2C and GRIN3A in SW1088 and SW1783, respectively (n = 3). (B): Transwell assay showed the effect of GRIN2C and GRIN3A on the migration and invasion of SW1088 (scale bar = 400 μm). (C) Transwell assay showed the effect of GRIN2C and GRIN3A on migration and invasion of SW1783 (scale bar = 400 μm). (D) EdU incorporation assay showed the effect of GRIN2C and GRIN3A on the proliferation of SW1088 (scale bar = 400 μm). EDU was shown in red, DAPI in blue. (E) EdU incorporation assay showed the effect of GRIN2C and GRIN3A on the proliferation of SW1783 (scale bar = 400 μm). EDU was shown in red, DAPI in blue Discussion Among LGG patients, the incidence of epilepsy range from 60% to 85% [31–33], which negatively impacted their quality of life, neurocognitive function and overall health status [34]. AEDs, as an important treatment of managing epilepsy, were widely used in patients with glioma-related epilepsy [35–37]. Previous studies have reported that the long-term usage of AEDs may improve the prognosis of glioma patients [14, 20, 38], but lack exploration of the underlying biological mechanisms. This study systematically evaluated the potential impacts of AEDs on the survival of LGG patients at molecular level via multi-dimensional approaches, revealed the bidirectional effects of AEDs on the OS, and provided an important theoretical reference for the clinical selection. Our study integrated AEDs targets, LGG transcriptomics data and relevant clinical information to construct a reliable AEDs-related LGG prognostic model via machine learning algorithms, providing evidence that the usage of AEDs can impact the survival of LGG patients via specific targets. The AEDs-related LGG prognostic model was consisted of 13 genes, including CA2, OGDH, HDAC1, CA3, CHRNA1, CD274, HGF, SCN9A, ALDH5A1, GABRD, GRIK4, GRIN2C and GRIN3A, exhibiting robust prediction value. Among these 13 prognostic AEDs target genes, CA2, OGDH, HDAC1, CA3, CHRNA1, CD274, HGF and SCN9A were shown to be hazard factors, while ALDH5A1, GABRD, GRIK4, GRIN2C and GRIN3A were shown to be protectors. As for their functions on glioma, previous studies have reported that CA2 was highly expressed in glioma and involved in TMZ resistance [39, 40]; HDAC1 participated in the maintenance of malignant phenotype and promote the mesenchymal transformation of glioma stem cells [41]; OGDH and HGF contributed to the progression of glioma through promoting malignant biological behaviors [42, 43]. On the contrary, the up-regulations of GABRD, GRIK4, GRIN2C and GRIN3A were reported to restrict glioma via glutamate excitotoxicity [28–30]; while the down-regulation of ALDH5A1 was associated with poor prognosis in patients with glioblastoma [44]. These studies provided reliable support and verification for our analysis results. Most importantly, the “drug-target-effect” axis displayed that only 7 of the 10 AEDs could target these 13 prognostic targets. Among these drugs, Carbamazepine, Lacosamide and Oxcarbazepine only acted on the hazardous factor: SCNA9, while Phenobarbital only acted on the protectors: GRIN2C and GRIN3A, suggesting the long-term usage of Carbamazepine, Lacosamide, and Oxcarbazepine could be beneficial for LGG patients, and the long-term usage of Phenobarbital may restrict the survival of LGG patients, highlighting the selection of types of AEDs should be more cautious. Molecular docking analysis displayed the strong affinity between the 4 AEDs and the prognostic targets. Meanwhile, the follow-up results of our retrospective cohort further verified the above analytical inferences. The usage of Carbamazepine, Lacosamide or Oxcarbazepine was associated with the prolonged OS of LGG patients, while the usage of Phenobarbital was associated with the decreased OS of LGG patients. These results systematically exhibited the effects of different AEDs on the survival of LGG patients and their potential function targets, providing references for the clinical selections of AEDs for LGG patients. Last but not least, considering SCN9A was a target for most AEDs and also acted as an independent prognostic factor for LGG, we conducted in vitro knockdown experiments to explore its biological functions in LGG. Functional assays demonstrated that the down-regulation of SCN9A could inhibit the malignant biological behavior of LGG cells, making it a promising therapy target for LGG. Consistent with our result, it has been reported that interfering the expression of SCN9A can damage the viability of glioblastoma cells both in vitro and in vivo by destroying the proliferation and neuronal projection pathways of glioblastoma cells [45]. These results suggested that the usage of AEDs that had inhibitory function on SCN9A may have potential benefits to LGG patients, further highlighting the usage of Carbamazepine, Lacosamide and Oxcarbazepine could be associated with the better survival of LGG patients. As for GRIN2C and GRIN3A, our study further validated their potential anti-glioma effects. The inhibiton of GRIN2C and GRIN3A resulted in the enhanced malignant behaviors of LGG cells, further highlighting the usage of Phenobarbital could be associated with the poorer survival of LGG patients. Limitation Although the results of bio-informatic analysis and in vitro assays supported our conclusion, several limitations needed to be clarified. First, the cell-based assays can not well reflect the effect of the target genes on glioma in vivo. Therefore, in the follow-up study, we will focus on the specific mechanism of the identified targets, especially SCN9A, on glioma in vivo. Second, because the included cohort was a retrospective cohort, it was hard to accurately know and quantify patients seizure severity at that time; and the medical record system only records the patients’ medication during hospitalization. However, epilepsy medication often changed the dosage, frequency and course of treatment due to possible changes in the patient’s condition after discharge, so it was difficult to accurately evaluate the impact of related variables on the patient’s prognosis. Therefore, a well designed prospective study was in urgent need. Third, simple molecular docking results were difficult to reflect the real biological activity. Therefore, experimental validation such as surface plasmon resonance or isothermal titration calorimetry were necessary. Last but not least, lacking enough studies explored how the identified key targets affect glioma progression, particularly through ion-channel signaling, neuron–glioma interactions, and the tumor microenvironment, making it difficult to provide in-depth discussion, more relevant studies would be carried out in our follow-up study. Conclusion In conclusion, by integrating bio-informatics analysis, retrospective cohort and functional assays, this study constructed a 13-gene AEDs-related LGG prognostic model, provided evidence for the potential benefits of long-term usage of Carbamazepine, Lacosamide and Oxcarbazepine in LGG patients, and identified SCN9A as a promising therapeutic target for LGG. This study provided a new insights into the impacts of AEDs on the prognosis of LGG patients at molecular level, and supplied a new understanding of AEDs selections for LGG patients in clinical. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2 Supplementary Material 3 Supplementary Material 3 Supplementary Material 4 Supplementary Material 4 Supplementary Material 5 Supplementary Material 5