F98 glioma cells were purchased from the American Type Culture Collection (ATCC). The cells were cultured in Dulbecco’s Modified Eagle Medium: Nutrient Mixture F-12 (DMEM/F12; from PAN Biotech, Aidenbach, Germany) with 10% fetal calf serum (FCS, Bio&SELL, Feucht, Germany). Cell-culturing was done at 37°C in a 5% CO₂ humidified atmosphere. At constant intervals, MycoSPY^®-PCR Mycoplasma Test Kit (Biontex, Munich, Germany) was used to test cell culture supernatants for mycoplasma contamination. During the project period, no contamination with mycoplasmas was detected. Cenobamate tablets (Ontozry, Angelini Pharma, Rome, Italy) were crushed thoroughly and dissolved in 98% ethanol. The suspension was centrifuged and filtered (40 μm) to yield a 120 mM stock solution. In our experiments, the concentration of ethanol in artificial cerebrospinal fluid (aCSF) was 5‰ (v/v).

All procedures were conducted according to national and international guidelines on the ethical use of animals (European Council Directive 86/609/EEC, approval of local authority LALLF M-V/TSD/7221.3–1-020/20). All efforts were made to minimize animal suffering and to reduce the number of animals used. The Fischer 344 rats were housed under environmentally controlled conditions (12 h light/dark cycles, lights switched on from 6 a.m. to 6 p.m., and 40–60% relative humidity). For the present study, a total of 29 Fischer 344 rats were included.

To imitate human brain cancer, F98 glioma cells were stereotactically implanted in the sensorimotor neocortex of male Fischer 344 rats (Charles River, Sulzfeld, Germany) at the age of 8–10 weeks as previously described (Lange et al., 2020). Briefly, the animals were anesthetized with ketamine (100 mg/kg i.p.) and xylazine (10 mg/kg i.p.), and the head was fixed in a stereotactic frame (Narishige, Tokyo, Japan). Following a scalp incision, the skull was freed from extracranial muscles, and a hole of 0.7 mm diameter was manually drilled into the skull in the left parasagittal position (relative to bregma: 1.8 mm posterior, 2.5 mm left, 2 mm deep). Next, F98 cells from a subconfluent-growing culture were prepared for injection with 1 × 10⁵ cells/μL phosphate-buffered saline (PBS; injection rate: 0.5 μL/min; total volume: 10 μL) using a Hamilton syringe (Model 701 N SYR; Hamilton, Reno, Nevada, United States). After injection of the cell suspension, the drill hole was covered with Heliobond (Ivoclar Vivadent, Schaan, Lichtenstein), and the scalp was closed with a suture. Sham-operated animals underwent the same procedure with 10 μL PBS instead of F98 cell suspension. Ten to twelve days after surgery, Fischer rats were sacrificed and subjected to electrophysiological investigations. Based on previous investigations, we know that glioma-bearing Fischer rats exhibit interictal spikes and seizures (Lange et al., 2020; Bouckaert et al., 2020).

For electrophysiological recordings, Fischer 344 rats were deeply anesthetized with diethyl ether (Thermo Fisher Scientific, Waltham, MA, United States) and subsequently decapitated. The brain was rapidly extracted and immediately transferred into a chilled, oxygenated dissection solution (95% O₂/5% CO₂) containing (in mmol/L) 87 NaCl, 25 NaHCO₃, 2.5 KCl, 1.25 NaH₂PO₄, 0.5 CaCl₂, 7 MgCl₂, 10 D-glucose, and 75 sucrose, adjusted to pH 7.4 with an osmolarity of 320–330 mosmol/L. After removing the cerebellum, coronal brain slices were prepared using a vibratome (LEICA VT1200, Wetzlar, Germany). 500-μm slices were prepared in a chilled solution continuously bubbled with carbogen. The slices were then transferred to a submerged storage chamber and maintained in aCSF equilibrated with carbogen. The aCSF was composed of (in mmol/L) 124 NaCl, 26 NaHCO₃, 3 KCl, 1.25 NaH₂PO₄, 2.5 CaCl₂, 1.5 MgCl₂, and 10 D-glucose, adjusted to pH 7.4 with an osmolarity of 304–312 mosmol/L. The slices were stored for one to 6 h before being transferred to the interface chamber for electrophysiological investigation.

The coronal slices were transferred into an interface chamber (BSC-HT, Harvard Apparatus, Holliston, MA, United States) and continuously perfused with aCSF bubbled with carbogen. The chamber temperature was kept constant at 31 ± 1°C (TC-10, npi electronic GmbH, Tamm, Germany). Under visual control, electrodes were positioned within a rim of 1 mm around the glioma, and field potentials were recorded using conventional aCSF-filled glass micropipette electrodes (Ag/AgCl, with a resistance of approximately 2–5 MΩ). The analog recordings were amplified, low-pass filtered at 30 Hz (EXT-10-2F, npi electronic GmbH), and digitized using a Micro1401 analog-to-digital converter (Cambridge Electronic Design, Cambridge, United Kingdom). Signal 2.16 (Cambridge Electronic Design) was used to save recordings with data sampled at 2,000 Hz. After being introduced into the measurement chamber, slices were perfused using aCSF. Since acute slices do not show spontaneous activity unless they are exposed to disinhibiting conditions, aCSF was modified to evoke spontaneous network disinhibitions. Two different protocols were tested: (I) aCSF with 8 mM KCl and 0 mM MgCl₂ and (II) aCSF containing 8 mM KCl, 0 mM MgCl₂ and 50 μmol/L 4-aminopyridine (4-AP; Tocris, Bristol, United Kingdom). The solutions were adjusted to a pH of 7.4 with an osmolarity of 304–312 mosmol/L. Robust occurrence of seizure-like events (SLEs) was typically reached by minute 80. Data recorded from 100 to 140 min were used for further analysis. We defined SLEs as clearly distinguishable, repetitive spiking events in local field potential (LFP) activity. These events had to exhibit similar morphology, appear in an almost rhythmic fashion for an extended period, last at least 0.2 s, and had to be separated by a minimum of 0.1 s from preceding and subsequent events.

The experimental timeline for naive rats is shown in Figure 1A; the timeline for glioma-bearing and sham-operated animals is illustrated in Figure 2A. We recorded LFPs to evaluate the effect of CNB at 60 and 120 μmol/L in both disinhibition models (Supplementary Figure 1; Figure 1). These comparisons led us to select the disinhibition model (II), containing additional 50 μmol/L 4-AP for all further experiments of the study. The first disinhibition model did not always yield clearly recognizable SLE onsets and terminations. Additionally, compared to initial recordings based on the 4-AP disinhibition model, the electrolyte-based model w/o 4-AP exhibited significantly more pauses in seizure-like activity lasting longer than 2 min (model (I): 10/18 slices vs. model (II): 1/27 slices; Fisher exact test: p < 0.001; Supplementary Figure 2).

We quantified the proportion of recording time occupied by SLEs (referred to as SLE fraction), SLE duration, spike frequency within SLEs, and SLE rate (Supplementary Figure 3). We employed a WaveNet-based deep learning model (Géron, 2019; van den Oord et al., 2016) to delineate SLE onsets and terminations. The model was trained on 30-s segments of human-labeled recordings, downsampled to 125 Hz. Each data point of these segments was labeled as part of an SLE or not. Deep learning was implemented in Python using Keras (3.3.3)/TensorFlow (2.16.1). SLEs, predicted by the model, which were shorter than 0.1 s were omitted, as those most often were interictal events or noise which the model had misidentified to be SLEs. SLEs with an SLE interval of less than 0.2 s were regarded as one SLE. All SLE endpoints were then algorithmically adjusted to align with their respective final extremum (final minimum or maximum, Supplementary Algorithm 1).

The model was used to sequentially label all recordings. We manually reviewed the algorithmically corrected outputs every 3–6 recordings using a custom Python script with an HTML and JavaScript-based frontend for each recording, allowing for manual correction of labels when necessary. We then included these recordings and their labels in the training data. The small model architecture allowed for rapid retraining with the manually corrected batch of recordings. This progressively reduced the amount of human intervention required for each batch of recordings. Training data were augmented by duplicating the original dataset and inverting the sign of the duplicated half. Synthetic noise, scaled to mimic the characteristics observed in actual recordings, was also added. Specifically, 25% of the training data remained unchanged; 25% had an additional, weak, approximately 50 Hz signal mixed with Gaussian noise; 25% had 5-s portions within each 30-s segment replaced by short bursts of Gaussian noise; and the remaining 25% had random sample points replaced with values drawn from a Gaussian distribution. Each 30-s segment was standardized by subtracting its mean and dividing by its standard deviation. Training and validation data were split such that each recording was exclusively assigned to either the training set or the validation set. This was necessary because SLE morphologies within each recording tended to be very similar. The total duration occupied by SLEs and the proportion of CNB-treated recordings were balanced between training and validation sets (Supplementary Table 1). The final model was trained for 100 epochs on approximately 3,500 min of data and validated on 860 min using the binary cross-entropy loss function, achieving an algorithmically corrected (as described above) validation accuracy of 97.4% (Supplementary Table 2). If no improvement in validation accuracy was observed for more than 2 epochs, the learning rate was divided by 10.

Within each event, spikes were detected by identifying local minima using SciPy’s argrelmin function. The order parameter was set so that each data point was evaluated within a 0.165-s window before and after to determine the presence of a local minimum.

We performed a slice-based analysis (2–6 slices from one animal in an experimental group), in which each slice was counted as one data point. In addition, the data were utilized in an animal-based manner (all ex vivo-stimulated slices of one experimental group obtained from one animal were merged).

For the final 40 min of each recording, the power spectral density (PSD) was computed using SciPy’s periodogram function with a Hann window and scaling was set to “density.” Frequency components below or equal to 0 Hz and above 30 Hz were removed, and the remaining PSD values were normalized by dividing each value by the total power within this frequency range. The normalized PSD was then binned into 0.2 Hz intervals. Bootstrapping generated a bootstrap distribution by repeatedly sampling with replacement from the observed distribution of PSDs (considering each recording as one sample). Confidence intervals were subsequently determined from this bootstrap distribution. The sampling process was repeated 10⁴ times.

Data analysis was performed using custom Python (3.11.8 for deep learning and 3.12.3 for all other purposes) scripts. Post hoc statistical tests were conducted in Python and custom graphical user interfaces (GUIs) were implemented either via Tkinter (8.6.13), Flask (3.1.0), HTML5 and JavaScript (≥ES6/ES2015). Statistical analyses were conducted using custom Python scripts, primarily utilizing the SciPy (1.13.1) and scikit-posthocs (0.7.0) libraries. The Fisher exact test was used to test whether the purely electrolyte-based disinhibition model showed more pauses in seizure-like activity. Mean group differences were tested for significance using the nonparametric Kruskal–Wallis test before, for multiple comparisons, subgroups were tested with the post hoc Dunn’s test. Post hoc test results and Fisher-exact test results are indicated as follows: *p < 0.05, **p < 0.01, ***p < 0.001.

The overall aim of the study was to analyze the effect of CNB on a preclinical glioma model, which is associated with tumor-related seizures (Bouckaert et al., 2020; Lange et al., 2021). To this end, coronal slices of naive animals were challenged with two doses of CNB, and seizure-like events (SLE) were analyzed (Figure 1), Kruskal-Wallis test, followed by post hoc Dunn’s test. CNB doses used in this ex vivo study mirror clinically relevant plasma levels used in the treatment of focal epilepsy (Steinhoff et al., 2024). CNB significantly reduced the SLE fraction (relative time occupied by SLEs) to 63% (60 μmol/L CNB, p < 0.01) and to 44% (120 μmol/L CNB, p < 0.001, Figure 1D₁). Additionally, the duration of SLEs was shortened after anticonvulsant exposure to 32% (60 μmol/L CNB) and 27% (120 μmol/L CNB, for both p < 0.001, Figure 1D₂), respectively. As illustrated in Figure 1D₃, spike frequency within SLEs was attenuated by the high dose of CNB to 79% (p < 0.05), an effect not observed at 60 μmol/L. In contrast to the diminishing impact of CNB on the fraction, duration, and spike frequency within an event, the SLE rate was determined to be elevated by anticonvulsant exposure (Figure 1D₄). In comparison to vehicle-treated controls, SLE rate was increased to 191% (p < 0.001) and 151% (p < 0.05), respectively, after treatment with 60 or 120 μmol/L CNB.

Since we found a hyperexcitability-suppressing effect of CNB in naive rats, the experiments were extended to the syngeneic Fischer/F98 glioma model (Figure 2). Here, we focused on experiments using only 120 μmol/L CNB. In slices from sham-operated and glioma-bearing rats, CNB reduced the fraction of time occupied by SLEs in all groups (Figures 2D₁,D₂). In controls, the fraction of SLEs was reduced to 39% (p < 0.01) and 42% (p < 0.05) in the ipsilateral and contralateral hemispheres, respectively. Similarly, in glioma-bearing slices, the proportion of SLE was reduced to 45% (ipsilateral, p < 0.01) and 50% (contralateral, p < 0.05) by CNB incubation.

Consistent with our findings in naive animals, incubation in CNB significantly reduced SLE duration in slices of sham-operated animals (ipsilateral to 29%; contralateral to 30%; p < 0.01; Figure 2E₁). In F98-bearing slices, CNB incubation similarly reduced SLE duration on the contralateral side to 32% (p < 0.01), although this reduction did not reach significance on the ipsilateral side (46%, p = 0.11, Figure 2E₂).

Next, we investigated whether CNB incubation affected spike frequency within SLEs. As illustrated in Figure 2F₁, no significant reduction was found in sham-operated animals (ipsilateral: 65%, p = 0.17, contralateral: 60%, p = 0.056). However, in tumor-bearing slices, spike frequency was significantly reduced in ipsilateral (62%, p < 0.05) and contralateral (57%, p < 0.05) hemispheres (Figure 2F₂). Regarding SLE rate, no significant differences in both experimental cohorts were determined (Supplementary Figure 4).

To verify whether our previous analyses based on SLEs might have missed critical frequency components, we conducted power spectral density (PSD) analyses. Slices were not differentiated according to their ipsilateral or contralateral origin but were classified based on the animals’ experimental conditions: naive, sham-operated, or glioma-implanted. Additionally, slices were categorized according to treatment status: vehicle control or 120 μmol/L CNB.

All three groups, not treated with CNB, exhibited at least two distinct peaks in the normalized, 0.2-Hz-binned PSD. The first peak appeared below 1 Hz, corresponding to the presence of SLEs, while the second peak ranged between 6 and 14 Hz, aligning with the spike frequency typically observed during SLEs (approximately 10 Hz). In contrast, slices treated with CNB showed only a single peak below 1 Hz, and the prominent 10 Hz peak was absent. Furthermore, at ~10 Hz, the mean normalized PSD values for control slices consistently surpassed those of treated slices, with non-overlapping bootstrapped 95% confidence intervals (Figures 3A–C). Interestingly, the non-normalized PSD with 0.2 Hz binning showed higher values between 1 and 5 Hz, with non-overlapping bootstrapped confidence intervals (Supplementary Figure 5).