Hippocampal low-frequency stimulation prevents seizure generation in a mouse model of mesial temporal lobe epilepsy

Mesial temporal lobe epilepsy (MTLE) is the most common form of focal, pharmacoresistant epilepsy in adults and is often associated with hippocampal sclerosis. Here, we established the efficacy of optogenetic and electrical low-frequency stimulation (LFS) in interfering with seizure generation in a mouse model of MTLE. Specifically, we applied LFS in the sclerotic hippocampus to study the effects on spontaneous subclinical and evoked generalized seizures. We found that stimulation at 1 Hz for 1 hr resulted in an almost complete suppression of spontaneous seizures in both hippocampi. This seizure-suppressive action during daily stimulation remained stable over several weeks. Furthermore, LFS for 30 min before a pro-convulsive stimulus successfully prevented seizure generalization. Finally, acute slice experiments revealed a reduced efficacy of perforant path transmission onto granule cells upon LFS. Taken together, our results suggest that hippocampal LFS constitutes a promising approach for seizure control in MTLE.


Introduction
Mesial temporal lobe epilepsy (MTLE) represents the most common form of acquired epilepsy in adults. MTLE is thought to arise from an initial precipitating insult in early childhood, such as status epilepticus (SE), complex febrile seizures, or head trauma (Engel, 2001). The most frequent histopathological hallmark of MTLE is hippocampal sclerosis, which is characterized by neuronal cell loss and gliosis (Blümcke et al., 2013). In addition, it is often associated with granule cell dispersion (GCD) and mossy fiber sprouting (Thom, 2014). MTLE is of particular clinical interest since it is frequently resistant to pharmacological treatment (Engel, 2001). Hence, surgical removal of the seizure focus is an effective therapeutic intervention for many MTLE patients (Englot and Chang, 2014). However, for patients with multiple seizure foci or for those at risk of resection-related impairments, this treatment is not an option.
In MTLE with hippocampal sclerosis, however, the efficacy of HFS is rather variable between patients (Boëx et al., 2011;Velasco et al., 2007). This is in line with the hypothesis that neuronal loss and/or altered electrical resistance in sclerotic neural tissue impair the efficacy of HFS since stimulation can only be successful when targeting a sufficiently preserved network (Cuéllar-Herrera et al., 2004;Velasco et al., 2007). MTLE patients with hippocampal sclerosis may therefore require specific stimulation parameters to achieve seizure control.
Interestingly, low-frequency stimulation (LFS) at 5 Hz was effective in MTLE patients in small cohort studies, including those with hippocampal sclerosis (Koubeissi et al., 2013;Lim et al., 2016). From a technical perspective, LFS would be favorable for clinical implementation due to its low duty cycle, resulting in less electric current injection and longer battery life. For a systematic assessment of seizure-suppressive effects of LFS in relation to disease parameters, studies in translational animal models are crucial.
Optogenetic stimulation offers cell-or pathway-specific modulation of neuronal activity and has been successfully applied to alleviate seizure burden in several rodent MTLE models Zhao et al., 2015). These optogenetic approaches targeting the hippocampus were either based on the inhibition of excitatory neurons or on the recruitment of inhibitory interneurons (Kim et al., 2020;Kokaia et al., 2013;Krook-Magnuson et al., 2013;Ladas et al., 2015;Ledri et al., 2014;Lu et al., 2016). In MTLE models and patients with strong hippocampal sclerosis, however, pyramidal cells and GABAergic interneurons are strongly diminished and therefore would be difficult to target (Bouilleret et al., 1999;Magló czky and Freund, 2005;Marx et al., 2013;Thom, 2014).
In the present study, we applied optogenetic LFS (oLFS) to systematically investigate the efficacy of different stimulation frequencies in seizure interference. We used the intrahippocampal kainate (KA) mouse model, which replicates the major hallmarks of human MTLE pathology, comprising the emergence of spontaneous recurrent seizures and robust unilateral hippocampal sclerosis (Bouilleret et al., 1999). In this model, dentate granule cells (DGCs) with their entorhinal inputs, i.e. the perforant path Janz et al., 2017a), and CA2 pyramidal cells are preserved (Häussler et al., 2016). Using different KA concentrations, we modified the severity of hippocampal sclerosis and applied oLFS to stimulate entorhinal afferents in the diseased hippocampus in vivo while recording local field potentials (LFP). We then assessed the responses of individual DGCs to oLFS in vitro by patch-clamp recordings. In addition, we probed the translational value of LFS by applying electrical LFS (eLFS) in vivo over several weeks. We present evidence that LFS is highly effective in preventing both subclinical epileptiform activity and behavioral seizures in experimental MTLE with severe hippocampal sclerosis.

Modification of the intrahippocampal KA mouse model
To achieve different degrees of hippocampal sclerosis and seizure burden as observed in human MTLE (Thom, 2014), we modified the established intrahippocampal KA mouse model by injecting three KA concentrations (10, 15, and 20 mM). To this end, we compared these KA groups at 35-40 days after KA injection with respect to GCD, cell loss in CA1 and hilus, and epileptiform activity ( Figures 1A, 2 and 3).
Next, we investigated the characteristics of spontaneous epileptiform activity in all mice using LFP recordings from electrodes in both dorsal hippocampi (idHC and cdHC, Figure 3). We used a custom algorithm as described in the Materials and methods and illustrated in the supplements (Figure 3-figure supplement 1) for data analysis of LFP recordings. In brief, epileptiform spikes were automatically detected and analyzed for bursts, which were classified as low-, medium-and highload bursts according to their spike load using a self-organizing map (SOM, Figure 3-figure supplement 1A-C). Bursts recorded in the 10 mM KA group covered the whole range of the SOM, but a larger fraction of these bursts matched nodes representing lower spike loads compared to the 15 and 20 mM KA group (Figure 3-figure supplement 1D).
Epileptiform activity occurred in both hippocampi of all KA mice ( Figure 3A,B). The idHC of 10 mM KA mice showed a smaller percentage of high-load bursts compared to the 15 or 20 Figure 1. Experimental design for in vivo LFS. Animals received intrahippocampal KA and a channelrhodopsin 2 (ChR2)-carrying virus into the entorhinal cortex to trigger epileptogenesis and the expression of ChR2-mCherry in entorhinal afferent fibers. After 16 days post-injection, recording electrodes, and (A) an optic fiber or (B) an optic fiber combined with a stimulation electrode were implanted. Following recovery from implantations, reference LFPs were recorded on 2 consecutive days for 3 hr each. (A) In the first group of experiments, the effect of oLFS on spontaneously occurring epileptiform activity was tested (week 4) in 4-hr recording sessions. A session consisted of 1 hr of 'pre' stimulus recording, followed by 1 hr of 'oLFS' pulses and 2 hr of post-stimulus recordings ('post 1' and 'post 2'). Three different oLFS frequencies (1, 0.5, or 0.2 Hz) were applied on successive days in each animal (two sessions per animal). Next, generalized seizures were induced by optogenetic (10 Hz) stimulation. To test the effects of oLFS on seizures, oLFS (1 and 0.5 Hz) was applied either immediately after (overwriting) or before the pro-convulsive (10 Hz) stimulation (preconditioning) (week 5). (B) The second group of in vivo experiments assessed the effects of eLFS on epileptiform activity. First, we tested the effects of eLFS on optogenetically induced seizures (week 4, preconditioning, see above). In weeks 5 and 6, animals were stimulated daily for 1 hr (1 Hz, eLFS) following the same 'pre', 'eLFS',' post 1', post 2' paradigm, as described above. In week 7, animals were stimulated twice (on 2 different days) over 3 hr continuously and twice in an on-off 'cycle' paradigm: after initial 30 min eLFS, eLFS stimulation was turned off for 10 min and then turned on again for 10 min. This was repeated four times, followed by another hour LFP recording ('post 1'). (A1, B1) All animals were perfused after the last LFP recording and brain sections were processed for immunohistological procedures. (B2) Implantation scheme for eLFS. DG, dentate gyrus; FISH, fluorescent in situ hybridization; IHC, immunohistochemistry; perf., perfusion. In each section, the electrode position is marked with a red asterisk. Epileptic hippocampi show GCD in the dentate gyrus (open arrowheads) and cell loss in CA1 (the region between filled arrowheads). Comparing KA concentration groups with respect to different markers of hippocampal sclerosis by quantification of (B) GCL volume of dispersed regions (i.e. GCD), (C) total length of cell loss in CA1, (D) % loss of NeuN + hilar cells, and (E) loss of Gad67 + hilar interneurons in the sclerotic vs. non-sclerotic hippocampus (i.e. (G, G1) ipsilateral vs (F, F1) contralateral). One-way ANOVA; Tukey's multiple comparison test; *p<0.05, **p<0.01 and ***p<0.001. All values are given as mean ± standard error of the mean (SEM). (H) Animals injected with higher KA concentrations (15 and 20 mM KA) display stronger hilar cell loss along with a higher degree of GCD. Scale bars 200 mm. mM KA groups. In the cdHC, however, the incidence of high-load events was similarly low for the 10 and 20 mM KA groups ( Figure 3C, Figure 3-source data 1). The mean ratio of time spent in high-load bursts (mean high-load burst ratio) and epileptic spike rate were significantly lower in the idHC 10 mM KA group compared to the 15 and 20 mM KA group ( Figure 3D-G, showing spontaneous epileptiform activity in the cdHC and idHC. (C) Automatic classification of epileptiform activity into low-load (blue), medium-load (green) and high-load bursts (orange). We used a custom algorithm as illustrated in Figure 3-figure supplement 1. In the 20 mM KA group, the percentage of high-load bursts in the cdHC is decreased, whereas for the 10 mM and 15 mM KA group the percentage of the burst classes is similar in both hippocampi. (D-G) Injections of 15 and 20 mM KA lead to an increased high-load burst ratio and a higher epileptic spike rate in the idHC but not in the cdHC. All values are given as mean ± SEM. Source data is provided in Figure 3-source data 1. (I, J) The high-load burst ratio is positively correlated with GCD and hilar NeuN + cell loss ((I): p<0.0001, two-tailed; Pearson's r = 0.84; (J): p<0.01, two-tailed; Pearson's r = 0.64). The online version of this article includes the following source data and figure supplement(s) for figure 3: Source data 1. Variable severity of epileptiform activity elicited by different KA concentrations.  3-source data 1). In the idHC the extent of GCD and hilar cell loss was positively correlated with the high-load burst ratio ( Figure 3H, p<0.0001, Pearson's r = 0.84; and Figure 3I, p<0.01, Pearson's r = 0.64; both n = 15 animals).
Taken together, we developed the intrahippocampal KA mouse model further, creating varying degrees of disease severity on both the anatomical and electrophysiological level. Thus, this model provides a valuable framework for the following stimulation experiments.

Application of different oLFS protocols during spontaneous epileptiform activity
Since the sclerotic hippocampus is considered as the focus of epileptiform activity Pallud et al., 2011), we targeted DGCs, the major surviving excitatory neurons, by photostimulation of entorhinal afferents. To this end, adult mice received KA into the hippocampus and a ChR2-carrying viral construct into the medial entorhinal cortex followed by LFP recordings and oLFS in the chronic epileptic phase ( Figure 4A; Figure 4-figure supplement 1A). Prior to photostimulation, reference LFPs were recorded for 1 hr in 'pre' sessions to confirm the occurrence of spontaneous epileptiform activity in the idHC ( Figure 4B One hour of optogenetic stimulation with pulsed light at 1 Hz or 0.5 Hz significantly decreased the high-load burst ratio and the epileptic spike rate, followed by a return to pre-stimulation levels within 2 hr independent of the KA concentration ( Figure 4D-I, Figure 4-source data 1). Photostimulation at 0.2 Hz had no significantly suppressive effect on the high-load burst ratio, but on the epileptic spike rate (Figure 4J-L; Figure 4-source data 1). 1 Hz oLFS was significantly more effective than 0.2 Hz in suppressing high-load bursts and epileptic spikes, whereas we found no significant difference between 1 Hz and 0.5 Hz ( Figure 4P,Q; Figure 4-source data 1). Looking at individual sessions in more detail, 1 Hz oLFS had a higher percentage of sessions with a suppression efficacy above 75% than 0.5 Hz and 0.2 Hz regarding high-load burst ratio and epileptic spike rate (1 Hz: 86.36% and 48.00%, 0.5 Hz: 56.25% and 23.08%, 0.2 Hz: 14.29% and 5.88%). As expected, oLFS (1 Hz) did not influence epileptiform activity in no-virus control mice ( Figure 4M-O, Figure 4source data 1).
To clarify whether the suppression of high-load bursts and epileptic spikes due to oLFS was locally restricted to the stimulation site (idHC), we analyzed LFPs contralaterally (cdHC), outside of the epileptic focus ( Looking at individual sessions in detail, 1 Hz oLFS had a higher percentage of sessions with a suppression efficacy above 75% than 0.5 Hz and 0.2 Hz regarding high-load burst ratio and epileptic spike rate (1 Hz: 87.50% and 66.67%, 0.5 Hz: 50.00% and 7.14%, 0.2 Hz: 23.53% and 5.88%).
Next, we tested whether the neuronal responses to oLFS were confined to the stimulated area by analyzing the spatial and temporal occurrence of evoked responses in the idHC and cdHC. In all animals, pulsed light delivery to the idHC did not only trigger local but also delayed responses in the cdHC (  2 Hz and no-virus control, 1 Hz) are shown. Automatic detection of epileptiform activity is marked for low-load (blue), medium-load (green), and highload bursts (orange). (D) Photostimulation at 1 Hz effectively decreases spontaneous epileptiform activity in the idHC. (E, F) Automatic quantification of epileptiform activity shows that oLFS reduces the high-load burst ratio as well as the epileptic spike rate in all animals independently of the KA concentration (10 mM: light gray; 15 mM: dark gray; 20 mM: black) followed by a return to pre-stimulation levels within 2 hr ('post 1' and 'post 2'). (G, J) oLFS with (H, I) 0.5 Hz or (K, L) 0.2 Hz has a weaker antiepileptic effect during stimulation. Single sessions (olive-green) were used to calculate the oneway ANOVA; Tukey's multiple comparison test (all KA concentrations pooled); *p<0.05, **p<0.01, and ***p<0.001. All mice were video recorded and the running behavior was analyzed during each session as shown in  So far, our findings demonstrated that 1 Hz oLFS of entorhinal afferents was highly effective in the suppression of spontaneous epileptiform activity within as well as outside of the epileptic focus. Moreover, this effect was independent of the degree of hippocampal sclerosis and seizure burden. Therefore, we pooled all KA groups in the following experiments.

Effects of oLFS and eLFS on induced behavioral seizures
In the intrahippocampal KA model, spontaneous epileptiform activity is mainly subclinical and rarely generalizes into behavioral seizures (Häussler et al., 2012;Janz et al., 2018;Klein et al., 2015). To assess the impact of oLFS on generalized seizures, we induced these seizures by 10 Hz photostimulation as described previously (Janz et al., 2018;Osawa et al., 2013). In addition to the evoked potentials during stimulation, high-amplitude epileptic spikes emerged in the LFP which gradually became rhythmic and dominant before progressing into full-blown behavioral seizures (Figure 5figure supplement 1A). These evoked seizures displayed electrographic features highly similar to those of spontaneous generalized seizures (Figure 5-figure supplement 1B) and were accompanied by the same stereotypic myoclonic movements (e.g. rearing, falling, and convulsion).
Since LFS and seizure induction were both driven by photostimulation, we performed additional experiments combining optogenetic 10 Hz stimulation with electrical 1 Hz preconditioning in the same animal. To this end, mice received intrahippocampal KA (15 mM) and the ChR2-carrying viral construct into the medial entorhinal cortex followed by a side-by-side implantation of an optic fiber and a stimulation electrode into the dentate gyrus. In addition, all animals were implanted with LFP recording electrodes as described above ( Figure 1B).     We determined the minimum optical stimulus duration sufficient to reliably trigger a generalized seizure. Then, we probed the seizure-suppressive action of 1 Hz eLFS prior to optogenetic 10 Hz seizure induction ( Figure 5H). Thirty minutes of 1 Hz eLFS was as effective as oLFS in reducing the probability of generalized seizures ( Figure 5I, w/o pre-eLFS: 93.75 ± 6.25%; with 1 Hz pre-eLFS:   16.67 ± 9.62%) and the associated behavior ( Figure 5J, w/o eLFS: RS 2.38 ± 0.40; with 1 Hz eLFS: RS 0.33 ± 0.33, n = 4 animals).
In conclusion, preconditioning by optogenetic and electrical 1 Hz LFS was highly effective in preventing evoked generalized seizures.

Cellular responses to oLFS
In order to investigate the underlying mechanisms of the anti-epileptic effects of oLFS, we quantified the cellular responses in the sclerotic hippocampus. For this, we calculated the area under the curve (AUC) of each evoked response over the 1-hr stimulation period. We only selected sessions with high stimulation efficacy (within the 95% confidence interval (CI), compare Figure 4P).
Photostimulation (1 Hz) evoked stable response waveforms ( Figure 6) which decreased slightly in amplitude over time in non-epileptic (saline-injected) mice ( Figure 6A1-3, B). In chronically epileptic mice, the cellular responses declined strongly and rapidly within the first 10 min ( Figure 6C1-3, D). Lower frequencies (0.5 and 0.2 Hz) altered the cellular response much less ( Figure 6E-H) as evident from AUC analysis. Similarly, AUCs of evoked responses of the pro-convulsive 10 Hz pulse-train were reduced by about 40% after pre-conditioning when compared to the responses without To investigate whether oLFS decreases the excitability of DGCs at the single-cell level, we studied their intrinsic properties and the synaptic strength of entorhinal inputs in acute slices obtained from chronically epileptic mice. We performed whole-cell recordings of DGCs located in the outer region of the dispersed GCL in the presence of GABA A and GABA B receptor blockers as described in the Methods. Cells were filled with biocytin during recordings for subsequent morphological identification ( Figure 7A, n = 14 animals). Photostimulation of afferent entorhinal fibers robustly induced depolarization of DGCs (5.1 ± 1.0 mV, n = 14 cells) and was occasionally sufficient to induce action potentials (À70 mV holding potential, n = 4 cells). Since cellular responses in vivo declined strongly and rapidly within the first 10 min, we applied 10 min oLFS at 1 Hz to acute slice preparations. During oLFS, evoked synaptic responses were strongly depressed ( Figure 7B, reduction of excitatory postsynaptic potential (EPSP) amplitude to 28.5 ± 9.9% of the original response, n = 14 cells). Next, Figure 7. Decrease of single-cell EPSPs and discharge probability after 10 min oLFS. (A) Representative confocal projection of a dentate gyrus slice from a KA-(15 mM) and AAV-treated mouse 28 days after SE. The two biocytinfilled DGCs (white) were recorded in this section. ChR2-expressing entorhinal fibers are visible in the middle molecular layer (red, mml). Cell bodies were stained with DAPI (blue). h, hilus; scale bar 50 mm. (B) Pulsed blue light delivery reliably induces EPSPs, which decline strongly during a 10-min stimulation protocol (50 ms pulses at 1 Hz). (C, D) Extracellular electrical stimulation (trial = 5 pulses, 50 Hz, arrows) of entorhinal fibers induces action potentials (APs) measured in DGCs (inset, gray traces, recorded in loose-patch). Photostimulation at 1 Hz over 10 min (inset, blue traces) significantly reduces the discharge probability of DGCs in epileptic KA (n = 19 cells) and saline (n = 11 cells) control slices. ANOVA on ranks with Dunn's Bonferroni post-hoc correction; *p<0.05. Values are given as mean ± SEM. (E) Spontaneous IPSCs in DGCs of saline (n = 6 cells) and KA (n = 7 cells) slices were analyzed for frequency and amplitude. The number of sIPSCs is strongly reduced in 'epileptic' DGCs compared to control but those that occur have a larger amplitude. Mann-Whitney Rank Sum Test; *p<0.05 both.
we evaluated the effect of oLFS on discharge probability upon electrical stimulation (trial = 5 pulses at 50 Hz) of entorhinal fibers. To maintain stable intracellular conditions in DGCs, we used cellattached recordings. In non-sclerotic control slices (saline mice), oLFS for 10 min reduced the discharge probability of DGCs ( Figure 7C, number of action potentials generated per trial at 100 V stimulation, no-oLFS: 1.8 ± 0.5 vs. after oLFS: 0.24 ± 0.15, p<0.05, n = 11 cells). Although DGCs were overall more excitable in KA mice, as reported previously in this model (Janz et al., 2017a), the number of action potentials in response to electrical stimulation was significantly reduced upon 10 min oLFS ( Figure 7D, number of action potentials generated per trial at 100 V stimulation, no-oLFS: 3.5 ± 0.6 vs. after oLFS: 2.1 ± 0.5, p<0.05, n = 19 cells).

Effects of repeated hippocampal eLFS on epileptiform activity
To assess whether the seizure-suppressive effect of 1 Hz LFS was stable over extended periods, we repeatedly applied eLFS over 3 weeks after the preconditioning experiments ( Figure 1B). In the first and second week (weeks 5 and 6 after SE), animals were stimulated and recorded daily for 4 hr (including 'pre'-, 'eLFS'-, 'post 1'-, and 'post 2'-sessions of 1 hr each). Three-hour reference LFP recordings were performed before, in between, and after these eLFS experiments ( Figure 8A). During these 2 weeks, spontaneous epileptiform activity was successfully suppressed in each eLFS session but returned to reference level during the following 2 hr ( Figure 8B-G, Figure 8-source data 1). Compared to the first reference recording, epileptiform activity remained unchanged in the intermittent reference recordings ( Figure 8H, Figure 8-source data 1; Figure 8J).
In the third week (week 7 after SE), we extended the eLFS period from 1 hr to 2 hr on 2 subsequent days ( Figure 8A). Here, we confirmed that stimulation was also effective for 3-hr stimulation periods. Within 1 hr following eLFS, high-load bursts gradually reappeared ( Figure 8I, Figure 8source data 1). Finally, we tested whether we could reduce the number of 1 Hz stimuli and still maintain the successful interference with epileptiform activity. Therefore, we introduced stimulationfree periods and stimulated mice on 2 consecutive days in an on-off manner: after an initial 30 min stimulation period, four 10 min 'off' and 'on' phases, and a 1-hr 'post' recording followed. Epileptiform activity was efficiently extinguished during the 2-hr discontinuous stimulation ('on-off' period, Figure 8K,L), showing the same suppressive effect as the continuous stimulation.
Interestingly, mice with a misplaced stimulation electrode (i.e. not in the dentate gyrus, Figure 8 In conclusion, eLFS in the sclerotic hippocampus has a strong seizure-suppressive effect, which (i) maintains its efficacy over several weeks without desensitization, (ii) can be prolonged for at least 3 hr, and (iii) remains stable in short stimulation-free phases.

Discussion
In the present study, we applied LFS in experimental MTLE to interfere with seizure generation in vivo. Our main findings can be summarized as follows: (1) we modified the intrahippocampal KA Figure 8. Stable suppression of epileptiform activity through eLFS in the sclerotic hippocampus. (A) Session design. We locally stimulated the dentate gyrus in the sclerotic idHC for 1 hr/day (1 Hz) over 2 weeks. In the third week, we stimulated animals 3 hr continuously and 2 hr in a 10 min 'on-off' manner on 4 successive days. (B, E) Representative LFP traces (15 mM KA, idHC electrode) for the 'pre' and 'eLFS' sub-sessions (weeks 5 and 6) are shown. Automatically detected high-load bursts are indicated by an orange bar. (C, D, F, G) Quantification of epileptiform activity shows that eLFS nearly extinguishes epileptiform activity in all animals followed by a return to pre-stimulation levels within 2 hr ('post 1' and 'post 2'). Single sessions were used to calculate the RM one-way ANOVA; Tukey's multiple comparison test ***p<0.001. Animals with a misplaced stimulation electrode did not show this remarkable suppression of epileptiform activity as shown in Figure 8-figure supplement 1. The correct positions of electrodes were confirmed by histology as shown in Figure 8-figure supplement 2. (H) Animals have a stable high-load burst ratio over 3 hr of reference recordings. (I) 3 hr continuous eLFS effectively reduces the high-load burst ratio over the whole stimulation period, reoccurring in a lower manner within 1 hr after stimulation (Two-way ANOVA; Tukey's multiple comparison test ***p<0.001). All values are given as mean ± SEM. Source data is provided in Figure 8-source data 1. (J) Epileptic spike rates for each minute (one data point, color-coded for each animal) in a 3-hr recording. (K, L) An initial 30 Figure 8 continued on next page mouse model to create different degrees of hippocampal sclerosis and seizure burden, (2) in this model, among several frequencies, 1 Hz stimulation exerts the strongest suppressive effects on both spontaneous subclinical epileptiform activity and evoked generalized seizures, (3) the suppressive action of LFS is stable during the applied stimulation time and does not fade over several weeks, and (4) this effect is most likely associated with reduced efficacy of perforant path transmission onto DGCs rather than downscaling their intrinsic excitability.

Modification of a mouse epilepsy model to obtain variable degrees of hippocampal sclerosis and seizure burden
To create an animal model that reflects the inter-individual variability of neurodegeneration and seizure frequency as observed in human MTLE (Blümcke et al., 2013;Thom, 2014), we modified the well-established intrahippocampal KA mouse model of MTLE by varying the KA dose. We found that different KA concentrations resulted in variable degrees of histopathological changes and the development of spontaneous recurrent epileptiform activity as previously described for 20 mM KA injection (Janz et al., 2017b). In particular, cell death of hilar neurons increased with the KA dose, whereas DGCs survived equally but dispersed more strongly. This is in line with previous reports from human patients, showing that GCD appeared to be related to the amount of cell loss in the hilus (Houser, 1990) and that DGCs survive in the sclerotic tissue (Thom, 2014). Similar to the human pathology, the sclerotic hippocampus in the intrahippocampal KA mouse model is thought to be critically involved in seizure generation Pallud et al., 2011;Walker, 2015). Accordingly, we observed a correlation between hilar cell loss, GCD, and epileptiform activity in the ipsilateral hippocampus.
Epileptiform activity generated in the sclerotic focus often spreads to the contralateral hippocampus in MTLE patients (Gloor et al., 1993;Mintzer et al., 2004) and KA-treated mice (Janz et al., 2018;Meier et al., 2007). We observed that epileptiform activity propagated less frequently to the contralateral hippocampus in mice with strong hippocampal sclerosis, than in those with mild sclerosis, pointing toward a relationship between network preservation and seizure spread.

oLFS of entorhinal afferents interferes with spontaneous epileptiform activity
We used the modified KA mouse model to stimulate the sclerotic hippocampus for seizure interference. Following KA injection, CA1 and CA3 pyramidal cells are extensively lost (Bouilleret et al., 1999;Marx et al., 2013), whereas DGCs and their entorhinal afferents are preserved (Janz et al., 2017a). Accordingly, by positioning the optic fiber as precisely as possible above the molecular layer of the dentate gyrus, we stimulated mainly the perforant path (Janz et al., 2018). Hence, CA2 neurons, surviving CA3 pyramidal cells, or temporoammonic fibers (in cases of mild hippocampal sclerosis) were most likely not activated by light delivery. Using this experimental design, we identified 1 Hz oLFS as a highly effective stimulation frequency to suppress spontaneous epileptiform activity in the sclerotic and also in the undamaged, contralateral hippocampus. Interestingly, we recorded population spikes at both the site of light delivery as well as contralaterally with a polysynaptic delay, pointing to a more widespread network effect.
LFS at 1 Hz has been used previously to alleviate epileptiform activity. In vitro approaches showed that optogenetic activation of calcium/calmodulin-dependent protein kinase type II (CaMKII)-positive neurons for 3 min in the entorhinal cortex and electrical stimulation of the ventral hippocampal commissure for 15 min disrupt seizure-like activity induced by 4-Aminopyridine (4-AP) (Shiri et al., 2017;Toprani and Durand, 2013). Also, electrical 1 Hz stimulation of Schaffer collaterals in acute hippocampal slices for 10 min diminished epileptiform activity induced by bicuculline (Albensi et al., 2004). In addition to this short-term efficacy of 1 Hz LFS, long-term effects over several hours (60 min continuous stimulation or alternating periods of 15 min on/off for 4 hr) were reported from in vivo studies using amygdala kindling or pilocarpine rodent models targeting the ventral hippocampal commissure or entorhinal cortex (Rashid et al., 2012;Xu et al., 2016). Although these reports give valuable insights into the seizure-suppressing effects of LFS, they differ from our study with respect to the epilepsy models and stimulation targets used. Our emphasis was to show that LFS can prevent seizure generation in chronically epileptic mice with hippocampal sclerosis.
Optogenetics allows us to selectively target entorhinal afferents in the hippocampus for our oLFS experiments. Optogenetic stimulation has been applied in the past to alleviate seizure-like activity in experimental epilepsy (for review see Christenson Wick and Krook-Magnuson, 2018;Zhao et al., 2015). Most optogenetic studies decreased network activity either by direct inhibition of hippocampal principal cells or by activating GABAergic interneurons (Kokaia et al., 2013;Krook-Magnuson et al., 2013;Ladas et al., 2015;Ledri et al., 2014). More recently, Bui et al., 2018 showed that optogenetic stimulation of mossy cells in the dorsal dentate gyrus can control electrographic seizures, but the effect was much weaker than with direct inhibition of DGCs.
In the present study, by indirectly stimulating excitatory DGCs we deviated from the approach of increasing inhibition. Accordingly, local 1 Hz photostimulation of entorhinal afferents effectively suppressed epileptiform activity in both hippocampi independent of the severity of hippocampal sclerosis. Lower stimulation frequencies (0.5 and 0.2 Hz) were less effective. This is in line with previous in vitro studies showing a frequency dependency for the reduction of the duration of 4-AP-induced seizure-like activity (D'Arcangelo et al., 2005;Toprani and Durand, 2013).

Preconditioning by LFS prevents induced generalized seizures
Next, we probed whether oLFS and eLFS can interfere with behavioral seizures induced by 10 Hz photostimulation of entorhinal afferents, as 5-20 Hz hippocampal photostimulation has previously been shown to be pro-convulsive (Janz et al., 2018;Osawa et al., 2013). We observed that 1 Hz stimulation right after a pro-convulsive stimulus, which is similar to an on-demand approach, was not able to interrupt an ongoing seizure. Similarly, on-demand photostimulation of DGCs at 7 or 20 Hz exacerbated spontaneous epileptiform activity to large behavioral seizures in the same model (Bui et al., 2018;, and on-demand inhibition of DGCs (or mossy cells) did not affect the occurrence of behavioral seizures (Bui et al., 2018).
However, preconditioning by 1 Hz (oLFS or eLFS) or 0.5 Hz (oLFS) for 30 min was able to prevent the occurrence of optogenetically-induced seizures. Our preconditioning results are in line with a technique called quenching that blocked the development and progression of amygdala-kindled seizures using 15-20 min 1 Hz stimulation right before or after a kindling stimulus or daily LFS in fully kindled rats (Jalilifar et al., 2018;Weiss et al., 1995). Taken together, timing and frequency of stimulation appear to be critical for successful seizure interference.
The suppressive effects of eLFS on epileptiform activity are stable over several weeks Considering the clinical limitations of optogenetics, we applied eLFS in the dentate gyrus over several weeks to test the suppressive action on epileptiform activity. We first repeated the 1-hr stimulation sessions daily over 2 weeks. Then, we applied eLFS for longer periods (3 hr) in the third week. Finally, since continuous LFS could potentially disrupt hippocampal function and shorten the battery life of implanted devices, we applied a discontinuous stimulation protocol. All of these stimulation protocols were repeatedly effective in suppressing spontaneous epileptiform activity. In the future, we plan to extend the LFS-off periods and develop a closed-loop system that initiates the 1 Hz stimulation when epileptiform activity returns. This will overcome the limitation of the current study in which the seizure-suppressive effects of LFS have been analyzed over relatively short durations (hours) as longer durations (days or weeks) would be closer to clinical applications.
Our results are promising when compared to the use of anti-epileptic drugs (AEDs) which only acutely exert seizure-suppressive effects in epileptic mice (Duveau et al., 2016;Klein et al., 2015;Riban et al., 2002) but can lose their effects during prolonged treatment (Roucard et al., 2010).
Also, MTLE is frequently pharmacoresistant in humans (Engel, 2001). From a translational point of view, selective stimulation of entorhinal afferents might be applied in these patients by placing electrodes in the clearly defined angular bundle (Zeineh et al., 2017).

LFS reduces the efficacy of perforant path transmission onto DGCs
Although the mechanism underlying the seizure-suppressive effect of LFS is not clear, several lines of evidence suggest that LFS reduces the excitability of the hippocampus and associated networks (i.e. the entorhinal cortex). Besides, retrograde firing of entorhinal cortex neurons may potentially lead to more widespread network effects. Mechanistically, it is rather remarkable that synchronization of the hippocampal network by LFS interferes with seizure generation, a process that is classically characterized by hypersynchronization. One explanation could be the local recruitment of GABAergic interneurons (Xu et al., 2016). However, this is unlikely, since the frequency of IPSCs was strongly reduced in our slice experiments and interneurons are often extensively diminished in the sclerotic hippocampus (Bouilleret et al., 2000;Marx et al., 2013).
Another hypothesis, investigated in a 4-AP in vitro slice model, suggests that electrical stimulation of the entorhinal cortex at low frequencies (e.g. at 0.5 Hz) results in less accumulation of extracellular potassium compared to the large potassium efflux associated with GABA A receptor-mediated interictal (between seizures [Fisher et al., 2014]) discharges that could trigger a seizure (Avoli et al., 2013). Clamping GABA-mediated potentials with 1 Hz LFS (directly via activation of GABAergic interneurons or indirectly via feedback inhibition upon principal cell activity) could therefore induce transient increases in extracellular potassium and restrain pro-epileptic discharges (Barbarosie et al., 2002). In other words, stimulating the network more frequently (i.e. at 1 Hz) leads to a lower accumulation of extracellular potassium than less frequent interictal discharges would induce (Shiri et al., 2017). Our results show that this mechanism cannot fully account for the seizuresuppressive effect of oLFS since we observed a decrease in the amplitude of evoked responses in the presence of GABA A and GABA B receptor blockers in whole-cell recordings.
Another explanation for the reduction of epileptiform activity achieved by our experimental design could be synaptic depression due to synaptic fatigue or long-term depression (LTD). Synaptic depression can be induced at the entorhinal-DGC synapse with stimulation frequencies equal to or below 1 Hz (Abrahamsson et al., 2005;Gonzalez et al., 2014). In line with this, we showed that the amplitude of evoked responses decreases over time on both the network and the single-cell level, which could explain the prolonged interval without spontaneous epileptiform activity following LFS. However, the intrinsic properties of DGCs did not change, thus there were no obvious signs of postsynaptic LTD. It is unlikely that downscaling of entorhinal inputs alone is sufficient to suppress epileptiform activity given that perforant path transection does not alleviate seizure-like activity (Meyer et al., 2016;Pallud et al., 2011). We therefore assume that the seizure-suppressive effects of our oLFS cannot fully rely on synaptic depression. Supporting this notion, LFS is not effective in inducing synaptic depression in sclerotic human hippocampal slices (Beck et al., 2000).
Concerning acute seizure-suppressive effects, we suggest that LFS drives the hippocampal network into a stable state, reducing the probability of seizures while LFS is ongoing. This hypothesis is supported by Chang et al., 2018 who showed that evoked interictal-like discharges elicit anti-epileptic effects, contingent on the network state as well as stimulus amplitude and frequency. This study investigated the influence of interictal discharges on seizure generation in vitro in local CA3 and CA1 circuits, two subnetworks that are typically lost in MTLE with hippocampal sclerosis. Our results suggest that the authors' observations also apply to the entorhinal-dentate network in vivo. In line with this interpretation, oLFS of entorhinal afferents may introduce glutamatergic synaptic perturbation followed by transient suppression of neuronal activity de Curtis and Avanzini, 2001;Muldoon et al., 2015) and thus may interfere with recurrent generation of epileptiform activity (Chang et al., 2018).
It is possible that oLFS could impair normal hippocampal function given its dramatic effect on glutamatergic transmission. However, Wang et al., 2020 demonstrated that LFS improved cognitive function in pilocarpine-injected rats. Future research will address whether hippocampal function is affected by long-term LFS in our model.
In conclusion, our study identified 1 Hz LFS in the dentate gyrus as a highly efficient approach to interfere with seizure generation in the MTLE mouse model with hippocampal sclerosis. We demonstrated that the effect was largely driven by repetitive activation of DGCs residing in the sclerotic seizure focus, and we have shed light on the associated cellular mechanisms. Considering the potential for clinical translation, our findings may pave the way for effective seizure control in one of the most common forms of drug-resistant epilepsy.

Animals
Experiments were conducted with adult (8-12 weeks) transgenic male mice (C57BL/6-Tg(Thy1-eGFP)-M-Line) (Feng et al., 2000). Each animal represents an individual experiment, performed once. In total, 96 mice were used for this study. Mice were kept in a 12 hr light/dark cycle at room temperature (RT) with food and water ad libitum. All animal procedures were carried out under the guidelines of the European Community's Council Directive of September 22, 2010 (2010/63/EU) and were approved by the regional council (Regierungsprä sidium Freiburg).

In vivo LFS experiments
After recovery from implantations, the freely behaving mice were first recorded on 2 successive days (3 hr each) to determine reference LFPs (Figure 1). Each mouse represents the biological replicate and the number of recordings per mouse the technical replicate. For LFP recordings, mice were connected to a miniature preamplifier (MPA8i, Smart Ephys/Multi-Channel Systems, Reutlingen, Germany). Signals were amplified 1000-fold, bandpass-filtered from 1 Hz to 5 kHz, and digitized with a sampling rate of 10 kHz (Power1401 analog-to-digital converter, Spike2 software, Cambridge Electronic Design, Cambridge, UK).

Optogenetic stimulation
On days 21-28 after KA and virus injection (10, 15, or 20 mM KA, AAV1.CamKIIa.ChR2-mCherry), photostimulation with pulsed blue light (460 nm; 50 ms pulse duration; 150 mW/mm 2 at the fiber tip; blue LED, Prizmatix Ltd.) was applied at low frequencies (1, 0.5, or 0.2 Hz) to test the effect of oLFS on spontaneous epileptiform activity. Taking into account the restricted penetration depth of 473 nm light in brain tissue (Yizhar et al., 2011) and an output of 150 mW/mm 2 at the tip of our optic fiber, we estimated that in our case the radius of light emission with sufficient power to activate ChR2 did not exceed 500 mm (Figure 4-figure supplement 2). During stimulation, we continuously recorded LFPs and videos. Recording sessions were divided into four sub-sessions: 'pre' -1 hr before oLFS; 'oLFS' -1 hr during oLFS; 'post 1' and 'post 2' -first hour and second hour after oLFS ( Figure 1A, week 4). For each frequency, we performed 2 trials on different days. To check for lightor heat-induced effects the 'pre' and 'oLFS' (1 Hz) sessions were performed in non-virus controls.
Since in the intrahippocampal mouse model spontaneous epileptiform activity is frequent but rarely generalizes we evoked behavioral seizures by 10 Hz photostimulation (25 ms pulse duration) and assessed the effect of oLFS before or after the pro-convulsive 10 Hz stimulus ( Figure 1A, week 5). To determine the minimum duration sufficient to trigger a behavioral seizure for each animal (identification of seizure threshold), we systematically increased the stimulation duration (in 1 s steps). Each evoked seizure was visually inspected on the electrophysiological as well as behavioral level. We assessed generic features as described by Jirsa et al., 2014 and motor symptoms according to the Racine scale (Racine, 1972). The identified seizure threshold was then validated for robust seizure induction for at least 3 times in each animal. In subsequent 'preconditioning' oLFS sessions, photostimulation at 1 or 0.5 Hz was performed for 30 min before applying the pro-convulsive 10 Hz stimulus. For both frequencies, a minimum of 3 trials was performed on different days. In a subset of mice (n = 5), the efficacy of the pro-convulsive 10 Hz stimulus was tested again after the preconditioning experiments to exclude confounding effects of habituation.

Fluorescent in situ hybridization (FISH)
Gad67 mRNA was localized by FISH with digoxigenin (DIG)-labeled cRNA probes generated by in vitro transcription as described earlier (Kulik et al., 2003). Slices were hybridized with DIG-labeled antisense cRNA probes and immunodetection of the DIG-labeled hybrids was performed with a peroxidase-conjugated anti-DIG antibody (1:2000; raised in sheep; Roche Diagnostics, Mannheim, Germany). The fluorescence signal was developed with tyramide signal amplification (TSA) Plus Cyanine 3 System kit (PerkinElmer, Waltham, Massachusetts, USA) as described previously (Tulke et al., 2019).
In detail, brain slices were pre-treated in a 1:1 mixture of hybridization buffer [50% formamide, 4xSSC, 5% dextran sulfate, 250 mg/ml heat-denatured salmon sperm DNA, 200 ml yeast t-RNA, 1% Denhardt's-reagent (Sigma-Aldrich, Steinheim, Germany)] and 2xSSC at RT for 15 min. Subsequently, the slices were pre-hybridized in a hybridization buffer for 60 min at 45˚C, followed by the addition of DIG-labeled antisense or sense Gad67 cRNA probe (100 ng/ml) and incubated overnight at 45˚C. Slices were washed in 2xSSC for 2 Â 15 min at RT and then successively rinsed at 55˚C for 15 min in 2x SSC with 50% formamide; 0.1x SSC with 50% formamide and twice in 0.1x SSC alone. Then the slices were rinsed in 0.1 M Tris-buffered saline (TBS) for 2 Â 10 min and transferred to the blocking buffer [1% blocking reagent (Roche Diagnostics, Mannheim, Germany) in TBS] for 60 min at RT. For fluorescent detection, tissue sections were treated with a horseradish peroxidase-conjugated DIG antibody overnight at 4˚C and developed in the presence of amplification buffer and tyramide working solution (1:50) for 6 min in the dark. The staining-reaction was stopped by rinsing in TBS for 3 Â 5 min and 1 Â 15 min. Slices were kept in the dark for further immunofluorescence staining.

Immunohistochemistry
For immunofluorescence staining, free-floating sections were pre-treated in 10% normal horse serum (Vectorlabs, Burlingame, California, USA) in phosphate buffer (PB) for 1 hr. Subsequently, slices were incubated first with guinea-pig anti-NeuN (1:500; Synaptic Systems, Gö ttingen, Germany) overnight at 4˚C and then with a donkey anti-guinea-pig Cy5-conjugated antibody for 2.5 hr at RT (1:200, Jackson ImmunoResearch Laboratories Inc, West Grove, Pennsylvania, USA) followed by extensive rinsing in PB. Sections were mounted on glass slides with an anti-fading mounting medium (DAKO, Hamburg, Germany).
To assess the extent of hippocampal sclerosis, we quantified the volume of the dispersed granule cell layer (GCL) and cell loss in the hilus and CA1 along the septo-temporal axis of the hippocampus using Fiji ImageJ software (Schindelin et al., 2012). Here, masking was performed since the evaluator was not aware of the respective KA treatment. In detail, a region-of-interest (ROI) was drawn visually comprising the dispersed parts of the GCL using the ImageJ polygon function in each slice (around 50 slices per animal). Afterwards, the volume was calculated based on the area measured in each slice (values are given in mm 3 ). For the quantification of cell loss, the summed length of pyramidal cell-free gaps in the CA1 region was measured in each section using the segmented line function. Furthermore, hilar cell loss was quantified by automated detection (Cell Counter plugin) of NeuN + cells (size parameter: 50-infinity pixel 2 ) and Gad67 mRNA + interneurons (size parameter: 100-infinity pixel 2 ) in the hilus of three dorsal sections for each animal. Here, we calculated the percentage of cell loss in the sclerotic hippocampus compared to the contralateral, non-sclerotic hippocampus (set to 100%). In detail, the Cy3 and Cy5 channels were split and individual images were converted to grayscale. Images were background-subtracted by manual adjustment of the threshold and further processed using the watershed function, which separates overlapping cell bodies. A ROI was then defined for the hilus with the polygon function, and the respective cell density was calculated. ChR2-mCherry expression analysis was performed using the mean gray value for each layer along the dorsoventral axis of the sclerotic hippocampus (Figure 4-figure supplement 2E) in Fiji ImageJ. We manually selected three ROIs in each layer (from alveus to the lower outer molecular layer of the dentate gyrus) in the image of the hippocampus where the optic fiber was positioned.

Analysis of epileptiform activity
Animals that showed abnormal hippocampal atrophy (i.e. extensive loss of DGCs) were excluded from analyses (n = 5). Recordings obtained from all electrodes were first visually inspected for epileptiform activity. Then, we used a custom algorithm developed in the intrahippocampal KA mouse model (Heining et al., 2019) to detect and classify epileptiform activity in LFP data from idHC and cdHC. The dataset used from Heining et al., 2019 served us as a reference for classification. This reference dataset consisted of data obtained from animals injected with 20 mM KA recorded at electrode positions comparable to ours. The detection and classification algorithm works as described in the following: Epileptiform spikes: The spectrogram per frequency bin was normalized to values between 0 and 1 and local maxima in the normalized (4-40 Hz) frequency band were identified as spikes. This was complemented by amplitude-based spike detection and spike sorting to reduce false negatives or positives. Bursts: Spikes within an inter-spike interval below 2.5 s were assigned to the same burst. Feature vectors: Each burst i was represented by a three-dimensional feature vector y i (= [log10 of spike count; log10 of mean inter-spike interval; standard deviation of inter-spike intervals]). To compare our data to the reference dataset, y i was normalized to the mean (x ) and standard deviation (s) of the feature vectors x of the reference dataset and multiplied by the weight vector w (=[2; 2; 1]): Burst classification: bursts were classified using a SOM that consists of nodes (visualized as hexagons) representing prototypical feature vectors. Nodes with similar feature vectors are neighbors on the SOM. Based on the SOM created with the reference dataset, bursts were classified according to their spike load into high-load, medium-load, and low-load bursts categories ( To assess the severity of epileptiform activity within a recording, we calculated the high-load burst ratio as the fraction of time spent in high-load bursts (sum of the high-load burst durations divided by total recording time). The automatic detection of high-load bursts was confirmed by visual inspection. To assess epileptic burden of individual mice in the three KA concentration groups, the average high-load burst ratio was calculated from a total of nine LFP recordings (15 hr) performed on different days (2x 3 hr before oLFS experiments, 6x 1 hr of 'pre' sub-sessions of oLFS experiments and 1x 3 hr after oLFS experiments; Figure 1A). In oLFS and eLFS experiments, suppressive effects on epileptiform activity were evaluated based on the high-load burst ratio and the epileptic spike rate calculated for the respective sub-session ('pre',' oLFS'/'eLFS',' post 1', post 2'). Whole sessions were excluded if the corresponding 'pre' recording had a high-load burst ratio below 0.05. Furthermore, we analyzed the AUC of evoked responses (integral of LFP traces in a defined time interval) during photostimulation. We used a 4th order low-pass Chebyshev filter type I with a cutoff frequency of 300 Hz to smoothen the signal. The time interval for calculating the AUCs was set relative to stimulus onset from À0.1 s to +0.2 s for oLFS (1, 0.5, and 0.2 Hz) and from À0.02 s to +0.06 s for the 10 Hz stimulation ( Figure 5-source code 1, Figure 6-source code 1). For responses during oLFS that were co-occurring with high-load bursts, the AUC was not determined. Recordings obtained from the cdHC site were used to measure pulse latencies between recording sites during photostimulation sessions (Figure 4-figure supplement 4-source code 1). In brief, we used the same time intervals and filters as applied for the AUC calculation and z-scored the voltage curve of each pulse response. We then calculated the delay between each idHC and cdHC pulse response using the first threshold crossing (one threshold was identified by visual inspection of the signal for all animals) and took the median for each oLFS sub-session.

Statistical analysis
Data were tested for significant differences with Prism 8 software (GraphPad Software Inc). Comparisons of two groups were performed with a paired (comparisons within animals) or unpaired (comparisons between animals) Student's t-test. When more than two groups were compared either oneway ANOVA, or two-way ANOVA followed by Tukey's post-hoc test was applied. In the case of missing values, statistics were calculated by fitting a restricted maximum likelihood model (REML). Significance thresholds were set to: *p<0.05, **p<0.01 and ***p<0.001. For all sample populations, mean and SEM are given, unless otherwise reported. Correlations were tested using Pearson's correlation (slope significantly non-zero, CI 95%). All statistical results are summarized in Supplementary file 1.

Data availability
The LFP dataset is available on Open Science Framework: https://osf.io/uk94m/. The source code files for the seizure detection algorithm is accessible at Zenodo (https://doi.org/10.5281/zenodo. 4110614). The source code for the seizure detection algorithm (Heining et al., 2019) was developed using previously published LFP data Janz et al., 2017b).
The following datasets were generated: