Cholinergic suppression of sharp wave-ripples impairs hippocampus-dependent spatial memory

The hippocampus plays a central role in long-term memory formation, and different hippocampal network states are thought to play distinct roles in this process. These network states are controlled by neuromodulatory inputs, in particular the cholinergic input from the medial septum. Here, we used optogenetic stimulation of septal cholinergic neurons to better understand how cholinergic activity affects different stages of spatial memory formation in a reward-based navigation task in mice. We found that optogenetic stimulation of septal cholinergic neurons (1) impaired memory formation when activated at goal location but not during navigation; (2) reduced sharp wave-ripple incidence at goal location; and (3) reduced ripple incidence and enhanced theta-gamma oscillations during sleep. These results underscore the importance of appropriate timing of cholinergic input in long-term memory formation, which might help explain why there has been limited success of cholinesterase inhibitor drugs in the treatment of memory impairment in Alzheimer’s disease.


Introduction
The role of the neuromodulator acetylcholine (ACh) in learning and memory is debated. On one hand, cholinergic neuron degeneration and a low tone of ACh correlate with memory impairment in normal and pathological ageing in humans and rodents (Bartus, 2000;Berger-Sweeney et al., 2001;Hasselmo and Sarter, 2011); indeed, drugs blocking the degradation of ACh are among the few that modestly ameliorate memory impairment in Alzheimer's disease (Ehret and Chamberlin, 2015); and extensive lesions of the medial septum (MS), where the cholinergic neurons projecting to the hippocampus are located, produce learning deficits (Hepler et al., 1985). On the other hand, selective lesions of MS cholinergic neurons, which account for ~ 5 % of neurons in network states are mutually exclusive (O'Keefe and Nadel, 1978;Buzsáki, 1986;Csicsvari et al., 2000) and it was suggested that memory encoding is associated with theta/gamma oscillations while memory consolidation relies on SWR activity Colgin, 2013).
Hippocampal network states are controlled by the local release of ACh.
In hippocampal CA3, cholinergic activation induces a slow gamma rhythm primarily by activating M1 muscarinic receptors (Fisahn et al., 1998;Betterton et al., 2017), while in the CA1, cholinergic activation promotes theta/gamma oscillations and suppresses ripple oscillations through the activation of M2/M4 muscarinic receptors (Vandecasteele et al., 2014;Zhou et al., 2019;Ma et al., 2020). This suggests that regulation of cholinergic tone allows the switching between online attentive processing (theta/gamma oscillations) and offline memory consolidation (SWRs) as described in the two-stage model of memory trace formation (Buzsaki, 1989). Evidence from microdialysis and electrophysiology experiments shows a high cholinergic tone during exploration, promoting theta activity, and a lower cholinergic tone during subsequent rest, permitting SWRs (Fadda et al., 2000;Giovannini et al., 2001;Fadel, 2011). Disruption of cholinergic activity at different stages of learning and memory impairs performance in memory tasks (for review see Hasselmo and Sarter, 2011;Solari and Hangya, 2018) and has improved our understanding of the dynamical aspect of cholinergic modulation. However, the differential effects of ACh in various phases of memory formation are still not well understood.
To clarify the function of the MS cholinergic system during hippocampus-dependent learning and memory, we investigated the phase-specific effect of optogenetic cholinergic stimulation in the appetitive Y-maze long-term memory task, a simple reward-based spatial memory task with two distinctive behavioural phases: a phase of navigation towards a reward and another after arriving in the goal area (Bannerman et al., 2012). We show that increased stimulation of cholinergic neurons does not affect learning of the task when applied during navigation towards a reward, but impairs learning if stimulation is applied at the goal location. Using electrophysiological recording in both freely behaving and anesthetized mice we show that activation of MS cholinergic neurons switches the CA3 and CA1 network states from ripple activity to theta/gamma oscillations. Our recordings of hippocampal local field potential (LFP) in freely moving animals indicate that impaired memory was related to disruption of ripples, supporting the two-stage model of memory trace formation (Buzsaki, 1989).

In vivo electrophysiology in anesthetized mice
Mice were anesthetized with intraperitoneal injections of 1.2 g . kg -1 urethane and their head was fixed in a stereotaxic frame. Body temperature was maintained at 35 ± 1 o C with a heating pad. The head was shaved, the Bregma and Lambda were aligned horizontally, and craniotomies were made above the MS and CA3. Simultaneous optical activation in the MS (AP: +1 mm, ML: 0 mm, DV: -3.6 mm, coordinates from Bregma) with a stripped optical fiber (200 µm, 0.22 NA; Doric Lenses) and electrical recordings in the MS or in the hippocampus (ML: +2.4 mm, AP: -2.46 mm, DV: -2.5 mm) using an extracellular parylene-C insulated tungsten microelectrode (127 µm diameter, 1 MΩ; A-M Systems) were performed.

Surgery
Surgeries were carried out following minimal standard for aseptic surgery. Mice were anesthetized with isoflurane (5% induction, 1-2% maintenance, Abbott Ltd, Maidenhead, UK) mixed with oxygen as carrier gas (flow rate 1.0-2.0 L . min -1 ) and placed in a stereotaxic frame (David Kopf Instruments, Tujunga, CA, USA). Then the skull was exposed after skin incision and Bregma and Lambda were aligned horizontally. A hole was drilled above the MS at coordinates AP: +1 mm and ML: 0 mm, and an optic fiber (200 µm, 0.22 NA; Doric Lenses) was lowered towards the MS (DV: -3.6 mm) at low speed (1 mm . min -1 ). Once positioned just above the MS, the optic fiber was secured to the skull using dental cement (Prestige Dental).
To perform recordings in freely moving animals, we implanted 5 mice with LFP duodes, consisting of two twisted 75 µm teflon-coated silver electrodes (AGT0510, World Precision Instruments, Hitchin, UK) with tips spaced 300-500 µm, one tip being implanted in the oriens and the other tip in the pyramidal cell layer. Mice were implanted bilaterally in CA1 (AP: -1.7, ML: ±1.2, DV: 1 & 1.3) and CA3 (AP: -2.3, ML: ±2.5, DV: 1.8 & 2.1), DV being taken from the surface of the brain. Ground and reference silver wires were connected to a stainless microscrew implanted over the cerebellum: AP: -5.3 ML: +/-1.5. To record the electromyogram activity. a 75 µm Teflon-coated silver electrode was implanted in the neck muscle. All wires were connected to a 32 pins Omnetics connector (Genalog, Cranbrook, UK). The exposed brain was covered with a protective dura gel (Cambridge NeuroTech, UK) to prevent damage upon cementing of the electrodes. Duodes were individually glued to the skull using UV-cured glue (Tetric EvoFlow) and the implant was secured to the skull using dental cement (Super-Bond C & B; Prestige Dental, Bradford, UK). At the end of the implantation, the mice were removed from the frame, Meloxicam (2 mg . kg -1 s.c.) was administered as analgesia and animals were left to recover in a post-surgery chamber and maintained at +34.0 °C. The mice were allowed to recover for 5 days before habituation started and during these 5 days were daily monitored and given Meloxicam as an analgesic.

Appetitive Y maze task
Long-term spatial memory was assessed using the appetitive Y-maze task, as described in full by Shipton et al. (2014). Briefly, mice had to learn to find a food reward (condensed milk) on a 3-arm maze that remained at a fixed location in relation to visual cues in the room. The 3-arm maze, elevated 82 cm from the floor, consisted of gray-painted 50×13 cm arms bordered by 1 cm high white plastic walls, which extended from a central triangular platform.
Plastic food wells (1.5 cm high) were positioned 5 cm from the distal end of the arms. Mice were kept on a restricted feeding schedule, allowing them to maintain at least 85% of their free food body weight. Before testing, the mice were habituated to the food reward and the maze in a different room to where behavioral testing would occur. During testing, mice were only allowed to make one arm choice each trial and were only allowed to consume the reward if the correct arm was chosen, otherwise, mice were removed from the maze and the trial was ended. Target arm assignments were counterbalanced such that at least one mouse of each experimental group was designated to each arm. Each mouse received ten trials per day for 6-10 consecutive days, five starts from the left of the target arm and five starts from the right in a pseudorandom order with no more than three consecutive starts from the left or right.
The maze was rotated either clockwise or anticlockwise after each trial to discourage the use of intra-maze cues to help learn the task. Optogenetic stimulation started either from the beginning of the trial (navigation and throughout cohorts), or when the mouse reached the goal zone (goal cohort). Light stimulation ceased when the mouse reached the goal zone for the navigation cohort. Light stimulation was performed using blue laser at 473 nm (Ciel, Laser Quantum, Cheshire, UK), powered 25 ±1 mW and we used 50-ms-long pulses at 10 Hz. Stimulation was controlled using custom made procedures in Igor Pro (WaveMetrics, Oregon, USA).

Optogenetic stimulation and electrophysiological recordings
Data were acquired from 5 ChAT-Ai32 male mice implanted with duodes for electrophysiological recordings and optic fiber for optogenetic stimulation. These mice were recorded during sleep and while performing the appetitive Y-maze task.
For recordings during sleep, after connecting the electrodes to the Whisper acquisition system (Neural Circuits, LLC, VA, USA) and optic fiber to the laser, the animals were placed in a cage (different to their home cage), to which the animal was habituated over a period of two days. The floor of the cage was covered with standard bedding. The recordings started after the mice visibly stopped moving and consisted of 30-s-long laser stimulation at 473 nm, power 25 ±1 mW using 50-ms-long pulses at 10 Hz alternating with 60-120 s interval without the stimulation. An overhead webcam camera tracked the movement and position of the animal. The videos were manually reviewed together with the recorded EMG signal to exclude trials that were interrupted by the mice moving. For Y-maze task, the mice underwent the same habituation and learning protocol as described above in the appetitive Y-maze task section. During learning, mice were connected to the laser and to the Whisper acquisition system and placed at the starting arm of the maze. The laser was activated in the goal zone on alternating trials to allow within-subject comparison. These 5 mice were not added to the behavioral cohort as the stimulation protocol (50 % of the trials) was different from that used in behavior only (stimulation performed for all trials).
The position of the animal was tracked with an overhead webcam and automatically extracted using custom procedures in MATLAB (MATLAB, 2019.
All recordings were performed using the Whisper acquisition system sampling at 25 kHz, laser stimulation was triggered using custom made procedures in Igor Pro and synchronized with the electrophysiological and webcam recordings.

Electrophysiology data analysis
Data analysis was performed in MATLAB Release 2019b (MATLAB, 2019) and R version 3.4.4 (R Core Team, 2018). To reduce contamination by volume conducted signal, we performed duode recordings. In the CA1, a single duode recorded from the oriens and the pyramidal cell layer; in the CA3, from the lacunosum moleculare and pyramidal cell layer. The duode signal was substracted, enhancing signal differences between the hippocampal layers. To remove noise artifacts caused by wire movement, changes in the consecutive samples of the EMG signal were detected. If the change exceeded a threshold set to 2 standard deviations, a 500 ms-long window of the signal centered on the noise timestamp was removed. We conducted the analysis on one duode from CA1 and one from CA3 based on the quality of signals for both theta oscillations and ripples.
For ripple detection, we adapted the method from (Vandecasteele et al., 2014). The signal was down-sampled to 1.25 kHz and 80-250 Hz bandpass filtered with Type II Chebyshev phase-preserving filter (filter order = 4, stopband attenuation = 20 dB). Next, the filtered signal was squared, meansubtracted, and smoothed by applying a moving average with 10-ms-long window. Ripples were detected when the squared signal crossed 2 standard deviations for 20-300 ms duration and its peak crossed 6 standard deviations in freely moving animals and 4 standard deviations in urethane-anesthetized animals. Ripple incidence was calculated as the number of detected ripples divided by the duration of the recording.
Power spectral density (PSD) was estimated using Welch's method (MATLAB built-in pwelch function with 0.5 s window and 0.25 s overlap) for frequencies spanning the range from 1 to 200 Hz. To visualize instantaneous changes in PSD during Y-maze trials, spectrogram was created with continuous wavelet transform using Morlet wavelets (MATLAB built-in cwt function with default parameters). Throughout the study, we defined theta and gamma frequency bands as the band ranging 2.5-5 Hz during anesthesia and 5-11 Hz otherwise for theta; 20-40 Hz during anesthesia and 25-45 Hz otherwise for slow gamma. The slow gamma frequency upper bound was chosen to exclude any line noise contamination at 50 Hz.
To estimate relative theta and slow gamma power, we used the FOOOF tool (https://github.com/fooof-tools/fooof) developed by Haller et al. (2018). It models the estimated PSD as the sum of background spectrum and Gaussian peaks in narrowband frequencies. The background spectrum was fitted on the PSD log-log plot with a straight line, which corresponds to a pink noise-like (1/f) background. To minimize the model error -the difference between the actual and the modeled PSD -the background spectrum fit was found in two frequency ranges separately (2 -20 Hz and 15 -150 Hz). Rela-tive theta and slow gamma peaks, as well as their peak frequencies, were taken from the Gaussian peaks fitted above the background spectrum.

Statistical analysis
Statistical analysis was performed in R version 3.4.4 (R Core Team, 2018). Data are reported as mean ± SEM unless otherwise stated. For significance testing, first, the normality of the data was assessed by Shapiro-Wilk test and by inspection of the quantile-quantile plot. If the normality criterion was satisfied (p > 0.05), a parametric test (t-test) was used, otherwise a nonparametric test (one-way ANOVA on ranks or Wilcoxon test) was used, as described in the Results section.
The effects of the optogenetic manipulation on ripple incidence and other quantities in freely moving mice were assessed with linear mixed-effects models. This method allows for correlations that exist in our data between samples coming from trials repeated in the same mouse and allows for an unbalanced number of samples between mice. Laser stimulation (L ∈ {0 for inactive, 1 for active}) was a fixed effect in the model; the random variable representing the animal (a) was treated as a random effect in the intercept and slope estimation. The quantity Y in the trial i for animal a was modeled as: • β0, β1 are linear regression coefficients for the fixed effects • R0a, R1a are random effects: normally distributed animal-specific corrections for linear regression coefficients with zero mean and maximumlikelihood standard deviation estimated by the model • εi is a random error with a normal distribution with zero mean and maximum-likelihood standard deviation estimated by the model.
The residual errors were checked for the assumptions of the linear models: mean of zero, no correlation with the predicted values and homoscedasticity. To satisfy these assumptions when modeling ripple duration, a log-linear model of variable Y was created as above.
The linear mixed-effects models were built in R with package 'lme4' and p values for the fixed effects were obtained using Sattherwaite estimation of degrees of freedom implemented in the 'lmerTest' R package.

Histological processing
Animals were terminally anesthetized by intra-peritoneal injection of pentobarbital (533 mg . kg -1 ) and then transcardially perfused with phosphatebuffered saline (PBS) followed by 4% paraformaldehyde (PFA). Brains were removed and post-fixed for 24-48 hours, then rinsed and subsequently cry- To identify the placement of the electrode (aided by Dil application) and optic fiber tracks for each mouse, sections containing evidence of the implant were selected and mounted in Fluoroshield (Sigma).
Sections were examined with a Leica Microsystems SP8 confocal microscope using the 10× and 20× magnification objectives. The eYFP + /GFP + and ChAT + cells were quantified manually using the ImageJ software. The location at which the implant appeared the deepest was determined and used to plot the implant location on the corresponding section in a mouse brain atlas (Franklin and Paxinos, 2007).

Data and code availability
Code used for the analysis and to generate the figures can be accessed on the authors' GitHub site: https://github.com/przemyslawj/ach-effecton-hpc. Electrophysiological data will be shared on any request.

Functional expression of ChR2 in cholinergic neurons
We aimed to investigate the effects of cholinergic modulation on hippocampal oscillations and performance in a spatial navigation task. To this end, we optogenetically controlled the activity of cholinergic neurons using ChAT-Ai32 mice that expressed enhanced YFP-tagged channelrhodopsin-2 (ChR2-eYFP) under the control of the choline-acetyl transferase promoter (ChAT).
We first checked the expression of ChR2 in MS cholinergic neurons by performing double immunostaining for ChAT and YFP ( Figure 1A). Out of 111 ChAT-immunopositive neurons in the MS, 98 (83.13%, n = 3 mice) were also immunopositive for YFP (YFP + ChAT + ). We did not detect YFP-immunopositive cells that were negative for ChAT (YFP + ChAT -), indicating a selective expression of ChR2-eYFP in ChAT cells.
We probed the functional expression of ChR2 in MS neurons by recording multi-unit activity in the MS of urethane-anesthetized mice. 473 nm light was delivered through an optic fiber implanted just above the MS while multi-unit activity was recorded from a co-assembled electrode whose tip protruded ~200 µm further than the optic fiber tip. Local light delivery (50 ms pulses at 10 Hz) resulted in an increase in multi-unit activity (baseline spike frequency: 7.9 ± 2.8 Hz vs. spike frequency during light delivery: 22.7 ± 7.3 Hz, two-tailed Wilcoxon matched pair signed-rank test: p = 0.03; n = 6 recordings from 2 mice; Figure 1B), confirming our ability to increase neuronal activity in the MS using optogenetics.

Activation of septal cholinergic neurons in the goal zone slows place learning
We investigated the effect of cholinergic activation during different phases of the appetitively-motivated Y-maze task, a hippocampus-dependent task commonly used to study long-term spatial memory (Bannerman et al., 2012;Shipton et al., 2014). Mice had to learn to find a food reward on an ele-vated 3-arm maze that remained at a fixed location in relation to visual cues in the room, while the mice pseudo-randomly started from one of the other two arms. As short-term memory errors caused by re-entry during a single trial have previously been shown to impair the acquisition of this spatial long term memory task (Schmitt et al., 2003), mice were only allowed to make one arm choice in each trial to isolate the long-term memory process (Figure 2A). Previous studies have reported sharp wave ripples, i.e. offline processing, at the reward location of different spatial navigation tasks (O'Neill et al., 2006;Dupret et al., 2010). Based on these studies, we defined a navigation phase corresponding to the arms of the maze except for the distal ends (20 cm from the edge), which we considered as goal zones. Cholinergic activation was achieved by light stimulation (473 nm, 25 mW, 50-ms-long pulses at 10 Hz) delivered via an optic fiber implanted in the MS of ChAT-Ai32 mice. ChAT-Ai32 mice were split into four groups to test four experimental conditions: (i) no stimulation (n = 13), (ii) optogenetic stimulation during navigation -from the start of the trial until they reached the goal zone (n = 9), (iii) optogenetic stimulation in the goal zone only -from the entry of the goal zone until the mice were removed from the maze either after they had eaten the food or they had reached the empty food well (n = 15), and (iv) optogenetic stimulation throughout the maze (n = 9; Figure 2B). Mice from all four cohorts learned the task but the acquisition was slower when cholinergic neurons were activated in the goal zone ( Figure 2C). To control for possible aversive effects of the illumination, we performed an additional experiment with MS-implanted wildtype (WT) mice split into 2 groups: no stimulation (n = 7) and light delivery in the goal zone (n = 9; Figure 2D).
We set a learning criterion of ≥80 % successful trials in a day. Comparison between the 6 cohorts revealed differences in reaching this criterion (one-way ANOVA on ranks χ 2 (5) = 17, p = 0.005; Figure 2E). Dunn posthoc tests indicated that the ChAT-Ai32 'goal' group was delayed at learning the task compared to the ChAT-Ai32 'no stimulation' group (4.47 ± 0.32 days vs 2.85 ± 0.25 days, p = 0.0003). Similarly the ChAT-Ai32 'goal' group was delayed compared to the ChAT-Ai32 'navigation' group (4.47 ± 0.32 days vs 3.11 ± 0.35 days, p = 0.01). Conversely we did not observe any learning difference between the 'goal' and 'no stimulation' groups in the WT mice cohorts (goal: 3.0 ± 0.37 days; no stimulation: 3.4 ± 0.37 days; p = 0.6).
These results suggest that cholinergic activation in the goal zone for 10-30 s slows learning of the appetitive Y-maze task. At the same time, optogenetic activation during navigation or throughout the maze had no significant effect on task acquisition. We confirmed that a long-term memory trace had formed by retesting the mice on the Y-maze task one week after the end of the acquisition period for each group of the ChAT-Ai32 mice (no stimulation: 100 ± 0 %; navigation: 99 ± 1 %; goal: 94 ± 3 %; throughout: 100 ± 0 %). After behavioral testing, implant placement and the level of eYFP expression were verified by immunohistochemistry, confirming that there were no significant differences in implant placement between the behavioral groups (

Activation of medium septum cholinergic neurons reduces the incidence of ripples via muscarinic receptor activation
To gain an understanding of why activating MS cholinergic neurons in the goal zone impairs task acquisition, we performed recordings from the hippocampal CA1 and CA3 areas of freely moving mice. A previous study showed that activation of septal cholinergic neurons suppresses SWRs in hippocampal CA1 (Vandecasteele et al., 2014). Here, we investigated how the cholinergic manipulation affects both CA3 and CA1 activity (Buzsáki, 1986;Csicsvari et al., 2000;Oliva et al. 2016).
We implanted recording electrodes in the CA1 and the CA3 and the optic fiber above the MS of the ChAT-Ai32 mice. We recorded LFP signal while the mice slept in a cage, to which they were familiarized over the two previous days. Within a single behavioral session, we alternated 30-s-long epochs with the stimulation on with 60 -120-s-long epochs without the stimulation. Only epochs when the mouse was asleep for their full duration were used for the analysis (10 ± 2 epochs per animal, n = 5 animals).
Optogenetic stimulation reduced the ripple incidence both in the CA1 and the CA3, and this reduction followed the duration of the stimulation (Figure 3A-B). To account for correlations between the epochs for the same mouse, we assessed this effect with a linear mixed-effects model (see Methods). Optogenetic stimulation significantly reduced ripple incidence in the CA1 from 0.26 ± 0.06 Hz to 0.05 ± 0.02 Hz (F(1, 3.7) = 0.24, p = 0.01, n = 5 animals) and in the CA3 from 0.26 ± 0.03 Hz to 0.08 Hz ± 0.02 Hz (F(1, 10) = 41, p < 10 -4 , n = 5 animals). We did not observe any effect of the laser stimulation on To investigate the mechanism by which the CA3 ripples are suppressed, we performed the same optogenetic manipulation in urethane-anesthetized mice (1.2 g . kg -1 ), which allowed us to combine recordings and pharmacological investigation. Under anesthesia, the ripples had an incidence of 0.24 ± 0.05 Hz (n = 7 mice) and were similar in appearance to those recorded in naturally sleeping animals: their peak frequency was 124 ± 2 Hz, compared to 132 ± 2 Hz in sleeping animals and their duration was 41 ± 2 ms compared to 33 ± 1 ms in sleeping animals. Stimulation of MS cholinergic neurons (473 nm light, 25 ms pulses, 20 Hz for 30 s) significantly reduced ripple incidence to 0.05 ± 0.03 Hz (median = 0.00 Hz; n = 7 mice; two-tailed Wilcoxon paired test: p = 0.02; Figure 3D), which was prevented by prior intra-peritoneal administration of scopolamine, a muscarinic receptor antagonist (2 mg . kg -1 ; control: 0.33 ± 0.03 Hz, median = 0.33 Hz; optogenetic stimulation: 0.33 ± 0.05 Hz, median = 0.29 Hz; n = 9 mice; two-tailed paired t-test: p = 0.81; Figure   3D).
Hence, our results demonstrate that optogenetic activation of MS cholinergic neurons suppresses ripple incidence both in the CA1 and in the CA3 where we confirmed that the suppression occurs through the activation of muscarinic receptors.

MS cholinergic neurons increased theta and slow gamma activity in sleeping animals
We next investigated the effect of optogenetic stimulation of MS cholinergic neurons on the other hippocampal network state: theta-gamma activity in the sleeping mouse.
We determined the power spectral density for frequencies ranging 1 -200 Hz in control and during stimulation. We observed a reduction of the power spectral density (PSD) across the full frequency range upon light stimulation ( Figure 4B-C, Figure 4 -figure supplement 1), as reported previously (Vandecasteele et al., 2014). PSD from electrophysiological recordings are the summation of a periodic signal and a background spectrum that is greatly influenced by aperiodic activity whose intensity on PSD is of pink noise (1/f) (Haller et al., 2018). Therefore, to quantify the power of theta and gamma oscillations, we estimated the background spectrum and looked at the relative peaks in the PSD above the fitted background spectrum (Haller et al., 2018, see Methods, Figure 4C).

Activation of septal cholinergic neurons in the goal zone slows place learning by reducing CA1 ripple incidence and CA3 slow gamma.
We last asked how the oscillatory regimes vary during the appetitive Y maze and how ACh neuron stimulation modulate the oscillations in the goal zone. To investigate whether the changes in the ripple activity and theta-gamma oscillations are responsible for the impaired place learning, we performed the Ymaze task using the same ChAT-Ai32 mice (n = 4) implanted with recording electrodes and the optic fiber. Optogenetic stimulation was applied when the mouse reached the goal zone ( Figure 5A) on alternating trials, allowing us to compare the CA1 and the CA3 activity between the stimulated and non-stimulated trials. Mice were trained and recorded over 6 days, and by day 5 had reached 70 ± 10 % successful trials. We analyzed only the successful trials (14 ± 2 trials per animal with no stimulation and 11 ± 1 trials with the stimulation at goal zone). SWR in CA1 and CA3 occurred at start and goal locations ( Figure 5B-C) and optogenetic stimulation significantly reduced the ripple incidence at goal in the CA1 from 0.12 ± 0.02 Hz to 0.06 ± 0.01 Hz. However, light did not significantly affect ripple incidence in CA3 (0.12 ± 0.02 Hz; linear mixed-effects model, CA1: F(1, 5.7) = 6.72, p = 0.04; CA3: F(1, 10) = 0.06, p = 0.81, n = 3 mice included in the analysis with ripple incidence at goal ≥ 0.03 Hz). The ripple incidence in the non-stimulated epochs was not significantly different between the early (before day 5) and late learning (linear mixed-effects model, CA1: F(1, 8.9) = 2.1, p = 0.18; CA3: F(1, 6.6) = 0.01, p = 0.74, n = 3 mice).
The mouse behavior strongly modulated hippocampal activity. We recorded the highest theta power intensity in the central section of the maze where the mice run the fastest ( Figure 5E) while at the goal location, where the mice consumed the reward, theta (5 -11 Hz) and high gamma frequency bands (80 -200 Hz) showed reduced power intensity ( Figure 5E, right panel).
Nevertheless, theta and slow gamma bands peaked above the background spectrum at the goal location in the control trials (88 ± 12% of trials had a theta peak in the CA1 signal and 85 ± 12 % in the CA3; 98 ± 1 % trials had a slow gamma peak in the CA1 signal and 88 ± 10 % in the CA3, Figure 5F).
Delivering light did not affect the relative theta power intensity (linear fixed-ef-  Figure 5I).
Overall, these results show that optogenetic stimulation of MS cholinergic neurons reduces ripple incidence in the CA1 and increases slow gamma power intensity in the CA3. Hence this suggests a causal relationship between the ripple incidence at the goal location and memory formation and highlights the need of a timely control of ACh release in the hippocampus during learning.

DISCUSSION
Using optogenetics, we investigated the effects of stimulating MS cholinergic neurons on hippocampal LFPs and memory when delivered at different phases of an appetitively-motivated spatial memory task. We found that (1) MS cholinergic activation at the goal location, but not during navigation, impairs spatial memory formation; (2) MS cholinergic stimulation during reward con-sumption reduces ripples incidence in the CA1 and promotes slow gamma in the CA3; and (3) cholinergic stimulation reduces hippocampal ripple incidence and promotes theta-gamma rhythm in both the CA1 and the CA3 in the sleeping mouse. These results show that timely control of cholinergic modulation is crucial for spatial learning on a time scale of seconds. Moreover, our results indicate that cholinergic stimulation suppresses ripple activity in both sleeping and awake, behaving animals, explaining the learning deficit induced by inappropriately timed cholinergic activity.

Cholinergic influence on hippocampal network activity
The correlation between hippocampal network activity and cholinergic tone or MS cholinergic neuron activity is well established. MS cholinergic neurons discharge at a maximal rate when the animal is running (Ma et al., 2020), which corresponds to highest theta power intensity in the CA1 and highest cholinergic tone measured in the pyramidal layer of CA1 (Fadda et al., 2000;Fadel, 2011). Conversely, cholinergic tone and MS cholinergic neuron discharge is at its lowest during slow wave sleep and wake immobility, which are associated with the highest ripple incidence (Fadda et al., 2000;Ma et al. 2020).
In accordance with these observations, we found that stimulation of MS cholinergic neurons strongly reduces ripple activity in both naturally sleeping and anesthetized animals, consistent with previous reports (Figure 3; Vandecasteele et al., 2014;Zhou et al., 2019;Ma et al., 2020). However, MS cholinergic stimulation did not significantly reduce the ripple occurrence in the CA3 when the mouse was in the goal zone of the Y-maze, while the incidence de-creased in the CA1 ( Figure 5). The lack of significant ripple incidence reduction in CA3 in the behaving mouse could be due to ripple detection failure or low numbers of animals tested but this seems unlikely as the same mice were used for sleep recording, during which ripple incidence reduction was observed upon light stimulation (Figure 3). Thus, it is possible that MS cholinergic stimulation differentially affects ripple induction or propagation depending on the mouse behavioral state. Interestingly, hippocampal area CA2 has recently been suggested to be crucial for initiating SWRs in awake mice (Oliva et al., 2016;Alexander et al., 2018), while both CA2 and CA3 could initiate them in sleeping animals, providing a possible explanation for this finding.
We show that stimulation of MS cholinergic neurons promotes theta-gamma rhythm in the CA3 and CA1 of sleeping and anesthetized animals, consistent with previous findings in the CA1 (Figure 4; Vandecasteele et al., 2014;Zhou et al., 2019;Ma et al., 2020). Similar to the report by Vandecasteele et al. (2014), only mild effects of MS cholinergic stimulation on theta and gamma peak power were seen in the sleeping mouse (Figure 3) due to a global reduction of the background aperiodic signal (1/f) also observed in Vandecasteele et al. (2014). However, the peak relative to the fitted background (Haller et al., 2018) revealed an increase in rhythmic theta-gamma power intensity (Figure 3 and 5).
Pharmacological evidence in vivo indicates the presence of two distinct mechanisms of theta oscillations in the hippocampus, an atropine-sensitive and an atropine-resistant component (Petsche et al., 1962;Buzsáki, 2002;Colgin, 2013). The atropine-sensitive component is mediated by the combination of cholinergic and GABAergic neurons in the MS (Buzsáki, 2002;Manseau et al., 2008) and is slower than the atropine-insensitive theta rhythm, which is generated primarily by the entorhinal cortex (Buzsáki, 2002;Colgin, 2013). Moreover, atropine-sensitive theta was best detected in the anesthetized animal while atropine-insensitive theta was shown to predominate in the running animal (Kramis et al., 1975;Newman et al., 2013). Consistent with this division, MS cholinergic stimulation enhanced a scopolaminesensitive theta oscillation in both anesthetized and sleeping mice (supplementary Figure 3) and shifted theta peak to a lower frequency in sleeping mice ( Figure 3), while theta power intensity was not affected by MS cholinergic stimulation in the behaving mouse ( Figure 5). The weak effect of MS cholinergic stimulation on theta-gamma rhythm in awake animals during the memory task could be explained either by the prominence of an atropine-resistant entorhinal-driven theta that would override any atropine-sensitive theta or, alternatively or additionally, by an already maximal cholinergic tone. Interestingly, we did observe an increase of slow gamma power in the CA3 of the behaving mouse but not in the CA1 ( Figure 5). This is likely due to the activation of the recurrent connections in the CA3, leading to a more prominent increase of the spectral power.

Importance of timely regulation of cholinergic tone for memory formation
We found that temporally controlled optogenetic stimulation of MS cholinergic neurons could affect learning of the appetitive Y-maze task. Stimulation of the cholinergic neurons during navigation had no effect on the mouse perfor-mance while, strikingly, cholinergic stimulation in the goal zone significantly impaired task acquisition (Figure 2).
The lack of behavioral effect of light stimulation during the navigation phase, when the cholinergic tone is naturally high (Fadda et al., 2000;Giovannini et al., 2001;Fadel, 2011), may suggest that release of ACh in the hippocampus is already maximal, or that ACh receptors are saturated. MS cholinergic neurons are slow spiking neurons with a maximal rate of ~ 10 Hz during active exploration (Ma et al., 2020), which was the stimulation frequency used here.
It is plausible that the ACh concentration in the hippocampus might have already reached a plateau naturally, on top of which our stimulation protocol would not permit a further increase.
The impairment of memory formation by cholinergic stimulation in the goal zone, when the cholinergic tone is naturally lower (Fadda et al., 2000;Giovannini et al., 2001;Fadel, 2011), suggests that any potential beneficial effect of increased excitability or synaptic plasticity is outweighed by a requirement of reduced cholinergic activity. There is evidence to suggest that CA1 SWRs, which occur during low cholinergic activity, play a crucial role in memory formation (Ego-Stengel and Wilson, 2010;Girardeau et al., 2004;Jadhav et al., 2012;Roux et al., 2017): disruption of SWRs in the first 15 to 60 minutes following training impairs learning of spatial navigation tasks (Ego-Stengel and Wilson, 2010;Jadhav et al., 2012). In exploring animals, SWRs occur during transient immobility periods, including periods at goal locations (O'Neill et al., 2006;Dupret et al., 2010;Roux et al., 2017). These SWRs stabilize spatial representations of the CA1 place cells supporting navigation towards the newly learned goals (Roux et al., 2017), and are predictive of performance in a spatial memory task (O'Neill et al., 2006;Dupret et al., 2010). During these SWRs, sequences of neuronal activation are replayed in reverse order with the activation first of cells whose place fields are close to the animal (Foster and Wilson, 2006;Csicsvari et al., 2007;Diba and Buzsáki, 2007;Karlsson and Frank, 2009;Ambrose et al., 2016). In contrast, SWRs occurring outside of the learning environment are not triggered by sensory stimuli, and sequences of cell activation are usually played in the forward direction (Diba and Buzsáki, 2007). These latter SWRs have been more studied and a series of experiments disrupting their occurrence have shown they are crucial for memory formation (Ego-Stengel and Wilson, 2010;Girardeau et al., 2004;Jadhav et al., 2012). We found that MS cholinergic activation for the brief time the mice spent in the reward zone, shorter than 30 s, is sufficient to significantly impair memory formation in the Y-maze task (Figure 2) likely to be caused by the disruption of SWR activity in CA1 (Figure 3). This might suggest that the reverse replay of cell sequence in the reward zone is important for learning. However, additional effects of MS cholinergic activation on intracellular signaling cascades and synaptic plasticity (Brzosko et al., 2019) or synaptic inhibition can not be ruled out at this stage (Hasselmo and Sarter, 2011;Haam and Yakel, 2017).

Possible implications for neurodegenerative disorders
Loss of cholinergic neurons in the basal forebrain is one of the hallmarks of AD (Whitehouse et al., 1982;Bowen et al., 1982), which is also associated with a reduction of ACh transporter expression in most cortical and subcortical areas (Davies and Maloney, 1976). These observations have led to the cholinergic hypothesis of AD, which suggests that loss of cholinergic inputs plays a role in the cognitive impairment of AD patients. However the association between the loss of basal forebrain cholinergic neurons and AD is not completely clear (Mesulam, 2004), and a growing body of anatomical and functional studies suggests that MS cholinergic neuronal loss occurs in both healthy ageing and AD brain (Schliebs and Arendt, 2011;Hampel et al., 2018). Therefore, drugs compensating for the decline of cholinergic tone are seen as a rational treatment of both ageing-related memory loss and AD.
However, so far, drugs targeting the cholinergic systems have shown limited beneficial effects on cognitive deficits of ageing and AD but the reason why is not entirely clear (Farlow et al., 2010;Ehret and Chamberlain, 2015). Our results shed some light on why cholinergic drugs have largely failed to improve the cognitive impairments in AD. Cholinesterase inhibitors prolong cholinergic activity by ~100 times and increase the basal cholinergic tone in the absence of spontaneous activity (Hay et al., 2015). In rodents, cholinergic tone is high during exploration, which maintains the 'online' hippocampal state dominated by theta and gamma oscillations (Buzsáki, 1989;Fadda et al., 2000;Giovannini et al., 2001). It is possible that cholinesterase inhibitors maintain a high cholinergic tone, preventing the network from transitioning into a SWR-dominant state. This could impair memory formation as our results suggest that artificially increasing ACh release for as short a time as 10-30 s during a lowcholinergic state is sufficient to impair task acquisition ( Figure 2). Moreover, enhancement of cholinergic activity during the 'online' state did not bring beneficial effects to memory formation. Thus, our results suggest that suboptimal timing of cholinergic activity impairs long-term memory formation and sup-ports the idea that appropriate timing of cholinergic modulation is crucial in learning and memory (Micheau and Marighetto, 2011).  Cholinergic stimulation in the goal zone slows learning of the appetitive Y-maze task. (A) Mice were trained on an elevated 3-arm maze to find a food reward (red dot) in an arm that remained at a fixed location relative to visual cues in the room. Mice were allowed to consume the reward if they chose the correct arm, but were removed from the maze if they chose the incorrect arm.