In the first set of experiments, five rats performed a hippocampus-dependent spatial non-match-to-sample task (NMTS, Table 1) in an eight-arm radial maze (Packard et al., 1989; Sasaki et al., 2018). Each trial in the task is independent and consists of an instruction (encoding) and test (retrieval) phase, separated by several seconds. The NMTS task tests if rats can memorize visited locations in the instruction phase for the duration of the trial (~1 min) and use this memory to navigate to the remaining unvisited locations in the test phase (Figure 1a and b). Prior to SWR manipulations, all rats were familiarized with the task until they reached a learning criterion of more than 50% of trials without errors in one session (15 trials per session) for 3 days in a row (Figure 1c). Note that each trial featured a randomly chosen set of locations for the instruction phase that was different from the previous trials.

Figure 1 with 1 supplement see all

Download asset Open asset

Following completion of the training, SWRs were detected in the extracellular field potentials recorded from hippocampal area CA1 (Buzsáki, 2015; Ciliberti and Kloosterman, 2017). In line with previous reports (Singer and Frank, 2009), most ripples occurred during reward consumption and immobility periods in the instruction and test phases (ripple rate, offline detected, instruction: 0.66  Hz [0.55,0.79], test: 0.58  Hz [0.47,0.71]).

In one third of the trials in every session, ripple detection was used to trigger closed-loop electrical stimulation of the ventral hippocampal commissure to transiently disrupt the ongoing SWR (same method as Michon et al., 2019; Figure 1d). Note that rats at this point have learned the general task rules, but still need to remember a new random set of trial-specific locations. The task performance of the rats in the disruption trials was compared to the performance in the remaining control trials with either delayed stimulation upon ripple detection (random delay between 150 and 250ms) or without stimulation. Ripple detection and disruption happened during all phases of the trial (instruction and test). The trial performance (i.e. whether or not a trial is completed without a single error) and number of correct visits in the test phase in the disruption trials were not significantly different from those in the control trials (performance; disruption: 0.68 [0.57,0.77], stim.control: 0.73 [0.63,0.81], no stim. 0.72 [0.62,0.80], Kruskal-Wallis test: H=0.93, p=0.63, correct visits test phase; disruption: 3.52 [3.30,3.67], stim.control: 3.66 [3.49,3.77], no stim.: 3.49 [3.24,3.66], H=1.17, p=0.56, Figure 1e and f).

For rats to correctly perform the task and identify which arms have not yet been visited, one strategy is to memorize all four visits from the instruction phase. This memory may be subject to temporal decay such that the last visited arm in the instruction phase is remembered best. Indeed, we observed a temporal order effect such that errors in the test phase are biased to the arms that were instructed furthest back in time (average order number instruction arm; disruption: 1.70 [1.45,2.02], stim.control: 1.93 [1.60,2.29], no stim.: 1.77 [1.47,2.11], Figure 1—figure supplement 1c). However, we observed no difference between trials with or without ripple disruption (Kruskal-Wallis test: H=1.84, p=0.4, Figure 1g), which taken together with the previous quantifications strongly suggests that performing the NMTS task is not impaired by SWR disruption.

It is possible that the effect of ripple disruption is not reflected in task performance measures, but might be observable in other behavioral parameters. For example, rats could linger on the central platform or arms for longer as an indication of their uncertainty about which arms to visit. For this, we first looked at the running speed at the central platform and arms across all three stimulation conditions. Overall, speed was lower on the central platform than on the arms. When running speed was analyzed separately for arms and center platform, no significant difference was found between the three stimulation protocols (Figure 1—figure supplement 1a). Likewise, the time per visit in the test phase was not impacted by ripple disruption (Figure 1—figure supplement 1b).

Due to a difference in experimental setup (see Methods), the inter-phase time for three animals was longer and more varied compared to the inter-phase times in the remaining two animals (Figure 1—figure supplement 1e). To test if trials with longer intervals may be more susceptible to ripple disruption, we compared the task performance between both groups of animals. We observed a tendency for rats to perform better with short inter-phase intervals, but there was no difference between stimulation conditions for either short or long intervals (Figure 1—figure supplement 1f).

To look for possible compensatory effects of ripple disruption on network activity, we analyzed the location of ripple detections. In line with the literature, the largest fraction of ripples were detected when the rats were on the reward platform (rewards; disruption: 0.85 [0.79,0.90], stim.control: 0.82 [0.77,0.87], no stim.: 0.84 [0.78,0.88], center; disruption: 0.10 [0.06,0.14], stim.control: 0.12 [0.09,0.16], no stim.: 0.10 [0.07,0.14], arms; disruption: 0.05 [0.03,0.08], stim.control: 0.06 [0.04,0.09], no stim.: 0.06 [0.04,0.08], Figure 1h). The bias for ripple detection on the reward platform was not significantly different between the different stimulation protocols (bias computed as the difference in ripple detections on the reward and central platforms divided by the total number of detections; Kruskal-Wallis test: H=1.53, p=0.46), showing that location of the ripple detections did not change across stimulation conditions. We also did not observe a difference in bias when looking at the instruction or test phase separately (Figure 1—figure supplement 1d).

The spatial distribution of ripple occurrence was not influenced by ripple disruption overall (Kruskal-Wallis test: H=1.53, P=0.46), nor in either instruction or test phase (Figure 1—figure supplement 1d). The online ripple detection rate, calculated using only periods of immobility (run speed <5 cm/s) in the trials was also not significantly different between the stimulation protocols in either the instruction and test phase (instruction: disruption: 0.61  Hz [0.49,0.72], stim.control: 0.58  Hz [0.50,0.65], no stim.: 0.64  Hz [0.54,0.76], Kruskal-Wallis test: H=0.73, p=0.7, test: disruption: 0.63  Hz [0.50,0.75], stim.control: 0.49  Hz [0.41,0.58], no stim.: 0.55  Hz [0.44,0.67], Kruskal-Wallis test: H=4.29, p=0.12, Figure 1i).

In a second set of experiments, five rats performed a hippocampus-dependent spatial match-to-sample task (MTS, Table 2) in an eight-arm radial maze (Okaichi and Oshima, 1990; Figure 2a). In this task, rats were required to memorize and return to a number of fixed rewarded locations across repeated trials. In daily training sessions, four randomly selected arms were baited and rats had to learn over the course of 25 trials to only visit the four baited arms in each trial. As learning progressed within each session, rats made fewer excursions to non-rewarded arms and the number of arm visits per trial decreased (Figure 2b and d). Rats were trained until the average number of visits per trial for the final 10 trials met the learning criterium (<5 visits) consistently for three sessions in a row (Figure 2c).

Figure 2 with 1 supplement see all

Download asset Open asset

Once the rats reached the learning criterium and were familiar with the general task to memorize a new set of locations every session, awake ripples were disrupted in one group of sessions (N=32). The task performance of the rats in these disruption sessions was compared to the performance in the remaining control sessions with either delayed stimulation upon ripple detection (N=31) or without stimulation (N=19). Detection and disruption of ripples was performed during all trials and intervening inter-trial intervals. As expected, during the trials most ripples occurred during periods of immobility (ripple rate, offline detected: mean [99% CI], 0.52  Hz [0.43,0.61]) and reward consumption (Figure 2d). The ripple rate in the inter-trial periods was lower than during the trials (mean [99% CI], 0.14  Hz [0.10,0.18]).

The ability of rats to learn the location of the four baited arms was not affected by SWR disruption. In the final 10 trials at the end of each session, neither the average number of arm visits to collect rewards (mean [99% CI]; disruption: 4.42 [4.27,4.71], stim. control: 4.32 [4.18,4.52], no stim.: 4.43 [4.27,4.71], Kruskal-Wallis test: H=2.43, p=0.3, Figure 2e) nor the fraction of trials without any error (disruption: 0.74 [0.64,0.83], stim. control: 0.80 [0.71,0.88], no stim.: 0.73 [0.63,0.83], Kruskal-Wallis test: H=2.56, p=0.28, Figure 2f) differed between disruption and control conditions. To test if the rate of learning over the course of a session was influenced by ripple disruption, we defined a learning trial as the first trial following three correct trials in a row. Rats learned the location of the rewarded arms equally fast in all conditions (learning trial, disruption: 10.78 [8.66,13.94], stim. control: 9.48 [8.26,11.33], no stim.: 9.37 [7.32,13.34], Kruskal-Wallis test: H=0.70, p=0.71, Figure 2g).

We quantified other behavioral parameters to investigate a possible impact of ripple disruption. Similar to the results in the NMTS task, rats ran faster on the arms than on the central platform suggesting that rats take time on the central platform to select the next arm to visit. If SWRs are important for identifying the baited arms, rats might be more uncertain about their choice in disruption sessions which could be reflected in increased lingering at the central platform. We observed no difference in running speed between between the three stimulation protocols (see Figure 2—figure supplement 1a). To assess in more detail if rats showed hesitation to perform the task, we divided each session into three stages (early [1-8], middle [9-17], and late [18-25] trials) and looked at the average time spent per arm visit. For all three stages, the average time per visit was not significantly different for the disruption condition as compared to the control conditions (Figure 2—figure supplement 1b).

Next, we considered that rats could employ a different strategy in the disruption sessions, which could serve as a compensatory mechanism for the missing SWRs. One well-known strategy for solving a match-to-sample task in an eight-arm radial maze is searching for rewarded arms in clockwise or counterclockwise order (Okaichi and Oshima, 1990). Rats did not engage in circular choice behavior more often than would be expected by chance for any of the stimulation conditions, and a direct comparison between ripple disruption and control conditions did not reveal a change in circular choice behavior (Figure 2—figure supplement 1c). We noticed that once the rats knew which four arms delivered reward, they tended to visit these arms each time in the same order. To investigate if SWR disruption influenced this stereotypical behavior, we computed a stereotypy index for every session by computing the average pairwise similarity between visit sequences of the final trials. The similarity measure was based on the Levenshtein distance (Levenshtein, 1996) that equals 0 if the visit sequences are the same, and adds 1 for every deletion, insertion or substitution necessary to map one visit sequence onto another. As expected, the stereotypy index was on average higher than expected by chance (disruption: 0.62 [0.55,0.70], stim.control: 0.65 [0.56,0.75], no stim.: 0.59 [0.51,0.68], Figure 2—figure supplement 1c) but was not significantly different between disruption and control sessions (Kruskal-Wallis test: H=0.64, p=0.72).

In all of the above analyses, we assumed that each configuration of four arms is equally difficult to learn. To confirm this assumption, we split the configurations in three different categories; (IIa) two sets of neighboring arm pairs, (IIb) one pair of neighboring arms and two arms without direct neighbors and (III) three neighboring arms and one arm without direct neighbors. For this analysis, we left out the sessions in which all four arms have no direct neighbor due to a low sample size (n=2 for each condition). For all three arm configurations, ripple disruption had no influence on any performance quantification, see Figure 2—figure supplement 1d.

Similar to the NMTS task, the largest fraction of ripples was detected when the rats were on the reward platforms (disruption: 0.91 [0.88,0.93], stim.control: 0.92 [0.89,0.94], no stim.: 0.93 [0.91,0.95]) followed by the center (disruption: 0.07 [0.05,0.09], stim.control: 0.05 [0.04,0.09], no stim.: 0.04 [0.03,0.06]) and arms (disruption: 0.02 [0.02,0.04], stim.control: 0.02 [0.01,0.03], no stim.: 0.03 [0.02,0.05], Figure 2h). The spatial distribution of ripple occurrence was not influenced by ripple disruption (Kruskal-Wallis test: H=1.80, p=0.41). The online detection rate in the trial or inter-trial periods, calculated using only periods of immobility (run speed <5 cm/s), was not different in disruption sessions versus control sessions (trial; disruption: 0.62  Hz [0.52,0.78], stim.control: 0.50  Hz [0.42,0.56], no stim.: 0.56  Hz [0.43,0.66], H=4.71, p=0.095, inter-trial: disruption: 0.19  Hz [0.15,0.27], stim.control: 0.13  Hz [0.11,0.16], no stim.: 0.15  Hz [0.11,0.19], H=5.90, p=0.052, Figure 2i).

The results from the first two experiments suggest that SWRs do not support the memorization of a set of locations within a short timeframe (single trial or session). Previous ripple manipulation studies used a spatial alternation task in which rats were required to go from a center arm to two outer arms in an alternating fashion, that is left, center, right, center, left,… To learn this rule, it is crucial to discover and remember the correct temporal order of arm visits, something that was not required in either the MTS or NMTS task.

To test if SWRs are required for memorizing the temporal order of two or more recent events, we designed a spatial sequence memory paradigm (Figure 3a). In this task, rats need to learn each day the order in which to visit a set of four pseudo-randomly selected arms by remembering which arm-arm transitions are rewarded. After extensive pre-training on the linear track and a three-arm radial maze, five rats were pre-trained on the sequence memory paradigm (Table 3) for several days until the final average sequence performance (average sequence performance in the last 100 arm visits) was above 50% for at least two sessions in a row. The sequence performance is calculated by multiplying the outcome of all individual visits (1=correct transition, 0=incorrect transition) in a moving window of five visits. A correct sequence performance (i.e. five correct visits in a row), will yield a sequence performance of one (Figure 3b and c).

Figure 3 with 1 supplement see all

Download asset Open asset

To test the involvement of awake SWRs in memorizing and learning a spatial sequence, we disrupted ripples in one third of the sessions (N=23), throughout the two run epochs and the rest epoch. The other sessions were used for stimulated (N=29) and non-stimulated (N=31) controls, similarly as defined for the MTS and NMTS task. Ripples were mostly detected on the reward platform when the animal is immobile (ripple rate, offline detected, 0.47  Hz [0.41,0.53], Figure 3c). The ripple rate during the interleaving rest epoch is comparably high (ripple rate, offline detected, 0.58  Hz [0.53,0.64]).

Quantification of the average visit and sequence performance in the final visits indicated no significant difference between the stimulation protocols (visit performance; disruption: 0.85 [0.71,0.94], stim.control: 0.87 [0.74,0.93], no stim.: 0.85 [0.73,0.91], Kruskal-Wallis test: H=2.61, p=0.27, sequence performance; disruption: 0.67 [0.44,0.83], stim.control: 0.71 [0.53,0.83], no stim.: 0.63 [0.47,0.75], Kruskal-Wallis test: H=2.55, p=0.28, Figure 3e and f). To test if rats learn at a different rate, we computed for every session the smoothed learning curve of both the visit and sequence performance and defined the visit where the curve is significantly above 0.5 as the learning visit. In every stimulation condition, there was a similar percentage of sessions in which the rat does not reach the learning criteria, which we left out for this quantification (visit performance; disruption: n=5, stim. control: n=3, no stim: n=4, sequence performance; disruption: n=8, stim. control: n=10, no stim: n=11). The learning visit for both the visit and sequence performance was not significantly different between sessions with a different stimulation protocol (learning visit performance; disruption: 117.72 [74.64,173.65], stim.control: 112.69 [76.92,154.99], no stim.: 154.22 [109.28,201.30], Kruskal-Wallis test: H=2.66, p=0.26, learning visit sequence performance; disruption: 238.80 [189.36,279.90], stim.control: 209.16 [156.94,263.22], no stim.: 273.15 [202.33,315.75], Kruskal-Wallis test: H=6.01, p=0.05, Figure 3g and h).

Quantification of general behavioral parameters did not reveal any change due to ripple disruption. Like in the previous experiments, we quantified the speed on the central platform and arms and looked at the time per visit as a measure for how certain rats were about their choice. The speed on the central platform and arms showed no difference in different stimulation conditions (Figure 3—figure supplement 1a). Rats did not take more or less time per visit in disruption versus control trials (Figure 3—figure supplement 1b).

When we looked at the smoothed learning curves for the individual arm-arm transitions, we observed that not all transitions were learned equally fast (Figure 3—figure supplement 1c). This is an indication that rats are probably not learning individual transitions, but are aware that there is a repeating pattern to be found. Namely, the possibilities for the new transitions are lower when using the information of learned transitions in the correct temporal order. By looking at the learning visits for the individual transition learning curves, we can investigate if the strategy employed by the rats changes due to ripple disruption. For all three stimulation protocols, learning visits of individual transitions vary and the distributions were significantly different between stimulation protocol (Figure 3—figure supplement 1d). Post-hoc Mann-Whitney tests, however, revealed only a significant difference between the two control conditions (Figure 3—figure supplement 1d). Next, we quantified the difference between the learning visits of the slowest and fastest learned transition in sessions where the rat had learned at least two transitions, as a measure of how much use had been made of the repeating pattern. This quantification of the learning trial difference in each session showed no significant difference due to ripple disruption (Figure 3—figure supplement 1e). Together these quantifications suggest that SWR disruption has no overall effect on the behavior in the sequence task.

In line with our expectations, ripples were detected mostly at reward sites (rewards; disruption: 0.82 [0.78,0.85], stim.control: 0.85 [0.82,0.87], no stim.: 0.85 [0.82,0.87], arms; disruption: 0.08 [0.06,0.10], stim. control: 0.07 [0.06,0.09], no stim.: 0.06 [0.05,0.07], center; disruption: 0.10 [0.09,0.12], stim. control: 0.08 [0.07,0.10], no stim.: 0.09 [0.08,0.11], Figure 3i), and this bias was not significantly different between stimulation protocols (Kruskal-Wallis test: H=0.33, p=0.85). The overall ripple detection rate calculated over the immobility periods in both run epochs and in the rest epochs did appear to be significantly different between sessions with a different stimulation protocol (RUN; disruption: 0.58 [0.52,0.67], stim.control: 0.46 [0.41,0.50]), no stim.: 0.49 [0.41,0.57], Kruskal-Wallis test: H=12.94, p=0.0016, (REST; disruption: 0.69 [0.58,0.78], stim. control: 0.49 [0.43,0.55]), no stim.: 0.64 [0.58,0.70], (Kruskal-Wallis test: H=21.02, p=$2.7\times {10}^{-5}$). Post-hoc Mann-Whitney tests for the run and rest epochs revealed a significant difference between the disruption and stimulated control sessions (run and rest) or between the stimulated control and non-stimulated control session (rest) and not between the other pairs (RUN: disruption - stim.control: U=1944.00, p=$6.6\times {10}^{-5}$, disruption - no stim.: U=1728.00, p=0.061, stim.control - no stim.: U=2011.00, p=0.26 REST: disruption - stim.control: U=520.00, p=0.00019, disruption - no stim.: U=435.00, p=0.17, stim.control - no stim.: U=695.00, p=$7.7\times {10}^{-5}$).

All experiments were conducted using the same implementation for online ripple detection and disruption as in our previous work (Michon et al., 2019), where it was shown that ripple-triggered VHC stimulation following learning reduced spatial memory for highly rewarded places. Here, we use the same post-hoc quality metrics for both the disruption and detection to show that we indeed performed the manipulations in a way that is comparable to those that impacted behavior in a different scenario.

Before each experimental session, the threshold for ripple detection and the stimulation current were manually adjusted by the experimenter to ensure accurate and complete detection and disruption. The disruption quality was quantified for all sessions in stimulated control and disruption condition by comparing the ripple amplitude in a 20ms window before and after the time of detection. The normalized post-detection ripple amplitude was significantly lower in the sessions with ripple disruption in all three tasks (Mann-Whitney test, NMTS: U=912.00, p=$5.9\times {10}^{-11}$, MTS: U=961.00, p=$7\times {10}^{-12}$, SEQ: U=2615.00, p=$2.6\times {10}^{-18}$), showing that detected ripples were successfully disrupted (Figure 4a).

Figure 4 with 2 supplements see all

Download asset Open asset

Since the stimulation electrode is placed in the ventral hippocampal commissure, a large white matter structure constructed from axons connecting the left and right hippocampi, ripples are disrupted in both hemispheres. Indeed, recordings from three rats with recording tetrodes placed in both the left and right hemisphere showed a clear disruption of ripples in both hemispheres using the same unilateral stimulation method as in the three behavioral experiments (Table 4; Figure 4—figure supplement 1).

The quality of the ripple detection was quantified based on data from the stimulated and non-stimulated control sessions. In each session, ripples were detected offline (see Methods) and compared to the online detected ripples by quantifying a true positive rate (TPR) and false discovery rate (FDR) (Figure 4b). The high TPR for almost all sessions suggest that a large fraction of all offline-defined ripples were detected online. We observed a wider distribution for the FDR in comparison to the distribution for the TPR, likely due to a variable signal-to-noise ratio across experiments. Ripple detection and disruption were quantified separately for the inter-trial times in the MTS task and the rest epochs in the SEQ task (Figure 4—figure supplement 2). In both cases, ripple power was significantly reduced following stimulation (Figure 4—figure supplement 2a). The TPR-FDR distribution for the inter-trial periods in the REF task was shifted toward lower TPR and higher FDR, indicating a worse detection (lower TPR and higher FDR, Figure 4—figure supplement 2b). In the inter-trial times, the animals were locked in the center of the maze and they frequently engaged with the closed doors, which resulted in movement artifacts in the LFP signal that were falsely identified as ripples. Quantifications for the sleep box epochs in the SEQ task were similar to those on the maze, that is high TPR and variable FDR (Figure 4—figure supplement 2c). As online detection of ripples can only occur after their initiation, there will always be a detection delay (Figure 4c). The absolute detection delay over all epochs in each task, and considering only the no stim. and stim.control epochs, is median [iqr], 36.69ms [34.68,39.19], in line with what is expected when using the Falcon software platform (Ciliberti and Kloosterman, 2017). When we compare this to the total length of the ripple, we obtain a relative delay (median [iqr]) of 33.32% [29.97,36.47] for long ripples (duration >100ms) and 57.79% [54.61,60.49] for the remaining ripples. The average peak ripple amplitude, quantified by filtering in the ripple frequency band (140–225 Hz) and taking the peak value of the envelope in the ripple interval, is 159.76 μV [117.61,195.45].

A total of 15 male Long Evans rats (supplier Janvier Labs), food restricted to 85–90% of the free-feeding weight, were used in this study. At the start of behavioral training procedures, rats were 7–10 weeks old. All rats received an implant for neural recordings and stimulation. Tables 1–3 provides an overview of the experimental sessions for each rat. All experiments were carried out in accordance with protocols approved by KU Leuven animal ethics committee (P119/2015 and P175/2020) and in accordance with the European Council Directive, 2010/63/EU. Animals in experiment were housed separately in individually ventilated cages (IVC) with ad libitum access to water and controlled intake of standard food pellets. Health status and body weight were checked daily by the experimenters and dedicated animal care personnel. To improve the well-being of the rats, a playpen (100x55 cm) was constructed in the course of this study that allowed rats to spend spent time in an enriched environment with their (former) cage mate. The playpen can be divided into two parts separated by a transparent wall with ‘sniffle holes’ to encourage social interaction between the two rats without accidental damage to their implants. The enrichment consisted of a small maze with hidden food, toys, a climbing rope and ladder that provide access to an upper story with a running wheel. Experiments for the delayed non-match-to-sample and match-to-sample tasks were started prior to completion of the playpen. All five rats used to test the memory of temporally ordered information (third experiment), however, spent at least 30 min per day in the playpen.

All animals were pretrained to run on elevated mazes for reward, habituated to automated doors (when used) and to a separate sleep box. For each of the three behavioral tasks, the pretraining and training procedures to teach the rats the task rules are described below.

In the first phase of pretraining (5–7days), rats are habituated to the eight-arm radial maze and learn to look for reward at the end of the arms. On days 1–2, rats are placed in the center of the maze and allowed to freely explore for 15–30min. Chocolate rewards are scattered throughout the maze to promote exploration. On the next 2–3days, rewards are progressively restricted to only the arms and finally the reward wells at the end of each arm. Once rats are sufficiently comfortable and motivated to run for the rewards, the second phase of the pretraining starts. This phase lasts three days and rats are required to perform at least three to six trials of the NMTS paradign. Rats are then taken off food restriction and undergo surgery. After 1 week of post-surgery recovery, the food restriction is resumed and once rats are at 85–90% of their post-surgery weight. Finally, rats run 6–15 trials per day until the average performance is above 50% for at least three days in a row.

The pretraining prior to surgery is the same as for the NMTS paradigm, except in the second phase where rats perform daily sessions of the MTS paradigm until they complete at least 15 trials in 30 min. Training is completed after recovery from surgery when rats are back at 85–90% of their post-surgery weight. In daily sessions (25 trials), rats perform the MTS paradigm until the average number of visits on the last 10 trials is less than 5 for 3 days in a row. In total rats take on average 14 days to reach the learning criterium.

Prior to surgery, rats are trained to run back and forth on a linear track (130 cm) for liquid reward. Animals perform two 15 min sessions per day, with a 15–20 min intervening rest in the sleep box. Training continues until rats collected at least 50 rewards in 15 min. Food restriction is then stopped to prepare for surgery. After one week of post-surgery recovery, food restriction is resumed, and rats are put back on the linear track until they perform at least 70 crossings in 15 min. Next, rats are trained on a continuous spatial alternation task in a three-arm radial maze following a single-day acquisition procedure that consist of eight 15 min learning sessions. Training on the alternation task is repeated on 3 days (separated by a rest day), and each day a different arm is chosen as the home arm. On the rest days, the rats perform two 15 min sessions with the same home arm as the day before. Finally, rats are introduced to the sequence memory paradigm as described above until their sequence performance is above 80% by the end of the session, for 3 days in a row. We found that pretraining rats on the alternation task improved the acquisition of the sequence task.

A custom-designed 3D-printed micro-drive array (Kloosterman et al., 2009), carrying up to eight tetrodes (four twisted 0.012 mm polyimide- insulated nickel-chrome wires, angled cut; Sandvik, Kista, Sweden) and three stimulation electrodes (two twisted 0.06 mm polyimide-insulated stainless-steel wires, angled cut; California Fine Wire, Grover Beach, CA), was surgically attached to the rat skull using standard aseptic techniques. To induce anesthesia, the rat was placed in an induction chamber filled with oxygen (0.5–1 L/m) and 5% isoflurane. Next, its head was securely mounted in a stereotaxic frame after shaving the head, and eyes were protected from drying out and prolonged light exposure with eye ointment and aluminum foil. During the surgery, anesthesia was maintained by administration of 0.5–2% isoflurane through a nose mask and adjusted if necessary, based on vital signs: blood oxygen level, heart rate and breathing rate. The body temperature (measured with a rectal probe) was kept constant using a heating pad. After disinfection of the skin with iodine and ethanol, an incision with a scalpel along the mid-line exposed the skull. Small scratches were carved in the bone plates to allow better adhesion of the dental cement used to fix the implant to the skull. Nine anchoring bone screws were inserted (three frontal, two left parietal, two occipital, and one right parietal) and gaps between the screws and the skull were filled up with bio compatible glue (VetBond) and after extra disinfection (10 min Baytril submersion) all screws were fixed together with dental cement. Next, two craniotomies were drilled for rats for the behavioral experiments, and the dura was removed to allow access to the brain above the HC and ventral hippocampal commissure (VHC) (HC-craniotomy center coordinates: 4 mm posterior to Bregma, 2.5 mm right from the midline; VHC-craniotomy center coordinates: 1.3 mm posterior to Bregma, 0.9 mm right from the midline). For bilateral recordings two extra craniotomies were drilled above the HC and VHC on the other hemisphere (HC-craniotomy center coordinates: 4 mm posterior to Bregma, 2.5 mm left from the midline; VHC-craniotomy center coordinates: 1.3 mm posterior to Bregma, 0.9 mm left from the midline). Before mounting of the implant, craniotomies were covered in silicon grease to seal the gaps between the edge of the craniotomies and the tetrode-carrying tubes. Additionally, mineral oil was applied to the tip of the cannulas prior to implantation which prevents back-filling (and possible clogging) of the tetrode-carrying tubes with cerebrospinal fluid or blood. The implant was fixed to the skull with light-curable dental cement (SDI, Bayswater, Australia) and one screw was wired to the electrode interface board to serve as electrical ground. The skin was sutured in the front and back of the implant with surgical threads (4–0 Silk wax coated braided silk, Sofsilk). While the rat was still under light anesthesia, all tetrodes and stimulation wires were lowered 1 mm into the cortex. Finally, 0.7 ml of saline (anti-dehydration) and 0.3 ml Metacam (anti-inflammatory and pain relieve) were administered subcutaneously. The Metacam injection was repeated in the three days following completion of the surgery.

After approximately one week of post-surgical recovery, four to five tetrodes were positioned in the CA1 pyramidal cell layer of the dorsal HC over the course of 3 days to minimize tissue damage. One tetrode was lowered in the white matter above the CA1 cell layer to serve as a reference and one tetrode was placed in the cortex above (parietal association cortex) to aid the online ripple detection. Wide-band (0.1–4 kHz) signals were sampled at 4 kHz, digitized using a 128- channel data acquisition system (Digilynx SX acquisition system with HS-36 analog headstage and Cheetah software; Neuralynx, Bozeman, MO) and saved to a hard disk for offline analysis.

A live network stream of digitized multi-channel samples from the Digilynx acquisition system was fed into a quad-core workstation that runs the real-time detection software Falcon (Ciliberti and Kloosterman, 2017). Falcon initiates TTL pulses via a microcontroller board (Arduino UNO) that is connected to a constant-current stimulator (MultiChannel System, Reutlingen, Germany) which generates a biphasic current for electrical stimulation of the VHC. Here, raw signals of 2–4 electrodes (with the clearest ripples) are first filtered in the ripple band (135–255 Hz) using a Chebyshev type-II IIR filter (order 20). Using the neuralynx acquisition system, the total round trip latency was below 1ms (Ciliberti and Kloosterman, 2017). After summing of the ripple power (RP) of the different electrodes, ripples were identified based on the following criteria.

In this equation, μ(t) and mad(t) are the running estimates of the mean and mean absolute deviation calculated on the ripple power, and f is a multiplier set to a value in the range 5–13. The value of f was adjusted for each daily session to maximize the ratio of positive over false ripple detections. The RP is the root mean square of the filtered signal. Estimates of μ(t) and mad(t) were computed using an exponential moving average filter with span set to 7 s. Estimates were not updated during a 50ms window after each detection to avoid multiple detections of the same event. To avoid false detections due to heavy head movements or chewing artifacts, the same ripple detection procedure was applied to signals from the cortical electrode. Detections were marked as false positive if a cortical detection occurred within a fixed time window ([–40, 1.5] ms) around a HC detection. While ripple-like events have been described in the parietal association cortex during sleep, only a small fraction is coupled to HC ripples and they have a much lower power compared to HC ripples (Khodagholy et al., 2017; Aleman-Zapata et al., 2022). In our experiments in the awake state, the limited cortical detections predominantly represent artifacts rather than joint hippocampal-cortical ripples.

The detected hippocampal ripples triggered the generation of a second TTL pulse by Arduino UNO either immediately (disruption condition) or after a random delay between 150 and 250ms. This TTL-pulse was sent to a constant-current stimulator (MultiChannel System, Reutlingen, Germany) to electrically stimulate the VHC. Both the detection event and the time of stimulation (=stimulation event) were sent to the Neuralynx acquisition system for logging. The amplitude of the biphasic electrical pulses (0.2ms duration) varied from 50 to 250 μA and was set in each session to the lowest amplitude that resulted in consistent disruption of hippocampal ripple events. To avoid over-stimulation that could lead to damage of the surrounding tissue, no stimulations could occur within a 150ms lockout period after a prior stimulation. In the disruption condition, this is achieved by discarding detections in a 150ms window after each detection. In the stimulated control condition, all detections that would trigger a stimulation that falls withing a 150ms window after the delayed stimulation are were discarded. Ripple detection and disruption started after rats acquired the task rules reached according to the learning criteria specific for the behavioral task. In the NMTS paradigm, ripple detection and/or disruption occurs only during a trial (i.e. both during and in between the instruction and test phase), but not in the inter-trial time. For every trial one of the control/disruption conditions was picked pseudo-randomly. In the MTS paradigm, ripple detection and/or disruption was applied during both the trials and the inter-trial times intervals. In the sequence memory paradigm, ripple detection and/or disruption was applied during both run epochs and rest epoch. In both the MTS and sequence memory paradigm, the stimulation condition was constant through the whole session.

Analysis of neural and behavioral data was performed using Python and its scientific extension modules (Millman and Aivazis, 2011), augmented with custom Python and C++toolboxes. Throughout the result section, all description summary statistics report mean [99% CI], unless stated otherwise.

Position and speed of the rat were tracked using LED lights mounted on one of the headstages and an overhead video camera (25 Hz). Entry into a maze arm was detected as soon as the animal moved past the first 20 cm of the arm (regardless of the rat actually reaching the reward platform). To assess stereotypical behavior in the MTS task, we defined for every session a stereotypy index. First, a string representing the sequence of arm visits in a trial was created for each of the last 10 trials in a session (for example, ‘1345’ or ‘5471’). The stereotypy index was defined as the average similarity (based on the normalized Levenshtein distance) between the strings of all trial pairs. The Levenshtein distance measures the minimum number of single character edits that are necessary to change one string into another, and for two strings a and b with lengths |a| and |b| is computed according to the following formula:

with $\text{tail}\left(x\right)$ equal to all of string x except for the first character. The similarity is then computed as one minus the normalized Levenshtein distance:

The shuffled distributions used to compute the z-scores for the circularity and stereotypy quantifications of every MTS session, were created by randomizing the visit order of each trial 2000 times.

For behavioral analysis of the sequence paradigm, only sessions in which the rat performed at least 150 visits in the first run epoch are considered. After this selection, data from run 1 and run 2 are merged and we only consider the first 400 visits to assess the (learning) behavior. To compute smoothed learning curves we use a Gaussian kernel with a bandwidth of 50 trials. Next, we calculated the corresponding binomial probability that the (smoothed) number of correct responses is significantly higher (alpha = 0.01) than would be expected by chance (0.5) and define the last trial at which this criteria is met. If the criteria were never met, no learning visit was defined.

Ripple events were detected in the hippocampal field potential for all trials and sessions of the non-stimulated and stimulated control conditions. For the stimulated control condition, the stimulation artifact is first removed and replaced by cubic interpolation of the signal in a 10ms window around the stimulation. To detect ripples offline, the local field potential recorded from 1 to 3 tetrodes was filtered in the ripple frequency band (140–225 Hz). The ripple envelope was computed as the absolute value of the Hilbert-transformed ripple signal, averaged across the recording sites and smoothed with a Gaussian kernel (bandwidth 15ms). Finally, ripple events were detected when the ripple envelope exceeded a high threshold of $\text{med}+9\cdot (\text{Q3}-\text{med})$. Here, med and Q3 represent the median and upper quartile range of the envelope. To determine the start and end time of each ripple, a second threshold was computed with the same method as before, but now only using the signal statistics of a 30ms window before each detected ripple (–100ms to –70ms) and a multiplier of 0.4 instead of 9. Ripple events that were separated by less than 20ms were merged into one, and events with a duration shorter than 40ms were excluded. Offline ripple detection was performed on the same task epochs when online ripple detection was performed. It should be noted that the multiplier for the percentile-based threshold, that we use, should not be compared directly to the multiplier for z-score based threshold that is used in previous studies because of the different statistics being used (median / upper quartile range vs mean / standard deviation). When computing the equivalent z-score multiplier that produces the same absolute threshold as our percentile multiplier, we find that these multiplier values are comparable to the 3–6 standard deviations reported in previous papers (Girardeau et al., 2009; Jadhav et al., 2012; Fernández-Ruiz et al., 2019; Figure 4—figure supplement 2d e).

Online ripple detection accuracy was verified for all trials and sessions of the non-stimulated and stimulated control conditions. Online detected ripples were compared to offline detected ripples to identify the fraction of offline ripples that were also detected online (true positive rate or TPR) and the fraction of online detections that did not correspond to an offline detected ripples (false discovery rate, FDR). To assess the disruption of ripples after a closed-loop stimulation of the VHC, the mean ripple envelope after each detection/stimulation was computed in a 30ms time window (from +20ms to +50ms) and normalized to the mean ripple envelope in a 30ms time window (from –50ms to –20ms) before detection/stimulation (=normalized post-detection ripple power). The ripple envelope was computed as described above, including the removal of the stimulation artifact for the disruption session. The artifact removal was also performed at the time of detection for the stimulated control sessions to have a fair comparison between the disruption and stimulated control sessions. The normalized post-detection MUA power was obtained in the same way but using the LFP filtered in 300–2000 Hz range.

To test a difference in means between two (unpaired) samples, we used the Mann-Whitney test. To test the difference in means between three (unpaired) samples we used the Kruskal-Wallis test.

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

We thank Jyh-Jang Sun, Frédéric Michon and Ta-Shun Su for help with the development of the sequence paradigm and Marine Chaput for setting up the automated maze setup and live video tracking. LD is funded by Research Foundation Flanders (FWO), Belgium as a PhD Fellow fundamental research, grant number 11D9322N. FK is funded by Research Foundation Flanders (FWO), Belgium under grant number G077321N and KU Leuven, Belgium C1 grant C14/17/042

All experiments were carried out in accordance with protocols approved by KU Leuven animal ethics committee (P119/2015 and P175/2020) and in accordance with the European Council Directive, 2010/63/EU.

1. Received: October 6, 2022
2. Preprint posted: November 3, 2022 (view preprint)
3. Accepted: March 25, 2024
4. Accepted Manuscript published: March 26, 2024 (version 1)
5. Version of Record published: April 15, 2024 (version 2)

© 2024, Deceuninck and Kloosterman

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.