Sleep loss diminishes hippocampal reactivation and replay

Summary Paragraph

Memories benefit from sleep¹ and the reactivation and replay of waking experiences during hippocampal sharp-wave ripples (SWRs) are considered to be crucial for this process². However, little is known about how these patterns are impacted by sleep loss. We recorded CA1 neuronal activity over ~12 h in rats across maze exploration, sleep, and sleep deprivation following by recovery sleep. We found that SWRs showed sustained or higher rates during sleep deprivation but with lower power and higher frequency ripples. Pyramidal cells exhibited sustained firing during sleep deprivation and reduced firing during sleep, yet their firing rates were comparable during SWRs regardless of sleep state. Despite the robust firing and abundance of SWRs during sleep deprivation we found that the reactivation and replay of neuronal firing patterns was diminished during these periods, and in some cases completely abolished compared to ad-libitum sleep. Interestingly, reactivation partially rebounded upon recovery sleep, but failed to reach the levels in natural sleep. These results delineate the adverse consequences of sleep loss on hippocampal function at the network level and reveal a dissociation between the large number of SWRs elicited during sleep deprivation and the few reactivations and replays that occur during these events.

Introduction:

Memories undergo continuous refinement following learning, in a process referred to as memory consolidation in which sleep plays a critical role. Sleep immediately after learning benefits memories¹ and memories can be disrupted by even a few hours of sleep loss³. Studies have highlighted the particular importance of the hippocampus for sleep-dependent memory consolidation. However, the mechanisms through which memories are impacted by sleep loss have yet to be understood. Hippocampal sharp-wave ripples (SWRs), which feature sharp-waves in the dendrites of CA1 pyramidal cells coupled with ripple oscillations (150–250 Hz) near the cell bodies, are widely considered to play a critical role in sleep-dependent memory processes. SWRs are observed more frequently in sleep after memory tasks⁴. Disrupting activity during these oscillations impairs memory^5,6, while enhancing them improves memory⁷.

Why are hippocampal sharp-wave ripples so important to memory? A key characteristic of these signals is that they are generated in the CA3 region of the hippocampus and then produce intense spiking activity throughout the hippocampal formation⁸ and beyond^9–11. Such synchronized activity drives synaptic plasticity in the network connections associated with individual memories, enhancing their storage and recall^12,13. In fact, both synaptic strengthening, via long-term potentiation^14,15 and synaptic weakening, via depotentiation or long-term depression^16,17, have been associated with SWRs. In particular, the spiking activity during SWRs can be highly patterned to reactivate and replay activities initially expressed during learning and behavior in a temporally compressed manner akin to rapid rehearsal¹⁸. By generating such rehearsals, SWRs can strengthen and stabilize spatial representations in the hippocampus^6,19, and broadcast this signal to cortical and subcortical brain regions^8,9. While reactivations and replays during SWRs are widely considered to play a key role in the memory consolidation process, remarkably nothing is known about how these events are impacted by sleep deprivation (SD).

Long-duration recordings during behavior, sleep, and sleep deprivation

We performed extracellular recordings using 128 channel high-density silicon probes implanted uni- and bilaterally in the CA1 region of the rat hippocampus (Methods) during behavior and sleep and tracked local field potentials and stable unit putatively classified into 754 pyramidal neurons (PN) and 96 interneurons (IN). Recordings initiated ~3.5 h before the onset of the light cycle with ~2.5 h of rest and sleep in a homecage (PRE). Animals were then placed in novel linear maze environments of differing shapes (MAZE) that they had not previously explored and allowed to run for ~1 h for water reward. Following the maze, animals were returned to the homecage for POST sessions that involved either natural (ad-libitum) sleep and rest (NSD) for ~9 h, or sleep deprivation (SD) via gentle handling for ~5 h followed by recovery sleep (RS) (Fig. 1A). We divided these periods into 2.5 h blocks (NS1–3 vs. SD1–2 & RS) aligned to zeitgeber time (ZT) = 0, the onset of the light cycle, and compared RS vs. NS1, the first blocks of ad-libitum sleep in each group, as well as SD2 vs. NS2, to reveal the effects of prolonged wakefulness relative to ad-libitum sleep. SD (8 sessions from 7 animals) and NSD sessions (7 sessions from 6 animals) were carried out in pseudo-random order on different days spaced > 24 h apart, in the same animals.

Figure 1: Sleep deprivation yields more sharp-wave ripples but with weaker power and higher frequency ripples.

(A) After PRE, animals were introduced to MAZE then allowed either undisturbed sleep (NS1-NS2), or 5 h sleep deprivation (SD1-SD2) followed by recovery sleep (RS). (B) Power spectral density in sample NSD (left) and SD (right) sessions with hypnogram (top) indicating brain state (active wake (AW), quiet wake (QW), rapid eye movement (REM), and non-REM (NREM) sleep) and spectrogram (bottom; z-scored over all frequencies for the time periods shown) of CA1 local field potential (LFP). (C) Average power spectral densities across all NSD & SD/RS sessions (black & red/blue with corresponding s.e.m. shaded). (D) The rate of delta waves is lower during SD vs. sleep but increases from SD1 to SD2 and RS (individual sessions superimposed with connected dots). (E) Sample sleep (left) with a high spontaneous rate of sharp-wave ripples (SWRs) with LFPs (2 shanks in black) and unit rasters (arbitrary color and sorting). The rate of SWRs (right) decreases with sleep but remains elevated during SD. (F) Power spectral densities in the ripple frequency band for the sessions in (B) with moving average ripple frequency (black). Sample SWRs (16-channel traces, white) at different time points (arrow heads). (G) Box plots showing population median and top/bottom quartiles (whiskers = 1.5 interquartile range), estimated using hierarchical bootstrapping, indicate higher frequency ripples in SD (n = 157964 ripples total from 8 sessions) vs. NSD (n = 143681 ripples total from 8 sessions), with a rebound in RS. Session means overlaid as connected dots. Rightmost panel highlights cross-group comparisons for the first block of sleep in each group (NS1 vs. RS) and the second block of SD vs. NSD. (H, I) Same as (G) for sharp-wave amplitude (H) and ripple band power (I). Statistics: All panels: two-sided within-group comparisons and one-sided cross-group comparisons; Panels D and E: t-tests; Panels G-I, comparisons of bootstrapped means; ns (not significant), #P <0.10, *P < 0.05, **P < 0.01, ***P < 0.001, with no corrections for multiple comparisons. See Supplementary Statistics Table for additional details.

Power spectral calculations (Fig. 1B, C) demonstrated strong delta (< 4 Hz) power in the hippocampal local field potential during natural slow-wave/NREM sleep and strong theta (5–10 Hz) during REM sleep (Extended Data Fig 1 & Extended Data Fig 2A,B). We saw evidence for neither prominent delta during sleep deprivation nor for prominent theta outside of REM periods²⁰. However, the rate of isolated delta waves (Fig 1D) increased throughout the SD period, indicative of micro- or local sleep^21,22 (see also Extended Data Figure 2C for detected OFF states). Overall, sleep deprivation was characterized by lower spectral power across frequencies (Fig. 1C). Recovery sleep following sleep deprivation subsequently featured a robust rebound in delta activity (Fig 1C,D, Extended Data Fig 1, and Extended Data Fig. 2C), consistent with sleep homeostasis^23,24.

A higher rate of sharp-wave ripples during sleep deprivation.

Previous studies have suggested that the incidence rate of ripples and associated population burst events play important homeostatic roles in hippocampal dynamics^16,17,25. We therefore asked how the rate of these events changes during sleep compared to a similar period during extended wakefulness (Fig. 1E). In naturally sleeping animals, we found that the incidence rate of SWRs decreased over time, consistent with a homeostatic effect from sleep (NS1 median = 0.57 Hz (interquartile range (IQR) = 0.06 Hz) vs. NS2 median = 0.46 Hz (IQR = 0.03 Hz), P = 1.86 × 10⁻³ , paired t-test (df = 8)). In contrast, the rate of SWRs remained high in animals during sleep deprivation (SD1 median = 0.5 Hz (IQR = 0.16) vs. SD2 median = 0.57 Hz (IQR = 0.03), P = 0.73, paired t-test (df = 8)) and was higher during the second block (ZT = 2.5–5 h) of SD compared to NSD (SD2 vs. NS2, P = 1.08 × 10⁻³, t-test (df1 = 8, df2 = 8); see also Extended Data Fig 2D,E). Once the SD animals were allowed to sleep (at ZT = 5 h), the rate of ripples dropped to levels lower than those in the early block of ad-libitum sleep (RS median = 0.45 Hz (IQR = 0.19) vs. NS1 median = 0.57 Hz (IQR = 0.06), P = 7.87 × 10⁻³, t-test (df1 = 8, df2 = 8)). In both NSD and SD, this ripple rate was consistently modulated by delta waves and the probability of an OFF state following a ripple increased over the course of SD (Extended Data Fig 2F–J), as expected¹¹. Overall, the number of SWRs was not negatively affected but was rather higher during sleep-deprivation compared to natural sleep.

Sleep loss alters the physiological properties of sharp-wave ripples

Given the prevalence of SWRs during both sleep and sleep deprivation, we hypothesized that other characteristics of these hippocampal events might differ across these periods (Fig. 1F). The peak frequency of ripples in our recordings (Fig. 1G) decreased over the course of sleep, (NS1, mean = 165.32 Hz (IQR = 1.75) vs. NS3 mean = 154.37 Hz (IQR = 2.93 ), P < 2 × 10⁻⁴ (Hierarchical bootstrap (HB) based on 10⁴ random samples with replacement across data levels²⁶), but during SD, ripple frequency remained elevated (SD2 mean = 170.13 Hz (IQR = 1.18) vs. SD1 mean = 171.55 Hz (IQR = 1.58) P = 0.14 , HB) and was significantly higher compared to ad-lib sleep, (SD2 mean vs. NS2 mean = 154.87 Hz (IQR = 1.24), P < 10⁻⁸, HB). The high frequency of ripples during SD (but not NSD) was also higher than those seen during MAZE (P = 0.0204, HB) or PRE sleep (P < 10⁻⁴). While changes in ripple frequency on the order of several Hz may be expected based on temperature differences across sleep and awake²⁷, we observed larger differences of up to ~15 Hz (e.g. SD2 vs. NS2); these differences remained significant when accounting for state dependence (Extended Data Fig 2K). Upon recovery sleep, ripple frequency dropped rapidly to levels lower than during the similar sleep period in NSD (RS mean = 155.40 Hz (IQR = 1.74) vs. NS1 mean =165.32 Hz (IQR = 1.75), P = 3.54 × 10⁻⁶, HB).

The sharp waves concurrent with ripples reflect synchronized Schaffer collateral input from CA3 converging on the apical dendrites of CA1 neurons. We measured changes in the amplitude of sharp waves using the difference between the most negative and most positive deflections (typically in stratum radiatum and stratum oriens, respectively) recorded on our CA1 spanning electrodes. In POST we found increased amplitudes of sharp waves compared to MAZE in both NSD (NS1 mean = 5.36 mV (IQR = 0.65) vs. MAZE mean = 4.33 mV (IQR = 0.67), P < 2 × 10⁻⁴, HB) and SD groups (SD1 mean = 4.62 mV (IQR = 0.53) vs. MAZE mean = 4.00 mV (IQR = 0.67), P = 1.40 × 10⁻³, HB). These amplitudes were also larger than those observed during PRE for NSD (NS1 vs. NSD PRE mean = 4.63 mV (IQR = 0.71), P = 3 × 10⁻⁴ HB) though not for SD (SD1 vs. SD PRE mean = 4.45 mV (IQR = 0.57), P = 0.13 HB), including when accounting for state-dependence (Extended Data Fig 2K). Sharp-wave amplitudes did not change further over the course of NSD or SD (Fig. 1H) but recovery sleep elicited a strong increase in sharp-wave amplitudes (RS mean = 5.32 mV (IQR = 0.69) vs. SD2 mean = 4.59 mV (IQR = 0.51), P < 2 × 10⁻⁴, HB). The power of ripples (100–250 Hz, z-scored over each entire session Fig. 1I) during SWRs varied similarly to sharp-wave amplitude, suggesting that higher amplitude sharp-waves in the stratum radiatum produce stronger ripples in the pyramidal layer. These increased in POST relative to MAZE (NS1 mean = 5.88 mV (IQR = 0.51) vs. MAZE mean = 4.39 mV (IQR = 0.56), P < 10⁻⁴, HB, and SD1 mean = 4.64 (IQR = 0.50) vs. MAZE mean = 4.06 mV (IQR = 0.39), P = 1.8 × 10⁻³, HB), and relative to PRE for NSD (NS1 mean = 5.88 mV (IQR=0.50) vs. PRE mean = 5.11 mV (IQR = 0.60), P < 2 × 10⁻⁴ HB) but not for SD (SD1 mean = 4.64 mV (IQR = 0.50) vs. PRE mean = 4.72 mV (IQR = 0.59), P = 0.65, HB) (see also Extended Data Fig 2K). Ripple power (z-scored over the session) then decreased over the course of sleep (NS3 mean = 5.57 (IQR = 0.44) vs. NS1 mean = 5.88 (IQR = 0.50), P = 1.8 × 10⁻², HB), but increased over SD (SD2 mean = 4.97 IQR (0.50) vs. SD1 mean = 4.64 (IQR = 0.50), P < 2 × 10⁻⁴, HB), and even further upon RS (RS mean = 5.67 (IQR = 0.59), P < 2 × 10⁻⁴, HB). While these measures showed high variability across subjects, cross-group comparisons (NS2 mean vs. SD2 mean) were significant for ripple power (P = 0.040) and demonstrated a trend for sharp-wave amplitude (P = 0.090). Overall, these results demonstrate that while the total number of ripples remains elevated during SD, sleep loss manifests with higher frequency ripples but at lower power and with smaller sharp-waves, potentially reflecting the physiological impact of fatigue on the pyramidal cell-interneuron interactions that give rise to these events²⁸

Sleep loss disturbs firing-rate dynamics in the hippocampal network

The firing rates of neurons are sensitive to changes in sleep states²⁹, serve as important signals of the homeostatic function of sleep^25,30, and can reflect the strength of synaptic connectivity among neurons^17,30. We therefore assessed the effects of sleep and sleep loss on hippocampal firing rate dynamics (Fig. 2). During active exploration on the maze, firing rates tended to increase from PRE for pyramidal cells (NSD: MAZE mean = 1.15 Hz (IQR = 0.14) vs. PRE mean = 0.94 Hz (IQR = 0.14), P = 0.0028 HB; but not significantly for SD: MAZE mean = 1.11 Hz (IQR = 0.26) vs. PRE mean = 0.95 Hz (IQR = 0.21), P = 0.128, HB), and for interneurons (a trend for NSD: MAZE mean = 23.35 Hz (IQR = 4.30) vs. PRE mean = 20.34 Hz (IQR = 5.23), P = 0.055 HB; and significantly for SD: MAZE mean = 21.96 Hz (IQR = 3.88) vs. PRE mean = 18.72 Hz (IQR = 4.12), P = 0.044, HB). However, following MAZE, sleep loss produced different dynamics from natural sleep. Pyramidal cell firing rates (Fig. 2A–C) dropped significantly within hours of natural sleep (NS1 mean = 0.90 Hz (IQR =0.12) vs. MAZE, P = 0.028, HB) and further over the course of NSD (NS2 mean = 0.80 Hz (IQR = 0.09) vs. NS1, P = 2.8 × 10⁻³, HB), but they remained elevated throughout the 5 h SD period (SD2 mean = 1.04 Hz (IQR = 0.23) vs. SD1 mean = 1.01 Hz (IQR= 0.20), P = 0.55, HB). Differences were also evident in the distributions of pyramidal cell firing rates; these were skewed during PRE and MAZE, subsequently became log-normal in sleep^29,31 but remained skewed in SD (NS2 mean IQR = 0.63 log(Hz), P = 0.24, vs. SD2 mean IQR = 0.81 log(Hz), P = 2.7 × 10⁻³, hierarchical bootstrapping of Shapiro-Wilk test, Fig. 2D), with a broader distribution (Fig. 2E). These broad and negatively skewed distributions indicate that during prolonged awake SD a few cells fire at elevated rates while other reduce firing, suggestive of competition among neurons mediated by inhibition²⁹. Consistent with this, interneurons firing rates (Fig. 2C) decreased upon natural sleep and remained so throughout sleep (NS1 mean = 17.91 Hz (IQR = 3.37) vs. MAZE mean = 23.35 Hz (IQR = 4.30), P = 4 × 10⁻⁴, HB; NS3 mean = 16.86 Hz (IQR = 3.02) vs. NS1, P = 0.21, HB) but, in contrast, remained elevated from MAZE to SD (SD1 mean = 20.28 Hz (IQR = 5.32) vs. MAZE mean = 21.96 Hz (IQR = 3.88), P = 0.34, HB), and the remainder of SD (SD2 mean = 20.03 Hz (IQR = 5.01) vs. SD1, P = 0.66, HB). Upon recovery sleep, firing rates decreased rapidly for pyramidal cell (RS mean = 0.72 Hz (IQR = 0.19) vs. SD2 mean = 1.04 Hz (IQR = 0.22), P < 2 × 10⁻⁴, HB) and interneurons (RS mean = 13.02 Hz (IQR = 3.41) vs. SD2 mean = 20.03 (IQR = 5.01), P < 2 × 10⁻⁴, HB). Due to large variability, cross-group comparisons were less salient, but demonstrated a trend toward higher firing rates in pyramidal cells in the second block of SD compared to NSD (NS2 mean = 0.80 Hz (IQR = 0.09) vs. SD2 mean = 1.04 Hz (IQR = 0.22) P = 0.074). Interneurons likewise trended toward lower firing rates in natural sleep compared to recovery sleep (NS1 mean = 17.91 Hz (IQR = 3.37) vs. RS mean = 13.03 Hz (IQR = 3.47), P = 0.090, HB). While these patterns were largely attributable to state-dependent effects of waking and NREM (Extended Data Fig. 3), overall, the increased firing rates and skewed distributions in sleep deprivation compared to ad-libitum natural sleep indicate a higher metabolic impact of prolonged waking on hippocampal activities, which confirm and extend previous observations^25,29.

Figure 2: Hippocampal firing-rates are elevated and are more dispersed during sleep deprivation.

(A) Example sessions from non-sleep deprivation (NSD, top) and sleep deprivation (SD, bottom) with recovery sleep (RS), showing firing rates of pyramidal units (5 min bins, units sorted by mean) and hypnograms (top; active wake (AW), quiet wake (QW), rapid eye movement (REM), and non-REM (NREM) sleep) during POST. Mean firing rate (right axis) superimposed (white, this session; black, across all sessions). (B) Pyramidal neurons (PNs) during NSD (black, left; n = 442 cells from 8 sessions) and SD/RS (red/blue, right; n = 312 cells from 8 sessions) in (PRE, MAZE, ZT 0–2.5, ZT 2.5–5, and ZT 5–7.5) show decreasing firing during sleep but elevated firing during SD. Individual session means superimposed (connected dots). (C) Same as (B) but for interneurons (IN, n = 48 cells from 8 NSD sessions and n = 48 cells from 8 SD sessions). (D) The full distribution of PN firing rates deviates from log-normal during SD1 and SD2 but not NS1 or NS2. (E) Log firing rate interquartile range shows greater variance for PNs in SD vs. NSD. (F) PN firing rates specifically within ripples decreased over sleep and remained stable during SD but with minimal cross-group differences. (G) IN firing during ripples decreased over sleep, but remained elevated during SD then dropped in RS, with significant differences between NS1 and RS. All box plots depict median and top/bottom quartiles (whiskers = 1.5 × interquartile range) of the hierarchically bootstrapped (HB) data. Statistics: Panels B, C, E, F, G: two-sided within-group comparisons and one-sided cross-group comparisons of HB means; Panel D: Shapiro-Wilk tests performed on each HB log distribution, with P obtained from the proportion with significant skew; one-sided cross-group comparisons performed on the HB Shapiro-Wilk test statistics; ns (not significant), #P < 0.10, *P < 0.05, **P < 0.01, ***P < 0.001, with no corrections for multiple comparisons. See Supplementary Statistics Table for additional details.

Interneurons of different types display a variety of firing response during SWRs and play an important role in determining the physiological characteristics of the ripple oscillation. Therefore, we also examined the firing responses of interneurons, alongside those of pyramidal cells, specifically within SWRs (Fig 2F,G). Interestingly, while firing rates within ripples varied across the periods we examined, we generally saw little difference between natural sleep and sleep deprivation (PN: NS2 mean = 3.44 Hz (IQR = 1.03) vs. SD2 mean = 3.71 Hz (IQR = 1.24), P = 0.41, HB; IN: NS2 mean = 45.16 Hz (IQR = 7.36) vs. SD2 mean = 37.38 Hz (IQR = 6.61), P = 0.16 HB). However, we observed a significant decrease in the ripple firing rates of interneurons during recovery sleep compared to the similar period in natural sleep (RS mean = 34.11 Hz (IQR =7.11) vs. NS1 mean = 49.37 Hz (IQR = 7.85), P = 0.033, HB). Some studies indicate that somatostatin positive interneurons, a subset of which are lacunosum-moleculare projecting interneurons that gate entorhinal cortical input to CA1³², generally fire at lower rates during SWRs than do other cells³³. The lowered firing rates we observe during recovery sleep may therefore reflect the differential impact of sleep loss specifically on this class of interneurons, consistent with a recent study employing immediate early genes³⁴.

Sleep loss attenuates memory reactivation

Given that our results thus far demonstrate a high rate of SWRs during SD with robust concurrent firing in in pyramidal cells, we next asked whether the specific content of SWRs may be impacted by sleep deprivation. We first examined the reactivation of neuronal ensembles, which have been linked to the memory function of the hippocampus^2,18. Such reactivations can persist for hours after a novel experience³⁵ and can broadcast the hippocampal signal to cortical regions^2,8,9. To measure reactivation, we calculated the partial correlation explained variance (EV; see Methods), which measures the similarity of pairwise correlations between MAZE and POST while controlling for pre-existing correlations in PRE³⁶ in 250-ms bins in sliding 15-min windows (5 min steps; Fig. 3A). A time-reversed EV (REV) was used to estimate the chance level for reactivation³⁷. In naturally sleeping animals following exposure to the novel maze we observed hours long reactivation, consistent with our previous study³⁵. During SD, however, we observed one of two scenarios: either virtually no reactivation (e.g. rats N and U, Fig. 3A; seen in 4 out of 7 sessions, Extended Data Fig. 4) or reactivation similar to NSD but with a faster rate of decay (e.g. rats S and V, Fig. 3A; seen in 3 out of 7 sessions, Extended Data Fig. 4). These differences were not due to discrepancies in the amount of time nor the distance covered on the MAZE, nor in the proportion of active versus quiet wake states in the home cage (Extended Data Fig. 5A–C). However, we observed a significant negative correlation (r = −0.9, P = 0.006, Pearson correlation coefficient) between the rate of delta waves during SD2 (Extended Data Fig. 5D, but not those during PRE or SD1) and the amount of reactivation (EV) during SD1 but not other periods. Give that delta waves represent accumulated sleep pressure^21,38, this indicates that variations in sleepiness or resilience to sleep loss could potentially explain the differences in the capacity for hippocampal reactivation in sleep deprived animals³⁹.

Figure 3: Reactivation attenuates during sleep deprivation and fails to be restored by recovery sleep.

(A) Explained variance (EV) of pairwise reactivation (NSD, black; SD, red) and its chance levels (REV, maize) during POST in ad-lib sleep (NSD; left column) and sleep deprivation (SD) with recovery sleep (RS; right column) sessions from 4 animals (sex on y-axis; hypnogram on top (active wake (AW), quiet wake (QW), rapid eye movement (REM), and non-REM (NREM) sleep); additional sessions in Extended Data Fig 4). Solid line shows mean EV/REV, and shaded regions indicate low standard deviations. NSD sessions featured robust reactivation lasting for hours while SD sessions showed either some (rats S and V) or little reactivation (rats N and U). (B) Proportion of time spent in WAKE during NSD/SD (top left) and in NREM during NS1/RS (top left). Calculated exclusively during WAKE (bottom left), mean EV (mean/s.d. shown in solid line/shading) shows a similar decrease in both NSD (n = 20544 cell-pairs from 6 sessions) and SD (n = 8114 cell-pairs from 7 sessions but calculated exclusively during NREM, bottom right), there is lower reactivation in RS compared to NSD. (C) The decay constant obtained from exponential fits to EV curves separated by brain state (individual sessions overlaid with connected dots, except when out of range). In NSD, EV decays more slowly during NREM vs. WAKE. Interestingly, WAKE EV decays more slowly in SD compared to NSD, but trends towards faster decay than during NSD NREM. (D) EV plots indicate lower reactivation during SD vs. NSD, with a rebound during RS to lower levels than in ad-lib sleep. All box plots depict median and top/bottom quartiles (whiskers = 1.5 × interquartile range) of hierarchically bootstrapped data. Statistics: two-sided within-group comparisons and one-sided cross-group comparisons of bootstrapped means, #P < 0.1, *P < 0.05, **P < 0.01, ***P < 0.001, with no corrections for multiple comparisons. See Supplementary Statistics Table for additional details.

Overall, when we compared EV calculated exclusively during the awake state (Fig. 3B, left), we found similarly low levels of reactivation that decreased over time in both NSD and SD, in contrast to the higher reactivation during sleep exclusively in NSD (Fig. 3B, right). Across subjects, the time constant of decay, estimated from an exponential fit to EV, was significantly larger in NREM compared to waking NSD (Fig. 3C), and interestingly, were also larger in waking SD than waking NSD, suggestive of compensation for the lack of sleep in the former group. Importantly, reactivation levels were significantly lower during SD (Fig 3D) when comparing SD1 with NS1 as well as comparing SD2 with NS2. Thus, compared to sleep, the awake state demonstrates a more limited capacity for reactivation, and prolonged waking in particular provides a lower level of reactivation when compared to ad-libitum sleep during the same period. Remarkably, however, while reactivation was nearly absent by the end of SD, it increased significantly with the onset of recovery sleep (Fig. 3A, D). This suggests that the hippocampus is capable of reprising ensemble patterns reactivation even after a pause, such as during SD. But critically, even with this compensatory increase, reactivation levels during recovery sleep were substantially reduced compared to a corresponding period from NSD (Fig, 3B, D; see also comparisons for 1-h blocks in Extended Data Fig. 6). This reveals a persisting consequence of sleep deprivation, that unlike other effects of sleep loss, is not restored even after lost sleep is reclaimed.

Sequence replay deteriorates during sleep deprivation and recovery sleep

While pairwise measures, such as EV, measure neuronal reactivation, finer scale analysis has revealed that neuronal activity during SWRs can also provide a temporally compressed replay of sequences of place cells that fired during maze behavior^2,18. While most studies of replay have been directed at rest and sleep within an hour of maze exposure, we took advantage of our long duration recordings to investigate how replay (Fig. 4A) unfolds over several hours of sleep compared with sleep deprivation. As quantification of these events relies on different assumptions about the nature of replay^40,41, we focused on using Bayesian methods (Fig. 4 A–B) to simply quantify the proportion of ripple events that decode continuous movement through the maze (i.e. “trajectory replays”). Ripple events featuring ≥ 5 active units, animal’s movement speed < 8 cm/s, and peak ripple power > 1 s.d. were considered candidates for further analyses (see Methods). We assessed trajectory structure using the distance between decoded locations in adjacent time steps, referred to as “jump distance”⁴²; ripple events with jump distance < 40 cm in at least three consecutive time bins, were classified as trajectory replays.

Figure 4: Trajectory replays deteriorate over sleep deprivation and recovery sleep.

(A) Hippocampal spike raster (arbitrary colors ordered by place-field location) and raw (black) and ripple-filtered (blue) local field (LFP) during a sample run (normalized track position overlaid in orange). Gray box (right) displays a sample sleep replay. (B) Sample trajectory replays from ad-lib sleep (NSD) and sleep deprivation (SD) from top 10 percentile of distance covered and lowest 10 percentile of mean jump distance (blue inset, normalized distance) in each epoch. Replays were observed in all epochs but became progressively shorter, particularly in SD, with fewer events meeting the replay criteria. (C) Fewer events qualified as replays by the second block of SD (out of n = 72584 candidate events from 7 sessions) and in recovery sleep (RS) compared to NSD (out of n = 64205 candidate events from 6 sessions). Critically, the proportion of replays in RS was significantly lower than in the equivalent period from ad-lib sleep (NS1). (D) Similar to (C) but for the rate of replays. Fewer replays were seen in the first block of SD compared to NSD and crucially, there were fewer replays in RS vs. NS1. (E) The durations of trajectory replays were significantly reduced from the first to second block of SD (n = 15005 replays from 7 sessions) but not NSD (n = 17742 replays from 6 sessions), with a further decrease upon RS. All box plots depict median and top/bottom quartiles (whiskers = 1.5 × interquartile range) of the hierarchically bootstrapped data with individual session means overlaid and connected. Statistics: two-sided within-group comparisons and one-sided cross-group comparisons of bootstrapped means, ns (not significant), #P < 0.1, *P < 0.05, **P < 0.01, ***P < 0.001, with no corrections for multiple comparisons. See Supplementary Statistics Table for additional details.

We observed that the proportion of ripples that qualified as trajectory replays was highest on the maze in both experimental groups, and was also higher in ad-lib sleep in NS1 compared to PRE (Extended Data Fig. 7), consistent with previous reports^43,44. However, the proportion of trajectory replays significantly decreased over the course of SD (SD1 mean = 0.21 (IQR = 0.034) vs. SD2 mean = 0.018 (IQR = 0.028), P = 0.024 HB) and was significantly lower from natural ad-libitum sleep by the second block of SD (NS2 mean ± SEM = 0.26 (IQR = 0.017) vs. SD2 mean = 0.018 (IQR = 0.18), P = 4.02 × 10⁻⁴). Importantly, even during recovery sleep, replays decreased further and did not rebound to the comparative levels in NSD (Fig. 4C; RS mean = 0.016 (IQR = 0.023) vs. NS1 mean = 0.027 (IQR = 0.034), P = 1.98 × 10⁻⁴ HB). The total rate of trajectory replays (Fig. 4D) was also lower in the first block of SD compared to NSD (SD1 mean = 590 (IQR = 84) vs. NS1 mean = 840 (IQR = 128), P =0.014 HB) and remained significantly lower in recovery sleep (RS mean = 290 (IQR = 64) vs. NS1 mean = 840 (IQR = 128), P < 10⁻⁴ HB). While the patterns in the effects of SD/RS vs. NSD on trajectory replays somewhat differ from those for reactivation, some discrepancies are expected due to the methodological differences in the measures used for these patterns⁴¹. Additional differences could also arise if pairwise co-activations during sleep reflect the maze experience without integrating into neural sequences that correspond to paths with momentum through the maze environment^45–47.

Finally, motivated by a recent study that reported a memory benefit for longer replays⁷, we measured and compared the durations of trajectory replays. Though we did not detect significant cross-group differences in replay durations, within SD (but not NSD) we observed a significant decrease from the first to the second block (SD1 mean 0.203 s (IQR = 0.0097 s) vs. SD2 mean = 0.186 s (IQR = 0.013 s), P < 2 × 10⁻⁴ HB) and a further decrease in recovery sleep (SD2 mean = 0.186 s (IQR = 0.013 s) vs. RS mean = 0.172 s (IQR = 0.014 s), P = 0.012 HB). Overall, these results demonstrate that the loss of sleep immediately following novel experience diminishes the hippocampal replay of place cell patterns following novel maze exposure and that this impairment persists even when sleep is regained.

Discussion:

During sleep deprivation compared to natural sleep, we observed lower amplitude sharp waves coupled with lower power ripples and higher frequency ripple oscillations at the CA1 pyramidal layer. Higher amplitude and power generally indicate greater synchrony of CA3 inputs to CA1 neurons, leading to greater spiking in CA1 neurons^2,48, and stronger resonance throughout the hippocampal formation⁸, although the animal’s sleep/awake was not separated in these studies. Nevertheless, one recent study reported that SWRs during waking, for which we observe lower power ripples, have a larger impact on prefrontal cortical neurons⁴⁹ than during sleep. Similarly, lower amplitude sharp-waves produced larger neuronal responses in extra-hippocampal regions⁸. These observations suggest that larger sharp-waves do not necessarily translate to greater activation in target regions. Additionally, hippocampal firing rates during SWRs remained comparable between sleep-deprived and sleeping animals despite differences in SWR features, indicating that SWRs of different power and amplitude generate similar responses in hippocampal neurons.

The increase in ripple power over SD but decreasing power over sleep, with parallel changes in ripple frequency, suggest that these SWR features can serve as indicators of sleep pressure. These indicators are measurable from the hippocampal LFP in both waking and sleep, which contrasts with cortical slow-wave activity in common models of sleep homeostasis²⁴, which can only be measured during sleep (see also ref(²³)). Higher frequency ripples potentially reflect the higher metabolism of the awake state⁵⁰ which is progressively lowered and reset in sleep²⁵. However, differences in ripple frequency can also reflect differences in neuromodulatory tone, such as activation of GABA-A⁵¹ or 5-HT receptors⁵², or different routing of inputs to CA1, with higher frequency ripples reflecting the influence of CA2 during waking⁵³, and lower frequency ripples reflecting input from the entorhinal cortex^2,54. Consistent with this notion, we noted an increase in ripple frequency, coupled with higher power and higher amplitude sharp-waves, following the novel maze exposure, particularly in the awake state (Extended Data Fig 2K; see also ref⁵⁵). Interestingly, lower frequency ripples have also been associated with aging⁵⁶, whereas ripple frequency increases after learning⁵⁷, consistent with the postulated correlation with higher metabolic cost.

In addition to differences in physiological features of SWRs, we reported firing rate patterns that appear generally consistent with “the synaptic homeostasis hypothesis”^25,50 which conjectures that waking drives strengthened connectivity between neurons, while sleep drives synaptic downscaling. The decrease in reactivation and replay over the course sleep may likewise be consistent with this hypothesis as the pathways providing reverberation of waking patterns are progressively reduced. On the other hand, the more rapid decline in replay and reactivation during SD versus sleep is not readily reconciled with a preferential role for waking in synaptic strengthening. If synaptic strengthening indeed occurs preferentially during the awake state, then it could be expected to elicit more robust reactivation than during sleep. Another possibility, however, is that the strengthening during awake activity is promiscuous rather than specific to the firing patterns evidenced on the maze. In this scenario, waking during sleep deprivation may actively interfere with hippocampal reactivation by provoking the hippocampus to generate and learn new patterns inconsistent with the maze experience. Similarly, whereas it has been conjectured that SWRs may serve to downscale synapses^17,25,50, reactivation and replay were longer lasting during sleep, even though SD featured a higher incidence rate of SWRs, potentially indicating a homeostatic drive⁵⁸. The background brain states against which SWRs occur, along with the specific hippocampal firing patterns that they produce, likely play an important role in determining their effects on the hippocampal circuit and other brain regions^6,8,17,49.

Most significantly, in this study we found reactivation during natural sleep that lasted for several hours, consistent with our previous report³⁵, but diminished reactivation during SD with only limited rebound when animals eventually regained lost sleep. The absence of a more complete rebound was remarkable because while most indices of brain health and function, including protein signaling⁵⁹ and gene transcription⁶⁰, return to normal levels following sufficient recovery sleep, memories compromised by sleep loss typically do not recover^3,59,60. Overall, our work calls attention to reactivation and replay, rather than simply the occurrence of SWRs, as potentially the crucial elements that mediate the role of sleep in memory and the negative impact of sleep loss. The disruption of these neuronal firing patterns could destabilize hippocampal spatial representations^19,47 and hippocampus-dependent spatial memories⁶. Furthermore, since SWRs provide privileged windows of communication between the hippocampus and other brain regions¹¹, the compromised nature of this exchange is likely to have repercussions on networks distributed throughout the brain^8,9.

Methods:

Animals and surgical procedures

Four male and three female Long-Evans rats (300–500 grams) were used in this study. All surgeries were performed on isoflurane anesthetized animals head fixed on a stereotaxic frame⁶¹. After removing hair from the head, the incision area was cleaned using alcohol and betadine. Next, an incision was made to expose the skull underneath. The skull was cleaned of tissues and blood, after which hydrogen peroxide was applied. Coordinates for probe implantation were marked above the dorsal hippocampus (AP: −3.36, ML: ±2.2) following measurement of bregma and lambda. Craniotomies were drilled at the marked location. Using a blunt needle, the dura was removed carefully to expose the brain surface. After cessation of bleeding, animals were implanted with 64 channel (8 shank “Buzsaki” probe; Neuronexus, MI; 1 animals) or 128 channel (8 shanks, Diagnostic Biochips, MD, 6 animals) silicon probes. Ground and reference screws were placed over the cerebellum. Craniotomy was covered with DOWSIL silicone gel (3–4680, Dow Corning, Midland, MI) and wax. A copper mesh was built around the implant for protection and electrical shielding. All procedures involving animals were approved by the Animal Care and Use Committee at the University of Michigan.

Behavior

Prior to the probe implant surgery animals were habituated to the experimenter for ≥ 40 mins for 5 days. Following habituation animals were water restricted and trained to associate water rewards with plastic wells. During the post-implant recovery period (7 days) animals were brought to the recording room for monitoring electrophysiology signals and probes were slowly lowered to the dorsal CA1 region of the hippocampus. In addition, animals were also habituated to sleep box for >1 h every day. Following this, animals were placed on a water restriction regiment for 24 h before experiments commenced. Each experimental session began by transferring animals to their sleep box ~4 h before the onset of light cycle. After 3 h of recording in the home cage, animals were transferred to a novel maze that they had not previously explored. These maze tracks were made distinct by the shape, color, and construction materials. Animals alternated for ~ 1 h between two water wells fixed at either ends of the maze to retrieve rewards from water wells. Following exploration, animals were transferred to the home cage and the recording continued for ≥ 10 h. Animals had access to ad libitum food and received ad libitum water for 30 mins per day.

Sleep deprivation protocol

Sleep deprivation was performed at the onset of the light cycle in the home cage using a standard ‘gentle handling’ procedure^62,63. Animals were extensively habituated to the experimenter conducting the sleep deprivation. During the initial hours of sleep deprivation, animals were kept awake by mild noises, tapping or gentle shaking of the cage when animals displayed signs of sleepiness. As sleep pressure built up over 5h sleep deprivation period, other techniques such as gently stroking the animal’s body with soft brush or disturbing bedding were increasingly employed to to ensure that animals stayed awake. Following sleep deprivation, animals were allowed to sleep and recover for 48 h before any further experiments.

Data Acquisition

Electrophysiology data was acquired using OpenEphys⁶⁴ or an Intan RHD recording controller sampled at 30 kHz. Analysis of local field potentials (LFP) was performed on signals downsampled to 1250 Hz. The animal’s position on the maze track was obtained using Optritrack (NaturalPoint, Inc, OR) hardware and Motive software (v2.0, https://optitrack.com/software/motive/), which uses infrared cameras to locate a 3d markers that were clipped to the animal’s crown. Position data was sampled at either 60 Hz or 120 Hz and later interpolated for aligning with electrophysiology. Water rewards during alternation on the maze track were delivered via solenoids interfaced with custom built hardware using Arduino. The timestamps for water delivery were recorded via TTLs.

Spike sorting, cell classification, and stability criteria

Raw data went through filtering, thresholding and automatic spike-sorting using SpyKING CIRCUS (v0.8.8-v1.1.0, https://github.com/spyking-circus/spyking-circus)⁶⁵, followed by manual inspection and reclustering using the Phy package (v2.0, https://github.com/cortex-lab/phy/). Only well isolated units were used for further analysis with the exception of decoding/sequence detection analysis where we used all clusters that satisfied the stability criteria.

LFP and unit analyses were performed using custom codes (NeuroPy) written in Python and are available in our lab’s GitHub repository (v0.1, https://github.com/diba-lab/NeuroPy) which utilizes the packages Numpy (v1.24.4, https://numpy.org), Scipy (v1.11.3, https://scipy.org), pingouin (v0.5.3, https://pingouin-stats.org) for data analysis and matplotlib (v3.8.1, https://matplotlib.org) and Seaborn (v0.11.2, https://seaborn.pydata.org) for visualization. Units were sorted into putative pyramidal cells and interneurons based on peak waveform shape, firing rate, and interspike-interval^66,67.

Stability criteria:

to ensure that a given neuron was reliably tracked across the recording duration, we divided each session into 5 equally sized bins (~2.5 h) and excluded any unit that fired below 25% of its overall mean in any given time bin.

Sharp wave ripple detection and related properties

For detecting ripples, one channel from each shank were selected based on the (highest) mean power in the ripple frequency band (125–250 Hz). The Hilbert amplitude was averaged across all selected channels, then smoothed using a Gaussian kernel ($\sigma =12.5ms$) and z-scored. Putative ripple epochs were identified from timepoints exceeding 2.5 standard deviations (s.d.) and the start/stop was associated with signals > 0.5 s.d..; Candidate ripples < 50 ms or > 450 ms were excluded from further analyses. The maximum z-score value within a ripple epoch was termed as its ripple power. Sharp wave amplitudes were obtained from a bandpass (2–30 Hz) filtered LFP using the difference between maximum and minimum value across all recorded channels within a given ripple. The peak frequency of each ripple was estimated using a complex wavelet transform. The LFP was first high-pass filtered > 100 Hz. This filtered signal was then convolved with complex Morlet wavelets with central frequencies selected from linearly spaced frequencies in the ripple frequency band (100 to 250 Hz). Within each ripple, the frequency with maximum absolute wavelet power was designated as the peak ripple frequency.

Sleep scoring

Sleep scoring was performed using correlation electromyogram (EMG), theta, and delta power. Correlation EMG was estimated by summing pairwise correlations across all channels calculated in 10 s time windows with a 1 s step^68,69. For theta power, a recording channel with the highest mean power in the 5–10 Hz theta frequency band was identified. Following theta channel selection, the power spectral density was calculated for each window. Periods with low and high EMG power were labeled as sleep and wake, respectively. The theta (5–10 Hz) over delta (1– 4 Hz) plus (10 –14 Hz) band ratio of the power spectral density was used to detect transitions between high theta and low theta, using custom python software based on hidden Markov models followed by visual inspection. Sleep states with high theta were classified as rapid eye movement (REM) and the remainder were classified as non-REM (NREM). Wake periods with high theta were labeled as “active” and the remaining were labeled “quiet”. These labels were merged in WAKE for the main figures. All detected states went through additional visual inspection to correct any misclassifications. Detailed, interactive sleep scoring plots for each session are available at: https://github.com/diba-lab/sleep_loss_hippocampal_replay.

Detection of delta-waves, delta power, and OFF states

To detect hippocampal delta waves⁷⁰, hippocampal LFP was filtered (0.5–4 Hz) and the resulting filtered signal was z-scored. The first order derivative of this signal was used to identify upward-downward-upward zero crossings, which corresponded to the beginning, peak, and end of the delta wave, respectively. Delta waves lasting < 150 ms or > 500 ms were discarded. In addition, we required the amplitude at peak to be either > 2 s.d., or the amplitude at peak > 1 s.d. and amplitude at end < −1.5 s.d..

Delta power spectral density was calculated by extracting LFP signal from a channel localized in the CA1 pyramidal cell layer with Welch’s method using 4 second bins.

The MUA smoothed with a Gaussian kernel of $\sigma$ = 20 ms was used to detected OFF periods, following a method adapted from Vyazovskiy et al.²¹ . Candidate OFF periods were identified when the MUA firing rate dropped below the session median. The surrounding timepoints when the firing rate reached the lowest 10 percentile were used to mark the onset, offsets, and corresponding duration of these events.

Explained variance measure for reactivation

Reactivation was assessed by the “explained variance” (EV) measure following previously described methods^35,36. This EV describes how much of the co-activity in a pair of neurons for a given window in POST is explained by the co-activity of those neurons during MAZE, while controlling for co-activity that was present during similar windows in PRE. Briefly, spike times were binned into 250 ms time bins, creating an N by T matrix, where N is the number of neurons and T is the number of time bins. Pearson’s correlations, R, were determined for spike counts from neuronal pairs in 15 min sliding windows (window length 15 min, sliding 5 min steps) to produce P, an M-dimensional vector, where M is the number of cell pairs. To reduce spurious correlations arising from cross contamination of units from the same shank⁷¹, only pairs with waveform similarity < 0.8 were used. Next, to assess similarity between P vectors from different windows, the Pearson correlation R of these vectors (i.e., the correlation between cell pair correlations) was determined (e.g., R_{[PRE, POST]}, R_{[PRE, MAZE]} and R_{[MAZE, POST]}). Controlling for preexisting correlations in a given sliding window (k) in PRE, the explained variance for a 15 min window (WIN) was calculated as:

$$\text{EV}\left(\text{WIN}\right)={\left(\frac{R_{\left[\text{MAZE},\text{WIN}\right]}-R_{\left[\text{MAZE},\text{PRE}\left(k\right)\right]}\times R_{\left[\text{PRE}\left(k\right),\text{WIN}\right]}}{\sqrt{1-R_{\left[\text{MAZE},\text{PRE}\left(k\right)\right]}^2}\sqrt{1-R_{\left[\text{PRE}\left(k\right),\text{WIN}\right]}^2}}\right)}^2$$

averaged over all windows in PRE. To get an estimate of the chance level for EV, we calculated a time-reversed explained variance (REV) for each WIN^37,72:

$$\text{REV}\left(\text{WIN}\right)={\left(\frac{R_{[\text{MAZE},\text{PRE}(k\left)\right]}-R_{\left[\text{MAZE},\text{WIN}\right]}\times R_{\left[\text{PRE}\right(k),\text{WIN}]}}{\sqrt{1-R_{[\text{MAZE},\text{PRE}(k\left)\right]}^2}\sqrt{1-R_{\left[\text{PRE}\right(k),\text{WIN}]}^2}}\right)}^2$$

similarly averaged over PRE.

To estimate the time constant of reactivation from each session³⁵, we fit the time course of bootstraps and sessions EV curves to an exponential function:

$$\text{EV}\left(t\right)=a{e}^{-t/\tau }$$

where $\tau$ provides the exponential decay constant.

Only sessions with > 15 stable units were used in the reactivation and replay analyses (13 out of 16 recorded sessions from 6 out of 7 animals).

Place field calculations

To calculate 1D place fields, animals’ 2D positions were linearized using ISOMAP⁷³ and visually inspected to ensure accuracy. For each unit, two firing rate maps were generated corresponding to each running direction. Occupancy within 2 cm spatial bins at timepoints when the animal’s speed exceeded 8 cm/s were calculated and smoothed with a Gaussian kernel ($\sigma$ = 4 cm). For each neuron, spike counts within each spatial bin were determined and also smoothed with the Gaussian kernel ($\sigma$ = 4 cm). Then, each neuron’s firing rate map was generated by dividing the smoothed spike counts by the smoothed occupancy map. Neurons with peak firing rate < 0.5 Hz were excluded from further analysis.

Decoding and trajectory replays

Multiunit activity (MUA) was used to detect population burst events that are concurrent with sharp-wave ripples. Within a session, the firing rate of MUA was derived from all putative spikes combined from all clusters then binned in 1 ms time bin and smoothed using a Gaussian kernel of $\sigma$ = 20 ms. Periods with peak MUA > 3 s.d. above the mean firing rate were considered candidate ripple events. The start and end times of these ripple events were defined by the first neighboring time points at which the MUA exceeded the mean. Ripple events occurring within 10 ms of each other were merged. Ripple events with duration < 80 ms or > 500 ms were discarded.

Before decoding, candidate ripple events were required to satisfy 1) ≥ 5 active units, 2) movement speed < 8 cm/s, and 3) concurrent peak ripple power > 1 s.d.. For these analyses alone, to minimize decoding error, we included all stable clusters with a mean firing rate < 10 Hz regardless of their isolation quality⁷⁴. Position decoding was carried out on ripple events using standard Bayesian decoding⁷⁵ methods. Probabilities of the animal occupying each position bin $x_P$ on the track were calculated according to:

$$P\left(x_p|n_t\right)=K_t\left\{\left.\prod _{i=1}^N{\lambda }_i{\left[x_p\right]}^{n_{i,t}}\right\}\right.e^{-\tau \sum _{i=1}^N{\lambda }_i\left[x_p\right]}$$

where $\tau$ is the duration of the time bin (20 ms) used, ${\lambda }_i\left[x_p\right]$ is the firing rate of the i-th neuron at $x_P$ on the maze, $K_t$ is a normalization constant to ensure that the sum of probabilities across all position bins equals to 1, and $n_t$ is the number of spikes fired by each neuron in that bin. The location with the maximum posterior probability was considered the `decoded location’ for each time bin. A candidate ripple event was classified as a ‘trajectory replay’ if it decoded a continuous trajectory across space for ≥ 60ms such that the distance between decoded locations in adjacent time bins was < 40cm. Posterior probability matrices for all ripple events that were classified as replay have been compiled in an interactive plot available in our GitHub repository (https://github.com/diba-lab/sleep_loss_hippocampal_replay).

Hierarchical bootstrapping

We employed hierarchical bootstrapping⁷⁶ to estimate confidence intervals and P-values for different variables following code found at https://github.com/soberlab/Hierarchical-Bootstrap-Paper. For each metric we generated a population of 10000 values by resampling with replacement at each level of the data hierarchy (first sessions, then for each session, the variable measured, e.g. frequency of the SWR, etc.) and pooled the values to calculate the test statistic and corresponding confidence interval or interquartile range. Two-tailed tests with α = 0.05 were used for within-group comparisons, where the P-value was determined from the proportion of bootstraps for which the test statistic in one group exceeded that of the other group. The within-group comparisons generated one-sided hypotheses for cross-group testing, performed at α = 0.05. For these cross-group comparisons, we used the joint probability distributions of the bootstrapped samples to determine the P-value: the likelihood that the mean of group one is $\ge$ the mean of group two. All hierarchical bootstrapped data was visualized using box and whisker plots generated using the boxplot function from the matplotlib (version 3.8.1, https://matplotlib.org/) and Seaborn (version 0.11.2, https://seaborn.pydata.org/) Python packages to depict the median and 1^st/3^rd quartiles, with whiskers extending to 1.5 × the interquartile range. For testing if firing rate distributions differed from log-normal, Shapiro-Wilk tests were performed on each bootstrapped log distribution and the P-value was determined from the proportion of bootstraps with significant skew at α = 0.05. Detailed statistics with estimated P-values for all performed comparisons are found in the Supplementary Statistics Table.

Parametric statistics

For instances in which parametric tests were more appropriate, the exact P-values, test-statistics, confidence intervals, and degrees of freedom are provided in Supplementary Statistics Table.

Extended Data

Extended Data Figure 1: Power spectra and delta for all recorded sessions.

Power spectral density of the CA1 local field potential (LFP), z-scored over 1–10 Hz for the time periods shown, with temporal evolution of delta (white) overlaid for each recorded session (similar to in Fig 1B). Hypnograms above each panel show the brain state (active wake (AW), quiet wake (QW), rapid-eye movement (REM) sleep, and non-REM sleep (NREM). State scoring was performed at 1-s resolution but for illustration purposes is provided averaged for 30-s periods (particularly due to rapid transitions between AW and QW during SD). Animal name initial, sex, and recording day are provided the left of the y-axes.

Extended Data Figure 2: Ripple and delta features and controls across sleep and sleep deprivation sessions.

(A) Local field potential spectrogram (1–10 Hz) from a sample theta channel during recovery sleep (RS) from three rats with corresponding hypnogram indicating the scored sleep/wake state above (active wake (AW), quiet wake (QW), rapid eye movement (REM), and non-REM (NREM) sleep). The Fourier spectrogram was calculated from the whitened LFP traces using 4 s windows with 1 s overlap. Z-scored delta power (1–4 Hz, smoothed with a 12 s gaussian kernel) is overlaid in white. More detailed sleep scored sessions are available at https://github.com/diba-lab/sleep_loss_hippocampal_replay. (B) The proportion of time spent in each brain state across all sessions. Individual session values overlaid in connected dots. We note that during sleep deprivation from ZT 0–2.5 (SD1) to ZT 2.5–5 (SD2), there was no significant change in the proportion of time in QW (P = 0.958, t(df = 7) =− 0.054) or AW (P = 0.769, t(df = 7) = 0.305). (C) The rate of OFF states compared across sessions. For the non sleep-deprived (NSD) group, OFF states were most prevalent during NS1 (ZT 0–2.5) and decreased over time, in NS2 (ZT 2.5–5) and NS3 (ZT 5–7.5). The rate of OFF states was initially lower in the SD group, but increased from SD1 to SD2, with a further large increase upon RS. (D) The rate of ripple events calculated in 5 min windows decreased over the first 5 h of NSD but remained stable during 5 h of SD. (E) Ripple rate calculated separately for NREM and WAKE states (individual sessions overlaid with connected dots). A decrease in ripple rates is observed in both NREM and WAKE in the NSD group, but there was no change in WAKE ripples from SD1 to SD2, and a decrease from SD2 to RS. Overall, NREM ripple rates were higher in NS1 vs. RS and WAKE ripple rates were higher in SD2 vs. NS2. (F) The ripple probability (solid line = mean, shaded region = s.e.m., n = 8) was modulated by delta waves. (G) However, the modulation depth of ripples by delta ((peak-trough/mean) was not significantly different across 2.5 h blocks. (H) OFF states were frequently preceded and followed by ripples⁶⁹. Modulation of OFF states by ripples did not change across NSD but the probability that OFF immediately followed a ripple increased over SD, from SD1 to SD2 and further in RS, with a significant difference between RS and NS1. The inducement of OFF states by ripples is similar to the rise in OFF states following bursts induced by sensory stimulation in the cortex⁷⁷. (I) Interventions needed to stop transitions to sleep during SD were tracked using piezo sensors on the sides of the home cage in 3 sessions. The number of interventions grew with time during SD. (J) Mean and 95% confidence intervals of ripple rate (left) and delta wave rate (right) relative to the onset of interventions. The rate of delta waves and concurrent ripples was higher immediately preceding interventions, consistent with signs of sleepiness that compel such interventions. (K) Ripple features (frequency, sharp wave amplitude, and ripple power) evaluated separately in NREM (n = 67007 ripples from 6 NSD sessions, n = 26798 ripples from 7 SD sessions) and WAKE states (n = 74363 ripples from 6 NSD sessions and 128957 ripples from 7 SD sessions). Rightmost panels in each row provide cross-group comparisons in NS1 vs. RS strictly during NREM and NS2 vs. SD2 strictly during WAKE. These results are largely consistent with patterns in Fig. 1G–I, except that here ripple power in NS2 vs. SD2 is not significantly different during WAKE, indicating state-dependence of this effect. Additionally, we note a significant increase in ripple frequency in WAKE from PRE to POST in both NSD and SD groups, indicating an effect of the novel maze exposure. All box plots show the median and top/bottom quartiles (whiskers = 1.5 × interquartile range) of the hierarchically bootstrapped data with individual session means overlaid with connecting dots. Statistics: panels C, E, G, two-sided paired t-tests (within group) and one-sided independent groups (across groups) t-tests; panel D, Pearson correlation coefficients with two-sided p-value; panel H, ${\chi }^2$ tests of independence; panel K, two-sided paired within group and one-sided cross-group comparisons with hierarchical bootstrapping; ns (not significant), *P < 0.05, **P < 0.01, ***P < 0.001, with no correction for multiple comparisons. See Supplementary Statistics Table for additional details.

Extended Data Figure 3: Firing rate changes within each state separately.

Mean firing rates calculated solely within the awake (WAKE) state (A) or solely within NREM (B) with individual sessions overlaid and connected. Differences calculated separately within wake or NREM were less pronounced than those shown in Fig. 2 B, C, consistent with the noted effect of background state on hippocampal firing rates^25,29. However, when estimating the metabolic cost of neuronal firing²³, comparisons that overlook the state and consider temporal variations in rates, such as those depicted in Figures 2B and C, are most appropriate. In WAKE (A), firing rates showed a trend towards decreased rates in pyramidal cells (top row) in the NSD group (n = 442 neurons from 8 sessions) but not in SD (n = 312 neurons from 8 sessions). The decrease in firing rates during brief wakings with the recovery sleep period (right panel) likewise showed a trend towards significance vs. a similar period in NSD. Interneuron firing rates (bottom row) within WAKE in recovery sleep showed a trend towards significance in comparison to the similar period in NSD (n = 48 cells from 8 NSD sessions and n = 48 cells from 8 SD sessions). In NREM (B) no significant differences were detected across groups or periods. (C) and (D) Same as (A) and (B) but for active wake (AW) and quiet wake (QW). (E) Firing rate distribution for all pyramidal cells recorded during SD sessions for AW vs. QW. Firing rates in both WAKE states remain skewed from log-normal distribution throughout SD. (F) Interquartile range (IQR) of the log firing rate of pyramidal cells reveals a trend toward a broader range of firing rates in AW vs. QW during SD. All box plots depict the median and top/bottom quartiles (whiskers = 1.5 × interquartile range) of the hierarchically bootstrapped data with individual session means overlaid with connecting dots. Statistics: A-D, F: two-sided paired within group and one-sided cross-group comparisons with hierarchical bootstrapping; E: Shapiro-Wilk tests performed on each bootstrapped log distribution, with P obtained from the proportion of bootstraps with significant skew; ns (not significant), #P < 0.1, *P < 0.05, **P < 0.01, ***P < 0.001, with no correction for multiple comparisons. See Supplementary Statistics Table for additional details.

Extended Data Figure 4: Temporal evolution of reactivation across recorded sessions.

Reactivation assessed using the explained variance (EV) metric (NSD (black), SD (red), and RS (blue)), in thirteen sessions from six different animals (3 male and 3 female, with 3 sessions from 2 animals (1 male, 1 female) excluded due to an insufficient number of stable neurons), as in Fig. 3A. Chance level (REV) is shown in maize. Solid lines show the mean and shaded regions show the standard deviation of EV/REV across all 15 min windows in POST. Each row provides session(s) from one animal, with number of putative pyramidal neurons and cell pairs used to calculate EV specified inside each panel. Hypnograms above panels depict sleep/wake history in active wake (AW), quiet wake (QW), rapid eye movement (REM) sleep and non-REM (NREM) sleep, with sleep deprivation/recovery sleep in red/blue and natural sleep in black. Animals’ tracked positions on the novel maze (purple) are depicted on the right of the panels along with the session recording day. See Supplementary Statistics Table for additional details.

Extended Data Figure 5: Accounting for the variability in reactivation during sleep deprivation.

We observed striking variability in reactivation across animals during the first block of sleep deprivation (SD1) in ZT0–2.5 (Fig. 3 and Extended Data Fig. 4). We conducted a series of analyses in an effort to account for this observation. Differences in (A) the distance run or (B) the total time spent running on the maze, did not account for the variance in EV during SD1. (C) Likewise, the variance in EV during SD1 cannot be attributed to differences in the proportion of time in active wake (left) or quiet wake (right) states during this period. (D) We next tested whether the rate of delta waves during sleep deprivation (top row), an indicator of sleep pressure, could explain the variance in EV during SD1. Remarkably, there was a strong significant negative correlation (P = 0.006) between the rate of delta from ZT 2.5–5 (SD2) and the reactivation (EV) during SD1. If delta during SD2 thus relates to animal’s level sleepiness, consistent with the sleep homeostasis model^24,38, the level of sleepiness correlates with the amount of hippocampal reactivation we observe during SD1. In contrast, we observed no correlation between EV and delta at any timepoint for NSD (bottom row) (E) A similar relationship was not evident between delta waves and EV in NS2. (F) Reactivation (EV) during SD1 was not predictive of the reactivation during RS. Statistics: All panels, Pearson correlation coefficients with two-sided P-values, **P < 0.01, with no correction for multiple comparisons. See Supplementary Statistics Table for additional details.

Extended Data Figure 6: Comparisons across 1-hour blocks.

Changes in ripple properties, firing rates, explained variance, and replays were assessed using 1-h blocks, based on the last hour of PRE, 1-h periods immediately after MAZE (ZT 0–1) and 1-h blocks immediately before and after recovery sleep (ZT 4–5 and ZT 5–6). All box plots depict the median and top/bottom quartiles (whiskers = 1.5 × interquartile range) of the hierarchically bootstrapped data with individual session means overlaid with connecting dots. Similar to our results for 2.5 h blocks in the main text, (A) ripple frequency (left) decreased over NSD (n = 143681 ripples total from 8 sessions) but increased in SD (n = 157964 ripples total from 8 sessions) relative to MAZE, with a rebound drop in RS (ZT 5–6). Rightmost panel highlights cross-group comparisons for the first block of sleep (NS1 vs. RS) and second block of SD vs. NSD. In both groups, sharp-wave amplitudes (middle) and ripple power (right) increased from MAZE to the first block of POST (ZT 0–1). Sharp-wave amplitude (middle) and ripple power (right) further increased in RS. Cross-group comparisons at ZT 4–5 showed increased ripple power in NSD compared to SD. (B) Firing rate of pyramidal neurons show decreasing firing rates during sleep but not during SD (n = 442 pyramidal neurons / 48 interneurons from 8 sessions NSD, 312 pyramidal neurons / 48 interneurons from 8 sessions SD). (C) EV was significantly lower in SD at ZT4–5 compared to NSD, with a modest but significant rebound during RS, but to lower levels than during the first hour of natural sleep. n = 20544 cell-pairs from 6 NSD sessions and n = 8114 cell-pairs from 7 SD sessions. (D) (left). The proportion of candidate ripple events that decoded continuous trajectories in different epochs (n = 65744 candidate events from 7 SD sessions and n = 56669 candidate events from 6 NSD sessions). SD sessions featured significantly fewer trajectory replays by ZT4–5. Critically, the proportion of replays in RS was significantly lower than in NS1. Similar results were observed for replay number (middle). A significant decrease was observed in mean replay event duration (right) for SD (n = 13911 replays from 7 sessions) from ZT0–1 to ZT4–5. Statistics: two-sided within-group comparisons and one-sided cross-group comparisons with hierarchical bootstrap, #P < 0.1, *P < 0.05, **P < 0.01, ***P < 0.001, with no correction for multiple comparisons. See Supplementary Statistics Table for additional details.

Extended Data Figure 7: Replay characterization during NREM and WAKE.

(A) Replays showed no bias in directionality. (B) The total number of candidate events decreased during POST in non sleep-deprivation (NSD, n = 64205 candidate events from 6 sessions) but remained elevated during sleep deprivation (SD, n = 72584 candidate events from 7 sessions) from the first to second block (SD1 to SD2), but dropping from SD2 to recovery sleep (RS). (C) The proportion of candidate events that scored as trajectory replays in NSD and SD groups, measured separately in WAKE (n = 30852 events from 6 NSD sessions and n = 59820 events from 7 SD sessions) and NREM (n = 32258 events from 6 NSD sessions and 11903 events from 7 SD sessions) states in each block. The rightmost panel provides comparisons between the first block of extended NREM sleep for each group (ZT 0–2.5 in the NSD group vs. ZT 5–7.5 in the SD group) and between WAKE during the second (late) block of POST (ZT 2.5–5 for both groups). There was a significantly lower proportion of trajectory replays in NREM recovery sleep (RS) compared to nature sleep (NS1) and fewer in WAKE SD2 vs. NS2, demonstrating that these results were significant when assessed within states as well as when compared across time blocks that involved pooled states as in Fig 4). Note also that there was a significant increase in the proportion of trajectory replays during NREM from PRE to POST, consistent with previous studies indicating increased replay following novel MAZE exposure^43,44. (D) Same as (C) but for the total number of trajectory replay events. Interestingly, the total number of trajectory replays decreased within WAKE in the NSD group, but did not change within SD, resulting in a greater total number of trajectory replays in SD2 compared to NS2. Importantly, however, there were significantly fewer trajectory replays in NREM RS vs. NS1. (E) Same as (C) but for duration of trajectory replay events (NREM: n = 8291 replays from 6 NSD sessions, n = 1869 replays from 7 SD sessions; WAKE: n = 9128 replays from 6 NSD sessions, n = 12940 replays from 7 NSD sessions). Note the decreased duration of these events during waking in SD2 vs. SD1. All box plots depict the median and top/bottom quartiles (whiskers = 1.5 × interquartile range) of the hierarchically bootstrapped data with individual session means overlaid with connecting dots. Statistics: Panel A: two-tailed, paired t-tests for within group comparisons and one-tailed Welch’s t-tests for cross-group comparisons; Panels B-E, two-sided within-group comparisons and one-sided cross-group comparisons with hierarchical bootstrap, #P < 0.01, *P < 0.05, **P < 0.01, ***P < 0.001, with no correction for multiple comparisons. See Supplementary Statistics Table for additional details.

Data availability

The processed group data for this study are available at https://doi.org/10.7302/73hn-m920 which includes Numpy .npy files used to generate most of the figures in this study. The remainder of the long-duration datasets generated during and analyzed for the current study will be made available by the corresponding author upon request.