To monitor the population activity of POA^GAD2→TMN neurons in vivo during spontaneous sleep, we performed fiber photometry recordings. GAD2-Cre mice were injected with retrograde adeno-associated viruses (AAVs) encoding Cre-inducible GCaMP8s (AAVretro-FLEX-jGCaMP8s) into the TMN (Tervo et al., 2016; Zhang et al., 2023), and an optic fiber was implanted into the POA (Figure 1A; virus expression and optic fiber tracts were located in the ventrolateral POA, lateral POA, and the lateral part of medial POA). The calcium activity of POA^GAD2→TMN neurons was significantly higher during REMs compared with that during wake and NREMs (Figure 1B and C; detailed statistical results are shown in Figure legends and Supplementary file 1).

We further analyzed the activity changes of the POA^GAD2→TMN neurons during NREMs→REMs or NREMs→wake transitions. We found that the calcium activity of POA^GAD2→TMN neurons during NREMs→ REMs transitions becomes significantly increased 40 s before the REMs onset and remains elevated throughout REMs (Figure 1D). During NREMs→wake transitions, the activity of POA^GAD2→TMN neurons started rising 10 s before the wake onset and remained elevated for 10 s after the onset (Figure 1D). The ΔF/F activity gradually increased throughout NREMs episodes (Figure 1E).

Given that both the θ and σ (6–9 and 10.5–16 Hz) power increased preceding the REMs onset (Figure 1D), we investigated the relationship between the POA^GAD2→TMN neuron activity and the spectral composition of the EEG in more detail (Figure 1F). We used a linear regression model to predict the current POA^GAD2→TMN neural activity based on the preceding and following spectral EEG features (Weber et al., 2010; Schott et al., 2023). We found that the activity of POA^GAD2→TMN neurons is preceded by an increase in the θ and σ power and reduction in the δ (0.5–4.5 Hz) power (Figure 1F). These changes in the EEG are characteristic for the stage of NREMs preceding REMs (Gottesmann, 1996) and their correlation with the activity of POA^GAD2→TMN neurons is consistent with a role of these neurons in promoting NREMs to REMs transitions.

In a complementary experiment, we monitored the calcium activity of POA^GAD2→TMN axonal fibers. GAD2-Cre mice were injected with AAV-FLEX-GCaMP6s into the POA and an optic fiber was implanted into the TMN (Figure 1—figure supplement 1A). Similar to the activity found for POA^GAD2→TMN neurons using retrograde AAVs (Figure 1B and C), POA^GAD2→TMN fibers were most active during REMs (Figure 1—figure supplement 1B and C). During NREMs, the activity of POA^GAD2→TMN axonal fibers gradually increased before transitioning to REMs (Figure 1—figure supplement 1D), demonstrating that the activity of POA^GAD2→TMN axonal fibers closely resembles that of cell bodies. In summary, these results demonstrate that the activity in POA^GAD2→TMN neurons increases prior to the onset of REMs episodes, following an increase in the θ and σ power in the EEG.

The TMN contains histamine neurons (TMN^HIS), and previous studies showed that POA^GAD2 neurons innervate TMN^HIS neurons (Chung et al., 2017; Saito et al., 2018). Consistent with previous electrophysiological studies (Steininger et al., 1999; Vanni-Mercier et al., 2003; John et al., 2004; Takahashi et al., 2006), we found in photometry recordings that TMN^HIS neurons are highly active during wake and less active during NREMs and REMs (Figure 1—figure supplement 2A–C). Consistent with this, the activity of TMN^HIS neurons became significantly activated after the transition from NREMs or REMs to wakefulness (Figure 1—figure supplement 2D). The TMN^HIS neuron activity gradually decreased during NREMs episodes (Figure 1—figure supplement 2E), possibly as a result of inhibitory inputs from POA^GAD2→TMN neurons (Chung et al., 2017). Moreover, examining the time course of TMN^HIS neurons between two successive REMs episodes (inter-REM interval), we found that their activity gradually decreases throughout the inter-REM interval, reaching its lowest level at the onset of REMs (Figure 1—figure supplement 2F). This finding suggests that a minimal activity of TMN^HIS neurons is required for entering REMs, and suppression of TMN^HIS activity therefore likely facilitates transitions to REMs.

To examine whether POA^GAD2→TMN neurons regulate REMs, we optogenetically inhibited these neurons using the bistable chloride channel SwiChR++ (Berndt et al., 2016; Smith et al., 2024; Stucynski et al., 2022). GAD2-Cre mice were bilaterally injected with retrograde AAVs encoding Cre-inducible SwiChR++ (AAVretro-DIO-SwiChR++-eYFP) or eYFP (AAVretro-DIO-eYFP) into the TMN followed by bilateral optic fiber implantation into the POA (Figure 2A). We compared SwiChR++ and eYFP recordings with and without laser stimulation (473 nm, 2 s step pulses at 60 s intervals for 3 hr, zeitgeber time [ZT] 2–5). SwiChR++-mediated inhibition of POA^GAD2→TMN neurons reduced the amount of REMs compared with recordings without laser stimulation in the same mice and eYFP mice with laser stimulation (Figure 2B and C). The overall time spent in wake and NREMs was not altered by SwiChR++-mediated inhibition (Figure 2C, Figure 2—figure supplement 1A and B), suggesting that POA^GAD2→TMN neuron activity specifically regulates the amount of REMs.

Next, we compared the spectral composition of the EEG throughout recordings with or without laser stimulation in SwiChR++ and eYFP mice. During REMs, SwiChR++-mediated inhibition of POA^GAD2→TMN neurons increased the δ and σ power in the EEG compared with SwiChR++-without laser and eYFP-laser groups (Figure 2D, Figure 2—figure supplement 1C). The δ, θ, and σ power in the EEG during NREMs and wake was indistinguishable between SwiChR++ and eYFP groups (Figure 2—figure supplement 1D).

Next, we investigated the effect of sustained inhibition of POA^GAD2→TMN neurons on the following 3 hr recording without laser stimulation (Figure 2E). Despite the reduction of REMs during the laser stimulation in SwiChR++ mice (Figure 2C), there were no differences in the amount of REMs, NREMs, and wake between SwiChR++ and eYFP groups during the post-laser recordings (Figure 2E, Figure 2—figure supplement 1E and F), suggesting that the loss of REMs during SwiChR++-mediated inhibition was not followed by a homeostatic increase in REMs.

In a complementary experiment, we also investigated how inhibition of TMN^HIS neurons regulates REMs using the same inhibitory optogenetics protocol. HDC-Cre mice were bilaterally injected with AAVs encoding Cre-inducible SwiChR++ (AAV₂-EF1a-DIO-SwiChR++-eYFP) or eYFP (AAV₂-Ef1α-DIO-eYFP) into the TMN followed by bilateral optic fibers implantation into the TMN (Figure 2—figure supplement 2A). Consistent with the previous experiment, we compared SwiChR++ and eYFP recording with and without laser stimulation (Figure 2—figure supplement 2A and B). SwiChR++-mediated inhibition of TMN^HIS neurons increased the amount of REMs compared with recordings without laser stimulation in the same mice and eYFP mice with laser stimulation (Figure 2—figure supplement 2C). Given that TMN also contains other types of neurons, inhibition of histamine neurons by POA^GAD2→TMN neurons may not be the sole source of the observed effect on REMs upon inhibition of POA^GAD2→TMN neurons.

Taken together, we first demonstrated that SwiChR++-mediated inhibition of POA^GAD2→TMN neurons reduced the amount of REMs, supporting a necessary role of these neurons in REMs regulation. Second, the lost amount of REMs was not compensated for in the subsequent sleep suggesting that their activity is involved in the homeostatic regulation of REMs.

To probe whether the activity of POA^GAD2→TMN neurons changes during periods of high REMs pressure resulting from REMs restriction, we performed fiber photometry recordings combined with a closed-loop REMs restriction protocol (Figure 3A). To detect the onset of REMs, we adapted a previously applied automatic REMs detection algorithm (Weber et al., 2015; Weber et al., 2018; Stucynski et al., 2022; Schott et al., 2023). Briefly, the animal’s brain state was classified based on real-time analysis of the EEG/EMG signals. As soon as a REMs episode was detected, a small vibrating motor attached to the animal’s head was turned on to terminate REMs by briefly awakening the animal (Figure 3A, Figure 3—figure supplement 1A; Cardis et al., 2021; Osorio-Forero et al., 2023). Compared with the baseline recordings from the same mice during the same circadian time, we found that 6 hr of REMs restriction (ZT 1.5–7.5) significantly reduced the amount of REMs and duration of REMs episodes, while increasing their frequency (Figure 3—figure supplement 1A, D-F; amount: Cohen’s d [d]=–1.621; duration: d=–2.669; frequency: d=1.294). The number of motor activations gradually increased as mice tried to enter REMs more frequently, indicating the accumulation of REMs pressure (Figure 3—figure supplement 1B). Before the onset of the motor vibration, we found a clear increase in the EEG θ band reflecting NREMs to REMs transitions (Figure 3—figure supplement 1C). We also investigated the effects of REMs restriction on the spectral composition of the EEG. We found an increased δ power for REMs during restriction (Figure 3—figure supplement 1G). During rebound (ZT 7.5–8.5), the amount of REMs was significantly increased compared with that during the same circadian time due to an increased frequency of REMs episodes (Figure 3—figure supplement 1H-K; amount: d=0.868; frequency: d=1.147), compensating for the amount of REMs lost during restriction.

During REMs restriction, we observed an increased number of calcium transients in the activity of POA^GAD2→TMN neurons as the REMs pressure increased (Figure 3A and B). To detect and quantify these transients, we applied an algorithm previously applied to detect calcium events in fiber photometry recordings (Antila et al., 2022; Smith et al., 2024). Using fiber photometry, we monitored the population activity of POA^GAD2→TMN neurons during the last hour of REMs restriction (ZT 6.5–7.5, referred as restr.) and the first hour of REMs rebound (ZT 7.5–8.5, referred as reb.) and compared it with the activity during baseline recordings of the same mice on separate days at the same circadian time (Figure 3A and B). To restrict REMs, we used a motor (ZT 1.5–5.5) as previously described or gently pulled a string attached to the animal’s head (ZT 5.5–7.5) during the photometry recordings (Figure 3A). The manual REMs deprivation was utilized during the last 2 hr of REMs restriction to avoid potential motion artifacts from the vibrating motor that could contaminate calcium signals. We verified that REMs was adequately restricted (Figure 3—figure supplement 2). During restriction, the number of calcium transients of the POA^GAD2→TMN neurons was significantly increased compared with that during baseline recordings (Figure 3C). In particular, the number of calcium peaks was significantly elevated during NREMs, likely reflecting an increased pressure to transition to REMs (Figure 3C). As the duration of REMs episodes was largely reduced as a result of the restriction (Figure 3—figure supplement 2A), the number of peaks during REMs was not changed.

We further examined the amplitude of calcium transients during both NREMs and REMs. During REMs restriction, the amplitude of NREMs and REMs calcium transients was significantly higher compared with that during the circadian baseline (Figure 3D and E) and remained elevated during NREMs in the rebound phase (Figure 3F and G). Overall, these findings show that during periods of high REMs pressure, the POA^GAD2→TMN neurons exhibit an increased number of calcium transients with higher amplitude compared with that during baseline levels suggesting that the activity of these neurons may reflect the heightened homeostatic need for REMs.

To investigate whether the activity of POA^GAD2→TMN neurons encodes REMs pressure and consequently facilitates the subsequent rebound in REMs, we optogenetically inhibited these neurons during the last 3 hr of REMs restriction. GAD2-Cre mice were bilaterally injected with retrograde AAVs encoding Cre-inducible SwiChR++ (AAVretro-DIO-SwiChR++-eYFP) or eYFP (AAVretro-DIO-eYFP) into the TMN followed by bilateral optic fiber implantation into the POA (Figure 4A). Mice underwent REMs restriction (6 hr, ZT 1.5–7.5), and laser stimulation (2 s step pulses at 60 s intervals) was applied during the last 3 hr (ZT 4.5–7.5), when the REMs pressure was highest (Figure 4A, Figure 3—figure supplement 1B). During restriction, SwiChR-mediated inhibition of POA^GAD2→TMN neurons decreased the percentage of REMs and reduced the frequency of REMs episodes with marginal significance (Figure 4B–D, Figure 4—figure supplement 1A and B amount: d=–1.254). We found that inhibition of POA^GAD2→TMN neurons during REMs restriction resulted in a reduced amount of REMs during the rebound compared with that in eYFP mice (Figure 4B, F, and G, Figure 4—figure supplement 1D, E, amount: d=–1.317). Thus, inactivating POA^GAD2→TMN neurons during heightened REMs pressure not only decreased the amount of REMs, but also prevented its homeostatic rebound during the following recovery sleep.

Finally, we examined the spectral composition of the EEG during the REMs restriction. We found that inhibition of POA^GAD2→TMN neurons during REMs restriction reduced the REMs δ power, NREMs θ and σ power compared with that in eYFP mice with laser stimulation (Figure 4E, Figure 4—figure supplement 1C). The reduced θ and σ power during NREMs could result from less attempts to enter REMs (Figure 4D). During rebound, there were no significant differences in EEG δ, θ, and σ power during REMs and NREMs (Figure 4H, Figure 4—figure supplement 1F).

Taken together, inhibiting POA^GAD2→TMN neurons during REMs restriction significantly decreased the amount of REMs and blocked the subsequent rebound in REMs. Our data suggest that the heightened activity of POA^GAD2→TMN neurons during sleep encodes the increased need for REMs and consequently plays an important role in the homeostatic response to REMs restriction.

All experimental procedures were approved by the Institutional Animal Care and Use Committee (IACUC reference # 806197) at the University of Pennsylvania and conducted in compliance with the National Institutes of Health Office of Laboratory Animal Welfare Policy. Experiments were performed in male and female GAD2-IRES-Cre mice (#010802, Jackson Laboratory, generously donated by Taniguchi et al., 2011) or HDC-IRES-Cre mice (#021198, Jackson Laboratory, generously donated by Zecharia et al., 2012) aged 10–18 weeks, weighing 18–25 g at the time of surgery. Animals were group-housed with littermates on a 12 hr light/12 hr dark cycle (lights on 7 am and off 7 pm) with ad libitum access to food and water.

Cre-dependent adeno-associated viral vectors were used to selectively express GCaMP, SwiChR++, or eYFP in POA^GAD2 →TMN neurons, POA^GAD2 →TMN projections, or TMN^HIS neurons. pGP-AAV-Syn-FLEX-jGCaMP8s-WPRE was developed from the GENIE Project (Zhang et al., 2023) (162377-AAVrg, Addgene). pAAV-Syn-FLEX-GCaMP6s-WPRE-SV40 was developed from Douglas Kim & GENIE Project (Chen et al., 2013) (Penn Vector Core or 100845-AAV1, Addgene).

rAAV₂-Retro-Ef1α-DIO-SwiChR++-eYFP (R47730, UNC Vector Core).

rAAV₂-Retro-Ef1α-DIO-eYFP (R49556, UNC Vector Core).

All procedures followed the IACUC guidelines for rodent survival surgery. Mice were anesthetized with isoflurane (1–2%) during the surgery, and placed on a stereotaxic frame (Kopf) while being on a heating pad to maintain body temperature. The skin was incised and small holes were drilled for virus injections and implantations of optic fibers and EEG/EMG electrodes.

For fiber photometry experiments to image POA^GAD2→TMN neurons (Figures 1 and 3), pGP-AAV-Syn-FLEX-jGCaMP8s-WPRE was injected (Nanoject II, Drummond Scientific) into the TMN (300 nl; AP –2.4 mm; ML –1 mm; DV –5.4 to –5.2 mm, relative to bregma) and an optic fiber (400 μm diameter) was implanted into the POA (AP 0.2 mm; ML –0.6 mm; DV –5.2 mm). For imaging the POA^GAD2 →TMN axonal fibers (Figure 1—figure supplement 1), pAAV-Syn-FLEX-GCaMP6s-WPRE-SV40 was injected into the POA (300 nl) and an optic fiber was implanted into the TMN. To image TMN^HIS neurons (Figure 1—figure supplement 2), pAAV-Syn-FLEX-GCaMP6s-WPRE-SV40 was injected into the TMN (300 nl) and an optic fiber was implanted into the TMN.

For optogenetic inhibition experiments (Figures 2 and 4), rAAV₂-Retro-Ef1α-DIO-SwiChR++-eYFP (for inhibition group) or rAAV₂-Retro-Ef1α-DIO-eYFP (for control group) was bilaterally injected into the TMN (300 nl) and bilateral optic fibers (200 μm diameter) were implanted into the POA (AP 0.2 mm; ML ±1.5 mm [angled at 10°]; DV –5.2 mm).

All mice were implanted with electroencephalogram (EEG) and electromyogram (EMG) electrodes. EEG signals were recorded with stainless steel wires attached to two screws, located in the skull on top of the parietal (AP –2 mm; ML 2 mm) and frontal cortex (AP 1.7 mm; ML 0.6 mm). A reference screw was inserted on top of the cerebellum. Two EMG electrodes were inserted into the neck musculature. The incision was closed with suture and the EEG/EMG electrodes and optic fibers were secured to the skull using dental cement (A-M Systems). We performed optogenetic and fiber photometry experiments at least 4 weeks after surgery.

Mice were deeply anesthetized and transcardially perfused with phosphate-buffered saline (PBS) followed by 4% paraformaldehyde (PFA) in PBS. Brains were removed and fixed overnight in 4% PFA in PBS and then stored in 30% sucrose in PBS. Brains were embedded with OCT compound (Tissue-Tek, Sakura Finetek) and frozen. 40 μm sections were cut using a cryostat (Thermo Scientific HM525 NX) and directly mounted onto glass slides. Brain sections were washed in PBS for 5 min, permeabilized using PBST (0.3% Triton X-100 in PBS) for 30 min, and incubated in blocking solution (5% normal donkey serum in 0.3% PBST; 017-000-001, Jackson ImmunoResearch Laboratories) for 1 hr. Brain sections were incubated with chicken anti-GFP antibody (1:1000; GFP8794984, Aves Lab) in the blocking solution overnight at 4°C. The following morning, sections were washed in PBS and incubated for 3 hr with the donkey anti-chicken secondary antibody conjugated to a green Alexa fluorophore 488 (1:500; 703-545-155, Jackson ImmunoResearch Laboratories). Afterward, sections were washed with PBS followed by counterstaining with Hoechst solution (#33342, Thermo Scientific). Slides were coverslipped with mounting medium (Fluoromount-G, Southern Biotechnic) and imaged using a fluorescence microscope (Microscope, Leica DM6B; Camera, Leica DFC7000GT; LED, Leica CTR6 LED) to verify virus expression and optic fiber placement. Animals were excluded if no virus expression is detected or the virus expression/optic fiber tips were not properly localized to the targeted area.

We generated heatmaps of the virus expression across mice as previously described (Smith et al., 2024; Stucynski et al., 2022; Schott et al., 2023). Coronal reference images for the corresponding AP coordinates were downloaded from the Allen Reference Atlas - Mouse Brain (atlas.brain-map.org). For a given AP reference atlas section, the corresponding histology section from each mouse was overlaid and regions in which GCaMP labeled cell bodies were present were manually outlined. Custom Python programs detected these outlines and determined for each location on the reference picture the number of mice with overlapping virus expression, which was encoded using different green color intensities.

All sleep recordings were performed in a cage to which the animal had been habituated for several days. All recordings were performed during the light phase between 8 am and 5 pm (ZT1–10) in sound-attenuating chambers. For sleep recordings, EEG and EMG signals were recorded using an RHD2132 amplifier (Intan Technologies, sampling rate 1 kHz) connected to an RHD USB interface board (Intan Technologies). For fiber photometry experiments, a calcium signal was recorded using an RZ5P amplifier (Tucker-Davis Technologies, sampling rate 1.5 kHz). EEG and EMG signals were referenced to a ground screw located on top of the cerebellum. At the start of each sleep recording, EEG and EMG electrodes were connected to flexible recording cables via small connectors. To determine the brain state of the animal, we first computed the EEG and EMG spectrogram for sliding, half-overlapping 5 s windows, resulting in 2.5 s time resolution. To estimate within each 5 s window the power spectral density (PSD), we performed Welch’s method with Hanning window using sliding, half-overlapping 2 s intervals. Next, we computed the time-dependent δ (0.5–4 Hz), θ (5–12 Hz), σ (12–20 Hz), and high γ (100–150 Hz) power by integrating the EEG power in the corresponding ranges within the EEG spectrogram. In addition, we calculated the ratio of the θ and δ power (θ/δ) and the EMG power in the range of 50–500 Hz. For each power band, we used its temporal mean to separate it into a low and high part (except for the EMG and θ/δ ratio, where we used the mean plus one standard deviation as threshold). REMs was defined by a high θ/δ ratio, low EMG, and low δ power. NREMs was defined by high δ power, a low θ/δ ratio, and low EMG power. In addition, states with low EMG power, low δ power, but high σ power were scored as NREMs. Wake was defined by low δ power, high EMG power, and high γ power (if not otherwise classified as REMs). Our automatic algorithm that has been previously used in Weber et al., 2015; Weber et al., 2018; Chung et al., 2017; Antila et al., 2022; Smith et al., 2024; Stucynski et al., 2022; Schott et al., 2023, has 90.256% accuracy compared with the manual scoring by expert annotators. We manually verified the automatic classification using a graphical user interface visualizing the raw EEG and EMG signals, EEG spectrograms, EMG amplitudes, and the hypnogram to correct for errors, by visiting each single 2.5 s epoch in the hypnograms. The software for automatic brain state classification and manual scoring was programmed in Python (https://github.com/tortugar/Lab/tree/tortugar-patch-1/PySleep copy archived at tortugar, 2024).

Prior to the recording, the optic fiber and EEG/EMG electrodes were connected to flexible patch cables. For calcium imaging, a first LED (Doric lenses) generated the excitation wavelength of 465 nm and a second LED emitted 405 nm light, which served as control for bleaching and motion artifacts. 465 and 405 nm signals were modulated at two different frequencies, 210 and 330 Hz respectively. Both lights traveled through dichroic mirrors (Doric lenses) before entering a patch cable attached to the optic fiber. Fluorescence signals emitted by GCaMP8s or GCaMP6s were collected by the optic fiber and traveled via the patch cable through a dichroic mirror and GFP emission filter (Doric lenses) before entering a photoreceiver (Newport Co.). Photoreceiver signals were relayed to an RZ5P amplifier and demodulated into two signals using the Synapse software (Tucker-Davis Technologies), corresponding to the 465 and 405 nm excitation wavelengths. To analyze the calcium activity, we used custom-written Python scripts. First, both signals were low-pass filtered at 2 Hz using a fourth order digital Butterworth filter. Next, we fitted the 405 nm to the 465 nm signal using linear regression. Finally, the linear fit was subtracted from the 465 nm signal to correct for photobleaching and/or motion artifacts, and the difference was divided by the linear fit yielding the ΔF/F signal. Both the fluorescence signals and EEG/EMG signals were simultaneously recorded using the RZ5P amplifier. Fiber photometry recordings were excluded if the signals suddenly shifted, likely due to a loose connection between the optic fiber and patch cable.

Sleep recordings were performed during the light phase (ZT2–8). Mice were tethered to bilateral patch cables connected with the lasers and a flexible recording cable to record EEG/EMG signals. The recording started after 30 min of habituation. For optogenetic inhibition experiments, 2 s step pulses (1–3 mW, 60 s intervals) were generated by a blue laser (473 nm, Laserglow) and sent through the optic fiber (200 µm diameter, Thorlabs) connected to the ferrule on the animal’s head for 3 hr (ZT2–5). This laser stimulation protocol was rationally designed based on previous reports of sustained inhibition and prior results that recapitulate similar findings as inhibitory chemogenetic techniques (Iyer et al., 2016; Kim et al., 2016; Wiegert et al., 2017; Stucynski et al., 2022). TTL pulses to trigger the laser were controlled using a Raspberry Pi, which was controlled by a custom-programmed user interface programmed in Python. Following sustained inhibition, an additional 3 hr (ZT5–8) recording was performed without laser stimulation (post-laser session). Baseline recordings were performed (without laser stimulation, ZT2–5), and counterbalanced across mice and days to avoid potential order effects. For each mouse, we collected two to three baseline and laser recordings each.

We employed an automatic REMs detection algorithm (Weber et al., 2015; Weber et al., 2018; Stucynski et al., 2022; Schott et al., 2023) and used a small vibrating motor to terminate/restrict REMs (Cardis et al., 2021; Osorio-Forero et al., 2023; https://github.com/luthilab/IntanLuthiLab, copy archived at luthilabcol, 2023). A small vibrating motor (DC 3 V Mini Vibration Motor, diameter: 10 mm, thickness: 3 mm, BOJACK) with a soldered lobster claw clasp was attached to a small wire hook secured in the dental cement of each mouse head. The motors had cables that were connected to a Raspberry Pi which controlled the motor onset and offset.

The automatic REMs detection algorithm determined whether the mouse was in REMs or not based on real-time spectral analysis of the EEG/EMG signals. The onset of REMs was defined as the time point where the EEG θ/δ ratio exceeded a threshold (mean + 1 std of θ/δ), which was calculated from the same mouse using previous recordings. As soon as a REMs episode was detected, the small vibrating motor turned on to terminate REMs and consequently woke up the mouse. The motor vibrated until the REMs episode was terminated, i.e., when the θ/δ ratio dropped below its mean value or if the EMG amplitude surpassed a threshold (mean + 0.5 std of amplitude). All REMs restriction experiments started at ZT1.5 and lasted until ZT7.5. Following the restriction, motors were turned off and mice were permitted to enter recovery sleep.

To monitor the calcium activity during REMs restriction using fiber photometry, we performed manual REMs restriction to avoid potential motion artifacts caused by the vibrating motor that could interfere with the integrity of the fiber photometry signals. Briefly mice underwent a 4 hr automatic REMs restriction protocol as described above. During the last 2 hr of restriction, REMs was detected manually based on the EEG and EMG signal by an experimenter who gently pulled on a string attached to the hook secured in the dental cement of the mouse head. To avoid potential photobleaching, the recording was performed during the last 1 hr of REMs restriction and 1 hr of REMs rebound. Each mouse also underwent a baseline recording on a separate day. The order of REMs restriction and baseline recordings were varied to minimize the impact of the experimental sequence on the results.

We also performed automatic REMs restriction combined with optogenetic inhibition. Each mouse underwent REMs restriction for 6 hr, and we continuously delivered 2 s step pulses (1–3 mW, 473 nm, Laserglow) at 60 s intervals during the last 3 hr of restriction (ZT4.5–7.5). Following REMs restriction, the recovery sleep was recorded.

To calculate the neural activity changes relative to brain state transitions (Figure 1D, Figure 1—figure supplements 1D and 2D), we aligned the ΔF/F signals for transitions across all mice relative to the time point of the transition (t=0 s). For each NREMs→ REMs or NREMs→wake transition, we ensured that the preceding NREMs episodes lasted for at least 60 s, only interrupted by short awakenings (≤20 s). To determine the time point at which the activity significantly started to increase or decrease, we used the first 10 s of NREMs as baseline. For REMs→wake and wake→NREMs transitions, the preceding REMs or wake episode was at least 30 s long. Using one-way repeated measures (rm) ANOVA, we tested whether the activity (downsampled to 10 s bins) within each 10 s bin was significantly modulated throughout the transition (from –60 to 30 s). Finally, using pairwise t-tests with Holm-Bonferroni correction, we determined the time bins for which the activity significantly differed from the baseline bin (activity for bin –60 to –50 s). The time point for a given 10 s bin was set to its midpoint. To account for multiple comparisons, we divided the significance level (α=0.05) by the number of comparisons (Bonferroni correction). To analyze the activity throughout NREMs, we normalized the duration of all NREMs episodes and the corresponding ΔF/F signals to the same length (Figure 1E, Figure 1—figure supplement 2D and E).

To identify features of the EEG spectrogram associated with POA^GAD2→TMN calcium activity, we adapted a receptive field model (Weber et al., 2010; Schott et al., 2023) to predict POA^GAD2→TMN neural activity from the spectrogram. Intuitively, we estimated a spectrotemporal filter using linear regression that predicts for each time point the POA^GAD2 →TMN calcium response. In more detail, we first computed the EEG spectrogram E(f_i, t_j) using 2 s windows with 80% overlap, resulting in a time resolution dt of 400 ms. Each spectrogram frequency was normalized by its mean power across the recording, and the parameter E(f_i, t_j) specifies the relative amplitude of frequency f_i for time point t_j. We then downsampled the ΔF/F response using the same time resolution as for the spectrogram, and extracted all time bins with REMs, NREMs, or wake for analysis.

To predict the calcium response, the EEG spectrogram was linearly filtered with the kernel H(f_i, t_l). In analogy to receptive fields estimated using similar approaches for sensory neurons, we used the term ‘spectral field’ for H(f_i, t_l). The spectral field can be described as the set of coefficients that optimally relate the neural activity at time t_j to the EEG spectrogram at time t_{j + l}, where l represents the lag between the time points of the spectrogram and the neural response. The estimated frequency components of H(f_i, t_l) ranged from f₁=0.5 Hz to f_nf = 20 Hz, while the time axis ranged from –n_T * dt = –50 s to n_T * dt = 50 s. Thus, H(f_i, t_l) comprises n_f frequencies and 2 * n_T+1 time bins over a window from –50 to +50 s relative to the neural response. The convolution of the EEG spectrogram with the spectral field can be expressed as(1)r^(tj)=r0+∑l=−nrnr∑i=1nfE(fi,tj+l)H(fiti)

The scalar parameter r₀ denotes a constant offset, and r(t||j) denotes the predicted neural response. The optimal spectral field H(f_i, t_l) minimizes the mean-squared error between the predicted and measured neural response at time t_j. To account for the large number of estimated parameters, we included a regularization term in the error function, which penalizes large kernel components and therefore guards against overfitting of the model. The spectral field for each recording was estimated using fivefold cross-validation to determine the regularization parameter (λ) that optimized the average model performance on the test sets. Kernels were averaged across all recordings for individual animals; the spectrotemporal correlation in Figure 1F represents the mean spectral field across all mice.

The PSD of the EEG was computed using Welch’s method with the Hanning window for sliding, half-overlapping 2 s intervals. To calculate the power within a given frequency band, we approximated the corresponding area under the spectral density curve using a midpoint Riemann sum. To compute the EMG amplitude, we calculated the PSD of the EMG and integrated frequencies between 5 and 100 Hz. To test whether laser stimulation changed the spectral density during a specific brain state, we determined for each mouse the δ, θ, or σ power for that state with and without laser. To compare the PSD between experimental and control mice, we calculated the δ, θ, and σ power with and without laser stimulation for each mouse and normalized the power values with laser by the power obtained for epochs without laser.

To detect calcium transients, we first filtered the ΔF/F signal with a zero-lag, fourth order digital Butterworth filter with cutoff frequency of 1/20 Hz. Next, prominent peaks in the signal were detected using the function scipy.find_peaks provided by the open source Python library scipy (https://scipy.org). As parameter for the peak prominence, we used 0.15 * distance between the 1st and 99th percentile of the distribution of the ΔF/F signal. A transient was defined as occurring during NREMs and REMs, if the peak overlapped with NREMs or REMs respectively. The calcium transient amplitude was calculated by using the values 10 s preceding the peak and subtracting that from the peak values at 0 s.

Statistical analyses were performed using the Python modules (scipy.stats, https://scipy.org; pingouin, https://pingouin-stats.org) and Prism v9.5.0.0 (GraphPad Software Inc). We did not predetermine sample sizes, but cohorts were similarly sized as in other relevant sleep studies (Jego et al., 2013; Ma et al., 2019). All data collection was randomized and counterbalanced. All data are reported as mean ± s.e.m. A (corrected) p-value <0.05 was considered statistically significant for all comparisons. Data were compared using unpaired t-tests, paired t-tests, one-way ANOVAs, or two-way ANOVAs followed by multiple comparisons as appropriate. The statistical results for the figures are presented in the Supplementary file 1 and Figure legends.

The code used for data analysis is publicly available under: https://github.com/tortugar/Lab, copy archived at tortugar, 2024. All the data have been deposited at Zenodo (https://zenodo.org) and are available as of the date of publication.