Weak representation of awake/sleep states by local field potentials in aged mice

Senescence affects various aspects of sleep, and it remains unclear how sleep-related neuronal network activity is altered by senescence. Here, we recorded local field potential signals from multiple brain regions covering the forebrain in young (10-week-old) and aged (2-year-old) mice. Interregional LFP correlations across these brain regions could not detect pronounced differences between awake and sleep states in both young and aged mice. Multivariate analyses with machine learning algorithms with uniform manifold approximation and projection and robust continuous clustering demonstrated that LFP correlational patterns at multiple frequency bands, ranging from delta to high gamma bands, in aged mice less represented awake/sleep states than those in young mice. By housing aged mice in an enriched environment, the LFP patterns were changed to more precisely represent awake/sleep states. Our results demonstrate senescence-induced changes in neuronal activity at the network level and provide insight into the prevention of pathological symptoms associated with sleep disturbance in senescence.

rior cingulate cortex (ACC), primary motor cortex (M1), striatum (STR), primary somatosensory cortex (S1), hippocampus (HPC), and medial parietal association cortex (MPtA) (Fig. 1B). These brain areas were selected so that they covered the wide range of the forebrain from anterior to posterior parts. Recording sites were verified by histological inspection. Figure 1C shows representative simultaneous recordings of LFP traces and an EMG trace obtained from a freely moving mouse. These electrophysiological signals were recorded from young (8-to 10-week-old; collectively termed 10-week) and aged (95-to 105-week-old; collectively termed 2-year) mouse groups. All recordings were performed in the familiar home cage that was transported to a novel soundproof room. Therefore, the mice were acclimated to recording environments. For each mouse, a recording time was 2.5-3.5 h.
First, we classified all recording periods into awake/sleep states based on the combination of EMG signals and cortical LFP signals ( Supplementary Fig. S1A). EMG signals were converted to root mean square (RMS) traces ( Fig. 1D) to define active awake states with a threshold of mean + 0.8-1.2 × SD ( Supplementary Fig. 1C, top raw), which represents apparent animal's movement ( Supplementary Fig. 1B). In addition, according to the awake/ sleep classification utilized in a previous study [19][20][21][22][23][24] . we defined slow wave sleep (SWS) states ( Supplementary  Fig. 1C, middle raw), rapid eye movement (REM) sleep states ( Supplementary Fig. 1C, bottom raw), and quiet awake states, based on the threshold defined by delta and theta power of cortical LFP signals and the EMG RMS traces ( Fig. 1D; for more detail, see Methods). For further analyses, active awake and quiet awake states were summarized as awake periods, whereas SWS and REM sleep states were summarized as sleep periods. In the representative 10-week and 2-year mice shown in Fig. 1D, sleep time accounted for 30.1% and 44.9% of the total recording time, respectively (individual data shown in Supplementary Fig. 1D). Overall, the percentage of sleep time in 2-year mice was not significantly different from that in 10-week mice ( Fig. 1E; n = 7 and 8 mice; Z = 1.79, P = 0.073, Mann-Whitney U test).
To confirm that our recording conditions were stable over time and not strongly affected by time-dependent effects (e.g. acclimatation over time), we compared sleep time observed from the first 20 min and the last 20 min of the recording period ( Supplementary Fig. 1E). No significant differences were found between the two time periods in both the 10-week (n = 7 mice; Z = 0.67, P = 0.49, Wilcoxon signed rank test) and 2-year-old groups (n = 8 mice; Z = 1.40, P = 0.17, Wilcoxon signed rank test). These results verify that the frequency of awake/sleep cycles did not prominently change over the entire recording time.
No pronounced differences in interregional LFP correlations between awake and sleep states both in young and aged mice. We next analyzed LFP signals from multiple brain regions. For each brain region, an LFP signal was converted to its power traces in six frequency bands (delta (δ), 1-4 Hz; theta (θ), 4-8 Hz; alpha (α), [8][9][10][11][12][13] Hz; beta (β), 13-30 Hz; low-gamma (low-γ (L)), 30-50 Hz; high-gamma (high-γ (H)), 50-100 Hz) ( Fig. 2A). First, we compared LFP power in these individual frequency bands averaged over all awake and sleep states. Several brain regions showed differences in LFP power at a single frequency between awake and sleep states in both 10-week and 2-year mice (Supplementary Fig. 2A). In all brain regions, delta   Fig. 2B), consistent with previous studies 19,25 . We next examined regional correlations of LFP power changes (i.e., functional connection) in awake/sleep states. To obtain an overview of functional connections from all datasets, we computed correlation coefficients of LFP power changes in individual frequency bands between two brain regions throughout the entire recording session. Correlation coefficients from all brain region pairs ( 6 C 2 = 15) from a representative mouse are summarized as a color-coded map and shown in Fig. 2B. In both the 10-week and 2-year mice, the correlation coefficients in all the pairs were not significantly different between awake and sleep states in the majority of frequency bands, except few frequency bands ( Fig. 2C; n = 7 and 8 mice; P > 0.99 (10-week, delta), P > 0.99 (10-week, theta), P = 0.032 (10-week, alpha), P = 0.0036 (10-week, beta), P > 0.99 (10-week, low-gamma), P > 0.99 (10-week, high-gamma), P > 0.99 (2-year, delta), P = 0.051 (2-year, theta), P = 0.29 (2-year, alpha), P > 0.99 (2-year, beta), P = 0.33 (2-year, www.nature.com/scientificreports/ low-gamma), P = 0.00090 (2-year, high-gamma), Wilcoxon signed rank test followed by Bonferroni correction). Next, we compared awake/sleep differences between the 10-week group and the 2-year group by calculating differences in correlation coefficients between awake and sleep states as Δcorr in individual brain region pairs and summarized as a color-coded map (Fig. 2D). Figure 2E shows all maps of Δcorr in the six frequency bands averaged over all mice. Overall, cumulative distributions revealed no significant differences in Δcorr between the 10-week group and the 2-year group in all frequency bands ( Fig. 2F; n = 7 and 8 mice; P > 0.99 (delta), P > 0.99 (theta), P = 0.22 (alpha), P > 0.99 (beta), P > 0.99 (low-gamma), P > 0.99 (high-gamma), Mann-Whitney U test followed by Bonferroni correction). These results demonstrate that LFP power changes at single frequency bands at single brain region pairs are not sufficient to differentiate awake and sleep states both in young and aged mice.
Weak representation of awake/sleep states by LFP patterns at multiple frequency bands in aged mice. The analyses in Fig. 2 focused on LFP signals at single frequency bands averaged over all awake/ sleep periods. Under natural conditions, awake/sleep states continuously vary across time (generally tens of seconds to several minutes). We next asked whether such time-varying awake/sleep state patterns could be differentiated by instantaneous brain LFP patterns. To this end, recording periods of 2.5-3.5 h were divided into bins of 10 s, and each bin was categorized as an awake or a sleep bin (for more detail, see Methods). In each bin, correlation coefficients of LFP power changes at each frequency band were computed across the 15 pairs of the six brain regions to obtain a 15-dimensional vector (Fig. 3A). Our analysis tested whether these correlational LFP patterns in each bin could represent an awake or sleep bin. For each frequency band, we applied UMAP (uniform manifold approximation and projection), a dimension reduction technique without any subjective bias, to these vectors. Figure 3B shows representative two-dimensional UMAP plots in the delta band for a www.nature.com/scientificreports/ 10-week and 2-year mouse. In these graphs, the plots appeared to be not clearly separated between awake and sleep bins. To quantify the degree of separation, the robust continuous clustering (RCC) algorithm was applied to all the plots in each graph. After identifying clusters by the RCC, each cluster was categorized as an awake or a sleep cluster, depending on whether there were more awake or sleep bins in the cluster, respectively. An F1 score was then computed as an index based on the percentage of awake/sleep bins assigned in awake/sleep clusters ( Fig. 3B,C). To assess the significance of an F1 score, shuffled datasets were constructed by randomly shuffling types (awake/sleep) to which the individual plots across all the plots were categorized. An F1 score was considered to be significant when it was higher than the top 5% of F1 scores computed from 1000 shuffled datasets (a distribution shown in the bottom panel in Fig. 3B). Of the seven 10-week mice, five (71.4%), three (42.8%), five (71.4%), two (28.6%), six (85.7%), and two (28.6%) animals did not show significant F1 scores for the delta, theta, alpha, beta, low-gamma, and high-gamma bands, respectively ( These results suggest that regional correlations of LFP power changes computed in a single frequency band do not perfectly represent awake/sleep states in both of the ages. Next, we asked whether the awake/sleep separations in LFP patterns were improved when the datasets in all six frequency bands were employed in an analysis. All vectors in the six frequency bands in individual bins were combined to construct 90-dimensional vectors (Fig. 3D), and the same procedures, plotting on UMAP dimensions and computing F1 scores, were applied to these 90-dimensional vectors (Fig. 3E). All 10-week mice showed significant F1 scores, whereas only 12.5% (1/8) of the 2-year mice showed significant F1 scores (Fig. 3F). Taken together, our multidimensional analyses demonstrate that awake and sleep states are less precisely represented by LFP correlational patterns representing entire brain networks in 2-year mice than in 10-week mice.
Enriched environment restores LFP correlations to represent awake/sleep states in aged mice. For rodent animals, housing in an enriched environment (EE) has been shown to exert positive effects on various behavioral patterns and sleep quality 26,27 . Here, we examined whether living in an EE could restore aging-related LFP changes (Fig. 4A). First, we tested whether our experimental EE conditions were sufficient to affect memory performance. An object location test was utilized in which a mouse was first introduced to two identical objects in an open field (acquisition phase) and then exposed to the same two objects, one of which was displaced to a new location (test phase) (Fig. 4B). If the mouse remembered the locations of the objects in the acquisition phase, it would spend more time exploring the object in the new position. In this behavioral test, we combined datasets from a subset of the mice that were used for electrophysiological recordings and a subset of www.nature.com/scientificreports/ new mice that were not used for electrophysiological recordings (for more detail, see Materials and Methods). The location index, representing the interaction time with the novel object in the test phase, was significantly higher than the chance level (50%) in 10-week mice (n = 6 mice; Z = 2.20, P = 0.028, Wilcoxon signed rank test) but not in 2-year mice without EE (termed 2-year non-EE mice; n = 9 mice; Z = 0.18, P = 0.86, Wilcoxon signed rank test) (Fig. 4B). On the other hand, 2-year mice housed in EE (termed 2-year EE mice) for more than 2 weeks exhibited a location index that was significantly higher than chance (n = 9 mice; Z = 2.00, P = 0.036, Wilcoxon signed rank test). These results confirmed that the 2-year mice housed in EE had sufficient memory that could be detected by the object location test, similar to the 10-week mice, which were not observed in the 2-year mice without EE. Furthermore, the percentage of sleep time in 2-year EE mice was significantly lower than that in 2-year non-EE mice ( Fig. 4C; n = 8 and 6 mice; Z = 2.65, P = 0.0081, Mann-Whitney U test). Then we confirmed gene expression patterns in the dorsal hippocampus in these mouse groups. Aged mice exhibited altered expression patterns of a subset of genes related to neuron development and synaptic signaling ( Supplementary Figs. 3A,C and 4). On the other hand, the 2-year EE mice exhibited increased or decreased expression levels of a subset of genes (as marked in blue or green in Supplementary Figs. 3A and 4), which became more similar to those observed from the 10-week-mice, confirming altered expression patterns of aging-related genes by EE.
Next, we applied the same analytical procedures as in Fig. 3F, plotting on UMAP dimensions and computing F1 scores, to the LFP datasets in all the frequency bands from the 2-year EE mice (Fig. 4D). Out of 6 mice tested, 5 (83.3%) mice exhibited significant F1 scores in the 2-year EE group, which was prominently higher than that (12.5% (1/8)) observed from the 2-year non-EE group (Fig. 4E). These results suggest that aged mice housed in EE conditions show brain network LFP patterns that distinguish awake/sleep states, which were not observed in the majority of the aged mice without EE conditions.

Discussion
To understand senescence-induced changes in neuronal activity at the network level, we developed a device that simultaneously records local field potential signals from multiple brain regions in the forebrain together with EMG signals by improving our previous methods [28][29][30] . We confirmed that the total sleep time was not different between young and aged mice. While some studies have shown that sleep time becomes longer in aged mice 19,31 , this discrepancy may be due to experimental conditions such as recording time in a day or manually defined parameters to detect sleep periods [19][20][21]31,32 . We demonstrated that single regional correlations of LFP power changes (i.e. functional connections) computed in a single frequency band do not perfectly represent awake/sleep states in both of the ages. Notably, machine learning algorithms with UMAP and RCC applied to multidimensional datasets consisting of LFP patterns at multiple frequency bands together, ranging from delta to high-gamma bands, revealed that they could differentiate awake/sleep states in young mice, suggesting the usefulness of these analytical methods for awake/sleep-related LFP pattern analyses. We note that our recording was acute and restricted to 3.5 h with a single day in each animal. Longer or chronic recordings might provide further insights into how awake/sleep states undergo changes across days.
The importance of focusing on interregional activity patterns is consistent with the suggestions that awake and sleep brain states are not simply sustained by activity levels of single brain regions but generated from complex interactions across widespread brain regions 33,34 . When the datasets in all six frequency bands were combined, the representations of awake/sleep states by LFP signals were improved in young mice. On the other hand, these LFP correlational patterns less differentiated awake/sleep states in aged mice. The aging-induced changes in neuronal dynamics at the network levels may be explained by cooperative interactions among a number of molecular and cellular factors that have been reported to alter with aging, including decreases in neuromodulators 5 , reduced expression of neuromodulator receptors 35 , and decreased synchronicity of neurons in the suprachiasmatic nucleus 36 . Further studies are needed to determine which these microscopic mechanisms most strongly contribute to awake/sleep-related LFP patterns.
Finally, our results demonstrated that housing aged mice in EE conditions reinstated the aging-related decreases in LFP patterns such that they more precisely represented awake/sleep states. In addition to the positive effects of EE on various behavioral patterns 26,27 and gene expression patterns 37,38 reported previously, our study verified that EE can restore awake/sleep-related neuronal activity patterns. Considering that declines in sleep quality in senescence are crucial risk factors for neurodegenerative and psychiatric diseases 4 , EE is expected to prevent these pathological symptoms through improvements in the quality of sleep and sleep-related neuronal activity. www.nature.com/scientificreports/ conditions of controlled temperature and humidity (22 ± 1 °C, 55 ± 5%) in a vivarium, maintained on a 12:12-h light/dark cycle (lights off from 7 am to 7 pm) with ad libitum access to food and water. All mice were housed individually. The numbers of mice used for electrophysiological recordings were seven (10-week), eight (2-year (non-EE)), and six (2-year (EE)). Of these recorded mice, four (10-week), five (2-year (non-EE)), and six (2-year (EE)) mice were used for the object location test. In addition, two (10-week), four (2-year (non-EE)), and three (2-year (EE)) mice were used for the object location test alone without electrophysiological recordings. No animals were excluded from our analyses after obtaining datasets.

Materials and methods
Enriched environment. All young 10-week-old mice were housed in a homecage. Two-year-old mice were randomly selected and exposed to either standard environment (non-EE) or enriched environment (EE) conditions for more than 2 weeks. For non-EE conditions, standard laboratory cages were used (24 cm × 17 cm × 12 cm), whereas mice maintained under EE conditions were housed in a rat cage (42 cm × 26 cm × 20 cm). The EE cages were equipped with a running wheel, a seesaw, a ball, and a climbing platform (Fig. 4A).
Object location test. The experimental apparatus used in this study was an open-field box (47 cm × 37 cm with walls with a height of 31 cm). The apparatus was placed in a sound-isolated room. Identical plastic cones (10 cm in height × 6 cm in diameter) created using a 3D printer (Form 2, Formlabs, MA) were used as objects. This test consisted of a habituation day and a test day. On a habituation day, the mice were allowed to freely explore the apparatus without objects for 1 h. On the next day, which was the test day, the same habituation procedure was first conducted for 30 min before starting the acquisition phase. In the acquisition phase, the mice were allowed to freely explore two identical objects that were placed symmetrically in the experimental apparatus for 5 min. The mouse was then removed from the apparatus and returned to its home cage. The objects were thoroughly cleaned with 70% ethanol. The open field box was cleaned with dry paper after each trial to ensure that it was saturated with the smell of the animals. A test phase was conducted 1 h after the acquisition phase. In the test phase, one of the objects (A2) was moved to a different location (A2′(D)), and the other object (A1) remained in the same position (A1(ND)) as in the acquisition phase. The terms D and ND indicate the displaced and the nondisplaced objects in the test phase, respectively. In the test phase, the mouse was allowed to freely explore the experimental apparatus for 5 min. Their behavior was recorded with a video camera mounted above the apparatus. The time spent exploring each object was measured using DeepLabCut 39 . Exploration of an object was defined as pointing the nose toward the object at a distance of < 1 cm and/or touching it with the nose. To analyze memory performance, a location index was calculated as follows: TA2′(D)/(TA1(ND) + TA2′(D)), where TA2′(D) is the time exploring the displaced object and TA1(ND) is the time exploring the nondisplaced object.
Preparation of an electrode array. The electrode array used in this study has been described elesewhere 40 .
An electrode array for brain LFP recording was assembled consisting of custom-made parts (Fig. 1A) and an electrical interface board (EIB) (Neuralynx, Bozeman, MT). A plastic core body, created by a 3D printer, contained multiple small holes with a dimeter of 0.7 mm distributed in space corresponding with the XY-coordinates of the targeted cortical areas, which served as a template to determine the locations of electrodes. An electrode assembly was created by setting metal tubes and electrodes into these holes so that the tips of the electrodes corresponded with the depth of individual brain regions. The other open ends of the electrodes were connected to metal holes of an EIB mounted on the top of a core body. An EIB consisted of 30 LFP channels and 2 ground channels, and all electrical signals from these channels were transferred to an Omentins connector. For brain LFP recording, a nichrome wire (A-M Systems, WA) or a tetrode that was constructed by bundling together four 17-μm polyimide-coated platinum-iridium (90/10%) wires (California Fine Wire, CA) and plated with platinum used to adjust the electrode impedances to 150-300 kΩ were used. The size of the electrode assembly was width 20 mm, length 20 mm, height 41 mm, and weight 2.2 g. The open edges of all electrodes were soldered to the corresponding channels on the EIB.

Surgery.
Standard surgical procedures were similar to those described previously 28,41,42 . Animals were anesthetized with 1-2% of isoflurane gas in air. The animal was then fixed in a stereotaxic instrument with two ear bars and a nose clamp. First, two craniotomies were made; one covering the coordinates for the anterior cingulate cortex (ACC; 1.2 mm anterior and 0.2 mm lateral to the bregma), primary motor cortex (M1: 1.2 mm anterior and 1.6 mm lateral to the bregma), and the striatum (STR: 1.2 mm anterior and 1.6 mm lateral to the bregma), and the other covering the coordinates for primary somatosensory cortex (S1; 1.9 mm posterior and 3.0 mm lateral to the bregma), the hippocampus (HPC; 1.9 mm posterior and 1.4 mm lateral to the bregma), and the medial parietal association cortex (MPtA; 1.9 mm posterior and 0.5 mm lateral to the bregma). The electrode array was directly implanted into the cortical tissue in the right hemisphere with electrodes inserted 0.5 mm into S1 and MPtA, 1.0 mm into ACC and M1, 1.4 mm into HPC, and 2.3 mm into STR. An electromyogram (EMG) electrode was implanted into the dorsal neck area. For the cerebellum, stainless steel screws were implanted on the skull attached to the brain surface to serve as ground/reference electrodes. Finally, all of the wires and the electrode array were secured to the skull using dental cement. After completing all surgical procedures, the anesthesia was terminated and the animals were spontaneously allowed to awake from the anesthesia. Following surgery, each animal was housed in a standard environment or an enriched environment with free access to water and food, with daily observation.
In vivo electrophysiology. For  www.nature.com/scientificreports/ the mice in the EE condition, all of the objects (e.g. the running wheel, the seesaw, the ball, and the climbing platform) were removed from the cage so that the recording conditions were consistent across all mouse groups. After transporting into the recording room, recordings started within 30 min. Under this condition, the mice were already acclimated to local environments in the homecage while they to some extent recognized the novelty of distal environments of the room. The EIB on the animal's head was connected to a digital headstage Cereplex M and Cereplex μ (Blackrock Microsystems), and signals were digitized and they were transferred to an electrophysiological data acquisition system Cereplex Direct (Blackrock Microsystems). Local field potential (LFP) recordings commenced at a sampling rate of 2 kHz for 2.5-3.5 h. For all recordings, electrophysiological signals were filtered between 0.1 Hz and 500 Hz. All electrophysiological recordings were performed between 10 am and 2 pm.
Histology of brain tissue. The procedure has been described elesewhere 29,43 . After the recordings, the mice were perfused intracardially with cold 4% paraformaldehyde (PFA) in 25 mM phosphate-buffered saline (PBS) and decapitated. The electrodes were carefully removed from the brain 6-8 h after the perfusion. The brains were placed in 30% sucrose until equilibrated and coronally sectioned at a thickness of 50 mm, and the slices were stained with cresyl violet.
Definition of awake/sleep states. All analyses were performed using custom-made MATLAB 2020a (MathWorks, MA, USA) and Python 3 routines. The awake/sleep states were determined using the same algorithm as in the previous study 19 . Briefly, active awake states were defined first based on EMG signals. The root mean square (RMS) of the EMG signals was computed every 1 s with a window of 5 s. Active awake states were defined as periods when EMG RMS was above a threshold ( Supplementary Fig. 1C, top raw). The threshold was manually adjusted for individual mice with the aid of visual inspection of EMG RMS traces 19 . The threshold was the mean + 0.8-1.2 × SD, which corresponded with an animal's moving speed of 3 cm/s ( Supplementary Fig. 1B), a speed generally considered to fully represent mouse's movement, not optical noise during sleep states. Periods with EMG RMS of the mean + 5.0 × SD was considered as massive noise and excluded from the analysis. For the other periods except active awake states, we computed the ratio of delta power to theta power in cortical LFP signals every 1 s with a window of 5 s and slow wave sleep (SWS) states were defined as periods when EMG RMS was above a threshold ( Supplementary Fig. 1C, middle raw). The threshold was manually adjusted for individual mice with the aid of visual inspection of LFP power traces 20,21 . The threshold was the mean + 0.2-0.5 × SD. For the other periods except active awake and SWS states, we computed an index, LFP theta power/(LFP delta power × EMG RMS), which has been utilized as an efficient measure to detect rapid eye movement (REM) sleep states 19 , every 1 s with a window of 5 s, and REM sleep states were defined as periods when this index was above a threshold ( Supplementary Fig. 1C, bottom raw). The threshold was manually adjusted for individual mice with the aid of visual inspection of LFP power traces 19 . The threshold was the mean + 1.0-1.2 × SD. The other periods that were not classified into these three states were classified into quiet awake periods. Active awake and quiet awake states were considered as awake periods, whereas SWS and REM sleep states were considered as sleep periods.
LFP power and correlation analysis. LFP signals were downsampled to 400 Hz, and LFP spectral power in six frequency bands (δ: 1-4 Hz, θ: 4-8 Hz, α: 8-13 Hz, β: 13-30 Hz, low-γ: 30-50 Hz, and high-γ: 50-100 Hz) was computed by Morlet wavelet analysis. Similar to the definition of awake/sleep states, the RMS of the EMG signals was computed every 1 s. For LFP analysis, LFP signals from 2.5-3.5 h were divided into 10-s bins. An awake bin was defined when a 10-s bin contained awake time shorter than 5 s, whereas a sleep bin was defined when a 10-s bin contained sleep time of 5 s or longer than 5 s. In each bin, correlations of LFP power changes in each frequency band were computed from the 15 pairs of brain regions to obtain a 15-dimensional vector. When specified, vectors in all five frequency bands were combined to obtain a 90-dimensional vector.
Machine learning algorithms. Uniform manifold approximation and projection (UMAP) was used for nonlinear dimensionality reduction 44 . With UMAP, 15-or 90-dimensional vectors were reduced to two dimensions with the following hyperparameters: n_neighbors = 3, min_dist = 0.1, n_components = 2, and metric = 'euclidean' . We verified that "n_neighbors" ranging from 3 to 5 and "min_dist" ranging from 0 to 0.1 yielded the same statistical results. Then, a robust continuous clustering (RCC) algorithm, an unsupervised clustering method, was applied to the plots from UMAP (Python implemented with the following hyperparameters: clus-tering_threshold = 60, k = 60, and measure = 'euclidean') ( Fig. 3) 45 . We verified that "clustering_threshold" ranging from 40 to 80 and "k" ranging from 40 to 80 yielded the same statistical results. If more than half of the bins in a cluster defined by the RCC were awake or sleep bins, the cluster was classified as an awake cluster or a sleep cluster, respectively. To quantify the degree of the separation between awake/ sleep bins in awake/sleep clusters, in each animal, F1 scores for awake and sleep bins (termed F1 awake and F1 sleep , respectively) were computed as the harmonic mean of precision and recall, where precision was the ratio of the number of awake/sleep bins included in awake/sleep clusters to the number of all bins included in the awake/ sleep cluster, and recall was the ratio of the number of awake/sleep bins included in awake/sleep clusters to the number of all awake/sleep bins. In each animal, an F1 score was computed as (F1 awake + F1 sleep )/2. To assess the significance of an F1 score, shuffled datasets were constructed for each original UMAP plot by randomly shuffling bin types (awake/sleep) assigned to the plots. For more limited conditions, shuffling was performed every 18 consecutive bins (i.e. 180 s), a duration representing the majority of sleep duration, which could keep temporal structures of the order and vicinity of epochs within this time period. An F1 score for the original data was considered to be significant when it was higher than top 5% of F1 scores computed from 1000 shuffled datasets.