Abstract
Sleep slow waves are known to participate in memory consolidation, but slow waves also occur in anesthetized states, with no positive effect on memory. Here, we shed light onto this paradox, based on a combination of analysis of extracellular recordings in vitro and in in vivo and computational models. We find two types of slow waves, based on analyzing the temporal patterns of successive slow-wave events. The first type of slow waves is seen during sleep, while the second type appears to prevail in anesthetized states. Network models of spiking neurons predict that these two slow-wave dynamics correspond to different levels of spike-frequency adaptation in excitatory cells. This prediction was tested in vitro by varying adaptation strength using an agonist of acetylcholine receptors, which demonstrated a neuromodulatory switch between the two types of slow waves. Finally, we show that the first type of slow-wave dynamics is more sensitive to external stimuli, which can explain how slow waves in sleep and anesthesia differentially affect memory consolidation, as well as provide a link between slow-wave dynamics and memory diseases.
Significance statement During sleep, cortical neurons display slow oscillations, which are believed to participate in memory consolidation. However, it remains unknown why the apparently similar slow-wave dynamics seen under anesthesia do not produce the same effect on memory. Here, we show distinctive features of slow oscillatory patterns in sleep versus anesthesia, robust across species and anesthetics. Using computational modeling and in vitro preparations, we show that in anesthesia, depressed neuromodulation stabilizes neural dynamics, limits communication between neural assemblies, and can explain why memory encoding is prevented. Because the same neuromodulatory system is depressed in both anesthesia and Alzheimers disease, our results offer a mechanistic link between the sleep and memory symptoms of the disease, with implications for the identification of novel therapeutics.
1 Introduction
In both natural sleep and anesthesia, cortical dynamics are characterized by slow, irregular oscillations (<1 Hz) [1]. However, certain cognitive processes, such as those involved in memory formation [2–4], are specific to deep sleep (also known as Slow-Wave Sleep, SWS), but not to anesthesia. While contrasts between sleep and anesthesia are relatively well-studied at the whole-brain scale [5–7], any underlying differences in the microscopic dynamics of cortical neural networks and their potential mechanisms remain to be identified. In fact, while anesthesia and sleep appear to produce similar collective dynamics, it it still unknown whether any differences exist in the correlation structure of network dynamics between the two states, and how they may affect the system’s ability to encode and remember stimuli.
At the neural population level, in both sleep and anesthesia, slow oscillations emerge from the alternation between transients of high neural firing (UP states) and transients of near silence (DOWN states) [1]. While certain regimes of anesthesia present very regular UP and DOWN states [8–10], in obvious contrast with the irregular patterns observed during sleep slow waves [7], irregular regimes also exist under anesthesia [10–12], and appear to present more ‘sleeplike’ dynamics. Additionally, UP-DOWN state activity has also been obtained in slice preparations in vitro [13]. Due to their general similarity in collective dynamics, slices and anesthesia, where direct pharmacological manipulation is possible, have often been used as models of natural sleep, paving the way to investigating mechanisms underlying UP-DOWN state activity.
Nevertheless, sleep and anesthesia present different levels of neuromodulation, known to affect the electrophysiological properties of neurons and how they respond to external inputs. In particular, as anesthesia is characterized by lower concentrations of acetylcholine (ACh) compared to sleep [14,15], the net effect is an enhancement of spike-frequency adaptation driven by K+ channels [16]. Following the increasing electrophysiological detail available on single neuron dynamics, computational models of spiking neurons have been employed to model spike-frequency adaptation and account for UP-DOWN state dynamics [17–19]. Indeed, comparison of models with neural recordings in anesthesia has recently uncovered a mechanistic interpretation for the emergence of UP and DOWN states, where background noise and neuromodulation, which controls spike-frequency adaptation, can account for the transitions between UP states and DOWN states [11]. While background noise is able to trigger a transition from a DOWN to UP state, spike-frequency adaptation on excitatory cells produces a self-inhibition that, destabilizing the UP state, causes a reset to the DOWN state [18]. However, it is plausible that subtle mechanistic particularities at the microscopic scale could be expressed as differences in the collective network dynamics that underlie distinct computational properties of sleep and anesthesia.
In the present paper, we find fundamental differences in correlation structure in slow waves during sleep and anesthesia, that can explain the different capacity for information encoding between the two states. We investigate computational models to search for plausible mechanisms underlying these differences. These mechanisms will be subsequently tested in in vitro preparations displaying UP-DOWN state dynamics, and where pharmacological manipulation is possible. In sum, we will show in which ways slow-wave dynamics differ between anesthesia and slow-wave sleep, robustly across different anesthetics, brain regions, and species, and propose possible mechanisms and functional consequences of these differences.
2 Results
To compare with previous anesthesia results [10,11], we consider the activity of a population of 102 neurons recorded from the temporal cortex of a human patient (Fig. S1A-C) during sleep. The dynamics is characterized by slow waves, as evident from Local Field Potentials (LFP) associated with an alternation of low (DOWN) and high (UP) activity periods, as visible in spiking activity. In all our analyses UP and DOWN states are defined based on neuron spiking activity (see Supplementary Materials and Methods). The time duration of both DOWN and UP states are variable, following an exponential, long-tailed distribution (see Supplementary Materials and Methods), similar to what has been reported for anesthesia recordings [11].
While the distributions of UP and DOWN state durations found in sleep and anesthesia are similar, surprisingly, the temporal distribution of UP and DOWN state durations is different in sleep compared to anesthesia. In anesthesia, DOWN state and next UP state durations are positively correlated, in other words long DOWN states are followed by long UP states, as shown in previous work [11]. However, we find that in sleep, DOWN state and next UP state durations are negatively correlated, as long DOWN states are followed by short UP states (Fig. 1B). Indeed, during sleep, while long UP states can occur after short DOWN states, UP states following long DOWN states are consistently short.
Because a difference in temporal correlation was found for adjacent UP and DOWN epochs between sleep and anesthesia, we next explored whether the network retains a memory of previous epochs further in time. To this end, correlations between the n-th DOWN state and the (n+k)-th UP state duration were explored. Here, k = 0 denotes the UP state following the nth DOWN state, as studied so far, while k = —1 denotes the UP state preceding the nth DOWN state in time. As shown in Fig.1C, the length of time correlations remain between DOWN and UP states lag k remains significantly negative up to a separation on the order of five UP/DOWN cycles in sleep data. In contrast, the correlations at lag k decay to zero immediately after one UP/DOWN state cycle in anesthesia [11]. Thus, the network retains a memory of previous cycles significantly longer in SWS than anesthesia.
In order to investigate whether the correlation between UP-DOWN state duration and its memory through time are specific to brain states across species and regions of cortex, data from primary visual cortex of animals under different anesthetics (monkey under sufentanil, rat under ketamine and medetomidine), and several animals sleeping (human temporal cortex, monkey premotor cortex, and rat prefrontal cortex) was analyzed.
The results of these analyses are reported in Fig.1 where in panels D-I we show the scatter plot of DOWN against following UP state duration. In all the sleep recordings a banana-shaped distribution is observed, indicating a robust negative UP-DOWN state correlation (for sleeping rats, 5/5 of animals analyzed showed a negative correlation, both before and after a navigation task).
Conversely, in anesthetized recordings the results are consistent with previously published results [11], an either positive or non-significant correlation is recorded (for anesthetized monkeys, 5/6 recordings presented a significant correlation and 1/6 a non-significant correlation; for anesthetized rats, 4/7 animals showed a significant positive correlation, 2/7 showed a positive non-significant correlation and the remaining 1/7 showed a negative significant correlation.)
Moreover, comparing within the same species (Fig.1F, I), the lag-correlation is verified to be different between sleep and anesthesia (we compare here the same species for the sake of coherence), with a much longer memory during sleep.
These observations suggest a clear difference in the correlation structure of the network dynamics during sleep and anesthesia, revealing fundamental differences in the dynamical mechanisms determining UP-DOWN state activity.
Spiking network model of UP and DOWN state dynamics
In order to investigate the mechanisms behind the UP and DOWN state duration correlations, we use a network of spiking neurons with conductance-based (COBA) synapses. The network is composed of 80% RS (regular spiking) excitatory and 20% fast spiking (FS) inhibitory neurons. Every neuron is modeled as an Adaptive Exponential integrate-and-fire cell (AdExp) [20]. In absence of adaptation the system is characterized by two stable states: a near-silent state (DOWN state) and a relatively high-activity state (UP state).
To allow for transitions between the two states, every neuron receives an independent identically distributed (i.i.d.) zero-mean noise of amplitude σ that permits a jump from the DOWN to the UP state. The presence of spike-frequency adaptation of strength b (see Supplementary Materials and Methods) for RS neurons [21] allows the system to transition back to the DOWN state. Indeed, RS neuron adaptation builds up as the neuron spikes, i.e. during UP states, and consequently reduces the firing rate of the excitatory population, which may cause the transition to a DOWN state. Adaptation decays exponentially throughout time when the neuron is silent, for instance during DOWN states (see Supplementary Materials and Methods for equations).
Such mechanisms for the emergence of UP and DOWN state dynamics has been so far established in the literature (see e.g. [11,22]). We observed the same mechanism in a spiking network model of Adaptive integrate and fire neurons with voltage-dependent synapses and a different gain between excitatory and inhibitory cells following experimental insights (see Supplementary Materials and Methods). The network dynamics is characterised by an alternation between UP and DOWN states whose durations follow an exponential distribution, in accordance with experimental data (see Fig 1C for data and Supplementary material for the model). For a fixed value of noise amplitude σ we observe a positive correlation between UP-DOWN dynamics, where the adaptation strength b changes the UP-DOWN duration, with no obvious effect on their correlation. The correlation between UP and DOWN state duration remains positive or non-significant over the all range of b values here investigated. This is consistent with adaptation having decayed after long DOWN states: following noise-triggered onset, the following UP state displays a high rate of activity that may sustain a long UP state. Consequently, long DOWN states tend to be followed by long UP states, hence the positive correlation. Exploring the parameter space by varying other parameters such as neurons’ excitability or synapses’ quantal conductances, positive or non-significant correlations were also always obtained.
In accordance with previously reported results [11], the model discussed so far is suitable for UP-DOWN dynamics during anesthesia but not for sleep, where we have shown a clear and robust inverse relationship. This shows additional elements are needed to accurately model the empirical UP-DOWN state dynamics during sleep.
Interplay between external fluctuation and adaptation strength
In order to investigate the mechanism enterting into play during sleep we observe that, additionally to adaptation strength, another natural parameter for affecting UP-DOWN state durations is the amount of noise σ. Indeed, the higher the fluctuations in the system, the shorter the DOWN states and the longer the UP states (Fig. 2A), implying that UP state and DOWN state durations vary in an anti-correlated fashion with σ. In other terms, if σ were to vary throughout time, a negative correlation could be observed between consecutive DOWN and UP state durations.
It is important to note, at this point, that our analyses are performed on relatively long recordings (from 20 minutes to several hours), and therefore one could assume that the properties of the system may change throughout the time of the recordings. The time scale T over which such changes occur seems relatively slow as compared to that of UP-DOWN dynamics (~ 1 s), i.e. in the range of tens of seconds.
To account for correlation inversion in our model, we introduce a parameter Δ describing the variability of noise amplitude σ throughout time. Here, the noise σ takes successive values within a range Δ of amplitudes σ (see Supplementary Materials and Methods), where each value is held constant over a time interval of duration T, as demonstrated in the inset of 2B. It should be noted that the resulting UP-DOWN correlations do not depend on the specific choice of T, as far as it is long enough to contain a sufficient number of UP-DOWN transitions in order to obtain well-defined UP DOWN statistics (in the plots shown in Fig 2 T = 100s).
By introducing variation of noise amplitude in time, a banana shape is observed in the scatter plot of UP and DOWN state duration, that disappears as the adaptation strength b is increased (Fig. 2B). Accordingly, a negative correlation between UP and DOWN state durations emerges increasing the range Δ of variation of the noise amplitude σ. Moreover, for sufficiently high Δ, an increase in the adaptation strength b is able to induce a transition from negative (sleep-like) to positive (anesthesia-like) correlation. In other words, adaptation is able to filter out noise variability, thus determining a positive correlation.
Apparent in the scatter plots of panel 2B, when adaptation is low, various values of noise amplitude σ (indicated by colors) cluster together in the scatter plot, altogether yielding a banana shape. Conversely, for high adaptation strength, data representing different values of noise amplitude overlap in the scatter plot, resulting in a non-significant or positive correlation. This can be understood as strong adaptation limiting the duration of UP states (green line in Fig. 2A, right), even in the presence of strong noise, and more generally controlling the transitions between UP and DOWN states. In sum, the model highlights the dominant mechanisms at work in each brain state, with the system being strongly adaptation driven in anesthesia, and fluctuation driven in sleep.
To investigate functional consequences of this difference in neuromodulation between sleep and anesthesia, we perform simulations to evaluate to what extent a stimulus affects collective dynamics. As shown in Fig. 3, the stimulus is simulated by delivering a spike train with Poisson statistics to all neurons, and the spike count after the stimulation is compared in the presence and in the absence of the stimulus (see Supplementary Materials and Methods). Immediately after the stimulation, anesthesia-like network is more responsive: as higher adaptation makes the network more silent, more spikes are evoked by the stimulus, relative to spontaneous firing rates. However, in the sleep-like network the difference between stimulated and non-stimulated dynamics diverges significantly faster, as lower adaptation makes the dynamics more sensitive to a stimulus. After tens of seconds, we therefore find that the difference between stimulated and non-stimulated networks is larger for low adaptation, suggesting a longer memory of the stimulus in the sleep-like case. This confirms that low adaptation strength renders sleep-like networks more sensitive to external noise and stimulations, thus allowing more encoding and memory of external inputs than in anesthesia-like networks.
Model prediction on non-stationarity in UP and DOWN state durations
A crucial ingredient of our model is the variability Δ in noise amplitude, especially in the sleep-like regime. Indeed, for lower adaptation strength, such as in SWS, noise fluctuations may play a larger role in shaping UP-DOWN dynamics. Accordingly, we expect to observe a higher variability in UP and DOWN state durations (as a direct outcome of noise variability, see Fig. 1) during sleep with respect to anesthesia. Comparing the mean empirical values of UP and DOWN state durations over relatively long time windows of 100 UP-DOWN cycles (i.e. 102 s), we observe a higher variability, of the order of 200%, under sleep as compared to under anesthesia, as predicted by the model (see Supplementary Materials and Methods). To further characterize the time scale of noise fluctuations, the time window size was varied, and the correlations across windows were studied. By collecting UP and DOWN durations in each window during sleep, we observe that, just as in our model, UP-DOWN state durations belonging to different windows have different correlation values. For short time windows (up to the order of 50 cycles), the Pearson coefficient is positive in the majority of windows, but becomes negative when computed over longer windows (see Supplementary Material). This suggests that fluctuations take place at a time scale T that can be as fast as the order of 10 seconds.
It can be noted this confirms the previous assumption that the time scale T of fluctuations is longer than the UP-DOWN cycle duration (of the order of 1 second). Conversely, T is much shorter than the duration of all our recordings (12 minutes to 3 hours) for either sleep or anesthesia, such that the absence of a negative correlation during anesthesia cannot be explained by too short recordings (unless T in anesthesia is not the same as in sleep, but much longer than the duration of the recordings studied here).
Additionally, the presence of background fluctuations at time scale T during sleep and not anesthesia is consistent with the apparent long memory of the UP-DOWN state duration correlation in SWS. Indeed, one may consider a period of time T over which background noise may be approximated as constant. During that period, if noise amplitude is high, UP states are long, and DOWN states short, and conversely if noise amplitude is low. Then whatever the lag, UP state durations are negatively correlated to the durations of DOWN states before or after them, provided that they occur within the same period of duration T. It is verified that the order of magnitude of T matches that of the time scale of the memory in sleep (a few UP-DOWN cycles, Fig. 1C, F, I).
Effect of in vitro modulation of adaptation strength on UP/DOWN correlation
Another strong prediction of our model is the ability of adaptation strength to modulate the correlations in UP and DOWN state durations. Spike-frequency adaptation models an effective action of activity-dependent potassium conductances, physiologically affected by neuromodulators [16].
It has been observed that neuromodulation is depressed during anesthesia [14], and thus the strength of adaptation should be increased [16]. This is consistent with our model prediction, where a transition to anesthesia (higher adaptation) yields a positive correlation between UP and DOWN states.
Nevertheless, a more direct experiment is preferable in order to validate our prediction. To this purpose we performed extracellular recordings of neural activity in acute slices of entorhinal cortex from wild type juvenile mice.
In control conditions, a certain variability was observed in both UP and DOWN state duration, sufficient to measure UP-DOWN states duration correlation.
The great advantage of in vitro preparations is the possibility to pharmacologically modulate adaptation [16] by dissolving neuromodulators in the artificial cerebrospinal fluid (ACSF) in which the slice is recorded.
To stimulate an effective decrease of adaptation strength, carbachol, an agonist of both nicotinic and muscarinic acetylcholine (ACh) receptors is used [16].
By increasing concentrations of carbachol (0, 0.025, 0.05 μM), we observe lengthening UP states and shortening DOWN states, Fig. 4, in accordance with a lower amount of spike-frequency adaptation, as in the model.
Finally, we report a clear tendency to more negatively correlated, sleep-like UP and DOWN states durations at higher carbachol concentrations. This confirms the predictions of our model, where an increase of adaptation strength permits a more positive correlation between UP and DOWN states durations.
3 Discussion
In the present paper, we have shown two types of slow waves by analyzing multi-electrode extracellular recordings in human and animal cerebral cortex. This conclusion was reached based on the temporal patterns of UP and DOWN states, an alternation of which causes slow waves to emerge. A first type of slow waves shows positively correlated consecutive UP and DOWN states, such as that found in previous work [11], while we find a second type of slow wave displaying negatively correlated consecutive UP and DOWN states. The first type of slow waves is seen during natural slow-wave sleep, while the second type appears to prevail in anesthetized states. Using computational models, we show that different levels of adaptation can explain the two different types of slow-waves. This prediction was tested in cortical slices, where the level of adaptation can be modulated by cholinergic agonists. These experiments demonstrate a cholinergic switch between the two types of slow waves. We also find that the two types of slow waves differ in their sensitivity to external inputs. We discuss below the significance of these findings, and perspectives for future work.
The difference between slow-wave sleep and anesthesia was robust across several species and brain regions. These results hold for slow waves produced by two different anesthetics, in contrast to natural slow waves observed in sleep. A larger effect of noise fluctuations throughout time on the network dynamics is observed during sleep, yielding a negative, long-memory correlation between UP and DOWN state durations. Conversely, during anesthesia the dynamics is more stable and characterized by short-memory and positive correlation between state durations. Employing a spiking neuron network model revealed that during anesthesia such noise fluctuations produce much smaller effects on population dynamics, as the fluctuations are filtered out by a strong modulatory effect (e.g. spike-frequency adaptation, which is controlled by acetylcholine levels). For the sake of simplicity, external noise was chosen to be the only effective source of variability in the model. Nevertheless, this is not the only possible choice: one could also consider variability in other parameters in time, that can be tuned to produce longer UP states and shorter DOWN states, like inhibitory conductance [23] or even adaptation strength (Fig. 4E). In sum, adaptation filters out time variation in the system’s parameters, in anesthesia but not in sleep. This mechanism is confirmed by modulation of spike-frequency adaptation strength in acute cortical slices. Indeed, addition of charbachol, a cholinergic receptor agonist, induced a switch between anesthesia-like and sleep-like slow-waves.
In addition, we show the implications of a switch in slow wave type for the network’s coding properties, with sleep-like slow waves associated with higher sensitivity to stimuli and greater capacity for network memory. This may help explain why key cognitive processes, such as memory consolidation, can take place during sleep [2–4], while anesthesia causes amnesia and memory impairment [24–26]. Our observation of two types of slow-waves suggests that slow-wave dynamics may be important to explain these differences. For example for memory consolidation to occur, the cortex should encode information by changing its dynamics upon receiving signals from the hippocampus. This implies that a non-trivial change in external input should be able to modulate the statistics of cortical activity. We showed (Fig. 3) that in anesthesia, unlike in sleep, strong adaptation filters out the effects of external perturbations on UP-DOWN state dynamics, so that any information encoded in the amplitude of inputs to the neural assembly does not affect the network dynamics, and consequently will not be encoded.
Finally, we emphasize that understanding the role of adaptation in filtering out external variability may shed light on pathological conditions where adaptation is disturbed. For example, the cholinergic system, that modulates spike-frequency adaptation [16], breaks down in Alzheimer’s disease [27]. A slowing down of slow oscillatory patterns during SWS [28,29] and a loss of memory [30] are also biomarkers of the disease. With loss of acetylcholine, spike-frequency adaptation will be increased, similarly to under anesthesia, and it is conceivable that the cortex of Alzheimer’s patients cannot encode fluctuating inputs from the hippocampus during sleep due to anesthesia-like slow-wave dynamics. This mechanism may contribute to explaining why new memories cannot be formed, and to better comprehending how treatments restoring acetylcholine levels alleviate Alzheimer’s sleep and memory symptoms [31]. Our study directly predicts that the anesthesia-type slow-wave dynamics should be observed in Alzheimer’s patients during their natural sleep, and provides an approach to modulating and quantifying the restoration of sleep type slow waves, perhaps a fruitful direction to explore in the design of new therapies.
4 Acknowledgments
This project/research has received funding from the European Unions Horizon 2020 Framework Programme for Research and Innovation under Grant Agreements No. 720270 (Human Brain Project SGA1) and No. 785907 (Human Brain Project SGA2). The authors thank C. Desbois, Z. Girones, M. Mattia, V. Medel, M.V. Sanchez-Vives, and Y. Zerlaut for useful discussion.