Early studies that aimed to clarify the functions of sleep in the brain focused on gene expression changes following brief (i.e., hours-long) periods of sleep or SD. These studies assessed changes in mRNA levels in different parts of the brain—hypothalamus (Terao et al., 2003), neocortex (Cirelli et al., 2004; Mackiewicz et al., 2007), cerebellum (Cirelli et al., 2004), and hippocampus (Vecsey et al., 2012)—using microarray analysis. Across these studies, a consistent finding is that the expression of genes involved in RNA, protein, and lipid biosynthetic pathways, the unfolded protein response (UPR), and synaptic plasticity change as a function of sleep and wake. More specifically, sleep is associated with increased expression of genes associated with protein and lipid synthesis, while SD is associated with increased expression of genes involved in mRNA transcription, cellular stress and the UPR. In support of SHY, in many of these studies, sleep decreases and wake increases expression of a subset of genes thought to be involved in activity-mediated synaptic plasticity—including arc, cfos, bdnf, narp, and homer1a. More recently, the Allen Brain Institute has made microarray and in situ hybridization data available from numerous regions in sleeping and sleep-deprived animals, revealing a more complex picture of gene regulation (i.e., across the entire brain) during different behavioral states (Wang et al., 2010). These gene expression data have been used as support for the idea that activity-mediated synaptic plasticity is widespread in the brain during wake, and generally reduced during sleep.

A critical unresolved question is whether the levels of protein translated from sleep- and wake-regulated mRNAs are similarly altered by state. Changes in protein levels appear to track transcript level changes in some of the cases where it has been carefully investigated (Cirelli et al., 1996; Simor et al., 2017). However, state-dependent changes in protein synthesis (Ramm and Smith, 1990) may compensate for some changes in gene expression during SD. For example, in the hippocampus, levels of arc and hspa5/BiP mRNA increase across a brief period of SD; however, due to alterations in translation rates during wake, levels of Arc and Hspa5/BiP protein abundance remain unchanged (Tudor et al., 2016).

Despite these caveats, recent studies have found evidence in support of SHY based on synaptic protein expression. In rats, for example, expression of GluA1 and active (phosphorylated) CaMKII is increased by roughly 20–40% in cortical and hippocampal synaptoneurosomes following a 6-h period of SD, relative to a similar period of sleep (Vyazovskiy et al., 2008). A more recent study (Diering et al., 2017) reported a similar ~20% decrease in GluA1 and GluA2 content in mouse forebrain synaptosomes during the circadian sleep phase relative to the wake phase. Critically, however, the authors were unable to replicate the previously-reported effects of SD on these targets (i.e., synaptic GluA1 and GluA2 levels were identical with sleep, SD, and SD + recovery sleep). Nonetheless, the authors conclude based on these data that a global downscaling mechanism acts on synapses during sleep (Diering et al., 2017).

Effects of sleep have also been seen at the level of dendritic structure in the developing brain. Yang and Gan (2012) recently used in vivo imaging of layer 5 pyramidal neurons' dendrites in the somatosensory cortex of juvenile mice, to investigate the effects of brief (i.e., 2-h) periods of sleep and SD on spine turnover. The authors found that across 2 h of SD, total dendritic spine/filopodia density increased by ~5%, while across 2 h of ad lib sleep, it decreased by ~5%. The difference was apparently due to increased elimination rates for existing spines and filopodia during sleep (there was no difference in the rate of new spine formation between sleep and wake). More recently, serial scanning electron microscopy (SEM) was used to reconstruct and measure dendritic spines (and apposed axon termini) in layer 2 of primary motor and somatosensory cortex of juvenile mice after periods of wake (spontaneous or enforced) vs. sleep (de Vivo et al., 2017). By quantifying the surface area of thousands of individual contacts between axon terminals and spines, the authors concluded that sleep leads to a small (~10–15%) but significant decrease in synaptic contact area. This effect is heterogeneous, with the largest synaptic contacts unaffected by sleep vs. wake. More modest effects of sleep are seen on the size of dendritic spines themselves (e.g., spine volume), with only a small subset of spines quantified showing any effect of sleep vs. wake. While the function consequence of these changes remains unclear, proponents of SHY have pointed to these findings as the most direct evidence that synaptic strength is reduced during sleep.

Numerous recent studies have used neuronal firing rates in freely behaving animals as surrogate measure for (or potential functional readout of) synaptic strength. While this measure is indirect, and can also be affected by changes in the intrinsic excitability, many laboratories have used it as a potential indicator of overall synaptic strength (Vyazovskiy et al., 2009; Aton et al., 2013, 2014; Hengen et al., 2013, 2016). For example, Vyazovskiy et al. tracked firing rates of rat barrel cortex neurons across periods of sleep and wake, and across the circadian day (Vyazovskiy et al., 2009). In this study, the authors found that neurons tended to fire at a lower rate at the end of the day (when “sleep pressure”—i.e., the propensity to fall asleep—is low) compared with the beginning of the day (when sleep pressure is high). Assuming that firing rate was directly proportional to (excitatory) synaptic strength, the authors concluded that these data indicated that greater synaptic strength is associated with greater sleep pressure, and that sleep reduces synaptic strength (in support of SHY). Importantly however, while these effects on firing rate were present both at the level of multiunit activity and in single neurons identified by the authors as fast-spiking interneurons, they were not statistically significant in putative pyramidal neurons. Nonetheless, this was the first demonstration of a functional change in neural circuits that could be related to the proposed mechanism in SHY.

If we assume (based on these findings) differential effects of sleep on firing in fast-spiking interneurons and pyramidal neurons, one possibility is that excitatory/inhibitory balance (i.e., the ratio of activity in glutamatergic and GABAergic neurons) is the major feature of cortical physiology that changes with sleep pressure. In support of this idea, a recent study of the hippocampal neurons' firing across states found the highest ratio of interneuron firing -to- pyramidal neuron firing during active wake (Miyawaki and Diba, 2016). Vanini et al. (2012) recently demonstrated that the rate of GABA release in the cortex increases steadily across periods of sustained wake. Another recent study found that while glutamate release in rat somatosensory and motor cortex also increases across brief periods of spontaneous wake, with SD, extracellular glutamate levels initially rise (over a period of 30 min- 2 h) and then fall (Dash et al., 2009). This supports the idea that sustained wake leads to a gradual decrease in excitatory/inhibitory balance associated with increasing inhibitory neurotransmission.

Additional evidence suggests that synaptic function per se may be altered after sleep vs. wake. For example, Liu et al. recently measured the frequency and amplitude of mEPSCs in layer 2/3 pyramidal neurons of juvenile rat and mouse frontal cortex after periods of sleep and wake (Liu et al., 2010). While the authors concluded that a 4-h period of SD significantly increased both mEPSC amplitude and frequency, it is worth noting that values for both sleep and wake groups were highly variable. For example, depending on the set of experiments in the study, sleep deprived and sleeping groups' mEPSC frequencies were either quite distinct, or completely overlapped (Liu et al., 2010). Furthermore, while frequency changes (presumably reflecting effects on presynaptic release of glutamate) were relatively large, mEPSC amplitude changes (which would be affected by post-synaptic changes in receptor expression) were minimal across sleep vs. wake or SD. However, to date, this is the most direct evidence of functional synaptic weakening across a period of sleep.

The data outlined above has been put forth by proponents of SHY as evidence of sleep-dependent downscaling, which renormalizes synapses following changes in neural circuits (i.e., synaptic potentiation) caused by wake-associated learning. One major caveat is that many of the studies described above (and all of the studies describing anatomical changes) were carried out in adolescent animals (see Table 1). As is true for humans (Tang et al., 2014), the rate of synapse elimination in both adolescent rats (corresponding to post-natal weeks 5–9; Drzewiecki et al., 2016) and mice (corresponding to post-natal weeks 4–8; Zuo et al., 2005a,b; Bian et al., 2015) is maximal, and significantly higher than that seen in the adult brain. An unanswered question is whether sleep plays a special role in promoting developmentally-regulated synapse downscaling and elimination in adolescence, or whether sleep-dependent synaptic effects are present across the lifespan. Effects of sleep on synaptic structure and function in the adult brain are still far from clear.

A second caveat is that in many of the studies supporting SHY, comparisons between sleeping and awake animals are confounded by one of two factors. Either (1) samples come from animals spontaneously asleep or awake at different circadian times, or (2) SD animals used for comparison have been deprived of sleep through environmental enrichment (e.g., novel object or running wheel presentation) that was not provided to sleeping animals (see footnotes in Table 1).

A third major caveat is that while SHY proponents have used a global downscaling mechanism to explain neural network performance improvement using computational models (Hill et al., 2008), biological data have not supported the idea of global downscaling during sleep. For example, cortical and hippocampal neurons show non-uniform changes in firing rate across bouts of sleep (Miyawaki and Diba, 2016; Watson et al., 2016). Specifically, neurons with the highest baseline firing rates (including interneurons) show decreases in spontaneous activity across periods of NREM sleep, while neurons with lower baseline firing rates show either no change, or an increase, in spontaneous firing across a period of sleep. This indicates that functionally, not all neurons are equally affected by sleep, and it stands to reason that not all synapses are equally affected. Indeed, as described above the available anatomical evidence indicates that only a subset of synapses show a reduction in size across a period of sleep (de Vivo et al., 2017). Based on these new findings, the use of the term “synaptic downscaling” may itself be questionable, as sleep does not appear to have truly global effects with regard to reducing synaptic strength (i.e., “scaling” may not be present).

For this reason, more recent descriptions of SHY have proposed that sleep leads to a decrease in the strength of only a subset of synapses, while preserving the strength of others. This preservation would be highly desirable for processes involved in learning and long-term memory formation, where information encoded by neural circuits prior to sleep needs to be retained or reinforced. Given these findings, a critical question is why during sleep, some synapses (and possibly some neurons) are apparently unaffected, while others undergo an apparent reduction in strength.

Finally, in none of these studies were the observed changes linked with sleep-dependent cognitive function. The animals under study were housed in standard (i.e., non-enriched, and presumably non-challenging) conditions, and were not being trained on specific learning tasks. While sleep affects numerous aspects of cognition (including experience-dependent sensory plasticity and memory consolidation, described in detail below) that are affected by sleep, sleep's effects on these processes have not been linked to synaptic weakening. Thus, while converging data suggest that under steady-state conditions, modest weakening of at least some synapses can been observed in multiple brain areas across periods of sleep, the function of this for information processing in the brain (if any) is still unknown.

Recent computational modeling studies from the Tononi lab (Olcese et al., 2010; Nere et al., 2013) invoked a modified STDP rule to explain reductions in synaptic strength during sleep, and the effects of this process on memory. The STDP rule employed dictated that during learning in the wake state, synapses with temporally correlated pre- and post-synaptic firing would be strengthened, while synapses with non-correlated firing would either be unaffected, or would be weakened. In contrast, during sleep, synapses with temporally correlated pre- and post-synaptic firing would be unaffected (i.e., their strength would be preserved), while synapses with non-correlated firing would be weakened. In the earlier study, this was implemented computationally by simply inverting the sign of STDP normally seen the cortex (Feldman, 2000). As implemented in this scheme, the same spike timing would cause LTD instead of LTP, for the same pre-post-activity pairing, if it was present in sleep instead of wake. The authors argued that the presence or absence of neuromodulators (a function of brain state) would result in the same pattern of firing having differential effects on synapses in the two states. This model was meant to illustrate the benefits of sleep-dependent reductions in synaptic strength, rather than to clarify the cellular mechanisms in operation in vivo. However, it is necessary to point out that the proposed cellular mechanism is at odds with neurobiological data in two important ways.

First, sleep and wake are not monolithic with regard to neuromodulation, nor is the neuromodulation state of the cortex binary. Dopamine, serotonin, acetylcholine, and norepinephrine release rates are differentially regulated by state, and these effects vary according to where in the brain release is being measured (Marrosu et al., 1995; Portas et al., 1998; de Saint Hilaire et al., 2000; Lena et al., 2005). Second and more importantly, the effects of the state-regulated neuromodulators dopamine, acetylcholine, and norepinephrine on STDP do not support the notion that STDP rules “flip” between wake and NREM sleep. Each neuromodulator has distinct effects on the relationship between spike timing and synaptic strength changes (Pawlak et al., 2010), however, none of these effects fit with the assumptions of the model's modified STDP rule. For example, acetylcholine (with cortical release highest during wake, intermediate during REM and lowest during NREM) (Marrosu et al., 1995) can block timing-based LTP and promote timing-based LTD of glutamatergic synapses in cortical pyramidal neurons (Seol et al., 2007). In contrast, noradrenergic signaling (with cortical release highest during wake, intermediate during NREM and lowest during REM) promotes timing-based LTP in both cortical pyramidal neurons and interneurons. These effects are independent of the relative timing of action potentials and EPSPs—i.e., neuromodulator tone, but not the ordering of pre- and post-synaptic activity, determines the outcome of spike pairing. Taken together, available data suggest that higher norepinephrine and acetycholine levels during wake would lower the threshold for inducing both STDP-based LTP and LTD. It is therefore unlikely that changes in neuromodulation alone would bias plasticity in favor of LTD during sleep.

Central to SHY is the concept of a globally-acting homeostatic mechanism which maintains synaptic strengths within a set physiological range. Homeostatic synaptic downscaling is a mechanism of plasticity that is thought to function in exactly this way, to counteract the network-level effects of excessive neuronal activity and synaptic excitation. Homeostatic downscaling differs from Hebbian synaptic weakening (e.g., LTD) with regard to both mechanism of induction and function. While LTD induction requires appropriately timed pre- and post-synaptic firing, and can lead to functional changes within minutes to hours, homeostatic downscaling appears to require increased neuronal firing and acts over a slower timescale of several hours to days. Homeostatic downscaling was first described in vitro by (Turrigiano et al., 1998), who described divisive shifts in neurons' mEPSC amplitude distributions in response to long-term increases in firing. Specifically, the authors found that 48 h of exposure to the GABAA receptor antagonist bicuculline led to a global reduction in neurons' mEPSC amplitude distribution (to ~66% of baseline) (Turrigiano et al., 1998). This study, along with numerous others since its publication, have led our current understanding of downscaling, wherein perturbations in either neuronal firing rate or neurotransmission leads to a global reduction of post-synaptic strength over several hours to days. Functional decreases in synaptic strength due to downscaling are accompanied by decreases in glutamatergic receptor (e.g., AMPAR) expression and spine volume (Turrigiano et al., 1998; Fernandes and Carvalho, 2016; Keck et al., 2017).

Only recently has sleep been implicated in regulating molecular pathways involved in homeostatic downscaling. Homeostatic reductions in AMPA receptor expression are mediated through multiple cellular pathways, and there is evidence that these pathways may be affected in parallel by sleep. Recent phosphoproteome profiling indicates that a kinase critical for downscaling, cyclin dependent kinase 5 (CDK5), is more active in the brain during the sleep phase of the rodent circadian cycle (Diering et al., 2017). CDK5 activity is increased in the nucleus of neurons in response to increased network activity (Liang et al., 2015), and is implicated in numerous cellular pathways that could promote synaptic downscaling. Within the nucleus CDK5 phosphorylates many targets including MeCP2. This phosphorylation event is critical for decreasing gluA2 mRNA expression in response to an increase in neuronal activity. CDK5 also interacts with polo like kinase 2 (PLK2) to promote downscaling via downstream effects on the Rap GTPase pathway. This leads to regulation of Rap-mediated changes in AMPA receptor trafficking and dendritic growth (Seeburg et al., 2008; Lee et al., 2011).

A second pathway which has received significant attention as a possible link between sleep and homeostatic downscaling is the Homer1a pathway. Homer1a is an immediate early gene and the short isoform of constitutively active Homer proteins. The constitutive Homer proteins act as scaffolds which bring together a complex including NMDA receptors and mGluR5 receptors at the post-synaptic density. In response to increased neuronal activity, the shorter Homer1a protein acts as a dominant negative isoform, which can disrupt this complex (Kammermeier and Worley, 2007). Loss of Homer1a disrupts homeostatic downscaling (Siddoway et al., 2014), and restoring its expression leads to decreased AMPA and metabotropic glutamate receptor expression at the post-synaptic density (Hu et al., 2010). Recent gene expression studies have shown that homer1a expression increases across the brain in response to SD (Nelson et al., 2004; Mackiewicz et al., 2007) and the genetic locus for homer1a has been implicated in the homeostatic regulation of NREM slow wave activity (Mackiewicz et al., 2008). Diering et al. (2017) recently found that Homer1a protein abundance at synapses rapidly increases during SD. If we assume that Homer1a localization at the synapse results in downscaling, this finding would suggest that downscaling occurs during wake. However, the authors also reported that reductions in synaptic GluA1 and GluA2 during the sleep phase of the circadian cycle were dependent on Homer1a. To reconcile these findings, the authors hypothesized that Homer1a mobilization to the synapse is gated by both norepinephrine and adenosine levels. They speculated that during wake, high levels of norepinephrine maintain Homer1a outside the synapse; reduced norepinephrine and increased adenosine levels lead to delivery of Homer1a to the post-synaptic density during sleep. In support of this idea, treating mice with the norepinephrine reuptake inhibitor d-amphetamine (or an A1 adenosine receptor antagonist) reduced synaptic Homer1a levels, while treating them with norepinephrine receptor antagonists increased synaptic Homer1a (Diering et al., 2017). The authors of the study argued that this represented a plausible mechanism whereby prolonged wakefulness could lead to subsequent synaptic downscaling during sleep. However, it is worth noting that in this study, the observed sleep-associated reduction in GluA1 levels preferentially occurred among spines with the highest baseline levels (i.e., it was not global). Indeed, some spines showed increases in GluA1 levels. Taken together with other evidence showing that synaptic weakening is heterogeneous during sleep (de Vivo et al., 2017), these data actually argue against true synaptic downscaling as a mechanism for sleep-dependent synaptic changes.

Intriguingly, sleep and wakefulness may have differential effects on so-called “upscaling”—which globally increases synaptic strengths in response to decreased network activity. Hengen et al. evoked homeostatic plasticity in freely behaving mice via monocular lid suture, leading to reduced visual cortex activity. The authors found that homeostatic increases in spontaneous firing rate after this treatment were primarily expressed across bouts of wake, with longer wake epochs resulting in greater firing rate increases (Hengen et al., 2016). The authors concluded that cellular mechanisms responsible for upscaling are active during wake, and inhibited by sleep.

Numerous studies have indicated that homeostatic responses to increased network activity may also involve modifications to GABAergic synapses, effecting a change in the balance of network excitation and inhibition. Following periods of overactivity, inhibitory synapses on pyramidal neurons have been shown to undergo presynaptic and post-synaptic enhancements, including increases in presynaptic GAD65 and GABAA receptor surface expression (Peng et al., 2010; Rannals and Kapur, 2011). Recent data suggest that GABAA receptor surface expression is increased on cortical pyramidal neurons in vivo in response to brief SD (Del Cid-Pellitero et al., 2017). Homeostatic increases in GABAA receptor expression have recently been linked to changes in the localization of gephyrin, a scaffolding protein that anchors GABAA receptors to the inhibitory PSD (Flores et al., 2015). Flores et al. found that the number and size of gephyrin clusters increase in pyramidal neurons following prolonged network activity. These clusters colocalize with GAD67 and are accompanied by increases in miniature inhibitory post-synaptic current (mIPSC) amplitude and frequency in response to prolonged depolarization of pyramidal neurons. Recent data suggest that this mechanism may be directly affected by sleep vs. wake. For example, gephyrin mRNA levels are higher in the brain after a period of sleep relative to a period of wake (Cirelli et al., 2004). Gephyrin is stabilized at the synapse by phosphorylation by CDK5 (Kalbouneh et al., 2014), which as mentioned above may be activated preferentially during sleep (Diering et al., 2017).

Glutamatergic synapses on inhibitory interneurons may also be potentiated in response to increased network activity, leading to increased feedback inhibition within the network. Chang et al. found that network overactivity results in significantly increased expression of the immediate early gene narp and NARP protein in pyramidal neurons. The authors found that subsequently, NARP is released presynaptically in parvalbumin-expressing interneurons, causing increases in surface expression of GluA4 containing AMPA receptors (Chang et al., 2010). Narp expression is increased throughout the brain after a period of wakefulness (Cirelli et al., 2004). Given the differential regulation of narp and gephyrin expression by wakefulness/sleep, it is possible they maintain network stability by modulating inhibitory activity at different time points to alter E/I balance. Whether these pathways are evoked in vivo as a consequence of learning-associated synaptic potentiation is unknown. However, sleep-associated changes in the number of inhibitory synapses have been observed in the cortex, as described above (Del Cid-Pellitero et al., 2017). Taken together, there are numerous alternate pathways by which sleep could regulate homeostatic changes in neural circuits in response to augmented network activity.

Proponents of SHY have speculated that synaptic weakening is mechanistically linked to the synchronous, low-frequency rhythms (slow wave activity; SWA) that synchronize thalamocortical and hippocampo-cortical activity patterns during NREM sleep (Buzsaki et al., 2003; Sirota et al., 2003). They argue that, like synaptic strength, SWA is homeostatically regulated. With increased time spent awake (and according to the hypothesis, more opportunity for synaptic potentiation), SWA during subsequent NREM sleep is significantly enhanced. After an initial period of recovery sleep, this enhanced SWA returns to baseline—a process which is speculated to reflect a renormalization of synaptic strength to levels seen before waking experience. Thus, according to SHY proponents, SWA homeostasis and synaptic homeostasis go hand in hand. Beyond this, numerous studies have also indicated that NREM SWA is selectively enhanced in cortical areas that are preferentially activated (e.g., by learning) during prior wake periods. Conversely, SWA is selectively decreased in cortical areas that are less active during prior waking experience (Huber et al., 2004, 2006). In the context of SHY, this has been interpreted as evidence for a causal role of SWA thalamocortical activity patterns in promoting synaptic weakening.

There is evidence that experimentally-generated firing patterns (analogous to those occurring during SWA) can cause LTD of glutamatergic synapses in vitro. A variety of paradigms have been used to emulate the activity patterns seen in thalamocortical and hippocampal circuits during NREM. One of these is low frequency stimulation—trains of single spikes or short bursts, occurring at frequencies between 1 and 3 Hz. This rhythmic pattern of activity mimics that generally seen in both hippocampal and cortical circuits during NREM SWA in vivo. However, numerous labs have reported that low frequency stimulation (i.e., 1 Hz trains or burst stimuli, which can induce LTD of in vitro) is insufficient for in vivo LTD induction in either the hippocampus (Errington et al., 1995; Abraham et al., 1999) or cortex (Jiang et al., 2003; Hager and Dringenberg, 2010). In contrast, higher-frequency stimulation can reliably induce LTP in hippocampal and thalamocortical circuits in vivo (Heynen and Bear, 2001; Whitlock et al., 2006; Cooke and Bear, 2010).

It is unclear why many stimulation protocols induce LTD less robustly in vivo, while LTP is more easily induced. In neural circuits where it has been studied, the level of spontaneous activity (which varies with brain state) seems to be a critical variable for both LTD induction and maintenance. For example, LTD can be induced more reliably in the cortex in vivo if animals are deeply anesthetized (Hager and Dringenberg, 2010). This effect of anesthesia can be blocked by stimulation of the pedunculopontine (PPT) nucleus (which is wake-active, and provides cholinergic input to the thalamus) (Stewart and Dringenberg, 2016). Because PPT activity is generally low during NREM relative to wake (Jones, 2005; Mena-Segovia et al., 2008), and because spontaneous thalamocortical activity is generally lower in NREM than in REM or wake (Vyazovskiy et al., 2009), it is tempting to speculate that NREM sleep provides ideal (and necessary) state conditions for in vivo LTD induction. NREM thalamocortical activation patterns also provide another feature that might be ideally suited for inducing LTD—burst mode firing. Bursts of presynaptic action potentials paired with post-synaptic EPSPs reliably induce LTD of cortical glutamateric synapses in vitro (Birtoli and Ulrich, 2004; Czarnecki et al., 2007). Bursts of action potentials with no post-synaptic EPSPs may also reduce subsequent glutamatergic neurotransmission by driving elimination of post-synaptic calcium-permeable AMPA receptors (Lante et al., 2011). EPSP-paired bursting can elicit LTD at any time of day (after periods of more sleep or more wake), while unpaired bursting can elicit synaptic depression throughout the day. This suggests that at least two forms of activity-dependent LTD may be expressed at cortical synapses, and these are differentially affected by sleep history. Since these studies were carried out ex vivo, and in cortical slices taken from juvenile animals, future studies will have to address how these mechanisms are affected in vivo and into adulthood (when rates of synaptic pruning are generally reduced).

There is also evidence that over the long term (24 h, vs. minutes for inducing LTD), low-frequency stimulation may also activate the same cellular pathways involved in homeostatic synaptic downscaling. (Goold and Nicoll, 2010) recently demonstrated that prolonged optogenetic low-frequency stimulation of individual hippocampal neurons led to both cell-autonomous downscaling of NMDA and AMPA receptor-mediated currents, and dramatic synaptic pruning. These effects were mediated post-synaptically (i.e., in optogenetically-stimulated neurons) via CaMKK and CaMKIV, and removal of GluA2-containing AMPA receptors and NMDA receptors (Goold and Nicoll, 2010).

Despite these data, it is worth noting that NREM sleep is characterized by other network activity features in addition to SWA. In thalamocortical circuits, sleep spindles emerge as 7–15 Hz coherent network oscillations, which are expressed as discrete waxing-and-waning events during NREM (Clawson et al., 2016). Recent ex vivo studies have aimed at mimicking patterns of activity during NREM spindles to determine effects on synaptic strength. Rosanova and Ulrich recorded activity from neurons in somatosensory cortex during spindles, and used this pattern to drive presynaptic activity in layer 2/3 while recording post-synaptic responses in layer 5 (Rosanova and Ulrich, 2005). The authors found that when this pattern was repeated at a frequency similar to the frequency of NREM spindle occurrence, post-synaptic responses were potentiated. Moreover, a synthetic spindle activity pattern (presynaptic bursts delivered at 10 Hz) likewise drove post-synaptic LTP. Thus, NREM network oscillations of different frequencies may have divergent effects on synaptic strength in cortical circuits.

Proponents of SHY have emphasized the potential mechanistic link between NREM SWA and synaptic weakening. However, most measurements of molecular, functional and structural synaptic changes have been measured after periods of sleep, which includes REM. REM sleep constitutes roughly 10–30% of total sleep time in adult mammals, depending on species. Intriguingly, the proportion of time spent in REM sleep across species has been linked to brain mass (Lesku et al., 2006, 2008). Studies evaluating sleep time across phylogeny have not found a similar link between NREM sleep time and brain size. This begs the question—could REM, rather than NREM SWA, mediate synaptic weakening across intervals of sleep? There are some experimental data that would suggest that this is possible. Firing rates in both cortical (Durkin and Aton, 2016; Watson et al., 2016) and hippocampal neurons (Grosmark et al., 2012; Miyawaki and Diba, 2016) decrease consistently across bouts of REM. Firing decreases are proportional to REM bout duration in the cortex (Watson et al., 2016) and to the amplitude of locally-generated theta (4–12 Hz) oscillations in the hippocampus (Grosmark et al., 2012; Miyawaki and Diba, 2016). A recent fMRI study (van der Helm et al., 2011) indicated that overnight decreases in amygdala functional responses to an emotionally arousing task are related to REM-associated EEG activity. More recently, a study measuring overall levels of cortical neural activity (with wide-field imaging of calcium signals) found that activity is globally reduced in the cortex (in all cortical layers) across bouts of REM (Niethard et al., 2016). In support of the idea that these functional changes are related to synaptic weakening, a recent in vivo imaging study demonstrated that the selective elimination of newly-formed dendritic spines is blocked by REM-targeted SD (but not NREM disruption) (Li et al., 2017). What features of REM could mediate synaptic weakening? Recent calcium imaging data indicates that the relative activity of fast spiking interneurons to pyramidal neurons is significantly higher during REM relative to NREM and wake (Niethard et al., 2016). Thus, REM may alter the E/I balance of neural networks, which could bias plasticity at glutamatergic synapses, to favor synaptic weakening. Alternatively, the relative high levels of cortical and hippocampal acetylcholine release (and simultaneous relative low levels of norepinephrine, serotonin and dopamine release) during REM may bias circuit plasticity in favor of spike timing-based LTD (see above).

Multiple lines of evidence have indicated that the biological support system surrounding neurons is significantly affected by sleep and wake states. Recent studies focused on the so-called “glymphatic” system have shown that interstitial space in the cortex increases significantly during NREM sleep, over a timescale of minutes (Xie et al., 2013). This process, mediated by astrocytic regulation of peri-arterial flow rates, is thought to promote both delivery of nutrients, and clearance of potentially harmful metabolic waste from the brain. Such a mechanism could affect synaptic function in myriad ways. For example, levels of extracellular glucose decline across bouts of wake and REM, and increase at the transition from wake to NREM sleep (Dash et al., 2013). At the same time, lactate accumulates in the brain (as a product of glycolysis) during wake (and also during REM sleep) (Naylor et al., 2012) and is cleared by the glymphatic system during NREM sleep (Lundgaard et al., 2017). Because at high enough concentrations lactate can potentiate NMDA receptor-mediated currents, leading to downstream changes in the expression of plasticity-related genes in the brain (Yang, J. et al., 2014), this mechanism could potentially mediate sleep-dependent synaptic weakening.

Sleep changes not only the volume of the brain's interstitial space, but also its ionic content. A more recent study demonstrated that the extracellular concentrations of calcium, magnesium, and protons increases (and the concentration of potassium decreases) in the cortex as animals transition from wake to NREM sleep (Ding et al., 2016). As might be expected, such changes directly impact the mode of firing in cortical neurons (and ECoG activity), but it remains unclear whether they also directly impact synaptic function and synaptic strength. Available data suggests that changes in the concentrations of these ions (like those that accompany wake-to-sleep transitions) can lead either to a selective increase in excitatory transmission (resulting in increased E/I balance) (Meeks and Mennerick, 2004) or to synaptic potentiation (Harsanyi and Friedlander, 1997; Hess, 2002; Abumaria et al., 2011; Du et al., 2014). Intriguingly, these extracellular ion concentration changes could all be mediated by astrocytes (Verkhratsky and Nedergaard, 2014) and could in turn impact the activity pattern of surrounding neurons (Ding et al., 2016). Indeed, recent experimental data has shown that optogenetic hyperpolarization of astrocytes leads to changes in local field potential (LFP) activity similar to that seen as animals transition to NREM sleep (Poskanzer and Yuste, 2016).

While the precise cellular mechanisms underlying all of these effects are generally unknown, it is clear from studies using cell type-specific mRNA profiling (i.e., translating ribosome affinity purification; TRAP) that sleep and wake affect a variety of cellular processes in both astrocytes (Bellesi et al., 2015) and oligodendrocytes (Bellesi et al., 2013). One speculative mechanism is based on the fact that ATP released from neurons during heightened network activity activates purinergic receptors on microglia, leading to release of interleukin 1 (IL1) and tumor necrosis factor-alpha (TNFα) (Hide et al., 2000; Shieh et al., 2014). Because IL1 and TNFα can induce NREM sleep, this signaling mechanism has been hypothesized to mediate both sleep homeostasis after extended wake, and local, use-dependent changes in NREM thalamocortical oscillations (Kreuger et al., 2011). Some have speculated that this same signaling pathway may also mediate sleep-associated synaptic weakening. However, because the in vitro effects of TNFα on glutamatergic (Beattie et al., 2002; Karrer et al., 2015) and GABAergic (Stellwagen et al., 2005; Pribiag and Stellwagen, 2013) synapses are diverse, it remains unclear whether glial-derived TNFα signaling offers a plausible molecular mechanism for synaptic weakening during sleep.