Sleep Promotes Learning-Induced Synapse Formation in the Motor Cortex

Sleep is widely believed to be important for memory consolidation, but the underlying processes remain elusive. There are conflicting views as to whether non-REM sleep contributes to memory consolidation by either promoting or down-regulating synaptic plasticity (19–22, 29). By directly imaging postsynaptic dendritic spines over time in the mouse cortex, our results indicate that sleep after learning promotes new spine formation on different sets of apical tuft branches of individual layer V pyramidal neurons. Furthermore, this sleep-dependent, branch-specific spine formation facilitates new spine survival when animals learn different tasks. These findings suggest that sleep promotes learning-induced synapse formation to aid long-term memory storage.

We examined how sleep affects the remodeling of postsynaptic dendritic spines induced by motor learning in the mouse primary motor cortex. Rotarod motor learning increases dendritic spine formation on apical tuft dendrites of layer V pyramidal neurons in the motor cortex within 2 days (18, 36). To investigate whether sleep is involved in this process, we first determined the time course of spine remodeling in mice that were trained to run forward on an accelerated rotating rod. Yellow fluorescent protein (YFP)–labeled dendrites in the hind limb region of the motor cortex were imaged in awake head-restrained mice before and in the hours after training with transcranial two-photon microscopy (18, 37). The formation rate of new spines in trained mice was significantly higher within 6 hours after training and continued to increase within the first day when compared to that in untrained controls (P < 0.05) (Fig. 1, A and B). In contrast, rotarod training had no significant effect on the elimination rate of existing spines within 6 to 48 hours (Fig. 1C).

We observed that, 24 hours after motor training, only a fraction (~30%) of apical tuft branches (average branch length: 62.7 ± 1.3 μm) in trained mice showed a higher rate of spine formation than the branches in untrained mice (Fig. 1D and fig. S1). When spine formation on two sibling branches sharing the same parent branch was compared, the difference in spine formation, but not spine elimination, between sibling branches was also significantly larger in trained mice than in untrained controls (Fig. 1, D to F) (P < 0.0001 for spine formation; P = 0.52 for spine elimination) (fig. S2). To investigate this branch-specific spine formation further, we classified the sibling branch with higher spine formation as a ‘high-formation branch’ (HFB) and the other as a ‘low-formation branch’ (LFB) (Fig. 1G). Twenty-four hours after training, the average rate of spine formation on HFBs in trained mice (15.3 ± 1.3%) was 2.4 to 3.5 times that of HFBs (6.4 ± 0.8%) or LFBs (4.4 ± 0.9%) in untrained control mice (P < 0.0001) (Fig. 1H). The difference in spine formation between HFBs and LFBs was statistically larger for sibling branches than for randomly paired branches (P < 0.0001) (Fig. 1I). However, spine formation on LFBs in trained mice (5.2 ± 0.5%) was not significantly different from that on either HFBs (P = 0.19) or LFBs (P = 0.49) in untrained controls. There was also no significant difference in spine elimination between HFBs and LFBs in both trained (P = 0.15) and untrained animals (P > 0.9) (Fig. 1J).

Different motor learning tasks often activate the same neurons in the motor cortex (38). We wondered whether different learning tasks lead to spine formation on the same or different dendritic branches. To address this question, we trained mice to run forward and, 12 hours later, to run either forward or backward (Fig. 1, K and L). When mice were subjected to the second session of forward running 12 hours after the initial forward-running session, new spines formed during 0 to 12 hours and 12 to 24 hours tended to occur on the same set of branches, although the effect was not statistically significant (Fig. 1K). In contrast, running backwards induced spine formation on a set of branches that showed little formation of new spines in response to the previous forward running (Fig. 1L). Furthermore, when sibling branches were classified as HFBs and LFBs based on the degree of spine formation induced by the initial forward training, we found that backward running, not forward running or no training, induced spine formation mainly on the LFBs but not on the HFBs during the second 12 hours (Fig. 1M).

Our results thus far have revealed task- and branch-specific spine formation over the course of 24 hours after motor skill learning. To test a potential role of sleep in this process, we examined spine formation in mice that were subjected to rotarod training (one 40-trial session of forward running, ~1 hour) and then sleep deprived (SD) for 7 hours by gentle handling (Fig. 2A). Electroencephalography (EEG) monitoring over 7 hours showed that SD mice were awake 97.0 ± 2.1% of the time, whereas mice with undisturbed sleep (non-SD) were awake only 26.4 ± 2.9% of the time (P < 0.05) (Fig. 2, B and C). There was a significant reduction in learning-induced spine formation over the entire 8 hours in SD mice when compared to non-SD mice (Fig. 2D). Sleep deprivation specifically reduced spine formation on HFBs (4.9 ± 0.7% versus 9.3 ± 0.7%; P < 0.0001), but not on LFBs (2.4 ± 0.4% versus 1.8 ± 0.4%; P = 0.16). To investigate whether the effect of sleep deprivation on spine formation might be stress-related, we administered the stress hormone corticosterone (2.5 mg/kg) to non-SD mice after motor training (fig. S3). Corticosterone administration had no significant effects on spine formation on either HFBs or LFBs in the course of 8 hours (Fig. 2D), which suggested that the elevation of stress hormones associated with sleep deprivation is not important for the reduction in spine formation after learning.

To better understand the importance of sleep in dendritic spine formation, we tested whether the reduced spine formation after sleep deprivation could be compensated for by additional training. Although spine formation on HFBs was significantly higher with intensive training (two 40-trial sessions) than with regular training (one 40-trial session) or no training in SD mice (P < 0.05) (Fig. 2, A and D), it remained significantly lower than in non-SD mice with regular training (P < 0.05). There was no significant difference in spine formation on LFBs among all five groups [P = 0.35, one-way analysis of variance (ANOVA)] (Fig. 2D). We also tested whether the reduction in spine formation could be compensated for by subsequent sleep by allowing animals to sleep during the next 16 hours after the initial 7-hour sleep deprivation (Fig. 2A). Over the subsequent 16 hours, the rate of spine formation on either HFBs or LFBs was found to be significantly lower in SD mice than non-SD mice (P < 0.05) (Fig. 2E and fig. S4). Thus, the reduction in spine formation after the 7-hour sleep deprivation could not be rescued by either an additional training session or subsequent sleep.

A fraction of learning-induced new spines persists over time, and the number of persisting new spines correlates with long-term retention of motor skills (18, 36). We followed the fate of all new spines that were formed during 8 hours with or without posttraining sleep (Fig. 3A). The survival of new spines on HFBs was significantly higher during the next day in mice with sleep after learning than without (P < 0.05) (Fig. 3B). In contrast, the survival of new spines on LFBs was not significantly different between mice with and without sleep (P = 0.97) (Fig. 3B). The performance improvement in mice with posttraining sleep, when tested 1 or 5 days after the initial training, was significantly larger when compared to that of SD mice (P < 0.05) (Fig. 3C and fig. S5). These results suggest that sleep contributes significantly to the formation of persistent new spines on HFBs, as well as motor skill retention.

Previous studies have shown that the survival of new spines is modulated by subsequent experiences (18, 36). To better understand the persistence of new spines formed during postlearning sleep, we examined how new spines induced by forward running are affected by subsequent motor learning experiences (Fig. 3D). The survival rate of new spines on HFBs was significantly higher when animals were trained again with the forward-running task than when animals were not trained or were subjected to backward running (Fig. 3E). Notably, the survival rate of new spines on LFBs was significantly lower in mice subjected to backward running when compared with mice subjected to either forward training or no training (P < 0.01) (Fig. 3, D and E). This reduction in new dendritic spine survival on LFBs could be related to the fact that backward training tended to promote new spine formation on LFBs (Fig. 1M). Because the majority (78%) of total new spines were formed on HFBs after forward running, the persistence of the total new spines induced by forward running was not significantly affected after backward running (Fig. 3E). The persistence of new spines formed during postlearning sleep may underlie a well-known feature of motor skill learning that, once a skill is learned, it persists for long periods of time with minimum interference by other learning tasks.

How does sleep promote branch-specific spine formation after learning? Sleep consists of two basic states, rapid eye movement (REM) sleep and non-REM (NREM) sleep. To explore the mechanisms underlying sleep-dependent spine formation, we first examined whether REM sleep is required for spine formation after rotarod learning. Mice were subjected to rotarod training (40 trials, ~1 hour) and deprived of REM sleep (REMD) for 7 hours (Fig. 4A). REM sleep was monitored continuously by EEG and electromyography (EMG) recordings and disrupted by gentle touching upon detection. EEG and EMG monitoring in the course of 7 hours showed that REM sleep in REMD mice was significantly reduced when compared to control mice (6.9 ± 1.1 min versus 32.1 ± 4.0 min; P < 0.01) (Figs. 2C and 4A). REM deprivation during 7 hours did not disrupt branch-specific spine formation induced by learning (Fig. 4B). Similar to mice with undisturbed sleep, spine formation during 8 hours after training was ~3.1 times as much on HFBs as on LFBs in REMD mice.

Neurons associated with wakeful experience are reactivated in multiple brain regions during subsequent NREM sleep, and this sleep reactivation occurs after the prior wakeful experience (6–11). Because neuronal activity is critical for regulating synaptic plasticity, neuronal reactivation during NREM sleep could be involved in promoting spine formation. We therefore examined whether motor task–related neurons are reactivated in the primary motor cortex during NREM sleep by performing calcium imaging of layer V pyramidal neurons expressing the genetically encoded calcium indicator GCaMP6 (39) (Fig. 4, C and D) (see methods). In this experiment, head-restrained mice were trained to run on a custom-built treadmill under a two-photon microscope. We found that, similar to rotarod motor learning, forward and backward running on the treadmill induced branch-specific spine formation in the course of 8 hours (fig. S6). Many layer V pyramidal neurons showed increased activity, as indicated by elevated levels of Ca2+ in cell somata, during forward running on the treadmill as compared to a state of quiet wakefulness (Fig. 4E). Over the 5-min recording period, ~41% (250 out of 617) of neurons showed a large increase (>50%) in somatic Ca2+ level (ΔFrunning/ΔFquiet > 1.5) and ~39% (242 out of 617) of neurons showed no or moderate increase (ΔFrunning/ΔFquiet = 1.0–1.5). When the same neurons were followed over the next 8 hours, neurons with >50% increase in somatic Ca2+ during running (ΔFrunning/ΔFquiet > 1.5, defined as task-related neurons) also showed higher levels of somatic Ca2+ during NREM sleep when compared to that under the quiet awake state (P < 0.0001) (Fig. 4F). To rule out the possibility that certain neurons active during postrunning sleep were not task-related, we removed neurons highly active during prerunning sleep from the analysis of sleep reactivation during postrunning sleep (fig. S7). We found that neurons highly activated during forward running but not during prerunning sleep (ΔFrunning/ΔFquiet > 1.5; ΔFprerun sleep/ΔFquiet < 1.5) were reactivated during the postrunning sleep episode (Fig. 4F). In contrast, neurons with no or moderate increase (<50%) in somatic Ca2+ level during running did not show a significant increase of Ca2+ activity during NREM sleep. These observations are consistent with previous electrophysiological studies of sleep replay in several brain regions (6–11) and suggest that neuronal reactivation of prior motor experience also occurs in the motor cortex during extended periods of time (>4 hours).

To test whether neuronal reactivation might be involved in branch-specific spine formation, we first blocked N-methyl-d-aspartate (NMDA) receptors with MK801 and examined branch-specific spine formation. MK801 (0.25 mg/kg) injection after training significantly reduced the activity of forward running–related neurons during NREM sleep within 8 hours after training (P < 0.001) (Fig. 4F). MK801 administration also blocked branch-specific spine formation after training (P < 0.0001) (Fig. 4H).

MK801 not only reduces neuronal activity during sleep but also alters the animals’ locomotion behavior in the first few hours after drug administration (40). Therefore, the effect of MK801 on spine formation may not be specifically related to altered neuronal activity during sleep. To manipulate the extent of neuronal reactivation more specifically, we took advantage of the findings that sleep reactivation is related to prior wakeful experience. We trained mice to run forward and allowed them to sleep for 4 hours. Subsequently, mice either received no further training (F-N) or were trained to run forward (F-F) or backward (F-B) (Fig. 4G). During the second 4-hour sleep period, the reactivation of neurons specific to forward running in the F-B group was significantly reduced when compared to neurons specific to backward-running or neurons activated during both forward and backward running in the same F-B group (P < 0.01) (Fig. 4G). The reactivation of neurons specific to forward running in the F-B group was also significantly less than neurons activated during forward running in the F-F and F-N groups (P < 0.01) (Fig. 4G). Notably, when spine formation on sibling branches was examined over the course of 8 hours, the rate of spine formation on HFBs was significantly reduced in the F-B group when compared to the F-F or F-N group (P < 0.05) (Fig. 4H). The ratio of spine formation rates between HFBs and LFBs was 1.8 in the F-B group, substantially lower than 3.5 and 5.6 in the F-F and F-N groups, respectively. Because all three experimental groups experienced a similar amount of sleep but differed in the extent of neuronal reactivation associated with forward training, these results provide further evidence for the role of sleep reactivation in branch-specific spine formation.

Sleep is widely believed to be important for memory consolidation, but the underlying processes remain elusive. There are conflicting views as to whether non-REM sleep contributes to memory consolidation by either promoting or down-regulating synaptic plasticity (19–22, 29). By directly imaging postsynaptic dendritic spines over time in the mouse cortex, our results indicate that sleep after learning promotes new spine formation on different sets of apical tuft branches of individual layer V pyramidal neurons. Furthermore, this sleep-dependent, branch-specific spine formation facilitates new spine survival when animals learn different tasks. These findings suggest that sleep promotes learning-induced synapse formation to aid long-term memory storage.