Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

The memory function of sleep

1,441 views

Published on

  • Be the first to comment

The memory function of sleep

  1. 1. REVIEWS sleep The memory function of sleep Susanne Diekelmann and Jan Born Abstract | Sleep has been identified as a state that optimizes the consolidation of newly acquired information in memory, depending on the specific conditions of learning and the timing of sleep. Consolidation during sleep promotes both quantitative and qualitative changes of memory representations. Through specific patterns of neuromodulatory activity and electric field potential oscillations, slow-wave sleep (SWS) and rapid eye movement (REM) sleep support system consolidation and synaptic consolidation, respectively. During SWS, slow oscillations, spindles and ripples — at minimum cholinergic activity — coordinate the re-activation and redistribution of hippocampus-dependent memories to neocortical sites, whereas during REM sleep, local increases in plasticity-related immediate-early gene activity — at high cholinergic and theta activity — might favour the subsequent synaptic consolidation of memories in the cortex.Declarative memory Although sleep is a systems-level process that affects finish by comparing two hypotheses that might explainMemories that are accessible the whole organism, its most distinctive features are the sleep-dependent memory consolidation on a mechanis-to conscious recollection loss of behavioural control and consciousness. Among tic level, that is, the synaptic homeostasis hypothesis andincluding memories for facts the multiple functions of sleep1, its role in the establish- the active system consolidation hypothesis.and episodes, for example,learning vocabulary or ment of memories seems to be particularly important:remembering events. as it seems to be incompatible with the brain’s normal Behavioural studiesDeclarative memories rely on processing of stimuli during waking, it might explain the Numerous studies have confirmed the beneficial effectthe hippocampus and loss of consciousness in sleep. Sleep promotes primarily of sleep on declarative and procedural memory in variousassociated medial temporal the consolidation of memory, whereas memory encoding tasks8–10, with practically no evidence for the oppositelobe structures, together withneocortical regions for and retrieval take place most effectively during waking. effect (sleep promoting forgetting)11. Compared with along-term storage. Consolidation refers to a process that transforms new wake interval of equal length, a period of post-learning and initially labile memories encoded in the awake state sleep enhances retention of declarative information3,12–16Procedural memory into more stable representations that become integrated and improves performance in procedural skills13,17–24.Memories for skills that resultfrom repeated practice and into the network of pre-existing long-term memories. Sleep likewise supports the consolidation of emotionalare not necessarily available Consolidation involves the active re-processing of ‘fresh’ information25–27. Effects of a 3-hour period of sleep onfor conscious recollection, for memories within the neuronal networks that were used emotional memory were even detectable 4 years later 28.example, riding a bike or for encoding them. It seems to occur most effectively However, the consolidating effect of sleep is not revealedplaying the piano. Procedural off-line, i.e. during sleep, so that encoding and consoli- under all circumstances and seems to be associated withmemories rely on the striatumand cerebellum, although dation cannot disturb each other and the brain does not specific conditions29 (see below).recent studies indicate that the ‘hallucinate’ during consolidation2.hippocampus can also be The hypothesis that sleep favours memory consolida- Sleep duration and timing. Significant sleep benefitsimplicated in procedural learning. tion has been around for a long time3. Recent research on memory are observed after an 8-hour night of sleep,University of Lübeck, in this field has provided important insights into the but also after shorter naps of 1–2 hours14,19,23,30, and evenDepartment of underlying mechanisms through which sleep serves an ultra-short nap of 6 minutes can improve memoryNeuroendocrinology, memory consolidation4–7. In this Review, we first discuss retention16. However, longer sleep durations yield greaterHaus 50, 2. OG, RatzeburgerAllee 160, 23538 Lübeck, findings from behavioural studies regarding the specific improvements, particularly for procedural memo-Germany. conditions that determine the access of a freshly encoded ries18,21,31. The optimal amount of sleep needed to benefitCorrespondence to J. B. memory to sleep-dependent consolidation, and regard- memory and how this might generalize across speciese‑mail: ing the way in which sleep quantitatively and qualitatively showing different sleep durations is unclear at present.born@kfg.uni‑luebeck.dedoi:10.1038/nrn2762 changes new memory representations. We then consider Some data suggest that a short delay betweenPublished online the role of slow-wave sleep (SWS) and rapid eye move- learning and sleep optimizes the benefits of sleep on4 January 2010 ment (REM) sleep in memory consolidation (BOX 1). We memory consolidation. For example, for declarative114 | FEbRuARy 2010 | VoluME 11 www.nature.com/reviews/neuro © 2010 Macmillan Publishers Limited. All rights reserved
  2. 2. REVIEWSSerial reaction time task information, sleep occurring 3 hours after learning finger sequence tapping task, which involves explicitA task in which subjects are was more effective than sleep delayed by more than procedural memory 17–19,24. For the serial reaction timerequired to rapidly respond to 10 hours32,33. However, these studies did not control for task (SRTT), which can be learnt implicitly or explicitly,different spatial cues by the confounding effects of forgetting during the wake the sleep-induced speeding of performance was morepressing correspondingbuttons. This task can be interval before the onset of sleep. For optimal benefit robust when people learnt the task explicitly thanperformed implicitly (that is, on procedural memory consolidation, sleep does not after implicit learning 34. These observations suggestwithout knowledge that there is need to occur immediately 18,19 but should happen on that explicit encoding of a memory favours access toa regularity underlying the the same day as initial training 17,22,24. sleep-dependent consolidation.sequence of cue positions) or The benefit of sleep is greater for memories formedexplicitly (by informing thesubject about this underlying Explicit versus implicit encoding. Whether memories from explicitly encoded information that was more dif-regularity). gain access to sleep-dependent consolidation depends ficult to encode or that was only weakly encoded35,36, and on the conditions of encoding. Encoding of declara- it is greater for memories that were behaviourally relevant. tive memories is typically explicit, whereas proce- Thus, sleep enhances the consolidation of memories for dural memory encoding can involve both implicit intended future actions and plans (D. S., I. Wilhelm, u. and explicit processes. Most robust and reliable sleep- Wagner, J. b., unpublished observations). Notably, this dependent gains in speed have been revealed for the enhancement could be nullified by letting the subject Box 1 | sleep architecture and neurophysiological characteristics of sleep stages Sleep is characterized by the cyclic occurrence of rapid a eye movement (REM) sleep and non-REM sleep, which Wake includes slow wave sleep (SWS, stages 3 and 4) and lighter REM sleep sleep stages 1 and 2 (see the figure, part a). In humans, the REM first part of the night (early sleep) is characterized by high amounts of SWS, whereas REM sleep prevails during the Stage 1 second half (late sleep). SWS and REM sleep are Stage 2 characterized by specific patterns of electrical field potential oscillations (part b) and neuromodulator activity Stage 3 (part c, BOX 3). SWS The most prominent field potential oscillations during SWS Stage 4 are the slow oscillations, spindles and sharp wave-ripples, Early sleep Late sleep whereas REM sleep is characterized by ponto-geniculo- occipital (PGO) waves and theta activity. The slow oscillations 23:00 0:00 1:00 2:00 3:00 4:00 5:00 6:00 7:00 originate in the neocortex with a peak frequency (in humans) Hours of ~0.8 Hz130,164. They synchronize neuronal activity into down-states of widespread hyperpolarization and neuronal b Field potential oscillations silence and subsequent up-states, which are associated with depolarization and strongly increased, wake-like neuronal Slow oscillation Spindle Sharp wave-ripple PGO wave Theta activity firing132,165,166 (part d). The hyperpolarization results from activation of a Ca -dependent K current and inactivation of 2+ + a persistent Na+ current, which dampens excitability165,167,168. The depolarizing up-state might be triggered by summation of miniature EPSPs (from residual activity from encoding information) and is formed by activation of T-type Ca2+ and c Neuromodulators persistent Na+ currents. Spindle activity refers to regular electroencephalographic Acetylcholine Acetylcholine oscillations of ~10–15 Hz, which are observed in human sleep stage 2 as discrete waxing and waning spindles, but are present Noradrenaline/ Noradrenaline/ serotonin serotonin at a similar level during SWS (although here they form less discrete spindles)169. Spindles originate in the thalamus from an Cortisol Cortisol interaction between GABAergic neurons of the nucleus reticularis, which function as pacemakers, and glutamatergic d Slow oscillations thalamo-cortical projections that mediate their synchronized and widespread propagation to cortical regions 132,168,169 . Hippocampal sharp waves are fast depolarizing events, generated in the CA3, on which high-frequency oscillations (100–300 Hz) originating from an interaction between inhibitory Field potential interneurons and pyramidal cells in CA1 (so-called ripples) are superimposed104,121. Sharp wave-ripples occur during SWS and also during waking, and accompany the re-activation Up-state of neuron ensembles that are active during a preceding wake experience70,71,121,122,170. PGO-waves are driven by intense bursts of synchronized activity that propagate from the pontine brainstem mainly to the lateral geniculate nucleus and visual cortex. They occur in temporal association with REM in rats and cats but are not reliably identified in humans. Theta oscillations (4–8 Hz) hallmark Single cell recording tonic REM sleep in rats and predominate in the hippocampus141. In humans, theta activity is less Down-state coherent144,145. 1s Nature Reviews | NeuroscienceNATuRE REVIEWS | NeuroscieNce VoluME 11 | FEbRuARy 2010 | 115 © 2010 Macmillan Publishers Limited. All rights reserved
  3. 3. REVIEWS execute the intended behaviour before sleep. Similarly, Interacting or competing memory systems? The behav- subjects who had been trained on two different finger- ioural findings described above show that sleep can tapping sequences showed greater sleep-dependent gains ‘re-organize’ newly encoded memory representations, in performance for the sequence for which they expected enabling the generation of new associations and the to be rewarded for optimal performance at re-testing after extraction of invariant features from complex stimuli, sleep37. Thus, a motivational tagging of memories, which and thereby eventually easing novel inferences and probably relies on the function of the prefrontal cortex38, insights. Re-organization of memory representations might signal behavioural effort and relevance and mediate during sleep also promotes the transformation of the preferential consolidation of these memories. implicit into explicit knowledge, as was shown in an In summary, a great number of studies indicate that SRTT which was implicitly trained but in which explicit sleep supports the consolidation of memory in all major knowledge about the underlying sequence was exam- memory systems, but preferentially those that are explicitly ined during the re-test 48. Following post-training sleep, encoded and that have behavioural relevance to the indi- subjects were better at explicitly generating the SRTT vidual. There is growing evidence that explicit encoding, sequence. Interestingly, subjects who developed explicit even in procedural tasks, involves a dialogue between sequence knowledge no longer showed the improve- the prefrontal cortex and the hippocampus38–40, which ment in implicit procedural skill (that is, faster reaction also integrates intentional and motivational aspects of times) that is normally observed after sleep, suggesting the task. Activity of this circuit may be crucial in mak- that procedural and declarative memory systems interact ing a memory susceptible to sleep-dependent memory during sleep-dependent consolidation. consolidation. Contrasting with this view of interacting memory systems, it has also been proposed that disengagement Sleep changes memory representations quantitatively of memory systems is an essential characteristic of sleep- and qualitatively. Consolidation of memory during dependent consolidation49. This idea derives mainly sleep can produce a strengthening of associations as from experiments showing that declarative learning of well as qualitative changes in memory representations. words immediately after training of a procedural skill Strengthening of a memory behaviourally expresses itself can block off-line improvement in that skill if the subject as resistance to interference from another similar task does not sleep between learning and re-testing, but not if (‘stabilization’) and as an improvement of performance the subject sleeps between learning and re-testing 50. This (‘enhancement’) that occurs at re-testing, in the absence suggests that memory systems compete and reciprocally of additional practice during the retention interval. interfere during waking, but disengage during sleep, The stabilizing effects of sleep have been observed in allowing for the independent consolidation of memories declarative41 and procedural19 memory tasks. Similarly, in different systems. The two views might be reconciled enhancements in performance after sleep have been by assuming a sequential contribution of interaction and shown for declarative information13,14,20 and in proce- disengagement processes to consolidation, which might dural tasks13,17,18,21,22,31. However, it is still controversial be associated with different sleep stages (REM sleep and to what extent these improvements reflect actual per- SWS), as discussed below. formance ‘gains’ induced by sleep, because the measured gains depend on the pre-sleep performance used as a Influence of sleep stages on consolidation reference, which itself can be subject to rapid changes Early studies in rats and humans investigating whether after training 42,43. different sleep stages have different roles in memory There is a long-standing debate about whether sleep consolidation mainly focused on REM sleep and the passively protects memories from decay and interfer- consequences of REM sleep deprivation (REMD) by ence or actively consolidates fresh memory represen- repeatedly waking subjects at the first signs of REM tations44 (for a review see Ref. 45). Importantly, a lack sleep. However, this approach is of limited value for logi- of enhancement of memory performance after sleep cal reasons and because the repeated awakenings cause does not preclude an active role of sleep in memory stress, which itself influences memory function51,52.Implicit learning consolidation. There is strong evidence for an active con- overall, these studies have provided mixed results52–55.Learning without being aware solidating influence of sleep from behavioural studies, of note is a recent study showing that pharmacologi-that something is being which indicate that sleep can lead to qualitative changes cal suppression of REM sleep by administration of anti-learned. in memory 46–48. For example, in one study, subjects depressant drugs (selective noradrenaline or serotoninExplicit learning learned single relations between different objects which, re-uptake inhibitors) did not impair consolidation ofLearning while being aware unknown to the subject, relied on an embedded hierar- procedural memory 56, which is in agreement with clini-that something is being chy 47. When learning was followed by sleep, subjects at a cal observations that antidepressant treatment does notlearned. re-test were better at inferring the relationship between affect memory function57. However, such substances alsoMemory systems the most distant objects, which had not been learned exert direct effects on synaptic plasticity and synapticDifferent types of memory, before. likewise, after sleep subjects more easily solved forms of consolidation that could compensate for a losssuch as declarative and a logical calculus problem that they were unable to solve of REM sleep58.non-declarative memory, are before sleep or after corresponding intervals of wakeful- Some studies performed in rats showed that REMDthought to be mediated bydistinct neural systems, the ness46. of note, sleep facilitated the gain of insight into is only effective during specific periods after learningorganization of which is still a the problem only if adequate encoding of the task was — the so-called ‘REM sleep windows’54. During post-topic of debate. ensured before sleep. learning sleep, increases in the amount and intensity116 | FEbRuARy 2010 | VoluME 11 www.nature.com/reviews/neuro © 2010 Macmillan Publishers Limited. All rights reserved
  4. 4. REVIEWS of REM sleep occur several hours or even days after place off-line after encoding relies on the re-activation learning, depending on the kind of task and amount of of neuronal circuits that were implicated in the encod- initial training 54, and memory is particularly impaired if ing of the information. This would promote both the REMD coincides with these periods. of note, the mem- gradual redistribution and re-organization of memory ory tasks used in rats are typically emotionally loaded. As representations to sites for long-term storage (that is, there is evidence that REM sleep preferentially benefits system consolidation; BOX 2) and the enduring synaptic the consolidation of emotional aspects of a memory 25,27, changes that are necessary to stabilize memories (syn- this could partly account for the strong REMD effect aptic consolidation). The conditions that enable these observed in many animal studies53,55. two processes during sleep differ strongly between SWS Studies in humans have compared the effects on and REM sleep. consolidation between sleep periods with different proportions of SWS and REM sleep. In humans, SWS Re-activation of memory traces during sleep. The finding and REM sleep dominate the early and late part of noc- that in rats the spatio-temporal patterns of neuronal turnal sleep, respectively (BOX 1). SWS-rich, early sleep firing that occur in the hippocampus during explora- consistently benefits the consolidation of declarative tion of a novel environment or simple spatial tasks are memories12,13,59, whereas REM-rich sleep benefits non- re-activated in the same order during subsequent sleep declarative types of memory (that is, procedural and was an important breakthrough in memory research69–74 emotional aspects of memory)13,25,59. These results are con- (fIG. 1a, see Ref. 75 for methodological considerations sistent with the ‘dual-process hypothesis’, which assumes on the identification of neuronal re-activations). Such that SWS facilitates declarative, hippocampus-depend- neuronal re-activation of ensemble activity mostly ent memory and REM sleep supports non-declarative, occurs during SWS (it is rarely observed during REM hippocampus-independent memory 6. sleep76,77) and during the first hours after learning (but other studies have shown that SWS can also see Ref. 78), and typically only in a minority of recorded improve procedural skill (that is, non-declarative) neurons69–74. Moreover, unlike re-activations that occur memories31,60,61 and that REM sleep can also improve during wakefulness, re-activations during SWS almost declarative memory 62,63. Although these divergent find- always occur in the order in which they were expe- ings could reflect that stimuli used in memory tasks are rienced79. Compared with activity during encoding often not of one type of memory system, they agree phases, re-activations during SWS seem to be noisier, with the ‘sequential hypothesis’, which argues that the less accurate and often happen at a faster firing rate71. optimum benefits of sleep on the consolidation of both They are also observed in the thalamus, the striatum declarative and non-declarative memory occur when and the neocortex 72–74,78. Sleep-dependent signs of re- SWS and REM sleep take place in succession31,64. Thus, activation in brain regions implicated in prior learning overnight improvements in visual texture discrimina- were also shown in human neuroimaging studies80,81. tion correlated with both the amount of SWS in the The first evidence for a causal role of re-activation first quarter of sleep and the amount of REM sleep in during SWS in memory consolidation came from a study the last quarter 21. Texture discrimination also improved in humans learning spatial locations in the presence of an following a short midday nap of 60–90 minutes con- odour 15. Re-exposure to the odour during SWS, but not taining solely SWS, but more so if the nap included REM sleep, enhanced the spatial memories (fIG. 1b) and both SWS and REM sleep23. Also, memory consolida- induced stronger hippocampal activation than during tion seems to be impaired by disruptions of the natural wakefulness, indicating that during SWS hippocampal SWS–REM sleep cycle that left the time spent in these networks are particularly sensitive to inputs that can sleep stages unchanged65. re-activate memories (fIG. 1c). It is assumed that the re- Intermediate sleep stages (non-REM sleep stage 2 in activations during system consolidation stimulate the humans, transitory sleep in rats) can also contribute to redistribution of hippocampal memories to neocorti- memory consolidation66,67. For example, pharmaco- cal storage sites, although this has not been directly logical suppression of REM sleep in humans produced demonstrated yet 82,83. an unexpected overnight improvement in procedural skill that was correlated with increased non-REM sleep Synaptic consolidation. In addition to system consoli- stage 2 spindle activity (see below)56. Such findings high- dation (BOX 2), consolidation involves the strengthening light the fact that it is not a particular sleep stage per se of memory representations at the synaptic level (syn- that mediates memory consolidation, but rather the aptic consolidation)84,85. long-term potentiation (lTP) neurophysiological mechanisms associated with those is considered a key mechanism of synaptic consolida- sleep stages, and that some of these mechanisms are tion, but it is unclear whether memory re-activation shared by different sleep stages. during sleep promotes the redistribution of memories by inducing new lTP (at long-term storage sites not Core features of off-line consolidation involved at encoding) or whether re-activation merelyTransitory sleep Since the publication of Hebb’s seminal book68, memory enhances the maintenance of lTP that was inducedShort transitory periods of formation has been conceptualized as a process in which during encoding.sleep in rats that, based oneeG criteria, can neither be neuronal activity reverberating in specific circuits pro- lTP can be induced in the hippocampus duringclassified as ReM sleep or motes enduring synaptic changes. building on this, it is REM sleep but less reliably so during SWS86. lTP induc-SWS. widely accepted that the consolidation process that takes tion in the hippocampus or neocortex during SWS isNATuRE REVIEWS | NeuroscieNce VoluME 11 | FEbRuARy 2010 | 117 © 2010 Macmillan Publishers Limited. All rights reserved
  5. 5. REVIEWSImmediate early genes Box 2 | The two-stage model of memory consolidationGenes that encodetranscription factors that are A key issue of long-term memory formation, the Long-term store (slow learning)induced within minutes of so-called stability–plasticity dilemma, is the problemraised neuronal activity without of how the brain’s neuronal networks can acquire newrequiring a protein signal. information (plasticity) without overriding olderImmediate-early gene Encoding knowledge (stability). Many aspects of eventsactivation is, therefore, used as experienced during waking represent unique andan indirect marker of neuronal irrelevant information that does not need to be storedactivation. The immediateearly genes Arc and Egr1 long term. The two-stage model of memory offers a(zif268) are associated with widely accepted solution to this dilemma2,7,85,152 (seesynaptic plasticity. the figure). The model assumes two separate memory stores: one store allows learning at a fast rate and Encoding ConsolidationHebbian plasticity serves as an intermediate buffer that holds theRefers to the functional information only temporarily; the other store learns atchanges at synapses that a slower rate and serves as the long-term store.increase the efficacy of Initially, new events are encoded in parallel in bothsynaptic transmission and stores. In subsequent periods of consolidation, theoccurs when the presynapticneuron repeatedly and newly encoded memory traces are repeatedlypersistently stimulates the re-activated in the fast-learning store, which drivespostsynaptic neuron. concurrent re-activation in the slow-learning store, Temporary store (fast learning) and thereby new memories become graduallySpike-time dependent redistributed such that representations in the slow-learning, long-term store are strengthened. Through the Nature Reviews | Neuroscienceplasticity repeated re-activation of new memories, in conjunction with related and similar older memories, the fast-learningRefers to the functional store acts like an internal ‘trainer’ of the slow-learning store to gradually adapt the new memories to thechanges at synapses that alter pre-existing network of long-term memories. This process also promotes the extraction of invariant repeatingthe efficacy of synaptic features from the new memories. As both stores are used for encoding information, in order to prevent interference,transmission depending on therelative timing of pre- and the re-activation and redistribution of memories take place off-line (during sleep) when no encoding occurs.postsynaptic firing (‘spiking’). Because in this model consolidation involves the redistribution of representations between different neuronalThe synaptic connection is systems that is, the fast- and slow-learning stores, it has been termed ‘system consolidation’. For declarativestrengthened if the presynaptic memories, the fast- and slow-learning stores are represented by the hippocampus and neocortex, respectively.neuron fires shortly before the Figure modified, with permission, from Ref. 85 © (2005) Macmillan Publishers Ltd. All rights reserved.postsynaptic neuron, but isweakened if the sequence offiring is reversed. probably temporally restricted to the up-states of the plasticity depends on the activation of glutamatergic slow oscillation and its concurrent phenomena of rip- NMDA (N-methyl-d-aspartate) receptors and associ- ples and spindles87,88 (see BOX 1 and below). Indeed, in ated cAMP-dependent protein kinase A (PKA), and neocortical slices, stimulation that mimicked neuronal on AMPA (α-amino-3-hydroxy-5-methyl-4-isoxazole activity during SWS could induce long-term depression propionic acid) receptor activation, that is, the post- (lTD)89 or lTP87 depending on the pattern of stimula- synaptic machinery that is crucial for the induction and tion (rhythmic bursts or spindle-like trains, respectively). maintenance of lTP99–102. These findings indicate that lTP maintenance in the rat hippocampus, but not in the local, off-line re-activation of specific glutamatergic medial prefrontal cortex, was impaired if induction was circuits supports both lTP induction and maintenance, followed by REMD90. In humans, sleep strengthened and the molecular processes underlying synaptic con- lTP-like plasticity that had been induced in the neocor- solidation. Moreover, these processes probably occur tex by transcranial magnetic stimulation (TMS) prior preferentially during REM sleep, although they are to sleeping 91. likely to be triggered by the re-activations that occur Globally (meaning measured in whole-brain or large during prior SWS (see below). Evidence about how lTP cortical samples) sleep suppresses the molecular sig- induction and maintenance is linked to specific sleep nals that mediate lTP-related synaptic remodelling but stages is presently scarce, but based on the available enhances lTD-related signalling, and this effect seems data it is tempting to speculate that SWS supports the to be mediated by SWS92–95. This observation, however, re-activation of new memories (system consolidation) does not preclude that lTP occurs during sleep (during and thus, could initialize lTP and prime the relevant SWS or REM sleep) in specific regions, for example in networks for synaptic consolidation during subsequent those that were engaged in memory encoding prior to REM sleep. This idea seems to be supported by elec- sleeping. In rats, both induction of hippocampal lTP troencephalographic (EEG) rhythms that characterize and exposure to a novel tactile experience during wak- these sleep stages. ing increased the expression of the plasticity-related immediate early genes (IEGs) Arc and Egr1 (which are sleep-specific field potential oscillations implicated in lTP) during subsequent sleep, mainly in Sleep stages are characterized by specific electrical cortical areas that were the most activated by the novel field potential rhythms that temporally coordinate experience, and this effect seemed to be mediated by information transfer between brain regions and REM sleep96–98. Investigations in visual cortex in cats might support Hebbian and spike-time-dependent and humans have demonstrated that sleep-dependent plasticity103,104.118 | FEbRuARy 2010 | VoluME 11 www.nature.com/reviews/neuro © 2010 Macmillan Publishers Limited. All rights reserved
  6. 6. REVIEWS a Cortex Hippocampus Field potentials associated with SWS. Neocortical slow Run Sleep Run Sleep oscillations, thalamo-cortical spindles and hippocampal 7 5 ripples have been associated with memory consolidation during SWS (BOX 1). The neocortical slow oscillations (of <1 Hz), by globally inducing up- and down-states of neu-Cell number Cell number 0 0 ronal activity, are thought to provide a supra-ordinate 7 5 temporal frame for the dialogue between the neocortex and subcortical structures that is necessary for redistrib- uting memories for long-term storage8,105,106. The ampli- 0 0 1s 0.5 s 1s 0.2 s tude and slope of the slow oscillations are increased when SWS is preceded by specific learning experi-b ences60,107,108 and decreased when the encoding of infor- Learning Sleep Odour/vehicle Retrieval mation was prevented109. These changes occur locally, in the cortical regions that were involved in encoding, and Wake can also be induced in humans by potentiating synap- REM Stage 1 tic circuits through TMS91,110,111. Inducing slow oscilla- Stage 2 tions during non-REM sleep by transcranial electrical Stage 3 Odour Stage 4 stimulation using slow (0.75 Hz) but not fast (5 Hz) oscillating potential fields improved the consolidation Odour of hippocampus-dependent but not hippocampus- independent (procedural) memories112, indicating that slow oscillations have a causal role in the consolidation of hippocampus-dependent memories.20:00 24:00 04:00 08:00 Thalamo-cortical spindles seem to prime cortical Time of day networks for the long-term storage of memory repre- sentations. Repeated spindle-associated spike dischargesOdour re-exposure During SWS During REM During waking % *** % % can trigger lTP87 and synchronous spindle activity Recalled card locations 100 100 100 occurs preferentially at synapses that were potentiated during encoding 113. Studies in rats and humans showed Retrieval 90 90 90 increases in spindle density and activity during non- performance REM sleep and SWS after learning of both declarative 80 80 80 tasks and procedural motor skills20,108,114–118. In some 0 0 0 studies these increases correlated with the post-sleep No odour Odour No odour Odour No odour Odour memory improvement 30,119,120 and were localized to the c cortical areas that were activated during encoding, for x = 11 y = -15 example, in the prefrontal cortex after encoding of dif- ficult word pairs117,119, the parietal cortex after a visuo- 6.0 spatial task120 and the contralateral motor cortex after finger motor-skill learning 30. t21 Hippocampal sharp wave-ripples accompany the sleep-associated re-activation of hippocampal neuron 3.5 ensembles that were active during the preceding awakeFigure 1 | Memory re-activation during slow wave sleep (sWs). a | In awake rats experience70,71,121,122. The occurrence of sharp wave-running on a circular track (Run), neurons in the sensory cortex and hippocampus fire Nature Reviews | Neuroscience ripples is facilitated in previously potentiated synap-in a characteristic sequential pattern. Each row represents an individual cell and each tic circuits123 and sharp wave-ripples might promotemark in the upper parts of the diagrams indicates a spike; the curves in the lower parts synaptic potentiation88,124. During an individual rippleindicate the respective average firing patterns of the cells. During subsequent slowwave sleep (SWS) (Sleep), temporal firing sequences observed in the cell assemblies event only a small subpopulation of pyramidal cells fireduring running re-appear both in the cortex and in the hippocampus72. b | Human — the subpopulation varies between successive ripples,subjects learned a two-dimensional object location task on a computer while an odour indicating modulation of select neuronal circuits121,125.was presented as a context stimulus. Re-exposure to the odour specifically during In rats, learning of odour–reward associations pro-subsequent SWS enhanced retention performance (recalled card locations) when duced a robust increase in the number and size of rippletested the next day. There was no enhancement in retention when no association was events for up to two hours during subsequent SWS126. Informed between object locations and odour (that is, odour presentation during SWS humans (epileptic patients) the consolidation of picturebut not during learning) or when odour re-exposure occurred during rapid eye memories that were acquired before a nap correlatedmovement (REM) sleep or waking15. c | When participants slept in an fMRI scanner after with the number of ripples recorded from the peri- andlearning in the presence of odour, re-exposure to the odour during SWS activated the entorhinal cortex, which are important output regionsleft anterior hippocampus (left) and neocortical regions like the retrosplenial cortex(right), which was not observed without odour presentation during prior learning7. Part of the hippocampus127. Selective disruption of ripplesa is modified, with permission, from Ref. 72 © 2007 Macmillan Publishers Ltd. All rights by electrical stimulation during the post-learning restreserved; part b is modified, with permission, from Ref. 15 © 2007 American periods in rats impaired formation of long-lasting spatialAssociation for the Advancement of Science; part c modified, with permission, from memories128, suggesting that ripples have a causal role inRef. 7 © 2007 Elsevier. sleep-associated memory consolidation.NATuRE REVIEWS | NeuroscieNce VoluME 11 | FEbRuARy 2010 | 119 © 2010 Macmillan Publishers Limited. All rights reserved
  7. 7. REVIEWS Interestingly, there is a fine-tuned temporal relation- sleep de-potentiates synaptic circuits that encode famil- ship between the occurrence of slow oscillations, spin- iar events but potentiates synaptic circuits that encode dles and sharp wave-ripples during SWS that coordinates novel episodes77. In humans, neocortical theta activity the bidirectional information flow between the neocor- was enhanced during REM sleep following learning of tex and the hippocampus. With some exceptions (which word pairs62. Theta activity specifically over the right are probably due to methodological differences 129) a prefrontal cortex was correlated with the consolidation consistent finding in humans, cats, rats and mice is that of emotional memories27. by contrast, mice exhibited spindle activity and ripples increase during the up-state reduced REM sleep theta activity after fear condition- and become suppressed during the down-state of a slow ing 143. Thus, although overall there is some evidence for oscillation105,129–132. The top–down control of neuronal an involvement of theta activity in memory processing activity by neocortical slow oscillations probably extends during sleep, its specific contribution to consolidation is to activity in other brain regions that are also relevant to obscure at present. memory consolidation, such as the noradrenergic Theta activity occurring in conjunction with activity burst activity of the locus coeruleus133,134. Sharp wave- in other EEG frequencies points to another important ripple complexes are also temporally coupled to sleep feature that is relevant to memory processing: during spindles105,135,136, with individual ripple events becom- REM sleep, EEG activity in a wide range of frequencies, ing nested in individual spindle troughs135. It has been including theta, shows reduced coherence between lim- suggested that such ripple-spindle events provide a bic-hippocampal and thalamo-cortical circuits than dur- mechanism for a fined-tuned hippocampal-neocortical ing SWS or waking 144,145. likewise, >40 Hz gamma band information transfer, whereby ripples and associated activity shows reduced coherence between CA3 and hippocampal memory re-activations feed exactly into CA1 during tonic REM sleep146. These findings suggest the excitatory phases of the spindle cycle8,105,137,138. In that memory systems become disengaged during REM this scenario, the feed-forward control of slow oscil- sleep49, possibly as a pre-requisite for establishing effec- lations over ripples and spindles enables transferred tive local processes of synaptic consolidation in these information to reach the neocortex during widespread systems (see below). depolarization (during the up-state), that is, a state that favours the induction of persistent synaptic changes, synaptic homeostasis versus system consolidation eventually resulting in the storage of the information in There are currently two hypotheses for the mecha- the cortex. The extent to which the grouping effect of the nisms underlying the consolidation of memory during slow oscillation on hippocampal activity is associated sleep (fIG. 2). The synaptic homeostasis hypothesis11,147 with transfer of memory-specific information in the assumes that consolidation is a by-product of the glo- opposite direction (from cortex to hippocampus), is bal synaptic downscaling that occurs during sleep. The currently unclear. active system consolidation hypothesis proposes that an active consolidation process results from selective Field potentials associated with REM sleep. Ponto- re-activation of memories during sleep 2,8. The two geniculo-occipital (PGo) waves and the EEG theta models are not mutually exclusive; indeed, the hypoth- rhythm seem to support REM sleep-dependent consoli- esized processes probably act in concert to optimize the dation processes (BOX 1). The significance of PGo-waves memory function of sleep. for memory consolidation is indicated by findings in rats of a robust increase in REM sleep PGo-wave density Synaptic homeostasis. According to the synaptic home- for 3–4 hours following training on an active avoid- ostasis hypothesis, information encoding during wake- ance task67,139,140. The increase was proportional to the fulness leads to a net increase in synaptic strength in the improvement in post-sleep task performance, and was brain. Sleep would serve to globally downscale synaptic associated with increased activity of plasticity-related strength to a level that is sustainable in terms of energy IEGs and brain-derived neurotrophic factor (bdnf ) and tissue volume demands and that allows for the re- in the dorsal hippocampus within 3 hours following use of synapses for future encoding 92,94. Slow oscillations training 140. are associated with downscaling: they show maximum The theta (4–8 Hz) oscillations that characterize REM amplitudes at the beginning of sleep when overall syn- sleep in rats are also thought to contribute to consolida- aptic strength is high, due to information uptake dur-Up- and down-states tion, based mainly on the finding that theta activity during ing encoding prior to sleep, and decrease in amplitudeThe slow oscillations that waking occurs during the encoding of hippocampus- across SWS cycles as a result of the gradual synaptic de-predominate eeG activity dependent memories141. However, evidence for this potentiation. Memories become relatively enhanced asduring SWS are characterized assumption is scarce. There is evidence of neuronal downscaling is assumed to be proportional in all syn-by alternating states of re-play of memories in the hippocampus during REM apses, nullifying weak potentiation and thus improv-neuronal silence with anabsence of spiking activity and sleep-associated theta activity 76,77. Place cells encoding a ing the signal-to-noise ratio for the synapses that weremembrane hyperpolarization familiar route were re-activated preferentially during the strongly potentiated during prior waking 147 (fIG. 2a).in all cortical neurons troughs of theta oscillations during post-training REM However, there is no clear evidence on how slow(‘down-state’) and strongly sleep, whereas cells encoding novel sites fired during the oscillations might induce synaptic downscaling. Theincreased wake-like firing oflarge neuronal populations and peaks77. As lTP induction in hippocampal CA1 cells low levels of excitatory neurotransmitters during SWSmembrane depolarization during theta activity depends on the phase of burst activ- (BOX 3) and the sequence of depolarization (up-states)(‘up-state’). ity 142, this finding is consistent with the idea that REM and hyperpolarization (down-states) of slow oscillations120 | FEbRuARy 2010 | VoluME 11 www.nature.com/reviews/neuro © 2010 Macmillan Publishers Limited. All rights reserved
  8. 8. REVIEWS at a frequency of <1 Hz might specifically promote the de- downscaling per se does not explain key features of sleep- potentiation of synapses148. Indeed, slow oscillations and dependent consolidation. However, the synaptic downs- the associated activation of T-type Ca2+ channels seem to caling model explains a second memory-related function favour lTD over lTP89; however, thalamo-cortical spindles of sleep, namely that sleep pro-actively facilitates the and hippocampal ripples nesting in depolarizing up-states encoding of new information during subsequent wake- of slow oscillations support lTP87,88,124. fulness through the de-potentiation of synapses that had In addition, although the expression of markers of become saturated during preceding wakefulness (this synaptic potentiation (such as plasticity-related IEGs) is topic is beyond the scope of this Review)151. globally reduced after a period of sleep, it is increased in specific regions, particularly if sleep was preceded by a Active system consolidation. This concept originated learning experience78,96,98, indicating that synaptic poten- from the standard two-stage model of consolidation pro- tiation might still take place during sleep. Consistent posed for declarative memory2,7,85,121,152 (BOX 2; fIG. 2b), but with downscaling, some neuroimaging studies (which might also account for consolidation in other memory measure relative changes in brain activation) have shown systems8. It is assumed that in the waking brain events reduced task-related activity in cortical regions after are initially encoded in parallel in neocortical networks sleep (e.g. Ref. 149), but these reductions were accom- and in the hippocampus. During subsequent periods of panied by increases in activity in other regions82,83,149,150. SWS the newly acquired memory traces are repeatedly Also, global synaptic downscaling implicates that weakly re-activated and thereby become gradually redistributed encoded memories are forgotten, which contrasts with such that connections within the neocortex are strength- behavioural evidence indicating either no or, under ened, forming more persistent memory representations. certain conditions, a greater benefit from sleep for Re-activation of the new representations gradually adapt weakly than strongly encoded memories35,36. Therefore, them to pre-existing neocortical ‘knowledge networks’, a Waking – Synaptic potentiation Sleep – Synaptic downscaling Synaptic strength W=100 W=150 W=5 W=80 W=120 W=100 W=100 Time b SWS REM Synaptic plasticity Neocortex LTP Slow oscillations Ca2+ NMDAR Synchronizing Hippocampus Synchronous CaMKII feed-forward Sharp wave-ripples feedback PKA AMPAR effect IEG Thalamus Spindles Figure 2 | synaptic homeostasis versus active system consolidation. The synaptic homeostasis hypothesis (a) Nature Reviews | Neuroscience proposes that due to encoding of information during waking, synapses become widely potentiated (large yellow nerve ending), resulting in a net increase in synaptic strength (W = synaptic weight). The small nerve ending represents a new synapse and the unfilled nerve ending is not activated and therefore does not increase in weight. The slow oscillations during subsequent SWS serve to globally downscale synaptic strength (burgundy nerve endings). Thereby, weak connections are eliminated, whereas the relative strength of the remaining connections is preserved. Thus, a memory is enhanced as a consequence of an improved signal-to-noise ratio after downscaling. The active system consolidation model (b) assumes that events during waking are encoded in both neocortical and hippocampal networks. During subsequent slow wave sleep (SWS), slow oscillations drive the repeated re-activation of these representations in the hippocampus, in synchrony with sharp wave-ripples and thalamo-cortical spindles (synchronizing feed-forward effect of the slow oscillation up-state). By synchronizing these events the slow oscillations support the formation of ripple-spindle events, which enable an effective hippocampus-to-neocortex transfer of the re-activated information. Arrival of the hippocampal memory output at cortical networks, coinciding with spindle activity during the depolarizing slow oscillation up-state predisposes these networks to persisting synaptic plastic changes (for example, expression of immediate early genes (IEG) through Ca2+/calmodulin-dependent protein kinase II (CaMKII) and protein kinase A (PKA) activation) that are supported primarily by subsequent rapid eye movement (REM) sleep. AMPAR, α-amino-3-hydroxy-5-methyl-4-isoxazole propionic acid receptor; LTP, long-term potentiation; NMDAR, N-methyl-d-aspartate receptor. Part a is modified, with permission, from Ref. 147 © 2006 Elsevier; part b is modified, with permission, from Ref. 5 © 2006 Sage publications.NATuRE REVIEWS | NeuroscieNce VoluME 11 | FEbRuARy 2010 | 121 © 2010 Macmillan Publishers Limited. All rights reserved
  9. 9. REVIEWS thereby promoting the extraction of invariant repeat- The depolarizing cortical up-states repetitively drive the ing features and qualitative changes in the memory re-activation of memory traces in hippocampal circuits representations2,7. in parallel with thalamo-cortical spindles and activity Corroborating this concept, studies showed that from other regions (for example, noradrenergic locus memory re-activation during post-learning SWS and coeruleus bursts, see BOX 3). This enables synchronous hippocampal ripples accompanying this re-activation feedback from these structures to the neocortex during have a causal role in consolidation15,128. Re-activation in the slow oscillation up-state, which is probably a pre- hippocampal networks seems to be enabled by the low requisite for the formation of more persistent traces in cholinergic tone that characterizes SWS153–155 (BOX 3). neocortical networks8,106. Consistent with this concept, Moreover, there is evidence that the re-activation and neuronal re-activations in the timeframe of cortical slow redistribution of memories during SWS is regulated oscillations have been demonstrated, in which hippocam- by a dialogue between the neocortex and the hippoc- pal re-play leads re-activation in the neocortex 72,122 (and ampus that is essentially under feed-forward control of also in other structures like the striatum156). Moreover, the slow oscillations, which provide a temporal frame. slow oscillations drive the ripples that accompany hip- pocampal re-activation, thus allowing for the formation of spindle-ripple events as a mechanism for effective Box 3 | Neuromodulators hippocampus-to-neocortex information transfer 105,137,138 The specific neurochemical milieu of neurotransmitters and hormones differs strongly (fIG. 2b). Spindles reaching the neocortex during slow between slow wave sleep (SWS) and rapid eye movement (REM) sleep. Some of these oscillation up-states probably act to prime specific neu- neuromodulators contribute to memory consolidation. Interestingly, the most ronal networks, for example, by stimulating Ca2+ influx, prominent contributions to memory processing seem to originate from the cholinergic for subsequent synaptic plastic processes87,157. and monoaminergic brainstem systems that are also involved in the basic regulation of The concept of active system consolidation during sleep171. SWS integrates a central finding from behavioural stud- sWs ies, namely that post-learning sleep not only strength- Cholinergic activity is at a minimum during SWS; this is thought to enable the ens memories but also induces qualitative changes in spontaneous re-activation of hippocampal memory traces and information transfer to their representations and so enables the extraction of the neocortex by reducing the tonic inhibition of hippocampal CA3 and CA1 feedback neurons8,154,155. Accordingly, increasing cholinergic tone during SWS-rich sleep (using invariant features from complex stimulus materials, the physostigmine) blocked the sleep-dependent consolidation of hippocampus-dependent forming of new associations and, eventually, insights word-pair memories153. Conversely, blocking the high cholinergic tone in awake into hidden rules46–48. The concept of a redistribution subjects improved consolidation but impaired the encoding of new information172, of memories during sleep has been corroborated by suggesting that acetylcholine serves as a switch between modes of brain activity, from human brain imaging studies82,83,149,150,158. Interestingly, in encoding during wakefulness to consolidation during SWS154,155. This dual function of these studies, hippocampus-dependent memories were acetylcholine seems to be complemented by glucocorticoids (cortisol in humans), the particularly redistributed to medial prefrontal cortex release of which is also at a minimum during SWS. Glucocorticoids block the regions82,83,122 that also contribute to the generation of hippocampal information flow to the neocortex, and if the level of glucocorticoids slow oscillations159,160. These regions not only have a key is artificially increased during SWS, the consolidation of declarative memories is role in the recall and binding of these memories once blocked173,174. Noradrenergic activity is at an intermediate level during SWS, and seems to be they are stored for the long term85, but also, together related to slow oscillations. In rats, phasic burst firing in the locus coeruleus (the brain’s with the hippocampus, form a loop that supports the main source of noradrenaline) can be entrained by slow oscillations in the frontal explicit encoding of information. As mentioned above, cortex, with a phase-delay of ~300 ms133. It is possible that such bursts enforce behavioural data indicate that sleep does not benefit all plasticity-related immediate early gene (IEG) activity in the neocortex93,95, and thereby memories equally, but seems to preferentially consoli- support at the synaptic level the stabilization of newly formed memory date explicitly encoded information34. In this context, the representations. In humans, the consolidation of odour memories was impaired after prefrontal–hippocampal system might provide a selec- pharmacological suppression of noradrenergic activity during SWS-rich sleep and tion mechanism that determines which memory enters improved after increasing noradrenaline availability (S. Gais, B. Rasch, J.C. Dahmen, S.J. sleep-dependent consolidation. Sara and J. B., unpublished observations). reM sleep A role for ReM sleep in synaptic consolidation Cholinergic activity during REM sleep is similar or higher than during waking. This high The active system consolidation hypothesis leaves open cholinergic activity might promote synaptic consolidation by supporting one challenging issue: although it explains a re-activation- plasticity-related IEG activity162 and the maintenance of long-term potentiation163. Accordingly, blocking muscarinic receptors in rats by scopolamine during REM sleep dependent temporary enhancement and integration impaired memory in a radial arm maze task175. In humans, blocking cholinergic of newly encoded memories into the network of pre- transmission during REM-rich sleep prevented gains in finger motor skill176. Conversely, existing long-term memories, active system consoli- enhancing cholinergic tone during post-training REM-rich sleep improved dation alone does not explain how post-learning sleep consolidation of a visuo-motor skill177. strengthens memory traces and stabilizes underlying Noradrenergic and serotonergic activity reaches a minimum during REM sleep, but it synaptic connections in the long term. Hence, sleep pre- is unclear whether this contributes to consolidation. It has been proposed that the sumably also supports a synaptic form of consolidation release from inhibitory noradrenergic activity during REM sleep enables the for stabilizing memories and this could be the function re-activation of procedural and emotional aspects of memory (in cortico-striatal and of REM sleep. amygdalar networks, respectively), thus supporting memory consolidation154,178. The view that synaptic consolidation is promoted However, enhancing noradrenergic activity during post-learning REM sleep in humans failed to impair procedural memory consolidation56. by REM sleep is supported by the molecular and elec- trophysiological events that characterize this stage.122 | FEbRuARy 2010 | VoluME 11 www.nature.com/reviews/neuro © 2010 Macmillan Publishers Limited. All rights reserved
  10. 10. REVIEWS Sequential Waking SWS REM sleep Time Long-term store Encoding Synaptic consolidation Active system consolidation Temporary store Figure 3 | sequential contributions of sWs and reM sleep to memory consolidation in a two-stage memory system. During waking, memory traces are encoded in both the fast-learning, temporary store and the slow-learning, Nature Reviews | Neuroscience long-term store (in the case of declarative memory these are represented by the hippocampus and neocortex, respectively). During subsequent slow wave sleep (SWS), active system consolidation involves the repeated re-activation of the memories newly encoded in the temporary store, which drives concurrent re-activation of respective representations in the long-term store together with similar associated representations (dotted lines). This process promotes the re-organization and integration of the new memories in the network of pre-existing long-term memories. System consolidation during SWS acts on the background of a global synaptic downscaling process (not illustrated) that prevents saturation of synapses during re-activation (or during encoding in the subsequent wake-phase). During ensuing rapid eye movement (REM) sleep, brain systems act in a ‘disentangled’ mode that is also associated with a disconnection between long-term and temporary stores. This allows for locally encapsulated processes of synaptic consolidation, which strengthen the memory representations that underwent system consolidation (that is, re-organization) during prior SWS (thicker lines). In general, memory benefits optimally from the sequence of SWS and REM sleep. However, declarative memory, because of its integrative nature (it binds features from different memories in different memory systems), benefits more from SWS-associated system consolidation, whereas procedural memories, because of their specificity and discrete nature, might benefit more from REM sleep-associated synaptic consolidation in localized brain circuits. Figure modified, with permission, from Ref. 85 © 2005 Macmillan Publishers Ltd. All rights reserved. Although any links between sleep phases of short dura- indicate that during REM sleep brain activation is as tion and gene expression are difficult to demonstrate high as during waking, but less coherent between differ- for methodological reasons, several studies suggest that ent regions and noisier 144–146. This high level of activation REM sleep, unlike SWS, is associated with an upregu- could act non-specifically to amplify local synaptic plas- lation of plasticity-related IEG activity (RefS 97,98,139). ticity in an environment that, compared with the awake The upregulation depends on learning experience dur- state, is almost entirely unbiased by external stimulus ing prior wakefulness and is localized to brain regions inputs. The disentangled, localized nature of synaptic involved in prior learning 97,98,139. Interestingly, this IEG consolidation might also explain why REM sleep alone activity is correlated with EEG spindle activity during fails to improve declarative memory consolidation: this preceding SWS98. Spindles (which, as discussed above, process essentially relies on the integration of features represent a candidate mechanism that tags networks for from different memories in different memory sys- the neocortical storage of memories during system con- tems and corresponding information transfer between solidation) per se do not induce IEG activity, but might widespread brain areas, that is, SWS-dependent system prime particular brain areas for it, possibly by enhanc- consolidation. ing Ca2+ concentrations in select subgroups of cortical neurons87,157. The activity of plasticity-related early genes Conclusions and future directions depends on cholinergic tone161,162, which is enhanced to SWS and REM sleep have complementary functions to wake-like levels during REM sleep (BOX 3). Cholinergic optimize memory consolidation (fIG. 3). During SWS — activation strengthens the maintenance of lTP in the hip- characterized by slow oscillation-induced widespread pocampus-medial prefrontal cortex pathway 163, a main synchronization of neuronal activity — active system route for transferring memories during SWS-dependent consolidation integrates newly encoded memories with system consolidation82,83,122,136. Electrophysiological sig- pre-existing long-term memories, thereby inducing con- natures of REM sleep, such as PGo waves, are increased formational changes in the respective representations. during post-learning sleep and might promote IEG System consolidation (which preferentially affects explic- activity and memory consolidation140. EEG recordings itly encoded, behaviourally relevant information) acts inNATuRE REVIEWS | NeuroscieNce VoluME 11 | FEbRuARy 2010 | 123 © 2010 Macmillan Publishers Limited. All rights reserved

×