Cholinergic modulation of neural networks supports NREM/REM memory consolidation

Across vertebrate species, sleep consists of repeating cycles of NREM followed by REM. However, their respective functions, and their stereotypic cycling pattern are not well understood. Using a simplified biophysical network model, we investigate the potential role of cholinergic modulation, acting via the muscarinic receptors, on network dynamics and memory consolidation. We show that low and high cholinergic levels associated with NREM and REM sleep, respectively, may play critical, sequential roles in memory consolidation. The network dynamics that facilitate these roles arise through alteration of neural excitability and changes to network-wide excitatory/inhibitory balance. At low acetylcholine (ACh) levels, reduced activation of inhibitory neurons leads to network-wide disinhibition and bursts of synchronized activity led by engram neurons, driving recruitment of additional excitatory neurons into the engram. In contrast, at high ACh levels, increased network inhibition suppresses firing in all but the most strongly recruited excitatory neurons, pruning the expanded engram population. Together, these results provide a testable hypothesis regarding the role of sleep state-specific cholinergic modulation in the process of memory consolidation.

Across animal species sleep stereotypically is composed of non-REM (non-rapid eye movement, NREM) and REM (rapid eye movement) phases, with NREM preceding REM. It remains not clear what differential functions those two phases play in sleep dependent memory consolidation – a process leading to better recollection of recently experienced events. Using a simplified computational model combined with analysis of in vivo recordings from mouse hippocampus, we show that changes in acetylcholine levels during NREM and REM sleep may drive these different roles in memory consolidation. Specifically, we show that low levels of acetylcholine during NREM disinhibit activity of otherwise quiescent subpopulations of excitatory neurons permitting the expansion of memory traces. In contrast, high levels of acetylcholine during REM create more competitive environment for neuronal activity leading to inhibition of less driven cells and preventing an overlap in memory representations when more than one memory traces are involved. We next further demonstrate advantages of the physiological structure of sleep in aiding this process.

Distinct network activity patterns are induced by ACh^– and ACh⁺ states

The hippocampal principal neuron population is heterogeneous with respect to firing rate [65]. In the context of de novo memory formation, a subset of memory-encoding “engram” neurons within the hippocampal network become more active, maintain a higher level of activity over time [66], and increase the strength of functional connections to one another [67]. To model effects of NREM and REM sleep on recently encoded hippocampal neurons, we constructed a reduced neural network composed of two excitatory cell populations: 1) a highly active memory encoding population representative of a single memory (“engram backbone”, EB population, 40 neurons) and 2) sparsely firing neurons with plastic connections from the EB cells (SF population, 80 neurons). These excitatory neurons form a recurrent network with ~10%, primarily random connectivity (see Methods), reminiscent of the hippocampal region CA3 [68]. We found that similar results were obtained when this reduced model was scaled up to 540 (total) neurons (S1 Fig). To model changes to the network initial memory engram formation, EB neurons’ connections to other EB neurons were selectively strengthened [64,69]. With the exception of synaptic connections between EB cells (which are assumed to have strengthened maximally during engram formation), all other excitatory synapses are subject to spike-timing dependent synaptic plasticity (STDP). Interneurons within the network inhibited both EB and SF neurons. To mimic connectivity within forebrain structures, this interneuron population was numerically smaller (20 neurons), but with a high connection density to the entire network (50%). Consistent with available data [66,70–73], these interneurons differentially targeted EB and SF populations, providing a higher level of inhibition to SF neurons via both fast (GABA_A-mediated) and slow (GABA_B - mediated) receptor kinetics (see Methods). While this led to initial differences in firing rate between EB and SF neurons, our results are not dependent upon this aspect of network connectivity, as similar network activity was achieved with equal inhibition to the two populations, when intrinsic current modulated their excitability instead – see supplemental data S2 Fig). To model effects of state-specific neuromodulation, all three populations were simultaneously influenced by changing ACh levels, modeled as a change in the slow, hyperpolarizing M-current conductance (see Methods) [74]. The low ACh neuromodulation during NREM sleep was modeled by high M-current conductance (, ACh^–), while higher, REM-associated ACh release was modeled by low M-current conductance (, ACh⁺).

We first investigated population-level activity dynamics as a function of M-current conductance (Fig 1) for a network with a single encoded memory (single EB population; Fig 1A). Mean spike frequency for neurons within each population (EB neurons, SF neurons and inhibitory interneurons) was measured for values between 0 and 1.5. At low ACh levels () mimicking NREM conditions (ACh^–), the I-F curve of individual neurons was flattened due to reduced excitability, lowering their firing frequency in response to input current [43,75]. We also observed heterogeneous responses among cell populations in the network, with the firing frequency of both inhibitory interneurons and EB neurons (Fig 1D, dark blue and red lines) decreasing. Lower firing rates among interneurons resulted in disinhibition of SF neurons, increasing their firing rate (Fig 1D, teal line). As a result, during the ACh^– state SF neurons are able to be recruited into spontaneous bursts of network activity. These bursts emerge as a result of low network inhibition, Type 2 neural excitability, and biphasic PRCs among neurons in the network [76,77] (Fig 1B, top; 1C, left). Rhythmic firing in the network in this low-ACh state is strongly influenced by the membrane resonance properties associated with Type 2 excitability, and less so by the time constants of inhibitory signaling. At high ACh levels () mimicking REM conditions (ACh⁺), both EB neurons and interneurons exhibited periodic bursts of activity, but at a higher frequency (Fig 1B bottom, 1C right panel). This periodic bursting resulted from slow-decaying inhibitory currents (GABA_B) in EB neurons (reminiscent of a slow version of the mechanism underlying gamma activity [76,77]). In an oscillatory mechanism distinct from the dynamics of the ACh^– state, strong activation of EB neurons in the ACh⁺ state recruits a heavy burst of firing among interneurons, which in turn synchronously inhibits all excitatory neurons until the slow inhibitory postsynaptic current decays, allowing excitatory neurons to spike again. Because interneurons are more active in the ACh⁺ condition and SF neurons are less able to self-excite than the strongly interconnected EB cell group, SF neurons remain relatively silent during this period of high inhibition (Fig 1D), with their firing driven solely by a random noise injected current (see Methods). These waves of excitation and inhibition manifest as network oscillations near theta frequency (Fig 1C, right), reminiscent of the theta rhythm seen in the hippocampus during REM sleep in rodents (and to a somewhat lesser extent in humans [78]). However, we note that while there is evidence that SST+ interneurons promote intrinsic theta-band oscillations in the hippocampus [79–82], the theta rhythm in the hippocampus is primarily understood to be driven by input from the medial septum, which we do not include in our model [83,84]. However, we believe that the source of the theta rhythm, as long as it is accompanied by an increase in inhibitory activity and ACh concentration, does not qualitatively alter our results. An additional observation from the model is a wider range of firing rates in the ACh⁺ state (Fig 1D), consistent with recent data showing that firing rates are more divergent among hippocampal neurons in REM sleep [85].

We next used a previously developed functional connectivity algorithm [86] to measure how the dynamics of firing in the ACh^– state and the ACh⁺ state affect network-wide functional connectivity. We first quantified changes to the mean functional connectivity between pairs of neurons by comparing pairwise spike timing relationships, either within the EB population or within the SF population (see Methods). We then systematically varied the strength of excitatory synapses from the EB population to both EB and SF neurons, and separately, EB excitatory synapses targeting inhibitory interneurons (Fig 1E). Strengthening synaptic connections from EB neurons to SF neurons (moving down the y-axis; Fig 1E) increased mean functional connectivity in the network, while strengthening backbone neurons’ synapses on inhibitory interneurons (moving along the x-axis from left to right; Fig 1E) decreased mean functional connectivity. However, these results differed between ACh^– and ACh⁺ states. We observed that within the EB population itself, functional connectivity was generally weaker during the NREM-like ACh^– state than in the REM-like ACh⁺ state, and did not vary as dramatically with incremental changes to synaptic strength. Conversely, mean functional connectivity between pairs of SF neurons was significantly higher during the ACh^– state than during the ACh⁺ state. Taken together, these findings suggest that both the strength of excitatory input from engram neurons and ACh-regulated network dynamics influence functional connectivity throughout the network.

Modeled effects of learning and sleep states on functional connectivity are recapitulated in the mouse hippocampus following learning

To compare these in silico results with in vivo neural data, we measured pairwise functional connectivity among individual neurons recorded in mouse hippocampal region CA1 during sleep in the hours following single-trial contextual fear conditioning (CFC) [14,16] (Fig 2A top). Functional connectivity was compared between 6 h of baseline recording prior to CFC and the first 6 h post-CFC (a time window critical for sleep-dependent memory consolidation [14,16]) for adjacent bouts of naturally-occurring NREM and REM sleep (i.e., for consecutive bouts of NREM and REM sleep that were separated by less than 10s). An example of such a comparison is shown for a representative mouse in the left panel of Fig 2B. As a negative control, we also compared pairwise functional connectivity for sham-conditioned mice, in which no association between the context and a foot shock was made (“sham”; Fig 2A (bottom). For sham-conditioned mice, the underlying assumption is that in the absence of associative memory, no substantial engram neuron population with increased activity is present in the hippocampus. Consistent with our modeling results, we observed that average functional connectivity between pairs of putative principal (i.e., non-fast spiking) neurons was lower in REM as compared to NREM sleep across all recording conditions. However, following CFC, the REM-associated drop in functional connectivity was less pronounced than during the corresponding baseline period. Indeed, following CFC, many neuronal pairs showed stronger functional connectivity during REM than NREM – something rarely seen during baseline sleep (Fig 2B, solid and dashed black lines; KS test comparison of baseline and post-CFC sleep p < 0.0001). This learning-associated change in state-dependent functional connectivity was not observed in sham conditioned mice (Fig 2B, solid and dashed blue lines), which actually showed a slightly greater decline in functional connectivity across NREM◊REM transitions after sham procedures (KS test comparison of baseline and post-sham p < 0.0014).

To better understand the mechanisms that could underlie these changes, we ran a similar analysis on simulated data using the network from Fig 1, comparing Z-scored pairwise functional connectivity changes between sequential ACh^– and ACh⁺ states. We made this measurement using two different excitatory connectivity strength values corresponding to connections originating from the EB population (as marked on Fig 1E). To simulate the “baseline” period before memory encoding, the EB neurons do not yet have their reciprocal connections (with weak EB synaptic outputs corresponding to the lack of a new memory trace; dashed lines on Fig 2C). For the “post-learning” simulations, EB neurons have their synapses set to a strengthened value (to represent the new memory trace present; solid lines on Fig 2C). Changes in functional connectivity across ACh^– and ACh⁺ were then measured separately for the EB and SF neuron populations. Both populations exhibited a modest drop in pairwise functional connectivity in the baseline condition from the ACh^– to the ACh⁺ state (Fig 2C) – similar to changes observed in mice between baseline NREM and REM states. In contrast, for the post-learning condition, the EB and SF groups exhibit dramatically different behavior. While pairwise functional connectivity of the SF population is significantly reduced at the transition from ACh^– to ACh⁺ (Fig 2C, left panel; p = 0, KS test), the EB neuron population becomes more strongly connected (Fig 2C, right panel; p < 4.75e-128, KS test).

Taken together, these results indicate that neurons with weak functional connectivity in NREM (in the sham/baseline condition in vivo, and the SF population in silico) undergo the large drop in functional connectivity during REM, while those with strong functional connectivity in NREM (in the post-CFC condition in vivo, and the EB population in silico) have a minimal drop or increase in connectivity. High co-firing in NREM (the ACh^– state) as a result of prior learning (modeled here as strengthened EB neuron synapses) enables the persistence of strong functional connectivity into REM (ACh⁺ state).

Further comparison of the experimental SHAM and CFC distributions (Fig 2B; left right panel) indicates that during exposure to CFC, there is rapid network reorganization leading to more heterogenous interactions between populations of putative excitatory neurons. This in turn leads to further selective strengthening of pairwise functional connectivity (i.e., recruitment) during REM – an effect observed to a much smaller extent during baseline and SHAM.

ACh^– and ACh⁺ states differentially recruit a heterogenous population of SF neurons

We next tested how the ACh^– and ACh⁺ state differentially recruit SF neuron populations into engrams, with varying degrees of excitatory input from the EB neurons (Fig 3). We divided the SF population into two groups - one receiving only weak synaptic input from EB neurons, and a second where EB neuronal input is strengthened (i.e., synapses originating from the EB to this SF population are multiplied by a factor ; Fig 3A). We monitored population responses of SF neurons as a function of this synapse multiplier during the ACh^– and ACh⁺ states. In ACh^– conditions, as observed previously, the SF population was disinhibited due to decreased firing rates among inhibitory interneurons (Fig 3B, top, red line). This allowed both SF groups (those with and without the input multiplier) to fire in synchronous bursts with EB neurons (Fig 3C, top), leading to their recruitment into the engram via STDP-mediated strengthening of inputs from EB neurons [87] (Fig 3D, left). Fig 3D depicts changes in mean synaptic strengths between the three populations of excitatory cells (EB neurons, SF neurons with , and SF neurons with ). During the ACh^– state, connections between EB neurons and both SF groups were strengthened; strengthening also occurred to a lesser extent at synapses between neurons in the SF groups. In contrast, during the ACh⁺ state, strong inhibitory input (Fig 3B, bottom, red line) suppressed the activity of those SF neurons with less excitatory input from EB neurons (i.e., only SF neurons with high are active; Fig 3B, bottom, teal line). The SF neurons receiving enough excitatory input to remain active in the ACh⁺ state synchronized with the EB population in coherent bursts (Fig 3C, bottom). This bursting drove STDP-mediated strengthening of connections from the EB neurons to the active SF neurons. At the same time, noisy uncorrelated activity among the strongly inhibited SF neurons induced weakening of their synaptic connections (Fig 3D, right).

ACh^– and ACh⁺ states play distinct roles when multiple engrams are consolidated simultaneously

Having characterized network and neuronal changes in the ACh^– and ACh⁺ states during consolidation, we next tested how ACh^– and ACh⁺ states affect STDP-mediated recruitment of SF neurons by EB populations for multiple (two) simultaneously encoded memories (Fig 4A). An additional 40 EB neurons were added to the network to comprise a second engram population (EB 2, green neurons, Fig 4A). Due to the limited size of our network, engram populations occupy a large fraction of the total number of cells and would be strongly directly connected to one another under completely random connectivity. Thus, connections between EB 1 and EB 2 were removed to prevent runaway excitation, and activity of one EB was always paired with injected current suppression of the other backbone, in an effort to mimic the temporally segregated offline reactivation of hippocampal ensembles corresponding to distinct waking experiences [88,89]. Connections from the EB populations to the SF neurons were random. Thus, the number of connections each SF neuron received from either EB1 or EB2 varied. For visualization purposes, we divided the SF population into four quartiles based on the number of connections to each EB (Fig 4B and 4BB): 1) those receiving many connections from EB 1 and few from EB 2 (blue with gray background), 2) those receiving many connections from EB2 and few from EB1 (green with gray background), 3) those receiving few connections from either EB 1 or EB2 (pink with gray background), and 4) those receiving many connections from both EB1 and EB2 (violet with gray background). To prevent chronic self-excitation and complete synchronous activity among the SF population, we applied a hyperpolarizing input () to both the pink and violet SF groups (3 and 4 above) so that they fired randomly and sparsely and were not recruited into either engram. In addition, the blue and green SF groups (1 and 2 above) had their inputs from EB1 and EB2 initially strengthened 20%, respectively ( see Methods). We tested the degree to which this selective strengthening impacted our results (S3 Fig), and found that our main results were robust to the removal of this bias. We then utilized this two-engram setup to illustrate how sequential ACh^– and ACh⁺ states impact the integration of EB1 or EB2-biased SF neurons into either engram. We monitored this network-wide integration during “test” bouts, which very loosely represent wake-state memory recall. These test bouts had a simulated ACh concentration nearly as high as the ACh⁺ (REM-like) state () and no synaptic plasticity. During testing, each of the two initial EB populations were selectively activated with constant intrinsic current, which serves as a proxy for external (e.g., sensory) input to the network during recall. To identify network-wide effects driven by NREM vs. REM-like conditions, these tests occurred at baseline (i.e., before the ACh^– state), after the ACh^– state, and finally after the ACh⁺ state (Fig 4B). Within the ACh^– and ACh⁺ states, STDP occurred for all excitatory synapses of SF neurons. We observed that during baseline testing, blue and green SF neuron populations fired only sparsely in response to activation of either EB1 or EB2. During the subsequent ACh^– state, a lower level of inhibition allowed both blue and green SF populations to be active, spiking in bursts triggered by either EB1 or EB2. This led to relatively broad STDP-mediated potentiation of connections between each EB population and both populations of SF neurons. This potentiation, in turn, led to increased and overlapping activation of the SF populations during post-ACh^– testing, where the blue and green SF populations were partially activated by either EB population, causing overlap between the two expanded engram representations. However, in the subsequent ACh⁺ state, greater network inhibition prevented most SF neurons receiving EB input from spiking. Only SF neurons with a sufficiently biased input from one of the EB populations underwent the necessary synaptic strengthening (i.e., in the preceding ACh^– state) to maintain high activity in the ACh⁺ state that followed. These ACh⁺ dynamics caused STDP-mediated pruning of connections between each EB population and most of the SF neurons in the network – with only the strongest connections being spared. Thus, during the subsequent post-ACh⁺ test, there was greater segregation of activity patterns between the blue and green SF populations – with SF neurons activated by EB1 vs. EB2 becoming largely non-overlapping, and resulting in two expanded, but separated, engram populations.

To quantify recruitment of SF neurons into each engram, we characterized changes to three observables during the three test bouts:1) the firing frequency of the four SF neuron groups (Fig 4C); 2) changes in connection weights between EB1 and EB2 populations and the SF neuron groups (Figs 4D and S4), and 3) changes in patterns of pairwise functional connectivity among the SF neurons (Fig 4E and 4F). The firing rate of SF neurons coactivated with EB populations during the ACh^– state increased substantially from the baseline test to the post-ACh^– test. This was true for both the weakly coupled group pairs (green SF driven by EB1, blue SF driven by EB2), and between the preferentially coupled group pairs (blue SF driven by EB1, green SF driven by EB2; Fig 4C, p < 0.001, two sample t-test for all comparisons). The subsequent ACh⁺ state raised some SF neuron firing rates and lowered others as measured in the post-ACh⁺ test (Fig 4C, post-ACh^– vs. post-ACh⁺). However, there was no significant change in mean spike frequency after the ACh⁺ state, compared to mean firing rates at testing after the ACh^– state (p > 0.21, two sample t-test; Fig 4B) S4 Fig (diagonal) depicts sample distributions of all connection strengths from EB1 and EB2 to the SF neuron subgroups during each test bout. The mean change in connection strength between test bouts (averaged over four simulation runs) is depicted in Fig 4D and the off-diagonal terms of S4 Fig, with colored vectors corresponding to changes in mean excitatory connection weight incurred during the ACh^– state (Fig 4D, left panel;), the ACh⁺ state (center panel) and as a result of both sequential states (right panel). Synaptic changes reflected the changes seen in firing frequency – the ACh^– state caused general synaptic strengthening, with preference for the more densely connected SF group, and ACh⁺ caused selective pruning of synapses, where weaker connections (EB1 to green SF, EB2 to blue SF) were further weakened and stronger connections (EB1 to blue SF, EB2 to green SF) were more variably altered (S4 Fig, compare post-ACh^– and post-ACh⁺). Thus, the ACh^– state induces strong potentiation to the engrams, and the following ACh⁺ state depresses some synapses and strengthens others. Ultimately, the ACh^–◊ACh⁺ sequence resulted in synaptic changes recruiting two separate populations of SF neurons, one to each of the two EB populations (the effects of which are visible in Fig 4E and 4F).

This effect is further illustrated in the coactivation patterns of the EB and SF populations. We analyzed patterns of functional connectivity within the SF neuron population, using the functional connectivity algorithm described for Fig 1E [81]. We used this metric to characterize pairwise coactivation patterns between SF neurons when EB1 and EB2 were activated separately during each test (Fig 4E). While no significant functional connectivity between individual SF neuron pairs was observed during the baseline test, strong functional connectivity was observed both between and within green and blue SF neuron groups after the ACh^– state. After the ACh^– ◊ACh⁺ sequence, functional connectivity between the two SF groups was reduced significantly, but intra-group connectivity was preserved. To further address this, we quantified the degree of overlap between the portion of these matrices corresponding to the blue and green SF neuron populations (upper left quadrant) by computing the dot product between test pairs (Fig 4F). The functional connectivity pattern overlap is zero during the baseline test, increases significantly after the ACh^– state, and decreases significantly after the ACh⁺ state. We also measured the network-wide distribution of pairwise functional connectivity among all SF neurons (S5 Fig) during test-associated activation of either EB1 (S5A Fig, top) or EB2 (S5A Fig, bottom). After the ACh^– state alone, a large portion of the SF population exhibited increased functional connectivity, skewing the distribution towards positive values (S5A Fig, left panels). In comparison, after the following ACh⁺ state, there was a strong tendency for functional connections to be weaker, although a small fraction of pairs had increased connectivity compared to the post-ACh^– test bout (S5A Fig, center panels). Thus, while synaptic weakening and loss of functional connectivity was the predominant effect of the ACh⁺ state, some functional connections became stronger during this time. Lastly, we compared changes in the functional connectivity between the baseline test and the post-ACh^–◊ACh⁺ test S5A Fig, right panels). While most functional connections among the SF population were not significantly changed, a small number were increased, corresponding to SF neurons that were selectively recruited into either of the two EB representations (Fig 4E). These results demonstrate the potential importance of a low ACh state followed by a high-ACh state in expanding coactive groups of cells in a manner that maintains separation between which new cells become functionally connected to which engram.

To test the robustness of these findings, the duration of depolarizing current pulses used for reactivation of the two EB populations during the ACh^– and ACh⁺ states, as well as the temporal separation of reactivation for the two populations, were systematically varied (S6 and S7 Figs). The duration of reactivation had little impact on either the normalized firing rate of the blue/green SF neurons (referred to as “activation”, see Methods), or the normalized difference in firing rate between the two SF populations (referred to as “segregation”, see Methods – value of one indicates total segregation S6 Fig). On the other hand, the temporal separation of reactivation had an impact on both properties, with longer delays between reactivation of EB1 and EB2 populations leading to decreases in both metrics. However, lengthening the total duration of reactivation (i.e., increasing the total number of reactivations) could compensate for the effects of longer delays on both engram activation and segregation (S7 Fig).

We find that there are network conditions which could interfere with these mechanisms. Our data suggest that the dynamics of the ACh^– state induce somewhat indiscriminate synaptic potentiation between EB neurons and the recruitable SF population. The resulting population of SF neurons with large numbers of connections from both EB populations, and potentially a high degree of excitability, could make engram segregation in the subsequent ACh⁺ state difficult. To investigate this, we ran a series of simulations suppressing the violet SF population (which initially receives strong input from both EB populations) to various degrees. We varied the difference in the between the violet SF population and the blue and green populations (S8 Fig). Not surprisingly, we observed that with strong activation of the violet population, the segregation of the blue and green SF groups is reduced significantly, whereas the activation of the blue and green groups remains high, with both populations being driven by elevated violet SF spiking (S8 Fig). This suggests that in networks with insufficient heterogeneity, such as highly recurrent networks, those with a small variance in the number of excitatory inputs per neuron, or those within highly overlapping engrams, maintaining separate representations during consolidation will be challenging.

Together, these data suggest that during ACh^–◊ACh⁺ sequences, multiple engrams are able to recruit additional network neurons, and can do so in a manner that preserves the separation of their neural representations. These results suggest that the sequential changes in cholinergic tone across the canonical NREM◊REM cycle may drive the recruitment of distinct neuronal populations into existing memory engrams, providing a mechanism for the simultaneous consolidation of multiple, distinct memories in a single network.

Sequential ACh^–◊ACh⁺ state ordering is essential for selective engram formation

Because wake◊NREM◊REM sleep state sequences are ubiquitous across animal species (and wake◊REM◊NREM sequences are not typically observed), we were curious how reversing the sequence would impact the consolidation process. We tested the effects of the ACh⁺ state preceding the ACh^– state (Fig 4BB-FF; right column). In this scenario, we observed diametrically different SF population dynamics (Fig 4BB). Increased synaptic inhibition during the ACh⁺ state led to suppressed activity among SF neurons across the entire simulation (Fig 4CC). Synaptic inputs from EB populations, once weakened during the suppressive ACh⁺ state, could no longer drive burst firing among SF neurons during the subsequent ACh^– state. This eliminated recruitment of the SF neuron population into engrams during the ACh^– state. This effect was evident as weakened EB input connections to SF neurons after the ACh^– state (Fig 4DD), effectively disconnecting SF neurons from the EB populations. Functional connectivity clusters did not form in this scenario (Fig 4EE) and the overlap between EB1 and EB2 functional connectivity remained at zero (Fig 4FF), indicating a complete lack of recruitment of SF neurons into the two engrams. There were also no observable changes in the functional connectivity distributions across the network (S5B Fig). Thus, if the typical ACh^–◊ACh⁺ sequence (corresponding to NREM◊REM) is reversed to ACh⁺◊ACh^– (REM◊NREM), the result is a lack of engram enlargement – interpretable as a failure of memory consolidation.

Multiple ACh^–◊ACh⁺ sequences drive optimal segregated engram enlargement

Sleep, across species, occurs in repeating cycles of NREM and subsequent REM. These cycles occur either intermittently throughout the day, or in the case of humans and some other species, in succession throughout a longer, consolidated sleep period. An unanswered question is whether, and why, it might be beneficial to have multiple NREM◊REM cycles, rather than a single, massed period of NREM and REM. To address this, we ran two sets of simulations. In the first, we divided the sleep period into four repeating ACh^–◊ACh⁺ cycles, with each cycle followed by a test in which EB1 and EB2 were separately activated as done previously (Fig 5A and 5B). In the second set of simulations, a single test was run after one long ACh^–◊ACh⁺ sequence, with the total duration of ACh^– and ACh⁺ states matched to the four cycle simulation (Fig 5C and 5D). For the tests interspersed between the four shorter cycles (Fig 5B), the green and blue SF neuron groups activated preferentially with their respective EB population, while maintaining separation. In contrast, after the longer-duration, single NREM◊REM cycle, there was significant co-activation of these SF neuron populations during the final test (Fig 5D). We applied the activation and segregation metrics to these simulation sets (Fig 5E, see Methods), and observed both activation and segregation increased steadily as a function of the number ACh^–◊ACh⁺ cycles. In contrast, the activation for the non-participating SF neuron groups (pink and violet) did not change throughout the cycles (green line). The outcome of a longer-duration, single ACh^–◊ACh⁺ cycle was qualitatively and quantitatively different from the four-cycle case. Activation of the blue and green SF neuron groups was significantly greater than after cycling ACh^–◊ACh⁺ states, but their segregation was decreased. This means that while these SF groups were more active with EB1 and EB2 after a single, long cycle compared to multiple short ACh^–◊ACh⁺ cycles, the overlap in activity across reactivation of both backbones was much higher. This is because during a single extended period of the ACh^– state, synaptic connections from the backbone populations to green and blue SF neuron groups underwent continuous strengthening, as well as the interconnections between the two SF groups. Once these connections became sufficiently strong, the subsequent ACh⁺ state was unable to silence most of the blue/green SF neurons. Activity in the ACh⁺ state became consistent co-activation of these populations, further strengthening most synapses and avoiding pruning. Taken together, these data suggest that under certain circumstances (e.g., when consolidating more than one memory in a network simultaneously), multiple NREM◊REM cycles can be more effective than a single, longer-duration cycle. In other words, multiple NREM◊REM iterations are most optimal for keeping engrams segregated.

Engram formation and maturation

Available data suggest that engram populations are formed rapidly in the context of waking learning experiences [94,95]. The engram neurons themselves are more excitable and form a strongly interconnected population [66,67,96–100]. The engram remains highly dynamic and undergoes maturation during consolidation, and is hypothesized to further expand (recruiting additional neurons into the memory representation) and/or strengthen connections within the memory trace [64,69], while at the same time increasing separation from other memory-encoding neurons, to allow for accurate recall in the future. It is established that during post-learning sleep, these engram neurons preferentially fire in association with ripples, show increased bursting and temporal coactivation [65], and selective place field remapping [59,101].

Differential roles of NREM and REM during memory consolidation

Behavioral studies in both human subjects and animal models have aimed to identify precise roles of NREM vs. REM sleep in specific aspects of memory consolidation, to whether the two states have differential roles. The importance of NREM sleep for consolidating many forms of memory (including declarative and procedural memory) has been well established, based on studies either disrupting NREM behaviorally, targeting state-specific features of NREM, or enhancing NREM-specific network activity patterns [13,15,17,22,102].

Effects of REM sleep-targeted deprivation studies in human subjects have been less clear, with many forms of memory consolidation (particularly declarative memory) proving robust to this manipulation. However, some sleep-dependent forms of consolidation are disrupted by prevention of REM, including studies where multiple associations (e.g., of different visual cues with an aversive shock vs. safety) must be made in more complex scenarios [103]. A converse approach, which has been used successfully to address REM’s cognitive functions in human subjects, is identification features of behavioral tasks where benefits are correlated with post-learning REM sleep. In one such study, participants were trained on two competing sensory discrimination tasks in rapid succession, followed by a nap opportunity. Only those who engaged in a post-learning nap containing REM showed perceptual improvement for both tasks, without interference between them [61]. Finally, a recent study used administration of ACh receptor antagonists to block ACh transmission during post-learning REM sleep in human subjects, and found subtle deficits in consolidation. Intriguingly, however, subjects in the study were trained on two separate memory tasks prior to the intervention [104]. Together, these data support the idea that REM sleep with high ACh signaling could provide an opportunity to segregate competing memories from one another, to improve performance on complex tasks and/or prevent interference between memories.

At the same time, it has been shown targeted memory reactivation (i.e., via sensory cueing of material learned prior to sleep) seems to ‘work’ only in NREM sleep and not REM [105]. Our modeling predicts that targeted memory reactivation (driving EB neuron reactivation in our model) would have less of an effect during REM due to the increased inhibition in the network.

REM also seems to facilitate other cognitive functions, which may not immediately appear related to a role in memory storage. Intriguingly, data from human subjects indicates that following a period of REM (but not NREM), participants show increased creative thinking, have greater ability to form new associations, and are more likely to have restructured learned information [106,107]. These features are not incompatible with the “pruning” aspects of REM described here, insofar as elimination of weak or irrelevant connections throughout a network may increase its capacity for forming new connections upon receiving new inputs. Experimental results from both humans and animals show that the elevated neural inhibition observed during REM protects memories from interference to permit continual learning [108].

Recent experimental data [109,110] indicate that memories carrying emotional silence specifically benefit from the integration of REM into sleep cycles, with NREM sleep consolidating memory while REM filters memory for continued processing (i.e., further consolidation) based on the saliency (e.g., emotional tone) of the memory. Our modeling is consistent with these results. Representations are pruned during the ACh⁺ state, leaving only the strongest excitatory connections, because of increased inhibition. This competitive environment could lead to reactivation of only more salient memories and competitively disadvantage weaker memories. To illustrate this effect, we ran simulations (S9 Fig) in which one EB population was reactivated for a comparatively longer period (S9A Fig) or with less excitatory drive (S9B Fig), than the other. We show that the EB populations with less or weaker reactivation have reduced neuron recruitment in the ACh^– state and increased pruning in the ACh⁺, state resulting in less SF co-activation during test periods. Based on these and other results, theories of consolidation involving both sleep stages are gaining support, positing roles for NREM and REM aligned to our modeling results [111,112].

Roles of cholinergic signaling in regulating network dynamics during memory consolidation

It is known that both principal cells and subpopulations of inhibitory interneurons express the M1 receptor, with especially prominent expression in the hippocampus and neocortex [39,48,113,114]. Altering state-specific ACh and/or M1-mediated signaling impairs consolidation [56,57,115–117]. For example, preventing the normal decrease in ACh signaling during NREM sleep disrupts memory consolidation, in both human subjects [54] and mice [13].

ACh preferentially activates interneuron populations in the hippocampus [49,50] and neocortex [118–120]; this selective sensitivity is mediated by both nicotinic and muscarinic receptors [51,121]. Because ACh release from medial septal inputs to hippocampus, and by basal forebrain inputs to neocortex, is higher during wake and REM than NREM sleep [122–125], this indicates that, as is true in our model, these networks may be gated by interneurons during REM and disinhibited during NREM. Recent findings also support the idea that altered inhibitory signaling may expand memory traces during post-learning sleep. In regions such as the dentate gyrus, suppression of inhibitory transmission during sleep seems to be an essential component of memory processing [13]. Available data suggests that reactivation and replay of memory traces occurs preferentially (and more broadly throughout the network) during NREM as a correlate of active memory consolidation [15,16]. Finally, there is accumulating evidence of synaptic strengthening occurring preferentially during post-learning NREM sleep. The analyses shown in Fig 2 illustrate that functional connectivity between CA1 neurons is stronger during post-learning NREM than post-learning REM, except for a smaller number of pairwise interactions that show competitive strengthening. In vivo electrophysiological recordings have demonstrated baseline [18,101] or learning-driven [17,19] firing rate increases that occur selectively across NREM bouts. In vivo recordings have shown that potentiation of thalamocortical synapses occurs selectively during NREM bouts [126]. Finally, in vivo imaging data have illustrated dendritic spine formation in the neocortex during post-learning NREM sleep [10].

There are also strong data to support the notion that REM facilitates inhibition-driven pruning of memory traces [111]. Recent in vivo imaging data have shown that in neocortex, interneuron-mediated suppression of pyramidal cell activity is higher during REM sleep compared with NREM sleep [8]. The idea that excitatory synapses are selectively pruned during REM (but not NREM) is supported by in vivo electrophysiological recordings from neocortex and hippocampus (where principal cell firing rates decline selectively across REM bouts) [101,127] and in vivo imaging studies of neocortex (where dendritic spines are lost preferentially during REM) [9]. More generally, recurrent feedback inhibition (which our modeling suggests is higher in REM) is thought to promote competition amongst principal neurons throughout the brain. Reliable spiking of interneurons is observed in the hippocampus and other structures after spontaneous pyramidal cell firing [128], or after optogenetic activation of one to several pyramidal cells [129–131]. On the other hand, optogenetic silencing of principal neurons can lead to strong disinhibition of neighboring principal cells [132,133]. This mechanism is known to be important in regions like the dentate gyrus, where interneuron populations are known to gate the population of principal neurons activated during memory consolidation [13] and recall [133]. It has also been shown that inhibition may play an important role in sculpting the activation patterns of excitatory neurons during sharp-wave ripples [134–136].

Ethics statement

All animal husbandry and experimental procedures were approved by the University of Michigan Institutional Animal Care and Use Committee (IACUC); protocol approval number: PRO00011982.

S1 Fig. Tripling the population of all neurons reproduces the results from smaller networks.

A) Differential activation of neural populations for ACh^- and ACh⁺ states (as in Fig 1). B) recruitment and pruning of SF neurons during reactivation of two engram backbones (as in Fig 4).

https://doi.org/10.1371/journal.pcbi.1013097.s001

(TIF)

S2 Fig. Qualitatively same results in terms of differential activation of SF cells for ACh⁺ and ACh- states is obtained while keeping inhibition level to EB and SF populations the same.

Synaptic multiplier, , for both populations; the for EB population and for SF population.

https://doi.org/10.1371/journal.pcbi.1013097.s002

(TIF)

S3 Fig. Activation and segregation of memory 1 and memory 2 in SF layer as a function of connection multiplier, (see methods).

When all SF groups have the same weight initially and the only difference between assigned SF subpopulations is a random fluctuation in the number of connections at baseline, from EB layer (as decribed in methods). Activity at baseline is measured before the sleep epochs occur. The multiplier does not affect significantly neighter activation or segregation of recruited SF neurons into memory 1 and memory 2.

https://doi.org/10.1371/journal.pcbi.1013097.s003

(TIF)

S4 Fig. Connection strength changes during simulated sleep memory consolidation.

DIAGONAL representative connection maps from backbones of both memories to individual neurons in SF layer. Each dot represents a total connection strength (i.e., sum of synaptic efficacies of neurons belonging to one of the backbones and targeting given SF neuron) from backbone of EB 1 (X-axis) and EB 2 (Y-axis) to individual SF neurons. The connectivity is set at random. The whole population of SF layer at baseline is divided into 4 quartiles (and remains the same for the rest of the simulation): SF neurons receiving stronger input (i.e., more connections) from the memory 1 backbone population are shown in blue; those receiving stronger input (i.e., more connections) from the memory 2 backbone population only are shown in green. Pink and violet SF neurons indicate populations receiving weak (i.e., least connections from both backbones) and strong input from both backbone populations (i.e., most connections from both backbones), respectively. During most of the simulations (except S5 Fig) these groups receive lower constant current () than green and blue SF groups, which leads to reduced, more random, firing patterns. Top left: representative map obtained at baseline. Center: representative map obtained post-ACh-. Bottom right: representative connection map obtained post-ACh^–◊ACh⁺. X - denotes mean connection strength for the given population. OFF-DIGONAL Change of mean connection strength between following timepoint tests: baseline to post-ACh^-, post-ACh^- to post-ACh⁺, baseline to post-ACh^–◊ACh⁺. Values indicate mean values of 4 simulation runs.

https://doi.org/10.1371/journal.pcbi.1013097.s004

(TIF)

S5 Fig.

A) Histograms of pairwise functional connectivity changes between the SF neurons after ACh^-, after ACh⁺, and after a ACh^–◊ACh⁺ cycle. ACh^- state indiscriminately recruits SF neurons into the engram by strengthening connections from backbone neurons to SF neuron populations. This results in a shift of the functional connectivity distributions towards stronger connections (higher z-scores) within the SF neuron population (left). ACh⁺ prunes these connections, leaving/strengthening only the strongest ones, shifting the distribution towards lower functional connectivity values (lower z-scores) and compared to ACh^- (center). When an ACh^- state is followed by an ACh⁺ state (right), a small group of functional connections are especially strengthened, illustrative of recruitment of some SF cells into the engram. B) Histograms of pairwise functional connectivity changes between the SF neurons after ACh⁺, after ACh^-, and after a (reversed) ACh⁺◊ACh^– state cycle (i.e., REM precedes NREM) Because ACh⁺ prunes backbone of SF connections before SF cells could be recruited into the engram, there are no significant shifts in the distributions of the connections.

https://doi.org/10.1371/journal.pcbi.1013097.s005

(TIF)

S6 Fig. Activation and segregation of the recruited SF representations as a function of length of reactivation bouts of memory 1 and memory 2 backbones.

Reactivation bouts were varied between 200–900 ms. A) Sample raster for 200 ms reactivation bouts. B) Sample raster for 800 ms reactivation bout. C) Calculation of activation and segregation (see Methods) as a function of length of reactivation bouts. Both activation and segregation are largely independent of reactivation bout length. Results in C averaged over 5 simulation runs.

https://doi.org/10.1371/journal.pcbi.1013097.s006

(TIF)

S7 Fig. Activation and separation of the recruited SF representations as a function of activation delay between reactivation bouts of memory 1 and memory 2 backbones.

The reactivation bout duration is set to 200 ms. The reactivation delay (i.e., dead space between two consecutive reactivations) is varied between 0–800 ms. A) Sample raster of simulation with an activation delay of 200 ms. B) Activation and segregation of SF representations (see Methods) as a function of length of reactivation delay. Both activation and segregation decrease as a function of activation delayed. However, when the simulation length is controlled for total reactivation time (which decreases as a function of reactivation delay), both functions recover. This indicates that total reactivation time controls activation and segregation magnitude rather than reactivation delay. Values in B indicate mean of 5 simulation runs.

https://doi.org/10.1371/journal.pcbi.1013097.s007

(TIF)

S8 Fig. Activation and segregation of SF neurons belonging to EB 1/EB 2 as a function of activation of neurons shared by both memories (i.e., SF sub-population that has largest number of connections with backbones of both memories; group denoted as violet on Figs 4, 5, S2, S4, and S5).

To activate this sub-population of SF cells, we changed its value from that of the other two (i.e., green and blue) SF populations; invariably, =-6 for the population having the least connections with the backbones of the two memories (i.e., group dentoted as pink on Figs 4, 5, S2, S4, and S5). The x-axis denotes difference in between the violet and green/blue SF population. When the is the same for all the groups the common subpopulation activates strongly (violet line), as do both memories, EB 1 and EB 2 (blue line). The segregation (orange line) is however impeded. As the SF common block population is progressively inactivated (larger values of difference) the segregation returns to normal levels. This intuitively indicates that if the memories share large common engram population the memories cannot be segregated and a single consolidated engram forms.

https://doi.org/10.1371/journal.pcbi.1013097.s008

(TIF)

S9 Fig. Bias in one of the engrams leads to its preferential consolidation.

EB 2 (denoted as green) receives longer reactivation (A) or stronger reactivation in terms of cell activation (B). In both cases EB 2 consolidates preferentially to show increased activation than EB 1 during post-test. A) EB1 (blue) is reactivated for 200ms per cycle whereas EB 2 is reactivated for 400ms per cycle. B) Weaker reactivation of one of the memories (lower constant current drive (I_Drive)).

https://doi.org/10.1371/journal.pcbi.1013097.s009

(TIF)