Hippocampal sharp wave-ripples and the associated sequence replay emerge from structured synaptic interactions in a network model of area CA3

  1. András Ecker
  2. Bence Bagi
  3. Eszter Vértes
  4. Orsolya Steinbach-Németh
  5. Mária R Karlócai
  6. Orsolya I Papp
  7. István Miklós
  8. Norbert Hájos
  9. Tamás F Freund
  10. Attila I Gulyás
  11. Szabolcs Káli

  • Curated by eLife

    eLife logo

    Evaluation Summary:

    The theoretical study by Ecker et al. designs a minimal yet biologically plausible spiking neural network model for hippocampal region CA3 in order to pinpoint the mechanistic sources of important features of population activity (namely sharp wave ripples and replay) observed in vivo. It reproduces many properties of these network events and offers explanations for the observed dynamics. In doing so it demonstrates that the synaptic connectivity patterns formed during spatial exploration may be crucial to the occurrence of these phenomena. The study will be of interest primarily to theoretical researchers because of the many innovative approaches fielded to design the network and analyse its dynamics, and potentially also to experimentalists investigating the hippocampus.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 and Reviewer #3 agreed to share their name with the authors.)

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

Hippocampal place cells are activated sequentially as an animal explores its environment. These activity sequences are internally recreated (‘replayed’), either in the same or reversed order, during bursts of activity (sharp wave-ripples [SWRs]) that occur in sleep and awake rest. SWR-associated replay is thought to be critical for the creation and maintenance of long-term memory. In order to identify the cellular and network mechanisms of SWRs and replay, we constructed and simulated a data-driven model of area CA3 of the hippocampus. Our results show that the chain-like structure of recurrent excitatory interactions established during learning not only determines the content of replay, but is essential for the generation of the SWRs as well. We find that bidirectional replay requires the interplay of the experimentally confirmed, temporally symmetric plasticity rule, and cellular adaptation. Our model provides a unifying framework for diverse phenomena involving hippocampal plasticity, representations, and dynamics, and suggests that the structured neural codes induced by learning may have greater influence over cortical network states than previously appreciated.

Article activity feed

  1. Version published to 10.7554/elife.71850 on eLife
  2. eLife

    Author Response

    Reviewer #1 (Public Review):

    Further manipulations of the network are performed in order to relate the learned network structure to functional properties (i.e. SWRs, replay). By shuffling the weight matrix, it is confirmed that the specific pattern of strong synaptic connections is necessary for replay (as opposed to the overall weight statistics). However, this seems unsurprising given the topological nature of the implicit cognitive map and its place cell representation.

    We agree that it is not surprising that destroying the spatial structure of the weight matrix (by shuffling the weights) eliminates sequence replay. The significant novel finding is that this manipulation (which preserves not just the mean but even the cell-by-cell statistics of the weights) also has a major effect on the population-level dynamics by eliminating sharp waves and ripple oscillations.

    The authors acknowledge and discuss in detail some significant limitations of their results, most prominently the apparently unending nature of the emergent replay (i.e., replay terminates only on encountering the end of the track). Otherwise, it could be noted that the results are demonstrated exclusively in a linear track environment. It would be interesting to establish whether the model is robust to learning from trajectories in an 2d open box environment for example.

    In principle, exploration in an open box (or any other environment) could be modeled by changing the spike statistics during the learning phase so that it mimics the corresponding experimental observations. All other properties and parameters of the model would remain unchanged. This would also allow us to test whether we can predict the replay dynamics in these novel situations, and we intend to examine this in a follow-up project.

    Overall, I find the conclusions of this paper to be supported by the results. Though this work provides a comprehensive account of several previous experimental results, it lacks specific experimental predictions thus limiting the significance of this manuscript for experimentalists who may be in a position to test any such novel hypotheses.

    Both our concrete mechanistic model and the general conceptual framework suggested by our simulation results can be used to derive novel experimental predictions. In fact, the consequences of all the manipulations that we employed to probe our in silico network can be regarded as experimental predictions. Although some of these manipulations (such as shuffling the weights) would be difficult to replicate in real experiments, others are probably feasible using modern tools (cell-type-specific targeting, optogenetics, chemogenetics, etc.). We have listed several specific predictions of our modeling work in a new section of the Discussion (“Experimental predictions”). These predictions include the following:

    • Blocking PVBC-PVBC synaptic interactions should eliminate ripples but not sharp waves or replay, while blocking recurrent excitation is expected to eliminate sharp waves, ripple oscillations and spontaneous replay as well.
    • The nature of replay (e.g., stereotypical vs. diffusive) should depend on the task and behavior during learning (e.g., 1D vs. 2D environment, goal-oriented vs. random foraging task).
    • It should be possible to influence the content of replay by providing structured input during learning or before/during SWRs.
    • Any manipulation changing the symmetry of the STDP rule should bias the direction of replay.

    Reviewer #2 (Public Review):

    The paper uses advanced methodology and the results are well presented. However, certain weaknesses remain, which are listed below.

    1. The main result lacks a mechanistic explanation. My understanding of the main result is that CA3 principal cell firing needs to be tightly controlled by plasticity (so that there are chains of strong bidirectional coupling guiding initial random activity, making it very unlikely that a cell fires outside of its place field) and strong adaptation (so that the participation of a cell in a chain is terminated after a short time, suppressing burst firing) for the model to generate replay events with physiological properties. Is it not possible at all to generate ripples without structured excitatory connectivity in the model?

    Our simulations explore the relationships among (structured) connectivity, the dynamics of population activity (sharp waves and ripples), and detailed activity dynamics (sequence replay). In essence, we find that periods of sequence reactivation (which indeed requires structured connections and adaptation) involve increased activity in all cell populations (corresponding to sharp waves), and that this high activity of PV+ interneurons is responsible for the generation of ripples (as PVBCs synchronize through mutual inhibitory connections). Manipulations of the model that eliminate sequence replay also eliminate sharp waves and ripple oscillations. However, if the excitatory drive to PVBCs is increased to a sufficiently high level by other means (e.g., by drastically increasing the excitatory recurrent weights (Figure 6 D-F), or by directly setting a high level of excitatory input to PVBCs (Figure 8 A-B)), ripple oscillations can be generated in the absence of structured recurrent excitation.

    Related to this, if the principal cells fire only once or twice during the whole replay event (as is observed in vivo and also in the present model), how can a spike-triggered adaptation current (that is required for replay, as the authors show) exert its influence on the dynamics of the network?

    First, in principle, a strong spike-triggered adaptation current could influence the dynamics of the network even after a single spike, by shifting the strongest net input (i.e., the sum of EPSCs, IPSCs, and adaptation current) to cells that have not been active recently but have nearby place fields. In addition, our place cells tend to fire more than once during individual sharp waves. We have amended the manuscript to clarify this fact in lines 188-191, and updated Supplementary Figure 4 (B1 insert and new panel C).

    1. Some of the modelling choices seem to be ad hoc and not well motivated: 2.1 It is not clear whether the symmetric STDP rule proposed by Mishra et al. is more biologically plausible for CA3 than other rules and/or whether there are other rules at work at all. Currently, it reads as if this is the only STDP rule that is plausible for CA3 recurrent excitatory connections.

    Importantly, the plasticity rule used in our paper was not just proposed by Mishra et al. Nature Comms., (2016), but it reflects their direct measurement of STDP in pairs of CA3 pyramidal neurons in hippocampal slices. To our best knowledge this is the only rule that has been experimentally measured in area CA3. On the other hand, we do investigate the effect of using a different type of STDP rule (one that has been measured in several other types of synapse), and find that this choice leads to altered dynamics and the absence of reverse replay events (Figure 5).

    Also, the time constant of the plasticity rule was fixed. It would be good to study the impact of different time constants if this is biologically realistic. The reason for this suggestion is that it might allow for a more mechanistic understanding of the model behavior. Is it possible that there is a certain spatial range for stronger-than-baseline connectivity that is required for there to be replay and if so, does it depend on the time constant?

    In our study, the time constant of STDP was fixed to the value that Mishra et al. Nature Comms., (2016) measured experimentally. The way in which the shape (magnitude and time constant) of the STDP rule influences the network dynamics is indeed an interesting topic for theoretical research; however, exactly this issue was recently explored by Theodoni et al. eLife, (2017), so we decided to focus our efforts on other questions.

    2.2 Plasticity between principal cells is switched off after the initial exploration phase and the networks are static thereafter. How realistic is this? It might be possible that spike-triggered synaptic plasticity during a ripple event interferes with the bidirectional coupling established during the initial learning phase, thus rendering replay unstable and/or terminating it before the end of the track is reached.

    We agree with the reviewer that the complete absence of plasticity in the second phase may not be realistic. This is a simplifying assumption of our model (listed as Assumption 5 in Table 1), which is shared by many models of hippocampal learning. On the other hand, plasticity is known to be strongly modulated by brain state (e.g., through the action of subcortical inputs on hippocampal cells and synapses), so plasticity may be substantially weaker or different in off-line states (such as slow-wave sleep; see Hasselmo, Current Opinion in Neurobiology, (2006), and Fuenzalida et al., Neuroscience, 2021, for reviews). Finally, since replay recapitulates activity correlations in the learning phase, the main effect of plasticity during replay would be to reinforce existing patterns of weights and activity (although some kind of homeostatic mechanism may be required to prevent uncontrolled positive feedback). These arguments have now been included in the Discussion section of the paper.

    1. The authors conclude that the FINO mechanism is at the heart of ripple generation in CA3. I found it confusing that apparently, no ripple oscillations are present in the population rate of the principal cells, but only in the calculated LFP, although the Figure Supplement for Fig. 3 seems to suggest that ripple oscillations are present also in the population rate of the principal cells. Maybe calculating the ripple frequency directly from the population rate of the principal cells, but not from the LFP, would change this conclusion (presented in Fig. 8), because the former still exhibits ripple oscillations, whereas the latter doesn't?

    This appears to be some sort of misunderstanding (possibly stemming from the relatively coarse binning of PC rate in some of our figures which are intended to illustrate long-term variations); ripple oscillations are indeed present in the PC rate (in our baseline model) as shown in Supplementary Figure 5. Regarding the modified model version presented in Figure 8 C, the PC rate is devoid of ripple frequency oscillations, just like the LFP.

    1. Is it possible to determine which initial conditions cause forward and which reverse replay? Would it be possible to compare this to different behavioural states exhibited by an awake animal traversing a track and then resting? The authors mention in the Introduction that forward replay is associated with memory recall, whereas reverse replay is associated with reward-based learning. How is this reflected in the statistics of these two types of replay events in the resting phase of the model? Is there an expectation that both events should occur equally often or are there any conditions that bias the direction of the replay in the model? This is related to the unstructured input that principal cells receive in the model, so maybe there is an interplay of these two sources of excitation (feedforward via the mossy fibers and recurrent via the learned connectivity) which determines the direction and frequency of replay events. I feel that addressing at least some of these points would allow the authors to gauge the realism of their replay model more finely.

    We explored some of the mechanisms that may bias the direction of replay in Figure 2C. In particular, we show that one can evoke either forward or backward replay via external input that corresponds to neural activity at the start and the end of the track, respectively. Without any input cues, forward and backward replays occur with the same frequency in our baseline model (we have verified this by running the simulation with different random seeds, which influences the random mossy fiber input).

    In general, the direction of replay in the model can depend on systematic biases and random variations in the weight structure, but also on (random or deterministic) spatio-temporal structure in the external inputs. We believe that the cued replay scenario that we simulated is not entirely artificial – even real environments may contain choice points where the animal pauses to plan its route, or reward sites that again make the animal stop and may also activate subcortical modulatory systems that affect plasticity. At intermediate sites, systematic bias could be introduced during exploration by factors that influence behavior (e.g., running speed), neuronal activity or plasticity (e.g., attention), especially if such factors can change the fully symmetric nature of the synaptic plasticity rule. Random variations in connectivity or the weights (e.g., due to stochastic spiking during learning) will lead to a location-dependent systematic bias towards forward or reverse replay. Finally, (randomly occurring or input-driven) sequences in the external input are also expected to bias the direction of replay even at intermediate points. However, a detailed analysis of the relative impact of these factors will require further efforts. We have included some of these arguments in lines 404-427 of the Discussion.

  3. eLife

    Evaluation Summary:

    The theoretical study by Ecker et al. designs a minimal yet biologically plausible spiking neural network model for hippocampal region CA3 in order to pinpoint the mechanistic sources of important features of population activity (namely sharp wave ripples and replay) observed in vivo. It reproduces many properties of these network events and offers explanations for the observed dynamics. In doing so it demonstrates that the synaptic connectivity patterns formed during spatial exploration may be crucial to the occurrence of these phenomena. The study will be of interest primarily to theoretical researchers because of the many innovative approaches fielded to design the network and analyse its dynamics, and potentially also to experimentalists investigating the hippocampus.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 and Reviewer #3 agreed to share their name with the authors.)

  4. eLife

    Reviewer #1 (Public Review):

    A neural network model of a subregion of the hippocampus is trained based on simulated experiences and examined in the context of reactivation during rest. The authors perform several manipulations in order to pinpoint the mechanistic sources of important features of population activity in this brain region, namely sharp wave deflections, ripple oscillations, and replay. This study enhances our understanding of area CA3 and hippocampal replay by pinpointing the anatomical and functional features of CA3 circuitry which contribute to SWR production and replay from experience.

    This study overcomes several significant barriers in the circuit modeling of SWR and replay in the hippocampus. First, the synaptic weights of the network are primarily learned from location inputs drawn from environment trajectories. This stands in contrast to many models of hippocampal activity which rely on synaptic weight matrices established through explicit calculation (e.g. based on the place cell covariance structure) or supervised learning. Second, the synaptic weights are learned from a symmetric STDP rule which has been established experimentally as opposed to an asymmetric STDP rule which has been implemented previously in order to generate sequential place reactivations. In the model presented here, it is shown (through an ablation simulation) that an adaptation mechanism causes sequential place reactivations to emerge and facilitates bidirectional replay. More broadly, previous network models trained online from simulated experiences lack much of the biological detail present here.

    In simulation it was noted that the network spontaneously shifted into high activity modes (generating sharp waves) accompanied by replay and emergent oscillatory activity in the ripple band. The authors were able to tease apart these phenomena in a manner consistent with empirical results. They showed that sharp waves and replay emerged from increased levels of activity in the the pyramidal cell population (containing place cells) spontaneously which then drive ripple oscillations in another subpopulation of basket cells supported by recurrent inhibition.

    Further manipulations of the network are performed in order to relate the learned network structure to functional properties (i.e. SWRs, replay). By shuffling the weight matrix, it is confirmed that the specific pattern of strong synaptic connections is necessary for replay (as opposed to the overall weight statistics). However, this seems unsurprising given the topological nature of the implicit cognitive map and its place cell representation.

    Although some of the results presented here appear in other models, it seems that this study represents an integrative, state-of-the-art account of area CA3 circuitry and the principles employed may be embedded in broader circuit models covering other areas of the hippocampus. Other neuroscientists may be facilitated in accomplishing this by the code repository associated to the manuscript. The authors acknowledge and discuss in detail some significant limitations of their results, most prominently the apparently unending nature of the emergent replay (i.e. replay terminates only on encountering the end of the track). Otherwise, it could be noted that the results are demonstrated exclusively in a linear track environment. It would be interesting to establish whether the model is robust to learning from trajectories in an 2d open box environment for example. Overall, I find the conclusions of this paper to be supported by the results. Though this work provides a comprehensive account of several previous experimental results, it lacks specific experimental predictions thus limiting the significance of this manuscript for experimentalists who may be in a position to test any such novel hypotheses.

  5. eLife

    Reviewer #2 (Public Review):

    The theoretical study by Ecker et al. uses advanced computational methods to show that the statistics of the connectivity between excitatory principal cells (PCs) determines the content (forward and reverse replay) and the structure (single cell and network activity levels) of replay events in the CA3 region of the hippocampus. The model assumes that a symmetric plasticity rule generates the connectivity between the principal cells in an initial exploratory learning phase. Before the start of this phase, excitatory neurons are coupled randomly and sparsely with 10% connection probability. In a second phase, akin to resting and consummatory behaviours, the model then generates both forward and reverse replay events with biologically plausible single cell and network activity properties. Importantly, this property is not a built-in feature of the model but emerges dynamically. The study uses single-compartment adaptive exponential integrate-and-fire models whose parameters are determined offline (that is, in a single-cell simulation instead of a network simulation) using an evolutionary algorithm (inspyred). In a second step, the synaptic weights, but not the connection probabilities, of the network are determined with another, separate evolutionary alogrithm (BluePyOpt) to yield dynamical properties typically observed during sharp wave/ripple events, such as realistic principal cell firing rates accompanied by significant ripple frequency oscillations. The authors show that the simultaneous presence of forward and reverse replay requires a symmetric plasticity kernel, structured and strong recurrent excitatory connectivity and a spike-triggered adaptation current for the principal cells. The model also supports learning two distinct environments and is robust to scaling the recurrent excitatory weights. It is also shown that ripples together with replay only occurs for sufficiently strong recurrent connections within the inhibitory basket cell (PVBC) population and that with such absent inhibitory connectivity, increasing the strength of the recurrent excitatory weights does not reinstate ripple oscillations, which supports the previously suggested FINO hypothesis for ripple oscillation generation. The manuscript finishes with a thoughtful and extensive discussion about the limitations and shortcomings of the model and its relationship to related studies of hippocampal network oscillations. For example, the authors discuss their main modelling assumption assuming 10% recurrent excitatory connectivity in light of recent experimental results (Guzmán et al. 2016) which suggests that 10% connectivity might actually be too large for CA3.

    The paper uses advanced methodology and the results are well presented. However, certain weaknesses remain, which are listed below.

    1. The main result lacks a mechanistic explanation. My understanding of the main result is that CA3 principal cell firing needs to be tightly controlled by plasticity (so that there are chains of strong bidirectional coupling guiding initial random activity, making it very unlikely that a cell fires outside of its place field) and strong adaptation (so that the participation of a cell in a chain is terminated after a short time, suppressing burst firing) for the model to generate replay events with physiological properties. Is it not possible at all to generate ripples without structured excitatory connectivity in the model? Related to this, if the principal cells fire only once or twice during the whole replay event (as is observed in vivo and also in the present model), how can a spike-triggered adapation current (that is required for replay, as the authors show) exert its influence on the dynamics of the network?

    2. Some of the modelling choices seem to be ad hoc and not well motivated.

    2.1) It is not clear whether the symmetric STDP rule proposed by Mishra et al. is more biologically plausible for CA3 than other rules and/or whether there are other rules at work at all. Currently, it reads as if this is the only STDP rule that is plausible for CA3 recurrent excitatory connections.

    Also, the time constant of the plasticity rule was fixed. It would be good to study the impact of different time constants if this is biologically realistic. The reason for this suggestion is that it might allow for a more mechanistic understanding of the model behavior. Is it possible that there is a certain spatial range for stronger-than-baseline connectivity that is required for there to be replay and if so, does it depend on the time constant?

    2.2) Plasticity between principal cells is switched off after the initial exploration phase and the networks are static thereafter. How realistic is this? It might be possible that spike-triggered synaptic plasticity during a ripple event interferes with the bidirectional coupling established during the initial learning phase, thus rendering replay unstable and/or terminating it before the end of the track is reached.

    1. The authors conclude that the FINO mechanism is at the heart of ripple generation in CA3. I found it confusing that apparently, no ripple oscillations are present in the population rate of the principal cells, but only in the calculated LFP, although the Figure Supplement for Fig. 3 seems to suggest that ripple oscillations are present also in the population rate of the principal cells. Maybe calculating the ripple frequency directly from the population rate of the principal cells, but not from the LFP, would change this conclusion (presented in Fig. 8), because the former still exhibits ripple oscillations, whereas the latter doesn't?

    2. Is it possible to determine which initial conditions cause forward and which reverse replay? Would it be possible to compare this to different behavioural states exhibited by an awake animal traversing a track and then resting? The authors mention in the Introduction that forward replay is associated with memory recall, whereas reverse replay is associated with reward-based learning. How is this reflected in the statistics of these two types of replay events in the resting phase of the model? Is there an expectation that both events should occur equally often or are there any conditions that bias the direction of the replay in the model? This is related to the unstructured input that principal cells receive in the model, so maybe there is an interplay of these two sources of excitation (feedforward via the mossy fibers and recurrent via the learned connectivity) which determines the direction and frequency of replay events. I feel that addressing at least some of these points would allow the authors to gauge the realism of their replay model more finely.

  6. eLife

    Reviewer #3 (Public Review):

    Computational neural network models are valuable tools for exploring mechanisms of brain dynamics, including generation of different frequencies of oscillation in neural activity and how synaptic plasticity can result in the storage and recall of patterns of information. These aspects are combined here to show that the connectivity patterns that arise when storing activity patterns through synaptic spike-timing-dependent plasticity (STDP) can actually result in the emergence of sharp waves and associated high-frequency ripples (SWR), as seen in the CA3 region of the mammalian hippocampus. This is a result that is of wide interest to neuroscience research.

    Though the model used is rather simple, it contains sufficient detail to be directly constrained by experimental data in key aspects, such as cell spiking responses and the STDP learning rule. The simplicity also allows the operation of the network to be examined and understood. The results clearly show that symmetry in the STDP rule is needed for both forward and backward replay of pattern sequences to emerge, which makes logical sense. Sequence replay depends on spike frequency adaptation in the pyramidal cells. The model confirms the results of other experimental and modelling studies that strong recurrent inhibition between inhibitory interneurons can underpin ripple oscillations.

    Appropriate analysis methods are used to identify sequence replay and the power of different oscillation frequencies in the neural activity.

    The authors are careful to discuss the limitations of their model and the mismatches with experimental data. Such mismatches do not invalidate their main conclusions, but do highlight that there are potentially other neural mechanisms that are absent from the model that contribute to shaping the spatio-temporal characteristics of sequence replay and SWRs.

  7. Version published to 10.1101/2021.02.18.431868v6 on bioRxiv
  8. Version published to 10.21203/rs.3.rs-345804/v1 on Research Square
  9. Version published to 10.1101/2021.02.18.431868v5 on bioRxiv
  10. Version published to 10.1101/2021.02.18.431868v4 on bioRxiv
  11. Version published to 10.1101/2021.02.18.431868v3 on bioRxiv
  12. Version published to 10.1101/2021.02.18.431868v2 on bioRxiv
  13. Version published to 10.1101/2021.02.18.431868v1 on bioRxiv