A cortical–hippocampal communication undergoes rebalancing after new learning

  1. Arron F Hall
  2. Dong V Wang

  • Curated by eLife

    eLife logo

    eLife Assessment

    This valuable study investigates the neural basis of bidirectional communication between the cortex and hippocampus during learning. The evidence supporting the identification of specific circuits and functional cell types involved is convincing. However, certain aspects of the behavioral analysis and statistical interpretation remain incomplete. Overall, the work will be of interest to neuroscientists studying learning and memory.

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

Abstract

The brain’s ability to consolidate a wide range of memories while maintaining their distinctiveness across experiences remains poorly understood. Sharp-wave ripples, neural oscillations that occur predominantly within CA1 of the hippocampus during immobility and sleep, have been shown to play a critical role in the consolidation process. More recently, evidence has uncovered functional heterogeneity of pyramidal neurons within distinct sublayers of CA1 that display unique properties during ripples, potentially contributing to memory specificity. Despite this, it remains unclear exactly how ripples shift the activity of CA1 neuronal populations to accommodate the consolidation of specific memories and how sublayer differences manifest. Here, we studied interactions between the anterior cingulate cortex (ACC) and CA1 neurons during ripples and discovered a reorganization of their communication following learning. Notably, this reorganization appeared specifically for CA1 superficial (CA1sup) sublayer neurons. Utilizing a generalized linear model decoder, we demonstrate the pre-existence of ACC-to-CA1sup communication, which is suppressed during new learning and subsequent sleep suggesting that ACC activity may reallocate the contribution of CA1sup neurons during memory acquisition and consolidation. Further supporting this notion, we found that optogenetic stimulations of the ACC preferentially suppressed CA1sup interneurons while activating a unique subset of CA1 interneurons. Overall, these findings highlight a possible role of the ACC in rebalancing CA1 neuronal populations’ contribution to ripple contents surrounding learning.

Article activity feed

  1. eLife

    eLife Assessment

    This valuable study investigates the neural basis of bidirectional communication between the cortex and hippocampus during learning. The evidence supporting the identification of specific circuits and functional cell types involved is convincing. However, certain aspects of the behavioral analysis and statistical interpretation remain incomplete. Overall, the work will be of interest to neuroscientists studying learning and memory.

  2. eLife

    Reviewer #1 (Public review):

    Summary:

    This work by Hall et al provides a novel and important new finding about communication between the anterior cingulate cortex (ACC) and the CA1 region of the dorsal hippocampus: there is a clear ability of ACC to predict CA1 activity, and that is modulated by learning/experience. Furthermore, they have some evidence that the modulation differs by whether the CA1 neurons were in the deep versus superficial sub-layer of CA1. The evidence is suggestive of new and exciting findings, but some gaps and weaknesses remain to be addressed before I believe all of the authors' claims can be supported. The figures also need to be slightly better organized, and the discussion is missing a major dimension in my opinion. Overall, this is a strong submission, but with some gaps to fill.

    Strengths:

    (1) This is a well-written manuscript - the introduction was especially clear, well-cited, and motivating.

    (2) The sub-layer specific communication between ACC and CA1 represents the discovery of a novel and functionally impactful piece of neurobiology.

    (3) Optogenetics was an important verification of ACC-CA1 communication, as was the analysis of neurons by waveform type.

    Weaknesses:

    (1) Figure 2: Why are the data separated into two groups from the outset? If all data are combined, is there a general drop in prediction gain from pre to post?

    (2) 2b and 2c are important since they are complementary means to show the same thing, and it is important that they cross-validate each other, especially since the non-significant task active neuron difference in 2b appears to be nearly as strong as the significant difference to its left. A more holistic analysis can be done to compare these dimensions.

    (3) Sup vs deep neuron definition: Did the authors have any means to validate this anatomical separation using histology or otherwise? I don't believe they described anything like that, and instead use physiology to infer anatomical location. I understand anatomy-based methods may be practically impossible with tetrodes, but this limitation should at least be mentioned, and it should be explained that without something like silicon probes or histological validation, anatomy had to be inferred from physiology.

    (4) Superficial vs deep differences in firing rate ratio based on PG: there are many fewer CAdeep neurons, but in 4c, the trends appear to be the same pre-training, top PG lower than others. It seems the lack of difference in CA1deep in 4c may be due to the much lower power/n. This should be discussed or addressed.

    (5) In Figure 5, the term "firing rate ratio" is used, and it sounds the same as in previous figures, but this is a different ratio (based on modulation by opto stim, not task).

    (6) I would like to learn more about these v-type neurons. I understand we do not yet know about their molecular or morphologic correlate, but more analysis can be done with the current data.

    (7) I would like more discussion of ACC-CA1 connectivity.

    (8) Some elements may be missing from the discussion, relating baseline functioning versus post-learning function.

  3. eLife

    Reviewer #2 (Public review):

    Summary:

    This study uncovers an inhibitory pathway from the anterior cingulate cortex (ACC) to pyramidal cells in the superficial sublayer of hippocampal area CA1 (CA1sup). As ACC neuron spiking tends to precede hippocampal ripples, this presents the intriguing possibility that ACC inputs are selectively inhibiting particular CA1sup neurons, which could play a role in the reactivation of task-related ensembles known to take place during hippocampal ripples. Indeed, through a generalized linear model (GLM) analysis, the authors demonstrate that the ACC activity within the 200ms immediately preceding the ripple is predictive of the ripple content.

    Strengths:

    The biggest strength of the work is the optogenetic manipulation experiments, which convincingly demonstrate that stimulation of ACC pyramidal neurons activates an interneuron population with symmetric spike waveforms, and inhibits parvalbumin interneurons and pyramidal cells in CA1sup but not CA1deep sublayer.

    An additional strength in the GLM analysis which consistently shows that ACC activity preceding the ripple is predictive of hippocampal activity during the ripple considerably more than in shuffled data for all cells and periods tested.

    Weaknesses:

    The major weakness of this work is that the link with learning and memory is not very well supported.

    The only evidence of rebalancing and reorganization appears to be a single statistical test (the test in Figure 1f, p=0.013) demonstrating a decrease of the GLM prediction gain from pre-task sleep to post-task sleep; the same test is repeated for subsets of the data in the rest of the figures. As the idea of rebalancing and reorganization is central to the paper as currently written, exploring it through another measure, independent of the GLM prediction gain, should be expected. The notion that this pathway is suppressed in sleep following learning can be supported by demonstrating a decrease in any of the following measures: ACC spike-triggered average CA1sup responses, cross-covariances (Wierzynski et al 2009) between ACC and CA1sup cells in post-task sleep, or ripple-triggered cross-correlations (Sirota et al. 2009).

    The differences between task-active and task-inactive neurons are not convincing. The separation between task-active and task-inactive neurons is to divide a distribution that is far from bimodal into what appears to be two arbitrary groups. Similarly, the authors divide cells relative to their prediction gain ("Top PG" and "Bottom PG" in Figure 2c), which fails to select for the population of significantly predicted cells (relative to the shuffle). Within CA1sup cells, after learning, there is a significant decrease in the prediction gain for "task-inactive" cells but not "task-active" cells, but it is important to keep in mind that the "task-active" group contains only 24 neurons, and there was no difference between the two groups of cells ("task-active" vs "task-inactive") when directly compared.

    Finally, it is not clear whether the identity of the pathway-responsive CA1sup neurons is fixed or whether it may change with learning. A deeper analysis into the cell pair cross-correlations or the weights of the GLM analysis may reveal whether there is a reorganization of CA1sup responses (some cells that were inhibited are no longer inhibited, and vice versa) or a dampening (the same CA1sup cells are inhibited in both cases, but the inhibition is less-pronounced in post-task sleep). The possibility of a rigid circuit dampened immediately following fear conditioning, is not discussed by the authors.

  4. eLife

    Reviewer #3 (Public review):

    Summary:

    In this study, Hall and colleagues investigate how the coupling of activity from ACC to CA1is altered by fear learning, showing that during sleep immediately before learning, there is evidence for increased coupling of ACC activity with neurons that will subsequently be inhibited during the learning process. They go on to show that this effect seems to be mediated most by a subpopulation of neurons in the superficial layer of CA1. This fits with previous reports suggesting that these superficial neurons are key for the flexible updating of memory. The authors then go on to show that artificial activation of ACC using optogenetics results in varied effects in CA1, including a subtle decrease in activity of superficial neurons that lasts longer than the stimulus itself. Finally, the authors present some preliminary data suggesting that different interneurons may be recruited by this optogenetic stimulation in different ways and at different times.

    Overall, this is an interesting paper, but much of the analysis is very preliminary, and much of the crucial data about the learning effects and alterations to cell firing are not presented clearly and fully. This is further confounded by a rather opaque description of the results and analysis in the text. Overall, there is something very interesting here, but there needs to be a substantial series of extra analyses to clearly say what this is. In many cases, more robust analysis may render the results underpowered, which could dramatically change the conclusions of the paper.

    Strengths:

    The authors performed difficult, dual-location recordings across a multi-day learning paradigm, which seems like it could be a really nice dataset. They delve into the circuit basis of an interesting finding regarding ACC to CA1 connectivity and how this changes before and after fear conditioning. They provide data to suggest this connectivity may be through specific and distinct subcircuits in CA1.

    Weaknesses:

    (1) There is essentially no information in the text or figures about what the actual learning was, how it was done, how individual animals performed, and how any of these metrics related to learning. Looking at the methods, the authors did a number of things never mentioned anywhere in the text or figures, including novel arena exposure, contextual reexposure in extinction after learning, etc. It seems that this is a very rich dataset that has not been presented at all. I would recommend at the very least:
    a) Plot all of the behavioural training data, and how each mouse relates to one another - did the mice learn? At this stage, we don't know!
    b) Explain in the text in detail exactly what was done and why, and what this tells us about the neuronal activity.
    c) If there is variance in learning and or conditioning, does this relate to features in the analysis, such as the GLM result.

    (2) Along similar lines, a key metric for most of the paper is that neurons most coupled with ACC are more likely to be inhibited during training. However, there is nothing anywhere in the paper showing these data. How do neurons in general respond to contextual shocks? The methods describe this as the average firing rate during training, normalised to pre-sleep activity. This metric seems a bit coarse and may obscure really important task-relevant dynamics. Are the neurons active at specific times, are they tuned to relevant parts of the task, and do any of these features of the cell activity also relate to the coupling with ACC? Similarly, how did the authors mitigate the influence of electrical artefacts caused by the foot shock in their recordings? Again, there is a huge amount of data here that is not being described, and likely holds very valuable information about what is actually happening. The paper would really benefit from the inclusion of these data in an accessible form, such as heatmaps of spiking, how these patterns change over time, and around e.g., foot shock, etc. Also key is how these features are altered by the variability of learning across subjects.

    (3) A number of the effects are presented by comparing a statistically significant effect to a non-statistically significant effect (e.g. in Figure 2b, Figure 2d, Figure 4 b,c, and others). This isn't really valid - the key test that the two groups are different is either with a direct test of the difference or an interaction term in an e.g., ANOVA test. In some places, I am not sure the same conclusions will be drawn from the data with these tests.

    (4) To what extent is defining superficial and deep CA1 neurons solely by ripple waveform an accepted method? Of the two papers referenced for this approach, one is a 2-photon calcium imaging paper that does not do electrical recordings (as far as I am aware), and the second uses this as a descriptor after defining the positions of units on an array. It would be good to clarify how accepted this is, and also how robust this is. At the very least, some kind of metric or walkthrough in the supplement as to how this was done, and how well each cell was classified and with what confidence, or some metric of how distinct and separate the two populations were (or was it just a smudge).

    (5) In the optogenetic experiment in Figure 5, the effect on the CA1 sup neurons seems to be driven by changes in a small subpopulation of this group, with no change in the others. Related to point 2, is there anything else in the data that can pull out what these cells are? More detailed analysis of the firing of these neurons might pull out something really interesting.

    (6) Related to this - a number of comparisons simply pool neurons across mice and analyse them as if independent. This is done a lot in the past, but it would be better if an approach that included the interdependence of neurons recorded from the same mouse at the same time were used (such as a hierarchical model). While this is complex, a simpler approach would just be to plot the summary data also per mouse. For example, in Figure 5, how do the neurons inhibited by ACC activation spread across the different mice? Is the level of inhibition related to how well the mice learned the CS-US association?

    (7) Figure 6 is interesting, but very preliminary. None of the effects are quantified, and one of the cell types is not identified. I think some proper analysis needs to be done, again across mice, to be able to draw conclusions from these data.

    (8) Finally, in general, I felt that the way the paper was written was very hard to follow, often relying on very processed levels of analysis that were hard to relate back to the raw traces and their biological meaning. In general taking more words to really simply and fully explain each analysis, and taking the words and figures to walk through how each analysis was done and what it tells us about the neuronal data/biology would be really beneficial, especially to someone who is not an extracellular electrophysiologist or immersed in the immediate field.

    In summary, while this manuscript explores an intriguing hypothesis about pre-learning circuit dynamics, it is currently held back by insufficient clarity in behavioural analysis, data presentation, and statistical quantification. Addressing these core issues would greatly improve interpretability and confidence in the findings.

  5. eLife

    Author response:

    We would like to thank the reviewers and the editorial team for all their thoughtful and constructive feedback. The reviewers provided many helpful comments which we will work to incorporate in our resubmission as we believe they will significantly enhance the quality of our manuscript.

    An overarching critique shared among reviewers was regarding limitations in our datasets. Namely, lower N-values for certain groups make some conclusions less reliable. We acknowledge this limitation and will add more experiments to address this concern. Additionally, attention was drawn to our reliance on using the generalized linear model (GLM) for making claims about rebalancing and learning-related changes. To address this, we will work to include additional analyses such as ACC spike-triggered average CA1sup responses, cross-covariances between ACC and CA1sup cells in post-task sleep, and ripple-triggered cross-correlations, among others as per reviewer recommendations. We will also provide a deeper analysis of the weights CA1 neuron in our GLM analysis and their specific features during learning. In accordance, we will provide a clearer description of our learning paradigm including performance data for each animal and how performance relates to our analyses. Overall, we will include more analyses of our datasets across various task events such as recall, to make more efficient use of the full repertoire of our recordings.

    Concerns were also raised regarding some aspects of our statistical analyses. During revision, we will ensure we select the most appropriate statistical measure for each of our tests. Our paper implements the use of tetrode recordings to assess sublayer identification. This approach comes with limitations, and in our resubmission, we will provide a more detailed explanation of those limitations along with a more thorough description of our measures to mitigate them.

    Lastly, in our follow-up submission we will work to improve the written clarity of findings. Specifically, we will simplify and better explain our findings and provide clearer justification for our interpretations and choice of analyses.

  6. Version published to 10.7554/elife.107370.1 on eLife
  7. Version published to 10.7554/elife.107370 on eLife
  8. Version published to 10.1101/2025.03.26.645547v1 on bioRxiv