Distributed subthreshold representation of sharp wave-ripples by hilar mossy cells

Article activity feed

1. 

   Version published to 10.7554/elife.97270 on eLife

   Oct 7, 2025
2. 

   eLife

   Mar 20, 2025

   Author response:

   Reviewer #1 (Public Review):

   We are grateful to this reviewer for her/his constructive comments, which have greatly improved our work. Individual responses are provided below.

   The authors recorded from multiple mossy cells (MCs) of the dentate gyrus in slices or in vivo using anesthesia. They recorded MC spontaneous activity during spontaneous sharp waves (SWs) detected in area CA3 (in vitro) or in CA1 ( in vivo). They find variability of the depolarization of MCs in response to a SW. They then used deep learning to parse out more information. They conclude that CA3 sends different "information" to different MCs. However, this is not surprising because different CA3 neurons project to different MCs and it was not determined if every SW reflected the same or different subsets of CA3 activity.

   Thank you for your …

   Author response:

   Reviewer #1 (Public Review):

   We are grateful to this reviewer for her/his constructive comments, which have greatly improved our work. Individual responses are provided below.

   The authors recorded from multiple mossy cells (MCs) of the dentate gyrus in slices or in vivo using anesthesia. They recorded MC spontaneous activity during spontaneous sharp waves (SWs) detected in area CA3 (in vitro) or in CA1 ( in vivo). They find variability of the depolarization of MCs in response to a SW. They then used deep learning to parse out more information. They conclude that CA3 sends different "information" to different MCs. However, this is not surprising because different CA3 neurons project to different MCs and it was not determined if every SW reflected the same or different subsets of CA3 activity.

   Thank you for your valuable comments. We agree that our finding that different MCs receive different information is unsurprising. These data are, in fact, to be expected from the anatomical knowledge of the circuit structure. However, as a physiological finding, there is a certain value in proving this fact; please note that it was not clear whether the neural activity of individual MCs received heterogeneous/variable information at the physiological level. It was therefore necessary to investigate this by recording neural activity. We believe this study is important because it quantitatively demonstrates this fact.

   The strengths include recording up to 5 MCs at a time. The major concerns are in the finding that there is variability. This seems logical, not surprising. Also it is not clear how deep learning could lead to the conclusion that CA3 sends different "information" to different MCs. It seems already known from the anatomy because CA3 neurons have diverse axons so they do not converge on only one or a few MCs. Instead they project to different MCs. Even if they would, there are different numbers of boutons and different placement of boutons on the MC dendrites, leading to different effects on MCs. There also is a complex circuitry that is not taken into account in the discussion or in the model used for deep learning. CA3 does not only project to MCs. It also projects to hilar and other dentate gyrus GABAergic neurons which have complex connections to each other, MCs, and CA3. Furthermore, MCs project to MCs, the GABAergic neurons, and CA3. Therefore at any one time that a SW occurs, a very complex circuitry is affected and this could have very different effects on MCs so they would vary in response to the SW. This is further complicated by use of slices where different parts of the circuit are transected from slice to slice.

   The first half of this paragraph is closely related to the previous paragraph. We propose that the variation in membrane potential of the simultaneously recorded MCs allows for the expression of diverse information. We also believe that this is highly novel in that no previous work has described the extent to which SWR is encoded in MCs. Our study proposes a new quantitative method that relates two variables (LFP and membrane potential) that are inherently incomparable. Specifically, we used machine learning (please note that it is a neural network, but not "deep learning") to achieve this quantification, and we believe this innovation is noteworthy.

   In the latter part of this article, you raise another important point. First, we would like to point out that this comment contains a slight misunderstanding. Our goal is not to reproduce the circuit structure of the hippocampus in silico but to propose a "function (or mapping/transformation)" that connects the two different modalities, i.e., LFP and Vm. This function should be as simple as possible, which is desirable from an explanatory point of view. In this respect, our machine learning model is a 'perceptron'-like 3-layer neural network. One of the simplest classical neural network models can predict the LFP waveform from Vm, which is quite surprising and an achievement we did not even imagine before. The fact that our model does not consider dendrites or inhibitory neurons is not a drawback but an important advantage. On the other hand, the fact that the data we used for our predictions were primarily obtained using slice experiments may be a drawback of this study, and we agree with your comments. However, we can argue that the new quantitative method we propose here is versatile since we showed that the same machine learning can be used to predict in vivo single-cell data.

   It is also not discussed if SWs have a uniform frequency during the recording session. If they cluster, or if MC action potentials occur just before a SW, or other neurons discharge before, it will affect the response of the MC to the SW. If MC membrane potential varies, this will also effect the depolarization in response to the SW.

   Thank you for raising an important point. We have done some additional analyses in response to your comment. First, we plotted how the SWR parameter fluctuated during our recording time (especially for data recorded for long periods of more than 5 minutes). As shown in the new Figure 1 - figure supplement 4, we can see that the frequency of SWRs was kept uniform during the recording time. These data ensure the rationale for pooling data over time.

   We also calculated the average membrane potentials of MCs before and after SWRs and found that MCs did not show depolarization or hyperpolarization before SWs, unlike Vm of CA1 neurons. These data indicate that the surrounding circuitry was not particularly active before SW, eliminating any concern that such unexpected preceding activity might affect our analysis. These data are shown in Figure 1 - figure supplement 2.

   In vivo, the SWs may be quite different than in vivo but this is not discussed. The circuitry is quite different from in vitro. The effects of urethane could have many confounding influences. Furthermore, how much the in vitro and in vivo SWs tell us about SWs in awake behaving mice is unclear.

   We agree with this point. Ideally, recording in vitro and in vivo under conditions as similar as possible would be optimal. However, as you know, patch-clamp recording from mossy cells in vivo is technically challenging, and currently, there is no alternative to conducting experiments under anesthesia. We believe that science advances not merely through theoretical discourse, but by contributing empirical data collected under existing conditions. However, as we mentioned in the paper, we believe that in vivo and in vitro SWR share some properties and a common principle of occurrence. We also observed that there are similar characteristics in the membrane potential response of MC to SWR. However, as you have pointed out, data derived from these limitations require careful interpretation, and we have explicitly stated in the paper that not only are there such problems, but that there are also common properties in the data obtained in vivo and in vitro (Page 12, Line 357).

   Also, methods and figures are hard to understand as described below.

   Thank you for all your comments. We have carefully considered the reviewers' comments and improved the text and legend. We hope you will take the time to review them.

   Reviewer #2 (Public Review):

   Thank you for the positive evaluations, which have encouraged us to resubmit this manuscript. We have revised our manuscript in accordance with your comments. Our point-by-point responses are as follows:

   • A summary of what the authors were trying to achieve

   Drawing from theoretical insights on the pivotal role of mossy cells (MCs) in pattern separation - a key process in distinguishing between similar memories or inputs - the authors investigated how MCs in the dentate gyrus of the hippocampus encode and process complex neural information. By recording from up to five MCs simultaneously, they focused on membrane potential dynamics linked to sharp wave-ripple complexes (SWRs) originating from the CA3 area. Indeed, using a machine learning approach, they were able to demonstrate that even a single MC's synaptic input can predict a significant portion (approximately 9%) of SWRs, and extrapolation suggested that synaptic input obtained from 27 MCs could account for 90% of the SWR patterns observed. The study further illuminates how individual MCs contribute to a distributed but highly specific encoding system. It demonstrates that SWR clusters associated with one MC seldom overlap with those of another, illustrating a precise and distributed encoding strategy across the MC network.

   We appreciate that this reviewer found scientific value in our manuscript. Thanks to the comments, we were pleased to be able to revise and improve the manuscript. Individual responses are listed below:

   • An account of the major strengths and weaknesses of the methods and results

   Strengths:

   (1) This study is remarkable because it establishes a critical link between the subthreshold activities of individual neurons and the collective dynamics of neuronal populations.

   (2) The authors utilize machine learning to bridge these levels of neuronal activity. They skillfully demonstrate the predictive power of membrane potential fluctuations for neuronal events at the population level and offer new insights into neuronal information processing.

   (3) To investigate sharp wave/ripple-related synaptic activity in mossy cells (MCs), the authors performed challenging experiments using whole-cell current-clamp recordings. These recordings were obtained from up to five neurons in vitro and from single mossy cells in live mice. The latter recordings are particularly valuable as they add to the limited published data on synaptic input to MCs during in vivo ripples.

   We appreciate the reviewer’s critical evaluations, which have encouraged us to revise and resubmit this manuscript. We have revised our manuscript in line with the reviewer’s comments. Our point-by-point responses are provided below:

   Weaknesses:

   (1) The model description could significantly benefit from additional details regarding its architecture, training, and evaluation processes. Providing these details would enhance the paper's transparency, facilitate replication, and strengthen the overall scientific contribution. For further details, please see below.

   Thank you for the suggestions. We have responded with model details based on the following comments.

   (2) The study recognizes the concept of pattern separation, a central process in hippocampal physiology for discriminating between similar inputs to form distinct memories. The authors refer to a theoretical paper by Myers and Scharfman (2011) that links pattern separation with activity backpropagating from CA3 to mossy cells. Despite this initial citation, the concept is not discussed again in the context of the new findings. Given the significant role of MCs in the dentate gyrus, where pattern separation is thought to occur, it would be valuable to understand the authors' perspective on how their findings might relate to or contribute to existing theories of pattern separation. Could the observed functions of MCs elucidated in this study provide new insights into their contribution to processes underlying pattern separation?

   Thank you for your valuable comment. The role of MCs in pattern separation is described in the discussion as follows:

   “It has been shown through theoretical models that MCs are a contributor to pattern separation (Myers and Scharfman, 2011). In general, the pathway of neural information is diverged from the entorhinal cortex through the larger granule cell layer and then compressed into the smaller CA3 cell layer. In this case, there is a high possibility of information loss during the transmission process. Thus, a backprojection mechanism via MCs has been proposed as a device to prevent information loss. Indeed, in theoretical models, such backprojection improves pattern separation and memory capacity, and the results are closer to experimental data than models without built-in backprojection. However, it was unclear what information individual MCs receive during backprojection. Our results show that CA3 SWR is distributed and encoded in the MC population, and that even though the number of MCs is smaller than in other regions, it is possible to reproduce about 30% of the SWR in CA3 from the membrane potential of only five MCs. Based on these results, it is believed that MCs not only play a role in preventing information loss, but also play a role in receiving some kind of newly encoded memory information in the CA3 region, and it is highly likely that the information contained in the backprojections is different from the neural information transmitted through conventional transmission pathways. Indeed, the fact that the information replayed in CA3 is reflected as SWR and propagated to each brain region suggests that the newly encoded memory information in CA3 is propagated to MC. If backprojection simply returned the information transmitted from DG to CA3, and to MC, this would be unrealistic and extremely inefficient. However, it is still unclear what kind of memory information is actually backprojected and distributed to the MC, and how it differs from the memory information transmitted in the forward direction. These are open questions that need to be addressed in future experiments in awake animals.” (Page 11, Line 333)

   (3) Previous work concluded that sharp waves are associated with mossy cell inhibition, as evidenced by a consistent ripple function-related hyperpolarization of the membrane potential in these neurons when recorded at resting membrane potential (Henze & Buzsáki, 2007). In contrast, the present study reveals an SWR-induced depolarization of the membrane potential. Can the authors explain the observed modulation of the membrane potential during CA1 ripples in more detail? What was the proportion of cases of depolarization or hyperpolarization? What were the respective amplitude distributions? Were there cases of activation of the MCs, i.e., spiking associated with the ripple? This more comprehensive information would add significance to the study as it is not currently available in the literature.

   Sorry for confusing the conclusion. First, we did not mention in the paper that in vivo MC depolarized during SWR. The following sentences have added to result:

   “Previous research has shown that the hyperpolarization of MC membrane potential associated with SWR indicates that SWR is related to the inhibition of mossy cells (Henze and Buzsáki, 2007). However, our data showed that the proportion of cases of depolarization or hyperpolarization was about the same, with a slight excess of depolarization. However, it should be noted that MCs are highly active and fluctuating cells, and the determination of whether they are depolarized or hyperpolarized is highly dependent on the method of analysis. Moreover, the firing rate of MCs that we recorded was 1.07 ± 0.93 Hz (mean ± SD from 6 cells, 6 mice), and 6.68 ± 4.79% (mean ± SD from 6 cells, 6 mice, n = 757 SWR events) of all SWRs recruited MC firing (calculated as firing within 50 ms after the SWR peak). ” (Page 5, Line 143)

   (4) In the study, the observation that mossy cells (MCs) in the lower (infrapyramidal) blade of the dentate gyrus (DG) show higher predictability in SWR patterns is both intriguing and notable. This finding, however, appears to be mentioned without subsequent in-depth exploration or discussion. One wonders if this observed predictability might be influenced by potential disruptions or severed connections inherent to the brain slice preparation method used. Furthermore, it prompts the question of whether similar observations or trends have been noted in MCs recorded in vivo, which could either corroborate or challenge this intriguing in vitro finding.

   As you pointed out, one cannot rule out the possibility that this predictability may be influenced by potential disruptions or disconnections inherent in the methods used to prepare the acute slices. And the number of cells is limited to six with respect to the anatomical location of the MC recorded in vivo, making SWR and MC patch clamp recording very difficult even under anesthesia. Therefore, it is difficult to find statistical significance in the current data. We have added following text in Discussion:

   “In addition, the finding that SWR is more predictive when the recorded location of the MC is near the lower blade of the DG is unexpected, so the possibility that this result is influenced by potential disruptions or severed connections during the preparation of the acute slice cannot be ruled out.” (Page 14, Line 405)

   (5) The study's comparison of SWR predictability by mossy cells (MCs) is complicated by using different recording sites: CA3 for in vitro and CA1 for in vivo experiments, as shown in Fig. 2. Since CA1-SWRs can also arise from regions other than CA3 (see e.g. Oliva et al., 2016, Yamamoto and Tonegawa, 2017), it is difficult to reconcile in vitro and in vivo results. Addressing this difference and its implications for MC predictability in the results discussion would strengthen the study.

   Thank you for your comment. We have added the following discussion to your comment:

   “In this study, we performed MC patch-clamp recording both in vivo and in vitro, and clarified that SWR can be predicted from _V_m of MC in both cases. However, there are three caveats to the interpretation of these data. First, the in vivo SWR cannot be said to be exactly the same as the in vitro SWR: note that in vitro SWR has some similarities to in vivo SWR, such as spatial and spectral profiles and neural activity patterns (Maier et al., 2009; Hájos et al., 2013; Pangalos et al., 2013). The same concern applies to MC synaptic inputs. The in vivo _V_m data may contain more information compared to the in vitro single MC data, because the entire projections that target MCs are intact, resulting in a complete set of synaptic inputs related to SWR activity, as opposed to slices where connections are severed. While we recognize these differences, it is also very likely that there are common ways of expressing information. Second, since the in vivo LFP recordings were obtained from the CA1 region, it is possible that the CA1-SWR receives input from the CA2 region (Oliva et al., 2016) and the entorhinal cortex (Yamamoto and Tonegawa, 2017). In addition, urethane anesthesia has been observed to reduce subthreshold activity, spike synchronization, and SWR (Yagishita et al., 2020), making it difficult to achieve complete agreement with in vitro SWR recorded from the CA3 region. Finally, although we were able to record MC _V_m during in vivo SWR in this study, the in vivo data set consisted of recordings from a single MC, in contrast to the in vitro dataset. To perform the same analysis as in the in vitro experiment, it would be desirable to record LFPs from the CA3 region and collect data from multiple MCs simultaneously, but this is technically very difficult. In this study, it was difficult to directly clarify the consistency between CA3 network activity and in vivo MC synaptic input, but the fact that the SWR waveform can be predicted from in vivo MC _V_m in CA1-SWR may be the result of some CA3 network activity being reflected in CA1-SWR. It is undeniable that more accurate predictions would have been possible if it had been possible to record LFP from the CA3 regions in vivo. ” (Page 12, Line 357)

   • An appraisal of whether the authors achieved their aims, and whether the results support their conclusions

   As outlined in the abstract and introduction, the primary aim is to investigate the role of MCs in encoding neuronal information during sharp wave ripple complexes, a crucial neuronal process involved in memory consolidation and information transmission in the hippocampus. It is clear from the comprehensive details in this study that the authors have meticulously pursued their goals by providing extensive experimental evidence and utilizing innovative machine learning techniques to investigate the encoding of information in the hippocampus by mossy cells (MCs). Together, this study provides a compelling account supported by rigorous experimental and analytical methods. Linking subthreshold membrane potentials and population activity by machine learning provides a comprehensive new analytic approach and sheds new light on the role of MCs in information processing in the hippocampus. The study not only achieves the stated goals, but also provides novel methodology, and valuable insights into the dynamics of neural coding and information flow in the hippocampus.

   We appreciate the reviewer’s critical evaluations, which have encouraged us to revise and resubmit this manuscript. We have revised our manuscript in line with the reviewer’s comments.

   • A discussion of the likely impact of the work on the field, and the utility of the methods and data to the community

   Impact: Both the novel methodology and the provided biological insights will be of great interest to the community.

   Utility of methods/data: The applied deep learning approach will be of particular interest if the authors provide more details to improve its reproducibility (see related suggestions below).

   We appreciate that this reviewer found scientific value in our manuscript. Thanks to the comments.

   Reviewer #3 (Public Review):

   We appreciate that this reviewer raised several important issues. We are pleased to have been able to revise the paper into a better manuscript based on these comments. Individual responses are listed below:

   Compared to the pyramidal cells of the CA1 and CA3 regions of the hippocampus, and the granule cells of the dentate gyrus (DG), the computational role(s) of mossy cells of the DG have received much less attention over the years and are consequently not well understood. Mossy cells receive feedforward input from granule cells and feedback from CA3 cells. One significant factor is the compression of the large number of CA3 cells that input onto a much smaller population of mossy cells, which then send feedback connections to the granule cell layer. The present paper seeks to understand this compression in terms of neural coding, and asks whether the subthreshold activity of a small number of mossy cells can predict above chance levels the shapes of individual SWs produced by the CA3 cells. Using elegant multielectrode intracellular recordings of mossy cells, the authors use deep learning networks to show that they can train the network to "predict" the shape of a SW that preceded the intracellular activity of the mossy cells. Putatively, a single mossy cell can predict the shape of SWs above chance. These results are interesting, but there are some conceptual issues and questions about the statistical tests that must be addressed before the results can be considered convincing.

   We appreciate that this reviewer found scientific value in our manuscript. Thanks to the comments, we were pleased to be able to revise and improve the manuscript. Individual responses are listed below:

   Strengths

   (1) The paper uses technically challenging techniques to record from multiple mossy cells at the same time, while also recording SWs from the LFP of the CA3 layer. The data appear to be collected carefully and analyzed thoughtfully.

   (2) The question of how mossy cells process feedback input from CA3 is important to understand the role of this feedback pathway in hippocampal processing.

   1. Given the concerns expressed below about proper statistical testing are resolved, the data appear supportive of the main conclusions of the authors and suggest that, to some degree, the much smaller population of mossy cells can conserve the information present in the larger population of CA3 cells, presumably by using a more compressed, dense population code.

   We appreciate the reviewer’s critical evaluations, which have encouraged us to revise and resubmit this manuscript. We have revised our manuscript in line with the reviewer’s comments. Our point-by-point responses are provided below:

   Weaknesses

   1. Some of the statistical tests appear inappropriate because they treat each CA3 SW and associated Vm from a mossy cell as independent samples. This violates the assumptions of statistical tests such as the Kolmogorov-Smirnov tests of Figure 3C and Fig 3E. Although there is large variability among the SWs recorded and among the Vm's, they cannot be considered independent measurements if they derive from the same cell and same recording site of an individual animal. This becomes especially problematic when the number of dependent samples adds up to the tens of thousands, providing highly inflated numbers of samples that artificially reduce the p values. Techniques such as mixed-effects models are being increasingly used to factor out the effects of within cell and within animal correlations in the data. The authors need to do something similar to factor out these contributions in order to perform statistical tests, throughout the manuscript when this problem occurs.

   Thank you for the insightful comment. As for the correlation between the animals, since they were brought in at the same age and kept in the same environment, we do not think it is necessary to account for the differences due to environmental factors. As the reviewer pointed out, we cannot completely rule out the possibility that within cell or within animal correlation might influence the results, so we plotted the differences in prediction accuracy between cells, slices, and animals (Figure 3 - figure supplement 7). The results showed that prediction accuracy of the real data was better than that of the shuffled data in 66 of the 87 MCs (75.9%). In response to the comment that measurements from the same animal do not constitute independent samples, we have indicated that the average ΔRMSE for each mouse were calculated and these values were significantly different from 0 (n = 14, *p = 0.0041, Student’s t-test). In other words, even if each animal is considered an independent sample, it is possible to obtain statistically significant differences.

   1. A separate statistical problem occurs when comparing real data against a shuffled, surrogate data set. From the methods, I gather that Figure 3C combined data from 100 surrogate shuffles to compare to the real data. It is inappropriate to do a classic statistical test of data against such shuffles, because the number of points in the pooled surrogate data sets are not true samples from a population. It is a mathematical certainty that one can eventually drive a p value to < 0.05 just by increasing the number of shuffles sufficiently. Thus, the p value is determined by the number of computer shuffles allowed by the time and processing power of a computer, rather than by sampling real data from the population. Figures such as 4C and 5A are examples that test data against shuffle appropriately, as a single value is determined to be within or outside the 95% confidence interval of the shuffle, and this determination is not directly affected by the number of shuffles performed.

   Thank you for raising a very good point. We understand the reviewer's comments, but we cannot fully agree with the part that says "It is mathematical certainty that one can eventually drive a p value to < 0.05 just by increasing the number of shuffles sufficiently". This is because when comparing data with no difference at all, no amount of shuffling will produce a significant difference. In this regard, we agree that increasing the number of shuffles will lower the p-value when comparing data with even a small difference. Based on the reviewer's comments, we used a paired t-test to test whether the difference between RMSEreal and RMSEsurrogate was significantly different from 0, and showed it was significantly different (Figure 3 - figure supplement 5). Even when a paired t-test was used for the test, as in Figure 3E, a significant difference in the prediction error of the real and shuffled data was observed for all MC number inputs and also for the in vivo data.

   1. The last line of the Discussion states that this study provides "important insights into the information processing of neural circuits at the bottleneck layer," but it is not clear what these insights are. If the statistical problems are addressed appropriately, then the results do demonstrate that the information that is reflected in SWs can be reconstructed by cells in the MC bottleneck, but it is not certain what conceptual insights the authors have in mind. They should discuss more how these results further our understanding of the function of the feedback connection from CA3 to the mossy cells, discuss any limitations on their interpretation from recording LFPs rather than the single-unit ensemble activity (where the information is really encoded).

   Thank you for your insightful comment. We have added the following text to the discussion:

   “Given that different SWRs may encode information that correlates with different experiences, it is also possible that the activity of individual MCs may play a role in encoding different experiences via SWRs. Indeed, several in vivo studies have confirmed that MC activity is involved in the space encoding (Bui et al., 2018; Huang et al., 2024). However, the relationship with SWRs has not been investigated. The significance of the fact that the SWR recorded from CA3 is reflected in the MC as synaptic input is that it not only shows the transmission pathway from CA3 to MC, but also reveals the information below the threshold that leads to firing, and in a broad sense, it approaches the mechanism by which information processing by neuronal firing. And the expression of synaptic input to the MC is not uniform, but varies in a variety of ways according to the pattern of SWR. Based on previous research showing that diversity is important for information representation (Padmanabhan and Urban, 2010; Tripathy et al., 2013), it is possible that this heterogeneity in membrane potential levels, rather than the all-or-none output of neuronal firing activity, is the key to encoding more precise information. In this respect, our research, which focuses on information encoding at the subthreshold level, may be able to extract even more information than information encoded by firing activity. ” (Page 14, Line 419)

   1. In Figure 1C, the maximum of the MC response on the first inset precedes the SW, and the onset of the Vm response may be simultaneous with SW. This would suggest that the SW did not drive the mossy cell, but this was a coincident event. How many SW-mossy cell recordings are like this? Do the authors have a technical reason to believe that these are events in which the mossy cell is driven by the CA3 cells active during the SW?

   Thank you for your insightful comment. Based on your comment, we have aligned all the MC EPSPs for each SWR onset and found that the EPSPs rise after the SWR onset (Figure 1 - figure supplement 2). This leads us to believe that the EPSP of the MC is most likely driven by the SWR.

   Read the original source
3. 

   eLife

   Mar 18, 2025

   eLife assessment

   Using intracellular in vitro and in vivo recordings and a deep learning approach, this study shows that mouse dentate gyrus mossy cells (MCs) and CA3 pyramidal cells process information from an important electrophysiological hall mark of hippocampus, sharp wave-ripples (SWRs). The innovative use of deep learning to predict SWR waveforms from MC membrane potentials represents an interesting methodological advance. While the key findings are potentially fundamental, some of the evidence is currently incomplete and should be revised to better support the findings.

   Read the original source
4. 

   eLife

   Mar 18, 2025

   Reviewer #1 (Public Review):

   The authors recorded from multiple mossy cells (MCs) of the dentate gyrus in slices or in vivo using anesthesia. They recorded MC spontaneous activity during spontaneous sharp waves (SWs) detected in area CA3 (in vitro) or in CA1 ( in vivo). They find variability of the depolarization of MCs in response to a SW. They then used deep learning to parse out more information. They conclude that CA3 sends different "information" to different MCs. However, this is not surprising because different CA3 neurons project to different MCs and it was not determined if every SW reflected the same or different subsets of CA3 activity.

   The strengths include recording up to 5 MCs at a time. The major concerns are in the finding that there is variability. This seems logical, not surprising. Also it is not clear how deep …

   Reviewer #1 (Public Review):

   The authors recorded from multiple mossy cells (MCs) of the dentate gyrus in slices or in vivo using anesthesia. They recorded MC spontaneous activity during spontaneous sharp waves (SWs) detected in area CA3 (in vitro) or in CA1 ( in vivo). They find variability of the depolarization of MCs in response to a SW. They then used deep learning to parse out more information. They conclude that CA3 sends different "information" to different MCs. However, this is not surprising because different CA3 neurons project to different MCs and it was not determined if every SW reflected the same or different subsets of CA3 activity.

   The strengths include recording up to 5 MCs at a time. The major concerns are in the finding that there is variability. This seems logical, not surprising. Also it is not clear how deep learning could lead to the conclusion that CA3 sends different "information" to different MCs. It seems already known from the anatomy because CA3 neurons have diverse axons so they do not converge on only one or a few MCs. Instead they project to different MCs. Even if they would, there are different numbers of boutons and different placement of boutons on the MC dendrites, leading to different effects on MCs. There also is a complex circuitry that is not taken into account in the discussion or in the model used for deep learning. CA3 does not only project to MCs. It also projects to hilar and other dentate gyrus GABAergic neurons which have complex connections to each other, MCs, and CA3. Furthermore, MCs project to MCs, the GABAergic neurons, and CA3. Therefore at any one time that a SW occurs, a very complex circuitry is affected and this could have very different effects on MCs so they would vary in response to the SW. This is further complicated by use of slices where different parts of the circuit are transected from slice to slice.

   It is also not discussed if SWs have a uniform frequency during the recording session. If they cluster, or if MC action potentials occur just before a SW, or other neurons discharge before, it will affect the response of the MC to the SW. If MC membrane potential varies, this will also effect the depolarization in response to the SW.

   In vivo, the SWs may be quite different than in vivo but this is not discussed. The circuitry is quite different from in vitro. The effects of urethane could have many confounding influences.

   Furthermore, how much the in vitro and in vivo SWs tell us about SWs in awake behaving mice is unclear.

   Also, methods and figures are hard to understand.

   Read the original source
5. 

   eLife

   Mar 18, 2025

   Reviewer #2 (Public Review):

   • A summary of what the authors were trying to achieve
   Drawing from theoretical insights on the pivotal role of mossy cells (MCs) in pattern separation - a key process in distinguishing between similar memories or inputs - the authors investigated how MCs in the dentate gyrus of the hippocampus encode and process complex neural information. By recording from up to five MCs simultaneously, they focused on membrane potential dynamics linked to sharp wave-ripple complexes (SWRs) originating from the CA3 area. Indeed, using a machine learning approach, they were able to demonstrate that even a single MC's synaptic input can predict a significant portion (approximately 9%) of SWRs, and extrapolation suggested that synaptic input obtained from 27 MCs could account for 90% of the SWR patterns observed. The study …

   Reviewer #2 (Public Review):

   • A summary of what the authors were trying to achieve
   Drawing from theoretical insights on the pivotal role of mossy cells (MCs) in pattern separation - a key process in distinguishing between similar memories or inputs - the authors investigated how MCs in the dentate gyrus of the hippocampus encode and process complex neural information. By recording from up to five MCs simultaneously, they focused on membrane potential dynamics linked to sharp wave-ripple complexes (SWRs) originating from the CA3 area. Indeed, using a machine learning approach, they were able to demonstrate that even a single MC's synaptic input can predict a significant portion (approximately 9%) of SWRs, and extrapolation suggested that synaptic input obtained from 27 MCs could account for 90% of the SWR patterns observed. The study further illuminates how individual MCs contribute to a distributed but highly specific encoding system. It demonstrates that SWR clusters associated with one MC seldom overlap with those of another, illustrating a precise and distributed encoding strategy across the MC network.

   • An account of the major strengths and weaknesses of the methods and results
   Strengths:
   (1) This study is remarkable because it establishes a critical link between the subthreshold activities of individual neurons and the collective dynamics of neuronal populations.
   (2) The authors utilize machine learning to bridge these levels of neuronal activity. They skillfully demonstrate the predictive power of membrane potential fluctuations for neuronal events at the population level and offer new insights into neuronal information processing.
   (3) To investigate sharp wave/ripple-related synaptic activity in mossy cells (MCs), the authors performed challenging experiments using whole-cell current-clamp recordings. These recordings were obtained from up to five neurons in vitro and from single mossy cells in live mice. The latter recordings are particularly valuable as they add to the limited published data on synaptic input to MCs during in vivo ripples.

   Weaknesses:
   (1) The model description could significantly benefit from additional details regarding its architecture, training, and evaluation processes. Providing these details would enhance the paper's transparency, facilitate replication, and strengthen the overall scientific contribution. For further details, please see below.
   (2) The study recognizes the concept of pattern separation, a central process in hippocampal physiology for discriminating between similar inputs to form distinct memories. The authors refer to a theoretical paper by Myers and Scharfman (2011) that links pattern separation with activity backpropagating from CA3 to mossy cells. Despite this initial citation, the concept is not discussed again in the context of the new findings. Given the significant role of MCs in the dentate gyrus, where pattern separation is thought to occur, it would be valuable to understand the authors' perspective on how their findings might relate to or contribute to existing theories of pattern separation. Could the observed functions of MCs elucidated in this study provide new insights into their contribution to processes underlying pattern separation?
   (3) Previous work concluded that sharp waves are associated with mossy cell inhibition, as evidenced by a consistent ripple function-related hyperpolarization of the membrane potential in these neurons when recorded at resting membrane potential (Henze & Buzsáki, 2007). In contrast, the present study reveals an SWR-induced depolarization of the membrane potential. Can the authors explain the observed modulation of the membrane potential during CA1 ripples in more detail? What was the proportion of cases of depolarization or hyperpolarization? What were the respective amplitude distributions? Were there cases of activation of the MCs, i.e., spiking associated with the ripple? This more comprehensive information would add significance to the study as it is not currently available in the literature.
   (4) In the study, the observation that mossy cells (MCs) in the lower (infrapyramidal) blade of the dentate gyrus (DG) show higher predictability in SWR patterns is both intriguing and notable. This finding, however, appears to be mentioned without subsequent in-depth exploration or discussion. One wonders if this observed predictability might be influenced by potential disruptions or severed connections inherent to the brain slice preparation method used. Furthermore, it prompts the question of whether similar observations or trends have been noted in MCs recorded in vivo, which could either corroborate or challenge this intriguing in vitro finding.
   (5) The study's comparison of SWR predictability by mossy cells (MCs) is complicated by using different recording sites: CA3 for in vitro and CA1 for in vivo experiments, as shown in Fig. 2. Since CA1-SWRs can also arise from regions other than CA3 (see e.g. Oliva et al., 2016, Yamamoto and Tonegawa, 2017), it is difficult to reconcile in vitro and in vivo results. Addressing this difference and its implications for MC predictability in the results discussion would strengthen the study.

   • An appraisal of whether the authors achieved their aims, and whether the results support their conclusions
   As outlined in the abstract and introduction, the primary aim is to investigate the role of MCs in encoding neuronal information during sharp wave ripple complexes, a crucial neuronal process involved in memory consolidation and information transmission in the hippocampus. It is clear from the comprehensive details in this study that the authors have meticulously pursued their goals by providing extensive experimental evidence and utilizing innovative machine learning techniques to investigate the encoding of information in the hippocampus by mossy cells (MCs). Together, this study provides a compelling account supported by rigorous experimental and analytical methods. Linking subthreshold membrane potentials and population activity by machine learning provides a comprehensive new analytic approach and sheds new light on the role of MCs in information processing in the hippocampus. The study not only achieves the stated goals, but also provides novel methodology, and valuable insights into the dynamics of neural coding and information flow in the hippocampus.

   • A discussion of the likely impact of the work on the field, and the utility of the methods and data to the community
   Impact: Both the novel methodology and the provided biological insights will be of great interest to the community.
   Utility of methods/data: The applied deep learning approach will be of particular interest if the authors provide more details to improve its reproducibility (see related suggestions below).

   Read the original source
6. 

   eLife

   Mar 18, 2025

   Reviewer #3 (Public Review):

   Compared to the pyramidal cells of the CA1 and CA3 regions of the hippocampus, and the granule cells of the dentate gyrus (DG), the computational role(s) of mossy cells of the DG have received much less attention over the years and are consequently not well understood. Mossy cells receive feedforward input from granule cells and feedback from CA3 cells. One significant factor is the compression of the large number of CA3 cells that input onto a much smaller population of mossy cells, which then send feedback connections to the granule cell layer. The present paper seeks to understand this compression in terms of neural coding, and asks whether the subthreshold activity of a small number of mossy cells can predict above chance levels the shapes of individual SWs produced by the CA3 cells. Using elegant …

   Reviewer #3 (Public Review):

   Compared to the pyramidal cells of the CA1 and CA3 regions of the hippocampus, and the granule cells of the dentate gyrus (DG), the computational role(s) of mossy cells of the DG have received much less attention over the years and are consequently not well understood. Mossy cells receive feedforward input from granule cells and feedback from CA3 cells. One significant factor is the compression of the large number of CA3 cells that input onto a much smaller population of mossy cells, which then send feedback connections to the granule cell layer. The present paper seeks to understand this compression in terms of neural coding, and asks whether the subthreshold activity of a small number of mossy cells can predict above chance levels the shapes of individual SWs produced by the CA3 cells. Using elegant multielectrode intracellular recordings of mossy cells, the authors use deep learning networks to show that they can train the network to "predict" the shape of a SW that preceded the intracellular activity of the mossy cells. Putatively, a single mossy cell can predict the shape of SWs above chance. These results are interesting, but there are some conceptual issues and questions about the statistical tests that must be addressed before the results can be considered convincing.

   Strengths
   (1) The paper uses technically challenging techniques to record from multiple mossy cells at the same time, while also recording SWs from the LFP of the CA3 layer. The data appear to be collected carefully and analyzed thoughtfully.
   (2) The question of how mossy cells process feedback input from CA3 is important to understand the role of this feedback pathway in hippocampal processing.
   (3) Given the concerns expressed below about proper statistical testing are resolved, the data appear supportive of the main conclusions of the authors and suggest that, to some degree, the much smaller population of mossy cells can conserve the information present in the larger population of CA3 cells, presumably by using a more compressed, dense population code.

   Weaknesses
   (4) Some of the statistical tests appear inappropriate because they treat each CA3 SW and associated Vm from a mossy cell as independent samples. This violates the assumptions of statistical tests such as the Kolmogorov-Smirnov tests of Figure 3C and Fig 3E. Although there is large variability among the SWs recorded and among the Vm's, they cannot be considered independent measurements if they derive from the same cell and same recording site of an individual animal. This becomes especially problematic when the number of dependent samples adds up to the tens of thousands, providing highly inflated numbers of samples that artificially reduce the p values. Techniques such as mixed-effects models are being increasingly used to factor out the effects of within cell and within animal correlations in the data. The authors need to do something similar to factor out these contributions in order to perform statistical tests, throughout the manuscript when this problem occurs.
   (5) A separate statistical problem occurs when comparing real data against a shuffled, surrogate data set. From the methods, I gather that Figure 3C combined data from 100 surrogate shuffles to compare to the real data. It is inappropriate to do a classic statistical test of data against such shuffles, because the number of points in the pooled surrogate data sets are not true samples from a population. It is a mathematical certainty that one can eventually drive a p value to < 0.05 just by increasing the number of shuffles sufficiently. Thus, the p value is determined by the number of computer shuffles allowed by the time and processing power of a computer, rather than by sampling real data from the population. Figures such as 4C and 5A are examples that test data against shuffle appropriately, as a single value is determined to be within or outside the 95% confidence interval of the shuffle, and this determination is not directly affected by the number of shuffles performed.
   (6) The last line of the Discussion states that this study provides "important insights into the information processing of neural circuits at the bottleneck layer," but it is not clear what these insights are. If the statistical problems are addressed appropriately, then the results do demonstrate that the information that is reflected in SWs can be reconstructed by cells in the MC bottleneck, but it is not certain what conceptual insights the authors have in mind. They should discuss more how these results further our understanding of the function of the feedback connection from CA3 to the mossy cells, discuss any limitations on their interpretation from recording LFPs rather than the single-unit ensemble activity (where the information is really encoded).

   1. In Figure 1C, the maximum of the MC response on the first inset precedes the SW, and the onset of the Vm response may be simultaneous with SW. This would suggest that the SW did not drive the mossy cell, but this was a coincident event. How many SW-mossy cell recordings are like this? Do the authors have a technical reason to believe that these are events in which the mossy cell is driven by the CA3 cells active during the SW?

   Read the original source
7. 

   Version published to 10.1101/2024.03.14.584957 on bioRxiv

   Mar 14, 2024