Single neurons detect spatiotemporal activity transitions through STP and EI imbalance

Reviewer #1 (Public review):

Summary:

This study uses optogenetics to activate CA3 while recordings from CA1 neurons and characterizing the excitation/inhibition (E/I) balance. They observe use-dependent alterations in the E/I balance as a result of STP and they develop a model to describe these observations. This is a very ambitious paper that deals with many issues using both experimental and modeling approaches.

Strengths:

This paper examines important principles regarding the manner in which synaptic circuitry and use-dependent synaptic plasticity can transform inputs and perform computations.

Weaknesses:

There are three issues that cause concern regarding the applicability of their slice recordings to physiological conditions and that make some aspects of their results difficult to interpret. First, they state that 2 mM added external calcium mimics calcium levels in CSF, but this is not the case. This will influence the plasticity they observe. Second, they indicate that there is a 2% decrease in activated fibers per stimulus and attribute this to ChR2 desensitization. Such use-dependent decreases in fiber activation are expected to build during their repetitive activation experiments and artifactually influence their results. Third, they do not know the responses of individual CA3 cells to stimulation. They do not know if each cell fires reliably during repetitive activation and whether each cell only fires once.

Reviewer #3 (Public review):

Summary:

This work shows experimentally and computationally that single CA1 neurons can perform mismatch detection on patterned CA3 inputs and that STP and EI balance underlie this detection.

Strengths:

It has been known that STP can enhance the EPSP when the corresponding presynaptic input exhibits abrupt changes in firing rate. This work provides experimental evidence and further computational support for the hypothesis that the basic computation through STP is useful for detecting abrupt changes in the spatial pattern of synaptic inputs at the Schaffer collaterals. Further, their results indicate the novel view that mismatch detection is most efficient when gamma-frequency bursting inputs exhibit mismatches between theta cycles. The authors included novel results in the revised manuscript to show that the effective frequency range of gamma oscillation is broad, including both slow and fast gamma bands.

In the initial submission, the dependence of mismatch detection performance on model parameters and experimental settings, such as pattern overlaps and other network parameters, was not sufficiently explored. In the revised manuscript, the authors extensively studied these points and summarized the novel results in Fig. 9. Furthermore, the authors clarified that jitters in input spikes can improve detection performance in some cases. These results show the robustness of their results against variations in external and internal conditions.

Weaknesses:

While this study shows an intriguing example of combined experimental and computational studies, some analytic results, for instance, regarding the complex contributions of jitters to detection performance, could have clarified the underlying mechanism deeper and further strengthened the manuscript.

Author response:

The following is the authors’ response to the original reviews.

We have made several major changes in response to the comments and we feel that the manuscript is considerably stronger. In brief: 1. We have added substantial content about homeostasis and EI balance to the introduction. 2. We have addressed concerns about physiological relevance by performing calculations to show that the free calcium in our solutions is well within the physiological range, by citing previous studies showing that short-term plasticity is consistent across 33-38 ℃, and by doing simulations scaled to physiological temperatures to show that the key computational effects are retained. 3. We have addressed concerns about readability by extensive text rewrites, reformatting most of the figures, and by splitting figures into smaller, more focussed ones. 4. We have organized over 20 statistical evaluations and comparisons between our model and experiments into a table. 5. We have carried out additional calculations to examine how the optimal frequency for mismatch detection depends on parameters, and to show that mismatch detection remains even in the presence of stimulus jitter. 6. We have stated more clearly how our proposed mechanism for mismatch detection is based on transient plasticity-mediated skewing of EI-balance, and have added a schematic for the last figure to show this.

Public Reviews:

Reviewer #1 (Public review):

Summary:

This study uses optogenetics to activate CA3, while recording from CA1 neurons and characterizing the excitation/inhibition (E/I) balance. They observe use-dependent alterations in the E/I balance as a result of STP, and they develop a model to describe these observations. This is a very ambitious paper that deals with many issues using both experimental and modeling approaches.

Strengths:

This paper examines important principles regarding the manner in which synaptic circuitry and use-dependent synaptic plasticity can transform inputs and perform computations.

Weaknesses:

The use of selective ChR2 expression in CA3 cells is a good approach, but there are numerous issues that cause concern regarding the applicability of their slice recordings to physiological conditions and that make some aspects of their results difficult to interpret. Experiments are not performed under physiological conditions (high external calcium and low temperature), which makes the interpretation of their findings difficult.

Calcium: We would like to reassure the reviewer that the free calcium levels in our solutions were at ~1.27 mM, well within the physiological range, since our aCSF solution used calcium buffers as well as CaCl2. We have added a section to the methods to show this calculation.

Temperature: Klyachko and Stevens (J. Neurosci 2006) show that the facilitation, augmentation and filtering properties of the CA3-CA1 network were consistent between 33 and 38 degrees C, thus spanning our conditions of ~33 degrees C. Additionally, we have performed simulations to show that the mismatch detection computations remain pronounced (or are even strengthened) when simulation rates for kinetics and channels are scaled to physiological temperatures. Using a Q10 of 2, the scaling term for kinetics is ~37% faster. The outcomes are presented in Figures 7 and 9. We now state these points at the start of the results section:

“Our bath solution had physiological levels of free ions including calcium (methods), and recordings were performed at 32-33 ℃ which has been shown in rats to yield similar short-term plasticity properties as at physiological temperatures (Klyachko and Stevens 2006b).”

We have added a new section to the discussion “Relevance to in-vivo computation” in which we enumerate the caveats but also the points of convergence between our study and physiological conditions, to strengthen the interpretability of our results.

In addition, the reliability of stimulating action potentials in CA3 pyramidal cells needs to be determined, particularly during high-frequency trains. If it is unreliable, there are alternative approaches that might prove to be superior, such as the use of somatically targeted ChR2.

We acknowledge that somatically targeted ChR2 might have slightly improved the sparseness of stimuli, but even such localized expression could lead to unreliability if the position of the soma with respect to the illumination is such that the stimulus is near threshold. Instead, we have adopted a data-driven estimation of CA3 reliability. We reanalyzed our optically-triggered field potential readouts from CA3, to estimate their reliability individually and over trains (Figure 1).

“Notably, the distribution of field amplitudes was very tight (Figure 1E), more so than the corresponding EPSPs (Figure 1H). Together with previous work using a similar optical stimulus system [6] we interpret this to say that the spiking responses from CA3 neurons to optical stimuli were consistent from trial to trial. The field response showed a slight decrease over the course of the pulse train of approximately 2% per pulse (regression fit slope=0.02, r2=0.05). We attribute this to ChR2 desensitization.”

As a further bound to any functional outcomes of CA3 spiking (un)reliability, we point out that CA3-CA1 release probability is low (p~0.2). Any reduction in CA3 reliability is equivalent to reducing the probability of synaptic release, which is already treated as a stochastic process in our simulations. We were able to compare this to experiment as follows: We explicitly modeled the effect of different synaptic volumes as a surrogate for changing p_release in Figure 6-figure supplement 1, and mapped this to our data in Figure 6 D.

“Then we compared the probability that each optical stimulus would elicit an EPSP (Figure 6 D). As expected, 15-square patterns (yellow dots) frequently gave an EPSP (77.5±11.7%), while 5-square patterns failed about half the time (51.4±16%). The simulated runs matched this (Table 1). The probability of failure reduced with increasing volume of the simulated presynaptic boutons, because larger volumes experienced smaller chemical noise (stochasticity) in synaptic release (Figure 6-figure supplement 1). We note that for the purposes of eliciting a postsynaptic response, any unreliability in optical stimulus-triggered firing of the CA3 neuron folds into the probability term for stochastic synaptic release. By matching this metric to experiment, we fine-tuned the volume scaling term for the presynaptic boutons to 0.2”

In addition, a clearer, more detailed discussion of their model that distinguishes it from previous modeling studies would be helpful (and would make it seem less incremental).

This is a good suggestion, as we regard our model as very substantially different from previous studies. We have incorporated this in the discussion as below:

“Our current model is distinct in that it is truly multiscale, closely constrained by experiment, yet runs on modest hardware. It incorporates the network, a conductance based model of a CA1 pyramidal neuron, and chemical kinetic models of a population of stochastic synapses on its dendrite.

Our network model is much reduced compared to models with exhaustive cellular and network-level detail44. Its simplicity enables extensive exploration of the network parameters and comparison with recorded activity under a series of well-controlled stimulus patterns (Figures 4-9).”

We also point out that our proposed mechanism for mismatch detection is an advance over previous ones:

“Leaving aside the obvious differences between auditory cortex and hippocampus, we frame our model as a transient differential tilt in EI balance (Figure 3, Figure 8A,B, Figure 10B), in distinction to the fresh-afferent model. This makes our model robust over a wide range of stimulus and network conditions (Figure 9), and has the functional implication that transient responses remain at about the same amplitude over a prolonged stimulus sequence (Figure 8B, Figure 10B), rather than declining.”

Reviewer #2 (Public review):

Summary:

The authors investigate EI balance in the CA3-CA1 projections, emphasizing synaptic depletion and the implied rebalancing of excitatory and inhibitory projections onto a single CA1 Pyramidal cell. They present physiological results with optical stimulation in CA3 and measuring various response features in CA1, showing signatures consistent with the adjustment of EI balance. In particular, the authors emphasize a transient effect where the neuron escapes from EI balance, which can be used for mismatch detection. They partially replicate these results in a computational model that looks at detailed properties of synaptic plasticity in CA1.

Strengths:

The authors provide compelling evidence that non-specific modulation of synaptic plasticity, combined with their differential effects on excitatory and inhibitory neurons, can be used by CA1 excitatory neurons to detect changes in the population activity of CA3 neurons. Indeed, they provide insight into the potential computational role of transient EI imbalance.

Weaknesses:

The authors observe that "little is known about how EI balance itself evolves dynamically due to activity-driven plasticity in sparsely active networks." This is an overstatement, or better an understatement, given the extensive literature on EI balance (e.g. Wen W, Turrigiano GG. Keeping Your Brain in Balance: Homeostatic Regulation of Network Function. Ann Rev Neurosci. 2024. https://doi.org/10.1146/annurev-neuro-092523-110001 PMID:38382543). This way of framing the question does a disservice to the field and fails to contextualize the current research properly.

We agree that we could have presented this better. Our focus was on short-term (<1 second) EI balance changes, but our statement did not set this context clearly. We rewritten and expanded the introduction to place our work in context of the substantial previous work on plasticity and homeostasis in EI balance.

The evidence is incomplete because the authors do not show a specific relationship between synaptic change in CA1 and EI balance adjustment, i.e., the alternative could be that this is an unspecific effect unrelated to the specific regulation of EI balance and its functional role in the hippocampus and the cortex.

We don’t quite follow this point. We have devoted Figures 2 and 3 to showing a specific relationship between short-term plasticity on CA3->CA1 synapses, and EI balance. In Figure 2 we show how E and I responses evolve over a pulse train. In Figure 3 we explicitly show the plasticity in E and I synapses, and then map it onto EI balance. In Panel 3E to G all these points come together and we show how gamma (the measure of nonlinearity of summation) evolves over a series of pulses in parallel with plasticity in E and I. We have added some new data in Figure 7A, B to show how E and I contribute to mismatch detection.

Indeed, the paper drifts from addressing EI balance to elucidating the mismatch detection.

We acknowledge that we did not sufficiently articulate the role of EI balance terms in our subsequent analysis of mismatch detection. We have added several figure panels (Figure 7A, B), added a summary schematic (Figure 10) and redone the text and discussion. With these changes we make the point that mismatch detection can be better framed as a transient shift in EI balance.

“we frame our model as a transient differential tilt in EI balance (Figure 3, Figure 8A,B, Figure 10B), in distinction to the fresh-afferent model. This makes our model robust over a wide range of stimulus and network conditions (Figure 9), and has the functional implication that transient responses remain at about the same amplitude over a prolonged stimulus sequence (Figure 8B, Figure 10B), rather than declining.”

The second shortcoming is that they do not show that the stimulation of the CA3 neurons occurs in a physiologically realistic regime.

We have responded to the concerns about calcium concentration and temperature above in the response to the first reviewer. From the text:

“Our bath solution had physiological levels of free ions including calcium (methods), and recordings were performed at 32-33 °C which has been shown in rats to yield similar shortterm plasticity properties as at physiological temperatures (Klyachko and Stevens 2006b).”

In addition, there is a concern about the mapping between physiological activity and our stimuli. It is true that the patterned stimuli we delivered were artificial. We make the point that they are nevertheless a much closer map to sparse physiological patterns than conventionally obtained through Schaffer collateral volleys:

“We use optical patterned stimuli to stimulate a cross-section of CA3 neurons with a variety of distributed patterns, theta, and other frequency rhythms. These stimuli are sparser and more dispersed than Schaffer collateral electrical stimuli which tend to stimulate adjacent fibres and in most cases are very strong.”

We have added a section to the discussion “Relevance to in-vivo computation” to more completely address these points.

Nor do they analyze what the impact will be of the excitatory transient in "mismatch detection", and CA1,

We are unsure what the reviewer means by the excitatory transient. At the level of CA3, we observe a narrow optically triggered field response for each light pulse. At the level of CA1, we monitor the responses due to activation of E and I synapses, and are able to observe peaks for each of the light pulses. We have analyzed all these features in figures 1 through 3, and they are also explicitly included in the model. Based on the reviewer’s comment we have further characterized the field responses in CA3:

“We observed a small amount of ‘ringing’ of the field response which we interpret as either CA3 spiking in a burst, or recurrent activation of the CA3 neurons (Figure 1 supplement 2). The ringing was down to ~5% within 8 ms, supporting our treatment of the optical input as a tightly time-delimited event, and setting a low bound to any contribution to patterns by recurrence.”

When this would occur at the level of the whole population, i.e., the physiological impossibility of triggering uncontrolled chaotic excitatory responses.

Again, we are unsure what population or chaotic responses the reviewer has in mind. As mentioned above we have further characterized the field readouts of population responses in CA3 and have established tight limits on recurrent activity (Figure 1-figure supplement 2). In case the reviewer is looking for the outcome at the entire CA1 network as a whole, our experiment figures 1GH,J,K,L show sharp, single peak CA1 neuronal responses.

In particular, when we consider CA3 as an attractor memory system, the range of deviations (mismatches) that a CA1 neuron can be exposed to and detect, given the model presented in this paper, might be below those generated due to CA3 pattern-completion dynamics.

While this is an interesting question for further work, our study focuses on a tighter question, that of mismatch detection downstream of the CA3. As indicated above and in Figure 1figure supplement 2, our field and patch recordings show that under our stimulus conditions, the internal dynamics of the CA3 produce minimal delayed or recurrent signals. Thus, by design, the CA3 layer in our system acts as an almost pure input layer with minimal internal dynamics. In the discussion we address some of the possibilities that may arise from pattern computations in CA3 and other upstream areas:

“We speculate that upstream areas may encode higher order stimulus features such as gaps, duration, intensity, localization, and frequency steps into distinct input patterns. Our proposed EI-balance shift mechanism could be a common end-point for all of these. This would transform quite complex mismatch detection tasks into a uniform computation of pattern change, generalizing the mechanism to stimuli which were previously considered to require a more complex network-level implementation”

In addition, the match between the model and the physiological results is not fully quantified, leaving it to the reader to make a leap of faith.

While the original version had numerous points of comparison between physiology and model, we agree that the values were scattered. In this revision we have tabulated them and performed additional statistical comparisons between model and data for a total of over 20 comparisons for the cell electrophysiology and network readouts (Table 1). We have also organized the preceding chemical kinetic comparisons in the supplements to Figure 4. We regard our study as one of very few to undertake quantitative experimental comparisons over such a range of readouts, experiments, and scales.

In addition, the manuscript suffers from poor analysis and presentation. The work could be improved by putting more effort into translating results into insightful metrics.

We acknowledge that the presentation needed improvement. We have performed a major rewrite and reorganized many of the figures. As mentioned above, we have tabulated numerous metrics (Table 1) and have characterized EI balance and its evolution due to plasticity in a pulse train (Figures 2 and 3). For higher-level metrics, the new figures now extensively explore how mismatch sensitivity depends on parameters, stimulus patterns, and repeat frequency (Figures 7, 8, 9). We have added a discussion section “Relevance to invivo computation”

Overall, the authors have not achieved their original aim to show that the observed phenomenon is relevant to computation in CA1 or the brain outside of a highly controlled in vitro setup and reductionist single cell model.

We feel that with this revision we have more clearly shown that our measurements are relevant to in-vivo computation, both through improved clarity and additional analysis. We have added a section “Relevance to in-vivo computation” in the discussion which enumerates the steps we have taken to support the relevance of our study. In the revision we have also performed several modelling extrapolations which encompass in-vivo conditions, such as testing jitter and frequency range. In a broader sense, in vitro work by design, is meant to be highly controlled so as to be able to get at mechanisms, and in our study we have delivered a range of physiologically relevant stimulus combinations to bridge the gap.

The authors combine several techniques for in vitro whole-cell patch-clamp recordings with patterned optical stimulation of the CA3 network in the mouse hippocampus, which is consistent with the state-of-the-art.

They introduce a metric of similarity between expected and observed response patterns, called gamma. The name is confusing given the wide use of the label gamma for oscillation frequencies above 20 Hz. Gamma is calculated as (E*O)/(E-O). This means that gamma approximates infinity as the difference goes to 0, to mention one of the problems. This metric is not interpretable, and it is not clear why the authors did not follow a standard approach, e.g., likelihood, correlation, or percent error.

We acknowledge the potential for confusion, however we felt it would be more confusing to change nomenclature. The metric gamma is derived from previous published work (Bhatia et al, eLife 2019) describing nonlinearities in summation, which is cited. In that study and the current one, there was no instance in which gamma became unreasonably large. It is true that the term gamma is used for many concepts, but we feel that the contexts are so different between summation nonlinearity and oscillation frequencies that confusion is unlikely. We have taken care with the wording in the text to further disambiguate the usage.

The authors aim to replicate the physiological results with an "abstract model of the hippocampal FFEI network. In practice, this is a conductance-based model of a single CA1 neuron, including chemical kinetics-based multi-step neurotransmitter vesicle release. This is an abstraction from the FFEI network that the paper starts with.

We stress that the full model was used for all simulations except synaptic chemistry parameter fitting. We have clarified this point in the text and discussion section. From the text following Figure 4:

“We used this full model, with optical stimulus, CA3, Interneurons, CA1 neuron, probabilistic connectivity, and presynaptic signaling chemistry, for all subsequent calculations in this study.”

The model has 256 integrate-and-fire CA3 neurons, 256 interneurons, plus 200 inhibitory and 100 excitatory synapses onto the CA1 neuron, in each of which we have distinct multistep transmitter release kinetics.

It raises the question whether this is the right level at which to model the computational impacts of EI imbalance on CA1 neurons. Given the highly reduced model they have elaborated, the generalization to the complete CA3-CA1 network that the authors suggest can be achieved in the discussion is overoptimistic. Network models of CA3 and C1 must be considered, together with afferents from the entorhinal cortex to accomplish this generalization.

We hope we have clarified that we do indeed base all our calculations on the full FFEI model converging onto the CA1 neuron whose connectivity influences circuit function, and we feel that this is necessary and sufficient for our goals in this study.

While the role of the recurrent CA3 network and EC would be interesting topics for future work, the scope of our study is to model the computational impact of EI imbalance in the FFEI network of CA3-> CA1 on CA1 neurons.

The authors reveal a potentially interesting physiological feature of CA1 excitatory neurons under very specific stimulus conditions.

We thank the reviewer for considering the work as interesting. We would like to clarify, however, that our stimulus conditions are actually multidimensional. Specifically, we have varied frequency, pattern, and number of inputs for burst stimuli, and we have also examined Poisson train inputs. In the model we have examined spiking responses, and theta modulated stimuli. In the revision we have also included jittered synaptic input, and obtained frequency dependence of the mismatch detection. To our knowledge this is among the more multidimensional stimulus-response and modeling studies on this system.

It could warrant follow-up studies to place EI imbalance in a physiologically realistic context.

Reviewer #3 (Public review):

Summary:

This work shows experimentally and computationally that single CA1 neurons can perform mismatch detection on patterned CA3 inputs and that STP and EI balance underlie this detection.

Strengths:

It has been known that STP can enhance the EPSP when the corresponding presynaptic input exhibits abrupt changes in firing rate. This work provides experimental evidence and further computational support for the hypothesis that the basic computation through STP is useful for detecting abrupt changes in the spatial pattern of synaptic inputs at the Schaffer collaterals. Further, their results indicate the novel view that mismatch detection is most efficient when gamma-frequency bursting inputs exhibit mismatches between theta cycles.

Weaknesses:

Their model assumes that patterned activities in CA3 do not have overlaps. However, overlaps between memory engrams have been shown. Therefore, this assumption may not hold, and whether the proposed mechanism is valid for overlapping CA3 inputs needs further clarification.

We see that our account of the methods needs clarification, since we explicitly incorporate overlap in our model. First, from the experiments themselves, we say that we expect overlap:

“This was also consistent with the observation of a wide field of excitability around individual CA3 neurons [6] (Figure 1-figure supplement 1). From this we expect that there is some overlap in the sets of CA3 neurons activated by different patterns, and this overlap increases with more stimulus squares.”

In the model, we systematically examine the effect of overlap and have added several figures to make the point (Figure 9 Bi, Figure 9Ci, Figure 4-figure supplement 6, Figure 7figure supplement 1, Figure 9-figure supplement 1).

Recommendations for the authors:

Reviewer #1 (Recommendations for the authors):

The use of selective ChR2 expression in CA3 cells is a good approach, but there are numerous issues that cause concern regarding the applicability of the slice recordings to physiological conditions and that make some aspects of the results difficult to interpret.

Weaknesses:

(1) Some aspects of this study seem somewhat incremental. There is a rich literature on the study of excitation and inhibitory synapses and the issue of EI balance. There are a great many related studies that are not cited (off the top of my head: Pouille and Scanziani 2001, Mittmann, Chadderton and Hausser 2004, Atallah and Scanziani 2009, but there are many, many more). A great many of the ideas presented in this study have already been published previously (Klyachko and Stevens, 2006, and numerous other related manuscripts).

We agree that the topic of EI balance has a very substantial literature. We have incorporated many of the mentioned articles and others in our introduction and discussion. Our study explicitly links several strands of work on EI balance with short-term plasticity and spatial patterning:

“The current study integrates several research themes of EI balance, short-term plasticity, and network computation to systematically characterize and model the properties of a network with feedforward inhibition. We complete the experiment-model-prediction-testing loop and show that differential changes on E and I synapses may provide a mechanism for single neurons to extract interesting features of spatiotemporal inputs through STP (Asopa and Bhalla 2023), while keeping mean activity steady.”

We find that our sparse optical stimulation protocol gives qualitatively distinct results, and is amenable to investigation of more complex spatial pattern dependent effects. We have explicitly discussed the mentioned paper by Klyachko and Stevens, and numerous others, to point out where our study differs. From the discussion:

“For example, studies using field electrode stimulation of the Shaffer collaterals report a sustained shift to excitation during burst input (Klyachko and Stevens 2006a). In contrast, our sparse optical patterned stimuli results in a small window of escape from EI balance around pulse 2 or 3 in a burst (Figure 3), following which both E and I undergo depression to restore balance (Figure 3, 8). Thus, spatial patterning intersects with short-term plasticity to add another layer of timing control through gating of E-I balance.”

(2) There are multiple technical issues that call into question the relevance of this study for physiological conditions and the study of STP.

(a) Their experiments were performed in elevated external calcium (2 mM) compared to physiological calcium (1.1-1.5 mM). This will have a major influence on the probability of release and short-term plasticity.

This concern does not take into account the composition of our solution, which incorporated calcium buffers to give free calcium levels of ~1.27 mM. We have provided detailed calculations in the methods section.

“Our bath solution had physiological levels of free ions including calcium (methods), and recordings were performed at 32-33 °C which has been shown in rats to yield similar short-term plasticity properties as at physiological temperatures (Klyachko and Stevens 2006b).”

(b) Their experiments were performed at reduced temperatures (32-33 {degree sign}C). This is alright for many studies, but this is an important deficiency for the particular issue of EI balance and STP, and the relevance of conclusions based on these conditions.

Klyachko and Stevens (J. Neurosci 2006) show that the facilitation, augmentation and filtering properties of the CA3-CA1 network were consistent between 33 and 38 degrees C, thus spanning our conditions of ~33 degrees C. Additionally, we have performed simulations to show that the mismatch detection computations remain pronounced (or are even strengthened) when simulation rates for kinetics and channels are scaled to physiological temperatures. Using a Q10 of 2, the scaling term for kinetics is ~37% faster. The outcomes are presented in Figures 7 and 9.

(c) I like the selective expression of ChR2 in CA3 pyramidal cells, but they have not provided any information on the effect of stimulation on the firing of CA3 cells (Extended Data Figure 1 is not enough). Is it reliable for single stimuli or stochastic?

We have used field recordings in the CA3 to put tight bounds on the properties of CA3 firing (Figure 1, Figure 1-figure supplementar 2.) The field recordings show that on average the firing is highly reliable. We explicitly characterize the probability of eliciting EPSPs through Poisson patterned stimuli in Figure 6 D. As discussed in the text (excerpted below) any stochasticity in firing folds into the parameters for p_release, and synaptic firing is itself stochastic.

We note that for the purposes of eliciting a postsynaptic response, any unreliability in optical stimulus-triggered firing of the CA3 neuron folds into the probability term for stochastic synaptic release.

Do CA3 cells fire once or multiple times?

This was a useful point, and we examined our field potential data more closely based on this.

“We observed a small amount of ‘ringing’ of the field response which we interpret as either CA3 spiking in a burst, or recurrent activation of the CA3 neurons (Figure 1-figure supplement 2). The ringing was down to ~5% by the third peak which occurred within 8 ms, supporting our treatment of the optical input as a single brief event, and setting a low bound to any contribution to patterns by recurrence.”

Are the spikes precisely timed, or do they vary?

Based on the field potentials, the spikes are precisely timed (Figure 1-figure supplement 2D, E).

“fEPSP Peak Width distribution centred around 1.2 ms, but no peak was wider than 1.6 ms, suggesting tight synchrony in case multiple CA3 neurons were spiking.”

Are there use-dependent changes in the ability of optogenetic stimulation to evoke spiking?

Yes, and this is characterized in figure 1 panel I. The decrement is about 2% per pulse.

The CA3 regions are highly interconnected with recurrent collaterals. Does stimulation during trains alter the activity in the CA3 region as a result of these collaterals?

Based on the CA3 field recordings, almost all CA3 activity is optically triggered (Figure 1figure supplement 2). Figure 1C shows narrow fEPSPs in a burst.

This would be a particularly important issue during trains. Would they have gotten more readily interpretable results if they had used a somatically targeted ChR2 variant?

We feel it is unlikely that a somatically targeted ChR2 would change outcomes. All our analysis assumes overlap of excitation of CA3 pyramidal neurons, that is, a given spot illuminates multiple cells to different degrees, and that there will be neurons which are activated by more than one spot. Somatic targeting does not eliminate activation due to scattering and out-of-focal-plane illumination.

In extended Figure 5, they show stimulus patterns used to stimulate. I need some more explanation. Are they stimulating in the cell body region only, or are they stimulating in the vicinity of dendrites?

Extended Figure 5 (now Figure 4-figure supplement 6) indicates the stimulus patterns in the model. The experimental illumination pattern was 336µm x 187.2µm oriented so that the long axis of the pattern lay along the CA3 cell body layer (methods). Sample stimulus patterns are illustrated in Figure 1 panels D, G and J. Given scatter and out-of-plane illumination we expect that dendrites will also be stimulated. This is corroborated in Figure 1figure supplement 1 where we find that in addition to a strong ‘receptive field’ at the soma, there is a dispersed region of weaker activation. We cannot say definitively whether this dispersed region is due to light scatter, out-of-plane illumination, or dendritic activation. However, even somatically targeted ChR2 would elicit multi-neuron activity due to scatter and out-of-plane soma activation.

If that is the case, there are a great many complications that arise, and it seems to be an approach that could unreliably activate a great many CA3 cells.

We have now put in a paragraph to discuss this, and to set bounds to the unreliability.

“To monitor the strength and consistency of the total resultant optogenetic activation of the CA3 layer, we used an extracellular field electrode in the CA3 stratum radiatum (Figure 1A, methods). The field response correlated well with optically-driven CA1 PC depolarization (Figure 1E-G), and scaled with the size of the pattern (Figure 1F). This was also consistent with the observation of a wide field of excitability around individual CA3 neurons (Bhatia et al. 2019) (Figure 1-figure supplement 1). From this we expect that there is some overlap in the sets of CA3 neurons activated by different patterns, and this overlap increases with more stimulus squares. Notably, the distribution of field amplitudes was very tight (Figure 1E), more so than the corresponding EPSPs (Figure 1H). Together with previous work using a similar optical stimulus system (Bhatia et al. 2019) we interpret this to say that the spiking responses from CA3 neurons to optical stimuli were consistent from trial to trial.”

We also note that any CA3 firing unreliability folds into the stochastic release terms, as discussed in an earlier point.

(d) As far as I can tell, they did not examine the effects of blocking NMDA receptors in their slice experiments. This seems like a very important experiment to perform if they really want to understand EI balance.

The reviewer is correct that we did not block NMDA receptors. While this would have teased apart contributions of NMDAR and AMPAR to the overall response, our analysis of EI balance required the intact synapse and hence this decomposition (which has been done in previous studies) was not needed for our analysis.

Based on a-d it is not clear that their conclusions regarding EI balance and STP are relevant under physiological conditions, and their findings are difficult to interpret.

We have addressed the concerns about physiological conditions when it comes to the Ca2+ levels and temperature. We do not feel that points c and d alter the interpretation of our findings.

Minor:

(3) Their model has only 1 type of interneuron, whereas there are many. CA3-interneuron synapse has very different plasticity for different types of interneurons, and different types of interneuron synapses onto different parts of the CA3 cell. They need to justify lumping all of these types of interneurons.

We agree that our model had a coarse-grained representation of interneurons as a single class. We feel this is an appropriate level of detail because it fits well for our experiments, and keeps the model tractable.

“We have, of course, simplified the network, most notably in the use of only one inhibitory interneuron class which maps to parvalbumin-positive fast-spiking interneurons with perisomatic connectivity. This level of detail was chosen as it was able to quantitatively fit a large number of observations with minimal circuit complexity.”

(4) How many parameters can they adjust in their model? It seems that with so many parameters, their model is not very good at times (extended Figure 3B and E, for example).

Our model has 6 free parameters for the network (Table 1), and another 7 parameters each for the E and I presynaptic plasticity models (Figure 4A and Supplementary Data). The presynapse plasticity parameters are directly assigned from the burst response recordings using the parameter fitting as described in the Methods. Normalized RMS differences between model and experiment for presynapse parameters are presented in Figure 4-figure supplements 1-3, panel F. Most traces lie below 0.3, which is a good fit. We have now tabulated numerous comparisons between model and experiment (Table 1). In all but 1 of 20 tests, the model value lies within the experimental range.

“Overall, we were able to quantitatively replicate almost all features of the experimental dataset in our multiscale model incorporating presynaptic signalling, postsynaptic electrophysiology, and abstracted network connectivity and responses. Between the datasets in Figure 4-figure supplements 1 to 3, Figure 5, and Figure 6, we were able to substantially constrain the parameters in our model, from chemical to cellular physiology to network.”

Additionally, we have included a new Figure 9 to systematically do parameter sweeps. From this we conclude:

“...mismatch detection in our model is robustly present and can be tuned over a wide range of network parameters and model assumptions, with the notable exception that it is absolutely dependent on the presence of STP.”

(5) They use the term short-term potentiation (STP), but plasticity is not just enhancement; there is also depression. That is why many others opt for the more inclusive "short-term plasticity".

We agree that this was unclear. We meant to use “Short Term Plasticity” and have now clarified this in the text.

Reviewer #2 (Recommendations for the authors):

The paper is poorly written and would benefit from a more careful preparation of the manuscript. In the opinion of this reviewer, it does not meet the expected quality for a paper of this type. Reviewing the paper was somewhat frustrating, requiring puzzling through details that were not well described. Also, failing to put clear labels on figures and their low quality did not help.

We have worked substantially on the readability in the revision. We have made numerous changes to the text and figure legends, and have reworked several figures, with the goal of addressing concerns about readability.

The introduction lacks proper context for EI balance and the hippocampus.

We have substantially rewritten the introduction to more clearly place our work in the context of the relevant literature. We touch upon short-term plasticity and computation, on homeostasis, on EI balance and on network correlates of plasticity such as mismatch detection.

The data analysis is superficial, and insufficient effort is put into compressing complex data into insightful metrics.

We have done substantial rewrites to address this concern. There are two kinds of metrics we have developed for this study: those that measure the goodness of fit between simulations and data (consolidated into Figure 4-figure supplements 1 to 3 and in Table 1), and those which capture high-level features such as sublinearity of summation due to EI balance (Figure 3), selectivity for mismatch detection (Figures 7 to 9), and peak frequency for mismatch selectivity (Figure 9). We have also performed additional simulations as per reviewer suggestions, which give metrics for dependence of transition detection on network parameters, and for sensitivity of mismatch detection to input spike jitter.

The only attempt to do this was the gamma measure, which left one wanting (see above).

We have responded to the points about the gamma measure above.

Figures are low-quality, labels are missing,

We have substantially reworked figures, their labels, and legends. The automated mapping from our high-resolution figures to PDF seems to have blurred many of the figures, however, links to the originals should be there in the revision.

And the analysis stays too close to the data without presenting a clear quantitative synthesis and insight.

Please see response above. We have tried to balance the process of characterizing numerous readouts and making a model that closely matches experiment, with the high-level insights by way of computational outcomes such as mismatch detection in a variety of more physiological contexts (pulse trains and theta patterned inputs, Figures 7 to 9).

Key results and mapping between physiology and the model are kept subjective and not quantified.

Please see response above. We have consolidated our comparisons between physiology and experiments into Figure 4-figure supplements 1 to 3 and Table 1.

In addition, the similarity measure gamma, which is introduced to express the relationship or the modulation of the response, is mathematically naïve and not well-motivated. It will approach infinity when expected and actual values become more and more similar. While this might be the range where sensitivity is required.

Please see response above. The metric gamma is derived from previous published work (Bhatia et al, eLife 2019) describing nonlinearities in summation, which is cited. In that study and the current one, there was no instance in which gamma became unreasonably large. It is true that the term gamma is used for many concepts, but we feel that the contexts are so different between summation nonlinearity and oscillation frequencies that confusion is unlikely. We have taken care with the wording in the text to further disambiguate the usage.

Some detailed observations:

P2: What is an "interesting" feature?

We have replaced the word “interesting” with “salient”:

“We complete the experiment-model-prediction-testing loop and show that differential changes on E and I synapses may provide a mechanism for single neurons to extract salient features of spatiotemporal inputs through STP (Asopa and Bhalla 2023), while keeping mean activity steady.”

P6 L110: However, over the pulse train, E and I underwent distinct STP profiles (Figure 1 M).

What makes them distinct?

This panel is now removed. A clearer account is presented in Figure 2D,E and F:

“The EPSC showed a trend of early potentiation followed by depression (Figure 2D, 2E), while the inhibition underwent depression from the start (Figure 2 D, Fi)”

P6 L115: Why can recurrent excitation in the CA3 segment be excluded?

We thank the reviewer for pointing us to a more detailed analysis, which is now presented in Figure 1-figure supplement 2. We have added the following text:

“We observed a small amount of ‘ringing’ of the field response which we interpret as either CA3 spiking in a burst, or recurrent activation of the CA3 neurons (Figure 1-figure supplement 2). The ringing was down to ~5% by the third peak which occurred within 8 ms, supporting our treatment of the optical input as a single brief event, and setting a low bound to any contribution to patterns by recurrence.”

P8 F2A: How are the responses normalized?

In the text we state:

“All the PSPs of an 8-pulse train were normalised to the probe pulse.”

We have added this line into the legend.

“Traces were normalised to a reference pulse 0, delivered 300ms before the burst.”

Explain why, given this normalization, the 15 square stimulation is less effective than the 5 square one.

We acknowledge this was unclear. In the revised text we explain:

“For the EPSCs, the 15-square trials had a higher reference pulse and higher stimulus overlap (discussed below), hence their normalised peak values were smaller (Figure 2E).”

F2D: Where do you show that the biphasic response is a statistically significant deviation?

Thank you for pointing out this missing analysis. We have added it in Figure 3A.

P9 149: E should be E&F.

Corrected.

P9 L150: Explain the "ii" indexing.

Corrected.

P10: It is a bit clumsy to call the measure gamma. For general observation on the equation, see the general remark above.

Please see discussion on this. We are reusing a published term.

P10 L175: How do your results and F3G show divisive inhibition?

In the current study we’re not setting out to show divisive inhibition, as that work has been published (Bhatia et al, eLife, 2019). We’ve corrected the text accordingly.

“Using responses from the reference pulse, we replicated earlier observations (Bhatia et al. 2019; Wehr and Zador 2003) showing divisive normalisation, and obtained a median gamma of 7.16 (95% CI = 4.76 - 10.2)(Figure 3G).”

Becomes

“By comparing observed vs. expected responses, we replicated earlier observations (Bhatia et al. 2019; Wehr and Zador 2003) showing sublinear summation, and obtained a median gamma of 7.16 (95% CI = 4.76 - 10.2) (Figure 3G).”

P16: How does F5 demonstrate a good match between model and physiology?

We acknowledge we left this out. In F5E we show the model and experiment distributions over different frequencies. Our previous analysis only reported frequency dependence, and now we have added the comparison of response amplitudes. We have inserted the analysis and consolidated the results into Table 1.

P18 l281: 15-square patterns (yellow dots) almost always gave an EPSP, while 5-square patterns frequently failed.

Where can I see this? It is mentioned in the caption, but legends are absent.

In the original source file and in the original confirmation pdf from eLife, the figure legend is present, and has an entry for panel C and D.

“C,D:probability of trigger to generate a peak in the EPSP trace”

In the revised version we have quantified these values and put the comparisons into Table 1:

“Then we compared the probability that each optical stimulus would elicit an EPSP (Figure 6 D). As expected, 15-square patterns (yellow dots) frequently gave an EPSP (77.5±11.7%), while 5-square patterns failed about half the time (51.4±16%). The simulated runs matched this (Table 1).”

P19 l307: Overall, we were able to replicate numerous features...

Please be specific. What exactly did you replicate? How is it statistically demonstrated?

This is a good point, we have updated the text to more systematically work through comparisons and metrics. We have also added some further metrics for features of the responses in Figures 5 and 6. As a way to organize all our comparisons we have added Table 1.

P22 l341: The transient responses must be proportional to the overlap. Please quantify this effect more precisely.

In Fig 7 panels L and O we had previously quantified the amplitude of transient responses with respect to two parameters closely related to overlap: pattern sparseness and probability of connections from CA3 to CA1. In Figure 9Bi we show that there is a complex and frequency-dependent relationship between overlap and mismatch responses. In the revision in figures 8 and 9 we have recast the “pattern sparseness” term as the more intuitive “overlap”. These are related almost linearly with a negative slope (Figure 9-figure supplement 1).

P22 l342: What does "in E" mean?

Should be Figure 7E for the original version. In the revised paper we have removed this panel.

l347: I cannot follow. How do these single traces (7C-E) show these effects?

We acknowledge that the figure and legend did not clearly indicate the timings of the transitions. We have completely redone and reduced figure 7 to simplify the presentation. The timing of transitions between patterns is now indicated using red triangles.

What does denser connectivity refer to?

Denser connectivity refers to a higher value for probability of connection between CA3 and CA1. In the revised version we have changed the figure to refer to stimulus overlap:

“None of the transitions in Figure 8D (dense stimuli, 34% overlap) were significant, but two transitions in Figure 8E were significant (sparse stimuli with 2.5% overlap, p = 1.53e-5 and 6.1e-5).”

P26: It is unreasonable to expect a reader to put this puzzle together.

We acknowledge that this is a large and complex figure. In response to the reviewer’s input we have split the figure between Figures 7 and 9, and removed some panels, so as to make it easier to navigate.

Reviewer #3 (Recommendations for the authors):

(1) Which parameters are crucial for determining the preferred frequency (i.e., gamma frequency) for mismatch detection? This point should be addressed further.

This is an interesting suggestion and we have performed additional simulations to address it. It turns out that the frequency tuning is very broad, over almost the entire gamma range from 40 to 200 Hz, and is indeed tuned by simulation parameters. We have placed these findings in Figure 9 in the new version of the paper.

(2) The meanings of horizontal and vertical color bars should be explained in the legend of Figure 2A. Do they show the average values over columns and rows? A similar question applies to Figure 3G.

We have removed the marginal heatmaps from Figures 2 and 3 as they were not contributing to the interpretation.

(3) I wonder whether the proposed mismatch detection is tolerant against timing jitters in repeated presynaptic spike patterns. This information allows us to infer the accuracy required for neural code using population spike patterns.

This is a good suggestion. We have run additional simulations to quantify this. It turns out that jitter has a clear effect on mismatch detection, and affects 5-square (low-overlap) patterns differently from high overlap (15 square) patterns. The latter see a boost in selectivity with 6 ms jitter. This comparison is now in Figure 7E ii and 7 Eiii