Inhibitory control of frontal metastability sets the temporal signature of cognition

  1. Vincent Fontanier
  2. Matthieu Sarazin
  3. Frederic M Stoll
  4. Bruno Delord
  5. Emmanuel Procyk

  • Curated by eLife

    eLife logo

    Summary: Both reviewers found this study's findings to be interesting and novel, and they appreciated the integration of intrinsic timescale analysis, coding of behavioral signals, and exploration of mechanisms in a biophysical circuit model. However, they both raised serious concerns about methodological and interpretational aspects of the analyses which would need to be satisfactorily addressed in subsequent review. In consultation, both reviewers were in agreement with all points raised in the other's review. I view the two reviewers' requests and suggestions as appropriate and complementary, and all of them needed response.

    I highlight here two of the major concerns raised by the reviewers (but all raised by the reviewers merit addressing). First is the need for intrinsic timescale analysis to not be confounded by slowly varying changes in firing rate coding task variables (Point 1.2 of Reviewer 2). This is important for interpretation on intrinsic timescale and its correlation with task variable coding as a major result. Second is the interpretation and implementation of Hidden Markov Models to non-simultaneously recorded neurons (Points 6 of Reviewer 1 and 3 of Reviewer 2). Here again the HMM states may be driven by task variable coding which is not corrected for, which could confound interpretation of results as in terms of meta-stable states and its link to the circuit model. Furthermore, HMM analysis of the circuit model does not match its methodology for experimental data, but could be through non-simultaneous model spike trains, which the manuscript does not justify. I will add my own suggestion that perhaps both of these methodological data analysis concerns could potentially be addressed through comparison to the null model of a non-homogeneous Poisson process with firing rate given by the variable-coding PSTH. I do not consider it necessary to study a network model that performs the task, as suggested for consideration by Reviewer 1.

Reviewed by

Read the full article

Discuss this preprint

Start a discussion What are Sciety discussions?

Listed in

Log in to save this article

Abstract

Cortical dynamics are organized over multiple anatomical and temporal scales. The mechanistic origin of the temporal organization and its contribution to cognition remain unknown. Here, we demonstrate the cause of this organization by studying a specific temporal signature (time constant and latency) of neural activity. In monkey frontal areas, recorded during flexible decisions, temporal signatures display specific area-dependent ranges, as well as anatomical and cell-type distributions. Moreover, temporal signatures are functionally adapted to behaviourally relevant timescales. Fine-grained biophysical network models, constrained to account for experimentally observed temporal signatures, reveal that after-hyperpolarization potassium and inhibitory GABA-B conductances critically determine areas’ specificity. They mechanistically account for temporal signatures by organizing activity into metastable states, with inhibition controlling state stability and transitions. As predicted by models, state durations non-linearly scale with temporal signatures in monkey, matching behavioural timescales. Thus, local inhibitory-controlled metastability constitutes the dynamical core specifying the temporal organization of cognitive functions in frontal areas.

Article activity feed

  1. Version published to 10.7554/elife.63795 on eLife
  2. eLife

    Reviewer #2:

    The paper investigates the temporal signatures of single-neuron activity (the autocorrelation timescale and latency) in two frontal areas, MCC and LPFC. These signatures differ between the two areas and cell classes, and form an anatomical gradient in MCC. Moreover, the intrinsic timescales of single neurons correspond with their coding of behaviorally relevant information on different timescales. The authors develop a detailed biophysical network model which suggests that after-hyperpolarization potassium and inhibitory GABA-B conductances may underpin the potential biophysical mechanism that explains diverse temporal signatures observed in the data. The results appear exciting, as the proposed relationship between the intrinsic timescales, coding of behavioral timescales, and anatomical properties (e.g., the amount of local inhibition) in the two frontal areas is novel. The use of the biophysically detailed model is creative and interesting. However, there are serious methodological concerns undermining the key conclusions of this study, which need to be addressed before the results can be credited.

    Major Concerns:

    1. One of the key findings is the correspondence between the intrinsic timescales of single neurons and their coding of information on different behavioral timescales (Fig. 4). However, the method for estimating the intrinsic timescales has serious problems which can undermine the finding.

      1.1 The authors developed a new method for estimating autocorrelograms from spike data but the details of this method are not specified. It is stated that the method computes the distribution of inter-spike-intervals (ISIs) up to order 100, which was "normalized", but how it was normalized is not described. The correct normalization is crucial, as it converts the counts of spike coincidences (ISI distribution) into autocorrelogram (where the coincidence counts expected by chance are subtracted) and can produce artifacts if not performed correctly.

      1.2 The new method, described as superior to the previous method by Murray et al, 2014, appears to have access to more spikes than the Murray's method (Fig. 2). Where is this additional data coming from? While Murray's method was applied to the pre-cue period, the time epoch used for the analysis with the new method is not stated clearly. It seems that the new method was applied to the data through the entire trial duration and across all trials, hence more spikes were available. If so, then changes in firing rates related to behavioral events contribute to the autocorrelation, if not appropriately removed. For example, the Murray's method subtracts trail-averaged activity (PSTH) from spike-counts, similar to shuffle-correction methods. If a similar correction was not part of the new method, then changes in firing rates due to coding of task variables will appear in the autocorrelogram and estimated timescales. This is a serious confound for interpretation of the results in Fig. 4. For example, if the firing rate of a neuron varies slowly coding for the gauge size across trials, this will appear as a slow timescale if the autocorrelogram was not corrected to remove these rate changes. In this case, the timescale and GLM are just different metrics for the same rate changes, and the correspondence between them is expected. Before results in Fig. 4 can be interpreted, details of the method need to be provided to make sure that the method measures intrinsic timescales, and not timescales of rate changes triggered by the task events. This is an important concern also because recent work showed that there is no correlation between task dependent and intrinsic timescales of single neurons, including in cingulate cortex and PFC (Spitmaan et al., PNAS, 2020).

    2. The balanced network model with a variety of biophysical currents is interesting and it is impressive that the model reproduces the autocorrelation signatures in the data. However, we need to better understand the network mechanism by which the model operates.

      2.1 The classical balanced network (without biophysical currents such as after-hyperpolarization potassium) generates asynchronous activity without temporal correlations (Renart et al., Science, 2010). The balanced networks with slow adaptation currents can generate persistent Up and Down states that produce correlations on slow timescales (Jercog et al., eLife, 2017). Since slow after-polarization potassium current was identified as a key ingredient, is the mechanism in the model similar to the one generating Up and Down states, or is it different? Although the biophysical ingredients necessary to match the data were identified, the network mechanism has not been studied. Describing this network mechanism and presenting the model in the context of existing literature is necessary, otherwise the results are difficult to interpret for the reader.

      2.2 Does the model operate in a physiologically relevant regime where the firing rates, Fano factor etc. are similar to the data? It is hard to judge from Fig. 5b and needs to be quantified.

      2.3 The latency of autocorrelation is an interesting feature in the data. Since the model replicates this feature (which is not intuitive), it is important to know what mechanism in the model generates autocorrelation latency.

    3. HMM analysis is used to demonstrate metastability in the model and data, but there are some technical concerns that can undermine these conclusions.

      3.1 HMM with 4 states was fitted to the data and model. The ability to fit a four-state HMM to the data does not prove the existence of metastable states. HMM assumes a constant firing rate in each "state", and any deviation from this assumption is modeled as state transitions. For example, if some neurons gradually increase/decrease their firing rates over time, then HMM would generate a sequence of states with progressively higher/lower firing rates to capture this ramping activity. In addition, metastability implies exponential distributions of state durations, which was not verified. No model selection was performed to determine the necessary number of states. Therefore, the claims of metastable dynamics are not supported by the presented analysis.

      3.2 HMM was fit to a continuous segment of data lasting 600s, and the data was pooled across different recording sessions. However, different sessions have potentially different trial sequences due to the flexibility of the task. How were different trial types matched across the sessions? If trial-types were not matched/aligned in time, then the states inferred by the HMM may trivially reflect a concatenation of different trial types in different sessions. For example, the same time point can correspond to the gauge onset in one session and to the work trial in another session, and vice versa at a different time. If some neurons respond to the gauge and others to the work, then the HMM would need different states to capture firing patterns arising solely from concatenating the neural responses in this way. This confound needs to be addressed before the results can be interpreted.

  3. eLife

    Reviewer #1:

    This is an interesting manuscript which covers an important topic in the field of computational neuroscience - the 'temporal signatures' of individual neurons. The authors set out to address several important questions using a single-neuron electrophysiology dataset, recorded from monkeys, which has previously been published. The behavioural paradigm is well designed, and particularly well suited to investigating the functional importance of different temporal signatures - as it simultaneously requires the subjects to monitor feedback across a short timescale, as well as integrate multiple outcomes across a longer timescale. The neural data are of high quality, and include recordings from lateral prefrontal cortex (LPFC) and mid-cingulate cortex (MCC). First, the authors modify an existing method to quantify the temporal signatures of individual neurons. This modification appears helpful, and an improvement on similar previously published methods, as the authors are able to capture the temporal signatures of the vast majority of neurons they recorded from. The temporal signatures differ across brain regions, and according to the neurons' spike width. The authors argue that the temporal signatures of a subset of neurons are modified by the subjects' degree of task engagement, and that neurons with different temporal signatures play dissociable roles in task-related encoding. However, I have several concerns about these conclusions which I will outline below. The authors then present a biophysical network model, and show that by varying certain parameters in their model (AHP and GABAB conductances) the temporal signatures of the monkey data can be reproduced. Although I cannot comment on the technical specifics of their models, this seems to be an important advance. Finally, they perform a Hidden-Markov Model analysis to investigate the metastability of activity in MCC, LPFC, and their network model. However, there are a few important differences between the model and experimental data (e.g. neurons recorded asynchronously, and the network model not performing a task) that limit the interpretation of these analyses. Overall, I found the manuscript interesting - and the insights from the biophysical modelling are exciting. However, in its current form, the conclusions drawn from the experimental data are not supported by sufficient evidence.

    Major Comments:

    1. The authors use a hierarchical clustering algorithm to divide neurons into separate groups according to their spike width and amplitude (Fig 1C). There are three groups: FS, RS1, and RS2. The authors ultimately pool RS1 and RS2 groups to form a single 'RS' category. They then go on to suggest that RS neurons may correspond to pyramidal neurons, and FS neurons to interneurons. I have a few concerns about this. Firstly, the suggestion that spike width determined from extracellular recordings in macaques can be used as an indicator of cell type is controversial. A few studies have presented evidence against this idea (e.g. Vigneswaran et al. 2011 JNeuro; Casale et al. 2015 JNeuro). The authors should at least acknowledge the limitation of the inference they are making in the discussion section. Secondly, visualising the data alone in Fig 1C, it is far from clear that there are three (or two) relatively distinct clusters of neurons to warrant treating them differently in subsequent analyses. In the methods section, the authors mention some analyses they performed to justify the cluster boundaries. However, this data is not presented. A recent study approached this problem by fitting one gaussian to the spike waveform distribution, then performing a model comparison to a 2-gaussian model (Torres-Gomez et al. 2020 Cer Cortex). Including an analysis such as this would provide a stronger justification for their decision to divide cells based on spike waveform.

    2. The authors conclude that the results in Fig 3 show that MCC temporal signatures are modulated by current behavioural state. However, this conclusion seems a bit of a stretch from the data currently presented. I can understand why the authors used the 'pause' periods as a proxy for a different behavioural state, but the experiment clearly was not designed for this purpose. As the authors acknowledge, there is only a very limited amount (e.g. a few minutes) of 'pause' data available for the fitting process compared with 'engage' data. Do the authors observe the same results if they constrain the amount of included 'engage' data to match the length of the 'pause' data? Also, presumably the subjects are more likely to 'pause' later on in the behavioural session once they are tired/sated. Could this difference between 'pause' and 'engage' data be responsible for the difference in taus? For instance, there may have been more across-session drift in the electrode position by the time the 'pause' data is acquired, and this could possibly account for the difference with the 'engage' data. Is the firing rate different between 'pause' and 'engage' periods - if so, this should be controlled for as a covariate in the analyses. Finally it is not really clear to me, or more importantly addressed by the authors, as to why they would expect/explain this effect only being present in MCC RS neurons (but not FS or LPFC neurons).

    3. At many points in the manuscript, the authors seem to be suggesting that the results of Fig 4 demonstrate that neurons with longer (shorter) timescales are more involved in encoding the task information which is used across longer (shorter) behavioural timescales (e.g. "long TAU were mostly involved in encoding gauge information", and "population of MCC RS units with short TAU was mostly involved in encoding feedback information"). However, I disagree that this conclusion can be reached based on the way the analysis has currently been performed. A high coefficient simply indicates that the population is biased to be more responsive depending on a particular trial type / condition - i.e. the valence of encoding. This does not necessarily tell us how much information the population of neurons is encoding, as the authors suggest. For instance, every neuron in the population could be extremely selective to a particular parameter (i.e. positive feedback), but if half the neurons encode this attribute by increasing their firing but the other half of neurons encode it by decreasing their firing, the effects will be lost in the authors' regression model (i.e. the beta coefficient would equal 0). I would suggest that the authors consider using an alternative analysis method (e.g. a percentage of explained variance or coefficient of partial determination statistic for each neuron) to quantify coding strength - then compare this metric between the high and low tau neurons.

    4. Similarly, in Fig 4 the authors suggest that the information is coded differently in the short and long tau neurons. However, they do not perform any statistical test to directly compare these two populations. One option would be to perform a permutation test, where the neurons are randomly allocated into the 'High TAU' or 'Low TAU' group. A similar comment applies to the different groups of neurons qualitatively compared in panel Fig 4C.

    5. The authors make an interesting and well-supported case for why changing the AHP and GABA-B parameters in their model may be one mechanism which is sufficient to explain the differences in temporal signatures they observed between MCC and LPFC experimentally. However, I think in places the conclusions they draw from this are overstated (e.g. "This suggests that GABA-B inhibitory - rather than excitatory - transmission is the causal determinant of longer spiking timescales, at least in the LPFC and MCC."). There are many other biophysical differences between different cortical regions - which are not explored in the authors' modelling - which could also account for the differences in their temporal signatures. These could include differences in extra-regional input, the position of the region in an anatomical hierarchy, proportion of excitatory to inhibitory neurons, neurotransmitter receptor/receptor subunit expression, connectivity architecture etc. I think the authors should tone down the conclusions a little, and address some more of these possibilities in more detail in their discussion.

    6. For the Hidden Markov Model, I think there are a couple of really important limitations that the authors only touch upon very briefly. Firstly, the authors are performing a population-level analysis on neurons which were not simultaneously recorded during the experiment (only mentioned in the methods). This really affects the interpretation of their results, as presumably the number of states and their duration is greatly influenced by the overall pattern of population activity which the authors are not able to capture. At this stage of the study, I am not sure how the authors can address this point. Secondly, the experimental data is compared to the network model which is not performing any specific task (i.e. without temporal structure). The authors suggest this may be the reason why their predictions for the state durations (Fig 7B) are roughly an order of magnitude out. Presumably, the authors could consider designing a network model which could perform the same task (or a simplified version with a similar temporal structure) as the subjects perform. This would be very helpful in helping to relate the experimental data to the model, and may also provide a better understanding of the functional importance of the metastability they have identified in behaviour.

    7. It is not clear to me how many neurons the authors included in their dataset, as there appear to be inconsistencies throughout the manuscript (Line 73, Fig 1A-B: MCC = 140, LPFC = 159; line 97-98: MCC = 294, LPFC=276; Fig2: MCC = 266, LPFC = 258; Methods section line 734-735 and Fig 2S2: MCC = 298, LPFC = 272). While this is likely a combination of typos and excluding some neurons from certain analyses, this will need to be resolved. It will be important for the authors to check their analyses, and also add a bit more clarity in the text as to which neurons are being included/excluded in each analysis, and justify this.

  4. eLife

    Summary: Both reviewers found this study's findings to be interesting and novel, and they appreciated the integration of intrinsic timescale analysis, coding of behavioral signals, and exploration of mechanisms in a biophysical circuit model. However, they both raised serious concerns about methodological and interpretational aspects of the analyses which would need to be satisfactorily addressed in subsequent review. In consultation, both reviewers were in agreement with all points raised in the other's review. I view the two reviewers' requests and suggestions as appropriate and complementary, and all of them needed response.

    I highlight here two of the major concerns raised by the reviewers (but all raised by the reviewers merit addressing). First is the need for intrinsic timescale analysis to not be confounded by slowly varying changes in firing rate coding task variables (Point 1.2 of Reviewer 2). This is important for interpretation on intrinsic timescale and its correlation with task variable coding as a major result. Second is the interpretation and implementation of Hidden Markov Models to non-simultaneously recorded neurons (Points 6 of Reviewer 1 and 3 of Reviewer 2). Here again the HMM states may be driven by task variable coding which is not corrected for, which could confound interpretation of results as in terms of meta-stable states and its link to the circuit model. Furthermore, HMM analysis of the circuit model does not match its methodology for experimental data, but could be through non-simultaneous model spike trains, which the manuscript does not justify. I will add my own suggestion that perhaps both of these methodological data analysis concerns could potentially be addressed through comparison to the null model of a non-homogeneous Poisson process with firing rate given by the variable-coding PSTH. I do not consider it necessary to study a network model that performs the task, as suggested for consideration by Reviewer 1.

  5. Version published to 10.1101/2020.08.20.259192 on bioRxiv