Cell type-specific connectome predicts distributed working memory activity in the mouse brain

  1. Xingyu Ding
  2. Sean Froudist-Walsh
  3. Jorge Jaramillo
  4. Junjie Jiang
  5. Xiao-Jing Wang

  • Curated by eLife

    eLife logo

    eLife assessment

    This paper presents valuable findings from whole-brain modeling of persistent activity states (underlying working memory) in the mouse brain. The most novel finding is that a spatial gradient of the density of inhibitory neurons supports a corresponding spatial gradient of propensity to support persistent activity. However, the evidence for this finding appears to be incomplete.

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

Recent advances in connectomics and neurophysiology make it possible to probe whole-brain mechanisms of cognition and behavior. We developed a large-scale model of the multiregional mouse brain for a cardinal cognitive function called working memory, the brain’s ability to internally hold and process information without sensory input. The model is built on mesoscopic connectome data for interareal cortical connections and endowed with a macroscopic gradient of measured parvalbumin-expressing interneuron density. We found that working memory coding is distributed yet exhibits modularity; the spatial pattern of mnemonic representation is determined by long-range cell type-specific targeting and density of cell classes. Cell type-specific graph measures predict the activity patterns and a core subnetwork for memory maintenance. The model shows numerous attractor states, which are self-sustained internal states (each engaging a distinct subset of areas). This work provides a framework to interpret large-scale recordings of brain activity during cognition, while highlighting the need for cell type-specific connectomics.

Article activity feed

  1. Version published to 10.7554/elife.85442 on eLife
  2. eLife

    Author Response

    Reviewer #1 (Public Review):

    This interesting manuscript sets out to develop for the mouse a series of important concepts and models that this group has previously developed for models of monkey brains, where they showed that in a large-scale model, anterior → posterior spatial gradients such as spine density (and thus inferred strength of local coupling) lead to a transition from transient stimulus responses to persistent responses, capable of supporting working memory (WM). No such spine density gradient is found in the mouse. Here, the authors propose and use modeling to explore the idea, that the corresponding gradient may be that of density of inhibitory PV cells in different regions of the brain.

    The goal of the study - a large-scale, anatomically-constrained model of WM - is an extremely valuable one, and the authors' efforts in this direction should be supported. That said, some of the main claims in the manuscript were not, at least as currently written, clearly supported by the data, a number of important clarifications need to be made, and some claims of novelty are made in a way that, for a typical reader, may obscure the actual contribution being made.

    The biggest issue is that one of the main claims, that together with cell-type specific long-range targeting, "density of cell classes define working memory representations" (abstract), is not terribly clear. For example, Figs. 2D and 2E show that a brain region's hierarchical location tightly predicts its persistent firing rate (2D), but that PV cell fraction has a far weaker correlation (2E). Is hierarchical location sufficient? If PV cell fraction were constant across model brain regions, would we still get persistent activity modes? It seems likely that the answer may be "yes", but the answer, easily within reach of the authors, is surprisingly not in the current version of the manuscript. Figure 3D, for the thalamocortical model, shows no significant correlation of firing rate with PV density.

    Given the claim about PV density (in the abstract and the first main point of the discussion), this is a big concern. Yet it seems easily addressable: e.g. if indeed the authors found that hierarchy was sufficient and PV density immaterial, the model would be no less interesting. And if the authors demonstrated clearly that a PV density gradient is required, that would make the claim a solid one. If, within the model, such a causal demonstration is present, this reader at least missed it.

    MAJOR CONCERNS:

    (1) The model appears to be a model of a single side of the brain. Perhaps each brain region in the model could be considered an amalgam of that region across both sides of the brain. Yet given results like Li et al. Nature 2016, who show that persistent activity is robust to inhibition of one side, but not both sides of ALM, at the very least discussion of the issue is warranted.

    The model is indeed a one-hemisphere model, and an expansion to a bihemispheric model is considered for future work. We have added the following sentence in the Discussion section:

    “Future versions of the large-scale model may consider different interneuron types to understand their contributions to activity patterns in the cortex (Kim et al,2017; Meng et al., 2023; Tremblay and Rudy, 2016; Nigro et al., 2022), the role of interhemispheric projections in providing robustness for short-term memory encoding (Ni et al., 2016), and the inclusions of populations with tuning to various stimulus features and/or task parameters that would allow for switching across tasks (Yang et al, 2018).”

    (2) The authors make an interesting attempt to distinguish core WM regions from other regions such as "readout" regions, defined as showing persistent activity yet not having an effect on persistent activity elsewhere in the network.

    However, this definition seemed problematic: for example, consider a network that consists of 20 brain regions, all interconnected to each other, and all equivalent to each other, capable of displaying persistent activity thanks to mutual connectivity. Imagine that inhibition of any one of these regions is not sufficient to significantly perturb persistent activity in the other 19. Then they would all be labeled as "readout". Yet, by construction in this thought experiment, they are all equivalent to each other and are all core areas. Such redundancy may well be present in the brain. How would the authors address this redundancy issue?

    We acknowledge the importance of this thought experiment. Although we initially restricted the definition of core area to how a single area contributes to working memory, we proceeded with concurrent inhibition of multiple readout areas (see Essential Revisions response 6 above).

    (3) Also important to discuss would be the fact that every brain region in this model is set up as composed of two populations, and when long-range interactions are strong and the attractors strongly coupled, the entire brain is set up as a 1-bit working memory. How would results and the approach be impacted by considering WM for more flexible situations?

    We have used a model of two populations as the simplest way to integrate large-scale connectivity and inhibitory gradients. Indeed, future work should consider more realistic connectivity and populations with various degrees of tuning to different task parameters. (see Reviewer 1 response 1 above)

    (4) Another concern that is important yet easily addressed is the authors' use of the term "novel cell-type specific graph theory measures". Describing in the abstract and elsewhere the fact that what they mean is to take into account the sign of connections, not just their magnitude, would transmit to readers the essence of the contribution in a manner very simple to understand. Most readers would fail to grasp the essential point of the current labeling, which sounds potentially very vague and complex.

    We have reworded the abstract - see also Essential reviews response 2 above.

    (5) Finally, the overall significance of the study, and advances over previous work, were not entirely clear. In the discussion, the authors identify three major findings: (1) WM function is shaped by the PV cell density gradient. But as above, further work is required to make it clear that this claim is supported by the model. (2) if local recurrent excitation is insufficient to generate persistent activity, then long-range recurrent excitation is needed to generate it. I had trouble understanding why a model was needed to reach this conclusion - it seems as if it is simply a question of straightforward logic. The discussion states that in this regard, the work here "offers specific predictions to be tested experimentally", but I had trouble identifying what these specific predictions are. (3) Taking into account sign, not only magnitude, of connections, is important. This last point once again seemed a matter of straightforward logic, making its novelty difficult to assess.

    We thank the reviewer, we have addressed these issues in the Essential Revisions 3) above.

    Reviewer #2 (Public Review):

    This paper uses the mouse mesoscale connectome, combined with data on the number and fraction of PV-type interneurons, to build a large-scale model of working memory activity in response to inputs from various sensory modalities. The key claims of the paper are two-fold. First, previous work has shown that there does not appear to be an increase in the number of excitatory inputs (spines) per pyramidal neuron along the cortical hierarchy (and this increase was previously suggested to underlie working memory activity occurring preferentially in higher areas along the cortical hierarchy). Thus, the claim is that a key alternative mechanism in the mouse is the heterogeneity in the fraction of PV interneurons. Second, the authors claim to develop novel cell type-specific graph theory.

    I liked seeing the authors put all of the mouse connectomic information into a model to see how it behaved and expect that this will be useful to the community at large as a starting point for other researchers wishing to use and build upon such large-scale models. However, I have significant concerns about both primary scientific claims. With regard to the PV fraction, this does not look like a particularly robust result. First, it's a fairly weak result to start, much smaller than the simple effect of the location of an area along the cortical hierarchy (compare Figs. 2D, 2E; 3C, 3D). Second, the result seems to be heavily dependent upon having subdivided the somatosensory cortex into many separate points and focusing the main figures of the paper (and the only ones showing rates as a function of PV cell fraction) solely on simulations in which the sensory input is provided to the visual cortex. With regards to the claim of novel cell type-specific graph theory, there doesn't appear to be anything particularly novel. The authors simply make sure to assign negative rather than positive weights to inhibitory connections in their graph-theoretic analyses.

    Major issues:

    1. Weakness of result on effect of PV cell fraction. Comparing Figures 2D and 2E, or 3C and 3D, there is a very clear effect of cortical hierarchy on firing rate during the delay period in Figures 2D and 3C. However, in Figure 2E relating delay period firing rate to PV cell fraction, the result looks far weaker. (And similarly for Figs. 3C, 3D, with the latter result not even significant). Moreover, the PV cell fraction results are dominated by the zero firing rate brain regions (as opposed to being a nice graded set of rates, both for zeros and non-zeros, as with the cortical hierarchy results of Figures 2D), and these zeros are particularly contributed to by subdividing somatosensory (SS) into many subregions, thus contributing many points at the lower right of the graph.

    Further, it should be noted that Figure 2E is for visual inputs. In the supplementary Figure 2 - supplement 1, the authors do apply sensory inputs to auditory and somatosensory cortex...but then only show the result that the delay period firing rate increases along the cortical hierarchy (as in Figure 2D for the visual input), but strikingly omit the plots of firing rate versus PV cell fraction. This omission suggests that the result is even weaker for inputs to other sensory modalities, and thus difficult to justify as a defining principle.

    We have now made an effort to exhaustively compare the contributions of PV versus hierarchy in defining the firing rate activity patterns in the model - see Essential Revisions response 1 above. Moreover, we included plots of firing rate versus PV cell fraction for other sensory modalities, and the results would still support a common architecture for short-term memory maintenance.

    1. Graph theoretic analyses. The main comparison made is between graph-theoretic quantities when the quantities account for or do not account for, PV cells contributing negative connection strengths. This did not seem particularly novel.

    See Essential Revisions response 2 above

    1. It was not clear to me how much the cell-type specific loop strength results were a result of having inhibitory cell types, versus were a result of the assumption ('counter-stream inhibitory bias') that there is a different ratio of excitation to inhibition in top-down versus bottom-up connections. It seems like the main results were more a function of this assumed asymmetry in top-down vs. bottom-up than it was a function of just using cell-type per se. That is, if one ignored inhibitory neurons but put in the top-down vs. bottom-up asymmetry, would one get the same basic results? And, likewise, if one didn't assume asymmetry in the excitatory vs. inhibitory connectivity in top-down versus bottom-up connections, but kept the Pyramidal and PV cell fraction data, would the basic result go away?

    We have addressed the issue of cell-type specific loop strength in Essential Revisions response 2 above.

    1. In the Discussion, there is a third 'main finding' claimed: "when local recurrent excitation is not sufficient to sustain persistent activity...distributed working memory must emerge from long-range interactions between parcellated areas". Isn't this essentially true by definition?

    We have addressed this important issue in Essential Revisions response 3 above.

    1. I don't know if it's even "CIB" that's important or just "any asymmetry (excitatory or inhibitory) between top-down vs. bottom-up directions along the hierarchy". This is worth clarifying and thinking more about, as assigning this to inhibition may be over-attributing a more basic need for asymmetry to a particular mechanism.

    We found that this asymmetry is indeed crucial, which may be provided by CIB or, in some regimes, it is sufficient that a PV gradient is present - see Essential Revisions response 1 above.

    Other questions:

    1. Is it really true that less than 2% of neurons are PV neurons for some areas? Are there higher fractions of other inhibitory interneuron types for these areas, and does this provide a confound for interpreting model results that don't include these other types?

    Maybe related to the above, the authors write in the Results that local excitation in the model is proportional to PV interneuron density. However, in the methods, it looks like there are two terms: a constant inhibition term and a term proportional to density. Maybe this former term was used to account for other cell types. Also, is local excitation in the model likewise proportional to pyramidal interneuron density (and, if not, why not?)?

    The reviewer is correct in pointing out that the ‘constant inhibition term’, which we interpret as a minimal inhibition, accounts for other cell types. We have added the respective explanation in the Methods section. Future versions of the model may include different interneuron types - see Reviewer 1 Response 1 above.

    1. Non-essential areas. The categorization of areas as 'non-essential' as opposed to, e.g. "inputs" is confusing. It seems like the main point is that, since the delay period activity as a whole is bistable, certain areas' contributions may be small enough that, alone, they can't flip the network between its bistable down and up states. However, this does not mean that such areas (such as the purple 'non-essential' area in Figure 5a) are 'non-essential' in the more common sense of the word. Rather, it seems that the purple area is just a 'weaker input' area, and it's confusing to thus label it as 'non-essential' (especially since I'd guess that, whether or not an area flips on/off the bistability may also depend on the assumed strength of the external input signal, i.e. if one made the labeled 'input area' a bit too weak to alone trigger the bistability, then the purple area might become 'essential' to cross the threshold for triggering a bistable-up state).

    This is an important point, and a similar point was also raised by Reviewer 1. For simplicity, we have restricted the definition of the function of an area (e.g., input, vs core vs non essential) to how a single area contributes to working memory. The existence of ‘subnetworks’ for any of these functions is indeed plausible - and potentially important, but we have left this for future modeling work. (see Essential Revisions response 6 above). The point that distinguishes ‘input’ and ‘non-essential’ areas is simply whether inhibiting said area during the stimulus period affects stimulus-specific persistent activity. Surely some of the areas that we have classified as ‘non-essential’ have important roles, even for the contents of working memory, however they are not essential to produce the activity pattern we observe here.

    1. Relation between 'core areas' and loop strength. The measure underlying 'prediction accuracy = 0.93' in Figure 6D and the associated results seems incomplete by being unidirectional. It captures the direction: 'given high cell-type specific loop strength, then core area' but it does not capture the other direction: 'given a cell is part of a core area, is its predicted cell-type specific loop strength strong?'. It would be good to report statistics for both directions of association between loop strength and core area.

    Indeed the prediction accuracy refers to the direction loop strength->core area, for which we estimate how well a continuous variable (loop strength) predicts a binary variable (whether core area or not). A prediction in the reverse direction is not well defined, namely to predict a continuous variable from a binary variable, so the reverse association may be only indirectly inferred from Figure 8D.

    1. More justification would be useful on the assumption that the reticular nucleus provides tonic inhibition across the entire thalamus.

    Relatively little is known about how specific this inhibition may be. We have included references in the Discussion section that speak to this fact. (Crabtree 2018, Hardinger et al., 2023).

    1. Is NMDA/AMPA ratio constant across areas and is this another difference between mice and monkeys? I am aware of early work in the mouse (Myme et al., J. Neurophys., 2003) suggesting no changes at least in comparing two brain regions' layer 2/3, but has more work been performed related to this?

    Recent anatomical in-vitro autoradiography work in the macaque shows that NMDA/AMPA ratio (in terms of receptor density) varies across the cortical hierarchy (Klatzmann et al., 2022). Functionally NMDA receptors seem important in PFC L2/3 for persistent activity, while in V1, they contribute relatively little to the stimulus response, which is dominated by AMPA-mediated excitation. This was shown by a recent physiological study in the macaque (Yang et al., 2018). This could indeed point to a species difference, although like-for-like comparisons of equivalent experiments across species are lacking in the literature.. We have included this and other related references in our Discussion - see Essential Revision 4 above.

    1. Are bilateral connections between the left and right sides of a given area omitted and could those be important?

    These potentially important connections were omitted for simplicity in the model, please see Reviewer 1 Responses 1, 3 above.

    Reviewer #3 (Public Review):

    Combining dynamical modelling and recent findings of mouse brain anatomy, Ding et al. developed a cell-type-specific connectome-based dynamical model of the mouse brain underlying working memory. The authors find that there is a gradient across the cortex in terms of whether mnemonic information can be sustained persistently or only transiently, and this gradient is negatively correlated to the local density of parvalbumin (PV) positive inhibitory cells but positively correlated with mesoscale-defined cortical hierarchy. In addition, weighing connectivity strength by PV density at target areas provides a more faithful relationship between input strength and delay firing rate. The authors also investigate a model where cortical persistent activity can only be sustained with thalamus input intact, although this result is rather separate from the rest of the study. The authors then use this model to test the causal contributions of different areas to working memory. Although some of the in silico perturbations are consistent with existing experimental data, others are rather surprising and need to be further discussed. Finally, the authors investigate patterns of attractor states as a result of different local and long-range connections and suggest that distinct attractor states could underlie different task demands.

    The importance of PV density as a predictor for working memory activity patterns in the mouse brain is in contrast to recent computational findings in the primate brain where the number of spines (excitatory synapses per pyramidal cell) is the key predictor. This finding reveals important species differences and provides complementary mechanisms that can shape distributed patterns of working memory representation across cortical regions. The method of biologically-based near-whole-brain dynamical modeling of a cognitive function is compelling, and the main conclusions are mostly well supported by evidence. However, some aspects of the method, result, and discussion need to be clarified and extended.

    1. Based on existing anatomical data, the authors reveal a negative correlation between cortical hierarchy (defined by mesoscale connectivity; this concept needs to be explicitly defined in the Results session, not just in the Method section) and local PV density (Fig. 1). In the dynamical model, the authors find that working memory activity is positively (and strongly) correlated with cortical hierarchy and negatively (and less strongly) correlated with PV cell density (Fig. 2), and conclude that working memory activity depends on both. But could the negative correlation between activity and PV density simply result from the inherent relationship between hierarchy and PV density across regions? To strengthen this result, the authors should quantify the predictive power of local PV density on working memory activity beyond the predictive power of cortical hierarchy.

    We have systematically compared the relationship between PV and hierarchy in generating delay-patterns of activity - see Essential Revisions response 1 above.

    1. In Fig. 4, the authors find that cell-type-specific graph measures more accurately predict delay-period firing rates. Specifically, the authors weigh connections with a cell-type-projection coefficient, which is smaller when the PV cell fraction is higher in the target area. Considering that local PV cell fraction is already correlated with delay activity patterns, weighing the input with the same feature will naturally result in a better input-output relationship. This result will be strengthened if there is a more independent measure of cell-type-projection coefficient, such as the spine density of PV vs excitatory cells across regions, or even the percentage of inhibitory versus excitatory cells targeted by upstream region (even just for an example set of brain regions).

    We have compared different measures of cell-type projection coefficients and how they predict delay-patterns of activity and whether an area is a core area - see Essential Revisions response 2 above.

    1. The authors aim to identify a core subnetwork that generates persistent activity across the cortex by characterising delay activity as well as the effects of perturbations during the stimulus and delay period. Consistent with existing data, the model identifies frontal areas and medial orbital areas as core areas. Surprisingly, areas such as the gustatory area are also part of the core areas. These more nuanced predictions from the model should be further discussed. Also surprisingly, the secondary motor cortex (MOs), which has been indicated as a core area for short-term memory and motor planning by many existing studies is classified as a readout area. The authors explain this potential discrepancy as a difference in task demand. The task used in this study is a visual delayed response task, and the task(s) used to support the role of MOs in short-term memory is usually a whisker-based delayed response task or an auditory delay response task. In all these tasks, activity in the delay period is likely a mixture of sensory memory, decision, and motor preparation signals. Therefore, task demand is unlikely the reason for this discrepancy. On the other hand, motor effectors (saccade, lick, reach, orient) could be a potential reason why some areas are recruited as part of the core working-memory network in one task and not in another task. The authors should further discuss both of these points.

    We have addressed this important point in Essential Revisions response 5 above.

    1. As a non-expert in the field, it is rather difficult to grasp the relationship between the results in Fig. 7 and the rest of the paper. Are all the attractor states related to working memory? If so, why are the core regions for different attractor states so different? And are the core regions identified in Fig. 5 based on arbitrary parameters that happen to identify certain areas as core (PL)? The authors should at least further clarify the method used and discuss these results in the context of previous results in this study.

    Attractor states that have a stable baseline are, by definition, related to working memory in that there is a baseline and a memory state associated with the model. Some areas, such as PL are more likely to be associated with different core subnetworks given its position in the hierarchy. In the current version of the manuscript, we provide a motivation for the different attractor states and how they may relate to cognitive function.

  3. eLife

    eLife assessment

    This paper presents valuable findings from whole-brain modeling of persistent activity states (underlying working memory) in the mouse brain. The most novel finding is that a spatial gradient of the density of inhibitory neurons supports a corresponding spatial gradient of propensity to support persistent activity. However, the evidence for this finding appears to be incomplete.

  4. eLife

    Reviewer #1 (Public Review):

    This interesting manuscript sets out to develop for the mouse a series of important concepts and models that this group has previously developed for models of monkey brains, where they showed that in a large-scale model, anterior → posterior spatial gradients such as spine density (and thus inferred strength of local coupling) lead to a transition from transient stimulus responses to persistent responses, capable of supporting working memory (WM). No such spine density gradient is found in the mouse. Here, the authors propose and use modeling to explore the idea, that the corresponding gradient may be that of density of inhibitory PV cells in different regions of the brain.

    The goal of the study - a large-scale, anatomically-constrained model of WM - is an extremely valuable one, and the authors' efforts in this direction should be supported. That said, some of the main claims in the manuscript were not, at least as currently written, clearly supported by the data, a number of important clarifications need to be made, and some claims of novelty are made in a way that, for a typical reader, may obscure the actual contribution being made.

    The biggest issue is that one of the main claims, that together with cell-type specific long-range targeting, "density of cell classes define working memory representations" (abstract), is not terribly clear. For example, Figs. 2D and 2E show that a brain region's hierarchical location tightly predicts its persistent firing rate (2D), but that PV cell fraction has a far weaker correlation (2E). Is hierarchical location sufficient? If PV cell fraction were constant across model brain regions, would we still get persistent activity modes? It seems likely that the answer may be "yes", but the answer, easily within reach of the authors, is surprisingly not in the current version of the manuscript. Figure 3D, for the thalamocortical model, shows no significant correlation of firing rate with PV density.

    Given the claim about PV density (in the abstract and the first main point of the discussion), this is a big concern. Yet it seems easily addressable: e.g. if indeed the authors found that hierarchy was sufficient and PV density immaterial, the model would be no less interesting. And if the authors demonstrated clearly that a PV density gradient is required, that would make the claim a solid one. If, within the model, such a causal demonstration is present, this reader at least missed it.

    MAJOR CONCERNS:

    (1) The model appears to be a model of a single side of the brain. Perhaps each brain region in the model could be considered an amalgam of that region across both sides of the brain. Yet given results like Li et al. Nature 2016, who show that persistent activity is robust to inhibition of one side, but not both sides of ALM, at the very least discussion of the issue is warranted.

    (2) The authors make an interesting attempt to distinguish core WM regions from other regions such as "readout" regions, defined as showing persistent activity yet not having an effect on persistent activity elsewhere in the network.
    However, this definition seemed problematic: for example, consider a network that consists of 20 brain regions, all interconnected to each other, and all equivalent to each other, capable of displaying persistent activity thanks to mutual connectivity. Imagine that inhibition of any one of these regions is not sufficient to significantly perturb persistent activity in the other 19. Then they would all be labeled as "readout". Yet, by construction in this thought experiment, they are all equivalent to each other and are all core areas. Such redundancy may well be present in the brain. How would the authors address this redundancy issue?

    (3) Also important to discuss would be the fact that every brain region in this model is set up as composed of two populations, and when long-range interactions are strong and the attractors strongly coupled, the entire brain is set up as a 1-bit working memory. How would results and the approach be impacted by considering WM for more flexible situations?

    (4) Another concern that is important yet easily addressed is the authors' use of the term "novel cell-type specific graph theory measures". Describing in the abstract and elsewhere the fact that what they mean is to take into account the sign of connections, not just their magnitude, would transmit to readers the essence of the contribution in a manner very simple to understand. Most readers would fail to grasp the essential point of the current labeling, which sounds potentially very vague and complex.

    (5) Finally, the overall significance of the study, and advances over previous work, were not entirely clear. In the discussion, the authors identify three major findings: (1) WM function is shaped by the PV cell density gradient. But as above, further work is required to make it clear that this claim is supported by the model. (2) if local recurrent excitation is insufficient to generate persistent activity, then long-range recurrent excitation is needed to generate it. I had trouble understanding why a model was needed to reach this conclusion - it seems as if it is simply a question of straightforward logic. The discussion states that in this regard, the work here "offers specific predictions to be tested experimentally", but I had trouble identifying what these specific predictions are. (3) Taking into account sign, not only magnitude, of connections, is important. This last point once again seemed a matter of straightforward logic, making its novelty difficult to assess.

  5. eLife

    Reviewer #2 (Public Review):

    This paper uses the mouse mesoscale connectome, combined with data on the number and fraction of PV-type interneurons, to build a large-scale model of working memory activity in response to inputs from various sensory modalities. The key claims of the paper are two-fold. First, previous work has shown that there does not appear to be an increase in the number of excitatory inputs (spines) per pyramidal neuron along the cortical hierarchy (and this increase was previously suggested to underlie working memory activity occurring preferentially in higher areas along the cortical hierarchy). Thus, the claim is that a key alternative mechanism in the mouse is the heterogeneity in the fraction of PV interneurons. Second, the authors claim to develop novel cell type-specific graph theory.

    I liked seeing the authors put all of the mouse connectomic information into a model to see how it behaved and expect that this will be useful to the community at large as a starting point for other researchers wishing to use and build upon such large-scale models. However, I have significant concerns about both primary scientific claims. With regard to the PV fraction, this does not look like a particularly robust result. First, it's a fairly weak result to start, much smaller than the simple effect of the location of an area along the cortical hierarchy (compare Figs. 2D, 2E; 3C, 3D). Second, the result seems to be heavily dependent upon having subdivided the somatosensory cortex into many separate points and focusing the main figures of the paper (and the only ones showing rates as a function of PV cell fraction) solely on simulations in which the sensory input is provided to the visual cortex. With regards to the claim of novel cell type-specific graph theory, there doesn't appear to be anything particularly novel. The authors simply make sure to assign negative rather than positive weights to inhibitory connections in their graph-theoretic analyses.

    Major issues:

    1. Weakness of result on effect of PV cell fraction. Comparing Figures 2D and 2E, or 3C and 3D, there is a very clear effect of cortical hierarchy on firing rate during the delay period in Figures 2D and 3C. However, in Figure 2E relating delay period firing rate to PV cell fraction, the result looks far weaker. (And similarly for Figs. 3C, 3D, with the latter result not even significant). Moreover, the PV cell fraction results are dominated by the zero firing rate brain regions (as opposed to being a nice graded set of rates, both for zeros and non-zeros, as with the cortical hierarchy results of Figures 2D), and these zeros are particularly contributed to by subdividing somatosensory (SS) into many subregions, thus contributing many points at the lower right of the graph.
      Further, it should be noted that Figure 2E is for visual inputs. In the supplementary Figure 2 - supplement 1, the authors do apply sensory inputs to auditory and somatosensory cortex...but then only show the result that the delay period firing rate increases along the cortical hierarchy (as in Figure 2D for the visual input), but strikingly omit the plots of firing rate versus PV cell fraction. This omission suggests that the result is even weaker for inputs to other sensory modalities, and thus difficult to justify as a defining principle.

    2. Graph theoretic analyses. The main comparison made is between graph-theoretic quantities when the quantities account for or do not account for, PV cells contributing negative connection strengths. This did not seem particularly novel.

    3. It was not clear to me how much the cell-type specific loop strength results were a result of having inhibitory cell types, versus were a result of the assumption ('counter-stream inhibitory bias') that there is a different ratio of excitation to inhibition in top-down versus bottom-up connections. It seems like the main results were more a function of this assumed asymmetry in top-down vs. bottom-up than it was a function of just using cell-type per se. That is, if one ignored inhibitory neurons but put in the top-down vs. bottom-up asymmetry, would one get the same basic results? And, likewise, if one didn't assume asymmetry in the excitatory vs. inhibitory connectivity in top-down versus bottom-up connections, but kept the Pyramidal and PV cell fraction data, would the basic result go away?

    4. In the Discussion, there is a third 'main finding' claimed: "when local recurrent excitation is not sufficient to sustain persistent activity...distributed working memory must emerge from long-range interactions between parcellated areas". Isn't this essentially true by definition?

    5. I don't know if it's even "CIB" that's important or just "any asymmetry (excitatory or inhibitory) between top-down vs. bottom-up directions along the hierarchy". This is worth clarifying and thinking more about, as assigning this to inhibition may be over-attributing a more basic need for asymmetry to a particular mechanism.

    Other questions:

    1. Is it really true that less than 2% of neurons are PV neurons for some areas? Are there higher fractions of other inhibitory interneuron types for these areas, and does this provide a confound for interpreting model results that don't include these other types?
      Maybe related to the above, the authors write in the Results that local excitation in the model is proportional to PV interneuron density. However, in the methods, it looks like there are two terms: a constant inhibition term and a term proportional to density. Maybe this former term was used to account for other cell types. Also, is local excitation in the model likewise proportional to pyramidal interneuron density (and, if not, why not?)?

    2. Non-essential areas. The categorization of areas as 'non-essential' as opposed to, e.g. "inputs" is confusing. It seems like the main point is that, since the delay period activity as a whole is bistable, certain areas' contributions may be small enough that, alone, they can't flip the network between its bistable down and up states. However, this does not mean that such areas (such as the purple 'non-essential' area in Figure 5a) are 'non-essential' in the more common sense of the word. Rather, it seems that the purple area is just a 'weaker input' area, and it's confusing to thus label it as 'non-essential' (especially since I'd guess that, whether or not an area flips on/off the bistability may also depend on the assumed strength of the external input signal, i.e. if one made the labeled 'input area' a bit too weak to alone trigger the bistability, then the purple area might become 'essential' to cross the threshold for triggering a bistable-up state).

    3. Relation between 'core areas' and loop strength. The measure underlying 'prediction accuracy = 0.93' in Figure 6D and the associated results seems incomplete by being unidirectional. It captures the direction: 'given high cell-type specific loop strength, then core area' but it does not capture the other direction: 'given a cell is part of a core area, is its predicted cell-type specific loop strength strong?'. It would be good to report statistics for both directions of association between loop strength and core area.

    4. More justification would be useful on the assumption that the reticular nucleus provides tonic inhibition across the entire thalamus.

    5. Is NMDA/AMPA ratio constant across areas and is this another difference between mice and monkeys? I am aware of early work in the mouse (Myme et al., J. Neurophys., 2003) suggesting no changes at least in comparing two brain regions' layer 2/3, but has more work been performed related to this?

    6. Are bilateral connections between the left and right sides of a given area omitted and could those be important?

  6. eLife

    Reviewer #3 (Public Review):

    Combining dynamical modelling and recent findings of mouse brain anatomy, Ding et al. developed a cell-type-specific connectome-based dynamical model of the mouse brain underlying working memory. The authors find that there is a gradient across the cortex in terms of whether mnemonic information can be sustained persistently or only transiently, and this gradient is negatively correlated to the local density of parvalbumin (PV) positive inhibitory cells but positively correlated with mesoscale-defined cortical hierarchy. In addition, weighing connectivity strength by PV density at target areas provides a more faithful relationship between input strength and delay firing rate. The authors also investigate a model where cortical persistent activity can only be sustained with thalamus input intact, although this result is rather separate from the rest of the study. The authors then use this model to test the causal contributions of different areas to working memory. Although some of the in silico perturbations are consistent with existing experimental data, others are rather surprising and need to be further discussed. Finally, the authors investigate patterns of attractor states as a result of different local and long-range connections and suggest that distinct attractor states could underlie different task demands.

    The importance of PV density as a predictor for working memory activity patterns in the mouse brain is in contrast to recent computational findings in the primate brain where the number of spines (excitatory synapses per pyramidal cell) is the key predictor. This finding reveals important species differences and provides complementary mechanisms that can shape distributed patterns of working memory representation across cortical regions. The method of biologically-based near-whole-brain dynamical modeling of a cognitive function is compelling, and the main conclusions are mostly well supported by evidence. However, some aspects of the method, result, and discussion need to be clarified and extended.

    1. Based on existing anatomical data, the authors reveal a negative correlation between cortical hierarchy (defined by mesoscale connectivity; this concept needs to be explicitly defined in the Results session, not just in the Method section) and local PV density (Fig. 1). In the dynamical model, the authors find that working memory activity is positively (and strongly) correlated with cortical hierarchy and negatively (and less strongly) correlated with PV cell density (Fig. 2), and conclude that working memory activity depends on both. But could the negative correlation between activity and PV density simply result from the inherent relationship between hierarchy and PV density across regions? To strengthen this result, the authors should quantify the predictive power of local PV density on working memory activity beyond the predictive power of cortical hierarchy.

    2. In Fig. 4, the authors find that cell-type-specific graph measures more accurately predict delay-period firing rates. Specifically, the authors weigh connections with a cell-type-projection coefficient, which is smaller when the PV cell fraction is higher in the target area. Considering that local PV cell fraction is already correlated with delay activity patterns, weighing the input with the same feature will naturally result in a better input-output relationship. This result will be strengthened if there is a more independent measure of cell-type-projection coefficient, such as the spine density of PV vs excitatory cells across regions, or even the percentage of inhibitory versus excitatory cells targeted by upstream region (even just for an example set of brain regions).

    3. The authors aim to identify a core subnetwork that generates persistent activity across the cortex by characterising delay activity as well as the effects of perturbations during the stimulus and delay period. Consistent with existing data, the model identifies frontal areas and medial orbital areas as core areas. Surprisingly, areas such as the gustatory area are also part of the core areas. These more nuanced predictions from the model should be further discussed. Also surprisingly, the secondary motor cortex (MOs), which has been indicated as a core area for short-term memory and motor planning by many existing studies is classified as a readout area. The authors explain this potential discrepancy as a difference in task demand. The task used in this study is a visual delayed response task, and the task(s) used to support the role of MOs in short-term memory is usually a whisker-based delayed response task or an auditory delay response task. In all these tasks, activity in the delay period is likely a mixture of sensory memory, decision, and motor preparation signals. Therefore, task demand is unlikely the reason for this discrepancy. On the other hand, motor effectors (saccade, lick, reach, orient) could be a potential reason why some areas are recruited as part of the core working-memory network in one task and not in another task. The authors should further discuss both of these points.

    4. As a non-expert in the field, it is rather difficult to grasp the relationship between the results in Fig. 7 and the rest of the paper. Are all the attractor states related to working memory? If so, why are the core regions for different attractor states so different? And are the core regions identified in Fig. 5 based on arbitrary parameters that happen to identify certain areas as core (PL)? The authors should at least further clarify the method used and discuss these results in the context of previous results in this study.

  7. Version published to 10.1101/2022.12.05.519094v4 on bioRxiv
  8. Version published to 10.1101/2022.12.05.519094v3 on bioRxiv
  9. Version published to 10.1101/2022.12.05.519094v2 on bioRxiv
  10. Version published to 10.1101/2022.12.05.519094v1 on bioRxiv