Excitatory and inhibitory neurons in the dorsal periaqueductal gray encode decisions to assess and escape natural threats
This article has been Reviewed by the following groups
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
- Evaluated articles (Review Commons)
Abstract
Prey species are able to engage hardwired neural pathways to rapidly escape from an imminent predator attack. However, when predator threat is less probable they typically show a stereotypical sequence of approach toward the threat aimed at gathering more information, followed by escape to safety when the threat threshold is reached. The brainstem dorsal periaqueductal gray (dPAG) is required for the expression of escape behavior to predator threats and stimulation of dPAG elicits goal-directed flight. However, in vivo neural recordings in dPAG have identified separate populations of neurons that are tuned to either the approach or escape phase of the behavior suggesting that the structure may also be involved in threat assessment. The genetic identity and connectivity of these Assessment + and Escape + neurons have not been defined, although optogenetic activation of glutamatergic, but not GABAergic neurons elicits high-speed flight, suggesting that Escape + neurons might be exclusively excitatory in nature. Moreover, it is not clear whether non-predator threats such as those elicited by conspecific or other animate threats are encoded by independent or overlapping neurons in dPAG. Here we report the activity pattern of ensembles of glutamatergic and GABAergic dPAG neurons during approach and escape from predator, social, and prey threats. Unexpectedly, we found that both glutamatergic and GABAergic neurons harbor Assessment + and Escape + neurons, suggesting that both cell-types are engaged in the approach-to-avoidance transition. Consistent with the functional involvement of both cell-types in approach-to-avoidance behavior, optogenetic activation of GABAergic cells elicited a reduction of risk assessment behavior towards the predator. Finally, we found that exposure to predator, social or prey threat recruited largely overlapping neurons in dPAG, demonstrating a convergence of threat processing in this structure. These findings point to a tightly coordinated role for dPAG excitatory and inhibitory neurons in the generalized control of innate threat assessment and avoidance behavior.
Article activity feed
-
Note: This response was posted by the corresponding author to Review Commons. The content has not been altered except for formatting.
Learn more at Review Commons
Reply to the reviewers
Reviewer #1* (Evidence, reproducibility and clarity (Required)):*
Animals ability to escape from threat is a crucial survival behaviour exhibited across the animal kingdom. In vertebrates, hard-wired circuits allow animals to escape from predators without the need for learning. One of the main vertebrate brain regions that controls escape from threat is the brainstem dorsal periaqueductal gray. Research in the last decades has shown that dorsal PAG glutamatergic neurons control the initiation of escape (from imminent threat) and escape vigour, whereas GABAergic neurons are spontaneously and tonically active and have been shown to provide an inhibitory …
Note: This response was posted by the corresponding author to Review Commons. The content has not been altered except for formatting.
Learn more at Review Commons
Reply to the reviewers
Reviewer #1* (Evidence, reproducibility and clarity (Required)):*
Animals ability to escape from threat is a crucial survival behaviour exhibited across the animal kingdom. In vertebrates, hard-wired circuits allow animals to escape from predators without the need for learning. One of the main vertebrate brain regions that controls escape from threat is the brainstem dorsal periaqueductal gray. Research in the last decades has shown that dorsal PAG glutamatergic neurons control the initiation of escape (from imminent threat) and escape vigour, whereas GABAergic neurons are spontaneously and tonically active and have been shown to provide an inhibitory threshold for eliciting escape and further signal escape termination (for reviews on this topic see for example: Gross and Canteras, 2012; Silva, Gross and Graeff, 2016; Motta, Carobrez and Canteras, 2017; Franklin, 2019; Silva and McNaughton, 2019; Lefler and Branco, 2020; Stempel, 2024). In addition to their role in the escape action, the dPAG has been suggested to have a broader function in threat processing. Specifically, a subset of neurons in the dorsal PAG has been shown have activity correlated to the approach/distance to a threat zone or predator. These neurons have been called 'risk assessment' neurons (Deng et al. 2016; Masferrer et al. 2020; Reis et al. 2021). Whether risk assessment neurons integrate threats across different modalities and contexts (here: social vs different predatory threats) is currently not known.
The present study addresses two important and long-standing questions in the field: what is the cell-type identity of dorsal PAG neurons encoding threat assessment versus escape?, and are different classes of threat processed by shared or dedicated neuronal populations? The study design is careful and methodologically performed well. While some of the experiments have been published in the past, the comparison of social and predator threat responses across contexts and a description of assessment+ cells across both main excitatory and inhibitory PAG neuron types is interesting.
Briefly, using miniaturized fluorescence microscopy (miniscope calcium imaging) in Vglut2::Cre and Vgat::Cre mice during a live predator (rat) exposure paradigm, the authors characterize neuronal activity in identified glutamatergic and GABAergic dPAG neurons across approach-escape cycles. The central finding is that both excitatory and inhibitory populations contain Assessment+ neurons (active during approach, silent at escape onset) and Escape+ neurons (suppressed during approach, activated at escape onset) and the authors propose that the activity profiles of Assessment+ and Escape+ cells may reflect local circuit wiring rules with putative GABAergic inhibition between excitatory neuronal subsets. Consistent with this, optogenetic activation of GABAergic dPAG neurons suppressed risk assessment behavior and promoted exploratory rearing but did not affect 'baseline' locomotion, building on Tsang et al. 2023 and Stempel et al. 2024. Population-level decoding using CEBRA (Schneider et al., 2023) confirmed that both Vglut2+ and Vgat+ ensembles independently encode behavioral state with high accuracy.
A second finding of this paper concerns threat generalization. Sequential exposure to a predator, an aggressive conspecific, and a prey insect (cockroach; paradigm from Rossier et al., 2021) revealed that more than half of responsive excitatory neurons and nearly half of responsive inhibitory neurons were activated by two or three threat types. This substantial overlap argues for at least partially convergent, rather than parallel, encoding of threat in dPAG, consistent with its role as a general trigger for defensive avoidance (Silva et al., 2013), and contrasts with the anatomically segregated upstream processing of predator and social threats in the medial hypothalamic defensive network. Taken together, this study suggests the dPAG as a site of coordinated excitatory-inhibitory computation in the control of innate threat assessment and avoidance across biologically diverse threat contexts.
Below are comments related to the figures/ data analysis and generally to the discussion part which we recommend should be expanded/changed to put the findings of the authors more in context of published literature and to discuss in more detail the proposed circuit mechanisms that align with the author's findings. Generally, this is a very nice and large dataset that could benefit from a more fine-grained and in-depth analysis of escape+ and risk assessment+ cells and their precise temporal activity profiles. We do not suggest to perform further experiments and think the work in this manuscript is publishable as is with some additional analyses and changes to the text.
Related Figure 1 and calcium imaging methods.
- The classification of neurons as 'Assessment+' and 'Escape+' positive is unclear and should be formally described in the methods, presumably these are the neurons with significantly increased "positive" or absolute slope?* __Author’s response: __We apologize that the classification was not sufficiently clear. We will add a more detailed description of the criteria used to classify both neuron classes in the methods section. Escape+ neurons were those with a significant increase of activity, whereas Assessment+ neurons had a significant decrease in activity (negative slope). Absolute maximum slope values were used for comparison to previously categorized Assessment+ and Escape+ neurons in Figure 2M.
- Can the authors clarify what they mean by escape? In the methods under "manual behavioral annotation", "Escape behavior" is defined as including retractions, retreats and flight. However, under "unsupervised behavioral annotation", escape is defined as "high-velocity locomotion aiming at increasing distance between threat source and subject". Are Escape+ neurons ones that are significantly modulated (presumably positively) at flight (latter definition), or throughout retractions, retreats and flight (former definition)? This also relates to the plots for 'escape+' neurons, where it would be useful to separately plot 'successful flights to shelter' to make the results comparable to previous studies and where trajectories are at least relatively stereotyped.*
__Author’s response: __We apologize for any confusion raised by the duplicated definition. We have time locked neural activity to ‘escape’ behaviors that we defined to include: retraction, retreat or flight. However, we agree with the reviewer that this includes a wide range of escape quality and that a finer analysis may be helpful. To assess any potential differences in neural encoding related to this variation we will include an analysis in the revision that separates high-intensity from low-intensity escapes.
- More example traces of different FOVs aligned to escape-to-shelter onset and to risk assessment onset would be useful, as well as a plot with the % of escape-active neurons and a reliability index (in how many trials is each neuron active?). Similarly, are the proportions of escape+ and assessment+ neurons similar across animals and FOVs recorded?*
__Author’s response: __We thank the reviewer for these excellent suggestions and will add additional example traces and nimiscope FOVs, together with a table with the proportions of each neuron class per mouse and reliability indices for all neurons.
- Some further basic analyses of the neurons' responses would be helpful to gage their activity profiles. Do these correspond to previously published descriptions of the two classes of escape+ neurons? Are they active during baseline locomotion? Do you observe the same clusters of GABAergic neurons that have been previously described where some dip at escape onset and some ramp up towards escape termination? Please add these plots as a supplement.*
__Author’s response: __We appreciate the reviewer's suggestions and request to link our findings better to published work. We will provide plots showing the correlations between neural activity and baseline locomotion, speed, distance to the threat, and escape termination. We expect that aligning neural activity to escape termination will allow us to observe the two subclasses of GABAergic neurons described by Stempel et al. (2024): neurons that gradually increase their activity, peaking at escape termination, and neurons that gradually decrease their activity, reaching a trough at escape onset.
- It is generally assumed that assessment+ cells 'map'/correlate to the distance to a threat zone. Plots with quantification of this correlation would be useful, and whether they are also speed modulated or not? If the animal stops on the way to the threat zone, does the risk assessment signal plateau for example?*
Author’s response: We will provide supplementary plots for correlations of neural activity with baseline locomotion, speed, and distance to threat. We will also examine the data for cases in which the approach was interrupted, as suggested by the reviewer.
- Both in Figures 1 and 2, both the single trial examples of individual neurons and averages across neurons in the Vgat+ and Vglut2+ recordings show very fast changes that seem to be shorter than Gcamp6s kinetics would allow, and that happen exactly at escape onset when there is presumable a fast head turn movement. What motion correction controls do the authors have in place to make sure that some of the fast changes they see are not motion artefacts? (e.g., see Figure 2, panel C bottom.) (Importantly, see also comment 8 below).*
__Author’s response: __We acknowledge the reviewer’s concern about motion artefacts. Several precautions were taken to minimize such artifacts. During experiments and prior to recording we carefully tapped the miniscope attached to the baseplate and looked at the image to make sure no obvious image movement occurred due to gross mechanical instability. Within the miniscope image analysis we applied five rounds of NoRMCorre motion correction implemented within CaImAn and confirmed the stability of the ROIs by manual scanning of all videos. While subtle motion artifacts could persist in our data despite these precautions, we believe that artifactual variations of GCaMP signal at escape onset (head turn followed by escape) are unlikely because similar changes in GCaMP signal are seen at the onset of risk assessment when sudden head movement occured. Nevertheless, to better address this potential confound we will: 1) manually annotate instances of head turning events outside of escapes and examine their neural correlations, and 2) provide frame-by-frame FOV images from examples of Assessment+ and Escape+ cells across head turns.
- Related to this, the overall escape velocity is extremely low (around 10cm/s). When the authors only analyze high speed escapes (>50cm/s), do they see different cell activity profiles emerge that they might miss with these very low speed escapes that presumably activate less neurons that high-speed escapes? While slow-speed escapes still elicit activity in both Vgat+ and Vglut2+ neurons, their calcium activity changes will be much lower, potentially hindering a more detailed analysis, as clear signals might be sparser for escape+ neurons.*
Author’s response: As discussed above, we used a relatively broad definition of escape so as to increase the number of trials and strengthen the power of our statistical analysis. However, to test this possibility more explicitly we will split our trials by high and low-intensity escapes and check for such correlations.
- With the min/max normalization that has been applied across the entire session it is hard to see 'local changes' in the heatmaps. Given that glutamatergic neurons are thought to only sparsely fire outside of escape episodes, the heatmaps are hard to read with the 'min/max' Z-scoring, and we would strongly encourage the authors to plot 'locally' Z-scored traces without a min/max normalization for each cell (at least for some examples). Importantly we would suggest to change the color scheme for the heatmaps to allow visual identification of the baseline / 'F0'.*
__Author’s response: __We agree that this could be a useful alternative way to visualize the data and will plot locally Z-scored traces and shift to standard colormaps that allow for easier visual identification of the baseline as suggested by the reviewer.
- The calculation of the slope estimated on the calcium signal is somewhat unclear in the methods section. The way it is described right now, it is not clear whether the window is centered at the behavioral event of interest or not. In the figures, it appears that this is an absolute slope, which is also not clear from the methods. Further, considering the half-decay time of GCamP6f, which is ~ 0.5s (Chen et al., 2013), it seems like the result of this calculation could be cancelation or a slope near-zero if the signal both rises and decays in this broader window. Why is not the peak (e.g. max) or the AUC used for this analysis? The authors should clarify the calculation and provide rationale for their choice? Is the sign (positive/negative) of the slope not of interest?*
__Author’s response: __We apologize that the description of the slope analysis was not clear. We will revise the Methods to specify the exact time window used – which is centered on the behavioral event – and whether signed or absolute slope values were used. We will also report whether the sign of the slope is used for classification and clarify the rationale for retaining the slope metric which derives from the argument that quantifying the slope around the behavioral event of interest is less sensitive to signal-to-noise ratio variability between cells within the same field of view. We also favored a slope-based, rather than AUC-based assessment because we were looking to identify cells with previously-identified properties (Masferrer et al. 2020). This criterion allowed for increasing the sensitivity of classification even in cases where peak or AUC criteria analyses were not significant.
- The GRIN lens placement in the example in Figure 1 is in the lateral PAG, whereas most others are located in the dorsolateral PAG. It would be useful to have a sentence in the introduction to state that the authors include the lateral, dorsolateral and dorsomedial PAG as 'dorsal PAG' and a rational for this placement.*
__Author’s response: __We agree that anatomical precision is important given potential functional differences across PAG columns. Based on our histological reconstruction and comparison with the anatomical atlas, we interpret the example shown in Figure 1 as being located within the dorsolateral PAG rather than the lateral PAG. To make this clearer we will revise the figure labeling and add PAG column boundary overlays. We will also add a statement in the Methods/Results clarifying which PAG subdivisions were included under the term “dorsal PAG” and provide a rationale for this grouping.
- Could the authors comment on how the proportions of Assessment+ and Escape+ neurons relate to previously published literature (e.g., Deng et al. 2016)?*
__Author’s response: __We will add a paragraph and table comparing our findings with those from published manuscripts (Deng et al. 2016; Masferrer et al. 2020).
Related to Figure 2
- In panel K of Figure 2, both Vgat+ and Vglut2+ assessment+ neurons seem to have a rise at escape onset in addition to the slow rise during their movement towards the threat zone. Also here, the offset kinetics of the signal seem to now correlate well to the slow decay kinetics you would expect for GCamp6f and a quantification of controls and motion correction quality metrics would be very helpful to add. Depending on the baselining the peaks during escape would probably be significant as well. The authors could try and cluster the neurons further to see if they can disentangle further 'sub classes/clusters'.*
Author’s response: In these time-warped analyses only cells with a significant correlation to escape onset were included. No statistical testing was performed to explore significance at the events highlighted by the reviewer. Nevertheless, we will add data from our clustering analyses as well as motion correction quality metrics and example FOV and ROIs.
- The calculation of the slope estimated on the calcium signal is somewhat unclear in the methods section. The way it is described right now, it is not clear to me whether the window is centered at the behavioral event of interest or not. In the figures, it appears that this is an absolute slope, which is also not clear from the methods. Further, considering the half-decay of GCamP6f which is ~ 0.5s (Chen et al., 2013) it seems like the result of this calculation could be cancelation or a slope near-zero if the signal both rises and decays in this broader window. Why is not the peak (e.g. max) or the AUC used for this analysis? The authors should clarify the calculation and provide rationale for their choice? Is the sign (positive/negative) of the slope not of interest?*
__Author’s response: __As discussed above, we chose to use a slope-based analysis because this feature best distinguishes the Assessment+ and Escape+ cell classes. This choice made it possible for us to maximize the identification of such cells even in cases in which AUC or peak analysis were not significant. We felt it was appropriate given that the aim was primarily to maximize the identification of previously well-described cell classes rather than de novo search. We will include a more detailed rationale for this approach in the manuscript.
- The authors find a significant increase in the accuracy of the classification of the two defensive behaviors (risk assessment versus escape) from models trained on the Vgat+ population calcium activity. This is an interesting finding which should be to the very least discussed in the discussion. Is there more information content in this population? Does this have anything to do with the slower dynamics of the signal? As mentioned above, would clustering these populations further reveal a more fine-grained detail of their population activity, e.g. see previously published work by Stempel et al. who suggest that there are at least two broad clusters of escape-active GABAergic neurons.*
__Author’s response: __We agree that the increase in classification accuracy is intriguing. While we do not have a precise explanation for this effect, we note that the higher decoding accuracy may reflect the fact that the Vgat+ population collectively contains richer information about defensive behaviors than the Vglut2+ population. However, because our recordings measure calcium activity rather than spiking, we cannot exclude the possibility that differences in signal-to-noise ratio, or other population-level recording characteristics like the number of recorded neurons contribute to this effect. Therefore, we have been careful not to overinterpret the improved decoding performance as necessarily reflecting greater information content. We also agree that further subdivision of the Vgat+ population could reveal additional functional organization. We are happy to incorporate further analysis for clustering (see above) into the manuscript as suggested.
Related to Figure 3
- In Figure 3A, the ChR2 seems to be significantly spread throughout the entire PAG and also the superior colliculus. As it is hard to see cell bodies with ChR2 at this magnification, we would recommend to add a supplemental figure with the outlines of the infection sites.*
__Author’s response: __We acknowledge that the viral spread of these infections is larger than the dPAG, but we are confident that we are recovering spatial specificity with the placement of the optic fibers in the dPAG. We are happy to provide outlines of the infection sites for all animals.
- The authors state that: "Interestingly, although stimulation of Vgat⁺ dPAG neurons significantly reduced the peak speed of escape, it did not change the overall time spent escaping". Could they comment on how this may relate to a previous study where the probability of escape is decreased upon activation of Vgat+ neurons and initiated escapes can be induced (Stempel et al. 2024)? Do they think this could be differences in location of stimulation fiber or one vs two injection sites (bigger vs smaller spread along the AP axis of the PAG?) - or do they think there could be a fundamental difference between fast escape from imminent looming stimuli to slower escapes to rat/prey/social predators?*
__Author’s response: __The publication mentioned by the reviewer found that Vgat+ activation decreased the probability of escape to a looming stimulus. We believe the discrepancy with our results could arise from differences in the temporal pattern of the stimulation. While Stempel and colleagues stimulated at escape onset, we stimulated for longer and regularly spaced, one minute long light pulses.
Related to Figure 4.
- Individual examples of cells responding to one threat vs multiple threats would be very useful to add to see their dynamics across threats.*
__Author’s response: __Thanks for the suggestion. We will provide full trace examples for the three tests for neurons responding to three, two or one threat.
- The Venn diagrams in panel I are very hard to read, and color-blind people might struggle with the red/green (also in panel C). We would suggest to change colors, and replace Venn diagrams to make results more interpretable/readable.*
__Author’s response: __We thank the reviewer for this useful feedback. We will remove the Venn diagrams and stick to the pie charts and make sure that all figure panels are color blind friendly.
- Have any statistical tests been performed to assess whether there is a difference between the proportions in figure D and H as well as Figure 4J?*
__Author’s response: __No. We will perform and include the results of these statistical tests.
General and other comments
- We strongly encourage the authors to add example videos of the behaviors tested and of all major findings including example calcium activity and optogenetic manipulations.*
__Author’s response: __We will include videos for approach and escape behaviors for representative mice performing the three tests as well as representative calcium traces. We will also include exemplary behavior videos of experimental and control animals in the optogenetic activation experiments.
- Relating to a more fine-grained analysis of escape-active cells (as recommended above), the authors state in the discussion that: 'A notable paradox of our findings is that while GABAergic and glutamatergic dPAG neurons showed nearly indistinguishable neural firing correlates of approach and avoidance (both harbored Assessment+ and Escape+ cells) [...]". Since previous works have already looked at escape-active cells in more detail, with differences between Vgat+ and Vglut2+ neurons having been described, we would encourage the authors to look at unsupervised clustering of the escape-active neuron populations, as the firing rates of these neurons should be distinguishable along the escape sequence and especially when taking into account speed correlations and activity profiles aligned to escape onset or offset. If they can't find differences these should be discussed, as major differences may relate to the paradigms used (rats vs visual 'looming' threats).*
Author’s response: As described above, we will correlate neural activity with speed and cluster the response types with unsupervised methods (e.g. K-Means clustering, hierarchical clustering) in order to uncover further differences between Vglut2+ and Vgat+ neurons. We will discuss the finding in the discussion section in more detail.
- We would ask the authors to make sure that the cited literature supports their statements. Some statements are in the manuscript are not clearly, only partially correct or the literature itself is inconsistent or citations are confusing/misleading.*
e.g.: "A functional cellular and circuit architecture of dorsal PAG is emerging in which stimulation of glutamatergic neurons in dorsal PAG promotes freezing at low intensity and flight at high intensity (Tovote et al. 2016, Deng et al. 2016, Evans et al. 2018, Tsang et al. 2023)."
Not all of the cited papers support that graded stimulation of glutamatergic neurons in the dorsal PAG results in escape/flight or freezing but rather that there is some (dis-)inhibitory interplay between dorsal and ventrolateral PAG which results in the selection of one or the other behaviors, and that the ventrolateral PAG more specifically drives freezing (e.g., Tovote et al. 2016). Some studies have shown that activation of the dorsal PAG and in particular of glutamatergic neurons elicits 'all-or-none' flight behavior. Additionally, studies where freezing has been observed with dorsal PAG stimulation have often used CamKIIa as a promoter (e.g. Deng et al. 2016*), which is expressed at significant levels in both excitatory and inhibitory neurons. Thus, we would argue it is not entirely clear whether freezing through 'low intensity' dPAG stimulation is a biological feature of the dPAG network, especially also as inhibition of the dPAG promotes freezing - probably through disinhibition of the vlPAG.
* In the results section on page 6, it is stated that Deng et al. 2016 use VGlut2-Cre to target ChR2 for eliciting escape, but they have used CamKIIa ("Previous studies have shown that optogenetic activation of Vglut2+ dPAG neurons elicited flight behavior[...]").
Similarly, on page 3, the following statement on GABAergic neurons isn't clear: "Stimulation of GABAergic neurons, on the other hand, does not elicit defensive behavior (Tsang et al., 2023) and recent evidence shows that they can inhibit looming stimulus-evoked flight behavior (Stempel et al., 2024) suggesting they might act as part of a tonic inhibitory circuit that receives primarily local inputs (Franklin et al., 2017; but see Wu et al. 2024)."
Multiple papers across all columns have shown that GABAergic neurons are tonically and spontaneously active (e.g., Chen et al., 2023; Wang et al., 2023; Stempel et al., 2024) and Stempel et al. estimated that GABAergic PAG neurons make up >50% of spontaneous inputs to glutamatergic PAG neurons, thus positioning them well to control their activity. Franklin et al. 2017 (and others) have shown that GABAergic neurons receive input from thalamic and hypothalamic regions (probably also from midbrain regions that were not analyzed in that study), thus arguing that they receive significant input from outside of the PAG to integrate - currently unknown - inputs from other brain regions to presumably guide their activity in a context-specific manner. The statement that they should thus be part of a circuit that receives primarily local inputs is not well supported by previous studies. Also, how this enables a 'push-pull' circuit and what is meant here exactly, could be clarified. It is not clear why the presence of risk assessment neurons would support such a circuit model rather than an integrative/threshold one. Further, Stempel et al. 2024 have also shown that GABAergic neurons ramp up activity during escape and are thus technically 'flight+' cells. We agree that there are interesting and presumably complicated intra-PAG dynamics to be studied where their connectivity will define different circuit model possibilities.
__Author’s response: __We thank the reviewer for this careful presentation and analysis of the literature and apologize where we may have inaccurately referenced previous work. We will carefully revise citations and make sure all statements are supported adequately by the literature and more explicitly declare the speculative tone where relevant for our conclusions.
**Referee cross-commenting**
We generally agree with most comments made by reviewer 2, and their concerns. While the novelty of this manuscript is limited, highlighting the advances better, also through more accurate analyses, should allow the authors to better build up their novel findings ('major concern 1'). We agree with the reviewers 'major concerns' 2-5 and that they should be addressed by the authors, for scientific reasons beyond novelty, as some of the analysis feels incomplete. For example, while we do not think that new experiments are necessary, the authors can discuss the issue of spatiotemporal overlap of variables that do not allow a definitive description of 'risk assessment' cells. This issue is in fact present in most published work on risk assessment cells (usually described with a rat predator that is continuously present). Concerns 7-9 are valid, and touch on similar points raised by our review. We do not think that modelling would add anything substantial to this manuscript that couldn't be hypothesised in the discussion beyond their proposed model. In summary, we agree with the points raised by reviewer too, but find that the data is useful and should be published, after some revising and additional more fine-grained analyses to strengthen the manuscript's claims.
Reviewer #1* (Significance (Required)):*
Overall, even though the overall novelty is somewhat limited considering the paper shows significant overlap with recent findings on the involvement of glutamatergic and GABAergic dorsal PAG neurons in risk assessment and escape, there are some important and novel findings that advance the knowledge in the field of neuronal circuits underlying defensive behaviors. For example, the direct comparison of how different threat contexts modulate the activity of individual neurons and the finding that both GABAergic and glutamatergic population activity can predict defensive behavioral output are interesting and important findings. While this study may of limited interest to a very broad audience, it is of importance to the fields of neuroethology and the study of the neural circuits underlying innate defensive behaviors.
Reviewer #2* (Evidence, reproducibility and clarity (Required)):*
This manuscript examines how excitatory and inhibitory neurons in the dorsal periaqueductal gray (dPAG) encode defensive behavior during exposure to natural threats. The question is important, and the experiments are technically ambitious. In particular, the attempt to compare Vglut2+ and Vgat+ populations across predator, social, and prey-related contexts is potentially valuable. However, in its current form, I do not think the manuscript yet establishes a sufficiently strong conceptual advance over the existing literature, and several of the main interpretations appear stronger than the data support.
Major Concerns:*
1)One main concern is novelty. The broader claim that dPAG contains neuronal populations related to threat assessment, approach/avoidance, and escape is not new. Prior work has already shown that PAG neurons differentiate distinct components of defensive behavior during predator exposure, including assessment-like and flight-related activity patterns, and that dPAG ensembles encode approach versus avoidance states across. In parallel, earlier circuit studies had already established a central role for excitatory dPAG neurons in driving escape and freezing, and for local GABAergic dPAG neurons in modulating instinctive escape. The study would benefit for better clarification for each of the main findi.ngs on where the novel contribution to the literature sits.
__Author’s response: __We agree with the reviewers that the novelty provided by our work is incremental and that our findings in some aspects overlap with those of previous studies that described Assessment+ and Escape+ cells in dPAG during approach and avoidance of a rat or that performed GCaMP miniscope recordings in glutamatergic and GABAergic neurons in dPAG under conditions of looming stimulus escape. However, we are pleased to see that the reviewers also acknowledge the novelty of our data and its usefulness for researchers in the field of innate defensive behavior. We will restructure the introduction and discussion sections to highlight the unique contributions of our work and better frame our findings with respect to the existing literature (see also our rebuttal introduction).
2)A second major issue concerns anatomical specificity. From the histology, it is not fully clear that all lens, fibers and injections are confined to the same functional PAG columns. This matters because the manuscript interprets the results at the level of "dorsal PAG," yet functional and input differences across PAG columns are well established in the literature (including dm/dl vs l), with a more prominent involvement of dm/dl in escape in respect to l. I therefore suggest the authors to repeat the key analyses using only neurons, injections and fiber placements clearly restricted dm/dl dPAG, and see if this generalises to l.
__Author’s response: __We agree that anatomical precision is important given potential functional differences across PAG columns. Based on our histological reconstruction and comparison with the anatomical atlas, we interpret the example shown in Figure 1 as being located within the dorsolateral PAG (dlPAG) rather than the lateral PAG (lPAG). To make this clearer we will revise the figure labeling and add PAG column boundary overlays as suggested by the reviewers. We will also repeat the core analyses using only cells from the dlPAG and incorporate those findings into the manuscript as relevant. We will also add a statement in the Methods/Results section clarifying which PAG subdivisions were included under the term “dorsal PAG” and provide a rationale for this grouping.
3)A third concern is the interpretation of "risk assessment" neurons recorded in this work. While this neurons have been previously reported in similar conditions, here the predator is continuously present and the behavioral space is highly constrained. Under those conditions, neuronal activity classified as related to risk assessment could in principle reflect a combination of position in the corridor, heading direction, distance from the safe chamber, distance to the threat compartment, body elongation, or locomotor state, etc rather than threat assessment per se. I do not think the current analyses are sufficient to separate these possibilities. To support the central interpretive claim, the manuscript would benefit from additional control analyses accounting for positional and kinematic variables, and additional experiments including control conditions that dissociates threat presence from spatial configuration.
__Author’s response: __We agree with the reviewer that at present it is not clear which particular aspect of the approach behavior, if any, is best correlated with Assessment+ cell activity. In order to better understand the activity patterns of these cells, we will correlate their activity also to speed, distance to threat, and baseline locomotion.
4)Relatedly, I found the treatment of behavioural annotation too qualitative for the strength of the neural claims. Behaviours such as risk assessment, retraction, retreat, and flight are described verbally, but the manuscript would be much more reproducible if the authors provided explicit operational criteria, quantitative thresholds, and inter-rater reliability. At present, the reader is asked to accept a fairly subjective labelling scheme, yet many of the neural conclusions depend directly on those labels. Critically no major differences have been observed between cell type responses, hinting that perhaps more rigorous behavioural quantification may be required.
__Author’s response: __For consistency across datasets our miniscope data were manually annotated by a single expert scorer and we did not use precise quantitative thresholds for each behavior nor quantify inter-rater reliability. However, in our optogenetic activation dataset we complemented our manually annotated data with automatically extracted measures as well as unsupervised behavior categorization with Keypoint MoSeq, with a satisfactory overlap for behaviors (see Figure 3F). We agree that there are differences in the pattern of correlations of Vglut2+ and Vgat+ neurons across behaviors. Unfortunately, despite major efforts on our part to test these differences – including looking at variations in behavioral vigor and quality across the datasets – we were unable to do so in a statistically reliable fashion. Thus, we do not think that this failure depended on a lack of consideration of the quality of behaviors involved. Instead, we conclude that we lacked the statistical power to see what may be subtle differences between these cell types.
5)The statistical framework used to define neuronal response classes also needs clarification. Responsive cells are identified relative to shuffled null distributions, but it is not clear that the analysis adequately controls for multiple comparison: multiple testing across neurons, behavioural epochs, and response classes.
__Author’s response: __We thank the reviewer for raising this important point and agree that multiple-comparison control was insufficiently described. We will implement a permutation-based maximum-statistic correction across behavioral epochs within each neuron, thereby controlling the family-wise error rate across the behavioral comparisons used to define responsivity. The resulting corrected permutation p-values will subsequently be controlled across neurons using the Benjamini–Hochberg FDR procedure (q = 0.05), separately for each cell type and experimental condition. We will also clarify that Assessment+ and Escape+ categories are assigned after significance testing based on the pattern and direction of behavioral modulation and therefore do not constitute additional independent statistical tests. We will update the Methods and corresponding analyses and figures accordingly and report whether these corrections affect the main conclusions.
6)I also think the interpretation of the CEBRA analyses is too strong. These analyses show that the recorded populations contain information sufficient to decode the annotated behaviors, but they do not demonstrate that the population encodes an abstract "behavioral state" rather than a mixture of posture, speed, position, and other correlated sensorimotor variables. Because the labelled behaviours are themselves associated with distinct kinematic structure, decoding alone is not enough to support the stronger conceptual claim. That interpretation would require nuisance-controlled analyses or matched comparisons showing that decoding persists beyond simple sensorimotor differences.
__Author’s response: __We agree that the CEBRA analysis is difficult to interpret and have been cautious in extracting actionable conclusions from this data analysis tool. Nevertheless, give the widespread interest in such advanced dimensionality reduction tools, we think it is a useful addition to the manuscript and will reframe our interpretation of the CEBRA results to adhere more closely to a strict statement of the observed correlations.
7)The optogenetic results in Vgat+ mice should also be interpreted more cautiously. I was not convinced by the conclusion that stimulation reduces risk-assessment behavior. Increased rearing does not straightforwardly imply reduced assessment, as in some contexts rearing itself can be part of assessment-related behavior. More generally, if stimulation alters locomotor structure, this could secondarily change the frequency of other scored behaviors and syllables without demonstrating that the manipulated neurons specifically control risk assessment. For example, if locomotion speed is reduced by the manipulation, then the mouse may engage in other behaviour that do not require locomotion such as rearing or grooming. I therefore think the claims in this section should be toned down and are not easily interpretable with the data provided.
__Author’s response: __We agree with the reviewer’s statement on the caveats of the optogenetic experiments and wholeheartedly appreciate the difficulty in distinguishing direct and indirect consequences of such manipulations. We will revise the Results and Discussion to be more cautious in our interpretations that prolonged activation of Vgat+ dPAG neurons altered the behavioral structure during predator exposure, reducing risk assessment behavior while increasing rearing/exploration and pointing out the difficulties inherent in interpreting the overall impact of such an artificial manipulation.
8)The functional interpretation of the inhibitory population is not sufficiently novel relative to prior work. Previous studies have already shown that GABAergic neurons in dPAG modulate instinctive escape behavior, so the present result that inhibitory neurons affect escape-related responding is, on its own, not a major conceptual advance. What would make the current study more compelling is clear evidence that these neurons specifically encode or regulate risk assessment. At present, however, that conclusion remains uncertain. Because the predator is continuously present in the assay, the timing of threat delivery is not well defined by construction, making it difficult to separate neural activity linked to threat assessment from activity linked to the suppression, gating, or delayed initiation of escape during approach. In other words, the observed slowing during "risk assessment" could reflect inhibition of escape-related motor output rather than a distinct effect of assessment itself.
__Author’s response: __We agree with the reviewer that our findings on manipulating Vgat+ neurons overlap in part with prior work. We started this work before those publications appeared and hope that the overlapping findings can, nevertheless, be a useful confirmation for researchers in the field. We completely agree with the reviewer on the difficulty inherent in interpreting neural activity correlations with the types of self-paced behaviors we are interested in here and the alternative explanations provided are eminently possible and even probable. Our data do not offer the possibility to distinguish these cases, and we apologize if our discussion appeared to draw definitive conclusions on a role of dPAG neurons in controlling risk assessment behaviors per se. We will take a more cautious approach in our discussion to argue for a role in controlling behavior during approach to threat and that this may be mediated by changes in risk assessment or other pre-escape behaviors.
9) The author states: "Unfortunately, we were not able to identify testing conditions under which the two cell types differed statistically, leaving open the question of whether GABAergic and glutamatergic neurons in dPAG reliably encode different aspects of defensive behavior. Our assessment of ensemble encoding of behavior also failed to shed light on cell-type specific differences in encoding, with both cell-types showing significant predictive correlations of approach and avoidance behaviors." This statement is surprising and substantially limits the conceptual advance, because the main distinction the study sets out to test is ultimately not resolved. In my view, one likely reason is the experimental design itself. Because the predator is continuously present, threat presentation, exploration, risk assessment, sensory sampling, escape decision, and escape onset are all temporally entangled. Under these conditions, it becomes very difficult to isolate which component of the defensive sequence is actually being encoded, and this may reduce the ability to detect meaningful differences between excitatory and inhibitory populations. This is especially important given that prior work has already shown that dPAG populations vglut and vgat population encode and modulate differently defensive behaviours such as escape. In that context, the present study would need a cleaner behavioral design to demonstrate a distinct contribution of cell type, rather than a mixed representation of overlapping sensory, motor, and defensive variables.
__Author’s response: __We thank the reviewer for raising the issue of the surprising lack of differences in encoding between excitatory and inhibitory neurons as the findings were equally unexpected for us. We will conduct the proposed analyses related to trial splitting according to speed and other kinematics, together with unsupervised clustering. In the eventuality of being unable to find them, we will add a paragraph in the discussion about the limitations of our experimental design.
- 10)Finally, the proposed model in which local GABAergic inhibition organizes Assessment+ and Escape+ excitatory populations is interesting, but at present it remains speculative. The data may be consistent with this idea, but they do not directly test it. Since computational modelling of this framework would not be too complex, I would suggest additional analyses and network modelling to substantiate this claim, for example by showing that a network model under these constraints would recapitulate the neural responses observed in vivo.*
Author’s response: We agree that the proposed local inhibitory circuit model is speculative. To make this more clear to the reader we will revise the manuscript to present the model as purely hypothetical. Because our current experiments do not measure synaptic connectivity or selectively manipulate Assessment+ versus Escape+ subpopulations we do not think that a network model alone would provide definitive mechanistic evidence. We will therefore tone down the circuit interpretation and clearly state which aspects are supported by the present data and which require future experiments, such as cell-type- and projection-specific recordings, connectivity mapping, or targeted manipulation of functionally defined neuronal subpopulations.
Overall, I think the manuscript contains useful data, but in its present form its novel contribution is unclear. Additional experiments and analyses may be needed to strengthen its central claims, as described above.
Reviewer #2 (Significance Required):*
This manuscript examines how excitatory and inhibitory neurons in the dorsal periaqueductal gray (dPAG) encode defensive behavior during exposure to natural threats. The question is important, and the experiments are technically ambitious. In particular, the attempt to compare Vglut2+ and Vgat+ populations across predator, social, and prey-related contexts is potentially valuable. However, in its current form, I do not think the manuscript yet establishes a sufficiently strong conceptual advance over the existing literature, and several of the main interpretations appear stronger than the data support.
-
Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.
Learn more at Review Commons
Referee #2
Evidence, reproducibility and clarity
This manuscript examines how excitatory and inhibitory neurons in the dorsal periaqueductal gray (dPAG) encode defensive behavior during exposure to natural threats. The question is important, and the experiments are technically ambitious. In particular, the attempt to compare Vglut2+ and Vgat+ populations across predator, social, and prey-related contexts is potentially valuable. However, in its current form, I do not think the manuscript yet establishes a sufficiently strong conceptual advance over the existing literature, and several of the main interpretations appear stronger than the data support.
Major Concerns:
1)One main …
Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.
Learn more at Review Commons
Referee #2
Evidence, reproducibility and clarity
This manuscript examines how excitatory and inhibitory neurons in the dorsal periaqueductal gray (dPAG) encode defensive behavior during exposure to natural threats. The question is important, and the experiments are technically ambitious. In particular, the attempt to compare Vglut2+ and Vgat+ populations across predator, social, and prey-related contexts is potentially valuable. However, in its current form, I do not think the manuscript yet establishes a sufficiently strong conceptual advance over the existing literature, and several of the main interpretations appear stronger than the data support.
Major Concerns:
1)One main concern is novelty. The broader claim that dPAG contains neuronal populations related to threat assessment, approach/avoidance, and escape is not new. Prior work has already shown that PAG neurons differentiate distinct components of defensive behavior during predator exposure, including assessment-like and flight-related activity patterns, and that dPAG ensembles encode approach versus avoidance states across. In parallel, earlier circuit studies had already established a central role for excitatory dPAG neurons in driving escape and freezing, and for local GABAergic dPAG neurons in modulating instinctive escape. The study would benefit for better clarification for each of the main findings on where the novel contribution to the literature sits.
2)A second major issue concerns anatomical specificity. From the histology, it is not fully clear that all lens, fibers and injections are confined to the same functional PAG columns. This matters because the manuscript interprets the results at the level of "dorsal PAG," yet functional and input differences across PAG columns are well established in the literature (including dm/dl vs l), with a more prominent involvement of dm/dl in escape in respect to l. I therefore suggest the authors to repeat the key analyses using only neurons, injections and fiber placements clearly restricted dm/dl dPAG, and see if this generalises to l.
3)A third concern is the interpretation of "risk assessment" neurons recorded in this work. While this neurons have been previously reported in similar conditions, here the predator is continuously present and the behavioral space is highly constrained. Under those conditions, neuronal activity classified as related to risk assessment could in principle reflect a combination of position in the corridor, heading direction, distance from the safe chamber, distance to the threat compartment, body elongation, or locomotor state, etc rather than threat assessment per se. I do not think the current analyses are sufficient to separate these possibilities. To support the central interpretive claim, the manuscript would benefit from additional control analyses accounting for positional and kinematic variables, and additional experiments including control conditions that dissociates threat presence from spatial configuration.
4)Relatedly, I found the treatment of behavioural annotation too qualitative for the strength of the neural claims. Behaviours such as risk assessment, retraction, retreat, and flight are described verbally, but the manuscript would be much more reproducible if the authors provided explicit operational criteria, quantitative thresholds, and inter-rater reliability. At present, the reader is asked to accept a fairly subjective labelling scheme, yet many of the neural conclusions depend directly on those labels. Critically no major differences have been observed between cell type responses, hinting that perhaps more rigorous behavioural quantification may be required.
5)The statistical framework used to define neuronal response classes also needs clarification. Responsive cells are identified relative to shuffled null distributions, but it is not clear that the analysis adequately controls for multiple comparison: multiple testing across neurons, behavioural epochs, and response classes.
6)I also think the interpretation of the CEBRA analyses is too strong. These analyses show that the recorded populations contain information sufficient to decode the annotated behaviors, but they do not demonstrate that the population encodes an abstract "behavioral state" rather than a mixture of posture, speed, position, and other correlated sensorimotor variables. Because the labelled behaviours are themselves associated with distinct kinematic structure, decoding alone is not enough to support the stronger conceptual claim. That interpretation would require nuisance-controlled analyses or matched comparisons showing that decoding persists beyond simple sensorimotor differences.
7)The optogenetic results in Vgat+ mice should also be interpreted more cautiously. I was not convinced by the conclusion that stimulation reduces risk-assessment behavior. Increased rearing does not straightforwardly imply reduced assessment, as in some contexts rearing itself can be part of assessment-related behavior. More generally, if stimulation alters locomotor structure, this could secondarily change the frequency of other scored behaviors and syllables without demonstrating that the manipulated neurons specifically control risk assessment. For example, if locomotion speed is reduced by the manipulation, then the mouse may engage in other behaviour that do not require locomotion such as rearing or grooming. I therefore think the claims in this section should be toned down and are not easily interpretable with the data provided.
8)The functional interpretation of the inhibitory population is not sufficiently novel relative to prior work. Previous studies have already shown that GABAergic neurons in dPAG modulate instinctive escape behavior, so the present result that inhibitory neurons affect escape-related responding is, on its own, not a major conceptual advance. What would make the current study more compelling is clear evidence that these neurons specifically encode or regulate risk assessment. At present, however, that conclusion remains uncertain. Because the predator is continuously present in the assay, the timing of threat delivery is not well defined by construction, making it difficult to separate neural activity linked to threat assessment from activity linked to the suppression, gating, or delayed initiation of escape during approach. In other words, the observed slowing during "risk assessment" could reflect inhibition of escape-related motor output rather than a distinct effect of assessment itself.
- The author states: "Unfortunately, we were not able to identify testing conditions under which the two cell types differed statistically, leaving open the question of whether GABAergic and glutamatergic neurons in dPAG reliably encode different aspects of defensive behavior. Our assessment of ensemble encoding of behavior also failed to shed light on cell-type specific differences in encoding, with both cell-types showing significant predictive correlations of approach and avoidance behaviors." This statement is surprising and substantially limits the conceptual advance, because the main distinction the study sets out to test is ultimately not resolved. In my view, one likely reason is the experimental design itself. Because the predator is continuously present, threat presentation, exploration, risk assessment, sensory sampling, escape decision, and escape onset are all temporally entangled. Under these conditions, it becomes very difficult to isolate which component of the defensive sequence is actually being encoded, and this may reduce the ability to detect meaningful differences between excitatory and inhibitory populations. This is especially important given that prior work has already shown that dPAG populations vglut and vgat population encode and modulate differently defensive behaviours such as escape. In that context, the present study would need a cleaner behavioral design to demonstrate a distinct contribution of cell type, rather than a mixed representation of overlapping sensory, motor, and defensive variables
10)Finally, the proposed model in which local GABAergic inhibition organizes Assessment+ and Escape+ excitatory populations is interesting, but at present it remains speculative. The data may be consistent with this idea, but they do not directly test it. Since computational modelling of this framework would not be too complex, I would suggest additional analyses and network modelling to substantiate this claim, for example by showing that a network model under these constraints would recapitulate the neural responses observed in vivo.
Overall, I think the manuscript contains useful data, but in its present form its novel contribution is unclear. Additional experiments and analyses may be needed to strengthen its central claims, as described above.
Significance
See previous section
-
Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.
Learn more at Review Commons
Referee #1
Evidence, reproducibility and clarity
Peer review of Ayuso-Jimeno et al. "Excitatory and inhibitory neurons in the dorsal periaqueductal gray encode decisions to assess and escape natural threats"
Animals ability to escape from threat is a crucial survival behaviour exhibited across the animal kingdom. In vertebrates, hard-wired circuits allow animals to escape from predators without the need for learning. One of the main vertebrate brain regions that controls escape from threat is the brainstem dorsal periaqueductal gray. Research in the last decades has shown that dorsal PAG glutamatergic neurons control the initiation of escape (from imminent threat) and escape …
Note: This preprint has been reviewed by subject experts for Review Commons. Content has not been altered except for formatting.
Learn more at Review Commons
Referee #1
Evidence, reproducibility and clarity
Peer review of Ayuso-Jimeno et al. "Excitatory and inhibitory neurons in the dorsal periaqueductal gray encode decisions to assess and escape natural threats"
Animals ability to escape from threat is a crucial survival behaviour exhibited across the animal kingdom. In vertebrates, hard-wired circuits allow animals to escape from predators without the need for learning. One of the main vertebrate brain regions that controls escape from threat is the brainstem dorsal periaqueductal gray. Research in the last decades has shown that dorsal PAG glutamatergic neurons control the initiation of escape (from imminent threat) and escape vigour, whereas GABAergic neurons are spontaneously and tonically active and have been shown to provide an inhibitory threshold for eliciting escape and further signal escape termination (for reviews on this topic see for example: Gross and Canteras, 2012; Silva, Gross and Graeff, 2016; Motta, Carobrez and Canteras, 2017; Franklin, 2019; Silva and McNaughton, 2019; Lefler and Branco, 2020; Stempel, 2024). In addition to their role in the escape action, the dPAG has been suggested to have a broader function in threat processing. Specifically, a subset of neurons in the dorsal PAG has been shown have activity correlated to the approach/distance to a threat zone or predator. These neurons have been called 'risk assessment' neurons (Deng et al. 2016; Masferrer et al. 2020; Reis et al. 2021). Whether risk assessment neurons integrate threats across different modalities and contexts (here: social vs different predatory threats) is currently not known.
The present study addresses two important and long-standing questions in the field: what is the cell-type identity of dorsal PAG neurons encoding threat assessment versus escape?, and are different classes of threat processed by shared or dedicated neuronal populations? The study design is careful and methodologically performed well. While some of the experiments have been published in the past, the comparison of social and predator threat responses across contexts and a description of assessment+ cells across both main excitatory and inhibitory PAG neuron types is interesting.
Briefly, using miniaturized fluorescence microscopy (miniscope calcium imaging) in Vglut2::Cre and Vgat::Cre mice during a live predator (rat) exposure paradigm, the authors characterize neuronal activity in identified glutamatergic and GABAergic dPAG neurons across approach-escape cycles. The central finding is that both excitatory and inhibitory populations contain Assessment+ neurons (active during approach, silent at escape onset) and Escape+ neurons (suppressed during approach, activated at escape onset) and the authors propose that the activity profiles of Assessment+ and Escape+ cells may reflect local circuit wiring rules with putative GABAergic inhibition between excitatory neuronal subsets. Consistent with this, optogenetic activation of GABAergic dPAG neurons suppressed risk assessment behavior and promoted exploratory rearing but did not affect 'baseline' locomotion, building on Tsang et al. 2023 and Stempel et al. 2024. Population-level decoding using CEBRA (Schneider et al., 2023) confirmed that both Vglut2+ and Vgat+ ensembles independently encode behavioral state with high accuracy.
A second finding of this paper concerns threat generalization. Sequential exposure to a predator, an aggressive conspecific, and a prey insect (cockroach; paradigm from Rossier et al., 2021) revealed that more than half of responsive excitatory neurons and nearly half of responsive inhibitory neurons were activated by two or three threat types. This substantial overlap argues for at least partially convergent, rather than parallel, encoding of threat in dPAG, consistent with its role as a general trigger for defensive avoidance (Silva et al., 2013), and contrasts with the anatomically segregated upstream processing of predator and social threats in the medial hypothalamic defensive network. Taken together, this study suggests the dPAG as a site of coordinated excitatory-inhibitory computation in the control of innate threat assessment and avoidance across biologically diverse threat contexts.
Below are comments related to the figures/ data analysis and generally to the discussion part which we recommend should be expanded/changed to put the findings of the authors more in context of published literature and to discuss in more detail the proposed circuit mechanisms that align with the author's findings. Generally, this is a very nice and large dataset that could benefit from a more fine-grained and in-depth analysis of escape+ and risk assessment+ cells and their precise temporal activity profiles. We do not suggest to perform further experiments and think the work in this manuscript is publishable as is with some additional analyses and changes to the text.
Related Figure 1 and calcium imaging methods.
- The classification of neurons as 'Assessment+' and 'Escape+' positive is unclear and should be formally described in the methods, presumably these are the neurons with significantly increased "positive" or absolute slope?
- Can the authors clarify what they mean by escape? In the methods under "manual behavioral annotation", "Escape behavior" is defined as including retractions, retreats and flight. However, under "unsupervised behavioral annotation", escape is defined as "high-velocity locomotion aiming at increasing distance between threat source and subject". Are Escape+ neurons ones that are significantly modulated (presumably positively) at flight (latter definition), or throughout retractions, retreats and flight (former definition)? This also relates to the plots for 'escape+' neurons, where it would be useful to separately plot 'successful flights to shelter' to make the results comparable to previous studies and where trajectories are at least relatively stereotyped.
- More example traces of different FOVs aligned to escape-to-shelter onset and to risk assessment onset would be useful, as well as a plot with the % of escape-active neurons and a reliability index (in how many trials is each neuron active?). Similarly, are the proportions of escape+ and assessment+ neurons similar across animals and FOVs recorded?
- Some further basic analyses of the neurons' responses would be helpful to gage their activity profiles. Do these correspond to previously published descriptions of the two classes of escape+ neurons? Are they active during baseline locomotion? Do you observe the same clusters of GABAergic neurons that have been previously described where some dip at escape onset and some ramp up towards escape termination? Please add these plots as a supplement.
- It is generally assumed that assessment+ cells 'map'/correlate to the distance to a threat zone. Plots with quantification of this correlation would be useful, and whether they are also speed modulated or not? If the animal stops on the way to the threat zone, does the risk assessment signal plateau for example?
- Both in Figures 1 and 2, both the single trial examples of individual neurons and averages across neurons in the Vgat+ and VGlut2+ recordings show very fast changes that seem to be shorter than Gcamp6s kinetics would allow, and that happen exactly at escape onset when there is presumable a fast head turn movement. What motion correction controls do the authors have in place to make sure that some of the fast changes they see are not motion artefacts? (e.g., see Figure 2, panel C bottom.) (Importantly, see also comment 8 below).
- Related to this, the overall escape velocity is extremely low (around 10cm/s). When the authors only analyze high speed escapes (>50cm/s), do they see different cell activity profiles emerge that they might miss with these very low speed escapes that presumably activate less neurons that high-speed escapes? While slow-speed escapes still elicit activity in both Vgat+ and Vglut2+ neurons, their calcium activity changes will be much lower, potentially hindering a more detailed analysis, as clear signals might be sparser for escape+ neurons.
- With the min/max normalization that has been applied across the entire session it is hard to see 'local changes' in the heatmaps. Given that glutamatergic neurons are thought to only sparsely fire outside of escape episodes, the heatmaps are hard to read with the 'min/max' Z-scoring, and we would strongly encourage the authors to plot 'locally' Z-scored traces without a min/max normalization for each cell (at least for some examples). Importantly we would suggest to change the color scheme for the heatmaps to allow visual identification of the baseline / 'F0'.
- The calculation of the slope estimated on the calcium signal is somewhat unclear in the methods section. The way it is described right now, it is not clear whether the window is centered at the behavioral event of interest or not. In the figures, it appears that this is an absolute slope, which is also not clear from the methods. Further, considering the half-decay time of GCamP6f, which is ~ 0.5s (Chen et al., 2013), it seems like the result of this calculation could be cancelation or a slope near-zero if the signal both rises and decays in this broader window. Why is not the peak (e.g. max) or the AUC used for this analysis? The authors should clarify the calculation and provide rationale for their choice? Is the sign (positive/negative) of the slope not of interest?
- The GRIN lens placement in the example in Figure 1 is in the lateral PAG, whereas most others are located in the dorsolateral PAG. It would be useful to have a sentence in the introduction to state that the authors include the lateral, dorsolateral and dorsomedial PAG as 'dorsal PAG' and a rational for this placement.
- Could the authors comment on how the proportions of Assessment+ and Escape+ neurons relate to previously published literature (e.g., Deng et al. 2016)?
Related to Figure 2
- In panel K of Figure 2, both Vgat+ and VGlut2+ assessment+ neurons seem to have a rise at escape onset in addition to the slow rise during their movement towards the threat zone. Also here, the offset kinetics of the signal seem to now correlate well to the slow decay kinetics you would expect for GCamp6f and a quantification of controls and motion correction quality metrics would be very helpful to add. Depending on the baselining the peaks during escape would probably be significant as well. The authors could try and cluster the neurons further to see if they can disentangle further 'sub classes/clusters'.
- The calculation of the slope estimated on the calcium signal is somewhat unclear in the methods section. The way it is described right now, it is not clear to me whether the window is centered at the behavioral event of interest or not. In the figures, it appears that this is an absolute slope, which is also not clear from the methods. Further, considering the half-decay of GCamP6f which is ~ 0.5s (Chen et al., 2013) it seems like the result of this calculation could be cancelation or a slope near-zero if the signal both rises and decays in this broader window. Why is not the peak (e.g. max) or the AUC used for this analysis? The authors should clarify the calculation and provide rationale for their choice? Is the sign (positive/negative) of the slope not of interest?
- The authors find a significant increase in the accuracy of the classification of the two defensive behaviors (risk assessment versus escape) from models trained on the Vgat+ population calcium activity. This is an interesting finding which should be to the very least discussed in the discussion. Is there more information content in this population? Does this have anything to do with the slower dynamics of the signal? As mentioned above, would clustering these populations further reveal a more fine-grained detail of their population activity, e.g. see previously published work by Stempel et al. who suggest that there are at least two broad clusters of escape-active GABAergic neurons.
Related to Figure 3.
- In Figure 3A, the ChR2 seems to be significantly spread throughout the entire PAG and also the superior colliculus. As it is hard to see cell bodies with ChR2 at this magnification, we would recommend to add a supplemental figure with the outlines of the infection sites.
- The authors state that: "Interestingly, although stimulation of Vgat⁺ dPAG neurons significantly reduced the peak speed of escape, it did not change the overall time spent escaping". Could they comment on how this may relate to a previous study where the probability of escape is decreased upon activation of Vgat+ neurons and initiated escapes can be induced (Stempel et al. 2024)? Do they think this could be differences in location of stimulation fiber or one vs two injection sites (bigger vs smaller spread along the AP axis of the PAG?) - or do they think there could be a fundamental difference between fast escape from imminent looming stimuli to slower escapes to rat/prey/social predators?
Related to Figure 4.
- Individual examples of cells responding to one threat vs multiple threats would be very useful to add to see their dynamics across threats.
- The Venn diagrams in panel I are very hard to read, and color-blind people might struggle with the red/green (also in panel C). We would suggest to change colors, and replace Venn diagrams to make results more interpretable/readable.
- Have any statistical tests been performed to assess whether there is a difference between the proportions in figure D and H as well as Figure 4J?
General and other comments
- We strongly encourage the authors to add example videos of the behaviors tested and of all major findings including example calcium activity and optogenetic manipulations.
- Relating to a more fine-grained analysis of escape-active cells (as recommended above), the authors state in the discussion that: 'A notable paradox of our findings is that while GABAergic and glutamatergic dPAG neurons showed nearly indistinguishable neural firing correlates of approach and avoidance (both harbored Assessment+ and Escape+ cells) [...]". Since previous works have already looked at escape-active cells in more detail, with differences between Vgat+ and Vglut2+ neurons having been described, we would encourage the authors to look at unsupervised clustering of the escape-active neuron populations, as the firing rates of these neurons should be distinguishable along the escape sequence and especially when taking into account speed correlations and activity profiles aligned to escape onset or offset. If they can't find differences these should be discussed, as major differences may relate to the paradigms used (rats vs visual 'looming' threats).
- We would ask the authors to make sure that the cited literature supports their statements. Some statements are in the manuscript are not clearly, only partially correct or the literature itself is inconsistent or citations are confusing/misleading.
e.g.: "A functional cellular and circuit architecture of dorsal PAG is emerging in which stimulation of glutamatergic neurons in dorsal PAG promotes freezing at low intensity and flight at high intensity (Tovote et al. 2016, Deng et al. 2016, Evans et al. 2018, Tsang et al. 2023)."
Not all of the cited papers support that graded stimulation of glutamatergic neurons in the dorsal PAG results in escape/flight or freezing but rather that there is some (dis-)inhibitory interplay between dorsal and ventrolateral PAG which results in the selection of one or the other behaviors, and that the ventrolateral PAG more specifically drives freezing (e.g., Tovote et al. 2016). Some studies have shown that activation of the dorsal PAG and in particular of glutamatergic neurons elicits 'all-or-none' flight behavior. Additionally, studies where freezing has been observed with dorsal PAG stimulation have often used CamKIIa as a promoter (e.g. Deng et al. 2016*), which is expressed at significant levels in both excitatory and inhibitory neurons. Thus, we would argue it is not entirely clear whether freezing through 'low intensity' dPAG stimulation is a biological feature of the dPAG network, especially also as inhibition of the dPAG promotes freezing - probably through disinhibition of the vlPAG.
In the results section on page 6, it is stated that Deng et al. 2016 use VGlut2-Cre to target ChR2 for eliciting escape, but they have used CamKIIa ("Previous studies have shown that optogenetic activation of Vglut2+ dPAG neurons elicited flight behavior[...]"). Similarly, on page 3, the following statement on GABAergic neurons isn't clear: "Stimulation of GABAergic neurons, on the other hand, does not elicit defensive behavior (Tsang et al., 2023) and recent evidence shows that they can inhibit looming stimulus-evoked flight behavior (Stempel et al., 2024) suggesting they might act as part of a tonic inhibitory circuit that receives primarily local inputs (Franklin et al., 2017; but see Wu et al. 2024)." Multiple papers across all columns have shown that GABAergic neurons are tonically and spontaneously active (e.g., Chen et al., 2023; Wang et al., 2023; Stempel et al., 2024) and Stempel et al. estimated that GABAergic PAG neurons make up >50% of spontaneous inputs to glutamatergic PAG neurons, thus positioning them well to control their activity. Franklin et al. 2017 (and others) have shown that GABAergic neurons receive input from thalamic and hypothalamic regions (probably also from midbrain regions that were not analyzed in that study), thus arguing that they receive significant input from outside of the PAG to integrate - currently unknown - inputs from other brain regions to presumably guide their activity in a context-specific manner. The statement that they should thus be part of a circuit that receives primarily local inputs is not well supported by previous studies. Also, how this enables a 'push-pull' circuit and what is meant here exactly, could be clarified. It is not clear why the presence of risk assessment neurons would support such a circuit model rather than an integrative/threshold one. Further, Stempel et al. 2024 have also shown that GABAergic neurons ramp up activity during escape and are thus technically 'flight+' cells. We agree that there are interesting and presumably complicated intra-PAG dynamics to be studied where their connectivity will define different circuit model possibilities.
Referee cross-commenting
We generally agree with most comments made by reviewer 2, and their concerns. While the novelty of this manuscript is limited, highlighting the advances better, also through more accurate analyses, should allow the authors to better build up their novel findings ('major concern 1'). We agree with the reviewers 'major concerns' 2-5 and that they should be addressed by the authors, for scientific reasons beyond novelty, as some of the analysis feels incomplete. For example, while we do not think that new experiments are necessary, the authors can discuss the issue of spatiotemporal overlap of variables that do not allow a definitive description of 'risk assessment' cells. This issue is in fact present in most published work on risk assessment cells (usually described with a rat predator that is continuously present). Concerns 7-9 are valid, and touch on similar points raised by our review. We do not think that modelling would add anything substantial to this manuscript that couldn't be hypothesised in the discussion beyond their proposed model. In summary, we agree with the points raised by reviewer too, but find that the data is useful and should be published, after some revising and additional more fine-grained analyses to strengthen the manuscript's claims.
Significance
Overall, even though the overall novelty is somewhat limited considering the paper shows significant overlap with recent findings on the involvement of glutamatergic and GABAergic dorsal PAG neurons in risk assessment and escape, there are some important and novel findings that advance the knowledge in the field of neuronal circuits underlying defensive behaviors. For example, the direct comparison of how different threat contexts modulate the activity of individual neurons and the finding that both GABAergic and glutamatergic population activity can predict defensive behavioral output are interesting and important findings. While this study may of limited interest to a very broad audience, it is of importance to the fields of neuroethology and the study of the neural circuits underlying innate defensive behaviors.
-
