Anterior cingulate cortex in complex associative learning: monitoring action state and action content
Curation statements for this article:-
Curated by eLife
eLife Assessment
Huang and colleagues examined neural responses in mouse anterior cingulate cortex (ACC) during a discrimination-avoidance task. The authors present useful findings that ACC neurons encode primarily post-action variables or "action content" over extended periods. Though the methodological approach was sound, the evidence in support of action state encoding, ruling out alternative explanations related to movement, is incomplete.
This article has been Reviewed by the following groups
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
- Evaluated articles (eLife)
Abstract
Environmental changes necessitate adaptive responses, and thus the ability to monitor one’s actions and their connection to specific cues and outcomes is crucial for survival. The anterior cingulate cortex (ACC) is implicated in these processes, yet its precise role in action monitoring versus outcome tracking remains unclear. To investigate this, we developed a novel discrimination–avoidance task for mice, designed with clear temporal separation between actions and outcomes. Our findings show that ACC neurons primarily encode post-action variables over extended periods, reflecting the animal’s preceding actions rather than the outcomes or values of those actions. Specifically, we identified two distinct subpopulations of ACC neurons: one encoding the action state (whether an action was taken) and the other encoding the action content (which action was taken). Importantly, increased post-action ACC activity was associated with better performance in subsequent trials. These findings suggest that the ACC supports complex associative learning through extended signaling of rich actionrelevant information, thereby bridging cue, action, and outcome associations.
Article activity feed
-
eLife Assessment
Huang and colleagues examined neural responses in mouse anterior cingulate cortex (ACC) during a discrimination-avoidance task. The authors present useful findings that ACC neurons encode primarily post-action variables or "action content" over extended periods. Though the methodological approach was sound, the evidence in support of action state encoding, ruling out alternative explanations related to movement, is incomplete.
-
Reviewer #1 (Public review):
Summary:
Huang et al. examined ACC response during a novel discrimination-avoid task. The authors concluded that ACC neurons primarily encode post-action variables over extended periods, reflecting the animal's preceding actions rather than the outcomes or values of those actions. The authors have made considerable revisions to address the raised concerns. However, it appears that some important issues remain unresolved.
Strengths:
The inclusion of new figures and analyses in response to the reviews is appreciated, such as Fig. 2 and 5.
Weaknesses:
Motion related signal in ACC: the new Fig. 2E looks good, but it is hard to visualize how it is just a reordering of the old Fig. 5C.
All categories in the new Fig. 4D appear to respond to shuttle initiation, with less than 1s latency. For example, type 2a/2b …
Reviewer #1 (Public review):
Summary:
Huang et al. examined ACC response during a novel discrimination-avoid task. The authors concluded that ACC neurons primarily encode post-action variables over extended periods, reflecting the animal's preceding actions rather than the outcomes or values of those actions. The authors have made considerable revisions to address the raised concerns. However, it appears that some important issues remain unresolved.
Strengths:
The inclusion of new figures and analyses in response to the reviews is appreciated, such as Fig. 2 and 5.
Weaknesses:
Motion related signal in ACC: the new Fig. 2E looks good, but it is hard to visualize how it is just a reordering of the old Fig. 5C.
All categories in the new Fig. 4D appear to respond to shuttle initiation, with less than 1s latency. For example, type 2a/2b consists of 40% of the population and their response to movement onset is apparent. Thus, it is not clear whether most neurons respond to shuttle crossing as described in the manuscript.
Could the authors use relatively simple analysis, such as comparing spike rate before and after crossing, or before and after initiation, to quantify the response properties of each neuron? This could also help validate the classification analysis performed in Fig. 4.
-
Reviewer #2 (Public review):
Summary:
Huang et al recorded anterior cingulate cortex activity in mice while they performed a shuttle escape task. The task utilized two auditory cues, each of which informed the mice to stay or escape depending on which side they were on, and incorrect responses were punished by shock administration. Analyses focused on ACC neurons that fired when mice crossed the shuttle box in either direction (A-->B or B-->A), coined "action state", or when mice crossed in one direction but not the other, coined "action content". The authors characterized these populations, and ACC firing changes mostly occurred around the time of shuttle crossing. This work will likely be of broad interest to those who are interested in neocortical neurophysiology broadly, anterior cingulate cortex specifically, and their …
Reviewer #2 (Public review):
Summary:
Huang et al recorded anterior cingulate cortex activity in mice while they performed a shuttle escape task. The task utilized two auditory cues, each of which informed the mice to stay or escape depending on which side they were on, and incorrect responses were punished by shock administration. Analyses focused on ACC neurons that fired when mice crossed the shuttle box in either direction (A-->B or B-->A), coined "action state", or when mice crossed in one direction but not the other, coined "action content". The authors characterized these populations, and ACC firing changes mostly occurred around the time of shuttle crossing. This work will likely be of broad interest to those who are interested in neocortical neurophysiology broadly, anterior cingulate cortex specifically, and their contributions to learning about actions. The task is well-designed and provides a nice background for neurophysiological recordings. The authors leveraged these strengths in characterizing the neural populations that fire to shuttle crossings in both directions vs one direction.
Strengths:
The factorial design nicely controls for sensory coding and value coding, since the same stimulus can signal different actions and values.
The figures are well presented, labeled, and easy to read.
Additional analyses, such as the 2.5/7.5s windows and place-field analysis, are nice to see and indicate that the authors were careful in their neural analyses.
The n-trial + 1 analysis where ACC activity was higher on trials that preceded correct responses is a nice addition, since it shows that ACC activity predicts future behavior, well before it happens.
The authors identified ACC neurons that fire to shuttle crossings in one direction or to crossings in both directions. This is very clear in the spike rasters and population scaled color images. While other factors such as place fields, sensory input, and their integration can account for this activity, the authors discuss this and provide additional supplemental analyses.
Weaknesses:
Some of the neural analyses could use the necessary and sufficient comparisons to strengthen the authors' claims.
Comment on revised version:
I think the authors did a very admirable job revising the manuscript. It is much improved. However, I believe a formal analysis of action-state versus action-content neurons on A-->B versus B-->A crossing is still warranted. I appreciate the fact that this analysis may not be as reliable with smaller ensemble sizes, but with careful pseudo-ensemble and resampling approaches, such an analysis would go a long way towards increasing the strength of evidence.
-
Reviewer #3 (Public review):
Summary:
The authors record from the ACC during a task in which animals must switch contexts to avoid shock as instructed by a cue. As expected, they find neurons that encode context, with some encoding of actions prior to the context, and encoding of neurons post-action. The primary novelty is dynamic encoding of action-outcome in a discrimination-avoidance domain, while this is traditionally done using operant methods.
Comments on revised version:
I appreciate the considerable work done on review, and additional details added throughout. I also noted the additional sessions included in analyses, and additional behavioral data in response to R1 and R2's insightful comments.
The only remaining comment that was not addressed pertains to anatomy and recording details. Some electrodes appear to be clearly in M2 …
Reviewer #3 (Public review):
Summary:
The authors record from the ACC during a task in which animals must switch contexts to avoid shock as instructed by a cue. As expected, they find neurons that encode context, with some encoding of actions prior to the context, and encoding of neurons post-action. The primary novelty is dynamic encoding of action-outcome in a discrimination-avoidance domain, while this is traditionally done using operant methods.
Comments on revised version:
I appreciate the considerable work done on review, and additional details added throughout. I also noted the additional sessions included in analyses, and additional behavioral data in response to R1 and R2's insightful comments.
The only remaining comment that was not addressed pertains to anatomy and recording details. Some electrodes appear to be clearly in M2 (Fig 2A), and the tetrodes were driven each day. I would strongly suggest that this be included as a further limitation, particularly given the statement on line 178.
-
Author response:
The following is the authors’ response to the original reviews.
Public Reviews:
Reviewer #1 (Public review):
Summary:
In the current study, Huang et al. examined ACC response during a novel discrimination-avoid task. The authors concluded that ACC neurons primarily encode post-action variables over extended periods, reflecting the animal's preceding actions rather than the outcomes or values of those actions. Specifically, they identified two subgroups of ACC neurons that responded to different aspects of the actions. This work represents admirable efforts to investigate the role of ACC in task-performing mice. However, in my opinion, alternative explanations of the data were not sufficiently explored, and some key findings were not well supported.
Strengths:
The development of the new discrimination-avoid task is …
Author response:
The following is the authors’ response to the original reviews.
Public Reviews:
Reviewer #1 (Public review):
Summary:
In the current study, Huang et al. examined ACC response during a novel discrimination-avoid task. The authors concluded that ACC neurons primarily encode post-action variables over extended periods, reflecting the animal's preceding actions rather than the outcomes or values of those actions. Specifically, they identified two subgroups of ACC neurons that responded to different aspects of the actions. This work represents admirable efforts to investigate the role of ACC in task-performing mice. However, in my opinion, alternative explanations of the data were not sufficiently explored, and some key findings were not well supported.
Strengths:
The development of the new discrimination-avoid task is applauded. Single-unit electrophysiology in task-performing animals represents admirable efforts and the datasets are valuable. The identification of different groups of encoding neurons in ACC can be potentially important.
Weaknesses:
One major conclusion is that ACC primarily encodes the so-called post-action variables (specifically shuttle crossing). However, only a single example session was included in Figure 2, while in Supplementary Figure 2 a considerable fraction of ACC neurons appears to respond to either the onset of movement or ramp up their activity prior to movement onset. How did the authors reach the conclusion that ACC preferentially respond to shuttle crossing?
We now include more example sessions and the main results from individual animals (Fig. 3; Figs. S2–S3; Fig. 8). Overall, the results are consistent across recording sessions and animals.
While shuttle crossings were the primary reference for most analysis, using shuttle initiation as a reference led to similar conclusions (Fig.4). Namely, we found that most ACC neurons exhibit either robust (22%; Types 1a & 2a) or moderate (51%; Types 1b & 2b) post-shuttle activity changes (Fig.4), while only a subset exhibits ramping pre-shuttle activity (16%; Types 3b & 3c). Therefore, our conclusion was intended to highlight the role of post-shuttle activity in learning. While we do not exclude the possibility that pre-shuttle ACC activity contributes to learning, its involvement is likely more limited.
In Figure 4, it was concluded that ACC neurons respond to action independent of outcome. Since these neurons are active on both correct and incorrect shuttle but not stay trials, they seem to primarily respond to overt movement. If so, the rationale for linking ACC activity and adaptive behavior/ associative learning is not very clear to me. Further analyses are needed to test whether their firing rates correlated with locomotion speed or acceleration/deceleration. On a similar note, to what extent are the action state neurons actually responding to locomotion-related signals? And can ACC activity actually differentiate correct vs. incorrect stays?
In this study, we highlight two distinct groups of ACC neurons: action-state and action-content neurons. Both groups of neurons tend to show sustained activity even when the animals remain immobile after completing shuttle behaviors, suggesting that their activity is not directly driven by locomotion. Furthermore, action-content neurons are selectively engaged in only one of the two shuttle categories, either rooms A→B or B→A shuttles. Therefore, differences in neuronal activity are unlikely to reflect locomotor differences, given that both shuttle types involve similar movement patterns. Finally, we analyzed ACC neuronal activity in relation to locomotion speed. Our results indicate that only a small fraction of neurons (<15%) show speed-correlated activity (Fig.5), suggesting that most ACC neurons do not encode movement-related information. Taken together, these findings support the distinction between ACC activity and locomotion encoding.
As for the small subset of speed-related neurons, it remains unclear whether these speed-related neurons represent a distinct subpopulation within the ACC or reflect recordings from the nearby motor cortex. Postmortem examination of the recording sites suggests that most neurons were recorded from the ACC, while a small subset may be located at the border between the ACC and motor cortex (Fig. S2). Therefore, it is possible that the small fraction of speed-related neurons originated from the motor cortex.
Lastly, given that the ACC neurons display no or limited activity during stay trials, their activity generally does not differentiate correct vs. incorrect stays (Fig.S7). However, ACC activity does show moderate differentiation between room-A vs. room-B stays (Fig.S7).
Given that a considerable amount of ACC neurons encode 'action content', it is not surprising that by including all neurons the model is able to make accurate predictions in Figure 6. How would the model performance change by removing the content neurons?
We thank the reviewer for this thoughtful analysis idea. Excluding action-content neurons drastically reduces decoding accuracy (Fig.8), suggesting that they are the main drivers for differentiating rooms A→B vs. B→A shuttles.
Moving on to Figure 7. Since Figure 4 showed that ACC neurons respond to movement regardless of outcome, it is somewhat puzzling how ACC activity can be linked to future performance.
As discussed earlier (point #2), ACC activity does not simply reflect locomotion itself. We interpret the post-shuttle ACC activity as encoding both the preceding shuttle state (shuttle or stay) and shuttle content (rooms A→B or B→A). Regardless of the outcome (safety or shock), such encoding is essential for cue–action–outcome associative learning, because both positive and negative feedback can drive learning. The level of post-shuttle ACC activity may reflect task engagement, with greater engagement facilitating learning and improving future performance.
Two mice contributed about 50% of all the recorded cells. How robust are the results when analyzing mouse by mouse?
We have added further analysis of highlighting the results of each mouse. Although the total number of recorded neurons varied across mice, the major findings were consistent. In every mouse, we observed sustained post-shuttle ACC activity (Fig.S2), and population-level ACC activity reliably decoded shuttle contents (rooms A→B vs. B→A; Fig.8).
Lastly, the development of the new discrimination-avoid task is applauded. However, a major missing piece here is to show the importance of ACC in this task and what aspects of this behavior require ACC.
We appreciate this feedback. We are currently conducting additional experiments to determine whether inhibiting ACC activity during distinct time windows disrupts task learning. We hope to publish a follow-up paper on these findings in the near future.
Reviewer #2 (Public review):
Summary:
The current dataset utilized a 2x2 factorial shuttle-escape task in combination with extracellular single-unit recording in the anterior cingulate cortex (ACC) of mice to determine ACC action coding. The contributions of neocortical signaling to action-outcome learning as assessed by behavioral tasks outside of the prototypical reward versus non-reward or punished vs non-punished is an important and relevant research topic, given that ACC plays a clear role in several human neurological and psychiatric conditions. The authors present useful findings regarding the role of ACC in action monitoring and learning. The core methods themselves - electrophysiology and behavior - are adequate; however, the analyses are incomplete since ruling out alternative explanations for neural activity, such as movement itself, requires substantial control analyses, and details on statistical methods are not clear.
Strengths:
(1) The factorial design nicely controls for sensory coding and value coding, since the same stimulus can signal different actions and values.
(2) The figures are mostly well-presented, labeled, and easy to read.
(3) Additional analyses, such as the 2.5/7.5s windows and place-field analysis, are nice to see and indicate that the authors were careful in their neural analyses.
(4) The n-trial + 1 analysis where ACC activity was higher on trials that preceded correct responses is a nice addition, since it shows that ACC activity predicts future behavior, well before it happens.
(5) The authors identified ACC neurons that fire to shuttle crossings in one direction or to crossings in both directions. This is very clear in the spike rasters and population-scaled color images. While other factors such as place fields, sensory input, and their integration can account for this activity, the authors discuss this and provide additional supplemental analyses.
Weaknesses:
(1) The behavioral data could use slightly more characterization, such as separating stay versus shuttle trials.
We appreciate this feedback. In the revised manuscript, we present data separating stay versus shuttle trials (Fig.1). Additionally, we provide new data from extended training sessions (Fig.S2).
(2) Some of the neural analyses could use the necessary and sufficient comparisons to strengthen the authors' claims.
We have now used the necessary and sufficient comparisons where applicable. In the SVM decoding analysis, we show that population ACC activity is sufficient to decode A→B or B→A shuttles. We also show that excluding action-content, but not other ACC neurons, drastically reduces decoding accuracy, suggesting that these neurons are necessary for the decoding (Fig.8).
(3) Many of the neural analyses seem to utilize long time windows, not leveraging the very real strength of recording spike times. Specifics on the exact neural activity binning/averaging, tests, classifier validation, and methods for quantification are difficult to find.
We chose to perform our neural analyses on a longer time scale, given the sustained activity we see in the data. To further justify that decision, we now provide additional results highlighting the sustained activity of ACC neurons in our task (Fig.2; Fig.S2). Additionally, we now provide more specifics of the neural analyses in Methods section.
(4) The neural analyses seem to suggest that ACC neurons encode one variable or the other, but are there any that multiplex? Given the overwhelming evidence of multiplexing in the ACC a bit more discussion of its presence or absence is warranted.
This is an interesting point of discussion, and we thank the reviewer for pointing this out. Overall, our results suggest that individual ACC neurons preferentially engage in only one of the proposed functions, rather than multiplexing across them. For example, action-state and action-content ACC neurons primarily engage in action monitoring, but not in decision-making, planning, or outcome tracking. Nevertheless, we cannot rule out the possibility that other ACC neurons, through their distinct connectivity or location in different ACC subregions, engage in other proposed functions. Thus, when considering the ACC as a whole, its function may still be multiplexed.
Another possible reason we do not see clear multiplexing of neurons may be due to the dynamic nature of our task. Unlike established tasks that often assign fixed positive or negative values to cues, the cues in our task are not inherently associated with valence. Instead, their meaning is dynamically determined by the animal’s location (context) at the time of cue presentation. Since values are not fixed and change based on context, value-related responses may not be reflected in the ACC in our tasks.
We have now incorporated the above discussions into our revised manuscript.
Reviewer #3 (Public review):
Summary:
The authors record from the ACC during a task in which animals must switch contexts to avoid shock as instructed by a cue. As expected, they find neurons that encode context, with some encoding of actions prior to the context, and encoding of neurons post-action. The primary novelty of the task seems to be dynamically encoding action-outcome in a discrimination-avoidance domain, while this is traditionally done using operant methods. While I'm not sure that this task is all that novel, I can't recall this being applied to the frontal cortex before, and this extends the well-known action/context/post-context encoding of ACC to the discrimination-avoidance domain.
While the analysis is well done, there are several points that I believe should be elaborated upon. First, I had questions about several details (see point 3 below). Second, I wonder why the authors downplayed the clear action coding of ACC ensembles. Third, I wonder if the purported 'novelty' of the task (which I'm not sure of) and pseudo-debate on ACC's role undermines the real novelty - action/context/outcome encoding of ACC in discrimination-avoidance and early learning.
Strengths:
Recording frontal cortical ensembles during this task is particularly novel, and the analyses are sophisticated. The task has the potential to generate elegant comparisons of action and outcome, and the analyses are sophisticated.
Weaknesses:
I had some questions that might help me understand this work better.
(1) I wonder if the field would agree that there is a true 'debate' and 'controversy' about the ACC and conflict monitoring, or if this is a pseudodebate (Line 34). They cite 2 very old papers to support this point. I might reframe this in terms of the frontal cortex studying action-outcome associations in discrimination-avoidance, as the bulk of evidence in rodents comes from overtrained operant behavior, and in humans comes from high-level tasks, and humans are unlikely to get aversive stimuli such as shocks.
We appreciate this feedback. We have revised the Introduction and Discussion.
(2) Does the purported novelty of the task undermine the argument? While I don't have an exhaustive knowledge of this behavior, the novelty involves applying this ACC. There are many paradigms where a shock triggers some action that could be antecedents to this task.
We argue our newly designed discrimination–avoidance task is unique for several reasons. First, it requires animals to discriminate both sensory cues and environment contexts. Unlike established tasks that often assign fixed positive or negative values to cues, the cues in our task are not inherently associated with valence. Instead, their meaning is dynamically determined by the animal’s location (context) at the time of cue presentation, which reflects a conceptual advance over previous techniques. Furthermore, by removing valence from the cues, this design helps disentangle the ACC’s potential role in value encoding from other cognitive functions.
Second, this task involves robust, ethologically relevant actions (i.e., shuttles), unlike many established paradigms that rely on less naturalistic behaviors such as saccades or lever presses. We view this as a key distinction from prior approaches, as even previous paradigms that utilize shutting responses or other naturalistic responses, fail to incorporate dynamic integration of cues and contexts.
Finally, the clear temporal separation between actions and outcomes further helps disentangle the ACC’s roles in action monitoring vs. outcome tracking.
(3) The lack of details was confusing to me:
(a) How many total mice? Are the same mice in all analyses? Are the same neurons? Which training day? Is it 4 mice in Figure 3? Five mice in line 382? An accounting of mice should be in the methods. All data points and figures should have the number of neurons and mice clearly indicated, along with a table. Without these details, it is challenging to interpret the findings.
We are sorry for the confusion. We now provide additional details and clear N numbers for each analysis to improve clarity.
(b) How many neurons are from which stage of training? In some figures, I see 325, in some ~350, and in S5/S2B, 370. The number of neurons should be clearly indicated in each figure, and perhaps a table.
All data were obtained from well-trained mice. For some analyses, the N is smaller because certain task sessions contained very few incorrect trials (≤3), which prevented us from examining ACC activity during those trials. We have modified figure legend so that neuron count is clear.
(c) Were the tetrodes driven deeper each day? The depth should be used as a regressor in all analyses?
Yes, the tetrodes were driven slightly deeper across task sessions (~80 µm per step; 2–4 depths per mouse). Given limited depth changes, preliminary analyses indicate no clear differences in ACC activity across these recording depths. However, we cannot rule out potential dorsal–ventral subregion differences if recordings were to span larger depth ranges.
(d) Was is really ACC (Figure 2A)? Some shanks are in M2? All electrodes from all mice need to be plotted as a main figure with the drive length indicated.
We have now included a supplementary figure showing all recording sites (Fig.S2). It is likely that a small subset of neurons was recorded at the ACC/M2 border area. Unfortunately, we are unable to separate them out due to blind recording design of our tetrode arrays.
(e) It's not clear which sessions and how many go into which analysis
We have now specified the number of task sessions for each analysis (see Methods).
(f) How many correct and incorrect trials (<7?) are there per session?
We have now specified the number of correct and incorrect trials per session (see Methods).
(g) Why 'up to 10 shocks' on line 358? What amplitudes were tried? What does scrambled mean?
We decided to use up to 10 mild shocks per trial because mice do not necessarily shuttle to the safe room after one or even a few shocks during the early stages of training. This design allows mice to efficiently learn the concept of the task (i.e., one room is safe while the other delivers shocks). Each shock was specified in the Methods section as 0.5 mA, 0.1 s. A “scrambled shock” refers to an electric shock delivered through multiple floor bars in a randomized pattern, effectively preventing the animal from avoiding the stimulus.
(4) Why do the authors downplay pre-action encoding? It is clearly evident in the PETHs, and the classifiers are above chance. It's not surprising that post-shuttle classification is so high because the behavior has occurred. This is most evident in Figure S2B, which likely should be a main figure.
We did not intend to downplay pre-action encoding. Our analysis shows that most ACC neurons exhibit either robust (22%; Types 1a & 2a) or moderate (51%;Types 1b & 2b) post-shuttle activity changes (Fig.4). Although a subset of ACC neurons exhibits ramping pre-shuttle activity, they represent a much smaller fraction (16%; Types 3b & 3c). Therefore, our conclusion was intended to highlight the role of post-shuttle activity in learning. While we do not exclude the possibility that pre-shuttle ACC activity contributes to learning, its involvement is likely more limited
(5) The statistics seem inappropriate. A linear mixed effects model accounting for between-mouse variance seems most appropriate. Statistical power or effect size is needed to interpret these results. This is important in analyses like Figure 7C or 6B.
We appreciate this feedback. We now use appropriate statistics and report effect size.
(6) Better behavioral details might help readers understand the task. These can be pulled from Figures S2 and S5. This is particularly important in a 'novel' task.
We now provide more details to help better understand the task and have added new figures (Fig.1; Figs. S1&S2).
(7) Can the authors put post-action encoding on the same classification accuracy axes as Figure 6B? It'd be useful to compare.
We appreciate the comment, but we are unsure what clarification is being requested.
(8) What limitations are there? I can think of several - number of animals, lack of causal manipulations, ACC in rodents and humans.
We now include discussions on limitation of our study. One caveat of our study is that the discrimination–avoidance task requires weeks of training in mice. By the time they master the task, ACC activity may reflect modified neural circuits. Investigating ACC activity during early phase of learning, such as by introducing a new pair of cues or contexts, could provide further insights into ACC’s role in learning and cognitive processes. Additionally, a limitation of the current study is the lack of evidence for the causal role of post-action ACC activity in complex associative learning. Future investigations using closed-loop strategies to selectively disrupt ACC activity during the post-action phase could help address this question.
Minor:
(1) Each PCA analysis needs a scree plot to understand the variance explained.
We have added a scree plot for each PCA analysis.
(2) Figure 4C - y and x-axes have the same label?
We have corrected the y-axis label.
(3) What bin size do the authors use for machine learning (Not clear from line 416)?
The bin sizes used were 2.5, 5, 7.5, or 10 sec which have now been discussed in the Methods section.
(4) Why not just use PCA instead of 'dimension reduction' (of which there are many?)
We have adjusted the phrasing where appropriate.
(5) Would a video enhance understanding of the behavior?
We appreciate this feedback. We now include a few videos to accompany our paper.
Recommendations for the authors:
Reviewer #1 (Recommendations for the authors):
(1) Is Figure 1C sufficiently powered?
We have now included data from additional mice and updated the figure accordingly.
(2) Task performance was not plateaued after 10 sessions in Figure 1B. How variable is task performance in the datasets with ephys recordings (session to session, mouse to mouse).
We have now included additional data from extended training (15 sessions; Fig.S2). Moderate variations across both sessions and mice are observed. Specifically, the total number of correct/incorrect shuttles used for ephys analysis are 19/5, 19/4, 21/5, 20/4 (mouse #1; 4 sessions); 20/7, 23/7, 20/7 (mouse #2; 3 sessions); 19/4, 16/2 (mouse #3; 2 sessions); 26/4, 23/4, 17/6, 25/5 (mouse #4; 4 sessions); 20/5, and 17/4 (mouse #5; 2 sessions), respectively.
(3) Please quantify the results in Figure 3, for both within individual mice and across mice.
We have calculated maximum trajectory length within the 3-D space (Fig. 3C).
(4) What is the effect size in Figure 7C?
We now report the effect size.
(5) Please provide more details for spike sorting.
We have now included more details in the Methods section.
(6) More detailed cell type or correlation analysis in Figures 4 and 5 may be helpful. For example, if putative regular and fast-spiking neurons were simultaneously recorded, did the FS directly inhibit the RS to give rise to the apparent encoding properties?
We recorded a small number of putative interneurons (n = 13) from only three mice, which precludes drawing meaningful conclusions, particularly given their heterogeneous responses during discrimination–avoidance tasks. Accordingly, we include only an example interneuron demonstrating discrimination between A→B vs. B→A shuttles (Fig. S5). Nevertheless, it is evident there are reciprocal monosynaptic connections between putative interneurons and certain pyramidal neurons, as indicated by short-latency (~2 ms) excitatory or inhibitory interactions (Fig. S5). That said, follow up studies with greater Ns are needed to parse out these details
Reviewer #2 (Recommendations for the authors):
(1) While I appreciate displaying the success rate for the sake of simplifying behavioral data in Figure 1B, it would be nice to also see these data broken out as correct vs incorrect for stay vs shuttle trials, since it is difficult to determine whether the performance increases are primarily driven by mice improving at stay vs shuttle responses
We appreciate this feedback. In the revised manuscript, we present data separating stay versus shuttle trials (Fig.1; Fig.S2).
(2) In Figure 2 the comparison between shuttle and stay is not particularly convincing, since the comparison is also essentially movement vs no movement and place1-->place2 vs place1-->place1. A more appropriate comparison might be action state neurons vs action content neurons during A-->B, B-->A, or both crossings. If it is true that these populations contain this information, then action state neurons should traverse a large component space in both directions, action content neurons only one direction, and so on.
We agree that the comparison is not ideal due to differences in locomotion. However, it provides valuable information suggesting that the ACC plays a limited role during stay trials, despite these trials involve mental and cognitive processes comparable to shuttle trials. While we appreciate the reviewer’s suggestion, the proposed analysis is not particularly reliable given the relatively small number of simultaneously recorded action-state or action-content neurons.
(3) I would say the above point applies to Figure 3 as well. I would also note that this reviewer greatly appreciates the rigor of showing ensemble activity in each subject.
We appreciate this comment. See our response above.
(4) In Figure 5 do these neurons show the same A-->B vs B-->A firing patterns during correct vs incorrect shuttles? The text describing the data in Figure 4 suggests this should be the case but even from a quick glance it sort of seems like the population dynamics during correct vs incorrect shuttles are not the same. My concern is that averaging neural activity over 5s windows washes out all these dynamics
Preliminary analysis suggests that these firing patterns apply to both correct and incorrect shuttles. However, the main reason we did not compare correct and incorrect trials is the limited amount of data. In many sessions, there are only a few (≤5) incorrect shuttles, which include both A→B or B→A shuttles (Fig.1C; Fig.S2), thus lacking the statistical power for a meaningful comparison.
(5) Some information on classifier validation is required - was this leave-out validation and if so how many trials were left-out vs tested? K-fold, and if so, how many folds? Was the trial order shuffled for each simulation? Classifiers will pick up within-session temporal information. In addition to this classifier accuracy during the different time points should be compared by a non-parametric test, and compared to the 95th percentile of the label-shuffled distribution.
Yes, we use standard 10-fold cross-validation. We appreciate the suggestion on trial-order shuffling, and implementing this procedure does not change our original conclusion. Additionally, we have applied a non-parametric test.
(6) How exactly were neurons classified as content vs state? Was it the average activity during the 5s following the shuttle? If this is stated I could not really find it easily so I might suggest clarifying.
We now use a new method for classification of the two neuron types (Fig.7). We have included detailed methods in the revised manuscript.
(7) Movement drives cortical neuron activity more than anything else I have ever seen. Really, more than anything else, it would be nice to demonstrate that it is not movement alone or movement multiplexed with place/sensory information/direction driving these responses.
We have analyzed ACC neuronal activity in relation to locomotion speed. Our results indicate that only a small fraction of ACC neurons (<15%) show speed-correlated activity (Fig.5). It remains unclear whether these speed-related neurons represent a distinct subpopulation within the ACC or reflect recordings from nearby motor cortex. Postmortem examination of the recording sites suggests that most neurons were recorded from the ACC, while a small subset may be located at the border between the ACC and motor cortex. Therefore, it is possible that the small fraction of speed-related neurons originated from the motor cortex.
Furthermore, we identify two distinct groups of ACC neurons: <iaction-state and action-content neurons, both of which tend to show sustained activity even when the animals remain immobile after completing shuttle behaviors. This prolonged activation in the absence of movement suggests that their activity is not directly driven by locomotion. Moreover, action-content neurons are selectively engaged in only one of the two shuttle categories, either rooms A→B or B→A shuttles. Therefore, differences in neuronal activity are unlikely to reflect locomotor differences, given that both shuttle types involve similar movement patterns.
(8) In addition to the above, the place-field analysis in Supplemental Figure 5 only shows 4 neurons. Was the whole population analyzed? Is it possible to decode place from the population during the ITI? The data in this figure sort of look exactly like place fields - many cortical neurons and also some hippocampal neurons have more than 1 place field
We have now provided additional place-field analysis. A comparison with hippocampal CA1 neurons (recorded during the same task) suggests that ACC neurons encode limited spatial information.
(9) "a simple Pavlovian association strategy is unlikely to be sufficient for learning the task" ... is Pavlovian occasion setting not a simple association? Tones and contexts both readily act as Pavlovian occasion setters. Similarly positive/negative patterning might also explain how the task is learned.
We appreciate this comment and have revised the sentence accordingly. It is possible that animals use multiple strategies to learn and perform the task effectively. In the early stages, animals may rely more heavily on sensory–spatial integration, whereas in later stages, sensory- or location-related Pavlovian associative strategies may contribute to performance, particularly when animals begin to show place preferences during inter-trial intervals.
(10) I might suggest softening this language and others like it. For example, 2x2 factorial designs are not really novel.
We have revised the language used to describe the task.
(11) Some of the color-scale bars and figures do not have labels. For example, Supplementary Figure 3, Supplementary Figure 5. Please add labels.
We have added the missing labels to all color bars.
Reviewer #3 (Recommendations for the authors):
(1) Some relevant papers that should be cited:
https://doi.org/10.1523/JNEUROSCI.4450-08.2008
10.1016/j.neuron.2018.11.016
We appreciate these suggestions.
(2) Where can we download the data and code?
We will upload the essential data and MATLAB code to GitHub to accompany the publication of the final version of this paper.
-
-
-
eLife Assessment
Huang and colleagues examined neural responses in mouse anterior cingulate cortex (ACC) during a discrimination-avoidance task. The authors present useful findings that ACC neurons encode primarily post-action variables over extended periods rather than the outcomes or values of those actions. Though the methodological approach was sound, the evidence ruling out alternative explanations is incomplete and requires substantial control analyses.
-
Reviewer #1 (Public review):
Summary:
In the current study, Huang et al. examined ACC response during a novel discrimination-avoid task. The authors concluded that ACC neurons primarily encode post-action variables over extended periods, reflecting the animal's preceding actions rather than the outcomes or values of those actions. Specifically, they identified two subgroups of ACC neurons that responded to different aspects of the actions. This work represents admirable efforts to investigate the role of ACC in task-performing mice. However, in my opinion, alternative explanations of the data were not sufficiently explored, and some key findings were not well supported.
Strengths:
The development of the new discrimination-avoid task is applauded. Single-unit electrophysiology in task-performing animals represents admirable efforts and …
Reviewer #1 (Public review):
Summary:
In the current study, Huang et al. examined ACC response during a novel discrimination-avoid task. The authors concluded that ACC neurons primarily encode post-action variables over extended periods, reflecting the animal's preceding actions rather than the outcomes or values of those actions. Specifically, they identified two subgroups of ACC neurons that responded to different aspects of the actions. This work represents admirable efforts to investigate the role of ACC in task-performing mice. However, in my opinion, alternative explanations of the data were not sufficiently explored, and some key findings were not well supported.
Strengths:
The development of the new discrimination-avoid task is applauded. Single-unit electrophysiology in task-performing animals represents admirable efforts and the datasets are valuable. The identification of different groups of encoding neurons in ACC can be potentially important.
Weaknesses:
One major conclusion is that ACC primarily encodes the so-called post-action variables (specifically shuttle crossing). However, only a single example session was included in Figure 2, while in Supplementary Figure 2 a considerable fraction of ACC neurons appears to respond to either the onset of movement or ramp up their activity prior to movement onset. How did the authors reach the conclusion that ACC preferentially respond to shuttle crossing?
In Figure 4, it was concluded that ACC neurons respond to action independent of outcome. Since these neurons are active on both correct and incorrect shuttle but not stay trials, they seem to primarily respond to overt movement. If so, the rationale for linking ACC activity and adaptive behavior/associative learning is not very clear to me. Further analyses are needed to test whether their firing rates correlated with locomotion speed or acceleration/deceleration. On a similar note, to what extent are the action state neurons actually responding to locomotion-related signals? And can ACC activity actually differentiate correct vs. incorrect stays?
Given that a considerable amount of ACC neurons encode 'action content', it is not surprising that by including all neurons the model is able to make accurate predictions in Figure 6. How would the model performance change by removing the content neurons?
Moving on to Figure 7. Since Figure 4 showed that ACC neurons respond to movement regardless of outcome, it is somewhat puzzling how ACC activity can be linked to future performance.
Two mice contributed about 50% of all the recorded cells. How robust are the results when analyzing mouse by mouse?
Lastly, the development of the new discrimination-avoid task is applauded. However, a major missing piece here is to show the importance of ACC in this task and what aspects of this behavior require ACC.
-
Reviewer #2 (Public review):
Summary:
The current dataset utilized a 2x2 factorial shuttle-escape task in combination with extracellular single-unit recording in the anterior cingulate cortex (ACC) of mice to determine ACC action coding. The contributions of neocortical signaling to action-outcome learning as assessed by behavioral tasks outside of the prototypical reward versus non-reward or punished vs non-punished is an important and relevant research topic, given that ACC plays a clear role in several human neurological and psychiatric conditions. The authors present useful findings regarding the role of ACC in action monitoring and learning. The core methods themselves - electrophysiology and behavior - are adequate; however, the analyses are incomplete since ruling out alternative explanations for neural activity, such as movement …
Reviewer #2 (Public review):
Summary:
The current dataset utilized a 2x2 factorial shuttle-escape task in combination with extracellular single-unit recording in the anterior cingulate cortex (ACC) of mice to determine ACC action coding. The contributions of neocortical signaling to action-outcome learning as assessed by behavioral tasks outside of the prototypical reward versus non-reward or punished vs non-punished is an important and relevant research topic, given that ACC plays a clear role in several human neurological and psychiatric conditions. The authors present useful findings regarding the role of ACC in action monitoring and learning. The core methods themselves - electrophysiology and behavior - are adequate; however, the analyses are incomplete since ruling out alternative explanations for neural activity, such as movement itself, requires substantial control analyses, and details on statistical methods are not clear.
Strengths:
(1) The factorial design nicely controls for sensory coding and value coding, since the same stimulus can signal different actions and values.
(2) The figures are mostly well-presented, labeled, and easy to read.
(3) Additional analyses, such as the 2.5/7.5s windows and place-field analysis, are nice to see and indicate that the authors were careful in their neural analyses.
(4) The n-trial + 1 analysis where ACC activity was higher on trials that preceded correct responses is a nice addition, since it shows that ACC activity predicts future behavior, well before it happens.
(5) The authors identified ACC neurons that fire to shuttle crossings in one direction or to crossings in both directions. This is very clear in the spike rasters and population-scaled color images. While other factors such as place fields, sensory input, and their integration can account for this activity, the authors discuss this and provide additional supplemental analyses.
Weaknesses:
(1) The behavioral data could use slightly more characterization, such as separating stay versus shuttle trials.
(2) Some of the neural analyses could use the necessary and sufficient comparisons to strengthen the authors' claims.
(3) Many of the neural analyses seem to utilize long time windows, not leveraging the very real strength of recording spike times. Specifics on the exact neural activity binning/averaging, tests, classifier validation, and methods for quantification are difficult to find.
(4) The neural analyses seem to suggest that ACC neurons encode one variable or the other, but are there any that multiplex? Given the overwhelming evidence of multiplexing in the ACC a bit more discussion of its presence or absence is warranted.
-
Reviewer #3 (Public review):
Summary:
The authors record from the ACC during a task in which animals must switch contexts to avoid shock as instructed by a cue. As expected, they find neurons that encode context, with some encoding of actions prior to the context, and encoding of neurons post-action. The primary novelty of the task seems to be dynamically encoding action-outcome in a discrimination-avoidance domain, while this is traditionally done using operant methods. While I'm not sure that this task is all that novel, I can't recall this being applied to the frontal cortex before, and this extends the well-known action/context/post-context encoding of ACC to the discrimination-avoidance domain.
While the analysis is well done, there are several points that I believe should be elaborated upon. First, I had questions about several …
Reviewer #3 (Public review):
Summary:
The authors record from the ACC during a task in which animals must switch contexts to avoid shock as instructed by a cue. As expected, they find neurons that encode context, with some encoding of actions prior to the context, and encoding of neurons post-action. The primary novelty of the task seems to be dynamically encoding action-outcome in a discrimination-avoidance domain, while this is traditionally done using operant methods. While I'm not sure that this task is all that novel, I can't recall this being applied to the frontal cortex before, and this extends the well-known action/context/post-context encoding of ACC to the discrimination-avoidance domain.
While the analysis is well done, there are several points that I believe should be elaborated upon. First, I had questions about several details (see point 3 below). Second, I wonder why the authors downplayed the clear action coding of ACC ensembles. Third, I wonder if the purported 'novelty' of the task (which I'm not sure of) and pseudo-debate on ACC's role undermines the real novelty - action/context/outcome encoding of ACC in discrimination-avoidance and early learning.
Strengths:
Recording frontal cortical ensembles during this task is particularly novel, and the analyses are sophisticated. The task has the potential to generate elegant comparisons of action and outcome, and the analyses are sophisticated.
Weaknesses:
I had some questions that might help me understand this work better.
(1) I wonder if the field would agree that there is a true 'debate' and 'controversy' about the ACC and conflict monitoring, or if this is a pseudodebate (Line 34). They cite 2 very old papers to support this point. I might reframe this in terms of the frontal cortex studying action-outcome associations in discrimination-avoidance, as the bulk of evidence in rodents comes from overtrained operant behavior, and in humans comes from high-level tasks, and humans are unlikely to get aversive stimuli such as shocks.
(2) Does the purported novelty of the task undermine the argument? While I don't have an exhaustive knowledge of this behavior, the novelty involves applying this ACC. There are many paradigms where a shock triggers some action that could be antecedents to this task.
(3) The lack of details was confusing to me:
a) How many total mice? Are the same mice in all analyses? Are the same neurons? Which training day? Is it 4 mice in Figure 3? Five mice in line 382? An accounting of mice should be in the methods. All data points and figures should have the number of neurons and mice clearly indicated, along with a table. Without these details, it is challenging to interpret the findings.
b) How many neurons are from which stage of training? In some figures, I see 325, in some ~350, and in S5/S2B, 370. The number of neurons should be clearly indicated in each figure, and perhaps a table.
c) Were the tetrodes driven deeper each day? The depth should be used as a regressor in all analyses?
d) Was is really ACC (Figure 2A)? Some shanks are in M2? All electrodes from all mice need to be plotted as a main figure with the drive length indicated.
e) It's not clear which sessions and how many go into which analysis
f) How many correct and incorrect trials (<7?) are there per session?
g) Why 'up to 10 shocks' on line 358? What amplitudes were tried? What does scrambled mean?
(4) Why do the authors downplay pre-action encoding? It is clearly evident in the PETHs, and the classifiers are above chance. It's not surprising that post-shuttle classification is so high because the behavior has occurred. This is most evident in Figure S2B, which likely should be a main figure.
(5) The statistics seem inappropriate. A linear mixed effects model accounting for between-mouse variance seems most appropriate. Statistical power or effect size is needed to interpret these results. This is important in analyses like Figure 7C or 6B.
(6) Better behavioral details might help readers understand the task. These can be pulled from Figures S2 and S5. This is particularly important in a 'novel' task.
(7) Can the authors put post-action encoding on the same classification accuracy axes as Figure 6B? It'd be useful to compare.
(8) What limitations are there? I can think of several - number of animals, lack of causal manipulations, ACC in rodents and humans.
Minor:
(1) Each PCA analysis needs a scree plot to understand the variance explained.
(2) Figure 4C - y and x-axes have the same label?
(3) What bin size do the authors use for machine learning (Not clear from line 416)?
(4) Why not just use PCA instead of 'dimension reduction' (of which there are many?)
(5) Would a video enhance understanding of the behavior?
-
Author response:
We thank the reviewers for their insightful feedback. Incorporating their recommendations will greatly enhance our manuscript for resubmission. Based on the review, it seems a major challenge to the interpretation of our study surrounds whether locomotion, itself, is responsible for increased ACC activity during our task. This was a shared concern for us during our analysis. We included data in our initial submission hoping to address these concerns. Specifically, we show that post-action activity outlasts movement termination, in most cases, on the order seconds after termination (Supplementary Fig 2). Likewise, post-action activity is not tied to shuttle initiations as ACC activity onset can vary greatly before and after initiation (Supplementary Fig 2). Lastly, the unique nature of action content neurons further …
Author response:
We thank the reviewers for their insightful feedback. Incorporating their recommendations will greatly enhance our manuscript for resubmission. Based on the review, it seems a major challenge to the interpretation of our study surrounds whether locomotion, itself, is responsible for increased ACC activity during our task. This was a shared concern for us during our analysis. We included data in our initial submission hoping to address these concerns. Specifically, we show that post-action activity outlasts movement termination, in most cases, on the order seconds after termination (Supplementary Fig 2). Likewise, post-action activity is not tied to shuttle initiations as ACC activity onset can vary greatly before and after initiation (Supplementary Fig 2). Lastly, the unique nature of action content neurons further supports a distinction from locomotor activity. They selectively fire for specific directions and, as a result, do not fire during movement in opposite directions. Despite these findings, we agree with reviews that inclusion of additional analyses, such as examining firing rates in respect to locomotion speed and acceleration/deceleration, will greatly strengthen our claim of ACC’s role in post-action activity. In our resubmission, we will seek to perform such an analysis, among others, to elucidate completely the role of locomotion in ACC post-action activity.
Reviewers also pointed out an overall lack of details surrounding our task, analysis, statistical methods and experimental approaches. We will consider all the recommendations from the reviewers and integrate them into our resubmission to provide more detailed information. Notably, we will adjust our approach in describing our task. Reviewers discussed some criticism regarding the perceived novelty of the task as it shares many similarities with previous discrimination-avoidance tasks. The distinction with our task is regarding the nuance of how the meaning (safety vs shock) of the context and sensory stimuli dynamically changes based on the current environment (context x sound). This requires not only the discrimination of contextual and sensory stimuli but also the inter-modal integration of stimuli, which varies throughout the task. Sound A/B leads to different outcomes depending on the context, and similarly, the meaning of the context shifts in a sound-dependent manner.
Lastly, in our follow-up submission we will work to include more robust analyses to utilize our temporal sensitivity of our recordings. We also will provide greater clarity on how each individual animal contributes to our overall findings. To conclude, we would like to once again thank our reviewers for their feedback and evaluation of our manuscript. We look forward to making the necessary adjustments for our future submission.
-
-
