Energy Landscape Analysis Reveals Thalamic Modulation of Brain State Transitions During Movie Watching

Reviewer #1 (Public review):

The authors analysed large-scale brain-state dynamics while humans watched a short video. They sought to identify the role of thalamocortical interactions.

Major concerns

(1) Rationale for using the naturalistic stimulus

In terms of brain state dynamics, previous studies have already reported large-scale neural dynamics by applying some data-driven analyses, like energy landscape analysis and Hidden Markov Model, to human fMRI/EEG data recorded during resting/task states. Considering such prior work, it'd be critical to provide sufficient biological rationales to perform a conceptually similar study in a naturalistic condition, i.e., not just "because no previous work has been done". The authors would have to clarify what type of neural mechanisms could be missed in conventional resting-state studies using, say, energy landscape analysis, but could be revealed in the naturalistic condition.

(2) Effects of the uniqueness of the visual stimulus and reproducibility

One of the main drawbacks of the naturalistic condition is the unexpected effects of the stimuli. That is, this study looked into the data recorded from participants who were watching Sherlock, but what would happen to the results if we analyzed the brain activity data obtained from individuals who were watching different movies? To ensure the generalizability of the current findings, it would be necessary to demonstrate qualitative reproducibility of the current observations by analysing different datasets that employed different movie stimuli. In fact, it'd be possible to find such open datasets, like www.nature.com/articles/s41597-023-02458-8.

(3) Spatial accuracy of the "Thalamic circuit" definition

One of the main claims of this study heavily relies on the accuracy of the localization of two different thalamic architectures: matrix and core. Given the conventional or relatively low spatial resolution of the fMRI data acquisition (3x3x3 mm^3), it appears to be critically essential to demonstrate that the current analysis accurately distinguished fMRI signals between the matrix and core parts of the thalamus for each individual.

(4) More detailed analysis of the thalamic circuits

In addition, if such thalamic localisation is accurate enough, it would be greatly appreciated if the authors perform similar comparisons not only between the matrix and core architectures but also between different nuclei. For example, anterior, medial, and lateral groups (e.g., pulvinar group). Such an investigation would meet the expectations of readers who presume some microscopic circuit-level findings.

(5) Rationale for different time window lengths

The authors adopted two different time window lengths to examine the neural dynamics. First, they used a 21-TR window for signal normalisation. Then, they narrowed down the window length to 13-TR periods for the following statistical evaluation. Such a seemingly arbitrary choice of the shorter time window might be misunderstood as a measure to relax the threshold for the correction of multiple comparisons. Therefore, it'd be appreciated if the authors stuck to the original 21-TR time window and performed statistical evaluations based on the setting.

(6) Temporal resolution

After identifying brain states with energy landscape analysis, this study investigated the brain state transitions by directly looking into the fMRI signal changes. This manner seems to implicitly assume that no significant state changes happen in one TR (=1.5sec), which needs sufficient validation. Otherwise, like previous studies, it'd be highly recommended to conduct different analyses (e.g., random-walk simulation) to address and circumvent this problem.

Reviewer #2 (Public review):

Summary:

In this study, Liu et al. investigated cortical network dynamics during movie watching using an energy landscape analysis based on a maximum entropy model. They identified perception- and attention-oriented states as the dominant cortical states during movie watching and found that transitions between these states were associated with inter-subject synchronization of regional brain activity. They also showed that distinct thalamic compartments modulated distinct state transitions. They concluded that cortico-thalamo-cortical circuits are key regulators of cortical network dynamics.

Strengths:

A mechanistic understanding of cortical network dynamics is an important topic in both experimental and computational neuroscience, and this study represents a step forward in this direction by identifying key cortico-thalamo-cortical circuits. The analytical strategy employed in this study, particularly the LASSO-based analysis, is interesting and would be applicable to other data types, such as task- and resting-state fMRI.

Weaknesses:

Due to issues related to data preprocessing, support for the conclusions remains incomplete. I also believe that a more careful interpretation of the "energy" derived from the maximum entropy model would greatly clarify what the analysis actually revealed.

(1) Major Comment 1:

I think the method used for binarization of BOLD activity is problematic in multiple ways.

a) Although the authors appear to avoid using global signal regression (page 4, lines 114-118), the proposed method effectively removes the global signal. According to the description on page 4, lines 117-122, the authors binarized network-wise ROI signals by comparing them with the cross-network BOLD signal (i.e., the global signal): at each time point, network-wise ROI signals above the cross-network signal were set to 1, and the rest were set to −1. If I understand the binarization procedure correctly, this approach forces the cross-network signal to be zero (up to some noise introduced by the binarization of network-wise signals), which is essentially equivalent to removing the global signal. Please clarify what the authors meant by stating that "this approach maintained a diverse range of binarized cortical states in data where the global signal was preserved" (page 4, lines 121-122).

b) The authors might argue that they maintained a diverse range of cortical states by performing the binarization at each time point (rather than within each network). However, I believe this introduces another problem, because binarizing network-wise signals at each time point distorts the distribution of cortical states. For example, because the cross-network signal is effectively set to zero, the network cannot take certain states, such as all +1 or all −1. Similarly, this binarization biases the system toward states with similar numbers of +1s and −1s, rather than toward unbalanced states such as (+1, −1, −1, −1, −1, −1). These constraints and biases are not biological in origin but are simply artifacts of the binarization procedure. Importantly, the energy landscape and its derivatives (e.g., hard/easy transitions) are likely to be affected by these artifacts. I suggest that the authors try a more conventional binarization procedure (i.e., binarization within each network), which is more robust to such artifacts.

Related to this point, I have a question regarding Figure S1, in which the authors plotted predicted versus empirical state probabilities. As argued above, some empirical state probabilities should be zero because of the binarization procedure. However, in Figure S1, I do not see data points corresponding to these states (i.e., there should be points on the y-axis). Did the authors plot only a subset of states in Figure S1? I believe that all states should be included. The correlation coefficient between empirical and predicted probabilities (and the accuracy) should also be calculated using all states.

c) The current binarization procedure likely inflates non-neuronal noise and obscures the relationship between the true BOLD signal and its binarized representation. For example, consider two ROIs (A and B): both (+2%, +1%) and (+0.01%, −0.01%) in BOLD signal changes would be mapped to (+1, −1) after binarization. This suggests that qualitatively different signal magnitudes are treated identically. I believe that this issue could be alleviated if the authors were to binarize the signal within each network, rather than at each time point.

(2) Major Comment 2:

As the authors state (page 5, lines 145-148), the "energy" described in the energy landscape is not biological energy but rather a statistical transformation of probability distributions derived from the Boltzmann distribution. If this is the case, I believe that Figure 2A is potentially misleading and should be removed. This type of schematic may give the false impression that cortical state dynamics are governed by the energy landscape derived from the maximum entropy model (which is not validated).

Reviewer #3 (Public review):

Summary:

In this study, Liu et al. analyze fMRI data collected during movie watching, applied an energy landscape method with pairwise maximum entropy models. They identify a set of brain states defined at the level of canonical functional networks and quantify how the brain transitions between these states. Transitions are classified as "easy" or "hard" based on changes in the inferred energy landscape, and the authors relate transition probabilities to inter-subject correlation. A major emphasis of the work is the role of the thalamus, which shows transition-linked activity changes and dynamic connectivity patterns, including differential involvement of parvalbumin- and calbindin-associated thalamic subdivisions.

Strengths:

The study is methodologically complex and technically sophisticated. It integrates advanced analytical methods into high-dimensional fMRI data. The application of energy landscape analysis to movie-watching data appears to be novel as well. The finding on the thalamus involved energy state transition and provides a strong linkage to several theories on thalamic control functions, which is a notable strength.

Weaknesses:

The main weakness is the conceptual clarity and advances that this otherwise sophisticated set of analyses affords. A central conceptual ambiguity concerns the energy landscape framework itself. The authors note that the "energy" in this model is not biological energy but a statistical quantity derived from the Boltzmann distribution. After multiple reads, I still have major trouble mapping this measure onto any biological and cognitive operations. BOLD signal is a measure of oxygenation as a proxy of neural activity, and correlated BOLD (functional connectivity) is thought to measure the architecture of information communication of brain systems. The energy framework described in the current format is very difficult for most readers to map onto any neural or cognitive knowledge base on the structure and function of brain systems. Readers unfamiliar with maximum entropy models may easily misinterpret energy changes as reflecting metabolic cost, neural effort, or physiological variables, and it is just very unclear what that measure is supposed to reflect. The manuscript does not clearly articulate what conceptual and mechanistic advances the energy formalism provides beyond a mathematical and statistical report. In other words, beyond mathematical description, it is very hard for most readers to understand the process and function of what this framework is supposed to tell us in regards to functional connectivity, brain systems, and cognition. The brain is not a mathematical object; it is a biological organ with cognitive functions. The impact of this paper is severely limited until connections can be made.

Relatedly, the use of metaphors such as "valleys," "hills," and "routes" in multidimensional measures lacks grounding. Valleys and hills of what is not intuitive to understand. Based on my reading, these features correspond to local minima and barriers in a probability distribution over binarized network activation patterns, but similar to the first point, the manuscript does not clearly explain what it means conceptually, neurobiologically, or computationally for the brain to "move" through such a landscape. The brain is not computing these probabilities; they are measurement tools of "something". What is it? To advance beyond mathematical description, these measurements must be mapped onto neurobiological and cognitive information.

This conceptual ambiguity goes back to the Introduction. At the level of motivation, the purpose and deliverables of the study are not defined in the Introduction. The stated goal is "Transitions between distinct cortical brain states modulate the degree of shared neural processing under naturalistic conditions". I do not know if readers will have a clear answer to this question at the end. Is the claim that state transitions cause changes in inter-subject correlation, that they index moments of narrative alignment, or that they reflect changes in attentional or cognitive mode? This level of explanation is largely dissociated from the methods in their current form.

Several methodological choices can use clarification. The use of a 21-TR window centered on transition offsets is unusually long relative to the temporal scale of fMRI dynamics and to the hypothesized rapidity of state transitions. On a related note, what is the temporal scale of state transition? Is it faster than 21 TRs?

The choice of movie-watching data is a strength. But, many of the analyses performed here, energy landscape estimation, clustering of states, could in principle be applied to resting-state data. The manuscript does not clearly articulate what is gained, mechanistically or cognitively, by using movie stimuli beyond the availability of inter-subject correlation.

Because of the above issues, a broader concern throughout the results is the largely descriptive nature of the findings. For example, the LASSO analysis shows that certain state transitions predict ISC in a subset of regions, with respectable R² values. While statistically robust, the manuscript provides little beyond why these particular transitions should matter, what computations they might reflect, or how they relate to known cognitive operations during movie watching. Similar issues arise in the clustering analyses. Clustering high-dimensional fMRI-derived features will almost inevitably produce structure, whether during rest, task, or naturalistic viewing. What is missing is an explanation of why these specific clusters are meaningful in functional or mechanistic terms.

Finally, the treatment of the thalamus, while very exciting, could use a bit more anatomical and circuit-level specificity. The manuscript largely treats the thalamus as a unitary structure, despite decades of work demonstrating big functional and connectivity differences across thalamic nuclei. A whole-thalamus analysis without more detailed resolution is increasingly difficult to justify. The subsequent subdivision into PVALB- and CALB-associated regions partially addresses this, but these markers span multiple nuclei with overlapping projection patterns.

Author response:

Reviewer #1 (Public review):

The authors analysed large-scale brain-state dynamics while humans watched a short video. They sought to identify the role of thalamocortical interactions.

Major concerns

(1) Rationale for using the naturalistic stimulus

In terms of brain state dynamics, previous studies have already reported large-scale neural dynamics by applying some data-driven analyses, like energy landscape analysis and Hidden Markov Model, to human fMRI/EEG data recorded during resting/task states. Considering such prior work, it'd be critical to provide sufficient biological rationales to perform a conceptually similar study in a naturalistic condition, i.e., not just "because no previous work has been done". The authors would have to clarify what type of neural mechanisms could be missed in conventional resting-state studies using, say, energy landscape analysis, but could be revealed in the naturalistic condition.

We appreciate your insightful comments regarding the need for a biological rationale in our study. As you mentioned, there are similar studies, just like Meer et al. utilized Hidden Markov Models to identify various activation modes of brain networks that included subcortical regions[1], Song et al. linked brain states to narrative understandings and attentional dynamics[2, 3]. These studies could answer why we use naturalistic stimuli datasets. Moreover, there is evidence suggesting that the thalamus plays a crucial role in processing information in a more naturalistic context while pointing out the vital role in thalamocortical communications[4, 5]. So, we tended to bridge thalamic activity and cortical state transition using the energy landscape description.

To address these gaps in conventional resting-state studies, we explored an alternative method—maximum entropy modeling based on the energy landscape. This allowed us to validate how the thalamus responds to cortical state transitions. To enhance clarity, we will update our introduction to emphasize the motivations behind our research and the significance of examining these neural mechanisms in a naturalistic setting.

(2) Effects of the uniqueness of the visual stimulus and reproducibility

One of the main drawbacks of the naturalistic condition is the unexpected effects of the stimuli. That is, this study looked into the data recorded from participants who were watching Sherlock, but what would happen to the results if we analyzed the brain activity data obtained from individuals who were watching different movies? To ensure the generalizability of the current findings, it would be necessary to demonstrate qualitative reproducibility of the current observations by analysing different datasets that employed different movie stimuli. In fact, it'd be possible to find such open datasets, like www.nature.com/articles/s41597-023-02458-8.

We appreciate your concern regarding the reproducibility of our findings. The dataset from the "Sherlock" study is of high quality and has shown good generalizability in various research contexts. We acknowledge the importance of validating our results with different datasets to enhance the robustness of our conclusions. While we are open to exploring additional datasets, we intend to pursue this validation once we identify a suitable alternative. Currently, we are considering a comparison with the dataset from "Forrest Gump" as part of our initial plan.

(3) Spatial accuracy of the "Thalamic circuit" definition

One of the main claims of this study heavily relies on the accuracy of the localization of two different thalamic architectures: matrix and core. Given the conventional or relatively low spatial resolution of the fMRI data acquisition (3x3x3 mm^3), it appears to be critically essential to demonstrate that the current analysis accurately distinguished fMRI signals between the matrix and core parts of the thalamus for each individual.

We acknowledge the importance of accurately localizing the different thalamic architectures, specifically the matrix and core regions. To address this, we downsampled the atlas of matrix and core cell populations from the previous study from a resolution of 2x2x2 mm³ to 3x3x3 mm³, which aligns with our fMRI data acquisition. We would report the atlas as Supplementary Figures in our revision.

(4) More detailed analysis of the thalamic circuits

In addition, if such thalamic localisation is accurate enough, it would be greatly appreciated if the authors perform similar comparisons not only between the matrix and core architectures but also between different nuclei. For example, anterior, medial, and lateral groups (e.g., pulvinar group). Such an investigation would meet the expectations of readers who presume some microscopic circuit-level findings.

We appreciate your suggestion regarding a more detailed analysis of thalamic circuits. We have touched upon this in the discussion section as a forward-looking consideration. However, we believe that performing nuclei segmentation with 3T fMRI may not be ideal due to well-documented concerns regarding signal-to-noise ratio and spatial resolution. That said, we are interested in exploring these nuclei-pathway connections to cortical areas in future studies with a proper 7T fMRI naturalistic dataset.

(5) Rationale for different time window lengths

The authors adopted two different time window lengths to examine the neural dynamics. First, they used a 21-TR window for signal normalisation. Then, they narrowed down the window length to 13-TR periods for the following statistical evaluation. Such a seemingly arbitrary choice of the shorter time window might be misunderstood as a measure to relax the threshold for the correction of multiple comparisons. Therefore, it'd be appreciated if the authors stuck to the original 21-TR time window and performed statistical evaluations based on the setting.

Thank you for your valuable feedback regarding the choice of time window lengths. We aimed to maintain consistency in window lengths across our analyses. In light of your comments and suggestions from other reviewers, we plan to test our results using different time window lengths and report findings that generalize across these variations. Should the results differ significantly, we will discuss the implications of this variability in our revised manuscript.

(6) Temporal resolution

After identifying brain states with energy landscape analysis, this study investigated the brain state transitions by directly looking into the fMRI signal changes. This manner seems to implicitly assume that no significant state changes happen in one TR (=1.5sec), which needs sufficient validation. Otherwise, like previous studies, it'd be highly recommended to conduct different analyses (e.g., random-walk simulation) to address and circumvent this problem.

Thank you for raising this important point regarding temporal resolution. Many fMRI studies, such as those examining event boundaries during movie watching, operate under similar assumptions concerning state changes within one TR. For example, Barnett et al. processed the dynamic functional connectivity (dFC) with a window of 20 TRs (24.4s). So, we do not think it is a limitation but is a common question related to fMRI scanning parameters. To strengthen our analysis of state transitions and ensure they are not merely coincidental, we plan to conduct random-walk simulations, as suggested, to validate our findings in accordance with methodologies used in previous research.

Reviewer #2 (Public review):

Summary:

In this study, Liu et al. investigated cortical network dynamics during movie watching using an energy landscape analysis based on a maximum entropy model. They identified perception- and attention-oriented states as the dominant cortical states during movie watching and found that transitions between these states were associated with inter-subject synchronization of regional brain activity. They also showed that distinct thalamic compartments modulated distinct state transitions. They concluded that cortico-thalamo-cortical circuits are key regulators of cortical network dynamics.

Strengths:

A mechanistic understanding of cortical network dynamics is an important topic in both experimental and computational neuroscience, and this study represents a step forward in this direction by identifying key cortico-thalamo-cortical circuits. The analytical strategy employed in this study, particularly the LASSO-based analysis, is interesting and would be applicable to other data types, such as task- and resting-state fMRI.

We thanks for this comment and encouragement.

Weaknesses:

Due to issues related to data preprocessing, support for the conclusions remains incomplete. I also believe that a more careful interpretation of the "energy" derived from the maximum entropy model would greatly clarify what the analysis actually revealed.

Thank you for your valuable suggestions, and we apologize for any misunderstandings regarding the interpretation of the energy landscape in our study. To address this issue, we will include a dedicated paragraph in both the methods and results sections to clarify our use of the term "energy" derived from the maximum entropy model. This addition aims to eliminate any ambiguity and provide a clearer understanding of what our analysis reveals.

(1) I think the method used for binarization of BOLD activity is problematic in multiple ways.

a) Although the authors appear to avoid using global signal regression (page 4, lines 114-118), the proposed method effectively removes the global signal. According to the description on page 4, lines 117-122, the authors binarized network-wise ROI signals by comparing them with the cross-network BOLD signal (i.e., the global signal): at each time point, network-wise ROI signals above the cross-network signal were set to 1, and the rest were set to −1. If I understand the binarization procedure correctly, this approach forces the cross-network signal to be zero (up to some noise introduced by the binarization of network-wise signals), which is essentially equivalent to removing the global signal. Please clarify what the authors meant by stating that "this approach maintained a diverse range of binarized cortical states in data where the global signal was preserved" (page 4, lines 121-122).

Thank you for highlighting the potential issue with our binarization method. We appreciate your insights regarding the comparison of network-wise ROI signals with the cross-network BOLD signal, as this may inadvertently remove the global signal. To address this, we will conduct a comparative analysis of results obtained from both our current approach and the original pipeline. If we decide to retain our current method, we will carefully reconsider the rationale and rephrase our descriptions to ensure clarity regarding the preservation of the global signal and the diversity of binarized cortical states.

b) The authors might argue that they maintained a diverse range of cortical states by performing the binarization at each time point (rather than within each network). However, I believe this introduces another problem, because binarizing network-wise signals at each time point distorts the distribution of cortical states. For example, because the cross-network signal is effectively set to zero, the network cannot take certain states, such as all +1 or all −1. Similarly, this binarization biases the system toward states with similar numbers of +1s and −1s, rather than toward unbalanced states such as (+1, −1, −1, −1, −1, −1). These constraints and biases are not biological in origin but are simply artifacts of the binarization procedure. Importantly, the energy landscape and its derivatives (e.g., hard/easy transitions) are likely to be affected by these artifacts. I suggest that the authors try a more conventional binarization procedure (i.e., binarization within each network), which is more robust to such artifacts.

Related to this point, I have a question regarding Figure S1, in which the authors plotted predicted versus empirical state probabilities. As argued above, some empirical state probabilities should be zero because of the binarization procedure. However, in Figure S1, I do not see data points corresponding to these states (i.e., there should be points on the y-axis). Did the authors plot only a subset of states in Figure S1? I believe that all states should be included. The correlation coefficient between empirical and predicted probabilities (and the accuracy) should also be calculated using all states.

Thank you for your thoughtful examination of our data processing pipeline. We agree that a comparison between the conventional binarization method and our current approach is warranted, and we appreciate your suggestion. Upon reviewing Figure S1, we discovered that there was indeed an error related to the plotting style set to "log10." As you correctly pointed out, the data should reflect that the probabilities for states where all networks are either activated or deactivated are zero. We are very interested in exploring the state distributions obtained from both the original and current approaches, as your comments highlight important considerations. We sincerely appreciate your insightful feedback and will make sure to address these points thoroughly in our first revision.

c) The current binarization procedure likely inflates non-neuronal noise and obscures the relationship between the true BOLD signal and its binarized representation. For example, consider two ROIs (A and B): both (+2%, +1%) and (+0.01%, −0.01%) in BOLD signal changes would be mapped to (+1, −1) after binarization. This suggests that qualitatively different signal magnitudes are treated identically. I believe that this issue could be alleviated if the authors were to binarize the signal within each network, rather than at each time point.

Thank you for your important observation regarding the potential inflation of non-neuronal noise in our current binarization procedure. We recognize that this process could lead to qualitatively different signal magnitudes being treated similarly after binarization, as you illustrated with your example. While we acknowledge your point, we believe that conventional binarization pipelines may also encounter this issue, albeit by comparing signals to a network's temporal mean activity. To address this concern and maintain consistency with previous studies, we will discuss this limitation in our revised manuscript. Additionally, if deemed necessary, we will explore implementing a percentile-based threshold above the baseline to further refine our binarization approach. Your suggestion provides a valuable perspective, and we appreciate your insights.

(2) As the authors state (page 5, lines 145-148), the "energy" described in the energy landscape is not biological energy but rather a statistical transformation of probability distributions derived from the Boltzmann distribution. If this is the case, I believe that Figure 2A is potentially misleading and should be removed. This type of schematic may give the false impression that cortical state dynamics are governed by the energy landscape derived from the maximum entropy model (which is not validated).

Thank you for your valuable feedback regarding Figure 2A. We apologize for any confusion it may have created. While we recognize that similar figures are commonly used in literature involving energy landscapes (maximum entropy model), we agree that Figure 2A may mislead readers into thinking that cortical state dynamics are directly governed by the energy landscape derived from the maximum entropy model, which has not been validated. In light of your comments, we will remove Figure 2A and instead emphasize the analytical strategy presented in Figure 2B. Additionally, we will provide a simplified line graph as an illustrative example to clarify the concepts without the potential for misinterpretation.

Reviewer #3 (Public review):

Summary:

In this study, Liu et al. analyze fMRI data collected during movie watching, applied an energy landscape method with pairwise maximum entropy models. They identify a set of brain states defined at the level of canonical functional networks and quantify how the brain transitions between these states. Transitions are classified as "easy" or "hard" based on changes in the inferred energy landscape, and the authors relate transition probabilities to inter-subject correlation. A major emphasis of the work is the role of the thalamus, which shows transition-linked activity changes and dynamic connectivity patterns, including differential involvement of parvalbumin- and calbindin-associated thalamic subdivisions.

Strengths:

The study is methodologically complex and technically sophisticated. It integrates advanced analytical methods into high-dimensional fMRI data. The application of energy landscape analysis to movie-watching data appears to be novel as well. The finding on the thalamus involved energy state transition and provides a strong linkage to several theories on thalamic control functions, which is a notable strength.

Thanks for your comments on the novelty of our study.

Weaknesses:

The main weakness is the conceptual clarity and advances that this otherwise sophisticated set of analyses affords. A central conceptual ambiguity concerns the energy landscape framework itself. The authors note that the "energy" in this model is not biological energy but a statistical quantity derived from the Boltzmann distribution. After multiple reads, I still have major trouble mapping this measure onto any biological and cognitive operations. BOLD signal is a measure of oxygenation as a proxy of neural activity, and correlated BOLD (functional connectivity) is thought to measure the architecture of information communication of brain systems. The energy framework described in the current format is very difficult for most readers to map onto any neural or cognitive knowledge base on the structure and function of brain systems. Readers unfamiliar with maximum entropy models may easily misinterpret energy changes as reflecting metabolic cost, neural effort, or physiological variables, and it is just very unclear what that measure is supposed to reflect. The manuscript does not clearly articulate what conceptual and mechanistic advances the energy formalism provides beyond a mathematical and statistical report. In other words, beyond mathematical description, it is very hard for most readers to understand the process and function of what this framework is supposed to tell us in regards to functional connectivity, brain systems, and cognition. The brain is not a mathematical object; it is a biological organ with cognitive functions. The impact of this paper is severely limited until connections can be made.

Thank you for your insightful and constructive comments regarding the conceptual clarity of our energy landscape framework. We appreciate your perspective on the challenges of mapping the statistical measure of "energy" derived from the Boltzmann distribution onto biological and cognitive operations. To address these concerns, we will revise our manuscript to clarify our expressions surrounding "energy" and emphasize its probabilistic nature. Additionally, we will incorporate a series of analyses that explicitly relate the features of the energy landscape to cognitive processes and key parameters, such as brain integration and functional connectivity. We believe these changes will help bridge the gap between our mathematical framework and its relevance to understanding brain systems and cognitive functions.

Relatedly, the use of metaphors such as "valleys," "hills," and "routes" in multidimensional measures lacks grounding. Valleys and hills of what is not intuitive to understand. Based on my reading, these features correspond to local minima and barriers in a probability distribution over binarized network activation patterns, but similar to the first point, the manuscript does not clearly explain what it means conceptually, neurobiologically, or computationally for the brain to "move" through such a landscape. The brain is not computing these probabilities; they are measurement tools of "something". What is it? To advance beyond mathematical description, these measurements must be mapped onto neurobiological and cognitive information.

Thank you for your valuable feedback. In our revisions, we would aim to link the concept of rapid transition routes in the energy landscape to cognitive processes, such as narrative understanding and related features. By exploring these connections, we hope to provide a clearer context for how our framework can enhance understanding of cognitive functions and their neural correlates.

This conceptual ambiguity goes back to the Introduction. At the level of motivation, the purpose and deliverables of the study are not defined in the Introduction. The stated goal is "Transitions between distinct cortical brain states modulate the degree of shared neural processing under naturalistic conditions". I do not know if readers will have a clear answer to this question at the end. Is the claim that state transitions cause changes in inter-subject correlation, that they index moments of narrative alignment, or that they reflect changes in attentional or cognitive mode? This level of explanation is largely dissociated from the methods in their current form.

Thank you for highlighting this important point regarding the conceptual clarity in our Introduction. We appreciate your feedback about the motivation and objectives of the study. To clarify the stated goal of investigating how transitions between distinct cortical brain states modulate shared neural processing under naturalistic conditions, we will revise the manuscript to explicitly define the specific claims we aim to address. We will ensure that these explanations are closely tied to the methods employed in our study, providing a clearer framework for our readers.

Several methodological choices can use clarification. The use of a 21-TR window centered on transition offsets is unusually long relative to the temporal scale of fMRI dynamics and to the hypothesized rapidity of state transitions. On a related note, what is the temporal scale of state transition? Is it faster than 21 TRs?

Thank you for your insightful questions regarding our methodological choices. Our focus on specific state transitions necessitated the use of a 21-TR window. While it’s true that other transitions may occur within this window, averaging across the same transitions at different times allows us to identify distinctive thalamic BOLD patterns that precede cortical state transitions. This methodology enables us to capture relevant dynamics while ensuring that we focus on the transitions of interest. We appreciate your feedback, and this clarification will be included in our revised manuscript. We would also add a figure that describe the dwell time of cortical states.

The choice of movie-watching data is a strength. But, many of the analyses performed here, energy landscape estimation, clustering of states, could in principle be applied to resting-state data. The manuscript does not clearly articulate what is gained, mechanistically or cognitively, by using movie stimuli beyond the availability of inter-subject correlation.

Thank you for your question, which closely aligns with a concern raised by Reviewer #1. Our core hypothesis posits that naturalistic stimuli yield a broader set of brain states compared to those observed during resting-state conditions. To support this assertion, we will clearly articulate the findings from previous studies that relate to this hypothesis. Additionally, if appropriate, we will provide a comparative analysis between our data and resting-state data to highlight the differences and emphasize the uniqueness of the brain states elicited by naturalistic stimuli.

Because of the above issues, a broader concern throughout the results is the largely descriptive nature of the findings. For example, the LASSO analysis shows that certain state transitions predict ISC in a subset of regions, with respectable R² values. While statistically robust, the manuscript provides little beyond why these particular transitions should matter, what computations they might reflect, or how they relate to known cognitive operations during movie watching. Similar issues arise in the clustering analyses. Clustering high-dimensional fMRI-derived features will almost inevitably produce structure, whether during rest, task, or naturalistic viewing. What is missing is an explanation of why these specific clusters are meaningful in functional or mechanistic terms.

Thank you for your questions. In our revisions, we will perform additional analyses aimed at linking state transitions to cognitive processes more explicitly. Regarding clustering, we will provide a thorough discussion in the revised manuscript.

Finally, the treatment of the thalamus, while very exciting, could use a bit more anatomical and circuit-level specificity. The manuscript largely treats the thalamus as a unitary structure, despite decades of work demonstrating big functional and connectivity differences across thalamic nuclei. A whole-thalamus analysis without more detailed resolution is increasingly difficult to justify. The subsequent subdivision into PVALB- and CALB-associated regions partially addresses this, but these markers span multiple nuclei with overlapping projection patterns.

This suggestion aligns with the feedback from Reviewer #1. We believe that performing nuclei segmentation with 3T fMRI may not be ideal due to well-documented concerns regarding signal-to-noise ratio and spatial resolution. Therefore, investigating core and matrix cell projections across different thalamic nuclei using 7T fMRI presents a promising avenue for further study.

(1) Van Der Meer J N, Breakspear M, Chang L J, et al. Movie viewing elicits rich and reliable brain state dynamics [J]. Nature Communications, 2020, 11(1): 5004.

(2) Song H, Park B Y, Park H, et al. Cognitive and Neural State Dynamics of Narrative Comprehension [J]. Journal of Neuroscience, 2021, 41(43): 8972-8990.

(3) Song H, Shim W M, Rosenberg M D. Large-scale neural dynamics in a shared low-dimensional state space reflect cognitive and attentional dynamics [J]. Elife, 2023, 12.

(4) Shine J M, Lewis L D, Garrett D D, et al. The impact of the human thalamus on brain-wide information processing [J]. Nature Reviews Neuroscience, 2023, 24(7): 416-430.

(5) Yang M Y, Keller D, Dobolyi A, et al. The lateral thalamus: a bridge between multisensory processing and naturalistic behaviors [J]. Trends in Neurosciences, 2025, 48(1): 33-46.