Functional brain reconfiguration during sustained pain

  1. Jae-Joong Lee
  2. Sungwoo Lee
  3. Dong Hee Lee
  4. Choong-Wan Woo

  • Curated by eLife

    eLife logo

    Evaluation Summary:

    This paper will be of great interest to researchers interested in the brain mechanisms of pain. It shows how the connectivity of brain networks associated with sustained pain change over time. These findings are conclusively supported by state-of-the-art fMRI analyses of a tonic pain paradigm in two cohorts of healthy human participants. These insights are important for the understanding of the brain mechanisms of sustained pain which is the hallmark of chronic pain as a major health care problem.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 and Reviewer #3 agreed to share their name with the authors.)

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

Pain is constructed through complex interactions among multiple brain systems, but it remains unclear how functional brain networks are reconfigured over time while experiencing pain. Here, we investigated the time-varying changes in the functional brain networks during 20 min capsaicin-induced sustained orofacial pain. In the early stage, the orofacial areas of the primary somatomotor cortex were separated from other areas of the somatosensory cortex and integrated with subcortical and frontoparietal regions, constituting an extended brain network of sustained pain. As pain decreased over time, the subcortical and frontoparietal regions were separated from this brain network and connected to multiple cerebellar regions. Machine-learning models based on these network features showed significant predictions of changes in pain experience across two independent datasets ( n = 48 and 74). This study provides new insights into how multiple brain systems dynamically interact to construct and modulate pain experience, advancing our mechanistic understanding of sustained pain.

Article activity feed

  1. Version published to 10.7554/elife.74463 on eLife
  2. eLife

    Author Response

    Reviewer #1 (Public Review):

    “This study investigates the dynamics of brain network connectivity during sustained experimental pain in healthy human participants. To this end, capsaicin was applied to the tongues of two cohorts of participants (discovery cohort, N=48; replication cohort, N=74). This procedure resulted in pain for several minutes. During sustained pain, pain avoidance/intensity ratings and fMRI scans were obtained. The analyses (i) compare the pain state with a resting state, (ii) assess the dynamics of brain networks during sustained pain, and (iii) aim to predict pain based on the dynamics of brain networks. To this end, the analyses focus on community structures of time-evolving networks. The results show that sustained pain is associated with the emergence of a brain network including somatomotor, frontoparietal, basal ganglia and thalamic brain areas. The somatomotor area of the tongue is particularly involved in that network while this area is decoupled from other parts of the somatomotor cortex. Moreover, the network configuration changes over time with the frontoparietal network decoupling from the somatomotor network. Frontoparietal-cerebellar connections were predictive of decreases of pain. Together, the findings provide novel and convincing insights into the dynamics of brain network during sustained pain.

    Strengths

    • The brain mechanisms of sustained pain is a timely and relevant topic with potential clinical implications.

    • Assessing the dynamics of sustained pain and relating it to the dynamics of brain networks is a timely and promising approach to further the understanding of the brain mechanisms of pain.

    • The study includes discovery and replication cohorts and pursues a cutting-edge analysis strategy.

    • The manuscript is very well-written and the results are visualized in an exemplary manner including a graphical outline and summary of the findings.”

    We thank the reviewer for the thoughtful summarization and evaluation of our study.

    “Weaknesses

    • It remains unclear whether the changes of brain networks over time simply reflect the duration of sustained pain or whether they essentially reflect different levels of pain intensity/avoidance.”

    We appreciate the editor and reviewer’s comment on this issue. With the current experimental paradigm, it is difficult to dissociate the pain duration from the level of pain because the delivery of oral capsaicin commonly induces initial bursting and then a gradual decrease of pain over time. That is, the pain duration is correlated with the pain intensity in our task.

    However, when we examined the time-course of the ratings at each individual level (as shown in Figure S2), the time duration explained 53.7% of the rating variance, R2 = 0.537 ± 0.315 (mean ± standard deviation). In addition, if we constrain the beta coefficient of the time duration to be negative (i.e., ratings should decrease over time), the explained variance decreases to 48.2%, R2 = 0.482 ± 0.457, leaving us enough variance (i.e., greater than 50%) for examining the distinct effects of time duration and ratings on the patterns of functional brain reorganization.

    Indeed, the two main analyses included in the manuscript—consensus community detection and predictive modeling—were designed to examine those two aspects of the task, i.e., time duration and pain avoidance ratings, respectively. First, through the consensus community detection analysis, we examined the community structure that changes over time, i.e., across the early, middle, and late periods (as shown in Figure 3). We then developed predictive models of pain avoidance ratings in the second main analysis (as shown in Figure 5).

    Though it is still a caveat that we cannot fully dissociate the effects of time duration versus pain ratings, we could interpret the first set of results to be more about time duration, while the second set of results is more about pain ratings.

    We now added a description of the implication of predictive modeling for isolating the effects of pain ratings. In addition, a discussion on the caveat of the current experimental design and relevant future direction.

    Revisions to the main manuscript:

    p. 25: Moreover, developing models to directly predict the pain ratings is helpful to complement the group-level analysis, because the changes in consensus community structure over the early, middle, and late periods only indirectly reflect the different levels of pain.

    p. 27: This study also had some limitations. First, with the current experimental paradigm, it is difficult to dissociate the pain duration from the level of pain because the delivery of oral capsaicin commonly induces initial bursting and then a gradual decrease of pain over time. Though we aimed to model the effects of pain duration and pain avoidance ratings with our two primary analyses, i.e., consensus community detection and predictive modeling, we cannot fully dissociate the impact of time duration versus pain ratings.

    “• Although the manuscript is very well-written it might benefit from an even clearer and simpler explanation of what the consensus community structure and the underlying module allegiance measure assesses.”

    We thank you for the suggestion. Now we added additional (but simple) descriptions of module allegiance and consensus community detection methods.

    Revisions to the main manuscript:

    pp. 8-9: Here, the consensus community means the group-level representative structures of the distinct community partitions of individuals. To determine the consensus community across different individuals and times, we first obtained the module allegiance (Bassett et al., 2011) from the community assignment of each individual. Module allegiance assesses how much a pair of nodes is likely to be affiliated with the same community label, and is defined as a matrix T whose element Tij is 1 when nodes i and j are assigned to the same community and 0 when assigned to different communities. This conversion of the categorical community assignments to the continuous module allegiance values allows group-level summarization of different community structures of individuals.

    p. 14: Here, high module allegiance indicates the voxels of two regions are likely to be in the same community affiliation, and vice versa.

    “• The added value of the assessment of the dynamics of brain networks remains unclear. Specifically, it is unclear whether the current analysis of brain networks dynamics allows for a clearer distinction between and prediction of pain and no-pain states than other measures of static or dynamic brain activity or static measures of brain connectivity.”

    The main goal (and thus, the added value) of the current study was to provide a “mechanistic” understanding of the brain processes of sustained pain, rather than the “prediction.” Even though we included the results from the predictive modeling, as in Figures 4-6, our focus was more on the interpretation of the model to quantitatively examine the functional changes in the brain, not on the maximization of the prediction performance.

    Indeed, maximizing the prediction performance was the main goal of our previous study (Lee et al., 2021), in which we developed a predictive model of sustained pain based on the patterns of dynamic functional connectivity. The model showed better prediction performances compared to the current study, but it was challenging to interpret the model because of the high dimensionality of the model and its features. In addition, functional connectivity itself provides only limited insight into how functional brain networks are structured and reconfigured over time.

    In this sense, the multi-layer community detection method has several advantages to achieving our goal. First, the community detection analysis allows us to summarize the complex, high-dimensional whole-brain connectivity patterns into neurobiologically interpretable subsystems. Second, the multi-layer community detection method allows us to study the temporal changes in community structure by connecting the same nodes across different time points.

    Now we added a description of the rationale behind the choice of the multi-layer community detection analysis over the conventional functional connectivity methods, and the added value of our study.

    Revisions to the main manuscript:

    p. 3: In this study, we examined the reconfiguration of whole-brain functional networks underlying the natural fluctuation in sustained pain to provide a mechanistic understanding of the brain responses to sustained pain.

    p. 7: In this study, we used this approach to examine the temporal changes of brain network structures during sustained pain, which cannot be done with conventional functional connectivity-based analyses (Lee et al., 2021).

    p. 27: However, the previous model provides a limited level of mechanistic understanding because of the high dimensionality of the model and its features. In addition, functional connectivity itself provides only limited insight into how functional brain networks are structured and reconfigured over time.

    Reviewer #2 (public Review):

    “The Authors J-J Lee et al., investigated cortical and subcortical brain networks and their organization in communities over time during evoked tonic pain. The paper is well-written, and the findings are interesting and relevant for the field. Interestingly, other than confirming well known phenomena (e.g., segregation within the primary somatomotor cortex) the Authors identified an emerging "pain supersystem" during the initial increase of pain, in which subcortical and frontoparietal regions, usually more segregated, showed more interactions with the primary somatomotor cortex. Decrease of pain was instead associated to a reconfiguration of the networks that sees subcortical and frontoparietal regions connected with areas of the cerebellum. The main novelty of the proposed analysis, lies in the resulting high performances of the classifier, that shows how this interesting link between frontoparietal network and subcortical regions with the cerebellum, is predictive of pain decrease. In summary, the main strengths of the present manuscript are: • Inclusion of subcortical regions: most of the recent papers using the Shaefer parcellation in ~200 brain areas1, do not consider subcortical areas, ignoring possible relevant responses and behaviors of those regions. Not only the Authors smartly addressed this issue, but most of their results showed how subcortical regions played a key role in the networks reconfiguration over time during evoked sustained pain.

    • Robust classification results: high accuracy obtained on training dataset (internal validation), using a leave-one-out approach, and on the available independent test dataset (external validation) of relatively large sample size (N=74).

    • Clarity in the description of aim and sub-aims and exhaustive presentation of the obtained results helped by appropriate illustrations and figures (I suggest less wording in some of them).

    • Availability of continuous behavioral outcome (track ball).”

    We appreciate the reviewer’s summary and positive evaluations.

    “Even though the results are mostly cohesive with previous literature, some of the results need to be discussed in relationship to recently published papers on the same topic as well as justifying some of the non-standard methodological procedures adding appropriate citations (or more detailed comments). The Authors do not touch upon the concept of temporal summation of pain, historically associated with tonic pain, especially when the study is finalized to better understanding brain mechanisms in chronic pain populations (chronic pain patients often exhibit increased temporal summation of pain2). I would suggest starting from the paper recently published by Cheng et al. that also shares most of the methodological pipeline3 to highlight similarities and novelties and deepen the comparison with the associated literature.”

    We thank the reviewer and editor for the comment on this important topic. Temporal summation of pain indicates progressively increased sensation of pain during prolonged noxious stimulation (Price, Hu, Dubner, & Gracely, 1977), and has been suggested as a hallmark of chronic pain disorders including fibromyalgia (Cheng et al., 2022; Price et al., 2002). In a recent study by Cheng et al. (2022), the authors induced tonic pain using constantly high cuff pressure and examined whether the participants experienced increased pain in the late period compared to the early period of pain. On the contrary, in our experimental paradigm, the capsaicin liquid initially delivered into the oral cavity is being cleaned out by saliva, and thus overall pain intensity was decreasing over time, not increasing (Figure 1B). Therefore, the temporal summation of pain may occur in a limited period (e.g., the early period of the run), but it is difficult to examine its effect systematically in our study.

    However, it is notable that Cheng et al.’s results overlap with our findings. For example, Cheng et al. reported the intra-network segregation within the somatomotor network and the inter-network integration between the somatomotor and other networks during the temporal summation of pressure pain in patients with fibromyalgia, which were similar to the findings we reported in Figure S9 and Figure 4. Although it is unclear whether these results reflect the temporal summation of pain, these network-level features shared across the two studies are likely to be an essential component of the sustained pain processes in the brain.

    Now we added a comment on the temporal summation of pain in the main manuscript.

    Revisions to the main manuscript (p. 26):

    Interestingly, a recent fMRI study on the temporal summation of pain in fibromyalgia patients reported results similar to ours (Cheng et al., 2022), including the intra-network dissociation within the somatomotor network and the inter-network integration between the somatomotor and other networks during pain. Although we cannot directly examine whether the temporal summation of pain gave rise to these network-level changes due to the limitation of our experimental paradigm, these consistent findings between the two studies may suggest that our findings could be generalized to clinical conditions.

    We thank the reviewer and editor for the information about this recent publication. Cheng et al. (2022) was not published at the time we wrote the manuscript, and we were surprised that Cheng et al. shares many aspects with our study, e.g., both used multilayer community detection and also reported similar findings, as described above.

    However, there were some differences between the two studies as well.

    First, the focus of our study was on the brain dynamics during the natural time-course of sustained pain from its initiation to remission in healthy participants, whereas the focus of Cheng et al. was on the temporal summation phenomenon of pain (TSP) and the enhanced TSP in patients with fibromyalgia patients. Because of this difference in the research focuses, our study and Cheng et al. are providing many nonoverlapping results and insights. For example, our study paid particular attention to the coping mechanisms of the brain (e.g., the network-level changes in the subcortical and frontoparietal network regions) and the brain systems that are correlated with the natural decrease of pain (e.g., the cerebellum in Figure 5). In contrast, Cheng et al. (2022) identified the brain connectivity and network features important for the increased TSP in fibromyalgia patients.

    Second, our great interest was in identifying and visualizing the fine-grained spatiotemporal patterns of functional brain network changes over the period of sustained pain. To utilize fine-grained brain activity information, we conducted our main analyses at a voxel-level resolution and on the native brain space, such as in Figures 2-3 and Figures S5, S7, and S8. With this fine-grained spatiotemporal mapping, we were able to identify small, but important voxel-level dynamics.

    We now cited Cheng et al. (2022) in multiple places and revised the manuscript accordingly.

    Revisions to the main manuscript (p. 26):

    Interestingly, a recent fMRI study on the temporal summation of pain in fibromyalgia patients reported results similar to ours (Cheng et al., 2022), including the intra-network dissociation within the somatomotor network and the inter-network integration between the somatomotor and other networks during pain. Although we cannot directly examine whether the temporal summation of pain gave rise to these network-level changes due to the limitation of our experimental paradigm, these consistent findings between the two studies may suggest that our findings could be generalized to clinical conditions.

    “Here the main significant weaknesses of the study:

    • The data analysis is entirely conducted on young healthy subjects. This is not a limitation per se, but the conclusion about offering new insights into understanding mechanisms at the basis of chronic pain is too far from the results. Centralization of pain is very different from summation and habituation, especially if all the subjects in the study consistently rated increased and decreased pain in the same way (it never happens in chronic pain patients). A similar pipeline has been actually applied to chronic pain patients (fibromyalgia and chronic back pain)3,4. Discussing the results of the present paper in relationship to those, could offer a more robust way to connect the Authors' results to networks behavior in pathological brains.”

    We are grateful for the opportunity to discuss the clinical implication of our study. First of all, we agree with the reviewer and editor that we cannot make a definitive claim about chronic pain with the current study, and thus, we revised the last sentence of the abstract to tone down our claim.

    Revisions to the main manuscript (p. 2, in the abstract):

    This study provides new insights into how multiple brain systems dynamically interact to construct and modulate pain experience, advancing our mechanistic understanding of sustained pain.

    However, as we noted above in E-4, some of our findings were consistent with the findings from a previous clinical study (Cheng et al., 2022), suggesting the potential to generalize our study to clinical pain conditions. In addition, we previously reported that a predictive model of sustained pain derived from healthy participants performed better at predicting the pain severity of chronic pain patients than the model derived directly from chronic pain patients (Lee et al., 2021), highlighting the advantage of the “component process approach.”

    The component process approach aims to develop brain-based biomarkers for basic component processes first, which can then serve as intermediate features for the modeling of multiple clinical conditions (Woo, Chang, Lindquist, & Wager, 2017). This has been one of the core ideas of the Research Domain Criteria (RDoC) (Insel et al., 2010) and the Hierarchical Taxonomy of Psychopathology (HiTOP) (Kotov et al., 2017). If the clinical pain of a patient group is modeled as a whole, it becomes unclear what is being modeled because of the multidimensional and heterogeneous nature of clinical pain (Melzack, 1999) as well as other co-occurring health conditions (e.g., mental health issues, medication use, etc.). The component process approach, in contrast, can specify which components are being modeled and are relatively free from heterogeneity and comorbidity issues by experimentally manipulating the specific component of interest in healthy participants.

    The current study was conducted on healthy young adults based on the component process approach. We used oral capsaicin to experimentally induce sustained pain, which unfolds over protracted time periods and has been suggested to reflect some of the essential features of clinical pain (Rainville, Feine, Bushnell, & Duncan, 1992; Stohler & Kowalski, 1999). Therefore, the detailed characterization of the brain processes of sustained pain will be able to serve as an intermediate feature of multiple clinical conditions in future studies.

    Now we added the discussion on the clinical generalizability issue in the discussion section.

    Revisions to the main manuscript:

    pp. 25-26: An interesting future direction would be to examine whether the current results can be generalized to clinical pain. Experimental tonic pain has been known to share similar characteristics with clinical pain (Rainville et al., 1992; Stohler & Kowalski, 1999). In addition, in a recent study, we showed that an fMRI connectivity-based signature for capsaicin-induced orofacial tonic pain can be generalized to chronic back pain (Lee et al., 2021). Therefore, a detailed characterization of the brain responses to sustained pain has the potential to provide useful information about clinical pain.

    p. 26: Interestingly, a recent fMRI study on the temporal summation of pain in fibromyalgia patients reported results similar to ours (Cheng et al., 2022), including the intra-network dissociation within the somatomotor network and the inter-network integration between the somatomotor and other networks during pain. Although we cannot directly examine whether the temporal summation of pain gave rise to these network-level changes due to the limitation of our experimental paradigm, these consistent findings between the two studies may suggest that our findings could be generalized to clinical conditions.

    “Vice versa, the behavioral measure used to assess evoked pain perception (avoidance ratings), has been developed for chronic pain patients and never validated on healthy controls5. It might not be an appropriate measure considering the total absence of pain variability in the reported responses over forty-eight subjects6,7.”

    We acknowledge that pain avoidance measures are not fully validated in the healthy population. Nevertheless, we used this measure in this study for the following two main reasons that outweigh the limitations.

    First, a pain avoidance rating provides an integrative measure that can reflect the multi-dimensional aspects of sustained pain. One of the essential functions of pain is to avoid harmful situations and promote survival, and the avoidance motivation induced by pain is composed of not only sensory-discriminative, but also cognitive components including learning, valuation, and contexts (Melzack, 1999). According to the fear-avoidance model (Vlaeyen & Linton, 2012), if the pain-induced avoidance motivation is not resolved for a long time and is maladaptively associated with innocuous environments, chronic pain is likely to develop, suggesting the importance and clinical relevance of pain avoidance measures. In addition, our experimental design is particularly suitable for the use of avoidance rating because the oral capsaicin stimulation is accompanied by the urge to avoid the painful sensation, but it cannot immediately be resolved similar to chronic pain. Moreover, capsaicin is sometimes experienced as intense but less aversive (or even appetitive) in some cases, e.g., spicy food craver (Stevenson & Yeomans, 1993). In this case, avoidance ratings can provide a more reasonable measure of pain compared to the intensity rating.

    Second, the avoidance measure provides a common scale on which we can compare different types of aversive experiences, allowing us to conduct specificity tests for a predictive model of pain. For example, a recent study successfully compared the brain representations of two types of pain and two types of aversive, but non-painful experiences (e.g., aversive auditory and visual experiences) using the same avoidance measure (Ceko, Kragel, Woo, Lopez-Sola, & Wager, 2022). These comparisons were possible because the avoidance measure provided one common scale for all the aversive experiences regardless of their types of stimuli.

    To provide a better justification for the use of the avoidance measure, we now included the specificity test results of our pain predictive models. More specifically, we tested our module allegiance-based SVM and PCR models of pain on the aversive taste and aversive odor conditions (Figure S13).

    Despite these advantages, the use of avoidance rating without thorough validation is a limitation of the current study, and thus future studies need to examine the psychometric properties of the avoidance rating, e.g., examining the relationship among pain intensity, unpleasantness, and avoidance measures. However, the current study showed that the predictive models derived with pain avoidance rating (Study 1) could be used to predict the pain intensity rating (Study 2). In addition, the overall time-course of pain avoidance ratings in Study 1 was similar to the time-course of pain intensity ratings in Study 2, providing some supporting evidence for the convergent validity of the pain avoidance measure.

    As to the following comment, “It might not be an appropriate measure considering the total absence of pain variability in the reported responses over forty-eight subjects,” there are pieces of evidence supporting that the low between-individual variability of ratings is due to the characteristics of our experimental design, not to the fact that we used the avoidance measure. As we discussed in more detail in our response to E-1, our experimental procedure based on capsaicin liquid commonly induces the initial burst of painful sensation and the subsequent gradual relief for most of the participants (Figure 1B, left). A similar time-course pattern of ratings was observed in Study 2 (Figure 1B, right), which used the pain “intensity” rating, not the pain avoidance rating. In addition, previous studies with a similar experimental design (i.e., intra-oral capsaicin application) (Berry & Simons, 2020; Lu, Baad-Hansen, List, Zhang, & Svensson, 2013; Ngom, Dubray, Woda, & Dallel, 2001) also showed a similar time-course of pain ratings with low between-individual variability regardless of the rating types (e.g., VAS or irritation intensity), confirming that this observation is not unique to the pain avoidance rating.

    Now we added descriptions on the small between-individual variability of pain ratings and the use of avoidance ratings.

    Revisions to the main manuscript:

    pp. 5-7: Note that the overall trend of pain ratings over time was similar across participants because of the characteristics of our experimental design, which has also been observed in the previous studies that used oral capsaicin (Berry & Simons, 2020; Lu et al., 2013; Ngom et al., 2001). However, also note that each individual’s time-course of pain ratings were not entirely the same (Figures S2 and S3).

    p. 26: However, there are also differences between the characteristics of capsaicin-induced tonic pain versus clinical pain. For example, clinical pain continuously fluctuates over time in an idiosyncratic pattern (Apkarian, Krauss, Fredrickson, & Szeverenyi, 2001), whereas capsaicin-induced tonic pain showed a similar time-course pattern across the participants—i.e., increasing rapidly and then decreasing gradually (Figure 1B). This typical time-course of pain ratings has been reported in previous studies that used oral capsaicin (Berry & Simons, 2020; Lu et al., 2013; Ngom et al., 2001).

    pp. 26-27: Note that Study 1 used a pain avoidance measure that is not yet fully validated in healthy participants. However, we chose to use the pain avoidance measure, which can provide integrative information on the multi-dimensional aspects of pain (Melzack, 1999; Waddell, Newton, Henderson, Somerville, & Main, 1993). It also has a clinical implication considering that the maladaptive associations of pain avoidance to innocuous environments have been suggested as a putative mechanism of transition to chronic pain (Vlaeyen & Linton, 2012). Lastly, the avoidance measure can provide a common scale across different modalities of aversive experience, allowing us to compare their distinct brain representations (Ceko et al., 2022) or test the specificity of their predictive models (Lee et al., 2021) (Figure S13). Although the psychometric properties of the pain avoidance measure should be a topic of future investigation, we expect that the pain avoidance measure would have a high level of convergent validity with pain intensity given the observed similarity between pain avoidance (Study 1) and pain intensity (Study 2) in their temporal profiles. The generalizability of our PCR model across Studies 1 and 2 also supports this speculation. However, there would also be situations in which pain avoidance is dissociated from pain intensity. For example, capsaicin can be experienced to be intense but less aversive or even appetitive in some contexts, such as cravings for spicy food (Stevenson & Yeomans, 1993). In addition, the gradual rise of avoidance ratings during the late period of the control condition in Study 1 would not be observed if the intensity measure was used. Future studies need to examine the relationship between pain avoidance and the other pain assessments and the advantage of using the pain avoidance measure.

    “• The dynamic measure employed by the Authors is better described from the term "windowed functional connectivity". It is often considered a measure of dynamic functional connectivity and it gives information about fluctuations of the connectivity patterns over time. Nevertheless, the entire focus of the paper, including the title, is on dynamic networks, which inaccurately leads one to think of time-varying measures with higher temporal resolution (either updating for every acquired time point, as the Authors did in their previous publication on the same dataset4, or sliding windows involving weighting or tapering8,9). This allows one to follow network reorganization over time without averaging 2-min intervals in which several different brain mechanisms might play an important role3,10,11. In summary, the assumption of constant response throughout 2-min periods of tonic pain and the use of Pearson correlations do not mirror the idea of dynamic analysis expressed by the Authors in title and introduction. I would suggest removing "dynamic" from the title, reduce the emphasis on this concept, address possible confounds introduced by the choice of long windows and rephrase the aim of the study in terms of brain network reconfiguration over the main phases of tonic pain experience.”

    Now we removed the word ‘dynamic’ from many places in the manuscript, including the title. In addition, we added a brief discussion on the reason we chose to use the long and non-overlapping windows for connectivity calculation.

    Revisions to the main manuscript (p. 8):

    Although the long duration of the time window without overlaps may obscure the fine-grained temporal dynamics in functional connectivity patterns, we chose to use this long time window based on previous literature (Bassett et al., 2011; Robinson, Atlas, & Wager, 2015), which also used long time windows to obtain more reliable estimates of network structures and their transitions.

    “• Procedure chosen for evoking sustained pain. To the best of my knowledge, capsaicin sauce on the tongue is not a validated tonic pain procedure. In favor of this argument is the absence of inter-subject variability in the behavioral results showed in the paper, very unusual for response to painful stimulations. The procedure is well described by the Authors, and some precautions like letting the liquid drying before the start of the scan, have helped reducing confounds. Despite this, the measures in figure 1B suggest that the intensity of the painful stimulation is not constant as expected for sustained pain (probably the effect washes out with the saliva). In this case, the first six-minute interval requires particular attention because it encapsulates the real tonic pain phase, and the following ones require more appropriate labels. Ideally the Author should cite previous studies showing that tongue evoked pain elicits a very specific behavioral response (summation, habituation/decrease of pain, absence of pain perception). If those works are missing, this response need to be treated as a funding rather than an obvious point.”

    We addressed this comment. Moreover, we could find previous studies that experimentally induced tonic pain through the application of capsaicin on the tongue (Berry & Simons, 2020; Boudreau, Wang, Svensson, Sessle, & Arendt-Nielsen, 2009; Green, 1991; Ngom et al., 2001), suggesting that our experimental procedure is in line with previous literature.

    Reviewer #3 (Public Review ):

    “In their manuscript, Lee and colleagues explore the dynamics of the functional community structure of the brain (as measured with fMRI) during sustained experimental pain and provide several potentially highly valuable insights into, and evaluate the predictive capacity of, the underlying dynamic processes. The applied methodology is novel but, at the same time, straightforward and has solid foundations. The findings are very interesting and, potentially, of high scientific impact as they may significantly push the boundaries of our understanding of the dynamic neural processes during sustained pain, with a (somewhat limited) potential for clinical translation.

    However (Major Issue 1), after reading the current manuscript version, not all of my doubts have been dissolved regrading the specificity of the results to pain. Moreover (Major Issue 2), some of the results (specifically, those related to the group level analysis of community differences) do not seem to be underpinned with a proper statistical inference in the current version of the manuscript and, therefore, their presentation and discussion may not be proportional to the degree of evidence. Next to these Major Issues (detailed below), some other, minor clarifications might also be needed before publications. These are detailed below or in the private part of the review ("Recommendations for the authors").

    Despite these issues, this is, in general, a high quality work with a high level of novelty and - after addressing the issues - it has a very high potential for becoming an important contribution (and a very interesting read) to the pain-research community and beyond.”

    We appreciate the reviewer’s thoughtful comments. We have revised the manuscript to address the Reviewer’s major concerns, as described below.

    “Major Issue 1:

    The main issue with the manuscript is that it remains somewhat unclear, how specific the results are to pain.

    Differences between the control resting state and the capsaicin trials might be - at least partially - driven by other factors, like:

    • motion artifacts
    • saliency, attention, axiety, etc.

    Differences between stages over the time-course might, additionally, be driven by scanner drifts (to which the applied approach might be less sensitive, but the possibility is still there ) or other gradual processes, e.g. shifts in arousal, attention shifts, alertness, etc.

    All the above factors might emerge as confounding bias in both of the predictive models.

    This problem should be thoroughly discussed, and at least the following extra analyses are recommended, in order to attenuate concerns related to the overall specificity and neurobiological validity of the results:

    • reporting of, and testing for motion estimates (mean, max, median framewise displacement or anything similar)
    • examining whether these factors might, at least partially, drive the predictive models.
    • e.g. applying the PCR model on the resting state data and verifying of the predicted timecourse is flat (no inverse U-shape, that is characteristic to all capsaicin trials).

    Not using the additional sessions (bitter taste, aversive odor, phasic heat) feels like a missed opportunity, as they could also be very helpful in addressing this issue.”

    We thank the reviewer for this comment on the important issue regarding the specificity of our results and the potential influences of noise. The effects of head motion and physiological confounds are particularly relevant to pain studies because pain involves substantial physiological changes and often causes head motion. To address the related concerns of specificity, we conducted additional analyses assessing the independence of our predictive models (i.e., SVM and PCR models) from head movement and physiology variables and the specificity of our models to pain versus non-painful aversive conditions (i.e., bitter taste and aversive odor) in Study 1.

    First, we examined the overall changes of framewise displacement (FD) (Power, Barnes, Snyder, Schlaggar, & Petersen, 2012), heart rate (HR), and respiratory rate (RR) in the capsaicin condition (Figure S11). For the univariate comparison between the capsaicin vs. control conditions (Figure S11A), the results showed that, as expected, the capsaicin condition caused significant changes in head motion and autonomic responses. The mean FD and HR were significantly higher, and the RR was lower in the capsaicin condition compared to the control condition (FD: t47 = 5.30, P = 2.98 × 10-6; HR: t43 = 4.98, P = 1.10 × 10-5; RR: t43 = -1.91, P = 0.063, paired t-test). In addition, the increased motion and autonomic responses were more prominent in the early period of pain (Figure S11B). The 10-binned (2 mins per time-bin) FD and HR showed a decreasing trend while the RR showed an increasing trend over time in the capsaicin condition. The comparisons between the early (1-3 bins, 0-6 min) vs. late (8-10 bins, 14-20 min) periods of the capsaicin condition showed significant differences both for FD and HR (FD: t47 = 6.45, P = 8.12 × 10-8; HR: t43 = 6.52, P = 6.41 × 10-8; RR: t43 = -1.61, P = 0.11, paired t-test). These results suggest that while participants were experiencing capsaicin tonic pain, particularly during the early period, head motion and heart rate were increased, while breathing was slowed down. Note that we needed to exclude 4 participants’ data in this analysis due to technical issues with the physiological data acquisition.

    Next, we examined whether the changes in head motion and physiological responses influenced our predictive model performance (Figure S12). We first regressed out the mean FD, HR, and RR (concatenated across conditions and participants as we trained the SVM model) from the predicted values of the SVM model with leave-one-subject-out cross-validation (2 conditions × 44 participants = 88) and then calculated the classification accuracy again (Figure S12A). The results showed that the SVM model showed a reduced, but still significant classification accuracy for the capsaicin versus control conditions in a forced-choice test (n = 44, accuracy = 89%, P = 1.41 × 10-7, binomial test, two-tailed). We also did the same analysis for the PCR model (10 time-bins × 44 participants = 440) and the PCR model also showed a significant prediction performance (n = 44, mean prediction-outcome correlation r = 0.20, P = 0.003, bootstrap test, two-tailed, mean squared error = 0.159 ± 0.022 [mean ± s.e.m.]) (Figure S12B). These results suggest that our SVM and PCR models capture unique variance in tonic pain above and beyond the head movement and physiological changes.

    Lastly, we examined the specificity of our predictive models to pain, by testing the models on the non-painful but aversive conditions including the bitter taste (induced by quinine) and aversive odor (induced by fermented skate) conditions (Figure S13). All the model responses were obtained using leave-one-participant-out cross-validation. The results showed that the overall model responses of the SVM model for the bitter taste and aversive odor conditions were higher than those for the control condition but lower than the capsaicin condition (Figure S13A). Classification accuracies for comparing capsaicin vs. bitter taste and capsaicin vs. aversive odor were all significant (for capsaicin vs. bitter taste, accuracy = 79%, P = 6.17 × 10-5, binomial test, two-tailed, Figure S13C; for capsaicin vs. aversive odor, accuracy = 83%, P = 3.31 × 10-6, binomial test, two-tailed, Figure S13E), supporting the specificity of our SVM model of pain. Similarly, the model responses of the PCR model for the bitter taste and aversive odor conditions were lower than the capsaicin condition, and their temporal trajectories were less steep and fluctuating compared to the capsaicin condition (Figure S13B). The time-course of the model responses for the control condition was flatter than all other conditions and did not show the inverted U-shape. Furthermore, the model responses of the bitter taste and aversive odor conditions did not show the significant correlations with the actual avoidance ratings (bitter taste: mean prediction-outcome correlation r = 0.05, P = 0.41, bootstrap test, two-tailed, mean squared error = 0.036 ± 0.006 [mean ± s.e.m.], Figure S13D; aversive odor: mean prediction-outcome correlation r = 0.12, P = 0.06, bootstrap test, two-tailed, mean squared error = 0.044 ± 0.004 [mean ± s.e.m.], Figure S13F), suggesting the specificity of PCR model to pain.

    Overall, we have provided evidence that our models can predict pain ratings above and beyond the head motion and physiological changes and that the models are more responsive to pain compared to non-painful aversive conditions.

    Now we added descriptions on the specificity tests to the main manuscript and also to the Supplementary Information.

    Revisions to the main manuscript (p. 20):

    Specificity of the module allegiance-based predictive models To examine whether the predictive models were specific to pain and the prediction performances were not influenced by confounding variables such as head motion and physiological changes, we conducted additional analyses as shown in Figures S11-13. The SVM and PCR models showed significant prediction performances even after controlling for head motion (i.e., framewise displacement) and physiological responses (i.e., heart rate and respiratory rate) (Figures S11 and S12) and did not respond to the non-painful but aversive conditions including the bitter taste and aversive odor conditions (Figure S13), supporting the specificity of our predictive to pain. For details, please see Supplementary Results.

    Revisions to the Supplementary Information (pp. 2-4):

    Specificity analysis (Figures S11-13) To examine whether the predictive models (i.e., SVM and PCR models) were specific to pain and not influenced by confounding noises, we conducted additional specificity analysis assessing the independence of the models from head movement and physiology variables and specificity of our models to pain versus non-painful aversive conditions (i.e., bitter taste and aversive odor) in Study 1. First, we examined the overall changes of framewise displacement (FD) (Power et al., 2012), heart rate (HR), and respiratory rate (RR) in sustained pain (Figure S11). For the univariate comparison between capsaicin vs. control conditions (Figure S11A), the results showed that, as expected, capsaicin condition caused significant changes in motion and autonomic responses. The mean FD and HR were significantly higher, and the RR was lower in the capsaicin condition compared to the control condition (FD: t47 = 5.30, P = 2.98 × 10-6; HR: t43 = 4.98, P = 1.10 × 10-5; RR: t43 = -1.91, P = 0.063, paired t-test). For the temporal changes of movement and physiology variables (Figure S11B), the results showed that the increased motion and autonomic responses are more prominent in the early period of pain. The 10-binned (2 mins per time-chunk) FD and HR showed decreasing trend while the RR showed increasing trend over time in capsaicin condition. Additional univariate comparisons between early (1-3 bins, 0-6 min) vs. late (8-10 bins, 14-20 min) period of capsaicin condition showed that differences were significant for FD and HR (FD: t47 = 6.45, P = 8.12 × 10-8; HR: t43 = 6.52, P = 6.41 × 10-8; RR: t43 = -1.61, P = 0.11, paired t-test). This suggests that while participants were experiencing tonic pain, particularly in the early period, motion and heart rate was increased but breathing was slowed. Note that we needed to exclude 4 participants’ data due to technical issues with physiological data acquisition. Next, we examined whether the head movement and physiological responses are the main driver of our predictive models (Figure S12). For all the original signature responses from SVM model (2 conditions × 44 participants = 88), we regressed out the mean FD, HR, and RR (concatenated across conditions and participants as the SVM model was trained) and calculated the classification accuracy (Figure S12A). Although the signature responses were controlled for movement and physiology variables, the SVM model still showed a high classification accuracy for the capsaicin versus control conditions in a forced-choice test (n = 44, accuracy = 89%, P = 1.41 × 10-7, binomial test, two-tailed). Similarly, for all the original signature responses from PCR model (10 time-bins × 44 participants = 440), we regressed out the 10-binned FD, HR, and RR (concatenated across time-bins and participants as the PCR model was trained) and calculated the within-individual prediction-outcome correlation (Figure S12B). Again, the PCR model showed a significantly high predictive performance (n = 44, mean prediction-outcome correlation r = 0.20, P = 0.003, bootstrap test, two-tailed, mean squared error = 0.159 ± 0.022 [mean ± s.e.m.]) while controlling for movement and physiology variables. These results suggest that our SVM and PCR models captures unique variance in tonic pain above and beyond the head movement and physiological changes. Lastly, we examined the specificity of our predictive models to pain, by testing the models onto the non-painful but tonic aversive conditions including bitter taste (induced by quinine) and aversive odor (induced by fermented skate) (Figure S13). All the signature responses were obtained using leave-one-participant-out cross-validation. The results showed that the overall signature responses of SVM model for bitter taste and aversive odor conditions were higher than those for control conditions, but lower than capsaicin condition (Figure S13A). Classification accuracy between capsaicin vs. bitter taste and vs. aversive odor were all significantly high (capsaicin vs. bitter taste: accuracy = 79%, P = 6.17 × 10-5, binomial test, two-tailed, Figure S13C; capsaicin vs. aversive odor: accuracy = 83%, P = 3.31 × 10-6, binomial test, two-tailed, Figure S13E), suggesting the specificity of SVM model to pain. Similarly, the temporal trajectories of the signature responses of PCR model for bitter taste and aversive odor conditions were not overlapping with that of the capsaicin condition (Figure S13B). Furthermore, the signature responses of bitter taste and aversive odor conditions do not have significant relationship with the actual avoidance ratings (bitter taste: mean prediction-outcome correlation r = 0.05, P = 0.41, bootstrap test, two-tailed, mean squared error = 0.036 ± 0.006 [mean ± s.e.m.], Figure S13D; aversive odor: mean prediction-outcome correlation r = 0.12, P = 0.06, bootstrap test, two-tailed, mean squared error = 0.044 ± 0.004 [mean ± s.e.m.], Figure S13F), suggesting the specificity of PCR model to pain. Overall, we have provided evidence that the module allegiance-based models can predict pain ratings above and beyond the movement and physiological changes, and are more responsive to pain compared to non-painful aversive conditions, which suggest the specificity of our results to pain.

    “Major Issue 2:

    Another important issue with the manuscript is the (apparent) lack of statistical inference when analyzing the differences in the group-level consensus community structures (both when comparing capsaicin to control and when analysing changes over the time-course of the capsaicin-challenge).

    Although I agree that the observed changes seem biologically plausible and fit very well to previous results, without proper statistical inference we can't determine, how likely such differences are to emerge just by chance.

    This makes all results on Figs. 2 and 3, and points 1, 4 and 5 in the discussion partially or fully speculative or weakly underpinned, comprising a large proportion of the current version of the manuscript.

    Let me note, that this issue only affects part of the results and the remaining - more solid - results may already provide a substantial scientific contribution (which might already be sufficient to be eligible for publication in eLife, in my opinion).

    Therefore I see two main ways of handling Major Issue 2:

    • enhancing (or clarifying potential misunderstandings regarding) the methodology (see my concrete, and hopefully feasible, suggestions in the "private part" of the review),
    • de-weighting the presentation and the discussion of the related results.

    I believe there are many ways to test the significance of these differences. I highlight two possible, permutation testing-based ideas.

    Idea 1: permuting the labels ctr-capsaicin, or early-mid-late, repeating the analysis, constructing the proper null distribution of e.g. the community size changes and obtain the p-values. Idea 2: "trace back" communities to the individual level and do (nonparametric) statistical inference there.”

    We appreciate this important comment. We did not conduct statistical inference when comparing the group-level consensus community affiliations of the different conditions (Figure 2) or different phases (Figure 3) because of the difficulty in matching the community affiliation values of the networks to be compared.

    For example, let us assume that the 800 out of 1,000 voxels of community #1 and 1,000 out of 4,000 voxels of community #2 in the control condition are commonly affiliated with the same community #3 in the capsaicin condition. To compare the community affiliation between two conditions, we should first match the community label of the capsaicin condition (i.e., #3) to that of the control condition (i.e., #1 or #2), and here a dilemma occurs; if we prioritize the proportion of the overlapping voxels for the matching, the common community should be labeled as #1, whereas if we prioritize the number of the overlapping voxels for the matching, the label of the common community should be #2. Although both choices look reasonable, none of them can be a perfect solution.

    As the example above, it is impossible to exactly match the community affiliation of the different networks. We must choose an imperfect criterion for the matching procedure, which essentially affects the comparison of network structure. This was the main reason that we limited our results of Figures 2-3 to a qualitative description based on visual inspection. Moreover, the group-level consensus community structures in Figures 2-3 are not a simple group statistic like sample mean; they were obtained from multiple steps of analyses including permutation-based thresholding and unsupervised clustering, which could further complicate the interpretation of statistical tests.

    Alternatively, there is a slightly different but more rigorous approach to the comparisons of the community structures, which is the Phi-test (Alexander-Bloch et al., 2012; Lerman-Sinkoff & Barch, 2016). Instead of direct use of the community labels, this method converts the community label of each voxel into a list of module allegiance values between the seed voxel and all the voxels of the brain (i.e., 1 if the seed and target voxels have the same community label and 0 otherwise). This allows quantitative comparisons of voxel-level community profiles between different conditions without an arbitrarily matching of the community labels. We adopted this Phi-test for our analyses to examine whether the regional community affiliation pattern is significantly different between (i) the capsaicin vs. control conditions and (ii) the early vs. late periods of pain (Figure S6), which correspond to the main findings of the Figures 2 and 3 in our manuscript, respectively.

    More specifically, to compare the group-level consensus community structures between the capsaicin vs. control conditions and the early vs. late periods, we first obtained a seed-based module allegiance map for each voxel (i.e., using each voxel as a seed). Then, we calculated a correlation coefficient of the module allegiance values between two different conditions for each voxel. This correlation coefficient can serve as an estimate of the voxel-level similarity of the consensus community profile. Because module allegiance is a binary variable, these correlation values are Phi coefficients. A small Phi coefficient means that the spatial pattern of brain regions that have the same community affiliation with the given voxel are different between the two conditions. For example, if a voxel is connected to the somatomotor-dominant community during the capsaicin condition and the default-mode-dominant community during the control condition, the brain regions that have the same community label with the voxel will be very different, and thus the Phi coefficient will become small. Moreover, the Phi coefficient can be small even if a voxel is affiliated as the same (matched) community label for both conditions, when the spatial patterns of the same community is different between conditions.

    To calculate the statistical significance of the Phi coefficient, we conducted permutation tests, in which we randomly shuffled the condition labels in each participant and obtained the group-level consensus community structure for each shuffled condition. Then, we calculated the voxel-level correlations of the module allegiance values between the two shuffled conditions. We repeated this procedure 1,000 times to generate the null distribution of the Phi coefficients, and calculated the proportion of null samples that have a smaller Phi coefficient (i.e., a more dis-similar regional community structure) than the non-shuffled original data.

    Results showed that there are multiple voxels with statistical significance (permutation tests with 1,000 iterations, one-tailed) in the area where the community affiliations of the two contrasting conditions were different (Figure S6). For example, the frontoparietal and subcortical regions for the capsaicin vs. control (c.f., Figure 2), and the frontoparietal, subcortical, brainstem, and cerebellar regions for the early vs. late period of pain (c.f., Figure 3) contain voxels that survived after thresholding with FDR-corrected q < 0.05, suggesting the robustness of our main results.

    Particularly, the somatomotor and insular cortices showed statistical significance in the permutation test, and this may reflect the large changes in other areas that are connecting to the somatomotor and insular cortices across different conditions. The statistical significance was also observed in the visual cortex, which was unexpected. We interpret that the spatial distribution of the visual network community is too stable across conditions, and thus the null distribution from permutation formed a very narrow distribution of Phi coefficients. Therefore, a small change in the community structure could achieve statistical significance.

    Now we added descriptions on the permutation tests.

    Revisions to the main manuscript:

    p. 9: Permutation tests confirmed that the community assignment in the frontoparietal and subcortical regions showed significant changes between the capsaicin versus control conditions (Figure S6A).

    p. 13: Permutation tests further confirmed that the community assignment in the frontoparietal, subcortical, and brainstem regions showed significant changes between the early versus late period of pain (Figure S6B).

    pp. 36-37: Permutation tests for regional differences in community structures. To test the statistical significance of the voxel-level difference of consensus community structures (Figures 2 and 3), we performed the following Phi-test (Alexander-Bloch et al., 2012; Lerman-Sinkoff & Barch, 2016). First, for each given voxel, we compared the community label of the voxel to the community label of all the voxels, generating a list of voxel-seed module allegiance values that allow quantitative comparison of voxel-level community profile (e.g., [1, 0, 1, 1, 0, 0, ...], whose element is equal to 1 if the seed and target voxels were assigned to the same community and 0 otherwise). Next, a correlation coefficient was calculated between the module allegiance values of the two different brain community structures (i.e., capsaicin versus control, and early versus late). This correlation coefficient is an estimate of the regional similarity of community profiles (here, the correlation coefficient is Phi coefficient because module allegiance is a binary variable). To estimate the statistical significance of the Phi coefficient, we performed permutation tests, in which we randomly shuffled the labels and then obtained the group-level consensus community structures from the shuffled data. Then, the Phi coefficient between the module allegiance values of the two shuffled consensus community structures was calculated. We repeated this procedure 1,000 times to generate the null distribution of the Phi coefficient for each voxel. Lastly, we examined the probability to observe a smaller Phi coefficient (i.e., a more dissimilar community profile) than the one from the non-shuffled original data, which corresponds to the P-value of the permutation test. All the P-values were one-tailed as the hypothesis of this permutation test is unidirectional.

  3. eLife

    Evaluation Summary:

    This paper will be of great interest to researchers interested in the brain mechanisms of pain. It shows how the connectivity of brain networks associated with sustained pain change over time. These findings are conclusively supported by state-of-the-art fMRI analyses of a tonic pain paradigm in two cohorts of healthy human participants. These insights are important for the understanding of the brain mechanisms of sustained pain which is the hallmark of chronic pain as a major health care problem.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 and Reviewer #3 agreed to share their name with the authors.)

  4. eLife

    Reviewer #1 (Public Review):

    This study investigates the dynamics of brain network connectivity during sustained experimental pain in healthy human participants. To this end, capsaicin was applied to the tongues of two cohorts of participants (discovery cohort, N=48; replication cohort, N=74). This procedure resulted in pain for several minutes. During sustained pain, pain avoidance/intensity ratings and fMRI scans were obtained. The analyses (i) compare the pain state with a resting state, (ii) assess the dynamics of brain networks during sustained pain, and (iii) aim to predict pain based on the dynamics of brain networks. To this end, the analyses focus on community structures of time-evolving networks. The results show that sustained pain is associated with the emergence of a brain network including somatomotor, frontoparietal, basal ganglia and thalamic brain areas. The somatomotor area of the tongue is particularly involved in that network while this area is decoupled from other parts of the somatomotor cortex. Moreover, the network configuration changes over time with the frontoparietal network decoupling from the somatomotor network. Frontoparietal-cerebellar connections were predictive of decreases of pain. Together, the findings provide novel and convincing insights into the dynamics of brain network during sustained pain.

    Strengths
    • The brain mechanisms of sustained pain is a timely and relevant topic with potential clinical implications.
    • Assessing the dynamics of sustained pain and relating it to the dynamics of brain networks is a timely and promising approach to further the understanding of the brain mechansims of pain.
    • The study includes discovery and replication cohorts and pursues a cutting-edge analysis strategy.
    • The manuscript is very well-written and the results are visualized in an exemplary manner including a graphical outline and summary of the findings.

    Weaknesses
    • It remains unclear whether the changes of brain networks over time simply reflect the duration of sustained pain or whether they essentially reflect different levels of pain intensity/avoidance.
    • Although the manuscript is very well-written it might benefit from an even clearer and simpler explanation of what the consensus community structure and the underlying module allegiance measure assesses.
    • The added value of the assessment of the dynamics of brain networks remains unclear. Specifically, it is unclear whether the current analysis of brain networks dynamics allows for a clearer distinction between and prediction of pain and no-pain states than other measures of static or dynamic brain activity or static measures of brain connectivity.

  5. eLife

    Reviewer #2 (Public Review):

    The Authors J-J Lee et al., investigated cortical and subcortical brain networks and their organization in communities over time during evoked tonic pain. The paper is well-written, and the findings are interesting and relevant for the field. Interestingly, other than confirming well known phenomena (e.g., segregation within the primary somatomotor cortex) the Authors identified an emerging "pain supersystem" during the initial increase of pain, in which subcortical and frontoparietal regions, usually more segregated, showed more interactions with the primary somatomotor cortex. Decrease of pain was instead associated to a reconfiguration of the networks that sees subcortical and frontoparietal regions connected with areas of the cerebellum. The main novelty of the proposed analysis, lies in the resulting high performances of the classifier, that shows how this interesting link between frontoparietal network and subcortical regions with the cerebellum, is predictive of pain decrease. In summary, the main strengths of the present manuscript are:
    • Inclusion of subcortical regions: most of the recent papers using the Shaefer parcellation in ~200 brain areas1, do not consider subcortical areas, ignoring possible relevant responses and behaviors of those regions. Not only the Authors smartly addressed this issue, but most of their results showed how subcortical regions played a key role in the networks reconfiguration over time during evoked sustained pain.
    • Robust classification results: high accuracy obtained on training dataset (internal validation), using a leave-one-out approach, and on the available independent test dataset (external validation) of relatively large sample size (N=74).
    • Clarity in the description of aim and sub-aims and exhaustive presentation of the obtained results helped by appropriate illustrations and figures (I suggest less wording in some of them).
    • Availability of continuous behavioral outcome (track ball).

    Even though the results are mostly cohesive with previous literature, some of the results need to be discussed in relationship to recently published papers on the same topic as well as justifying some of the non-standard methodological procedures adding appropriate citations (or more detailed comments). The Authors do not touch upon the concept of temporal summation of pain, historically associated with tonic pain, especially when the study is finalized to better understanding brain mechanisms in chronic pain populations (chronic pain patients often exhibit increased temporal summation of pain2). I would suggest starting from the paper recently published by Cheng et al. that also shares most of the methodological pipeline3 to highlight similarities and novelties and deepen the comparison with the associated literature. Here the main significant weaknesses of the study:
    • The data analysis is entirely conducted on young healthy subjects. This is not a limitation per se, but the conclusion about offering new insights into understanding mechanisms at the basis of chronic pain is too far from the results. Centralization of pain is very different from summation and habituation, especially if all the subjects in the study consistently rated increased and decreased pain in the same way (it never happens in chronic pain patients). A similar pipeline has been actually applied to chronic pain patients (fibromyalgia and chronic back pain)3,4. Discussing the results of the present paper in relationship to those, could offer a more robust way to connect the Authors' results to networks behavior in pathological brains. Vice versa, the behavioral measure used to assess evoked pain perception (avoidance ratings), has been developed for chronic pain patients and never validated on healthy controls5. It might not be an appropriate measure considering the total absence of pain variability in the reported responses over forty-eight subjects6,7.
    • The dynamic measure employed by the Authors is better described from the term "windowed functional connectivity". It is often considered a measure of dynamic functional connectivity and it gives information about fluctuations of the connectivity patterns over time. Nevertheless, the entire focus of the paper, including the title, is on dynamic networks, which inaccurately leads one to think of time-varying measures with higher temporal resolution (either updating for every acquired time point, as the Authors did in their previous publication on the same dataset4, or sliding windows involving weighting or tapering8,9). This allows one to follow network reorganization over time without averaging 2-min intervals in which several different brain mechanisms might play an important role3,10,11. In summary, the assumption of constant response throughout 2-min periods of tonic pain and the use of Pearson correlations do not mirror the idea of dynamic analysis expressed by the Authors in title and introduction. I would suggest removing "dynamic" from the title, reduce the emphasis on this concept, address possible confounds introduced by the choice of long windows and rephrase the aim of the study in terms of brain network reconfiguration over the main phases of tonic pain experience.
    • Procedure chosen for evoking sustained pain. To the best of my knowledge, capsaicin sauce on the tongue is not a validated tonic pain procedure. In favor of this argument is the absence of inter-subject variability in the behavioral results showed in the paper, very unusual for response to painful stimulations. The procedure is well described by the Authors, and some precautions like letting the liquid drying before the start of the scan, have helped reducing confounds. Despite this, the measures in figure 1B suggest that the intensity of the painful stimulation is not constant as expected for sustained pain (probably the effect washes out with the saliva). In this case, the first six-minute interval requires particular attention because it encapsulates the real tonic pain phase, and the following ones require more appropriate labels. Ideally the Author should cite previous studies showing that tongue evoked pain elicits a very specific behavioral response (summation, habituation/decrease of pain, absence of pain perception). If those works are missing, this response need to be treated as a funding rather than an obvious point.

    References
    1. Schaefer, A. et al. Local-Global Parcellation of the Human Cerebral Cortex from Intrinsic Functional Connectivity MRI. Cereb. Cortex N. Y. N 1991 28, 3095-3114 (2018).
    2. Price, D. D. et al. Enhanced temporal summation of second pain and its central modulation in fibromyalgia patients. Pain 99, 49-59 (2002).
    3. Cheng, J. C. et al. Dynamic functional brain connectivity underlying temporal summation of pain in fibromyalgia. Arthritis Rheumatol. Hoboken NJ (2021) doi:10.1002/art.42013.
    4. Lee, J.-J. et al. A neuroimaging biomarker for sustained experimental and clinical pain. Nat. Med. 27, 174-182 (2021).
    5. Vlaeyen, J. W. S. & Linton, S. J. Fear-avoidance model of chronic musculoskeletal pain: 12 years on. Pain 153, 1144-1147 (2012).
    6. Asmundson, G. J., Norton, P. J. & Norton, G. R. Beyond pain: the role of fear and avoidance in chronicity. Clin. Psychol. Rev. 19, 97-119 (1999).
    7. Beebe, J. A. et al. Gait Variability and Relationships With Fear, Avoidance, and Pain in Adolescents With Chronic Pain. Phys. Ther. 101, pzab012 (2021).
    8. Hindriks, R. et al. Can sliding-window correlations reveal dynamic functional connectivity in resting-state fMRI? NeuroImage 127, 242-256 (2016).
    9. Lurie, D. J. et al. Questions and controversies in the study of time-varying functional connectivity in resting fMRI. Netw. Neurosci. 4, 30-69 (2020).
    10. Allen, E. A. et al. Tracking Whole-Brain Connectivity Dynamics in the Resting State. Cereb. Cortex N. Y. NY 24, 663-676 (2014).
    11. Hutchison, R. M. et al. Dynamic functional connectivity: promise, issues, and interpretations. NeuroImage 80, 360-378 (2013).

  6. eLife

    Reviewer #3 (Public Review):

    In their manuscript, Lee and colleagues explore the dynamics of the functional community structure of the brain (as measured with fMRI) during sustained experimental pain and provide several potentially highly valuable insights into, and evaluate the predictive capacity of, the underlying dynamic processes.

    The applied methodology is novel but, at the same time, straightforward and has solid foundations. The findings are very interesting and, potentially, of high scientific impact as they may significantly push the boundaries of our understanding of the dynamic neural processes during sustained pain, with a (somewhat limited) potential for clinical translation.

    However (Major Issue 1), after reading the current manuscript version, not all of my doubts have been dissolved regrading the specificity of the results to pain.
    Moreover (Major Issue 2), some of the results (specifically, those related to the group level analysis of community differences) do not seem to be underpinned with a proper statistical inference in the current version of the manuscript and, therefore, their presentation and discussion may not be proportional to the degree of evidence.

    Despite these issues, this is, in general, a high quality work with a high level of novelty and - after addressing the issues - it has a very high potential for becoming an important contribution (and a very interesting read) to the pain-research community and beyond.

    Major Issue 1:

    The main issue with the manuscript is that it remains somewhat unclear, how specific the results are to pain.

    Differences between the control resting state and the capsaicin trials might be - at least partially - driven by other factors, like:
    - motion artifacts
    - saliency, attention, axiety, etc.
    Differences between stages over the time-course might, additionally, be driven by scanner drifts (to which the applied approach might be less sensitive, but the possibility is still there ) or other gradual processes, e.g. shifts in arousal, attention shifts, alertness, etc.

    All the above factors might emerge as confounding bias in both of the predictive models.

    This problem should be thoroughly discussed, and at least the following extra analyses are recommended, in order to attenuate concerns related to the overall specificity and neurobiological validity of the results:
    - reporting of, and testing for motion estimates (mean, max, median framewise displacement or anything similar)
    - examining whether these factors might, at least partially, drive the predictive models.
    - e.g. applying the PCR model on the resting state data and verifying of the predicted timecourse is flat (no inverse U-shape, that is characteristic to all capsaicin trials).

    Not using the additional sessions (bitter taste, aversive odor, phasic heat) feels like a missed opportunity, as they could also be very helpful in addressing this issue.

    Major Issue 2:

    Another important issue with the manuscript is the (apparent) lack of statistical inference when analyzing the differences in the group-level consensus community structures (both when comparing capsaicin to control and when analysing changes over the time-course of the capsaicin-challenge).

    Although I agree that the observed changes seem biologically plausible and fit very well to previous results, without proper statistical inference we can't determine, how likely such differences are to emerge just by chance.

    This makes all results on Figs. 2 and 3, and points 1, 4 and 5 in the discussion partially or fully speculative or weakly underpinned, comprising a large proportion of the current version of the manuscript.

    Let me note, that this issue only affects part of the results and the remaining - more solid - results may already provide a substantial scientific contribution.

    Therefore I see two main ways of handling Major Issue 2:
    - enhancing (or clarifying potential misunderstandings regarding) the methodology (see my concrete, and hopefully feasible, suggestions in the "private part" of the review),
    - de-weighting the presentation and the discussion of the related results.

    I believe there are many ways to test the significance of these differences. I highlight two possible, permutation testing-based ideas.

    Idea 1: permuting the labels ctr-capsaicin, or early-mid-late, repeating the analysis, constructing the proper null distribution of e.g. the community size changes and obtain the p-values.

    Idea 2: "trace back" communities to the individual level and do (nonparametric) statistical inference there.

  7. Version published to 10.1101/2021.10.16.464642v1 on bioRxiv