Open reviews
November 4, 2021 — Open review of preprint by Di and colleagues: PubPeer
Di, X., Zhang, Z., Xu, T., & Biswal, B. B. (2021). Dynamic and stable brain connectivity during movie watching as revealed by functional MRI. bioRxiv. DOI
Di and colleagues evaluate the intersubject consistency of dynamic connectivity during several naturalistic movies, and demonstrate that this consistency varies across movies (more so than static connectivity). This is an interesting set of results, but the terminology is sometimes hard to follow and I have a few questions about how this approach relates to similar methods. I provide some high-level comments on the approach and terminology, followed by a few minor comments/typos.
Major comments:
On my first glance through the manuscript, I was confused about how “dynamic connectivity” and “intersubject correlation”/“inter-individual consistency” implemented here are related to intersubject functional connectivity (ISFC) analysis (Simony et al., 2016). Here the authors are computing within-subject dynamic connectivity between a pair of regions or networks, resulting in a cofluctuation time series for each subject for a given pair of networks. They aggregate these within-subject cofluctuation time series across subjects and apply PCA; the variance accounted for by the first PC then serves as a measure of how consistent these within-subject cofluctuation time series are across subjects. A related, but different, approach would be to directly compute the cofluctuation across subjects—i.e. the compute the cofluctuation between the response time series for a pair of networks A and B between subjects 1 and 2 (or between subject 1 and the mean of N – 1 subjects). This would be the “dynamic” analog of the typical “static” ISFC. If computed in leave-one-out fashion, this would result in a n_timepoints x n_networks x n_networks dynamic ISFC array per subject; if you averaged across the network-by-network matrices across time points, you would recover the traditional static ISFC matrix. I’m not asking the authors to run this sort of dynamic ISFC analysis (although I’m curious how it would compare), but I think the distinction between these two related approaches needs to be treated carefully. For example, in the abstract, the authors say they use “inter-individual correlation,” but as far as I can tell, the correlations (both dynamic cofluctuations and static connectivity) are computed within participants (then the within-subject cofluctuations are aggregated across subjects and PCA is applied). Another example: at line 43 (page 2), you mention how intersubject analyses capture stimulus-driven information—but I’m not sure this “filtering” effect of ISC/ISFC applies in the same way when correlations are computed within subjects.
In a similar vein, the authors find that the static functional connectivity matrices are highly similar across movies (Figure 4). But I suspect that this is because these connectivity matrices are computed within subjects, and therefore do not isolate stimulus-driven information. Based on the results from Simony et al., 2016 (Figure 3), I suspect these connectivity matrices would be better differentiated if they were computed using ISFC analysis.
The authors should make clear in the the abstract, introduction, etc, that this paper is about the inter-subject consistency of dynamic connectivity. For example, in the abstract you say “a court drama clip showed higher dynamic connectivity with the auditory cortex and temporoparietal junction,” but it’s not clear what “higher dynamic connectivity” means. I think what you really mean here is “more consistent dynamic connectivity across subjects”—right? Similarly, the authors should be clear up front that they’re not really going to visualize the actual dynamics (i.e. cofluctuation time series) or relate points in these time series to the contents of the movies. The analysis computes dynamic connectivity, but then collapses the dynamics back down into a measure of consistency using PCA. It might be worth plotting some more cofluctuation time series alongside the consistency matrices, just so readers understand where this consistency metric is coming from.
At line 183 (page 8), you introduce the term “time-by-time interactions” to refer to the “unwrapped” correlation coefficient from the referenced article by Faskowitz et al., 2020 (the product of two z-scored vectors prior to the averaging step of the typical correlation). The term “interaction” has a lot of baggage and seems suboptimal here. I understand that there’s some variability in how people refer to this metric for time-resolved dynamic connectivity (e.g. I’ve seen “instantaneous correlation” as well), but I would encourage the authors to go with the term “cofluctuation” used in the Faskowitz paper (as well as a related paper by Esfahlani et al., 2020).
Is there a precedent in the literature for using the term “stable” to refer to traditional connectivity computed over the entire time series in contrast to dynamic functional connectivity? For example, Hutchinson et al., 2013, Zhang et al., 2018, and Lurie et al., 2020, use the term “static functional connectivity” instead of “stable.” The term “static” seems more appropriate here.
I didn’t fully understand the classification analysis. Where do the number of features 1431 and 54 come from? When classifying movies based on dynamic connectivity, is the input to the classifier the actual dynamic connectivities (i.e. cofluctuation time series) or just the consistency of dynamic connectivity? In the latter case, I’m not sure how to interpret the classification analysis. Might also be worth making clear that this is a 3-way classification where (theoretical) chance would be ~33%. Finally, I understand this is a very simple matching classifier, but classifying only 9 samples of 1431-dimensional vectors should be interpreted with caution.
Minor comments:
In figures 2–5, I think it would be worthwhile to more explicitly label the color bars; e.g. “consistency (PC1 VAF)”.
In the discussion, the authors should be careful to not go overboard with reverse inference (Poldrack, 2006); for example, the paragraph starting at line 421 (page 20).
In Figure S1 in the supplementary material, I believe the x-axes should be labeled “independent components” instead of “principal components”—is this correct? My understanding was that the figure shows for each IC(!) how much variance across subjects is accounted for by the first PC. When I first read the figure as PCs along the x-axis, I was very confused as to why earlier PCs did not account for more variance than later PCs.
Page 1, line 9: “higher consistent” > “more consistent”
Page 2, line 31: “functional meaningful” > “functionally meaningful” or “meaningful functional”
Page 2, line 35: “The time-varying” > “Time-varying”
Page 4, line 84: “Partly cloudy” > “Partly Cloudy”
Page 6, line 141: “striped” > “stripped”
Line 328: “few” > “Few”
Line 397: “carton” > “cartoon”
Line 429: “virtue” > “virtual”
References:
Esfahlani, F. Z., Jo, Y., Faskowitz, J., Byrge, L., Kennedy, D. P., Sporns, O., & Betzel, R. F. (2020). High-amplitude cofluctuations in cortical activity drive functional connectivity. Proceedings of the National Academy of Sciences, 117(45), 28393–28401. DOI
Poldrack, R. A. (2006). Can cognitive processes be inferred from neuroimaging data? Trends in Cognitive Sciences, 10(2), 59-63. DOI
Simony, E., Honey, C. J., Chen, J., Lositsky, O., Yeshurun, Y., Wiesel, A., & Hasson, U. (2016). Dynamic reconfiguration of the default mode network during narrative comprehension. Nature Communications, 7, 12141. DOI