Intrinsic dynamic shapes responses to external stimulation in the human brain

Reviewer #1 (Public review):

This manuscript presents an interesting new framework (VARX) for simultaneously quantifying effective connectivity in brain activity during sensory stimulation and how that brain activity is being driven by that sensory stimulation. The core idea is to combine the Vector Autoregressive model that is often used to infer Granger-causal connectivity in brain data with an encoding model that maps the features of a sensory stimulus to that brain data. The authors do a nice job of explaining the framework. And then they demonstrate its utility through some simulations and some analysis of real intracranial EEG data recorded from subjects as they watched movies. They infer from their analyses that the functional connectivity in these brain recordings is essentially unaltered during movie watching, that accounting for the driving movie stimulus can protect one against misidentifying brain responses to the stimulus as functional connectivity, and that recurrent brain activity enhances and prolongs the putative neural responses to a stimulus.

This manuscript presents an interesting new framework (VARX) for simultaneously quantifying effective connectivity in brain activity during sensory stimulation and how that brain activity is being driven by that sensory stimulation. Overall, I thought this was an interesting manuscript with some rich and intriguing ideas. That said, I had some concerns also - one potentially major - with the inferences drawn by the authors on the analyses that they carried out.

Main comments:

(1) My primary concern with the way the manuscript is written right now relates to the inferences that can be drawn from the framework. In particular, the authors want to assert that, by incorporating an encoding model into their framework, they can do a better job of accounting for correlated stimulus-driven activity in different brain regions, allowing them to get a clearer view of the underlying innate functional connectivity of the brain. Indeed, the authors say that they want to ask "whether, after removing stimulus-induced correlations, the intrinsic dynamic itself is preserved". This seems a very attractive idea indeed. However, it seems to hinge critically on the idea of fitting an encoding model that fully explains all of the stimulus-driven activity. In other words, if one fits an encoding model that only explains some of the stimulus-driven response, then the rest of the stimulus-driven response still remains in the data and will be correlated across brain regions and will appear as functional connectivity in the ongoing brain dynamics - according to this framework. This residual activity would thus be misinterpreted. In the present work, the authors parameterize their stimulus using fixation onsets, film cuts, and the audio envelope. All of these features seem reasonable and valid. However, they surely do not come close to capturing the full richness of the stimuli, and, as such, there is surely a substantial amount of stimulus-driven brain activity that is not being accounted for by their "B" model and that is being absorbed into their "A" model and misinterpreted as intrinsic connectivity. This seems to me to be a major limitation of the framework. Indeed, the authors flag this concern themselves by (briefly) raising the issue in the first paragraph of their caveats section. But I think it warrants much more attention and discussion.

(2) Related to the previous comment, the authors make what seems to me to be a complex and important point on page 6 (of the pdf). Specifically, they say "Note that the extrinsic effects captured with filters B are specific (every stimulus dimension has a specific effect on each brain area), whereas the endogenous dynamic propagates this initial effect to all connected brain areas via matrix A, effectively mixing and adding the responses of all stimulus dimensions. Therefore, this factorization separates stimulus-specific effects from the shared endogenous dynamic." It seems to me that the interpretation of the filter B (which is analogous to the "TRF") for the envelope, say, will be affected by the fact that the matrix A is likely going to be influenced by all sorts of other stimulus features that are not included in the model. In other words, residual stimulus-driven correlations that are captured in A might also distort what is going on in B, perhaps. So, again, I worry about interpreting the framework unless one can guarantee a near-perfect encoding model that can fully account for the stimulus-driven activity. I'd love to hear the authors' thoughts on this. (On this issue - the word "dominates" on page 12 seems very strong.)

(3) Regarding the interpretation of the analysis of connectivity between movies and rest... that concludes that the intrinsic connectivity pattern doesn't really differ. This is interesting. But it seems worth flagging that this analysis doesn't really account for the specific dynamics in the network that could differ quite substantially between movie watching and rest, right? At the moment, it is all correlational. But the dynamics within the network could be very different between stimulation and rest I would have thought.

(4) I didn't really understand the point of comparing the VARX connectivity estimate with the spare-inverse covariance method (Figure 2D). What was the point of this? What is a reader supposed to appreciate from it about the validity or otherwise of the VARX approach?

(5) I think the VARX model section could have benefitted a bit from putting some dimensions on some of the variables. In particular, I struggled a little to appreciate the dimensionality of A. I am assuming it has to involve both time lags AND electrode channels so that you can infer Granger causality (by including time) between channels. Including a bit more detail on the dimensionality and shape of A might be helpful for others who want to implement the VARX model.

(6) A second issue I had with the inferences drawn by the authors was a difficulty in reconciling certain statements in the manuscript. For example, in the abstract, the authors write "We find that the recurrent connectivity during rest is largely unaltered during movie watching." And they also write that "Failing to account for ... exogenous inputs, leads to spurious connections in the intrinsic "connectivity".

Reviewer #2 (Public review):

Summary:

The authors apply the recently developed VARX model, which explicitly models intrinsic dynamics and the effect of extrinsic inputs, to simulated data and intracranial EEG recordings. This method provides a directed method of 'intrinsic connectivity'. They argue this model is better suited to the analysis of task neuroimaging data because it separates the intrinsic and extrinsic activity. They show: that intrinsic connectivity is largely unaltered during a movie-watching task compared to eyes open rest; intrinsic noise is reduced in the task; and there is intrinsic directed connectivity from sensory to higher-order brain areas.

Strengths:

(1) The paper tackles an important issue with an appropriate method.

(2) The authors validated their method on data simulated with a neural mass model.

(3) They use intracranial EEG, which provides a direct measure of neuronal activity.

(4) Code is made publicly available and the paper is written well.

Weaknesses:

It is unclear whether a linear model is adequate to describe brain data. To the author's credit, they discuss this in the manuscript. Also, the model presented still provides a useful and computationally efficient method for studying brain data - no model is 'the truth'.

Appraisal of whether the authors achieve their aims:

As a methodological advancement highlighting a limitation of existing approaches and presenting a new model to overcome it, the authors achieve their aim. Generally, the claims/conclusions are supported by the results.

The wider neuroscience claims regarding the role of intrinsic dynamics and external inputs in affecting brain data could benefit from further replication with another independent dataset and in a variety of tasks - but I understand if the authors wanted to focus on the method rather than the neuroscientific claims in this manuscript.

Impact:

The authors propose a useful new approach that solves an important problem in the analysis of task neuroimaging data. I believe the work can have a significant impact on the field.