An improved neuroanatomical model of the default-mode network reconciles previous neuroimaging and neuropathological findings

The brain is constituted of multiple networks of functionally correlated brain areas, out of which the default-mode network (DMN) is the largest. Most existing research into the DMN has taken a corticocentric approach. Despite its resemblance with the unitary model of the limbic system, the contribution of subcortical structures to the DMN may be underappreciated. Here, we propose a more comprehensive neuroanatomical model of the DMN including subcortical structures such as the basal forebrain, cholinergic nuclei, anterior and mediodorsal thalamic nuclei. Additionally, tractography of diffusion-weighted imaging was employed to explore the structural connectivity, which revealed that the thalamus and basal forebrain are of central importance for the functioning of the DMN. The contribution of these neurochemically diverse brain nuclei reconciles previous neuroimaging with neuropathological findings in diseased brains and offers the potential for identifying a conserved homologue of the DMN in other mammalian species.

I have one major and several minor comments.

Major:
The definition of regions of interest was globally lacking transparency and/or clarity: o For the DMN individual maps, the authors cited Gordon et al 2016 to justify the selection of seeds. By quickly having a look at this paper, it is not obvious for me what are the 20 seeds. The author should provide a map (or a table at least) to show them. o In Figure 2e, we see the 3 "hypothesized" subcortical regions. For consistency across figures, the authors could represent those three structures in a more complete diagram in Figure 1c.
o The selection of region of interest for the functional connectivity comparison is not explained. I believe it has been done in the same way as the description lines 250-255 in the tractography section. However, this last description is not clear: are they defined according to the literature (Andrews-Hanna et al., 2010;Buckner et al., 2008) or based on the results? ("The map of the DMN created through functional alignment was used to define the regions of interest"). None of this selection is a problem for the tractography analysis but the latter might be a problem for the functional connectivity comparison. Indeed, if we select ROI based on the DMN aligned map, then it seems logical to find higher correlations in the functional space. If the ROI selection had been done in the structural space, isn't it exactly the opposite result that we would observe? Except if I miss something here I am afraid it could be some double dipping. Could the authors show the same analysis but with ROI based in the structural space? Or explain how it is not double dipping? Why not taking the 40 seeds of interest from Gordon et al 2016? o For the ROI defined according to the DMN functionally aligned map, with which statistical threshold? Line 329, it is not clear how the 9 additional regions are defined (regions that are found in the functionally aligned but not on the structurally aligned?). Minor: 1. In the introduction, the authors could mention Glasser et al (2016), in which a new method of brain data alignment is proposed, based on multiple brain measures such as functional connectivity, sulcal depth of topography. This study focuses on cortical areas (and on surfaces) but has proven how efficient a functional alignment could be to investigate brain networks. I think it supports even more the need for a similar work on subcortical regions.
3.5 The graph theory part of the analyses is of interest and should be expanded. The corresponding figure is illegible, please correct. The path length/connectedness parameters should be shown more clearly for different structures (and in individual subjects), to convince readers of the high centrality of thalamus and BF DMN nodes.
p.12: It would be useful to highlight briefly other studies that have used functional alignment, and discuss the limitations / advantages of this method. Two or three specific examples from other systems neuroscience areas would be useful for illustration.
p.14: It should be mentioned that the BF also contains GABAergic and glutamatergic populations of neurons with diverse projection patterns in addition to the cholinergic cells.
p.15: The considerations of the role of DMN in various disorders is a little too superficial to be useful. Can the authors focus on the putative specific role of the DMN, for example hypo-or hyperactivation in specific disorders mentioned? p.15: I would also encourage the authors to speculate why BF functional activations (gamma band LFP) are strongly coupled to the medial frontal cortex (i.e. cingulate) in the rat? It seems like the Cingulate is not a major target of the projection in humans? What is the most likely avenue for this BF-ACC coupling given the present data? Some thoughts on this would be appreciated.
Reviewer #3 (Remarks to the Author): Title: Subcortical Anatomy of the Default Mode Network: a functional and structural connectivity study

GENERAL COMMENTS
The authors present a study in which they attempt to demonstrate the feasibility of including several subcortical areas --including basal forebrain, anterior and mediodorsal thalamic nuclei, and cholinergic nuclei --in the definition of the "default mode network" (DMN). They do this on the basis of common functional covariance and deterministic DTI tractography observed in these brain regions.
In general, the approach is a novel attempt to improve our understanding of DMN function, by including subcortical regions that have traditionally not been considered in such analyses. The move to areas beyond the neocortex is a welcome one. It also appears promising to use functional rather than structural coregistration to improve the precision with which the subcortical regions of interest can be localized (but these conclusions have issues; see below). Overall, this study is a nice idea and with some work could be an important addition to the scientific discussion about what the DMN is, and the neural/mental processes it subserves.
I have some criticisms of the approaches used, however, which are detailed below. These should be addressed either by clarification of the methods, re-analysis of the data, or a discussion of the potential pitfalls of the methods where the authors are not inclined to change them (in particular, I am not convinced by the DTI approach).
Additionally, a number of important details are missing from the Methods section, which make it difficult to either reproduce the results reported, properly interpret them, or assess the degree to which they may be accounted for by confounds or statistical artifacts. Specific comments follow.

METHODS [line numbers in square brackets]
1.
[178] "Resting state functional MRI images were corrected for artefacts" -Which artifacts, what steps were followed? You need to provide detail here, it's not sufficient to say you passed data through a pipeline, without specifying which steps were taken. It is especially important to explain how you dealt with motion correction. Was a global signal regression done? A censoring approach? Removal of temoral derivatives? Because you use functional data to co-align subjects and then find increased functional correlations in the same sample, how do we know this improvement is not due to motion artifacts, which have been demonstrated to produce systematic spurious correlations and changes to apparent topological organization (e.g., Power et al., Neuroimg, 2012 [197] "To obtain a group DMN map in the structural space, the median of the twenty individual mpas was derived." --Not sure what this means. What is "the structural space"? It becomes clearer later in the text, but you should introduce these terms early and not assume the reader knows what they mean. Also should be more clear that this is a "group median map".

5.
[213] "different regions of interest identified in the DMN maps" --how were these ROIs derived? Was thresholding applied, and if so what was the rationale? Were time-series averaged or was each voxel correlated separately? 6.
[231] "A percentage of volume overlap between the DMN map and each nucleus of interest" --Were the volumes comparable? What is this quantity meant to show? How dependent is this quantity on the threshold used to define the DMN map (which hasn't been specified)? The rationale for thresholding becomes even more important since you are using these values to draw inferences...

[237]
Why use a deterministic tractography approach instead of probabilistic? How do you account for variability in streamline distribution? Does this method account for crossing fibres or different anisotropy profiles generally? 9.
[250] "Regions of interest were defined according to the previous anatomical models of the DMN" --It should be explained how these were derived, what their granularity is (and how many there are), why they were chosen over other alternatives, etc. More detail on these ROIs is required as they are a critical aspect of this method. 10.
[256] "In order to have a group-representative tract volume for each tract, one sample T-test of individual tract volumes was calculated using FSL randomise" --How are "tract volumes" computed? What does this computation depend on? What does it represent in biological terms? 12.
[271] "The matrices were binarised (no threshold) for structural connection or no structural connection between regions." --What does "no threshold" mean? On what basis was this binarization performed? Since you are not using probabilistic tractography, and have no estimate of uncertainty how are you confident this is the "correct" connectome (i.e., no false negatives)? Particularly important since you only have 20 subjects. Would adding more subjects change this connectome (and its topological characterization)?

RESULTS
13. Fig 2d: Higher FC for functionally aligned maps: You use the same subjects to functionally co-align as to compute FC. FC is used as the feature for coalignment. Doesn't it seem trivial (i.e., doubledipping) then that you find stronger FC for the functional versus structural maps? It would be more convincing had you shown this for an independent sample. This should be discussed. 14.
[305] "statistically significant in 95% and in 18% of pairs, whether or not correcting and correcting for multiple comparisons was applied" --If you correct for multiple comparisons (which you should), there's no need to report uncorrected statistics. [Also, the sentence is poorly constructed]. Table 2 represent the MNI coordinates of all the regions of interest in the DMN map" --Should specify that these coordinates are the center of mass of the ROI.

[306] "
16. Fig 3b is confusing because it is not clear whether you are showing partial correlation coefficients or differences in these coefficients between structural and functional maps. The significance levels appear to refer to the latter, but the caption indicates these are coefficients only for the functional map. Why use asterisks, instead of changing the edge colours to indicate significance? Why show uncorrected significance, when you clearly need to apply a FWE correction here? Also, the labels are not all readable.
17. Fig 4: The colour map here has red/orange on both ends of the scale, which makes interpretation difficult. This should be fixed to have unique (preferably hot/cool) colours for each end of the scale. Also, units would make more sense as percentages, especially since these are discussed in the corresponding text.
18. The percent overlap metric relies on a threshold, and how this threshold was determined is unclear (statistical significance at 0.05? corrected/uncorrected? FWE? Cluster-level?). Instead of a (necessarily) arbitrary threshold approach, why not simply compute the mean/std correlation values for these voxels? Surely this would highlight the overlapping areas in a less binarized fashion...? 19.
[329] "nine additional regions of interest were defined including..." --how were these defined? Was a threshold set? These details are important! DISCUSSION 20.
[447] "Hence, integrating present and previous findings, it appears that the DMN, as defined by functional connectivity, is at the interplay between a cholinergic and a dopaminergic system dedicated to memory and emotion." --It is not clear how the present findings elucidate this system, and neurotransmitter systems in particular, beyond what we already knew. The methods used here do not allow the level of resolution (temporal or spatial) to permit any sort of direct inference with respect to these systems. It is not clear what "at the interplay" refers to. I'd suggest removing these speculative statements, or make it clearer that they are purely speculative and not in any way supported by the evidence presented in this study. It is important to put your work in context, but equally important to avoid unwarranted conclusions.  Figure   S4. Pearson's correlation was higher in the functional space in 92% of cells of the correlation matrix, and the difference was statistically significant in 5% of cells (p-value< 0.00004, after Bonferroni's correction for multiple comparisons). Accordingly, the following sentences were added to the methods and to the results section respectively: "Two alternative correlation matrices using cortical DMN areas according to Gordon et al. (2016), instead of the ones defined in classical anatomical models of Andrews-Hanna et al. (2010) and Buckner et al. according to Gordon et al. (2016) are presented in the Supplementary Materials ( Figure   S4)." We agree that demonstrating stronger functional connectivity after performing functional alignment may bear some aspect of circularity. However, we originally performed this analysis to demonstrate that the method of functional alignment was valid at the subcortical level. We have now clarified this point in the limitations section of the discussion: "Furthermore, demonstrating that the additional step of functional alignment results in higher correlation values is somewhat circular. However, our purpose was to demonstrate that this increase in correlation revealed subcortical structures that were previously R: We set the statistical threshold at a Pearson correlation of r=0.3 to define the ROIs as it corresponds to a medium effect size (Cohen, 1992). This point was clarified in the methods: "Functional and structural-based DMN maps were thresholded at r=0.3 which, corresponds to a medium effect size (Cohen, 1992)." Regarding the nine additional regions, they were defined as mentioned in the commentary. R: We now added this reference to the manuscript. "The thalamus has also been shown to be structurally and functionally connected to DMN regions (Cunningham et al., 2016;Fransson, 2005)." Concerning the partial correlations of the basal forebrain, we did not find any areas with statistically significant stronger or weaker connectivity. Our sample may not have enough power to detect differences. Tables with the median, and with the range and interquartile range of the partial correlation between all the areas is now available in the Supplementary Materials -Tables S3 and S4. In these tables, the partial correlations for the basal forebrain and for the thalamus were highlighted. In addition, we included a table with the p-values of the comparison of partial correlation between the anatomical and functional alignment method (Supplementary Materials - Table S5).
Regarding Pearson's correlations, the left basal forebrain had significantly higher correlations with the right antero-median prefrontal cortex, with the right posterior parietal cortex and with the midbrain area, while the right basal forebrain had significantly higher correlations with the right temporal pole and with the left cerebellar hemisphere. A table with the p-values of the comparison of Pearson's correlations between the two methods of alignments was also included in the Supplementary Materials (Table S2).
To show the correspondence of these results at individual level, we added a  (Tables S3, S4, (Table 4). In addition, this table was appended with the interquartile range of node degrees and betweenness centrality of each node in order to provide a measure of the distribution between individual subjects.
Individual examples were also included in the Supplementary Materials (Table S7) with graph theory measures of subject 1 and 2 (the same subjects that were referred in reviewer's commentaries in A and B).  (Glasser et al., 2016;Langs et al., 2015)." Additionally, to state the limitations of functional alignment, the following text was included: "The intra-individual functional connectivity variation across time may also be a source of bias (Braga and Buckner, 2017). However, this factor and the lack of uniformity regarding the optimal functional alignment method have not precluded the achievement of higher accurate results in previous studies (Glasser et al., 2016;Langs et al., 2015;Mueller et al., 2013;Robinson et al., 2014)." We appreciate the suggestion to mention similar approaches from other fields of neuroscience. The following sentences and references were added regarding the consistency of our results with the mouse model: "For instance, the medial thalamus is part of mouse's DMN (Gozzi and Schwarz, 2016), as found in rs-fMRI studies (Bertero et al., 2018;Liska et al., 2015;Sforazzini et al., 2014 (Markello et al., 2018). We speculate that the putative differences in the basal forebrain projections may be one explanation for the more diffuse DMN activation at the midline level in rats, when compared to humans (Lu et al., 2012)."

G) p.14: It should be mentioned that the BF also contains GABAergic and glutamatergic populations of neurons with diverse projection patterns in addition to the cholinergic cells.
R: The following sentence was added in the discussion: "The basal forebrain also contains GABAergic and glutamatergic neurons that mediate hippocampal theta synchronization through an indirect septo-hippocampal pathway (Dannenberg et al., 2015)."

H) p.15: The considerations of the role of DMN in various disorders is a little too superficial to be useful. Can the authors focus on the putative specific role of the DMN, for example hypo-or hyperactivation in specific disorders mentioned?
R: We thank the reviewer for the suggestion. We now detail the altered connectivity and activations mentioned with respect to diseases in the text and in Table 5.
To complement the presented information, we also added a new references reviewing this  ICA-Aroma (Pruim et al., 2015). This method performs an ICA decomposition of the data and estimates which components reflect motion-related noise in the fMRI signal on the basis of a robust set of spatial and temporal features. This is made possible due the distinctiveness of the motion-related components isolated by ICA on the fMRI signal (Salimi-khorshidi et al., 2014). This approach outperforms other methods such as the regression of the motion parameter estimates, while limiting in the same time the loss in degrees of freedom (Pruim et al., 2015). Compared to spike removal methods such as scrubbing (Power et al., 2012), ICA-Aroma has the advantage of preserving the temporal structure of the fMRI signal."

B) [180] Was normalization linear or also nonlinear?
R: The normalization was linear (affine) and nonlinear (diffeomorphic). This information is now clarified in the text.  (Table S1). In addition a surface map representing these ROIs was created (Supplementary Materials - Figure S1). The following sentence was included:

C) Section 2.2.1. is poorly written. It could be rewritten to provide a step-by-
"The DMN nodes as defined by Gordon et al. are

available in the Supplementary Materials
( Figure S1 and Table S1, respectively)" We also completed the following section to better describe the method used to obtain DMN distributions: "These 40 regions were used as seeds. The correlation map for each seed was obtained using the Funcon-Connectivity tool of the Brain Connectivity and Behaviour toolkit (http://toolkit.bcblab.com;Foulon et al., 2018). This tool calculates the Pearson's correlation between the mean rs-fMRI time-course of the seed with the rest of the brain to generate the functional connectivity map for every seed (Foulon et al., 2018). The DMN map of each participant was obtained by calculating the median of the 40 seed-based correlation maps of each individual using FSL (Jenkinson et al., 2012)." The median was chosen as the centrality measure, instead of the mean, because it is less sensitive to outliers. This is especially relevant in small samples sizes (40 maps in this case).
The following sentence was included: "The median was used, instead of the mean, because it is less sensitive to outliers, being a better centrality measure for small sample sizes (Kenney, 1939)." R: To specify how the ROIs were derived, the following text was added: "Cortical regions of interest were defined according to the previous anatomical models of the DMN (Andrews-Hanna et al., 2010;Buckner et al., 2008). Subcortical regions of interest were defined manually based on the experienced judgment of the neuroanatomists among the authors combined with the careful comparison with previously published atlases (Catani and Thiebaut de Schotten, 2012;Nieuwenhuys et al., 2008)." In addition, a map with these ROIs was now made available in the Supplementary Materials ( Figure S2).

D) [197] "To obtain a group DMN map in
We applied a threshold of 0.3 because it corresponds to a medium effect size. The rationale for this choice was now added to this section: "For the functional and structuralbased DMN maps, a prespecified threshold of 0.3 was applied. This value was chosen because it corresponds to a medium effect size (Cohen, 1992)." For the analysis, the time-series were averaged. We now clarified this point in the text: R: This corrected p-value corresponds to 528 comparisons. Since we have 33 regions of interest, the number of correlations was 33x32/2.

G) [231] "A percentage of volume overlap between the DMN map and each nucleus of interest" --Were the volumes comparable? What is this quantity meant to show?
How dependent is this quantity on the threshold used to define the DMN map (which hasn't been specified)? The rationale for thresholding becomes even more important since you are using these values to draw inferences...

R:
The threshold used was the same and for the same reasons specified in E). We have chosen to present percentage to account for the differences in the volumes of the nuclei. For the purpose of transparency, an additional column was inserted in the Table 3 with the absolute volumes of intersection. The legend of this table was revised accordingly:

"Proportion of individual DMN maps intersecting, average proportion of the intersection and average volume of the intersection for each nucleus."
In addition, the density map of proportion of individuals overlapping the hypothesized regions presented in Figure 4 was made available at neurovault.org To clarify this aspect, the following paragraph was reformulated: "Cortical regions of interest were defined according to previous anatomical models of the DMN (Andrews-Hanna et al., 2010;Buckner et al., 2008). Subcortical regions of interest were defined manually based on the experienced judgment of the neuroanatomists among the authors combined with the careful comparison with previously published atlases (Nieuwenhuys et al. 2008;Catani and Thiebaut de Schotten 2012)." In addition, a map with these ROIs was now made available in the Supplementary Materials ( Figure S2).
The threshold used for the anatomical delimitation was a Pearson correlation of 0.3, because it corresponds to a medium effect size. This point was now clarified accordingly in the text: "For the functional and structural-based DMN maps, a prespecified threshold of 0.3 was applied. This value was chosen because it corresponds to a medium effect size (Cohen, 1992).  Figure S4. Accordingly, the following sentences were added to the methods and to the results section respectively: "Two alternative correlation matrices using cortical DMN areas according to Gordon et al. (2016), instead of the ones defined in classical anatomical models of Andrews-Hanna et al. (2010) andBuckner et al. (2008), were calculated.". "The alternative correlation matrices using cortical DMN areas according to Gordon et al. (2016)

2011)."
As previously stated in commentary H), we also added statistical maps of this calculation in the Supplementary Materials - Figure S7.
In biological terms, it has been shown that the described method to create grouprepresentative tract maps has a good anatomical correspondence, when compared to the histological atlas of white matter tracts. The sentence "The resulting group-representative tract maps using this method were shown to have a good anatomical correspondence with histological atlases of white matter tracts (Bürgel et al., 2006;Thiebaut de Schotten et al., 2011b)." was added to the manuscript. To better explain this point, the following text was changed and the corresponding references included: "The matrices were binarised depending on the existence or absence of streamlines connecting two regions of interest (Gong et al., 2009;Shu et al., 2011). The defined threshold for binarisation was 1 because the number of streamlines does not reflect the connectivity strength or the true number of axonal projections between two brain regions (Gong et al., 2009;Jones et al., 2013), and previous evidence has shown that changing the streamline count threshold for binarisation (between 1 and 5) does not change the overall results of the network analysis." We are aware of the limitations of tractography analysis, and have further detailed the potential caveats of the tractography analysis in the limitation section. R: The sentence was changed to: " Table 2  The font size of the labels was also increased. We also detailed how the subcortical regions were defined: "Subcortical regions of interest were defined manually based on the experienced judgment of the neuroanatomists among the authors combined with the careful comparison with previously published atlases (Catani and Thiebaut de Schotten, 2012;Nieuwenhuys et al., 2008)."

M) Fig
The threshold was a Pearson's correlation of 0.3: "For the functional and structural-based DMN maps, a prespecified threshold of 0.3 was applied. This value was chosen because it corresponds to a medium effect size (Cohen, 1992)." T) [447] "Hence, integrating present and previous findings, it appears that the DMN, as defined by functional connectivity, is at the interplay between a cholinergic and a dopaminergic system dedicated to memory and emotion." --It is not clear how the present findings elucidate this system, and neurotransmitter systems in particular, beyond what we already knew. The methods used here do not allow the level of resolution (temporal or spatial) to permit any sort of direct inference with respect to these systems. It is not clear what "at the interplay" refers to. I'd suggest removing these speculative statements, or make it clearer that they are purely speculative and not in any way supported by the evidence presented in this study. It is important to put your work in context, but equally important to avoid unwarranted conclusions.
R: We completely agree with the reviewer. The point made is purely speculative and it was mentioned only to put our research in context and to foster future research.
Therefore, the paragraph was reformulated to reinforce its speculative nature: "Hence, combining present and previous findings, we speculate that the DMN, as defined by functional connectivity, might have a putative role in the integration of cholinergic and dopaminergic systems dedicated to memory and emotion."