Brain dysconnectivity relates to disability and cognitive impairment in multiple sclerosis

Abstract The pathophysiology of cognitive dysfunction in multiple sclerosis (MS) is still unclear. This magnetoencephalography (MEG) study investigates the impact of MS on brain resting‐state functional connectivity (rsFC) and its relationship to disability and cognitive impairment. We investigated rsFC based on power envelope correlation within and between different frequency bands, in a large cohort of participants consisting of 99 MS patients and 47 healthy subjects. Correlations were investigated between rsFC and outcomes on disability, disease duration and 7 neuropsychological scores within each group, while stringently correcting for multiple comparisons and possible confounding factors. Specific dysconnections correlating with MS‐induced physical disability and disease duration were found within the sensorimotor and language networks, respectively. Global network‐level reductions in within‐ and cross‐network rsFC were observed in the default‐mode network. Healthy subjects and patients significantly differed in their scores on cognitive fatigue and verbal fluency. Healthy subjects and patients showed different correlation patterns between rsFC and cognitive fatigue or verbal fluency, both of which involved a shift in patients from the posterior default‐mode network to the language network. Introducing electrophysiological rsFC in a regression model of verbal fluency and cognitive fatigue in MS patients significantly increased the explained variance compared to a regression limited to structural MRI markers (relative thalamic volume and lesion load). This MEG study demonstrates that MS induces distinct changes in the resting‐state functional brain architecture that relate to disability, disease duration and specific cognitive functioning alterations. It highlights the potential value of electrophysiological intrinsic rsFC for monitoring the cognitive impairment in patients with MS.


| INTRODUCTION
Multiple sclerosis (MS) is a chronic, autoimmune, inflammatory, demyelinating but also degenerative disorder that affects both the white and gray matters of the CNS (for reviews, see, e.g., Compston & Coles, 2008;Geurts, Calabrese, Fisher, & Rudick, 2012;Ciccarelli et al., 2014). MS is the leading cause of nontraumatic neurological disability in young adults, especially in women.
More than 50% of patients with MS are also affected by cognitive impairments (CIs) and fatigue. CIs are mainly characterized by alterations in executive, attentional and memory functions, and are encountered throughout all disease stages (for a review, see, e.g., Chiaravalloti & DeLuca, 2008). Fatigue is defined as a lack of motivation, an overall feeling of exhaustion and behavioral performance decrements. It is divided into motor, psychosocial, and cognitive fatigue (for a review, see, e.g., Linnhoff, Fiene, Heinze, & Zaehle, 2019). Both CI and fatigue are important contributors to employment status, quality of life, and social functioning in patients with MS. Pharmacological and rehabilitation strategies are currently insufficient to alleviate these symptoms and require the development of novel therapeutic approaches (for a review, see, e.g., Benedict & Zivadinov, 2011). Providing a better understanding of the mechanisms involved in CIs and fatigue in MS is therefore of major importance (Di Filippo, Portaccio, Mancini, & Calabresi, 2018).
Although structural neuroimaging based on, for example, cerebral MRI, has been used extensively in the diagnosis and monitoring of MS (Polman et al., 2011), it has failed at explaining the degree and the variety of CIs observed in this disorder (Mollison et al., 2017). Apart from a link with gray matter atrophy, only weak associations have indeed been reproducibly found between structural MRI parameters and CIs/fatigue (Andreasen et al., 2019). Functional neuroimaging therefore offers a unique opportunity to better understand the pathophysiology of cognitive and fatigue symptoms in MS (for a review, see Van Schependom & Nagels, 2017).
MS has traditionally been considered as a disease affecting white matter tracts forming the structural connections between CNS gray matter structures (Compston & Coles, 2008). MS-related gray matter involvement has also been clearly established (Mandolesi et al., 2015).
To characterize the functional changes that accompany MS-related alterations in structural white matter connectivity and gray matter lesions, imaging functional brain connectivity appears highly relevant to better understand the brain-behavior relationship in this major brain disorder (Di Filippo et al., 2018). In a clinically heterogeneous disorder like MS, investigating functional brain connectivity at rest (i.e., in the absence of any goal directed task) has some key advantages over task-based studies, that is, it is free of any performance bias and requires minimal patient cooperation, no task-related training beforehand, and no complex experimental paradigm. Furthermore, previous studies have demonstrated a strong anatomical correspondence between task-based and resting-state functional connectivity (rsFC) (Cole, Bassett, Power, Braver, & Petersen, 2014;Mennes, Kelly, Colcombe, Xavier Castellanos, & Milham, 2013).
Functional MRI (fMRI) is the most widely used technique to investigate rsFC both in healthy subjects and patients with brain disorders.
While these results are valuable, MS-related alterations in cerebrovascular reactivity might impact the neurovascular coupling at the basis of the fMRI signal (Marshall et al., 2014), potentially limiting the usefulness of fMRI in this disorder. Its low temporal resolution also precludes the study of neural oscillations, which support short-and long-range functional brain connectivity and underlie a wide range of cognitive functions (for a review, see, e.g., Siegel, Donner, & Engel, 2012). Furthermore, fMRI failed to demonstrate alterations in neural network organization in patients with early MS, while electrophysiological investigations did in the same patients (Tewarie et al., 2015). For these reasons, investigations of electrophysiological, spectrally-resolved rsFC with magnetoencephalography (MEG) or electroencephalography (EEG) have become increasingly popular (Hall, Robson, Morris, & Brookes, 2014). Previous EEG (Gschwind et al., 2016;Leocani et al., 2000;Van Schependom et al., 2014) and MEG (Cover et al., 2006;Schoonheim et al., 2013;Tewarie et al., 2013Tewarie et al., , 2015 studies have investigated rsFC alterations in patients with MS using phase-based measures (i.e., synchronization between neural populations as assessed through the time-delayed correlation of their oscillations or closely-related measures, see Figure 1). Although a few studies suggested that phase-based rsFC is related to fMRI-based RSNs Vidaurre et al., 2018), their main electrophysiological correlate is rsFC based on band-limited power envelope correlation, that is, synchronization between neural populations as assessed through the correlation of the amplitude of their oscillations (Brookes et al., 2011;Colclough et al., 2017;de Pasquale et al., 2010;Garcés et al., 2016;Hipp, Hawellek, Corbetta, Siegel, & Engel, 2012;Hipp & Siegel, 2015;Liu, Farahibozorg, Porcaro, Wenderoth, & Mantini, 2017;Liu, Ganzetti, Wenderoth, & Mantini, 2018;Siems, Pape, Hipp, & Siegel, 2016;Tewarie et al., 2016;Wens et al., 2014;Zhang et al., 2009), see also Figures 1 and 2b. This rsFC index also has the critical advantage of being more robust on a test-retest basis than phase-based measures (Colclough et al., 2016).
Although scarcely done with phase-based rsFC , possible MS-related reorganizations of electrophysiological RSNs estimated with envelope correlation, and their relationship with individual cognitive and clinical parameters, have not been assessed per se. Such investigation might prove crucial to achieve a better understanding of the brain-behavior relationship in MS as it might involve specific alterations of within-and cross-RSN interactions. This MEG study therefore investigates the impact of MS on human brain RSNs and its relationship with various factors (e.g., motor disability, disease duration, CIs, and fatigue) in a large population of patients with MS. For that purpose, we designed a comprehensive, prior-free analysis of rsFC based on a functional parcellation of the human brain into major RSNs extended with investigations of both within-and cross-frequency couplings as developed by Brookes et al. (2016). We hypothesized that MS would lead to definite alterations of rsFC within and between specific RSNs compared with matched healthy subjects, and that those changes would be associated with disability, CIs and fatigue.

| Participants
One hundred patients with MS (69 females, 31 males; age: 47.8 ± 9.8 years (mean ± SD) were recruited from the National MS Center Melsbroek with the following inclusion criteria: (i) diagnosis of MS according to the revised 2011 McDonald criteria (Polman et al., 2011), (ii) age between 18 and 60 years, (iii) disability score (Expanded Disability Status Scale; EDSS; Kurzke (1983)) ≤6.5, and (iv) no relapse or treatment with corticosteroids within the 6 weeks preceding participation to the study. Eighty-five patients had relapsing-remitting MS, while 15 patients had progressive MS. We also recruited fifty-four healthy subjects (33 females, 21 males; age: 47.5 ± 11.7 years) matched in terms of gender and age. For both groups, participants were excluded if they took recreational psychoactive drugs, had any implanted ferromagnetic materials and if they had any prior neurologic or psychiatric disorder (except MS in the patients' group). Twenty patients took benzodiazepines (such as alprazolam, clonazepam, flurazepam, lorazepam, or Triazolam) at the time of the study. Data from 7 healthy controls and 1 patient were not included in the final analyses due to quality issues with the MRI (1 patient with MS, 1 healthy subject), MEG data (4 healthy subjects) and due to being severe outliers across the cognitive tests (2 healthy subjects). Therefore, ninety-nine patients and forty-seven healthy participants were included in the final analyses. Demographic and clinical details of the final included participants are presented in Table 1.
The study was approved by the ethics committees of the Universitair Ziekenhuis Brussel (Commissie Medische Ethiek UZ Brussel, B.U.N. 143201423263, 2015/11) and the National MS Center Melsbroek (February 12, 2015). All participants gave their express written consent to participate in the study prior to their inclusion. Participants' consent was obtained according to the Declaration of Helsinki.  Penner et al., 2009), and cognitive fatigue (cognitive part of FSMC), as well as on upper extremity function (9-Hole Peg Test; 9-HPT; Mathiowetz et al., 1985). This led to a total of seven F I G U R E 1 Illustration of envelope and phase coupling. Each column shows two signals both separately (orange and blue; top and middle rows) and superimposed (bottom row). Envelope correlation (a,b): The two oscillations have correlated envelopes (black dotted curves). This can occur both when their carrying frequencies (f 1 , f 2 ) are equal (within-frequency coupling, a) or different (cross-frequency coupling, b), and independently of any phase coupling. Phase locking (c,d): The two equal-frequency oscillations exhibit a phase relationship (illustrated by vertical dotted lines). This can occur when their phases (φ 1 , φ 2 ) are equal (c) or maintain a constant difference (d) neuropsychological scores. Depressive symptoms were evaluated using the Beck Depression Inventory (BDI; Beck, Steer, & Brown, 1996). Additionally, experienced neurologists performed a standard EDSS test in patients with MS. netometer. There was no difference in acquisition parameters or MEG F I G U R E 2 Illustration of the functional connectivity pipeline. (a). Overview of the locations and labels of the 32 nodes included in the connectome, color-coded according to the network they belong to (see legend at the bottom). (b). Schematic illustration of the power envelope correlation used as rsFC measure. The envelope (red curve) of the neural oscillations at each node is used for the correlation analyses. (c). Matrix representation of all rsFC estimates across all node pairs (left) and all pairs of frequency bands, leading to the "multi-layer" rsFC matrix as described in Brookes et al. (2016) (Coquelet et al., 2020;Naeije et al., 2019).

| Data acquisition
Recordings took place at the CUB Hôpital Erasme (Brussels, Belgium).

| Source reconstruction
The MEG forward model was computed for each participant based on their MRI, which was anatomically segmented using FreeSurfer (Fischl, 2012). MEG and MRI coordinate systems were manually coregistered within the proprietary software MRIlab™ (MEGIN, Croton Healthcare, Helsinki, Finland) based on the acquired anatomical fiducials and head-surface points. A common source space (5-mm rectangular grid) was defined in the Montreal Neurological Institute (MNI) brain volume and deformed onto the participants' MRIs using a nonlinear spatial normalization scheme (Ashburner & Friston, 1999) as implemented in SPM8 (Friston, 2006). Subject-specific forward models were then computed using the single-layer boundary element method implemented in MNE-C (Gramfort et al., 2014).
Source projection of band-specific MEG signals over a grid of the whole brain volume relied on minimum norm estimation (MNE, Dale & Sereno, 1993) based on the implementation detailed in Wens et al. (2015). The noise covariance was estimated from 5 min of empty room recordings and the regularization parameter was adapted to the Lesion load (ml, mean) 9.4 7.8 0.8-34.7 Note: Mean, standard deviation (SD), and range (min-max) are given where appropriate (and just count for Gender), as well as a p value for the comparison between the two groups (age, education: unpaired Welch's t tests, gender: two-sided χ 2 test for two proportions). Education: total number of years in school since start of primary school. EDSS: Extended Disability Status Scale. Duration: disease duration, that is, number of years since initial diagnosis. RR/PP/SPMS: relapsing-remitting/primary progressive/secondary progressive multiple sclerosis.
MEG signal-to-noise ratio via the prior consistency condition derived in Wens et al. (2015). The reconstructed source time series were further projected onto their direction of maximum variance (Brookes et al., 2011;Wens et al., 2014).

| Connectivity analysis
Our pipeline for intrinsic rsFC analysis is illustrated in Figure 2. used cortical sources for our analysis). We then computed the slow envelope (i.e., Hilbert envelope low-pass filtered to 1 Hz) correlation between the band-specific source time courses of each node pair (which will be referred to as nodewise rsFC in the results; Figure 2b).
Spatial leakage was reduced prior to rsFC computation using pairwise static orthogonalization (Brookes, Woolrich, & Barnes, 2012). We did not need to use multivariate symmetrical orthogonalization (Colclough, Brookes, Smith, & Woolrich, 2015) here because spatial leakage is inherently symmetrical with MNE source reconstruction (Hauk & Stenroos, 2014). To investigate both within-and cross-frequency coupling, we used a "multi-layer" network design  by allowing the band of each node signal to be the same or different from each other ( Figure 2c). To control for possible power-induced effects in our rsFC, we also estimated source signal absolute power (i.e., their temporal variance) at each node with noise standardization to correct for the depth bias (Pascual-Marqui, 2002). We focused here on absolute power to ensure that rsFC changes are not merely due to modulations in signal-to-noise ratio (Muthukumaraswamy & Singh, 2011), but it is noteworthy that relative power changes associated to peak fre-  Figure 2d). In this context, power estimates were also averaged within each network.

| Statistical analyses
Neuropsychological and clinical test scores were compared between patients with MS and healthy subjects using two-tailed unpaired t tests. Because of our unbalanced design (99 patients vs. 47 healthy subjects), we used the Welch's version of the t statistic throughout this work, as it is more resilient to population heterogeneity (Ruxton, 2006 Within each group, we also analyzed the correlation between nodewise/mean network rsFC and clinical or neuropsychological scores (patients: disease duration, as measured from first clinical diagnosis, and 9 behavioral scores (listed in Section 2.2); healthy subjects: 8 behavioral scores (the same as for patients except EDSS)). We estimated a multiple regression model of nodewise/mean network rsFC values with the score of interest and the confounding factors as regressors and extracted the regression coefficient β corresponding to the relevant score. The permutation distribution of the resulting β matrices was generated by randomly shuffling the participants' order (within each group) against their respective scores before the regression analysis.
Significance levels were established using two-tailed maximum statistic testing (Nichols & Hayasaka, 2003) to simultaneously correct for the multiple comparisons across all pairs of nodes/networks and frequency band interactions. The null distribution of the maximum absolute value over all matrix entries was estimated from the permutation schemes described above (number of permutations: 10 5 ), and a significance threshold at significance level p was defined as the (1-p) th percentile of this permutation distribution. We set p = .05 Bonferroni corrected for the total number of tests, that is, 38 (1 contrast, 10 correlations for patients, and 8 correlations for healthy subjects = 19 tests, performed for both nodewise and mean network rsFC). All supra-threshold values were deemed to exhibit a significant effect (Nichols & Hayasaka, 2003). The p value of each maximum statistic, that is, the null probability of exceeding the observed maximum absolute value, was also estimated with its permutation distribution.
We also performed posthoc analyses to further check our results.
To confirm the absence of power differences that could drive rsFC contrasts, we applied an analogous statistical t contrast on power estimates (now within frequency only and without power regression).
Additionally, to see whether correlations within the patients' and healthy subjects' groups themselves were significantly different, we examined the group contrast of the corresponding β coefficients.
Their permutation distribution was obtained by shuffling participants (i.e., patients vs. healthy subjects) before regression, as for the t contrasts. These β contrasts were only investigated for clinical or neuropsychological scores exhibiting significant correlations. These extra analyses were performed at the same significance level as above (i.e., p = .05/38).

| Structural versus functional modeling of CI
An important question brought by the analysis of MS-related rsFC changes and their correlation to cognitive scores, is whether electrophysiological rsFC provides an added value to describe MS-related CI compared to structural markers that are often used (e.g., brain atrophy, cortical lesions, see e.g., Van Schependom & Nagels, 2017). To address this issue, we focused on cognitive scores that were significantly altered in patients with MS, and we first built for each of them a purely structural regression model with two well-established markers of neurostructural damage as dependent variables, that is, global lesion load and normalized thalamic volume (see e.g., Barkhof, 2002;Tewarie et al., 2013). They were obtained individually from 3D T1-weighted imaging and FLAIR MRIs using lesion detection and tissue segmentation implemented in the icobrain software (version 3.1; for details, see Jain et al., 2015). We then considered similar regression models that further included functional electrophysiolgical markers, and assessed whether this led to a significant increase in the amount R 2 of explained variance. Specifically, we added as dependent variables all mean network connectivity values (across RSNs and frequency bands, see Figure 2d) exhibiting a significant MS-related contrast. Given that structural and functional parameters are related Van Schependom & Nagels, 2017), these rsFC variables were orthogonalized with respect to the structural parameters before being used in the regression. This merely amounts to a reparametrization that emphasizes the independent information brought by rsFC but leaves the R 2 statistic unchanged.
By design, the structuro-functional regression model is biased toward a larger R 2 than the purely structural model as it contains more dependent variables. To assess whether rsFC actually increases the R 2 beyond this bias, we performed nonparametric statistical model comparisons. The permutation distribution of R 2 was generated under the null hypothesis that rsFC does not bring more information about patients' cognition than what is already entailed by structural markers, that is, that its orthogonal part is statistically independent of cognitive scores. Specifically, 10 4 null samples of R 2 were obtained by randomly shuffling the patients' order in the orthogonalized rsFC parameters before estimating the regression model and its R 2 value. Importantly, this permutation approach avoids the aforementioned bias and preserves orthogonality with the structural parameters. The p value on the R 2 statistic was derived as the fraction of null samples exceeding the R 2 value of the original structuro-functional regression model.

| RESULTS
We investigated the differences in electrophysiological power envelope rsFC among six major RSNs within and across five frequency bands (i.e., δ, θ, α, β L , and β H ) in patients with MS and matched healthy subjects, both at node and mean network levels. Correlations between clinical/neuropsychological test scores and both nodewise and mean network rsFC were also performed in the two cohorts of participants. For all the results henceforth described (Figures 3-8), significance was determined through a maximum-statistic permutation approach after controlling for signal power, age, gender, education level, and MEG system type as well as benzodiazepine status for patients, with additional correction for the total number of tests that F I G U R E 3 Significant nodewise rsFC changes between multiple sclerosis patients and healthy subjects. All are lower in patients. The bands in which significant differences were detected are indicated on the left. Only nodes with significant connections are shown (node color corresponds to their network, see Figure 2a), and their labels are superimposed were performed. Importantly, only the significant results will be henceforth reported, that is, all other tests were nonsignificant.
Among these, notably, no significant power differences were found in the contrast between patients with MS and healthy subjects (jtj < 1.2, permutation p value > .3) in any of the frequency bands.

| Comparison of clinical and neuropsychological scores
The neuropsychological test scores revealed that verbal fluency was significantly lower (COWAT, t = −2.9, p = .0044), while both cognitive and physical fatigues were significantly higher (cognitive and motor part of FSMC, t = 6.9, p = 9.2 × 10 −11 and t = 5.4, p = 6.5 × 10 −9 , respectively) in patients with MS compared to healthy subjects. Of

| Correlation between rsFC and EDSS score in multiple sclerosis patients
The multiple regression between nodewise rsFC data in patients with MS and their EDSS score ( Figure 5) highlighted a significant correlation for a subset of the band-specific connections that were observed in the contrast analysis detailed in Section 3.2 (Figure 3), i.e., a significant correlation in the β H band with nodes of the SMN (S2 and contralateral CS and FEF bilaterally) (permutation p = 3.12 × 10 −5 ). All these significant correlations were negative, that is, lower rsFC was associated with higher disability status. Accordingly, the EDSS score was also significantly negatively correlated with mean network SMN rsFC in the β H band as well in the β L band. Of note, these results appear to identify a topological reorganization of the brain-behavior relation with verbal fluency associated with MS. This was confirmed by the statistical contrast between correlation matrices obtained in patients and healthy subjects, which demonstrated that correlations were significantly higher in patients compared with healthy subjects for the exact same connections as in  Table 2).

| Correlations between rsFC and cognitive fatigue
In patients, negative correlations between nodewise rsFC and the cognitive part of the FSMC score were disclosed in the δ-β H cross- As for verbal fluency (Section 3.5.1), the apparent topological reorganization of brain-behavior relation associated with MS was confirmed by considering correlation contrasts. This analysis demonstrated that correlations were significantly lower in patients compared with healthy subjects for the exact same connections than in Figure 8 (left), and significantly higher for the exact same connections than in Critically, all significant results were obtained after stringent corrections for multiple comparisons and after regressing out possible confounding effects of multiple variables (e.g., power of oscillatory brain activity, age, sex, educational level, MEG system type, and benzodiazepine status). These data therefore demonstrate that MS is associated with changes in RSNs that mainly involve within-and cross-network Importantly, none of the included patients had an EDSS score ≥ 6 (i.e., intermittent or unilateral constant assistance [cane, crutch or brace] required to walk 100 m with or without resting) and most were within the range of 2-5. This indicates that almost all included Note: Model parameter R 2 represents the percentage of explained variance of cognitive outcomes (Verbal fluency and Cognitive fatigue). R 2 (structural): Model with structural regressors only. R 2 (structural, adjusted): Structural model R 2 corrected for the number of additional dependent (functional) variables.
Obtained as the mean of the null distribution generated by permutation of functional (but not structural) variables. R 2 (structuro-functional): Full model including both structural and functional dependent variables. Model comparison p value: Null probability that R 2 values generated by permutation exceeds the observed structuro-functional R 2 .
patients were able to walk unaided, and none were wheelchair-bound.
It therefore suggests that the observed reduction in SMN functional connectivity is not merely a consequence of sedentary living imposed by the disease. Still, we lack a specific measure of the patients' daily activity to confirm this hypothesis.
A positive correlation between global β-band MEG-derived phase rsFC and MS-related disability has also been reported , further suggesting that it represents an underlying electrophysiological correlate of the MS-related sensorimotor disability.
This hypothesis is in line with data showing that rehabilitation strategies specifically improving central integration of afferent proprioceptive inputs are more effective in improving balance disorders than conventional training in patients with MS with similar EDSS scores as those included in the present study (Gandolfi et al., 2015). These data suggest that electrophysiological rsFC could be a reliable method, free of performance bias, to properly assess the effects of therapeutic or rehabilitation strategies in MS.

| Network-level reduction of functional connectivity of the default-mode network
Mean network functional connectivity corresponded to the average of the correlation strengths between node pairs belonging to a given RSN (i.e., mean within-network rsFC) or between all nodes of one RSN and those of another RSN (i.e., mean cross-network rsFC). It thus gave an estimate of the global rsFC level within or between the considered RSNs. This approach was used to reveal rsFC changes associated with MS that were subtle at the nodewise connectivity level but consistent across connections within or between RSNs. This was confirmed by the analysis comparing nodewise and mean network effect sizes, which showed a dramatic increase of the latter compared to the former especially for intra-DMN, cross DMN-SMN, and cross DMN-LAN α-band rsFC.
Significant decrease of α-band mean network rsFC was found within the DMN, and between the DMN and the SMN/LAN in patients with MS compared to healthy subjects.
This MEG study therefore provides novel findings that complement previous fMRI studies (see Introduction for a summary) using an electrophysiological method that gives direct information about neural activity and is therefore free of any neurovascular coupling bias. Considering the clinical characteristics of our patients' cohort, it supports rs-fMRI papers showing reduced DMN rsFC in patients with rather advanced multiple sclerosis. Indeed, it demonstrates that the diseaserelated alterations in within-and cross-network DMN rsFC actually involve a rather global disruption of within-and of some specific cross-network (i.e., with the SMN and the LAN) DMN connections that is not detectable at the nodewise connectivity level. This concurs with the recognized role of the DMN as a core region for the functional integration with other RSNs (de Pasquale et al., 2012). It also demonstrates that these mean network rsFC alterations are specifically observed in the α frequency band, which is perfectly in line with the main carrying frequency of electrophysiological DMN rsFC previously reported (Brookes et al., 2011;Sjøgård et al., 2019;Vidaurre et al., 2018;Wens et al., 2014). These findings might also explain the more random (i.e., less structured and hierarchical) organization of functional brain networks that has been previously reported in α-band phase-based rsFC in patients with MS .

| Disease duration is correlated with reduced language network functional connectivity
A strong association between disease duration and intrinsic functional connectivity within the LAN was observed in patients with MS in the β frequency band in the form of a negative correlation (i.e., the longer the disease duration, the lower the rsFC within the LAN). This was the case for both mean within-LAN rsFC and for specific nodewise connections.
The occurrence and severity of language impairments in MS are actually poorly defined (for a review, see Renauld, Mohamed-Saïd, & Macoir, 2016). Various verbal language impairments (picture naming, reading comprehension) as well as phonemic and semantic verbal fluency have been repeatedly reported in patients with MS (for a review, see Henry & Beatty, 2006). However, the heterogeneity of the methods used renders the elaboration of definite conclusions difficult (Renauld et al., 2016). Still, given the sensorimotor and various cognitive deficits characterizing MS, verbal language functions should be substantially affected (Renauld et al., 2016).
The strong negative correlation between disease duration and LAN rsFC found in this study might represent a neural correlate of the verbal language dysfunctions observed with the evolution of the disease. Unfortunately, no proper verbal language assessment was performed in our patients' cohort. Still, verbal fluency scores were significantly lower in patients than in healthy subjects and correlated with some language-related network connections in patients.
Although verbal fluency does not specifically assess verbal language function, some studies suggested that language processing is a critical component for this task (Whiteside et al., 2016). This finding therefore suggests that verbal language function might indeed be impaired in our patients' cohort. Still, disease duration did not significantly cor- As discussed in Section 4.3., verbal fluency is not a languagespecific cognitive measure but rather reflects the integrity of various high-level cognitive functions such as working memory, executive functions and semantic cognition (Ralph, Jefferies, Patterson, & Rogers, 2017;Whiteside et al., 2016). In healthy subjects, we found a positive correlation between verbal fluency scores and the strength of MS is often referred to as a structural disease (Compston & Coles, 2008;Mandolesi et al., 2015). For functional measures like MEG rsFC to be of additional utility as markers for MS-related CI, they should add some information not already explained by the available structural measures. This study showed that, in fact, including functional connectivity in a regression model significantly added explanatory power over purely structural measures. Previous studies have shown a correlation between MEG rsFC and overall cognition , and we here show that it can independently explain a significant amount of variation in specific MS-related CIs as well.

| Methodological considerations and limitations
This study is the first to investigate rsFC in MS using band-limited power envelope correlation, and as such, lacks direct comparability to the existing literature. All previous MEG/EEG studies of MS used phase-based rsFC, which measures different aspects of electrophysiological brain interactions (Engel, Gerloff, Hilgetag, & Nolte, 2013 (Siems & Siegel, 2020). Furthermore, both of them have been shown to be related to fMRI rsFC in some ways (e.g., Tewarie et al., , 2016. There is also evidence that, as rsFC based on power envelope correlation, phase coupling also displays some intrinsic (i.e., task independent) properties, although to a lesser degree than envelope correlation (Sjøgård et al., 2020). However, MEG power envelope correlation is closely related to rs-fMRI functional connectivity, which is why we have split out discussion between the fMRI literature and the (phase-based) MEG literature. That said, the fact that the neurovascular coupling seems to be altered in MS (Marshall et al., 2014) may also limit the validity of the comparison to fMRI. To mitigate this lack of comparability, we used here a large cohort of patients, a comprehensive analysis of a large-scale multifrequency connectome, and a stringent control of confounding factors and false positives (as further discussed below).
It is noteworthy that, on general grounds, the interpretation of MS-related rsFC increases or decreases is not straightforward. Both directions of change have been reported with fMRI rsFC and could either represent beneficial or maladaptive processes (Schoonheim, Meijer, & Geurts, 2015). Additionally, pathological white matter damage may lead to both increases and decreases in rsFC (Tewarie et al., 2018). That said, our MEG data consistently identified rsFC decreases only in patients with MS, and these decreases were related to worse clinical and cognitive scores.
Another limitation is that we only considered static rsFC estimates, which encapsulates the time-averaged, temporally stable brain It is also worth cautioning about a possible interpretational pitfall regarding the cross-frequency approach  due to our use of conventional frequency bands, which may not reflect the data at play in the rsFC changes reported here (although Vidaurre et al., 2018 showed that MEG data-driven bands converge with the conventional ones). Cross-frequency coupling identified across adjacent frequency bands could therefore merely reflect a broadband process overlapping the two bands and be artifactually coined as cross-frequency. The β L -β H ( Figure 6) and α-β L couplings (Figure 7) disclosed in our brain-behavior correlations may be of that type. On the other hand, it is more likely that couplings across nonadjacent frequency bands such as the δ-β H and α-β H couplings disclosed in Figure 8 reflect genuinely cross-frequency interactions. On a similar note, our rsFC analyses focused on the frequency bands typically considered to carry the electrophysiological RSNs, so we did not include gamma-band rsFC (see, for example, Hipp et al., 2012). This means that some disease-related rsFC changes might have been missed, either within the gamma band or between the lower frequency bands and the gamma band.
Last, and critically, the results obtained in this study generally err on the conservative side and may thus be fraught with false negatives.
First, the low-density, 32-node brain parcellation used here was limited to nodes of well-known RSNs (based on a meta-analysis of fMRI rsFC, as in de Pasquale et al., 2012). Although the spatial smoothness inherent to MEG partially mitigates this coarseness, some brain areas were poorly sampled at best in our connectome analysis (see Figure 2a) and functional connections outside these RSN regions were not considered here. Further, a connectome design based on fMRI may not be optimal to investigate the multi-spectral signatures of rsFC, to which fMRI is insensitive. Given the close anatomical correspondence between electrophysiological and fMRI RSNs (see, for example, Brookes et al., 2011), using this type of design is well justified in the context of within-frequency rsFC but it may have a more limited value for crossfrequency couplings. These two limitations mean that our analyses might miss MS-related alterations in network configurations outside these RSNs. Second, our statistical design included stringent control for several confounding factors, and also for the large number of comparisons (496 connections for nodewise rsFC or 15 for mean network rsFC estimates, 10 pairs of frequency bands, 38 tests) so as to ensure a family-wise false positive rate below 5%. The price to pay is a lessened sensitivity to true differences and correlations. This was presumably mitigated by our inclusion of a large number of participants, and the negative impact of our unbalanced design in contrasts was further alleviated by the use of Welch's t statistic. Still, given the lack of comparable works, we chose to focus on conservative statistics, without overinterpreting nonsignificance as a lack of genuine effect. The results reported here thus represent the most robust effects of MS on electrophysiological RSNs and its brain-behavior correlates.

| Conclusion
This MEG study demonstrates that MS entails several robust frequency-specific network-level and regional changes within and between RSNs (mainly the SMN, DMN, and LAN) that are related to motor disability, disease duration and specific CIs. It also shows that MEG rsFC relying on power envelope correlation brings significant independent information in specific MS-related CIs outside structural brain abnormalities. This shows that frequency-specific RSN changes may be suitable candidates for electrophysiological markers of both clinical and cognitive aspects of the disease with the remarkable advantages of being totally noninvasive, free of performance bias and free of any neurovascular issue. The ability of EEG to uncover similar RSNs as MEG (Coquelet et al., 2020;Liu et al., 2017;Siems et al., 2016;Sockeel et al., 2016) should facilitate the dissemination of the proposed approach in MS and other brain disorders.

DATA AVAILABILITY STATEMENT
Data and code used for this research project will be shared in the context of an academic and scientific collaboration formalized by a collab-