Brain functional network integrity sustains cognitive function despite atrophy in presymptomatic genetic frontotemporal dementia

The presymptomatic phase of neurodegenerative disease can last many years, with sustained cognitive function despite progressive atrophy. We investigate this phenomenon in familial frontotemporal dementia (FTD).


INTRODUCTION
Across the adult healthy lifespan, the structural and functional properties of brain networks are coupled, and both are predictive of cognitive ability. 1,2 The connections between structure, function, and performance have been influential in developing current models of aging and neurodegeneration. [3][4][5] However, this work contrasts with the emerging evidence of neuropathological and structural changes many years before the onset of symptoms of Alzheimer's disease (AD) and frontotemporal dementia (FTD). [6][7][8] Genetic FTD with highly penetrant gene mutations provides the opportunity to examine the precursors of symptomatic disease. Three main genes account for 10% to 20% of FTD cases: chromosome 9 open reading frame 72 (C9orf72), granulin (GRN), and microtubule-associated protein tau (MAPT). These genes vary in their phenotypic expression and in the age at onset. 9 Despite pleiotropy 10 and environmental and secondary genetic moderation 11,12 all three mutations cause significant structural brain changes in key regions over a decade before the expected age at disease onset, 7,13 confirmed by longitudinal studies. 14,15 The divergence between early structural change and late cognitive decline provokes the question: how do presymptomatic mutation carriers stay so well in the face of progressive atrophy? We propose that the answer lies in the maintenance of network dynamics and functional organisation. 16 Across the lifespan, functional brain network connectivity predicts cognitive status, 17 and this connectivity-cognition relationship becomes stronger with age. [18][19][20] Our overarching hypothesis is that for those at genetic risk of dementia, the maintenance of network connectivity prevents the manifestation of symptoms despite progressive structural changes. A challenge is that neither the anatomical and functional substrates of cognition nor the targets of neurodegenerative disease are mediated by single brain regions: They are distributed across multi-level and interactive networks. We therefore used a multivariate data-driven approach to identify differences in the multidimensional brainbehavior relationship between presymptomatic carriers and noncarriers of mutations in FTD genes. We identified key brain networks 21 from a large independent population-based age-matched data set. 22 We tested three key hypotheses: (1) presymptomatic carriers differ from non-carriers in brain structure and brain function, but not in cognitive function, (2) brain structure and function correlate with performance in both groups, but functional network indices are stronger predictors of cognition in carriers, and (3) the dependence on network integrity for maintaining cognitive functioning increases as carriers approach the onset of symptoms.

Participants
Thirteen research sites across Europe and Canada recruited participants as part of an international multicenter partnership, the Genetic Frontotemporal Initiative (GENFI). A total of 313 participants had usable structural and resting state functional magnetic resonance imaging (fMRI) data. 7,13 The study was approved by the institutional review boards for each site, and participants providing written informed consent. Inclusion criteria included anyone over the age of 18 who is symptomatic or an asymptomatic first-degree relative. Five participants were excluded due to excessive head motion (see below), resulting in 308 data sets for further analysis.
Participants were genotyped based on whether they carried a pathogenic mutation in MAPT, GRN Participants and site investigators were blinded to the research genotyping, although a minority of participants had undergone predictive testing outwith the GENFI study. See Table 1 for demographic information and Table 2 Figure 1 provides a schematic representation of imaging data processing pipeline and the analysis strategy for linking brain-behavior data.

Neuroimaging assessment
MRI data were acquired using 3T scanners and 1.5T where no 3T scanning was available from various vendors, with optimized scanning protocols to maximize synchronization across scanners andsites. 7 26 To quantify the total motion for each participant, the root mean square volume-to-volume displacement was computed using the approach of Jenkinson et al. 27 Participants with 3.5 or more standard deviations (SD) above the group mean motion displacement were excluded from further analysis (N = 5). To further ensure that potential group bias in head motion did not affect later analysis of connectivity, we took three further steps: F I G U R E 1 Schematic representation of data processing and analysis pipeline to test for brain-behavior differences between presymptomatic carriers (PSCs) and non-carriers (NCs) as a function of expected years to onset (EYO) of symptoms, while controlling for covariates of no interest (Covs). Brain structural measures were based on the mean gray matter volume (GMV) in 246 nodes, as defined in the Brainnetome atlas. 35 Brain functional measures were based on the functional connectivity between 15 nodes as part of four large-scale networks, which were defined in an independent cohort of 298 age-matched individuals part of the Cam-CAN data set (1) fMRI data were further postprocessed using whole-brain independent component analysis (ICA) of single subject time-series denoising, with noise components selected and removed automatically using a priori heuristics and the ICA-based algorithm, 28 (2) postprocessing of network node time-series (see below), and (3) a subject-specific estimate of head movement for each participant 27 included as a covariate in group-level analysis. 29

Network definition
The location of the key cortical regions in each network was identified by spatial-ICA in an independent data set of 298 age-matched healthy individuals from a large population-based cohort. 22 Full details about preprocessing and node definition have been described previously. 30 Four networks commonly affected by neurodegenerative diseases including FTD 21 were identified by spatially matching to pre-existing templates. 31 The default mode network (DMN) contained five nodes:

Group differences in brain structure, function, and cognition
To assess the group differences in neuroimaging and behavioral data set we used multiple linear regression with a well-conditioned shrink-F I G U R E 2 Visualization of spatial localization of the nodes part of the four large-scale networks and their mean functional connectivity (circular plot) across all participants in this study. Nodes and networks were defined in an independent cohort of 298 age-matched individuals part of the Cam-CAN data set. 30 The default mode network (DMN) contained five nodes: the ventral anterior cingulate cortex (vACC), dorsal and ventral posterior cingulate cortex (vPCC and dPCC), and right and left inferior parietal lobes (rIPL and lIPL). The salience network (SN) was defined using right and left anterior insular (rAI and lAI) and dorsal anterior cingulate cortex (dACC). The frontoparietal network (FPN) was defined using right and left anterior superior frontal gyrus (raSFG and laSFG) and right and left angular gyrus (rAG and lAG). The dorsal attention network (DAN) was defined using right and left intraparietal sulcus (rIPS and lIPS) age regularization 32,33 and 10-fold cross-validation. 34 In the analysis of brain structure we used as independent variables the mean GM volume (GMV) of the 246 brain nodes in the Brainnetome atlas. 35 The Brainnetome atlas was developed to link functional and structural characteristics of the human brain 35 and provides a fine-grained wholebrain parcellation with a superior representation of age-related differences in brain structure compared to other cortical parcellation schemes. 36,37 In the analysis of brain function, we used the functional connectivity between 15 nodes, which were part of the four large-scale functional networks described earlier. In the analysis of cognitive function, the independent variables comprised the performance measures on the 13 neuropsychological tests performed outside of the scanner.
In all three analyses the dependent variable was the genetic status (PSC vs NC) including age as a covariate of no interest. GENFI's largesampled cohort was created using harmonized multi-site neuroimaging data. Although, scanning protocols were optimized to maximize comparability across scanners and sites, 7,13 different scanning platforms can introduce systematic differences that might confound true effects of interest. 38 Therefore, in the analysis of neuroimaging data we included scanner site and head motion as additional covariates of no interest.

Brain-behavior relationships
For the brain-behavior analysis, we adopted a two-level procedure.
In the first-level analysis, we assessed the multidimensional brainbehavior relationships using partial least squares. 39  group, brain scores x years to expected onset, and so on). The dependent variable was subjects' cognitive scores from the first level analysis in the corresponding PLS (Cognition-LV). Given that the interaction effects were derived from continuous variables, we tested and interpreted interactions based on simple slope analysis and slope difference tests. [40][41][42] Covariates of no interest included gender, handedness, head movement, and education ( Figure 1). In addition, we included average GMV across all 15 nodes as a covariate in the FCbehavior analysis to ensure that the observed effects are over and above differences in the level of atrophy.

Brain structure
The multiple linear regression model testing for overall group differences in GMV between PSC and NC was significant (r = .14, P = .025), reflecting expected presymptomatic differences in brain-wide atrophy.
The frontal, parietal, and subcortical regions had most atrophy in PSC ( Figure 3). As expected, the group difference in GMV of these regions increased as EYO decreased (see Supplementary Materials).

Brain function
The multiple linear regression model testing for overall group differences in functional connectivity between PSC and NC was marginally significant (r = .12, P = .049). The pattern of connectivity indicated mainly increased connectivity between SN-DMN and SN-FPN in presymptomatic carriers, coupled with decreased connectivity within the networks and DMN-FPN connectivity ( Figure 3).

Cognitive function
We did not identify group differences in cognition and behavior (r = .002, P = .807), confirming the impression of "healthy" status among presymptomatic carriers. However, in the next section, we consider the relationships between structure, function, and cognition that underlie this maintenance of cognitive function.

Structure-cognition
Partial least squares analysis of GMV and cognition identified one  The scatter plots in the middle and right hand-side represents function-cognition LV relationship as a function of expected years to onset (EYO split in two groups, near and far, see text) in each genetic status group separately. This is also represented using a bar chart in (D), where continuous and dashed lines indicate significance of effect differences and difference in differences, respectively. † and * denote significant tests at P-value < .05 (one-and two-sided, respectively) ance in Cognition-LV. We used simple slope analysis and slope difference tests [40][41][42] to test formally for differences in the relationship between Function-LV and Cognition-LV for PSC and NC. The relationship between Function-LV and Cognition-LV was stronger for PSC relative to NC (r = .16, P = .002), indicating the increasing importance of functional connectivity between the large-scale networks for PSC participants to maintain performance ( Figure 5).

Connectivity-cognition
For ease of interpretation and illustration, we also computed the correlation between Cognition-LV and Function-LV for high and low levels of expected years to onset (or EYO) within each group separately, where the levels were taken to be 1 SD above and below the mean values of EYO following the simple slopes approach. [40][41][42] The two EYO subgroups were labeled "near" and "far," with "near" for EYO values close to zero (ie, participant's age is "near" the age at which disease symptoms were demonstrated in the family), and "far" for EYO being a largely negative value (ie, participant's age is "far" from the age at which disease symptoms were demonstrated in the family). The analysis indicated that as the EYO decreases (ie, participant's age is reaching the years of onset of symptoms) the relationship between functional connectivity and performance becomes stronger. This effect was highly significant in presymptomatic carriers (r = .31, P < .001) and tended towards significance in non-carriers (r = .12, P = .038, one-sided). The differences in effects between presymptomatic carriers and non-carriers was qualified by a significant interaction term (t = 2.27, P = .024, ie, the effect in presymptomatic mutation carriers was statistically stronger than the effect detected in non-carriers). These findings indicate that the relationship between FC and cognition is stronger in PSC relative to NC, and that this relationship increases as a function of EYO.

DISCUSSION
In the present study, we confirmed previous findings of group differences in brain structure and function, in the absence of differences in cognitive performance between non-carriers and presymptomatic carriers of FTD-related genetic mutations. But, although the relationship between structure and cognition was similar in both groups, the coupling between function and cognition was stronger for presymptomatic carriers, and increased as they approached the expected onset of disease.
These results suggest that people can maintain good cognitive abilities and successful day-to-day functioning despite significant neuronal loss and atrophy. This disjunction between structure and function is a feature of healthy aging, but we have shown that it also characterizes presymptomatic FTD, over and above the age effects in their other family members, despite widespread progressive atrophy. The multivariate approach reveals two key findings: (1) stronger within-network and weaker between-network functional connectivity is associated with better cognition, more strongly in presymptomatic carriers than in agematched non-carriers, and (2) as carriers approach their estimated age of symptom onset, and atrophy becomes evident, the maintenance of good cognition is increasingly associated with sustaining balance of within-and between-network integration.
This balance of within-and between-network connectivity is characteristic of segregated and specialized network organization of brain systems. Such functional segregation varies with physiological aging, 17,18,43 with cognitive function, 18 and in individuals at risk for AD. 44 Graph-theoretic quantification of network organization confirms the relevance of modularity and efficiency to function in FTD. 16 Conversely, the loss of neural systems' modularity mirrors the loss of functional specialization with age 45 and dementia. 44 Here, we show the significance of the maintenance of this functional network organization, with a progressively stronger correlation with cognitive performance as seemingly healthy adults approach the age of expected onset of FTD.
The uncoupling of brain function from brain structure indicates that there may be independent and synergistic effects of multiple factors leading to cognitive preservation. This is consistent with a previous work in healthy aging where brain activity and connectivity provide independent and synergistic predictions of performance across the lifespan. 19 Therefore, future studies need to consider the independent and synergistic effects of many possible biomarkers, based on MRI, computed tomography, positron emission tomography, CSF, blood, and brain histopathology. For example, functional network impairment may be related to tau expression and tau pathology, amyloid load, or neurotransmitter deficits in neurodegenerative diseases, independent of atrophy. 30,[46][47][48] It is important to note that studies need to recognize the rich multivariate nature of cognition and of neuroimaging in order to improve stratification procedures, for example, based on integrative approaches that explain individual differences in cognitive impairment. 30,49 On a clinical level, this may facilitate future studies to establish whether presymptomatic carriers who maintain such connectivity profiles and thereby neuropsychological function in the presence of atrophy may have a lower risk of progression and better prognosisinformation that will be important for future triallists, patients, and carers.
We also recognize the difficulty in determining a unique contribution of each factor (eg, brain structure and brain function), given the increasing interaction between factors in advanced stages of disease. 50 This is further complicated by these alterations becoming irreversible with progression of neurodegeneration. 51 This suggests that the critical interplay between multiple factors (including brain structure and function) may be better studied in the asymptomatic and preclinical stages as well as across the healthy lifespan, which could still be modifiable and their influences are likely to be more separable.
Our findings agree with the model of compensation in the presymptomatic and early phases of Huntington disease, where network coupling predicted better cognitive performance. 52 In a recent longitudinal study, a non-linear concave-down pattern of both brain activity and behavior was present, despite a linear decline in brain volume over time. 53 Similar effects have been observed also in healthy aging and amnestic mild cognitive impairment, where greater connectivity with the default-mode network and weaker connectivity between default-mode network and dorsal-attention network was associated with higher cognitive status in both groups. 54 Network integrity may also play a role in compensatory mechanisms in non-cognitive symptoms, such as motor impairment in Parkinson disease. 55 Accordingly, increased network efficiency and connectivity have been shown in prodromal phases, followed by decreased local connectivity in symptomatic phases, suggesting the emergence and dissipation of neural compensation. 56 The current study has several limitations. First, despite the large size of the overall GENFI cohort, we did not analyze each genetic group sep- The current study is cross-sectional. Therefore, we cannot infer longitudinal progression within subjects as the unambiguous cause of the effects we observe in relation to expected years of onset. Accumulating evidence suggests that network integrity serves to maintain performance with either physiological ageing or pathological conditions. However, longitudinal mediation studies and pharmacological or electroceutical interventions would be needed to prove its causal role in cognitive preservation. Finally, our findings are limited to autosomal dominant FTD, which represents a minority of FTD: Generalization to sporadic forms of disease would be speculative.
In conclusion, we used a multivariate data-driven approach to demonstrate that brain functional integrity may facilitate presymptomatic carriers to maintain cognitive performance in the presence of progressive brain atrophy for years before the onset of symptoms. The multivariate approach to cognition and brain function is well-suited to address the effects of multiple interacting risk factors on biomarkers of the progression of neurodegeneration, ahead of clinical conversion to dementia. The approach and our findings have implications for the design of presymptomatic disease-modifying therapy trials, which are likely to rely initially on surrogate markers of brain health rather than clinical end points.

CONFLICT OF INTEREST
There were no financial or other conflicts of interest requiring declaration.