The functional brain favours segregated modular connectivity at old age unless affected by neurodegeneration

Brain’s modular connectivity gives this organ resilience and adaptability. The ageing process alters the organised modularity of the brain and these changes are further accentuated by neurodegeneration, leading to disorganisation. To understand this further, we analysed modular variability—heterogeneity of modules—and modular dissociation—detachment from segregated connectivity—in two ageing cohorts and a mixed cohort of neurodegenerative diseases. Our results revealed that the brain follows a universal pattern of high modular variability in metacognitive brain regions: the association cortices. The brain in ageing moves towards a segregated modular structure despite presenting with increased modular heterogeneity—modules in older adults are not only segregated, but their shape and size are more variable than in young adults. In the presence of neurodegeneration, the brain maintains its segregated connectivity globally but not locally, and this is particularly visible in dementia with Lewy bodies and Parkinson’s disease dementia; overall, the modular brain shows patterns of differentiated pathology.

optimal densities reached at 1.9% (global) and 3.24% (local). b Same as (a) for the NKI cohort; optimal densities were reached at ~2.04% (global) and 3.8% (local). c Same as (a) for the 1000 functional connectome (TFC) cohort; optimal edge densities were reached at 3.39% (global) and 5.11% (local). Only the NKI cohort with global construction did not reach optimal density at the same value for both groups; for older adults, this was reached at 2.26% and for young adults at 1.81%.

Validation analysis
The previous set of experiments were implemented using 451-ROI atlas at optimal edge density; the density at which the modularity index difference − was found to be maximum ( Supplementary Fig 2). However, in order to observe if our analyses were also valid at other densities, we estimated MV and MD at 10% and 20% edge density. Additionally, three brain parcellations were implemented in addition to the 451-ROI one; these were 100, 200 and 247 ROIs.
Results from this validation analysis are presented as supplementary material. Validation results for MD in the NCL cohort using the different atlas parcellations and edge densities are shown in Supplementary Fig 6. Here, there was a high variability on MD amplitude values across the range of densities and parcellations, however, the pattern of higher MD at insular and basal regions and lower MD at motor-sensory, occipital cortices and cerebellum remained invariant. Additionally, high contrast between regions of high and low MD is seen for results at optimal edge density. Validation on the NKI cohort showed similar patterns (not published) and this analysis was not implemented on the TFC cohort because only connectivity matrices at a fixed parcellation were available for this cohort. Supplementary Fig 6. Validation of modular dissociation (MD) results across different parcellation resolutions (atlases) and densities; optimal edge density, 10% and 20% of the strongest edges within the Newcastle University (NCL) database. a1-a4) 100-ROI atlas at optimal edge density of 6%, a1) older adults (NCL healthy controls), a2 Alzheimer's disease dementia (ADD), a3 dementia with Lewy bodies (DLB) and a4 Parkinson's disease dementia (PDD). c and e same as (a) for 10% and 20% edge density respectively same 100-ROI atlas. b1-b4 200-ROI atlas at optimal edge density of 6%, b1 older adults, b2 ADD, b3 DLB, b4 PDD. d and f the same as (b) for densities 10% and 20%. g1-g3 ADD MD values for the 247-ROI atlas at an optimal density of 5.28%, 10% and 20% in the same order. h1-h3 same as g1-g3 for the DLB group. i1-i3 same as g1-g3 for the PDD group. j1-j3 same as g1-g3 for the older adult (NCL healthy control) group. k1-k3 ADD MD values for the 451-ROI atlas at an optimal density of 3.24%, 10% and 20% in the same order. l1-l3 Same as k1-k3 for the DLB group. m1-m3 same as k1-k3 for the PDD group. n1-n3 Same as k1-k3 for the older adult (NCL healthy control) group.

Motion analysis
Participant motion within the MRI scanner may drive spurious results when comparing young vs older adults; it is known that older adults tend to move more than young adults during acquisition of neuroimages. We assessed motion by estimating Subjects with a large head motion were excluded with a mean FD threshold of 0.5mm.
The demographics of cohorts with FD exclusion are shown in Supplementary  Table 2). Nevertheless, the eight consensus modules for NKI cohort remained (ref. Supplementary Fig 11b vs Fig 3a in the main text), while for NCL a total of eight modules (ref. Supplementary Fig 11c vs  MD. The patterns from healthy ageing to dementia were also similar to the results before FD correction (ref. Supplementary Fig 14 vs Fig 6, 7 in the main text).
Compared with OA, we found the PDD group had low differences for MV in most From the global level, the mean MD across all nodes in dementia groups didn't change from OA either (Supplementary Fig 15b). The YA also had significantly higher mean MD than OA, mainly in four of the eight modules: superior frontal-occipital, ventral frontal, temporal and cerebellum ( Supplementary Fig 15c), which were mostly overlapped by results before mean FD regression. Locally, the MD patterns varied, however, were invariant to some degree. For instance, the MD change in the frontal related modules was on average negative in the dementia groups ( Supplementary Fig   16e, h). And for the motor-sensory module, all groups remained a positive change relative to OA (Supplementary Fig 16a). Therefore, the additional motion processing was proved not to change the main patterns of how MV and MD change from healthy ageing to a dementia state due to neurodegeneration.
Overall, the above replication analysis indicated that our primary findings seemed robust to head motion although it did influence FC.

Standard parcellation analysis
In the current study, we estimated and used four different functional parcellations derived from an independent study that was scanned in the same MRI. The use of inhouse parcellations assure there is no bias introduced by including the same cohorts under analysis (NCL and NKI) for atlas estimation and no bias introduced by study site.
Although evidence has shown that the measures of network segregation and integration seem to be robust to the underlying parcellations 4 , it's unclear if our findings are also obvious in well-established brain parcellations. To this end, we conducted a replication analysis using a multi-modal atlas--Human Brainnectome Atlas 5 which in total included 210 cortical, 36 subcortical and 28 cerebellar subregions. The mean FD, sex, and age were added as covariates of no interest in analysis.
As shown in Supplementary Fig 17a, the optimal edge density was selected at 4.40% where there was a big bias as well between ADD and other groups because of high motion correction. The modular definitions were different from in-house parcellation studies (ref. Supplementary Fig 17b, c). For example, the frontal-parietal module in

451-ROI case was divided into two separate modules on the basis of Human
Brainnectome Atlas for NKI cohort, and two separate modules in 451-ROI case: cerebellum and occipital modules were combined into one integrated module based on Human Brainnectome Atlas for NCL cohort. Nevertheless, the MV patterns were consistent with 451-ROI parcellation findings in which high MV in association cortices and low MV in primary cortices. More specifically, higher MV was found at inferior/superior frontal, parietal, basal structures, superior temporal gyri, and lower MV at occipital, superior motor-sensory, temporal poles, cerebellum. The MD also showed similar patterns (ref. Supplementary Fig 18): low values in superior frontal, occipital, motor-sensory cortices and high values located primarily in cerebellum, basal structures. The temporal cortex, in contrast, showed significantly high MD, especially for NKI cohort. Although this phenomenon was not obvious in 451-ROI NKI results, it was also found in the 177-ROI TFC (ref. Supplementary Fig 3) and 200-ROI NKI cases (ref. Supplementary Fig 6 b1-b4). Thus, despite differences derived from the standard atlas, the general MV and MD distributions of low and high values were maintained.
MV and MD comparisons between OA and YA groups were performed in NKI cohort, Supplementary Fig 19. Contrary to the 451-ROI studies, the patterns of high MV for OA group shrunk, only the regions in insular, inferior frontal cortices were reserved.
Several nodes of occipital cortex had a trend of high MV in OA as well but not significantly after FDR corrections while it was obvious in 451-ROI results. However, the patterns of low MV in OA which were primarily located in parietal, motor-sensory, cerebellum and part of the superior frontal cortices were consistent. Moreover, the results of MD comparison between OA and YA groups on the standard atlas were similar to 451-ROI NKI and 177-ROI TFC results, showing high MD in insulo-opercular cortex, part of frontal and occipital cortices although they were not significant after FDR correction. In general, the overall MV across the brain, on the contrary,  Fig 20), we came to the same conclusion that the PDD group had globally low differences for MV but significantly high differences for MD, especially in cerebellum, insulo-opercular and motor-sensory cortices. The ADD and DLB groups showed differentiated patterns even though the attached features were different from those in 451-ROI analysis. From the perspective of the global MD across all nodes, we found the mean MD higher in ADD group than in OA ( Supplementary   Fig 21b), which was not shown in 451-ROI case. But the mean MD in DLB and PDD groups didn't change significantly from healthy ageing in the same way. Besides, the YA also showed higher global MD than in OA which was mainly in four of the seven modules: frontal-parietal, ventral frontal, temporal and temporal-occipital modules ( Supplementary Fig 21c). Locally, the MD patterns varied but to some degree kept invariant compared to 451-ROI studies. For example, the MD change in the frontal related modules was also on average negative in the dementia groups ( Supplementary Fig 22e, f). And for the motor-sensory module, all groups remained a positive change relative to OA (Supplementary Fig 22a). The standard-parcellation analysis consistently indicated a segregated modular structure with ageing and this pattern didn't continue from healthy ageing to a dementia state, except for ADD which showed a further dissociation.
The differences between standard-parcellation and in-house parcellation analysis mainly existed in the MV comparison between YA and OA ( Supplementary Fig 19a), and the mean MD comparison between ADD and OA (Supplementary Fig 21b) Fig 19a-b vs Supplementary Fig 19c-d) can support this. The later differences, on the one hand, could be owing to the confounding effect of head motion and the standard atlas. The global mean MD/OA for ADD group increased to be positive after regressing out head motion even though this was still not significant ( Supplementary Fig 15b vs Fig 8b). The use of the Human Brainnetome Atlas promoted a further increase of the global mean MD/OA for ADD group which was even shown to be significantly greater than zero.
The above results demonstrated that the standard parcellation did affect the comparison results but the primary findings generally remained, especially for the large and small MV/MD distribution across the brain, the high MV/MD of OA group in the insulo-opercular cortex compared with YA, and local but not global changes of neurodegenerative dementia groups: DLB and PDD groups.

Limitations and considerations
Our investigation has some limitations and considerations that deserve to be mentioned. Previous research has pointed out that disease populations present with lower weights in their connectivity matrices 6 . We believe that our investigation is not affected by this factor because we compared communities (their variance and dissociation) and not connectivity strength or measures that are directly influenced by strength. Community estimation is only influenced by the distribution of weights within the connectivity matrix but not their global strength. Another related concern is proportional thresholding which is implemented to binarise weighted connectivity previous work has proved that this increases sensitivity in group comparisons and that it can reduce the influence of movement in fMRI time series 9 . However, it is also known that this regression heavily modifies the weight distribution in the connectivity matrix which does affect community estimation 9 . For our in-house pre-processed databases, NKI and NCL, we did not implement global signal regression (we only regressed average CSF signal from bilateral ventricles). However, the TFC matrices downloaded from the UMCD had this regression. This may have contributed to the differences between the NCL-NKI cohorts and the TFC cohort (Fig 3-Fig 5 in the main text). In support of this, a recent investigation by Turchi, et al. 10 reported that the global signal is highly driven by basal activity. The authors performed pharmacological inactivation of two subregions of the nucleus basalis of Meyner (NBM). By this inactivation, Turchi, et al. 10 found that regional "global" signal (ipsilateral to the inactivated region) was suppressed, however, the presence of the resting state networks, such as the default mode network, remained invariant, proving that global signal in resting-state fMRI is of neuronal origin. It is possible therefore that global signal regression modifies brain connectivity at the basal brain by regressing out its activity. By this experiment, the authors have presented compelling evidence that discourages regression of the global signal in future resting-state fMRI studies.
Another consideration is the multi-site nature of the TFC database. The downloaded matrices came from nine different sites, and these were exclusive for older and younger adults; five sites comprised YA and four sites OA only. Because of this, we were only able to correct for within-group studies but not when comparing OA vs YA.
We believe that the global higher MV (Fig 4 in the main text) and Q ( Supplementary   Fig 2) observed in OA compared with YA may be partially driven by study-site differences. However, the pattern of MV, MD and their differences between groups were remarkably similar in both NKI and TFC cohorts.
One last consideration in our investigation is that in order to observe the effects of neurodegenerative dementia in the old age brain, we used two independent databases; NKI for ageing and NCL for dementia. We normalised modular MD effects by their respective OA groups and studied the relative MD changes from the perspective of OA. This strategy allowed us to see that in healthy ageing, modular dissociation is increased at insular and occipital regions but decreased everywhere else and that in dementia there are deviations of MD at the modular level but not globally. This question could be ideally investigated with a longitudinal neuroimaging database that follows patients from youth to old age and then to dementia (if acquired). Databases