NETWORK CONNECTIVITY PREDICTS CORTICAL THINNING AND COGNITIVE DECLINE IN EARLY PARKINSON’S DISEASE

Parkinson′s Disease (PD) is a progressive neurodegenerative disorder characterized by motor and cognitive deficits. The neurodegenerative process is thought to move stereotypically from the brainstem up to the cerebral cortex, possibly reflecting the spread of toxic alpha-synuclein molecules. Using a large, longitudinal, multi-center database of de novo PD patients, we tested whether focal reductions in cortical thickness could be explained by disease spread from a subcortical “disease reservoir” along the brain9s connectome. PD patients (n=105) and matched controls (n=57) underwent T1-MRI at entry and one year later. Over this period, PD patients demonstrated significantly greater loss of cortical thickness than healthy controls in parts of the left occipital and bilateral frontal lobes and right somatomotor-sensory cortex. Cortical regions with greater connectivity (measured functionally or structurally) to a “disease reservoir” evaluated via MRI at baseline demonstrated greater atrophy one year later. The atrophy pattern in the frontal lobes resembled that described in Alzheimer′s disease. Moreover, path models suggest that cerebrospinal fluid amyloid-β42 predicted left frontal cortical thinning, which in turn was associated with reduced cognitive scores. Our findings suggest that disease propagation to the cortex follows neural connectivity, and that cognitive impairments occur earlier than previously thought in PD.


INTRODUCTION
values are expressed as mean ± standard deviation; * = p<.01; t 1 = baseline; t 2 = one-year follow-up . CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/147611 doi: bioRxiv preprint

Changes in Cortical Thickness at One-Year Follow-Up
Over the one-year period, mean whole-brain cortical thickness significantly decreased amongst both PD patients (t1 = 3.055mm±0.013; t2 = 3.027mm±0.013) and controls (t1 = 3.055mm±0.016; t2 = 3.036mm±0.017) (both p<0.01). Regional cortical thinning was greater in PD than HC ( Figure 1).  The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/147611 doi: bioRxiv preprint Figure 2. Differences in cortical thickness between PD and HC over one-year. Four peak clusters of cortical thinning in PD were identified in parts of (1)  In the context of the seven intrinsic brain networks derived from resting state fMRI by Yeo, et al. 27 , cortical regions belonging to the limbic, frontoparietal, and ventral attention networks on average demonstrated the greatest cortical thinning respectively, as well as anterior parts of the default mode network (Figure 3). There were no areas showing greater loss of cortical thickness in HC than PD. 27

Connectivity and Disease Propagation
The disease propagation hypothesis predicts that cortical brain atrophy at follow-up should depend on connectivity of cortical areas to affected areas at baseline. We had previously shown predominant subcortical atrophy in de novo PD in this population 8 . Connectivity between any cortical area and this putative subcortical "disease reservoir" should therefore predict the degree . CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/147611 doi: bioRxiv preprint of cortical thinning relative to age-matched controls when comparing follow-up to baseline. We define the disease reservoir as the areas showing atrophy in PD patients relative to controls at the initial evaluation, as determined previously using DBM 8 . For the current analysis, the brain was segmented into 463 cortical and subcortical regions of interest 28 . Each surface vertex from the cortical thickness images was mapped to the nearest volumetric parcel (see Methods). Then to examine the propagation hypothesis, the connectivity (Connij) between any cortical parcel i and any parcel belonging to the disease reservoir j was measured. Here we calculated Connij based on normal brain connectomes. The analysis was performed for the whole brain, and for each hemisphere separately. Disease exposure was defined as the product of connectivity and atrophy measure in the corresponding disease reservoir region j at the onset of the disease, summed over all possible connections.
Both functional and structural connectivity were used to test the propagation hypothesis ( Figure   5). Connectomes were generated using data from young healthy individuals to calculate the Connij. Atrophy(j) was set equal to the value at each region j of the independent component demonstrating atrophy in PD compared to HC from the previously computed DBM map from the baseline MRI 8 . Note that for areas not showing significant atrophy at baseline the value of First, a functional connectome was generated using resting state fMRI to define Connij.
Significant correlation was observed between regional cortical thinning and disease exposure (r=0.18, p<0.0002). Significance was confirmed by permutation testing where we kept the . CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/147611 doi: bioRxiv preprint connectivity structure (i.e. subcortical-cortical connections) intact and permuted the cortical thinning values (n=10,000) and the 95% confidence interval was measured using bootstrapping (n=10,000) (r ∈ [0. 10 Second, the same hypothesis was tested using structural connectivity as measured by diffusionweighted MRI (DW-MRI). Though the correlation was not significant at the whole-brain level  The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/147611 doi: bioRxiv preprint The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/147611 doi: bioRxiv preprint

DISCUSSION
The present study investigated 105 PD and 57 control participants over a one-year period following initial diagnosis. Early PD progression was associated with significant cortical changes. While aging was also associated with cortical thinning, PD patients demonstrated significantly greater reduction in cortical thickness than controls (0.028mm vs 0.019mm reduction on average). This was observed mostly in parts of the left occipital and bilateral frontal lobes and right somatomotor-sensory cortexregions largely belonging to the limbic, frontoparietal, ventral attention and default mode networks. Within PD, greater reduction in cortical thickness of the left frontal cluster over the one-year period was associated with worsening cognition (as assessed by the MoCA). Cortical neuronal degeneration could be occurring earlier in the clinical course of PD than once thought. Although PD is primarily thought of as a movement disorder, it is recognized to be a brain-wide neurodegenerative process that spreads up from brainstem into cortex, as originally suggested by Braak 29 . Indeed, dementia in PD has a point prevalence of roughly 30% and incidence rates four to six times higher than controls 30 . Our results suggest that even in a dementia-free cohort of de novo PD patients, cortical atrophy and cognitive impairment may begin within one to two years of diagnosis.
Baseline CSF Aβ42 predicted the loss of cortical thickness in the frontal clusters. The path model ( Figure 5) suggests that, among CSF measures, only reduced Aβ42 predicted left frontal cortical thinning, which was in turn associated with reduced MoCA score. Lower CSF α-syn was associated only with occipital thinning, which did not correlate with cognitive decline. The frontal clusters demonstrating greater loss of cortical thickness in PD involved lateral and medial frontal cortex, predominantly in ventral and orbitofrontal areas. These regions of frontal lobe . CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/147611 doi: bioRxiv preprint belong to the limbic system, the frontoparietal systems implicated in focused attention, and the default mode network, implicated in memory retrieval. This pattern of frontal atrophy is similar to that described in a subgroup of AD patients identified by hierarchical clustering 31 , and in patients with behavioral variant frontotemporal dementia 1 . A notable difference between the pattern of cortical atrophy described here and that commonly seen in AD is the relative absence of more posterior involvement (precuneus, posterior cingulate and parietal lobe) in our PD cohort. The effect seen here is much more anterior, possibly reflecting the propagation of disease via dopamine projections. Nonetheless, the default mode network is also disrupted in AD 32,33 , and a target of amyloid plaques 34 . The anterior default mode network may be an avenue through which dementia develops in PD. We note also that, in the Braak model of AD, the initial cortical site of amyloid-β involvement is the ventral frontal cortex 35 , and that the observed ventral frontal atrophy pattern in PD shows a high degree of similarity to early Aβ42 deposition measured by positron emission tomography in Alzheimer's Disease 36 .
While cognitive impairment is a recognized consequence of PD 9 , its underlying mechanism is not well understood. Our findings suggest that, in the PD group, higher Aβ42 deposition in the brain at baseline (indicated by lower CSF Aβ42 levels 37 ) predicted greater cortical thinning of the left frontal cluster over the one-year period. Aβ42 may make the PD brain more susceptible to cortical thinning and in turn, more prone to cognitive decline. Indeed, recent positron emission tomography studies have revealed that cortical amyloid deposits are associated, in a dosedependent manner, with increased risk for cognitive decline in PD 38 . Additionally, low CSF Aβ42 has previously been associated with risk of subsequent dementia in PD 39 . These findings provide evidence in favor of potentially overlapping mechanisms for dementia in PD and AD. Previous The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/147611 doi: bioRxiv preprint possible explanation is that synergy between Aβ42 and α-syn deposition promotes cognitive impairment in PD 41 . Aβ42 has been found to promote gray matter atrophy in distributed regions of the parietal and frontal lobe in the normal aging brain 42 , akin to the atrophy pattern observed in the present study; α-syn may accelerate the neurodegenerative process associated with normal age-related Aβ42 deposition, especially in regions connected to the PD disease reservoir.
We propose that neurodegeneration in PD results from a disease spreading process that enters the cerebral cortex via dopaminergic and basal ganglia projections to the frontal lobes. Per the network propagation hypothesis, neurodegenerative diseases result from the aggregation and propagation of misfolded proteins, which in turn results in neuronal death and brain atrophy 2,5 .
We postulate that focal cortical thinning is contingent on connectivity to a disease reservoir, by analogy to the spread of epidemics in populations: connectivity determines disease exposure and spread. In the present study, we tested this hypothesis in PD with longitudinal data. We found that cortical regions with greater connectivity (measured functionally or structurally) to the mostly subcortical disease reservoir demonstrated greater cortical atrophy over the one-year period. This is in line with our previous work using DBM 8 , which indicated that regions demonstrating the greatest atrophy were those belonging to a connectivity network with an epicenter in substantia nigra -the hypothesized origin of misfolded protein spread to the supratentorial central nervous system in PD. However, we did not observe structural interhemispheric propagation between disease reservoir and cortical areas, perhaps due to inherent limitations in DW-MRI to capture these connections 43,44 Mathematical models of this dynamic spread through brain networks, accounting for factors such as protein accumulation and clearance, have been proposed in recent years for Alzheimer's Disease 45,46 . However, further research is needed to extend these models to PD.   The current study is the largest to date to assess longitudinal measures of cortical thickness using 3T MRI in a de novo PD cohort. Our findings suggest that disease propagation to the cortex occurs earlier than previously thought. Focal cortical thinning may be contingent on connectivity to a disease reservoir: the greater the connections, the higher the disease exposure, and the more the cortical atrophy. PD progression involves subcortical α-syn spreading through intrinsic brain networks to the cortex, possibly in synergy with Aβ42, which may be the harbinger of dementia.

METHODS
Eligibility criteria and study procedures have been previously detailed 7 . Each participating PPMI site received approval from their local institutional review board, and obtained written informed consent from all subjects. Analysis of this dataset was approved by the Montreal Neurological Institute Research Ethics Board. Only participants with 3T MRI data at both initial visit and one-. CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/147611 doi: bioRxiv preprint year follow-up were included in the present study. Of the 237 PD patients with 3T acquisition at baseline, 122 had MRI data at the one-year follow-up. For controls, 118 subjects had 3T acquisition at baseline, and 62 had MRI data at the one-year follow-up.
All non-categorical data were normally distributed as per Shapiro-Wilk tests. Group difference in sociodemographic and neuropsychological variables were analyzed using the χ² test for categorical data, and t-test for normally distributed data.
Acquisition site was used as a covariate in all analyses.

Cortical Thickness
Cortical thickness models were generated from the T1-weighted 3T MRI scans using the CIVET2.1 preprocessing pipeline 50 . A summary of the steps involved follows; further details can be found online (http://www.bic.mni.mcgill.ca/ServicesSoftware/CIVET). The T1-weighted MRIs were linearly registered to the MNI -ICBM152 volumetric template 51,52 . Images were then . CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/147611 doi: bioRxiv preprint corrected for signal intensity non-uniformity 53 , and a brain mask was calculated from each input image. Images were then segmented into grey matter, white matter, CSF, and background 50 .
After the CIVET pipeline, quality control was carried out by 2 independent reviewers using previously described criteria 54 to ensure adequate quality of the T1-wieghted volume images, such as linear and nonlinear registration, and to exclude distortions in grey and white matter surfaces, and motion artifacts. 17 of 122 PD and 5 of 62 HC failed quality control.
The distance between corresponding vertices of inner and outer cortical surfaces were evaluated by a fully automated method (Constrained Laplacian Anatomic Segmentation using Proximity algorithm) 55 to provide a measure of cortical thickness at each vertex. The white and grey matter interface was fitted across 40,692 vertices in each hemisphere to create the inner cortical surface, then expanded to fit the grey matter and CSF interface to create the outer cortical surface. The surfaces of each subject were nonlinearly registered to an average surface created from the volumetric ICBM152 template. In order to test the propagation hypothesis, the brain was segmented into 448 cortical and 15 subcortical regions 28  The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/147611 doi: bioRxiv preprint correlations. Within PD patients, the difference in cognitive performance over the one-year interval was examined using path models -these tested the association between CSF measures and changes in cortical thickness, and whether cortical thickness of a specific cluster was in turn linked to cognitive performance. All data was scaled to have mean of 0 and SD of 1 such that regression coefficients were standardized and comparable with correlation statistics. Path models were built with R and lavaan, and visualised with semPlot [57][58][59] . Age at baseline and sex were used as confounding covariates for all analyses.

Deformation Based Morphometry and Independent Component Analysis
Data were analyzed from the baseline visits of newly diagnosed PD patients (n = 232) and an age-matched control group (n = 117) obtained from the PPMI database. All subjects' initial visit 3T high-resolution T1-weighted MRI scans underwent preprocessing steps including denoising 60 , intensity non-uniformity correction 53 , and intensity range normalization. Next, each subject's T1weighted image was first linearly, and then nonlinearly, registered to the MNI-ICBM152 template. Using the resulting non-linear transformation fields, we calculated deformation morphometry maps (i.e. the determinant of the Jacobian matrix of transformation). DBM maps were then concatenated and independent component analysis (ICA) was employed to define independent sources of deformation in the brain. For the resulting ICA components (n=30), the PD group was compared to healthy controls (unpaired t-test). The atrophy level significantly differed between PD and control groups in only one of these components after correcting for multiple comparisons (Bonferroni correction) suggesting this component captures PD-specific alterations in the brain. We used this component (referred to as PD-ICA here) as our atrophy . CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/147611 doi: bioRxiv preprint map. The component score (atrophy measure) in PD-ICA for each subject was significantly correlated with disease severity as measured by UPDRS-III and SBR 8 .

Connectivity and Disease Propagation
The difference in cortical thinning in PD versus controls was used to define disease progression.
DBM on the baseline MRI was employed to assess disease-related brain atrophy and to define a source, or reservoir, for disease propagation. As previously, based on a parcellation that is a subdivision of the common FreeSurfer implementation of the Desikan-Killiany Atlas, grey matter was segmented into 463 regions of interest of approximately similar size. Of these parcels, 448 were cortical and 15 were subcortical (7 in each hemisphere and 1 bilateral brainstem) 28 . Each surface vertex was then mapped to the corresponding whole-brain parcel.
To generate brain networks and connectomes, data-driven brain parcellations from young healthy individuals were used 28 . The functional connectivity map was derived from resting state fMRI data of 40 young healthy subjects (25.3±4.9yr old, 24 males) 61 . The scans were done using Siemens Medical Trio 3T MRI scanner, consisting of a high resolution T1-weighted image, as well as T2*-weighted images with BOLD contrast (3.3 mm isotropic voxels, TR 1.9s). The structural connectivity map was generated using the Illinois Institute of Technology Human Brain Atlas v.3 constructed from high resolution DW-MRI data from 72 healthy subjects (26.6±4.8yr old, 30 males) (http://www.psy.cmu.edu/~coaxlab/data.html). Both datasets were pre-processed using conventional steps consisting of rigid body motion correction, correcting for nuisance variables (white matter, cerebrospinal fluid, motion parameters), and low pass filtering. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/147611 doi: bioRxiv preprint The PD-ICA was used to assign a disease severity measure for each region at the first visit to determine a disease reservoir available for propagation 61 .
Pearson correlation was used to investigate the relationship between disease progression in cortical areas (cortical thinning) and the 'disease exposure' for each region in the PD-ICA (defined by the average values of each region weighted by connectivity). The same procedure was performed to test propagation hypotheses along the functional and structural connectomes.
The propagation analysis was performed for the whole brain and for each hemisphere separately.

CSF Measures
CSF Aβ42, α-syn, t-tau, and p-tau181 from aliquots were assessed as previously described 62,63 . The MAP Luminex platform (Luminex Corp, Austin, TX) with Innogenetics (INNO-BIA AlzBio3; Ghent, Belgium; for research use-only reagents) immunoassays was used for Aβ42, t-tau and p-tau181. CSF α-syn assay was performed at Covance using a commercially available ELISA assay kit (Covance, Dedham, MA) 64 . To evaluate the possible contamination of blood in CSFa factor thought to influence the level of some proteins including α-syn and Aβ42 65,66the relationship between CSF hemoglobin and other CSF measures was assessed. No significant association was observed (all p<0.05), suggesting that there is no effect of added blood on concentration of CSF measures. Due to the ongoing nature of the study and incomplete data, only CSF data from baseline was used in this analysis 64 .   The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/147611 doi: bioRxiv preprint