Brain atrophy progression in Parkinson’s disease is shaped by connectivity and local vulnerability

Abstract Brain atrophy has been reported in the early stages of Parkinson’s disease, but there have been few longitudinal studies. How intrinsic properties of the brain, such as anatomical connectivity, local cell-type distribution and gene expression combine to determine the pattern of disease progression also remains unknown. One hypothesis proposes that the disease stems from prion-like propagation of misfolded alpha-synuclein via the connectome that might cause varying degrees of tissue damage based on local properties. Here, we used MRI data from the Parkinson Progression Markers Initiative to map the progression of brain atrophy over 1, 2 and 4 years compared with baseline. We derived atrophy maps for four time points using deformation-based morphometry applied to T1-weighted MRI from 120 de novo Parkinson’s disease patients, 74 of whom had imaging at all four time points (50 Men: 24 Women) and 157 healthy control participants (115 Men: 42 Women). In order to determine factors that may influence neurodegeneration, we related atrophy progression to brain structural and functional connectivity, cell-type expression and gene ontology enrichment analyses. After regressing out the expected age and sex effects associated with normal ageing, we found that atrophy significantly progressed over 2 and 4 years in the caudate, nucleus accumbens, hippocampus and posterior cortical regions. This progression was shaped by both structural and functional brain connectivity. Also, the progression of atrophy was more pronounced in regions with a higher expression of genes related to synapses and was inversely related to the prevalence of oligodendrocytes and endothelial cells. In sum, we demonstrate that the progression of atrophy in Parkinson’s disease is in line with the prion-like propagation hypothesis of alpha-synuclein and provide evidence that synapses may be especially vulnerable to synucleinopathy. In addition to identifying vulnerable brain regions, this study reveals different factors that may be implicated in the neurotoxic mechanisms leading to progression in Parkinson’s disease. All brain maps generated here are available on request.


Introduction
Atrophy has been reported in multiple brain regions, even in the early stages of Parkinson's disease. [1][2][3] However, less is known about the progression of tissue loss after diagnosis. 4 There have been a few studies of MRI-derived measures over time, but they were limited by small sample sizes, 5,6 short follow-up durations 7,8 or use of global volumetric measures. [9][10][11] The most consistent findings are subcortical tissue loss early on, notably affecting striatum, thalamus, amygdala and hippocampus, and more widespread cortical atrophy when longer durations or larger sample sizes are used. With few exceptions, 7 no study has attempted to relate the spatial pattern of atrophy progression to intrinsic properties of the brain, such as anatomical connectivity, cellular composition or regional gene expression. Such an analysis could provide information about pathophysiological mechanisms in Parkinson's disease.
There is increasing evidence that brain connectivity may shape tissue loss in a multitude of neurodegenerative and psychiatric diseases, including schizophrenia, frontotemporal dementia, amyotrophic lateral sclerosis, Alzheimer's and Parkinson's disease. 7,[12][13][14][15][16][17][18] This pattern points to the possibility of pathogenic agents spreading from neuron to neuron as a common mechanism. In Parkinson's disease, converging evidence points to misfolded a-synuclein as the propagating agent. [19][20][21][22][23][24] Most a-synuclein aggregates are localized in synapses where they are thought to interfere with neurotransmitter release and cause dendritic spine loss 25,26 that may be detectable using MRI-based measures, such as cortical thickness or deformation-based morphometry (DBM). Indeed, using a single time point or follow-up, we previously demonstrated that the MRIderived baseline atrophy pattern and the atrophy progression after 1 year in de novo Parkinson's disease are consistent with a connectome-based propagation, 2,7,16 a finding that has been replicated computationally in animal models. 24 However, the role of connectivity in explaining the longitudinal progression of atrophy in Parkinson's disease has not yet been tested in human patients followed longitudinally with multiple MRI time points.
The progression of brain atrophy is likely not only explained by brain connectivity but also by regional variations in tissue vulnerability. 16,27 Indeed, some brain areas with higher a-synuclein concentration appear to be more susceptible to injury but also more likely to act as disease propagators. 16,28,29 Regional vulnerability may additionally depend on the local cellular distribution or gene expression patterns that render the milieu to be more neurotoxic or neuroprotective. [30][31][32] For instance, microglia have been implicated in both neuroprotection and neurotoxicity via regulation of inflammation. 32,33 Astrocytes have been shown to accumulate a-synuclein and produce proinflammatory cytokines and chemokines. 31 Oligodendrocytes may be targeted by neurodegeneration but can also play a role in neuroprotection via the synthesis of neurotrophic factors. 34 Endothelial cells have also been implicated in the progression of neural damage in Parkinson's disease 35 and blood-brain barrier (BBB) dysfunction is thought to play a role in the pathology associated with Parkinson's disease and other neurodegenerative diseases. 36,37 On the other hand, endothelial cells have also been shown to contribute to neuron survival under certain physiological and inflammatory conditions. 38,39 Finally, the ratio of excitatory to inhibitory neurons could be relevant for disease pathology, as excitotoxicity has also been implicated in Parkinson's disease, possibly in synergy with glial dysfunction. 40 A virtual histology approach may help explain how cellular composition relates to local vulnerability to atrophy. 41 How brain architecture, local cell-type distribution and gene expression combine to determine the pattern of disease progression remains unknown. Here, we first mapped the progression of brain atrophy using DBM and then related it to structural and functional networks, celltype composition and gene ontology (GO) enrichment analyses. We hypothesized that the constraints imposed by structural and functional networks as well as the regional differences in cell-type composition would jointly shape the pattern of atrophy found in Parkinson's disease. We show that the brain's connectome significantly shapes the course of atrophy over 4 years and that oligodendrocytes and endothelial cells may play a role in neuroprotection, while GO analysis points to synapses as an important target of neurodegeneration.

Data acquisition
Participants Table 1 describes the clinical characteristics of the participants with Parkinson's disease included in the present study at each time point (i.e. baseline, 1, 2 and 4 years). Participants met the inclusion criteria for PPMI (http://www. ppmi-info.org/study-design/, Last accessed, February 2021) described in the Supplementary material (p. 1). Participants with a history of Parkinson's disease medication use, including L-Dopa, dopamine agonists, monoamine oxidase B inhibitors or amantadine, within 60 days of the baseline visit, or with a diagnosis of dementia 43 were excluded from the study. HC participants in the PPMI database were aged !30 years old at the screening visit with no current neurological disorder. Exclusion criteria for HC were a Montreal Cognitive Assessment (MoCA) score 26 or a first-degree relative with idiopathic Parkinson's disease. There were a total of 403 patients with Parkinson's disease and 184 HCs with a T 1 -weighted MRI at baseline. After visual inspection to exclude images with abnormalities, 74 patients (50 men and 24 women) had both MRI and clinical evaluation at each of the three follow-up time points and 157 HCs (115 men and 42 women) at baseline. Most exclusions were due to missing assessments at one of the time points. In addition, neuroimaging data passing quality control were available for 120 Parkinson's disease participants at baseline and 1 year and 109 at 2 years, and these were used to confirm our findings with larger sample sizes. We ensured that the control group was similar to the Parkinson's disease group (n ¼ 74) relative to age at baseline (HC mean ¼ 60.1 years, range: 31-83; Parkinson's disease patients mean ¼ 60.2 years, range: 38-82 years; P ¼ 0.96) and sex (v 2 ¼ 0.09, P ¼ 0.77).
We also assessed the sample for attrition bias, whereby more severely affected individuals tend to drop out of a longitudinal study. We used our previous classification of the patients in this dataset into three severity groups based on their baseline brain atrophy, which yielded mild, intermediate and severe subtypes. 44 The drop-out rates were 38, 39 and 58% for mild, intermediate and severe subtypes at 2 years (v2 ¼ 5.8, P ¼ 0.05) and 60,57, and 70% at 4 years (v2 ¼ 2.5, P ¼ 0.30). This suggests that our dataset may have included less severely affected participants (compared to the entire de novo sample), especially at the 2-year time point. Such a bias could obscure associations (e.g. between symptom severity and brain atrophy). 45

Brain structural analysis
Deformation-based morphometry DBM quantifies voxel-wise brain tissue atrophy by performing non-linear transformations from the participant's brain to a template brain. DBM maps were derived from each participant's T 1 -weighted MRI image at each time  46 implemented in Statistical Parametric Mapping software (SPM12) (see Supplementary material for the detailed protocol, p. 2). The preprocessing and DBM calculation were performed separately for each T 1 -weighted MRI image. Every voxel value represents the factor by which each voxel of the participant's brain has to expand (positive value) or shrink (negative value) to be registered to the MNI template.
W-score maps were generated to account for age and sex on brain deformation. 47,48 A W-score map was computed for each time point (baseline, 1, 2 and 4 years) using the following formula, at each voxel: where W i is the W-score at voxel i for a participant, DBM i is the computed DBM value at voxel i, DBM exp is the expected DBM value for that participant defined by (b1*age þ b2*sex þ b3), and derived from HC data. All statistical analyses involving DBM were performed using the W-scored DBM maps. Negative W-score values indicate atrophy (reduced volume) whereas positive W-scores indicate expansion, with normal ageing and sex effects accounted for. Only the baseline HC data were used as a reference since only a limited number of HC had longitudinal measurements. 49

Brain atrophy progression analysis
The pattern of atrophy progression was investigated voxel-wise using two SPM12 toolboxes: CAT12 46 and the probabilistic approach for threshold-free cluster enhancement (pTFCE), 50 written in MATLAB (R2018b). A two-tailed repeated measures ANOVA (F-contrast) with time as covariate followed by post hoc paired comparisons [t-contrasts between baseline (Bl) and 1 year, Bl and 2 years, Bl and 4 years] combined with the pTFCE based on Bayes' rule were used to compare atrophy differences between time points in the Parkinson's disease group. In all cases, W-score maps were used. Although sex was accounted for in the W-score calculation, we added it as a covariate in the repeated measures ANOVA to regress out its effect, as sex differences in atrophy were recently demonstrated in Parkinson's disease. 49 Scanner site was also added as a covariate. Family-wise error (FWE) rate (threshold: P FWE < 0.05) was used to correct for multiple comparisons. Only significant clusters with 10 voxels or more were retained. This analysis was computed for the 74 participants with a T 1 -MRI at each of the four time points. Additional analyses (t-contrasts between Bl and 1 year or Bl and 2 years) were done with all the patients who had a T 1 -MRI at either baseline and 1 year (N ¼ 120) or baseline and 2 years (N ¼ 109). Afterwards, the Cammoun atlas 51 was used to find the anatomical location of the significant clusters in the brain.
To understand how tissue changes relate to brain function, the mean atrophy progression was computed for each of the seven cortical resting-state networks defined by Yeo et al. 52 These networks are thought to represent the distributed neural systems that support diverse cognitive domains. 53 One-sample permutation tests were computed to compare the mean atrophy progression of each network to its null distribution (two-sided P-value). Statistical significance was estimated by permutation testing using the netneurotools toolbox (https://netneurotools. readthedocs.io/en/latest/index.html, Last accessed, August 2021): network labels were permuted 1000 times while preserving the spatial autocorrelation, and network-specific means were recalculated to generate a null distribution for each network. 54,55 To investigate the relationship between atrophy progression and the synaptic organization and hierarchy of different zones of the cortex, the cortical regions were defined following the nomenclature of Mesulam. 56 The mean atrophy progression scores for each cortical tissue type (paralimbic, heteromodal, unimodal and idiotypic) were compared using a one-way ANOVA and post hoc tests with Bonferroni corrections.
Correlations were computed to investigate the relationship between clinical measures (motor and non-motor symptoms, mood and CSF biomarkers) and atrophy progression in the whole brain, in the significant clusters and in each of the seven resting-state networks (threshold: P FDR < 0.05, two-tailed). Partial Spearman's correlations were used due to some data being non-normally distributed, with age, sex and education as covariates (only age and sex were used for motor symptoms).

Structural and functional network analysis
We next investigated the network spread hypothesis of Parkinson's disease. We tested whether the atrophy progression observed in each region was correlated with the atrophy progression of its structurally and functionally connected neighbourhood. In brief, (i) the brain was parcellated into equally sized cortical regions at four different spatial resolutions 51 ; (ii) the mean atrophy progression was calculated for each region; (iii) matrices of structural and functional connectivity between regions were constructed, using diffusion-weighted MRI tractography and resting state functional MRI acquired on a different dataset of 70 healthy adults 57 ; (iv) the collective neighbourhood atrophy progression of each region was calculated, using the structural and functional networks to define the neighbours; (v) correlations were computed between the atrophy progression in each region and its collective neighbourhood atrophy progression; and (vi) the significance of the correlations was tested against a null model preserving spatial autocorrelation. (See the detailed protocol in the Supplementary material, p. 2-4.). This network analysis approach was previously used in schizophrenia. 15 Complementary analysis on the role of functional connectivity Additional analyses were performed to verify whether FC, without considering the structural connections between regions, is related to atrophy progression. Pearson's correlations were calculated between the regional atrophy progression and the collective deformation weighted by the FC of (i) the non-structurally connected direct neighbours (i.e. only region pairs not demonstrating structural connectivity, as defined above) and (ii) all the regions irrespective of their structural connections. These correlations were compared against spatial autocorrelation null models (1000 spins, two-tailed). The cocor package was used to statistically compare the magnitude of the correlations and calculate a Zou's confidence interval. 58 Cell-type analysis

Virtual histology
We next investigated if the atrophy progression was associated with the prevalence of specific cell types in the cortex, notably astrocytes, endothelial cells, microglia, excitatory and inhibitory neurons, oligodendrocytes and oligodendrocyte precursors. A virtual histology approach was used to correlate the neuroimaging data with cellspecific gene expression across brain regions. 59,60 Every class of cell type was associated with a gene list first derived by Seidlitz et al., 41 from five single-cell RNA sequencing studies of post-mortem human cortical samples [61][62][63][64][65] (for the detailed protocol, see Hansen et al. 66 ). To generate cortical maps for each cell type, the spatial expression patterns of these gene lists were derived from post-mortem brain data from six donors available in the Allen Human Brain Atlas (AHBA) genetics dataset. 67 Cell-type distribution was computed for each region of the Cammoun atlas 51 at four different resolutions (68,114, 219 and 448 cortical regions) using the abagen toolbox (https://github.com/rmarkello/abagen, Last accessed, August 2021 ) and the recommendations laid out by Arnatkeviciute et al. 68,69 Pearson's correlations were calculated between the atrophy progression of each region and the region's average gene expression of each cell class. All the correlations were tested against null models preserving spatial autocorrelation (1000 spins; twotailed). 55

Null model
The null models used in this study preserved the spatial autocorrelation between regions using the netneurotools toolbox. 55 This approach generates a null model by projecting the brain regions onto a sphere and randomly rotating the sphere (see Markello and Misic 70 for a comparison of different spatially constrained null models). Briefly, a surface-based representation of the different resolution parcellations of the Cammoun atlas on a FreeSurfer (release v6.0.0) surface was created. The spherical projection of the surface was then used to define spatial coordinates for each region by selecting the vertex closest to the center of mass of the region. The spatial coordinates were used to create null models by rotating and reassigning region values 1000 times. The spatial rotation was performed at the parcel resolution and in one hemisphere before being mirrored to the other. For the correlation analysis, the Pearson's correlation coefficients (observed values) were tested against the distributions of correlation coefficients derived from the null networks (permuted values).

GO enrichment analysis
A GO enrichment analysis was performed to explore the biological processes related to atrophy progression over 2 and 4 years using brain regional gene expression. To do this, we extracted the average gene expression value for all genes available in the AHBA genetics dataset (i.e. 15 633) 67 for each of the 448 cortical regions of the Cammoun atlas using the abagen toolbox and following previous recommendations. 68,69 For investigating the functions of the genes associated with atrophy, we only selected the genes whose expression significantly correlated with atrophy progression after false discovery rate (FDR) correction and when compared to a null model preserving the spatial autocorrelation (1000 spins; twotailed). This yielded lists of genes whose expression pattern was positively or negatively correlated with atrophy progression at 2 (positive correlation: N ¼ 619 genes, negative correlation: 1058 genes) and 4 years (positive correlation: N ¼ 900 genes, negative correlation: 1546 genes). We next investigated if the proportion of GO terms for the genes correlated to atrophy (i.e. the target gene list) significantly differed from the proportion of GO terms found with all genes extracted from the AHBA (i.e. background gene list). To ensure that the results were not due to the choice of a particular classification system, two publicly available GO platforms were used to obtain GO terms: the Gene Ontology enRIchment anaLysis and visuaLizAtion tool (GOrilla) 71 and the PANTHER Classification System. 72 Of the 15 633 genes available in the AHBA genetics dataset, 13 992 and 14 657 genes were associated with a GO term in the GOrilla (GO Process) and PANTHER (GO biological process) platforms, respectively. Supported gene IDs are available from the GOrilla (http://cbl-gorilla.cs.technion. ac.il, Last accessed, February 2021) and PANTHER (www.pantherdb.org, Last accessed, February 2021) websites. For both platforms, a statistical overrepresentation analysis was conducted with Bonferroni correction to control for multiple comparisons. Whereas a hypergeometric model was implemented in GOrilla, the Fisher's exact test was used in PANTHER.

Clinical measures and CSF biomarkers
For each feature and CSF biomarker presented in Table 1, two-tailed repeated measures ANOVA with post hoc comparisons and Bonferroni-Holm correction 73 were used to evaluate the progression after 1, 2 and 4 years. All measures of motor severity worsened over time, namely Hoehn and Yahr stage (P < 0.00005), total UPDRS-III (P ¼ 0.002), tremor (P ¼ 0.02) and rigidity (P < 0.00005). At each follow-up, the Hoehn and Yahr stage and total rigidity score were significantly different from baseline while the total UPDRS-III showed a significant difference only after 4 years (P-adjusted ¼ 0.004). Of the non-motor measures, only the Scale for outcomes in PD ( SCOPA)-autonomic and phonemic fluency scores (Letter F from MoCA) showed a significant difference with time (P-adjusted < 0.00005). Both scores were significantly different from baseline at each follow-up. While the autonomic score worsened with time, an improvement was observed in phonemic fluency possibly reflecting a medication and/or learning effect (note that participants were not on anti-parkinsonian medications at baseline). Among the CSF biomarkers, only the neurofilament light polypeptide, a possible biomarker of neurodegeneration, 74,75 presented a significant increase with time (P ¼ 0.0001), more specifically after 1 (P-adjusted ¼ 0.002) and 2 years (P-adjusted ¼ 0.001).

Brain atrophy
The DBM W-Score maps were used to analyse voxel-wise atrophy progression after 1, 2 and 4 years in the participants with Parkinson's disease. This analysis showed an effect of time (F-contrast, two-tailed) in 24 clusters. The Cammoun atlas 51 was used to localize the significant clusters ( Supplementary Tables 1-3), which were widely distributed throughout the brain. Atrophy progression was mostly found in the caudate, nucleus accumbens, hippocampus, amygdala and the temporal, parietal, occipital and cingulate cortex.

Post hoc analysis
Post hoc tests (t-contrasts, one-tailed) were also computed between baseline and the three follow-up time points. Figure 1A shows the regions with a significant atrophy progression after 2 and 4 years. No significant clusters presented significantly greater atrophy after 1 year. After 2 years, 11 clusters were significant (Supplementary Table 2), located in the bilateral temporal lobe, left precuneus, right caudate and angular gyrus. After 4 years, significant atrophy progression was found in 13 clusters (Supplementary Table 2), located in the right caudate, bilateral nucleus accumbens and temporal, parietal, occipital and cingulate cortex. Small clusters in the superior and inferior frontal gyrus were also found.
The previous results were derived from the participants who had scanning sessions at all time points. We repeated the 1-and 2-year analyses using the larger sample sizes of individuals who have not undergone scanning at all time points. A high overlap was observed between the regions with a significant atrophy progression at 2 years in the full 4-year sample (described above) and the regions with atrophy progression found after including participants with a T 1 -MRI at baseline and 1 year (N ¼ 120), or baseline and 2 years (N ¼ 109) (Supplementary Fig. 1). Small clusters of atrophy progression were observed after 1 year in the additional analysis with 120 participants and larger clusters were found after 2 years with 109 participants (Supplementary Table 3).
Additional analysis, excluding subjects (HC and patients with Parkinson's disease) <50 years old was also performed to verify the impact of including younger subjects, but similar patterns of atrophy progression after 2 and 4 years were found ( Supplementary Fig. 2).
To investigate the relationship with the synaptic hierarchy in the cortex, the average atrophy progression was compared between the paralimbic, heteromodal, unimodal and idiotypic cortical areas (Fig.  1B) as defined by Mesulam. 56 There was a significant effect of cortex type for the atrophy progression Post hoc comparisons indicated that this was due to significantly less atrophy progression in idiotypic (primary sensory and motor) cortex than other types. The mean atrophy progression values of the idiotypic areas after both 2 (M ¼ 0.0009, SD ¼ 0.02) and 4 years (M ¼ 0.008, SD ¼ 0.03) were significantly lower compared to the atrophy progression in other areas (P < 0.00005).
To further localize the disease process, the average atrophy progression was computed for each of the seven resting-state networks defined by Yeo et al. 52 and compared against a null distribution preserving the spatial autocorrelation (1000 spins; two-tailed) (Fig. 1C). The mean atrophy progression was significant after 2 years in the limbic (P spin ¼ 0.001), default mode (P spin ¼ 0.002) and visual (P spin ¼ 0.002) networks. After 4 years, the mean atrophy progression was also significant in the dorsal attention (P spin ¼ 0.001), frontoparietal (P spin ¼ 0.001), limbic (P spin ¼ 0.001), default mode (P spin ¼ 0.001) and visual (P spin ¼ 0.001) networks. Only the somatomotor (P spin ¼ 0.42) and ventral attention (P spin ¼ 0.10) networks did not present a significant atrophy progression after 4 years. Overall, the limbic and default mode networks were most affected.

Relationship with clinical measures and CSF biomarkers
Correlation analyses were performed to investigate how the mean atrophy progression after 2 and 4 years related to 10 clinical characteristics [MDS-UPDRS-III (total), the non-motor features and depressive symptoms] and all the CSF biomarkers presented in Table 1. Additional correlations with the different clusters of significant atrophy progression after 2 and 4 years were also computed (Supplementary Tables 4-6). Partial Spearman's correlations controlling for the effect of age, sex, and education were not significant before (P < 0.01; two-tailed) or after FDR corrections (P FDR < 0.05; twotailed). When investigating the association between the mean atrophy progression of the seven resting-state networks after 2 and 4 years and the change in clinical measures over the same time period, no significant correlations were found after correcting for multiple comparisons.

The influence of structural and functional connectivity
We next investigated if there was a network-specific distribution of atrophy progression. If this is the case, brain regions structurally and/or functionally connected with Cortical regions were defined following the nomenclature of Mesulam based on synaptic organization and hierarchy. 56 The mean atrophy progression score and the standard error of the mean were calculated for each cortical zone. The idiotypic areas (primary sensory/ motor cortex) showed a lower atrophy progression than the other cortical areas (*P < 0.05). (C) The atrophy progression pattern after 2 and 4 years was assigned to the seven resting-state networks defined by Yeo et al. 52 The mean atrophy progression score and the standard error of the mean was calculated for each network and compared to a null distribution (1000 spins; two-tailed) preserving spatial autocorrelation (*P < 0.05).
neighbours showing more atrophy progression should also present more atrophy progression after 2 or 4 years. For both structural and functional connectivity, significant correlations against spatial null models were found between the atrophy progression of a region and the atrophy progression of its connected neighbours after 2 years (structural: r ¼ 0.61, P-value spin ¼ 0.0001; FC: r ¼ 0.62, P-value spin ¼ 0.0001; two-tailed) and 4 years (structural: r ¼ 0.61, P-value spin ¼ 0.0001; FC: r ¼ 0.64, P-value spin ¼ 0.0001; two-tailed) at a parcellation of 448 cortical regions (Fig. 2). The results were replicated at all other spatial resolutions (Supplementary Fig. 3).
FC from resting state fMRI is constrained by structural connectivity. 76 To verify more specifically its influence on atrophy progression, we investigated the relationships between the regional atrophy progression and the collective deformation weighted by the FC of (i) the non-structurally connected neighbours and (ii) all the regions irrespective of their structural connections. Table 2 shows that, in both cases, correlations were markedly reduced. These results were replicated at the other spatial resolutions. Altogether, this finding suggests that both structural and functional connectivity are associated with atrophy progression, but that neuronal activity modulates co-atrophy between structurally connected regions.

Relationship with cell-type distribution
In addition to connectivity, local factors may also influence atrophy progression in Parkinson's disease. Here, we investigated the relationship between the relative regional prevalence of different cell types and cortical atrophy progression. Two cell types were associated with relatively less atrophy progression. Using the Cammoun atlas with 448 regions, significant negative correlations were found between the prevalence of endothelial cells and atrophy progression after 2 (r ¼ À0.15, P-value spin ¼ 0.04, twotailed) and 4 years (r ¼ À0.19, P-value spin ¼ 0.01, twotailed) (Fig. 3). The atrophy progression was also inversely correlated with the prevalence of oligodendrocytes after 2 (r ¼ À0.11, P-value spin ¼ 0.049, two-tailed) and 4 years (r ¼ À0.11, P-value spin ¼ 0.04, two-tailed). In addition, negative correlations were obtained with three lower resolutions [68,114 and 219 regions (R)] (see Supplementary  Fig. 4). Only the correlations between the atrophy Figure 2 Atrophy progression relationship with structural and functional connectivity. (A, C) Regional atrophy progression was significantly correlated with the atrophy progression of the structurally connected (SC) neighbouring regions after 2 (A) and 4 years (C) compared to their spatial null coefficients distribution. (B, D) Regional atrophy progression was also related to the atrophy progression of the neighbouring regions weighted by their FC after 2 (B) and 4 years (D). These correlations were also tested against a null coefficients distribution using a model preserving the spatial autocorrelation between regions. All the correlations showed were calculated using the Cammoun atlas with 448 regions, but similar results were obtained with three other resolutions (Supplementary Fig. 3). (A, C) The regional prevalence of endothelial cells in the brain was significantly correlated with lower atrophy progression observed after 2 (A) and 4 years (C) compared to their spatial null coefficients distribution. (B, D) A negative correlation was also found with the prevalence of oligodendrocytes for atrophy progression after 2 (B) and 4 years (D) after testing against their spatial null coefficients distribution. All the correlations showed were calculated using the Cammoun atlas with 448 regions, but negative correlations were also obtained with three lower resolutions ( Supplementary Fig. 4). progression after 4 years and the endothelial cells were significant at the three other resolutions (68R: r ¼ À0.33, Pvalue spin ¼ 0.04; 114R: r ¼ À0.33, P-value spin ¼ 0.003; 219R: r ¼ À0.25; P-value spin ¼ 0.049) while the correlations between the atrophy progression after 4 years and oligodendrocytes were near statistical significance (68R: r ¼ À0.23; P-value spin ¼ 0.06; 219R: r ¼ À0.14; P-value spin ¼ 0.06) or significant (114R: r ¼ À0.21; P-value spin ¼ 0.049). No significant correlation was found for the other five cell types (astrocytes, microglia, excitatory and inhibitory neurons and oligodendrocyte precursors) ( Supplementary Fig. 5), and no cell type was associated with greater progression. These results seem to indicate slower cortical atrophy progression in regions with a higher prevalence of endothelial cells or oligodendrocytes.

Relationship with specific biological processes
To determine the functions of the genes whose expression was spatially associated with atrophy progression, a GO enrichment analysis was done. Two platforms, Gorilla and PANTHER, were used. The results from each platform consistently implicated terms related to synaptic function. Figure  4 shows the significant GO terms (Pvalue Bonferonni < 0.05, two-tailed) from the genes positively associated with the atrophy progression after 2 (N ¼ 23) and 4 years (N ¼ 17), and their average fold enrichment. Individual platform results are shown in Supplementary  Figs 6 and 7. The GO enrichment analysis revealed processes related to synaptic activity (regulation of synaptic plasticity Fig. 4A and B, fold enrichment ¼ 3.61 and 3.33), chemical synaptic transmission ( Fig. 4A and B, fold enrichment ¼ 3.16 and 2.67) and cell signalling, namely trans-synaptic signalling ( Fig.  4A and B, fold enrichment ¼ 3.12 and 2.62), and cell-cell signalling ( Fig. 4A and B, fold enrichment ¼ 2.21 and 2.05) from the genes related to the atrophy progression after 2 (13 GO terms) and 4 years (10 GO terms). In sum, the GO analysis revealed that the regions showing greater atrophy progression tend to have greater expression of genes implicated in synaptic activity and cell signalling.
Additional analysis shows that 33 genes uncovered in this analysis were located in known Parkinson's disease GWAS loci 77 (Supplementary Tables 7 and 8), but future studies would be needed to examine whether these specific genes are driving the associations of the relevant loci with Parkinson's disease atrophy.

Discussion
Progressive brain atrophy was observed after 2 and 4 years in de novo Parkinson's disease after regressing out the effects of sex and normal ageing. The atrophy distribution was widespread, affecting five of the seven canonical resting-states networks 52 at 4 years. In line with the Braak hypothesis, 19 cortical atrophy progression was determined by brain connectivity, supporting trans-neuronal propagation of the pathogenic process. Atrophy was greater in regions enriched for genes related to synaptic activity and signalling. In addition, cortical regions containing relatively more oligodendrocytes and endothelial cells had reduced vulnerability to atrophy. These findings may point to the biological mechanisms at play in neurodegeneration in Parkinson's disease. More generally, they support a model according to which neurodegeneration results from the interaction of a propagating agent with regional vulnerability.

Brain atrophy progression
Evaluating the progression of atrophy in the whole brain using DBM measures, we found small regions over the temporal and parietal lobe in addition to the right caudate nucleus presenting an atrophy progression after 2 years, while a more widespread atrophy progression in the caudate nucleus, nucleus accumbens and the temporal, parietal, occipital and cingulate cortex was observed after 4 years. This mostly posterior cortical pattern of atrophy progression is similar to the atrophy distribution observed in the ENIGMA-Parkinson's Study including 2367 patients with Parkinson's disease and 1183 HC. 78 Our findings showed alterations in the default mode, limbic, dorsal attention, frontoparietal and visual networks over this time. No relationship was found between the atrophy progression in any of the resting-state networks and the change in clinical symptoms. One explanation may be that the clinical measures available did not specifically target the regions showing atrophy progression. Indeed, motor deficits are mainly due to dopamine deficiency in the substantia nigra, 79 while autonomic dysfunction is mostly associated with brainstem and spinal neuronal loss in Parkinson's disease. 80 The substantia nigra may not show atrophy progression due to a floor effect (cell loss being already severe at presentation), while the brainstem is difficult to assess with DBM. Cortical atrophy patterns may be expected to relate to cognitive impairment; however, no cognitive decline was observed in this patient group, possibly due to a form of attrition bias affecting the PPMI cohort, whereby more severely affected individuals tended not to undergo the later MRI sessions. This may also explain why atrophy did not impinge on frontal areas in this sample. It is expected that more broadly distributed atrophy progression in later stages of the disease will eventually lead to cognitive deficits. 81

Structural and functional connectivity mediates atrophy progression
We have previously shown that atrophy in de novo Parkinson's disease targets an intrinsic brain network, 2 supporting the network spread hypothesis. Here, atrophy progression significantly affected five of the seven restingstate networks; however to test the role of connectivity, we asked whether cortical regions with greater atrophy progression were more likely to be structurally and functionally connected with each other. Correlations were computed against spatial null models to control for the intrinsic relationships between proximal cortical regions.
Our results indicate that both structural and functional connectivity are related to the atrophy progression after 2 and 4 years, even after controlling for regional autocorrelation in the brain. This is in line with studies suggesting that atrophy in Parkinson's disease progresses via neuronal connectivity, 2,7,14,16,21 most likely reflecting the spread of misfolded a-synuclein via neuronal projections. 19,22,23,82 Moreover, our findings are in accordance with previous work showing that the atrophy pattern in other neurodegenerative diseases follows structural and functional brain network architecture. 12,15,23,83 Our previous work with an agent-based spreading model demonstrated that the pattern of brain atrophy depends on both neuronal connectivity and regional factors. 16 Local vulnerability to neurodegeneration could depend on several factors including the prevalence of specific cell types. [30][31][32] We next used gene-expression data to test for these local factors.
Cellular composition may reflect regional vulnerability Using gene expression associated with different cell types in the healthy brain, 41 the relationships between the prevalence of seven cell types (astrocytes, endothelial cells, microglia, excitatory and inhibitory neurons, oligodendrocytes and oligodendrocyte precursors) and atrophy progression were investigated. Only significant negative correlations were observed between regional atrophy progression and the prevalence of oligodendrocytes and endothelial cells in the cortex, suggesting a possible neuroprotective effect of both of these cell types. Our finding is in line with studies showing that oligodendrocytes promote neuronal survival by secreting trophic factors, such as brain-derived neurotrophic factor (BDNF), enhancing neuronal survival and axon regeneration, insulin-like growth factor 1 (IGF1), increasing the survival of young and aged cortical neurons and glial cell line-derived neurotrophic factor (GDNF), known for its neuroprotective effect on dopaminergic neurons. [84][85][86][87][88] However, little is known about the role of oligodendrocytes in Parkinson's disease. A small number of a-synuclein inclusions has been observed in non-myelinating oligodendrocytes that may eventually contribute to neuronal death, 31 but the production of neurotrophic factors by both myelinating and non-myelinating oligodendrocytes may override this neurotoxic effect. Oligodendrocyte genes have also been related to cortical synaptic density and likely play a role in synaptic elimination and maturation. 89 The role of endothelial cells and the BBB in the progression of Parkinson's disease is unclear. Some studies suggest that BBB alterations could contribute to neuropathological damage 36,37 while others showed that endothelial cells contribute to neuronal survival under both physiological and inflammatory conditions. 38,39 Endothelial cells are part of a neurovascular niche that supports neurogenesis in the adult brain. [90][91][92] In addition, many angioneurins (including BDNF, IGF1 in addition to nerve, vascular endothelial, hepatocyte and epidermal growth factors) have receptors expressed by endothelial cells and have been shown to have neuroprotective effects in different neurodegenerative disease models, including Parkinson's disease. 93 Angioneurins affect both neural and vascular cell function and regulate angiogenesis, BBB integrity, neuroregeneration, neuroprotection and synaptic plasticity. 93 Even though statistically significant, the correlations observed with both oligodendrocytes and endothelial cells were modest, indicating a need for further studies with animal models, larger samples of Parkinson's disease patients and later disease stages.

Regions prone to atrophy are enriched for synaptic genes
Regions with greater atrophy progression were enriched for genes implicated in synaptic plasticity, chemical synaptic transmission, trans-synaptic signalling and cell-cell signalling. This is consistent with post-mortem evidence that a-synuclein aggregates are mostly located in synapses. 25 Indeed, Lewy neurites, consisting of a-synuclein aggregates in presynaptic terminals, are a hallmark of Parkinson's disease pathology, 94 and one of the many functions of a-synuclein is to regulate synaptic homeostasis. Cell-culture experiments show that pathogenic a-synuclein fibrils first target pre-synaptic terminals and alter synaptic protein levels. 95 Thus, it is likely that misfolded a-synuclein first accumulates at synapses, causing impaired neurotransmission followed by synaptic and eventual axonal and neuronal death. 26 Pathological findings are mirrored by human imaging evidence showing that diffusion weighted MRI measures sensitive to neuritic damage are consistently abnormal in Parkinson's disease. 49,[96][97][98] Finally, synaptic and neuritic death should be associated with reduced neuropil and hence tissue loss, which should be reflected in the DBM measures used here.
It has also been demonstrated that a-synuclein can be transmitted from cell to cell in a prion-like manner. 23,99 Two findings from our work support the a-synuclein synaptic dysfunction and synaptic spreading hypothesis: (i) the regions with greater atrophy progression were enriched for genes related to synaptic activity, cell-cell communication and signalling and (ii) atrophy was dependent on connectivity.
The link to synaptic proteins may also explain the relative distribution of tissue loss along the sensory-limbic cortical hierarchy. We found that primary sensory and motor cortical areas exhibited significantly less progressive atrophy than transmodal or paralimbic cortex, in keeping with Braak's observations. 19 Primary areas have lower synaptic density, 100 but they also exist at the periphery of the brain connectome, 56 and are the most distant from limbic and default mode areas that appear to be affected early in the course of the disease. Thus, connectivity and local features may both account for the fact that primary areas are affected in the last stage of Parkinson's disease.

Strengths and limitations
Strengths of the study include the use of DBM, which allows the detection of both cortical and subcortical changes and has been shown to be more sensitive than VBM to subcortical volume loss in early Parkinson's disease. 2,101 Secondly, the atrophy progression measure accounts for the expected effects of sex and normal ageing by using the W-score approach. Thirdly, using a longitudinal design with the same participants at all time points limits drop-out effects that may affect brain measures derived from a different number of participants at each time point. 9,44,102 Given the well-known clinical heterogeneity in participants with Parkinson's disease, [103][104][105] looking at the data from the same participants in a longitudinal study ensures greater consistency and reduces confounding factors. Fourthly, having multiple follow-up time points (1, 2 and 4 years) also gives a more precise estimate of the trajectory of atrophy progression through time.
Some limitations should also be acknowledged. Followup duration for the clinical and neuroimaging data is relatively short (4 years) and longer follow-up should help to determine the relationships between atrophy progression and cognitive dysfunction, mood symptoms and CSF biomarkers. Moreover, the PPMI database includes measures from multiple centres which may lead to site or scanner specific biases. However, PPMI has strict guidelines and protocols to acquire the clinical and imaging data to ensure standardization 42 and scanner site was used as covariate in all our analyses. Also, we combined data acquired at two different MRI field strengths. But it is noteworthy that, in a previous study, the DBM pattern observed at baseline in PPMI using 3 T T 1 weighted MRI was replicated at 1.5 T, 2 suggesting that the use of different magnetic field intensities in the current analysis (3 T and 1.5 T) did not impact our findings. In addition, only participants with Parkinson's disease with neuroimaging data at all four points were included in the main analysis, possibly excluding those with more severe symptoms and/or more atrophy (who are more predisposed to drop out of the study) and raising the possibility of survivor and collider bias. 45 The former could mask disease progression, while the latter could lead to biased estimates of associations. Note, however, that additional analyses were performed with larger sample sizes including participants with data at baseline and 1 year, and baseline and 2 years, and results were comparable. Also, since only a limited number of HC had longitudinal measurements, it was not possible to control for the intra-subject variability in normal ageing. Nevertheless, the normal ageing effect was regressed out at every time point using W-Scores. Finally, other structural neuroimaging measures (such as cortical thickness or diffusionweighted MRI) should also be used to investigate the longitudinal progression of atrophy.

Conclusion
Widespread atrophy was found after 4 years in the early stages of Parkinson's disease. This pattern identified vulnerable brain regions that could eventually be used as a guide for other Parkinson's disease longitudinal studies. We also showed factors associated with disease progression, including connectivity and regional tissue composition. This may help identify biological processes implicated in the progression of Parkinson's disease. In addition, our results provide further support for the a-synuclein network spread hypothesis first proposed by Braak.

Supplementary material
Supplementary material is available at Brain Communications online.