LRRK2 and GBA Variants Exert Distinct Influences on Parkinson’s Disease-Specific Metabolic Networks

Abstract

The natural history of idiopathic Parkinson’s disease (PD) varies considerably across patients. While PD is generally sporadic, there are known genetic influences: the two most common, mutations in the LRRK2 or GBA1 gene, are associated with slower and more aggressive progression, respectively. Here, we applied graph theory to metabolic brain imaging to understand the effects of genotype on the organization of previously established PD-specific networks. We found that closely matched PD patient groups with the LRRK2-G2019S mutation (PD-LRRK2) or GBA1 variants (PD-GBA) expressed the same disease networks as sporadic disease (sPD), but PD-LRRK2 and PD-GBA patients exhibited abnormal increases in network connectivity that were not present in sPD. Using a community detection strategy, we found that the location and modular distribution of these connections differed strikingly across genotypes. In PD-LRRK2, connections were gained within the network core, with the formation of distinct functional pathways linking the cerebellum and putamen. In PD-GBA, by contrast, the majority of functional connections were formed outside the core, involving corticocortical pathways at the network periphery. Strategically localized connections within the core in PD-LRRK2 may maintain PD network activity at lower levels than in PD-GBA, resulting in a less aggressive clinical course.

Introduction

Idiopathic Parkinson’s disease (PD) is heterogeneous in its clinical presentation and highly variable in its disease course (Espay et al. 2017). There has thus been considerable interest in identifying genetic influences, even though the vast majority of PD patients do not have a known hereditable cause for the disorder (Deng et al. 2018). Mutations in LRRK2 (leucine-rich repeat kinase-2) and variants of the GBA1 (glucocerebrosidase) gene both increase the risk of developing the disease and influence its course (Sidransky et al. 2009; Klein and Westenberger 2012). At the group level, PD patients with LRRK2 mutations (PD-LRRK2) exhibit slower motor progression than sporadic PD (sPD), and cognitive impairment is rare in these individuals (Saunders-Pullman et al. 2018). By contrast, PD patients with GBA1 variants (PD-GBA) experience more aggressive disease with substantial cognitive dysfunction (Alcalay et al. 2012; Cilia et al. 2016; Davis et al. 2016). Yet individual genotypic patients cannot be clinically or histopathologically distinguished from their sporadic counterparts (Healy et al. 2008; Neumann et al. 2009). There is a clear need to understand the functional basis for the differences in progression observed in genotypic PD.

PD is associated with attrition of the nigrostriatal dopamine system that leads to a cascade of stereotyped changes in cerebral function in downstream regions (Eidelberg 2009; Politis 2014). We have used metabolic brain imaging with [¹⁸F]-fluorodeoxyglucose (FDG) positron emission tomography (PET), which measures regional glucose metabolism as an index of local synaptic activity (Patel et al. 2014; Stoessl 2017), to identify and validate a disease-specific metabolic network known as the PD-related pattern (PDRP) (Schindlbeck and Eidelberg 2018). Expression levels for this validated pattern are consistently elevated in multiple PD populations (Niethammer and Eidelberg 2012; Schindlbeck and Eidelberg 2018) and correlate with disease progression (cf. Niethammer and Eidelberg 2012; cf. Woo et al. 2017). We have also characterized and validated an analogous PD-related cognitive pattern (PDCP), as a biomarker of cognitive decline in PD (Mattis et al. 2016; Schindlbeck and Eidelberg 2018). These measures are specific and sensitive enough to identify treatment effects in the underlying brain networks (Niethammer and Eidelberg 2012; Ko et al. 2014b; Niethammer et al. 2018). Beyond their utility as biomarkers, they convey information about the structure and function of the underlying networks even before the onset of clinical manifestations (Ko et al. 2018; Niethammer et al. 2018).

Thus far, no studies have explicitly compared functional metabolic networks in subjects with inherited forms of PD and sPD. We were particularly interested in whether genotypic patients exhibit disease networks not seen in sPD or whether they exhibit the same networks but with different patterns of functional connectivity. To answer these questions, we used graph theory to analyze patterns of connectivity within known disease networks.

Materials and Methods

Study Sample

We studied 40 genotyped subjects with akinetic-rigid PD along with 14 age- and gender-matched healthy volunteer subjects. PD subjects were recruited through the Movement Disorders Centers at Northwell Health (Manhasset, NY) and Mount Sinai Beth Israel (New York, NY); healthy control (HC) subjects were recruited from hospital staff and the surrounding community. PD patients were diagnosed according to UK Parkinson Disease Society Brain Bank criteria (Hughes et al. 1992). Ethical permission was obtained from the Institutional Review Boards of the participating institutions; written consent was obtained from each participant after detailed explanation of the procedures. To allow for group comparisons in this cross-sectional design, the three PD groups (PD-LRRK2, PD-GBA, and sPD) were matched for age, gender, symptom duration, and motor disability as gauged by off-state Unified Parkinson’s Disease Rating Scale (UPDRS) motor ratings. The subjects had no evidence of severe cognitive dysfunction based on the Mini-Mental State Examination (MMSE) (total score > 26) or the Mattis Dementia Rating Scale (total score > 140). Clinical and demographic criteria for the study participants are presented in Supplementary Table 1.

Genomic DNA was extracted from peripheral blood samples of all participants using standard methods. The entire GBA coding region was sequenced in all subjects to capture variants including pathogenic mutations as well as the E326K polymorphism that is associated with increased PD risk (Sidransky and Lopez 2012). The LRRK2 region was sequenced for the G2019S and R1441C/G/H mutations using conventional methods (Sanger sequencing and TaqMan genotyping). The PD-GBA group included 12 patients: eight had pathogenic mutations on one allele (1 L444P, 4 N370S, 1 R496H, 1 V394L, 1 D409H); two had the E326K polymorphism on one allele; and two patients had Gaucher’s disease, due either to homozygous or compound heterozygous pathogenic mutations (1 N370S/N370S, 1 N370S/R496H). All GBA mutations are shown in Figure 1. The PD-LRRK2 group comprised 14 PD patients who all carried the LRRK2-G2019S mutation. The remaining 14 PD patients were classified as sPD; they had no family history of disease and lacked GBA or LRRK2 mutations on genotyping. Fourteen age- and gender-matched healthy volunteer subjects without GBA variants or LRRK2 mutations comprised a HC group for comparison with the PD samples.

Figure 1.

Effect of genetic mutations on PD-related network expression. (A) Top: The PDRP displayed as voxel weights thresholded at Z = 3.1 (P < 0.001) and overlaid on T1-weighted MRI template. Bottom: PDRP expression was significantly elevated in the 3 PD groups relative to HC subjects (PD-LRRK2 P < 0.011; PD-GBA P < 0.0001; sPD P < 0.0006) and in PD-GBA patients relative to PD-LRRK2 (P < 0.0001) and sPD (P < 0.006). All PD-LRRK2 subjects had the same G2019S mutations; all GBA1 variants are listed. Two PD-GBA patients with Gaucher’s disease due to homozygous or compound heterozygous mutations expressed the highest levels of PDRP; even after excluding these patients, the PD-GBA group still had the highest PDRP expression relative to all other groups (P < 0.009 for all comparisons). (B) Top: The PDCP displayed as voxel weights thresholded at Z = 3.1 (P < 0.001) and overlaid on T1-weighted MRI template. Bottom: PDCP expression levels were elevated relative to HC subjects in PD-GBA (P < 0.0001) but not in the other PD groups. (^****P < 0.0001, ^***P < 0.001, ^**P < 0.01, ^*P < 0.05, relative to HC subjects; one-way ANOVA, Bonferroni post hoc comparison. Individual data points show the 12 PD-GBA subjects with their genotypes represented by different symbols/colors.)

PET Imaging and Network Analysis

All subjects underwent metabolic imaging with [¹⁸F]-FDG PET on a GE Advance tomograph (General Electric) at Northwell Health as described in detail elsewhere (e.g., Mattis et al. 2016; Niethammer et al. 2018). All subjects fasted overnight prior to PET imaging; anti-parkinsonian medications were withheld for at least 12 hours before the scanning session. Scans from each subject were realigned and spatially normalized to a standard Montreal Neurological Institute (MNI)-based FDG PET template and smoothed with an isotropic Gaussian kernel (10 mm) in all directions to improve the signal-to-noise ratio. Image processing was performed using Statistical Parametric Mapping (SPM5) software (Wellcome Trust Centre for Neuroimaging, Institute of Neurology).

We used spatial covariance mapping to identify and measure the expression of disease-related functional brain networks in resting-state metabolic brain scans from each of the PD groups (Spetsieris and Eidelberg 2011; Spetsieris et al. 2015). This automated routine (software available at http://www.feinsteinneuroscience.org, implemented in MATLAB (Mathworks) version 7.3.0) performs voxel-based principal component analysis (PCA) on the combined data from patients and control subjects to obtain a set of orthogonal PCs, that is, linearly uncorrelated spatial covariance patterns. Each pattern is associated with a set of subject scores (PC scalars), representing network expression in each individual. Subject scores for the leading patterns (the PCs with the highest eigenvalues, i.e., those accounting for the largest portion of the subject × voxel variance in the data) are then entered into a logistic regression model to identify the individual PC, or the linear combination of PCs, that best discriminates patient from control scores on permutation testing (Spetsieris and Eidelberg 2011). In addition, to identify potential other genotype-related networks, we did a voxel-wise whole brain search in the space orthogonal to the PDRP and PDCP, to minimize the contribution of known PD-related networks. We evaluated the correspondence between the resulting topographies and previously identified disease patterns using voxel-wise correlational analysis, incorporating a correction for spatial autocorrelation (Ko et al. 2014c).

We also computed expression values for the previously validated PD-related topographies [PDRP and PDCP (Niethammer and Eidelberg 2012; Schindlbeck and Eidelberg 2018)] in all of the subjects using an automated algorithm blind to genotypic classification (Spetsieris and Eidelberg 2011; Spetsieris et al. 2013). The resulting values were standardized (z-scored) with respect to values from the HC subjects used in the original pattern derivation (Mattis et al. 2016; Niethammer et al. 2018) and compared across groups.

Graph Theory

Defining Nodes and Edges

We used graph theory to assess differences in network parameters within the PDRP and PDCP spaces for the PD-LRRK2, PD-GBA, sPD, and HC groups. We used the metabolic data from the nodes that contributed most to network activity—that is, those with absolute region weights ≥1.0 on the z-scale—to construct a matrix of node-to-node pairwise correlations for each group. Nodes were defined as metabolically active for region weights ≥+1.0 and relatively underactive for region weights ≤−1.0 (Ko et al. 2018). To mitigate potential oversampling bias in network clusters identified in the original voxel-wise derivation, we analyzed connectivity using standardized regions-of-interest (ROIs) defined according to the Anatomic-Automatic Labeling (AAL) atlas (Niethammer et al. 2018). This transformed the PDRP into a set of 38, and the PDCP into a set of 35, nodal ROIs (Supplementary Table 2A,B); 20 ROIs were common to both networks (Supplementary Table 2B).

Functional connections (graphical edges) linking the shared nodes were included in the PDRP space, along with those linking the unshared nodes to other (shared or unshared) PDRP regions. The PDCP space, on the other hand, contained connections linking the shared nodes to unshared PDCP regions, as well as those linking unshared PDCP regions to one another. Thus, although the PDRP and PDCP share nodes, the corresponding network graphs that we analyzed were composed of distinct, nonoverlapping sets of functional connections.

For each node (Supplementary Table 2), we computed globally normalized metabolic activity for the corresponding anatomically defined AAL region. We used bootstrap methods (in-house Matlab script; MATLAB R2017a) to generate 100 samples for each group for subsequent analysis. In each iteration, we estimated pairwise correlation coefficients and used median values from 100 bootstrap correlation estimates to create an adjacency matrix that defined a graph for each group. Correlation coefficients for nodal pairs (graphical edges) were computed using the Statistics and Machine Learning Toolbox in MATLAB R2017a. Graphical displays were produced using the igraph R package (https://igraph.org/r/).

In this scheme, the magnitude of the correlation (r) provided a measure of connectivity between network nodes for each patient group. For a given pair of nodes, group differences in connectivity (a graphical edge) were denoted by a distance measure (dr) that was defined as the absolute difference in the two connectivity measures (cf. Niethammer et al. 2018). For the gain or loss of a connection in one of the patient groups to be considered significant relative to HC subjects, we required that the magnitude of the correlation coefficient (|r|) that defined the edge in the patient group be greater than or equal to 0.6 (P < 0.05; Pearson correlation) and that the corresponding change (dr) from the HC group be greater than 0.4 (P < 0.05; permutation test, 1000 iterations). The threshold for a particular edge to show a significant change in connectivity across groups was determined using the HC graph and permuting regional labels 1000 times, thereby creating a set of pseudo-random correlations with each iteration. Under these conditions, the distance dr between the actual HC correlation and the simulated values for each edge is itself a random variable. Examining the distribution of dr values for |r| ≥ 0.6, we found that for dr ≥ 0.4, the proportion of sample permutations exceeded the 95% significance level and therefore could not be explained by chance. For graphical edges, the pairwise connections that satisfied these criteria were confirmed by bootstrap resampling (100 iterations).

Network Parameters

To assess group differences in network-level information processing for the PDRP and PDCP spaces and for subspaces defined by the metabolic valence (region weight sign) for each network (Ko et al. 2018), we computed:

• (1) “Degree centrality,” the total number of connections (edges) within the network/community space, which is a measure of overall functional connectivity within a given space.

• (2) Random-graph normalized “clustering coefficients,” which reflect the likelihood that the nearest neighbors of a network node will also be directly connected.

• (3) Random-graph normalized “characteristic path length,” an indicator of the global efficiency of the network (Newman 2010; Rubinov and Sporns 2010).

• (4) “Small worldness,” the ratio of clustering coefficient to path length, normalized to corresponding parameters from an equivalent random graph (Bassett and Bullmore 2016).

For each group and network space, we also computed:

• (5) “Modularity” (the underlying community structure), defined by the number of connections (edges) falling within communities minus the number that would be expected for an equivalent network with edges placed at random (Newman 2006); and

• (6) “Synchronization coefficient,” a measure of functional connectivity between modules, defined here as the ratio of the first non-zero eigenvalue (FNZE) to the maximum eigenvalue of the corresponding graph Laplacian matrix for each group and network subspace (Chen et al. 2012).

These calculations were performed using the Brain Connectivity Toolbox (Rubinov and Sporns 2010). The resulting data were presented at varying graph thresholds ranging from r = 0.30 to 0.60, at 0.05 increments. At these thresholds, the graph densities range from 25% to 60% on average in the PDRP/PDCP space. Permutation tests (5000 iterations) were used to determine the significance of differences between groups (PD-LRRK2, PD-GBA, sPD, and HC) for each measure at the various correlation thresholds. Differences were considered significant for P < 0.05, incorporating a Bonferroni correction for multiple network parameters (n = 6) and cost levels (n = 7). Group effects were reported only when significant differences were evident at four (graph density range 25–50%) or more (graph density range 25–60%) consecutive correlation thresholds.

Genotypic Differences in Network Connectivity

For each group (sPD, PD-LRRK2, and PD-GBA), we computed the gain-to-loss ratio (GLR) for connections within the PDRP and PDCP spaces that were gained or lost relative to HCs. In each network, connections were further classified according to edge type: active–active, underactive–underactive, or active–underactive (see above). The gain of connections in each edge category was computed as a percentage of the total; bootstrap resampling (100 iterations) was used to estimate the mean and standard error of this measure for each genotype and edge category. Genotypic differences were evaluated using ANOVA with the post hoc Bonferroni correction for multiple comparisons.

In previous work, we noted a close correspondence between the metabolically active (underactive) PDRP subgraph and the core (periphery) of the network (Correa et al. 2012; Ko et al. 2018). We considered the possibility that PD-LRRK2 and PD-GBA differed as to whether the gain in functional connectivity was more pronounced at the network’s core or in its periphery. To this end, we computed the difference between the connections gained inside (active–active) versus outside (active–underactive and underactive–underactive) the PDRP core for each genotype (100 bootstrap iterations; Student’s t-test). A similar analysis was conducted for the PDCP.

Community Detection

To examine network structure in a data-driven manner that is independent of metabolic valence, we employed two community detection algorithms. We used the network graphs from the PD-LRRK2 group as the basis for community detection, since this group is genetically homogeneous (each subject bears the same LRRK2-G2019S mutation).

The first method was Newman’s Modularity, a well-established modularity maximization algorithm (Newman 2006). Partitioning was done using the “modularity_und” function in the Brain Connectivity Toolbox (Rubinov and Sporns 2010), which maximizes the number of within-group edges while minimizing the number of between-group edges. Because of the well-recognized resolution limit on community size, Newman’s Modularity tends to partition graphs into clusters with similar numbers of nodes, which may not reflect true network structure (Fortunato and Barthelemy 2007). To circumvent this problem, we also partitioned the networks using Asymptotic Surprise (Nicolini et al. 2017), a different algorithm based on information theory and partitioning cost optimization that is not bound by a resolution limit (Nicolini et al. 2017). Partitioning by this method was conducted using the PACO Matlab mex function (https://brainetlab.github.io/sections/software/software.html).

For each network, community structure was based on node-to-node correlation coefficients (|r| ≥ 0.5), without constraining the number of nodes (community size) in each module. Neither the sign nor the magnitude of PDRP or PDCP region weights influences the algorithm’s assignment of a node to a given community. In both procedures, partitioning revealed a tightly connected indivisible community, termed Module A, which could be considered as a core zone (Da Silva et al. 2008). The remaining nodes formed a second community, termed Module B, which was further divisible and included elements of the periphery.

For both partitioning algorithms, we assessed the stability of the observed community structure by measuring its similarity to that realized in each of 100 bootstrap iterations of the data. This was done by computing the Rand Index (Rand 1971; Severiano et al. 2011) for each iteration. Similarity of communities across iterations was determined by the mean and 95% confidence interval (CI) of the Rand Index for partitions conducted using Newman’s Modularity as well as Asymptotic Surprise.

Statistical Analysis

Network expression values were compared across groups using one-way analysis of variance (ANOVA) with post hoc tests corrected for multiple comparisons according to Bonferroni. Relationships between individual subject pattern expression values and the UPDRS scores were evaluated separately in each group by calculating the Pearson product–moment correlations. All statistical analyses were performed with either R (version 3.3.2, 2016-10-31) or GraphPad Software (Version 7.0, La Jolla California, USA) and considered significant at P < 0.05 (two-tailed).

Data Availability

Deidentified data will be made available on reasonable request from interested investigators for the purpose of replicating results.

Results

PD Networks are Expressed in the sPD, PD-LRRK2, and PD-GBA

The three PD groups were closely matched for subject age, symptom duration, and off-medication motor disability ratings (Supplementary Table 1). We computed expression values for the previously validated PD-related topographies, PDRP and PDCP (Schindlbeck and Eidelberg 2018). In addition, we used spatial covariance mapping to determine whether other metabolic patterns were present in the various PD groups (see Methods).

Each group expressed elevated levels of PDRP compared to HCs (Fig. 1A). PDRP expression correlated with motor ratings in individual subjects, in the combined PD sample (r = 0.60, P < 0.0001) and in each patient group (sPD: r = 0.71, P < 0.005; PD-LRRK2: r = 0.58, P < 0.05; PD-GBA: r = 0.69, P < 0.02). Despite matched motor disability ratings, the PD-GBA group showed higher PDRP subject scores than the other patient groups (PD-LRRK2: P < 0.0001; sPD P < 0.0008). PD-GBA was also the only group to show elevated PDCP expression relative to HCs (Fig. 1B). These results suggest greater underlying disease activity in PD-GBA. No network topographies apart from PDRP and PDCP were identified in the PD-LRRK2 or PD-GBA groups (Supplementary Fig. 1).

Network Parameters in sPD, PD-LRRK2, and PD-GBA

To explore whether specific network properties differed across groups, we used graph theory to compute the following network parameters: “degree centrality,” a measure of overall functional connectivity within the network space; “clustering coefficient,” a measure of a node’s local connection density; “characteristic path length,” an indicator of the efficiency of information transfer; and “small worldness,” as a measure of efficient communication that combines short path length and high clustering (Bassett and Bullmore 2016) (see Methods for more details).

As we had previously shown, the PDRP space in sPD exhibits a greater than normal number of connections linking metabolically active nearest-neighbor nodes at the network core, particularly in the putamen, globus pallidus, and thalamus, thereby exaggerating the network’s small worldness (Ko et al. 2018; Niethammer et al. 2018). To our surprise, both PD-GBA and PD-LRRK2 overcorrect these network parameter abnormalities (Supplementary Fig. 2, first column). Both show lower clustering and greater characteristic path length in PDRP than either sPD or HC; their small worldness was even lower than that of healthy subjects (P < 0.001; permutation test, corrected for multiple comparisons at four or more consecutive thresholds). Equally unexpectedly, these measures did not differ between the LRRK2 and GBA groups (P > 0.05, corrected).

We therefore asked whether LRRK2 and GBA variants alter network parameters in pre-defined subnetworks. We had previously conceptualized PDRP as containing a “core” of 20 strongly linked, metabolically active nodes (defined by region weights ≥1.0) and a “periphery” of 18 less strongly connected, metabolically less-active nodes (region weights ≤−1.0) (Supplementary Table 2A) (Ko et al. 2018). When we computed graphical parameters for just the core subnetwork, PD-LRRK2 diverged from the other groups in every measure (Supplementary Fig. 2, middle column). In the PDRP periphery, however, PD-GBA essentially switched places with PD-LRRK2 (Supplementary Fig. 2, right column). For example, PD-LRRK2 had the greatest degree centrality in the active subnetwork (core), but PD-GBA had the greatest degree centrality in the underactive subnetwork (periphery). PD-LRRK2 showed reduced clustering, greater characteristic path length, and reduced small worldness in the core (P < 0.001; permutation test corrected for multiple comparisons); PD-GBA took on the same characteristics, but in the periphery.

We were curious to know whether the PDCP space underwent similar changes in network characteristics. We divided the PDCP space into nonoverlapping subnetworks based on region weight sign. One subnetwork comprised 19 metabolically active nodes with region weights ≥1.0, while the other comprised 16 relatively underactive nodes defined by region weights ≤−1.0 (Supplementary Table 2B). Considering PDCP as a whole, PD-LRRK2 values overcorrected all parameters relative to sPD (Supplementary Fig. 3, left column). PD-LRRK2 also differed the most from the other groups in both the active and the underactive subnetworks (Supplementary Fig. 3, middle and right columns).

To better understand the communication between modules within the PDRP and PDCP spaces, we computed “modularity,” the propensity for nodes to form densely interconnected clusters (modules), and “synchronization,” a measure of modular communication within a given network. Relative to the other groups, PD-LRRK2 showed less modularity and greater synchronization within the active PDRP subspace (core) (P < 0.001, corrected; Fig. 2A,B). PD-GBA showed analogous differences, but in the relatively underactive PDRP subspace (periphery). In the PDCP space, PD-LRRK2 again showed lower modularity, in both the active and underactive subnetworks (P < 0.001, corrected) (Fig. 2C). PD-LRRK2 showed greater synchronization than the other groups (P < 0.001, corrected) in the metabolically underactive PDCP subnetwork but not in the active subnetwork (Fig. 2D).

Figure 2.

Distinct genotypic effects on PDRP and PDCP subnetworks. PDRP (top): In the active subnetwork (left), PD-LRRK2 (orange) showed (A) lower modularity and (B) and increased synchronization relative to PD-GBA (green), sPD (black), and HC subjects (gray). In the underactive subnetwork (right), PD-GBA showed the lowest modularity and greatest synchronization. PDCP (bottom): Analogous measurements in the metabolically active (left) and underactive (right) PDCP subnetworks revealed lower modularity (C) in the PD-LRRK2 group in both zones. Increased synchronization (D) in this group was seen only in the underactive subnetwork. (Threshold levels 1–7 correspond to cutoff thresholds r = 0.3–0.6 with increments of 0.05 [see Methods]. ^***P < 0.001 relative to HC, sPD, and PD-GBA; ^*P < 0.001 relative to sPD.)

Patterns of Connectivity Gains and Losses in the PDRP Space

We next identified the functional connections gained and lost in the PDRP space for each PD group relative to the healthy subjects (Supplementary Table 3A–C). Both the PD-LRRK2 and PD-GBA groups showed greater GLR (see Methods) than sPD group (P < 0.0001; post hoc Bonferroni tests), with PD-LRRK2 showing the largest GLR value (P < 0.0001; post-hoc Bonferroni test) in this space (Fig. 3A, left). PD-LRRK2 and PD-GBA showed a strikingly different distribution of gained connections for the different PDRP subnetworks (Fig. 3A, middle). We observed a significant genotype × edge category interaction effect on connectivity gain in the PDRP space (F(2, 594) = 244.4; P < 0.0001; two-way ANOVA). PD-LRRK2 gained significantly more connections in the metabolically active subnetwork compared to PD-GBA (P < 0.0001; post hoc Bonferroni test); nearly half (44%) of the new connections linked metabolically active nodes or linked active with underactive nodes (42%), very few connections linked underactive PDRP nodes in this genotype (14%). By contrast, PD-GBA gained more connections in the underactive subnetwork (P < 0.0001), as well as connections linking active and underactive nodes (P < 0.0001). The majority of gained connections in PD-GBA linked underactive nodes with one another (26%) or linked underactive and active nodes (51%). Lastly, the two genotypes differed in the overall number of connections gained in the core versus the periphery (Fig. 3A, right): PD-GBA gained significantly more connections outside relative to inside the core (P < 0.0001).

Figure 3.

Patterns of connectional gain in PD subtypes. (A) Left: Ratio of GLR of functional connections in the PDRP space. For this network, GLR values were increased in both PD-LRRK2 (orange) and PD-GBA (green) relative to sPD (black) and in PD-LRRK2 relative to both PD-GBA and sPD. Middle: Effects of genotype on the gain of connections (% total) in the different PDRP edge categories. In PD-LRRK2, the majority of the gained PDRP connections linked metabolically active nodes to other active nodes or to underactive nodes. In PD-GBA, gained connections linked mostly underactive nodes to each other or to active nodes. Right: Effects of genotype on the difference between the connections gained (% total) outside vs. inside the PDRP core. This measure was significantly greater in PD-GBA compared to PD-LRRK2. (B) Genotypic effects on: left: GLR values, middle: gain of connections in the different edge categories, and right: gain of connections outside vs. inside the core, in the PDCP space. Genotypic effects on PDCP connective patterns resembled those observed on the PDRP. (Error bars represent mean ± SE, 100 bootstrap iterations, see Methods. ^**P < 0.01, ^***P < 0.001, and ^****P < 0.0001.)

Whereas the patterns of edge gain differed markedly across the three groups, the pattern of edge loss was quite similar. Irrespective of subtype, the majority of lost connections in PD patients were between metabolically active and underactive nodes. At the individual edge level, all three PD groups lost connections between (1) the thalamus and the inferior occipital cortex, (2) the superior parietal cortex and the putamen, and (3) the hippocampus with the midfrontal, posterior parietal, and occipital association regions (Supplementary Table 3A–C).

Connectional Gains and Losses in the PDCP Network

Functional connections gained and lost in the PDCP space were measured for each PD group relative to HCs (Supplementary Table 4A–C). Analogous to the PDRP space, both the PD-LRRK2 and PD-GBA groups showed a greater GLR than the sPD group (Fig. 3B, left). The two genotypes showed a markedly different distribution of gained connections in the PDCP space (Fig. 3B, middle). We observed a significant genotype × edge category interaction effect on connectivity gain in the PDCP space (F(2, 594) = 60.7; P < 0.0001; two-way ANOVA). PD-LRRK2 gained significantly more connections in the underactive subnetwork than PD-GBA (P < 0.0001; post hoc Bonferroni test), linking underactive nodes (31%) or active and underactive nodes (44%). By contrast, PD-GBA gained more connections in the active subnetwork (P < 0.0001) and between these two subnetworks (P < 0.002) relative to PD-LRRK2, linking active nodes with each other (35%) or linking active and underactive nodes (46%). As with the PDRP, the gain in PDCP connections outside relative to inside the core was substantially greater in PD-GBA than in PD-LRRK2 (P < 0.0001; Fig. 3B, right).

PDRP Community Structure and Connectional Gain in PD Subtypes

The division of the PDRP and PDCP graphs into subnetworks based upon relative metabolic valence, that is, the sign of region weights on the PDRP or PDCP, is based on graph-theoretical inferences from sPD (Ko et al. 2018), which may not apply to genotypic forms of the disorder. Therefore, we sought to partition the networks based on community structure rather than region weight sign. According to this approach, networks are decomposed into communities or modules, internally dense clusters of neural elements that are more sparely linked to one another (Newman 2006). To this end, we used two data-driven community detection algorithms: Newman’s Modularity (Newman 2006), which has a well-known resolution limit, and Asymptotic Surprise (Nicolini et al. 2017), an information-theoretical alternative that is free of resolution limits (see Methods).

For the PDRP, Newman’s method yielded two distinct communities. The first, Module A, consisted of 12 interconnected metabolically active nodes (red circles; Fig. 4). The second, Module B, consisted of the remaining 8 active core nodes (red circles), as well as the 18 relatively underactive nodes (blue circles). The metabolically active nodes in Module B clustered together and had lower component values on the partitioning vector (represented by a smaller radius in Fig. 4). These nodes reflect comparatively small contributions to network modularity indicating that they may move between communities with little penalty. By contrast, the underactive nodes along the outer rim of Module B had comparatively large component values signifying relatively greater contributions to network modularity; such nodes are firmly anchored to their module. Repeat partitioning of PDRP graphs from the PD-LRRK2 group (100 bootstrap iterations) revealed this parcellation to be stable (0.73 ± 0.02 [mean and 95% of CI]; Rand index). Analogous partitions of similar stability were seen with Asymptotic Surprise (0.71 ± 0.03 [mean and 95% of CI]; Rand index).

Figure 4.

PDRP community structure and connectional gain in PD subtypes. Two major nodal communities were detected within the PDRP space: Module A was composed of 10 interconnected metabolically active nodes (red circles), while Module B contained the remaining 10 metabolically active nodes that comprised the network core and the 18 relatively underactive nodes (blue circles) that comprised the network periphery. Significant functional connections (gray lines) linking nodes are mapped for sPD, PD-LRRK2, and PD-GBA (A–C). Bold lines denote connections gained with respect to HCs (see Methods). In PD-LRRK2 (B), approximately half of the connections gained in the PDRP space linked metabolically active nodes within or between modules. In PD-GBA (C), the PDRP connections gained linked underactive Module B nodes or linked active Module A nodes with underactive Module B nodes. [The radius of each circle (node) is proportional to the corresponding regional component of the leading eigenvector of the modularity matrix (Newman 2006).]

We then mapped the significant connections gained within and between these communities in each group. In sPD, few connections were gained relative to HC subjects, mostly linking Module A with Module B. Many more connections were gained in PD-LRRK2, such as those linking the cerebellum to the putamen via the thalamus in Module A, as well as connecting active nodes in Modules A and B through the amygdala, insula, and parahippocampal gyrus (Fig. 4B and Supplementary Table 3B). The majority of connections gained in PD-GBA were within Module B or linked active and underactive nodes in the two modules. The GBA group showed significant new links between frontal, parietal, and occipital association regions in Module B or between limbic and frontal-basal ganglia pathways spanning Modules A and B (Fig. 4C and Supplementary Table 3C).

Relationship to PDCP Community Structure

The PDCP space also contained two distinct communities according to the abovementioned partitioning algorithms. The first, Module A, contained 16 nodes, of which 15 were metabolically underactive; while the majority of these (10 nodes; squares) corresponded to specific PDCP regions, there were 6 nodes (circles) that were common to both the PDRP and PDCP networks (Fig. 5). The underactive PDCP nodes in Module A (blue squares) had larger component values (corresponding to the size of the symbol) on the partitioning vector, which were situated along the outskirts of this community. Module B contained 19 nodes, of which 18 are metabolically active. In contrast to Module A, the majority of Module B (14 nodes; circles) were PDRP regions that linked to PDCP nodes. As with PDRP, repeat parcellation of the PDCP graph (100 bootstrap resamples) revealed this community structure to be reproducible in the PD-LRRK2 group using the two partitioning algorithms (Newman’s Modularity: 0.69 ± 0.02 [mean and 95% CI]; Asymptotic Surprise: 0.71 ± 0.02 [mean and 95% CI]; Rand index).

Figure 5.

PDCP community structure and connectional gain in PD subtypes. Two major nodal communities were detected within the PDCP space: Module A was composed of 16 interconnected nodes (15 metabolically underactive [blue] and 1 active [red]) and Module B contained 19 nodes (18 active [red] and 1 underactive [blue]). Significant functional connections (gray lines) linking nodes are mapped for sPD, PD-LRRK2, and PD-GBA (A–C). Bold lines denote connections gained with respect to HCs (see Methods). (A) Few gains of PDCP connections were seen in the sPD group. (B) PD-LRRK2 exhibited a greater percentage of gained connections linking nodes of the same valence in Module A than PD-GBA. (C) PD-GBA exhibited connectional gains linking nodes of the same metabolic valence in Module B or linking nodes of opposite valence in the two modules. [The PDCP space contains connections linking PDCP nodes to one another (squares), as well as linking PDCP and PDRP nodes (circles); see Methods. The size of each symbol (node) is proportional to the corresponding regional component of the leading eigenvector of the modularity matrix (Newman 2006).]

Very few new PDCP connections were seen in sPD (Fig. 5A and Supplementary Table 4A). In PD-LRRK2, the majority of gained PDCP connections linked underactive nodes in Module A, such as those between prefrontal nodes or between underactive prefrontal and parietal nodes (Fig. 5B and Supplementary Table 4B). By contrast, PD-GBA patients exhibited a prominent gain of connections between active PDCP nodes in Module B, such as those linking nodes in the cerebellar vermis with the amygdala and medial temporal cortex (Fig. 5C and Supplementary Table 4C). In both genotypes, a greater number of connections were gained between PDCP and PDRP nodes (PD-LRRK2: 30.9 ± 1.4; PD-GBA: 23.2 ± 0.8; mean ± SD) than between PDCP nodes (PD-LRRK2: 7.5 ± 0.5; PD-GBA: 6.1 ± 0.4; P < 0.0004; two-way ANOVA; 100 bootstrap iterations). Both PD-LRRK2 and PD-GBA exhibited minor loss of normal connections within the PDCP space (Supplementary Table 4B,C).

Discussion

While the two genetic forms of PD express the same disease-specific networks as sPD, information flow through these networks differs profoundly across patient groups. PD-LRRK2 showed increased functional connectivity within the metabolically active PDRP core zone. This was confirmed using hypothesis-free community detection methods, which revealed that a large portion of the gained connections were in Module A, a nondivisible community, which included the majority of the metabolically active nodes that composed the network core (Ko et al. 2018). In PD-LRRK2, preferential gain in connectivity within the PDRP core is associated with lower disease network expression, indicating less severe underlying functional pathology in PD patients carrying this mutation. Maintenance of PDRP expression at such levels is consistent with the more benign clinical course observed in this genotype. By contrast, in PD-GBA the gains in connectivity extend outside the core, along with increased expression of the whole network. These findings are consistent with the more aggressive natural history seen in PD-GBA patients.

In an earlier longitudinal study of sPD, we tracked patients with early disease of two years’ duration and followed them for four years (Huang et al. 2007; Tang et al. 2010). We found that changes in the metabolically active PDRP core, which involves regions connected to the substantia nigra, basal forebrain, and paralimbic structures (Braak stage 3–4), preceded those in the metabolically underactive regions of the network periphery (Ma et al. 2009). This suggests that the disease’s functional “weather front” moves from the core to the periphery over time. The PD-LRRK2 group had more connections within the core, and PD-GBA within the periphery, even though both groups had symptoms for roughly 7 years. This suggests that the PDRP “weather front” has progressed less in PD-LRRK2, and further in PD-GBA, over the same disease duration (Davis et al. 2016; Saunders-Pullman et al. 2018).

Although comprehensive neuropsychological testing was not available for all our patients, brief assessments revealed no evidence of severe cognitive dysfunction in any of these individuals. Nonetheless, changes in network parameters analogous to those seen in the PDRP were documented in the PDCP, particularly in its metabolically underactive subnetwork. Connections between metabolically underactive PDCP nodes may have “core-like” features analogous to those seen in the PDRP (Ko et al. 2014a; Ko et al. 2018). As with the PDRP, the PDCP in PD-LRRK2 is distinguished by comparatively dense functional connections between nodes of the same valence within Module A, which behaves as the network core. These connections, which link underactive PDCP nodes with one another and with nearby nodes of the PDRP periphery (Fig. 5B), may be more efficient by virtue of their relatively short geographical distances (Khambhati et al. 2018). By contrast, in PD-GBA, the PDCP space displays connections that link nodes of opposite valence in different modules (and networks), which are at greater distance from each other. The comparative length of these pathways may limit the ability to stabilize functional connectivity within the PDCP core. It is important to keep in mind that these findings rely on small cohorts that need to be validated in independent patient populations.

Implications for Disease Modification

Without longitudinal data from young carriers of LRRK2 mutations, we cannot distinguish between those connections in the PDRP and PDCP space that are compensatory, those that are pathogenic, and those that developed very early in life and contribute either to later vulnerability to PD or resilience in the face of the disease. Evidence from both humans and animal models suggests that LRRK2 mutants are antagonistically pleiotropic: increased LRRK2 activity early in life protects against opportunistic infection and increases synaptic connectivity and efficiency (Sweet et al. 2015; Matikainen-Ankney et al. 2016), but later in life it increases the risk of developing PD, Crohn’s disease (Hui et al. 2018), and possibly autoimmune disease (Alessi and Sammler 2018). One recent study found that LRRK2-G2019S carriers, although more likely to develop PD at an earlier age than noncarriers, also had higher educational level and global cognitive function, and lower motor disability ratings compared to noncarriers (Ben Romdhan et al. 2018). This suggests that early changes may indeed benefit G2019S mutation carriers by inducing more robust functional connectivity to begin with.

Gains in corticostriatal and nigrocortical functional connectivity have already been described in clinically non-manifesting carriers of the LRRK2-G2019S mutation and in G2019S mice (Helmich et al. 2015; Vilas et al. 2016; Benson et al. 2018). Moreover, young clinically unaffected carriers of LRRK2 mutations show increased activation responses in some of the same regions in which our PD-LRRK2 subjects showed increased resting-state functional connectivity (Thaler et al. 2013). The pronounced connectional gains we observed may involve abnormalities in multiple neurotransmitter systems: some of the gains in PDRP coincide with increased serotonin transporter binding in LRRK2 mutation carriers (Wile et al. 2017), while some we observed in the PDCP space could relate to increased cholinergic function (Liu et al. 2018). It is unknown whether similar neurotransmitter changes are present in carriers of GBA1 variants or whether GBA1 mutations affect neurodevelopment.

Could the induction of circuit changes similar to those seen in PD-LRRK2 modify the disease course in sPD by stabilizing the activity of pathological PD networks? In PD-LRRK2, cerebello–thalamo–putamen pathways form discrete, closed subcircuits within the network core. This pathway represents a major interface between cerebellar and basal ganglia circuitry involved in programming and executing movement (Bostan and Strick 2018). The centromedian and/or ventrolateral thalamic nuclei may be worthwhile targets of therapeutic network modulation in patients with early stage sPD or PD-GBA. Analogous interventions to stabilize PDCP activity by inducing functional connections between frontal and parietal nodes could also prove beneficial. Before any circuit-level interventions can be considered, however, it will be necessary to distinguish between compensatory and pathological changes in connectivity within network spaces. This will require longitudinal studies conducted in manifesting and nonmanifesting carriers of these mutations.

Even so, disease-specific network patterns can already provide valuable information as to whether a given therapy alters the underlying disease process, either in the context of a clinical trial (Niethammer et al. 2018) or in devising personalized therapeutic strategies for individual patients. Clinical trials of LRRK2 and glucosylceramide inhibitors are already underway, and it would be useful to determine whether these agents modify the underlying disease process by altering community structure. Our approach may also be beneficial in discerning genotypic network alterations in other neurodegenerative diseases. Even with limited knowledge of genetic risk factors, graph theoretic approaches may disclose distinct functional connectivity patterns of prognostic and therapeutic relevance.