Disentangling the Molecular Pathways of Parkinson’s Disease using Multiscale Network Modeling

Abstract

Parkinson’s disease (PD) is a complex neurodegenerative disorder. The identification of genetic variants has shed light on the molecular pathways for inherited PD, while the disease mechanism for idiopathic PD remains elusive, partly due to a lack of robust tools. The complexity of PD arises from the heterogeneity of clinical symptoms, pathologies, environmental insults contributing to the disease, and disease comorbidities. Molecular networks have been increasingly used to identify molecular pathways and drug targets in complex human diseases. Here we review recent advances in molecular network approaches and their application to PD. We discuss how network modeling can predict functions of PD genetic risk factors through network context and assist in the discovery of network-based therapeutics for neurodegenerative diseases.

The complexity of Parkinson’s disease

Parkinson’s disease (PD) is the most common motor neurodegenerative disorder, and is characterized by selective loss of dopaminergic (DA) neurons in the substantia nigra (SN). PD is a complex syndrome, with the manifestation of heterogeneous motor and non-motor impairments. In most patients, PD is pathologically associated with α-synuclein-containing Lewy bodies and Lewy neurites, which can spread throughout brain regions that underlie disease comorbidities such as olfactory dysfunction, cognitive impairment, depression, sleep disorders, and fatigue that further lower the patients’ quality of life [1]. Some clinically diagnosed PD cases do not have Lewy pathology [2, 3]. While treatments are available for alleviating PD’s symptoms in some PD patients, currently, no cure is available for the disease.

Over the past two decades, geneticists have identified more than 20 causal genes from the familial PD cases and over 90 risk loci from large GWAS studies that contribute to approximately 20% of the total PD cases[4]. The PD causal genes, SNCA, LRRK2, VPS35, PINK1, DJ-1, Parkin, FBXO7, DNAJC6, ATP13A2, DCTN1, and SYNJ1, are involved in distinct molecular pathways, including mitochondrial function, autophagy-lysosome pathways, protein and membrane trafficking, and the ubiquitin-proteasome system [5]. Moreover, many genetic risk factors are found to reside in nearby loci of these PD genes. These genetic variants could potentially regulate PD gene expressions to increase the disease penetrance and/or exacerbate the disease phenotypes. Functional interactions between PD gene products were also reported. For example, LRRK2 directly phosphorylates and modulates the function of endophilin [6], synaptojanin1 [7], and auxilin [8], which are all associated with Parkinsonism and required for the synaptic vesicle trafficking pathway. PINK1-mediated phosphorylation of Parkin primes its mitochondrial translocation and regulates mitophagy [9]. Parkin mediates the ubiquitination of VPS35 and modulates retromer-dependent endosomal sorting [10]. The identification of genetic and functional interplay among multiple genetic factors enhances our understanding of the complexity of inherited PD. Whether the majority of PD cases share similar molecular pathways, however, remains unclear. The consensus is that the complexity of PD arises from the interaction of genetic factors, environment, aging, and disease comorbidity.

Furthermore, a growing body of evidence suggests that PD is not simply a disease of the brain. In some patients, disease pathogenesis may occur or possibly even originate outside the brain [11]. For example, in some cases dysfunction of the enteric nervous system (ENS) is believed to precede the lesions in the brain [12]. Recent research has further demonstrated that patients with Crohn’s disease have a higher incidence of PD than the healthy population [13], highlighting the possible link between peripheral chronic inflammation and PD pathogenesis.

Given the high complexity of PD, systems biology approaches are needed to integrate multiple layers of biological data and existing knowledge to derive a holistic view as well as detailed circuits of interacting signaling pathways in PD. Here we review recent systems biology, especially molecular network based studies of PD and elaborate on key components and outcomes of network modeling of PD pathogenesis. We discuss how network modeling can provide a functional context for genetic risk factors of PD. We hypothesize that network modeling can not only predict regulators of PD molecular pathways underlying disease progression, but also pave the way for developing novel therapeutic strategies for PD.

Network biology approaches to dissect complex diseases

Complex diseases such as cancer and neurodegenerative diseases are caused by an array of genetic, epigenetic, and environmental factors. Molecular network analysis offers a systematic view of the complexity of such diseases. A network can be defined as a group of nodes connected by edges. However, the indication of nodes and edges in biological networks vary depending on the context. In the cases of interaction networks such as the human protein interactome [14], protein and DNA/RNA interactions identified by ENCODE [15, 16] and microRNA and RNA interactions assembled in miRBase databases [17, 18], edges represent physical interactions between the molecules. In gene co-expression networks [19, 20], however, edges indicate intrinsic correlations between gene expression profiles. Further, edges can represent drug and target interactions [21], metabolic reactions [22], neural circuit connectivity [23, 24], and information propagation from genetic variants to gene expression and ultimately to disease traits [25].

Biological networks take into account multiple co-occurring biological events from a simple ligand-receptor interaction to complex signaling cascades. From the perspective of input data, such networks can be constructed via low-throughput perturbation experiments or high-throughput molecular profiling of a large number of samples. Although the two types of network constructions differ in coverage and context specificity [26], they often cross-validate and complement each other. Integrating multiple types of networks also leads to more comprehensive signaling maps of cellular physiology. For example, a network-extracted ontology (NeXO) was inferred from integration of large gene and protein interaction networks constructed in yeast with coverage and power equivalent to those of the existing Gene Ontology (GO). Moreover, it identified new biologically meaningful terms that were not previously recognized by domain experts/curators [27].

Over the past decade, network biology approaches have been increasingly used to decode complex diseases such as cancer and neurodegenerative diseases. MAster Regulator INference algorithm (MARINa) was first developed for the identification of molecular determinants for various cancer types [28, 29] and later applied to other pathological conditions including AD [28–31] and PD [32], resulting in the identification of driver genes for pathogenic pathways. These early studies, however, reported methods focused primarily on transcription factors (TF) and were incapable of pinpointing non-TF regulators that also contribute to disease onset and progression. Large-scale GWAS, by contrast, has been fruitful in uncovering genetic risk factors, providing critical information about the genetic contribution to diseases. However, disease relevance and mechanisms of the vast majority of risk loci are poorly understood, and functional mapping of those gene variants remains a daunting task. Hence, novel and effective analytic tools are urgently needed to meet the major challenges in dicephering disease pathogenesis.

We and others have shown how gene regulatory network analysis complements GWAS by providing functional contexts for disease associated genetic variants. Different from the pathway-wide association studies (PWAS) [33] and network-wide association studies (NetWAS) [34], which heavily rely on the existing generic (not disease specific) knowledge, integrative network biology analysis (INBA)[35] infers disease assocociated networks from large-scale molecular profiling data of disease tissues as we showed in PD[37] and AD[38, 39]. Various network analyses have been used to identify molecular mechanisms of AD. For example, Weighted Gene Co-expression Analysis (WGCNA) [40] and multiscale embedded gene coexpression network analysis (MEGENA)[41] were applied to identify coexpressed gene modules; Bayesian network analysis was used to predict gene regulatory relationship [42, 43] [44, 45]; and brain imaging analyses were used to establish the connectome [46–50]. Experimental validation of several novel regulators of the AD-associated networks supports their key roles in cell non-autonomous regulation of disease pathways relevant to AD pathogenesis [52–55] . The success of network approaches in AD has thus paved a way for their application to PD.

Molecular network models of Parkinson’s disease

Network biology approaches were utilized in several idiopathic and LRRK2-related PD studies [56–59]. Additionally, a protein-protein interaction network (PPIN)-based enrichment analysis was applied to identify synaptic transmission and dopamine metabolism as the top disrupted pathways in PD [60]. Integration of transcriptional profiles and network diffusion model demonstrated that the gene expression of immune-related and lysosomal risk factor genes can be used to predict seed region location and the pattern of disease propagation from the substantia nigra[61]. A meta-proteomic analysis of PD identified proteolysis, mitochondria organization, and mitophagy as the most affected pathways in PD [62]. However, data-driven network modeling of PD was not done (to our knowledge) until our recent publication[37].

By leveraging the previous studies of molecular profiling data in PD, our group assembled eight transcriptomic datasets from the human substantia nigra of normal control and idiopathic PD brains and performed Multiscale Embedded Gene co-Expression Network Analysis (MEGENA) [41] of the combined data from 83 PD and 70 control samples. Our analysis identified modular structures of gene co-regulation/co-expression through MEGENA and gene-gene causal relationships through Bayesian causal network (BCN) inference as well as key molecular regulators of PD through key driver analysis [37]. Different from WGCNA, MEGENA constructs planar filtered networks (PFNs) by filtering out redundant correlations using a planarity constraint and then identifies coexpressed gene modules at multiple compactness scales. Comparison of WGCNA and MEGENA showed that MEGENA can identify modules with meaningful biological functions that are inaccessible through WGCNA [37]. In particular, PFNs from MEGENA enables determination of key regulators based on network topology.

One crucial application of the molecular networks approach is to provide functional context for genes with unknown functions. GWAS has identified a large number of PD variants that may affect over 200 genes, the majority of which have uncharacterized functions. Networks underlying PD pathogenesis permit the prediction of the putative function of the known GWAS risk variants. For example, AAK1 was suggested to play a key role in synaptic signaling based on the PFN in PD [37], and indeed this gene was previously reported to regulate clathrin-mediated endocytosis [63]. Such network-based functional annotation can be extended to every single gene in the network. Our Bayesian causal network analysis identified STMN2 as one top key regulator of PD [37]. STMN2 encodes a stathmin family protein and is down-regulated in the substantia nigra of PD brains in comparison with normal controls [37]. STMN2, a hub gene in a synaptic transmission associated module which is most enriched for the molecular changes (i.e., differentially expressed genes) in the substantia nigra between PD and normal control brains, was subsequently validated for its role in PD using mouse models [37]. The gene network-based annotation is complementary to GWAS-based annotations, such as GARFIELD[64], which ties GWAS findings to the regulatory or functional information primarily from ENCODE and Roadmap epigenomic data [65] in order to identify features relevant to a particular disease phenotype. Integration of GARFIELD and transcriptome-wide association (TWAS) shows that neurons and oligodendrocytes are the most affected cell types in PD, where gene expression patterns, RNA splicing, and histone modification were modified in both the central nervous system (CNS) and peripheral monocytes [66]. The network analysis not only elucidates the landscape of PD transcriptomic networks but also pinpoints potential regulators of pathogenic pathways associated with idiopathic PD.

Concluding Remarks and Future Perspectives

In 2013, the National Institute of Aging (NIA) at the National Institutes of Health (NIH) launched the Accelerating Medicines Partnership - Alzheimer’s Disease (AMP-AD) venture to transform basic and translational research of AD through generation and modeling of unprecedented large-scale molecular data (e.g., whole-genome sequencing, RNA-sequencing, DNA methylation, and metabolite profiling) from brains and peripheral tissues. Several AD cohorts [67–69] have been developed to facilitate pathogenic mechanism discoveries and multiscale network analysis. The studies of these cohorts identified increased viral activity in AD brains [70], numerous neuron-specific gene subnetworks dysregulated in AD, and cortical circulated RNAs associated with AD clinicopathological traits [71]. Recently, and along similar lines, the NIH/NINDS launched the Accelerating Medicines Partnership - Parkinson’s Disease (AMP-PD) venture to generate and analyze large-scale sequencing data from PD brains and peripheral tissues. Despite the challenges (see Outstanding Questions), it is imperative to establish appropriate analytic pipelines and experimental platforms to iterate prediction and validation in order to gain a comprehensive understanding of the molecular mechanisms (e.g., signaling networks and key regulators) of PD. As the first multiscale gene network modeling of PD, our recent study sheds light on the landscape of gene co-expression and regulatory networks as well as key molecular drivers. Our study, however, is limited by small sample size, the quality of RNA array data, and the lack of proteomic, epigenetic, or metabolic data. The integration of the large-scale multi-Omics data from the AMP-PD program will significantly empower the identification of reliable biomarkers and better therapeutic targets for PD.

Outstanding questions.

1. Given the complexity and heterogeneity of PD, can multiscale network modeling determine subtypes of idiopathic PD based on correlated clinical diagnoses, pathological data, and molecular pathways?

2. Current network modeling of PD is based on a limited number and variable quality of postmortem PD brains. Can we improve sample quality and size and validate the network models using independent cohorts?

3. Current network modeling of PD informs the functional connectivity of disease pathways through an endpoint snapshot in postmortem brains. How can we integrate longitudinal data from brain imaging, PD-related changes in peripheral tissues and clinical symptoms to increase the prediction power of network modeling?

4. Functional validation of an increasing number of GWAS-genes or risk factors has become challenging. How can we develop a robust, better network platform to validate the interactions between GWAS-genes or risk factors and specific cellular functions at scale?

5. How do we integrate emerging multi-omics data, including single-cell omics, to enhance the power of network analysis and strength of prediction for more robust molecular regulators of PD?

6. Molecular biomarkers that predict PD onset and progression are lacking, hindering early therapeutic intervention. Can multiscale network modeling identify reliable molecular biomarkers to inform the development of subtype PD?

7. Environmental factors contribute significantly to the etiology of PD through gene-environment interactions. How can we integrate datasets of environmental factors into network modeling to improve the prediction power?

8. Large gene expression datasets from drug treatments of human cell lines are currently available. How can multiscale network modeling leverage these resources to identify small chemicals/compounds for therapeutic intervention for normalizing impaired networks in PD?

Emerging single-cell RNA sequencing (scRNA-seq) techinques have enhanced our understanding of complex diseases like PD at the single-cell resolution. For example, studies in humans, mice, and human stem cell models identified subtypes of DA neurons with distinct spatial projections and differential vulnerability to PD [72–76]. scRNA analyses in mice and other species also revealed dynamic microglia programs associated with neurodegenerative diseases such as AD and PD [77, 78]. A recent single-nucleus RNA sequencing analysis has profiled human postmortem SN and proposed an oligodendrocyte-specific association with PD risk [79]. To fully dissect these large-scale, information-rich data, a number of network approaches for analyzing bulk RNA sequencing data have been adapted for scRNA-seq data. By projecting a single-cell gene expression profile onto a global, context-independent reference interactome, SCINET unravels specific disease cell type modularity in neurodegenerative diseases and neuropsychiatric disorders [80]. The Module detection via Topological information and GO knowledge (MTGO) [81], which was originally designed to utilize protein-protein interaction network topology for module identification, was extended as MTGO-SC [82] to perform co-expression network analysis of each cell cluster in scRNA-seq data. Another approach, the single-cell differential network analysis (scdNet), identifies gene pairs whose correlation strength is significantly altered between two different conditions or two single-cell groups [83]. Network analysis of single-cell-based transcriptomic analysis will advance our understanding of disease pathways through cell non-autonomous mechanisms.

High throughput sequencing/profiling technologies are expected to facilitate the development of therapeutics for complex diseases. Indeed, network medicine has emerged to assist in drug discovery [84]. The premise is that restoring molecular states of key disease regulators and networks to normal would potentially reverse the disease phenotypes. Combinations of drugs that target multiple drivers of a disease simultaneously are likely to improve therapeutic effects, as shown in an anti-cancer treatment [85]. Similar efforts have been made in approaching other complex diseases such as PD [86, 87]. Methods have been developed to identify drug candidates to reverse gene expression changes in diseases [88, 89] using high-throughput profiling data based on the cell lines treated with tens of thousands of drugs in the Connectivity Map (CMap) [90] and Library of Integrated Network-based Cellular Signatures (LINCs) [21]. In an initial clinical trial of patients with multiple myeloma, personalized treatment recommendation was provided based on the individual’s genetic and transcriptomic background. Notably, a clinical benefit rate of 76% and an overall response rate of 66% were achieved [91], which was very encouraging. One avenue to further improve the accuracy of drug repurposing is to incorporate disease-specific multiscale network models that carry more information about disease pathogenesis and progression. Network biology approaches that utilize disease-relevant multi-omics data have the potential to bridge the gap between basic research and clinical practice towards the development of precision medicine.

Highlights.

1. Multiscale network modeling offers a holistic, system-level landscape of disease-associated, interacting cellular pathways through an unbiased approach that dissects gene-gene co-expression and co-regulation structures in complex diseases such as Parkinson’s disease (PD).

2. Multiscale molecular network modeling provides informative, functionally relevant context and annotation for genetic risk variants of PD identified from genetic and genome-wide association studies.

3. Multiscale network modeling predicts key molecular regulators for disease progression as potential therapeutic targets.

4. Multiscale network modeling paves the way for developing novel therapeutic strategies for complex diseases by incorporating network structure and connectivity information.