A Combined Analysis of Genetically Correlated Traits Identifies Genes and Brain Regions for Insomnia

Abstract

Aims:

Previous studies have inferred that there is a strong genetic component in insomnia. However, the etiology of insomnia is still unclear. This study systematically analyzed multiple genome-wide association study (GWAS) data sets with core human pathways and functional networks to detect potential gene pathways and networks associated with insomnia.

Methods:

We used a novel method, multitrait analysis of genome-wide association studies (MTAG), to combine 3 large GWASs of insomnia symptoms/complaints and sleep duration. The i-Gsea4GwasV2 and Reactome FI programs were used to analyze data from the result of MTAG analysis and the nominally significant pathways, respectively.

Results:

Through analyzing data sets using the MTAG program, our sample size increased from 113,006 subjects to 163,188 subjects. A total of 17 of 1,816 Reactome pathways were identified and showed to be associated with insomnia. We further revealed 11 interconnected functional and topologically interacting clusters (Clusters 0 to 10) that were associated with insomnia. Based on the brain transcriptome data, it was found that the genes in Cluster 4 were enriched for the transcriptional coexpression profile in the prenatal dorsolateral prefrontal cortex (P = 7 × 10⁻⁵), inferolateral temporal cortex (P = 0.02), medial prefrontal cortex (P < 1 × 10⁻⁵), and amygdala (P < 1 × 10⁻⁵), and detected RPA2, ORC6, PIAS3, and PRIM2 as core nodes in these 4 brain regions.

Conclusions:

The findings provided new genes, pathways, and brain regions to understand the pathology of insomnia.

Introduction

The prevalence of insomnia has been defined as around 6% to 10% of the general population.¹ Several studies have identified that insomnia was associated with a multitude of negative mental disorders, including posttraumatic stress disorder² and suicide,³ as well as adverse long-term health outcomes, including diabetes⁴ and cardiovascular disease,⁵ providing the need for research into the etiology of the disorder.

Twin studies have revealed that the sleep characteristics and insomnia had a highly heritable range from 22% to 59% in adults and 14% to 71% in children.^6,7 Both Galbiati and colleagues (2018)⁸ and Fan and colleagues (2018)⁹ found that primary insomnia was associated with decreased objective sleep duration. Previous genome-wide association studies (GWASs) identified massive susceptible genetic variants, particularly single-nucleotide polymorphisms (SNPs), in developing insomnia with different phenotypes, such as insomnia symptoms¹⁰ and insomnia complaints.¹¹ These studies also showed the genetic correlations among insomnia symptoms, insomnia complaints, and sleep duration. Current research developed a new technique, MTAG,¹² that performs the meta-analysis of summary statistics from genetically related traits and can increase the sample size to GWASs of insomnia based on the genetic variance that is shared with other phenotypes.

In addition, a prior study found that genetic variants, which passed with nominal statistical significance in GWAS, could offer useful information on the pathological mechanism of the disease, particularly the variations aggregating within a functional cluster or pathway.¹³ Furthermore, a number of methodologies have been developed to identify biological pathways with a risk of disease. For instance, the gene set enrichment analysis (GSEA) was initially designed to analyze genome-wide expression data. This method could currently be used in a common pathway-based approach. Wang and colleagues (2007)¹⁴ developed a modified version of the GSEA to handle genome-wide SNP associations, and this algorithm could detect the combined multiple genetic markers/gene effects of a disease. Therefore, the algorithm may be applied to explore the biological mechanisms of disease at the system level. It is advantageous to extending the pathway-based analysis into molecular networks or coexpression networks to reveal the etiology of insomnia. A network-based analysis could identify the molecular network that interacts with biomolecules, such as genes, proteins, and metabolites (e.g., in functional interactions (FIs), protein–protein interactions (PPIs), and gene regulation). Observably, prior researchers have found that genes associated with the same or similar disorders are inclined to cluster in the same gene network and cling to physical or functional modules.^15,16 Hence, based on newly developed methods and programs, investigating GWAS data could help identify genes and gene pathways as the risk of insomnia development.

In the present study, we systematically analyzed multiple results of GWASs to increase the power to detect potential gene pathways and networks associated with insomnia. The analyses we carried out included (1) MTAG to combine the meta-analysis of summary statistics from insomnia symptoms/complaints and sleep duration; (2) the i-GSEA4GWAS-v2 program to integrate MTAG results for insomnia and the Reactome pathway data set, and the genes in the significant pathways to build the functional interacting network and clusters sets to provide further insights into potential mechanism of insomnia; (3) the brain transcriptome data set from BrainSpan to analyze and obtain gene coexpression networks based on these clusters.

Materials and Methods

Retrieval of GWASs Data

In this study, we obtained the GWASs data from insomnia symptoms (32,155 cases and 26,973 controls),¹⁰ insomnia complaints (113,006 subjects),¹¹ and sleep duration (111,975 participants).¹⁰ For sleep duration and insomnia symptoms, we performed GWAS analyses using logistic/linear regression with age, sex, 10 principal components, and genotyping array as covariates. For insomnia complaints, the association tests were performed using logistic regression adjusted for sex, age, top 5 principal components, and genotyping array. The ethics approval of insomnia studies can be found in the original articles.^10,11

MTAG Analysis

In order to further increase genetic signals, we applied a newly developed method, MTAG,¹² which combines linkage disequilibrium (LD) score regression and meta-analysis methodology from genetically related traits. The software was used to combine summary statistics from the GWASs of sleep duration,¹⁰ insomnia symptoms,¹⁰ and insomnia complaints.¹¹ MTAG software can increase the genetic signals via related traits and has been shown to be effective in handling uncharted sample overlaps. Thus, this method assessed the gain in average power from MTAG relative to the GWAS results by the increase in the mean χ²-statistic and uses this increase to calculate how much larger the GWAS sample size would have to be to obtain an equivalent increase in expected χ². The MTAG program aligned all alleles based on different summary statistics and ensured that SNPs were present in all data sets. SNPs that were not present in any data set were removed. The final count of SNP for MTAG analyses was 8,270,035.

Gene-Based GWAS

In the study, we used the Multi-marker Analysis of GenoMic Annotation software¹⁷ to conduct the gene-based analysis based on the P value of SNPs into a gene test. Gene boundaries and locations were used from the National Center for Biotechnology Information build 37, and the LD among the SNPs in the gene was then assessed according to the reference data (1,000 genomes phase 3 release).¹⁸ Multiple testing was performed using Bonferroni correction for 18,199 autosomal genes (0.05/18,199 = 2.75 × 10⁻⁶).

Pathway-Based Analysis

We applied the i-GSEA4GWAS-v2 software¹⁹ and the Reactome pathway data set to carry out a pathway-based GSEA for MTAG results. The program obtained a gene across the max-log (P value) from all the SNPs that belonged to that gene within a 20 kb interval. The genes were then ranked via the P value, and the list of ranked genes was then used to calculate the enrichment score (ES) for the pathways. The ES showed a trend in which the genes in the pathways were located at the top of the list of entire ranked genes. The phenotype label permutation and normalization were then performed to obtain the ES distribution and correct the gene and gene set variations. False discovery rate (FDR) analyses were then performed for multiple tests based on the ES distributions obtained by the permutation test.

Network Clustering Analyses of Genes Associated with Insomnia

Pathway-based analysis was used to identify genes and pathways associated with insomnia. However, obtaining the genes in each pathway may not be sufficient to comprehend the topological and functional nexuses of genes in the pathway because some genes may be important in a cluster to construct molecular networks and modulate the biological mechanism of insomnia. Therefore, the study applied Reactome FI software²⁰ to form networks using genes in the significant pathways. The Reactome FI data set contained the Reactome, Panther, BioCyc, KEGG, Pathway Interaction Database, Cancer Cell Map, and other pairwise interactions collected from physical PPIs, protein domain–domain interactions, gene coexpression data, Gene Ontology (GO) annotation, and text mining.

Furthermore, a large molecule network may comprise diverse small modules in which the edges among nodes within a module are tight, whereas the connections among modules are loose. Therefore, network cluster analyses were used to identify the architecture of each module in a large molecular network based on the algorithm as used in a prior study.²¹ Each cluster may be involved in different biological functions and provide novel insight to better understand the pathogenesis of insomnia.

Gene Prioritization across Network Topology

Network topology is capable of providing vital information to understand the structure of a network across identifying the core node using the topological properties within the network. In the present study, we used the CentiScaPe 2.0 software²² to research the topological characteristics of the networks that originated from gene networks. We focused on 2 key node centrality measures for gene networks, namely degree and betweenness, because these are important for biological function.²³ We first imported genes in the significant pathways into Reactome FI networks and then analyzed the core nodes (genes) based on these 2 node centrality measurements.

Analysis of the Patterns of Gene Expression and Coexpression Network for Significant Clusters

We obtained gene expression data on 26 brain regions from BrainSpan (http://www.brainspan.org). The pattern analyses of gene expression and coexpression for each brain region at 3 developmental periods (a prenatal stage of 8 to 37 postconception weeks [PCW], infancy to late childhood at 4 months to 11 years, and adolescence to adulthood at 13 to 23 years) according to Gulsuner et al.²⁴ We then constructed coexpression networks for each brain region at these different developmental periods and defined the absolute value of Spearman rank correlation coefficient (|R|) of expression levels for gene pairs ≥ 0.8 as “connected” and extracted them as a node and edge to form a coexpression network.²⁴ We then applied the CentiScaPe 2.0 program²² to indicate the key nodes based on the topological characteristics of the betweenness and degree.

Moreover, we compared the interconnections of gene networks in Clusters 0 to 6 versus gene networks from 100,000 simulations in the significant pathways according to each brain region and developmental periods. In each simulated data set, we randomly sampled the same number of genes in Clusters 0 to 6 without replacing the genes in the significant pathways and then computed the number of connections as Mn. This process was repeated 100,000 times, and 100,000 instances of Mn were gained to obtain the empirical distribution of Mn under the null hypothesis. The empirical P value was assessed by calculating the scale of Mn that are more than or equal to the number of connections in the real data sets among the 100,000 simulated data sets.

PPI Analysis

In this study, we carried out the permutation test using the Disease Association Protein-Protein Link Evaluator²⁵ and evaluated whether genes in each module had significant physical interactions with each other or with other proteins via network connectivity parameters (number of edges and degree) versus the random networks with similar size and degree distribution.

Partitioned Heritability

Partitioned heritability for MTAG results was implemented across stratified LD regression to identify whether heritability of insomnia was enriched in specific cell types or tissues. Specific method details for calculating heritability estimates for 52 functional categories, GTEX tissue expression, or cell types can be found in the study of Finucane et al,²⁶ in which the stratified LD score was computed using the European subjects in the 1,000 Genomes project, and only SNPs with a minor allele frequency (MAF) > 0.05 in the HapMap 3 were included. Multiple testing was carried out using a Bonferroni correction for 52 functional categories (P < 0.00096), 13 brain regions (P < 0.0038), and 3 neuronal cell types (P < 0.016).

Results

Through LD score regression analysis, we found that insomnia symptoms/complaints and sleep duration have strong genetic correlations (Supplementary Tables 1 to 3). Using the MTAG to combine GWASs of insomnia symptoms/complaints with sleep duration showed higher statistical power for SNP-level and gene-based analysis (Supplementary Table 4; Supplementary Figure 1A to F). From GWAS to MTAG, the number of genome-wide significant genes increased from 2 to 10 for insomnia complaints, from 9 to 14 for insomnia symptoms, and from 4 to 6 for sleep duration (Supplementary Table 5; Supplementary Figure 1D to F).

We used a minimum of 3 sets of MTAG SNP P values (Bonferroni adjusted for the number of traits) as a total MTAG results and conducted a pathway-based analysis for total MTAG results and obtained 17 of the 1,816 Reactome pathways with a gene set enrichment (P < 0.05) and a FDR (P < 0.25; Table 1). We then compared the significant pathways based on the full pathways in Reactome data set, and at least one significant pathway was identified from 10 of 21 core biological processes, including activated TLR4 signaling, cell cycle, metabolism, metabolism of RNA, gene expression, signal transduction, transmembrane transport of small molecules, disease and immune system, and DNA repair (Supplementary Figure 2).

Table 1.

Significant Reactome Pathways Enriched for Association with Insomnia.

Pathway	P	P FDR
IKK complex recruitment mediated by RIP1	0.001	0.081
Vasopressin regulates renal water homeostasis via aquaporins	0.001	0.1505
CLEC7A (Dectin-1) signaling	0.003	0.1954
TLR4 cascade	0.003	0.2105
TP53 regulates transcription of cell death genes	0.005	0.2107
Insulin receptor recycling	0.003	0.2143
Activated TLR4 signaling	0.004	0.2206
Diseases associated with the TLR signaling cascade	0.004	0.2317
Diseases of immune system	0.004	0.2317
DNA replication preinitiation	0.011	0.2323
M/G1 transition	0.011	0.2323
PI3K/AKT signaling in cancer	0.002	0.2392
rRNA modification in the nucleus	0.009	0.2407
Phase 1—functionalization of compounds	0.007	0.2451
Aquaporin-mediated transport	0.001	0.246
Cyclin A: Cdk2-associated events at S phase entry	0.014	0.2489
Formation of incision complex in GG-NER	0.011	0.2498

Note. AKT = protein kinase B; Cdk2, cyclin-dependent kinase 2; FDR = false discovery rate; PI3K = phosphatidylinositol 3-kinase; rRNA, ribosomal RNA; TLR, Toll-like receptor.

A total of 638 genes were obtained from the significant pathways and imported into the larger molecular network (Supplementary Figure 3) and then clustered into several subnetworks. We identified 11 clusters with 7 clusters containing >10 gene members (Figure 1; Supplementary Table 6). We also obtained a series of gene interactions across the edge connection of the genes within each cluster and detected some tight links among or between the clusters (Figure 1). For example, Figure 1 demonstrates the relationships between Cluster 0 and other clusters. Using the GO data set, we conducted enrichment analyses for genes in each module and found several interesting biological processes (Figure 1), such as nerve growth factor receptor signaling pathway (Clusters 0, 1, 2, 3, and 5), axon guidance (Cluster 1), and dopamine catabolic process (Cluster 4). Supplementary Table 7 illustrates all significant biological processes with P _FDR < 0.01.

Figure 1.

Functionally interacting network modules. These were constructed from genes in the significant pathways. Functional interactions between the genes from significant pathways were analyzed and clustered using the ReactomeFI tool and visualized in Cytoscape. Genes are represented as nodes, while the interactions between genes are represented as edges. The parent network was further analyzed to yield subnetwork clusters, and each cluster is separately shown and the color encoded for clarity.

We only found that the interconnections of network and genes in Cluster 4 had more edges than the networks of genes in the significant pathways if |R| ≥ 0.8 in the fetal dorsolateral prefrontal cortex (DFC; P = 7 × 10⁻⁵), inferolateral temporal cortex (ITC; P = 0.02), medial prefrontal cortex (MFC; P < 1 × 10⁻⁵), and amygdala (AMY; P < 1 × 10⁻⁵; Figure 2A to D). RPA2 was the key node in the AMY coexpression network (betweenness = 224.7; Supplementary Figure 4A). ORC6, PIAS3, and PRIM2 were the core nodes in DFC, ITC, and MFC coexpression networks (betweenness = 246.2, 516.5, and 296.1) (Supplementary Figure 4B to D), respectively. Supplementary Figure 5A to D illustrates the average expression level of genes in Cluster 4 with the DFC, ITC, AMY, and MFC in different developmental stages. The average expression of genes in Cluster 4 was high in these 4 brain regions during early prenatal development (8 to 37 PCW).

Figure 2.

The normalized gene expression data for 26 brain regions at multiple developmental stages from BrainSpan were obtained. Coexpression networks of genes in Cluster 4 in AMY (A), DFC (B), ITC (C), and MFC (D) for all 3 developmental stages (a prenatal stage of 8 to 37 PCW, infancy to late childhood at 4 months to 11 years, and adolescence to adulthood at 13 to 23 years). The coexpression of genes in Cluster 4 was analyzed using RNA-seq data from the BrainSpan atlas. Gene pairs were defined as coexpressed if the Spearman rank correlation coefficiency (|R|) > 0.8 for RNA-seq expression levels in the 4 brain regions of a given developmental stage. Histograms represent distributions of the numbers of edges in 100,000 simulated networks using genes in the significant pathways. Dotted lines refer to the numbers of connections (edges) in the networks created using genes in Cluster 4. The significant enrichment in coexpressed genes in Cluster 4 was observed in DFC (P = 7 × 10⁻⁵), ITC (P = 0.02), MFC (P < 1 × 10⁻⁵), and AMY (P < 1 × 10⁻⁵) during prenatal development. AMY = amygdala; DFC = dorsolateral prefrontal cortex; ITC = inferolateral temporal cortex; MFC = medial prefrontal cortex; PCW, postconception weeks.

Moreover, we imported 72 genes in Cluster 4 into the PPI network. The resulting network was significantly different from any random networks (Supplementary Figure 6), in that there were 395 direct edges in the network compared with only 27.63 edges expected by chance in Cluster 4 (P < 0.001). Moreover, the observed average connectivity per gene was 13.39 compared with an expected 1.66 from random networks in Cluster 4 (P < 0.001). These results suggest that the networks generated from genes in Cluster 4 did not occur by chance.

Partitioned heritability analysis for total MTAG results showed that significant enrichments of cell type expression for neuronal tissues were neuronal (P = 1.2 × 10⁻⁴) and astrocytes (P = 3.09 × 10⁻³), not in oligodendrocytes (P = 0.059; Figure 3). Results also demonstrated that functional groups for Conserved LindbladToh (P = 1.34 × 10⁻⁶), TSS Hoffman.extend.500 (P = 9.36 × 10⁻⁵), and Conserved LindbladToh.extend.500 (P = 7.69 × 10⁻⁴; Supplementary Figure 7) were significantly enriched for the heritability of insomnia. We also found significant enrichment of heritability for the frontal cortex (BA9; P = 4.58 × 10⁻⁵), anterior cingulate cortex (BA24; P = 4.92 × 10⁻⁵), nucleus accumbens (P = 5.98 × 10⁻⁴), cerebellar (P = 8.37 × 10⁻⁴), caudate (P = 1.03 × 10⁻³), cortex (P = 1.06 × 10⁻³), hypothalamus (P = 1.13 × 10⁻³), cerebellum (P = 2.32 × 10⁻³), and putamen (P = 2.45 × 10⁻³; Figure 4).

Figure 3.

Enrichment analysis for insomnia using 3 cell types (neuron, astrocyte, and oligodendrocyte).

Figure 4.

Enrichment analysis for insomnia using the 13 brain regions (GTEx data).

Discussion

The present study used the GWAS summary statistics of insomnia symptoms/complaints and sleep duration, combined with a new tool (MTAG), identified the high levels of statistical power that required more sample size and attempted to identify potential gene pathways and networks associated with insomnia. Our current analysis obtained 17 human pathways that participate in at least 10 distinct biological processes that may be related to insomnia risk or susceptibility.

Indeed, we identified several pathways that may be associated with insomnia. Both Gudmundsdottir and colleagues²⁷ and Elder et al.²⁸ revealed that Dectin-1/CLEC7A signaling plays important roles in regulating the composition of the cytokine milieu generated from dendritic cells. Vaure et al. identified that the activation of Toll-like receptor 4 signaling leads to an intracellular signaling pathway nuclear factor-κB (NF-κB) and inflammatory cytokine production, which is responsible for activating the innate immune system.²⁹ During sleep disturbance, the sympathetic nervous system releases norepinephrine into primary and secondary lymphoid organs and stimulates the adrenal gland to release epinephrine into the systemic circulation.^30,31 Neuromodulators, such as epinephrine, acidic fibroblast growth factor, and epidermal growth factor, stimulate leukocyte adrenergic receptors to further activate NF-κB, which subsequently regulates immune response gene transcription, such as tumor necrosis factor (TNF) and interleukin 6, ultimately leading to the translation and production of proinflammatory cytokines that serve to regulate the inflammatory response.^31–33 Furthermore, TNF acts within a complex biochemical network, leading to the stimulation of NF-κB and the modification of the transcription of hundreds of gene products, each of which promotes sleep³⁴ and promotes a positive feedback loop.³¹ These findings support a bidirectional influence of the brain and periphery, enabling the brain to modulate inflammatory activity, and inflammatory activity, in turn, to affect neural processes in the brain, including those related to sleep. However, when this bidirectional communication is disturbed by sustained insomnia, a feed-forward dysregulation of sleep may occur.

The current study identified several significant enrichments within Cluster 4 in prenatal DFC, ITC, MFC, and AMY, and the average expression of genes in Cluster 4 was high in these brain regions during 8 PCW-2 years old. These results suggest that insomnia may be related to neurodevelopmental abnormalities. Insomnia is the most common sleep problem in children with neurodevelopmental disorders, with more than 85% reported to meet the required criteria.³⁵ For instance, attention deficit hyperactivity disorder, Autism spectrum disorder, cerebral palsy, and fetal alcohol spectrum disorder are all associated with higher typical rates of insomnia despite their different manifestations and etiologies.^36,37 These brain regions are likely involved with the biological mechanism of insomnia. For example, a previous study identified that the patients with primary insomnia showed significantly reduced gray matter concentrations in the DFC and decreased gray matter volumes in the MFC compared with normal controls, suggesting that the abnormal DFC and MFC is an intermediate phenotype linked to insomnia risk. Another study of 27 primary insomnia patients and 30 healthy controls conducted by Dae Lim Koo et al.³⁸ found that an atrophic change in the AMY correlated negatively with frontal function in these patients. Li et al.³⁹ also detected that primary insomnia patients showed higher ALFF values in the right inferior temporal cortex. Meanwhile, we also identified some significant enrichments of heritability in the frontal cortex, anterior cingulate cortex, nucleus accumbens, caudate, hypothalamus, cerebellum, and putamen. Suh et al.⁴⁰ found that patients with persistent insomnia symptoms have cortical thinning in the anterior cingulate cortex and lateral prefrontal cortex compared to good sleepers. The patients with fatal familial insomnia also showed marked hypometabolism in the bilateral thalami, caudate nuclei, hypothalamus, and posterior cingulate gyri.⁴¹ Therefore, LD score regression analysis further supports the network analysis results, which suggest that insomnia may be closely related to prenatal abnormal connectivity of multiple brain regions, disturbance of circulation activity, and decline of integration function.

Using the stratified LD score regression, we detected significant enrichments for neuronal and astrocytes. Yamashita et al.⁴² provided novel evidence that astrocytic activation in the anterior cingulate cortex can imitate sleep disturbance based on the increased release of glutamate to decrease the extracellular concentration of γ-aminobutyric acid, and the patients with fatal familial insomnia have severe neuronal loss and prominent astrocytic gliosis in the mediodorsal and anterior nuclei of the thalamus.⁴³

The current study also had some limitations. For instance, we may have found some false-positive results based on the selection of significant pathways with FDR less than 0.25. Therefore, future studies will continue to explore these findings. In conclusion, the present study may provide potential associations among pathways, brain regions, and insomnia risk.