High-throughput sequencing of the synaptome in major depressive disorder

Abstract

Major depressive disorder (MDD) is among the leading causes of worldwide disability. Despite its significant heritability, large-scale genome-wide association studies (GWASs) of MDD have yet to identify robustly associated common variants. Although increased sample sizes are being amassed for the next wave of GWAS, few studies have as yet focused on rare genetic variants in the study of MDD. We sequenced the exons of 1742 synaptic genes previously identified by proteomic experiments. PLINK/SEQ was used to perform single variant, gene burden and gene set analyses. The GeneMANIA interaction database was used to identify protein–protein interaction-based networks. Cases were selected from a familial collection of early-onset, recurrent depression and were compared with screened controls. After extensive quality control, we analyzed 259 cases with familial, early-onset MDD and 334 controls. The distribution of association test statistics for the single variant and gene burden analyses were consistent with the null hypothesis. However, analysis of prioritized gene sets showed a significant association with damaging singleton variants in a Cav2-adaptor gene set (odds ratio = 2.6; P = 0.0008) that survived correction for all gene sets and annotation categories tested (empirical P = 0.049). In addition, we also found statistically significant evidence for an enrichment of rare variants in a protein-based network of 14 genes involved in actin polymerization and dendritic spine formation (nominal P = 0.0031). In conclusion, we have identified a statistically significant gene set and gene network of rare variants that are over-represented in MDD, providing initial evidence that calcium signaling and dendrite regulation may be involved in the etiology of depression.

INTRODUCTION

Major depressive disorder (MDD) is a leading cause of suffering and morbidity worldwide.¹ MDD is associated with an increased risk of mortality from both suicide and early morbid sequelae from common medical illnesses.² Despite this overall burden, little is known about its pathophysiology. Broad risk factors such as family history and early adversity have been well established to increase the risk of MDD in adulthood; yet, how these risk factors give rise to the often dynamic phenotype of MDD remains unknown.³ Twin and molecular-based studies have found consistent evidence for a moderate heritable component, with heritability estimates typically ranging from 30% to 40%.⁴ However, genome-wide linkage studies have struggled with inconsistent replication³ and meta-analyses of large-scale genome-wide association studies (GWAS) have so far failed to identify findings with genome-wide significance.⁵ Together, these results suggest that a major variant of large effect is unlikely to exist and are consistent with the presence of prominent genetic heterogeneity.

Current GWAS identify rare variation in the form of large copy-number variants but do not, by design, assay rare single-nucleotide variants (SNVs), which form the bulk of genetic variation in individuals.⁶ Large-scale surveys of rare variation have recently been made possible by ‘next-generation’ sequencing studies, which have been applied to the study of a number of psychiatric disorders, including intellectual disability,⁷ autism,⁸ schizophrenia (SCZ)^9,10 and bipolar disorder (BP).¹¹ Although the precise genetic architecture may differ by phenotype, an emerging theme from these early sequencing efforts has been the presence of locus heterogeneity, suggesting the need for initial efforts to focus on gene sets and pathways, as these analyses provide greater power by collating larger numbers of rare variants.

Few studies have so far focused on high-throughput sequencing in MDD. To our knowledge, only one such study of MDD has been performed, but exome sequencing was carried out on only 10 subjects with an ‘extreme’ pharmacological phenotype.¹² In the current study, we sought to perform a hypothesis-driven exploration of rare variants in genes found within the synapse, a complex organelle-like structure that has been implicated in the pathophysiology of major mental illness and is the target of almost all currently active psychotropic medications, including antidepressants.¹³ We leverage the results of recent proteomic studies that have comprehensively catalogued the constituent proteins of the synapse¹⁴ to perform high-throughput sequencing on a mutational target with a greater a priori likelihood of being involved in the pathophysiology of major mental illness, such as MDD.

MATERIALS AND METHODS

Sample

Cases were selected from the Genetics of Early-Onset, Recurrent Depression (GenRED) study, which previously ascertained a sample of European-American families with early-onset (age at onset <30 years) depression across six US sites.¹⁵ Informed consent was obtained from all cases and original Institutional Review Board's approval was provided by each of the recruiting sites, including Johns Hopkins (IRB approval number: NA_00035775). Cases were interviewed with the Diagnostic Instrument for Genetic Studies and underwent best-estimate diagnostic procedures. For the current study, we selected cases from independent families who had at least one first-degree relative also diagnosed with early-onset depression. These cases had an average age of onset of 18.9 years and had suffered a mean of 7.4 lifetime depressive episodes. Controls of European-American ancestry were selected from the NIMH Genetics Initiative, excluding subjects with a history of psychotic symptoms, BP or sufficient symptoms to meet a Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition, depressive episode criteria.

Synaptic gene target selection

At the onset of this study (2010), we performed an evaluation of all published or publically available proteomic studies focused on the synapse. A comprehensive list of these synaptic proteins was compiled into SynaptomeDB (http://metamoodics.org/SynaptomeDB).¹⁶ Our search strategy identified 1886 individual synaptic proteins, consisting of proteins found in synaptic vesicles (n = 107), the presynaptic membrane (n = 336), the presynaptic active zone (n = 209) and the postsynaptic density (n = 1,755).

Capture design

We used an Agilent Sure Select Custom Panel to design baits targeting the synaptome genes (n = 50 079 baits) based on the bait sequences of the Agilent Sure Select All Exome Version 2 Capture Kit (Agilent, Santa Clara, CA, USA). We added additional baits to provide further coverage of GC-rich genes (n = 5,251) and miRNA-binding sites (n = 973). In all, the capture targeted 6.7 Mb of genomic sequence.

High-throughput sequencing

We used the Illumina TruSeq DNA Sample Preparation Kit (Illumina, CA, USA) with 12 multiplexed adapters. Multiplexed DNA libraries were pooled at equal molar ratio before hybridization with the custom Agilent Sure Select probes. Following elution of captured sequence, we performed 6–8 cycles of PCR-based amplification. Samples passing standard quality-control (QC) metrics were sequenced on the Illumina HiSeq2000 (Illumina, CA, USA). Twelve multiplexed samples were sequenced per lane using 75 base pair-end read module.

Data processing and QC

The various QC steps are highlighted in Supplementary Figure S1. After separating the reads into individual samples based on the index barcode, reads were aligned to the human reference genome (UCSC hg19) using the BWA aligner,¹⁷ allowing for two mismatches in the 30-base seed. Picard (http://picard.sourceforge.net/) was used to fix mate pair mismatches and to remove reads with identical outer mapping coordinates, which represent likely PCR artifacts. Target coverage for the Agilent Sure Select capture was assessed using Picard’s HSmetrics utility. Genome Analysis Toolkit (GATK) was used to generate SNV and small indel calls within the targeted regions.¹⁸ We combined all cases and controls and performed multisample variant calling using GATK’s Unified Genotyper with variant recalibration. SNV clusters (defined as >3 SNVs per 10 bases) and SNVs falling within a called indel region were masked. We used variant quality score recalibration to generate a final list of high-confidence SNVs and indels as recommended by GATK documentation.^19,20 We also added three additional QC steps to reduce the rate of false-positive calls: genotype quality >20; a depth-of-coverage of at least 10; and an average allele balance for heterozygous genotypes between 0.2 and 0.8.

Variant Call Format (VCF) files were converted to PLINK file format using VCF tools and custom scripts. PLINK was subsequently used to remove individuals with >50% missing genotype, followed by the removal of variants with a missing rate >10% and variants in Hardy–Weinberg Disequilibrium (P < 1 × 10⁻⁶). After these QC steps, the study-wide missing rate was 97.6% (Supplementary Table S1). Principal component analysis (PCA) of the case–control sample was performed using Eigenstrat²¹ to assess for potential population stratification using common sequenced variants (minor allele frequency (MAF) >5%) in approximate linkage equilibrium (Supplementary Figure S2). We inspected the top axes of variation in each PCA component and removed three outlier individuals, with the remaining samples showing appropriate clustering consistent with a European-American sample. After applying all the above QC steps to an initial sample of 308 cases and 353 controls, the final sample size was 259 cases and 334 controls.

Annotation

Identified variants were annotated with ANNOVAR²² using Ensembl version 63 as the reference database. We considered two functional categories of variants in the analyses described below: loss of function and damaging. We categorized loss-of-function variants as those that were stopgain, stoploss, canonical splicing or indel frameshift variants. Damaging variants included loss-of-function variants, in addition to the broader class of missense variants that were assessed as potentially damaging by either SIFT or Polyphen-2 using default thresholds (SIFT >0.95 or PolyPhen >0.85). We also considered three different frequency cutoff categories: singletons (or mac1), MAF <1% (maf01), and MAF <5% (maf05). We defined allelic thresholds based on estimates from the European-American and ALL populations of the NHLBI-ESP and the European-American and ALL populations of the 1000 Genomes April 2012 release.

Statistical analysis

We carried out a coordinated series of genetic association tests at the level of the individual variant(s), individual gene(s) and previously defined gene sets. For variant-level tests, we used logistic regression with Firth's penalized likelihood method,²³ controlling for the top five ancestry-based principal components. Firth’s method provides unbiased estimates of the logistic regression model when there is sparse data,^24,25 as is often the case with rare variants. For the gene-level tests, we used two complementary approaches. We used PLINK/SEQ to calculate gene burden with the SMP (statistic/matrix/permutation) utility as described in Purcell et al.,⁹ to control for any study-wide differences in the rates of variation across cases and controls. Significance of the gene burden results is evaluated with 10 000 permutations. We also used EPACTS (http://genome.sph.umich.edu/wiki/EPACTS) to calculate SKAT-O tests.²⁶ The PLINK/SEQ burden test is a one-sided that assumes rare functional alleles increase risk for disease, while SKAT-O allows for both risk and protective alleles. We carried out these gene-level tests using the predefined functional and allele frequency categories.

Finally, we used two complementary approaches for gene set-level tests. First, we tested for a burden of risk variants in sets of genes that were based on prior evidence. We followed the approach of a recent SCZ exome sequencing study of SCZ,⁹ which defined a number of gene sets from the literature, including prior studies of GWAS, copy-number and de novo variants, as well as proteome-based studies of the postsynaptic density and the Cav2 calcium channel. Although the prior evidence implicating any specific type of genetic variation to MDD is much more limited, evidence of shared variation across multiple psychiatric phenotypes, including MDD and SCZ,²⁷ led us to consider the same primary gene sets used in the SCZ exome sequencing study,⁹ with the added inclusion of two gene sets that include the top published findings from Psychiatric Genomics Consortium GWAS of MDD⁵ and BP²⁸ (genes within 300 kb of associated loci with meta-analysis association P < 10⁻⁴). Further details, including gene set membership, are available in Supplementary Table S4. We tested only predefined gene sets with at least five genes sequenced in our study. To obtain a study-wide level of significance accounting for all gene sets tested and the different analyses carried out with these gene sets, we repeated the entire gene set testing procedure after randomly permuting case–control labels 1000 times.

We also used a more agnostic gene set strategy by adapting an approach used in a previous study²⁹ to examine whether genes with nominal evidence of association from our PLINK/SEQ-based gene-level test were more likely to interact with each other than expected by chance, thereby suggesting a common molecular pathway. We downloaded the direct protein–protein interaction data set from GeneMANIA³⁰ and calculated the number of nodes and edges between the top associated genes in our observed data, where nodes represents proteins and edges represent interactions between any two proteins. To evaluate the significance of the observed network, we randomly selected an equal number of genes from the GeneMANIA database that were also sequenced in our study and were matched by gene size (±10 kb) with our observed genes. We repeated this 10 000 times to calculate the number of times that the number of nodes or edges in the randomly drawn genes was greater than what was found in the observed data.

RESULTS

After extensive QC (summarized in Supplementary Figure S1), the final cleaned data set included 259 cases with early-onset, recurrent major depression and 334 controls sequenced to an average depth of 50 ×, with 88.1% of bases being covered with at least 6 × reads. In total, there were 1742 genes with ≥70% of bases covered at 6 ×, which was used as a minimum coverage threshold for downstream analyses. Basic QC metrics are shown in Supplementary Table S1. Controls were sequenced to slightly higher depth, but there were otherwise minimal differences across QC metrics between cases and controls, including a comparable distribution of variants across annotation classes in cases and controls.

Single variant and gene-based analyses

Our results for the single-variant analyses were consistent with the null hypothesis (Supplementary Table S2 and Supplementary Figure S3). We subsequently performed gene-based tests using three frequency cutoffs (maf05, maf01 and singletons or mac1) and two functional categories representing a ‘damaging’ (defined by SIFT or Polyphen2) or ‘loss-of function’ effect. We implemented gene-based testing using PLINK/SEQ’s one-sided burden test or SKAT-O, a kernel-based test that is more sensitive to the presence of both protective and risk variants. The results, shown in Supplementary Table S3 and Supplementary Figure S4, show a number of genes with nominally significant enrichment for rare damaging variants, although no single gene-based association result survived correction for the full number of genes tested (best gene burden association P = 0.002).

Gene set and network analyses

Given the limited power of our sample to detect effects at the variant or gene level³¹, our primary focus was on the analysis of gene sets and gene networks, which could aggregate a sufficient number of rare variants for an enrichment analysis to reach statistical significance. Based on evidence suggesting that the genetic etiology of MDD partially overlaps with that of BP and SCZ,²⁷ we used prior findings from these related psychiatric phenotypes, as well as two gene sets comprised of the top loci from the published Psychiatric Genomics Consortium GWAS meta-analysis of BP and MDD (genes within 300 kb of a locus with a meta-analysis association P < 10⁻⁴). The 41 prioritized gene sets (described in Supplementary Table S4) generally fall into four main classes: (1) statistically defined associations from prior studies of common and rare variation in SCZ, BP and MDD; (2) proteome-based gene sets of the postsynaptic density and its substructures; (3) proteome-based gene sets of the Cav2 calcium channel; and (4) annotation-based gene sets of the major ion channels. We tested for the enrichment of risk variants in gene sets that contained at least five genes in our post-QC sequenced sample, using the same three allele frequency (maf05, maf01 and mac1) and two functional annotation (damaging and loss of function) categories from the gene-level analyses. The top results are shown in Table 1, with the full results shown in Supplementary Table S5. The strongest association was in the calcium channel Cav2-channel ‘adaptor’ gene set, which showed an odds ratio (OR) = 2.6 (P = 0.0008) in the damaging singleton category and showed at least nominal evidence of association across the various annotation classes. This proteomically defined gene set consisted of 20 genes, including 11 that were sequenced in our study, of which 7 had over-representation of rare variants in cases compared with controls (STXBP5, RIMS1, CTNNB1, DMXL2, SYN1, YWHAB, YWHAH). This gene set remained significant after accounting for the multiple testing of the different frequency and annotation categories via permutation (empirical P = 0.049).

Table 1.

Top results from the prespecified gene set analysis showing all associations with a nominal P < 0.05

Gene/set	Annotation/frequency category	Variants in affected/unaffected	OR	P-value	P-value (adjusted)
Cav2 adaptors	Damaging/mac1	22/11	2.58	0.0008	0.049
Neuronal synaptic vesicle	Lof/maf05	10/3	4.30	0.0171	0.780
Cav2 adaptors	Damaging/maf05	41/35	1.51	0.0174	0.785
Neuronal synaptic vesicle	Lof/maf01	10/3	4.30	0.0187	0.801
Neuronal PSD	Lof/mac1	32/32	1.29	0.0211	0.834
Cav2 adaptors	Damaging/maf01	32/25	1.65	0.0221	0.843
Neuronal synaptic vesicle	Lof/mac1	8/3	3.44	0.0224	0.848
Neuronal early endosomes	Damaging/mac1	4/1	5.16	0.0240	0.866
Neuronal plasma membrane	Damaging/mac1	20/11	2.34	0.0244	0.875
Neuronal plasma membrane	Damaging/maf01	29/23	1.63	0.0279	0.907
Neuronal plasma membrane	Damaging/maf05	38/34	1.44	0.0390	0.954
Neuronal Golgi	Damaging/mac1	11/9	1.58	0.0421	0.966
MDD GWAS (P < 1 × 10−4)	Lof/mac1	8/4	2.58	0.0495	0.979
Neuronal ARC	Damaging/maf05	46/40	1.48	0.0499	0.981

Abbreviations: GWAS, genome-wide association studies; MDD, major depressive disorder; OR, odds ratio; PSD, postsynaptic density.

In an alternative network-based approach, we selected all genes with a PLINK/SEQ gene burden association P-value < 0.05 (n = 28 for maf05 and n = 16 for maf01) and tested whether these genes were more interconnected than expected by chance. We evaluated how many times these nominally significant genes were observed to interact in the curated protein–protein interaction database from GeneMANIA and compared this with 10 000 randomly drawn sets of an equal number of genes, which were sequenced in our study and matched by gene size (±10 kb). This analysis found an interconnected network of proteins under in maf05 frequency threshold (shown in Figure 1), which was statistically significant for the number of proteins found to interact with each other (nodes; P = 0.0031) and near significant for the number of pairwise interactions between these proteins (edges; P = 0.054). The analysis of the smaller list of nominally associated genes with maf01 also showed an excess number of nodes and edges, but these were not statistically significant, with empirical P-values of 0.053 and 0.11, respectively.

Figure 1.

Protein–protein interaction network identified using GeneMANIA (direct interaction database). The histograms showed the null distribution, with the line showing the analysis’ observed results and its empirical P-value.

DISCUSSION

MDD remains one of the few prominently heritable phenotypes that have yet to yield reliable genetic findings, suggesting the presence of extensive etiological and phenotypic heterogeneity. In this study, we have performed an initial study of rare variation in a large subset of genes with a priori relevance to MDD and have attempted to improve our chances of detecting a meaningful association by focusing on subjects with early-onset, recurrent MDD, which may be a more heritable form of the illness.¹⁵ Although our sample size was insufficient to identify a statistically significant rare variant burden at the single gene level, we were able to identify a significantly associated gene set and a gene network with enriched protein–protein interactions. Together, these findings provide initial evidence that calcium channels and dendrite regulation may be involved in the pathophysiology of depression.

We tested a number of prespecified gene sets with prior evidence for involvement in mental disorders such as SCZ and BP, reasoning that shared genetic risk across common variants between these disorders and MDD²⁷ would also be seen for rare variants. Indeed, out of the 41 prioritized gene sets, only one gene set (damaging, singleton variants in the Cav2-adapter) showed sufficient evidence for association (OR = 2.6; P = 0.0008) to remain significant after a permutation procedure accounting for all the annotation and frequency classes tested (empirical P = 0.049). This gene set showed nominal evidence for association in both the damaging maf5 (OR = 1.51, P = 0.0174) and the damaging maf1 (OR = 1.65, P = 0.0221) categories, but its stronger evidence for association with the singleton category is consistent with findings from theoretical models³² and from studies of other neuropsychiatric disorders^9,10 that point to rarer mutations being potentially more penetrant. The Cav2-adaptor gene set, as part of a broader set of Cav2-related genes, was initially reported in a proteomic analysis of proteins associated with the central Cav2 α1 subunit and its four β accessory subunits.³³ The Cav2-adaptor gene set consists of a diverse group of genes encoding proteins involved in scaffolding, exocytosis and intracellular signaling. Among the genes contributing to the gene set association signal, there is a notable convergence on the process of synaptic vesicle exocytosis (RIMS1, STXBP5, CTNNB1, DMXL2, SYN1).^34–38 Not unexpectedly, a number of rare variants within these genes have also been found in brain disorders, such as autism^39,40 and epilepsy.⁴¹ RIMS1 is also notable for being one of the 108 recently identified genome-wide significant loci in SCZ.⁴²

The complimentary protein–protein interaction analysis was driven by a network of 14 genes that included several of actin-based cytoskeleton genes (GSN, ARPC3, SPTBN1, CALP1, MYO1C). Gelsolin (GSN) is an actin-binding protein, which regulates actin polymerization by ‘severing’ actin polymers when activated by calcium signaling.^43,44 ARPC3 encodes a critical subunit of the actin-related protein 2/3 complex, which is a regulator of actin polymerization and is involved in the formation of dendritic spines.⁴⁵ SPTBN1, also known as βII-spectrin, is a scaffolding protein that binds the actin cytoskeleton to stabilize axonal projections⁴⁶ and induce neurite formation.⁴⁷ Intriguingly, βIIspectrin is also a substrate of calpain 1 (CALP1), which was also identified in the network analysis, and is a calcium-activated protease and potential mediator of corticotropin-releasing hormone-induced loss of dendritic spines.⁴⁸ Finally, MYO1C, is a widely expressed myosin molecule with binding affinity for actin and the cell-signaling membrane inositide PIP₂.⁴⁹ Its role in the synapse remains unknown, although it has been found to have a role in axonal guidance⁵⁰ and in mechanical amplification of electrical stimuli in vestibular hair cells.⁵¹

The genes identified in this network analysis converge on the role of actin polymerization and regulation of dendritic spines, which may underlie the key mechanisms responsible for synaptic plasticity.⁵² More broadly, abnormalities of size, density and shape of dendritic spines have been hypothesized to have a role in various brain disorders, including intellectual disability,⁵³ autism,⁵⁴ Alzheimer’s disease,⁵⁵ SCZ and BP,⁵⁶ as well as MDD.⁵⁷ There is also increasing evidence to suggest that they may additionally be involved in the regulation of mood: lowered spine densities have been found in postmortem brains with two prototypical mood disorders (BP and MDD)^56,57; in addition, chronic stress, long considered one of the most robust risk factors for mood disorders, has been associated with excessive pruning of dendritic spines;⁵⁸ and finally, the antidepressant ‘effect’ of both traditional and novel pharmacological agents has been proposed to be mediated by increased formation of dendritic spines in the prefrontal cortex.^58,59

As one of the first attempts to investigate the role of rare variants in MDD, our study should be seen in the context of a number of limitations. First, the sample size is modest for genetic studies of rare variation, leading us to focus the primary analyses on gene sets and pathways, as associations in these aggregate ‘analysis units’ are more likely to be detected in smaller sample sizes.⁹ Power analyses (Supplementary Figure S5a) based on this study’s sample size revealed sufficient power (>80%) to detect gene burden results with ORs > 2.5 under cumulative burden frequencies of approximately 10%. Yet, in scenarios where the burden frequencies were <5%, this study only had sufficient power to detect gene burden results with relatively high effect sizes (OR > 3). Power calculation of the gene set analyses, however, showed significantly higher power to detect associations with gene sets with OR > 2.0 given the gene set cumulative frequencies that were seen in our study (Supplementary Figure S5b). Nevertheless, even in the gene set analyses, our study sample was significantly underpowered to detect associations with ORs < 2.0. A second important limitation is that, although we surveyed 1742 genes hypothesized to have greater a priori likelihood of being associated with MDD, an exome or whole genome approach would have offered a more comprehensive overview of genes potentially associated with MDD, albeit one with a greater burden of multiple testing and, therefore, a need for larger sample sizes. Third, our sequencing yielded a ‘medium’ level of coverage, which may have missed or miscalled particularly rare variants, such as singletons. To allay these concerns, we applied stringent QC procedures and used the Integrative Genomics Viewer⁶⁰ to individually visualize all the variants in the significant gene set and in protein–protein interaction network. In addition, we performed Sanger sequencing and successfully validated all variants found in cases within the ‘singleton Cav2-adaptor’ gene set category for which we had additional DNA (20 out of 22 variants). Fourth, our sequencing study was likely limited by what has been termed the ‘annotation gap,’ where the known and unknown imprecision of bioinformatics annotation may have led to the inclusion of false positives and false negatives in our analysis of ‘damaging’ variants.⁶¹ As sequencing becomes more mainstream, functional annotation will likely become increasingly accurate; however, for the time being, we have, by necessity, resorted to the most widely used tools. Finally, our sequencing strategy has focused only on the more ‘interpretable’ exonic variations, with no systematic sequencing of non-coding regions, which represent an even more significant annotation challenge.

In conclusion, this study suggests that the evaluation of rare variants may be a promising endeavor even for phenotypes as heterogeneous as depression. Limited success in GWAS may not imply lack of success in rare variant studies, as highlighted by the successful identification of rare, but not yet common, variants in autism.⁶² Although our findings need to be replicated in independent samples, the presence of a statistically significant gene set and network analysis suggests that focusing on rare variants may be a tractable way to elucidate at least part of the genetic underpinnings of MDD.