Associations between DNA methylation and schizophrenia-related intermediate phenotypes a gene set enrichment analysis

Abstract

Multiple genetic approaches have identified microRNAs as key effectors in psychiatric disorders as they post-transcriptionally regulate expression of thousands of target genes. However, their role in specific psychiatric diseases remains poorly understood. In addition, epigenetic mechanisms such as DNA methylation, which affect the expression of both microRNAs and coding genes, are critical for our understanding of molecular mechanisms in schizophrenia.

Using clinical, imaging, genetic, and epigenetic data of 103 patients with schizophrenia and 111 healthy controls of the Mind Clinical Imaging Consortium (MCIC) study of schizophrenia, we conducted gene set enrichment analysis to identify markers for schizophrenia-associated intermediate phenotypes. Genes were ranked based on the correlation between DNA methylation patterns and each phenotype, and then searched for enrichment in 221 predicted microRNA target gene sets.

We found the predicted hsa-miR-219a-5p target gene set to be significantly enriched for genes (EPHA4, PKNOX1, ESR1, amongst others) whose methylation status is correlated with hippocampal volume independent of disease status. Our results were strengthened by significant associations between hsa-miR-219a-5p target gene methylation patterns and hippocampus-related neuropsychological variables. IPA pathway analysis of the respective predicted hsa-miR-219a-5p target genes revealed associated network functions in behaviour and developmental disorders.

Altered methylation patterns of predicted hsa-miR-219a-5p target genes are associated with a structural aberration of the brain that has been proposed as a possible biomarker for schizophrenia. The (dys)regulation of microRNA target genes by epigenetic mechanisms may confer additional risk for developing psychiatric symptoms. Further study is needed to understand possible interactions between microRNAs and epigenetic changes and their impact on risk for brain-based disorders such as schizophrenia.

1 Introduction

Schizophrenia is a highly heritable mental disorder [1], and after the decoding of the human genome, efforts to identify schizophrenia-associated genes have intensified. However, schizophrenia is genetically complex and likely involves thousands of genetic risk variants with small contributions to the disorder [2]. Recent meta-analyses of large genetic studies have identified novel genetic variants associated with schizophrenia or schizophrenia-related phenotypes [3, 4] but investigations have largely failed in finding replicable schizophrenia-specific susceptibility genes due to the lack of an etiologically-related phenotypic definition and the aforementioned polygeneity [5]. Furthermore, modulating environmental factors, which can affect gene expression through epigenetic mechanisms, have not been well studied. At present, DNA methylation is the most frequently studied epigenetic mechanism [6]. The covalent binding of a methyl group to the 5′ end of cytosine in DNA CpG dinucleotides can silence the gene and inhibit gene expression [7]. It has been proposed that measuring DNA methylation changes may provide a better measure of environmental impacts than evaluating environmental factors themselves [8, 9]. Due to their dependence on genetic and environmental factors, epigenetic markers integrate information from both causal contributors to phenotypes. Therefore, understanding epigenetic signatures may help to improve our understanding of the pathogenesis of schizophrenia [10, 11].

Since intermediate phenotypes are thought to be more proximal to the underlying biology of schizophrenia than varying clinical constructs [12, 13], in this study we analyzed DNA methylation in conjunction with three different widely acknowledged schizophrenia-related intermediate phenotypes: reduced hippocampal volume, reduced superior temporal gyrus (STG) thickness, and dorsolateral prefrontal cortex (DLPFC) working memory (WM) load-dependent neural activity (% BOLD signal change). A reduction of hippocampal volume in schizophrenia patients has repeatedly been demonstrated [14–17] and abnormalities of hippocampal structure and function in schizophrenia have been associated with deficits in memory and executive function [18]. The STG is involved in auditory processing [19] and social cognitive processes [20]. Previous studies have shown structural alterations of the STG in schizophrenia patients [21–24], which have also been associated with auditory hallucinations [25–27]. Another well-studied intermediate phenotype is DLPFC activity during WM processing [28–31]. Schizophrenia patients are thought to require additional neural resources to achieve a comparable WM performance as healthy individuals. Depending on the WM load this can result in increased DLPFC response, which has been termed “neural inefficiency” [32–36].

An important class of molecules that we expect to affect risk for schizophrenia possibly via modulation of intermediate phenotypes for schizophrenia are microRNAs (miRNAs) and their target genes [37]. MiRNAs are small non-coding RNAs that can post-transcriptionally silence the expression of a large number of target genes either through translational inhibition or degradation of target mRNAs [38]. A set of miRNA target genes regulated by a single miRNA likely constitutes a biological network of functionally-associated molecules. Consequently, dysregulation of a single miRNA could be powerfully influential for a polygenic disorder such as schizophrenia. In line with that, the largest schizophrenia genome-wide association studies (GWAS) conducted to date identified variants within the hsa-miR-137 host gene, MIR137HG, as well as variants within validated miR-137 target genes to be among the top risk variants for schizophrenia [39, 40], suggesting a potential role for the miRNA in the disorder. Several studies suggest that one of the identified MIR137HG variants is associated with schizophrenia-related intermediate phenotypes such as dorsolateral prefrontal cortex hyperactivation [41] and dorsolateral prefrontal-hippocampal functional connectivity [42]. Furthermore, Wright et al. [43] found significant enrichment for schizophrenia-associated genes among the list of potential and experimentally validated miR-137 targets, as well as significant enrichment of targets within schizophrenia-relevant canonical pathways, such as those involved in neuronal function and development. In an imaging genetics approach, Potkin et al. [44] discovered gene regulatory networks of GWAS-identified risk variants for schizophrenia that are assumed to be regulated by several miRNAs, including miR-137 and others (miR-448, miR-218, miR-182, miR-518C, miR-200B, miR-429, miR-374, miR-369-3P, miR-27A, and miR-27B). Apart from miR-137, other miRNAs (such as miR-15, miR-219, miR-508) also have extensive evidence of their potential involvement in the pathophysiology of mental disorders [45, 46].

MiRNA-mediated regulation of target genes is highly correlated with miRNA target-gene specific promoter methylation [47]. Simultaneous changes of DNA methylation combined with miRNA dysregulation could thus potentiate effects on “downstream” genes (i.e. genes in the regulatory pathway of a miRNA) and various phenotypes. So far, there is little knowledge about the cooperative regulation of gene expression through miRNA targeting and DNA methylation. Analyzing miRNA target gene networks (instead of single gene analyses) and their epigenetic alterations may further deepen our understanding of the biological pathways underlying a complex illness such as schizophrenia.

In the present study we conducted gene set enrichment analyses (GSEA) using the predicted¹ miRNA target gene sets provided by the Molecular Signatures Database v4.0 (http://www.broadinstitute.org/gsea/msigdb/) of the GSEA toolbox [48]. In contrast to recent studies that mostly investigated enrichment in gene expression data sets [49–51], we explored DNA methylation in schizophrenia patients and healthy controls to identify potential associations between network level epigenetic changes in predicted miRNA target gene sets and widely studied intermediate phenotypes for schizophrenia. Gene set enrichment analysis holds the advantage that pathways can be reliably detected even when effect sizes of individual genes are small or signal-to-noise ratio is low, which is of importance especially for polygenic disorders such as schizophrenia. To the best of our knowledge, this approach combining DNA methylation and intermediate phenotypes in a gene set enrichment analysis has not been applied previously in the field of schizophrenia. Since we were interested in phenotypes associated with a brain disorder, we only included CpG sites (and corresponding genes) for which at least moderate correlation in DNA methylation between blood and brain tissue can be assumed [52].

2 Material and Methods

2.1 Participants

Imaging, genetic, epigenetic and behavioral data from participants of the Mind Clinical Imaging Consortium (MCIC) study of schizophrenia from four participating sites (the University of New Mexico (UNM), the University of Minnesota (UM), Massachusetts General Hospital (MGH), and the University of Iowa (UI)) were used to determine DNA methylation and genetic polymorphisms in cryo-conserved blood samples and to analyze structural and functional intermediate phenotypes. Out of a total of 328 participants, blood samples were available for 234 participants. DNA methylation and genetic data of 214 participants passed epi-/genetic quality control procedures, resulting in a final dataset of 103 schizophrenia patients and 111 healthy controls after imaging quality control steps (see below). Patients had a Diagnostic and Statistical Manual of Mental Disorders (DSM-IV) diagnosis of schizophrenia, established using a Structured Clinical Interview for DSM disorders (SCID) [53] and a review of case files by trained clinicians. In the initial cohort, controls were matched to the patient group for age, sex, and parental education and were excluded if they had a history of a medical or Axis I psychiatric diagnosis. For additional details about the participants and clinical measures, see Ehrlich et al. [14], Gollub et al. [54] and Supplementary Material (SM) 1.1.

2.2 Behavioral task

The Sternberg Item Recognition Paradigm (SIRP) is a working memory task, previously shown to consistently activate the DLPFC in healthy controls and schizophrenia patients [35]. The SIRP was administered during six 46 second blocks per run for three 360 second runs. In each block, a memory set composed of one (load 1), three (load 3), or five (load 5) digits was presented (two blocks per load condition). The Encode phase was followed by a presentation of 14 digits, one at a time (the Probe phase) and participants responded to each probe to indicate whether or not the probe digit was in the memory set. For additional details about the paradigm, see Roffman et al. [55] and SM 1.2. The stimuli and responses were presented and collected using E-prime software (EPrime v1.1, Psychology Software Tools, Inc., Pittsburgh, PA). Participants were excluded from further analysis, if they completed a block with less than a 75% accuracy rate and/or with more than 6 probes not answered within a block.

2.3 Image Acquisition and Processing

Structural magnetic resonance imaging (MRI) data was acquired with either a 1.5T Siemens Sonata (UNM, MGH, and UI) or a 3T Siemens Trio (UMN). Cortical reconstruction and volumetric segmentation based on high resolution structural MRI scans was performed with the FreeSurfer surface reconstruction software (http://surfer.nmr.mgh.harvard.edu). Hippocampal volume and STG thickness are part of the FreeSurfer standard output, mean volume and thickness values for each participant were calculated using measures from the left and right hemisphere, respectively.

Functional MRI data was acquired with either a 1.5T Siemens Sonata (UNM) or a 3T Siemens Trio (UMN, MGH, and UI) and was analyzed using fMRIB Software Library (FSL) (http://www.fmrib.ox.ac.uk/fsl). A Functional Imaging Linear Model [56] was fit to model the Probe phases of each subject’s preprocessed functional time series. We used the linear Contrasts Of Parameter Estimate (COPE) Probe-load 5 versus Probe-load 1 and refer to responses to this condition as ‘load-dependent’ activation [55]. For a region-of-interest (ROI) analysis, we derived DLPFC ROIs from FreeSurfer cortical parcellations based on each participants individual cortical anatomy as described previously [14, 57]. We obtained indices of activation for the left and right DLPFC ROI using the COPE obtained from the second-level fixed-effects analysis for each participant and calculated mean DLPFC activation in percent signal change. For additional details about data acquisition, quality control and preprocessing, see Walton et al. [58], SM 1.3 and 1.4.

2.4 DNA methylation data: Sample processing and quality control

Blood samples were obtained from 234 participants and sent to the Harvard Partners Center for Genetics and Genomics for DNA extraction. All DNA extraction, bisulphite modification and hybridization steps were done blinded to group assignment, the latter two were performed at the Mind Research Network Neurogenetics Core Lab on an Infinium HumanMethylation27 BeadChip using Illumina Infinium Methylation Assay. Built-in controls were used to evaluate the quality of individual arrays.

Samples and CpG sites were filtered according to coverage (see SM 1.5). After excluding one intensity outlier identified by a quantile color bias adjustment of the red and green fluorescence channel intensities, data were normalized using the R package “wateRmelon” of the Bioconductor project [59]. Finally, we removed five samples displaying sex mismatch resulting in a dataset of 229 participants. After excluding participants, who had no imaging data or imaging/behavioral/genetic data of insufficient quality (see above), genome-wide methylation beta-values were extracted from a final dataset of 103 schizophrenia patients and 111 healthy controls. For more information on DNA methylation preprocessing and intensity data extraction, see SM 1.5.

We corrected our data statistically for the effects of chip and bisulphite sequencing (BS) conversion efficiency by first estimating coefficients in a linear regression model and then computing residuals (corrected beta-values). In a similar way, we removed the effects of sex and age [60–62] as well as population stratification as follows: DNA methylation has been shown to be moderately influenced by population structure, although to a much lesser degree than genetic data [63]. By means of principal and independent component analyses, Liu et al. [63] found that the population structure present in DNA methylation data is very well captured by the first genetic principal component. Hence, to avoid confounding effects due to population stratification, we applied principal component analysis (PCA) to our genotype data (Illumina HumanOmni-Quad BeadChip) using EIGENSTRAT of the EIGENSOFT 3.0 software package [64, 65], extracted ten principal components (PC) and then included the first component in our regression models. For more details, see [15].

Since DNA methylation is known to be tissue-specific [66–68], the question arises whether we can draw conclusions on brain-based intermediate phenotypes based on methylation values measured in peripheral blood cells. Walton et al. (unpublished results) established a list of CpG sites with well correlated DNA methylation beta-values by comparing brain and blood samples from patients with treatment-resistant epilepsy in a within-subject design. Accordingly, we excluded all CpG sites with a low correlation (Spearman’s rank correlation coefficient < 0.5) between brain and blood tissue-specific methylation beta-values. The final methylation data set comprised 214 subjects and 3,086 CpG sites.

2.5 Statistics

Basic demographic characteristics were compared across diagnostic groups and acquisition site-specific scanner field strengths using Welch’s t tests. Chi-square statistics were used to examine differences in categorical variables. Alpha was set to 0.05 for all analyses. Sample characteristic analyses were carried out with IBM SPSS Statistics 21.

Each of the three intermediate phenotypes (hippocampal volume, STG thickness, and WM load-dependent neural activity (% BOLD signal change) in the DLPFC) was controlled for the effects of age, sex and site-specific scanner field strength in a linear regression model. Standardized residuals of each phenotype were used in all (gene set enrichment) analyses. For hippocampal volume, we also controlled for intracranial volume[69].

We used GSEA [48] to identify predicted miRNA target gene sets whose methylation patterns correlated with the intermediate phenotypes (Figure 1): First, each gene’s methylation pattern was correlated with the respective phenotype across all patients and controls. If there was more than one CpG site per gene, i.e. more than one beta-value associated with a gene name, median methylation was used (GSEA standard procedure). Correlation was determined using the Pearson product-moment correlation coefficient (GSEA standard procedure for continuous phenotypes). Next, genes were ranked based on those correlations and tested if the ones targeted by a given miRNA had methylation-phenotype correlations larger than expected by chance. For this, predicted target gene sets of 221 miRNAs were obtained from the Molecular Signatures Database (http://www.broadinstitute.org/gsea/msigdb/) by selecting genes containing the respective 3′-UTR miRNA binding motifs, and then filtered for genes with available methylation data. We adopted the current annotation and terminology of miRNAs according to miRBase v21.

Figure 1. Workflow of statistical analysis.

GSEA (gene set enrichment analysis) procedure. ES = enrichment score.

There are three key elements of the GSEA method: (1) Calculation of an Enrichment Score (ES), that reflects the degree to which a predicted miRNA target gene set is overrepresented at the top or bottom of the list of genes ranked according to the positive or negative correlation between methylation and phenotype; (2) Estimation of Significance Level of ES using an empirical phenotype-based permutation test procedure; (3) Adjustment for Multiple Hypothesis Testing, including normalization of the ES to account for different gene set sizes and calculation of the false discovery rate (FDR) to control the proportion of false positives. For further details about the GSEA method, please see Subramanian et al. [48]. Since we applied the gene set enrichment analysis as an exploratory approach, we used an initial FDR cutoff of 0.25 to determine significance of our results (GSEA default setting for permutation), which is consistent with many previous analyses using GSEA [70–73]. Using a more stringent FDR cutoff may substantially increase the chances to overlook potentially significant results [74]. To control for the number of tested intermediate phenotypes (n = 3), we lowered the FDR threshold to 0.083 (= 0.25/3).

3 Results

3.1 Sample characteristics

Patients and controls did not significantly differ in demographic variables such as age, sex, parental socio-economic status and Annett handedness score. Schizophrenia patients had a significantly lower WRAT3-RT score and showed significantly lower mean hippocampal volume and mean STG thickness. We found no significant difference in DLPFC activity during SIRP, but schizophrenia patients performed significantly lower in the WM paradigm (Table 1). For site-specific demographics, clinical variables and intermediate phenotypes see SM Table 1.

Table 1.

Demographic data, clinical variables and intermediate phenotypes

Sample	Size	Sex (female)		Age	WRAT3-RT	Parental SES	Handedness	Medication (cum. dose years)	Hippocampal volume	STG thickness	DLPFC activity	SIRP performance
	N	N	%	Mean (SD)	Mean (SD)	Mean (SD)	Mean (SD)		Mean (SD)	Mean (SD)	Mean (SD)	Mean (SD)
SZ	103	28	27.2	34.55 (10.60)	46.25 (7.01)*	2.87 (1.01)	1.51 (3.52)	57.30 (123.07)	4139.64 (471.04)*	2.79 (0.16)*	0.05 (0.10)	95.21 (4.99)*
HC	111	40	36.0	32.05 (10.83)	51.03 (3.97)	2.68 (0.72)	0.79 (2.38)	-	4439.49 (427.69)	2.90 (0.16)	0.06 (0.10)	98.46 (1.62)
Total	214	68	31.8	33.25 (10.77)	48.78 (6.09)	2.77 (0.87)	1.14 (2.99)	-	4295.17 (472.51)	2.85 (0.17)	0.06 (0.10)	96.90 (3.99)

Means and standard deviations (SD) are given. SZ = patient with schizophrenia; HC = healthy control. WRAT3-RT = Wide Range Achievement Test 3 - Reading Test. Parental SES (socioeconomic status) was classified according to Hollingshead, and handedness determined using the Annett Scale of Hand Preference. DLPFC activity = mean working memory load-dependent neural activity (% BOLD signal change) in the DLPFC during the Sternberg Item Recognition Paradigm (SIRP). SIRP performance represents the mean accuracy during the working memory task across all loads (1, 3, 5).

^*significantly different between SZ and HC on basis of Welch’s t test (p<0.05). For DLPFC activity and SIRP performance, data were only available for 203 and 209 participants, respectively.

3.2 Gene set enrichment analysis of DNA methylation

We correlated each gene’s methylation pattern with the phenotype variation across the study population and subsequently identified miRNAs whose predicted target genes are more correlated than expected by chance (see Materials and Methods). This approach assumes that target genes of a miRNA are functionally linked and thus more likely to be methylated in concert than random sets of genes.

For hippocampal volume, we found three predicted miRNA target gene sets significantly enriched (FDR < 25%) based on a positive correlation between methylation data and the phenotype (see Table 2). Hsa-miR-219a-5p² was the target gene set with the highest normalized ES (NES = 1.9368). The associated FDR q-value was 0.0498. Given this q-value, the hsa-miR-219a-5p gene set was the only target gene set passing multiple testing correction. See Figure 2 for a graphical view of the ES (with detailed information on how to interpret enrichment plots in SM 2.1) and refer to Table 3 for a full list of genes included in the predicted hsa-miR-219a-5p target gene set.

Table 2.

MicroRNA target gene sets significantly enriched for hippocampal volume (positively ranked list metric; Pearson correlation)

Gene Set	Size	ES	NES	NOM p-val	FDR
GACAATC,HSA-MIR-219A-5P	22	0.60	1.94	0.002	0.050
GGCCAGT,MIR-193A,MIR-193B	8	0.73	1.72	0.004	0.246
TCCAGAT,MIR-516-5P	12	0.63	1.72	0.014	0.194

Names of gene sets as given by the Molecular Signatures Database of GSEA. Size = number of genes included in each gene set (restricted by the number of CpG sites and associated genes included in the MCIC data); ES = enrichment score; NES = normalized enrichment score (adjusted for different sizes of gene sets); NOM p-val = nominal p-value; FDR = false discovery rate q-value. FDR threshold = 0.083 (= 0.25/3;corrected for number of tested intermediate phenotypes).

Figure 2. Enrichment plot: Graphical view of the enrichment score (ES) for the predicted hsa-miR-219a-5p target gene set.

Profile of the running ES and positions of gene set members on the rank ordered list (according to Pearson correlation with hippocampal volume). See SM 2.1 and GSEA documentation (http://www.broadinstitute.org/gsea/doc/GSEAUserGuideFrame.html) for detailed information on how to interpret enrichment plots.

Table 3.

Full list of genes included in the hsa-miR-219a-5p gene set (as analyzed by GSEA)

Gene Name	Location	Rank in Gene List	Rank Metric Score	Running ES	HGNC Gene Name /Aliases	Experimental Evidence
KIAA0182	16q24.1	31	0.148	0.108	GSE1; Gse1 coiled-coil protein	predicted
PKNOX1	21q22.3	48	0.140	0.215	PBX/knotted 1 homeobox 1	predicted
TRAF7	16p13.3	55	0.137	0.323	TNF receptor-associated factor 7, E3 ubiquitin protein ligase	predicted
CBFA2T3	16q24	98	0.118	0.403	core-binding factor, runt domain, alpha subunit 2; translocated to, 3	predicted
EPHA4	2q36.3	179	0.102	0.456	EPH receptor A4	predicted
MKNK2	19p13.3	373	0.076	0.448	MAP kinase interacting serine/threonine kinase 2	starBase CLIP-seq
DDAH1	1p22	413	0.072	0.492	dimethylargininedimethylaminohydrolase 1	starBase CLIP-seq
ESR1	6q24-q27	416	0.072	0.549	estrogen receptor 1	Luciferase Reporter Assay [105]
AKAP13	15q24-q25	426	0.071	0.603	A kinase (PRKA) anchor protein 13	starBase CLIP-seq
PHACTR2	6q24.1	788	0.039	0.505	phosphatase and actin regulator 2	starBase CLIP-seq
SLC31A1	9q32	1028	0.022	0.438	solute carrier family 31 (copper transporter), member 1	starBase CLIP-seq
KBTBD8	3p14	1086	0.019	0.433	kelch repeat and BTB (POZ) domain containing 8	starBase CLIP-seq
ADD2	2p13.3	1148	0.016	0.423	adducin 2 (beta)	predicted
ERGIC1	5q35.1	1223	0.011	0.406	endoplasmic reticulum-golgi intermediate compartment (ERGIC) 1	starBase CLIP-seq
RBM24	6p22.3	1260	0.009	0.401	RNA binding motif protein 24	predicted
KIAA0240	6p21.1	1318	0.005	0.384	GLTSCR1L; GLTSCR1-like	predicted
NEK6	9q33.3-q34.11	1451	−0.003	0.339	NIMA-related kinase 6	starBase CLIP-seq
THRB	3p24.2	1551	−0.009	0.311	thyroid hormone receptor, beta	starBase CLIP-seq
SCN5A	3p21	1646	−0.015	0.290	sodium channel, voltage-gated, type V, alpha subunit	predicted
TGFBR2	3p22	1907	−0.032	0.222	transforming growth factor, beta receptor II (70/80kDa)	starBase CLIP-seq
FZD4	11q14-q21	2324	−0.063	0.124	frizzled class receptor 4	predicted
SEMA4G	10q24.31	2354	−0.066	0.167	sema domain, immunoglobulin domain (Ig), transmembrane domain (TM) and short cytoplasmic domain, semaphoring) 4G	predicted

Genes contributing to the core enrichment of the predicted hsa-miR-219a-5p gene set (leading edge subset) are bold. HGNC = HUGO (HUman Genome Organization) Gene Nomenclature Committee. ES = enrichment score. The last column indicates whether target genes were predicted based on bioinformatic algorithms (labelled as “predicted”) or are experimentally validated, e.g. by the starBase [126, 127] prediction intersection tool using CLIP-Seq (cross-linking immunoprecipitation-high-throughput sequencing) data.

For STG thickness, the predicted miR-513 target gene set was significantly enriched (NES = 1.7636) with a FDR q-value of 0.2404, again based on a positive correlation between methylation data and phenotype. None of the predicted miRNA target gene sets was significantly enriched when analyzing the methylation data in association with working-memory load-dependent neural activity in the DLPFC.

3.3 Additional analyses

We carried out separate gene set enrichment analyses for schizophrenia patients and healthy controls to investigate potential subgroup effects. Again, the predicted hsa-miR-219a-5p target gene set was enriched for hippocampal volume in schizophrenia patients as well as in healthy controls (nominal p-value of 0.016 and 0.038, respectively; based on a positively ranked list metric).

In a subsequent model, we also adjusted for lifetime exposure to antipsychotic drugs in schizophrenia patients (by including this variable as an additional covariate in the phenotype regression analysis), and again found the predicted hsa-miR-219a-5p target gene set enriched with a nominal p-value of 0.016.

Since our main finding is related to hippocampal volume, we aimed to further explore this association by investigating hippocampus-related neuropsychological variables as additional phenotypes. In particular, we tested for a methylation-diagnosis interaction effect (using ANOVA) on indices of episodic memory (Logical Memory II subtest of the WMS-III [75]) and manipulation of material in working memory (Letter Number Sequencing [LNS] subtest of the WAIS-III [76]), both of which have been found to be important to memory formation and medial temporal lobe function [77]. We found a significant interaction effect of EPHA4 methylation (cg15808558) and diagnosis on LNS (F(1) = 7.120, p = 0.008). Similarly, we found an interaction effect of trend level significance for PKNOX1 methylation (cg14204791) and diagnosis on LNS (F(1) = 3.799, p = 0.053).

Ingenuity pathway analysis (IPA, Ingenuity^® Systems, CA, USA, www.ingenuity.com; see SM 1.6 for method details) of predicted hsa-miR-219a-5p target genes that were most highly correlated with hippocampal volume (Table 3), revealed a functional role of these genes in cellular development (p = 2.19 × 10⁻⁴), gene expression (p = 2.38 × 10⁻⁴) and cell-to-cell signaling and interaction (p =4.70 × 10⁻⁴), as well as associated network functions in behavior and developmental disorders (Ingenuity network score = 21; according to the fit of the network to the hsa-miR-219a-5p leading edge genes) [78]. Including all predicted hsa-miR-219a-5p target genes of Table 3, there is supporting evidence for an association with axonal guidance signaling pathways (p = 0.0128), and again associated network functions in behavior (Ingenuity network score = 24) as well as cell-to-cell signaling and interaction (Ingenuity network score = 16).

4 Discussion

Results of this study, which employed a GSEA approach to methylation data in schizophrenia for the first time, suggests that the methylation state of hsa-miR-219a-5p targets correlates with hippocampal volume an intermediate phenotype for schizophrenia. By focusing on CpG sites with highly correlated blood and brain tissue-specific methylation beta-values, we were able to ensure that the identified potential targets impact brain-related biological processes, which was confirmed by the IPA pathway analysis.

As the first identified functional mature product of the miR-219 precursor family, hsa-miR-219a-5p was originally referred to as hsa-miR-219. Hsa-miR-219a-5p is coded by two different stem loop sequences, hsa-miR-219a-1 and hsa-miR-219a-2, in two locations, 6p21 (which is a putative susceptibility locus for schizophrenia [79]) and 9q34³, respectively (see SM Table 2 for details). This may add to the complexity of regulation of hsa-miR-219a-5p expression and subsequently its target genes. Identified as miR-219 in previous studies, hsa-miR-219a-5p has also been implicated in circadian clock regulation [81], oligodendrocyte differentiation and myelination [82] as well as disruption of NMDA receptor signaling [83] all processes that have been associated with schizophrenia: A disruption of circadian rhythm, in particular sleep-onset and maintenance insomnia, is a common phenomenon in schizophrenia [84]. In line with Zhao et al. [82], a large number of diffusion tensor imaging studies provide evidence for reduced white matter integrity in schizophrenia [85–90] and there is ample evidence for NMDA receptor hypofunction in schizophrenia [91, 92]. Furthermore, hsa-miR-219-5p was found to be the most highly enriched miRNA in synaptic fractions (5-fold), while at the same time it is the most strongly down-regulated miRNA in schizophrenia synaptosomes (70% decrease) [46]. Taken together, these findings and our own data suggest a major role of the miR-219 precursor family miRNAs, and especially hsa-miR-219a-5p, in brain development, abnormal behavior and schizophrenia in particular.

One of hsa-miR-219a-5p’s predicted target genes that was identified as a leading edge gene (i.e. genes that contribute most to the ES) in our gene set enrichment analysis, is the ephrin type-A receptor 4 (EPHA4) gene, which belongs to the ephrin receptor subfamily of the protein-tyrosine kinase family and is expressed predominantly in the hippocampus [93, 94]. EPH and EPH-related receptors have been implicated in mediating developmental events, particularly in the nervous system, including axon guidance [95, 96]. EPH signaling was found to be crucial for hippocampal circuit plasticity [97] and to be associated with synaptic long-term potentiation [98]. In addition, we found a significant interaction effect between EPHA4 methylation (cg15808558) and diagnosis on memory performance in our sample, which suggests a potential influence of an altered hsa-miR-219a-5p target methylation also at the behavioral level. A reduced performance on memory tests is one of the most stable cognitive findings in schizophrenia [99, 100].

Another predicted target gene of hsa-miR-219a-5p is PKNOX1 (PBX/knotted 1 homeobox 1) and its overexpression in the hippocampus was associated with facilitated learning and memory in a mouse model for the trisomy of the Abcg1-U2af1 region [101]. Again, we found an interaction effect between PKNOX1 methylation (cg14204791) and diagnosis on memory performance implying that an altered PKNOX1 expression in schizophrenia patients or individuals at risk (i.e. related to an aberrant methylation pattern) may further add to the cognitive deficits which are hallmarks of schizophrenia.

The estrogen receptor 1, encoded by the ESR1 gene and another predicted hsa-miR-219a-5p target, is a ligand-activated transcription factor [102] regulating cellular proliferation and differentiation in target tissues [103, 104]. Only recently, ESR1 has been experimentally validated to be a target gene of hsa-mir-219-5p [105]. Although not explicitly in hippocampus, Knickmeyer et al. [106] observed smaller volumes for an ESR1 variant in multiple regions of frontal and parietal lobes. Weickert et al. [107] found several variants of the ESR1 gene and ESR1 expression being associated with schizophrenia (although a negative genetic study does exist as well [108]). Additional studies conducted by the same group also found reduced ESR1 mRNA levels in cortical and subcortical brain regions in patients suffering from major mental illness including a reduction in hippocampal ESR1 mRNA in patients with schizophrenia [109, 110]. Estrogen, which binds to estrogen receptor 1 encoded by the ESR1 gene, modulates human emotion [111] and frontal cortical activity during cognitive tasks [112, 113], both of which are altered in schizophrenia and other mental disorders [107]. Furthermore, sex differences have been described for nearly all features of schizophrenia [114]. Clinical studies suggest that estrogens exert psychoprotective effects in schizophrenia patients [115] which may be mediated by an interaction effect between estrogen and brain-derived neurotrophic factor (BDNF) [114].

To further identify possible functional mechanisms, we performed IPA pathway analysis including all predicted hsa-miR-219a-5p target genes of Table 3, and found several of those potential target genes to be involved in axonal guidance signaling, a process known to be involved in neuronal development and a prerequisite for normal cognitive functioning which has been repeatedly shown to be disrupted in schizophrenia [116, 117]. Furthermore, IPA pathway analysis revealed predicted hsa-miR-219a-5p target genes to be associated with cellular development, gene expression and cell-to-cell signaling and cell-cell interaction, all of which are of substantial interest in the etiology of psychiatric disorders. However, the exact mechanisms of how the expression of schizophrenia-associated risk genes may be concurrently regulated through miRNA activity and DNA methylation in terms of possible synergistic effects remain to be clarified and could be studied in vitro or using animal models.

The findings of our study have to be considered in the light of the following limitations: First, complete epigenetic and phenotypic data was available for only 214 participants after quality assurance steps, lowering the statistical power of our analyses. Our findings warrant replication using larger samples. In addition, we analyzed a limited number of CpG sites following the approach of Walton et al. (unpublished results). Although this approach may lead to false negative results, it avoids inclusion of CpG sites with methylation levels that cannot be predicted from blood samples. Second, antipsychotic medication has been shown to affect DNA methylation [118]. We accounted for the effect of medication on DNA methylation in the sample of schizophrenia patients by using cumulative lifetime exposure to antipsychotic drugs as an additional covariate (see 3.3. Additional analyses). Results of this sub-analysis strengthen our main finding. Third, the motif gene sets of the Molecular Signature Database used in our GSEA procedure, in particular the miRNA target gene sets, contain bioinformatically predicted target genes (based on 3′-UTR miRNA binding motifs) that are not experimentally validated for the most part (see also Table 3). Therefore, some of the miRNA target genes which we have scrutinized here may in fact not be regulated by the predicted miRNAs [119]. Fourthly, gene set enrichment analysis as carried out by GSEA ignores gene-gene correlation, which may lead to an overestimation of significance levels and eventually to false positives. Finally, Geeleher et al. [120] stated, that gene set analysis could be severely biased when applied to genome-wide methylation data because lengths and expression levels of genes frequently differ between gene sets. To overcome this potential bias, GSEA applies sample label permutations which was shown to be robust to the differences in the number of probes per gene [121, 122]. In summary, we suggest GSEA as a very useful method for hypothesis generation as it is robust to a low signal-to-noise ratio and our subsequent statistical analyses substantiated the GSEA main findings.

5 Conclusions

We found hsa-miR-219a-5p target methylation to be associated with hippocampal volume – a schizophrenia-related intermediate phenotype. The (dys)regulation of miRNAs and miRNA targets by epigenetic mechanisms has been found to interfere with physiological brain functions [123–125] and might convey additional risk for structural brain changes and, possibly, psychiatric symptoms. However, further study is needed to deepen our understanding of the complex genetic and epigenetic architecture underlying neuropsychiatric disorders. An integrated analysis of miRNA and epigenetic control could highlight genes that are regulated by DNA methylation and are targeted by miRNAs with a potential use as clinical markers, eventually leading to the identification of novel targets for therapeutic approaches.

Highlights.

• We used gene set enrichment analysis for DNA methylation data in schizophrenia.

• Predicted target genes of microRNA miR-219 were associated with hippocampus volume.

• This association was independent of disease status.

• Epigenetic (dys)regulation of miR targets may confer risk for psychiatric disorders.