Cross-Tissue Exploration of Genetic and Epigenetic Effects on Brain Gray Matter in Schizophrenia

Abstract

Closely linking genetics and environment factors, epigenetics has been of increasing interest in psychiatric disease studies. In this work, we integrated single nucleotide polymorphisms (SNPs), DNA methylation of blood and saliva, and brain gray matter (GM) measures to explore the role of genetic and epigenetic variation to the brain structure changes in schizophrenia (SZ). By focusing on the reported SZ genetic risk regions, we applied a multi-stage multivariate analysis to a discovery dataset (92 SZ patients and 110 controls, blood) and an independent replication dataset (93 SZ patients and 99 controls, saliva). Two pairs of SNP-methylation components were significantly correlated (r = .48 and .35) in blood DNA, and replicated (r = .46 and .29) in saliva DNA, reflecting cross-tissue SNP cis-effects. In the discovery data, both SNP-related methylation components were also associated with one GM component primarily located in cerebellum, caudate, and thalamus. Additionally, another methylation component in NOSIP gene with significant SZ patient differences (P = .009), was associated with 8 GM components (7 with patient differences) including superior, middle, and inferior frontal gyri, superior, middle, and inferior temporal gyri, cerebellum, insula, cuneus, and lingual gyrus. Of these, 5 methylation-GM associations were replicated (P < .05). In contrast, no pairwise significant associations were observed between SNP and GM components. This study strongly supports that compared to genetic variation, epigenetics show broader and more significant associations with brain structure as well as diagnosis, which can be cross-tissue, and the potential in explaining the mechanism of genetic risks in SZ.

brain gray matter, cross-tissue, DNA methylation, genetics, meQTL, schizophrenia

Introduction

Schizophrenia (SZ) is a highly heritable complex disease with a lifetime prevalence of around 1%.^1–3 Recent landmark work from the Psychiatric Genomics Consortium (PGC) has identified 128 genome-wide significant SZ-risk SNPs⁴ from 108 highly linkage genomic regions (R² > .6), providing reliable guidance for investigating the etiology and pathophysiology of SZ.

To uncover the biological mechanisms of such SZ-risk variants, a valuable strategy is to explore the genetic influence on brain-based phenotypes through magnetic resonance imaging (MRI) techniques.⁵ Imaging-derived structural or functional measures of the brain (eg, gray matter [GM] density) have been shown to be reliable and highly heritable.^6–8 Considerable success in the imaging genetics field has enriched our understanding of the putative neurobiological pathways from genetic mutations to system level and behavioral changes resulting in psychiatric disorders.^9–12 However, a recent study of genetic influences on SZ and subcortical brain volumes found no clear evidence of shared genetic variants.¹³ Although this study is limited to the selected subcortical regions, it suggests that more complex biological pathways may involve bridging the gap from DNA variation to brain system changes in SZ.

Epigenetics, beyond genetic sequence variation, regulates gene expression and thus influences biological functions.^14,15 Epigenetics can be modulated by genetics,¹⁶ potentially mediating genetic risks,¹⁷ and also integrate environmental factors in the etiopathogenesis of SZ.^18–20 DNA methylation, as one of the main epigenetic modifications, has shown to be related to SZ symptoms,²¹ disease onset,²² cognitive deficit,²³ and be significantly changed in schizophrenic brain^24,25 and peripheral tissues.²⁶ Although methylation tends to be tissue specific in general, recent studies have noted a set of methylation sites significantly correlated between blood and brain.^27–29 This cross-tissue mechanism is hypothesized to be through genetics, called methylation quantitative trait loci (meQTL) effect.^30,31 In parallel, accumulating evidence has shown the relationship of peripheral DNA methylation with brain structure and function.^23,32–35 These findings support the potential of peripheral methylation in association with brain structure or function to further understand psychiatric disorders.

Two recent studies have demonstrated the significant role of meQTL in brain development and SZ.^31,36 However, the underlying mechanisms of how epigenetic factors interplay with genetic factors, and further influence brain structural variation and psychiatric disorders are far from clear. In this study, we explored the interconnections between the 3 modalities—single nucleotide polymorphisms (SNPs), DNA methylation from blood and saliva, and whole brain GM—from SZ patients and healthy controls (HCs) in 2 independent datasets. We focused on SNPs and methylation sites within the 108 SZ-risk regions identified by PGC.⁴ Given the dependence of loci in both genotype and methylation, a multivariate analysis based on parallel independent component analysis (ICA),³⁷ which has been suggested to identify clusters of correlated loci in genetic applications,^38,39 was performed as shown in figure 1 to capture the intrinsic genetic–epigenetic relationships. Brain networks were extracted by ICA using GM images from structural MRI.⁴⁰ Finally, we tested the relationships among components of three modalities, as well as their associations with clinical diagnosis, symptoms, and cognition.

Materials and Methods

More details of methods and materials were provided in the supplementary Materials and methods section.

Participants

The discovery dataset consisted of 202 subjects recruited by the Mind Clinical Imaging Consortium study.⁴¹ Patients met criteria for DSM-IV-TR schizophrenia, schizoaffective, or schizophreniform disorders. HCs had no current or past history of psychiatric illness including substance abuse or dependence. Following a similar recruitment protocol, the replication data were collected from the Center for Biomedical Research Excellence study⁴² and Glutamate and Outcome in Schizophrenia study.⁴³ All participants of these studies as shown in table 1 provided written consent forms.

Table 1.

Demographic of Participants in the Both Discovery and Replication Data

Variables	Discovery data (n = 202)		Replication data (n = 192)
	SZ (schizoaffective/ schizophreniform)	HC	SZ (schizoaffective/ schizophreniform)	HC
Diagnosis	92 (1/4)	110	93 (11/0)	99
Gendera (women, %)	26.09	37.27	14.13	28.28
Age, y (SD)	34.14 (10.23)	32.66 (10.91)	37.96 (14.33)	37.09(12.16)
Racea (White/Black/Asian/Other)	70/15/3/4	99/3/4/4	42/7/0/44	51/6/0/42
Alcohol depend/abuse	27	0	0	0

^aIndicates significant differences between SZ and HC. The differences were controlled or corrected in the analyses.

In the discovery data, symptoms were assessed with the Scale of the Assessment of Positive Symptoms⁴⁴ and the Scale of the Assessment of Negative Symptoms.⁴⁵ Alcohol abuse/dependence status was assessed by DSM-IV and 101 participants provided reliably nicotine use information measured by pack per year.⁴⁶ In the replication data, participants had no current diagnosis of alcohol abuse or dependence, and nicotine use was assessed by Fagerstrom Test for Nicotine Dependence.⁴⁷ For the cognitive assessment, we focused on 2 comparable tests between studies: (1) letter-number sequencing test of the WAIS-III and MATRICS for measuring working memory,^48,49 and (2) computerized version of the Tower of London (TOL) test⁵⁰ and the NAB maze test⁵¹ for measuring planning and problem solving.

Data Preprocessing

DNA was extracted from whole blood in the discovery data. Genotyping was performed using the Illumina Infinium HumanOmni1-Quad assay. Data was quality controlled including checking for missing data, sex mismatch, relatedness, Hardy–Weinberg equilibrium (<1 × 10⁻⁶), minor allele frequency (>0.01) and population structure correction. DNA methylation was assayed by the Illumina Infinium HumanMethylation27 assay. Quality control included removal of bad samples and probes, batch correction, and removal of nonvariant sites. Age and gender were further adjusted by linear regression model. Finally, we narrowed down our analyses to the 108 SZ-susceptible regions,⁴ resulting in 4475 SNPs and 102 cytosine-phosphate-guanine (CpG) sites.

Saliva samples were used for DNA extraction in the replication data. Genotyping was performed using both Illumina Infinium HumanOmni1-Quad and Omni5 assays. The SNP data was quality controlled by following the same procedure as that in MCIC data. Missing genotypes were imputed with the 1000 genomes reference panel. DNA methylation was assayed by the Infinium MethylationEPIC beadchip. Quality control was performed using package “minfi.”⁵² Batch, age, and gender were adjusted on each site prior to analysis. Cell type proportions in blood and saliva samples were estimated using the algorithm by Houseman et.al.⁵³

T1-weighted MRI images in both datasets were preprocessed using SPM12 (http://www.fil.ion.ucl.ac.uk/spm/software/spm12/). Unmodulated GM images were smoothed by a full width half maximum Gaussian kernel of 8 mm, followed by the correction for scanning sites (for the discovery data), age and gender on each voxel as suggested in previous studies.^54,55 No significant differences on GM networks were observed when using modulated images (data not shown).

Primary Analysis

A multilevel association analysis was applied as shown in figure 1. Parallel ICA was first applied to SNP and DNA methylation data to derive independent genetic and methylation components, while maximizing inter-modality correlations of the components’ loadings. Independent components of each modality were linear combinations of individual variants, capturing interactions among SNPs or CpG sites. A loading vector presented the expression of each component across subjects. Ten-fold subsample and permutation tests were conducted to verify the stability and significance of the SNP-methylation component links. For the linked SNP-methylation components, meQTL analysis was applied to the contributing SNPs and CpG sites to verify SNP effect on methylation at individual site level. In parallel, independent GM components were derived by a regular ICA on whole brain GM images. Then, the SNP and methylation components were tested for associations with GM components using a linear regression model. For any 3 modality links, a mediation model was further tested for the methylation component mediating the genetic component’s effect on the GM component. In addition, all SNP, methylation, and GM components were tested for the associations with diagnosis, symptoms, and cognition to help elucidate their relation to SZ disease. Medication, cell type proportion, alcohol abuse, and smoking were also assessed with their influences on the results as additional covariates in regression model. All the tests were corrected for multiple comparison using the false discovery rate (FDR ≤ 0.05).⁵⁶ For replication, we extracted the same SNP, methylation and GM components identified in the discovery data using a projection method and assessed their associations. The significance P <.05 and the same direction of correlation were set for replication.

Fig. 1.

The analysis flow of the study.

Results

SNP, DNA Methylation Components and Intermodality Relationship

In the discovery data, 15 SNP components and 10 methylation components were extracted. Among them, 3 methylation components, but no SNP component showed significant disease association. After correction for multiple comparisons, 2 pairs of SNP-methylation components were significantly correlated with r = .48 (P = 5.46 × 10⁻¹¹) and 0.35 (P = 3.28 × 10⁻⁵). The correlations were consistent in 10-fold stability tests with the average correlations of 0.50 ± 0.04 and 0.25 ± 0.11, respectively. No significant effects of race and alcohol abuse were observed on the SNP-methylation associations. Permutation tests showed the significance of the two links with empirical P-values of 1 × 10⁻³ and 0.01.

Figures 2B and 2E shows the loadings of the 2 pairs of SNP and methylation components across subjects. Both pairs indicated increased SNP loading along with the higher methylation loading. A higher loading indicates more presentation of positively weighted SNPs, and hyper-methylation of positively weighted CpG sites. The HC and SZ groups exhibited comparable correlations for both pairs (1st pair: 0.49 and 0.46; 2nd pair: 0.39 and 0.23), respectively. Excluding the schizoaffective and schizophreniform patients did not change the correlation by much. Marginal group differences were found in components of the second pair (P_snp = .09, P_methylation = .04, uncorrected). After correcting for clinical diagnosis, the SNP-methylation associations remained significant for both pairs (P = 4.41 × 10⁻¹⁰ and 2.36 × 10⁻⁴). By thresholding the contribution weights of loci in the components with |z-score| >2.5, there were 94 significant SNPs from 20 genes or their nearby regions in the first linked SNP component (supplementary table S1), and 5 CpG sites in the genes NT5C2, CSMD1, FES, and ARL3 in the corresponding methylation component. In the second pair, 60 highly weighted SNPs from 19 genes and one CpG site in the gene ITIH4 were involved in the correlation (supplementary table S2).

Fig. 2.

(A) and (C) show meQTL mapping of the contributing SNPs and methylation sites (|z| ≥ 2.5) in both pairs of significantly linked SNP and DNA methylation components. (B) and (E) plot their loadings with correlations r = .48 and .35, shown separately by HC and SZ groups. (C) and (F) give examples of the strongest meQTL in each pair (located on chr.10 around gene NT5C2 and chr.3 around gene ITIH4, respectively).

meQTL analysis on the contributing SNPs and CpG sites found 69 meQTLs in the first pair and 59 meQTLs in the second pair after FDR correction, mostly showing cis effects (SNP effect of CpG site within 500 kbp distance³¹). In the first pair cis-acting meQTLs were located on chromosome 10 around 3 CpG sites in NT5C2 and ARL3 (figure 2A), where the strongest cis-effect resides in SNP rs1926030 and CpG cg00035347 with partial η² (% of variance explained) = 50.3%, P = 3.6 × 10⁻³² (figure 2C), indicating that an increasing minor allele A count in rs1926030 was associated with hypomethylation at cg00035347 located in the promoter of NT5C2. The cis-effecting meQTLs in the second pair was found in chromosome 3 around ITHI4 (figure 2D). The strongest meQTL effect was between SNP rs2071041 and CpG cg17890764 with partial η² = 13.9%, P = 4.5 × 10⁻⁸ (figure 2F). Other meQTL relations were listed in supplementary tables S3 and S4. Overall, the cis-effecting meQTLs strongly affected the methylation change (partial η² median value: 10.88%, interquartile change IQR: 6.95%–21.24%).

We further compared our meQTLs with 2 recent meQTL studies in fetal³¹ and adult brain tissues.³⁶ Among 69 meQTLs identified in the first SNP-methylation pair, 14 were also separately identified in both brain tissues with 12 in common. For the 59 meQTLs in the second pair, 13 were found in the fetal brain tissue and 11 were in adult brain tissue with 7 in common.

In the replication data 72.21% of SNPs and 52% of methylation sites used in the discovery data were present, including over 96% of contributing loci and CpG sites (|z| > 2.5) in the identified significant SNP-methylation pairs. The replicated correlations for the 2 linked SNP-methylation pairs were r = .46, P = 1.59 × 10⁻¹¹ and r = .29, P = 5.9 × 10⁻⁵, respectively, and the correlations (r = .53, P = 4.93 × 10⁻⁸; r = .26, P = 1.08 × 10⁻²) were consistent in SZ group only (no significant difference after excluding schizoaffective patients). In the first SNP-methylation pair, 72 meQTLs were identified including 67 (97.1%) found in the discovery set. Similarly, for the second SNP-methylation pair, 57 meQTLs were identified and replicated 98.28% meQTLs in the discovery data.

GM Components and Their Associations With DNA Methylation

GM data were decomposed into 15 independent components by ICA, of which 10 components showed significant associations with diagnosis (supplementary figure S2). Nine components (figure 3) were significantly associated with 3 methylation components, as listed in table 2. Interestingly, 7 out of 9 GM components showed significant group differences, and remained significant after controlling for alcohol abuse except component 7 with P = .06. It was worth noting that in the GM components 1 and 2 with the largest group differences, SZ patients showed substantial GM loss in the areas of superior/middle/medial/inferior frontal gyri, and superior/middle/inferior temporal gyri.

Fig. 3.

Brain mapping of 9 GM components significantly associated with 3 methylation components. (A) Seven GM components have significant group differences while the other 2 in (B) show no significant differences.

Table 2.

The Relationship Between GM and DNA Methylation Components in the Discovery Data

Methylation components		Associated gray matter imaging components
No.	Methylation- diagnosis association, Cohen’s d (P, direction); significant genes	No.	Methylation-GM association, R2 (T-score, P)	GM-diagnosis association, Cohen’s d, (P, direction)	Brain regions (|z| > 2.5)	L/R volume (mm3)
1	−0.348 (.009, HC > SZ); NOSIP	1	0.1 (4.77, 9.1 × 10−5)	−0.863 (7.7 × 10−8, HC > SZ)	Positive: Superior/middle/medial/inferior frontal gyri	8.6/11.8
					Negative: Precentral/postcentral gyri	3.5/0.8
		2	0.11 (5.02, 3.8 × 10−5)	−0.79 (5.6 × 10−7, HC > SZ)	Positive: Superior/middle/inferior temporal gyri	7.7/7
		3	0.078 (4.11, 9.7 × 10−4)	−0.551 (4.9 × 10−4, HC > SZ)	Positive: Inferior semi-lunar lobule	3.2/2.6
					Uvula	1.0/1.0
					Pyramis	1.3/1.4
		4	0.071 (−3.89, 2.0 × 10−3)	0.532 (5.7 × 10−4, HC < SZ)	Positive: Lingual gyrus	2.6/1.5
					Negative: Inferior/middle Frontal gyri	1/1.6
					Lingual gyrus	1.0/0.2
					Insula	0.3/0.6
		5	0.13 (5.45, 1.5 × 10−5)	−0.418 (7.5 × 10−3, HC > SZ)	Positive: Lingual gyrus	1.7/2.0
					Cuneus	3.9/4.6
					Superior/Middle/inferior occipital gyri	4.9/4.6
					Precuneus	0.9/1.8
		6	0.098 (−4.65, 1.2 × 10−4)	0.401 (9.4 × 10−3, HC < SZ)	Positive: Insula	2.3/2.1
					Precentral/postcentral gyrus	1.2/1.4
					Negative: Inferior parietal lobule	0.4/0.4
					Cingulate gyrus	1.7/2.0
					Superior/middle temporal gyri	0.4/0.7
		7	0.11 (5.07, 4.5 × 10−5)	−0.353 (0.02, HC > SZ)	Positive: Superior/middle/inferior temporal gyri	5.5/4.7
					Middle/inferior frontal gyri	0.6/0.4
		8	0.05 (−3.25, 0.015)	0.024 (nonsignificance)	Positive: Superior/middle temporal gyri	3.6/2.7
					Supramarginal gyrus	0.6/0.8
					Superior/inferior parietal lobule	1.4/2.1
					Precuneus	0.5/0.3
2	−0.211 (.12, HC > SZ); NT5C2, CSMD1, FES, ARL3	9	0.038 (−2.82, 0.05)	−0.091 (nonsignificance)	Positive: Declive	1.4/1.5
					Pyramis	0.6/0.7
					Vermis	0.5/0.5
					CulmenThalamusCaudate	1.7/1.60.3/0.61.1/1.3
3	−0.296 (.06, HC > SZ); ITIH4		0.056 (3.44, 8.7 × 10−3)

Among the 3 methylation components, methylation component 1, mainly involving the NOSIP gene, showed a significant group difference (Cohen’s d = −0.348; P = 0.009). Consistently, the CpG site (cg05696092) located in the gene promoter region was hypomethylated in SZ (P = 6.4 × 10⁻³, uncorrected), and showed 5.17% mean methylation beta difference between 2 groups. This component was positively associated with GM components 1, 2, 3, 5, and 7, and negatively associated with GM components 4, 6, and 8 after multiple comparisons correction, as listed in table 2. These GM-methylation associations remained significant after controlling for diagnosis or in SZ group only (supplementary table S8), suggesting that the GM-methylation relationship is more general rather than driven by the disease. The interaction between methylation component 1 and diagnosis was significant in 4 pairs of GM-methylation components with stronger correlation in the patient group than in the HC group (supplementary figure S3 and supplementary table S5), likely reflecting larger variations (eg, much lower loadings in the GM and methylation components) only observed in patients.

Methylation components 2 and 3 were the ones linked to the SNP components aforementioned, and also significantly associated with GM component 9, mainly located at cerebellum, thalamus, and caudate (R² = .038 and .056; P = .05 and 8.7 × 10⁻³, respectively). Although connected to the same GM component, increasing GM in the component 9 regions was associated with decreased loadings of methylation component 2, and increased loadings of methylation component 3.

With respect to medication and symptoms, no significant effects were observed on the 3 methylation and 9 GM components using the threshold FDR ≤0.05 in SZ patients. However, significant associations were found between methylation component 1 with the working memory score (R² = .053, P = 1.27 × 10⁻³) and the TOL score (R² = .023, P = .04). Seven out of 8 GM components related to the methylation component 1 were also significantly associated with the working memory score while 4 of them were related to the TOL score after correcting for multiple testing (supplementary table S7).

In the replication data, we replicated significant correlations between methylation component 1 and 5 GM components 1, 2, 3, 5, 7 (supplementary table S6).The correlations remained significant after controlling for diagnosis (P < .05) and in SZ group only (supplementary table S8). We also replicated the associations of GM components 1, 2, 4, 7 with the working memory score and GM components 1, 2, 5, 7 with the TOL score (P < .06; supplementary table S7). GM components 1, 2, 3, 7 showed significant group differences with the same direction of GM changes as in the discovery data, indicating reduced GM in SZ patients on the corresponding brain regions. The correlations between methylation components 2 and 3 with the cerebellum GM component were not significant in the replication data (P = .15), but showed similar effects in HC group (R² = .05, P = .027).

Three-Way Link of SNP, DNA Methylation and GM

There was no significant association between SNP components and GM components in both datasets. However, the SNP component linked to methylation component 3 (the second SNP-methylation pair) was also marginally associated with GM component 9 (P = 3.38 × 10⁻⁴, uncorrected). A mediation test showed that methylation component 3 significantly mediated the SNP component’s effect on GM component 9 (average mediation effect = 0.0039, the proportion of SNP effect mediated: 23.63%, P = .02).

Discussion and Conclusions

In this work, we observed, in blood and saliva samples, 2 pairs of significantly related SNP-methylation components by multivariate analysis. The correlations were consistent in both SZ and HC groups and mainly contributed by cis-meQTLs (supplementary figure S1), suggesting that such intrinsic SNP-methylation relationship was driven by the effects of the SNPs on DNA methylation in close proximity, consistent with previous findings.^4,18,31,36 The benefit of identifying component based SNP-methylation correlations, instead of individual meQTLs, lies in the fact that multiple SNP or methylation loci can be functionally related and be grouped into one factor.³⁸ Component analysis not only reduced the number of statistical test but also derived a region of interest. The main regions covered by both SNP-methylation component pairs (Chr.10 around NT5C2, ARL3, and Chr.3 around ITIH4 respectively) were also identified as prominent meQTL regions in previous studies using brain tissues.^31,36

More than 97% of meQTLs were replicated in saliva, and about 20% were also reported in studies of brain tissues passing genome-wide significance.^31,36 We speculated more meQTLs might be replicated in brain if less strict significance thresholds were used. Given meQTL has been proposed to contribute to cross-tissue similarity,^30,57 we used data of blood and brain samples from the same 12 epilepsy subjects²⁹ to test cross-tissue similarity of the identified CpG sites. Two CpG sites (cg00035347 on NT5C2 and cg10872209 on ARL3) involving the majority of significant cis-effect, showed strong blood–brain correspondence with correlation r = 0.73 and 0.76, respectively. These findings support that the meQTL effects are relatively consistent across tissue types,^31,36 and likely mediate cross-tissue methylation similarity. Hence, it is plausible to study effects of this specific set of methylation sites, assessed from peripheral tissues, on brain structure in psychiatric disorders.¹⁷

SNP-affected methylation components 2 and 3 in the blood sample were significantly associated with one cerebellum GM component, and showed an interesting mediation effect, even though these correlations were replicated significantly only in HC group. Although SNP-effect on methylation was consistent cross tissues, different types of peripheral tissues may have different surrogate ability for brain. Given the contributing genes of these 2 components such as ITIH4, NT5C2 have shown expressional changes in brain tissues in SZ patients (P = .0025 and .01, respectively),⁵⁸ we speculate these methylation components are somewhat relevant to GM in brain. This hypothesis warrants further investigation using a larger sample.

Methylation component 1 from the gene coding for nitric oxide synthase interacting protein (NOSIP) was significantly associated with 5 GM components in both datasets. Further examination found a significant negative blood–brain correlation of the CpG site cg05696092 (r = −.63).²⁹ Considering that there was no meQTL identified on this site, a negative cross-tissue correlation suggests another mechanism, more likely environment-related, which may underlie the cross-tissue similarity as negative correlations were also observed in another study.⁵⁹ The site cg05696092 was located in the promoter region of NOSIP and was hypomethylated in blood and saliva samples of SZ patients, suggesting hypermethylation in brain of SZ. Testing for the change of NOSIP gene expression in brain showed significant decreased expression in SZ patients,⁵⁸ suggesting the hypermethylation may repress NOSIP expression. The protein encoded by the gene NOSIP mainly sequesters and inhibits the activity of nitric oxide synthase (NOS) and reduces nitric oxide (NO) production.^60,61 Several studies have found higher levels of NOS expression or NO in subjects with psychiatric disorders cross brain including the frontal cortex,^62,63 caudate,⁶⁴ hippocampus,⁶⁵ and cerebellum.⁶⁶ Additionally, NO or NOS have been shown to relate to neural cell apoptosis, synaptic plastic and remodeling, neurodegeneration, and the development of neuronal structures.^67–69 Taken together, we speculate that hypermethylation in the promoter of the NOSIP gene induces a decrease of NOSIP expression and an increase of NO production in the brain of patients, and further affects neuron activity which may be partially reflected by GM signal.

No significant SNP-GM associations were detected, and no SNP component showed group difference, perhaps due to the small sample size. Even though we focused on 108 SZ-risk regions, genetic effects on the brain may not be as large as expected.³⁶ In contrast, methylation component 1 showed altered methylation level in SZ in both datasets, and its associated 4 GM components also had replicated group differences. The most significance GM density reduction included frontal, temporal, and cerebellum regions, and an increase of GM from small regions of pre/post-central gyri, consistent with SZ brain structural abnormalities reported in a meta-analysis.⁵⁴ Additionally, cognitive ability presented replicated, strong relation with GM components in various regions. Methylation in NOSIP also showed an interesting relationship with GM components, though it was not replicated. Thus, regarding the relationship between SZ and cognition, our data suggest that GM measures show strong association, followed by methylation, and no SNP association.

The findings of this study should be interpreted with consideration of several limitations. First, methylation assay in the discovery data only covers promoter regions, which significantly limits our discovery power. Although the replication data have much higher coverage, we chose to use it to replicate our findings. Secondly, we were unable to confirm the methylation measures via a different platform. However, some publications^70,71 have reported high reproducibility of Illumina array for DNA methylation assay, in line with our own test–retest stability (Spearman rank correlation: .96–.99). Furthermore, we used an independent cohort from a different tissue to replicate the identified associations, strengthening the credibility of our findings. Thirdly, DNA from whole blood reflects a mixture of cell types. The estimated cell type proportion from blood⁵³ was correlated with methylation components 2 and 3 but did not alter the significance of the association findings. Fourthly, except alcohol abuse or dependence, other substance uses or comorbidities in SZ may have influence on epigenetics or GM. We tested the effect of nicotine use and further top 8 PCs of methylation data as surrogate of other potential confounding factors, and no significant change on the findings was observed. Finally, the findings in this study were mainly built on correlation analyses. We specifically intended to identify SNP-methylation associations, and then explore their effects on GM which could be further linked to the disease. While these associations suggest underlying biological pathways for SZ, we are unable to assess a causal relationship among the features. Other ICA-based methods (eg, 3-way parallel ICA⁷²) can be applied to the analysis of more than 2 modalities, but these methods may not optimally estimate the hierarchical relationship.

In summary, we first identified cross tissue SNP-methylation relationships within the known genetic SZ-risk regions, which were largely driven by cis-meQTL effects. We also found significant associations between methylation and GM components as well as SZ diagnosis in both blood and saliva tissues. These findings support a plausible mechanism of genetic risk to influence brain GM variation through epigenetic mediation. Furthermore, we identified cross-tissue similarities of methylation patterns unlikely to be regulated by SNP variants. Methylation-brain associations either regulated by SNPs or other factors, suggest the integral role epigenetics plays on brain structural variation and in the pathogenesis of SZ conveyed by both genetic and environmental influences.

Supplementary Material

Supplementary data are available at Schizophrenia Bulletin online.

Funding

This study was funded by the National Institutes of Health (P20GM103472 and R01EB005846) and National Science Foundation (1539067).

Acknowledgments

The authors declare that they have no competing interests.

Author notes

These authors contributed equally to the work