Introduction

Schizophrenia (SZ) is thought to be a disorder of cerebral cortical circuitry disruption. Particular SZ symptoms are associated with dysfunction of certain cortical circuits.1 For example, cortical circuit abnormalities in the superior temporal gyrus (STG), a brain region critical for auditory processing, are associated with auditory verbal hallucinations and impaired auditory sensory processing. Impaired auditory processing further contributes to phonologic dyslexia and difficulty recognizing and expressing spoken emotional tone (prosody) in SZ.2

Reduced dendritic spine density (DSD) in cortical STG layer 3, and other brain regions, is observed in postmortem studies of SZ.3, 4, 5, 6 We have previously demonstrated reduced DSD in STG layer 3 of SZ subjects in multiple cohorts.5, 6 We have also shown that the DSD reduction in SZ is of a similar magnitude in both the Heschl's Gyrus and planum temporale of the STG.6

Reduced DSD has several features indicating it is an intermediate phenotype for SZ. An intermediate phenotype is a heritable quantitative biological trait that is correlated with a disorder due, in part, to shared genetic architecture.7 A number of genes contribute to regulation of dendritic spine features including DSD8 and several of these are also SZ risk genes.9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22

The most useful intermediate phenotypes are functionally associated with aspects of the core clinical deficits of the disorder. DSD is intimately linked to neuronal function and changes in DSD are essential for normal cognition and sensory processing.8, 23, 24 Many disorders characterized, in part, by impaired cognition are also characterized by DSD abnormalities,8, 25 thus suggesting that reduced DSD likely contributes to cognitive deficits in SZ. For example, in the auditory cortex, dendritic spines on layer 2/3 neurons segregate frequency inputs to the neurons.26 Thus reduced DSD on these neurons would likely lead to impaired frequency discrimination, a deficit that has been observed in SZ.2

Elucidating the mechanisms of this intermediate phenotype is important for understanding SZ pathophysiology, identifying SZ treatment targets and developing animal models. DNA methylation (DNAm), the addition of a methyl group to a cytosine nucleotide, regulates gene transcription and is a strong candidate for such a mechanism. DNAm is altered in the brain27, 28, 29, 30, 31, 32, 33, 34 of SZ subjects and DNAm alterations are present in other contexts characterized by DSD abnormalities including neurodevelopmental disorders, models of addiction, and activity-dependent plasticity.8 In this study, we evaluated the hypothesis that DNAm correlates with DSD in the human STG and that this relationship is disrupted in SZ.

Materials and methods

Postmortem brains

Brains were recovered and processed as described previously.35 Briefly, brains were recovered during routine autopsies at the Allegheny County Medical Examiner’s Office, Pittsburgh, PA, USA following informed consent from next-of-kin. DSM-IV diagnoses were made based on clinical records and structured interviews with surviving relatives. The right hemisphere was blocked coronally and the resultant slabs snap frozen and stored at −80 °C. Slabs containing the STG were identified and the STG was removed as a single block. Samples containing all six cortical layers of STG (planum temporale), but excluding the adjacent white matter, were harvested. All procedures were approved by the University of Pittsburgh Committee for the Oversight of Research and Clinical Training Involving the Dead and the Institutional Review Board for Biomedical Research.

Cohort membership

The cohort was comprised of 22 subjects with either schizophrenia (N=16) or schizoaffective disorder.5 Schizophrenia and schizoaffective disorder were considered together because studies of DSD have not found differences between them.5, 6 Each SZ subject was matched with a non-psychiatric control (NPC) subject for sex, hemisphere, and as closely as possible for postmortem interval (PMI), age and other characteristics (Table 1 and Supplementary Table 1). DSD measures from STG layer 3 were available for 17 SZ-NPC pairs (NPC=0.036±0.0019 spines per μm3, SZ=0.028±0.0021 spines per μm3, t=2.8, P=0.084; Supplementary Figure 6). The 17 SZ-NPC pairs are a subset of the cohort studied in Shelton et al. (2015)5 and there is no overlap between the 17 SZ-NPC pairs and subjects studied in Sweet et al.6 Each pair was processed together to minimize experimental variability, and experimenter was blinded to subject’s diagnosis throughout.

Table 1 Cohort characteristics
Full size table

DNA preparation and bisulfite conversion

DNA (~10 μg) was isolated from STG gray matter (~20 mg) using AllPrep DNA/RNA/Protein Mini Kit (Qiagen, Valencia, CA, USA) and bisulfite converted using EZ-96 DNA Methylation Kit (Zymo Research, Irvine, CA, USA), both as per manufacturer’s protocol.

DNAm arrays

DNAm is the addition of a methyl group to a cytosine nucleotide within the context of a cytosine-phosphate-guanine (CpG) dinucleotide, usually, but also within the context of a cytosine-phosphate-H dinucleotide (CpH; H=adenine, cytosine or thymine).36 CpGs and CpHs are referred to as ‘DNAm sites’ or ‘sites’ in this manuscript. DNAm was measured at 485 577 sites (482 421 CpG dinucleotides, 3091 CpH dinucleotides and 65 SNPs) using Infinium HumanMethylation450 Beadchip Array (HM450; Illumina, San Diego, CA, USA) as per manufacturer’s protocol. β-values, the ratio of signal from a methylated probe relative to the sum of both methylated and unmethylated probes, were calculated. A β-value corresponds to the proportion of a particular site that is methylated in a sample.

Data preprocessing and filtering

Data analyses were performed using the R software environment (www.r-project.org). Color adjustment and background correction were performed using the bgAdjust2C method.37 Normalization was performed using the β-mixture quantile normalization method.38

Multidimensional scaling (MDS)39 was used to visualize the degree of similarity among subjects using HM450 data. Prior to data filtering, samples from four subjects were run in replicate and replicate samples from each of the four subjects clustered together (Supplementary Figure 1A). The β-values for each replicate pair were averaged for the remaining analyses. Samples also segregated by sex (Supplementary Figure 1A) and this segregation remained after filtering out data from SNP probes (N=65) and probes with detection P-values>0.01 in any sample (N=3390; Supplementary Figure 1B). After filtering out probe data from sex chromosomes (N=11 648), samples no longer segregated by sex (Supplementary Figure 1C), but segregation by race became evident (Supplementary Figure 1D). Filtering data from invariable probes (s.d.<5th percentile; N=23 547), did not alter similarity among samples (Supplementary Figure 1E). Data from 447 392 probes remained for downstream analysis.

Because samples did not segregate by sex on MDS analysis after filtering out probe data from sex chromosomes, sex was not considered a covariate in downstream analyses. Others have shown that DNAm sex differences on autosomal chromosomes are often small in magnitude and inconsistently reproduced.40, 41 Given segregation by race on the MDS analysis, race was included as a covariate in downstream analyses. One subject in this study was of Asian Indian ancestry. This subject, consistent with known genetic architecture,42 clustered with the subjects of European ancestry (Supplementary Figure 1D) and was thus combined with this group for analyses. Although the samples did not segregate by age on the MDS analyses (Supplementary Figure 1F), age was considered as a covariate in downstream analyses given the overwhelming evidence that DNAm varies extensively by age.43, 44, 45 Similarly, the samples did not segregate by PMI (data not shown) on the MDS analysis but because many factors that may have an impact on DSD in postmortem brain have been found to be particularly sensitive to PMI,46, 47 PMI was considered as a covariate for downstream analysis. All analyses presented in the body of this paper adjust for race, age and PMI. Results of analyses adjusting only for race and age can be found in Supplementary Tables 3–6.

Cell population estimation

DNAm differs markedly between neurons and glia.48 The proportion of neurons to glia in samples was estimated using a model based on β-values from cell-type-specific sites.49 Neuronal proportion did not differ between SZ and NPC subjects (Supplementary Figure 2A).

Site-specific DNAm-DSD correlations

Pearson correlations between DNAm (normalized β-values) at each site and DSD (spines per μm3) were calculated for all subjects. Examination of the linear DNAm-DSD correlations was performed for each of 3 groups (NPC and SZ subjects, NPC subjects and SZ subjects) using linear regression models with race, age and PMI adjusted.

Diagnosis-dependent differences in the DNAm-DSD correlations

Differences in the slopes DNAm-DSD correlations were assessed using linear regression models. For each site, two models were fitted: (1) DSD~β0+β1 × DNAm+β2 × diagnosis and (2) DSD~β0+β1 × DNAm+β2 × diagnosis+β3 × (DNAm × diagnosis). The likelihood-ratio test (LRT) was then used to test whether the DNAm-DSD correlation differed between SZ and NPC subjects.

Candidate genes for mediating reduced DSD in SZ

For each candidate gene, a permutation-based test was performed to assess whether the difference in slope of the DNAm-DSD correlation (SZ-NPC; across all sites in that gene) is significant. The diagnosis label (SZ or NPC) was permuted 1000 times for all 34 samples. Within each permutation, the differences in slopes of the DNAm-DSD correlation (SZ-NPC) for all the sites in that gene were computed, and then a one-sample t-test statistic was computed for these difference values. Finally, the permutation-based P-value was generated by comparing this one-sample t-test statistic under the true diagnosis to those under the permutation.

Results

There are more DNAm-DSD correlations than would be expected by chance in NPC, but not SZ, subjects

When all subjects are combined for analysis, there are more DNAm-DSD correlations than would be expected by chance (Figures 1a and b). This is true when NPC subjects only (Figures 1c and d) are considered. The number of DNAm-DSD correlations in SZ subjects is no more than would be expected by chance (Figures 1e and f). In the combined group, no DNAm-DSD correlations reached significance (P<1 × 10−7) and 84 reached a suggestive level of significance (P<1 × 10-4; Figure 1b). In NPC subjects, one DNAm-DSD correlation reached significance and 150 reached a suggestive level of significance (Figure 1d). In SZ subjects, no DNAm-DSD correlations reached significance and 51 reached a suggestive level of significance (Figure 1f and Table 3). After adjusting for potential confounders, DNAm-DSD correlations were, in general, less statistically significant (Table 2).

Figure 1
figure 1

Q-Q plots showing that DNAm-DSD correlation analysis is enriched in small P-values for (a) the group comprised of NPC and SZ subjects and (c) the group comprised of NPC subjects only, but not (e) the group comprised of SZ subjects only. Manhattan plots showing that DNAm at many sites correlate with DSD at a suggestive level of significance (P<1 × 10−4) in (d) NPC subjects and that the number of such sites is fewer in (f) SZ subjects and (b) when NPC and SZ subjects are considered together. DNAm, DNA methylation; DSD, dendritic spine density; NPC, non-psychiatric control; SZ, schizophrenia.

Full size image
Table 2 DNAm sites of DNAm-DSD correlations
Full size table

DNAm-DSD correlations at multiple sites differ between NPC and SZ subjects

Not only were there many fewer strong DNAm-DSD correlations in SZ subjects, the slopes of the linear DNAm-DSD correlations differed between NPC and SZ subjects at more sites than would be expected by chance (Figure 2a). The slopes of the DNAm-DSD correlations at two sites differed significantly (P<1x10-7), and at 269 sites suggestively (P<1 × 10-4), between SZ and NPC subjects (Table 3, Supplementary Table 4 and Figure 2b).

Figure 2
figure 2

(a) Q-Q plot showing that the differential DNAm-DSD correlation analysis is enriched in small P-values. (b) Manhattan Plot showing that the slopes of DNAm-DSD correlation at two sites significantly differ (P<1 × 10−7) between NPC and SZ subjects and that the slopes of DNAm-DSD correlation differ at an additional 269 DNAm sites at a level of suggestive genome-wide significance (P<1 × 10-4). DNAm, DNA methylation; DSD, dendritic spine density; NPC, non-psychiatric control; SZ, schizophrenia.

Full size image
Table 3 DNAm sites at which slope of DNAm-DSD correlations differed most between NPC and SZ subjects
Full size table

Candidate genes for mediating reduced DSD in SZ

We selected for more detailed follow-up genes meeting three criteria: (1) it was a gene for which there is evidence that one of its variants, either rare or common, genetically associates with SZ; (2) it was a gene in which a role in regulation of dendritic spines is established; and (3) it was one of the genes (or closest genes) to a site with a DNAm-DSD correlation reaching at least the suggestive level of significance, P<1 × 10−4. For this latter criterion the more liberal suggestive level of significance was chosen so as not to preclude detection of potentially causally related genes due to the limited power inherent in a study of the current sample size. Two genes met all three criteria: Brain-specific angiogenesis inhibitor 1-associated protein 2 (BAIAP2) and Discs Large, Drosophila, Homolog of, 1 (DLG1).

BAIAP2

DNAm-DSD correlations at two BAIAP2 sites (cg01276536 and cg23261327) reached a suggestive level of significance (Supplementary Table 3; criterion 1). Multiple rare BAIAP2 mutations have been associated with SZ16, 18 (criterion 2). BAIAP2 encodes a scaffolding and adaptor protein that regulates membrane and actin dynamics in dendritic spines and Baiap2 null mice exhibit reduced DSD17 (criterion 3).

DNAm (normalized β-values) at 120 of 176, or 68.2%, of the BAIAP2 sites were relatively hypomethylated in SZ subjects (Figure 3a), significantly more than the proportion of such sites observed among all sites analyzed (43.2%, Pearson χ2-test, P=0.00013).

Figure 3
figure 3

(a) BAIAP2 is hypomethylated in SZ subjects relative to NPC subjects and (b) the slopes of the DNAm-DSD correlations at most BAIAP2 sites analyzed differ between SZ and NPC subjects but DNAm at the sites associated with both cg01276536 and cg23261327 positively correlates with DSD independent of diagnosis. (c) Q-Q plot showing that the differential DNAm-DSD correlation analysis is enriched in small P-values compared to what would be expected by chance for the DNAm sites analyzed in BAIAP2. DNAm, DNA methylation; DSD, dendritic spine density; NPC, non-psychiatric control; SZ, schizophrenia.

Full size image

DNAm was correlated with DSD at many sites across BAIAP2 in both SZ and NPC subjects but the direction and magnitude of correlation often differed by diagnosis (Figure 3b). The slopes of the linear DNAm-DSD correlations differed between NPC and SZ subjects using the LRT (Figure 3c). Further, the difference in slope of the DNAm-DSD correlation (SZ-NPC; across all 176 BAIAP2 sites) is significant (permutation-based P=0.011).

The BAIAP2-AS1 gene is an antisense-oriented long non-coding RNA with a head-to-head orientation with respect to the 5′ region of BAIAP2. Like BAIAP2, BAIAP2-AS1 is characterized by DNAm-DSD correlations at multiple sites, which differ between NPC and SZ subjects (Figure 3b). The LRT performed to assess whether the correlations differ by diagnosis showed an excess of small P-values compared to what would be expected by chance (data not shown). The difference in slopes of the DNAm-DSD correlation (SZ-NPC; across 13 BAIAP2-AS1 sites) is significant (permutation-based P<0.001). Notably, the slope of the correlation was negative at all sites in NPC subjects and positive in 11 of 13 of the sites in SZ subjects (Supplementary Table 5 and Figure 3b).

DLG1

One site for which the DNAm-DSD correlation differed significantly between SZ and NPC subjects (cg07756562) is located in the region 5′ to DLG1 (Table 3; criterion 1). Studies have found common DLG1 variants to be associated with SZ.20, 21 Further, studies of copy-number variation have found a significant excess of deletions at the chromosomal position 3q29, which includes the DLG1 gene, in SZ50, 51 (criterion 2). DLG1 encodes a scaffolding protein that participates in the localization of glutamate receptors to the post-synaptic membrane and overexpression of DLG1 in organotypic slice cultures alters dendritic spine morphology 19, 52 (criterion 3).

DNAm sites in DLG1 and the genomic region immediately 5′ to DLG1 were characterized by a wide range of DNAm levels (Supplementary Figure 3B). DNAm levels did not exhibit any discernible pattern with respect to DLG1 gene features, though such an assessment is limited by the fact that no data for DNAm sites at the 3′ end of DLG1 were available in the data set. No overall hypo- or hypermethylation in SZ is evident in DLG1 (Pearson χ2-test, P=1).

The linear DNAm-DSD correlation at site cg07756562 differed between SZ and NPC subjects (P=6.22x10-8). At this site, DNAm correlates positively with DSD in SZ subjects and negatively with DSD in NPC subjects (Supplementary Table 6, Figure 4b).

Figure 4
figure 4

(a) DLG1 DNAm does not differ between subjects with SZ and NPC subjects. (b) The DNAm site cg0775662 is 5′ of DLG1 and is one of two DNAm sites in which the differential correlation between DNAm and DSD reached significance. DNAm, DNA methylation; DSD, dendritic spine density; NPC, non-psychiatric control; SZ, schizophrenia.

Full size image

Discussion

To our knowledge, this is the first postmortem brain study of the relationship of DNAm to DSD in SZ subjects. We evaluated the hypothesis that DNAm correlates with DSD in the human STG and that this relationship is disrupted in SZ subjects. Consistent with our hypothesis, we found DNAm to correlate with DSD at more sites than expected by chance in NPC, but not SZ, subjects. We also found that the slopes of DNAm-DSD correlations often differed between NPC and SZ subjects. We identified BAIAP2 and DLG1 as candidate genes for mediating DSD abnormalities in SZ.

DNAm-DSD correlations

Our findings suggest that DNAm is an important upstream mechanism for generating normal DSD and that this mechanism is disrupted in SZ subjects. Although, to our knowledge, this is the first time a DNAm-DSD relationship has been demonstrated in SZ, DNAm is altered in a number of contexts characterized by abnormal DSD.53 Perhaps the most convincing evidence for a causal effect of DNAm on DSD comes from the study of addiction models where overexpression of a DNA methyltransferase, and downstream DNAm, alone, is sufficient to alter DSD.54

The DNAm alterations observed in SZ, and thus the disrupted DNAm-DSD relationships, are likely to result from a combination of both genetic and environmental factors.36 Notably, common risk variants for SZ have been shown to regulate local DNAm,31, 32 but none of the sites in Tables 2 and 3 have been identified as targets of methylation quantitative trait loci (mQTLs) that overlap with SZ risk loci32 or neurodevelopmental mQTLs (http://epigenetics.essex.ac.uk/mQTL/).31 A number of environmental factors have been implicated in the pathogenesis of SZ,55 and many of them have been shown to alter DNAm.56 It is also important to consider that alterations in DNAm and DNAm-DSD correlations observed in SZ subjects may be the result of treatment-induced changes in the brain. We have previously found that antipsychotic treatment does not alter STG DSD in an animal model6 but accumulating evidence suggests that antipsychotics do alter DNAm.57 However, DNAm alterations are observed in peripheral blood from early SZ subjects with only brief (<16 weeks) antipsychotic treatment58 thus suggesting that not all DNAm alterations in SZ are explained by antipsychotic treatment. In some cases, SZ-associated DNAm alterations are normalized by antipsychotic drugs,59 perhaps suggesting that the therapeutic effect of antipsychotics are mediated, in part, by affecting DNAm.

Putative mechanisms underlying DNAm-DSD correlation

The study of DNAm function has historically focused on its role in promoter regions. In this context, DNAm usually blocks transcription. It is now recognized that DNAm function is context dependent60 and that intragenic and intergenic DNAm affects transcription. Notably, DNAm affects alternative promoter usage, regulation of short and long non-coding RNAs, alternative splicing and enhancer activity.61, 62

DNAm-DSD correlations in BAIAP2 and DLG1 were distributed throughout intragenic, associated non-coding RNAs, and promoter regions, suggesting that DNAm may alter DSD in SZ by affecting transcription via both canonical and non-canonical mechanisms. DNAm at two BAIAP2 sites within intron 7–8 are strongly and positively correlated with DSD. These two sites are in a CCCTC-binding factor (CTCF) binding site. CTCF binds unmethylated DNA and, in intragenic contexts, promotes exon inclusion into the mature transcript.62 We predict that DNAm at these sites leads to an increase in BAIAP2 transcript variants with exclusion of exons local to intron 7–8 and that these transcript variants positively regulate DSD. Consistent with this prediction, multiple BAIAP2 transcript variants have been identified which differ with respect to the composition of their 3′ end (Miyahara et al., 2003) and primary data including mRNA and EST alignments suggest that there are transcripts that do not contain exon 7 and/or 8.63

Most of the BAIAP2 DNAm-DSD correlations in NPC subjects, however, were negative ones in which lower DNAm was correlated with higher DSD. DNAm at a site 5′ to DLG1 (cg07756562) is similarly correlated with DSD in NPC subjects. Decreased DNAm in 5′ regions is usually associated with increased total transcription.64 We suggest that lower DNAm at these sites allows for increased BAIAP2 and DLG1 transcription, promoting dendritic spine formation. Supporting this interpretation is evidence that BAIAP2 overexpression promotes DSD17 and DLG1 overexpression promotes dendrite growth and complexity.65, 66

Other DNAm-DSD correlations are annotated to BAIAP2-AS1 and are also relatively hypomethylated in SZ subjects. We anticipate increased BAIAP2-AS1 transcription in SZ as a result of this hypomethylation. It is difficult to know how higher levels of BAIAP2-AS1 might affect expression of BAIAP2. Antisense long non-coding RNAs, like BAIAP2-AS1, often regulate local gene transcription at multiple levels 67 but BAIAP2-AS1 has not been studied.

DNAm differences between SZ and NPC subjects at particular sites is one mechanism by which the DNAm-DSD correlations may be disrupted in SZ. Indeed, our data suggest that there are many sites where DNAm differs between SZ and NPC subjects (Supplementary Table 2 and Supplementary Figure 4). Global DNAm, however, does not appear to differ between SZ and NPC subjects (Supplementary Figure 5). Disruptions of DNAm-DSD correlations in SZ that do not result from a change in DNAm may reflect an abnormality downstream of DNAm (e.g., disrupted binding of a DNAm-dependent transcription factor and so on).

Limitations

Despite plausible relationships between DNAm in multiple genes (including BAIAP2 and DLG1) and DSD, the findings, like those of any postmortem brain study, are only correlative and cannot establish a mechanistic relationship. Our use of SZ risk gene and DSD regulator criteria to define candidate genes limits the ability to identify novel genes important in SZ pathophysiology or dendritic spine regulation. However, because of the large number of sites tested, there is a likelihood that some DNAm-DSD correlations are spurious and not relevant to the DSD phenotype in SZ. We chose to use these criteria to increase the probability of identifying DNAm alterations that may contribute causally to the DSD phenotype in SZ. Another potential technical limitation is that the sites studied were constrained by the use of the HM450 array. It only measures a fraction of the >28 million DNAm sites in the human genome and coverage is biased toward CpG islands, promoters and genic regions.

Conclusions and future directions

The study of reduced DSD as an intermediate phenotype in SZ across different levels of analysis including genetics,68 transcriptomics69 and proteomics35 has provided valuable insights into SZ. This study suggests that epigenetic alterations, specifically disrupted DNAm-DSD correlations, in SZ may be a mechanism for SZ-related reductions in DSD and justify future studies probing this relationship.

Studies to confirm the DNAm-DSD relationship and the effect of DNAm on candidate gene transcription in additional, larger cohorts will be a critical next step. Also, DNAm varies widely between cell types in the human cerebral cortex, with studies indicating that DNAm in GABA neurons is more extensive by several fold than in glutamatergic neurons.70 Thus, future cell-type-specific studies, using laser capture microdissection,71 fluorescent-activated nuclei sorting72 or similar methods, may increase the likelihood of detecting diagnosis-specific DNAm alterations by decreasing variability and revealing findings that were masked by opposing DNAm changes in different cell types. Studies in model systems to evaluate the correlative versus causal nature of the DNAm-DSD relationship will be also important. Understanding the DNAm-DSD relationship may facilitate the development of new, and/or the repurposing of existing, DNAm modifying drugs for SZ treatment.