Common DNA methylation alterations in multiple brain regions in autism

Abstract

Autism spectrum disorders (ASD) are increasingly common neurodevelopmental disorders defined clinically by a triad of features including impairment in social interaction, impairment in communication in social situations, and restricted and repetitive patterns of behavior and interests, with considerable phenotypic heterogeneity among individuals. Although heritability estimates for ASD are high, conventional genetic-based efforts to identify genes involved in ASD have yielded only few reproducible candidate genes that account for only a small proportion of ASDs. There is mounting evidence to suggest environmental and epigenetic factors play a stronger role in the etiology of ASD than previously thought. To begin to understand the contribution of epigenetics to ASD, we have examined DNA methylation (DNAm) in a pilot study of post-mortem brain tissue from 19 autism cases and 21 unrelated controls, among three brain regions including dorsolateral prefrontal cortex, temporal cortex, and cerebellum. We measured over 485,000 CpG loci across a diverse set of functionally relevant genomic regions using the Infinium HumanMethylation450 BeadChip and identified 4 genome-wide significant differentially methylated regions (DMRs) using a novel bumphunting approach and a permutation-based multiple testing correction method. We replicated 3/4 DMRs identified in our genome-wide screen in a different set of samples and across different brain regions. The DMRs identified in this study represent suggestive evidence for commonly altered methylation sites in ASD and provide several promising new candidate genes.

INTRODUCTION

Autism Spectrum Disorders (ASD) are common neurodevelopmental disorders defined by three core clinical features: (1) impairment in social interaction; (2) impairment in communication in social situations; and (3) restricted and repetitive patterns of behaviors and interests^{1, 2}. Given the striking recent increase in ASD prevalence and its associated social and economic impact³, there has been considerable interest in understanding the biological basis of ASD⁴; however the molecular underpinnings of ASD still remain elusive.

Familial aggregation, rare chromosomal abnormality, and twin studies suggest there is a strong genetic basis to ASD. Numerous genetic studies have examined common variants, rare mutations, and copy number variants (CNVs) across many thousands of autistic children but have had limited success. The strongest genetic candidates identified to date involve rare mutations. For example, rare variants have been identified in only 6/461 ASD cases for CACNA1H^{5, 6} and CNTN4 deletions have been identified in 7/~2000 cases^7–10. These genes represent two of the five leading candidate genes in ASD to date¹¹. Several promising recent studies have revealed an association between de novo mutations and CNVs with ASD; however de novo events do not directly explain the high heritability OF ASD¹². A recent review suggests that at most 10–15% of all non-syndromic cases of ASD are associated with genetic alterations¹³, and to date, whole-genome and in-depth exon sequencing studies have suggested ASD is minimally explained by genetics alone.

Although epigenetic involvement has been a persistent minority view of ASD etiology, there is mounting evidence to suggest the contribution of epigenetics to ASD is stronger than previously thought. Direct evidence¹⁴ comes from identification of ASD-associated chromosomal abnormalities in imprinted regions, ASD linkage and association in areas of known imprinting, and evidence of parent-of-origin effects in association and linkage signals^{15, 16}. Additional support stems from the observation of autistic features, sometimes reaching the diagnostic threshold, in individuals affected with known epigenetic disorders such as Angelman, fragile X, and Rett Syndromes. Rett Syndrome is caused by mutations in the MECP2¹⁷, a gene that encodes a protein responsible for recognizing DNA methylation marks throughout the genome. Angelman syndrome involves alterations of the chromosome 15q11-q13 imprinted gene cluster¹⁸. Lastly, fragile X syndrome is caused by the expansion of a methylated CGG repeat in the 5′ untranslated region of the FMR1 gene¹⁹, resulting in inappropriate gene silencing.

Previous genome-wide DNA methylation (DNAm) assays, applied to ASD^{20, 21} and other diseases, have focused on measuring DNAm at CpG sites assumed to be functionally important, i.e. CpG islands (CGI) and promoters. However, we recently demonstrated that most functionally important changes in DNAm occur outside of CGI and promoter regions, specifically at CpG island shores^{22, 23}. Non-island and promoter regions are the predominant sites of aberrant DNAm for many diseases including several types of cancers ^{22, 24–28} and rheumatoid arthritis^{29, 30}. Previous limited studies of specific candidate genes and CpG island or promoter arrays have shown epigenetic differences in lymphoblastoid cell lines and brain samples from individuals with ASD^{20, 21}. However, these studies were limited either in scope, focused on particular regions of the genome, or by sample size, and/or by cell type (lymphoblast cell lines). One recent study that examined trimethylated histone 3 lysine 4 (H3K4me3) marks in post-mortem prefrontal cortex brain tissue identified a slight spreading of H3K4me3 marks, normally restricted to promoter regions, into gene bodies among 4/16 autistic individuals³¹. Nonetheless, to date, no study has examined DNAm on a genome-scale, not limited to CpG islands, in brain samples of ASD cases.

Here, we report the first genome-wide examination of DNAm in ASD among 40 post-mortem brain samples, including 19 cases and 21 controls, across 3 brain regions, using the Infinium HumanMethylation450 BeadChip (450K). Availability of ASD brain samples is extremely limited³², with the numbers examined in this study being comparable to other molecular analyses of ASD post-mortem brain samples^{31, 33}. The 450K quantitatively measures DNAm at 485,577 genomic loci at high value content regions including all CpG island and promoter regions as well as CpG shores, cancer and tissue-differentially methylated regions (DMRs), non-coding RNAs, and DNase hypersensitive sites^{34, 35}. We applied a new methodology termed “bump hunting”³⁶ to our analysis which allows effective modeling of measurement error, detection of differentially methylated regions (DMRs), and assessment of genome-wide statistical significance (via permutation testing). Identifying regions of differential methylation is advantageous to single site analysis for multiple reasons including better protection from technical artifacts associated with individual probes and functional significance of regional methylation change versus at a single CpG.

MATERIALS AND METHODS

Human post-mortem brain tissue samples

We acquired 41 post-mortem brain tissue samples, including 20 cases and 21 controls, representing three brain regions: temporal cortex (TC, n=16), prefrontal cortex (PFC, n=12), and cerebellum (CBL, n=13) from the Autism Tissue Program and Harvard Brain Tissue Resource Center at McLean Hospital, Belmont, MA (TC and PFC samples) and the National Institute of Child Health and Human Development (NICHD) Brain and Tissue Bank for Developmental Disorders at the University of Maryland, Baltimore (CBL samples). These 41 samples represent 36 individuals; 5 individuals have both a TC and PFC sample included in our analyses (Figure 1a). For each brain region, autistic and control samples were matched as best as possible for age, sex, and post-mortem interval; no significant differences were detected (t-test, p < 0.05 and Supplementary Table 1). Furthermore, we examined the relationship between DNAm and age/PMI for all 4 DMRs identified in our genomic screen; no correlation between the covariates and DNAm was observed (Supplementary Figures 1 and 2). All autistic individuals included in our analysis were diagnosed using scores from the Autism Diagnostic Interview (ADI-R) and/or the Autism Diagnostic Observation Schedule (ADOS). In Supplementary Table 2, we provide a list of post-mortem brain samples and their associated features.

Figure 1.

(a) Summary of post-mortem brain samples and analytic procedures utilized. (b and c) Modified region-based (“Dry”) Manhattan plots for epigenome-based analysis. Region-specific DNA methylation was calculated from 450K data, and p values were determined by permutation analysis to accommodate the genome-wide screen. (b) Three differentially methylated regions (DMRs) in the temporal cortex brain region are significantly associated with ASD. (c) One DMR in the cerebellum is significantly associated with ASD. Each red segment represents the adjusted p-value for a genomic region on the Illumina Infinium 450K Methylation BeadChip. The significance threshold was based on a family-wise error rate of 0.1.

Illumina Infinium 450K methylation measurements

Genomic DNA was extracted from TC and PFC samples using Trizol (Invitrogen) and from CBL samples using the MasterPure DNA purification kit (Epicentre Biotechnologies) according to the manufacturer’s specifications for each kit. Each DNA sample was bisulfite treated, 500ng of gDNA, using the EZ DNA methylation kit (Zymo Research) according to the manufacturers specifications, optimized for the 450K. All of the samples per brain region were processed on the same plate with cases and controls randomized across the wells to minimize potential plate or batch effects.

Statistical analysis

Overview and quality control measures

All statistical analyses were performed using R 2.15 and Bioconductor 2.9. Raw intensity files (.idat) were obtained and processed using minfi package³⁷ to obtain log ratios of methylation percentage (M-values). We applied several quality control measures to remove any spurious samples or probes. First, we examined 450K control probes to assess bisulfite conversion, extension, hybridization, staining, specificity, negative control, and others; no outlier samples were detected. Next, we checked for sex-discrepancies by comparing self-reported genders against data derived sex values. No sex-discrepancies were identified. We then looked for poorly performing arrays by calculating the total array intensity (sum of the methylated and unmethylated signals across all probes on the array) for each sample; no substantial differences were detected. Unsupervised clustering of samples identified one sample that was inappropriately clustering with a different brain region (PFC sample clustering with CBL samples); this sample was removed from downstream analysis. We removed probes that had an annotated SNP (dbSNP134) at the single base extension or CpG site since it is possible that SNP differences in these locations may manifest as differential methylation on the 450K, leaving a total of 428,526 probes.

Preprocessing and normalization

The signal in the methylated and unmethylated channels was first computed without background correction. Each channel was normalized separately, by first quantile normalizing the signal for all autosomal probes and second by quantile normalizing the signal for the sex chromosomes for each sex separately. After normalization, log ratios of methylation percentages were computed (M-values). Previous work³⁸ has shown M-values to perform better than Beta-values for Illumina methylation arrays.

Identification of differentially methylated regions (DMRs)

We adapted the bump hunting technique previously described³⁶ to the 450K. Probes were assigned to clusters so that two neighboring probes in the same cluster are separated by at most 500bp. For each probe, we estimated the difference in average log-ratios between cases and controls, controlled for sex, and smoothed these estimated differences using running medians. Next, the smoothed estimated differences were thresholded based on the 97.5% percent quantile of the empirical distribution of the smoothed estimated differences. This yielded a set of putative DMRs. Significance was assigned by permutation testing. For each of 1000 permutations of case-control status, a new list of putative DMRs was obtained. The genome-wide family-wise error rate for each observed DMR was calculated as the proportion of null-derived DMRs across the genome with more CpGs and greater difference between cases and controls than the observed DMR. Note that most of the genome is not inside a DMR, and is therefore not assigned an empirical significance via this approach.

Replication analysis

For the 3 DMRs identified in our genome-wide screen of TC tissue, we performed replication analyses in PFC and CBL tissues using methylation data obtained using the 450K. For replication purposes, the PFC and CBL data was restricted to include only the 74 probes that were located within the 3 significant TC DMRs. Similarly, for the DMR we identified in CBL tissue, we performed replication analyses by restricting our analysis in PFC and TC tissues to the 12 sites located within the cerebellum DMR. P values were computed using a two-sided t-test with a significance threshold of 0.05. Any DMR identified in our genome-wide screen that also had at least 2 significant differentially methylated probes in a replication set was deemed a positive replication.

Assessment of cell composition differences

Brain cell type specific DNAm data and cell epigenotype specific (CETS) patterns ³⁹ were acquired using the CETS package (http://psychiatry.igm.jhmi.edu/kaminsky/software.htm) in R/Bioconductor. Overlap between the 450K probes identified in our brain samples and the top 10,000 CETS marks for brain cell types was computed by simply comparing the two lists of 450K probe names. Using the CETS package, we then estimated the proportion of neuronal and glial cells for each post-mortem brain tissue sample in our study, as specified previously³⁹.

RESULTS

Figure 1a depicts an overview of our experimental design and the brain samples utilized in our analyses. Briefly, we performed individual methylome analyses for each of three brain regions: dorsolateral prefrontal cortex (PFC, 6 cases/5 controls), temporal cortex (TC, 6 cases/10 controls), and cerebellum (CBL, 7 cases/6 controls) using a “bump hunting” approach.

Genome-wide screen

For each of the three brain regions we performed a genome-wide screen and identified a total of four genome-wide significant (adjusted p < 0.1) differentially methylated regions (DMRs), three in the superior temporal cortex and one in the cerebellum (Figure 1b, c and Table 1). Because bump-hunting identifies a differentially methylated region (DMR) rather than a differentially methylated position (DMP), the data display is similar but with a somewhat more austere appearance than a conventional Manhattan plot. Therefore, we have termed it a “dry Manhattan” plot suitable for this type of DNAm analysis. As described in detail in the Methods section, empirical P values reported in our dry Manhattan plots (Figure 1b, c) were calculated by permutation with a family-wise error rate of 0.1.

Table 1.

Genome-wide significant DMRs associated with ASD

Brain region	Genomic location	Nearest gene	Gene distance (bp)	ΔMa	Adjusted p-valueb	Replication (p < 0.05)c
						TC	PFC	CB
TC	chr6: 32115964-32117379	PRRT1	0	−7.8%	0.001	na	yes	yes
TC	chr11: 2321770-2323272	C11orf21	0	−6.6%	0.013	na	yes	no
TC	chr6: 29648379-29649092	ZFP57	3449	13.9%	0.019	na	no	yes*
CB	chr5: 1594021-1595048	SDHAP3	0	15.8%	0.073	no	no	na

Abbreviations: TC, temporal cortex; PFC, prefrontal cortex; CB, cerebellum; M, DNA methylation

^aMean methylation difference between autism and control groups for the differentially methylated region. Positive values reflect relative hypermethylation and negative values relative hypomethylation in autistic individuals compared to controls.

^bEmpirical significance value, computed using permutation testing (n=1000).

^cReplication is equal to “yes” if any Illumina Infinium 450K probes located within the DMR (identified in the genome-screen) have a nominal p-value less than 0.05 in a different brain tissue, otherwise it is reported as “no”.

^*Sex-specific replication. Significance is not reached when males and females are examined as one set.

As shown in Figure 2a, we identified a DMR at the 3′ UTR of PRRT1, proline-rich transmembrane protein 1, with relative hypomethylation in the temporal cortex tissue of autistic individuals (shown in green) compared to controls (shown in purple). On average, the cases are 7.8% less methylated at this DMR (empirical P = 0.001) than controls (Table 1). It is possible that changes in copy number variants (CNVs) may manifest as DNAm changes on the 450K. We would expect CNVs to result in different total intensities, i.e. the sum of the methylated and unmethylated channels, on the 450K platform. Therefore to assess potential CNV changes, in the second panel of Figure 2a, we plot the difference between the total probe intensity for a given individual and the mean total intensity across all individuals for a given probe, with purple and green points denoting control and ASD, respectively. There are no clear differences between the case and control groups with respect to CNV status; thus, it is likely the DMR does not reflect a CNV.

Figure 2.

Two genome-wide significant differentially methylated regions (DMRs) located at PRRT1 (a–c) and C11orf21/TSPAN32 (d–f) are relatively hypomethylated in temporal cortex tissue from autistic individuals. (a and d) DMR plots for temporal cortex brain samples. (b and e) DMR plots for prefrontal cortex brain samples. (c and f) DMR plots for cerebellum brain samples. For (a and d), methylation data is shown in the top panels. Smoothed lines denote mean methylation levels for individuals with (green) and without (purple) ASD across the DMR. Each point represents the methylation level of a particular individual, green for those with autism and purple for those without autism, at a specific genomic location. The methylation values range from 0 to 1 with 0 equal to 0% methylation and 1 equal to 100% methylation. The upper middle panels provide information about potential copy number changes. Each point, green for autistic and purple for normal individuals, represents the difference between total probe intensity for a given individual and the mean total intensity across all individuals for a given probe. The black points in the lower middle panel are individual probe-based nominal p-values for differences in methylation between the two groups (autism and normal). The dashed magenta line indicates the significance threshold used (nominal p < 0.05) such that points above the line are significant in replication analyses. The bottom panel displays relevant genome annotation information. For (b, c, e, and f), plots are as described above but blue and orange points represent individuals with and without autism, respectively. Dotted boxes describe the location of methylation values that correspond to probes with significant nominal p-values.

We also identified a DMR (empirical P=0.013) that is on average 6.6% less methylated in ASD temporal cortex tissue, shown in green, compared to controls, shown in purple (Table 1 and Figure 2b). This DMR is located within the promoter regions of tetraspanin 32 (TSPAN32), and chromosome 11 open reading frame 21 (C11orf21), and continues into the gene body of C11orf21 (bottom panel of Figure 2b). We do not observe substantial changes in CNV levels between cases and controls for this DMR (second panel of Figure 2b).

The third DMR (empirical P=0.019) in temporal cortex tissue is on average 13.9% more methylated in ASD, shown in green, compared to controls, shown in purple (Table 1 and Figure 3). The relatively hypermethylated DMR is located in an intergenic region; the nearest gene is ZFP57, located 3.5kb away (Figure 3a). Although a few control individuals appear to have a slight increase in copy number for a few probes, across the region there is no substantial difference in CNV state between cases and controls (Figure 3a, second panel).

Figure 3.

A genome-wide significant differentially methylated region (DMR) located 3.45 Kb upstream of ZNF57 is relatively hypermethylated in temporal cortex tissue from autistic individuals. (a) DMR plots for temporal cortex brain samples. (b) DMR plots for temporal cortex brain samples from male individuals. (c) DMR plots for temporal cortex brain samples from female individuals. For (a–c), methylation data is shown in the top panel. Lines denote mean methylation levels for individuals with (green) and without (purple) autism across the DMR. Each point represents the methylation level of a particular individual, green for those with autism and purple for those without autism, at a specific genomic location. The methylation values range from 0 to 1 with 0 equal to 0% methylation and 1 equal to 100% methylation. The upper middle panel provides information about potential copy number changes. Each point, green for autistic and purple for normal individuals, represents the difference between total probe intensity for a given individual and the mean total intensity across all individuals for a given probe. The black points in the lower middle panel are individual probe-based nominal p-values for differences in methylation between the two groups (autism and normal). The dashed magenta line indicates the significance threshold used (nominal p < 0.05) such that points above the line are significant in replication analyses. The bottom panel displays relevant genome annotation information. (d) DMR plots for all prefrontal cortex brain samples. (e) DMR plots for prefrontal cortex tissue samples from male individuals. (f) DMR plots for prefrontal cortex tissue samples from female individuals. (g) DMR plots for all cerebellum brain samples. (h) DMR plots for cerebellum samples from male individuals. (i) DMR plots for cerebellum samples from female individuals. For (d–h), plots are as described above but blue and orange points represent individuals with and without autism, respectively. Dotted boxes describe the location of methylation values that correspond to probes with significant nominal p-values.

Among 13 cerebellum samples, including 7 cases and 6 controls, we identified 1 significant DMR located within SDHAP3, succinate dehydrogenase complex, subunit A, flavoprotein pseudogene 3. Cases are, on average, 15.8% more methylated than controls across this region (Table 1), although there is considerable variation among the ASD group that seems to be, at least in part, related to CNV (Figure 4a). The cases with higher methylation levels also have higher total intensities, representing potential CNVs. Thus, the methylation change detected in this region may be influenced by an underlying CNV change. Nonetheless, it still implicates this particular region as important in the etiology of ASD.

Figure 4.

A genome-wide significant differentially methylated region (DMR) at SDHAP3 is relatively hypermethylated in cerebellum tissue from autistic individuals. (a) DMR plots for cerebellum brain samples. (b) DMR plots for prefrontal cortex brain samples. (c) DMR plots for temporal cortex brain samples. For (a), methylation data is shown in the top panel. Lines denote mean methylation levels for individuals with (green) and without (purple) autism across the DMR. Each point represents the methylation level of a particular individual, green for those with autism and purple for those without autism, at a specific genomic location. The methylation values range from 0 to 1 with 0 equal to 0% methylation and 1 equal to 100% methylation. The upper middle panel provides information about potential copy number changes. Each point, green for autistic and purple for normal individuals, represents the difference between total probe intensity for a given individual and the mean total intensity across all individuals for a given probe. The black points in the lower middle panel are individual probe-based nominal p-values for differences in methylation between the two groups (autism and normal). The dashed magenta line indicates the significance threshold used (nominal p < 0.05) such that points above the line are significant in replication analyses. The bottom panel displays relevant genome annotation information. For (b and c), plots are as described above but blue and orange points represent individuals with and without autism, respectively. Dotted boxes describe the location of methylation values that correspond to probes with significant nominal p-values.

Replication analyses

Finally, we sought to confirm DMRs identified in our genome-wide screen in an independent set of samples. In total 74 probes are located within the 3 TC DMRs and 12 probes are located within the 1 cerebellar DMR. Given the limited availability of post-mortem brain samples from autistic individuals, we considered replication of TC DMRs via evidence of association at those probes in the PFC and CBL tissue samples, and used the TC and PFC sample sets to replicate our CBL DMR finding. Since the samples for each brain region were obtained from different individuals, with the exception of 5 individuals that are represented in both the TC and PFC sample sets, this design is suitable to evaluate how well our genome-wide findings replicate in different individuals and across different brain regions. Figure 1 provides a summary of the total number of individuals and probes examined in our replication analyses.

For the PRRT1 associated DMR discovered in TC, we found significant methylation differences between cases and controls in PFC tissue for 14 of 33 sites (Figure 2b, blue). For clarity, we use dashed boxes to highlight methylation values that correspond to probes with significant p-values. We did not observe any differences related to copy number for this region in PFC samples (upper middle, Figure 2b). We also observed two significant CpG sites and others with suggestive significance among the cerebellar tissue (Figure 2c) in an independent sample of 7 cases and 6 controls. Consistent with both our genome-wide screen and PFC replication results, the cerebellum tissue from autistic individuals showed less methylation than control individuals at the PRRT1 locus (Figure 2c) and no differences in CNV between cases and controls.

Replication analysis of the TC DMR associated with TSPAN32 and C11orf21 revealed 7 of 26 and 1 of 26 significant (P < 0.05, third panel of Figures 2e and f) differentially methylated loci in the PFC and cerebellum tissue samples, respectively (Figures 2e and f). However, since we would expect to see one significant difference by chance, we do not report the CBL as significant in our replication result (Table 1, Figure 1a). As shown in the top panels of Figures 2e and 2f, autistic individuals (blue points) are less methylated than the controls (orange points) for both the PFC and CBL tissues, consistent with our genome-wide result in TC tissue. In addition, no CNV changes are observed at the significant loci (second panel, Figures 2e, f).

The third DMR identified from TC, located 3.5kb upstream of ZFP57, showed sex-specific replication in the CBL samples (Figure 3i). We decided to stratify by sex for this DMR because we observed two clusters of methylation values that were correlated with sex. Interestingly, for both TC and CBL tissues, upon stratification by sex we found that normal females are less methylated than autistic females. Since there was only one female, we could not assess significant differences between cases and controls for females in PFC tissue. However, in the one female (autistic) with PFC data, we observed high methylation levels (Figure 3f); consistent with our findings in the other 2 regions for autistic females (Figures 3c, i). We also observed suggestive evidence that autistic males have more DNAm variability in TC and CBL tissues than control individuals (Figures 3b, h) for this genomic region.

Lastly, although none of the loci located in the CBL DMR in SDHAP3 reached statistical significance in the replication sets (Figures 4b, c), we observed the same trend in PFC as we did in CBL. More specifically, a few autistic samples in PFC show methylation levels near 50% that are potentially related to CNV changes (Figure 4b).

Brain tissue heterogeneity

We evaluated the possibility that our ASD-associated DMRs may reflect underlying cell type heterogeneity between ASD cases and controls using neuronal and glial cell type specific methylation data ³⁹. None of the 86 probes located within the ASD DMRs we identify are among the list of 10,000 probes with neuron-specific methylation values from a recent report ³⁹. Furthermore, for each of our brain samples, we computed the proportion of neuronal and glial cells using the Guintivano et al. ³⁹ algorithm and found no significant differences between the ASD cases and controls for either brain region (Supplementary Table 3). Thus, our findings do not simply reflect a shift in the proportion of neuronal and glial cells between ASD case and control samples.

DISCUSSION

We provide the first evidence for significant DNAm changes in post-mortem brain tissue from ASD patients. After adjusting for multiple testing, we identified a total of four differentially methylated regions, three in TC tissues and one in CBL. Three of the four DMRs identified via genome-wide screen replicated in independent samples from different brain regions. Although the fourth DMR did not meet significance criteria for replication, it did show the same trend in the PFC samples as we observed in our genome-wide screen using CBL samples. Lastly, we demonstrated that the DMRs we identified are not related to underlying brain tissue cell composition differences between ASD and control individuals.

There is substantial evidence to support the biological plausibility of the regions we identified in the etiology of ASD. The DMR within the PRRT1 3′ UTR overlaps two DNase hypersensitive sites and an alternative transcript finish site. Given the location it is possible that this DMR is an important regulatory site^{40, 41}. Although little is known about the function or expression patterns of PRRT1 in humans it has been shown to be specifically expressed in the hippocampus of marmoset⁴², a nonhuman primate often used as an ideal model for a wide-variety of central nervous system disorders. The hippocampus brain region has been previously implicated in human studies of ASD demonstrating macroscopic and microscopic anatomical differences^43–46 and altered synaptic function and plasticity⁴⁷ in autistic individuals. Mutations in other genes in the PRRT family have been shown to cause several neurological disorders in humans, some of them developmental in nature, including familial infantile seizures⁴⁸, paroxysmal kinesigenic dyskinesia (PKD)⁴⁹, hemiplegic migraine⁵⁰, PKD combined with infantile seizures (ICCA)^{51, 52}, and benign familial infantile seizures (BFIS)^{53, 54}.

The second DMR identified in TC tissue is located in the promoter regions of TSPAN32 and C11orf21 and extends into the gene body of C11orf21. While there is no literature to describe the function of C11of21, TSPAN32 is important in cellular immunity⁵⁵ and acts as a structural and cell signaling scaffold protein⁵⁶. Functional mutations in other tetraspanins have been identified in schizophrenia and bipolar pateints ⁵⁷. Mutations in several other scaffolding proteins such as SHANK3^58–60, SHANK2⁶¹, and NBEA⁶², have been identified in autistic individuals. Interestingly, this DMR is also located within an imprinted region of the genome. Imprinting ⁶³ and parent-of-origin specific effects^{15, 16} have been previously implicated in ASD.

The third DMR is located about 3.5kb upstream of ZFP57 and overlaps the 5′ end of an alternatively spliced EST. ZFP57 is instrumental in maintaining imprinting marks during development^{64, 65} by providing a mechanism for targeting DNA methyltransferase⁶⁶, responsible for transferring methyl groups to cytosines, to specific locations in the genome⁶⁷. This is directly relevant to ASD since other methylation machinery proteins are mutated in neurodevelopmental syndromes associated with ASD, e.g. Rett Syndrome¹⁷.

Lastly, in cerebellar tissue we identified a DMR at an alternative promoter for SDHAP3, which is associated with a non-coding RNA and a small coding RNA. This DMR falls directly on a CTCF binding site and an active regulatory element site identified by the ENCODE project⁶⁸. Thus, it is likely to be an important regulatory site; however, currently there is no literature describing the function of this particular gene. SDHAP3 is a member of the succinate dehydrogenase gene family that includes genes that are critical components of the metabolic machinery. Mitochondrial respiratory chain complex protein dysfunction has been previously associated with ASD^69–71.

For one of the DMRs, near ZPF57, we identified a striking difference in methylation between autistic and control females in cerebellar tissue. We performed a BLAST search of the DNA sequence for this DMR and did not find alignment to any other regions of the genome, including sex-chromosomes. While intriguing, we examined a relatively small number of female samples in our analyses and the biological significance of sexually dimorphic DNAm differences in the brain is unclear at this time. This result suggests future epigenetic studies may benefit from inclusion of females and additional sex-specific analyses.

While this study is the first to identify commonly altered DNAm in brain tissue from autistic individuals, we recognize there are several limitations that need to be carefully considered. First, from an epidemiological and statistical perspective we examined a relatively small number of samples, 40 in total. It is difficult to ascertain a large number of reliably phenotyped postmortem brain tissues from autistic individuals given the scarcity of the resource³². This is a problem the field of ASD faces, generally. While this study focused on brain tissue, due to disease relevance and a lack of DNAm data in previous autism brain studies, there is also a need for complementary autism epigenomic studies in peripheral tissues, such as blood. Blood-based samples are particularly useful in ASD because large numbers of samples can be collected and because useful biomarkers for disease are needed and most likely will come from this tissue, as opposed to brain tissue, for practical reasons. Second, the DMR we identified at SDHAP3 in cerebellum showed a DNAm change in several autistic individuals that was potentially also related to a CNV. The 450K array is now the primary platform utilized for large epidemiology studies and this result highlights the need to examine and recognize the potential contribution of CNVs to methylation signals in these large studies. In addition, this finding demonstrates the need for new studies to clarify the relationships between CNVs and DNA methylation.

Despite these limitations, the findings presented in this study are of significance to the field of ASD for several reasons. To our knowledge, this is the first study to examine DNAm in ASD brains at genome-scale in regions outside of CpG islands and promoters, identifying 4 DMRs associated with the disorder. Second, while traditional genetic studies have identified rare variants in a minority of ASD cases, here, we identified genomic regions that commonly show differential methylation between autistic and control individuals. Lastly, the DMRs themselves are useful candidate regions for follow up studies.