Decoding the non-coding genome: elucidating genetic risk outside the coding genome

Abstract

Current evidence emerging from genome-wide association studies indicates that the genetic underpinnings of complex traits are likely attributable to genetic variation that changes gene expression, rather than (or in combination with) variation that changes protein-coding sequences. This is particularly compelling with respect to psychiatric disorders, as genetic changes in regulatory regions may result in differential transcriptional responses to developmental cues and environmental/psychosocial stressors. Until recently, however, the link between transcriptional regulation and psychiatric genetic risk has been understudied. Multiple obstacles have contributed to the paucity of research in this area, including challenges in identifying the positions of remote (distal from the promoter) regulatory elements (e.g. enhancers) and their target genes and the under-representation of neural cell types and brain tissues in epigenome projects – the availability of high-quality brain tissues for epigenetic and transcriptome profiling, particularly for the adolescent and developing brain, has been limited. Further challenges have arisen in the prediction and testing of the functional impact of DNA variation with respect to multiple aspects of transcriptional control, including regulatory-element interaction (e.g. between enhancers and promoters), transcription factor binding and DNA methylation. Further, the brain has uncommon DNA-methylation marks with unique genomic distributions not found in other tissues – current evidence suggests the involvement of non-CG methylation and 5-hydroxymethylation in neurodevelopmental processes but much remains unknown. We review here knowledge gaps as well as both technological and resource obstacles that will need to be overcome in order to elucidate the involvement of brain-relevant gene-regulatory variants in genetic risk for psychiatric disorders.

Changes in gene expression are now acknowledged as a major contributor to the genetic risk for complex genetic traits. In this article, we review the key evidence supporting this, as well as mechanisms by which DNA variation influences gene expression and methodologies for studying these relationships. We note that a comprehensive review of the genetic findings for psychiatric disorders is beyond the scope of this review, as is a comprehensive review of epigenetic regulation. Instead, we focus this review on classes of transcripts and epigenetic marks that are prevalent in brain and that may make unique contributions to brain function. We also focus on methods that, in our opinion, are likely to make the greatest contributions to moving the field forward. Finally, we outline a number of obstacles that currently stand in the way of fully elucidating the relationships between gene expression and DNA variation underlying psychiatric disorders.

Transcriptional regulation of gene expression in genetic disease

Gene-regulatory elements controlling gene transcription include the promoter, enhancers, silencers, locus control regions and insulators. With the exception of the promoter that is located in the immediate vicinity of the transcription start site of the gene (Butler & Kadonaga 2002), the other elements can be located anywhere within a gene or even up to megabases away (Dean 2006; Nobrega et al. 2003; Spilianakis & Flavell 2004; Steidl et al. 2007), often within neighboring genes (Lettice et al. 2003). Genetic variation altering gene expression is a documented cause of genetic disease, and until recently, the majority of risk alleles studied were in the promoter, as this region is easily defined. However, genetic changes (duplications, deletions, single nucleotide changes and translocations) in remote regulatory elements – i.e. outside of the proximal promoter – resulting in altered gene expression have also been documented as disease mechanisms, both in Mendelian disorders (e.g. thalassemias, X-linked deafness and facioscapulohumeral muscular dystrophy) and in complex genetic disorders (e.g. cancer, diabetes, rheumatoid arthritis and systemic lupus erythematosus) (Dathe et al. 2009; De Gobbi et al. 2006; Driscoll et al. 1989; Furniss et al. 2008; Gabellini et al. 2002; Hatton et al. 1990; Lettice et al. 2003; Loots et al. 2005; Naranjo et al. 2010; Prokunina et al. 2002; Sharpe et al. 1992; Steidl et al. 2007; Sun et al. 2008; Tokuhiro et al. 2003).

Recent results from genome-wide association studies (GWAS) provide further evidence that variation in gene expression contributes to genetic risk. Genome-wide association studies have recently been completed for a number of psychiatric disorders, including bipolar disorder (Ferreira et al. 2008; Green et al. 2013; Muhleisen et al. 2014; Psychiatric GWAS Consortium Bipolar Disorder Working Group 2011), attention-deficit hyperactivity disorder (ADHD) (Elia et al. 2011; Franke et al. 2009; Hinney et al. 2011; Lasky-Su et al. 2008, 2010; Lesch et al. 2008; Neale et al. 2008, 2010a; Stergiakouli et al. 2012), schizophrenia (Hamshere et al. 2013; O’Donovan et al. 2008, 2009; Schizophrenia Psychiatric Genome-Wide Association Study (GWAS) Consortium 2011; Schizophrenia Working Group of the Psychiatric Genomics Consortium 2014), autism (Anney et al. 2012; Casey et al. 2012), major depression (Major Depressive Disorder Working Group of the Psychiatric GWAS Consortium1 2013; The Psychiatric GWAS Consortium Steering Committee 2009), Tourette syndrome (Scharf et al. 2013) and obsessive-compulsive disorder (OCD) (Mattheisen et al. 2015; Stewart et al. 2013), as well as for adverse effects of therapeutic agents, including antipsychotic-induced weight gain (Malhotra et al. 2012). While the majority of the early studies did not meet the significance threshold for genome-wide evidence of association, studies with larger sample sizes are now providing such evidence, with some of the findings replicating across samples and some showing evidence of shared signals across disorders (Cross-Disorder Group of the Psychiatric Genomics Consortium et al. 2013).

Compared with GWAS analyses of other complex traits that have examined over 100 000 cases, investigations of psychiatric disease using GWAS are in their infancy. This is especially true for the majority of childhood-onset psychiatric disorders, for which GWAS sample sizes have been relatively small, numbering only a few thousand subjects (Anney et al. 2012; Neale et al. 2010b; Scharf et al. 2013). Nevertheless, with GWAS sample sizes for some psychiatric disorders now reaching the necessary numbers, replicated GWAS signals have emerged. Of note, the recent GWAS of 36 989 schizophrenia cases identified 108 GWAS-significant regions (Schizophrenia Working Group of the Psychiatric Genomics Consortium 2014).

Moving forward to identify the functional DNA variants (risk alleles) tagged by disorder-associated GWAS signals is now the focus of intense investigation, as we enter what has been coined as the post-GWAS era. In many instances, no change in the protein-coding region could be identified within the associated gene; in other instances, coding region changes were identified but could not fully explain the association (Dwyer et al. 2010; Schizophrenia Working Group of the Psychiatric Genomics Consortium 2014). Some associated markers are located at large distances (>10–100 kb) from an obvious candidate gene (Malhotra et al. 2012; Wang et al. 2010); others are located in gene ‘deserts’, with no identifiable genes within kilobases or even megabases of the positive markers (Franke et al. 2009; Moraes et al. 2012; Stewart et al. 2013; Zandi et al. 2007). Of note, only 11 of the 108 schizophrenia loci identified in the recent GWAS could be explained by a coding region single nucleotide polymorphism (SNP) (Schizophrenia Working Group of the Psychiatric Genomics Consortium 2014). Further, direct screening of a number of key, replicated candidate genes for psychiatric disorders could not identify coding region changes that would account for the association (e.g. ZNF804A; Dwyer et al. 2010). It appears likely that the vast majority of risk variants underlying psychiatric disorders will turn out to be located in non-coding regions of the genome.

That the landscape of genetic risk for psychiatric disorders will be largely non-coding aligns with overwhelming evidence that changes in gene-regulatory regions are a major contributor to risk for complex genetic traits: The majority (~93%) of disease-associated markers emerging from GWAS findings lie within the non-coding genome (Andersson et al. 2014; Farh et al. 2014; Maurano et al. 2012). The GWAS signals are enriched for expression quantitative trait loci (eQTL), i.e. loci that correlate with gene expression (Ellinghaus et al. 2012; Franke et al. 2010; Khor et al. 2011; Nicolae et al. 2010; Ramasamy et al. 2014; Zhong et al. 2010), and are enriched for markers located within gene-regulatory regions (Farh et al. 2014; International Genetics of Ankylosing Spondylitis Consortium et al. 2013; Maurano et al. 2012; Ricano-Ponce & Wijmenga 2013; Roadmap Epigenomics Consortium et al. 2015; Schaub et al. 2012).

Enhancers and genetic risk

Among the various types of regulatory elements in the genome, GWAS results for the majority of complex traits point increasingly to variants in one type in particular – namely, enhancers (Andersson et al. 2014; Ernst et al. 2011; Farh et al. 2014; Hnisz et al. 2013; Ward & Kellis 2012). Enhancers can reside anywhere within a gene and can even be millions of nucleotides distant from the genes they regulate, i.e. their target genes (Leveille et al. 2015). Enhancer activation, resulting in stimulation of target-gene transcription, involves the binding of transcription factors (TFs; proteins specialized for transcriptional regulation) to the enhancer sequence, modifying the epigenetic environment, with the enhancer looping to physically contact the target-gene promoter. Importantly, enhancers are the major genomic determinant of cell-type/tissue–specific gene expression (Jin et al. 2011). Further, DNA variants associated with a specific disease are more frequently found in enhancers specific to cell types relevant to that disease (Gerasimova et al. 2013), a finding that supports their contribution to disease risk.

Another important property of enhancers is that their target genes are typically not the nearest gene (only 27–40% target the nearest gene) (Andersson et al. 2014; Sanyal et al. 2012). This could explain GWAS signals in regions where the nearest gene(s) in the associated region are not likely candidates for involvement in the trait or disorder being investigated. Further, a single enhancer can regulate multiple genes (on average, 2.5) (Andersson et al. 2014; Sanyal et al. 2012). Thus, a genetic variant in an enhancer could have a profound impact on disease risk by dysregulating more than one gene (Verlaan et al. 2009). The identification of such multiple-gene targets will be critical to the subsequent modeling of disease mechanisms, as knockout or knockdown models of a single gene may not recapitulate the disease phenotype.

Despite the wealth of evidence that many GWAS signals could be pointing to remote regulatory elements in general, and enhancers in particular, a number of obstacles have stood in the way of easily moving forward to pinpoint the enhancers involved and identify the risk alleles. Systematic and comprehensive functional screening of the typically large genomic regions flagged by GWAS signals is prohibitively labor-intensive using standard reporter assays. On the other hand, functional testing approaches that assay only a randomly selected subset of DNA fragments spanning a large region of interest may miss the sought-after element(s). Multispecies comparisons have been used successfully to identify regulatory regions on the basis of sequence conservation, but principally for a restricted subset of genes (those expressed early in development with strict constraints on temporal and spatial expression) (Boffelli et al. 2004; Nobrega et al. 2003; Pennacchio et al. 2006); for other genes with only short stretches of conserved sequence, this approach may miss key regions. Further, conservation of regulatory region function can be maintained even without conservation of sequence, due to compensatory changes at other sites (Ludwig et al. 2000), and conversely, conservation of sequence may be maintained without conservation of function (Shen et al. 2012). DNA sequence differences between species may also result in new functions conferring evolutionary advantages (Dermitzakis & Clark 2002; Rockman et al. 2003) – including sequence variants that are human-specific (King & Wilson 1975; Prabhakar et al. 2006). For example, comparison of known functional elements in humans indicated that only 60–68% are functional in rodents (Dermitzakis & Clark 2002). In addition, non-conserved elements may be relevant to genes contributing to human higher order complex cognitive traits (e.g. reading and language ability, verbal working memory and executive function).

Mapping gene regulatory elements

Histone modification, movement and/or replacement play a dynamic role in gene regulation (Eberharter & Becker 2002; Higgs et al. 2006; Peters & Schubeler 2005), and specific histone modifications (e.g. H3 and H4 acetylation, methylation) mark regulatory elements (Bernstein et al. 2005; Schubeler et al. 2004). This property of histones can be leveraged to locate regulatory regions, focusing the search for risk alleles within these regions (Couto et al. 2010; Darabi et al. 2015; Elbert et al. 2011; Heintzman et al. 2007; Ni et al. 2008). By chemically cross-linking DNA to histones, antibodies specific for histone modifications can be used to capture the modified-histone–associated DNA regions; coupling this chromatin immunoprecipitation (ChIP) approach with DNA identification methods, such as genomic tiling arrays (ChIP-chip) (Ren et al. 2000) or next-generation sequencing (ChIP-seq) (Barski et al. 2007; Johnson et al. 2007; Mikkelsen et al. 2007; Robertson et al. 2007), flags modified chromatin across genomes, creating genome-wide maps of regulatory regions. Comparative analyses found that the locations of sites marked by methylated histones are conserved between mouse and human genomes even though most of the genomic sequences at those sites are not (Bernstein et al. 2005). Thus, modified chromatin identifies regulatory regions where function, but not necessarily sequence, is conserved.

The ChIP-seq data, most notably from the ENCODE project and Epigenomic Roadmap (Bernstein et al. 2010, 2012; Kim et al. 2007; Roadmap Epigenomics Consortium et al. 2015), have recently resulted in an explosion of information on the genomic positions of modified histones and other epigenomic features, as markers for regulatory elements. Much of this information is now publicly available on websites such as UCSC (https://genome.ucsc.edu/cgi-bin/hgGateway) and through the Epigenome Road-map (http://www.roadmapepigenomics.org).

The available ChIP-seq maps of regulatory regions allow us to begin studies of these regions in relation to disease; even so, the maps are – for the study of psychiatric disease – still quite limited in two major respects. First, as noted earlier, there is considerable tissue/cell-type specificity inherent in transcriptional regulation. Thus, the applicability of the ChIP-seq information obtained is limited by the cell/tissue types examined, and unfortunately, little information is available on cell types relevant to the nervous system. Early on, there was little data in ENCODE for human neural cell types or human brain tissue, and this tissue is still underrepresented, given that brain tissue has high levels of gene expression with the majority of annotated genes expressed in the brain (Datson et al. 2001; Miller et al. 2014). For example, on the basis of microarray data, it was estimated that 84% of Reference Sequence (RefSeq) genes are expressed in adult neocortex (Hawrylycz et al. 2012) and 95% in the developing neocortex (Miller et al. 2014). Second, brain-relevant maps that have been produced are far from complete. The Epigenomic Roadmap has mapped histone modifications (both acetylations and methylations) in brain tissues from only one adult individual (eight brain regions) and from fetal brain at only a few gestational time points (Roadmap Epige-nomics Consortium et al. 2015). Until recently, the majority of histone-modification data available were for H3K4me3, a mark of promoters, with little information on remote regulatory regions. However, the number of epigenomic marks has now expanded to include six additional histone marks, including those that mark enhancers (Roadmap Epigenomics Consortium et al. 2015), a start to providing more information on the epigenome of brain.

Functional studies of enhancers

Impact of DNA variants on enhancer function

The mapping of gene-regulatory regions allows us to correlate the positions of genetic association signals with the positions of regulatory elements. However, the data do not provide information on the actual function of specific alleles within these elements. One cannot assume simply on the basis of location that a SNP within a regulatory region changes gene expression – functional studies will still be required. Here again, the cell/tissue-type specificity of transcriptional regulation is an important consideration. Function can only be tested reliably in cells/tissues that have the ‘correct’ components/configuration of transcriptional machinery. High-throughput functional screening of putative regulatory regions can be performed using animal models that test expression in all tissues (Pennacchio et al. 2006), but this is expensive. Further, regulatory elements functional in humans may not be functional in rodents (Dermitzakis & Clark 2002). Thus, functional studies with appropriate disease-relevant human cell types are required to show the impact of DNA variation on transcriptional regulatory function.

High-throughput enhancer testing

Standard methods using in vitro transfection assays can be used to test individual regulatory regions and the impact of DNA variation on transcription. However, even when the locale has been narrowed down (through overlay of the ChIP-seq and GWAS maps), both the construction of vectors and the assays themselves are time-consuming and tedious to perform for each individual gene/regulatory region/DNA-variant combination. A recent breakthrough in molecular biology has been the development of high-throughput methods to screen for enhancer function using next-generation sequencing (Arnold et al. 2013; Melnikov et al. 2012; Patwardhan et al. 2012; Vanhille et al. 2015). These protocols make use of ‘bar-coded’ transfection vectors that allow for the transfection of cells with hundreds to thousands of enhancers in separate reporter vectors – the different bar codes are used to tag the individual enhancers and quantify the expression levels generated from each of the corresponding vectors.

Super-enhancers in genetic disease

Despite the methodological breakthrough described above, there is a recently identified class of enhancers, termed stretch or super-enhancers (Hnisz et al. 2013; Parker et al. 2013; Whyte et al. 2013), that poses additional challenges for study. Super-enhancers are clusters of enhancers larger (>3 kb) than typical enhancers, some spanning as much as 50 kb (Whyte et al. 2013), and they are associated with key genes that control cell identity. Disease-associated alleles are enriched in super-enhancers (Hnisz et al. 2013) and tend to occur in the super-enhancers of disease-relevant cell types, more so than in the typical enhancers (Hnisz et al. 2013). Unfortunately, however, unlike typical enhancers, the super-enhancers cannot be tested for function in their entirety – generally, they are too big to be efficiently cloned into traditional test vectors and transfected into cells. Given their key role in cell identity and their association with risk alleles, alternative methods to test for functional effects of genetic variation in super-enhancers will be required such as genome editing methods (CRISPR/Cas9) (Charpentier & Doudna 2013; Jinek et al. 2012; Ran et al. 2013). Genome editing is a powerful tool for the study of putative functional variants; however, the availability of neural cell types for these studies will be a limitation for psychiatric disease.

DNA variation and TF binding

The transcriptional activity brought about by enhancers depends, fundamentally, on the concerted activities of TFs –DNA-binding proteins that orchestrate the appropriate selection and activation of cell-type–specific and/or signal-dependent enhancers to spawn new enhancer–promoter interaction loops or modulate existing ones, ultimately stimulating target-gene transcription (Heinz et al. 2015). Bioinformatics approaches can be used to predict the binding of a particular TF to a particular DNA sequence (Whitaker et al. 2015) and to predict the impact of DNA variation on that binding, but because of the short length and degenerate nature of consensus TF-binding sequences, such predictions are not precise (Schug 2008; Weirauch et al. 2013; Whitaker et al. 2015). Prediction methods that take the three-dimensional structures of the DNA motifs into account (in addition to their simple linear sequences) may help to improve the situation, but such methods have only begun to emerge and it remains to be seen whether they will be useful in predicting the impact of DNA variation on TF binding (Gao & Ruan 2015; Maienschein-Cline et al. 2012; Zhou et al. 2015).

A further complication when it comes to predicting the impact of motif variation on TF-binding activity is that TFs bind collaboratively, and the larger genomic context influences the probability of binding (Heinz et al. 2013, 2015). The presence of multiple homotypic TF-binding sites can buffer the loss of a single site (Kilpinen et al. 2013). Further, there can be interdependency of collaborating heterotypic sites: direct loss of TF binding (due to sequence variation) at one site can indirectly abolish binding of another TF to its site, without any direct change of sequence there (Heinz et al. 2013).

The limitations described above can be overcome, to some extent, using methods that assay TF binding directly. For example, the in vitro method of electrophoretic mobility shift assay (EMSA) has been used for decades; it is, however, cumbersome and limited by low throughput. Addressing this problem, more-recent, high-throughput in vitro methods that assay TF binding to immobilized oligonucleotides have been developed (e.g. using ELISA-based, bead-based and protein-binding microarray technologies) to automate screening for hundreds of DNA regions simultaneously (Berger & Bulyk 2009; Bhattacharya et al. 2010; Chorley et al. 2008; Weirauch et al. 2013). Further, the in vivo methods of ChIP – using TF-specific antibody – can be used to capture and map the genomic sequences to which a specified TF is bound in the cell type or tissue type of interest. Some of this in vivo TF-binding information is available from the ENCODE and Epigenomic Roadmap projects. However, because of the issue of cell-type specificity described earlier and because of the possibility that the genotype(s) of the samples used may be influencing TF binding at some site(s), evidence of non-binding in TF-ChIP is not necessarily conclusive. Even when efforts to overcome these shortcomings are implemented, however (e.g. using appropriate cell/tissue types and genotypes), there may remain a technical obstacle: the availability of ChIP-grade TF-specific antibodies is currently limited.

Despite the drawbacks encountered in ascertaining the impacts of DNA variation on TF binding, the need to further develop and pursue this line of research in relation to psychiatric disease cannot be overstated – identifying disease-relevant transcriptional networks predicated on the coordinate actions of TFs at their cognate DNA binding sites will likely be the key to ultimately designing more-specific and effective therapeutic interventions. Highlighting this scenario is the recent example where the analysis of motifs predicted to be changed by GWAS alleles in regulatory regions across diverse autoimmune and inflammatory disorders pointed to a network of TFs in the JAK/STAT-mediated type I interferon response (Maurano et al. 2012); JAK inhibitors are currently showing promise in clinical trials for immune-mediated disorders (Garber 2011; Sandborn et al. 2012; Tanaka et al. 2012). Information regarding the networks of TFs and second messengers regulating risk genes may indicate direct targets for interventions. Thus, we will need to understand the TF networks that regulate risk genes underlying psychiatric disorders and how risk alleles perturb these networks.

Analysis of the transcriptome

Expression quantitative trait locus mapping

Within linkage disequilibrium (LD) regions, post-GWAS searches for functional risk alleles can be expedited by leveraging eQTL data. As noted earlier, eQTL are genetic markers that correlate with gene expression – thus presumably tagging functional variants altering gene expression. The production of genome-wide eQTL maps follows the same principle as GWAS, except that rather than assessing panels of genome-wide markers for associations with a single trait or disease/disorder, the markers are assessed for associations with the expression levels of RNA species in the transcriptomes of cells/tissues. Thus, eQTL markers identified in a disease-relevant cell/tissue type that are found in high LD with disease-associated markers from a GWAS can be extremely useful – they can help to focus on functional (gene-regulatory) risk alleles for the disease in question and can help to shed light on the identity of the target-gene ‘culprits’ whose dysregulation is causal to the disease (Darabi et al. 2015; Ramasamy et al. 2014; Zou et al. 2012).

Recently, there has been a great deal of progress in the availability of public databases containing transcriptome and genotype information (eQTL databases) (Colantuoni et al. 2011; Kang et al. 2011; Mele et al. 2015; Ramasamy et al. 2014; Zou et al. 2012); however, these resources are currently limited in terms of either the number of individuals represented or the developmental periods or brain regions covered. Early resources used microarrays [BrainCloud (Colantuoni et al. 2011), Brain eQTL Almanac (http://ukbec.wordpress.com/) (Ramasamy et al. 2014; Zou et al. 2012)]; thus, the data were limited to annotated genes and, as a result, were missing unannotated exons and transcripts, particularly non-coding transcripts [e.g. long non-coding RNAs, small RNAs including microRNAs (miRNAs)] (Batista & Chang 2013; De Santa et al. 2010; Derrien et al. 2012; Kocerha et al. 2015; Orom & Shiekhattar 2013; Orom et al. 2010; van de Leemput et al. 2014).

Non-coding RNA

Over the past several years, it has become increasingly apparent that the genome produces an abundance of non-coding transcripts (De Santa et al. 2010; Derrien et al. 2012; Kim et al. 2010; Mele et al. 2015; van de Leemput et al. 2014): non-coding RNAs now outnumber annotated coding RNAs (Derrien et al. 2012). Non-coding RNAs have been documented to downregulate (Batista & Chang 2013; Kocerha et al. 2015) as well as upregulate gene expression (Batista & Chang 2013; De Santa et al. 2010; Lai et al. 2013; Orom & Shiekhattar 2013), with some species recently shown to function in an ‘enhancer-like’ fashion (Lai et al. 2013; Orom & Shiekhattar 2013; Orom et al. 2010). Non-coding RNAs play key roles in neurogenesis, neurodevelopment and activity-dependent brain plasticity (Barry 2014; Kocerha et al. 2015; Lipovich et al. 2012). Further, specific miRNA genes have been documented as genetic risk factors in psychiatric disorders (Kocerha et al. 2015) and others will likely be implicated as genetic studies progress.

Despite these important findings, a relative paucity of detailed information on the non-coding transcriptome in the brain continues to hinder our ability to interpret GWAS signals in psychiatric disorders. In some instances, genetic association signals may be pointing to an unannotated non-coding transcript or to a regulatory region controlling the expression of a non-coding transcript locus as opposed to a coding gene (Farh et al. 2014). Recent eQTL resources such as the GTEx project (GTEx Consortium 2013, 2015; Mele et al. 2015), BrainSpan (http://brain-map.org/) and other databases (Jaffe et al. 2015) are beginning to make some headway in this area by using RNA-seq data. However, the RNA libraries for some of these were polyadenylation-selected and non-polyadenylated transcripts are not represented (Derrien et al. 2012; Kim et al. 2010). It is not clear how this omission will impact our understanding of gene regulation in brain, but non-polyadenylated transcripts are abundant in brain. One category of RNA, the long non-coding RNAs (lncRNAs), are enriched for non-polyadenylated transcripts compared with mRNA (Derrien et al. 2012), and the highest number of tissue-specific lncRNAs are found in brain (Derrien et al. 2012; Francescatto et al. 2014). Further, there is a typically non-polyadenylated species of RNA transcribed at enhancers (eRNAs) (Kim et al. 2010; Koch et al. 2011) and these eRNAs themselves function as key transcriptional regulators (Leveille et al. 2015).

A further complication when it comes to the detection of non-coding RNA species in RNA-seq datasets is that many non-coding RNAs exhibit lower expression (~10-fold lower for lncRNAs) and more tissue specificity than the mRNAs of coding genes (Cabili et al. 2011; Djebali et al. 2012). Also, some are expressed only transiently (Cabili et al. 2011). Thus, annotating these transcripts will require combining deep sequencing with the extensive brain regional representation that was achieved previously using microarrays (Hawrylycz et al. 2012). Current RNA-seq methods require libraries to be constructed differently for long and short transcripts (e.g. microRNAs) and this need for two separate libraries increases the sequencing expense. Because of costs, it has been argued that it currently may be impractical to obtain full annotation of rare transcripts; more-sensitive and/or less-expensive sequencing methods will be required.

Despite the advances afforded by SNP–expression (eQTL) datasets in narrowing the field of candidates in post-GWAS searches for functional risk alleles (Darabi et al. 2015; Ramasamy et al. 2014), these studies still cannot distinguish the functional alleles from those in strong LD with them. That is, the datasets identify correlations between particular genetic variants and the expression of particular transcripts, but do not identify the actual functional allele(s) among all the DNA variants in strong LD. Given that LD can span megabases from the associated allele(s) identified in a GWAS, many DNA variants may lie within the region of LD and correlate with expression. Functional studies will still be required to distinguish – among these correlated markers – the risk allele(s) that are actually influencing expression and that, in turn, are likely to be causal in the disease/disorder. The most illuminating studies will be those that not only identify eQTL located in regulatory elements but also shed light on the relevant target gene(s) – even when the target lies outside the LD region.

Direct determination of enhancer targets through looping analyses

Enhancers work by looping to physically contact the promoter(s) of their target gene(s) and stimulate transcription. Chromosome conformation capture (3C) methods (Miele & Dekker 2008; Ni et al. 2008; Sajan & Hawkins 2012) are looping assays that detect these physical interactions between enhancers and their target(s) and thus can be used to identify the disease-relevant gene(s) influenced by risk alleles within the enhancers.

In 3C studies, cells are first fixed to stabilize – by cross-linking – the physical contact points between regulatory elements within the nucleus. The cells are then lysed and the chromatin digested with a restriction enzyme, leaving the ends of what were once linearly distant but conformationally close DNA sequences available for ligation to one another; religation under dilute conditions promotes the desired ligation events, positioning the originally distant DNA sequences adjacent to each another in the ligation products (Lieberman-Aiden et al. 2009; Miele & Dekker 2008; Rao et al. 2014; Shen et al. 2012). The cross-links on the chromatin are then reversed, and the ligation products analyzed either by quantitative PCR or by DNA sequencing (depending on the application) (Dixon et al. 2012; Hughes et al. 2014; Rao et al. 2014).

In its most basic form, 3C is combined with quantitative PCR for targeted interrogation of individual prespecified candidate promoter–enhancer interactions, and is the least expensive method available (Abou El Hassan & Bremner 2009; Miele & Dekker 2008). However, this method requires a priori prediction as to which promoters and enhancers pair, and hence, which DNA sequence pairs are most likely to interact. Given that promoter targets may be hundreds of kilobases away from the enhancer, even outside the associated LD region found in a GWAS, either all known promoters within a large area must be tested or the promoter targets must be predicted. Thus, this method is limited. Nevertheless, the recently released FANTOM5 dataset provides some basis for predictions (Andersson et al. 2014). Leveraging the observation that active enhancers are, typically, themselves transcribed, an analysis of the FANTOM5 dataset used pairwise expression correlations of the transcriptional activity of putative enhancers with the transcriptional activity of transcription start sites to predict the targets of enhancers across a diverse set of cell types.

Given the current numbers of GWAS signals for psychiatric disorders and the expected increase in these numbers as sample sizes increase, high-throughput genomic methods for the parallel interrogation of multiple potential promoter–enhancer interactions are required. The state of the art is Hi-C, a method that combines 3C with deep sequencing of all points of chromatin interaction, genome-wide (Dixon et al. 2012; Lieberman-Aiden et al. 2009). The resolution to identify interactions is dependent on sequencing depth. Unfortunately, because the costs of reaching the sequencing depths required for whole-genome coverage are currently high, the typical resolution achievable does not allow for the identification of specific promoter–enhancer interactions for most datasets (Zuin et al. 2014). However, the results can provide information on chromatin domains that can be used to estimate the extent of potential interactions (Dixon et al. 2012).

Bridging the gap between the limited ‘one-to-one’ scope of basic 3C and the limited resolution of the whole-genome (‘all-to-all’) Hi-C alternative, there are a number of other 3C-related methods that combine next-generation sequencing with targeted investigation of elements or regions of interest. These include methods such as 5C, a ‘many-to-many’ approach that identifies all interactions occurring within a particular region (as opposed to genome-wide as in Hi-C) (Dostie et al. 2006), ChIA-PET, a hybrid method that incorporates ChIP to interrogate the subset of all chromatin looping points that are bound by a protein of interest (Fullwood et al. 2009; Heidari et al. 2014), and 4C, a ‘one-to-all’ approach that investigates all (genome-wide) interactions with one particular element (e.g. an enhancer or a promoter) – the element of interest is designated the bait or viewpoint (Simonis et al. 2006; Splinter et al. 2012; van de Werken et al. 2012a,b; Zhao et al. 2006). Among these methods, the 4C strategy is particularly well-suited to the task of finding the target(s) of an enhancer that has been singled out for post-GWAS investigation. There are also some adaptations of 4C that can be further advantageous. For example, enhanced 4C (e4C) includes a bait/viewpoint-enrichment step (using biotin-coupled bait-specific primer extension) that results in at least 100-fold enrichment of bait-sequence representation in the library – improving the sensitivity to detect weaker (yet important) interactions that might otherwise go undetected (Sexton et al. 2012). Other adaptations, Capture-C and Capture Hi-C (CHi-C), use a hybridization capture approach to enrich for many different bait sequences (viewpoints) simultaneously. These multiplex enrichment strategies followed by high-throughput sequencing allow for interrogation of the interaction points of hundreds to many thousands of different bait elements (e.g. enhancers or promoters) simultaneously (Hughes et al. 2014; Mifsud et al. 2015).

Ultimately, one must show that the putative risk alleles identified in gene-regulatory elements alter such interactions – either destroy them or create new interactions (Visser et al. 2012). When appropriate cell types are available, specific nucleotides can be altered using genome editing (Charpentier & Doudna 2013; Jinek et al. 2012) and the impact on these interactions observed. This approach will not be possible, however, when the interactions are found only in human brain tissues. In that case, a large multi-individual panel of brain tissues screened by genotype would be necessary to correlate the presence of the putative risk allele with altered interactions. Such studies would, however, require careful attention to tissue quality – particularly post-mortem (or post-surgical) intervals, as some interactions may be prone to degradation (Mitchell et al. 2014).

DNA variation and methylation

DNA methylation is known to play a critical role in brain function and behavior, and deregulated methylation results in disease (Davies et al. 2007; Labrie et al. 2012; Lyst & Bird 2015; Shin et al. 2014); however, the impact of specific DNA variants on methylation and the impact of the resultant methylation changes on gene function have been studied for relatively few loci (Dempster et al. 2013; Labrie et al. 2012). Genetic changes influencing methylation may result in the creation of a repressive or permissive chromatin environment and/or may impact the binding sites of specific TFs (Dantas Machado et al. 2015; Lyst & Bird 2015; Sun et al. 2014), thereby altering transcriptional regulation.

CG methylation

Until recently, the focus of DNA methylation studies was on methylated C (5mC) in CG dinucleotides (mCG). DNA variation may directly, through removal or addition of a cytosine nucleotide, or indirectly, through removal or addition of a guanine nucleotide, cause a loss or gain of methylation that could alter gene expression. Further, nucleotides surrounding the CG may also influence the probability that the cytosine is methylated. Allele-specific methylation analysis in mouse frontal cortex indicated that methylation of CG is biased by the sequence of adjacent nucleotides on either side of the CG, with some motifs preferentially hypermethylated and others hypomethylated (Xie et al. 2012). For example, the motif CTCGCG was found to be hypermethylated, whereas the motif CACGTG was found to be hypomethylated. Analysis of human methylomes also showed bias in the methylation of these motifs (Xie et al. 2012), indicating cross-species conservation of sequence preference of the methylation machinery; DNA variation of these motifs may affect methylation and impact gene expression.

Non-CG methylation

Non-CG methylation was first reported in mouse frontal cortex (Xie et al. 2012), challenging dogma that in differentiated mammalian tissues, C is only methylated (5mC) in the context of an adjacent G (CG methylation). Up until this discovery, non-CG methylation had been documented only in embryonic stem (ES) cells, oocytes and preimplantation embryos (He & Ecker 2015; Lister et al. 2009). Non-CG methylation disappeared upon differentiation of ES cells and reappeared in the creation of induced pluripotent stem (iPS) cells (Lister et al. 2009), and thus was thought to be a marker of pluripotent cell types. The finding of non-CG methylation in adult brain challenged this view and paved the way to better understanding the role of this epigenetic modification.

Additional studies confirmed the prevalence of non-CG methylation in mouse brain (Guo et al. 2014; Lister et al. 2013) and showed that it is also present in human brain (Guo et al. 2014; Lister et al. 2013). In adult brain tissue, 2.9% of cytosines genome-wide were found in methylated CGs and 1.3–1.5% were found in methylated non-CGs (He & Ecker 2015; Lister et al. 2013). Differentiating between brain cell types, non-CG methylation appears to be prevalent in neurons (~6% of cytosines methylated in the non-CG context) but low in glial cells (<1%) (He & Ecker 2015; Lister et al. 2013; Schultz et al. 2015).

Non-CG and CG methylation differ in their developmental timing: CG methylation is established during fetal development, whereas non-CG methylation is low in fetal brain (Lister et al. 2013; Spiers et al. 2015) and accumulates in neurons in the course of postnatal development, gradually increasing during brain maturation to late adolescence before declining slightly during aging (Guo et al. 2014; Lister et al. 2013). The most rapid neuronal accumulation of non-CG methylation occurs from birth to 2 years of age, with slower accumulation continuing through adolescence till it becomes the dominant form of methylation in adult neurons (Lister et al. 2013). This trajectory may be related to the dramatic rates of gene expression changes seen in frontal cortex during development: the highest rates are in fetal brain, and then during infancy, the rates of changes in gene expression become much slower, slowing further still during childhood (1–10 years of age) to adolescence, after which low rates of change persist into mid-adulthood (Colantuoni et al. 2011).

The genomic distribution of non-CG methylation is non-random and differs from CG methylation, indicating unique functions; there are, for example, large (in the order of megabases) regions that are largely devoid of non-GC methylation but retain mCG (He & Ecker 2015). Elsewhere, non-CG methylation is enriched in repeat elements, inactive enhancers and gene bodies (He & Ecker 2015). Like CG methylation, non-CG methylation in brain and neurons is influenced by surrounding sequence: A preference for methylation of CA motifs (Lister et al. 2013; Mo et al. 2015; Xie et al. 2012) compared with other nucleotides is similar to that seen in oocytes (Tomizawa et al. 2011) and ES cells (Lister et al. 2009). But when motif analysis was expanded to include additional flanking nucleotides, differences in sequence preference between tissues were evident – the motif TACAC is preferentially methylated in neurons and glia, whereas TACAG is preferentially methylated in pluripotent cells (Schultz et al. 2015). This difference suggests that the function of non-CG methylation in neurons may be distinct from that in pluripotent cells, and analysis of gene expression supports this view (He & Ecker 2015). Methylation of CAC across the gene body is negatively correlated with expression in neurons, whereas methylation of CAG across the gene body is positively correlated with expression in pluripotent cells (He & Ecker 2015). Genetic variation in the surrounding motif could influence the probability of methylation, and hence gene expression, in a tissue-specific manner, and possibly the resultant phenotype.

Recently, non-CG methylation was reported in several other differentiated tissues, but at much lower levels than in brain – less than 0.4% of cytosines methylated in the non-CG context (Schultz et al. 2015) compared with 1–2% in frontal cortex (He & Ecker 2015). The mark seems to be in a subpopulation of cells within the tissues, as replicate samples excised from the same tissue differed in their percentage of non-CG methylation, indicating non-homogenous distribution (Schultz et al. 2015). A motif, TNCAC, similar to the TACAC motif methylated in neurons was observed, as well as a NNCAN motif (where N is any nucleotide). The authors speculate that these non-CG methylation patterns may represent adult stem cells present in the tissues (Schultz et al. 2015). Another possibility to consider is that neural cells within these tissues could be contributing to the observed non-CG methylation.

The specific role(s) of non-CG methylation in the brain and whether this epigenetic mark functions in disease are currently unknown (He & Ecker 2015). One intriguing possibility is that non-CG methylation is directly related to Rett syndrome, a neurodevelopmental disorder resulting from deleterious mutations in the gene encoding methyl-CpG-binding protein 2 (MeCP2) (Lyst & Bird 2015). Non-CG methylation is correlated with repressed transcription, and MeCP2, in addition to binding mCG, binds methylated non-CG with both high affinity and a preference for methylated CA (mCA) (Chen et al. 2015; Gabel et al. 2015; Guo et al. 2014). It is speculated that the timing of onset of Rett syndrome symptoms – typically not until 6–18 months after birth – is directly related to the acquisition of increased levels of non-CG methylation in a key subset of neuronal genes that begins after birth in the maturing brain (Chen et al. 2015; Gabel et al. 2015). The genes implicated are typically long (>100 kb) and mCA-rich and appear to be subject to MeCP2-binding–mediated transcriptional repression. In the absence of functional MeCP2, however (i.e. in Rett syndrome), derepression ensues, resulting in the pathological overexpression of these heavily non-CG–methylated genes (Chen et al. 2015; Gabel et al. 2015). In cultured neurons lacking MeCP2, long-gene derepression could be reversed pharmacologically, using a topoisomerase inhibitor, and this rebalancing of gene expression was associated with an improvement in cellular health (Gabel et al. 2015). The authors speculate that the rebalancing of long-gene expression in neurons could attenuate neuronal dysfunction in Rett syndrome and may be a strategy for treatment (Gabel et al. 2015).

5-Hydroxymethylation

First identified in the early 1970s, 5-hydroxymethylation (5-hydroxymethylcytosine, 5hmC) was thought to be an artifact until its presence in brain and ES cells was confirmed in 2009 (Kriaucionis & Heintz 2009; Kriukiene et al. 2012; Tahiliani et al. 2009). It was then thought that 5hmC was a transient intermediary in the demethylation of 5mC to demethylated C (Kinde et al. 2015). However, 5hmC has since been shown to be a bona fide epigenetic mark, with some loci stable across the life span and others dynamic, changing during development (Kinde et al. 2015; Sun et al. 2014; Szulwach et al. 2011). In brain tissue, 5hmC is abundant, accounting for 25–40% of modified cytosines (Hahn et al. 2013; Lister et al. 2013; Munzel et al. 2010; Nestor et al. 2012; Wen & Tang 2014) and increasing during neuronal differentiation (Hahn et al. 2013). In mouse and human brain, 5hmC was found to markedly increase from the early postnatal stage to adulthood, suggesting a role in neurodevelopment (Lister et al. 2013; Song et al. 2011; Szulwach et al. 2011; Wang et al. 2012). The vast majority (97–99.98%) of 5hmC is found in the CG context (Lister et al. 2013; Wen & Tang 2014; Wen et al. 2014a; Yu et al. 2012); however, studies of 5hmC during fear extinction have also pointed to changes in non-CG methylation sites (Li et al. 2014). 5hmC is enriched in enhancers and gene bodies and is positively correlated with gene expression in brain (Wen & Tang 2014). Further, 5hmC is more enriched on the sense strand than on the antisense strand, whereas the antisense strand is more enriched for 5mC (Wen & Tang 2014).

As with non-CG methylation, the role of 5hmC modification in brain is not entirely clear (Kinde et al. 2015; Wen & Tang 2014). This modification has been implicated in multiple functions, including exon splicing, learning and memory, hippocampal neurogenesis and DNA-methylation–related plasticity (Kinde et al. 2015; Li et al. 2014; Santiago et al. 2014; Sun et al. 2014; Wen & Tang 2014). Some of these functions have been indicated indirectly, through manipulation of the ten-eleven translocation (Tet) family of dioxygenases (Kinde et al. 2015; Santiago et al. 2014; Sun et al. 2014; Wen & Tang 2014) – the enzymes responsible for the conversion of 5mC to 5hmC, and also for the conversion of 5mC and/or 5hmC to 5-formylcytosine (5fC) and 5-carboxylcytosine (5caC) (He et al. 2011; Ito et al. 2011). 5-Formylcytosine and 5caC are found at much lower levels in the brain than 5hmC (Liu et al. 2013a), thus the assumption that these functions are likely attributable to 5hmC. However, the potential role of either 5fC or 5caC vs. the specific function of 5hmC in these processes is unknown. 5-Formylcytosine shows a different distribution in brain compared with 5mC and 5hmC (Wagner et al. 2015). Further, 5fC levels are higher in non-neuronal populations than in neurons and 5fC shows a different pattern of age-related distribution, its levels decreasing with age starting at early developmental stages (Wagner et al. 2015). These findings suggest different functions, but further studies are required.

Detection of DNA variation effects on DNA methylation

A number of sequence-based methods can be used to interrogate DNA variation and methylation (Harris et al. 2010; Plongthongkum et al. 2014; Shull et al. 2015; Ulahannan & Greally 2015). The state of the art is whole-genome single-nucleotide–resolution methylation analysis (methylC-seq) (Lister et al. 2009, 2013; Xie et al. 2012), which relies on bisulfite conversion of the DNA. Currently, it is cost-prohibitive to perform methylC-seq for more than a few individuals because of the depth of sequencing required to accurately call methylation events in the bisulfite-treated DNA. Enrichment methods paired with sequencing (without bisulfite conversion), such as methylated DNA immunoprecipitation sequencing (MeDIP-seq), allow genome-wide coverage at relatively low cost, but have lower ability to identify regions with intermediate levels of methylation and cannot distinguish allele-specific methylation (Harris et al. 2010; Plongthongkum et al. 2014; Shull et al. 2015). It is possible, however, to achieve single-nucleotide resolution for detection of allele-specific methylation within the methylation-enriched DNA by subsequently interrogating the captured (methylation-enriched) DNA fragments with bisulfite sequencing (Choi et al. 2010; Shull et al. 2015). Another well-established genome-wide approach with single-nucleotide resolution for subgenome-scale methylation study is reduced representation bisulfite sequencing (RRBS) (Meissner et al. 2005); however, RRBS relies on restriction enzyme sites and is estimated to cover only 8–14% of CpGs (Harris et al. 2010; Shull et al. 2015). In contrast to these genome-wide approaches, targeted bisulfite-sequencing protocols that provide single-nucleotide resolution for methylation analysis of selected regions of the genome have also been developed. Such methods are dependent on targeted oligonucleotide capture either before or after bisulfite conversion of the DNA and can be used to investigate hundreds of individuals for thousands of selected 5mCs (Plongthongkum et al. 2014; Shull et al. 2015). For example, a method that uses padlock probes (Deng et al. 2009; Diep et al. 2012) (annealed to the bisulfite-treated genomic DNA, circularized and amplified by PCR for subsequent high-throughput sequencing) has been optimized such that a genome-scale set of over 330 000 padlock probes can be used in a single-tube reaction to cover >500 000 selected CG sites (Diep et al. 2012) (http://genome-tech.ucsd.edu/public/Gen2_BSPP/); it also allows for the development of additional probes for more in-depth coverage of regions of particular interest. This method is advantageous because DNA variants are called from the sequence and the impact on methylation identified.

The analysis of 5hmC presents further challenges: 5hmC cannot be distinguished from 5mC by standard bisulfite sequencing but can be distinguished by other methods, such as Tet-assisted bisulfite sequencing (TAB-seq) (Yu et al. 2012). 5-Formylcytosine and 5caC also cannot be distinguished using bisulfite sequencing and are too rare to be identified using immunoprecipitation-enrichment techniques. A number of methods for the identification of 5fC have been published including 5fC-selective chemical labeling (fC-Seal) for genome-wide profiling of 5fC and a 5fC chemically assisted bisulfite sequencing (fCAB-seq) method for the base-resolution detection of 5fC (Song et al. 2013).

DNA methylome maps

The quantitation of DNA methylation across the genome, through the types of approaches described above, produces methylome maps that can be further investigated through multiple types of association analysis, including methylome-wide association study (MWAS) and methylation QTL (meQTL) and expression methylation-site (eMS) analyses. The MWAS approach analyzes DNA methylation sites in relation to a complex trait or disorder (Sun 2014), the meQTL approach analyzes the methylation sites in relation to genome-wide genetic markers (Luijk et al. 2015) and the eMS approach analyzes them in relation to genome-wide gene/RNA expression (Liu et al. 2013b). Post-GWAS studies of complex disorders are increasingly integrating these methylomics analyses, not only with the GWAS signals themselves but also with eQTL and regulatory-region (e.g. enhancer) maps, in order to uncover functional risk variants – including ones that may be involved in brain function and/or psychiatric disorders (Clark et al. 2015; Gamazon et al. 2013; Gibbs et al. 2010; Kumar et al. 2015; van Eijk et al. 2015; Wockner et al. 2014; Zhao et al. 2014).

Obstacles to the study of gene regulation in psychiatric disorders

The study of gene regulation involved in psychiatric disorders has unique limitations not encountered in the study of many other genetic disorders. Specifically, the field is limited by the availability of neural cell types that can be cultured as well as by the availability of high-quality post-mortem brain tissues from both patients and controls. These limitations are particularly problematic for the study of sensitive developmental time periods (prenatal, early childhood and adolescence), sex differences and/or for the study of specific genotypes. We review these obstacles below.

Representative neural cell types for molecular studies

A major obstacle in moving genetic association findings for brain disorders forward to understanding altered gene function has been the limited ability to culture neurons and glial cells for molecular studies. Human iPS cells (Ho et al. 2015; Maury et al. 2015; Wright et al. 2014), differentiated human ES cells (Espuny-Camacho et al. 2013; Maury et al. 2015; Nicholas et al. 2013) and olfactory neurosphere-derived cells (Matigian et al. 2010) can be used as a source of neural cell types. Among these, iPS cells from patients are particularly valuable in the examination of cellular phenotypes (Brennand et al. 2011; Chen et al. 2014; Madison et al. 2015; Robicsek et al. 2013; Wang et al. 2014), especially when derived from patients with known risk genotypes (Pedrosa et al. 2011; Wen et al. 2014b; Yoon et al. 2014). Currently, these cellular models can capture a number of neural cell types, but as yet cannot fully capture the diversity of cells found in brain (Ho et al. 2015; Maury et al. 2015; Merkle et al. 2015; Nicholas et al. 2013). There are also multiple sources of variability, such as cell origin, cell-derivation methods, cell-type heterogeneity within the culture and cell-culture conditions, that limit the interpretation of cellular characteristics and outcomes in cell cultures derived from patient samples (Ho et al. 2015; Nityanandam & Baldwin 2015; Sandoe & Eggan 2013). Appropriate representation in terms of neuronal maturity is also an issue (Ho et al. 2015; Nicholas et al. 2013; Sandoe & Eggan 2013). For example, human iPS neural precursor cells and 6-week-old neurons exhibit a gene expression pattern similar to that of first-trimester fetal forebrain (Brennand et al. 2015). These cells could serve as a model for neurodevelopmental disorders that are primarily prenatally influenced but may not model disorders where the critical period is later in life. Because of these limitations, there is a crucial need for the development of methods to create mature neurons (Nityanandam & Baldwin 2015).

Representation of key developmental risk periods

Currently, key developmental risk periods for psychiatric disorders are not well represented in transcriptome and epigenome datasets with the majority of data obtained from adult tissues. This represents a major limitation because a number of the psychiatric disorders are considered neurodevelopmental (Estes et al. 2015; Owen et al. 2011; Weinberger 1987; Zwaigenbaum et al. 2005), with symptoms seen very early in development. A critical time point for some of these may be prenatal. Fetal brain is represented in the Epigenomic Roadmap, but for only a very limited number of gestational ages (http://www.ncbi.nlm.nih.gov/geo/roadmap/epigenomics/?search=fetal+brain&display=50). Gene expression dramatically and continually changes over the course of fetal development (Jaffe et al. 2015; Kang et al. 2011) and such a limited sample set may not capture the totality of the changes (Kang et al. 2011).

The most important developmental time period for the emergence of many psychiatric disorders is late childhood through adolescence (Kessler et al. 2005). Adolescence is a time of dramatic physiological, cognitive and emotional changes and is associated with an increased risk of psychiatric disorders, suicide and aggression among adolescent youth. This increased risk of disturbance is most likely attributable to both psychosocial and hormonal factors – with learning and experience interacting with biologically programmed developmental influences to modify structure and function in the brain (Sisk & Zehr 2005). First documented in post-mortem studies, recent neuroimaging studies confirm major structural changes occurring in the brain (e.g. decreases in gray matter and increases in white matter, myelination, synaptogenesis and synaptic pruning) from late childhood to early adolescence, and functional imaging studies have also revealed differences in regional activation in the brain during this period (Blakemore 2012). Analysis of brain transcriptome data reflects these changes, with adolescent brain showing the second highest number of differentially expressed transcripts after fetal brain (Jaffe et al. 2015).

Maps of regulatory regions functioning in brain during this critical period are not currently represented in the Epigenomic Roadmap (Roadmap Epigenomics Consortium et al. 2015) and are underrepresented in current transcriptome databases (http://help.brain-map.org/download/attachments/3506181/Human_Brain_Seq_Stages.pdf?version=1&modificationDate=1433436980032) (Colantuoni et al. 2011; Jaffe et al. 2015; Kang et al. 2011). This is because of the paucity of brain tissues from these age groups available in brain banks.

Representation of male and female brain

There is a sex bias in the prevalence of most psychiatric disorders (Baron-Cohen et al. 2011; Kigar & Auger 2013; Ngun et al. 2011). Many of the early childhood-onset disorders have a male predominance (e.g. ADHD, Tourette syndrome and autism) (Zahn-Waxler et al. 2008). Depression in children with onset at less than 14 years of age is found in almost equal proportions in boys and girls, with a slight predominance of male cases (Hankin et al. 1998; Kovacs et al. 2003). During midpuberty, at around Tanner Stage III (~age 13–14), this ratio changes, with a dramatic increase in incidence among girls and a shift to female predominance (Angold et al. 1998).

In brain, sex differences are evident by structural neuroimaging as well as by neural activation patterns in functional imaging paradigms (Blakemore 2012; Ingalhalikar et al. 2014). Estrogen is present in the adult, postnatal and, possibly, prenatal brain, influencing brain organization and assisting in the feminization of both the brain and behavior (see Collaer & Hines 1995). Testosterone is present in the fetus as early as 8 weeks of gestation, with a surge reaching almost pubertal levels (Hines 2008). There is another surge after birth, before dropping to low levels at 4–6 months (Forest et al. 1974). Gonadal hormones influence histone modifications as well as DNA methylation, and levels of mRNA for methyltransferases differ in specific regions of the male and female brain (Nugent et al. 2012), likely contributing to differential gene expression. Sex-chromosome–linked genes also display sex-biased gene expression and these may contribute to sex differences in brain (Deng et al. 2014; Loke et al. 2015; Mele et al. 2015; Ngun et al. 2011; Trabzuni et al. 2013; Vawter et al. 2004; Weickert et al. 2009). Notable is the Y chromosome gene SRY (Sex-determining Region on the Y chromosome), the male sex determination and differentiation gene that plays a key role in regulating the catecholaminergic system (Loke et al. 2015; Ngun et al. 2011). A number of autosomal genes exhibit sex-specific or sex-biased gene expression and/or exon usage as borne out in comparative analyses of exon-level transcriptome data from human male and female brain samples taken at various stages – from fetal to adult (Kang et al. 2011; Trabzuni et al. 2013). Thus, epigenetic changes in the male and female developing and adult brain are of extreme interest for the understanding of psychiatric disorders and resources should be increased for more comprehensive study in this area.

Outlook

The GWAS findings clearly point to variants in the non-coding genome as being significant contributors to genetic risk. While large sequencing studies are underway for several of the psychiatric disorders, the interpretation of the findings for variants identified in the non-coding genome will be hampered by our ignorance of the function of these regions and the impact of genetic variants within them. Although resources providing transcriptome and epigenomic maps for brain tissue are increasing, there are still gaps in the developmental stages covered by these resources as well as gaps in the available epigenetic marks and transcriptome coverage. As these resources continue to expand, integrating the multiple layers of ‘omics’ data produced will yield a more complete understanding of the complex relationships that exist among DNA variants and the diversity of epigenetic features involved in transcriptional regulation (Banovich et al. 2014; Baubec & Schubeler 2014; Zhi et al. 2013) – this will require the ongoing development of analytic methods that can handle the increasingly multilayered datasets (Huang 2014; Pineda et al. 2015). Even with this advanced profiling of the epigenetic landscape, however, the links between specific variants and function will require experimental validation. The development of neural cell types from pluripotent cells to closely approximate cellular functions in brain will be critical for these studies.