Abstract
Early-life adversity (ELA), including child abuse and other forms of early-life maltreatment, is a major predictor of negative mental health outcomes. ELA is thought to increase lifetime risk of psychopathology by epigenetically regulating genomic regions that in turn adjust different brain systems. Here, focusing on the lateral amygdala, a major brain site for emotional homeostasis, we comprehensively describe molecular cross-talk across multiple epigenetic mechanisms, including 6 histone marks, DNA methylation and the transcriptome, in subjects with a history of ELA and healthy controls. We first provide evidence for previously unknown interactions among epigenetic layers in the healthy brain. Focusing on non-CG methylation, and particularly on CAC, our results further suggest that the immune system and small GTPase signaling are the most consistently impaired pathways in the amygdala of ELA individuals. Overall, the present work provides new insight into epigenetic regulation of brain plasticity as a function of early-life experience.
Introduction
Early-life adversity (ELA), including sexual and physical abuse, as well as other forms of child maltreatment, is a major public-health problem that affects children of all socio-economic backgrounds1. ELA is a strong predictor of increased lifetime risk of negative mental health outcomes, including depressive disorders2. Among other findings, a growing number of studies suggest an association between ELA and morphological and functional changes in the amygdala3, a brain structure critically involved in emotional regulation4. It is possible, thus, that amygdala changes observed in individuals who experienced ELA may contribute to increase risk of psychopathology.
The amygdala is composed of inter-connected nuclei, among which the basal and lateral sub-divisions are responsible for receiving and integrating external information. In turn, these nuclei strongly innervate the central amygdala, the primary nucleus projecting outside the amygdalar complex to mediate behavioural outputs4. While specific functional properties of these nuclei remain difficult to assess in humans, animal studies indicate that the basal and lateral sub-divisions exhibit differential responsivity to stress, in particular as a function of the developmental timing of exposure (adolescence versus adulthood)5, 6. Here, we focused on homogeneous, carefully dissected tissue from the human lateral amygdala.
Childhood is a sensitive period, during which the brain is more responsive to the effect of life experiences7. Proper emotional development is contingent on the availability of a supportive caregiver, with whom children develop secure attachments8. On the other hand, ELA signals an unreliable environment that triggers adaptive responses, and deprives the organism from essential experience. A growing body of evidence now supports the hypothesis that epigenetic mechanisms play a major role in the persistent impact of ELA on gene expression and behaviour9. While DNA methylation has received considerable attention, available data also points towards histone modifications as another critical and possibly interacting factor9. Therefore, in this study we conducted a comprehensive characterization of epigenetic changes occurring in individuals with a history of severe ELA, and carried out genome-wide investigations of multiple epigenetic layers, and their cross-talk. Using post-mortem tissue from a well-defined cohort of depressed individuals with histories of ELA, and controls with no such history, we characterized 6 histone marks, DNA methylation, as well as their final endpoint at gene expression level.
We first generated data for six histone modifications: H3K4me1, H3K4me3, H3K27ac, H3K36me3, H3K9me3, and H3K27me310, using chromatin-immunoprecipitation sequencing (ChIP-Seq). This allowed us to create high-resolution maps for each mark, and to define chromatin states throughout the epigenome. In parallel, we characterized DNA methylation using Whole-Genome Bisulfite Sequencing (WGBS). While previous studies in psychiatry focused on the canonical form of DNA methylation that occurs at CG dinucleotides (mCG), here we investigated both CG and non-CG contexts. Indeed, recent data has shown that non-CG methylation is not restricted to stem cells, and can be detected in brain tissue at even higher levels11. Available evidence also indicates that it progressively accumulates, preferentially in neurons, during the first decade of life12, 13, a period when ELA typically occurs. Thus, we postulated that changes in non-CG methylation might contribute to life-long consequences of ELA, and focused in particular on the CAC context, where non-CG methylation is most abundant. Parallel analyses combining all epigenetic layers and transcriptomes converged to identify immune system processes and small GTPases as critical pathways associated with ELA, and suggested that ELA leaves distinct, albeit equally frequent, traces in CG and CAC sites. Altogether, our results suggest previously unforeseen sources of epigenetic and transcriptomic plasticity, which may likely contribute to the severe and lifelong impact of ELA on behavioural regulation, and the risk of psychopathology.
Results
Histone landscapes
Six histone modifications were assessed in subjects with histories of ELA (n=21) and healthy controls (C) with no such history (n=17; Supplementary Tables1,2). Following the International Human Epigenome Consortium (IHEC) procedures, we achieved >60 and >30 million reads for broad (H3K4me1, H3K36me3, H3K27me3 and H3K9me3) and narrow (H3K27ac and H3K4me3) marks, respectively (4.0 billion reads total; Fig.S1a, Supplementary Table3). Quality controls confirmed that all samples for the 2 narrow marks showed relative and normalized strand cross correlations that were, respectively, greater than 0.8 and 1.05 (Fig.S1b), according to expectations14. Read distribution within genes showed expected patterns (Fig.1a,b): reads were strongly enriched in Transcription Start Sites (TSS) regions for H3K27ac, H3K4me3 and H3K4me1, while H3K27me3 and H3K36me3 showed antagonistic distributions, consistent with results obtained in other tissues10. Samples strongly clustered by the type of mark, with a large distinction between activating and repressive marks (Fig.1c). To investigate tissue specificity of our dataset, we then compared it with data from other brain regions and blood tissue (from the NIH Roadmap Epigenomics; Fig.S2). For each modification, we observed higher correlations among amygdalar samples (r=0.75-0.92 across the 6 marks) than when compared with samples from other brain regions (r=0.51-0.81), and even lower correlations with peripheral blood mononuclear cells (r=0.35-0.64), consistent with the role of histones in tissue identity. We next investigated relationships between histone marks and gene expression (Fig.1d). As expected, we observed activating functions for H3K27ac, H3K4me1, H3K36me3 and H3K4me3, and repressive functions for H3K27me3 and H3K9me3. Distinct correlation profiles were observed between marks along the spectrum of gene expression levels, indicating that multiple marks are likely to better predict gene expression than individual ones. Comparisons between ELA and C groups found no significant overall differences in terms of read distribution (Fig.1b) or relationship to gene expression (Fig.1d), indicating that ELA, as expected, does not globally reconfigure amygdalar histone landscapes.
Considering that different combinations of histone modifications define so-called chromatin states15, we then conducted a combined analysis of all marks using ChromHMM machine-learning. Maps of chromatin states were generated as described previously16, with each state corresponding to a distinct combination of individual marks. This unbiased approach defined a consensus map corresponding to genomic regions showing ≥70% agreement across samples (Fig.1e, and see Methods), and consistent with studies in the brain and other tissues16-18: for example, regions defined by a combination of H3K27ac and H3K4me1 corresponded to enhancers19. As detailed below, these maps allowed us to characterize cross-talks between chromatin states and DNA methylation, and differences between groups.
CG and non-CG methylation patterns
We used WGBS to characterize methylomes in both groups, generating a total of 6.2 billion reads. Rates of bisulfite conversion and over-conversion, sequencing depth, and library diversity met the IHEC standards and were similar across groups (Fig.S3a-d). In this large dataset, >13 million individual CGs showed an average coverage ≥5 in the entire cohort (Fig.S3e), which favourably compares with recent human brain studies in terms of sample size20 or number of CGs covered21, 22.
Because recent studies suggest that non-CG methylation is enriched in mammalian brains11, 23, we first computed average genome-wide levels of methylation in multiple cytosine contexts. Focusing on 3-letter contexts (Fig.2a), we observed that, as expected, methylation levels were highly variable among the 16 possibilities (2-way ANOVA; context effect: [F(15,540)=196283; p<0.0001]), with much higher methylation levels in the CGA, CGC, CGG, and CGT contexts than in the 12 non-CG contexts. Of note, no difference was found in overall methylation between groups ([F(1,36)=0.12; p=0.73]), indicating, as expected, that ELA does not associate with a global dysregulation of the methylome. Among non-CG contexts, methylation was the highest at CACs (4.1±0.1%), followed by a group of contexts exhibiting between 1.8 to 1.1% average methylation (CTC, CAG, CAT, and CAA), while low levels (<0.4%) were observed for remaining contexts. This ordering was strikingly similar to what was recently described in the mouse brain24 (Fig.S4a-b), suggesting a robust and conserved distribution of methylation according to sequence context. Considering that methylation at CA25 or CAC26 sites may have specific functions in the brain, and because CAC methylation (hereafter mCAC) was most abundant, we focused on this context in follow-up analyses.
We first compared the abundance of mCG and mCAC. While CG sites were highly methylated, CAC (Fig.2b-c) or other non-CG (Fig.S4c) sites were mostly unmethylated, with a minority of them showing methylation levels between 10 to 20%, consistent with mouse data27. Regarding distinct genomic features and chromosomal location, we confirmed that, while mCG is lower within promoters (where CGs frequently cluster in CG islands), this effect is much less pronounced for mCAC (Fig.S5a)11. Further, we observed that (i) compared with CGs28, depletion of methylation from pericentromeric regions is even stronger at CACs, and that (ii) methylation levels were extremely low in both contexts in the mitochondrial genome (Fig.S5b). We then confronted methylation data with measures of gene expression, regardless of group status, and found the expected anti-correlation in both contexts (Fig.2d; CG: [F(1,37)=557; p=6.7E-24]; CAC: [F(1,37)=3283; p=9.7E-38]). Because CAC sites, in contrast with CGs, are by definition asymmetric on the two DNA strands, we wondered whether this anti-correlation would be different when contrasting gene expression with mCAC levels on the gene’s (sense) or the opposite (antisense) strand. No difference was found (Fig.S6), indicating that gene expression is predicted to the same extent by mCAC on either strand, at least for the coverage achieved here. These data emphasize noticeable differences and similarities between mCG and mCAC, and are consistent with results previously obtained for all non-CG contexts combined, in the mouse and in human27, 29.
Regarding histone modifications, while mechanisms mediating their interactions with mCG have been documented30, no data is currently available to describe such relationship for non-CG contexts. To address this gap, we confronted our consensus model of chromatin states (see Fig.1e) with DNA methylation. For both mCG ([F(1,36)=0.36; p=0.55]) and mCAC ([F(1,36)=0.07; p=0.80]; Fig.2e-f), genome-wide methylation levels were similar between groups across the 10 states, indicating that ELA does not associate with a global disruption of the cross-talk between DNA methylation and chromatin. Nevertheless, methylation levels strongly differed between states, in both CG ([F(9,324)=5127; p<0.0001]) and CAC ([F(9,324)=910.7; p<0.0001]) contexts, unravelling previously uncharacterized dissociations. First, in the CG context (Fig.2e) we observed a strong anti-correlation between DNA methylation and both forms of H3K4 methylation (me1, me3), consistent with findings in other cell types30: lowest mCG was observed in the 3 promoter states defined (Fig.1e) by high levels of H3K4me3 in combination with either high H3K4me1 (Flanking Promoter, Flk Prom; p<0.0001 for every post-hoc comparison, except against the Polycomb repressed state, PcR), high H3K27ac (Active Promoter, Act-Prom; p<0.0001 for every comparison against other states), or intermediate levels of both H3K27ac and H3K4me1 (Weak Promoter, Wk-Prom; p<0.0001 against other states). In contrast, among these 3 promoter states mCAC was particularly enriched in Wk-Prom regions (p<0.0001 against Act-Prom and Flk-Prom; Fig.2f). Second, mCG was abundant in transcribed regions defined by either intermediate H3K36me3 (Weak Transcription, Wk-Trans), or high H3K36me3 and low H3K27ac (Strong Transcription, Str-Trans). By contrast, mCAC was selectively decreased in the Str-Trans state (p<0.0001 against Wk-Trans). Third, while mCG levels were high in heterechromatin (Heteroch, defined by high H3K9me3 alone), consistent with its role in chromatin condensation, mCAC appeared depleted from these regions (p<0.0001 for every comparison against other states, except PcR and Flk-Prom). Overall, results indicate that interactions between DNA methylation, histones, and chromatin strikingly differ across mCG and mCAC, possibly as a result of brain-specific epigenetic processes in the latter 3-letter context29.
Changes in individual histone marks or chromatin states as a function of ELA
We investigated local adaptations in histone profiles of ELA subjects using diffReps31. A total of 5126 differential sites (DS) were identified across the 6 marks (Fig.3a-b, Fig.S7, Supplementary Table4) using consensus significance thresholds32 (p<10−4, FDR-q<0.1). Interestingly, H3K27ac contributed to 30% of all DS, suggesting a meaningful role of this mark in epigenetic changes associated with ELA. Annotation to genomic features revealed distinct distributions of DS across marks (df=25, χ2 =1244, p<0.001; Fig.S8a): H3K4me1- and H3K4me3-DS were equally found in promoter regions and gene bodies, while H3K36me3- and H3K27ac-DS were highly gene-body enriched, and H3K27me3- and H3K9me3-DS found in intergenic/gene desert regions. Sites showing enrichment (up-DS) or depletion (down-DS) of reads in ELA subjects were found for each mark, with an increased proportion of down-DS associated with changes in H3K4me1, H3K4me3, H3K36me3 and H3K27me3 (Fig.S8b).
We then used GREAT (Supplementary Table5), a tool that maps regulatory elements to genes based on proximity, to test whether ELA subjects had perturbations in histone modifications that affected genes in specific pathways33. We performed this GO analysis for biological processes and molecular functions (Fig.3c-d) on each mark, and found significant enrichments for DS involving 3 marks: H3K27ac, H3K27me3 and H3K36me3. Importantly, overlaps between enriched GO terms were observed across marks: notably, terms related to immune processes, as well as small GTPases, were enriched for H3K36me3- and H3K27ac-DS, suggesting these pathways may play a significant role in ELA.
To strengthen these findings, a joint analysis of all marks was conducted using maps of chromatin state15. First, we identified genomic regions where a switch in chromatin state (state transitions, ST; n=61,922) occurred between groups (Supplementary Table6). Across the 90 possible ST in our 10-state ChromHMM model, only 56 were observed, with a high proportion (50.2%, indicated by * in Fig.4a) involving regions in quiescent (Quies), Wk-Trans or Enh states in the C group that mostly turned into Quies, Str-Trans, Wk-Trans, and Heteroch states in the ELA group. Furthermore, 17% and 59% of ST occurred in regions within 3kb of a promoter or in gene bodies (Fig.4b), respectively, suggesting that ELA-associated changes affected selected chromatin states, and mostly occurred within genes.
We next investigated GO enrichment of ST, using GREAT (Fig.4c-d, Supplementary Table7). As described previously32, a co-occurrence score reflecting both the significance of GO terms and their recurrence across multiple STs was computed. Importantly, biological processes (Fig.4c) with highest co-occurrence scores were similar to those found from the GO analysis of individual histone marks, and clustered in two main categories: immune system, and small GTPases. These terms were significant for state transitions involving transcription, quiescent and enhancer states. Regarding molecular functions (Fig.4d), most enriched gene categories were related to GTPases, and involved the same types of ST. Therefore, analyses at the level of individual histone marks and chromatin state converged to suggest global impairments in similar pathways.
Differential DNA methylation in ELA
We next sought to identify changes in DNA methylation. As the abundance of mCG and mCAC were very different, and considering data suggesting possible mCAC-specific processes26, we used the BSmooth algorithm34 to identify DMRs separately in each context, using strictly similar parameters (see Methods). DMRs were defined as regions of five or more clustered cytosines that each exhibited a significant difference in methylation (p<0.001), and an absolute methylation difference ≥1% between groups. Surprisingly, we found that as many DMRs could be identified in the CAC context (n=866) as in the canonical CG context (n=878, Fig.5a-b). These 2 categories of DMRs were distributed throughout the genome and in every chromosome, while there was no direct overlap between genomic regions that they covered. Compared with CG-DMRs, CAC-DMRs were composed of slightly fewer cytosines (Fig.S10a, p=2.9E-04) and were smaller (Fig.S10b, p<2.2E-16). In addition, consistent with the overall lower abundance of mCAC, CG-DMRs affected sites showing a wide range of methylation levels, while CAC-DMRs were primarily located in lowly methylated regions (Fig.5c-d). Further, methylation changes detected in the ELA group were less pronounced in the CAC context, as shown by smaller percentage changes in methylation (p<2.2E-16; Fig.5e-f, Fig.S10c) and smaller areaStat values (the BSmooth measure of the statistical strength of methylation changes within DMRs34; p=5.8E-08, Fig.S10d). Overall, these results suggest that cytosines in the CAC context may represent a significant form of plasticity that may contribute to long-term consequences of ELA.
We next characterized genomic features where DMRs occurred, and observed that their distribution strikingly differed (p<2.2E-16; Fig.6a-b, Supplementary Table8): CG-DMRs were located in promoters (38.5% in proximal promoter, promoter1k and promoter3k) and gene bodies (35.4%), while CAC-DMRs were mostly found in gene bodies (53%) and intergenic regions (28.1%). Second, we characterized histone modifications around DMRs (Fig.6c-d, Fig.S11): CG-DMRs were enriched with H3K4me1, H3K4me3 and H3K27ac (Fig.6c), reinforcing our previous observations that these histone marks (Fig.1e) and DMRs (Fig.6a) were preferentially located in promoters. In sharp contrast, the 2 main features characterizing CAC-DMRs were an enrichment in H3K36me3 and a depletion in H3K9me3 (Fig.6d, Fig.S11). These differences were further supported by the analysis of chromatin states (p<2.2E-16; Fig.6e, Fig.S12, Supplementary Table9). CAC-DMRs were largely absent from promoter (Act-Prom, Flk-Prom, and Wk-Prom) and enhancer (Str-Enh and Enh) states that were all defined, to varying degrees, by the 3 marks that primarily characterize CG-DMRs: H3K4me1, H3K4me3 and H3K27ac (Fig.1e). In addition, CAC-DMRs were (i) enriched in the Wk-Trans state, defined by the presence of H3K36me3, and (ii) depleted from the 2 states (PcR, Heteroch) characterized by the H3K9me3 mark. These effects were consistent with the enrichment and depletion previously observed individually for each of these 2 histone modifications, respectively.
Finally, we conducted a GREAT analysis of GO terms enriched for DMRs: CG-DMRs notably associated with terms related to the regulation of neuronal transmembrane potential (Fig.6f), in agreement with histone results (see Fig.4c), while CAC-DMRs were enriched for terms related to glial cells (glial cell differentiation; leukocyte migration, Fig.6g), consistent with the immune dysregulation previously observed with histone DS and ST. Altogether, while ELA associates with similar numbers of mCG and mCAC adaptations, these 2 types of plasticity occur in genomic regions characterized by different histone marks, chromatin states, and GO categories, possibly reflecting the implication of distinct molecular mechanisms.
Combining GO analyses
Analyses of histone modifications and DNA methylation identified GO terms consistently affected in ELA individuals. To determine how these epigenetic adaptations may ultimately modulate amygdalar function, we characterized gene expression in C and ELA groups, using RNA-Sequencing. Samples with similar RNA integrity across groups were sequenced at high depth (>50 million reads/sample), yielding good quality data (Fig.S13). Quantification of gene expression was conducted using HTSeq-count, as described previously35, and validated by an alternative pseudo-alignment approach, Kallisto36, generating very similar results (r=0.82, p<2.2E-16; Fig.S14a). A differential expression analysis between groups was then performed using DESeq2 (see full results in Supplementary Table10). Similar to our analyses of epigenetic datasets, we searched for patterns of global functional enrichment using GO and Gene Set Enrichment Analysis (GSEA)35. Enrichment of GO categories using the 735 genes that showed nominal differential expression in the ELA group (p<0.05, Fig.7a, Supplementary Table11) identified numerous terms consistent with our previous analyses at epigenetic level, including immune and small GTPase functions (Fig.7b). As a complementary approach, we used GSEA37, which does not rely on an arbitrary threshold for significance, and takes directionality of gene expression changes into account. GSEA identified 163 genome-wide significant sets, among which 109 were related to immune processes and negatively correlated with ELA (Supplementary Table12, Fig.7c-d). Therefore, analysis of transcriptomic data identified gene pathways that in part overlap with those identified at the level of histone marks and DNA methylation.
To combine analyses conducted for histone modifications, chromatin states, DNA methylation and gene expression, we finally grouped GO terms enriched at each level to identify biological mechanisms most consistently affected (Fig.S14c). Overall, a clear pattern emerged whereby the highest number of genome-wide significant terms (n=122 GO terms) were related to immune processes, with contributions from each of the 4 types of data. Second came terms related to small GTPases, which were documented by histone modifications, chromatin states and gene expression (n=22), followed by terms related to neuronal physiology (n=19, mostly linked with neuronal excitability and sensory processing; Supplementary Table13), cellular adhesion (n=9), and the cytoskeleton (n=6). Overall, these combined analyses defined major epigenetic and transcriptomic pathways affected by ELA in the lateral amygdala.
Discussion
This study investigated 6 histone marks, DNA methylation and gene expression in, to our knowledge, one of the most comprehensive comparison of canonical mCG with the brain-enriched mCAC. In the healthy brain, striking differences in the relationship that the 2 forms of methylation exhibit with histone modifications were uncovered, providing avenues for mechanistic molecular studies. Also, going beyond previous studies of ELA9, this work represents the first analysis of its consequences across multiple epigenetic layers. Results indicated that most extensive changes affect immune-related genes and small GTPases, in part through the reprogramming of chromatin and the methylome. Finally, results suggest that mCAC is plastic in the human brain, and may be developmentally regulated by ELA to a similar extent to what can be observed in the reference CG context, uncovering a potential new molecular substrate for the embedding of early-life experience.
We first integrated DNA methylation data with histone marks and gene expression to identify biological pathways most significantly associated with ELA in the lateral amygdala. Genome-wide analyses showed converging evidence for significant enrichment in immune-related GO terms (including genes encoding the complement system, Toll-like receptors, clusters of differentiation, the major histocompatibility complex), across all molecular layers, suggesting a meaningful contribution to psychopathological risk. Over the last two decades, considerable evidence has associated enhanced inflammation with stress-related phenotypes such as depression, in particular based on measures of cytokines and inflammatory factors in blood samples38. Limited molecular data, however, is available to understand how this pro-inflammatory state may translate at brain level, with few studies reporting inconsistent findings in cortical structures (with mostly increases but also decreases in the expression of inflammation-related genes38). Our findings therefore represent, to our knowledge, the first indication for altered immune processes in the lateral amygdala in relation to ELA and associated depression. While the global pattern of downregulation we observed at gene expression level was surprising considering the general view that stress-related psychopathology associates with higher inflammation, it is nevertheless consistent with the lower density of glial cells described in the amygdala of depressed individuals (in some39, 40 but not all41 studies), which may notably concern microglial and astrocytic cells, the main immune actors in brain tissue.
The second pathway most significantly altered in ELA subjects was related to small GTPases, a large family of GTP hydrolases that among other processes regulate synaptic structural plasticity through interactions with the cytoskeleton42. Interestingly, the association observed for small GTPases was accompanied by significant changes affecting GO terms related to the cytoskeleton. Overall, this indicates that some form of synaptic plasticity occurs in the lateral amygdala as a function of ELA, and reveals part of the underlying epigenetic mechanisms. While very few molecular post-mortem studies support this hypothesis43, it strongly resonates with the wealth of human imaging and animal data documenting structural and functional plasticity in this brain region as a function of stressful experiences4.
We next used our extensive data set to conduct a detailed analysis of non-CG methylation. Over the last few years, the significance of this type of DNA methylation, and the possibility that it may fulfill biological functions, have been supported by several lines of evidence, including: (i) distinct methylation patterns shown to preferentially affect CAG sites in embryonic stem cells as opposed to CACs in neuronal and glial cells11, (ii) the particular abundance of non-CG methylation in genes with a higher genomic size in the human brain25, and (iii) the specific binding of the methyl-CpG-binding domain protein MeCP2 to both mCG and mCAC in the mouse brain26, 44. Here, we provide additional evidence reinforcing this notion, as we found that the relative frequency of non-CG methylation across 3-letter cytosine contexts is conserved in the human brain compared to mouse and, importantly, that mCAC exhibits peculiar interactions with the histone code as well as quantifiable plasticity in relation to ELA.
We first compared mCG and mCAC in distinct genomic features and chromatin states, regardless of clinical grouping. While interactions between DNA methylation and histone marks were described previously in other tissues45, 46, and in the brain for mCG45, here we unravel unforeseen specificities regarding mCAC. First, among the 3 chromatin promoter states (Act-Prom, Wk-Prom, Flk-Prom), mCAC was selectively enriched in Wk-Prom, which was not observed for mCG. Considering that Wk-Prom was relatively depleted in H3K27ac and H3K4me1 compared to the 2 other promoter states, it is possible to hypothesize that these 2 histone modifications may potentially repress mCAC accumulation in brain tissue. A second dissociation consisted in the fact that lower mCAC levels were measured in Str-Trans compared with Wk-Trans regions, while no such difference was observed in the CG context. This may result at least in part from higher levels of H3K36me3 observed in the Str-Trans state. Third, among the 2 tightly compacted chromatin states defined by the repressive mark H3K9me3, PcR and Heteroch, the latter state was characterized by higher DNA methylation in the CG, but not in the CAC, context, as well as by a relative increase in H3K9me3 and a decrease in H3K27me3. While there is currently no data, to our knowledge, supporting a potential interaction between H3K27me3 and non-CG methylation11, a role for H3K9me3 can be speculated considering studies of cellular reprogramming. Indeed, the in vitro dedifferentiation of fibroblasts into induced pluripotent stem cells associates with the restoration of non-CG methylation patterns characteristic of stem cells, except in genomic regions characterized by high levels of H3K9me347. Therefore, the possibility exists that H3K9me3 may be implicated in the regulation of mCAC in the brain, an hypothesis that warrants further investigation.
Recently, a molecular pathway that may contribute to such differences among mCG and mCAC in the brain has started to be unravelled: the methyltransferase Dnmt3a was shown in the mouse to mediate progressive post-natal accumulation of DNA methylation in the CA context13, while in vivo recruitment of MeCP2 was demonstrated to rely primarily on mCG and mCAC levels rather than by methylation at other contexts, including CAT, CAA, or CAG26. In combination with these rodent studies, our results therefore suggest that DNMT3a and MeCP2 may be implicated in the particular cross-talk that seem to emerge between mCAC and histone modifications during brain maturation, and suggest that future investigations should focus on their putative interaction with aforementioned specific marks (H3K27ac, H3K4me1, H3K36me3, and H3K9me3), and related histone-modifying enzymes.
Finally, we wondered whether mCAC might show some degree of plasticity in the human brain. We found that similar amounts of differential methylation events could be detected across CAC and CG contexts in ELA subjects, suggesting that mCAC might be sensitive to behavioral regulation. While previous studies clearly showed that ELA associates with widespread effects on mCG throughout the brain’s genome, they were conducted using methodologies (methylated DNA immunoprecipitation coupled to microarrays48, reduced representation bisulfite sequencing35) primarily designed for the investigation of mCG. In comparison, the present study using WGBS provides a more comprehensive and unbiased assessment of the overall methylome, and represents, to our knowledge, the first indication in humans that the brain-specific mCAC form of DNA methylation might be affected by ELA. This result is consistent with recent mouse work49 that provided evidence for an effect of environmental enrichment, during the adolescence period, on non-CG methylation, suggesting that both positive and negative early-life experiences may have the capacity to modulate this non-canonical epigenetic mechanism. Importantly, our combined investigation of DNA methylation and histone marks provides further characterization of this form of plasticity. Strikingly, mCAC and mCG changes occurred in genomic regions that appeared distinguishable at virtually any level of analysis, including genomic features, individual histone marks, chromatin states, and GO categories. Accordingly, CG-DMRs primarily located among promoter regions and gene bodies, were enriched in H3K4me1, H3K4me3 and H3K27ac, and were represented across all chromatin states. In comparison, CAC-DMRs were less frequently found in promoter, were enriched in H3K36me3 and depleted in H3K9me3, and mostly associated with 2 chromatin states, Quiescent and Wk Trans. Overall, these results suggest a model whereby the complex cascades of neurobiological adaptations associated with ELA may result from, or alternatively contribute to, distinct pathophysiological phenomenon that differentially manifest at the level of CG and CAC sites. It is also possible to speculate that part of these adaptations may result from the impact of ELA on mechanisms that drive the developmental emergence of mCAC. In the future, animal models will be instrumental in testing this hypothesis and in deciphering underlying molecular cross-talk among epigenetic layers.
In conclusion, the epigenetic and transcriptomic landscape of the lateral amygdala exhibit targeted reconfigurations as a function of ELA. This reprogramming can be detected consistently across multiple epigenetic layers, including the newly recognized form of DNA methylation affecting CAC sites. Future studies will hopefully define the extent to which non-CG methylation at CACs, and potentially at other cytosine contexts, contribute to adaptive and maladaptive encoding of life experiences in the brain.
Methods
Methods and any associated references are available in the online version of the paper.
Authors contributions
PEL and GT conceived the study. GGC, EM, JY and PEL performed RNA and DNA extractions. TK and TP provided library preparations and next-generation sequencing through the IHEC consortium, and AR prepared ChIP-Seq libraries. MAC and PEL analyzed histone data. ZA and PEL analyzed methylation data. JFT and PEL analyzed RNA-Seq data, while the Kallisto analysis was conducted by JCG. AP conducted the BSmooth differential methylation analysis. MA, LCV, CE, IY, NM and GT provided lab resources and equipment. PEL, MAC and GT prepared the manuscript, and all authors approved the final version of the manuscript.
Acknowledgments
PEL was supported by fellowships from the ‘Fondation Fyssen’, the Canadian Institutes of Health Research, the American Foundation for Suicide Prevention, the ‘Fondation pour la Recherche Médicale’, and by fundings from the Bettencourt-Schueller Foundation, the Brain Canada Foundation, the ‘Fondation Deniker’, the ‘Congrès Français de Psychiatrie’, the ‘Fondation de France’ (N°Engt:00081244) and the ‘Union Nationale de Familles et Amis de Personnes Malades et/ou Handicapées Psychiques’.