The Methylation Landscape and its role in Domestication and Gene Regulation in the Chicken

Abstract

Domestication is one of the strongest examples of artificial selection, and has produced some of the most extreme within-species phenotypic variation known. In the case of the chicken, it has been hypothesised that DNA methylation may play a mechanistic role in the domestication response. By inter-crossing wild-derived Red Junglefowl with domestic chickens we mapped Quantitative Trait Loci for hypothalamic methylation (methQTL), gene expression (eQTL) and behaviour. We find large, stable methylation differences, with 6179 cis and 2973 trans methQTL identified. Over 46% of the trans effects were genotypically controlled by five loci, mainly associated with increased methylation in the Junglefowl genotype. In a third of eQTL we find that there is a correlation between gene expression and methylation, whilst statistical causality analysis reveals multiple instances where methylation is driving gene expression, as well as the reverse. We also show that methylation is correlated to some aspects of behavioural variation in the intercross. In conclusion, our data suggest a role of methylation in the regulation of gene expression underlying the domesticated phenotype of the chicken.

Introduction

Domestication has occurred in a range of different plants and animals, and generally leads to extreme phenotypic variation from the original wild progenitor phenotype ^1,2. In the case of the chicken, domestication first occurred around 8k years ago^3,4. The greatest changes in phenotypic variation, however, occurred only in the past few hundred years, a common occurrence among domesticated species ^5–7, with the chicken separately selected in the last ~50 years as broiler (for increased meat production) and layer (for increased egg production) breeds. In the case of the layer chicken, this recent selection has led to genetic changes in reproduction⁸, brain composition⁹, bone allocation¹⁰, growth¹¹, sexual ornament size¹², and anxiety-related and other behaviours, amongst others^13–15. More recently, methylation differences between Red Junglefowl (the wild progenitor to the domestic chicken) selected for high and low tameness¹⁶ and an excess of mutations in CpG sites ¹⁷have also been reported in chickens. In the case of the chicken, evidence exists for transgenerational effects on behaviour, with a potential epigenetic basis¹⁸. However, no direct evidence is available to our knowledge on how epigenetic and genetic components interact to modulate gene expression in the chicken. Disentangling DNA methylation-related epigenetic modifications and their effects on transcription will provide valuable information to understand the domestication process in the chicken and the potential role of DNA methylation.

Epigenome alterations and DNA methylation in particular are key regulators of eukaryotic genomes, and can lead to heritable phenotypic and transcriptomic changes. The relationship between DNA methylation and gene expression is complex. The canonical view is that CpG methylation in promoter regions directly represses transcription factor binding, thereby inhibiting transcription ^19,20. Recent work, however, has reported that DNA methylation can also increase transcription factor binding ²¹. Additionally, DNA methylation in other genomic regions such as enhancers, exons and splicing sites has also been found to yield complex patterns of transcriptomic expression. In vertebrates, enhancers regulated by DNA methylation are linked to the coordinated transcriptional and epigenomic regulation of developmental genes ²². Additionally, there are genomic regions that concentrate mediator and cohesion components called ‘super-enhancers’²³. These super enhancers have altered methylation status in tumours, with methylation being inversely correlated to gene expression ²⁴. In mammals and honey bees exonic regions which are hypermethylatedin relation to adjacent regions are correlated with increased gene expression ²⁵. In humans, CpG methylation is shown to be higher in alternative splicing sites ²⁶.

Local DNA variation can also modify DNA methylation ^27–30, with sequence polymorphisms modifying the extent of CpG methylation. In this manner, epigenetic modifications can act as the mechanism by which DNA variants influence gene expression, and ultimately phenotypic variation³¹. By using genome-wide inter-population variation it is possible to identify DNA polymorphisms that underpin the extent oflocal DNA methylation. Quantitative Trait Loci (QTL) studies with DNA methylation can be performed in this manner, by either inter-crossing populations with different levels of DNA methylation (a linkage-based approach) or by using within-population variation in DNA methylation in a single population (a linkage disequilibrium based approach). Both methods work by assessing local DNA methylation in discrete windows throughout the genome, then correlating these with molecular SNP markers also spaced throughout the genome to identify genomic regions that underpin quantitative variation in DNA methylation. Although powerful, very few studies have used this approach, which has only been performed in human cell lines ³², blood ^33,34 and brain tissue ³⁵, and Arabidopsis³⁶. By combining methylation QTL studies with simultaneous expression QTL studies using the same individuals, it is possible to detect regions that control both methylation and transcriptomic variation, and even statistically assess the likelihood that methylation status is potentially causal to gene expression and vice-versa for each individual region and gene.

To address the extent to which DNA methylation and related gene expression is affected by domestication, we conducted a DNA methylation quantitative trait locus (methQTL) analysis using an advanced intercross between domestic (White Leghorn) chickens and Red Junglefowl. The hypothalamic transcriptome and methylome of 124 individuals were assayed and DNA methylation and expression QTL were jointly mapped and compared, along with a variety of domestication-related anxiety behaviours. In this manner, not only was methylation over the entire autosomal genome covered, but this was performed on an individual basis, allowing inter-individual variation in DNA methylation for any given region to be assessed.

Results

I. The Regulation of Methylation, Gene Expression and Behaviour

Abundant Genetic Regulation of Methylation in the Chicken

Both local (cis) and trans methQTL were detected in the intercross, with 6179 cis and 2973 trans methQTL identified (see Supplementary Table 1 and Figure 1). In comparison, 1123 expression QTL were detected in the same tissue from the same intercross individuals, indicating that genetic variation in the regulation ofDNAmethylation between wild and domestic genotypes appears to be highly abundant in the chicken in both cis and trans forms. Of the cis methQTL, there was a modest but significant enrichment of alleles with a greater effect on methylation being derived from the Red Junglefowl genotype (3235 methQTL with the greater effect derived from RJF alleles, 2944 methQTL with the greater effect from WL alleles, chi-squared test, p-value=0.0002). In the case of cis regulation, this was not only limited to small discrete regions, but in some cases stretched over several kilobases to form consistently methylated blocks (i.e. multiple adjacent 1kb regions were controlled by the same genetic region -see Supplementary Table 2). In total 879 cis methQTL mapped to 77 different SNP markers, with these regions stretching over ~13kb on average. The largest of these consistently methylated blocks consisted of 186 cis methQTL mapping to a SNP marker on chromosome 1 in three discrete blocks, covering a 204kb region at 137Mb (137.574 Mb - 137.778 Mb). All these blocks were associated with increased methylation from the RJF allele. In general, where these blocks occur they are significantly more likely to show an increased methylation in the wild (RJF) genotype compared to the domestic White Leghorn (WL) genotype (58 regions to 24; chi-square p-value < 2e-5). Similarly, the methQTL within these blocks were significantly more likely to have a greater effect from the RJF allele (675 methQTL with a greater effect from RJF alleles, 183methQTL with a greater effect from WL alleles; chi-square p-value < 3e-63).

Figure 1.

Circle plot showing the position of (A) continuous cis methQTL blocks, and (B-D) three of the transmethQTL hotspots. In (A) blocks in the grey shaded area indicate a locus with a larger effect from the domestic allele, whilst blocks in the red shaded area indicates a larger effect from the RJF allele. (B-D) show the trans effects arising from three trans methQTL hotspots on chromosomes 15, 2 and 3 (one hotspot per frame), with the lines connecting the different trans genomic regions regulated via the hotspot.

Trans MethQTL Hotspots

Trans acting methQTL were less abundant than cis methQTL, however thetrans methQTL were found to cluster together into a small number of regions (referred to as hotspots hereafter). Thirteen hotspots were identified on nine chromosomes (see Supplementary Table 3 and Figure 1), with these accounting for ~60% of all trans methQTL (1792 of the 2973 trans methQTL were present in these 13 hotspots). The largest of these hotspots contained 561 trans methQTL, while the top five trans hotspots accounted for 1375 trans methQTL in total (see Supplementary Table 3). There was a significant bias towards trans methQTL possessing a larger allelic effect derived from Red Junglefowl in these hotspots (986 RJF methQTL, 806 WL methQTL, p-value= 2.2e-5). To more finely analyse the trans methQTL hotspot regions, we firstly checked if there were any specific epistatic interactions between the trans QTL in the hotspot and the methylation phenotypes they control. Very little epistasis was apparentin the hotspot regions, with less than 7% of trans methQTL in these hotspots possessing a nominally significant (P<0.05) epistatic interaction, and less than 1%of trans methQTLwith a significant epistatic interaction(see Supplementary Table 3). One hotspot on chromosome 5 did have more significant epistasis (with 12 of the 98 trans methQTL exhibiting a significant epistatic interaction). We also calculated the minimum common overlap region between the confidence intervals of every trans methQTL in the hotspot to identify any candidate genes that could be present (see Supplementary Table 3). We performed a GO enrichment analysis using all the genes present in each hotspot, and also identified which of these genes possessed an eQTL in the hypothalamus, with 11 eQTL found spread over these 13 hotspots. These genes were PATZ1, LIMK2, CAPN3, ACOT1L, GANC, ERH, ADAM20, DAP, ROPN1L, CMBL, EPGN. Taken as a whole, the hotspots were significantly enriched for five GO terms (alpha-linolenic acid metabolism, linoleic acid metabolism, arachidonic acid metabolism, ether lipid metabolism, glycerophospholipid metabolism), related to acid and lipid metabolism, though this was principally driven by just six genes.

Methylation Regulation of Gene Expression

By overlapping autosomal methQTL (n=8689) with autosomal expression QTL (n=1123) and then using methylation to predict gene expression for the resulting overlaps, we identified candidate regions where methylation putatively iscontrolling gene expression and vice versa. A total of 3694 significant overlaps were identified using this approach, with these representing 435 unique expression microarray probesets and 1988 unique methylation windows (with multiple methylation windows often overlapping a given probeset, see Supplementary table 4). Where multiple methQTL overlapped with an eQTL, we then used a multiple regression to combine all methylation windows associated with the expression of a given probeset(see Supplementary Table 5). All non-significant methylation windows were then dropped sequentially (starting with the least significant) until only significant correlations remained. This resulted in1093 overlaps being retained, representing 795 methylation windows that significantly correlated with one or more of the 435 expression probesets.

To compare the relative impact of genomic variation on DNA methylation and gene expression, all of the 795 methylation windows that were significant in the multiple regression model were then modelled with genotype at the eQTL included as a covariate. In this case 136 of the methylation windows no longer explained any additional variation in gene expression (no significant difference in a likelihood ratio test between the linear models of genotype with and without methylation), leaving 890 overlaps with 659 methylation windows that significantly increased the explained variance of the combined gene expression model. These represented 306 unique probesets (i.e. methylation significantly increased the total variance explained in a probeset model that also included eQTL genotype, see Supplementary table 5), with these comprising of 201annotated genes. We can see that methylation explains much less variation in gene expression than genomic variation does (average proportion of gene expression attributable to methylation is 7.7%, average proportion of gene expression attributable to genotype plus sex is 15%), but there is nevertheless an appreciable effect (see Extended Data 1).

Statistical Directionality/ Correlation Based Tests of Causal Models

When two traits share a common genetic marker (SNP), it is possible to use these markers as a causal ‘anchor’ to assess how these traits relate both to one another and to the anchor marker ³⁷. Analyses can then be performed which weigh the correlations between the traits themselves and between the traits and marker to assess the probability of how these traits are linked. In this way, it is possible to weigh the probability that one of the traits is causal to the other, or whether they are both independently linked to the anchor SNP. Using the Network Edge Orientation (NEO) method devised by Aten et al³⁷that uses a network-based approach to look at these different correlations and assess their relationship, it is possible to try and define the relationships between the 306probesets and 659 methylation windowsidentified above. Using the NEO software we inferred the likelihood of whether the probeset expression in question was statistically causally associated to methylation, the reverse, or whether methylation and probeset expression were independent of one another. Note that when we refer to statistical causation, this is not the same as true functional validation, but rather whether the correlation networks between the traits and the SNP genotypes support the statistical models of SNP genotypes controlling the methylation level of the window in question, andthis methylation, in turn, controlling the probeset expression. Alternatively, the SNP genotype may control probeset expression, which in turn controls methylation at the window (see methods). To distinguish between these two types, we refer to the statistical causality analysis technique devised by Aten et al as ‘statistical directionality’ to reduce the confusion with actual functional validation using wet-lab techniques. Of the 890 overlaps that were assessed, we could infer statistical directionality in110 of the overlaps, representing 98 probesets (69 annotated genes and 23ESTs) and 95 methylation windows (see Supplementary Table 6). Of these, 36 probe sets supported gene expression causing changes in methylation and 74probesetssupported methylation causing gene expression changes. Of these 110 suggestiveor significantmodels, 88 were local (cis) methylation and eQTL interactions and 22 involved either trans eQTL or methQTL.

Identification of Putative Causal SNPs

For the 88 methylation-probeset regions that were significant using statistical directionality and that were also cis acting it was possible to scan the methylation sequence reads for each region for putatively causal SNPs. Using an allelic imbalance test these SNPs should display an excess of the ‘high’ methylation allele in heterozygotes, whilst also being in linkage with the QTL locus. A total of 55 SNPs called from the sequence readshada cis-methQTL effect (i.e. were significantly correlated in a linear model with the respective DNA methylation phenotype). Of these,36 SNPs (covering 14 of the methylation windows) also displayed a significant allelic imbalance (see Supplementary Table 7). These SNPs can be further refined by assessing the genotypic effect of SNPs and comparing to the effectof the imputed QTL genotype. If these SNPs are causal the SNP should be a similar or more significant predictor of methylation. These are then either the causal quantitative trait nucleotides for the methQTLand the corresponding eQTL, or are in linkage with the actual causal SNP and also closely mirror the allele ratio of the causal SNP. Twelve genes contained one or more SNPs where heterozygote individuals (as determined by either the local methylation reads directly or the genotype at the methQTL locus) displayed allelic imbalance (see Supplementary Table 7). As an example, in Figure 2 we show the location of the SNPs for the 1kb region on chromosome 7 at 27835000-27836000 bp, their proximity to CpG islands, and their proximity to selected predicted transcription factor binding sites. The four strongest candidates are NEK3 (SNP 182303631bp on chromosome 1), ENSGAL0000023564 (SNP 99945095bp on chromosome 2), MYLK (SNP 27344378bp on chromosome 7), and DPP10 (27835914bp on chromosome 7). In each instance, the genotypic effect taken directly from the SNP allele genotype is more significant than the imputed QTL genotype from the methQTL mapping procedure for those reads (see Supplementary Table 7).

Figure 2.

Significant NEO Causal methQTL/eQTL complex on chromosome 7 at 27Mb. A) Mean +/- s.e. for eQTL and methQTL genotype effects at this locus (n=124). B) Comparison of methQTL effects with the genotype taken directly from the SNP present within the methQTL 1kb region (vcf) and the genotype taken from the predicted QTL peak (peak), with the phenotype restricted to the sequencing reads overlapping the SNP region, median +/- quartiles shown (n=124). C) MethQTL 1kb region, showing the normalised read coverage for the three different genotypes (corrected for number of individuals in each genotype class). Locations of the significant SNPs are marked, as are several classes of predicted Transcription Factor Binding Sites and CpG locations. D) Allelic imbalance for heterozygous individuals for each of the putatively causal SNPs, mean +/- s.e. shown (n=52 for SNP 27835914).

Methylation and Behaviour

Previously in this intercross we have identified 25 genes as being significant candidate genes to behavioural phenotypes, using the same method of statistical directionality analysis demonstrated here^9,13–15, with these behaviours representing anxiety-related behaviours, as well as predictability/ habituation behaviour. Predictability behaviour refers to the degree of intra-individual variation present (i.e how repeatable/ predictable a given individual is in their behavioural response to a particular test). We have previously shown that domestic WL genotypes display more intra-individual variability than wild RJF genotypes ³⁸. Of these 25 genes, 14 were present in the subset of 435 probesets that possessed both eQTL and methQTL that overlapped and were correlated, with this being a close to significant enrichment (hypergeometric test, p-value= 0.058). Therefore, we estimated the association of the behaviours linked with these genes with gene expression and methylation in a multiple linear regression model. Of the 14 genes, 5 genes had significant associations between methylation and behaviour (see Supplementary Table 8). Two genes (ITGBL, LOC10013671/Loc770352) that correlated with predictability behavioural phenotypes were correlated with methylation, whilst one of the genes that was a candidate for freezing behaviour/ tonic immobility (CA8), one gene that was a candidate for social reinstatement behaviour (TTRAP) and two genes that correlated with open field activity (ADAM10 and LOC10013671/Loc770352 once again) also showed a correlation between behaviour and methylation. We then used the statistical directionality analysis to test the association between behaviour and methylation in the 12 behavioural traits that correlated with these five genes. Although none were significant, they did pass the 0.3 suggestive threshold suggested by ^39,40 in 8 instances (see Supplementary Table 9, encompassing four social reinstatement traits, one open field trait, one tonic immobility trait and one predictability trait).

Comparison of Cis methQTLwith Red Junglefowl Selection Lines

To assess whether the methylation differences observed between genotypes in our intercross were replicated in a separate population of selected Red Junglefowl individuals, we used data previously available from Belteky and colleagues ¹⁶. These RJF (n=12) had been selected for either high fear or low fear of humans for five generations, as a means to try and mimic the effects of early domestication. If the differences we observe in our wild x domestic intercross are also present in these selected individuals it could imply that very early tameness selection could be responsible for the differences we observe. Genome-wide methylation data for 6 high and 6 low selected individuals was generated previously and the data available online¹⁶. The locations of suggestive and significant cis methQTLfrom the most variable methylated regions we identified in the wild x domestic intercross were then lifted from the selected Red Junglefowl data. For each 1kb window, a t-test was used to identify significant differences between high fear and low fear selected Red Junglefowl individuals. Of the 1073 regions analysed, 57 1kb windows showed nominally (p<0.05) significant differences between high fear and low fear groups, however none of the p-values passed an experiment-wide significance threshold (see Supplementary Table 10).

II. Methylation Landscape in the Chicken

Typically, when examining patterns of methylation with gene expression on a genome-wide level gene expression is considered with regard to the rest of the genome. I.e. methylation patterns of genes with greater or lesser expression relative to the rest of the transcriptome present in the tissue under consideration are investigated. This has shown that methylation at the Transcription Start Site (TSS) is decreased in highly expressed genes (for example see the Great Tit genome⁴¹). To verify that the chicken has a similar pattern we divided the transcriptome into high and low expressed genes, relative to the overall gene expression present. Methylation both immediately up-and down-stream from these genes was then averaged across all genes in these categories. A similar pattern to the Great Tit is seen in our data, with decreased methylation found at the TSS of highly expressed genes (see Figure 3A). The large number of individual methylomes in our dataset also allows us to examine the inter-population variation in methylation and its relationship with gene expression on a gene-by-gene basis. To test this, a +/-2Mb region was considered around each gene in the transcriptome, with all methylation windows present individually correlated with expression of the gene under consideration. The most significant window was then retained, with a permutation test used to determine experiment-wide significance. Using this method, we identified 6689 genes with a significant correlation with a local 1kb methylation window. Of these, 3460 genes were positively correlated and 3229 were negatively correlated with methylation (see Figure 3E). By overlapping these results with the analysis of methylation patterns in high and low expressed genes we see that bothcategories (high and low expressed genes) havenegative and positive local correlations (see Figure 3D). We also find that genes with a negative correlation with local methylation have lower methylation at the TSS than those with a positive local correlation (Figure 3B, C). No effect was found if the local correlation was located upstream or downstreamof the gene body (figure 3C).

Figure 3.

Median methylation for regions upstream of the TSS, in the gene body, and downstream of the TTS. (A) Genes subdivided into the highest expressed (1^st quartile) and lowest expressed genes (4^th quartile), n= 5201. (B) The degree of methylation subdivided into those that possess a positive correlation and those that possess a negative correlation with methylation when considering inter-individual variation (genes with a significant correlation with methylation used, n= 5425). (C) Positively and negatively correlated genes further subdivided by location of methylation-gene expression correlation (genes with a significant correlation with methylation used, n= 5425). (D) Positively and negatively correlated genes subdivided into highly expressed (1^st quartile) and low expressed genes (4^th quartile) (genes with a significant correlation with methylation used, with high and low expressed genes then subsetted out, n=2650). (E) Location of correlation and sign (positively or negatively correlated) between methylation and gene expression for all experiment-wide significantly correlated genes, n=6689.

Discussion

Using a wild x domestic intercross to map the effects of chicken domestication, we find that selection appears to haveinduced extreme changes not only in the phenotype and transcriptome, but also in itsregulation by DNA methylation, with continuous stretches of the genome being hypo-methylated via local cis-acting elements. The presence of large trans-acting methylation hotspots is also particularly striking. This is evidence that certain key regions control a huge amount of the variation in DNA methylation inthe domestic chicken, at least in the case of the intercross presented here. Although to our knowledge thisis the first evidence that chicken domestication may be regulated via such trans hotspots, this pattern of massive pleiotropy affecting trans loci has also been found in plants. Studies inArabidopsis²⁷demonstrate that the gene CMT2(a DNA (cytosine-5)-methyltransferase) accounts for approximately 21% of significant associations ³⁶. Overall, these results show that domestication appears to have had several large-scale effects on the methylome of the chicken, with the caveat that QTL results pertain directly to the actual individuals used in the inter-cross used, and may not be entirely representative of the overall populations they are taken from, depending on the degree of genetic variation present in the founders. These methylation changes also appear to affect gene expression, with around one third of all the eQTL detected in the intercross showing a significant correlation with a methylated region. Using a more stringent statistical directionality analysis we identify74probesets that support a role for local methylation altering gene expression, and 36 probesets that support a role for gene expression altering local methylation. Although such tests are still not the same as true functional assessments of causality, they do help us to start assessing which genes are either partially regulated by or are regulating DNA methylation in the chicken. In this way, we demonstrate an important role for DNA methylation in chicken domestication, and identify genes that appear to be affected as well asthe potentially causal SNPs that underpin such variation in DNA methylation. Finally, we also find evidence that DNA methylation is associated with the regulation of behavioural traits in the chicken.

By looking at inter-individual variation in methylation and how it correlates with gene expression, we find that around half of the correlations between DNA methylation and gene expression were positive, with the locations of these peak correlated windows being highly dispersed around the genes and not just limited to the promoter regions. This supports recent findings regarding the role of DNA methylation ingene expression, with DNA methylation being involved in increasing transcription factor binding activity ²¹ and also being present in enhancer regions ⁷, therefore outside promoter regions ⁴². Here we use inter-individual variation to show that such complex regulation is present through-out the genome, and can lead to both positive and negative correlations between DNA methylation and gene expression. Such a large amount of positive correlations between DNA methylation and gene expression are also seen in the few other methylation QTL studies that have been performed. Using human blood as a tissue, Bonder and colleagues ³⁴ found that one third of all methylation/ transcription correlations were positive. A similar split between positive and negative correlations was also seen in a methylation QTL study using Arabidopsis ⁴³.

The results shown here based on inter-individual variation do have some caveats however. Unlike the statistical directionality analysis performed on the overlapping methylation and expression QTL, these only represent correlations between gene expression and local methylation. Therefore, in many cases these are not the actual causal region regulating gene expression, but are in linkage or linkage disequilibrium with the causal region(s). Furthermore, the methylation windows used represent the total methylation observed in a 1kb region. In the case of windows where multiple CpGs are methylated it is possible that the actual causal methylated SNP does not have a similar methylation pattern compared to surrounding CpGs (see ⁴⁴ for a classical example of such an effect), either causing the true association to be missed, or to reverse the correlation from positive to negative or vice-versa. However, in general, methylated DNA marks tend to strongly correlate with one another in local regions of around 400-1000bp ^45–47, whilst a more recent study found differentially methylated regions to be around 1700bp on average ⁴⁸. This suggests that such isolated CpGs may be less of a problem in our data.

The large trans methQTL hotspots contained 13 differentially expressed genes that could be considered as potential candidates for these hotspots. Five of these genes were related to cancer susceptibility (ACOT1L⁴⁹, ERH⁵⁰, DAP⁵¹, CMBL ⁵², EPGN^53,54), whilst five have previously been identified as having their expression modified by methylation (CAPN3⁵⁵, DAP ⁵¹, ROPN1L^56–58, EGF⁵⁹, EPGN⁵³). However, two of the candidates (CAPN3 and ADAM20 at two neighbouring chromosome 5 hotspots) were also identified as being significant using the statistical directionality analysis, with gene expression driving methylation changes in the significant models in both cases. This would be consistent with a gene driving methylation changes throughout the genome, making them strong candidates for further analysis. CAPN3 has been associated with multiple production traits in domestic animals, including in geese ⁶⁰. ADAM20 encodes a membrane-anchored protein involved in cell-cell and cell-matrix interactions that has also been associated with fertilization in domestic animals and embryonic development ^61,62.

Gene expression has the potential to modify DNA methylation andvice-versa. Genetic variants may recruit DNA methylation or affect histone modifications^27–29, with these being potential mechanisms by which phenotypic change is effected. We see that around one third of our strongest candidate genes statistically support this pattern, with gene expression driving methylation in 36 instances, and methylation driving gene expression in 74 instances. Examples of DNA methylation changes as a consequence of gene expression are not solely restricted to the chicken, with Meng and colleagues ⁴³ finding a similar pattern of causality in Arabidopsis methQTL candidates, with 20% of their causal candidates appearing to show gene expression driving methylation. Similarly, the transcription factor PU.1 forms a complex with Dnmt3a/b to induce changes in DNA methylation around the binding sites ⁶³. Examples of gene expression changes as a cause of methylation are found in transcription factor binding sites that preferentially bind methylated DNA ²¹, whilst DNA methylation can also act to stabilize gene regulation ⁶⁴.

The combination of statistical directionality analysis with allelic imbalance identified candidate SNPs for 14 of the methylation/gene expression complexes, with 12of these having an annotated gene. Of these, six genes (DPP10, PAPSS2, KLF12, MYLK, NSUN4, DDX18) have previous links with DNA methylation as a regulator or a role in methylation regulation^65–70. Of note, NSUN4 is a rRNA methyltransferase of the mitochondrial ribosome, whilst PAPSS2 has been shown to contain a methylation QTL in a human GWAS assay⁴⁵. Further, four genes have roles in neurological diseases or cognitive phenotypes (DPP10 – autism, ADHD and brain development, NEK3 – neurodegeneration⁷¹, DDX18 – opioid susceptibility⁷², FAM196B – epilepsy susceptibility⁷³). In particular, DPP10 has an association between DNA methylation and ADHD⁷⁴, and adverse cognitive effects due to prenatal alcohol exposure⁷⁵. Structural variation at this gene is associated with autism ⁷⁶. It has also been found thathistone H3 lysin 4 trimethylated in humans, chimpanzees and macaques, potentially playing a role in cognition⁶⁵. Two of the genes were also previously implicated in the regulation of production traits in domestic animals, with MYLK in particular being differentially methylated between fast and slow-growing broiler chickens⁴². NEK3 was one of the strongest candidates in the candidate SNP analysis and plays a role in microtubule acetylation in neurons⁷⁷, modulates signaling in the prolactin receptor⁷⁸ and is also associated with neuronal degeneration⁷¹. The links with cognition also further indicates a role for DNA methylation in the regulation of behaviour, bolstering the correlations observed between DNA methylation and behaviour in this study. Such a role for methylation is also seen in other studies on birds. Temperature-related changes related to onset of seasonal breeding also alters blood DNA methylation in Great Tits ⁷⁹, whilst the gene DRD4 that is correlated with exploratory behaviour also shows differential methylation in selected lines ⁸⁰.

To date no studies have shown direct evidence of genomic variation causing epigenetic inheritance in wild birds ⁸¹. Although the study presented here is not performed on birds in a natural environment, it does show that domestication has led to a strong degree of heritable epigenetic variation via genomic regulation in the chicken. These strong inter-population differences could beindicative of a role for DNA methylation in speciation. However, an important caveat that applies to all QTL mapping, especially linkage-based QTL mapping, is that the results are most strictly applicable to the actual populations (and even just the individuals) used in the actual analysis. As such, the differences observed may be specific to the particular Red Junglefowl and White Leghorn sub-populations used, rather than all domestic birds (in the case of the White Leghorn) or wild Red Junglefowl in general (in the case of the Red Junglefowl subpopulation used). For example, in the twelve albeit highly inbred and selected Red Junglefowl birds that were selected for high and low fear of humans, no evidence of common cis methQTL effects at the loci identified in the intercross were seen between the high and low fear birds. This could indicate that the methQTL identified in the intercross presented here are the product of long-term not short-term selection, or that the populations used are sufficiently different that they have different patterns of methylation due to either recent environmental and selection pressures or long-standing genetic differences.

Other evidence for a role of DNA methylation in speciation comes from analyses with house sparrow populations, with increased global methylation seen in older populations ⁸². House sparrows also exhibit a correlation between geographic distances between populations, bill length and global DNA methylation ⁸³. Other indirect evidence for genetic variation underlying methylation variation comes from early exploratory behaviour in selected lines of Great Tits ⁸⁴. Although these birds exhibited heritable variation in exploration behaviour, no QTL were identified in two separate GWAS studies ^85,86. However, these lines do differ in DNA methylation at the DRD4 promoter ⁸⁰.

Methods

Materials & Methods

Population sample

A total of 124 adult chickens (55 females and 69 males) from an 8th generation advanced intercross were used. The parental of the cross were 3 White Leghorn females that have been maintained since the 1960s and 1 Red Junglefowl male originating from Thailand. A detailed description of rearing and housing conditions can be found elsewhere (seeJohnsson et al¹²). A seven-piece brain dissection was performed on day 212 and the tissues flash-frozen in liquid nitrogen and stored at -80°C until RNA and DNA isolation. The study was approved by the Local Ethical Committee of the Swedish National Board of Laboratory Animals.

RNA & DNA isolation

To assess gene expression levels mRNA isolation was performed on hypothalamus tissue by homogenizing the samples with Ambion TRI Reagent (Life Technologies) and following the manufacturer's protocol. A detailed description of cDNA synthesis and microarray preparations can be found elsewhere (seeJohnsson et al¹⁵). To assess DNA methylation levels DNA was isolated from the TRI Reagent-homogenate from the RNA extraction, using DNeasy Blood & Tissue Kit (Qiagen) with these additional steps prior to following the manufacturer’s protocol: 125 µl ice cold 99% ethanol was added to 250 µl TRI Reagent-homogenate; the samples were vortexed, incubated on ice for 5 min and centrifuged at 12 000 RPM for 10 min at room temperature. Past this step the manufacturer’s protocol was followed.

Methylated DNA Immunoprecipitation (MeDIP)

4 µg DNA was fragmented on a Bioruptor (Diagenode) with the following sonication settings; high for 30s and off 90s, for 4-6 cycles. Fragmentation verification was done on a 2%-agarose gel with GeneRuler Low Range ladder (Thermofisher), with a desired fragmentation size of 500 - 1000bp. To exclude any un-fragmented genomic DNA a spin-column (100 - 10000bp) from PCR Purification Kit (Qiagen) was used following the manufacturer’s protocol. The MeDIP protocol developed by Guerrero-Bosagana and colleagues⁸⁷was used with the following modifications; 10 µg monoclonal mouse anti-5-methylcytidine (Diagenode) antibody was incubated for 2h at 4°C on a rotating platform. Post MeDIP the methylated DNA was whole genome amplified using GenomPlex WGA2 Kit (Sigma-Aldrich) following the manufacturer's protocol and lastly purified with PCR Purification Kit (Qiagen). The WGA protocol used has been developed to reduce bias by using only a low number of cycles (14). In addition, a standardized starting amount of DNA was taken from all samples before proceeding with the WGA. Regions that are resistant to WGA may be missed, however this will be the same for all samples (i.e. some methQTL may be missed, but the ones that are detected will be unaffected). This protocol therefore means that only CpG methylation is measured, and not other more rare methylation motifs. It does cover both dense, less dense, and repeat regions throughout the genome. However, due to the nature of the antibody enrichment it does not give single base resolution, but is around 150bp at best and more usually around 500-800bp, leading to assessment of methylation in windows instead of by single base⁸⁸. Sequencing was performed by the National Genomics Infrastructure (NGI) in Uppsala Sweden on an IonProton machine (Thermo Fisher Scientific) using a fragment read sequencing approach and aligned using the Torrent Suite software (version 4.4.1). The sequence data had an average of 3.4X ± 0.97 coverage (based on the total genome size), 136 bp ± 15 average read length, 23.8 million ± 5.2 raw reads and 22 quality score ± 1.

Data cleanup

Because the lack of post-alignment tools working with IonTorrent Suite BAM alignment files, the initial data cleanup step was done using a custom script. We remove reads that map to multiple places in the chicken genome and mark PCR duplicates.

We considered multi-mapped reads as those with 0 MAPQ score and they are accounted for but not written to the new BAM output.

The algorithm to filter and mark PCR duplicates in Ion Torrent data handles forward and reverse strands separately. In the forward strand, first, collect all strands that start at the given position; second, after sorting by 3' position, the longest strand will not count as a duplicate; finally, we marked all shorter reads without a B adapter as duplicated in the BAM file within this strand except the one that was kept with the same 3’ length. A schematic depiction of this forward strand process is as follows:

------------------------------------------------>3'

-------------------------------------------->B3'

---------------------------------------->B3'

---------------------------------------->B3'           Dup

------------------------------------>3'          Dup

---------------------------------->3'         Dup

In the reverse strand, we build a table of reads given that in a certain position, the longer read is not the duplicate one. This table facilitates the process because only the 0-base position, marking the 3’ end of the alignment, and the CIGAR line are known from the BAM file. A schematic depiction of this reverse strand process for reads starting in the same 5’ position is as follows:

3'<-----------------------------------------

B3'<-------------------------------------

B3'<-------------------------------------           Dup

3'<-------------------------------------

After processing all the BAM files, on average, 58% percent of reads remained following the removal of multi-mapped reads and mark PCR duplicates.

Data QC

Reads with low quality map scores (q-values less than 10) were filtered out using SAMtools (version 1.3) ⁸⁹ prior to data QC. The extend, shift and window size parameters were set to 800, 0 and 100 respectively. All reads mapping to the same genomic loci were retained by setting the 'uniq' parameter to 0 because PCR duplicates had already been removed in the data cleanup step. Default values for the 'nit' and 'nrit' parameters were used, and the 'empty_bins' and 'rank' parameters were set to 'TRUE'.

DNA methylation phenotyping

In order to QTL map DNA methylation differences in the study population, the chicken reference genome was divided into 1000bp windows, starting at the first nucleotide on each chromosome adding 1000bp until the end of the chromosome. Sequence read aligning with each methylation window were counted. To normalize the phenotype values between individuals each window's count was divided by the total number of sequence read for each individual, respectively, then multiplied by 1 x10⁷ for readability purposes. The GRCg6areference genome was used, generating 1050176 windows (https://www.ncbi.nlm.nih.gov/assembly/GCF_000002315.6). Outlier removal within each window accounting all individuals was performed with the R-package 'outliers' version 0.14 using the Chi-square test. Any test with a p-value less than 1e-9 was used to reject values, no more than two values per window.

methQTL mapping

615 autosomal SNP markers were used to generate a map of length ~9668cM, average distance of ~16cM between markers. Location and genotypes of the markers can be found elsewhere (seeJohnsson et al¹⁰. DNA methylation QTL mapping, methQTL mapping, was performed with Haley-Knott regression using the R-package R/qtl⁹⁰. Cis-acting methQTL were mapped in an interval 100cM around each methylation window, expanding to a flanking marker at least 50cM both upstream and downstream. Trans-acting methQTL were mapped to all markers genome-wide. Rearing batch was included as an additive covariate and sex was included as an additive and (where significant) interactive covariate. Through permutation tests LOD-thresholds were generated using the nperm-function in R/qtl set to 1000 for 1000 randomly chosen methylation windows in 10 different seed-iterations, for sex and non-sex-interactions respectively. The 95th percentile LOD-score was used as the significant threshold and the 80th percentile was used as the suggestive threshold. methQTL peak position was determined using refineqtl-function and the confidence intervals using lodint-function (R/qtl package), with a 1.8 LOD-score drop from the peak position expanding to the closest markers. The threshold for suggestive cis-methQTL was 4.1 LOD and 5.53 LODfor standard and sex interaction, respectively, whilst the threshold for significance for cis methQTL was 4.48 LOD and 6.03 LOD (non-sex and sex-interaction). The suggestive threshold for trans methQTL was 5.6 LOD and 7.27 LOD (nonsex and sex interaction respectively), whilst the threshold for significant trans methQTL was 6.8 LOD and 9.01 LOD (nonsex and sex interaction).

methQTLcontinuous blocks and trans-hotspots

Suggestive cis methQTL originating from the same marker were named cis methQTL blocks. A minimum of four cis methQTL methylation windows had to be adjacent to each other within a distance of up to 5000bp to count as a block. Suggestive and significant trans methQTL originating from the same marker, all with an overlapping confidence interval and with a minimum of 30 methylation windows were considered to be trans methQTL hotspots. Only methQTL on the autosomes were considered. Additionally, for the trans methQTL methylation windows an epistatic QTL interaction was performed to assess if the underlying genomic region had any influence on the methylation levels. This was tested by using the closest genetic markers genotype to the methylation window and the QTL marker using the makeqtl and fitqtl functions (R/qtl package). Interactions with p-value < 0.05 were considered as suggestive, whilst p-values <0.001 were significant.

Overlap between eQTL and methQTL

From a previous study (see Johnsson et al., 2016), 1123eQTL (1064 cis, 59 trans) were identified from the same individuals and tissue used in this study. Overlapping regions between eQTL and methQTL were investigated to elucidate the relationship between DNA methylation and gene expression levels. Regions chosen from the eQTL were based on the confidence intervals of each eQTL with a 1.8 LOD-score drop from the peak position expanded to the closest markers. The intervals from the methQTL were the intervals for the 1kb methylation windows. The overlapping regions were tested for correlation (see model 1).

(1) methylation ~ gene expression + batch + sex + error

To control for multiple testing correction for these overlapsaPrincipal Component analysis was performed on all the gene expression probesets overlapping with any methylation window, 1091 in total. 16 PCs explained 1% or more of the variation, therefore the significance threshold for these initial overlaps between methQTL and eQTL phenotypes was set at 0.05/16 = 0.003. Additionally, each significant association was tested with and without DNA methylation to elucidate how much the DNA methylation affected the model (models2a and b below):

(2a) gene_expression~ methylation + eqtl_genotype + sex + batch + error

(2b) gene_expression~eqtl_genotype + sex + batch + error

Combined multiple methylation phenotypes model

As multiple methylation phenotypes were frequently correlated with an expression probeset phenotype, all such methylation predictors were combined in a single model for each of these expression phenotypes. The model tested the methylation phenotypes, with the weakest non-significant methylation phenotypes discarded one by one until only those that were significant predictors in the combined model remained. Furthermore, the significant multiple methylation models were tested with and without all methylation predictors to elucidate the total effect of DNA methylation on gene expression for that particular probeset. A likelihood ratio test was used on the two models to test if there was a significant increase in the R-squared of the full model (3a and 3b).

(3a) gene_expression ~ methylation1 + methylation2 + … + eqtl_genotype + sex + batch + error

(3b) gene_expression ~eqtl_genotype + sex + batch + error

Statistical Directionality/ Causality Assessment using NEO

All of the methylation phenotypes that were significant in the combined model were then taken forward to test for statistical causality with gene expression. This was performed using a network edge orienting (NEO) method based on the NEO R-package (R version 3.3.5) (Aten et al., 2008). This method not only assesses causality between gene expression and methylation, but also enables determination of whether methylation is causing changes in gene expression or whether gene expression is causing changes in methylation. NEO uses the genetic SNP markers from each QTL as anchors for orientating the edges for the network and integrating the traits of interest, gene expression andDNA methylation, with Structural Equation Model comparisons (SEM). The edge orientation is summarized with a Local SEM-based Edge Orienting score (LEO), comparing it to the next best model as well (NB). In the cases where both the eQTL and methQTL shared the genetic marker a LEO. NB. CPA threshold of > 0.8 was used as significant. When the QTL had separate genetic markers LEO. NB. OCA score of > 0.3 was used as significant. Direction of causality was tested as follows:

(4) eQTL marker -> gene expression ->DNA methylation <- methQTL marker

(5) eQTL marker -> gene expression <- DNA methylation <- methQTL marker

(6) eQTL marker -> gene expression ->DNA methylation

(7) gene expression <- DNA methylation <- methQTL marker

(8) gene expression <- shared QTL marker ->DNA methylation

(9) gene expression ->DNA methylation <- methQTL marker

(10) eQTL marker -> gene expression <- DNA methylation

Allelic imbalance/ Epi-allelic Imbalance analysis

Allelic imbalance (AI) occurs when regulatory elements affect gene expression where alleles are expressed in different ratios in the same individual. In RNA sequencing it is possible to detect AI in heterozygote regions. As the MeDIP-seq protocol bind to methylated cytosine in a CpG it should be possible to detect epiallelic imbalance when comparing methylated regions in heterozygote state, i.e. a heterozygote locus should show a 50:50 ratio if the DNA methylation levels are the same, any deviation from that ratio would indicate an epiallelic imbalance.

From the NEO-analysis any cis methQTL phenotype (1kb methylation window) was used to test for epiallelic imbalance. The genotype for the regions was assessed in two ways; (1) from the meQTL peak QTL position - imputed with the sim.geno function (R/qtl R-package) with step = 1 and error.prob = 10e-9, and (2) using the bcftools call -c function. Polymorphic positions (SNPs) and the allele counts within the methylation windows were located using the function scanForHeterozygotes(AllelicImbalance R-package). A multiple linear regression was performed on the observed genotypes from both the genotype calculated from both the vcf calling and the QTL peak, sex and batch as predictors and the observed allele count as response variable (i.e. two regressions were performed for each SNP, one using the QTL-peak genotype, the other using the genotypes called from the SNP itself). Regression with p-values < 0.05 were kept as candidates. From these candidates the heterozygote genotypes were used to calculate an epiallelic imbalance with a chi-square test (chi-square p-value < 0.05), with heterozygotes again called from both the QTL-peak and the SNP itself.

Ascertainment bias can be an issue with sequence data whereby SNPs are more likely to be called when they are the same as the genomic reference allele rather than the alternate genomic allele, with this being a particular problem with low coverage SNPs. To test for this, the SNP with the strongest allelic imbalance from each of the 14 discrete methylation windows containing one or more SNPs in allelic imbalance were tested. Of these 14 SNPs, five had the higher frequency allele derived from the alternate allele(chr1.119143447, chr2.99945095, chr6.22860234, chr7.27344378, chr7.28598812), so ascertainment bias was not an issue for these five SNPs. Of the remaining nine SNPs, these were tested for allelic imbalance both with all individuals and also with individuals that only had at least seven reads, to remove the issue of low coverage individuals biasing the estimate. In all cases, the SNPs without low coverage individuals still displayed significant allelic imbalance, implying that ascertainment bias is not an issue with the SNP data (see Supplementary Table 11).

DNA methylation levels ±2Mb up- and downstream of each gene

Initially, an assessment of gene expression with methylation was performed for all genes that were 4kb or larger and more than 10kb away from the start or end of a chromosome (Figure 3A). All applicable genes were subdivided into high expression genes (1st quartile) and low expression genes (4th quartile), with the average methylation for each of these two classes plotted from 10kb upstream of the TSS and from the TTS to 10kb downstream (n=5201 genes). To further assess the relationship between gene expression and methylation, we tested for correlations between DNA methylation levels and gene expression levels genome-wide on a gene-by-gene basis using a general linear model, with the DNA methylation levels as predictor and the probeset expression levels as the response variable. Regions tested were divided into 2Mb upstream, 2Mb downstream and within gene. These regions were further divided into 1kb methylation windows. The analysis was performed forall available autosomes. The most significant correlation was retained for each expression probeset, with a total of 29740 probesets tested. As some genes are represented by multiple probesets, one gene could be composed of both positive and negative correlations with methylation in certain instances. A permutation tests was performed to generate a significant p-value threshold by selecting 1000 probesets randomly and shuffling their expression values before running the regression test retaining the top p-value. This was done 500 times. The 95th percentile was used as the significant threshold, generating a p-value < 6.9e-4.

Behavioural Correlations with Methylation

Several behavioural assays relating to anxiety were performed on the individuals used in this study, with these being open field behaviour (degree of movement when in a novel arena), social reinstatement behaviour (time spent associating with conspecifics in a novel arena), and tonic immobility behaviour (a form of freezing behaviour). Further details of the assays can be found in ^13–15. A fourth behavioural assay was based on the previous three, and measured the degree of within-individual variation for each of the measurements (termed predictability, though more strictly within-individual variation), as each test was performed twice for each individual³⁸. These behaviours have previously been mapped for QTL, with these QTL then overlapped with eQTL, followed by correlating gene expression and behaviour andto identify a total of 25 candidate genes that putatively different aspects of these behaviours. These genes and the behavioural traits that they are correlated with are shown in Supplementary Table 8. Any of these genes that were also identified as being correlated with methylation were therefore tested for potential effects between the methylation phenotype that correlated with gene expression and the behavioural measure that correlated with gene expression. The following three linear models were used to assess the relationship between behaviour and methylation:

(11) behaviour ~ methylation + gene expression + sex + batch + error

(12) behaviour ~methylation+ sex + batch + error

(13) behaviour ~gene expression + sex + batch + error

In addition, a NEO causal analysis was also performed using the five genes that possessed one or more significant correlations. In this case the same procedure as outlined above was performed, but with the behavioural phenotype in place of the gene expression phenotype.

Comparison of Cis methQTL with Red Junglefowl Selection Lines

MeDip-seq data was available from Belteky and colleagues ¹⁶, consisting of RJF (n=12) selected for either high fear or low fear of humans for five generations. Genome-wide methylation data for 6 high and 6 low selected individuals was taken from an online repository ¹⁶. The locations of suggestive and significant cis methQTL from the 200k most variable methylated regions identified in the wild x domestic intercross(n= 1073) were then lifted from the selected Red Junglefowl data. For each 1kb window, a t-test was used to identify significant differences between high fear and low fear selected Red Junglefowl individuals. Suggestive significance was assumed with a nominal p<0.05, whilst an experiment-wide p-value (incorporating a Bonferroni multiple testing correction) occurred with p <4.7 x 10^-5.

Conclusion

In conclusion, we detect large andwide-spread genetic influences on DNA methylation in our wild x domestic inter-cross, indicating an important role for methylation in chicken domestication. Our population-level analysis of DNA methylation reveals a rich landscape of regulatory effects in the chicken, where DNA methylation is not just suppressing expression at the transcription start site, but also positively associated with expression through other regulatory elements. In this fashion, DNA methylation is highly influential in regulating the variation in gene expression present in this intercross. The regulation of DNA methylation variation under domestication in the chicken is genetically complex, but key signatures of domestication are observed and are functionally important in the methylome of the chicken. There are large trans-acting hotspots that affect DNA methylation of many regions in concert, whilst relatively long continuous methylated regions are also de-methylated during domestication in the chicken. Furthermore, DNA methylation appears to be the driver of expression at some loci, but also responsive to gene expression at other loci. Brain DNA methylation in the chicken is thus both complex and rife for further exploration. To our knowledge, this is the first evidence of the extent polymorphisms affect DNA methylation in domestication, and of theimportant role DNA methylationplays as a mechanism of transcriptional modulationin domestication.

Extended Data

Extended Data Fig. 1. Percentage of variation in gene expression explained by methylation and genotype in the significantly correlated methQTL/ eQTL candidates.

Percentage of variation in gene expression explained by methylation and genotype in the significantly correlated methQTL/ eQTL candidates. Genes possessing an eQTL are subdivided into the total % of variation that can be accounted for in a model including eQTL genotype, sex and methylation from any significant methQTL (i.e. genotype, sex and methylation together explain 10-20% of variation in gene expression in the smallest category, and over 50% of gene expression variation in the largest category). The explained gene expression variation in each category is then separated into the % of variation explained by the eQTL genotype and sex and the % variation explained by methylation. In the case of the category for >50% variation explained in gene expression, only four methQTL/ eQTL models fell into this bracket.

Data Availability

Microarray data for the chicken hypothalamus tissue are available at E-MTAB-3154 in ArrayExpress. DNA methylation and behavioural phenotypes (https://doi.org/10.6084/m9.figshare.12803873), genotypes (https://doi.org/10.6084/m9.figshare.12803876) and a readymade QTL cross-file (https://doi.org/10.6084/m9.figshare.12803870) are available via Figshare